Datasets:

math-ai

/

AutoMathText

like 148

Follow

math-ai 19

https://arxiv.org/abs/1311.6238

Exact post-selection inference, with application to the lasso

We develop a general approach to valid inference after model selection. At the core of our framework is a result that characterizes the distribution of a post-selection estimator conditioned on the selection event. We specialize the approach to model selection by the lasso to form valid confidence intervals for the selected coefficients and test whether all relevant variables have been included in the model.

\section{Introduction} \label{sec:intro} As a statistical technique, linear regression is both simple and powerful. Not only does it provide estimates of the ``effect'' of each variable, but it also quantifies the uncertainty in those estimates, allowing inferences to be made about the effects. However, in many applications, a practitioner starts with a large pool of candidate variables, such as genes or demographic features, and does not know \emph{a priori} which are relevant. This is especially problematic when there are more variables than observations; in the usual regression setting where predictors are assumed to be fixed, the model is unidentifiable. In such settings, it is tempting to use the data to decide which variables to include in the model. For example, one common approach when the number of variables is not too large is to fit a linear model with all variables included, observe which ones are significant at level $\alpha$, and then refit the linear model with only those variables included. The problem with this is that the $p$-values can no longer be trusted, since the variables that are selected will tend to be those that are significant. Intuitively, we are ``overfitting'' to a particular realization of the data. To formalize the problem, consider the standard linear regression setup, where the response ${\bm y} \in\mathbb{R}^n$ is generated \begin{equation} {\bm y} \sim N({\bm\mu}, \sigma^2I_n) \label{eq:model} \end{equation} and ${\bm\mu}$ is modeled as a linear function of predictors ${\bm x}_1, ..., {\bm x}_p \in \mathbb{R}^n$, and $\sigma^2$ is assumed known. (We consider the more realistic case where $\sigma^2$ is unknown in Section~\ref{sec:estimate:sigma}.) We choose a subset $M \subset \{1, ..., p\}$ and ask for the linear combination of the predictors in $M$ that minimizes the expected error, i.e., \begin{equation} {\bm\beta}^M \equiv \argmin_{{\bm b}^M}\ \mathbb{E} ||{\bm y} - X_M {\bm b}^M ||^2 = X_M^\dagger {\bm\mu}, \label{eq:target} \end{equation} where $X_M^\dagger \equiv (X_M^T X_M)^{-1} X_M^T$ is the pseudo-inverse of $X_M$. Notice that \eqref{eq:target} implies that the targets $\beta^M_j$ and $\beta^{M'}_j$ in different models $M \neq M'$ are in general different. This is simply a restatement of the well-known fact that a regression coefficient describes the effect of a predictor, {\em adjusting for the other predictors in the model}. In general, the coefficient of a predictor cannot be compared across different models. Thus, ``inference after selection'' is ambiguous in linear regression because the target of inference changes with the selected model \citep{berk2013posi}. In the next section, we discuss several ways to resolve this ambiguity. \section{Post-Selection Inference in Linear Regression} \label{sec:goals} At first blush, the fact that the target ${\bm\beta}^M$ changes with the model is deeply troubling, since it seems to imply that the parameters are random. However, the randomness is actually in the {\em choice} of which parameters to consider, not in the parameters themselves. Imagine that there are {\em a priori} $p2^{p-1}$ well-defined population parameters, one for each coefficient in all $2^p$ possible models: \[ \{ \beta^M_j: M \subset \{1, ..., p\}, j \in M \}. \] We only ever form inferences for the parameters $\beta^{\hat M}_j$ in the model $\hat M$ we select. This adaptive choice of which parameters to consider can lead to inferences with undesirable frequency properties, as noted by \citet{benjamini2005false} and \citet{benjamini2009selective}. To be concrete, suppose we want a confidence interval $C^{\hat M}_j$ for a parameter $\beta^{\hat M}_j$. What frequency properties should $C^{\hat M}_j$ have? By analogy to the classical setting, we might ask for $$\mathbb{P}(\beta^{\hat M}_j \in C^{\hat M}_j) \geq 1-\alpha,$$ but this criterion is ill-posed because $\beta^M_j$ is undefined when $j \notin M$. Two ways around this issue are suggested by \citet{berk2013posi}: \begin{enumerate} \item Conditional Coverage: Since we form an interval for $\beta^M_j$ if and only if model $M$ is selected, i.e., $\hat M = M$, it makes sense to condition on this event. Hence, we might require that our confidence interval $C^M_j$ satisfy \begin{equation} \mathbb{P}(\beta^M_j \in C^M_j\ |\ \hat M = M) \geq 1-\alpha. \label{eq:conditional_coverage} \end{equation} The benefit of this approach is that we avoid ever having to compare coefficients across two different models $M \neq M'$. Another way to understand conditioning on the model is to consider {\em data splitting} \citep{cox1975note}, an approach to post-selection inference that most statisticians would agree is valid. In data splitting, the data is divided into two halves, with one half used to select the model and the other used to conduct inference. \citet{fithian2014optimal} shows that if one accepts data splitting as valid, then one is led naturally to consider frequency properties conditional on the model, such as conditional coverage and selective type I error. \item Simultaneous Coverage: It also makes sense to talk about events that are defined simultaneously over all $j \in \hat M$. \citet{berk2013posi} propose controlling the familywise error rate \begin{equation} FWER \equiv \mathbb{P}(\beta^{\hat M}_j \notin C^{\hat M}_j\ \text{for any $j \in \hat M$}), \label{eq:fwer} \end{equation} but this is very stringent when many predictors are involved. Instead of controlling the probability of making {\em any} error, we can control the expected proportion of errors---although ``proportion of errors'' is ambiguous in the edge case that we select zero variables. \citet{benjamini2005false} simply declare the error to be zero when $|\hat M| = 0$: \begin{equation} FCR \equiv \mathbb{E}\left[ \frac{\left|\{ j \in \hat M: \beta^{\hat M}_j \notin C^{\hat M}_j \}\right|}{ |\hat M | }; |\hat M| > 0 \right], \label{eq:fcr} \end{equation} while \citet{storey2003positive} suggests conditioning on $|\hat M| > 0$: \begin{equation} pFCR \equiv \mathbb{E}\left[ \left. \frac{\left|\{ j \in \hat M: \beta^{\hat M}_j \notin C^{\hat M}_j \}\right|}{ |\hat M | } \right| |\hat M| > 0 \right]. \label{eq:pfcr} \end{equation} The two criteria are closely related. Since $FCR = pFCR \cdot \mathbb{P}(|\hat M| > 0)$, $pFCR$ control implies $FCR$ control. \end{enumerate} The two ways above are related: conditional coverage \eqref{eq:conditional_coverage} implies pFCR \eqref{eq:pfcr} (and hence, FCR) control. \begin{lemma} Consider a family of intervals $\{ C^{\hat M}_j \}_{j\in\hat M}$ that each have conditional $(1-\alpha)$ coverage: \[ \mathbb{P}(\beta^{\hat M}_j \notin C^{\hat M}_j | \hat M = M ) \leq \alpha,\ \text{for all $M$ and $j \in M$}. \] Then, $FCR \leq pFCR \leq \alpha$. \begin{proof} Condition on $\hat M$ and iterate expectations. \begin{align*} pFCR &= \mathbb{E}\left[\left.\mathbb{E}\left[ \left. \frac{\left|\{ j \in \hat M: \beta^{\hat M}_j \notin C^{\hat M}_j \}\right|}{|\hat M|} \right| \hat M \right] \ \right|\ |\hat M| > 0 \right] \\ &= \mathbb{E}\left[\left. \frac{\sum_{j\in\hat M} \mathbb{P}\left( \beta^{\hat M}_j \notin C^{\hat M}_j \big|\hat M\right)}{|\hat M|}\ \right|\ |\hat M| > 0 \right] \\ &\leq \mathbb{E}\left[\left.\frac{\alpha |\hat M|}{|\hat M|}\ \right|\ |\hat M| > 0\right] \\ &= \alpha. \end{align*} \end{proof} \label{lem:fcr} \end{lemma} Theorem 2 in \cite{selection:benjamini} proves a special case of Lemma~\ref{lem:fcr} for a particular selection procedure, and Proposition 11 in \citet{fithian2014optimal} provides a more general result, but this result is sufficient for our purposes: to establish that conditional coverage implies other important properties. Although the criterion is easy to state, how do we construct an interval with conditional coverage? This requires that we understand the conditional distribution $$ {\bm y}\ |\ \{\hat{M}({\bm y}) = M\}, \qquad {\bm y} \sim N(\mu, \sigma^2 I). $$ One of the main contributions of this paper is to show that this distribution is indeed possible to characterize, and valid inference conditional on selection is computationally possible (and indeed simple) in the context of model selection in linear regression. \section{Outline of Our Approach} We have argued that post-selection intervals for regression coefficients should have $1-\alpha$ coverage conditional on the selected model: \[ \mathbb{P}(\beta^M_j \in C^M_j\ |\ \hat M = M) \geq 1-\alpha, \] both because this criterion is interesting in its own right and because it implies FCR control. To obtain an interval with this property, we study the conditional distribution \begin{equation} {\bm\eta}_M^T {\bm y}\ |\ \{\hat M = M\}, \label{eq:cond_dist} \end{equation} which will allow, more generally, conditional inference for parameters of the form ${\bm\eta}_M^T {\bm\mu}$. In particular, the regression coefficients $\beta^M_j = {\bm e}_j^T X_M^\dagger {\bm\mu}$ can be written in this form, as can many other linear contrasts. Our paper focuses on the specific case where the lasso is used to select the model $\hat M$. We begin in Section \ref{sec:lasso-selection} by characterizing the event $\{ \hat M = M \}$ for the lasso. As it turns out, this event is a union of polyhedra. More precisely, the event $\{ \hat M = M, \hat{\bm s}_M = {\bm s}_M \}$, that specifies the model {\em and} the signs of the selected variables, is a polyhedron of the form \[ \{{\bm y} \in \mathbb{R}^n: A(M, {\bm s}_M) {\bm y} \leq {\bm b}(M, {\bm s}_M) \}. \] Therefore, if we condition on both the model and the signs, then we only need to study \begin{equation} {\bm\eta}^T {\bm y}\ |\ \{A{\bm y} \leq {\bm b}\}. \end{equation} We do this in Section \ref{sec:polyhedra}. It turns out that this conditional distribution is essentially a (univariate) truncated Gaussian. We use this to derive a statistic $F^{\bm z}({\bm\eta}^T {\bm y})$ whose distribution given $\{A{\bm y} \leq {\bm b}\}$ is $\unif(0, 1)$. \subsection{Related Work} The resulting post-selection test has a similar structure to the pathwise significance tests of \cite{lockhart2012significance} and \cite{taylor2014post}, which also are conditional tests. However, the intended application of our test is different. While their significance tests are specifically intended for the path context, our framework allows more general questions about the model the lasso selects: we can test the model at any value of $\lambda$ or form confidence intervals for an individual coefficient in the model. There is also a parallel literature on confidence intervals for coefficients in high-dimensional linear models based on the lasso estimator \citep{van2013asymptotically,zhang2011confidence,javanmard2013confidence}. The difference between their work and ours is that they do not address post-selection inference; their target is ${\bm\beta}^0$, the coefficients in the true model, rather than ${\bm\beta}^{\hat M}$, the coefficients in the selected model. The two will not be the same unless $\hat M$ happens to contain all non-zero coefficients of ${\bm\beta}^0$. Although inference for ${\bm\beta}^0$ is appealing, it requires assumptions about correctness of the linear model and sparsity of ${\bm\beta}^0$. \cite{potscher2010confidence} consider confidence intervals for the hard-thresholding and soft-thresholding estimators in the case of orthogonal design. Our approach instead regards the selected model as a linear approximation to the truth, a view shared by \citet{berk2013posi} and \citet{miller2002subset}. The idea of post-selection inference conditional on the selected model appears in \citet{potscher1991effects}, although the notion of inference conditional on certain \emph{relevant subsets} dates back to \citet{fisher1956test}; see also \citet{robinson1979conditional}. \citet{leeb2005model,leeb2006can} obtained a number of negative results about estimating the distribution of a post-selection estimator, although they note their results do not necessarily preclude the possibility of post-selection inference. The problem was most recently considered by \cite{berk2013posi}, who cast post-selection inference in terms of simultaneous inference over all possible submodels. \citet{benjamini2005false} also consider conditioning on the selection event, although they argue that this is too conservative. To the contrary, we show that conditioning on the selected model can produce reasonable confidence intervals in a wide variety of situations. Inference conditional on selection has also appeared in literature on the {\em winner's curse}: \cite{sampson2005drop,sill2009drop,zhong2008bias,zollner2007overcoming}. These works are not really associated with model selection in linear regression, though they employ a similar approach to inference. \section{The Lasso and Its Selection Event} \label{sec:lasso-selection} In this paper, we apply our post-selection inference procedure to the model selected by the lasso \citep{tibshirani:lasso}. The lasso estimate is the solution to the usual least squares problem with an additional $\ell_1$ penalty on the coefficients: \begin{equation} \label{eq:lasso} \hat{\bm\beta} \in \argmin_{{\bm\beta}} \, \frac{1}{2} \|{\bm y}-X{\bm\beta}\|^2_2+ \lambda \|{\bm\beta}\|_1. \end{equation} The $\ell_1$ penalty shrinks many of the coefficients to exactly zero, and the tradeoff between sparsity and fit to the data is controlled by the penalty parameter $\lambda \geq 0$. However, the distribution of the lasso estimator $\hat{\beta}$ is known only in the less interesting $n \gg p$ case \citet{knight2000lasso}, and even then, only asymptotically. Inference based on the lasso estimator is still an open question. Because the lasso produces sparse solutions, we can define model ``selected'' by the lasso to be simply the set of predictors with non-zero coefficients: \[ \hat M = \{ j: \hat\beta_j \neq 0 \}. \] Then, post-selection inference seeks to make inferences about ${\bm\beta}^M$, given $\{\hat M = M\}$, as defined in \eqref{eq:target}. The rest of this section focuses on characterizing this event $\{ \hat M = M \}$. We begin by noting that in order for a vector of coefficients $\hat{\bm\beta}$ and a vector of signs $\hat{\bm s}$ to be solutions to the lasso problem \eqref{eq:lasso}, it is necessary and sufficient that they satisfy the Karush-Kuhn-Tucker (KKT) conditions: \begin{gather} X^T(X\hat{\bm\beta} - {\bm y}) + \lambda\hat{\bm s} = 0, \label{eq:lasso-kkt} \\ \hat s_i = \sign(\hat\beta_j) \ \ \text{if $\hat\beta_j \neq 0$} \nonumber\\ \hat s_i \in [-1, 1] \ \ \ \ \ \text{if $\hat\beta_j = 0$} \end{gather} Following \citet{tibshirani2013lasso}, we consider the {\em equicorrelation set} \begin{equation} \label{eq:equicor} \hat M \equiv \left\{i\in\{1,\dots,p\} : |\hat s_i| = 1 \right\}. \end{equation} Notice that we have implicitly defined the model $\hat M$ to be equicorrelation set. Since $|\hat s_i| = 1$ for any $\hat\beta_i \neq 0$, the equicorrelation set does in fact contain all predictors with non-zero coefficients, although it may also include some predictors with zero coefficients. However, for almost every $\lambda$, the equicorrelation set is precisely the set of predictors with non-zero coefficients. It turns out that it is easier to first characterize $\{ (\hatM, \hat {\bm s}) = (M, {\bm s}) \}$ and obtain $\{ \hatM = M \}$ as a corollary by taking a union over the possible signs. The next result is an important first step. \begin{lemma} \label{lem:equiv_sets} Assume the columns of $X$ are in general position \citep{tibshirani2013lasso}. Let $M \subset \{1, \dots, p\}$ and ${\bm s} \in \{-1,1\}^{|M|}$ be a candidate set of variables and their signs, respectively. Define the random variables \begin{align} {\bm w}(M, {\bm s}) &:= (X_M^T X_M)^{-1} (X_M^T {\bm y} - \lambda {\bm s}) \label{eq:beta-active} \\ {\bm u}(M, {\bm s}) &:= X_{-M}^T(X_M^T)^\dagger {\bm s} + \frac{1}{\lambda} X_{-M}^T (I - P_{M}) {\bm y} \label{eq:inactive_subgradient}. \end{align} where $P_M \equiv X_M(X_M^T X_M)^{-1} X_M$ is projection onto the column span of $X_M$. Then the selection procedure can be rewritten in terms of ${\bm w}$ and ${\bm u}$ as: \begin{align} \left\{ (\hatM, \hat {\bm s}) = (M, {\bm s}) \right\} = \left\{ \sign({\bm w}(M, {\bm s})) = {\bm s},\phantom{\Big|} ||{\bm u}(M, {\bm s})||_\infty < 1 \right\} \label{eq:equiv_sets} \end{align} \begin{proof} First, we rewrite the KKT conditions \eqref{eq:lasso-kkt} by partitioning them according to the equicorrelation set $\hatM$, adopting the convention that $-\hatM$ means ``variables not in $\hatM$.'' \begin{align*} X_{\hatM}^T (X_{\hatM}\hat{{\bm\beta}}_{\hatM} - {\bm y}) + \lambda\hat {\bm s}_{\hatM} &= 0 \\ X_{-\hatM}^T (X_{\hatM}\hat{{\bm\beta}}_{\hatM} - {\bm y}) + \lambda\hat {\bm s}_{-\hatM} &= 0 \\\sign(\hat{\bm\beta}_{\hatM}) &= \hat {\bm s}_{\hat M} \\ ||\hat{\bm s}_{-\hatM}||_\infty &< 1. \end{align*} Since the KKT conditions are necessary and sufficient for a solution, we obtain that ${\{ (\hatM, \hat{\bm s}) = (M,{\bm s}) \}}$ if and only if there exist ${\bm w}$ and ${\bm u}$ satisfying: \begin{align*} X_M^T (X_M {\bm w} - {\bm y}) + \lambda{\bm s} &= 0\\ X_{-M}^T (X_M {\bm w} - {\bm y}) + \lambda {\bm u} &= 0\\ \sign({\bm w}) &= {\bm s} \\ ||{\bm u}||_\infty &< 1. \end{align*} We can solve the first two equations for ${\bm w}$ and ${\bm u}$ to obtain the equivalent set of conditions \begin{gather*} {\bm w} = (X_M^T X_M)^{-1} (X_M^T {\bm y} - \lambda {\bm s}) \\ {\bm u} = X_{-M}^T(X_M^T)^\dagger {\bm s} + \frac{1}{\lambda} X_{-M}^T (I - P_{M}) {\bm y} \\ \sign({\bm w}) = {\bm s}\\ ||{\bm u}||_\infty < 1, \end{gather*} where the first two are the definitions of ${\bm w}$ and ${\bm u}$ given in \eqref{eq:beta-active} and \eqref{eq:inactive_subgradient}, and the last two are the conditions on ${\bm w}$ and ${\bm u}$ given in \eqref{eq:equiv_sets}. \end{proof} \end{lemma} Lemma \ref{lem:equiv_sets} is remarkable because it says that the event ${\{ (\hatM, \hat{\bm s}) = (M, {\bm s}) \}}$ can be rewritten as affine constraints on ${\bm y}$. This is because ${\bm w}$ and ${\bm u}$ are already affine functions of ${\bm y}$, and the constraints $\sign(\cdot) = {\bm s}$ and $||\cdot ||_\infty < 1$ can also be rewritten in terms of affine constraints. The following proposition makes this explicit. \begin{proposition} \label{prop:A_b} Let ${\bm w}$ and ${\bm u}$ be defined as in \eqref{eq:beta-active} and \eqref{eq:inactive_subgradient}. Then: \begin{align} \left\{ \sign({\bm w}) = {\bm s} , \norm{{\bm u}}_\infty < 1\right \} &= \left\{ \begin{pmatrix} A_0(M, {\bm s}) \\ A_1(M, {\bm s}) \end{pmatrix} {\bm y} < \begin{pmatrix} {\bm b}_0(M, {\bm s}) \\ {\bm b}_1(M, {\bm s}) \end{pmatrix} \right\} \label{eq:polyhedron} \end{align} where $A_0, {\bm b}_0$ encode the ``inactive'' constraints $\{\norm{{\bm u}}_\infty < 1\}$, and $A_1, {\bm b}_1$ encode the ``active'' constraints $\{\sign({\bm w}) = {\bm s}\}$. These matrices have the explicit forms: \begin{align*} A_0(M, {\bm s}) &= \frac{1}{\lambda}\begin{pmatrix}X_{-M}^T (I - P_M) \\- X_{-M}^T (I - P_M) \end{pmatrix} & {\bm b}_0(M, {\bm s}) &= \begin{pmatrix}\mathbf 1 - X_{-M}^T (X_M^T)^\dagger {\bm s} \\ \mathbf 1 + X_{-M}^T (X_M^T)^\dagger {\bm s} \end{pmatrix} \\ A_1(M, {\bm s}) &= -\diag({\bm s}) (X_M^T X_M)^{-1} X_M^T & {\bm b}_1(M, s) &= -\lambda \diag({\bm s})(X_M^T X_M)^{-1} {\bm s} \end{align*} \end{proposition} \begin{proof} First, substituting expression \eqref{eq:beta-active} for ${\bm w}$, we rewrite the ``active'' constraints as \begin{align*} \{ \sign({\bm w}) = {\bm s} \} &= \{ \diag({\bm s}) {\bm w} > 0 \} \\ &= \{ \diag({\bm s}) (X_M^T X_M)^{-1}(X_M^T {\bm y} - \lambda {\bm s}) > 0 \} \\ &= \{ A_1(M, {\bm s}) {\bm y} < {\bm b}_1(M, {\bm s}) \}. \end{align*} Next, substituting expression \eqref{eq:inactive_subgradient} for ${\bm u}$, we rewrite the ``inactive'' constraints as \begin{align*} \{ ||{\bm u}||_\infty < 1 \} &= \left\{ -\mathbf 1 < X_{-M}^T(X_M^T)^\dagger {\bm s} + \frac{1}{\lambda} X_{-M}^T (I - P_{M}) {\bm y} < \mathbf 1 \right\} \\ &= \{ A_0(M, {\bm s}) {\bm y} < {\bm b}_0(M, {\bm s}) \} \end{align*} \end{proof} Combining Lemma \ref{lem:equiv_sets} with Proposition \ref{prop:A_b}, we obtain the following. \begin{theorem} Let $A(M, {\bm s}) = \begin{pmatrix} A_0(M, {\bm s}) \\ A_1(M, {\bm s}) \end{pmatrix}$ and $b(M, {\bm s}) = \begin{pmatrix} {\bm b}_0(M, {\bm s}) \\ {\bm b}_1(M, {\bm s}) \end{pmatrix}$, where $A_i$ and $b_i$ are defined in Proposition \ref{prop:A_b}. Then: $$\displaystyle\{ \hatM = M, \hat{\bm s} = {\bm s} \} = \{ A(M, {\bm s}){\bm y} \leq {\bm b}(M, {\bm s}) \}.$$ \label{thm:lasso_partition} \end{theorem} As a corollary, $\{ \hatM = M \}$ is simply the union of the above events over all possible sign patterns. \begin{corollary} $\displaystyle\{ \hatM = M \} = \bigcup_{{\bm s} \in \{-1, 1\}^{|M|}}\{ A(M, {\bm s}){\bm y} \leq {\bm b}(M, {\bm s}) \}.$ \label{cor:lasso_partition} \end{corollary} Figure \ref{fig:lasso_partition} illustrates Theorem~\ref{thm:lasso_partition} and Corollary~\ref{cor:lasso_partition}. The lasso partitions of $\mathbb{R}^n$ into polyhedra according to the model it selects and the signs of the coefficients. The shaded area corresponds to the event $\{\hat M = \{1, 3\}\}$, which is a union of two polyhedra. Notice that the sign patterns $\{+, -\}$ and $\{-, +\}$ are not possible for the model $\{1, 3\}$. \begin{figure} \begin{tikzpicture}[scale=1.2] \defgray!70{gray!70} \defblue!20{blue!20} \node at (-1.5,-1.5) {}; \node[gray!70] (x1) at (.6,.6) {${\bf x}_1$}; \node[gray!70] (x2) at (0,.8) {${\bf x}_3$}; \node[gray!70] (x3) at (-.8,0) {${\bf x}_2$}; \draw[gray!70,->] (0,0) -- (x1); \draw[gray!70,->] (0,0) -- (x2); \draw[gray!70,->] (0,0) -- (x3); \draw[gray!70] (-1,1) -- (.4,1) -- (1,.4) -- (1,-1) -- (-.4,-1) -- (-1,-.4) -- cycle; \draw[fill=blue!20,blue!20] (.4,1) -- (.4,3) -- (2.4,3) -- cycle; \draw[fill=blue!20,blue!20] (-.4,-1) -- (-.4,-3) -- (-2.4,-3) -- cycle; \draw[gray!70] (.4,1) -- (.4,3); \draw[gray!70] (.4,1) -- (2.4,3); \draw[gray!70] (1,.4) -- (2.4,1.8); \draw[gray!70] (1,.4) -- (2.4,.4); \draw[gray!70] (1,-1) -- (2.4,-1); \draw[gray!70] (1,-1) -- (1,-3); \draw[gray!70] (-.4,-1) -- (-.4,-3); \draw[gray!70] (-.4,-1) -- (-2.4,-3); \draw[gray!70] (-1,-.4) -- (-2.4,-1.8); \draw[gray!70] (-1,-.4) -- (-2.4,-.4); \draw[gray!70] (-2.4,1) -- (-1,1) -- (-1,3); \node[blue!80] at (1.1,2.5) {\parbox{1in}{\begin{align*}\hat M = \{1, 3\}\\ \hat{\bm s} = \{ +, + \}\end{align*}}}; \node[blue!80] at (-1.1,-2.4) {\parbox{1in}{\begin{align*}\hat M = \{1, 3\} & \\ \hat{\bm s} = \{ -, - \} & \end{align*}}}; \end{tikzpicture} \caption{A geometric picture illustrating Theorem \ref{thm:lasso_partition} for $n=2$ and $p = 3$. The lasso partitions $\mathbb{R}^n$ into polyhedra according to the selected model and signs.} \label{fig:lasso_partition} \end{figure} \section{Polyhedral Conditioning Sets} \label{sec:polyhedra} In order to obtain inference conditional on the model, we need to understand the distribution of \[ {\bm\eta}_M^T {\bm y}\ |\ \{ \hatM = M \}. \] However, as we saw in the previous section, $\{ \hatM = M \}$ is a union of polyhedra, so it is easier to condition on both the model {\em and the signs}, \[ {\bm\eta}_M^T {\bm y}\ |\ \{ \hatM = M, \hat{\bm s} = {\bm s} \}, \] where the conditioning event is a single polyhedron $\{ A(M, {\bm s}){\bm y} \leq {\bm b}(M, {\bm s})\}$. Notice that inferences that are valid conditional on this finer event will also be valid conditional on $\{ \hatM = M \}$. For example, if a confidence interval $C^M_j$ for $\beta^M_j$ has $(1-\alpha)$ coverage conditional on the model and signs \[ \mathbb{P}(\beta^M_j \in C^M_j\ |\ \hatM = M, \hat{\bm s} = {\bm s}) \geq 1-\alpha, \] then it will also have $(1-\alpha)$ coverage conditional only on the model: \begin{align*} \mathbb{P}(\beta^M_j \in C^M_j\ |\ \hatM = M) &= \sum_{{\bm s}} \mathbb{P}(\beta^M_j \in C^M_j\ |\ \hatM = M, \hat{\bm s} = {\bm s}) \mathbb{P}(\hat{\bm s} = {\bm s}\ |\ \hatM = M) \\ &\geq \sum_{{\bm s}} (1-\alpha) \mathbb{P}(\hat{\bm s} = {\bm s}\ |\ \hatM = M) \\ &= 1-\alpha. \end{align*} This section is divided into two subsections. First, we study how to condition on a single polyhedron; this will allow us to condition on $\{ \hatM = M, \hat{\bm s} = {\bm s} \}$. Then, we look at how to extend the framework to condition on a union of polyhedra, which will allow us to condition only on the model $\{ \hatM = M \}$. The inferences obtained by conditioning on the model will in general be more efficient (i.e., narrower intervals, more powerful tests), at the price of more computation. \subsection{Conditioning on a Single Polyhedron} Suppose we observe ${\bm y} \sim N({\bm\mu}, \Sigma)$, and ${\bm\eta} \in \mathbb{R}^n$ is some direction of interest. To understand the distribution of \begin{equation} {\bm\eta}^T {\bm y}\ |\ \{ A{\bm y} \leq {\bm b} \}, \end{equation} we rewrite $\{ A{\bm y} \leq {\bm b} \}$ in terms of ${\bm\eta}^T {\bm y}$ and a component ${\bm z}$ which is independent of ${\bm\eta}^T {\bm y}$. That component is \begin{equation} {\bm z} \equiv (I_n - {\bm c}{\bm\eta}^T){\bm y}, \label{eq:z} \end{equation} where \begin{equation} {\bm c} \equiv \Sigma{\bm\eta} ({\bm\eta}^T \Sigma {\bm\eta})^{-1}. \label{eq:c} \end{equation} It is easy to verify that ${\bm z}$ is uncorrelated with, and hence independent of, ${\bm\eta}^T {\bm y}$. Although definition \eqref{eq:z} may seem unmotivated, in the case where $\Sigma = \sigma^2 I_n$, ${\bm z}$ is simply the residual $(I_n - P_{{\bm\eta}}){\bm y}$ from projecting ${\bm y}$ onto ${\bm\eta}$. We can now rewrite $\{A{\bm y} \leq {\bm b}\}$ in terms of ${\bm\eta}^T {\bm y}$ and ${\bm z}$. \begin{lemma} \label{lem:conditional} Let ${\bm z}$ be defined as in \eqref{eq:z} and ${\bm c}$ as in \eqref{eq:c}. Then, the conditioning set can be rewritten as follows: \[ \{A{\bm y} \leq {\bm b}\} = \{{\mathcal V}^-({\bm z}) \leq {\bm\eta}^T {\bm y} \leq {\mathcal V}^+({\bm z}), {\mathcal V}^0({\bm z}) \geq 0 \} \] where \begin{align} {\mathcal V}^-({\bm z}) &\equiv \max_{j: (A{\bm c})_j < 0} \frac{b_j - (A{\bm z})_j} {(A{\bm c})_j} \label{eq:v_minus} \\ {\mathcal V}^+({\bm z}) &\equiv \min_{j: (A{\bm c})_j > 0} \frac{b_j - (A{\bm z})_j}{(A{\bm c})_j} \label{eq:v_plus} \\ {\mathcal V}^0({\bm z}) &\equiv \min_{j: (A{\bm c})_j = 0} b_j - (A{\bm z})_j. \label{eq:v_zero} \end{align} Note that ${\mathcal V}^-$, ${\mathcal V}^+$, and ${\mathcal V}^0$ refer to functions. Since they are functions of ${\bm z}$ only, \eqref{eq:v_minus}--\eqref{eq:v_zero} are independent of ${\bm\eta}^T {\bm y}$. \end{lemma} \begin{figure} \begin{tikzpicture}[scale=1.5] \defgray!70{gray!70} \defblue!20{blue!20} \draw[fill=blue!20,blue!20] (.4,1) -- (.4,2.6) -- (2,2.6) -- cycle; \node[blue!80] at (1.1,2.4) {$\{A{\bm y} \leq {\bm b}\}$}; \node (y) at (1.1,1.7) {$\bf y$}; \draw[->] (0,0) -- (1,1.9); \draw[thick, |-|] (.4,1.9) -- (1.3,1.9); \draw[->] (0,0) -- (.7,0); \node at (.65,.15) {${\bm\eta}$}; \draw[red,thick,dashed,->] (0,0) -- (1,0); \node[red] at (1.1,.15) {${\bm\eta}^T {\bm y}$}; \draw[red,thick,dashed,->] (0,0) -- (0,1.9); \node[red] at (-.2,1.8) {${\bm z}$}; \draw[dotted] (.4,1.9) -- (.4,-.2); \draw[dotted] (1.3,1.9) -- (1.3,-.2); \node at (.4,-.4) {$V^-({\bm z})$}; \node at (1.4,-.4) {$V^+({\bm z})$}; \end{tikzpicture} \caption{A geometric interpretation of why the event $\left\{A{\bm y} \leq {\bm b} \right\}$ can be characterized as $\{ {\mathcal V}^-({\bm z}) \leq \eta^T y \leq {\mathcal V}^+({\bm z})\}$. Assuming $\Sigma = I$ and $||\eta||_2 = 1$, ${\mathcal V}^-({\bm z})$ and ${\mathcal V}^+({\bm z})$ are functions of ${\bm z}$ only, which is independent of ${\bm\eta}^T {\bm y}$.} \label{fig:polyhedron} \end{figure} \begin{proof} We can decompose ${\bm y} = {\bm c} ({\bm\eta}^T {\bm y}) + {\bm z}$ and rewrite the polyhedron as \begin{align*} \{ A {\bm y} \leq {\bm b} \} &= \{ A({\bm c} ({\bm\eta}^T {\bm y}) + {\bm z}) \leq {\bm b} \} \\ &= \{ A{\bm c} ({\bm\eta}^T {\bm y}) \leq {\bm b} - A {\bm z} \} \\ &= \left\{ (A{\bm c})_j ({\bm\eta}^T {\bm y}) \leq b_j - (A{\bm z})_j\ \text{for all $j$} \right\} \\ &= \left.\begin{cases} {\bm\eta}^T {\bm y} \leq \frac{b_j - (A{\bm z})_j}{(A{\bm c})_j} & \text{for $j: (A{\bm c})_j > 0$} \\ {\bm\eta}^T {\bm y} \geq \frac{b_j - (A{\bm z})_j}{(A{\bm c})_j} & \text{for $j: (A{\bm c})_j < 0$} \\ 0 \leq b_j - (A{\bm z})_j & \text{for $j: (A{\bm c})_j = 0$} \end{cases}\right\}, \end{align*} where in the last step, we have divided the components into three categories depending on whether $(A{\bm c})_j \gtreqless 0$, since this affects the direction of the inequality (or whether we can divide at all). Since ${\bm\eta}^T {\bm y}$ is the same quantity for all $j$, it must be at least the maximum of the lower bounds and no more than the minimum of the upper bounds, which is precisely the definition of ${\mathcal V}^-({\bm z})$ and ${\mathcal V}^+({\bm z})$. Finally, $b_j - (A{\bm z})_j \geq 0$ for all $j: (A{\bm c})_j = 0$ is encoded by ${\mathcal V}^0({\bm z}) \geq 0$. \end{proof} Lemma \ref{lem:conditional} tells us that \begin{equation} \left[{\bm\eta}^T {\bm y}\ |\ \{A{\bm y} \leq {\bm b}\}\right]\ \overset{d}{=}\ \left[{\bm\eta}^T {\bm y}\ |\ \{{\mathcal V}^-({\bm z}) \leq {\bm\eta}^T {\bm y} \leq {\mathcal V}^+({\bm z}), {\mathcal V}^0({\bm z}) \geq 0\}\right] \end{equation} Since ${\mathcal V}^+({\bm z}),{\mathcal V}^-({\bm z}), {\mathcal V}^0({\bm z})$ are independent of ${\bm\eta}^T {\bm y}$, they behave as ``fixed'' quantities. Thus, ${\bm\eta}^T{\bm y}$ is conditionally like a normal random variable, truncated to be between ${\mathcal V}^-({\bm z})$ and ${\mathcal V}^+({\bm z})$. We would like to be able to say \begin{equation*} \text{`` }{\bm\eta}^T {\bm y}\ |\ \{A{\bm y} \leq {\bm b}\} \sim TN({\bm\eta}^T {\bm\mu}, \sigma^2 {\bm\eta}^T \Sigma{\bm\eta}, {\mathcal V}^-({\bm z}), {\mathcal V}^+({\bm z})), \text{''} \end{equation*} but this is technically incorrect, since the distribution on the right-hand side changes with ${\bm z}$. By conditioning on the value of ${\bm z}$, ${\bm\eta}^T {\bm y}\ |\ \{A{\bm y} \leq {\bm b}, {\bm z}={\bm z}_0\}$ is a truncated normal. We then use the probability integral transform to obtain a statistic $F^{\bm z}({\bm\eta}^T{\bm y})$ that has a $\unif(0,1)$ distribution for any value of ${\bm z}$. Hence, $F^{\bm z}({\bm\eta}^T{\bm y})$ will also have a $\unif(0,1)$ distribution marginally over ${\bm z}$. We make this precise in the next theorem. \begin{theorem} \label{thm:truncated-gaussian-pivot} Let $F_{\mu, \sigma^2}^{[a, b]}$ denote the CDF of a $N(\mu, \sigma^2)$ random variable truncated to the interval $[a, b]$, i.e.: \begin{equation} F_{\mu, \sigma^2}^{[a, b]}(x) = \frac{\Phi((x-\mu)/\sigma) - \Phi((a-\mu)/\sigma)}{\Phi((b-\mu)/\sigma) - \Phi((a-\mu)/\sigma)} \label{eq:U} \end{equation} where $\Phi$ is the CDF of a $N(0, 1)$ random variable. Then: \begin{equation} F_{{\bm\eta}^T{\bm\mu},\ {\bm\eta}^T \Sigma {\bm\eta}}^{[{\mathcal V}^-({\bm z}), {\mathcal V}^+({\bm z})]}({\bm\eta}^T {\bm y})\ \big|\ \{A{\bm y} \leq {\bm b}\} \sim \unif(0,1) \label{eq:pivot} \end{equation} where ${\mathcal V}^-$ and ${\mathcal V}^+$ are defined in \eqref{eq:v_minus} and \eqref{eq:v_plus}. Furthermore, \[ \left[{\bm\eta}^T {\bm y} \big| A{\bm y} \leq {\bm b}, {\bm z} = {\bm z}_0 \right]\sim TN({\bm\eta}^T{\bm\mu}, \sigma^2 ||{\bm\eta}||^2, {\mathcal V}^-({\bm z}_0), {\mathcal V}^+({\bm z}_0)). \] \end{theorem} \begin{proof} First, apply Lemma \ref{lem:conditional}: \begin{align*} \left[{\bm\eta}^T {\bm y} \big| A{\bm y} \leq {\bm b}, {\bm z} = {\bm z}_0 \right] &\overset{d}{=} \left[{\bm\eta}^T {\bm y} \big| {\mathcal V}^-({\bm z}) \leq {\bm\eta}^T{\bm y} \leq {\mathcal V}^+({\bm z}), {\mathcal V}^0({\bm z})\geq 0, {\bm z} = {\bm z}_0 \right] \\ &\overset{d}{=} \left[{\bm\eta}^T {\bm y} \big| {\mathcal V}^-({\bm z}_0) \leq {\bm\eta}^T{\bm y} \leq {\mathcal V}^+({\bm z}_0), {\mathcal V}^0({\bm z}_0)\geq 0, {\bm z} = {\bm z}_0 \right] \end{align*} The only random quantities left are ${\bm\eta}^T {\bm y}$ and ${\bm z}$. Now we can eliminate ${\bm z} = {\bm z}_0$ from the condition using independence: \begin{align*} \left[{\bm\eta}^T {\bm y} \big| A{\bm y} \leq {\bm b}, {\bm z} = {\bm z}_0 \right] &\overset{d}{=} \left[{\bm\eta}^T {\bm y} \big| {\mathcal V}^-({\bm z}_0) \leq {\bm\eta}^T{\bm y} \leq {\mathcal V}^+({\bm z}_0) \right] \\ &\sim TN({\bm\eta}^T{\bm\mu}, \sigma^2 ||{\bm\eta}||^2, {\mathcal V}^-({\bm z}_0), {\mathcal V}^+({\bm z}_0)) \end{align*} Letting $F^{\bm z}({\bm\eta}^T {\bm y}) \equiv F_{{\bm\eta}^T{\bm\mu},\ {\bm\eta}^T \Sigma {\bm\eta}}^{[{\mathcal V}^-({\bm z}), {\mathcal V}^+({\bm z})]}({\bm\eta}^T {\bm y})$, we can apply the probability integral transform to the above result to obtain \begin{align*} \left[F^{{\bm z}}({\bm\eta}^T {\bm y}) \big| A{\bm y} \leq{\bm b}, {\bm z}={\bm z}_0 \right] &\overset{d}{=} \left[F^{{\bm z}_0}({\bm\eta}^T {\bm y}) \big| A{\bm y} \leq {\bm b}, {\bm z}={\bm z}_0 \right] \\ &\sim \unif(0,1) \end{align*} If we let $p_X$ denote the density of a random variable $X$ given $\{ A{\bm y} \leq {\bm b} \}$, what we have just shown is that \[ p_{F^{\bm z}({\bm\eta}^T {\bm y}) | {\bm z}}(t | {\bm z}_0) \equiv \frac{p_{F^{\bm z}({\bm\eta}^T {\bm y}), {\bm z}}(t, {\bm z}_0)}{p_{{\bm z}}({\bm z}_0)} = 1_{[0,1]}(f) \] for any ${\bm z}_0$. The desired result now follows by integrating over ${\bm z}_0$: \begin{align*} p_{F^{\bm z}({\bm\eta}^T {\bm y})}(t) &= \int p_{F^{\bm z}({\bm\eta}^T {\bm y})|{\bm z}}(t|{\bm z}_0)\, p_{{\bm z}}({\bm z}_0)\,d{\bm z}_0 \\ &= \int 1_{[0,1]}(t)\, p_{{\bm z}}({\bm z}_0)\,d{\bm z}_0 \\ &= 1_{[0,1]}(t). \end{align*} \end{proof} \subsection{Conditioning on a Union of Polyhedra} \label{sec:union} We have just characterized the distribution of ${\bm\eta}^T {\bm y}$, conditional on ${\bm y}$ falling into a single polyhedron $\{ A{\bm y} \leq {\bm b} \}$. We obtain such a polyhedron if we condition on both the model and the signs $\{ \hatM = M, \hat{\bm s} = {\bm s} \}$. If we want to only condition on the model $\{ \hatM = M \}$, then we will have to understand the distribution of ${\bm\eta}^T {\bm y}$, conditional on ${\bm y}$ falling into a union of such polyhedra, i.e., \begin{equation} {\bm\eta}^T {\bm y}\ \Bigg|\ \bigcup_{\bm s}\ \{ A_{\bm s}{\bm y} \leq {\bm b}_{\bm s} \}. \end{equation} \begin{figure} \begin{tikzpicture}[scale=1.2] \defgray!70{gray!70} \defblue!20{blue!20} \node at (-1.5,-1.5) {}; \draw[gray!70] (-1,1) -- (.4,1) -- (1,.4) -- (1,-1) -- (-.4,-1) -- (-1,-.4) -- cycle; \draw[fill=blue!20,blue!20] (.4,1) -- (.4,3) -- (2.4,3) -- cycle; \draw[fill=blue!20,blue!20] (-.4,-1) -- (-.4,-3) -- (-2.4,-3) -- cycle; \node[rotate=63.43495] (y) at (.7, 2.6) {$\bf y$}; \draw[->] (0,0) -- (.7, 2.4); \node[rotate=63.43495] (y) at (0, 1.9) {${\mathcal V}^-_{\{+,+\}}({\bm z})$}; \draw[|-,thick] (.4,1.8) -- (1, 3); \draw[dotted] (0.8, 2.6) -- (-1.6, -2.2); \draw[|-,thick] (-1.6, -2.2) -- (-2, -3); \node[rotate=63.43495] (y) at (-2, -2.0) {${\mathcal V}^+_{\{-,-\}}({\bm z})$}; \draw[->] (0,0) -- (.4,.8); \node at (.6, .8) {${\bm\eta}$}; \draw[red,thick,dashed,->] (0,0) -- (1.1, 2.2); \node[red, rotate=63.43495] at (1.25, 2.5) {${\bm\eta}^T {\bm y}$}; \draw[red,thick,dashed,->] (0,0) -- (-.4, .2); \node[red] at (-.6, .3) {${\bm z}$}; \draw[gray!70] (.4,1) -- (.4,3); \draw[gray!70] (.4,1) -- (2.4,3); \draw[gray!70] (1,.4) -- (2.4,1.8); \draw[gray!70] (1,.4) -- (2.4,.4); \draw[gray!70] (1,-1) -- (2.4,-1); \draw[gray!70] (1,-1) -- (1,-3); \draw[gray!70] (-.4,-1) -- (-.4,-3); \draw[gray!70] (-.4,-1) -- (-2.4,-3); \draw[gray!70] (-1,-.4) -- (-2.4,-1.8); \draw[gray!70] (-1,-.4) -- (-2.4,-.4); \draw[gray!70] (-2.4,1) -- (-1,1) -- (-1,3); \end{tikzpicture} \caption{When we take the union over signs, the conditional distribution of ${\bm\eta}^T {\bm y}$ is truncated to a union of disjoint intervals. In this case, the Gaussian is truncated to the set $(-\infty, {\mathcal V}^+_{\{-, -\}}({\bm z})] \cup [{\mathcal V}^-_{\{+, +\}}({\bm z}), \infty)$.} \label{fig:union} \end{figure} As Figure \ref{fig:union} makes clear, the argument proceeds exactly as before, except that ${\bm\eta}^T {\bm y}$ is now truncated to a union of intervals, instead of a single interval. There is a ${\mathcal V}^-$ and a ${\mathcal V}^+$ for each possible sign pattern ${\bm s}$, so we index the intervals by the signs. This leads immediately to the next theorem, whose proof is essentially the same as that of Theorem \ref{thm:truncated-gaussian-pivot}. \begin{theorem} \label{thm:union} Let $F_{\mu, \sigma^2}^S$ denote the CDF of a $N(\mu, \sigma^2)$ random variable truncated to the set $S$. Then: \begin{equation} F_{{\bm\eta}^T{\bm\mu},\ {\bm\eta}^T\Sigma{\bm\eta}}^{\bigcup_{{\bm s}} [{\mathcal V}_{{\bm s}}^-({\bm z}), {\mathcal V}_{{\bm s}}^+({\bm z})]}({\bm\eta}^T {\bm y})\ \Bigg|\ \bigcup_{\bm s}\ \{ A_{\bm s}{\bm y} \leq {\bm b}_{\bm s} \}. \sim \unif(0,1), \label{eq:minimal-pivotal-quantity} \end{equation} where ${\mathcal V}_{{\bm s}}^-({\bm z})$ and ${\mathcal V}_{{\bm s}}^+({\bm z})$ are defined in \eqref{eq:v_minus} and \eqref{eq:v_plus} and $A = A_{\bm s}$ and $b = b_{\bm s}$. \end{theorem} \section{Post-Selection Intervals for Regression Coefficients} \label{sec:lasso} In this section, we combine the characterization of the lasso selection event in Section \ref{sec:lasso-selection} with the results about the distribution of a Gaussian truncated to a polyhedron (or union of polyhedra) in Section \ref{sec:polyhedra} to form post-selection intervals for lasso-selected regression coefficients. The key link is that the lasso selection event can be expressed as a union of polyhedra: \begin{align*} \{\hatM = M \} &= \bigcup_{{\bm s} \in \{-1, 1\}^{|M|}} \{\hatM = M, \hat{\bm s} = {\bm s}\} \\ &= \bigcup_{{\bm s} \in \{-1, 1\}^{|M|}} \{A(M, {\bm s}){\bm y} \leq {\bm b}(M, {\bm s})\}, \end{align*} where $A(M, {\bm s})$ and ${\bm b}(M, {\bm s})$ are defined in Theorem \ref{thm:lasso_partition}. Therefore, conditioning on selection is the same as conditioning on a union of polyhedra, so the framework of Section \ref{sec:polyhedra} applies. Recall that our goal is to form confidence intervals for $\beta^M_j = {\bm e}_j^T X_M^\dagger {\bm\mu}$, with $(1-\alpha)$-coverage conditional on $\{\hatM = M\}$. Taking ${\bm\eta} = (X_M^\dagger)^T {\bm e}_j$, we can use Theorem \ref{thm:union} to obtain \[ F^{\bigcup_{\bm s} [{\mathcal V}_{\bm s}^-({\bm z}), {\mathcal V}_{\bm s}^+({\bm z})]}_{\beta^M_j, \sigma^2 ||{\bm\eta}||^2}({\bm\eta}^T {\bm y})\ \big|\ \{\hat{M} = M\} \sim \unif(0,1). \] This gives us a test statistic for testing any hypothesized value of ${\bm\beta}^M_j$. We can invert this test to obtain a confidence set \begin{equation} C^M_j \equiv \left\{ \beta^M_j: \frac{\alpha}{2} \leq F^{\bigcup_{\bm s} [{\mathcal V}_{\bm s}^-({\bm z}), {\mathcal V}_{\bm s}^+({\bm z})]}_{\beta^M_j, \sigma^2 ||{\bm\eta}||^2}({\bm\eta}^T {\bm y}) \leq 1-\frac{\alpha}{2} \right\}. \label{eq:conf_int} \end{equation} In fact, the set $C^M_j$ is an {\em interval}, as formalized in the next result. \begin{theorem} \label{thm:conf_int} Let ${\bm\eta} = (X_M^\dagger)^T {\bm e}_j$. Let $L$ and $U$ be the (unique) values satisfying \begin{align*} F_{L,\ \sigma^2 ||{\bm\eta}||^2}^{\bigcup_{\bm s} [{\mathcal V}^-_{\bm s}({\bm z}), {\mathcal V}^+_{\bm s}({\bm z})]}({\bm\eta}^T {\bm y}) &= 1-\frac{\alpha}{2} & F_{U,\ \sigma^2 ||{\bm\eta}||^2}^{\bigcup_{\bm s} [{\mathcal V}^-_{\bm s}({\bm z}), {\mathcal V}^+_{\bm s}({\bm z})]}({\bm\eta}^T {\bm y}) &= \frac{\alpha}{2} \end{align*} Then $[L, U]$ is a $(1-\alpha)$ confidence interval for $\beta^M_j$, conditional on $\{ \hatM = M \}$, i.e., \begin{equation} \label{eq:coverage} \mathbb{P} \left( \beta^M_j \in [L, U]\ \big|\ \hatM = M \right) = 1-\alpha. \end{equation} \begin{proof} By construction, $\mathbb{P}_{\beta^M_j}(\beta^M_j \in C^M_j | \hatM = M) = 1-\alpha$, where $C^M_j$ is defined in \eqref{eq:conf_int}. The claim is that the set $C^M_j$ is in fact the interval $[L, U]$. To see this, we need to show that the test statistic $F_{L,\ \sigma^2 ||{\bm\eta}||^2}^{\bigcup_{\bm s} [{\mathcal V}^-_{\bm s}({\bm z}), {\mathcal V}^+_{\bm s}({\bm z})]}({\bm\eta}^T {\bm y})$ is monotone decreasing in $\beta^M_j$ so that it crosses $1-\frac{\alpha}{2}$ and $\frac{\alpha}{2}$ at unique values. This follows from the fact that the truncated Gaussian distribution has monotone likelihood ratio in the mean parameter. See Appendix \ref{appendix:monotone} for details. \end{proof} \end{theorem} Alternatively, we could have conditioned on the signs, in addition to the model, so that we would only have to worry about conditioning on a single polyhedron. We also showed in Section \ref{sec:polyhedra} that \[ F^{[{\mathcal V}^-_{\bm s}({\bm z}), {\mathcal V}^+_{\bm s}({\bm z})]}_{{\bm\beta}^M_j, \sigma^2 ||{\bm\eta}||^2}({\bm\eta}^T {\bm y})\ |\ \{\hat{M} = M, \hat{\bm s} = {\bm s}\} \sim \unif(0,1). \] Inverting this statistic will produce intervals that have $(1-\alpha)$ coverage conditional on $\{\hat{M} = M, \hat{\bm s} = {\bm s}\}$, and hence, $(1-\alpha)$ coverage conditional on $\{\hat{M} = M\}$. However, these intervals will be less efficient; they will in general be wider. However, one may be willing to sacrifice statistical efficiency for computational efficiency. Notice that the main cost in computing intervals according to Theorem \ref{thm:conf_int} is determining the intervals $[{\mathcal V}^-_{\bm s}({\bm z}), {\mathcal V}^+_{\bm s}({\bm z})]$ for each ${\bm s} \in \{-1, 1\}^{|M|}$. The number of such sign patterns is $2^{|M|}$. While this might be feasible when $|M|$ is, say, less than 15, it is not feasible when we select hundreds of variables. Conditioning on the signs means that we only have to compute the interval $[{\mathcal V}^-_{\bm s}({\bm z}), {\mathcal V}^+_{\bm s}({\bm z})]$ for the sign pattern ${\bm s}$ that was actually observed. Figure \ref{fig:ci_comparison} shows the tradeoff in statistical efficiency. When the signal is strong, as in the left-hand plot, there is virtually no difference between the intervals obtained by conditioning on just the model, or the model and signs. On the other hand, in the right-hand plot, we see that we can obtain very wide intervals when the signal is weak. The widest intervals are for actual noise variables, as expected. \begin{figure} \includegraphics[width=.5\textwidth]{ci_comparison_strong.pdf}\includegraphics[width=.5\textwidth]{ci_comparison_weak.pdf} \caption{Comparison of the confidence intervals by conditioning on the model only (statistically more efficient, but computationally more expensive) and conditioning on both the model and signs (statistically less efficient, but computationally more feasible). Data were simulated for $n=25$, $p=50$, and 5 true non-zero coefficients; only the first 20 coefficients are shown. (Variables with no intervals are included to emphasize that inference is only on the selected variables.) Conditioning on the signs in addition to the model results in no loss of statistical efficiency when the signal is strong (left) but is problematic when the signal is weak (right).} \label{fig:ci_comparison} \end{figure} To understand why post-selection intervals are sometimes very wide, notice that when a truncated Gaussian random variable $Z$ is close to the endpoints of the truncation interval $[a, b]$, there are many means $\mu$ that would be consistent with that observation---hence, the wide intervals. Figure \ref{fig:intervals} shows confidence intervals for $\mu$ as a function of $Z$. When $Z$ is far from the endpoints of the truncation interval, we basically recover the nominal OLS intervals (i.e., not adjusted for selection). \begin{figure}[!h] \includegraphics[width = .48\textwidth]{intervals_finite.pdf} \includegraphics[width = .48\textwidth]{intervals_infinite.pdf} \caption{Upper and lower bounds of 90\% confidence intervals for $\mu$ based on a single observation $x/\sigma \sim TN(0, 1, -3, 3)$. We see that as long as the observation $x$ is roughly $0.5\sigma$ away from either boundary, the size of the intervals is comparable to the unadjusted OLS confidence interval.} \label{fig:intervals} \end{figure} The implications are clear. When the signal is strong, ${\bm\eta}^T {\bm y}$ will be far from the endpoints of the truncation region, so we obtain the nominal OLS intervals. On the other hand, when a variable just barely entered the model, then ${\bm\eta}^T {\bm y}$ will be close to the edge of the truncation region, and the interval will be wide. \subsection{Optimality} We have derived a confidence interval $C^M_j$ whose conditional coverage, given $\{\hatM = M\}$, is at least $1-\alpha$. The fact that we have found such an interval is not remarkable, since many such intervals have this property. However, given two intervals with the same coverage, we generally prefer the shorter one. This problem is considered in \cite{fithian2014optimal} where it is shown that $C^M_j$ is, with one small tweak, the shortest interval among all {\em unbiased} intervals with $1-\alpha$ coverage. An {\em unbiased} interval $C$ for a parameter $\theta$ is one which covers no other parameter $\theta'$ with probability more than $1-\alpha$, i.e., \begin{equation} \mathbb{P}_{\theta}(\theta' \in C) \leq 1-\alpha,\ \text{for all $\theta$, $\theta' \neq \theta$}. \end{equation} Unbiasedness is a common restriction to ensure that there is an optimal interval or test at all \citep{TSH}. The shortest unbiased interval for $\beta^M_j$, among all intervals with conditional $1-\alpha$ coverage, is similar to the interval $[L, U]$ in Theorem \ref{thm:conf_int}. The only difference is that the critical values $L$ and $U$ were chosen symmetrically so that the pivot has $\alpha/2$ area in either tail. However, allocating the $\alpha$ probability equally to either tail may not be optimal in general. Theorem 5 of \citet{fithian2014optimal} provides the general recipe for constructing optimal intervals. \section{Data Example} \label{sec:examples} We apply our post-selection intervals to the diabetes data set from \citet{efron2004least}. Since $p < n$ for this data set, we can estimate $\sigma^2$ using the residual sum of squares from the full regression model with all $p$ predictors. After standardizing all variables, we chose $\lambda$ according to the strategy in \cite{negahban2012unified}, $\lambda = 2 \Expect(\|X^T\epsilon\|_{\infty})$. This expectation was computed by simulation, where $\epsilon \sim N(0, \hat\sigma^2)$, resulting in $\lambda \approx 190$. The lasso selected four variables: \verb\BMI\, \verb\BP\, \verb\S3\, and \verb\S5\. The post-selection intervals are shown in Figure \ref{fig:diabetes}, alongside the nominal confidence intervals produced by fitting OLS to the four selected variables, ignoring selection. The nominal intervals do not have $(1-\alpha)$ coverage conditional on the model and are not valid post-selection intervals. Also depicted are the confidence intervals obtained by data splitting, as discussed in Section \ref{sec:goals}. This is a competitor method that also produces valid confidence intervals conditional on the model. The lasso selected the same four variables on half of the data, and then nominal intervals for these four variables using OLS on the other half of the data. We can make two observations from Figure \ref{fig:diabetes}. \begin{enumerate} \item The adjusted intervals provided by our method essentially reproduces the OLS intervals for the strong effects, whereas data splitting intervals are wider by a factor of $\sqrt{2}$ (since only $n/2$ observations are used in the inference). For this dataset, the POSI intervals are $1.36$ times wider than the OLS intervals. For all the variables, our method produces the shortest intervals among the methods that control selective type 1 error. \item One variable, \verb\S3\, which would have been deemed significant using the OLS intervals, is no longer significant after accounting for selection. Data splitting, our selection-adjusted intervals, and POSI intervals conclude that \verb\S3\ is not significant. This demonstrates that taking model selection into account can have substantive impacts on the conclusions. \end{enumerate} \begin{figure}[!h] \includegraphics[width=.72\textwidth]{diabetes_confint_simul.pdf} \caption{Inference for the four variables selected by the lasso ($\lambda = 190$) on the diabetes data set. The point estimate and adjusted confidence intervals using the approach in Section \ref{sec:lasso} are shown in black. The OLS intervals, which ignore selection, are shown in red. The green lines show the intervals produced by splitting the data into two halves, forming the interval based on only half of the data. The blue line corresponds to the POSI method of \cite{berk2013posi}.} \label{fig:diabetes} \end{figure} \section{Extensions} \subsection{Estimation of $\sigma^2$} \label{sec:estimate:sigma} The above results rely on knowing $\sigma^2$ or at least having a good estimate of it. If $n > p$, then the variance $\hat\sigma^2$ of the residuals from fitting the full model is a consistent estimator and in general can be substituted for $\sigma^2$ to yield asymptotically valid confidence intervals. Formally, the condition is that the pivot is smooth with respect to $\sigma$. Geometrically speaking, the upper and lower truncation limits ${\mathcal V}^+$ and ${\mathcal V}^-$ must be well-separated (with high probability). We refer the interested reader to Section 2.3 in \cite{tian2015asymptotics} for details. In the setting where $p > n$, obtaining an estimate of $\sigma^2$ is more challenging, but if the pivot satisfies a monotonicity property, plugging in an overestimate of the variance gives conservative confidence intervals. We refer the reader to Theorem 11 in \cite{tibshirani2015uniform} for details. \subsection{Elastic Net} One problem with the lasso is that it tends to select one variable out of a set of correlated variables, resulting in estimates that are unstable. One way to stabilize them is to add an $\ell_2$ penalty to the lasso objective, resulting in the elastic net \citep{zou2005regularization}: \begin{align} \tilde{\bm\beta} = \underset{{\bm\beta}}{\text{argmin}}\ \frac{1}{2} \norm{{\bm y}-X {\bm\beta}}_2^2 +\lambda \norm{{\bm\beta}}_1+ \frac{\gamma}{2} \norm{{\bm\beta}}_2 ^2. \label{eq:elastic-net} \end{align} Using a nearly identical argument to Lemma \ref{lem:equiv_sets}, we see that $\{ \hatM = M, \hat{\bm s} = {\bm s} \}$ if and only if there exist $\tilde{\bm w}$ and $\tilde{\bm u}$ satisfying \begin{align*} (X_M^T X_M + \gamma I) \tilde{\bm w} - X_M^T {\bm y} + \lambda {\bm s} &= 0 \\ X_{-M}^T (X_M \tilde{\bm w} - {\bm y}) + \lambda \tilde{\bm u} &= 0 \\ \sign(\tilde{\bm w}) &= {\bm s} \\ ||\tilde{\bm u}||_\infty &< 1. \end{align*} These four conditions differ from those of Lemma \ref{lem:equiv_sets} in only one respect: $X_M^T X_M$ in the first expression is replaced by $X_M^T X_M + \gamma I$. Continuing the argument of Section \ref{sec:lasso-selection}, we see that the selection event can be rewritten \begin{equation} \{ \hatM = M, \hat{\bm s} = {\bm s} \} = \left\{ \begin{pmatrix} \tilde A_0(M, {\bm s}) \\ \tilde A_1(M, {\bm s}) \end{pmatrix} {\bm y} < \begin{pmatrix} \tilde{\bm b}_0(M, {\bm s}) \\ \tilde {\bm b}_1(M, {\bm s}) \end{pmatrix}\right\} \label{eq:enet_A_b} \end{equation} where $\tilde A_k$ and $\tilde{\bm b}_k$ are analogous to $A_k$ and ${\bm b}_k$ in Proposition~\ref{prop:A_b}, except replacing $(X_M^T X_M)^{-1}$ by $(X_M^T X_M + \gamma I)^{-1}$ everywhere it appears. Notice that $(X_M^T X_M)^{-1}$ appears explicitly in $A_1$ and ${\bm b}_1$, and also implicitly in $A_0$ and ${\bm b}_0$, since $P_M$ and $(X_M^T)^\dagger$ both depend on $(X_M^T X_M)^{-1}$. Now that we have rewritten the selection event in the form $\{A{\bm y} \leq {\bm b}\}$, we can once again apply the framework in Section \ref{sec:polyhedra} to obtain a test for the elastic net conditional on this event. \section{Conclusion} Model selection and inference have long been regarded as conflicting goals in linear regression. Following the lead of \citet{berk2013posi}, we have proposed a framework for post-selection inference that {\em conditions on which model was selected}, i.e., the event $\{\hatM = M\}$. We characterize this event for the lasso and derive optimal and exact confidence intervals for linear contrasts ${\bm\eta}^T {\bm\mu}$, conditional on $\{\hatM = M\}$. With this general framework, we can form post-selection intervals for regression coefficients, equipping practitioners with a way to obtain ``valid'' intervals even after model selection. \section*{Acknowledgements} We thank Will Fithian, Sam Gross, and Josh Loftus for helpful comments and discussions. In particular, Will Fithian provided insights that led to the geometric intuition of our procedure shown in Figure \ref{fig:polyhedron}. J. Lee was supported by a National Defense Science and Engineering Graduate Fellowship and a Stanford Graduate Fellowship. D. L. Sun was supported by a Ric Weiland Graduate Fellowship and the Stanford Genome Training Program (SGTP; NIH/NHGRI). Y. Sun was partially supported by the NIH, grant U01GM102098. J.E. Taylor was supported by the NSF, grant DMS 1208857, and by the AFOSR, grant 113039. \begin{appendix} \section{Monotonicity of $F$} \label{appendix:monotone} \begin{lemma} Let $F_{\mu}(x) := F_{\mu, \sigma^2}^{[a,b]}(x)$ denote the cumulative distribution function of a truncated Gaussian random variable, as defined as in \eqref{eq:U}. Then $F_\mu(x)$ is monotone decreasing in $\mu$. \end{lemma} \begin{proof} First, the truncated Gaussian distribution with CDF $F_{\mu} := F_{\mu, \sigma^2}^{[a,b]}$ is a natural exponential family in $\mu$, since it is just a Gaussian with a different base measure. Therefore, it has monotone likelihood ratio in $\mu$. That is, for all $\mu_1 > \mu_0$ and $x_1 > x_0$: $$ \frac{f_{\mu_1}(x_1)}{f_{\mu_0}(x_1)} > \frac{f_{\mu_1}(x_0)}{f_{\mu_0}(x_0)}$$ where $f_{\mu_i} := dF_{\mu_i}$ denotes the density. (Instead of appealing to properties of exponential families, this property can also be directly verified.) This implies \begin{align*} f_{\mu_1}(x_1) f_{\mu_0}(x_0) &> f_{\mu_1}(x_0) f_{\mu_0}(x_1) & x_1 > x_0. \end{align*} Therefore, the inequality is preserved if we integrate both sides with respect to $x_0$ on $(-\infty, x)$ for $x < x_1$. This yields: \begin{align*} \int_{-\infty}^{x} f_{\mu_1}(x_1) f_{\mu_0}(x_0)\,dx_0 &> \int_{-\infty}^{x} f_{\mu_1}(x_0) f_{\mu_0}(x_1)\,dx_0 & x < x_1 \\ f_{\mu_1}(x_1) F_{\mu_0}(x) &> f_{\mu_0}(x_1) F_{\mu_1}(x) & x < x_1 \end{align*} Now we integrate both sides with respect to $x_1$ on $(x, \infty)$ to obtain: \begin{align*} (1 - F_{\mu_1}(x)) F_{\mu_0}(x) &> (1 - F_{\mu_0}(x)) F_{\mu_1}(x) \end{align*} which establishes $F_{\mu_0}(x) > F_{\mu_1}(x)$ for all $\mu_1 > \mu_0$. \end{proof} \end{appendix} \frenchspacing \bibliographystyle{ims}

https://arxiv.org/abs/1808.00596

Ergodic Theorems for the Shift Action and Pointwise Versions of The Abért--Weiss Theorem

Let $\Gamma$ be a countably infinite group. A common theme in ergodic theory is to start with a probability measure-preserving (p.m.p.) action $\Gamma \curvearrowright (X, \mu)$ and a map $f \in L^1(X, \mu)$, and to compare the global average $\int f \,\mathrm{d}\mu$ of $f$ to the pointwise averages $|D|^{-1} \sum_{\delta \in D} f(\delta \cdot x)$, where $x \in X$ and $D$ is a nonempty finite subset of $\Gamma$. The basic hope is that, when $D$ runs over a suitably chosen infinite sequence, these pointwise averages should converge to the global value for $\mu$-almost all $x$.In this paper we prove several results that refine the above basic paradigm by uniformly controlling the averages over specific sets $D$ rather than considering their limit as $|D| \to \infty$. Our results include ergodic theorems for the Bernoulli shift action $\Gamma \curvearrowright ([0;1]^\Gamma, \lambda^\Gamma)$ and strengthenings of the theorem of Abért and Weiss that the shift is weakly contained in every free p.m.p. action of $\Gamma$. In particular, we establish a purely Borel version of the Abért--Weiss theorem for finitely generated groups of subexponential growth. The central role in our arguments is played by the recently introduced measurable versions of the Lovász Local Lemma, due to the current author and to Csóka, Grabowski, Máthé, Pikhurko, and Tyros.

\section{Introduction} The \emph{Lov\'asz Local Lemma} \ep{the \emph{LLL} for short} is a powerful tool in probabilistic combinatorics, introduced by Erd\H os and Lov\'asz \cite{EL}. The LLL is mostly used to obtain existence results, and it is particularly well-suited for showing that a given structure $X$ admits a coloring satisfying some ``local'' constraints. Roughly speaking, in order for the LLL to apply in this context, two requirements must be met: First, a \emph{random} coloring should be ``likely'' to fulfill each individual constraint; second, the constraints must not interact with each other ``too much.'' For the precise statement, see \S\ref{subsec:SLLL}. It has been a matter of interest to determine if the LLL can be used to derive conclusions that are, in some sense, ``constructive'' \ep{as opposed to pure existence results}. A decisive breakthrough was made by Moser and Tardos \cite{MT}, who developed an \emph{algorithmic} approach to the LLL. \ep{The work of Moser and Tardos was preceded by a line or earlier results, starting with Beck's paper \cite{Beck}; for more details, see the references in \cite{MT}.} The Moser--Tardos method proved quite versatile and was adapted to establish ``constructive'' analogs of the LLL in a variety of different contexts. For example, Rumyantsev and Shen \cite{RSh} proved a \emph{computable} version of the LLL. Here we will be focused on the \emph{measurable} versions of the LLL that were studied in \cite{MLLL} by the current author and in \cite{CGMPT} by Cs\'{o}ka, Grabowski, M\'ath\'e, Pikhurko, and Tyros \ep{see also \cite{Kun} for related work by Kun}. Measurable analogs of the LLL are designed to apply in the following framework. Let $(X, \mu)$ be a standard probability space and let $C$ be a set of colors \ep{we will only consider the case when $C$ is finite}. Suppose we are looking for a coloring $f \colon X \to C$ that fulfills a family ${\mathscr{B}}$ of constraints. Under suitable assumptions, the ordinary LLL implies that such a coloring $f$ exists; however, this $f$ need not behave well with respect to the measurable structure on $(X, \mu)$. In contrast to that, measurable versions of the LLL can provide a \emph{$\mu$-measurable} \ep{or sometimes even \emph{Borel}} function $f \colon X \to C$ that satisfies the constraints ${\mathscr{B}}$, or at least does so on a ``large'' subset of $X$. Such results appear to be particularly relevant in ergodic theory, since many concepts pertaining to measure\-/preserving group actions are phrased in terms of measurable partitions of the underlying probability space---which can naturally be thought of as measurable colorings. Some ergodic\-/theoretic applications of the LLL can be found in \cite{MLLL, LSS}. Here we present further consequences of the LLL in measurable dynamics, specifically in the study of ergodic averages and of weak containment of measure\-/preserving group actions. Our arguments employ a general approach that is standard in combinatorics, in particular in graph coloring theory \ep{see, e.g., the book \cite{MR} for many examples}. The first step is to use \emph{concentration of measure} to obtain strong upper bounds on probabilities of certain ``bad'' random events; the LLL is then invoked to eliminate all the ``bad'' events. Nontrivial results can also be derived by combining the concentration of measure bounds with more classical tools, such as the Borel--Cantelli lemma \ep{Theorem~\ref{theo:weak_erg} below is as an example}. Roughly speaking, using the LLL instead of the Borel--Cantelli lemma results in replacing pointwise convergence with approximation in the $\infty$-norm. \subsubsection*{{Acknowledgement}} I am very grateful to Anush Tserunyan for many insightful discussions and to the anonymous referee for helpful suggestions. \section{Statements of results} Throughout, $\G$ denotes a countably infinite group with identity element $\mathbf{1}$. We study \emphd{probability measure\-/preserving \ep{$\text{p.m.p.}$\xspace} actions} of $\G$, i.e., actions of the form $\alpha \colon \G \curvearrowright (X, \mu)$, where $(X, \mu)$ is a standard probability space and the measure $\mu$ is $\alpha$-invariant. We also consider, more generally, \emphd{Borel actions} $\alpha \colon \G \curvearrowright X$, i.e., actions of $\G$ on a standard Borel space $X$ by Borel automorphisms. Given a set $A$, the \emphd{shift action} $\sigma_A \colon \G \curvearrowright A^\G$ on the set of all maps $x \colon \G \to A$ is defined by \[ (\gamma \cdot x)(\delta) \coloneqq x(\delta \gamma) \qquad \text{for all } x \in A^\G \text{ and } \gamma,\ \delta \in \G. \] We are particularly interested in the case when $A$ is the unit interval $[0;1]$ equipped with the Lebesgue probability measure $\lambda$. \ep{Owing to the measure isomorphism theorem \cite[Theorem~17.41]{K_DST}, any other atomless standard probability space could be used instead.} To unclutter the notation, set \[(\Omega, {\bm{\lambda}}) \coloneqq ([0;1]^\G, \lambda^\G)\] and $\sigma \coloneqq \sigma_{[0;1]}$. Note that the action $\sigma \colon \G \curvearrowright (\Omega, {\bm{\lambda}})$ is measure\-/preserving. \subsection{Ergodic theorems for the shift action}\label{subsec:erg} Let $\alpha \colon \G \curvearrowright (X, \mu)$ be a $\text{p.m.p.}$\xspace action. Given $f \in L^1(X, \mu)$, we can compute its \emphd{global average}: \begin{equation* \mathbb{E}_\mu f \coloneqq \int_X f \,\mathrm{d} \mu, \end{equation*} and compare it to the \emphd{pointwise averages} of the form \begin{equation* \mathbb{E}_D f(x) \coloneqq \frac{1}{|D|} \sum_{\delta \in D} f(\delta \cdot x), \end{equation*} where $x \in X$ and $D$ is a nonempty finite subset of $\G$. Note that $\mathbb{E}_D \colon L^1(X, \mu) \to L^1(X, \mu)$ is a linear operator of norm $1$: The lower bound on $\|\mathbb{E}_D\|_{\mathrm{op}}$ is witnessed by the constant $1$ function, while the upper bound follows from the fact that, since $\mu$ is $\alpha$-invariant, we have $\mathbb{E}_\mu \mathbb{E}_D f = \mathbb{E}_\mu f$, and hence \[ \|\mathbb{E}_D f\|_1 \,=\, \mathbb{E}_\mu |\mathbb{E}_D f| \,\leq\, \mathbb{E}_\mu \mathbb{E}_D |f| \,=\, \mathbb{E}_\mu |f| \,=\, \|f\|_1. \] Assuming the action $\alpha$ is ergodic, one hopes to show that the pointwise averages $\mathbb{E}_D f$ converge, in a suitable sense, to $\mathbb{E}_\mu f$, as $D$ ranges over a given infinite family of finite subsets of $\G$. Results of this kind are usually referred to as \emph{ergodic theorems} \ep{often with adjectives indicating the mode of convergence, such as ``pointwise''}. Two prototypical examples are \emph{von Neumann's} \cite{vN} and \emph{Birkhoff's} \cite{Birkhoff} \emph{ergodic theorems}. Both of these classical results apply when $\G = \mathbb{Z}$ and $D$ ranges over the sets of the form $\set{0, 1, \ldots, n-1}$ with $n \in {\mathbb{N}}^+$. Von Neumann's theorem yields convergence in the $2$-norm \ep{assuming $f \in L^2(X, \mu)$ to begin with}, while Birkhoff's result ensures pointwise convergence almost everywhere. An extension of Birkhoff's pointwise ergodic theorem to all amenable $\G$ was obtained by Lindenstrauss \cite{Lin}; there $D$ ranges over a \emph{tempered F\o lner sequence} \ep{the special case of Lindenstrauss's result for $f \in L^2(X, \mu)$ follows from the earlier work of Shulman, see \cite[\S5.6]{Temp}}. Generalizing ergodic theorems beyond the realm of amenable groups is a major challenge; for further background, see, e.g., \cite{AAB, BufKlim, BowNev} and the references therein. Here we work with an \emph{arbitrary} group $\G$; moreover, the only condition on the sequence $(D_n)_{n \in {\mathbb{N}}}$ of averaging sets is that $|D_n|$ grows sufficiently quickly with $n$. On the other hand, instead of studying arbitrary ergodic actions, we focus our attention on the shift action $\sigma \colon \G \curvearrowright (\Omega, {\bm{\lambda}})$ in the hope of exploiting its mixing properties. Our first result is a pointwise ergodic theorem for \emph{continuous} functions $f \colon \Omega \to \C$: \begin{theo}[\textls{Pointwise ergodic theorem for continous maps on the shift}]\label{theo:weak_erg} Let $(D_n)_{n \in {\mathbb{N}}}$ be a sequence of finite subsets of $\G$ such that $|D_n|/\log n \to\infty$. Then, for all continuous $f \colon \Omega \to \C$, \[ \lim_{n \to \infty} \mathbb{E}_{D_n} f(x) \,=\, \mathbb{E}_{{\bm{\lambda}}} f, \qquad \text{for ${\bm{\lambda}}$-a.e.} \ x \in \Omega. \] \end{theo} Since the set of all continuous functions is dense in $L^1(\Omega, {\bm{\lambda}})$ and $\|\mathbb{E}_D\|_{\mathrm{op}} = 1$ for all nonempty finite $D \subset \G$, Theorem~\ref{theo:weak_erg} has the following immediate corollary: \begin{corl}[\textls{Mean ergodic theorem for the shift}]\label{corl:mean_erg} Let $(D_n)_{n \in {\mathbb{N}}}$ be a sequence of finite subsets of $\G$ such that $|D_n|/\log n \to\infty$. Then, for all $f \in L^1 (\Omega, {\bm{\lambda}})$, we have \[ \lim_{n \to \infty} \mathbb{E}_{D_n} f \,=\, \mathbb{E}_{{\bm{\lambda}}} f \qquad \text{in}\ L^1(\Omega, {\bm{\lambda}}). \] \end{corl} It is natural to ask whether Theorem~\ref{theo:weak_erg} can be extended to all $f \in L^1(\Omega, {\bm{\lambda}})$. The answer turns out to be negative even if the lower bound on the growth rate of the averaging sets is raised, as the constructions of Akcoglu and del~Junco \cite{AkcogluJunco} and del~Junco and Rosenblatt \cite{JunRos} \ep{with minor modifications} demonstrate: \begin{theo}[{$\text{ess.}$ Akcoglu--del Junco \cite{AkcogluJunco} and del Junco--Rosenblatt \cite{JunRos}}]\label{theo:bad} Suppose that $\G = \mathbb{Z}$ and let $h \colon {\mathbb{N}} \to {\mathbb{N}}$ be an arbitrary function. There exists a sequence $(D_n)_{n \in {\mathbb{N}}}$ of finite subsets of $\mathbb{Z}$ with the following properties: \begin{itemize}[label=--,wide] \item each $D_n$ is an interval, i.e., a set of the form $\set{s, s+1, \ldots, s + \ell - 1}$ for $s \in \mathbb{Z}$ and $\ell \in {\mathbb{N}}^+$; \item $|D_n| \geq h(n)$ for all $n \in {\mathbb{N}}$; \item for every free $\text{p.m.p.}$\xspace action $\mathbb{Z} \curvearrowright (X, \mu)$, there is a Borel set $A \subseteq X$ such that \[ \liminf_{n \to \infty} \,\mathbb{E}_{D_n} \mathbbm{1}_A(x) \,=\, 0 \quad \text{and} \quad \limsup_{n \to \infty} \, \mathbb{E}_{D_n} \mathbbm{1}_A(x) \,=\, 1, \qquad \text{for $\mu$-a.e.}\ x \in X, \] where $\mathbbm{1}_A \colon X \to \set{0,1}$ is the indicator function of $A$. Moreover, the family of such sets $A$ is comeager in the measure algebra $\operatorname{MAlg}(X, \mu)$. \end{itemize} \end{theo} For completeness, we sketch a proof of Theorem~\ref{theo:bad} using Rokhlin's lemma in the \hyperref[sec:app]{appendix}. As mentioned in the introduction, Theorem~\ref{theo:weak_erg} follows by combining a concentration of measure inequality with the Borel--Cantelli lemma. We now turn to further results that can be obtained if the Borel--Cantelli lemma is replaced by the LLL. For a $\text{p.m.p.}$\xspace action $\alpha \colon \G \curvearrowright (X, \mu)$, $f \in L^1(X, \mu)$, and a nonempty finite set $D \subset \G$, define the \emphd{discrepancy norm} of $f$ with respect to $D$ by the formula \[ \|f\|^{\mathrm{disc}}_D \coloneqq \| \mathbb{E}_D f \,-\, \mathbb{E}_\mu f\|_\infty. \] \ep{Here $\| \cdot \|_\infty$ is the $\infty$-norm in the sense of $L^\infty(X, \mu)$.} Even if $f \colon \Omega \to \C$ is continuous, its discrepancy norm may be separated from $0$. For instance, consider the continuous map \[f \colon \Omega \to [-1;1] \colon x \mapsto -1 + 2\cdot x(\mathbf{1}).\] Then $\mathbb{E}_{\bm{\lambda}} f = 0$ yet $\| \mathbb{E}_D f\|_\infty = 1$, and hence $\|f\|^{\mathrm{disc}}_D = 1$, for all nonempty finite $D \subset \G$. However, we show that any $f \in L^1(\Omega, {\bm{\lambda}})$ can be written as a {sum} of two functions $g$, $h \in L^1(\Omega, {\bm{\lambda}})$, where $g$ is small in the discrepancy norm, while $h$ is small in the $1$-norm: \begin{theo}[\textls{$L^\infty$-ergodic theorem for the shift}]\label{theo:strong_erg} For all $f \in L^1(\Omega, {\bm{\lambda}})$ and $\epsilon > 0$, there exists $C > 0$ with the following property: \smallskip Let $(D_n)_{n \in {\mathbb{N}}}$ be a sequence of finite subsets of $\G$ with $|D_n|\geq C \log(n+2)$ for all $n \in {\mathbb{N}}$. Then there exist $g$, $h \in L^1(\Omega, {\bm{\lambda}})$ such that $f = g + h$, $\|h\|_1 \leq \epsilon$, and $\|g\|^{\mathrm{disc}}_{D_n} \leq \epsilon$ for all $n \in {\mathbb{N}}$. \end{theo} Note that Theorem~\ref{theo:strong_erg} also yields Corollary~\ref{corl:mean_erg}. Our ultimate goal in this subsection is to sharpen Theorem~\ref{theo:strong_erg} by considering other statistical properties of the function $f$, beside its average $\mathbb{E}_{\bm{\lambda}} f$. This is made precise by the following formalism. Let $K$ be a compact metric space. We use $\Prob(K)$ to denote the space of all probability Borel measures on $K$ equipped with the usual weak-$\ast$ topology \ep{see, e.g., \cite[\S{}17.E]{K_DST}}. Given a $\text{p.m.p.}$\xspace action $\alpha \colon \G \curvearrowright (X, \mu)$ and a Borel function $f \colon X \to K$, define \[ \mathbb{M}_\mu f \coloneqq f_\ast(\mu), \] where $f_\ast \colon \Prob(X) \to \Prob(K)$ is the pushforward map. \ep{This notation is intended to be reminiscent of $\mathbb{E}_\mu f$, while the letter ``$\mathbb{M}$'' emphasizes that $\mathbb{M}_\mu f$ is a \emph{measure}.} For $x \in X$ and a nonempty finite set $D \subset \G$, let $\upsilon_{x, D}$ be the probability measure on $X$ with \ep{finite} support $D \cdot x$ given by \begin{equation* \upsilon_{x, D}(\set{y}) \coloneqq \frac{1}{|D|} \cdot |\set{\delta \in D \,:\, \delta \cdot x = y}| \qquad \text{for all } y \in D \cdot x. \end{equation*} If the $\alpha$-stabilizer of $x$ is trivial, then $\upsilon_{x, D}$ is simply the uniform probability measure on $D \cdot x$. Let \[ \mathbb{M}_D f(x) \coloneqq f_\ast(\upsilon_{x, D}). \] The measures $\mathbb{M}_\mu f$ and $\mathbb{M}_D f(x)$ are points in $\Prob(K)$ that encode the global and the pointwise statistics of $f$, respectively. In particular, when $K$ is a subset of $\C$, $\mathbb{M}_\mu f$ and $\mathbb{M}_D f(x)$ contain the information about $\mathbb{E}_\mu f$ and $\mathbb{E}_D f(x)$; explicitly, \begin{equation}\label{eq:star} \mathbb{E}_\mu f = \int_K z \,\mathrm{d} (\mathbb{M}_\mu f)(z) \qquad \text{and} \qquad \mathbb{E}_D f(x) = \int_K z \,\mathrm{d} (\mathbb{M}_D f(x))(z). \end{equation} We wish to also take into account more detailed information about the interaction of $f$ with the action $\alpha$. Toward that end, let $\code{f} \colon X \to K^\G$ denote the equivariant map given by \[ \code{f}(x)(\gamma) \coloneqq f(\gamma \cdot x) \qquad \text{for all } x \in X \text{ and } \gamma \in \G. \] The map $\code{f}$ is called the \emphd{symbolic representation}, or the \emphd{coding map}, of the dynamical system $(X, \G, \alpha, f)$. Notice that the projection function $\mathsf{p} \colon K^\G \to K \colon \kappa \mapsto \kappa(\mathbf{1})$ satisfies $f = \mathsf{p} \circ \code{f}$ and gives rise to a continuous map $\mathsf{p}_\ast \colon \Prob(K^\G) \to \Prob(K)$ such that $f_\ast = \mathsf{p}_\ast \circ (\code{f})_\ast$. This observation shows that, by considering $\code{f}$, we achieve greater generality than just by working with $f$ itself. Given a standard probability space $(X, \mu)$ and a compact metric space $(K, {\mathfrak{d}})$, let $\mathfrak{B}(X, K)$ denote the set of all Borel functions from $X$ to $K$. We equip $\mathfrak{B}(X, K)$ with a psedometric ${{\mathfrak{d}}}_\mu$ given by \[ {{\mathfrak{d}}}_\mu(f, g) \coloneqq \int_X {\mathfrak{d}}(f(x), g(x))\, \mathrm{d}\mu(x). \] If $K$ is a subset of $\C$ equipped with the metric ${\mathfrak{d}}(z_1, z_2) = |z_1 - z_2|$, then ${{\mathfrak{d}}}_\mu(f, g) = \|f - g\|_1$. \begin{theo}[\textls{Pushforward-ergodic theorem for the shift}]\label{theo:ult_erg} Let $(K, {\mathfrak{d}})$ be a compact metric space and let $f \colon \Omega \to K$ be a Borel function. For any $\epsilon > 0$ and an open neighborhood $U$ of the measure $\mathbb{M}_{\bm{\lambda}}\code{f}$, there exists $C > 0$ with the following property: \smallskip Let $(D_n)_{n \in {\mathbb{N}}}$ be a sequence of finite subsets of $\G$ with $|D_n|\geq C \log(n+2)$ for all $n \in {\mathbb{N}}$. Then there is a Borel map $g \colon \Omega \to K$ such that ${{\mathfrak{d}}}_{\bm{\lambda}}(f, g) \leq \epsilon$ and \[ \mathbb{M}_{D_n} \code{g}(x) \in U, \qquad \text{for all}\ n \in {\mathbb{N}}\ \text{and for ${\bm{\lambda}}$-a.e.}\ x \in X. \] \end{theo} In the light of \eqref{eq:star}, it is clear that Theorem~\ref{theo:strong_erg} is a special case of Theorem~\ref{theo:ult_erg}. We end this subsection with a simple application of Theorem~\ref{theo:ult_erg}. Recall that a group $\G$ is called \emphd{residually finite} if the intersection of all its subgroups of finite index is trivial. The following is an easy observation: \begin{prop}\label{prop:res_fin} A countable group $\G$ is residually finite if and only if every open neighborhood $U \subseteq \Prob(\Omega)$ of ${\bm{\lambda}}$ contains a finitely supported measure $\nu$ that is shift-invariant. \end{prop} \begin{coolproof} Let $(\nu_n)_{n \in {\mathbb{N}}}$ be a sequence of finitely supported shift-invariant measures on $\Omega$ that converges to ${\bm{\lambda}}$. This gives us a sequence of actions of $\G$ on the finite sets $X_n \coloneqq \mathrm{supp}(\nu_n)$ and, since $\nu_n \to {\bm{\lambda}}$, each nonidentity group element $\gamma \in \G$ acts on $X_n$ nontrivially for all large enough $n$. This shows that $\G$ is residually finite. Conversely, suppose that $\G$ is residually finite and let $(\Delta_n)_{n \in {\mathbb{N}}}$ be a decreasing sequence of finite index subgroups of $\G$ with trivial intersection. For $k \in {\mathbb{N}}^+$, let \[ Q_k \coloneqq \set{0,\, 1/k,\, \ldots,\, (k-1)/k} \subset [0;1]. \] Let $P(k, n)$ denote the set of all maps $x \colon \G \to Q_k$ that are constant on the right cosets of $\Delta_n$. Then the set $P(k, n)$ is finite and shift-invariant, and, letting $\nu_{k, n}$ be the uniform probability measure on $P(k,n)$, we see that $\nu_{k, n} \to {\bm{\lambda}}$ as $k$, $n \to \infty$. \end{coolproof} Motivated by Proposition~\ref{prop:res_fin}, we say that a group $\G$ is \emphd{approximately residually finite} if for every open neighborhood $U$ of ${\bm{\lambda}}$, there is a finitely supported measure $\nu$ such that $\gamma \cdot \nu \in U$ for all $\gamma \in \G$. Proposition~\ref{prop:res_fin} implies that every residually finite group is approximately residually finite, so our terminology is consistent. An intuitive way of thinking about approximate residual finiteness is as follows: To show that a group $\G$ is approximately residually finite, we have to find finite subsets $X \subset \Omega$ that are ``almost uniformly distributed'' over the space $(\Omega, {\bm{\lambda}})$ and also remain such when shifted by any $\gamma \in \G$. We remark that a \emph{random} finite set $X$ fails to have this property: For any $n \in {\mathbb{N}}^+$, the product action $\sigma^n \colon \G \curvearrowright (\Omega^n, {\bm{\lambda}}^n)$ is ergodic, and hence if $x_1$, \ldots, $x_n \in \Omega$ are chosen randomly and independently from each other, then, with probability $1$, for every open $V \subseteq \Omega$ there is some $\gamma \in \G$ such that $\gamma \cdot x_1$, \ldots, $\gamma \cdot x_n \in V$. Nevertheless, we have the following: \begin{corl}[to Theorem~\ref{theo:ult_erg}]\label{corl:approx_res_fin} Every countable group is approximately residually finite. \end{corl} \begin{coolproof} Let $U$ be an open neighborhood of ${\bm{\lambda}}$. It suffices to exhibit a finitely supported measure $\nu \in U$ such that $\nu \cdot \gamma \in U$ for all $\gamma \in \G$, where the \emphd{right shift action} $\Omega \curvearrowleft \G$ is given by \[ (x \cdot \gamma)(\delta) \coloneqq x(\gamma\delta) \qquad \text{for all } x \in \Omega \text{ and } \gamma,\ \delta \in \G. \] Applying Theorem~\ref{theo:ult_erg} with $K = [0;1]$ and $f = (x \mapsto x(\mathbf{1}))$, we obtain a nonempty finite set $D \subset \G$ and a Borel map $g \colon \Omega \to [0;1]$ such that $\mathbb{M}_D \code{g}(x) \in U$ for ${\bm{\lambda}}$-a.e.\ $x \in \Omega$. Since $\G$ is countable and the measure ${\bm{\lambda}}$ is right-shift-invariant, there is $x \in \Omega$ such that $\mathbb{M}_D \code{g}(x \cdot \gamma) \in U$ for all $\gamma \in \G$. Set $\nu \coloneqq \mathbb{M}_D \code{g}(x)$. Then $\nu$ is finitely supported; furthermore, it is straightforward to verify, using the \ep{left-}equivariance of $\code{g}$ and the fact that the left and the right shift actions of $\G$ on $\Omega$ commute with each other, that $\nu \cdot \gamma = \mathbb{M}_D\code{g}(x \cdot \gamma)$ for all $\gamma \in \G$. Hence, $\nu$ is as desired. \end{coolproof} Since the above argument only involves the properties of $g$ on a countable subset of $\Omega$, Corollary~\ref{corl:approx_res_fin} can also be derived directly from the classical LLL, without using its measurable analogs. \subsection{Pointwise versions of the Ab\'ert--Weiss theorem}\label{subsec:AW} So far we have considered the action $\sigma \colon \G \curvearrowright (\Omega, {\bm{\lambda}})$ on its own. Now we would like to discuss the relationship between $\sigma$ and other actions of $\G$. The concepts of \emph{weak containment} and \emph{weak equivalence} of $\text{p.m.p.}$\xspace actions were introduced by Kechris in~\cite[\S10(C)]{K_book}. They are inspired by the analogous notions for unitary representations and are closely related to the so-called \emph{local\-/global convergence} in the theory of graph limits~\cite{LocalGlobal}. The relation of weak equivalence is much coarser than the isomorphism relation, which makes it relatively well-behaved. On the other hand, several interesting parameters associated with $\text{p.m.p.}$\xspace actions---such as their cost, type, etc.---turn out to be invariants of weak equivalence. Due to these favorable properties, the relations of weak containment and weak equivalence have attracted a considerable amount of attention in recent years. For a survey of the topic, see \cite{BK}. Roughly speaking, a $\text{p.m.p.}$\xspace action $\alpha \colon \G \curvearrowright (X, \mu)$ is weakly contained in another $\text{p.m.p.}$\xspace action $\beta \colon \G \curvearrowright (Y, \nu)$ if for every compact metric space $K$ and for any Borel map $f \colon X \to K$, the interaction of $f$ with $\alpha$ can be arbitrarily well ``simulated'' by a Borel map $g \colon Y \to K$ interacting with $\beta$. Here is a precise definition: \begin{defn}[{\textls{Weak containment}; {\cite[\S{}2.2(2)]{BK}}}] Let $\alpha \colon \G \curvearrowright (X, \mu)$ and $\beta \colon \G \curvearrowright (Y, \nu)$ be $\text{p.m.p.}$\xspace actions of $\G$. We say that $\alpha$ is \emphd{weakly contained} in $\beta$, in symbols $\alpha \preceq \beta$, if for any compact metric space $K$, a Borel function $f \colon X \to K$, and an open neighborhood $U$ of the measure $\mathbb{M}_\mu \code{f}$, there exists a Borel map $g \colon Y \to K$ such that $\mathbb{M}_\nu \code{g} \in U$. If both $\alpha \preceq \beta$ and $\beta \preceq \alpha$, then $\alpha$ and $\beta$ are said to be \emphd{weakly equivalent}, in symbols $\alpha \simeq \beta$. \end{defn} Weak containment can be defined in a number of equivalent ways, several of which can be found in~\cite[\S\S2.1, 2.2]{BK}. The characterization given above is due to Ab\'ert and Weiss~\cite[Lemma~8]{AW} \ep{see also \cite[Proposition~3.5]{T-D}}. We sometimes write $(\alpha, \mu) \preceq (\beta, \nu)$ instead of $\alpha \preceq \beta$ in order to emphasize the dependence of weak containment on the invariant measures $\mu$ and $\nu$. Burton~\cite[Corollary~4.2]{B} (see~\cite[Theorem~3.3]{BK}) proved that there exist continuum many distinct weak equivalence classes of \ep{not necessarily ergodic} $\text{p.m.p.}$\xspace actions of $\G$. Glasner, Thouvenot, and Weiss~\cite{GTW} and independently Greg Hjorth (unpublished) proved that the pre-order of weak containment has a \emph{maximum} element (see also~\cite[Theorem~3.1]{BK}). A complementary result of Ab\'ert and Weiss~\cite[Theorem~1]{AW} (see also~\cite[Theorem~3.5]{BK}) asserts that the shift action $\sigma \colon \G \curvearrowright (\Omega, {\bm{\lambda}})$ is \emph{minimum} among all $\text{p.m.p.}$\xspace actions $\alpha \colon \G \curvearrowright (X, \mu)$ that are \emphd{\ep{almost everywhere} free}, i.e., such that the $\alpha$-stabilizer of $\mu$-a.e.\ $x \in X$ is trivial: \begin{theo}[{Ab\'ert--Weiss~{\normalfont\cite[Theorem~1]{AW}}}]\label{theo:AW} Let $\alpha \colon \G \curvearrowright (X, \mu)$ be an almost everywhere free $\text{p.m.p.}$\xspace action of $\G$. Then $(\sigma, {\bm{\lambda}}) \preceq (\alpha, \mu)$; or, explicitly, the following statement holds: \smallskip Let $K$ be a compact metric space and let $f \colon \Omega \to K$ be a Borel function. Then, for any open neighborhood $U$ of the measure $\mathbb{M}_{\bm{\lambda}} \code{f}$, there is a Borel map $g \colon X \to K$ such that $\mathbb{M}_\mu \code{g} \in U$. \end{theo} We strengthen Theorem~\ref{theo:AW} by replacing the measure $\mathbb{M}_\mu \code{g}$ by its pointwise analogs of the form $\mathbb{M}_D \code{g} (x)$. Moreover, our result applies to actions that are not necessarily free but only ``close enough'' to being free. Specifically, for a set $S \subseteq \G$, we say that an action $\alpha \colon \G \curvearrowright X$ is \emphd{$S$-free} if for all $\gamma$, $\delta \in S$ and $x \in X$, $\gamma \cdot x = \delta \cdot x$ implies $\gamma = \delta$. \ep{Thus, ``free'' is the same as ``$\G$-free.''} Given a sequence of sets $S_1$, \ldots, $S_n \subseteq \G$, we say that $\alpha$ is \emphd{$(S_1, \ldots, S_n)$-free} if $\alpha$ is $S_i$-free for each $1 \leq i \leq n$. \begin{theo}[\textls{Pointwise Ab\'ert--Weiss}]\label{theo:main_delta} Let $K$ be a compact metric space and let $f \colon \Omega \to K$ be a Borel function. For any open neighborhood $U$ of the measure $\mathbb{M}_{\bm{\lambda}} \code{f}$, there exist $C > 0$ and a finite set $S \subset \G$ with the following property: \smallskip Let $D$ be a finite subset of $\G$ with $|D| \geq C$ and let $\alpha \colon \G \curvearrowright X$ be an $(S, D)$-free Borel action of $\G$. Then, for any $\mu \in \Prob(X)$ and $\delta > 0$, there is a Borel map $g \colon X \to K$ such that \[ \mu(\set{x \in X \,:\, \mathbb{M}_D \code{g} (x) \in U}) \,\geq\, 1 - \delta. \] \end{theo} \begin{remks}\label{remk:delta} Let us make a few comments about the statement of Theorem~\ref{theo:main_delta}. \begin{enumerate}[label={\ep{\roman*}},wide] \item To see that Theorem~\ref{theo:main_delta} is a strengthening of the \hyperref[theo:AW]{Ab\'ert--Weiss theorem}, let $\alpha \colon \G \curvearrowright (X, \mu)$ be a free $\text{p.m.p.}$\xspace action. Given a compact metric space $K$ and a Borel function $f \colon \Omega \to K$, we can apply Theorem~\ref{theo:main_delta} to obtain a finite set $D \subset \G$ and a Borel map $g \colon X \to K$ such that the pushforward measure $\mathbb{M}_D \code{g} (x)$ is arbitrarily close to $\mathbb{M}_{\bm{\lambda}} \code{f}$, for all points $x \in X$ away from a set of arbitrarily small measure. The $\alpha$-invariance of $\mu$ yields \[ \mu = \int_X \upsilon_{x,D} \,\mathrm{d} \mu(x), \quad \text{hence} \quad \mathbb{M}_\mu \code{g} = \int_X \mathbb{M}_D \code{g} (x) \,\mathrm{d} \mu(x), \] and thus $\mathbb{M}_\mu \code{g}$ is also close to $\mathbb{M}_{\bm{\lambda}} \code{f}$, as desired. \smallskip \item The measure $\mu$ in Theorem~\ref{theo:main_delta} is not required to be $\alpha$-invariant \ep{or even $\alpha$-quasi-invariant} and is only used to bound the set of all $x \in X$ with $\mathbb{M}_D \code{g} (x) \not \in U$. \smallskip \item We emphasize that the averaging set $D$ in Theorem~\ref{theo:main_delta} is independent of the choice of $\delta > 0$; that is what makes this result particularly interesting. It is possible that the conclusion of Theorem~\ref{theo:main_delta} also holds with $\delta = 0$, but we do not know how to prove \ep{or disprove} that in general; see Problem~\ref{prob:delta_zero} in Section \ref{sec:remks}. \ep{However, we can make $\delta$ be zero under some additional assumptions---see \ref{item:shift} and Theorem~\ref{theo:main_Borel} below.} \smallskip \item\label{item:shift} In contrast to the \hyperref[theo:AW]{Ab\'ert--Weiss theorem}, the conclusion of Theorem~\ref{theo:main_delta} is nontrivial even if $(X, \mu) = (\Omega, {\bm{\lambda}})$ and $\alpha = \sigma$. This case, however, is already covered by the ergodic Theorem~\ref{theo:ult_erg}, in fact even with $\delta = 0$. \smallskip \item\label{item:derivation} For actions $\alpha$ that are free and measure\-/preserving, Theorem~\ref{theo:main_delta} follows relatively straightforwardly by combining Theorem~\ref{theo:ult_erg} with the usual \hyperref[theo:AW]{Ab\'ert--Weiss theorem}. We sketch the argument here. Let $\alpha \colon \G \curvearrowright (X, \mu)$ be a free $\text{p.m.p.}$\xspace action. Let $K$ be a compact metric space and let $f \colon \Omega \to K$ be a Borel function. Fix an open neighborhood $U$ of the measure $\mathbb{M}_{\bm{\lambda}} \code{f}$. By Theorem~\ref{theo:ult_erg}, for any sufficiently large finite set $D \subset \G$, there is a Borel map $h \colon \Omega \to K$ with \begin{equation}\label{eq:h} \mathbb{M}_D \code{h} (x) \in U, \qquad \text{for ${\bm{\lambda}}$-a.e.}\ x \in \Omega. \end{equation} The equivariance of $\code{h}$ yields $\mathbb{M}_D \code{h}(x) = (\code{h})_\ast(\upsilon_{x, D}) = \upsilon_{\code{h}(x), D}$, and hence \eqref{eq:h} is equivalent to \[ \upsilon_{\kappa, D} \in U, \qquad \text{for $\mathbb{M}_{\bm{\lambda}} \code{h}$-a.e.}\ \kappa \in K^\G. \] Now we can use the \hyperref[theo:AW]{Ab\'ert--Weiss theorem} to obtain a Borel map $g \colon X \to K$ for which the pushforward measure $\mathbb{M}_\mu \code{g}$ is so close to $\mathbb{M}_{\bm{\lambda}} \code{h}$ that \[ \mu(\set{x \in X \,:\, \mathbb{M}_D \code{g} (x) \in U}) \,=\, \mathbb{M}_\mu \code{g} (\set{\kappa \in K^\G \,:\, \upsilon_{\kappa, D} \in U}) \,\geq\, 1-\delta, \] for any given $\delta > 0$, as desired. For non-free actions $\alpha$, a different, more direct proof is necessary. \smallskip \item The results of \S\ref{subsec:erg} apply to an infinite sequence of averaging sets $(D_n)_{n \in {\mathbb{N}}}$, while in Theorem~\ref{theo:main_delta} we only consider a single set $D$. Our approach can be routinely adapted to extend Theorem~\ref{theo:main_delta} to the case of finitely many averaging sets; however, when the family of averaging sets is infinite, our methods are not applicable---see Remark \ref{remk:amenable}. \end{enumerate} \end{remks} Notice that the pointwise operator $\mathbb{M}_D$ is well-defined for an arbitrary Borel action $\alpha \colon \G \curvearrowright X$ and does not require fixing a probability measure $\mu$ on $X$. Therefore, it makes sense to ask for a \emph{purely Borel} version of the \hyperref[theo:AW]{Ab\'ert--Weiss theorem}, with the last line of Theorem~\ref{theo:main_delta} replaced by \[ \text{$\mathbb{M}_D \code{g} (x) \in U$, \qquad \textsl{for all} $x \in X$.} \] Here we establish such a version for finitely generated groups of subexponential growth and, more generally, for uniformly subexponential Borel actions. Let $\alpha \colon \G \curvearrowright X$ be a Borel action of $\G$. We say that $\alpha$ is \emphd{uniformly subexponential} if for every finite set $S \subset \G$ and for all $\epsilon > 0$, there is $n_0 \in {\mathbb{N}}$ such that for all $n \geq n_0$ and for all $x \in X$, $\left|S^n \cdot x\right| \leq (1 + \epsilon)^n$, where $S^n \coloneqq \set{\gamma_1 \cdots \gamma_n \,:\, \gamma_i \in S \text{ for all }1 \leq i \leq n}$. For example, if $\G$ is a finitely generated group of subexponential growth, then every action of $\G$ is uniformly subexponential. \begin{theo}[\textls{Borel Ab\'ert--Weiss for uniformly subexponential actions}]\label{theo:main_Borel} Let $K$ be a compact metric space and let $f \colon \Omega \to K$ be a Borel function. For any open neighborhood $U$ of the measure $\mathbb{M}_{\bm{\lambda}} \code{f}$, there exist $C > 0$ and a finite set $S \subset \G$ with the following property: \smallskip Let $D$ be a finite subset of $\G$ with $|D| \geq C$ and let $\alpha \colon \G \curvearrowright X$ be a uniformly subexponential $(S, D)$-free Borel action of $\G$. Then there is a Borel map $g \colon X \to K$ such that \[\mathbb{M}_D \code{g} (x) \in U, \qquad \text{for all}\ x \in X.\] \end{theo} Note that, even though groups of subexponential growth are amenable, the averaging set $D$ in the statement of Theorem~\ref{theo:main_Borel} is {not} assumed to be a F\o{}lner set. \subsection{Outline of the remainder of the paper} This paper is organized as follows. Section~\ref{sec:prelim} contains a few definitions and some preliminary results concerning the continuity of various basic operations, such as $f \mapsto f_\ast$. We commence the proofs of Theorems~\ref{theo:weak_erg}, \ref{theo:ult_erg}, \ref{theo:main_delta}, and \ref{theo:main_Borel} in Section~\ref{sec:red}, where they are reduced to their special cases with a more ``combinatorial'' flavor. Then, in Section~\ref{sec:concentration}, we state and prove a certain concentration of measure inequality. At this point, we already have all the tools needed to derive Theorem~\ref{theo:weak_erg}, which is done in \S\ref{subsec:weak_erg}. In Section~\ref{sec:LLL} we review the LLL and its measurable analogs, and in Section~\ref{sec:proofs} we complete the proofs of Theorems \ref{theo:ult_erg}, \ref{theo:main_delta} and \ref{theo:main_Borel}. It turns out that in order to prove Theorem~\ref{theo:ult_erg}, it is not enough to simply apply a known measurable version of the LLL---we actually have to go through the \emph{proof} of one of them to obtain some additional information; this is done in \S\ref{subsec:ult_erg}. We conclude the paper with some open problems in Section \ref{sec:remks}. The \hyperref[sec:app]{appendix} contains a proof of Theorem~\ref{theo:bad}. \section{Preliminaries}\label{sec:prelim} \subsection{Further notation} \subsubsection*{Integers} We use ${\mathbb{N}}$ to denote the set of all nonnegative integers and identify each $k \in {\mathbb{N}}$ with the $k$-element set $\set{i \in {\mathbb{N}} \,:\, i < k}$. Let ${\mathbb{N}}^+ \coloneqq {\mathbb{N}} \setminus \set{0}$. All finite sets \ep{including each $k \in {\mathbb{N}}$} are assumed to carry discrete topologies. \subsubsection*{Sets and functions} Each function $f$ is identified with its graph, i.e., the set $\set{(x, y) \,:\, y = f(x)}$. This enables the use of set-theoretic notation, such as $\subseteq$, $|\cdot|$, etc., for functions. For a function $f$ and a set $S$ of its domain, $\rest{f}{S}$ denotes the restriction of $f$ to $S$. For sets $A$ and $B$, \[ \begin{array}{rll} \text{--} & \fins{B} & \text{denotes the set of all finite subsets of $B$;}\\ \text{--} & [B \to A] & \text{denotes the set of all partial functions $\phi \colon B \rightharpoonup A$;}\\ \text{--} & \finf{B}{A} & \text{denotes the set of all partial functions $\phi \colon B \rightharpoonup A$ with $\mathrm{dom}(\phi) \in \fins{B}$.} \end{array} \] The identity function $X \to X$ on a set $X$ is denoted by $\operatorname{id}_X$. \subsubsection*{Symbolic dynamics} Let $A$ be a set and let $\alpha \colon \G \curvearrowright X$ be an action of $\G$. We extend the definition of the coding map to partial functions $f \colon X \rightharpoonup A$ by letting $\pi_f(x) \colon \G \rightharpoonup A$ be given by \[ \pi_f(x)(\gamma) \coloneqq \begin{cases} f(\gamma \cdot x) &\text{if } \gamma \cdot x \in \mathrm{dom}(f);\\ \text{undefined} &\text{otherwise}, \end{cases} \qquad \text{for all}\ x \in X \text{ and } \gamma \in \G. \] We similarly extend the shift action $\sigma_A \colon \G \curvearrowright A^\G$ to an action $\G \curvearrowright [\G \to A]$ in the obvious way. \subsubsection*{The free part of an action} For an action $\alpha \colon \G \curvearrowright X$ of $\G$, let $\operatorname{Free}(X) \subseteq X$ denote the set of all $x \in X$ whose $\alpha$-stabilizer is trivial and let $\operatorname{Free}(\alpha) \colon \G \curvearrowright \operatorname{Free}(X)$ denote the induced action of $\G$ on $\operatorname{Free}(X)$; we call $\operatorname{Free}(\alpha)$ the \emphd{free part} of $\alpha$. \subsubsection*{Miscellaneous} For a metric space $(K, {\mathfrak{d}})$, $a$, $b \in K$, and $\epsilon > 0$, we write $a \approx_\epsilon b$ to mean ${\mathfrak{d}}(a,b) < \epsilon$. \subsection{Topological preliminaries} \subsubsection*{Continuity of the coding map} Fix an arbitrary enumeration $\set{\gamma_n}_{n \in {\mathbb{N}}}$ of the elements of $\G$. If $(K, {\mathfrak{d}})$ is a compact metric space, then the product topology on $K^\G$ is induced by the metric $\hat{{\mathfrak{d}}}$: \[ \hat{{\mathfrak{d}}}(\kappa, \eta) \coloneqq \sum_{n=0}^\infty \frac{{\mathfrak{d}}(\kappa(\gamma_n), \eta(\gamma_n))}{2^{n+1}}. \] Recall that if $X$ is a standard Borel space and $\mu \in \Prob(X)$, then the space $\mathfrak{B}(X, K)$ is endowed with the pseudometric ${\mathfrak{d}}_\mu$. Additionally, we shall consider the uniform metric ${\mathfrak{d}}_{\mathsf{uni}}$ given by \[ {\mathfrak{d}}_{\mathsf{uni}} (f, g) \coloneqq \sup_{x \in X} {\mathfrak{d}} (f(x), g(x)). \] \begin{lemma}\label{lemma:cont_of_code} Let $(K, {\mathfrak{d}})$ be a compact metric space and let $\alpha \colon \G \curvearrowright X$ be a Borel action of $\G$. \begin{enumerate}[label=\ep{\normalfont\alph*},wide] \item\label{item:mu} If $\mu \in \Prob(X)$ is $\alpha$-invariant, then the function $(\mathfrak{B}(X, K), {\mathfrak{d}}_\mu) \to (\mathfrak{B}(X, K^\G), \hat{{\mathfrak{d}}}_\mu) \colon f \mapsto \pi_f$ is distance\-/preserving, hence continuous. \smallskip \item\label{item:uni} The function $(\mathfrak{B}(X, K), {\mathfrak{d}}_{\mathsf{uni}}) \to (\mathfrak{B}(X, K^\G), \hat{{\mathfrak{d}}}_{\mathsf{uni}}) \colon f \mapsto \pi_f $ is $1$-Lipschitz, hence continuous. \end{enumerate} \end{lemma} \begin{coolproof} \ref{item:mu} For all $f$, $g \in \mathfrak{B}(X, K)$, we have \[ \hat{{\mathfrak{d}}}_\mu(\code{f}, \code{g}) \,=\, \int_X \hat{{\mathfrak{d}}}(\code{f}(x), \code{g}(x)) \,\mathrm{d}\mu(x) \,=\, \int_X \sum_{n=0}^\infty \frac{{\mathfrak{d}}(f(\gamma_n \cdot x), g(\gamma_n \cdot x))}{2^{n+1}} \,\mathrm{d}\mu(x). \] Switching the order of integration and summation, we rewrite the last expression as \[ \sum_{n=0}^\infty \frac{1}{2^{n+1}}\int_X {\mathfrak{d}}(f(\gamma_n \cdot x), g(\gamma_n \cdot x)) \,\mathrm{d}\mu(x). \] Since $\mu$ is $\alpha$-invariant, this is equal to \[ \sum_{n=0}^\infty \frac{1}{2^{n+1}} \int_X {\mathfrak{d}}(f(x), g(x)) \,\mathrm{d} \mu(x) \,=\, \sum_{n=0}^\infty\frac{{\mathfrak{d}}_\mu(f, g)}{2^{n+1}} \,=\, {\mathfrak{d}}_\mu(f, g). \] \ref{item:uni} For all $f$, $g \in \mathfrak{B}(X, K)$ and $x \in X$, we have \[ \hat{{\mathfrak{d}}}(\code{f}(x), \code{g}(x)) \,=\, \sum_{n = 0}^\infty \frac{{\mathfrak{d}}(f(\gamma_n \cdot x), g(\gamma_n \cdot x))}{2^{n+1}} \,\leq\, \sum_{n = 0}^\infty \frac{{\mathfrak{d}}_{\mathsf{uni}}(f, g)}{2^{n+1}} \,=\, {\mathfrak{d}}_{\mathsf{uni}}(f, g), \] and the desired conclusion follows. \end{coolproof} \subsubsection*{Continuity of the pushforward operator} For a Polish space $X$, let $C_\mathsf{b}(X)$ denote the set of all bounded continuous real-valued functions on $X$. By definition, the weak-$\ast$ topology on $\Prob(X)$ is generated by the maps $\Prob(X) \to \mathbb{R} \colon \mu \mapsto \int \xi \,\mathrm{d} \mu$, where $\xi \in C_\mathsf{b}(X)$. \begin{lemma}\label{lemma:push_of_cont} Let $X$ and $K$ be Polish spaces and let $f \colon X \to K$ be continuous. Then $f_\ast \colon \Prob(X) \to \Prob(K)$ is also continuous. \hfill \qedsymbol \end{lemma} Now we turn to the continuity properties of the mapping $f \mapsto f_\ast$. \begin{lemma}\label{lemma:push_mu} Let $(K, {\mathfrak{d}})$ be a compact metric space and let $(X, \mu)$ be a standard probability space. Then the function $ (\mathfrak{B}(X, K), {{\mathfrak{d}}}_\mu) \to \Prob(K) \colon f \mapsto f_\ast(\mu) $ is continuous. \end{lemma} \begin{coolproof} Let $f$, $f_0$, $f_1$, \ldots\ $\in \mathfrak{B}(X, K)$ be such that $f_n \to f$ in $(\mathfrak{B}(X, K), {\mathfrak{d}}_\mu)$. To demonstrate that $(f_n)_\ast(\mu) \to f_\ast(\mu)$, let $\xi \in C_\mathsf{b}(X)$; we have to show that \begin{equation}\label{eq:xi} \int_K \xi \,\mathrm{d}(f_n)_\ast(\mu) \,\to\, \int_K \xi \,\mathrm{d} f_\ast(\mu). \end{equation} We may scale $\xi$ if necessary to make it bounded in absolute value by $1$. Take any $\epsilon > 0$. Since $K$ is compact, $\xi$ is uniformly continuous, so we can let $\delta > 0$ be such that $\xi(a) \approx_\epsilon \xi(b)$ whenever $a \approx_\delta b$. For $n \in {\mathbb{N}}$, let $X_n$ denote the set of all $x \in X$ with ${\mathfrak{d}}(f_n(x), f(x)) < \delta$. Since ${\mathfrak{d}}_\mu(f_n, f) \to 0$, we have $\mu(X_n) \to 1$, and hence, for all large enough $n \in {\mathbb{N}}$, \[ \left|\int_K \xi \,\mathrm{d}(f_n)_\ast(\mu) \,-\, \int_K \xi \,\mathrm{d} f_\ast(\mu)\right| \,\leq\, \int_X |\xi \circ f_n \,-\, \xi \circ f|\,\mathrm{d}\mu \,\leq\, \epsilon \mu(X_n) \,+\, 2(1 - \mu(X_n)) \,\leq\, 2\epsilon. \] Since $\epsilon$ was chosen arbitrarily, \eqref{eq:xi} follows. \end{coolproof} If $K$ is a compact metric space, then the space $C_\mathsf{b}(K)$, equipped with the uniform norm, is separable. Therefore, there exists a countable set $\set{\xi_n}_{n \in {\mathbb{N}}}$ of continuous real-valued functions on $K$ bounded in absolute value by $1$ such that $\set{a\xi_n \,:\, a \in \mathbb{R},\ n\in {\mathbb{N}}}$ is a dense subset of $C_\mathsf{b}(K)$. With this choice of $\set{\xi_n}_{n \in {\mathbb{N}}}$, the topology on $\Prob(K)$ is induced by the metric $\Delta^K$: \[ \Delta^K(\mu, \nu) \coloneqq \sum_{n = 0}^\infty \frac{\left|\int_K \xi_n \,\mathrm{d} \mu \,-\, \int_K \xi_n \,\mathrm{d} \nu\right|}{2^{n+1}}. \] \begin{lemma}\label{lemma:push_uni} Let $X$ be a Polish space and let $(K, {\mathfrak{d}})$ be a compact metric space. Then the map $(\mathfrak{B}(X, K), {\mathfrak{d}}_{\mathsf{uni}}) \to (\mathfrak{B}(\Prob(X), \Prob(K)), \Delta^K_\mathsf{uni}) \colon f \mapsto f_\ast$ is continuous. \end{lemma} \begin{coolproof} Let $\set{\xi_n}_{n \in {\mathbb{N}}}$ be the set of functions used to define $\Delta^K$. Take any $N \in {\mathbb{N}}^+$ and $\epsilon > 0$. Since $K$ is compact, each $\xi_n$ is uniformly continuous, hence we can choose $\delta > 0$ so that $\xi_n(a) \approx_\epsilon \xi_n(b)$ for all $n \leq N$, whenever $a \approx_\delta b$. Let $f$, $g \in \mathfrak{B}(X, K)$ and suppose that ${\mathfrak{d}}_\mathsf{uni}(f, g) < \delta$. Then, for any $\mu \in \Prob(K)$, we have \begin{align*} \Delta^K(f_\ast(\mu), g_\ast(\mu)) \,&=\, \sum_{n = 0}^\infty \frac{\left|\int_K \xi_n \,\mathrm{d} f_\ast(\mu) \,-\, \int_K \xi_n \,\mathrm{d} g_\ast(\mu)\right|}{2^{n+1}} \\ &\leq\, \sum_{n = 0}^{N} \frac{1}{2^{n+1}}\int_X \left|\xi_n \circ f - \xi_n \circ g\right| \,\mathrm{d} \mu \,+\, \frac{1}{2^{N-1}} \,<\, \epsilon + \frac{1}{2^{N-1}}. \end{align*} Hence, $\Delta^K_\mathsf{uni}(f_\ast, g_\ast) < \epsilon + 2^{-N+1}$. Since $\epsilon$ and $N$ are arbitrary, this completes the proof. \end{coolproof} \subsubsection*{Density of continuous functions} Recall that a topological space $X$ is \emphd{zero-dimensional} if it has a basis consisting of clopen sets. \begin{lemma}\label{lemma:zero-dim} Let $X$ be a zero-dimensional Polish space and let $(K, {\mathfrak{d}})$ be a compact metric space. If $\mu \in \Prob(X)$, then the set of all continuous maps $f \colon X \to K$ is dense in $(\mathfrak{B}(X, K), {\mathfrak{d}}_\mu)$. \end{lemma} \begin{coolproof} Without loss of generality, assume that the metric ${\mathfrak{d}}$ is bounded by $1$. Let $f \in \mathfrak{B}(X, K)$ and $\epsilon > 0$. Since $K$ is compact, it contains a finite $\epsilon$-net $Z = \set{z_0, \ldots, z_{n-1}} \subseteq K$. Let $g \colon X \to Z$ be the map that sends each $x \in X$ to the point $z \in Z$ that is closest to $f(x)$ \ep{ties may be broken arbitrarily}. Then ${\mathfrak{d}}_\mathsf{uni}(f, g) < \epsilon$ by construction. Since the measure $\mu$ is regular \cite[Theorem 17.10]{K_DST} and the space $X$ is zero-dimensional, for each $0 \leq i < n$, there is a clopen set $U_i \subseteq X$ such that $\mu(U_i \bigtriangleup g^{-1}(z_i)) < \epsilon/n$. For every $x \in X$, set $h(x) \coloneqq z_i$ if $x \in V_i \coloneqq U_i \setminus (U_0 \cup \ldots \cup U_{i-1})$ for some $0 \leq i < n$, and $h(x) \coloneqq z_0$ if $x \in V \coloneqq X \setminus (U_0 \cup \ldots \cup U_{n-1})$. Since the sets $V_0$, \ldots, $V_{n-1}$, $V$ are clopen, the map $h$ is continuous. If $h(x) \neq g(x)$, then either $x \in V_i \setminus g^{-1}(z_i)$ for some $0 \leq i < n$, in which case $x \in U_i \setminus g^{-1}(z_i)$; or else, $x \in V \setminus g^{-1}(z_0)$, in which case $x \in g^{-1}(z_i) \setminus U_i$ for some $1 \leq i< n$. Since the metric ${\mathfrak{d}}$ is bounded by $1$, we conclude that \[ {\mathfrak{d}}_\mu(g, h) \,\leq\, \sum_{i = 0}^{n-1} \mu(U_i \setminus g^{-1}(z_i)) \,+\, \sum_{i=1}^{n-1} \mu(g^{-1}(z_i) \setminus U_i) \,\leq\, \sum_{i=0}^{n-1} \mu(U_i \bigtriangleup g^{-1}(z_i)) \,<\, \epsilon. \] Therefore, we have found a continuous function $h \colon X \to K$ with ${\mathfrak{d}}_\mu(f, h) < 2\epsilon$. As $\epsilon$ is arbitrary, the proof is complete. \end{coolproof} \section{Combinatorial reductions}\label{sec:red} For $k \in {\mathbb{N}}^+$, let $u_k$ denote the uniform probability measure on $k$, i.e., let $u_k(\set{i}) \coloneqq 1/k$ for all $i < k$. Set $\Omega_k \coloneqq k^\G$ and $\bm{u}_k \coloneqq u_k^\G$. Recall that the space $(\Omega_k, \bm{u}_k)$ is equipped with the shift action $\sigma_k$. Given $\phi \in \finf{\G}{k}$ and a partial map $c \colon \G \rightharpoonup k$, we say that $\gamma \in \G$ is an \emphd{occurrence} of $\phi$ in $c$ if $\gamma \cdot c \supseteq \phi$. The set of all {occurrences} of $\phi$ in $c$ is denoted by $\mathcal{O}_\phi(c)$. By definition, if $\gamma \in \mathcal{O}_\phi(c)$, then, in particular, $\mathrm{dom}(\phi) \gamma \subseteq \mathrm{dom}(c)$. Define \[ \Omega_k(\phi) \coloneqq \set{c \in \Omega_k \,:\, \mathbf{1} \in \mathcal{O}_\phi(c)} = \set{c \in \Omega_k \,:\, \phi \subset c}. \] Note that $\bm{u}_k(\Omega_k(\phi)) = k^{-|\phi|}$ \ep{where $|\phi|$ is the cardinality of the domain of $\phi$}. The family of sets $\set{\Omega_k(\phi) \,:\, \phi \in \finf{\G}{k}}$ forms a basis for the topology on $\Omega_k$ consisting of clopen sets. From this fact and \cite[Theorem~17.20]{K_DST}, we obtain the following: \begin{lemma}\label{lemma:basic} Let $k \in {\mathbb{N}}^+$ and $\mu$, $\mu_0$, $\mu_1$, \ldots\ $\in \Prob(\Omega_k)$. Then $\lim_{n \to \infty}\mu_n = \mu$ if and only if, for all $\phi \in \finf{\G}{k}$, we have $\lim_{n \to \infty} \mu_n(\Omega_k(\phi)) \,=\, \mu(\Omega_k(\phi))$. \hfill \qedsymbol \end{lemma} We also consider the space $\tilde{\Omega}_k \coloneqq (k^{\mathbb{N}})^\G$, equipped with the product measure $\tilde{\bm{u}}_k \coloneqq (u_k^{\mathbb{N}})^\G$ and the shift action $\sigma_{k^{\mathbb{N}}}$ of $\G$. To simplify the notation, given $x \in \tilde{\Omega}_k$, $\gamma \in \G$, and $n \in {\mathbb{N}}$, we write $x(\gamma, n)$ to mean $x(\gamma)(n)$ \ep{however, $x(\gamma)$ still denotes the corresponding element of $k^{\mathbb{N}}$}. If $k \geq 2$, then, by the {measure isomorphism theorem} \cite[Theorem~17.41]{K_DST}, the standard probability spaces $([0;1], \lambda)$ and $(k^{\mathbb{N}}, u_k^{\mathbb{N}})$ are Borel isomorphic, which allows us to replace $\sigma \colon \G \curvearrowright (\Omega, {\bm{\lambda}})$ by $\sigma_{k^{\mathbb{N}}} \colon \G \curvearrowright (\tilde{\Omega}_k, \tilde{\bm{u}}_k)$ in the statements of Theorems~\ref{theo:ult_erg}, \ref{theo:main_delta}, and \ref{theo:main_Borel}. This gives us two main advantages. First, the space $\tilde{\Omega}_k$ is zero\=/dimensional; in particular, Lemma~\ref{lemma:zero-dim} applies to it. Second, the structure of $\tilde{\Omega}_k$ will be explicitly used in the proof of Theorem~\ref{theo:ult_erg} presented in \S\ref{subsec:ult_erg}. \subsection{Reduction for Theorem~\ref{theo:weak_erg}} In this subsection we reduce Theorem~\ref{theo:weak_erg} to the following statement: \begin{theobis}{theo:weak_erg}\label{theo:weak_erg_bis} Let $k \in {\mathbb{N}}^+$ and let $(D_n)_{n \in {\mathbb{N}}}$ be a sequence of finite subsets of $\G$ with $|D_n|/\log n \to\infty$. Then, for all $S \in \fins{\G}$ and $\phi \colon S \to k$, we have \[ \lim_{n \to \infty} \frac{|D_n \cap \mathcal{O}_\phi(c)|}{|D_n|} \,=\, \frac{1}{k^{|S|}}, \qquad \text{for $\bm{u}_k$-a.e.}\ c \in \Omega_k. \] \end{theobis} \begin{lemma}\label{lemma:red1} Theorem \ref{theo:weak_erg_bis} implies Theorem~\ref{theo:weak_erg}. \end{lemma} \begin{coolproof} \stepcounter{ForClaims} \renewcommand{\theForClaims}{\ref{lemma:red1}} Assume Theorem~\ref{theo:weak_erg_bis}. Fix a sequence $(D_n)_{n \in {\mathbb{N}}}$ of nonempty finite subsets of $\G$ such that $|D_n|/\log n \to \infty$. Notice that Theorem~\ref{theo:weak_erg} is equivalent to the following assertion: \begin{equation}\label{eq:weak_erg} \lim_{n \to \infty} \upsilon_{x, D_n} = {\bm{\lambda}}, \qquad \text{for ${\bm{\lambda}}$-a.e.}\ x \in \Omega. \end{equation} On the other hand, by Lemma~\ref{lemma:basic}, the conclusion of Theorem~\ref{theo:weak_erg_bis} is equivalent to \begin{equation}\label{eq:assump} \lim_{n \to \infty} \upsilon_{c, D_n} \,=\, \bm{u}_k, \qquad \text{for $\bm{u}_k$-a.e.}\ c \in \Omega_k. \end{equation} \begin{claim}\label{claim:factor} If $\pi \colon (\Omega, {\bm{\lambda}}) \to (\Omega_k, \bm{u}_k)$ is a factor map, then \[ \lim_{n \to \infty} \mathbb{M}_{D_n} \pi(x) \,=\, \mathbb{M}_{\bm{\lambda}} \pi \,=\, \bm{u}_k, \qquad \text{for ${\bm{\lambda}}$-a.e.}\ x \in \Omega. \] \end{claim} \begin{claimproof} From the equivariance of $\pi$, it follows that for all $x \in \Omega$ and $D \in \fins{\G} \setminus \set{\varnothing}$, \[ \mathbb{M}_D \pi(x) \,=\, \pi_\ast(\upsilon_{x, D}) \,=\, \upsilon_{\pi(x), D}. \] Using \eqref{eq:assump} and the fact that, since $\pi$ is a factor map, $\mathbb{M}_{\bm{\lambda}}(\pi) = \pi_\ast({\bm{\lambda}}) = \bm{u}_k$, we conclude that \[ \mathbb{M}_{D_n} \pi(x) \,=\, \upsilon_{\pi(x), D_n} \,\xrightarrow[n \to \infty]{} \bm{u}_k, \qquad \text{for ${\bm{\lambda}}$-a.e.}\ x \in \Omega. \qedhere \] \end{claimproof} Define a function $\mathsf{p} \colon \Omega \to [0;1]$ by $\mathsf{p}(x) \coloneqq x(\mathbf{1})$. Notice that $\code{\mathsf{p}} = \operatorname{id}_\Omega$. For each $k \in {\mathbb{N}}^+$, let $Q_k \subset [0;1]$ be the set of all fractions of the form $i/k$, $0 \leq i < k$, and define $f_k \colon \Omega \to Q_k$ by \[ f_k(x) \coloneqq \max \set{q \in Q_k \,:\, q \leq \mathsf{p}(x)}. \] Let $\pi_k \coloneqq \code{f_k}$. By definition, $f_k(x) \approx_{1/k} \mathsf{p}(x)$ for all $x \in \Omega$; in other words, the sequence $(f_k)_{k \in {\mathbb{N}}}$ converges to $\mathsf{p}$ uniformly. By Lemmas~\ref{lemma:cont_of_code}\ref{item:uni} and \ref{lemma:push_uni}, this implies that \[ \pi_k \to \operatorname{id}_\Omega \quad \text{and} \quad (\pi_k)_\ast \to \operatorname{id}_{\Prob(\Omega)} \qquad \text{uniformly}. \] By construction, $(f_k)_\ast({\bm{\lambda}})$ is the uniform probability measure on $Q_k$, and $(\pi_k)_\ast({\bm{\lambda}})$ is the corresponding product measure on $Q_k^\G$. Thus, we may apply Claim~\ref{claim:factor} to $\pi_k$ and conclude that \[ \lim_{n \to \infty} \mathbb{M}_{D_n} \pi_k(x) \,=\, \mathbb{M}_{\bm{\lambda}} \pi_k, \qquad \text{for ${\bm{\lambda}}$-a.e.}\ x \in \Omega. \] We can put all of these facts together as follows: \begin{figure}[H] \begin{tikzpicture} \node (a) at (0,0) {$\mathbb{M}_{D_n} \pi_k (x)$}; \node (b) at (0,-2.3) {$\mathbb{M}_{\bm{\lambda}} \pi_k$}; \node (c) at (5,0) {$\upsilon_{x, D_n}$}; \node (d) at (5,-2.3) {${\bm{\lambda}}$}; \draw[->] (a) -- node[midway, anchor=south, rotate=-90] {$n \to \infty$} (b); \draw[->] (b) -- node[midway, anchor=south] {$k \to \infty$} (d); \draw[->] (a) -- node[midway, anchor=south] {$k \to \infty$} node[midway, anchor=north] {uniformly in $n$} (c); \end{tikzpicture} \end{figure} \noindent It is clear from the above diagram that $\upsilon_{x, D_n}$ converges to ${\bm{\lambda}}$ as $n \to \infty$, proving \eqref{eq:weak_erg}. \end{coolproof} \subsection{Reductions for Theorems~\ref{theo:ult_erg}, \ref{theo:main_delta}, and \ref{theo:main_Borel}} Theorem~\ref{theo:ult_erg} reduces to the following statement: \begin{theobis}{theo:ult_erg}\label{theo:ult_erg_bis} For all $k \in {\mathbb{N}}^+$, $S \in \fins{\G}$, and $\epsilon > 0$, there is $C > 0$ with the following property: \smallskip Let $(D_n)_{n \in {\mathbb{N}}}$ be a sequence of finite subsets of $\G$ with $|D_n| \geq C \log(n+2)$ for all $n \in {\mathbb{N}}$. Then there exists a Borel map $g \colon \tilde{\Omega}_k \to k$ such that \[ \tilde{\bm{u}}_k(\set{x \in \tilde{\Omega}_k \,:\, g(x) \neq x(\mathbf{1}, 0)}) \,\leq\, \epsilon, \] and, for all $\phi \colon S \to k$, we have \[ \frac{|D_n \cap \mathcal{O}_\phi(\code{g}(x))|}{|D_n|}\,\approx_\epsilon\, \frac{1}{k^{|S|}}, \qquad \text{for all $n \in {\mathbb{N}}$ and for $\tilde{\bm{u}}_k$-a.e.}\ x \in \tilde{\Omega}_k. \] \end{theobis} \begin{lemma}\label{lemma:red2} Theorem~\ref{theo:ult_erg_bis} implies Theorem~\ref{theo:ult_erg}. \end{lemma} \begin{coolproof}\stepcounter{ForClaims} \renewcommand{\theForClaims}{\ref{lemma:red2}} Assume Theorem~\ref{theo:ult_erg_bis}. Taking advantage of the measure isomorphism theorem, we will prove the statement of Theorem~\ref{theo:ult_erg} with $(\tilde{\Omega}_2, \tilde{\bm{u}}_2)$ in place of $(\Omega, {\bm{\lambda}})$. We equip the Cantor space $2^{\mathbb{N}}$ with the metric $\mathfrak{m}$ given by \[ \mathfrak{m}(a, b) \coloneqq \sum_{n = 0}^\infty \frac{\mathbbm{1}_{a(n) \neq b(n)}}{2^{n+1}}. \] Define $\mathsf{p} \colon \tilde{\Omega}_2 \to 2^{\mathbb{N}}$ by $\mathsf{p}(x) \coloneqq x(\mathbf{1})$. Note that $\code{\mathsf{p}} = \operatorname{id}_{\tilde{\Omega}_2}$ and $\mathbb{M}_{\tilde{\bm{u}}_2} \code{\mathsf{p}} = \tilde{\bm{u}}_2$. \begin{claim} It suffices to prove Theorem~\ref{theo:ult_erg} with $K = 2^{\mathbb{N}}$ and $f = \mathsf{p}$. \end{claim} \begin{claimproof} Let $(K, {\mathfrak{d}})$ be a compact metric space. Without loss of generality, assume that the metric ${\mathfrak{d}}$ is bounded by $1$. Fix a Borel function $f \colon \tilde{\Omega}_2 \to K$, $\epsilon > 0$, and an open neighborhood $U$ of $\mathbb{M}_{\tilde{\bm{u}}_2}\code{f}$. The space $\tilde{\Omega}_2$ is zero-dimensional, so Lemmas \ref{lemma:zero-dim}, \ref{lemma:cont_of_code}\ref{item:mu}, and \ref{lemma:push_mu} allow us to assume that $f$ is continuous \ep{after replacing $\epsilon$ by, say, $\epsilon/2$}. Since $\tilde{\Omega}_2$ is compact, $f$ is uniformly continuous, so we can pick $\delta \in (0; \epsilon/2)$ such that $f(x) \approx_{\epsilon/2} f(y)$ whenever $x \approx_\delta y$. By Lemma~\ref{lemma:push_of_cont}, the set $U' \coloneqq (\code{f})_\ast^{-1}(U)$ is an open neighborhood of $\tilde{\bm{u}}_2$. Let $\mathsf{q} \colon \tilde{\Omega}_2 \to 2^{\mathbb{N}}$ be a Borel map and consider the function $g \coloneqq f \circ \mathsf{q}$. Note that if $\mathfrak{m}_{\tilde{\bm{u}}_2}(\mathsf{p}, \mathsf{q}) \leq \delta^2$, then ${\mathfrak{d}}_{\tilde{\bm{u}}_2}(f, g) \leq \epsilon$. Indeed, if $\mathfrak{m}_{\tilde{\bm{u}}_2}(\mathsf{p}, \mathsf{q}) \leq \delta^2$, then, by Markov's inequality, \[ \tilde{\bm{u}}_2(\set{x \in \tilde{\Omega}_2 \,:\, \mathsf{p}(x) \not \approx_\delta \mathsf{q}(x)}) \,\leq\, \delta, \] and, by the choice of $\delta$ and since ${\mathfrak{d}}$ is bounded by $1$, we have ${\mathfrak{d}}_{\tilde{\bm{u}}_2}(f, g) \leq \epsilon/2 + \delta < \epsilon$. Additionally, if $D\in \fins{\G}\setminus \set{\varnothing}$ and $x \in \tilde{\Omega}_2$ satisfy $\mathbb{M}_D \code{\mathsf{q}}(x) \in U'$, then \[ \mathbb{M}_{D} \code{g}(x) \,=\, (\code{f})_\ast (\mathbb{M}_{D} \code{\mathsf{q}}(x)) \,\in \, U. \] Therefore, if Theorem~\ref{theo:ult_erg} holds for $\mathsf{p}$, $\delta^2$, and $U'$, then it also holds for $f$, $\epsilon$, and $U$, as desired. \end{claimproof} The remainder of the argument is similar to the last part of the proof of Lemma~\ref{lemma:red1}. For each $n \in {\mathbb{N}}^+$, let $Q_n$ be the set of all $a \in 2^{\mathbb{N}}$ such that $a(i) = 0$ for all $i \geq n$, and define $\mathsf{p}_n \colon \tilde{\Omega}_2 \to Q_n$ by \[ \mathsf{p}_n(x)(i) \coloneqq \begin{cases} x(\mathbf{1}, i) &\text{if } i < n;\\ 0 &\text{if } i \geq n. \end{cases} \] Then $\mathsf{p}_n \to \mathsf{p}$ uniformly, so to prove Theorem~\ref{theo:ult_erg} for $\mathsf{p}$, it is enough to prove it for each $\mathsf{p}_n$. Due to Lemma~\ref{lemma:basic}, Theorem~\ref{theo:ult_erg} for $\mathsf{p}_1$ is equivalent to Theorem~\ref{theo:ult_erg_bis} applied with $k = 2$. For larger $n$, consider the mapping $\theta_n \colon 2^{\mathbb{N}} \to (2^n)^{\mathbb{N}}$ given by \[ \theta(a)(i) \coloneqq (a(i n),\ a(i n + 1),\ \ldots,\ a(i n + n - 1)) \qquad \text{for all } a \in 2^{\mathbb{N}} \text{ and } i \in {\mathbb{N}}, \] where we identify the natural numbers less than $2^n$ with the $n$-tuples of zeros and ones. This mapping induces an equivariant isomorphism between $(\tilde{\Omega}_2, \tilde{\bm{u}}_2)$ and $(\tilde{\Omega}_{2^n}, \tilde{\bm{u}}_{2^n})$ and shows that Theorem~\ref{theo:ult_erg} for $\mathsf{p}_n$ is equivalent to Theorem~\ref{theo:ult_erg_bis} applied with $k = 2^n$. \end{coolproof} Similarly, Theorems \ref{theo:main_delta} and \ref{theo:main_Borel} reduce to the following statements: \begin{theobis}{theo:main_delta}\label{theo:main_delta_bis} For all $k \in {\mathbb{N}}^+$, $S \in \fins{\G}$, and $\epsilon > 0$, there is $C > 0$ with the following property: \smallskip Let $D$ be a finite subset of $\G$ with $|D| \geq C$ and let $\alpha \colon \G \curvearrowright X$ be an $(S, D)$-free Borel action of $\G$. Then, for any $\mu \in \Prob(X)$ and $\delta > 0$, there is a Borel map $g \colon X \to k$ such that, for all $\phi\colon S \to k$, \[ \mu\left(\left\{x \in X \,:\, \frac{|D \cap \mathcal{O}_\phi(\code{g}(x))|}{|D|} \approx_\epsilon \frac{1}{k^{|S|}}\right\}\right) \,\geq\, 1 - \delta. \] \end{theobis} \begin{theobis}{theo:main_Borel}\label{theo:main_Borel_bis} For all $k \in {\mathbb{N}}^+$, $S \in \fins{\G}$, and $\epsilon > 0$, there is $C > 0$ with the following property: \smallskip Let $D$ be a finite subset of $\G$ with $|D| \geq C$ and let $\alpha \colon \G \curvearrowright X$ be a uniformly subexponential $(S, D)$-free Borel action of $\G$. Then there is a Borel map $g \colon X \to k$ such that, for all $\phi \colon S \to k$, \[ \frac{|D \cap \mathcal{O}_\phi(\code{g}(x))|}{|D|} \,\approx_\epsilon\, \frac{1}{k^{|S|}} \qquad \text{for all } x \in X. \] \end{theobis} The proof of the following lemma is essentially the same as of Lemma~\ref{lemma:red2}, and we omit it. \begin{lemma} Theorem~\ref{theo:main_delta_bis} implies Theorem~\ref{theo:main_delta}, while Theorem~\ref{theo:main_Borel_bis} implies Theorem~\ref{theo:main_Borel}. \hfill \qedsymbol \end{lemma} \section{Using concentration of measure}\label{sec:concentration} \subsection{The main probabilistic bound} The following inequality is the main probabilistic input for our arguments: \begin{lemma}\label{lemma:concentrated} Let $k \in {\mathbb{N}}^+$, $S \in \fins{\G}$, and $\epsilon > 0$. Let $D$ be a nonempty finite subset of $\G$ and let $\alpha \colon \G \curvearrowright X$ be an $(S, D)$-free action of $\G$. Take any $x \in X$ and pick a function $c \colon (SD \cdot x) \to k$ uniformly at random. Then, for all $\phi \colon S \to k$, \[ \P \left[\frac{|D \cap \mathcal{O}_\phi(\code{c}(x))|}{|D|} \not\approx_\epsilon \frac{1}{k^{|S|}}\right] \,\leq\, 2\exp\left(-\epsilon^2 \frac{|D|}{2 |S|^3}\right). \] \end{lemma} \begin{coolproof} We shall apply the following concentration of measure result, which is a consequence of Azuma's inequality for Doob martingales \cite[\S{}7.4]{AS}: \begin{theo}[{\textls{Simple Concentration Bound}; see \cite[79]{MR}}]\label{theo:SCB1} Let $\xi$ be a random variable determined by $s$ independent trials such that changing the outcome of any one trial can affect $\xi$ at most by~$b$. Then \[ \P\left[\xi \not \approx_t \mathbb{E} \xi\right] \,\leq\, 2\exp\left(-\frac{t^2}{2b^2s}\right). \] \end{theo} \noindent Let $k$, $S$, $\epsilon$, $D$, $\alpha$, and $c$ be as in the statement of Lemma~\ref{lemma:concentrated}. Since the action $\alpha$ is $S$-free, for all $\phi \colon S \to k$, we have \[ \mathbb{E}\left[\left|D \cap \mathcal{O}_\phi(\code{c}(x))\right|\right] \,=\, \sum_{\delta \in D} \P\left[\delta \in \mathcal{O}_\phi(\code{c}(x))\right] \,=\, \frac{|D|}{k^{|S|}}. \] Consider any $y \in SD \cdot x$ and let $c_1$, $c_2 \colon (SD\cdot x) \to k$ be two maps that agree on $(SD \cdot x) \setminus \set{y}$. Let $\phi \colon S \to k$ and suppose that some $\delta \in D$ belongs to $\mathcal{O}_\phi(\code{c_1}(x)) \bigtriangleup \mathcal{O}_\phi(\code{c_2}(x))$. Then $y \in S \cdot (\delta \cdot x)$, i.e., $\delta \cdot x \in S^{-1} \cdot y$. Since $\alpha$ is $D$-free, there are at most $|S^{-1} \cdot y| = |S|$ possible values for $\delta$. Thus, we may apply the \hyperref[theo:SCB1]{Simple Concentration Bound} with parameters \[ s \coloneqq |SD \cdot x| \leq |S||D|, \qquad b \coloneqq |S|, \qquad \text{and} \qquad t \coloneqq \epsilon|D|, \] to obtain \[ \P \left[|D \cap \mathcal{O}_\phi(\code{c}(x))| \not\approx_{\epsilon |D|} \frac{|D|}{k^{|S|}}\right] \,\leq\, 2\exp\left(-\epsilon^2 \frac{|D|}{2 |S|^3}\right), \] as desired. \end{coolproof} \subsection{Proof of Theorem~\ref{theo:weak_erg}}\label{subsec:weak_erg} We are now ready to prove Theorem~\ref{theo:weak_erg} \ep{or rather Theorem~\ref{theo:weak_erg_bis}}. Let $k \in {\mathbb{N}}^+$ and let $(D_n)_{n \in {\mathbb{N}}}$ be a sequence of finite subsets of $\G$ such that $|D_n|/\log n \to\infty$. Take any $S \in \fins{\G}$, $\phi \colon S \to k$, and $\epsilon > 0$. We will show that for $\bm{u}_k$-a.e.\ $c \in \Omega_k$ and for all sufficiently large $n \in {\mathbb{N}}$, \begin{equation}\label{eq:approx_epsilon} \frac{|D_n \cap \mathcal{O}_\phi(c)|}{|D_n|} \,\approx_\epsilon\, \frac{1}{k^{|S|}}, \end{equation} which will imply the conclusion of Theorem~\ref{theo:weak_erg_bis}. For each $n \in {\mathbb{N}}$, let $X_n$ denote the set of all $c \in \Omega_k$ for which \eqref{eq:approx_epsilon} fails. By Lemma~\ref{lemma:concentrated}, we have \[ \sum_{n \in {\mathbb{N}}} \bm{u}_k(X_n) \,\leq\, \sum_{n \in {\mathbb{N}}} 2\exp\left(-\epsilon^2 \frac{|D_n|}{2 |S|^3}\right) \,<\, \infty, \] since $\epsilon^2|D_n|/(2|S|^3) > 2 \log n$ for all sufficiently large $n$. An application of the Borel--Cantelli lemma completes the proof. \section{The Lov\'asz Local Lemma and its measurable versions}\label{sec:LLL} \subsection{The classical LLL}\label{subsec:SLLL} The reader is referred to~\cite[Chapter~5]{AS} and \cite{MR} for background on the~LLL and its applications in combinatorics. The presentation below follows, with slight modifications,~\cite[Section~1.2]{MLLL}. Let $X$ be a set and let $k \in {\mathbb{N}}^+$. A \emphd{bad \ep{$k$-}event} over $X$ is a nonempty subset $B \subseteq \finf{X}{k}$ such that for all $\phi$, $\phi' \in B$, $\mathrm{dom}(\phi) = \mathrm{dom}(\phi')$. If a bad event $B$ is nonempty, then its \emphd{domain} is the set $\mathrm{dom}(B) \coloneqq \mathrm{dom}(\phi)$ for any \ep{hence all} $\phi \in B$; the domain of the empty bad event is, by definition, the empty set. The \emphd{probability} of a bad $k$-event $B$ with domain $F$ is defined to be \[ \P[B] \coloneqq {|B| \over k^{|F|}}. \] We say that a map $f \colon X \to k$ \emphd{avoids} a bad $k$-event $B$ if there is no $\phi \in B$ such that $\phi \subseteq f$. Note that if $X$ is finite and $f \colon X \to k$ is chosen uniformly at random, then $\P[B]$ is the probability that $f$ does not avoid $B$. A~\emphd{\ep{$k$\=/}instance \ep{of the LLL}} over a set $X$ is an arbitrary set ${\mathscr{B}}$ of bad $k$-events. A~\emphd{solution} to a $k$-instance~${\mathscr{B}}$ is a function $f \colon X \to k$ that avoids all $B \in {\mathscr{B}}$. For an instance ${\mathscr{B}}$ and a bad event $B \in {\mathscr{B}}$, the \emphd{neighborhood} of $B$ in ${\mathscr{B}}$ is the set \[ \mathrm{N}_{\mathscr{B}}(B) \coloneqq \set{B' \in {\mathscr{B}} \setminus \set{B} \,:\, \mathrm{dom}(B') \cap \mathrm{dom}(B) \neq \varnothing}. \] The \emphd{degree} of $B$ in ${\mathscr{B}}$ is defined to be $\deg_{\mathscr{B}}(B) \coloneqq |\mathrm{N}_{\mathscr{B}}(B)|$. Let \[ p({\mathscr{B}}) \coloneqq \sup_{B \in {\mathscr{B}}} \P[B] \qquad \text{and} \qquad d({\mathscr{B}}) \coloneqq \sup_{B \in {\mathscr{B}}} \deg_{\mathscr{B}}(B). \] An instance ${\mathscr{B}}$ is \emphd{correct for the Symmetric LLL} (the \emphd{SLLL} for short) if \[ e\cdot p({\mathscr{B}}) \cdot (d({\mathscr{B}}) + 1) < 1, \] where $e = 2.71\ldots${} denotes the base of the natural logarithm. Note that if ${\mathscr{B}}$ is correct for the SLLL, then, in particular, $\deg_{\mathscr{B}}(B) < \infty$ for all $B \in {\mathscr{B}}$ \ep{instances ${\mathscr{B}}$ with this property are called \emphd{locally finite} in \cite{MLLL}}. \begin{theo}[{Erd\H os--Lov\'asz~\cite{EL}; \textls{Symmetric Lov\'asz Local Lemma}}]\label{theo:SLLL} Let $k \in {\mathbb{N}}^+$ and let ${\mathscr{B}}$ be a $k$-instance of the LLL over a set $X$. If ${\mathscr{B}}$ is correct for the SLLL, then ${\mathscr{B}}$ has a solution. \end{theo} The Symmetric LLL was introduced by Erd\H os and Lov\'asz \ep{with $4$ in place of $e$} in their seminal paper~\cite{EL}; the constant was subsequently improved by Lov\'asz (the sharpened version first appeared in~\cite{Spencer}). Theorem~\ref{theo:SLLL} is a special case of the~SLLL in the so-called \emph{variable framework} (the~name is due to Kolipaka and Szegedy~\cite{KolipakaSzegedy}), which encompasses most typical applications. For the full statement of the~SLLL, see~\cite[Corollary~5.1.2]{AS}. Deducing Theorem~\ref{theo:SLLL} from \cite[Corollary~5.1.2]{AS} is routine when $X$ is finite (see, e.g., \cite[41]{MR}); the case of infinite~$X$ then follows by compactness. A more general version of Theorem~\ref{theo:SLLL} for infinite $X$, with $k$ replaced by an arbitrary standard probability space, was proved by Kun~\cite[Lemma~13]{Kun}. Theorem~\ref{theo:SLLL} can be extended to instances ${\mathscr{B}}$ with $d({\mathscr{B}}) = \infty$, provided that the probability $\P[B]$ of a bad event $B \in {\mathscr{B}}$ decays sufficiently quickly as $|\mathrm{dom}(B)|$ increases. An instance ${\mathscr{B}}$ is \emphd{correct for the General LLL} (the~\emphd{GLLL} for short) if $\mathrm{N}_{\mathscr{B}}(B)$ is countable for every $B \in {\mathscr{B}}$, and there is a function $\omega \colon {\mathscr{B}} \to [0;1)$, called a \emphd{witness} to the correctness of ${\mathscr{B}}$, such that for all $B \in {\mathscr{B}}$, \[ \P[B] \leq \omega(B) \prod_{B' \in \mathrm{N}_{\mathscr{B}}(B)} (1 - \omega(B')). \] \begin{theo}[{\textls{General Lov\'asz Local Lemma}; \cite[Lemma~5.1.1]{AS}}]\label{theo:GLLL} Let $k \in {\mathbb{N}}^+$ and let ${\mathscr{B}}$ be a $k$-instance of the~LLL over a set $X$. If ${\mathscr{B}}$ is correct for the~GLLL, then ${\mathscr{B}}$ has a solution. \end{theo} A standard calculation (see~\cite[proof of Corollary~5.1.2]{AS}) shows that if an instance ${\mathscr{B}}$ is correct for the~SLLL, then it is also correct for the~GLLL (hence the name ``General LLL''). \subsection{Measurable versions of the LLL}\label{subsec:MLLL} Let $X$ be a standard Borel space and let $k \in {\mathbb{N}}^+$. Then the set of all bad $k$-events is also naturally equipped with the structure of a standard Borel space (indeed, each bad event is a finite set, so the set of all bad $k$-events is a Borel subset of the space $\fins{\finf{X}{k}}$). Thus, it makes sense to talk about \emphd{Borel} instances of the~LLL, i.e., Borel sets of bad events. Given a Borel $k$-instance ${\mathscr{B}}$ over $X$ that is correct for the~SLLL, it is natural to wonder if it has a Borel solution $f \colon X \to k$. Although the answer is negative in general (see~\cite[Theorem~1.6]{CJMST-D}), Cs\'oka, Grabowski, M\'ath\'e, Pikhurko, and Tyros~\cite{CGMPT} answered the question in the affirmative for \emph{uniformly subexponential} instances. Given an instance ${\mathscr{B}}$ over a set $X$, an element $x \in X$, and an integer $n \in {\mathbb{N}}$, let $R^n_{\mathscr{B}}(x)$ denote the set of all $y \in X$ such that either $y = x$, or there exists a sequence $B_1$, \ldots, $B_m \in {\mathscr{B}}$ with $m \leq n$ satisfying \[ x \in \mathrm{dom}(B_1), \qquad \mathrm{dom}(B_i) \cap \mathrm{dom}(B_{i+1}) \neq \varnothing \text{ for all } 1 \leq i < m,\qquad\text{ and }\qquad y \in \mathrm{dom}(B_m). \] The instance ${\mathscr{B}}$ is \emphd{uniformly subexponential} if for every $\epsilon > 0$, there exists $n_0 \in {\mathbb{N}}$ such that for all $n \geq n_0$ and for all $x \in X$, $|R^n_{\mathscr{B}}(x)| < (1+\epsilon)^n$. \begin{theo}[{Cs\'oka--Grabowski--M\'ath\'e--Pikhurko--Tyros~\cite[Theorem~1.3]{CGMPT}, \textls{Borel SLLL for uniformy subexponential instances}}]\label{theo:LLL_Borel} Let $k \in {\mathbb{N}}^+$ and let ${\mathscr{B}}$ be a Borel $k$-instance of the~LLL over a standard Borel space $X$. If ${\mathscr{B}}$ is correct for the~SLLL and uniformly subexponential, then ${\mathscr{B}}$ has a Borel solution $f \colon X \to k$. \end{theo} For a $k$-instance ${\mathscr{B}}$ over a set $X$ and a map $f \colon X \to k$, we define the \emphd{defect} $\mathrm{Def}(f; {\mathscr{B}})$ of $f$ with respect to ${\mathscr{B}}$ by \begin{equation}\label{eq:defect1} \mathrm{Def}(f,{\mathscr{B}}) \coloneqq \set{x \in X \,:\, x \in \mathrm{dom}(\phi) \text{ for some } \phi \in B \in {\mathscr{B}} \text{ with } \phi \subseteq f}. \end{equation} Evidently, $f$ is a solution to ${\mathscr{B}}$ if and only if $\mathrm{Def}(f, {\mathscr{B}}) = \varnothing$. Thus, in the absence of a Borel solution to ${\mathscr{B}}$, it is natural to seek a Borel map $f \colon X \to k$ whose defect is ``small'' in some sense. The next result was proved by the current author in~\cite{MLLL}: \begin{theo}[{\cite[Theorem~5.1]{MLLL}, \textls{approximate SLLL}}]\label{theo:LLL_approx} Let $k \in {\mathbb{N}}^+$ and let ${\mathscr{B}}$ be a Borel $k$-instance of the~LLL over a standard Borel space $X$. If ${\mathscr{B}}$ is correct for the~SLLL, then for any $\mu \in \Prob(X)$ and $\delta > 0$, there is a Borel function $f \colon X \to k$ with $\mu(\mathrm{Def}(f, {\mathscr{B}})) \leq \delta$. \end{theo} It is an open question whether the conclusion of Theorem~\ref{theo:LLL_approx} holds with $\delta = 0$; see Problem~\ref{prob:meas_LLL} in Section~\ref{sec:remks}. Also, Theorem~\ref{theo:LLL_approx} fails for instances that are correct for the GLLL instead of the SLLL \ep{see \cite[Theorem~7.1]{MLLL} and Remark~\ref{remk:amenable} below}. However, when the underlying structure is in a certain sense induced by the shift action $\sigma$, even instances that are only correct for the~GLLL can be solved with a null defect---see Theorem~\ref{theo:LLL_meas} in the next subsection. \subsection{Using the LLL over group actions} Now we describe a convenient set-up for applying the LLL to problems in ergodic theory. Let $\alpha \colon \G \curvearrowright X$ be an action of $\G$ and let $\Phi \subseteq \finf{\G}{k}$ be a bad $k$-event over~$\G$ with domain $F \in \fins{\G}$. For each $x \in X$, define a bad $k$-event $B_x(\Phi, \alpha)$ over $X$ via \[ B_x(\Phi, \alpha) \coloneqq \set{\phi \colon (F \cdot x) \to k \,:\, \rest{\code{\phi}(x)}{F} \in \Phi}. \] Note that if $B_x(\Phi, \alpha) \neq \varnothing$, then $\mathrm{dom}(B_x(\Phi, \alpha)) = F \cdot x$. \ep{If $\alpha$ is not $F$-free, then $B_x(\Phi, \alpha)$ may be empty even if $\Phi$ is not.} By construction, a function $f \colon X \to k$ avoids $B_x(\Phi, \alpha)$ precisely when $\code{f}(x)$ avoids $\Phi$. Define an instance ${\mathscr{B}}(\Phi, \alpha)$ of the LLL over $X$ as follows: \[ {\mathscr{B}}(\Phi, \alpha) \coloneqq \set{B_x(\Phi, \alpha) \,:\, x \in X}. \] Clearly, if $X$ is a standard Borel space and $\alpha \colon \G \curvearrowright X$ is a Borel action, then the instance ${\mathscr{B}}(\Phi, \alpha)$ is Borel. A function $f \colon X \to k$ is a solution to ${\mathscr{B}}(\Phi, \alpha)$ if and only if $\code{f}(x)$ avoids $\Phi$ for all $x \in X$. Hence, it is somewhat more convenient to define the \emphd{defect} of a map $f \colon X \to k$ as the set of all $x \in X$ such that $\code{f}(x)$ does \emph{not} avoid $\Phi$: \[ \mathrm{Def}(f, \Phi, \alpha) \coloneqq \set{x \in X \,:\, \rest{\code{f}(x)}{F} \in \Phi}. \] There is a straightforward relationship between this definition and the one in \eqref{eq:defect1}; namely, \begin{equation}\label{eq:two_defects} \mathrm{Def}(f, {\mathscr{B}}(\Phi, \alpha)) = F \cdot \mathrm{Def}(f, \Phi, \alpha). \end{equation} Using the above notation, we can formulate the following corollaries of Theorems~\ref{theo:LLL_Borel} and \ref{theo:LLL_approx}: \begin{corl}[to Theorem~\ref{theo:LLL_Borel}]\label{corl:LLL_Borel} Let $\alpha \colon \G \curvearrowright X$ be a uniformly subexponential Borel action of $\G$ and let $k \in {\mathbb{N}}^+$. Let $\Phi$ be a bad $k$-event over $\G$ and suppose that the instance ${\mathscr{B}}(\Phi, \alpha)$ is correct for the SLLL. Then ${\mathscr{B}}(\Phi, \alpha)$ has a Borel solution $f \colon X \to k$. \end{corl} \begin{corl}[to Theorem~\ref{theo:LLL_approx}]\label{corl:LLL_approx} Let $\alpha \colon \G \curvearrowright X$ be a Borel action of $\G$ and let $k \in {\mathbb{N}}^+$. Let $\Phi$ be a bad $k$-event over $\G$ and suppose that the instance ${\mathscr{B}}(\Phi, \alpha)$ is correct for the SLLL. Then, for any $\mu \in \Prob(X)$ and $\delta > 0$, there is a Borel function $f \colon X \to k$ with $\mu(\mathrm{Def}(f,\Phi, \alpha)) < \delta$. \end{corl} \begin{remk} In the statement of Corollary~\ref{corl:LLL_approx}, the measure $\mu$ is not assumed to be $\alpha$-invariant. Because of that, to derive Corollary~\ref{corl:LLL_approx}, one has to apply Theorem~\ref{theo:LLL_approx} not to $\mu$ itself, but to the measure obtained by shifting $\mu$ by one of the elements of $\mathrm{dom}(\Phi)$, and then use \eqref{eq:two_defects}. \end{remk} More generally, let $(\Phi_n)_{n \in {\mathbb{N}}}$ be a sequence of bad $k$-events over $\G$. For an action $\alpha \colon \G \curvearrowright X$ and a map $f \colon X \to k$, define \[ {\mathscr{B}}((\Phi_n)_{n \in {\mathbb{N}}}, \alpha) \coloneqq \bigcup_{n = 0}^\infty {\mathscr{B}}(\Phi_n, \alpha) \qquad \text{and} \qquad \mathrm{Def}(f, (\Phi_n)_{n \in {\mathbb{N}}}, \alpha) \coloneqq \bigcup_{n=0}^\infty \mathrm{Def}(f, \Phi_n, \alpha). \] When $\alpha = \sigma$, we have the following strengthening of Corollary~\ref{corl:LLL_approx}: \begin{theo}[{\cite[Corollary 6.7]{MLLL}, \textls{measurable GLLL over the shift}}]\label{theo:LLL_meas} Let $k \in {\mathbb{N}}^+$ and let $(\Phi_n)_{n \in {\mathbb{N}}}$ be a sequence of bad $k$-events over $\G$. Suppose that the instance ${\mathscr{B}}((\Phi_n)_{n \in {\mathbb{N}}}, \operatorname{Free}(\sigma))$ is correct for the GLLL. Then there is a Borel function $f \colon \Omega \to k$ with ${\bm{\lambda}}(\mathrm{Def}(f, (\Phi_n)_{n \in {\mathbb{N}}}, \sigma)) = 0$. \end{theo} \begin{remk}\label{remk:amenable} Theorem \ref{theo:LLL_meas} can fail for actions other than $\sigma$: According to \cite[Theorem~7.1]{MLLL}, if $\G$ is amenable, then the analog of Theorem~\ref{theo:LLL_meas} holds for a free ergodic $\text{p.m.p.}$\xspace action $\alpha \colon \G \curvearrowright (X, \mu)$ if and only if there is a factor map $\pi \colon (X, \mu) \to (\Omega, {\bm{\lambda}})$. \end{remk} Theorem~\ref{theo:LLL_meas} is a special case of \cite[Theorem~6.6]{MLLL}, whose full statement is rather technical and will not be needed here. Roughly speaking, \cite[Theorem~6.6]{MLLL} asserts that any combinatorial argument proceeding via a series of iterative applications of the~GLLL can be performed in a measurable fashion over the shift action $\sigma \colon \G \curvearrowright (\Omega, {\bm{\lambda}})$. \section{Proofs of Theorems \ref{theo:ult_erg}, \ref{theo:main_delta}, and \ref{theo:main_Borel}}\label{sec:proofs} \subsection{Proofs of Theorems \ref{theo:main_delta} and \ref{theo:main_Borel}}\label{subsec:pointwise} We first establish Theorems~\ref{theo:main_delta} and \ref{theo:main_Borel}, as their proofs are somewhat more straightforward than that of Theorem~\ref{theo:ult_erg} \ep{for instance, they only use the Symmetric LLL rather than the more technical General LLL}. Let $k \in {\mathbb{N}}^+$, $S \in \fins{\G}$, and $\epsilon > 0$. For a nonempty finite subset $D \subset \G$, let $\Phi(k, S, \epsilon, D)$ denote the bad $k$-event over $\G$ with domain $SD$ consisting of all maps $c \colon SD \to k$ such that \[ \frac{|D \cap \mathcal{O}_\phi(c)|}{|D|} \,\not\approx_\epsilon\, \frac{1}{k^{|S|}} \qquad \text{for some } \phi \colon S \to k. \] By definition, if $\alpha \colon \G \curvearrowright X$ is a Borel action of $\G$ and $g \colon X \to k$ is a Borel map, then we have \begin{equation}\label{eq:defect2} x \in \mathrm{Def}(g, \Phi(k, S, \epsilon, D), \alpha) \quad\Longleftrightarrow\quad \frac{|D \cap \mathcal{O}_\phi(\code{g}(x))|}{|D|} \,\not\approx_\epsilon\, \frac{1}{k^{|S|}} \ \text{for some}\ \phi \colon S \to k. \end{equation} \begin{lemma}\label{lemma:correct} Let $k \in {\mathbb{N}}^+$, $S \in \fins{\G}$, and $\epsilon > 0$. There exists $C > 0$ such that for all $D \in \fins{\G}$ with $|D| > C$ and for every $(S, D)$-free action $\alpha \colon \G \curvearrowright X$, the instance ${\mathscr{B}}(\Phi(k, S, \epsilon, D), \alpha)$ is correct for the SLLL. \end{lemma} \begin{coolproof} Let $D \in \fins{\G} \setminus \set{\varnothing}$ and let $\alpha \colon \G \curvearrowright X$ be $(S, D)$-free. Set \[ \Phi \coloneqq \Phi(k, S, \epsilon, D), \qquad {\mathscr{B}} \coloneqq {\mathscr{B}}(\Phi, \alpha), \qquad \text{and} \qquad B_x \coloneqq B_x(\Phi, \alpha) \text{ for all } x \in X. \] Due to Lemma~\ref{lemma:concentrated}, we have \[ p({\mathscr{B}}) \,\leq\, 2k^{|S|} \exp\left(-\epsilon^2 \frac{|D|}{2 |S|^3}\right). \] To upper bound $d({\mathscr{B}})$, note that for each $x \in X$, \[ \mathrm{N}_{\mathscr{B}}(B_x) = \set{B_y \in {\mathscr{B}} \setminus \set{B_x} \,:\, (SD \cdot y) \cap (SD \cdot x) \neq \varnothing}. \] Since $(SD \cdot y) \cap (SD \cdot x) \neq \varnothing$ if and only if $y \in (SD)^{-1}SD \cdot x$, we obtain \[ \deg_{\mathscr{B}}(B_x) \leq |(SD)^{-1}SD| - 1 \leq |S|^2|D|^2 - 1. \] (We subtracted $1$ since $y$ cannot be equal to $x$.) Hence, $d({\mathscr{B}}) \leq |S|^2|D|^2 - 1$, and ${\mathscr{B}}$ is correct for the SLLL as long as \[ e \cdot 2k^{|S|} \exp\left(-\epsilon^2 \frac{|D|}{2 |S|^3}\right) \cdot |S|^2|D|^2 \,<\, 1, \] which holds whenever $|D|$ is sufficiently large. \end{coolproof} Theorems~\ref{theo:main_delta_bis} and \ref{theo:main_Borel_bis} now follow immediately by combining \eqref{eq:defect2} and Lemma~\ref{lemma:correct} with Corollaries \ref{corl:LLL_approx} and \ref{corl:LLL_Borel} respectively. \subsection{Proof of Theorem~\ref{theo:ult_erg}}\label{subsec:ult_erg} For the purposes of proving Theorem~\ref{theo:ult_erg}, the role of Lemma~\ref{lemma:correct} is played by the following fact: \begin{lemma}\label{lemma:correct2} Let $k \in {\mathbb{N}}^+$, $S \in \fins{\G}$, and $\epsilon > 0$. There exists $C > 0$ with the following property: \smallskip Let $(D_n)_{n \in {\mathbb{N}}}$ be a sequence of finite subsets of $\G$ with $|D_n|\geq C \log(n+2)$ for all $n \in {\mathbb{N}}$ and let $\alpha \colon \G \curvearrowright X$ be a free action of $\G$. Set \[ \Phi_n \coloneqq \Phi(k, S, \epsilon, D_n) \text{ for all } n \in {\mathbb{N}}, \] \[ {\mathscr{B}} \coloneqq {\mathscr{B}}((\Phi_n)_{n \in {\mathbb{N}}}, \alpha), \qquad \text{and} \qquad B_{n,x} \coloneqq B_x(\Phi_n, \alpha) \text{ for all } n \in {\mathbb{N}} \text{ and } x \in X. \] Then the instance ${\mathscr{B}}$ is correct for the GLLL. Moreover, there is a function $\omega \colon {\mathbb{N}} \to [0;1)$ such that \begin{equation}\label{eq:small_sum} \sum_{n = 0}^\infty |SD_n| \cdot \frac{\omega(n)}{1-\omega(n)} \,<\, \epsilon, \end{equation} and the mapping $\tilde{\omega} \colon {\mathscr{B}} \to [0;1) \colon B_{n,x} \mapsto \omega(n)$ is a witness to the correctness of ${\mathscr{B}}$. \end{lemma} \begin{coolproof} Fix any $0 < a < \epsilon^2/(2|S|^3)$. We claim that if $C$ is large enough, then the function \[\omega(n) \coloneqq \exp(-a|D_n|)\] has the desired properties. To begin with, we are going to assume that $C$ is so large that \begin{equation* \exp(-a \cdot C\log(2)) \,<\, 1/2, \end{equation*} and that the function $\xi \mapsto \xi\exp(-a\xi)$ is decreasing for all $\xi \geq C\log 2$. For any such $C$, we have \begin{align*} \sum_{n =0}^\infty |SD_n| \cdot \frac{\omega(n)}{1 - \omega(n)} \,&\leq\, |S|\sum_{n=0}^\infty |D_n| \cdot \frac{\exp(-a|D_n|)}{1 - \exp(-a|D_n|)} \\ &\leq\, 2|S| \sum_{n= 0}^\infty |D_n| \exp(-a|D_n|) \,\leq\, 2|S|C \sum_{n= 0}^\infty \frac{\log(n+2)}{(n+2)^{Ca}}. \end{align*} The last expression approaches $0$ as $C \to \infty$, so we can guarantee \eqref{eq:small_sum}. Consider any $n \in {\mathbb{N}}$ and $x \in X$. By Lemma~\ref{lemma:concentrated}, we have \[ \P[B_{n,x}] \,\leq\, 2k^{|S|} \exp\left(-\epsilon^2 \frac{|D_n|}{2 |S|^3}\right). \] If $\mathrm{dom}(B_{n,x}) \cap \mathrm{dom}(B_{m,y}) \neq \varnothing$ for some $m \in {\mathbb{N}}$ and $y \in X$, then $y \in (SD_m)^{-1}SD_n \cdot x$, and hence for any particular $m \in {\mathbb{N}}$, there are at most $|S|^2|D_m||D_n|$ choices of such $y$. Therefore, the mapping $\tilde{\omega} \colon {\mathscr{B}} \to [0;1)$ is a witness to the correctness of ${\mathscr{B}}$ as long as we have \begin{equation}\label{eq:correct1} 2k^{|S|} \exp\left(-\epsilon^2 \frac{|D_n|}{2 |S|^3}\right) \,\leq\, \omega(n) \prod_{m = 0}^\infty (1 - \omega(m))^{|S|^2 |D_m||D_n|}, \end{equation} for all $n \in {\mathbb{N}}$. Using the definition of $\omega$ and then taking the logarithm of both sides of \eqref{eq:correct1} and dividing them by $(-|D_n|)$, we rewrite \eqref{eq:correct1} as \begin{equation}\label{eq:correct2} -\frac{\log(2k^{|S|})}{|D_n|} \,+\, \frac{\epsilon^2}{2 |S|^3} \,\geq\, a \,-\, |S|^2\sum_{m=0}^\infty |D_m|\log(1 - \exp(-a|D_m|)). \end{equation} Let us first look at the left-hand side of \eqref{eq:correct2}. We have \[ -\frac{\log(2k^{|S|})}{|D_n|} \,+\, \frac{\epsilon^2}{2 |S|^3} \,\geq\, -\frac{\log(2k^{|S|})}{C \log2} \,+\, \frac{\epsilon^2}{2 |S|^3} \,\xrightarrow[C \to \infty]{}\, \frac{\epsilon^2}{2 |S|^3}. \] As for the right-hand side of \eqref{eq:correct2}, note that $-\log(1 - \xi) < 2\xi$ for all $0 < \xi < 1/2$, so \begin{align*} a \,-\, |S|^2\sum_{m=0}^\infty &|D_m|\log(1 - \exp(-a|D_m|)) \,<\, a \,+\, 2|S|^2\sum_{m=0}^\infty |D_m| \exp(-a|D_m|) \\ &\leq\, a \,+\, 2|S|^2 C \sum_{m=0}^\infty \frac{\log(m+2)}{(m+2)^{Ca}} \,\xrightarrow[C \to \infty]{}\, a. \end{align*} Since $a$ was chosen to be less than $\epsilon^2/(2|S|^3)$, we conclude that \eqref{eq:correct2} holds for all large $C$. \end{coolproof} From \eqref{eq:defect2}, Lemma~\ref{lemma:correct2}, and Theorem~\ref{theo:LLL_meas}, we can derive most of Theorem~\ref{theo:ult_erg_bis}. The only part that is missing is that the map $g \colon \tilde{\Omega}_k \to k$ can be chosen so that \[ \tilde{\bm{u}}_k(\set{x \in \tilde{\Omega}_k \,:\, g(x) \neq x(\mathbf{1}, 0)}) \,\leq\, \epsilon. \] To argue this, we have to review the {proof} of Theorem~\ref{theo:LLL_meas}. As mentioned in the introduction, the tool used to prove Theorem~\ref{theo:LLL_meas} is the \emph{Moser--Tardos algorithm}, developed by Moser and Tardos in \cite{MT}. Here we outline only the most relevant elements of the Moser--Tardos theory when applied to our current situation. For further details, see \cite{MT} and \cite[\S3]{MLLL}. For the rest of this subsection, fix $k \in {\mathbb{N}}^+$ and a sequence $(\Phi_n)_{n \in {\mathbb{N}}}$ of bad $k$-events over $\G$. For each $n \in {\mathbb{N}}$, set $F_n \coloneqq \mathrm{dom}(\Phi_n)$. Define \[ {\mathscr{B}} \coloneqq {\mathscr{B}}((\Phi_n)_{n \in {\mathbb{N}}}, \operatorname{Free}(\sigma_{k^{\mathbb{N}}})), \qquad \text{and} \qquad B_{n,x} \coloneqq B_x(\Phi_n, \sigma_{k^{\mathbb{N}}}) \text{ for all } n \in {\mathbb{N}} \text{ and } x \in \tilde{\Omega}_k. \] Consider the following inductive construction: \begin{leftbar} \noindent Set $t_0(x) \coloneqq 0$ for all $x \in \tilde{\Omega}_k$. \medskip \noindent {\sc Step $i \in {\mathbb{N}}$}: Define \begin{align*} g_i(x) &\coloneqq x(\mathbf{1}, t_i(x)) \quad \text{for all } x \in \tilde{\Omega}_k;\\ \medskip A_i' &\coloneqq \set{(n,x) \in {\mathbb{N}} \times \tilde{\Omega}_k \,:\, g_i \text{ does not avoid } B_{n,x}}. \end{align*} Choose $A_i \subseteq A_i'$ to be an arbitrary Borel maximal subset of~$A_i'$ with the property that \[(F_n \cdot x) \cap (F_m \cdot y) = \varnothing \qquad \text{for all distinct pairs } (n,x),\ (m,y) \in A_i. \] (Such $A_i$ exists by, e.g., \cite[Lemma 7.3]{KechrisMiller}.) Let \[ T_i \coloneqq \bigcup_{(n,x) \,\in\, A_i} (F_n \cdot x) \qquad \text{and} \qquad t_{i+1}(x) \coloneqq \begin{cases} t_i(x) + 1 &\text{if } x \in T_i;\\ t_i(x) &\text{otherwise}. \end{cases} \] \end{leftbar} \noindent By definition, $g_0(x) = x(\mathbf{1}, 0)$ for all $x \in \tilde{\Omega}_k$. We call a sequence $\mathcal{A} \coloneqq (A_i)_{i=0}^\infty$ obtained via the above procedure a \emphd{Borel Moser--Tardos process}. Note that there is not a unique Borel Moser--Tardos process, as there is some freedom in the choice of the Borel maximal subset $A_i \subseteq A_i'$. Let $\mathcal{A} = (A_i)_{i=0}^\infty$ be a Borel Moser--Tardos process. For $x \in \tilde{\Omega}_k$, define $t(x) \in {\mathbb{N}} \cup \set{\infty}$ by \[ t(x) \coloneqq \lim_{i \to \infty} t_i(x). \] We say that $x$ is \emphd{$\mathcal{A}$-stable} if $t(x) < \infty$, i.e., if the corresponding sequence $t_0(x)$, $t_1(x)$, \ldots{} is eventually constant. Let $\mathrm{St}(\mathcal{A}) \subseteq \tilde{\Omega}_k$ denote the set of all $\mathcal{A}$-stable elements. For $x \in \mathrm{St}(\mathcal{A})$, we can define \begin{equation}\label{eq:g} g(x) \coloneqq x(\mathbf{1}, t(x)). \end{equation} It is easy to verify \ep{see \cite[Proposition 3.3]{MLLL}} that if $F_n \cdot x \subseteq \mathrm{St}(\mathcal{A})$, then $x \not \in \mathrm{Def}(g, \Phi_n, \sigma_{k^{\mathbb{N}}})$. The \emphd{index} $\mathrm{Ind}(n, x, \mathcal{A}) \in {\mathbb{N}} \cup \set{\infty}$ of a pair $(n,x) \in {\mathbb{N}} \times \tilde{\Omega}_k$ in $\mathcal{A}$ is defined by the formula \[ \mathrm{Ind}(n, x, \mathcal{A}) \coloneqq |\set{i \in {\mathbb{N}}\,:\, (n,x) \in A_i}|. \] Note that for all $x \in \operatorname{Free}(\tilde{\Omega}_k)$, we have \begin{equation}\label{eq:stabilizing1} t(x) \,=\, \sum_{n = 0}^\infty \,\sum_{\delta \in F_n} \mathrm{Ind}(n, \delta^{-1} \cdot x, \mathcal{A}), \end{equation} and hence such $x$ is $\mathcal{A}$-stable if and only if the expression on the right hand side of \eqref{eq:stabilizing1} is finite. The following theorem is the central result of the Moser--Tardos theory: \begin{theo}[{Moser--Tardos~\cite{MT}; see also \cite[Theorem~3.5]{MLLL}}]\label{theo:MoserTardos1} Let $\omega \colon {\mathbb{N}} \to [0;1)$ be a function such that the mapping $\tilde{\omega} \colon {\mathscr{B}} \to [0;1) \colon B_{n,x} \mapsto \omega(n)$ is a witness to the correctness of ${\mathscr{B}}$. Then, for any Borel Moser--Tardos process $\mathcal{A}$ and for all $n \in {\mathbb{N}}$, we have \begin{equation*} \int_{\tilde{\Omega}_k} \mathrm{Ind}(n, x, \mathcal{A}) \, \mathrm{d} \tilde{\bm{u}}_k(x) \,\leq\, \frac{\omega(n)}{1 - \omega(n)}. \end{equation*} \end{theo} \begin{corl}[to Theorem~\ref{theo:MoserTardos1}]\label{corl:dist} Let $\omega \colon {\mathbb{N}} \to [0;1)$ be such that $\tilde{\omega} \colon {\mathscr{B}} \to [0;1) \colon B_{n,x} \mapsto \omega(n)$ is a witness to the correctness of ${\mathscr{B}}$. Then there is a Borel function $g \colon \tilde{\Omega}_k \to k$ such that \[ \tilde{\bm{u}}_k(\mathrm{Def}(g, (\Phi_n)_{n \in {\mathbb{N}}}, \sigma_{k^{\mathbb{N}}})) = 0 \qquad \text{and} \qquad \tilde{\bm{u}}_k(\set{x \in \tilde{\Omega}_k \,:\, g(x) \neq x(\mathbf{1}, 0)}) \,\leq\, \sum_{n=0}^\infty |F_n| \cdot \frac{\omega(n)}{1 - \omega(n)}. \] \end{corl} \begin{coolproof} First we show that the sum \[ S \,\coloneqq\, \sum_{n=0}^\infty |F_n| \cdot \frac{\omega(n)}{1 - \omega(n)} \] is finite. Without loss of generality, assume that $\Phi_0 \neq \varnothing$. Consider any $x \in \operatorname{Free}(\tilde{\Omega}_k)$. Since $\tilde{\omega}$ is a witness to the correctness of ${\mathscr{B}}$, we have $\P[B_{0, x}] \leq \omega(0) < 1$, so $F_0 \neq \varnothing$. Hence, for every $n \in {\mathbb{N}}^+$, there exist at least $|F_n|$ distinct $y$ with $B_{n,y} \in \mathrm{N}_{\mathscr{B}}(B_{0,x})$. Therefore, $ \prod_{n=1}^\infty (1 - \omega(n))^{|F_n|} \geq \P[B_{0,x}] > 0, $ which implies that $\sum_{n=0}^\infty |F_n| \omega(n)$ is finite. In particular, for all sufficiently large $n$ we have $\omega(n) \leq 1/2$ and $\omega(n)/(1- \omega(n)) \leq 2\omega(n)$, and hence $S$ is also finite. Let $\mathcal{A} = (A_i)_{i=0}^\infty$ be an arbitrary Borel Moser--Tardos process and let $g$ be given by \eqref{eq:g}. From \eqref{eq:stabilizing1} and the \hyperref[theo:MoserTardos1]{Moser--Tardos theorem}, we get \begin{align*} \int_{\tilde{\Omega}_k} t(x) \,\mathrm{d} \tilde{\bm{u}}_k(x) \,&=\, \sum_{n = 0}^\infty \sum_{\delta \in F_n} \int_{\tilde{\Omega}_k} \mathrm{Ind}(n, \delta^{-1} \cdot x, \mathcal{A}) \, \mathrm{d} \tilde{\bm{u}}_k(x) \\ [\text{$\tilde{\bm{u}}_k$ is shift-invariant}] \qquad \,&=\, \sum_{n = 0}^\infty |F| \cdot \int_{\tilde{\Omega}_k} \mathrm{Ind}(n, x, \mathcal{A}) \, \mathrm{d} \tilde{\bm{u}}_k(x) \,\leq\, S\,<\,\infty. \end{align*} In particular, $t(x) < \infty$ for $\tilde{\bm{u}}_k$-a.e.\ $x \in \tilde{\Omega}_k$, i.e., $\tilde{\bm{u}}_k(\mathrm{St}(\mathcal{A})) = 1$, so \[ \tilde{\bm{u}}_k(\mathrm{Def}(g, (\Phi_n)_{n \in {\mathbb{N}}}, \sigma_{k^{\mathbb{N}}})) = 0. \] Furthermore, if $x \in \mathrm{St}(\mathcal{A})$ and $g(x) \neq x(\mathbf{1}, 0) = g_0(x)$, then $t(x) \geq 1$; thus, \[ \tilde{\bm{u}}_k(\set{x \in \tilde{\Omega}_k \,:\, g(x) \neq x(\mathbf{1}, 0)}) \,\leq\, \tilde{\bm{u}}_k(\set{x \in \tilde{\Omega}_k \,:\, t(x) \geq 1}) \,\leq\, \int_{\tilde{\Omega}_k} t(x) \,\mathrm{d} \tilde{\bm{u}}_k(x) \,\leq\, S, \] as desired. \end{coolproof} Since the domain of $\Phi(k, S, \epsilon, D)$ is, by definition, $SD$, \eqref{eq:small_sum} in the statement of Lemma~\ref{lemma:correct2} and Corollary~\ref{corl:dist} yield the remaining part of Theorem~\ref{theo:ult_erg_bis}. \section{Open problems}\label{sec:remks} The following is perhaps the central open question regarding the behavior of the LLL in the measurable setting: \begin{prob}\label{prob:meas_LLL} Does the SLLL hold measurably with a null defect? In other words, can one replace $\mu(\mathrm{Def}(f, {\mathscr{B}})) \leq \delta$ by $\mu(\mathrm{Def}(f, {\mathscr{B}})) = 0$ in the conclusion of Theorem~\ref{theo:LLL_approx}? \end{prob} \noindent A positive solution to Problem~\ref{prob:meas_LLL} would allow one to strengthen Theorem~\ref{theo:main_delta} by taking $\delta=0$. For now, we leave this potential strengthening as an open problem. \begin{prob}\label{prob:delta_zero} Does Theorem~\ref{theo:main_delta} hold with $\delta = 0$? \end{prob} \noindent As mentioned in \S\ref{subsec:MLLL}, the SLLL {fails} in the purely Borel context~\cite[Theorem~1.6]{CJMST-D}. However, it is still conceivable that a purely Borel pointwise version of the \hyperref[theo:AW]{Ab\'ert--Weiss theorem}, similar to Theorem~\ref{theo:main_Borel}, holds in full generality, in which case a different proof approach might be needed to establish it. We state it here as another open question. \begin{prob}\label{prob:Borel} Let $K$ be a compact metric space and let $f \colon \Omega \to K$ be a Borel function. Fix an open neighborhood $U$ of the measure $\mathbb{M}_{\bm{\lambda}} \code{f}$. Does there always exist a nonempty finite set $D \subset \G$ such that the following statement holds? \smallskip Let $\alpha \colon \G \curvearrowright X$ be a free Borel action of $\G$. Then there is a Borel map $g \colon X \to K$ such that \[\mathbb{M}_D \code{g} (x) \in U, \qquad \text{for all}\ x \in X.\] \end{prob} \printbibliography

https://arxiv.org/abs/0802.4115

Fast Directional Computation for the High Frequency Helmholtz Kernel in Two Dimensions

This paper introduces a directional multiscale algorithm for the two dimensional $N$-body problem of the Helmholtz kernel with applications to high frequency scattering. The algorithm follows the approach in [Engquist and Ying, SIAM Journal on Scientific Computing, 29 (4), 2007] where the three dimensional case was studied. The main observation is that, for two regions that follow a directional parabolic geometric configuration, the interaction between the points in these two regions through the Helmholtz kernel is approximately low rank. We propose an improved randomized procedure for generating the low rank representations. Based on these representations, we organize the computation of the far field interaction in a multidirectional and multiscale way to achieve maximum efficiency. The proposed algorithm is accurate and has the optimal $O(N\log N)$ complexity for problems from two dimensional scattering applications. We present numerical results for several test examples to illustrate the algorithm and its application to two dimensional high frequency scattering problems.

\section{Introduction} \label{sec:intro} \subsection{Problem statement} In this paper, we consider the two dimensional $N$-body problem for the high frequency Helmholtz kernel. Let $\{f_i, 1\le i \le N\}$ be a set of charges located at points $\{p_i, 1\le i\le N\}$ in $\mathbb{R}^2$. We assume that the points $\{p_i\}$ belong to a square centered at the origin with size $K$. The problem is to evaluate the potentials $\{u_i,1\le i \le N\}$ defined by \begin{equation} u_i = \sum_{j=1}^N G(p_i,p_j) \cdot f_j \label{eq:nbody} \end{equation} where $G(x,y) = \frac{\i}{4} H^{(1)}_0(2\pi |x-y|)$ is the fundamental solution of the 2D Helmholtz equation. In this paper, we use $\i$ to denote $\sqrt{-1}$. This computational problem mostly arises from the numerical solution of 2D time harmonic scattering problems \cite{colton-1983-iemst}. For example, suppose that $D \subset \mathbb{R}^2$ is a compact object with a smooth boundary and $u^{inc}$ is the incoming field. If $D$ represents a sound soft scatterer, the scattering field $u$ satisfies the following Helmholtz equation with the Dirichlet boundary condition: \[ -\Delta u - (2\pi)^2 u = 0 \quad\mbox{in}\; \mathbb{R}^d \setminus \bar{D} \] \[ u(x) = - u^{inc}(x) \quad\mbox{for}\; x\in\partial D \] \[ \lim_{r\rightarrow\infty} r \left( \frac{\partial u}{\partial r} - 2\pi\i u \right) = 0 \] where the wave number is set to be $2\pi$. The last condition is the Sommerfeld radiation condition and guarantees that the scattering field $u$ propagates to infinity. One highly efficient way to solve this problem is to reformulate it into an equivalent boundary integral equation (BIE) \begin{equation} \frac{1}{2} \phi(x) + \int_{\partial D} \left( \frac{\partial G(x,y)}{\partial n(y)} - \i \eta G(x,y) \right) \phi(y) d y = - u^{inc}(x) \label{eq:bie} \end{equation} where $n(y)$ is the exterior normal of $\partial D$ at $y$, $\eta$ is some fixed constant, and $\phi(x)$ for $x\in \partial D$ is the unknown charge distribution on the boundary $\partial D$. Once $\phi$ is solved from \eqref{eq:bie}, the scattering field $u$ can be simply computed with an integral formula \cite{colton-1983-iemst}. The BIE approach has the advantage of reducing the number of unknowns. The discrete version of \eqref{eq:bie}, however, is a dense linear system which usually requires an iterative algorithm like GMRES \cite{saad-1986-gmres} for its solution. At each step of the iterative solver, we then need to evaluate the computational problem in \eqref{eq:nbody}, with $\{p_i\}$ being the appropriate quadrature points. It is well known that the complexity of a scattering problem often scales with the size of scatterer in terms of the wavelength. Since the wavelength is taken to be 1 in our setup, the complexity of \eqref{eq:nbody} depends on the number $K$, which can be of order $10^4$ for a typical large scale scattering problem. Since one often uses a constant number of points per wavelength when discretizing \eqref{eq:bie}, the number of points $N$ is proportional to $K$. \subsection{Previous work} Direct computation of \eqref{eq:nbody} takes $O(N^2)$ operations. This can be quite time consuming when $N$ is large. Various fast algorithms have been proposed to reduce this complexity in the past two decades. Among them, the most popular approach is the high frequency fast multipole method (HF-FMM) developed by Rokhlin et al. \cite{cheng-2006-riwfmm2d,rokhlin-1990-rsiest}. In the HF-FMM, the whole computational domain is partitioned into a quadtree and one associates with each square of the quadtree two expansions: the far field expansion and the local field expansion \cite{cheng-2006-riwfmm2d}. These expansions allow one to accelerate the computation in the low frequency region. In the high frequency region, the Fourier transforms of these expansions are used instead to achieve optimal efficiency since the translations between them become diagonal operators under the Fourier basis. The HF-FMM has an optimal $O(N\log N)$ complexity and has been widely used. A different approach is to discrete the integral equation \eqref{eq:bie} under the Galerkin framework using local Fourier bases or wavelet packets. The stiffness matrix becomes approximately sparse under these bases since most of the entries are close to zero and can be safely discarded. Early algorithms \cite{averbuch-2000-ecoi,bradie-1993-fnc,canning-1992-sasieok,deng-1999-fseie,deng-1999-cpwpb,golik-1998-wpfseie} of this approach focus on finding the correct one dimensional basis, while a recent development \cite{huybrechs-2006-twtmci} considers the use of two dimensional wave packets which can offer more flexibility and better compression rate. Another early development is the multilevel matrix decomposition by Michielssen and Boag \cite{michielssen-1996-mmda}. The three stage multiplication algorithm, which is later named the butterfly algorithm by \cite{oneil-2007-ncabft}, is quite similar to the FFT and brings the overall complexity down to $O(N \log^2 N)$. In \cite{engquist-2007-fdmaok}, we proposed an algorithm for the three dimensional $N$-body problem of the high frequency Helmholtz kernel. It relies on a low rank property of the 3D Helmholtz kernel for certain geometric configurations. The algorithm organizes the computation in a multidirectional and multilevel fashion and has an optimal $O(N \log N)$ complexity. \subsection{A multidirectional approach} In this paper, we adapt the approach in \cite{engquist-2007-fdmaok} to the two dimensional $N$-body problem of the Helmholtz kernel. The main idea is a similar low rank property of the 2D Helmholtz kernel. We say that two sets $Y$ and $X$ satisfy the {\em directional parabolic separation condition} if $Y$ is a disk of radius $r$ and $X$ is the set of points that belong to a cone with spanning angle $1/r$ and are at least $r^2$ away from $Y$ (see Figure \ref{fig:xryr}). \begin{figure} \begin{center} \includegraphics[height=1.5in]{Slide1.jpg} \end{center} \caption{Two sets $Y$ and $X$ that satisfy the directional parabolic separation condition.} \label{fig:xryr} \end{figure} Once $Y$ and $X$ satisfy the directional parabolic separation condition, one can show that for any fixed accuracy the interaction between $X$ and $Y$ via the Helmholtz kernel $G(x,y)$ is approximately of low rank and the rank is independent of $r$. More precisely, for any accuracy $\varepsilon$, there exist a constant $T(\varepsilon)$ and two sets of functions $\{\alpha_i(x), 1\le i \le T(\varepsilon) \}$ and $\{\beta_i(y), 1\le i \le T(\varepsilon) \}$ such that for any $x\in X$ and $y\in Y$ \[ \left| G(x,y) - \sum_{i=1}^{T(\varepsilon)} \alpha_i(x) \beta_i(y) \right| \le \varepsilon \] (see Theorem \ref{thm:dlr}). Notice that $\{\alpha_i(x)\}$ and $\{\beta_i(y)\}$ are only functions of $x$ and $y$ respectively. We call such an approximation a {\em directional separated representation}. One major component of our approach is to use these representations to build equivalent charges for well-separated interaction. Similar to the 3D algorithm in \cite{engquist-2007-fdmaok}, our 2D algorithm starts by generating a quadtree for the whole computational domain. In the low frequency region where the squares are of size less than 1, the interactions are accelerated using the kernel independent FMM algorithm in \cite{ying-2004-kiafmm}. In the high frequency region where the squares are of size greater than or equal to 1, the far field of each square is partitioned into wedges which follow the directional parabolic separation condition (see Figure \ref{fig:onewedge}). Between the square and each of its wedges, the computation is accelerated via the directional separated representation associated with the wedge. \begin{figure} \begin{center} \includegraphics[height=1.5in]{tree.jpg} \hspace{0.1in} \includegraphics[height=1.5in]{Slide2.jpg} \end{center} \caption{Left: the quadtree constructed for a point distribution supported on a curve. Right: for each square $B$ in the high frequency region, its far field is partitioned into multiple wedges. We construct a low rank representation of the interaction between $B$ and each of its wedges. This representation is further used to accelerate the computation between $B$ and all the squares in the wedge. } \label{fig:onewedge} \end{figure} Apart from extending the multidirectional algorithm of \cite{engquist-2007-fdmaok} to the 2D Helmholtz kernel, this paper also contains two new contributions: \begin{itemize} \item We provide an improved randomized procedure for the construction of the directional separated representations. The new procedure is more efficient and generates representations with smaller ranks. \item Our algorithm has been applied to the solution of \eqref{eq:bie}. This allows us to study large scatterers that are thousands of wavelengths wide. \end{itemize} The rest of this paper is organized as follows. In Section \ref{sec:direct}, we briefly summarize the theoretical result on which our approach is based and describe the new improved procedure for constructing the separated representations. After describing our algorithm for \eqref{eq:nbody} in detail in Section \ref{sec:algo}, we present in Section \ref{sec:results} the numerical results for several test examples. Finally, Section \ref{sec:concl} provides some comments on future research directions. Though this paper focuses on the two dimensional Helmholtz kernel, we would like to point out that our algorithm is also applicable to other 2D oscillatory kernels such as $e^{2\pi\i |x-y|}$. \section{Directional Separated Representations} \label{sec:direct} \begin{definition} Let $f(x,y)$ be a function for $x \in X$ and $y \in Y$. We say $f(x,y)$ has a $T$-term $\varepsilon$-expansion for $X$ and $Y$ if there exist functions $\{ \alpha_i(x), 1 \le i \le T\}$ and $\{ \beta_i(y), 1 \le i \le T\}$ such that \[ \left| f(x,y) - \sum_{i=1}^{T} \alpha_i(x) \beta_i(y) \right| \le \varepsilon \] for all $x \in X$ and $y \in Y$. \end{definition} Since the two sets of functions $\{\alpha_i(x)\}$ and $\{\beta_i(y)\}$ depend only on $x$ and $y$ respectively, the above expansion is called {\em separated}. Suppose $r \ge \sqrt{2}$. For our problem, we take \begin{equation} Y = B(0,r) \quad\mbox{and}\quad X = \{ x: \theta(x,\ell) \le 1/r, |x| \ge r^2\} \label{eq:xryr} \end{equation} where $\ell$ is a given unit vector and $\theta(a,b)$ is the spanning angle between vectors $a$ and $b$. The geometric relationship between $Y$ and $X$ is illustrated in Figure \ref{fig:xryr}. The following theorem serves as the theoretical foundation of our approach. \begin{theorem} For any $\varepsilon>0$, there exists a number $T(\varepsilon)$ which is independent of $r$ such that \[ G(x,y) = \frac{\i}{4} H^{(1)}_0(2\pi |x-y|) \] has a $T(\varepsilon)$-term $\varepsilon$-expansion for any $X$ and $Y$ given by \eqref{eq:xryr}. \label{thm:dlr} \end{theorem} The representation guaranteed by Theorem \ref{thm:dlr} is called a {\em directional separated representations} for the obvious reason. One way to prove this theorem is to use the asymptotic behavior of $H^{(1)}_0$ for large arguments \cite{abramowitz-1992-hmf,bronshtein-1997-hm}: \[ H^{(1)}_0 (r) = \sqrt{ \frac{2}{\pi r} } \left( e^{\i (r-\pi/4)} + O\left(\frac{1}{r}\right) \right), \] and then follow the same path as the proof for Theorem 2.2 in \cite{engquist-2007-fdmaok}. \subsection{Construction of directional separated representation} A procedure based on random sampling has been described in \cite{engquist-2007-fdmaok} for the construction of these directional separated representations. In the rest of this section, we propose an improved version which gives lower separation ranks and better accuracy based on our numerical experience. For a given pair $Y$ and $X$ that satisfy the directional parabolic separation condition, our new procedure takes the following steps: \begin{enumerate} \item Sample $Y$ randomly with a set of samples $\{y_j, 1\le j \le N_Y\}$. In our implementation, we use 2 to 3 points per wavelength and the number of samples $N_Y$ grows linearly with the area of $Y$. Sample $X$ similarly with a set of samples $\{x_i, 1\le i \le N_X\}$. Let $A$ be the matrix defined by \[ A_{ij} = G(x_i, y_j) = \frac{\i}{4} H^{(1)}_0(2\pi |x_i-y_j|), \] for $ 1\le i \le N_X$ and $1\le j \le N_Y$. In the language of linear algebra, Theorem \ref{thm:dlr} states that $A$ can be factorized, within error $O(\varepsilon)$, into the product of two matrices, the first containing $T(\varepsilon)$ columns and the second containing $T(\varepsilon)$ rows. \item Let $A_1$ be the submatrix of $A$ containing a set of $N_1$ randomly selected rows. Here we set $N_1 \approx 3 \cdot T(\varepsilon)$ in practice. Our goal is to find a set of $T(\varepsilon)$ columns of $A_1$ that has the largest $T(\varepsilon)$-dimensional volume. Since $A_1$ is only of size $O(T(\varepsilon)) \times N_Y$, one can use either the interpolative decomposition \cite{cheng-2005-clrm} or the greedy standard pivoted QR factorization to find these columns. Both algorithms have an $O(N_Y)$ complexity. Suppose the pivoted QR factorization is used. We then have the decomposition \[ A_1 P_1 = Q_1 R_1, \] where $P_1$ is a permutation matrix, $Q_1$ is orthonormal, and $R_1$ is upper triangular. Now identify the diagonal elements of $R_1$ which are less than $\varepsilon$ and truncate the associated columns of $Q_1$ and rows of $R_1$. Denote the resulting matrices by $Q_{1,c}$ and $R_{1,c}$. Since $A_1$ itself has an $O(T(\varepsilon))$-expansion, $Q_{1,c}$ contains only $O(T(\varepsilon))$ columns in practice. Moreover, it is clear that \[ Q_{1,c} R_{1,c} = A_{1,c}, \] where $A_{1,c}$ is the submatrix containing the columns of $A_1$ from which the matrix $Q_{1,c}$ is generated. We denote by $A_c$ the submatrix of $A$ that consists of the same columns. The $O(T(\varepsilon))$ samples of $Y$ associated with these columns are denoted $\{b_q\}$. \item Let $A_2$ be a submatrix of $A$ containing a set of $N_2$ randomly selected columns. We again set $N_2 \approx 3 \cdot T(\varepsilon)$. Repeat the previous step on $A_2^*$. As a result, we have two matrices $Q_{2,r}$ and $R_{2,r}$. $Q_{2,r}$ is orthonormal and has $O(T(\varepsilon))$ columns again, while $R_{2,r}$ is upper triangular. They satisfy the relationship \[ R_{2,r}^* Q_{2,r}^* = A_{2,r}, \] where $A_{2,r}$ is a submatrix containing appropriate rows of $A$. We denote by $A_r$ the submatrix of $A$ that consists of the same rows and by $\{a_p\}$ the $O(T(\varepsilon))$ samples of $X$ associated with these rows (see Figure \ref{fig:samples}). \item We randomly pick a set $S$ of $N_S$ rows and a set $T$ of $N_T$ columns. In practice, we choose $N_S$ and $N_T$ to be equal to $10 \cdot T(\varepsilon)$. Set $A_3$ to be the minor containing the elements from rows in $S$ and columns in $T$, $A_{c,S}$ to be the submatrix of $A_c$ containing the rows in $S$, and $A_{r,T}$ to be the submatrix of ${A_r}$ containing the columns in $T$. Next, we choose $ D = (A_{c,S})^+ A_3 (A_{r,T})^+, $ where $(\;)^+$ stands for pseudoinverse. We claim that \[ \left| A - A_c D A_r \right| = O(\varepsilon). \] Such an approximate factorization is often called a pseudoskeleton approximation of $A$ in the literature (see \cite{goreinov-1997-tpa,goreinov-1997-pasgs}). Notice that the matrix $D$ has only $O(T(\varepsilon))$ rows and columns. Denoting the entries of $D$ by $d_{qp}$, we can rewrite the previous statement in the form \[ \left| G(x_i,y_j) - \sum_{p,q} G(x_i,b_q) \cdot d_{qp}\cdot G(a_p,y_j) \right| = O(\varepsilon) \] for all $x_i$ and $y_j$. \item Finally, since $\{x_i\}$ and $\{y_j\}$ sample the sets $X$ and $Y$ with a constant number of points per wavelength, it is reasonable to expect \begin{equation} \left| G(x,y) - \sum_{p,q} G(x,b_q) \cdot d_{qp}\cdot G(a_p,y) \right| = O(\varepsilon) \label{eq:fnlexp} \end{equation} for any $x \in X \cap B(0,K)$ and $y \in Y$. \end{enumerate} \begin{figure} \begin{center} \includegraphics[height=1.5in]{Slide6.jpg} \end{center} \caption{Constructions of the separated representation between $X$ and $Y$. $\{b_q\}$ are the samples associated with the columns in $A_c$ (Step 2). $\{a_p\}$ are the samples associated with the columns in $A_r$ (Step 3). } \label{fig:samples} \end{figure} Since both $\{a_p\}$ and $\{b_q\}$ are of order $O(T(\varepsilon))$, it is clear that \eqref{eq:fnlexp} is a low rank separated representation. Moreover, we only need to store $\{a_p\}$, $\{b_q\}$, and $D$ for \eqref{eq:fnlexp}, thus reducing the storage requirement dramatically. We would like to point out that recently there has been a lot of research devoted to problems similar to \eqref{eq:fnlexp} (see \cite{bebendorf-2003-alacm,drineas-2006-fmcam2,drineas-2006-fmcam3,martinsson-2006-rafam} for details). This randomized procedure performs quite well in practice as we will see from the numerical results in Section \ref{sec:results}. Though we do not yet have a proof, the following heuristic argument provides some useful insights. In the standard pseudoskeleton approximation \cite{goreinov-1997-tpa,goreinov-1997-pasgs}, an $m \times n$ matrix $A$ has the following approximation: \[ A \approx A_c G A_r, \] where $A_c$, $G$, and $A_r$ are of size $m \times k$, $k \times k$, and $k \times n$ respectively. Often $A_c$ contains the columns of $A$ that have the largest $k$-dimensional volume and, similarly, $A_r$ contains the rows with the largest $k$-dimensional volume. Finding these columns and rows are quite expensive if both $m$ and $n$ are large. Suppose now that we can project the columns (or rows) of $A$ onto a $p$ dimensional subspace $L$ which is randomly selected from all $p$-dimensional subspaces with the uniform rotational invariant probability measure. As long as $p$ is adequately larger than $k$, the volume spanned by any set of $k$ columns (or rows) is preserved to a good accuracy \cite{dasgupta-2003-eptjl,magen-2002-dr}. Therefore, one efficient method to find the columns of $A$ with the largest volume would be to \begin{enumerate} \item project $A$ onto a random $p$ dimensional subspace, \item find the columns of the projected matrix that have the largest $k$-dimensional volume, \item pick the corresponding columns of $A$ to be the answer. \end{enumerate} The only difference between this approach and the second and third steps of our randomized procedure is that we only project to a random set of coordinates, which is much more restrictive than the uniform random projection. However, since both the columns and the rows of our matrix $A$ is highly oscillatory and incoherent with the Dirac functions, our procedure works well in practice. \subsection{Equivalent charges} The directional separated representation \eqref{eq:fnlexp} provides a way to represent the potential in $X$ generated by the charges inside $Y$ in a compact way. Suppose that $X$ is centered around the unit direction $\ell$ and $\{f_i\}$ are the charges located at points $\{y_i\}$ in $Y$. After applying \eqref{eq:fnlexp} to $y=y_i$ for each $y_j$ and summing them up with weight $f_i$, we have \[ \left| \sum_i G(x,y_i) f_i - \sum_q G(x,b_q) \left( \sum_p d_{qp} \sum_i G(a_p,y_i) f_i \right) \right| = O(\varepsilon). \] This states that we can place a set of charges \begin{equation} \left\{ \sum_p d_{qp} \sum_i G(a_p,y_i) f_i \right\} \label{eq:doed} \end{equation} at points $\{b_q\}$ in order to reproduce the potential generated by the charges $\{f_i\}$ located at points $\{y_i\}$. We call the charges in \eqref{eq:doed} the {\em directional outgoing equivalent charges} of $Y$ in direction $\ell$ and the points $\{b_q\}$ the {\em directional outgoing equivalent points} of $Y$ in direction $\ell$. In addition, we refer to the quantities \begin{equation} \left\{ \sum_i G(a_p,y_i) f_i \right\} \label{eq:docp} \end{equation} as the {\em directional outgoing check potentials} of $Y$ in direction $\ell$ and the points $\{a_p\}$ as the {\em directional outgoing check points} of $Y$ in direction $\ell$. Given the check potentials, the equivalent charges can be computed easily by a multiplication with $D$. Let us now reverse the role of $X$ and $Y$. Suppose we have a set of charges $\{f_i\}$ located at points $\{x_i\}$ in $X$. Since $G(x,y) = G(y,x)$, \[ \left| \sum_i G(y,x_i) f_i - \sum_p G(y,a_p) \sum_q d_{qp} \sum_i G(b_q,x_i) f_i \right| = O(\varepsilon). \] This states that we can put a set of charges \begin{equation} \left\{ \sum_q d_{qp} \sum_i G(b_q,x_i) f_i \right\} \label{eq:died} \end{equation} at points $\{a_p\}$ and they reproduce the potential generated by the charges $\{f_i\}$ located at points $\{x_i\}$. Therefore, we call the charges in \eqref{eq:died} the {\em directional incoming equivalent charges} of $Y$ in direction $\ell$ and the locations $\{a_p\}$ the {\em directional incoming equivalent points} of $Y$ in direction $\ell$. In analogy to the previous terminology, \begin{equation} \left\{ \sum_i G(b_q,x_i) f_i \right\} \label{eq:dicp} \end{equation} are called the {\em directional incoming check potentials} of $Y$ in direction $\ell$ and the location $\{b_q\}$ are called the {\em directional incoming check points} of $Y$ in direction $\ell$. \section{Algorithm Description} \label{sec:algo} Without loss of generality, we assume that the size of the domain $K = 2^{2L}$ for a positive integer $L$. \subsection{Data structure} We start by constructing a quadtree which contains the whole computational domain. We often use $B$ to denote a square in the quadtree and $w$ for its width. A square $B$ is said to be in the low frequency regime if $w < 1$ and in the high frequency regime if $w \ge 1$. In the high frequency regime of the quadtree, no adaptivity is used, i.e., every non-empty square is further partitioned until the width of the square is less than $1$. In the low frequency regime, a square $B$ is partitioned as long as the number of points in $B$ is greater than a fixed constant $N_p$. The value of $N_p$ is chosen to optimize the computational complexity and, in practice, we pick $N_p = 50$. For a square $B$ in the low frequency regime, its data structure follows the description of the kernel independent FMM in \cite{ying-2004-kiafmm}. The near field $N^B$ is the union of the squares $A$ that satisfies $dist(A,B) = 0$, where $dist(A,B) = \inf_{x\in A,y\in B} |x-y|$. The far field $F^B$ is the complement of $N^B$. The interaction list $I^B$ contains all the squares in $N^P \backslash N^B$ on $B$'s level, where $P$ is the parent square of $B$. \begin{itemize} \item $\{y^{B,o}_k\}$, $\{f^{B,o}_k\}$, $\{x^{B,o}_k\}$ and $\{u^{B,o}_k\}$ are, respectively, the {\em outgoing} equivalent points, equivalent charges, check points, and check potentials. \item $\{y^{B,i}_k\}$, $\{f^{B,i}_k\}$, $\{x^{B,i}_k\}$ and $\{u^{B,i}_k\}$ are, respectively, the {\em incoming} equivalent points, equivalent charges, check points, and check potentials. \end{itemize} For a square $B$ in the high frequency region, the near field $N^B$ is the union of all the squares $\{A\}$ that satisfy $dist(A,B) \le w^2$. The far field $F^B$ is the complement of $N^B$. The interaction list $I^B$ contains all the squares in $N^P \backslash N^B$ on $B$'s level, where $P$ is $B$'s parent square. Notice that the far field of a square $B$ in the high frequency region is pushed away in order to be compatible with the directional parabolic separation condition. The far field $F^B$ is further partitioned into a group of directional wedges, each belonging to a cone with spanning angle $O(1/w)$. We denote the set of all the wedges of $B$ by $\{W^{B,\ell}\}$. In Figure \ref{fig:six}, we illustrate the case for for $w=1,2,4$. \begin{figure} \begin{center} \includegraphics[height=1.8in]{Slide3.jpg} \includegraphics[height=1.8in]{Slide4.jpg} \includegraphics[height=1.8in]{Slide5.jpg} \end{center} \caption{The far field is partitioned into wedges. From left to right, $w=1,2,4$. The radii are 1,4, and 16, respectively. } \label{fig:six} \end{figure} For each square $B$ and each direction $\ell$, we summarize the relevant quantities as follows: \begin{itemize} \item $\{y^{B,o,\ell}_k\}$, $\{f^{B,o,\ell}_k\}$, $\{x^{B,o,\ell}_k\}$, and $\{u^{B,o,\ell}_k\}$ are the {\em outgoing directional} equivalent points, equivalent charges, check points and check potentials respectively. \item $\{y^{B,i,\ell}_k\}$, $\{f^{B,i,\ell}_k\}$, $\{x^{B,i,\ell}_k\}$, and $\{u^{B,i,\ell}_k\}$ are the {\em incoming directional} equivalent points, equivalent charges, check points and check potentials respectively. \end{itemize} \subsection{Translation operators} Following the convention in \cite{greengard-1987-afaps,rokhlin-1990-rsiest}, we name these operators M2M, L2L, and L2L translations, though no multipole or local expansions are involved in our algorithm. The translation operators for squares in the low frequency regime are detailed already in \cite{ying-2004-kiafmm}. The operators in the high frequency regime are more complicated. The main reason is that the computations are now directional. For a square $B$ in the high frequency regime, the {\em M2M translation} constructs the outgoing directional equivalent charges of $B$ from the outgoing equivalent charges of $B$'s children. There are two cases to consider. In the first case, $w=1$. The children squares have only nondirectional equivalent charges. The M2M translation iterates over all of the directional indices $\{\ell\}$ of $B$, and the steps for a fixed direction $\ell$ are as follows: \begin{enumerate} \item Use $\bigcup_{C} \{y^{C,o}_k\}$ as source points in $B$ and $\bigcup_{C} \{f^{C,o}_k\}$ as source charges. Here the union is taken over all of the children squares of $B$. \item Compute $\{u^{B,o,\ell}_k\}$ at points $\{x^{B,o,\ell}_k\}$ with kernel evaluation, and then obtain $\{f^{B,o,\ell}_k\}$ by multiplying $\{u^{B,o,\ell}_k\}$ with the matrix $D$ associated with the wedge $W^{B,\ell}$. \end{enumerate} In the second case, $w>1$. Now the children squares have directional equivalent charges as well. The M2M translation iterates over all of the directional indices $\{\ell\}$ of $B$. The steps for a fixed direction $\ell$ are as follows: \begin{enumerate} \item Pick $\ell'$, a direction associated with the squares of width $w/2$, such that the wedge $W^{B,\ell}$ is contained in the wedge $W^{C,\ell'}$ where $C$ stands for anyone of $B$'s children. The existence of $\ell'$ is ensured by the way we partition $F^B$ (see Figure \ref{fig:ellell}). \item Use $\bigcup_{C} \{ y^{C,o,\ell'}_k\}$ as source points in $B$ and $\bigcup_{C} \{f^{C,o,\ell'}_k\}$ as source charges. Here the union is taken over all the children squares of $B$. \item Compute $\{u^{B,o,\ell}_k\}$ at $\{x^{B,o,\ell}_k\}$ with kernel evaluation and then obtain $\{f^{B,o,\ell}_k\}$ by multiplying $\{u^{B,o,\ell}_k\}$ with the matrix $D$ associated with the wedge $W^{B,\ell}$. \end{enumerate} \begin{figure} \begin{center} \includegraphics[height=1.5in]{Slide7.jpg} \end{center} \caption{$B$ is a square with width $w>1$. For any fixed $\ell$, there exists $\ell'$ such that $W^{B,\ell}$ is contained in $W^{C,\ell'}$ where $C$ is any one of $B$'s children.} \label{fig:ellell} \end{figure} The {\em L2L translation} constructs the incoming check potentials of $B$'s children from the incoming directional check potentials of $B$. Again there are two cases to consider. In the first case $w=1$. The children squares have only nondirectional check potentials. The L2L translation iterates over all of the directional indices $\{\ell\}$ of $B$, and the steps for a fixed direction $\ell$ are as follows: \begin{enumerate} \item Compute $\{f^{B,i,\ell}_k\}$ from $\{u^{B,i,\ell}_k\}$ by multiplying it with the appropriate $D$ matrix. \item For each child $C$ of the square $B$, add to $\{u^{C,i}_k\}$ the potentials evaluated at $\{x^{C,i}_k\}$ using $\{f^{B,i,\ell}_k\}$ as the source charges at $\{y^{B,i,\ell}_k\}$. \end{enumerate} In the second case, $w>1$. Now the children squares have directional equivalent charges. The L2L translation iterates over all of the directional indices $\{\ell\}$ of $B$. The steps for a fixed direction $\ell$ are as follows: \begin{enumerate} \item Pick $\ell'$, a direction associated with the squares of width $w/2$, such that the wedge $W^{B,\ell}$ is contained in the wedge $W^{C,\ell'}$ where $C$ stands for anyone of $B$'s children. \item Compute $\{f^{B,i,\ell}_k\}$ from $\{u^{B,i,\ell}_k\}$ by multiplying it with the appropriate $D$ matrix. \item For each child $C$ of the square $B$, add to $\{u^{C,i,\ell'}_k\}$ the potentials evaluated at $\{x^{C,i,\ell'}_k\}$ using $\{f^{B,i,\ell}_k\}$ as the source charges at $\{y^{B,i,\ell}_k\}$. \end{enumerate} Finally, the {\em M2L translation} is applied to pairs of squares $A$ and $B$ on the same level of the quadtree. They need to be on each other's interaction lists. Suppose $B$ falls into the wedge $W^{A,\ell}$ of $A$ while $A$ falls into the wedge $W^{B,\ell'}$ of $B$. The implementation of the M2L translation contains only one step: \begin{enumerate} \item Add to $\{u^{B,i,\ell'}_k\}$ the potentials evaluated at $\{x^{B,i,\ell'}_k\}$ using the charges $\{f^{A,o,\ell}_k\}$ at points $\{y^{A,o,\ell}_k\}$. \end{enumerate} To summarize the discussion on the transition operators, we would like to emphasize that all of these operators involve only kernel evaluation and matrix-vector multiplication with precomputed matrices. Therefore, they are simple to implement and highly efficient. \subsection{Algorithm} \label{sec:algo-algo} \begin{figure} \begin{center} \includegraphics[height=3.2in]{graph.jpg} \end{center} \caption{A small part of the quadtree used in the computation. Each rectangular region stands for a square of the quadtree. The diagram shows how the outgoing nondirectional equivalent charges from a leaf square have been transformed into incoming nondirectional check potentials at other leaf squares. Far field interaction involves directional computation in the high frequency regime.} \label{fig:algo} \end{figure} Now we are ready to give the overall structure of our new algorithm. It has exactly the same structure as the 3D algorithm in \cite{engquist-2007-fdmaok} and we simply reproduce it here: \begin{enumerate} \item Construct the quadtree. In the high frequency regime, the squares are partitioned uniformly. In the low frequency regime, a leaf square contains at most $N_p$ points. \item Travel up in the quadtree and visit the squares in the low frequency regime. These squares have width less than 1. For each square $B$, compute its outgoing nondirectional equivalent charges $\{f^{B,o}_k\}$. This is done using the low frequency nondirectional M2M translation. \item Travel up in the quadtree and visit the squares in the high frequency regime. For every such square $B$, use the high frequency directional M2M translation to compute the outgoing directional equivalent charges $\{f^{B,o,\ell}_k\}$ for each outgoing direction $\ell$. We skip the squares with width greater than $\sqrt{K}$ since their interaction lists are empty. \item Travel down in the quadtree and visit the squares in the high frequency regime. For every such square $B$ and for each direction $\ell$, perform the following two steps: \begin{enumerate} \item Transform the outgoing directional equivalent charges $\{f^{A,o,\ell}_k\}$ of all of the squares $\{A\}$ in $B$'s interaction list and in direction $\ell$ via the high frequency directional M2L translation. Next, add the result to the incoming directional check potentials $\{u^{B,i,\ell}_k\}$. \item Perform the high-frequency directional L2L translation to transform $\{u^{B,i,\ell}_k\}$ to the incoming check potentials for $B$'s children. \end{enumerate} Again, we skip the squares with width greater than $\sqrt{K}$. \item Travel down in the quadtree. For every square $B$ in the low frequency regime, we perform the following two steps: \begin{enumerate} \item Transform the outgoing nondirectional equivalent charges $\{f^{A,o}_k\}$ of all of the squares $\{A\}$ in $B$'s interaction list via the low frequency nondirectional M2L operator. Next, add the result to the incoming nondirectional check potentials $\{u^{B,i}_k\}$. \item Perform the low frequency directional L2L translation. Depending on whether $B$ is a leaf square or not, add the result to the incoming check potentials of $B$'s children or to the potentials at the original points inside $B$. \end{enumerate} \end{enumerate} An illustration of the various components of the algorithm is given in Figure \ref{fig:algo}. The following theorem summarizes the complexity of the proposed algorithm. \begin{theorem} Let $\S$ be a rectifiable curve in $B(0,1/2)$. Suppose that for a fixed $K$ the points $\{p_i, 1 \le i \le N\}$ are samples of $K\S$, where $N=O(K)$ and $K\S = \{K \cdot p, p\in \S\}$ (the surface obtained by magnifying $\S$ by a factor of $K$). Then, for any prescribed accuracy, the proposed algorithm has a computational complexity $O(K \log K) = O(N\log N)$. \end{theorem} The proof of this theorem follows closely the steps of Theorem 4.1 of \cite{engquist-2007-fdmaok}. The main step of the proof is the observation that, for any fixed $w>1$, there are at most $O(K/w)$ squares of size $w$ and, for each of them, there are at most $O(w)$ squares for which we apply the M2L operator. \section{Numerical Results} \label{sec:results} In this section, we provide some numerical results to illustrate the properties of our new algorithm. All of the computational results below are obtained on a desktop computer with a 2.8 GHz CPU. Let us first study the performance of the randomized procedure presented in Section \ref{sec:direct}. In Table \ref{tbl:rank}, we list the number of terms in the separated representation for two sets $X$ and $Y$ for different choices of accuracy $\varepsilon$ and square width $w$. Here $r$, the radius of $Y$, is set to be $\sqrt{2} w$ so that the square of width $w$ is contained in $Y$. We can see from Table \ref{tbl:rank} that the separation rank is bounded by a constant which is independent of the values of $w$. This is consistent with our theoretical estimate in Theorem \ref{thm:dlr}. In fact, as $w$ grows, it seems that the separation rank decays slightly. \begin{table} \begin{center} \begin{tabular}{|c|cccccccc|} \hline & $w=1$ & $w=2$ & $w=4$ & $w=8$ & $w=16$ & $w=32$ & $w=64$ & $w=128$ \\ \hline $\varepsilon$=1e-4 & 14 & 11 & 11 & 10 & 9 & 9 & 9 & 9\\ $\varepsilon$=1e-6 & 19 & 16 & 14 & 13 & 12 & 12 & 12 & 11\\ $\varepsilon$=1e-8 & 27 & 20 & 16 & 15 & 15 & 15 & 14 & 14\\ \hline \end{tabular} \end{center} \caption{The separation rank of the directional separated representation for different choices of requested accuracy $\varepsilon$ and square size $w$.} \label{tbl:rank} \end{table} Next, we applied our algorithm to the $N$-body problems on several objects. In our experiments, the boundary of each object is represented by a piecewise smooth curve. For these tests, the point set $\{p_i\}$ is generated by sampling the curve randomly with about $20$ points per wavelength. The densities $\{f_i\}$ are generated from a random distribution with mean $0$. We use $\{u_i\}$ to denote the true discrete potentials and $\{u_i^a\}$ to denote the approximations obtained through our algorithm. We estimate the relative error by picking a set $S$ of $200$ points from $\{p_i\}$. The true potentials $\{u_i, i\in S\}$ are computed by using direct evaluation. The error is then estimated to be \[ \sqrt{ \frac{ \sum_{i\in S} |u_i - u_i^a |^2 } { \sum_{i\in S} |u_i|^2 } }. \] Before reporting the results, let us summarize the notations we use here: $N$ is the number of points, $K$ is the size of the problem in terms of the wavelength, $\varepsilon$ is the prescribed error threshold such that the final error is to be bounded by a constant multiple of $\varepsilon$, $T_a$ is the running time of our algorithm in seconds, $T_d$ is the running time of the direct evaluation in seconds, $T_d / T_a$ is the speedup factor, and $\varepsilon_a$ is the resulting error of our algorithm. The first example is a circle and the results are summarized in Table \ref{tbl:circ}. The second example is an airfoil and the results are shown in Table \ref{tbl:foil}. The final example is a kite-shaped object and we report the numbers in Table \ref{tbl:kite}. These numbers demonstrate clearly that our algorithm scales exactly like $O(N\log N)$ in terms of the number of points. Furthermore, the error seems to grow only slightly as we increase the number of points. \begin{table}[h] \begin{center} \includegraphics[height=2in]{circ.pdf} \includegraphics[height=2in]{circ_tree.jpg}\\ \vspace{0.1in} \begin{tabular}{|c|ccccc|} \hline $(K,\varepsilon)$ & $N$ & $T_a$(sec) & $T_d$(sec) & $T_d / T_a$ & $\varepsilon_a$\\ \hline (2048,1e-4) & 1.13e+5 & 3.40e+1 & 8.05e+3 & 2.37e+2 & 1.25e-4\\ (8192,1e-4) & 4.50e+5 & 1.56e+2 & 1.28e+5 & 8.21e+2 & 1.31e-4\\ (32768,1e-4)& 1.80e+6 & 7.07e+2 & 2.06e+6 & 2.91e+3 & 1.80e-4\\ \hline (2048,1e-6) & 1.13e+5 & 5.30e+1 & 8.00e+3 & 1.51e+2 & 7.88e-7\\ (8192,1e-6) & 4.50e+5 & 2.39e+2 & 1.28e+5 & 5.37e+2 & 9.98e-7\\ (32768,1e-6)& 1.80e+6 & 1.08e+3 & 2.06e+6 & 1.91e+3 & 1.00e-6\\ \hline (2048,1e-8) & 1.13e+5 & 8.20e+1 & 8.05e+3 & 9.82e+1 & 8.48e-9\\ (8192,1e-8) & 4.50e+5 & 3.57e+2 & 1.29e+5 & 3.60e+2 & 1.18e-8\\ (32768,1e-8)& 1.80e+6 & 1.58e+3 & 2.07e+6 & 1.31e+3 & 1.30e-8\\ \hline \end{tabular} \end{center} \caption{Results of a circle with the Helmholtz kernel. $N$ is the number of points, $K$ is the size of the problem in terms of the wavelength, $\varepsilon$ is the prescribed error threshold such that the final error is to be bounded by a constant multiple of $\varepsilon$, $T_a$ is the running time of our algorithm in seconds, $T_d$ is the running time of the direct evaluation in seconds, $T_d / T_a$ is the speedup factor, and $\varepsilon_a$ is the estimated error of our algorithm.} \label{tbl:circ} \end{table} \begin{table}[h] \begin{center} \includegraphics[height=2in]{foil.pdf} \includegraphics[height=2in]{foil_tree.jpg}\\ \vspace{0.1in} \begin{tabular}{|c|ccccc|} \hline $(K,\varepsilon)$ & $N$ & $T_a$(sec) & $T_d$(sec) & $T_d / T_a$ & $\varepsilon_a$\\ \hline (2048,1e-4) & 7.82e+4 & 2.00e+1 & 3.87e+3 & 1.94e+2 & 1.15e-4\\ (8192,1e-4) & 3.13e+5 & 8.80e+1 & 6.17e+4 & 7.02e+2 & 1.21e-4\\ (32768,1e-4)& 1.25e+6 & 3.90e+2 & 9.90e+5 & 2.54e+3 & 1.07e-4\\ \hline (2048,1e-6) & 7.82e+4 & 3.20e+1 & 3.87e+3 & 1.21e+2 & 1.04e-6\\ (8192,1e-6) & 3.13e+5 & 1.38e+2 & 6.20e+4 & 4.50e+2 & 9.65e-7\\ (32768,1e-6)& 1.25e+6 & 6.05e+2 & 1.01e+6 & 1.67e+3 & 1.20e-6\\ \hline (2048,1e-8) & 7.82e+4 & 4.70e+1 & 3.87e+3 & 8.24e+1 & 8.58e-9\\ (8192,1e-8) & 3.13e+5 & 2.03e+2 & 6.22e+4 & 3.06e+2 & 1.69e-8\\ (32768,1e-8)& 1.25e+6 & 8.78e+2 & 9.95e+5 & 1.13e+3 & 1.33e-8\\ \hline \end{tabular} \end{center} \caption{Results of an airfoil with the Helmholtz kernel.} \label{tbl:foil} \end{table} \begin{table}[h] \begin{center} \includegraphics[height=2in]{kite.pdf} \includegraphics[height=2in]{kite_tree.jpg}\\ \vspace{0.1in} \begin{tabular}{|c|ccccc|} \hline $(K,\varepsilon)$ & $N$ & $T_a$(sec) & $T_d$(sec) & $T_d / T_a$ & $\varepsilon_a$\\ \hline (2048,1e-4) & 1.13e+5 & 4.00e+1 & 8.11e+3 & 2.03e+2 & 1.08e-4\\ (8192,1e-4) & 4.53e+5 & 1.77e+2 & 1.30e+5 & 7.36e+2 & 1.33e-4\\ (32768,1e-4)& 1.81e+6 & 8.04e+2 & 2.09e+6 & 2.60e+3 & 1.41e-4\\ \hline (2048,1e-6) & 1.13e+5 & 6.10e+1 & 8.11e+3 & 1.33e+2 & 9.35e-7\\ (8192,1e-6) & 4.53e+5 & 2.72e+2 & 1.30e+5 & 4.78e+2 & 9.15e-7\\ (32768,1e-6)& 1.81e+6 & 1.24e+3 & 2.10e+6 & 1.70e+3 & 8.80e-7\\ \hline (2048,1e-8) & 1.13e+5 & 9.20e+1 & 8.16e+3 & 8.87e+1 & 1.45e-8\\ (8192,1e-8) & 4.53e+5 & 4.05e+2 & 1.30e+5 & 3.22e+2 & 1.31e-8\\ (32768,1e-8)& 1.81e+6 & 1.80e+3 & 2.11e+6 & 1.17e+3 & 1.52e-8\\ \hline \end{tabular} \end{center} \caption{Results of a kite-shaped model with the Helmholtz kernel.} \label{tbl:kite} \end{table} Compared with the results presented in \cite{cheng-2006-riwfmm2d}, our algorithm is slower by a factor of 8. The reason is that we heavily use the kernel evaluation formula in our algorithm. The 2D Helmholtz kernel involves the Hankel functions and the current computational procedure for their evaluation is rather slow. On the other hand, all of the high frequency translations in \cite{cheng-2006-riwfmm2d} are precomputed and stored in the diagonal form and no special function evaluation is required during the computation. Finally, we apply our algorithm to the solution of the BIE formulation \[ \frac{1}{2} \phi(x) + \int_{\partial D} \left( \frac{\partial G(x,y)}{\partial n(y)} - \i \eta G(x,y) \right) \phi(y) d y = - u^{inc}(x) \] of the 2D scattering problem mentioned in Section \ref{sec:intro}. Here, we report the numerical results for the smooth objects in Tables \ref{tbl:circ} and \ref{tbl:kite}. In our experiments, we use a uniform discretization of about 20 points per wavelength. We pick $\eta = \pi$ and set the incoming field $u^{inc}(x)$ to be $e^{2\pi\i x \cdot d}$ with $d=(1,0)$. We discretize the integral equation with the Nystr\"{o}m method \cite{colton-1983-iemst,kress-1999-lie} and use the endpoint-corrected trapezoidal rules from \cite{kapur-1997-hctqr} to integrate the weakly singular part of the integral. The system is solved iteratively using the GMRES algorithm and the restarted number is set to be $80$. Within each iteration of the GMRES solver, the application of the integral operator is accelerated using our multidirectional algorithm with $\varepsilon=$1e-4. Table \ref{tbl:sccirc} summarizes the results for the circle with wavelengths from 1024 to 8192. Here $T_i$ is the averaged time of each iteration, $N_i$ is the number of iterations, and $T_t$ is the total time. Table \ref{tbl:sckite} reports the results of the kite-shaped object in Table \ref{tbl:kite}. In Figure \ref{fig:kite}, we display the scattering field of the kite-shaped object in a region with caustics. \begin{table}[h] \begin{center} \begin{tabular}{|c|cccc|} \hline $K$ & $N$ & $T_i$(sec) & $N_i$ & $T_t$(sec) \\ \hline 1024 & 65536 & 22 & 72 & 1.60e+3\\ 2048 & 131072 & 45 & 93 & 4.32e+3\\ 4096 & 262144 & 99 & 118 & 1.20e+4\\ 8192 & 524288 & 202 & 150 & 3.12e+4\\ \hline \end{tabular} \end{center} \caption{ Timings of computing the scattering field of the circle. $K$ is the size of the problem in terms of the wavelength, $N$ is the number of quadrature points, $T_i$ is the averaged time of each iteration, $N_i$ is the number of iterations, and $T_t$ is the total time. } \label{tbl:sccirc} \end{table} \begin{table}[h] \begin{center} \begin{tabular}{|c|cccc|} \hline $K$ & $N$ & $T_i$(sec) & $N_i$ & $T_t$(sec) \\ \hline 1024 & 65536 & 22 & 227 & 5.11e+3\\ 2048 & 131072 & 46 & 314 & 1.49e+4\\ 4096 & 262144 & 99 & 435 & 4.42e+4\\ 8192 & 524288 & 204 & 604 & 1.25e+5\\ \hline \end{tabular} \end{center} \caption{ Timings of computing the scattering field of the kite-shaped object. } \label{tbl:sckite} \end{table} \begin{figure} \begin{center} \includegraphics[height=2in]{sctr1.pdf} \includegraphics[height=3in]{sctr2.pdf} \end{center} \caption{Scattering field of the kite-shaped object with $K=1024$. Top: a square region that contains the caustics. Bottom: the real part of the scattering field inside the square. The field is sampled at 8 points per wavelength. } \label{fig:kite} \end{figure} \section{Conclusions} \label{sec:concl} In this paper, we described a directional multiscale algorithm for computing the $N$-body problem for the high frequency Helmholtz kernel in two dimensions. The approach follows the framework described in \cite{engquist-2007-fdmaok}. Our algorithm is accurate and works well for problems in all scales. By using the directional low rank representations for regions that follow the directional parabolic separation condition, our algorithm achieves the optimal $O(N\log N)$ complexity. A new and more efficient randomized technique compared to the one in \cite{engquist-2007-fdmaok} has also been introduced for the construction of the low rank separated representations. The numerical results have shown that our algorithm is capable of addressing very large scale problems in high frequency scattering. For future work, we would like to have a rigorous proof for the randomized procedure proposed in Section \ref{sec:direct}. Another interesting direction for future research is to apply this kind of directional multiscale idea to other problems with oscillatory behavior, in both two and three dimensions. One typical example is the computation of the far field pattern of a scattering field \cite{colton-1983-iemst, ying-2007-sftba}. {\bf Acknowledgments.} The authors would like to thank P.G. Martinsson for helpful discussions. B.E. is partially supported by an NSF grant DMS 0714612 and a startup grant from the University of Texas at Austin. L.Y. is partially supported by an Alfred P. Sloan Research Fellowship and a startup grant from the University of Texas at Austin. \bibliographystyle{abbrv}

https://arxiv.org/abs/1210.2707

Analytical mechanics of a relativistic particle in a positional potential

We propose a form for the action of a relativistic particle subject to a positional force that is invariant under time reparametrization and therefore allows for a consistent Hamiltonian formulation of the dynamics. This approach can be useful in the study of phenomenological models. Also the Dirac and Klein-Gordon equation differ from the standard formulation, with corrections of order (E-m)/m in the energy spectra.

\section{1. Introduction} Analytical mechanics is at the basis of modern theoretical physics. In particular, Hamiltonian methods are largely employed in several applications and are essential for the formulation of the quantum theory. However, not all systems have been consistently described in Hamiltonian terms. For example, if one excludes fundamental interaction, like Maxwell theory, whose coupling to particles is dictated by gauge invariance, in special relativity it is not easy to introduce particle interactions that preserve the Lorentz and the reparametrization invariance of the action. While for phenomenological interactions the request of Lorentz invariance is not compelling, the breakdown of the reparametrization invariance has serious consequences on the possibility of defining a consistent Hamiltonian for the model under study. The lack of a Hamiltonian formulation also implies that in quantum theory the Klein-Gordon or Dirac equations for generic interactions have to be postulated rather than derived from the correspondence principle. These problems arise especially for models like that of an external central force acting on a particle, when the force is fixed a priori and is not determined by field equations derived from a variational principle. An important case is that of the harmonic oscillator, whose relativistic formulation is problematic [1]. Usually, in such cases the Hamiltonian is defined only for a specific choice of the time coordinate. In this paper, we propose a different solution that, slightly modifying the coupling, makes the action reparametrization invariant. This allows to obtain a Hamiltonian by means of the usual Dirac formalism for constrained systems. The corrections with respect to the standard formalism are of order $(E-m)/m$, where $E$ is the relativistic energy and $m$ the mass of the particle. Moreover, a really covariant Klein-Gordon or Dirac equation can be obtained. In the case of the harmonic oscillator, a proposal leading to equivalent results was advanced long ago in [1], starting from different considerations. The paper is organized as follows: in sect.\ 2 we review the Lagrangian and Hamiltonian formalism for a free particle. In sect.\ 3 we describe our proposal and compare it with the standard formulation of the relativistic interacting particle. In sect.\ 4 we give some elementary instances of application of the formalism. In sect.\ 5 the changes in the Klein-Gordon equations are discussed and illustrated with simple examples in sect.\ 6. \section{2. Free particle} In special relativity, the motion of a free particle in \coo $x^\m$ is determined by the variation of the action $$S=\int{\cal L}}\def\cH{{\cal H}\,d\t=-m\int\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}\,d\t=-m\int ds,\eqno(1)$$ where ${\cal L}}\def\cH{{\cal H}$ is the Lagrangian density, $\t$ an arbitrary evolution parameter, a dot denotes a derivative with respect to }\def\ie{i.e.\ }\def\tls{transformation laws $\t$, $m$ is the rest mass and $s$ the proper time on the trajectory. The action is invariant under Lorentz transformations and under reparametrizations $\t\to\t'(\t)$. The Euler-Lagrange equations for the action (1) read $$m\,{d\over d\t}\,{\dot x^\m\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}=0.\eqno(2)$$ The dynamics can also be written in Hamiltonian form [2]. The momentum conjugated to $x^\m$ is defined as $$p_\m={\de{\cal L}}\def\cH{{\cal H}\over\de\dot x^\m}=m{\dot x^\m\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}=m{dx^\m\over ds},\eqno(3)$$ and satisfies the constraint $$p_\m^2=m^2.\eqno(4)$$ Since the action is reparametrization invariant, the Hamiltonian $\cH$ vanishes, and the action can be written in first-order form as $$S=\int d\t(\dot x^\m p_\m-\l\cH),\qquad\cH=p_\m^2-m^2,\eqno(5)$$ where $\l$ is a Lagrange multiplier enforcing the constraint (4). The Hamilton equations ensuing from the action (5) are $$\dot x_\m=\{x_\m,\l\cH\}=2\l p_\m,\qquad\dot p_\m =\{p_\m,\l\cH\}=0,\eqno(6)$$ from which, comparing with (3), follows that $$\l={\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}\over2m}.\eqno(7)$$ \medskip In Dirac formalism, the constraint (4) is first order. One can therefore reduce the system by choosing a gauge, \ie fixing the time coordinate. The standard choice is $t=x^0$, namely one identifies the evolution time $t$ with the coordinate time $x_0$. Since $\{t,\l\cH\}=2\l p_0$, no secondary constraints arise and one easily checks that $$x_i'={p_i\over p_0}={p_i\over\sqrt{p_k^2+m^2}},\qquad p_i'=0, \qquad\l={1\over2m}\xdt={1\over2p_0}\eqno(8)$$ where $i=1,2,3$ and a prime denotes a derivative with respect to }\def\ie{i.e.\ }\def\tls{transformation laws $t$. The reduced action is then $$S=\int dt\(x'_ip_i-\sqrt{p_i^2+m^2}\),\eqno(9)$$ and the 3-dimensional effective Hamiltonian $H=p_0=\sqrt{p_i^2+m^2}$ can be identified with the energy of the particle in the laboratory frame. Other choices of the gauge are however possible. For example, with the choice $t=x^\m p_\m$, the evolution time coincides with the proper time of the particle. \section{3. Particle in an external potential} The addition of an external potential acting on the free particle presents some problems and to our knowledge has not been discussed in depth. In the standard formalism, one simply adds to the action a potential term $$S_{int}=-\int V(x,\dot x)\,d\t.\eqno(10)$$ In general the potential breaks the Lorentz invariance. Moreover, unless $V$ is homogeneous of degree one in the velocity, as in the Maxwell case when $V=\dot x^\m A_\m(x)$, $S_{int}$ is not invariant under reparametrization, and the Dirac formalism cannot be consistently applied for writing down the \eom in Hamiltonian form. In the following we discuss the case of positional potentials $V=V(x)$ that do not depend on the velocity of the particle. These usually arise as phenomenological potentials, like those associated to elastic forces. For positional potentials, the interaction term (10) leads to the standard \eom $${d\over d\t}\ {m\,\dot x^\m\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}- \de^\m V=0.\eqno(11)$$ Multiplying eq.\ (11) by $\dot x_\m$ and integrating, one can show that the total energy $$E={m\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}+V\eqno(12)$$ is conserved. Since $V=V(x)$, the momenta are given by (3) as for the free particle and satisfy the constraint (4). If one tries to apply the Dirac formalism to the action so defined, one obtains that $V$ must be a constant. This is a consequence of the lack of reparametrization invariance of (10). \medskip In order to avoid these problems and to preserve reparametrization invariance, we propose an interaction term of the form $$S_{int}=-\int\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}\ V(x)\,d\t,\eqno(13)$$ leading to the total action $$S=-\int\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}\,(m+V)\,d\t.\eqno(14)$$ The Euler-Lagrange equations for the action (14) read then $${d\over d\t}\ {(m+V)\,\dot x^\m\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}-\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}\ \de^\m V=0.\eqno(15)$$ If $V$ does not depend on $\dot x^\m$, eq.\ (15) can also be written $$(m+V){d\over d\t}\ {\dot x^\m\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}+{\dot x^\m\dot x^\n\de_\nu V- (\dot x^\n)^2\de^\m V\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}=0.\eqno(16)$$ One can again find the energy integral by multiplying eq.\ (15) by $\dot x_\m$ and integrating, obtaining $$E={m+V\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}.\eqno(17)$$ To pass to the Hamiltonian formulation, we compute the momentum conjugate to $x^\m$, $$p_\m={\de{\cal L}}\def\cH{{\cal H}\over\de\dot x^\m}=(m+V){\dot x^\m\over\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}}=(m+V){dx^\m\over ds}, \eqno(18)$$ which is subject to the constraint $$p_\m^2=(m+V)^2.\eqno(19)$$ It follows that the action can be written in Hamiltonian form as $$S=\int d\t(\dot x^\m p_\m-\l\cH),\qquad\cH=p_\m^2-(m+V)^2\eqno(20)$$ where $\l$ is a Lagrange multiplier enforcing the constraint (19). The Hamilton equations read $$\dot x_\m=\{x_\m,\l\cH\}=2\l p_\m,\qquad\dot p_\m =\{p_\m,\l\cH\}=-2\l(m+V)\de_\m V, \eqno(21)$$ from which follows that $$\l={\sqrt{(\dot x^\r)^2}}\def\xdt{\sqrt{1-x_k'^2}\over2(m+V)}.\eqno(22)$$ In our formalism, the potential essentially plays the role of a variable mass added to the rest mass of the particle. This is in accord with special relativity, since the external potential is fixed and does not change with the motion of the particle. So its field cannot contribute to the energy balance in any other way than effectively modifying the mass of the particle. \medskip Again, one can choose a gauge in order to reduce to a three-dimensional problem with external time. Taking $t=x^0$, one has $\{t-x_0,\l\cH\}=2\l p_0$, and hence no secondary constraints arise if the potential does not depend on $x_0$. Moreover $$x_i'={p_i\over p_0}={p_i\over\sqrt{p_i^2+(m+V)^2}},\qquad p_i'=-{m+V\over p_0}\ \de_i V, \qquad\l={\xdt\over2(m+V)}={1\over2p_0}.\eqno(23)$$ In this case the reduced action is $$S=\int dt\(x'_ip_i-\sqrt{p_i^2+(m+V)^2}\).\eqno(24)$$ The 3-dimensional Hamiltonian $H=\sqrt{p_i^2+(m+V)^2}$ represents the energy in the laboratory frame. It should be compared with the expression adopted in the standard formalism, namely $H=\sqrt{p_i^2+m^2}+V$. This is obtained from the reduced lagrangian of the standard formalism in the gauge $t=x_0$ and has no covariant meaning. The two Hamiltonians differ by terms of order $(E-m)/m$. For $m\gg p_i$, $m\gg V$, the expansion of both Hamiltonians gives rise to the classical expression plus higher-order corrections, $$E\sim m+{p_i^2\over2m}+V+\dots\eqno(25)$$ However, with our Hamiltonian higher-order corrections are present in the expansion also for the potential term. Like for the free particle, also in presence of a potential term, one can choose a different gauge. For example, in 1+1 dimensions, the evolution time can be made to coincide with the proper time of the particle for the choice $t=(m+V)\(1+\int^x{mdx\over(m+V)^3}\)p$. \section{4. Examples} We give two elementary examples that we solve in the Lagrangian formalism, and compare them with the results obtained using the standard formalism. The potentials are not Lorentz invariant and therefore the results hold only in the rest reference frame. However the \eom are reparametrization invariant. \subsect{4.1 Harmonic oscillator} For a 1-dimensional system, the \eom are immediately integrated. Choosing $t=x_0$, the conserved energy (17) can be written $$E={m+V\over\sqrt{1-{x'}^2}}.\eqno(26)$$ Easy calculations lead to the differential equation $$x'={1\over E}\ \sqrt{E^2-(m+V)^2}.\eqno(27)$$ that can be immediately integrated. In the case of a harmonic oscillator, $V=\ha kx^2$, the same equation has been obtained in [1], starting from a different perspective. The solution of (27) is given by [1] $$x=\sqrt{E^2-m^2\over kE}\ {\rm sd}\!\!\(\sqrt{k\over E}\ t,{E-m\over2E}\),\eqno(28)$$ with period $$T=4\sqrt{E\over k}\ {\rm K}\!\!\({E-m\over2E}\)\sim2\pi}\def\q{\psi}\def\r{\rho}\def\s{\sigma}\def\t{\tau}\def\u{\upsilon\sqrt{m\over k} \(1+{5\over8}\ {E-m\over m}+\dots\),\eqno(29)$$ where sd and K are elliptic functions. Contrary to nonrelativistic mechanics, the period of the oscillations is not constant, but depends on the energy. We have written down the first term of its expansion in $(E-m)/m$ around the nonrelativistic value. Note that in the standard formalism, the equation of motion would have read $$x'={\sqrt{(E-V)^2-m^2}\over E-V}\eqno(30)$$ Also this equation can be solved in terms of elliptic functions [3]. In particular, the period of the solution is $$T\sim2\pi}\def\q{\psi}\def\r{\rho}\def\s{\sigma}\def\t{\tau}\def\u{\upsilon\sqrt{m\over k} \(1+{3\over8}\ {E-m\over m}+\dots\)\eqno(31)$$ whose dependence on energy differs from (29). \subsect{4.2 Kepler problem} It is well known that the Kepler problem can be reduced to a one-dimensional problem because of the presence of two integrals of the motion related to the conservation of the angular momentum. This implies that the motion occurs on a plane and that the component of the angular momentum orthogonal to the plane is conserved. Taking $t=x_0$ and polar \coo $r$ and $\varphi}\def\w{\varpi}\def\y{\eta}\def\x{\xi}\def\z{\zeta$ on the plane of the motion, from (17) one has $$E={m+V\over\sqrt{1-{r'}^2-r^2{\varphi}\def\w{\varpi}\def\y{\eta}\def\x{\xi}\def\z{\zeta'}^2}},\eqno(32)$$ while from the conservation of the norm of the angular momentum follows that $$l={mr^2\varphi}\def\w{\varpi}\def\y{\eta}\def\x{\xi}\def\z{\zeta'\over\sqrt{1-{r'}^2-r^2{\varphi}\def\w{\varpi}\def\y{\eta}\def\x{\xi}\def\z{\zeta'}^2}}\eqno(33)$$ is a constant. After standard computations from (32) and (33) one obtains the equation of the orbits, $$\({dr\over d\varphi}\def\w{\varpi}\def\y{\eta}\def\x{\xi}\def\z{\zeta}\)^2+r^2={r^4\over l^2}\ [E^2-(m+V)^2]\eqno(34)$$ Defining $u={1\over r}$ and $V=-{\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon\over r}=-\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon u$, eq.\ (34) becomes $$\({du\over d\varphi}\def\w{\varpi}\def\y{\eta}\def\x{\xi}\def\z{\zeta}\)^2+u^2={1\over l^2}\ [E^2-(m-\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon u)^2]\eqno(35)$$ whose solution is given by $$r={1\over u}={p\over1+\varepsilon\cos q(\varphi}\def\w{\varpi}\def\y{\eta}\def\x{\xi}\def\z{\zeta-\varphi}\def\w{\varpi}\def\y{\eta}\def\x{\xi}\def\z{\zeta_0)},\eqno(36)$$ with $$q=\sqrt{1+{\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon^2\over l^2}},\qquad p={l^2+\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon^2\over\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon m},\qquad \varepsilon={E\over m}\ \sqrt{1+{(E^2-m^2)\,l^2\over\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon^2E^2}}.\eqno(37)$$ These values should be compared with those obtained using the standard formalism, namely $$q=\sqrt{1-{\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon^2\over l^2}},\qquad p={l^2-\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon^2\over\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon E},\qquad \varepsilon={m\over E}\ \sqrt{1+{(E^2-m^2)\,l^2\over\alpha}\def\b{\beta}\def\c{\chi}\def\d{\delta}\def\e{\epsilon^2m^2}}.\eqno(38)$$ In the formulae above, $\varepsilon$ is the eccentricity of the orbit, while the parameter $p$ is related to the size of the orbit and $q$ to the perihelion advance $\d=2\pi}\def\q{\psi}\def\r{\rho}\def\s{\sigma}\def\t{\tau}\def\u{\upsilon(q^\mo-1)$. The angle $\d$ has opposite value in our formalism with respect to }\def\ie{i.e.\ }\def\tls{transformation laws the standard one. Of course, the correct value is that given by general relativity and is three times the standard one. The other parameters of the orbit calculated in our formalism differ as usual for terms of order $(E-m)/m$ from the standard ones. \section{5. Quantization} A first quantized relativistic equation can be obtained as usual starting from the Hamiltonian (20). The Klein-Gordon equation is obtained with the substitution $p_\m\to i\hbar\de_\m$, namely $$\hbar^2\({\de^2\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta\over\de t^2}-{\de^2\q\over\de x_i^2}\)+(m+V)^2\q=0.\eqno(39)$$ Setting $\q=\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta\,e^{-iEt/\hbar}$, it reduces to $$-\hbar^2{\de^2\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta\over\de x_i^2}+[(m+V)^2-E^2]\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta=0,\eqno(40)$$ that has the form of a \schr equation with $E_{eff}=E^2-m^2$, $V_{eff}=2mV+V^2$. Note that if the potential $V$ is positive definite, also $V_{eff}$ is. This is not true in the standard formalism, leading to problems with stability. In fact, in the standard formalism, the Klein-Gordon equation is assumed to be $$-\hbar^2{\de^2\q\over\de x_i^2}+[m^2-(E-V)^2]\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta=0.\eqno(41)$$ It also has the form of a \schr equation with $E_{eff}=E^2-m^2$, $V_{eff}=2EV-V^2$. Note that this equation cannot be derived from a classical Hamiltonian, since that is not well defined. Moreover, the presence of $E$ in the effective potential makes it difficult to find the spectrum of the energy explicitly. Finally, we notice that also the Dirac equation can be modified in accord with (39), as $$[i\hbar\g^\m\de_\m-(m+V)]\q=0.\eqno(42)$$ \section{6. Examples} We give two simple examples of solution of the Klein-Gordon equation. In the first case, only a perturbative solution is possible, while the second is constructed so that an exact solution can be found. For simplicity, in this section we set $\hbar=1$. \subsect{6.1. Quantum harmonic oscillator} The Klein-Gordon equation (39) for a one-dimensional harmonic oscillator with potential $V=\ha kx^2$ reads $${d^2\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta\over dx^2}=(\m x^4+mk\,x^2-\e)\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta,\eqno(43)$$ where $\e=E^2-m^2$, $\m=k^2/4$, and has the form of a nonrelativistic biquadratic oscillator. Its energy spectrum can be obtained using standard perturbation theory, taking the quartic term as a perturbation of the nonrelativistic harmonic oscillator. The calculation has been performed in [1] and gives at first order $$E_n=\pm\[m+\sqrt{k\over m}\(n+\ha+{3\over4}\,\m\(n^2+n+\ha\)\)\],\eqno(44)$$ where $n$ is an integer. In the standard formalism, the Klein-Gordon equation (38) reads instead $${d^2\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta\over dx^2}=(-\m x^4+Ek\,x^2-\e)\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta,\eqno(45)$$ with $\e$ and $\m$ as before, and the effective potential is no longer positive definite, leading to possible instabilities [4]. Neglecting this problem, one can obtain as before the energy spectrum, which at first order in $\hbar$, and neglecting corrections of order $(E-m)/m$, yields the opposite sign for the correction, namely $$E_n=\pm\[m+\sqrt{k\over m}\(n+\ha-{3\over4}\,\m\(n^2+n+\ha\)\)\].\eqno(46)$$ \vfil\eject \subsect{6.2. Exactly solvable potential} An instance in which a solution of the Klein-Gordon equation (39) can be found explicitly is $$V=m\({1\over\cos x}-1\).\eqno(47)$$ This is a potential well whose value is zero at the origin and infinite at $x=\pm{\pi}\def\q{\psi}\def\r{\rho}\def\s{\sigma}\def\t{\tau}\def\u{\upsilon\over2}$. The Klein-Gordon equation reads $${d^2\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta\over dx^2}-\[{m^2\over\cos^2x}-E^2\]\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta=0.\eqno(48)$$ Defining $z=\sin x$, eq.\ (48) can be put in the form $${d^2\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta\over dz^2}-z\,{d\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta\over dz}-\[{m^2\over(1-z^2)^2}-{E^2\over1-z^2}\]\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta=0, \eqno(49)$$ that in turn can be reduced to a standard hypergeometric equation with solution $$\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta=\cos^\n\!x\ {\rm F}\!\!\(\n+E,\n-E,\n+\ha,{1+\sin x\over2}\),\eqno(50)$$ where F is a hypergeometric function and $\n=\ha\(1+\sqrt{1+4m^2}\)$. Imposing the vanishing of $\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta$ at $x=\pm{\pi}\def\q{\psi}\def\r{\rho}\def\s{\sigma}\def\t{\tau}\def\u{\upsilon\over2}$, one finds the eigenfunctions $$\phi}\def\g{\gamma}\def\h{\theta}\def\i{\iota}\def\j{\vartheta_n=\cos^\n\!x\ C_n^{(\n)}(\sin x),\eqno(51)$$ with $C_n^{(\n)}$ Chebyshev polynomials, and the energy spectrum $$E=\pm\[m^2+\(n+\ha\)\(1+\sqrt{1+4m^2}\)+n^2\]^{1/2},\eqno(52)$$ where $n$ is an integer. In the standard formalism, the Klein-Gordon equation cannot be solved analytically and therefore we do not pursue its investigation here. \section{7. Conclusions} We have shown that it is possible to formulate the problem of the motion of a particle in a positional potential in special relativity in such a way to preserve reparametrization invariance, and hence to give a consistent Hamiltonian formulation in terms of the Dirac formalism for constrained systems. Of course, potentials of the kind investigated in this paper do not occur in fundamental interactions, whose action is already invariant under reparametrization, but can be of interest for phenomenological models. \beginref \ref [1] A.L. Harvey, \PR{D6}, 1474 (1972). \ref [2] A. Hanson, T. Regge and C. Teitelboim, {\sl "Constrained Hamiltonian systems"}, Accademia Nazionale dei Lincei, Rome 1976. \ref [3] L.A. MacColl, Am. J. Phys. {\bf 25}, 535 (1957). \ref [4] P.O. Lipas, Am. J. Phys. {\bf 38}, 85 (1970). \par\endgroup \end Another possibility is to choose $t={x^\m p_\m\over m}$. In this case one easily checks that $$x_i'={p_i\over m},\qquad p_i'=0,\qquad\l={1\over2m}.\eqno(10)$$ Comparing with (), it is evident that in this gauge $t$ coincides with the proper time $s$ along the trajectory. In this case the reduced action is $$S=\int dt\[\dot x_ip_i-{1\over m}\(p_i^2+m^2\)\],\eqno(11)$$ and $H={1\over m}\(p_i^2+m^2\)$ assumes the nonrelativistic form. A second possibility is to choose $t={x^\m p_\m\over m}$. In this case one easily checks that $$x_i'={mp_i\over(m+V)(m+V-x^\m\de_\m V)},\qquad p_i'=-{\de_i V\over m+V-x^\m\de_\m V}, \qquad\l={1\over2(m+V-x^\m\de_\m V)},\eqno(23)$$ and the reduced action reads $$S=\int dt\[x'_ip_i-m{p_i^2+(m+V)^2\over(m+V)(m+V-x^\m\de_\m V)}\].\eqno(24)$$ In this case the reduced Hamiltonian differs from its classical equivalent. Usually these interactions break the Lorentz invariance (for example they depend on the spatial distance or have a fixed center). What is more serious from the point of view of analytical mechanics is that the usual formulation obtained by simply adding to the action the potential energy, is not reparametrization invariant, and hence a consistent Hamiltonian formulation is not possible. In this paper we propose a simple modification of the usual formalism that solves this problem.

https://arxiv.org/abs/1911.09139

Certain hybrid polynomials associated with Sheffer sequences

Inspired by the framework of operational methods and based on the generating functions of Legendre-Gould Hopper polynomials and Sheffer sequences, we discuss certain new mixed type polynomials and their important properties. We show that the use of operational nature allows the relevant polynomials to be unified and general in nature. It is illustrated how the polynomials, we develop, provide an easy derivation of a wide class of new and known polynomials, and their respective properties.

\section{Introduction and preliminaries} Let the two types of series $f(t)$ and $g(t)$ be delta series and invertible series respectively. Let us consider a polynomial sequence $ s_{n} (x)$ satisfying the conditions \begin{equation}\label{1.6} \big<g(t)~f(t)^{k}|~s_{n}(x)\big>=n!~\delta_{n,k}~,\quad \forall~ n, k\ge 0. \end{equation} This sequence is called the Sheffer sequence for $(g(t),f(t))$ which is usually denoted as $s_{n}(x)\sim(g(t)\,f(t))$. If $s_{n}(x)\sim(1,f(t))$, then $s_{n}(x)$ is called the associated sequence for $f(t)$. If $s_{n}(x)\sim(g(t),t)$, then $s_{n}(x)$ is called the Appell sequence (see \cite [ p. 17]{2SRom}). The polynomial sequence $\{s_{n} (x)\}_{n=0}^{\infty }$ ($ s_{n} (x)$ being a polynomial of degree $n$) is also called Sheffer A-type zero \cite [ p.222 (Theorem 72)]{2Rain}, (which we shall hereafter call Sheffer-type), if $s_{n} (x)$ whose exponential generating function is of the form \begin{equation}\label{1.1} A(t)\exp\left(xH(t)\right)=\sum _{n=0}^{\infty}s_{n} (x)\frac{t^{n}}{n!}, \end{equation} where $A(t)$ and $H(t)$ have (at least the formal) expansions: \begin{equation}\label{1.2} A(t)=\sum _{n=0}^{\infty }A_{n} \frac{t^{n} }{n!},\quad A_{0} \ne 0 \end{equation} and \begin{equation}\label{1.3} H(t)=\sum _{n=1}^{\infty }H_{n} \frac{t^{n} }{n!},\quad H_{1} \ne 0, \end{equation} According to Roman \cite [p.18 (Theorem 2.3.4)]{2SRom}, the polynomial sequence $s_{n}(x)$ has the generating function \begin{equation}\label{1.7} \frac {1}{g\big(f^{-1}(t)\big)} \exp \big(xf^{-1}(t)\big)=\sum _{n=0}^{\infty }s_{n} (x)\frac{t^{n} }{n!}, \end{equation} for all $x$ in $\mathbb C$, where $f^{-1}(t)$ is the compositional inverse of $f(t)$.\\ In view of equations \eqref{1.1} and \eqref{1.7}, we have \begin{equation}\label{1.8} A(t)=\frac {1}{g\big(f^{-1}(t)\big)} \end{equation} and \begin{equation}\label{1.9} H(t)=f^{-1}(t). \end{equation} The idea of Quasi-Monomial (QM) treatment of special polynomials emerged within the context of poweroid, suggested by J. F. Steffenson \cite{2JFStef}. The QM principle has been proved to be a powerful tool for the investigation of the properties of a wide class of polynomials like Sheffer polynomials, see for example \cite{2GGDat}. The series of papers developed during the last years, for example the work of Dattoli (see \cite{2Dat}) deepend their roots in QM theory and reformulated the so called study of special polynomials.\\ According to such a point of view, we recall the following definitions:\\ A family of polynomials $p_{n}(x),~({n \in \mathbb{N},~\forall x\in\mathbb{R}})$ is said to be a QM, if a couple of operators, $\hat{M}$ and $\hat{P}$, hereafter called respectively, the multiplicative and derivative operators, do exist and satisfy the rules: \begin{equation}\label{mcap} \hat{M}\, \{p_{n} (x)\}=p_{n+1} (x) \end{equation} and \begin{equation}\label{pcap} \hat{P}\, \{p_{n} (x)\}=n ~p_{n-1} (x). \end{equation} For $n \in \mathbb{N}, x\in\mathbb{R}$, the combinations of the two identities gives \begin{equation}\label{mpcap} \hat{M}\,\hat{P} \{p_{n} (x)\}=n~p_{n} (x) \end{equation} and \begin{equation} \hat{P}\,\hat{M} \{p_{n} (x)\}=(n+1)~p_{n} (x), \end{equation} which allows the commutation relation of the operators to be \begin{equation}\label{p,m} [\hat{P},\hat{M}]=\hat{P}\hat{M}-\hat{M}\hat{P}=\hat{1} \end{equation} and thus can be viewed as the generators of a Weyl group structure. Assuming hereafter, the vacuum is such that $p_{0} (x)=1$, then, from \eqref{mcap}, the family of polynomials $p_{n}(x)$ can be can be generated as: \begin{equation}\label{1.14} \hat{M}^{n}\{1\}=p_{n}(x),\quad \forall n\in\mathbb{N},~\forall x\in\mathbb{R}. \end{equation} The generating function of $p_{n}(x)$ straightforwardly follows from the identity \eqref{1.14}, namely \begin{equation}\label{1.15} \sum _{n=0}^{\infty }p_{n}(x)\frac{t^{n}}{n!}=\sum _{n=0}^{\infty }\frac{t^{n}\hat{M}^{n}}{n!}\{1\}=\exp{(t\hat{M})}\{1\},\quad|t| < \infty. \end{equation} The previous definitions summed up briefly, gives enough content of QM theory to exploit their use on special polynomials. \vspace{.25cm} We recall the Gould Hopper polynomials [\cite{2HWG}, p. 58, (6.2)] (also called sometimes as higher-order Hermite or Kamp\'{e} de F\'{e}riet polynomials) $H_{n}^{(s)(x,y)}$, which are defined as \begin{equation}\label{GHP} H_{n}^{(s)}(x,y)=n!\sum_{k=0}^{[\frac{n}{s}]}\frac{y^{k}x^{n-sk}}{k!\,(n-sk)!}, \end{equation} and their generating function reads \begin{equation}\label{GHPgen} \exp(xt+yt^{s})=\sum_{n=0}^{\infty}\,H_{n}^{(s)}(x,y)\frac{t^{n}}{n!}. \end{equation} They appear as the solution of the generalized heat equation (see \cite{2SGDat}) \begin{equation}\label{heat} \frac{\partial}{\partial y}\,f(x,y)=\frac{\partial^{s}}{\partial{x}^{s}}\,f(x,y)\quad f(x,0)=x^{n}, \end{equation} and under the operational formalism, they are defined as \begin{equation}\label{1.11} H_{n}^{(s)}(x,y)=\exp\left(y\frac{\partial^{s}}{\partial{x}^{s}}\right)\{x^{n}\}. \end{equation} These polynomials are quasi-monomial under the action of the operators (see \cite{2SGDat}) \begin{equation}\label{GHPoper} \aligned &\hat{M}_{H}:=x+sy\frac{\partial^{s-1}}{\partial x^{s-1}},\\ &\hat{P}_{H}:=\frac{\partial}{\partial x}. \endaligned \end{equation} Now we recall the the Legendre polynomials $S_{n}(x,y)$ and $\frac{R_{n}(x,y)}{n!}$ which are defined by the generating functions (see \cite{2GPED}): \begin{equation}\label{leg1} \exp(yt)\,C_{0}(-xt^{2})=\sum_{n=0}^{\infty}S_{n}(x,y) \frac{t^{n}}{n!} \end{equation} and \begin{equation}\label{leg2} C_{0}(xt)\,C_{0}(-yt)=\sum_{n=0}^{\infty}\,\frac{R_{n} (x,y)}{n!}\frac{t^{n}}{n!}, \end{equation} respectively, where $C_0(x)$ denotes the Bessel-Tricomi function of order $0$. The $nth$ order Bessel-Tricomi function $C_{n}(x)$ defined by the following series \cite [p.150]{2Dat}: \begin{equation}\label{tricomi} C_n(x)=x^{-\frac{n}{2}}J_n(2 \sqrt {x})~=~\sum_{k=0}^\infty \frac{(-1)^k\ x^k}{k!\ (n+k)!},\quad n=0,1,2,\ldots\ , \end{equation} with $J_n(x)$ being the ordinary cylindrical Bessel function of first kind \cite{2And}. The operational definition of $0^{th}$-order Bessel-Tricomi function $C_{0}(x)$ is given by: \begin{equation}\label{tricomi-exp} C_{0}(\alpha x)=\exp \left(-\alpha D_{x}^{-1}\right)\{1\}, \end{equation} where $D_{x}^{-1}$ denotes the inverse derivative operator and \begin{equation} D_{x}^{-n}\{1\}=\frac{x^{n}}{n!}. \end{equation} Lately, by taking as base, the Legendre polynomials $S_{n}(x,y)$ and $\frac{R_{n}(x,y)}{n!}$ in the generating function \eqref{GHPgen} of Gould Hopper polynomials (GHP), the author in \cite{ghazala} introduced and studied the Legendre-Gould Hopper polynomials (LeGHP) ${}_S{H}_n^{(s)}(x,y,z)$ and ${}_R{H}_n^{(s)}(x,y,z)$, which are defined, respectively, by the following generating function: \begin{equation}\label{leGHP1} C_0(-xt^2)\exp(yt+zt^{r})=\sum_{n=0}^{\infty}{}_SH_n^{(r)}(x,y,z)\frac{t^n}{n!} \end{equation} and \begin{equation}\label{leGHP2} C_0(xt)\,C_0(-yt)\,\exp(zt^{r})=\sum_{n=0}^{\infty}\frac{{}_RH_n^{(r)}(x,y,z)}{n!}\frac{t^n}{n!}. \end{equation} Systematically, they were found to be quasi-monomial under the action of operators \begin{equation}\label{leghpcap1} \aligned &\hat{M}_{SH}:=y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}},\\ &\hat{P}_{SH}=\frac{\partial}{\partial y} \endaligned \end{equation} and \begin{equation}\label{leghpcap2} \aligned &\hat{M}_{RH}:=-D_x^{-1}+D_y^{-1}+rz\frac{\partial^{r-1}}{\partial y^{r-1}},\\ &\hat{P}_{RH}=-\frac{\partial}{\partial x}x\frac{\partial}{\partial x}, \endaligned \end{equation} respectively. \vspace{.25cm} The LeGHP ${}_SH_{n}^{(s)}(x,y,z)$ and $\frac{{}_RH_{n}^{(s)}(x,y,z)}{n!}$ contain several special polynomials as their particular cases. Their generality can be viewed from the below table:\\ \noindent \textbf{Table 1. Particular cases of the LeGHP ${}_SH_{n}^{(s)}(x,y,z)$ and $\frac{{}_RH_{n}^{(s)}(x,y,z)}{n!}$ }\\ \\ {\tiny{ \begin{tabular}{lllll} \hline \bf{S.} & \bf{Values of } & \bf{Relation Between the }& \bf{Name of }& \bf{Series} \\ \bf{No.}& \bf{the Indices} & \bf{LeGHP~${}_SH_{n}^{(r)}(x,y,z)$,$\frac{{}_RH_{n}^{(r)}(x,y,z)}{n!}$ }&\bf{the Known}& \bf{Definitions} \\ & \bf{and Variables}& \bf{and its Special Case}& \bf{Polynomials}&\\ \hline I.& & & & \\ & $x=0$ & ${}_SH_{n}^{(r)}(0,y,z)={}_SH_{n}^{(r)}(y,z)$ & Gould Hopper \cite{2GGGA} &${}_SH_{n}^{(r)}(x,y)=n!\sum_{k=0}^{\frac{n}{r}}\frac{y^{k}x^{n-rk}}{k!(n-rk)!}$\\ &&& &\\ \hline II. & & &2-Variable & \\ &~$z=0$ &${}_SH_n^{(r)}(x,y,0)={}_{2L}L_{n}(x,y)$ & Legendre - &${}_{2L}L_{n}(x,y)=n!\sum_{s=0}^{\frac{n}{2}}\frac{x^{s}y^{n-2s}}{(s!)^{2}(n-2s)!}$ \\ &&&type \cite{2GAMDat}&\\ \hline III.&i. $r=m;~x=0,$ & & 2-variable -& \\ & ~$y\rightarrow-D_{x}^{-1}$,$z\rightarrow y$ & ${}_SH_{n}^{(m)}(0,D_{x}^{-1},y)={}_{[m]}L_n(x,y)$& generalized Lagurre&${}_{[m]}L_n(x,y)=n!\sum_{k=0}^{\frac{n}{m}}\frac{y^{k}x^{n-mk}}{k![(n-mk)!]^{2}}$ \\ &ii.$r=m; y=0,z\rightarrow y$& ${}_RH_{n}^{(m)}(x,0,y)={}_{[m]}L_n(x,y)$ & type \cite{2GGADat}&\\ \hline IV.& $r=m-1;~x=0,$ & & generalized& \\ & ~$y\rightarrow x$,$z\rightarrow y$ & ${}_SH_{n}^{(m-1)}(0,x,y)=U_n^{(m)}(x,y)$& Chebyshev \cite{2HWG}&$U_n^{(m)}(x,y)=\sum_{k=0}^{\frac{n}{m}}\frac{(n-k)!y^{k}x^{n-mk}}{k!(n-mk)!}$ \\ &\\ \hline V.&i. $r=1;~x=0,z\rightarrow-D_{x}^{-1}$ & & 2-variable -& \\ & ~$y\rightarrow-D_{x}^{-1}$,$z\rightarrow y$ & ${}_SH_{n}^{(1)}(0,y,-D_{x}^{-1})=L_n(x,y)$& Lagurre \cite{2SGDat}&$L_n(x,y)=n!\sum_{s=0}^{n}\frac{(-x)^{s}y^{n-s}}{(s!)^{2}(n-s)!}$ \\ &ii.$r=1; y=0,z\rightarrow y$& ${}_RH_{n}^{(1)}(x,0,y)=L_n(x,y)$&\\ \hline VI. & & &2-Variable & \\ &~$z=0$ &${}_RH_n^{(r)}(x,y,0)=R_{n}(x,y)$ & Legendre \cite{2GDat} &$R_{n}(x,y)=(n!)^{2}\sum_{s=0}^{n}\frac{y^{s}(-x)^{n-s}}{(s!)^{2}[(n-2s)!]^{2}}$ \\ &&&\\ \hline VII. & $x=0,y\rightarrow x, z\rightarrow yD_{y}y$, & ${}_SH_{n}^{(r)}(0,x,yD_{y}y)=e_n^{(r)}(x,y)$ &truncated of order r &$e_n^{(r)}(x,y)=n!\sum_{k=0}^{\frac{n}{r}}\frac{x^{n-rk}y^{K}}{(n-rk)!}$ \\ &&2-variable \cite{2DAT}&\\ \hline VIII. & $r=2,x=0$, & ${}_SH_{n}^{(2)}(0,y,z)=H_n(y,z)$ &2-variable &$H_n(x,y)=n!\sum_{k=0}^{\frac{n}{2}}\frac{x^{n-2k}y^{K}}{k!(n-2k)!}$ \\ &&&Hermite Kamp\'{e} de F\'{e}riet \cite{2GADat}&\\ \hline IX.&i. $r=2$; $x=0$, & ${}_SH_{n}^{(2)}(0,-D_x^{-1},y)=G_{n}(x,y)$ & Hermite& \\ & $y\rightarrow D_{x}^{-1},z\rightarrow y$&${}_RH_{n}^{(2)}(0,x,y)=G_{n}(x,y)$&type \cite{2GPED} & $G_{n}(x,y)=n!\sum_{k=0}^{\frac{n}{2}} \frac{y^{k}x^{n-2k}}{k![(n-2k)!]^{2}}$\\ &$ii. r=2; x=0$,&\\&$y\rightarrow x,z\rightarrow y$&&&\\ \hline X.&i. $x\rightarrow\frac{(x^{2}-1)}{4}$, & ${}_SH_{n}^{(r)}(\frac{(x^{2}-1)}{4},x,0)=P_{n}(x,y)$ \\ & $y\rightarrow x, z=0$&${}_RH_{n}^{(1)}(\frac{(1-x)}{2},\frac{(x+1)}{2},0)=P_{n}(x,y)$ & Legendre \cite{2GDat} & $P_{n}(x,y)=n!\sum_{k=0}^{\frac{n}{2}} \frac{(x^{2}-1)^{k}x^{n-2k}}{2^{2k}(k!)^{2}(n-2k)!}$\\ &$ii. r=1; x\rightarrow\frac{(1-x)}{2}$,&\\&$y\rightarrow \frac{(1+x)}{2},z= 0$&&&\\ \hline XI.& $r=3$; $x\rightarrow zD_{z}z$, & ${}_SH_{n}^{(3)}(zD_{z}z,x,y)=H_{n}^{(3,2)}(x,y,z)$ & Bell-type \cite{2GDattoli}& $ H_{n}^{(3,2)}(x,y,z)=n!\sum_{k=0}^{\frac{n}{3}}\frac{y^{k}H_{n-3k}^{(2)}(x,y)}{k!r!(n-3k-2r)!}$ \\ & $y\rightarrow x$, $z\rightarrow y$ & &&$=n!\sum_{k=0}^{\frac{n}{3}}\sum_{s=0}^{\frac{n-3k}{2}}\frac{y^{k}z^{s}x^{n-3k-2s}}{k!s!(n-3k-2s)!}$ \\ \hline \end{tabular}}}\\ \\ The special cases mentioned in Table 1 will be exploited later to derive the relevant results.\\ In this paper, the families of Legendre-Gould Hopper based Sheffer polynomials are introduced by using the concepts and the methods associated with monomiality principle. In Section 2, we introduce the Legendre-Gould Hopper based Sheffer polynomials (LeGHSP) ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and frame these polynomials within the context of monomiality principle formalism. In Section 3, we consider some examples of these polynomials. In Section 4, the operational and integral representations for the Laguerre-Gould Hopper based Sheffer polynomials are established. In Section 5, results are obtained for the members of Legendre-Gould Hopper based Sheffer and Legendre-Gould Hopper based associated Sheffer polynomial families by considering some members of the Sheffer and associated Sheffer families respectively. \\ \\ \section{Legendre-Gould Hopper based Sheffer polynomials} To introduce the Legendre-Gould Hopper based Sheffer polynomials ( LeGHSP) denoted by ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and ${}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)$ we prove the following results:\\ \begin{theorem}\label{t2.1} The generating function of the Legendre-Gould Hopper based Sheffer polynomials ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ reads \begin{equation}\label{2.1a} \frac {1}{g\big(f^{-1}(t)\big)}C_0\left(-x\big(f^{-1}(t)\big)^2\right)\exp\left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right)=\sum_{n=0}^{\infty}{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)\frac{t^n}{n!}, \end{equation} or \begin{equation}\label{2.1b} A(t)C_0\left(-x\big(H(t)\big)^2\right)\exp\left(yH(t)+z\big(H(t)\big)^r\right)=\sum_{n=0}^{\infty}{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)\frac{t^n}{n!}. \end{equation} \end{theorem} \begin{proof} We follow the operational technique of replacing $x$ in equation \eqref{1.7} by the multiplicative operator ${}_s\hat {M}_{LeG}$ of the LeGHP ${}_S{H}_n^{(s)}(x,y,z)$, we have indeed \begin{equation}\label{2.3} \frac {1}{g\big(f^{-1}(t)\big)} \exp\big({}_s\hat {M}_{LeG}f^{-1}(t)\big)=\sum _{n=0}^{\infty }s_{n}\big({}_s\hat {M}_{LeG}\big) \frac{t^{n} }{n!}, \end{equation} which by using the expression in \eqref{leghpcap1} and the Crofton-type identity \cite [p. 12]{2GPDat} \begin{equation}\label{crofton} f\left(y+m\lambda\frac{d^{m-1}}{dy^{m-1}}\right)\big\{1\big\}=\exp\left(\lambda\frac{d^m}{dy^m}\right)\Big\{f(y)\Big\}, \end{equation} yields \begin{equation} \frac {1}{g\big(f^{-1}(t)\big)}\exp{\Big(z\frac{\partial^{r}}{\partial y^{r}}\Big)} \exp{\left(\left(y+2D_x^{-1}\frac{\partial}{\partial y}\right)f^{-1}(t)\right)}=\sum _{n=0}^{\infty }s_{n}\Big(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\Big) \frac{t^{n} }{n!}. \end{equation} Further using the identity \eqref{crofton} gives \begin{equation}\label{2.6} \frac {1}{g\big(f^{-1}(t)\big)}\exp{\Big(z\frac{\partial^{r}}{\partial y^{r}}\Big)}\exp{\Big(D_x^{-1}\frac{\partial^{2}}{\partial y^{2}}\Big)} \exp({yf^{-1}(t)})=\sum _{n=0}^{\infty }s_{n}\Big(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\Big) \frac{t^{n} }{n!}. \end{equation} Now, using the exponential expansion and \eqref{tricomi-exp} in \eqref{2.6}, gives \begin{equation} \frac {1}{g\big(f^{-1}(t)\big)}C_0\left(-x\big(f^{-1}(t)\big)^2\right)\exp{\Big(z\frac{\partial^{r}}{\partial y^{r}}\Big)} \exp({yf^{-1}(t)})=\sum _{n=0}^{\infty }s_{n}\Big(y+mD_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\Big) \frac{t^{n} }{n!}, \end{equation} which, on further simplification and by denoting the resultant LeGHSP by ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$, namely \begin{equation}\label{2.8} {}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)=s_{n}({}_s\hat{M}_{LeG})=s_{n}\Big(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\Big), \end{equation} the assertion \eqref{2.1a} directly follows. Also, the equivalent generating function in \eqref{2.1b} follows easily in view of equations \eqref{1.8} and \eqref{1.9} . \end{proof} \begin{theorem} The generating function of Legendre-Gould Hopper based Sheffer polynomials $\frac{{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ reads \begin{equation}\label{2.9} \frac {1}{g\big(f^{-1}(t)\big)}C_0\left(xf^{-1}(t)\right)C_{0}(-yf^{-1}(t))\exp\left(z\big(f^{-1}(t)\big)^r\right)=\sum_{n=0}^{\infty}{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)\frac{t^n}{(n!)^{2}}, \end{equation} or \begin{equation}\label{2.10} A(t)C_0\left(xf^{-1}(t)\right)C_{0}(-yf^{-1}(t))\exp\left(z\big(H(t)\big)^r\right)=\sum_{n=0}^{\infty}{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)\frac{t^n}{(n!)^{2}}. \end{equation} \end{theorem} \begin{proof} Following the same procedure as of Theorem \ref{t2.1}, that is, using the multiplicative operator ${}_r\hat{M}_{LeG}$ of the LeGHP $\frac{{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{(n!)}$ given in \eqref{leghpcap2}, we rewrite generating function \eqref{1.7} as \begin{equation} \frac {1}{g\big(f^{-1}(t)\big)} \exp\big({}_r\hat {M}_{LeG}f^{-1}(t)\big)=\sum _{n=0}^{\infty }s_{n}\big({}_r\hat {M}_{LeG}\big) \frac{t^{n} }{n!}. \end{equation} Using the expression of ${}_r\hat{M}_{LeG}$ given in equation \eqref{leghpcap2} and then decoupling the exponential operator in the l.h.s. of the resultant equation by using the Crofton-type identity \cite [p. 12]{2GPDat} \begin{equation} f\left(y+m\lambda\frac{d^{m-1}}{dy^{m-1}}\right)\big\{1\big\}=\exp\left(\lambda\frac{d^m}{dy^m}\right)\Big\{f(y)\Big\}, \end{equation} After some caculations, we find \begin{equation} \frac {1}{g\big(f^{-1}(t)\big)}C_0\left(xf^{-1}(t)\right)C_{0}(-yf^{-1}(t))\exp{\Big(z\frac{\partial^{r}}{\partial y^{r}}\Big)} \exp({f^{-1}(t)})=\sum _{n=0}^{\infty }s_{n}\Big(-D_x^{-1}+D_y^{-1}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\Big) \frac{t^{n} }{n!}, \end{equation} from which the assertion \eqref{leghpcap2} follows. Also, in view of equations \eqref{1.8} and \eqref{1.9}, generating function \eqref{leghpcap2} can be expressed equivalently as equation \eqref{2.10}. \end{proof} In order to show that the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and $\frac{{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ satisfy the monomiality principle, we prove the following results: \begin{theorem} The Legendre-Gould Hopper based Sheffer polynomials ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ are quasi-monomial under the action of the following multiplicative and derivative operators: \begin{equation}\label{2.14} {}_s\hat {M}_{LeGs}=\left(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}-\frac {g^{\prime} \left(\partial_{y}\right)} {g\left(\partial_{y}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{y}\right)}, \end{equation} or \begin{equation}\label{2.15} {}_s\hat {M}_{LeGs}=\left(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}+\frac {A^{\prime} \left(H^{-1}\left(\partial_{y}\right)\right)}{A\left(H^{-1}\left(\partial_{y}\right)\right)}\right)H'\left(H^{-1}\left(\partial_{y}\right)\right) \end{equation} and \begin{equation}\label{2.16} {}_s\hat {P}_{LHs}=f\left(\partial_{y}\right), \end{equation} or \begin{equation}\label{2.17} {}_s\hat {P}_{LHs}=H^{-1}\left(\partial_{y}\right), \end{equation} respectively, where $\partial_{y}:= \frac{\partial}{\partial y}$. \end{theorem} \begin{proof} Consider the following identity: \begin{equation}\label{dy} \partial_{y}~\left\{\exp \left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right)\right\}=f^{-1}(t)~\exp \left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right). \end{equation} Since $f^{-1}$ denotes the compositional inverse of the function $f$ and $f(t)$ has an expansion (1.3a) in powers of $t$, therefore we have \begin{equation}\label{dy2} f\left(\partial_{y}\right)~\left\{\exp\left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right)\right\}=t~\exp\left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right). \end{equation} Differentiating equation \eqref{2.3} partially with respect to $t$ and in view of relation \eqref{2.8}, we find \begin{equation} \left(\left({}_s\hat {M}_{LeG}-\frac{g'\left(f^{-1}(t)\right)}{g\left(f^{-1}(t)\right)}\right) \frac{1}{f'(f^{-1}(t))}\right)\frac {1}{g\big(f^{-1}(t)\big)} \exp{\big({}_s\hat {M}_{LeG} f^{-1}(t)\big)}=\sum _{n=0}^{\infty }{}_{{}_{s}LeG^{(r)}}s_{n+1}(x,y,z)\frac{t^{n} }{n!}, \end{equation} which on using monomiality principle equation \eqref{1.15} with $t=f^{-1}(t)$ gives \begin{equation} \aligned \left(\left(\hat {M}_{LeG}-\frac{g'\left(f^{-1}(t)\right)}{g\left(f^{-1}(t)\right)}\right) \frac{1}{f'(f^{-1}(t))}\right)&\frac {1}{g\big(f^{-1}(t)\big)}C_0\left(-x\big(f^{-1}(t)\big)^2\right)\exp\left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right)\\ =&\sum _{n=0}^{\infty }{}_{{}_{s}LeG^{(r)}}s_{n+1}(x,y,z)\frac{t^{n} }{n!}. \endaligned \end{equation} Since $g(t)$ is an invertible series and $f(t)$ is a delta series of $t$ therefore $\frac{g'\left(f^{-1}(t)\right)}{g\left(f^{-1}(t)\right)}$ and $\frac{1}{f'(f^{-1}(t))}$ possess power series expansions of $f^{-1}(t)$. Thus, in view of relation \eqref{dy}, the above equation becomes \begin{equation} \aligned \left(\left({}_s\hat {M}_{LeG}-\frac{g'\left(\partial_{y}\right)}{g\left(\partial_{y}\right)}\right) \frac{1}{f'(\partial_{y})}\right)&\left\{\frac {1}{g\big(f^{-1}(t)\big)}C_0\left(-x\big(f^{-1}(t)\big)^2\right)\exp\left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right)\right\}\\ =&\sum _{n=0}^{\infty }{}_{{}_{s}LeG^{(r)}}s_{n+1}(x,y,z)\frac{t^{n} }{n!}, \endaligned \end{equation} which on using generating function \eqref{2.1a} becomes \begin{equation} \left(\left({}_s\hat {M}_{LeG}-\frac{g'\left(\partial_{y}\right)}{g\left(\partial_{y}\right)}\right) \frac{1}{f'(\partial_{y})}\right)\left \{\sum _{n=0}^{\infty }{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)\frac{t^{n} }{n!}\right \}=\sum _{n=0}^{\infty }{}_{{}_{s}LeG^{(r)}}s_{n+1}(x,y,z)\frac{t^{n} }{n!}, \end{equation} or, in an equal manner, it becomes \begin{equation} \sum _{n=0}^{\infty}\left(\left({}_s\hat {M}_{LeG}-\frac{g'\left(\partial_{y}\right)}{g\left(\partial_{y}\right)}\right) \frac{1}{f'(\partial_{y})}\right)\left\{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)\right\}\frac{t^{n} }{n!}=\sum _{n=0}^{\infty }{}_{{}_{s}LeG^{(r)}}s_{n+1}(x,y,z)\frac{t^{n} }{n!}. \end{equation} Now, equating the coefficients of like powers of $t$ in the above equation, we find \begin{equation} \left(\left({}_s\hat {M}_{LeG}-\frac{g'\left(\partial_{y}\right)}{g\left(\partial_{y}\right)}\right) \frac{1}{f'(\partial_{y})}\right)\{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)\}={}_{{}_{s}LeG^{(r)}}s_{n+1}(x,y,z), \end{equation} which, in view of equation \eqref{mcap} shows that the multiplicative operator for ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ is given as: $${}_s\hat{M}_{LeGs}=\left({}_s\hat {M}_{LeG}-\frac{g'\left(\partial_{y}\right)}{g\left(\partial_{y}\right)}\right) \frac{1}{f'(\partial_{y})}.$$ Finally, using equation \eqref{leghpcap1} in the r.h.s of above equation, we get assertion \eqref{2.9}. \\ Again, in view of identity \eqref{dy2}, we have \begin{equation} \aligned f \left(\partial_{y}\right)& \left\{\frac {1}{g\big(f^{-1}(t)\big)}C_0\left(-x\big(f^{-1}(t)\big)^2\right)\exp\left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right)\right\}\\ =&t \frac {1}{g\big(f^{-1}(t)\big)}C_0\left(-x\big(f^{-1}(t)\big)^2\right)\exp\left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right), \endaligned \end{equation} which on using generating function \eqref{2.1a} becomes \begin{equation} f \left(\partial_{y}\right) \left \{\sum _{n=0}^{\infty }{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)\frac{t^{n} }{n!}\right \}=\sum _{n=1}^{\infty }{}_{{}_{s}LeG^{(r)}}s_{n-1}(x,y,z)\frac{t^{n} }{(n-1)!}, \end{equation} or \begin{equation} \sum _{n=0}^{\infty }f \left(\partial_{y}\right) \{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)\} \frac{t^{n} }{n!}=\sum _{n=1}^{\infty }{{}_{{}_{s}LeG^{(r)}}s_{n-1}(x,y,z)}\frac{t^{n} }{(n-1)!}. \end{equation} Equating the coefficients of like powers of $t$ in the above equation, we get \begin{equation} f \left(\partial_{y}\right) \{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)\}=n~{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z),\quad(n \ge 1), \end{equation} which in view of equation \eqref{pcap} yields assertion \eqref{2.16}. Also, in view of relations \eqref{1.8} and \eqref{1.9}, assertions \eqref{2.14} and \eqref{2.16} can be expressed equivalently as equations \eqref{2.15} and \eqref{2.17}, respectively. \end{proof} \begin{theorem} The Legendre-Gould Hopper based Sheffer polynomials $\frac{{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ are quasi-monomial under the action of the following multiplicative and derivative operators: \begin{equation}\label{2.30} {}_r\hat {M}_{LeGs}=\left(-D_x^{-1}+D_y^{-1}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}-\frac {g^{\prime} \left(\partial_{y}\right)} {g\left(\partial_{y}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{y}\right)}, \end{equation} and \begin{equation}\label{2.31} {}_r\hat {P}_{LHs}=f\left(\frac{-\partial}{\partial x}x\frac{\partial}{\partial_{y}}\right), \end{equation} respectively, where $\partial_{y}:= \frac{\partial}{\partial y}$. \end{theorem} \begin{remark} In view of equation \eqref{1.14} and using equations \eqref{2.9} and \eqref{2.10}, we deduce the following consequence of Theorem 2.2. \end{remark} \begin{corollary} The Legendre-Gould Hopper based Sheffer polynomials ${}_{{}_{s}LeG^{(r)}}s_{n-1}(x,y,z)$ have the following explicit representations: \begin{equation*} {}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)={}_s\hat {M}_{LeGs}^n\{1\} \end{equation*} that is \begin{equation} {}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)=\left(\left(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}-\frac {g^{\prime} \left(\partial_{y}\right)} {g\left(\partial_{y}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{y}\right)}\right)^{n}\{1\},\end{equation} or, \begin{equation} {}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)=\left(\left(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\right)H'\left(H^{-1}\left(\partial_{y}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(\partial_{y}\right)\right)}{A\left(H^{-1}\left(\partial_{y}\right)\right)}\right)^{n}\{1\}. \end{equation} \end{corollary} \begin{theorem} The Legendre-Gould Hopper based Sheffer polynomials ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and $\frac{{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ satisfy the following respective differential equation: \begin{equation}\label{2.34} \left(\left(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}-\frac {g^{\prime} \left(\partial_{y}\right)} {g\left(\partial_{y}\right)}\right)\frac {f \left(\partial_{y}\right)} {f^{\prime}\left(\partial_{y}\right)}-n\right){}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)=0 \end{equation} and \begin{equation}\label{2.35} \aligned \left(\left(-D_x^{-1}+D_y^{-1}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}-\frac {g^{\prime} \left(\partial_{y}\right)} {g\left(\partial_{y}\right)}\right)\frac {f\left(\frac{-\partial}{\partial x}x\frac{\partial}{\partial_{y}}\right)} {f^{\prime}\left(\partial_{y}\right)}-n\right){}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)=0. \endaligned \end{equation} \end{theorem} \noindent {\bf{Proof.}} Using equations \eqref{2.14} and \eqref{2.16} in equation \eqref{mpcap}, we get assertion \eqref{2.34} and similarly using equations \eqref{2.30} and \eqref{2.31} in equation \eqref{mpcap}, we get assertion \eqref{2.35}. \begin{remark} Since the Sheffer sequence $s_{n}(x)$ for $g(t)=1$ becomes the associated Sheffer sequence ${\mathfrak s}_{n}(x)$ for $f(t)$. (For our convenience, we denote the associated Sheffer sequence by ${\mathfrak s}_{n}(x)$). Therefore, for $g(t)=1$, we deduce the following consequences of Theorems 2.1-2.5: \end{remark} \begin{corollary} The Legendre-Gould Hopper based associated Sheffer polynomials (LeGHASP) ${}_{{}_{s}LeG^{(r)}}\mathfrak{s}_{n}(x,y,z)$ and $\frac{{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ are defined by the generating function respectively as \begin{equation} C_0\left(-x\big(f^{-1}(t)\big)^{2}\right)\exp\left(yf^{-1}(t)+z\big(f^{-1}(t)\big)^r\right)=\sum_{n=0}^{\infty}{}_{{}_{s}LeG^{(r)}}\mathfrak{s}_{n}(x,y,z)\frac{t^n}{n!}, \end{equation} or \begin{equation} C_0\left(-x\big(H(t)\big)^2\right)\exp\left(yH(t)+z\big(H(t)\big)^r\right)=\sum_{n=0}^{\infty}{}_{{}_{s}LeG^{(r)}}\mathfrak{s}_{n}(x,y,z)\frac{t^n}{n!} \end{equation} and \begin{equation} C_0\left(xf^{-1}(t)\right)C_{0}(-yf^{-1}(t))\exp\left(z\big(f^{-1}(t)\big)^r\right)=\sum_{n=0}^{\infty}{}_{{}_{r}LeG^{(r)}}\mathfrak{r}_{n}(x,y,z)\frac{t^n}{(n!)^{2}}, \end{equation} or \begin{equation} C_0\left(xH(t)\right)C_{0}(-yH(t))\exp\left(z\big(H(t)\big)^r\right)=\sum_{n=0}^{\infty}{}_{{}_{r}LeG^{(r)}}\mathfrak{r}_{n}(x,y,z)\frac{t^n}{(n!)^{2}}. \end{equation} \end{corollary} \begin{corollary} The Legendre-Gould Hopper based associated Sheffer polynomials ${}_{{}_{s}LeG^{(r)}}\mathfrak{s}_{n}(x,y,z)$ and $\frac{{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ are quasi-monomial under the action of the following multiplicative and derivative operators: \begin{equation} {}_s\hat {M}_{LeGs}=\left(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\right)\frac {1} {f^{\prime}\left(\partial_{y}\right)}, \end{equation} \begin{equation} \hat {P}_{LH{\mathfrak s}}=f\left(\partial_{y}\right) \end{equation} and \begin{equation} {}_r\hat {M}_{LeGs}=\left(-D_x^{-1}+D_y^{-1}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\right)\frac {1} {f^{\prime}\left(\partial_{y}\right)}, \end{equation} \begin{equation} {}_r\hat {P}_{LHs}=f\left(\frac{-\partial}{\partial x}x\frac{\partial}{\partial_{y}}\right), \end{equation} respectively. \end{corollary} \begin{corollary} The Legendre-Gould Hopper based associated Sheffer polynomials ${}_{{}_{s}LeG^{(r)}}\mathfrak{s}_{n}(x,y,z)$ and $\frac{{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ satisfy the differential equations \begin{equation} \left(\left(y+2D_x^{-1}\frac{\partial}{\partial y}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\right)\frac {f \left(\partial_{y}\right)} {f^{\prime}\left(\partial_{y}\right)}-n\right){}_{{}_{s}LeG^{(r)}}\mathfrak{s}_{n}(x,y,z)=0 \end{equation} and \begin{equation} \left(\left(-D_x^{-1}+D_y^{-1}+rz\frac{\partial^{r-1}}{\partial y^{r-1}}\right)\frac {f\left(\frac{-\partial}{\partial x}x\frac{\partial}{\partial_{y}}\right)} {f^{\prime}\left(\partial_{y}\right)}-n\right){}_{{}_{r}LeG^{(r)}}\mathfrak{r}_{n}(x,y,z)=0 \end{equation} respectively. \end{corollary} \begin{remark} Since, for $f(t)=t$, the Sheffer polynomials $s_{n}(x)$ reduce to the Appell polynomials $A_{n}(x)$ \cite{2App}. Therefore, by taking $f(t)=t$ in Theorems 2.1-2.5, we can obtain the corresponding results for the Legendre-Gould Hopper based Appell polynomials (LeGHAP). \end{remark} In the next section, we consider certain new and known families of special polynomials related to the Sheffer sequences and obtain the results for these mixed type special polynomials. \\ \vspace{.50cm} \section{Examples} In Table 1, we have mentioned special cases of the LGHP ${}_{L}H_{n}^{(m,r)}(x,y,z)$. In order to obtain the results for the corresponding new or known special polynomials related to the Sheffer sequences, we consider the following examples: \noindent {\bf Example 1.} Since, for $x=0$, the LeGHP ${}_SH_n^{(r)}(x,y,z)$ reduce to the Gould Hopper polynomials (GHP) $H_{n}^{(r)}(x,y,z)$ (Table 1(I)). Therefore, for the same choice of $x$, the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ reduce to the Gould Hopper based Sheffer polynomials (GHSP) ${}_SH_{n}^{(r)}(x,y,z)$. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we get the corresponding results for the GHSP ${}_SH_{n}^{(r)}(x,y,z)$ (see \cite{2SubMum})\\ \noindent {\bf Example 2.} For $z=0$, the LeGHP ${}_SH_n^{(r)}(x,y,z)$ reduce to the 2-variable Legendre-type polynomials (2VLebP) ${}_{2L}L_{n}(x,y)$ (Table 1(II)). Therefore, for the same choice of $z$, the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ reduce to the 2-variable Legendre-type based Sheffer polynomials (2VLebSP) $_{{}_2Leb}s_{n}(x,y)$. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we get the following results for the 2VLebSP $_{{}_2Leb}s_{n}(x,y)$: \vspace{.25cm} \noindent \noindent \textbf{Table 1. Results for the 2VLebSP $_{{}_2Leb}s_{n}(x,y)$}\\ \\ {\tiny{ \begin{tabular}{lll} \hline \bf{S.} && \\ \bf{No.} &~~~~~~~~~{\bf Results}&~~~~~~~~~~~~~~~~~~~~~~~{\bf Mathematical Expressions}\\ \hline \bf{1.} & \bf{Generating } & $\frac {1}{g\big(f^{-1}(t)\big)}C_0\left(x\big(f^{-1}(t)\big)^{2}\right)\exp\left(yf^{-1}(t)\right)$ \\ \bf{}& \bf{function} & $~~~~~~~~~~~~~=A(t)C_0\left(-x\big(H(t)\big)^{2}\right)\exp\left(yH(t)\right)=\sum_{n=0}^{\infty}\,_{{}_2Leb}s_{n}(x,y)\frac{t^n}{n!}$\\ \hline \bf{2.} & \bf{Multiplicative and } & $\hat {M}=\left(y+2D_x^{-1}\frac{\partial}{\partial y}-\frac {g^{\prime} \left(\partial_{y}\right)} {g\left(\partial_{y}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{y}\right)}$\\ \bf{}& \bf{derivative operators} & $~~~~~~~~~~=\left(y+2D_x^{-1}\frac{\partial}{\partial y}\right)H'\left(H^{-1}\left(\partial_{y}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(\partial_{y}\right)\right)}{A\left(H^{-1}\left(\partial_{y}\right)\right)}$,~~~~$\hat {P}=f\left(\partial_{y}\right)=H^{-1}\left(\partial_{y}\right)$\\ \hline \bf{3.} & \bf{Differential } & $\left(\left(y+2D_x^{-1}\frac{\partial}{\partial y}-\frac {g^{\prime} \left(\partial_{y}\right)} {g\left(\partial_{y}\right)}\right)\frac {f \left(\partial_{y}\right)} {f^{\prime}\left(\partial_{y}\right)}-n\right)\,_{{}_2Leb}s_{n}(x,y)=0$ \\ \bf{}& \bf{equation} & \\ \hline \bf{4.} & \bf{Explicit} & $_{{}_2Leb}s_{n}(x,y)=\left(\left(y+2D_x^{-1}\frac{\partial}{\partial y}-\frac {g^{\prime} \left(\partial_{y}\right)} {g\left(\partial_{y}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{y}\right)}\right)^{n}\{1\}$\\ \bf{}& \bf{representation} & ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\ \hline \end{tabular}}}\\ \\ \noindent {\bf Example 3.} Since, for $r=m$, $x=0$, $y\rightarrow -D_{x}^{-1}$, $z\rightarrow y$ and $r=m$; $y=0$,$z\rightarrow y$, the LeGHP ${}_SH_n^{(r)}(x,y,z)$ and $\frac{{}_RH_n^{(r)}(x,y,z)}{n!}$ respectively reduce to the 2-variable generalized Laguerre type polynomials (2gLTP) ${}_{[m]}L_n(x,y)$ (Table 1(III)). Therefore, for the same choice of $r$, $x$, $y$ and $z$, the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ reduce to the 2-variable generalized Laguerre type based Sheffer polynomials (2gLTSP) ${_{{}_{[m]}L}s_{n}(x,y)}$. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we get the following results: \noindent \textbf{Table 2. Results for the 2gLTSP ${_{{}_{[m]}L}s_{n}(x,y)}$}\\ \\ {\tiny{ \begin{tabular}{lll} \hline \bf{S.} && \\ \bf{No.} &~~~~~~~~~{\bf Results}&~~~~~~~~~~~~~~~~~~~~~~~{\bf Mathematical Expressions}\\ \hline \bf{1.} & \bf{Generating } & $\frac {1}{g\big(f^{-1}(t)\big)}C_0\left(xf^{-1}(t)\right)\exp\left(y\big(f^{-1}(t)\big)^m\right)$ \\ \bf{}& \bf{function} & $~~~~~~~~~~~~~~~~~~~~~~~~~~~~=A(t)C_0\left(xH(t)\right)\exp\left(y\big(H(t)\big)^m\right)=\sum_{n=0}^{\infty}{_{{}_{[m]}L}s_{n}(x,y)}\frac{t^n}{n!}$\\ \hline \bf{2.} & \bf{Multiplicative and } & $\hat {M}=\left(-D_x^{-1}+(-1)^{m}my\frac{\partial^{m-1}}{\partial x^{m-1}}x^{m-1} \frac{\partial^{m-1}}{\partial x^{m-1}}-\frac {g^{\prime} \left(-\partial_{x}x\partial_{x}\right)} {g\left(-\partial_{x}x\partial_{x}\right)}\right)\frac {1} {f^{\prime}\left(-\partial_{x}x\partial_{x}\right)}$\\ \bf{}& & $~~~=\left(-D_x^{-1}+(-1)^{m}my\frac{\partial^{m-1}}{\partial x^{m-1}}x^{m-1} \frac{\partial^{m-1}}{\partial x^{m-1}}\right)H'\left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right)}{A\left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right)}$,\\ &\bf{derivative operators} &~~~$\hat {P}=f\left(-\partial_{x}x\partial_{x}\right)=H^{-1}\left(-\partial_{x}x\partial_{x}\right)$\\ \hline \bf{3.} & \bf{Differential } & $\left(\left(-D_x^{-1}+(-1)^{m}my\frac{\partial^{m-1}}{\partial x^{m-1}}x^{m-1} \frac{\partial^{m-1}}{\partial x^{m-1}}-\frac {g^{\prime} \left(-\partial_{x}x\partial_{x}\right)} {g\left(-\partial_{x}x\partial_{x}\right)}\right)\frac {f \left(-\partial_{x}x\partial_{x}\right)} {f^{\prime}\left(-\partial_{x}x\partial_{x}\right)}-n\right){_{{}_{[m]}L}s_{n}(x,y)}=0,$ or\\ \bf{}& \bf{equation} & $\left(\left(\left(y+mD_x^{-1}\frac{\partial^{m-1}}{\partial y^{m-1}}\right) H'\left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right) +\frac {A^{\prime} \left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right)}{A\left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right)}\right) H^{-1}\left(-\partial_{x}x\partial_{x}\right)-n\right)$\\ &&~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~$ {_{{}_{[m]}L}s_{n}(x,y)}=0$\\ \hline \bf{4.} & \bf{Explicit} & $_{{}_{[m]}L}s_{n}(x,y)=\left(\left(-D_x^{-1}+(-1)^{m}my\frac{\partial^{m-1}}{\partial x^{m-1}}x^{m-1} \frac{\partial^{m-1}}{\partial x^{m-1}}-\frac {g^{\prime} \left(-\partial_{x}x\partial_{x}\right)} {g\left(-\partial_{x}x\partial_{x}\right)}\right)\frac {1} {f^{\prime}\left(-\partial_{x}x\partial_{x}\right)}\right)^{n}\{1\}$\\ \bf{}& \bf{representation} & $=\left(\left(-D_x^{-1}+(-1)^{m}my\frac{\partial^{m-1}}{\partial x^{m-1}}x^{m-1} \frac{\partial^{m-1}}{\partial x^{m-1}}\right)H'\left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right)}{A\left(H^{-1}\left(-\partial_{x}x\partial_{x}\right)\right)}\right)^{n}\{1\}$\\ \hline \end{tabular}}}\\ \\ \noindent {\bf Example 4.} Since, for $r=m-1$; $x=0$, $y\rightarrow x$, $z\rightarrow y$ the LeGHP ${}_SH_n^{(r)}(x,y,z)$ reduce to the generalized Chebyshev polynomials (GCP) $U_{n}^{(m)}(x,y)$ (Table 1(IV)). Therefore, for the same choice of $r$, $x$, $y$ and $z$, the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ reduce to the generalized Chebyshev based Sheffer polynomials (GCSP) $_{{}_{gC}}s_{n}(x,y)$ \cite{2SubMum}. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we obtain the corresponding results for the GCSP $_{{}_{gC}}s_{n}(x,y)$ \cite{2SubMum}.\\\\ \noindent {\bf Example 5.} Since, for $r=1$;~$x=0$, $z\rightarrow-D_{x}^{-1}$ and for $r=1$; $y=0$, $z\rightarrow y$ the LeGHP ${}_SH_n^{(r)}(x,y,z)$ and $\frac{{}_RH_n^{(r)}(x,y,z)}{n!}$ respectively reduce to the 2-variable Laguerre polynomials (2VLP) $L_n(x,y)$ (Table 1(V)). Therefore, for the same choice of variables and parameters, the LeGHSP reduce to the 2-variable Laguerre based Sheffer polynomials (2VGLSP) $_{{}_L}s_{n}(x,y)$. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we obtain the corresponding results for the 2VLSP \cite{2SubNus}.\\ \noindent {\bf Example 6.} Since, for $z=0$, the LeGHP $\frac{{}_RH_n^{(r)}(x,y,z)}{n!}$ reduce to the 2-variable Legendre polynomials (2VLeP) $\frac{R_{n}(x,y)}{n!}$ (Table 1(VI)). Therefore, for the same choice of $z$, the LeGHSP ${}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)$ reduce to the 2-variable Legendre based Sheffer polynomials (2VLeSP) $\frac{{}_{R}s_{n}(x,y)}{n!}$ \cite{2SubNusR}. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we obtain the corresponding results for the 2VLeSP $\frac{{}_{R}s_{n}(x,y)}{n!}$ \cite{2SubNusR}.\\ \noindent {\bf Example 7.} Since, for $x=0,y\rightarrow x, z\rightarrow yD_{y}y$, the LeGHP ${}_SH_{n}^{(r)}(0,x,yD_{y}y)$ reduce to the 2-variable truncated polynomials of order $r$ (2VTP) $=e_n^{(r)}(x,y)$ (Table 1(VII)). Therefore, for the same choice of $x$, $y$ and $z$, the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ reduce to the 2-variable truncated exponential based Sheffer polynomials (2VTESP) $_{e^{(r)}}s_{n}(x,y)$ \cite{2SubGaz}. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we obtain the corresponding results for the 2VTESP $_{e^{(r)}}s_{n}(x,y)$ \cite{2SubGaz}.\\ \noindent {\bf Example 8.} Since, for $r=2$, $x=0$, the LeGHP ${}_SH_{n}^{(r)}(0,y,z)$ reduce to the 2-variable Hermite Kamp\'{e} de F${\acute{e}}$riet polynomials (2VHKdFP) $H_{n}(y,z)$ (Table 1(VIII)). Therefore, for the same choice of $r$ and $x$, the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ reduce to the 2-variable Hermite Kamp${\rm\acute e}$ de F${\acute{e}}$riet based Sheffer polynomials (2VHKdFSP) ${}_{H}s_{n}(y,z)$ \cite{2SubSaad}. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we obtain the corresponding results for the 2VHKdFSP ${}_{H}s_{n}(y,z)$ \cite{2SubSaad}.\\ \noindent {{\bf Example 9.}} Since, for $r=2$; $x=0$ and for $r=2; x=0$,$y\rightarrow x, z\rightarrow y$ the LeGHP ${}_SH_n^{(r)}(x,y,z)$ and $\frac{{}_RH_n^{(r)}(x,y,z)}{n!}$ respectively reduce to the Hermite type polynomials (HTP) $G_{n}(x,y)$ (Table 1(IX)). Therefore, for the same choices, the LeGHSP reduce to the Hermite type based Sheffer polynomials (HTSP) ${}_{G}s_{n}(x,y)$. Thus, by using these substitutions in Theorems 2.1, 2.3 and 2.5, we get the following results for the HTSP ${}_{G}s_{n}(x,y)$:\\ \noindent \textbf{Table 3. Results for the HTSP ${}_{G}s_{n}(x,y)$}\\ \\ {\tiny{ \begin{tabular}{lll} \hline \bf{S.} && \\ \bf{No.} &~~~~~~~~~{\bf Results}&~~~~~~~~~~~~~~~~~~~~~~~{\bf Mathematical Expressions}\\ \hline \bf{1.} & \bf{Generating } & $\frac {1}{g\big(f^{-1}(t)\big)}C_0\left(-xf^{-1}(t)\right)\exp\left(y\big(f^{-1}(t)\big)^2\right)$ \\ \bf{}& \bf{function} & $~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~=A(t)C_0\left(-xH(t)\right)\exp\left(y\big(H(t)\big)^2\right) =\sum_{n=0}^{\infty}{}_{G}s_{n}(x,y)\frac{t^n}{n!}$\\ \hline \bf{2.} & \bf{Multiplicative and } & $\hat {M}=\left(D_x^{-1}+2y\frac{\partial}{\partial x}x\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}x\partial_{x}\right)} {g\left(\partial_{x}x\partial_{x}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{x}x\partial_{x}\right)}$\\ \bf{}&& $~~~~=\left(D_x^{-1}+2y\frac{\partial}{\partial x}x\frac{\partial}{\partial x}\right)H'\left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right)}$,\\ &\bf{derivative operators}&~~~~~~~~$\hat {P}=f\left(\partial_{x}x\partial_{x}\right)=H^{-1}\left(\partial_{x}x\partial_{x}\right)$\\ \hline \bf{3.} & \bf{Differential } & $\left(\left(D_x^{-1}+2y\frac{\partial}{\partial x}x\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}x\partial_{x}\right)} {g\left(\partial_{x}x\partial_{x}\right)}\right)\frac {f \left(\partial_{x}x\partial_{x}\right)} {f^{\prime}\left(\partial_{x}x\partial_{x}\right)}-n\right){}_{G}s_{n}(x,y)=0$, or equivalently\\ \bf{}& \bf{equation} & $\left(\left(\left(D_x^{-1}+2y\frac{\partial}{\partial x}x\frac{\partial}{\partial x}\right) H'\left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right) +\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right)}\right) H^{-1}\left(\partial_{x}x\partial_{x}\right)-n\right)$\\ &&~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~$ {}_{G}s_{n}(x,y)=0$\\ \hline \bf{4.} & \bf{Explicit} & ${}_{G}s_{n}(x,y)=\left(\left(D_x^{-1}+2y\frac{\partial}{\partial x}x\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}x\partial_{x}\right)} {g\left(\partial_{x}x\partial_{x}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{x}x\partial_{x}\right)}\right)^{n}\{1\}$\\ \bf{}& \bf{representation} & $~~~~~~~~~~~~=\left(\left(D_x^{-1}+2y\frac{\partial}{\partial x}x\frac{\partial}{\partial x}\right)H'\left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}x\partial_{x}\right)\right)}\right)^{n}\{1\}$\\ \hline \end{tabular}}}\\ \\ \noindent {{\bf Example 10.}} Since, for $x\rightarrow\frac{(x^{2}-1)}{4}$, $y\rightarrow x, z=0$ and for $r=1; x\rightarrow\frac{(1-x)}{2}$, $y\rightarrow \frac{(1+x)}{2},z= 0$, the LeGHP ${}_SH_{n}^{(r)}(\frac{(x^{2}-1)}{4},x,0)$ and ${}_RH_{n}^{(1)}(\frac{(1-x)}{2},\frac{(x+1)}{2},0)$ respectively reduce to the Legendre polynomials (LeP) $P_{n}(x)$ (Table 1(X)). Therefore for the same choices, the LeGHSP respectively reduce to the Legendre based Sheffer polynomials (LeSP) ${}_{P}s_{n}(x)$. Thus, by Theorems 2.1, 2.3 and 2.5, we get the following results: \noindent \textbf{Table 4. Results for the LeSP ${}_{P}s_{n}(x)$} \\ {\tiny{ \begin{tabular}{lll} \hline \bf{S.} && \\ \bf{No.} &~~~~~~~~~{\bf Results}&~~~~~~~~~~~~~~~~~~~~~~~{\bf Mathematical Expressions}\\ \hline \bf{1.} & \bf{Generating } & $\frac {1}{g\big(f^{-1}(t)\big)}C_0\left(-\left(\frac {x^{2}-1}{4}\right)\big(f^{-1}(t)\big)^2\right)\exp\left(xf^{-1}(t)\right)$ \\ \bf{}& \bf{function} & $~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~=A(t)C_0\left(-\left(\frac {x^{2}-1}{4}\right)\big(H(t)\big)^2\right)\exp\left(xH(t)\right) =\sum_{n=0}^{\infty}{}_{P}s_{n}(x)\frac{t^n}{n!}$\\ \hline \bf{2.} & \bf{Multiplicative and } & $\hat {M}=\left(x+2D_{\left(\frac {x^{2}-1}{4}\right)}^{-1}\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}\right)} {g\left(\partial_{x}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{x}\right)}$\\ \bf{}&& $~~~~=\left(x+2D_{\left(\frac {x^{2}-1}{4}\right)}^{-1}\frac{\partial}{\partial x}\right) H'\left(H^{-1}\left(\partial_{x}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}\right)\right)}$,\\ &\bf{derivative operators}&~~~~~~~~$\hat {P}=f\left(\partial_{x}\right)=H^{-1}\left(\partial_{x}\right)$\\ \hline \bf{3.} & \bf{Differential } & $\left(\left(x+2D_{\left(\frac {x^{2}-1}{4}\right)}^{-1}\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}\right)} {g\left(\partial_{x}\right)}\right)\frac {f \left(\partial_{x}\right)} {f^{\prime}\left(\partial_{x}\right)}-n\right){}_{P}s_{n}(x)=0$, or equivalently\\ \bf{}& \bf{equation} & $\left(\left(\left(x+2D_{\left(\frac {x^{2}-1}{4}\right)}^{-1}\frac{\partial}{\partial x}\right) H'\left(H^{-1}\left(\partial_{x}\right)\right) +\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}\right)\right)}\right) H^{-1}\left(\partial_{x}\right)-n\right){}_{P}s_{n}(x)=0$~~~~~~~~~~~~~\\ \hline \bf{4.} & \bf{Explicit} & ${}_{P}s_{n}(x)=\left(\left(x+2D_{\left(\frac {x^{2}-1}{4}\right)}^{-1}\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}\right)} {g\left(\partial_{x}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{x}\right)}\right)^{n}\{1\}$\\ \bf{}& \bf{representation} & $~~~~~~~~~~~~=\left(\left(x+2D_{\left(\frac {x^{2}-1}{4}\right)}^{-1}\frac{\partial}{\partial x}\right)H'\left(H^{-1}\left(\partial_{x}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}\right)\right)}\right)^{n}\{1\}$\\ \hline \end{tabular}}}\\ \\ \noindent {{\bf Example 11.}} Since, for $r=3$, $x\rightarrow z\partial_{z}z$, $y\rightarrow x$, $z\rightarrow y$, the LeGHP ${}_{S}H_{n}^{(r)}(x,y,z)$ reduce to the Bell type polynomials (BTP) $H_{n}^{(3,2)}(x,y,z)$ (Table 1(XI)). Therefore, for the same substitutions, the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ reduce to the Bell type based Sheffer polynomials (BTSP) $_{H^{(3,2)}}s_{n}(x,y,z)$. Thus, by Theorems 2.1, 2.3 and 2.5, we get the following results:\\ \noindent \textbf{Table 5. Results for the BTSP $_{H^{(3,2)}}s_{n}(x,y,z)$}\\ \\ {\tiny{ \begin{tabular}{lll} \hline \bf{S.} && \\ \bf{No.} &~~~~~~~~~{\bf Results}&~~~~~~~~~~~~~~~~~~~~~~~{\bf Mathematical Expressions}\\ \hline \bf{1.} & \bf{Generating } & $\frac {1}{g\big(f^{-1}(t)\big)}\exp\left(xf^{-1}(t)+y\big(f^{-1}(t)\big)^3+z\big(f^{-1}(t)\big)^2\right)$ \\ \bf{}& \bf{function} & $~~~~~~~~~~~~~~~~~~~=A(t)\exp\left(xH(t)+y\big(H(t)\big)^3+z\big(H(t)\big)^2\right) =\sum_{n=0}^{\infty}{_{H^{(3,2)}}s_{n}(x,y,z)}\frac{t^n}{n!}$\\ \hline \bf{2.} & \bf{Multiplicative and } & $\hat {M}=\left(x+3y\frac{\partial^{2}}{\partial x^{2}}+2z\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}\right)} {g\left(\partial_{x}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{x}\right)}$\\ \bf{}&& $~~~~=\left(x+3y\frac{\partial^{2}}{\partial x^{2}}+2z\frac{\partial}{\partial x}\right) H'\left(H^{-1}\left(\partial_{x}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}\right)\right)}$,\\ &\bf{derivative operators}&~~~~~~~~$\hat {P}=f\left(\partial_{x}\right)=H^{-1}\left(\partial_{x}\right)$\\ \hline \bf{3.} & \bf{Differential } & $\left(\left(x+3y\frac{\partial^{2}}{\partial x^{2}}+2z\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}\right)} {g\left(\partial_{x}\right)}\right)\frac {f \left(\partial_{x}\right)} {f^{\prime}\left(\partial_{x}\right)}-n\right){_{H^{(3,2)}}s_{n}(x,y,z)}=0$, or equivalently\\ \bf{}& \bf{equation} & $\left(\left(\left(x+3y\frac{\partial^{2}}{\partial x^{2}}+2z\frac{\partial}{\partial x}\right) H'\left(H^{-1}\left(\partial_{x}\right)\right) +\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}\right)\right)}\right) H^{-1}\left(\partial_{x}\right)-n\right){_{H^{(3,2)}}s_{n}(x,y,z)}=0$\\ \hline \bf{4.} & \bf{Explicit} & $_{H^{(3,2)}}s_{n}(x,y,z)=\left(\left(x+3y\frac{\partial^{2}}{\partial x^{2}}+2z\frac{\partial}{\partial x}-\frac {g^{\prime} \left(\partial_{x}\right)} {g\left(\partial_{x}\right)}\right)\frac {1} {f^{\prime}\left(\partial_{x}\right)}\right)^{n}\{1\}$\\ \bf{}& \bf{representation} & $~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~=\left(\left(x+3y\frac{\partial^{2}}{\partial x^{2}}+2z\frac{\partial}{\partial x}\right) H'\left(H^{-1}\left(\partial_{x}\right)\right)+\frac {A^{\prime} \left(H^{-1}\left(\partial_{x}\right)\right)}{A\left(H^{-1}\left(\partial_{x}\right)\right)}\right)^{n}\{1\}$\\ \hline \end{tabular}}}\\ \\ \noindent {\bf{Remark~3.1.}}~For $\frac{1}{g(f^{-1}(t))}=A(t)=1$, the above mentioned special cases (Examples 1-11) of the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and $\frac{{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ yield the corresponding results for the LGHASP ${}_{{}_{s}LeG^{(r)}}\mathfrak{s}_{n}(x,y,z)$ and $\frac{{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$.\\ \noindent {\bf{Remark~3.2.}}~For $f^{-1}(t)=H(t)=t$, the above mentioned special cases of the LeGHSP yield the corresponding results for the Legendre-Gould Hopper based Appell polynomials LeGHAP ${}_{{}_{s}LeG^{(r)}}\mathcal{A}_{n}(x,y,z)$ and $\frac{{}_{{}_{r}LeG^{(r)}}\mathcal{A}_{n}(x,y,z)}{n!}$.\\ In the next section, we derive certain operational and integral representations for the LGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and $\frac{{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$.\\ \vspace{.50cm} \section{Operational and integral representations} To establish the operational representations for the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and $\frac{{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$, we prove the following results:\\ \noindent \begin{theorem} The following operational representation connecting the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and $\frac{{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ with the Sheffer polynomials $s_{n}(x)$ holds true: \begin{equation} {}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)=\exp\left(D_{x}^{-1}\frac{\partial^{2}}{\partial y^{2}}+z\frac{\partial^{r}}{\partial y^{r}}\right)s_{n}(y) \end{equation} and \begin{equation} {}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)=\exp\left(-D_{x}^{-1}\frac{\partial}{\partial y}+D_{y}^{-1}\frac{\partial}{\partial y}+z\frac{\partial^{r}}{\partial y^{r}}\right)s_{n}(0). \end{equation} \end{theorem} \noindent {\bf{Proof.}}~ In view of equation \eqref{2.8}, the proof is direct use of identity \eqref{crofton}.\\ \begin{theorem} The following operational representation connecting the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ with the 2VLebSP $_{{}_2L}s_{n}(x,y)$ holds true: \begin{equation}\label{4.3} {}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)=\exp\left(z\frac{\partial^{r}}{\partial y^{r}}\right){_{{}_2Leb}s_{n}(x,y)} \end{equation} \end{theorem} \begin{proof} From equation \eqref{2.1a}, we have \begin{equation}\label{4.4} \frac{\partial^{r}}{\partial y^{r}}{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)}=\frac{\partial}{\partial z}{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)}. \end{equation} Since, in view of Table 1(II), we have \begin{equation} _SH^{(r)}_{n}(x,y,0)={}_2{L}_{n}(x,y). \end{equation} Therefore, from Example 2 of Section 3, we have \begin{equation}\label{4.6} {}_{{}_{s}LeG^{(r)}}s_{n}(x,y,0)={_{{}_2Leb}s_{n}(x,y)} \end{equation} Now, solving equations \eqref{4.4} subject to initial condition \eqref{4.6}, we get assertion \eqref{4.3}. \end{proof} \noindent \begin{theorem} The following operational representation connecting the LGHSP $_{{}_LH^{(m,r)}}s_{n}(x,y,z)$ with the GHSP ${}_{H^{(r)}}s_{n}(y,z)$ holds true: \begin{equation} _{{}_LH^{(m,r)}}s_{n}(x,y,z)=\exp\left(D_{x}^{-1}\frac{\partial^{m}}{\partial y^{m}}\right){}_{H^{(r)}}s_{n}(y,z). \end{equation} \end{theorem} \noindent {\bf{Proof.}}~From equations (1.15) and (2.1) (or (2.2)), we have\\ $$\frac{\partial^{m}}{\partial y^{m}}~{_{{}_LH^{(m.r)}}s_{n}(x,y,z)}=\frac{\partial}{\partial D_{x}^{-1}}~{_{{}_LH^{(m,r)}}s_{n}(x,y,z)},\eqno(4.7)$$ where (\cite{2GHMDat}; p. 32 (8)):\\ $$\frac{\partial}{\partial D_{x}^{-1}}:=\frac{\partial}{\partial x}x\frac{\partial}{\partial x}.$$ Since, in view of Table 1(IV), we have $${}_LH_{n}^{(m,r)}(0,y,z)=H^{(r)}(y,z).\eqno(4.8)$$ Therefore, from Example 4 of Section 3, we have $$_{{}_LH^{(m,r)}}s_{n}(0,y,z)=_{H^{(r)}}s_{n}(y,z).\eqno(4.9)$$ Solving equation (4.7) subject to initial condition (4.9), we get assertion (4.6).\\ \newpage \noindent Next, we prove the integral representations for the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ and $\frac{{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ in the form of following theorems:\\ \begin{theorem}\label{t4.4} The following integral representation for the LeGHSP ${}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)$ holds true: \begin{equation}\label{4.8} {}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)=\int_{0}^{\infty} e^{-s} {{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,sD_{x}^{-1})}\,ds. \end{equation} \end{theorem} \begin{proof} In view of \eqref{leGHP1} and \eqref{2.1a}, we write \begin{equation}\label{4.11} \sum_{n=0}^{\infty}{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)}\frac{t^n}{n!}=\frac {1}{g\big(f^{-1}(t)\big)}\sum_{n=0}^{\infty}{}_{S}H_{n}^{(r)}(x,y,z)\frac{\big(f^{-1}(t)\big)^n}{n!}. \end{equation} Using the following integral representation of LeGHP ${}_{S}H_{n}^{(r)}(x,y,z)$ \cite{ghazala}: \begin{equation} {}_SH_{n}^{(r)}(x,y,z)~=~\int_{0}^{\infty} e^{-s} ~{}_SH_{n}^{(r)}(x,y,sD_{z}^{-1}) ~ds~, \end{equation} in the r.h.s. of equation \eqref{4.11}, we find \begin{equation} \sum_{n=0}^{\infty}{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)}\frac{t^n}{n!}=\frac {1}{g\big(f^{-1}(t)\big)}\int_{0}^{\infty} e^{-s} \left(\sum_{n=0}^{\infty}{}_{S}H_{n}^{(r)}(x,y,sD_{z}^{-1})\frac{\big(f^{-1}(t)\big)^n}{n!}\right) ~ds. \end{equation} Now, making use of the generating function \eqref{leGHP1} in the r.h.s. of the above equation, we have \begin{equation*} \sum_{n=0}^{\infty}{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)}\frac{t^n}{n!}=\int_{0}^{\infty} e^{-s} \left(\frac {1}{g\big(f^{-1}(t)\big)}\,C_0(-x(f^{-1}(t))^{2})\exp(y(f^{-1}(t))+sD_{z}^{-1}(f^{-1}(t))^{r}\right) ~ds, \end{equation*} which in view of\eqref{2.1a} yields \begin{equation} \sum_{n=0}^{\infty}{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,z)}\frac{t^n}{n!}=\sum_{n=0}^{\infty}\,\left(\int_{0}^{\infty} e^{-s}{{}_{{}_{s}LeG^{(r)}}s_{n}(x,y,sD_{z}^{-1})}\,ds \right)\frac{t^{n}}{n!} \end{equation} Finally, equating the coefficients of like powers of $t$ above, we get assertion \eqref{4.8}. \end{proof} \begin{theorem}\label{t4.5} The following integral representations for $\frac{{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)}{n!}$ hold true: \begin{equation} {}_{{}_{r}LeG^{(r)}}s_{n}(x,y,z)=\int_{0}^{\infty} e^{-s} {{}_{{}_{r}LeG^{(r)}}s_{n}(x,y,sD_{x}^{-1})}\,ds. \end{equation} \end{theorem} \begin{proof} The proof of Theorem \ref{t4.4} is same as of Theorem \ref{t4.5}. So we omit all the details. \end{proof} \vspace{.35cm} In Section 3, we have obtained the results for the new and known families of special polynomials related to Sheffer sequences by taking the special cases of the LGHP $_LH^{(m,r)}_{n}(x,y,z)$. In the Appendix section, we consider certain special polynomials belonging to the Sheffer and associated Sheffer families and obtain the results for the corresponding mixed special polynomials.\\ \noindent {\bf{5.~~Appendix}}\\ The Sheffer class contains important sequences such as the Hermite, Laguerre, Bernoulli, Poisson-Charlier polynomials {\it etc}. These polynomials are important from the view point of applications in physics and number theory. Also, the associated Sheffer family contains Mittag-Leffler, exponential, lower factorial polynomials {\it etc}.\\ We present the lists of some known members of the Sheffer and associated Sheffer families in Tables 12 and 13 respectively.\\ \noindent \textbf{Table 12. Some members of the Sheffer family}\\ \\ {\tiny{ \begin{tabular}{lllll} \hline &&&&\\ \bf{S. No.} & \bf{$g(t)$; $A(t)$} & \bf{$f(t)$; $H(t)$} & \bf{Generating Functions} & \bf{Polynomials}\\ \hline I. & $e^{(\frac{t}{\nu})^{k}}$; ~$e^{-t^{k}}$ &$\frac{t}{\nu}$; ~$\nu t$ & $\exp(\nu xt-t^k)$&Generalized Hermite \\ &&&$=\sum\limits_{n=0}^{\infty} H_{n,k,\nu}(x)\frac{t^n}{n!}$&polynomials \\ &&&&$H_{n,k,\nu}(x)$ \cite{2Lah}\\ \hline II. & $(1-t)^{-\alpha-1}$; ~$(1-t)^{-\alpha-1}$ & $\frac{t}{t-1}$; ~$\frac{t}{t-1}$ &$\frac{1}{(1-t)^{\alpha+1}}\exp (\frac{xt}{t-1})$ &Generalized Laguerre \\ &&&$=\sum\limits_{n=0}^{\infty} L_{n}^{(\alpha)}(x)t^{n}$&polynomials\\ &&&&$n!L_{n}^{\alpha}(x)$ \cite{2And, 2Rain}\\ \hline III.& $\frac{2}{e^t-1}$; ~$\frac{t}{1-t}$ & $\frac{e^t-1}{e^t+1}$; ~$\ln (\frac{1+t}{1-t})$ & $\frac{t}{1-t}(\frac{1+t}{1-t})^{x}$& Pidduck polynomials\\ &&&$=\sum\limits_{n=0}^{\infty} P_{n}(x)\frac{t^n}{n!}$ &$P_{n}(x)$ \cite{2Boas, 2Erd}\\ \hline IV.& $(1-t)^{-\beta}$; ~$e^{\beta t}$ & $\ln (1-t)$; ~$1-e^t$ & $\exp (\beta t+x(1-e^t))$& Acturial polynomials \\ &&&$=\sum\limits_{n=0}^{\infty} a_{n}^{(\beta)}(x)\frac{t^n}{n!}$&$a_{n}^{(\beta)}(x)$ \cite{2Boas}\\ \hline V.& $\exp (a(e^t-1))$; ~$e^{-t}$ & $a(e^t-1)$; ~$\ln (1+\frac{t}{a})$ & $e^{-t}(1+\frac{t}{a})^{x}$& Poisson-Charlier \\ &&&$=\sum\limits_{n=0}^{\infty} c_{n}(x;a)\frac{t^n}{n!}$&polynomials\\ &&&&$c_{n}(x;a)$ \cite{2Erde, 2Jor, 2Sze}\\ \hline VI. & $(1+e^{\lambda t})^{\mu}$; ~$(1+(1+t)^{\lambda})^{-\mu}$ & $e^t-1$; ~$\ln (1+t)$ & $(1+(1+t)^{\lambda})^{-\mu}(1+t)^{x}$& Peters polynomials\\ &&&$=\sum\limits_{n=0}^{\infty} S_{n}(x;\lambda,\mu)\frac{t^n}{n!}$&$S_{n}(x;\lambda,\mu)$ \cite{2Boas}\\ \hline VII.& $\frac{t}{e^t-1}$; ~$\frac{t}{\ln (1+t)}$ & $e^t-1$; ~$\ln (1+t)$ & $\frac{t}{\ln (1+t)}(1+t)^{x}$ & Bernoulli polynomials \\ &&&$=\sum\limits_{n=0}^{\infty} b_{n}(x)\frac{t^n}{n!}$&of the second kind \\ &&&&$b_{n}(x)$ \cite{2Jor}\\ \hline VIII.& $\frac{1}{2}(1+e^t)$; ~$\frac{2}{2+t}$ &$e^t-1$; ~$\ln (1+t)$ & $\frac{2}{2+t}(1+t)^{x}$& Related polynomials\\ &&&$=\sum\limits_{n=0}^{\infty} r_{n}(x)\frac{t^n}{n!}$&$r_{n}(x)$ \cite{2Jor}\\ \hline IX.& $\sec t$; ~$\frac{1}{\sqrt{1+t^{2}}}$ & $\tan t$; ~$\arctan(t)$ & $\frac{1}{\sqrt{1+t^{2}}}\exp (x \arctan(t))$& Hahn polynomials\\ &&&$=\sum\limits_{n=0}^{\infty} R_{n}(x)\frac{t^n}{n!}$&$R_{n}(x)$ \cite{2Bend}\\ \hline X.& $\frac{1+t}{(1-t)^{a}}$; ~$(1-4t)^{-\frac{1}{2}}\left(\frac{2}{1+\sqrt{1-4t}}\right)^{a-1}$ & $\frac{1}{4}-\frac{1}{4}\left(\frac{1+t}{1-t}\right)^{2}$; ~$\frac{-4t}{(1+\sqrt{1-4t})^{2}}$ & $(1-4t)^{-\frac{1}{2}}\left(\frac{2}{1+\sqrt{1-4t}}\right)^{a-1} $& Shively's psedo-Laguerre\\ &&&$\times\exp \left(\frac{-4xt}{(1+\sqrt{1-4t})^{2}}\right)$&polynomials\\ &&&$=\sum\limits_{n=0}^{\infty} R_{n}(a,x)t^n$ & $R_{n}(a,x)$ \cite{2Rain}\\ \hline \end{tabular}}}\\ \\ \noindent \textbf{Table 13. Some members of the associated Sheffer family}\\ \\ {\tiny{ \begin{tabular}{llll} \hline &&&\\ \bf{S. No.} & ~~~~~~~\bf{$f(t)$; $H(t)$} &~~~~~~~~ \bf{Generating Functions}& ~~~~~~~~\bf{Polynomials}~~~~~~~~\\ \hline I.& $\frac{e^t-1}{e^t+1}$; $\ln\left(\frac{1+t}{1-t}\right)$ & $\left(\frac{1+t}{1-t}\right)^x=\sum_{n=0}^{\infty}M_{n}(x)\frac{t^n}{n!}$ & Mittag-Leffler polynomials $M_{n}(x)$ \cite{2Bat} \\ \hline &&&\\ II.& $\ln(1+t)$; $e^t-1$ & $\exp(x(e^t-1))=\sum_{n=0}^{\infty}\phi_{n}(x)\frac{t^n}{n!}$ & Exponential polynomials $\phi_{n}(x)$ \cite{2Bell} \\ \hline &&&\\ III.& $e^t-1$ ; $\ln(1+t)$& $(1+t)^{x}=\sum_{n=0}^{\infty}(x)_{n}\frac{t^n}{n!}$ & Lower factorial polynomials $(x)_{n}$ \cite{2SRom} \\ \hline &&&\\ IV.& $-\frac{1}{2}t^{2}+t$ ; $1-\sqrt {1-2t}$& $\exp \left(x(1-\sqrt {1-2t})\right)=\sum_{n=0}^{\infty}p_{n}(x)\frac{t^n}{n!}$ & Bessel polynomials $p_{n}(x)$ \cite{2Car, 2Krall} \\ \hline \end{tabular}}}\\ \\ \noindent {\bf{Remark~5.1.}}~We remark that corresponding to each member belonging to the Sheffer (or associated Sheffer) family, there exists a new special polynomial belonging to the LeGHSP (or LeGHASP) family. The generating function and other properties of these special polynomials can be obtained from the results derived in Section 2.\\ Thus, by taking $g(t)$ (or $A(t)$) and $f(t)$ (or $H(t)$) of the special polynomials belonging to Sheffer family (Table 12 (I to X)) in equations \eqref{2.1a} (or \eqref{2.1b}), \eqref{2.9} (or \eqref{2.10}), \eqref{2.14} (or \eqref{2.15}), \eqref{2.16} (or \eqref{2.17}), \eqref{2.30} and \eqref{2.31}, we can get the generating function, multiplicative and derivative operators for the corresponding members belonging to the LeGHSP family.\\ \noindent

https://arxiv.org/abs/1201.3946

The Torelli group and congruence subgroups of the mapping class group

These are the lecture notes for my course at the 2011 Park City Mathematics Graduate Summer School. The first two lectures covered the basics of the Torelli group and the Johnson homomorphism, and the third and fourth lectures discussed the second cohomology group of the level p congruence subgroup of the mapping class group, following my papers "The second rational homology group of the moduli space of curves with level structures" and "The Picard group of the moduli space of curves with level structures".

\section*{Introduction} Let $\Sigma_{g,n}$ be a compact oriented genus $g$ surface with $n$ boundary components. The {\em mapping class group} of $\Sigma_{g,n}$, denoted $\Mod_{g,n}$, is the group of orientation-preserving diffeomorphisms of $\Sigma_{g,n}$ that restrict to the identity on $\partial \Sigma_{g,n}$, modulo isotopies that fix $\partial \Sigma_{g,n}$. The group $\Mod_{g,n}$ plays a fundamental role in many areas of mathematics, ranging from low-dimensional topology to algebraic geometry. At least for degrees less than about $2g/3$, the cohomology of $\Mod_{g,n}$ is well-understood due to the resolution of the Mumford conjecture by Madsen and Weiss \cite{MadsenWeiss} (together with Harer's unpublished improved version of his homological stability theorem; see \cite{Boldsen} for an exposition of this). However, the cohomology of finite-index subgroups of $\Mod_{g,n}$ remains a mystery. In these notes, we will focus on one low-degree calculation. Consider $n \in \{0,1\}$. For an integer $p$, the {\em level $p$ congruence subgroup} of $\Mod_{g,n}$, denoted $\Mod_{g,n}(p)$, is the subgroup of $\Mod_{g,n}$ consisting of mapping classes that act trivially on $\HH_1(\Sigma_{g,n};\Z/p)$. Another description of $\Mod_{g,n}(p)$ is as follows. The action of $\Mod_{g,n}$ on $\HH_1(\Sigma_{g,n};\Z)$ preserves the algebraic intersection pairing. Since $n \leq 1$, this is a nondegenerate alternating form, so we obtain a representation $\Mod_{g,n} \rightarrow \Sp_{2g}(\Z)$. Classically this representation was known to be surjective (see \S \ref{appendix:symplectic}). Let $\Sp_{2g}(\Z,p)$ be the subgroup of $\Sp_{2g}(\Z)$ consisting of matrices which equal the identity modulo $p$. Then $\Mod_{g,n}(p)$ is the pullback of $\Sp_{2g}(\Z,p)$ to $\Mod_{g,n}$. These notes will discuss the calculation of $\HH^2(\Mod_{g,n}(p);\Z)$. One motivation for this is the study of line bundles on the finite cover of the moduli space of curves associated to $\Mod_{g,n}(p)$, which is known as the moduli space of curves with level $p$ structures. The first Chern class of such a line bundle lies in $\HH^2(\Mod_{g,n}(p);\Z)$, and the determination of $\HH^2(\Mod_{g,n}(p);\Z)$ is the heart of the paper \cite{PutmanPicardGroupLevel}, which gives a complete classification of such line bundles. However, in these notes we will ignore this connection to algebraic geometry. Instead, we will use the computation of this cohomology group as an excuse to discuss a number of interesting topics related to the mapping class group. The universal coefficients exact sequence for $\HH^2(\Mod_{g,n}(p);\Z)$ takes the form \begin{align*} 0 \longrightarrow \text{Ext}(\HH_1(\Mod_{g,n}(p);\Z),\Z) &\longrightarrow \HH^2(\Mod_{g,n}(p);\Z) \\ &\ \ \ \ \ \ \ \ \longrightarrow \Hom(\HH_2(\Mod_{g,n}(p);\Z),\Z) \longrightarrow 0. \end{align*} The third and fourth lecture will be devoted to calculating the kernel and cokernel of this exact sequence. They will be proceeded by two lectures on necessary background. Let us now give a more detailed description of the four lectures. \begin{itemize} \item Lecture 1 will be devoted to the {\em Torelli group}. Denoted $\Torelli_{g,n}$, this is the subgroup of $\Mod_{g,n}$ consisting of mapping classes that act trivially on $\HH_1(\Sigma_{g,n};\Z)$. There are short exact sequences $$1 \longrightarrow \Torelli_{g,n} \longrightarrow \Mod_{g,n} \longrightarrow \Sp_{2g}(\Z) \longrightarrow 1$$ and $$1 \longrightarrow \Torelli_{g,n} \longrightarrow \Mod_{g,n}(p) \longrightarrow \Sp_{2g}(\Z,p) \longrightarrow 1,$$ and the structure of $\Mod_{g,n}(p)$ is a sort of mixture of the structure of $\Torelli_{g,n}$ and $\Sp_{2g}(\Z,p)$. \item Lecture 2 will be devoted to the {\em Johnson homomorphism}. Set $\HHH = \HH_1(\Sigma_{g,n};\Z)$. The Johnson homomorphism is a surjective homomorphism $\tau : \Torelli_{g,1} \longrightarrow \wedge^3 \HHH$. A deep theorem of Johnson shows that the Johnson homomorphism gives the ``rational part'' of the abelianization of $\Torelli_{g,1}$. More precisely, $\HH_1(\Torelli_{g,1};\Z) \cong W \oplus \wedge^3 \HHH$, where $W$ consists of torsion (in fact, $2$-torsion). We will also construct a ``mod $p$'' version of the Johnson homomorphism which takes the form $\tau_p : \Mod_{g,1}(p) \longrightarrow \HHH_p$, where $\HHH_p = \HH_1(\Sigma_{g,n};\Z/p)$. \item Lecture 3 is devoted to calculating $\HH_1(\Mod_{g,1}(p);\Z)$ for odd $p$. See the beginning of that lecture for why we restrict to odd $p$ and do not consider the closed case. There are two basic pieces. The first comes from the mod $p$ Johnson homomorphism and the second comes from the abelianization of $\Sp_{2g}(\Z,p)$. \item Lecture 4 is devoted to proving that $\HH_2(\Mod_{g}(p);\Q) \cong \Q$. Of course, this implies that $$\Hom(\HH_2(\Mod_{g,n}(p);\Z),\Z) \cong \Z.$$ The major work here is related to homological stability. \end{itemize} \section*{Lecture 1 : The Torelli group} \addtocounter{section}{1} \setcounter{theorem}{0} \setcounter{figure}{0} The Torelli group was first considered by Nielsen and Magnus in the early 20th century. However, its study only really took off in the late '70's and early '80's thanks to work of several people, most especially Birman and Johnson. Johnson's work has proven particularly fundamental and influential, and his survey \cite{JohnsonSurvey} cannot be recommended enough. Throughout this lecture, we will fix some $n \in \{0,1\}$. \ParagraphHeading{The symplectic representation.} Recall that $\Torelli_{g,n}$ is the kernel of the representation $\Mod_{g,n} \rightarrow \Sp_{2g}(\Z)$ arising from the action of $\Mod_{g,n}$ on $\HH_1(\Sigma_{g,n};\Z)$. We will need the following fact about this action. If $x$ is a simple closed curve on $\Sigma_{g,n}$, then let $T_x$ denote the right Dehn twist about $x$. Also, let $i_a(\cdot,\cdot)$ denote the algebraic intersection pairing on $\HH_1(\Sigma_{g,n};\Z)$. \begin{lemma} \label{lemma:transvection} Let $x$ be a simple closed curve on $x$. Orient $x$ in an arbitrary way, and let $[x] \in \HH_1(\Sigma_{g,1};\Z)$ denote the associated homology class. Then for $v \in \HH_1(\Sigma_{g,1};\Z)$, we have $$T_x(v) = v + i_a([x],v) \cdot [x].$$ \end{lemma} \begin{remark} The Dehn twist $T_x$ does not depend on an orientation on $x$. As a sanity check, you should verify that despite its appearance, the formula in Lemma \ref{lemma:transvection} does not depend on the orientation of $x$. \end{remark} \Figure{figure:generators}{Generators}{A separating twist $T_x$ and a bounding pair map $T_y T_{y'}^{-1}$} \begin{exercise} Prove Lemma \ref{lemma:transvection}. \end{exercise} \ParagraphHeading{Basic elements of Torelli.} Lemma \ref{lemma:transvection} allows us to construct some important elements of $\Torelli_{g,1}$. First, if $[x]=0$, then $T_x$ acts trivially on $\HH_1(\Sigma_{g,1};\Z)$. This will happen exactly when $x$ bounds an embedded subsurface of $x$ (see Figure \ref{figure:generators}). We will call such elements of Torelli {\em separating twists}. Next, the formula Lemma \ref{lemma:transvection} only depends on the homology class of the simple closed curve. Thus if $y$ and $y'$ are homologous, then $T_y$ and $T_{y'}$ act the same on $\HH_1(\Sigma_{g,1};\Z)$, so $T_y T_{y'}^{-1} \in \Torelli_{g,1}$. If $y$ and $y'$ are disjoint and homologous, then their union bounds an embedded subsurface (see Figure \ref{figure:generators}). If in addition to being disjoint neither $y$ nor $y'$ is separating, then we will call $T_y T_{y'}^{-1} \in \Torelli_{g,1}$ a {\em bounding pair map}. \ParagraphHeading{Generating sets for Torelli.} We have the following theorem. \begin{theorem} \label{theorem:torelligen} For all $g \geq 1$, the group $\Torelli_{g,1}$ is generated by bounding pair maps and separating twists. For $g \geq 3$, only bounding pair maps are needed. \end{theorem} \pagebreak \begin{remarks} \mbox{} \begin{enumerate} \item The fact that $\Torelli_{g,1}$ is generated by bounding pair maps and separating twists was originally proven by Powell \cite{Powell}, using earlier work of Birman \cite{BirmanSiegel}. This proof depended on some heroic calculations in the symplectic group whose details were omitted from the published papers. More recently, Putman \cite{PutmanCutPaste} gave a modern proof using the curve complex. Even more recently, Hatcher and Margalit \cite{HatcherMargalit} have given an even shorter proof. \item The fact that for $g \geq 3$ only bounding pair maps are needed is due to Johnson \cite{JohnsonBP}. He later proved a remarkable theorem which says that $\Torelli_{g,1}$ is finitely generated for $g \geq 3$ (see \cite{JohnsonFinite}). The size of Johnson's generating set grows exponentially in $g$. Answering a conjecture of Johnson, Putman \cite{PutmanSmallGensetTorelli} has recently constructed a generating set for $\Torelli_{g,1}$ that grows cubically in $g$. As we will discuss below, the abelianization of $\Torelli_{g,1}$ has rank cubic in the genus, so one cannot do better. \item McCullough and Miller \cite{McCulloughMiller} proved that $\Torelli_{2,n}$ is not finitely generated for $n \in \{0,1\}$. Later, in his thesis Mess \cite{MessThesis} proved that $\Torelli_{2}$ is an infinite rank free group. \item It is not known if $\Torelli_{g,n}$ is finitely presentable for $g \geq 3$. \end{enumerate} \end{remarks} \ParagraphHeading{The Birman exact sequence.} We will need to make several calculations in $\Torelli_{g,n}$. For us, the key tool for making such calculations is the fact that $\Torelli_{g,n}$ contains a large number of groups derived from surface groups. This follows from the {\em Birman exact sequence}, which takes the following form. Let $U\Sigma_{g}$ be the unit tangent bundle of $\Sigma_g$. For $g \geq 2$, we then have an exact sequence $$1 \longrightarrow \pi_1(U\Sigma_g) \longrightarrow \Mod_{g,1} \longrightarrow \Mod_g \longrightarrow 1.$$ The terms here have the following meanings. Let $\beta$ be the boundary component of $\Sigma_{g,1}$. \begin{itemize} \item The map $\Mod_{g,1} \rightarrow \Mod_g$ comes from gluing a disc to $\beta$ and extending mapping classes over this disc by the identity. \item The subgroup $\pi_1(U\Sigma_g)$ of $\Mod_{g,1}$ is known as the ``disc-pushing subgroup''. The mapping class associated to $\gamma \in \pi_1(U\Sigma_g)$ ``pushes'' the boundary component $\beta$ around the surface while allowing it to rotate. \end{itemize} Of course, the original version of the Birman exact sequence goes back to work of Birman \cite{BirmanBraids}. The version here first appeared in \cite{JohnsonFinite}; see \cite{FarbMargalitPrimer} for a textbook treatment. \Figure{figure:birmanexactseq}{BirmanExactSeq}{Pushing the boundary component around a simple closed curve induces a bounding pair map} \ParagraphHeading{The loop around the fiber.} The fiber $F_0$ of $\U\Sigma_g$ over the basepoint satisfies $F_0 \cong S^1$. The orientation on $\Sigma_g$ determines an orientation on $F_0$, so it makes sense to talk about ``clockwise'' and ``counterclockwise'' directions on $F_0$. The group $\pi_1(U\Sigma_g)$ contains a distinguished element $\delta_0$ which goes once around $F_0$ in the clockwise direction. The element of the disc-pushing subgroup of $\Mod_{g,1}$ corresponding to $\delta_0$ rotates the boundary component $\beta$ by a full turn in the clockwise direction. Clearly this is simply $T_{\beta}$. Observe that $T_{\beta} \in \Torelli_{g,1}$. \ParagraphHeading{Calculating in the disc-pushing subgroup.} Consider some $\gamma \in \pi_1(\Sigma_g)$ that can be realized by a smoothly embedded simple closed curve. The derivative of a smooth simple representative of $\gamma$ is a path in the tangent bundle of $\gamma$ which does not contain any zero vectors. For some fixed Riemannian metric on the surface, we can reparametrize $\gamma$ so that its derivative is a loop $\tilde{\gamma}$ in the unit tangent bundle. If $\gamma \neq 1$, then the element $\tilde{\gamma} \in \pi_1(U\Sigma_g)$ is independent of the choice of a smoothly embedded representative of $\gamma$. Indeed, any two such realizations are smoothly homotopic (this can be proved using the techniques in \cite{Epstein}; to test your understanding, you should verify that this fails if $\gamma = 1$). Let $\tau_{\gamma} \in \Mod_{g,1}$ be the element of the disc-pushing subgroup corresponding to $\tilde{\gamma} \in \pi_1(U\Sigma_g)$. As is shown in Figure \ref{figure:birmanexactseq}, the mapping class $\tau_{\gamma}$ is a bounding pair map, and hence lies in $\Torelli_{g,1}$. Since the loop $\delta_0$ around the fiber also corresponds to an element of $\Torelli_{g,1}$, we deduce that the disc-pushing subgroup lies in $\Torelli_{g,1}$. This implies that relations in $\pi_1(U\Sigma_{g})$ yield relations in $\Torelli_{g,1}$. Even more relations can be obtained by embedding $\Torelli_{g,1}$ into $\Torelli_{g',n}$ via a subsurface inclusion $\Sigma_{g,1} \hookrightarrow \Sigma_{g',n}$. \Figure{figure:lantern}{Lantern}{The lantern relation is $\tilde{\gamma}_1 \cdot \tilde{\gamma}_2 \cdot \tilde{\gamma}_3 = \delta_0^k$. To make the figure more attractive, the curves $\gamma_i$ have what appears to be a singularity at the basepoint, but in reality one should imagine them rounded and smooth there. In terms of Dehn twists, the lantern relation is $(T_{x_3} T_{y_3}^{-1}) (T_{x_2} T_{y_2}^{-1}) (T_{x_1} T_{y_1}^{-1}) = T_{\beta}$} As an example, consider the relation $\gamma_1 \cdot \gamma_2 \cdot \gamma_3 = 1$ in $\pi_1(\Sigma_{g,1})$ depicted in Figure \ref{figure:lantern}. We have $\tilde{\gamma}_1 \cdot \tilde{\gamma}_2 \cdot \tilde{\gamma}_3 = \delta_0^k$ for some $k \in \Z$. \begin{exercise} Prove that $k = 1$. \end{exercise} \noindent The associated relation $\tau_{\gamma_3} \tau_{\gamma_2} \tau_{\gamma_1} = \delta_0$ in $\Torelli_{g,1}$ is the {\em lantern relation} \begin{equation} \label{eqn:lantern} (T_{x_3} T_{y_3}^{-1}) (T_{x_2} T_{y_2}^{-1}) (T_{x_1} T_{y_1}^{-1}) = T_{\beta}; \end{equation} here the curves $x_i$ and $y_i$ are as depicted in Figure \ref{figure:lantern}. \begin{remark} The order of the terms in \eqref{eqn:lantern} is the opposite of what one might expect because elements in the fundamental group are composed left to right but mapping classes are composed right to left. \end{remark} \noindent Observe that if $g' \geq 3$, then this relation can be embedded in $\Torelli_{g',n}$ to express a separating twist as a product of bounding pair maps (c.f.\ Theorem \ref{theorem:torelligen}). \ParagraphHeading{Killing off separating twists.} The lantern relation gives numerous ways of expressing $T_{\beta}$ as a product of bounding pair maps. For $1 \leq i \leq 3$, let $x_i,y_i \in \pi_1(\Sigma_g)$ be the curves in Figure \ref{figure:killsep}. Observe that $x_1 x_2 x_3 = 1$ and $y_1 y_2 y_3=1$, so we have two different lantern relations $$T_{\beta} = \tau_{x_3} \tau_{x_2} \tau_{x_1} \quad \text{and} \quad T_{\beta} = \tau_{y_3} \tau_{y_2} \tau_{y_1}.$$ These curves have the property that $x_i$ is homologous to $y_i^{-1}$ for $1 \leq i \leq 3$. The group $\Torelli_{g,1}$ acts on $\pi_1(\Sigma_g)$, and it is not hard to see that there exists some $f_i \in \Torelli_{g,1}$ such that $f_i(x_i) = y_i^{-1}$. To use this, we will need the following exercise. \begin{exercise} \label{exercise:conjugation} If $\gamma \in \pi_1(\Sigma_g)$ can be realized by a simple closed curve and if $f \in \Mod_{g,1}$, then $\tau_{f(\gamma)} = f \tau_{\gamma} f^{-1}$. \end{exercise} \noindent Applying Exercise \ref{exercise:conjugation} several times, we obtain the following relation in $\Torelli_{g,1}$. \begin{align} T_{\beta}^2 &= (\tau_{x_3} \tau_{x_2} \tau_{x_1}) (\tau_{y_3} \tau_{y_2} \tau_{y_1})\notag\\ &= (\tau_{x_3} \tau_{x_2} \tau_{x_1}) (\tau_{f_3(x_3)}^{-1} \tau_{f_2(x_2)}^{-1} \tau_{f_1(x_1)}^{-1})\notag\\ &= (\tau_{x_3} \tau_{x_2} \tau_{x_1})(f_3 \tau_{x_3}^{-1} f_3^{-1} f_2 \tau_{x_2}^{-1} f_2^{-1}f_1 \tau_{x_1}^{-1} f_1^{-1})\label{eqn:killsep} \end{align} Upon abelianizing $\Torelli_{g,1}$, the right hand side of \eqref{eqn:killsep} vanishes. Letting $[T_{\beta}] \in \HH_1(\Torelli_{g,1};\Z)$ be the associated element of the abelianization, we obtain that $2 [T_{\beta}] = 0$. \Figure{figure:killsep}{KillSep}{$x_1 x_2 x_3 = 1$ and $y_1 y_2 y_3=1$} If $x$ is a separating curve on a surface of genus at least $3$, then we can embed the above relation into the surface to get that $T_x^2$ has to vanish upon abelianizing the Torelli group. We have proven the following. \begin{lemma} \label{lemma:septwiststorsion} Fix $g \geq 3$ and $n \in \{0,1\}$. Let $T_x$ be a separating twist in $\Torelli_{g,n}$. Then the image $[T_x]$ of $T_x$ in $\HH_1(\Torelli_{g,n};\Z)$ satisfies $2 [T_x]=0$. \end{lemma} \noindent Lemma \ref{lemma:septwiststorsion} first appeared in \cite{JohnsonAbel}. The above is a version of Johnson's proof. For an alternate exposition of that proof which arranges the details a little differently, see \cite[\S 7.2]{PutmanJohnsonHomo}. \ParagraphHeading{A preview.} In Lecture 2, we will construct the important Johnson homomorphism. Letting $\HHH = \HH_1(\Sigma_{g,n};\Z)$, this is a surjective homomorphism $$\tau : \Torelli_{g,1} \longrightarrow \wedge^3 \HHH.$$ There is also a version for closed surfaces, but we will not discuss it. The key property of the Johnson homomorphism is that its kernel is exactly the subgroup generated by separating twists. Lemma \ref{lemma:septwiststorsion} will then allow us to deduce the following theorem of Johnson \cite{JohnsonAbel}. \begin{theorem} \label{theorem:torelliabel} For $g \geq 3$, we have $\HH_1(\Torelli_{g,1};\Z) \cong W \oplus \wedge^3 \HHH$, where $W$ consists of $2$-torsion. \end{theorem} \begin{remark} Johnson also calculated the $2$-torsion $W$. The associated $\Z/2$-quotients of $\Torelli_{g,1}$ come from the Rochlin invariants of homology $2$-spheres. They were originally constructed by Birman and Craggs \cite{BirmanCraggs}. Later, in \cite{JohnsonBirmanCraggs} Johnson packaged all of Birman and Craggs's homomorphisms together into a single homomorphism and determined exactly how many linearly independent quotients they had constructed. \end{remark} \noindent As a prologue for the construction, we recommend performing the following exercise, which explains the appearance of $\wedge^3 \HHH$ in the Johnson homomorphism. \begin{exercise} \label{exercise:torus} Let $\Torus^n$ denote the $n$-torus $(S^1)^n$. \begin{enumerate} \item Prove that the cohomology ring $\HH^{\ast}(\Torus^n;\Z)$ is isomorphic to the exterior algebra $\wedge^{\ast} \Z^n$. \item Let $G$ be an abelian topological group. Define a product $$\HH_i(G;\Z) \otimes \HH_j(G;\Z) \longrightarrow \HH_{i+j}(G;\Z)$$ via the composition $$\HH_i(G;\Z) \otimes \HH_j(G;\Z) \stackrel{\phi}{\longrightarrow} \HH_{i+j}(G \times G;\Z) \stackrel{\psi}{\longrightarrow} \HH_{i+j}(G;\Z),$$ where $\phi$ is the map coming from the K\"{u}nneth exact sequence and $\psi$ is induced by the group product $G \times G \rightarrow G$. Prove that with this product structure, $\HH_{\ast}(G;\Z)$ is a graded-commutative algebra. We remark that this product is known as the {\em Pontryagin product}. \item The space $\Torus^n$ is an abelian topological group. Prove that the resulting graded-commutative ring $\HH_{\ast}(\Torus^n;\Z)$ is isomorphic to the exterior algebra $\wedge^{\ast} \Z^n$. \end{enumerate} \end{exercise} \subsection{Appendix to lecture 1 : the surjectivity of the symplectic representation} \label{appendix:symplectic} Recall that the action of $\Mod_{g,1}$ on $\HH_1(\Sigma_{g,1};\Z)$ preserves the algebraic intersection pairing and thus gives a representation $\pi : \Mod_{g,1} \rightarrow \Sp_{2g}(\Z)$. In this appendix, we will give a sequence of exercises about the surjectivity of $\pi$. A {\em symplectic basis} for $\HH_1(\Sigma_{g,1};\Z)$ is a basis $\{a_1,b_1,\ldots,a_g,b_g\}$ for $\HH_1(\Sigma_{g,1};\Z)$ such that $$i_a(a_i,b_j) = \delta_{ij} \quad \text{and} \quad i_a(a_i,a_j) = i_a(b_i,b_j) = 0$$ for all $1 \leq i,j \leq g$. Let $\mathcal{S}$ be the set of symplectic bases for $\HH_1(\Sigma_{g,1};\Z)$. The following exercise should be straightforward. \begin{exercise} \label{exercise:sptran} $\Sp_{2g}(\Z)$ acts simply transitively on $\mathcal{S}$. \end{exercise} \Figure{figure:symplecticbasis}{SymplecticBasis}{A geometric symplectic basis} If $x$ and $y$ are simple closed curves on $\Sigma_{g,1}$, then let $i_g(x,y)$ be their {\em geometric intersection number}; i.e.\ the minimal cardinality of $x' \cap y'$ as $x'$ and $y'$ range over all simple closed curves homotopic to $x$ and $y$, respectively. A {\em geometric symplectic basis} (see Figure \ref{figure:symplecticbasis}) is a collection $\{\alpha_1,\beta_1,\ldots,\alpha_g,\beta_g\}$ of simple closed curves on $\Sigma_{g,1}$ such that $$i_g(\alpha_i,\beta_j) = \delta_{ij} \quad \text{and} \quad i_g(\alpha_i,\alpha_j) = i_g(\beta_i,\beta_j) = 0$$ for all $1 \leq i,j \leq g$. Let $\mathcal{G}$ be the set of geometric symplectic bases on $\Sigma_{g,1}$. We then have the following. \begin{exercise} \label{exercise:modtran} $\Mod_{g,1}$ acts transitively on $\mathcal{G}$. Hint : given two geometric symplectic bases, prove using the Euler characteristic that you get homeomorphic surfaces when you cut along them. \end{exercise} The following lemma is the heart of the fact that $\pi(\Mod_{g,1}) = \Sp_{2g}(\Z)$. \begin{lemma} \label{lemma:realizebasis} If $\{a_1,b_1,\ldots,a_g,b_g\}$ is a symplectic basis for $\HH_1(\Sigma_g;\Z)$, then there exists a geometric symplectic basis $\{\alpha_1,\beta_1,\ldots,\alpha_g,\beta_g\}$ on $\Sigma_g$ such that $[\alpha_i] = a_i$ and $[\beta_i] = b_i$ for $1 \leq i \leq g$. \end{lemma} \noindent Proofs of Lemma \ref{lemma:realizebasis} can be found in \cite[Lemma A.3]{PutmanCutPaste} and \cite[3rd proof of Lemma 6.4]{FarbMargalitPrimer}; however, it is worthwhile to contemplate how one might prove it (though it is probably too hard for an exercise). \begin{exercise} Combine Lemma \ref{lemma:realizebasis} with Exercises \ref{exercise:sptran} and \ref{exercise:modtran} to deduce that $\pi(\Mod_{g,1}) = \Sp_{2g}(\Z)$. \end{exercise} \section*{Lecture 2 : The Johnson homomorphism} \addtocounter{section}{1} \setcounter{theorem}{0} \setcounter{figure}{0} Let $\HHH = \HH_1(\Sigma_g;\Z)$. In this lecture, we will construct the {\em Johnson homomorphism}, which is a surjective homomorphism $$\tau : \Torelli_{g,1} \longrightarrow \wedge^3 \HHH.$$ This homomorphism can be constructed in a number of completely different ways. It was originally constructed in \cite{JohnsonHomo} by examining the action of $\Torelli_{g,1}$ on the second nilpotent truncation of $\pi_1(\Sigma_{g,1})$. We explain this original construction in an appendix. In his survey \cite{JohnsonSurvey}, Johnson outlined several alternate constructions. We will use a definition in terms of mapping tori which was introduced in \cite{JohnsonSurvey} and was first shown to be equivalent to the original definition by Hain \cite{HainTorelli}. Our exposition will follow the paper \cite{ChurchFarbTorelli} of Church and Farb, which gives a more direct proof of this equivalence. \ParagraphHeading{The construction.} Consider $f \in \Torelli_{g,1}$. Though it is an abuse of notation, we will regard $f$ as a homeomorphism of $\Sigma_{g,1}$. Glue a disc to the boundary component of $\Sigma_{g,1}$ and extend $f$ over this disc by the identity to obtain a homeomorphism $F$ of $\Sigma_g$. Let $p_0 \in \Sigma_g$ be the center of the glued-in disc, so $F(p_0)=p_0$. Now let $M_{F}$ be the {\em mapping torus} of $F$, i.e.\ the quotient $\Sigma_g \times I / \sim$, where $(x,1) \sim (F(x),0)$. Give $M_F$ the basepoint $q_0 = (p_0,0)$. There is a distinguished element $t \in \pi_1(M_F,q_0)$ which traverses the embedded loop $p_0 \times I / \sim$ in $M_F$ in the positive direction. Fix a standard generating set $S=\{s_1,\ldots,s_{2g}\}$ for $\pi_1(\Sigma_g)$ that satisfies the surface relation $$[s_1,s_2] \cdots [s_{2g-1},s_{2g}] = 1.$$ Since $F(p_0)=p_0$, the map $F$ acts on $\pi_1(\Sigma_g,p_0)$. For $1 \leq i \leq 2g$, let $w_i$ be an expression for $F_{\ast}(s_i)$ in terms of the generating set $S$. We then have a presentation $$\pi_1(M_F,q_0) = \GroupPres{s_1,\ldots,s_{2g},t}{[s_1,s_2] \cdots [s_{2g-1},s_{2g}]=1, t s_i t^{-1} = w_i \text{ for } 1 \leq i \leq 2g}.$$ For $\gamma \in \pi_1(\Sigma_g,p_0)$, let $[\gamma] \in \HHH$ be the associated element of the abelianization. Since $F \in \Torelli_{g}$, we have $[s_i] = [w_i]$ for $1 \leq i \leq 2g$. This implies that we can define a homomorphism $\phi_{\ast} : \pi_1(M_F,q_0) \rightarrow \HHH$ such that $\phi_{\ast}(s_i) = [s_i]$ for $1 \leq i \leq 2g$ and such that $\phi_{\ast}(t)=0$. The space $M_{F}$ is clearly a $K(\pi_1(M_{F}),1)$. Let $\Torus^{2g}$ be the $2g$-torus. Fix an identification of $\pi_1(\Torus^{2g})$ with $\HHH$. Since $\Torus^{2g}$ is a $K(\HHH,1)$, the standard properties of Eilenberg-MacLane spaces show that there is a canonical homotopy class of continuous maps $\phi : M_{F} \rightarrow \Torus^{2g}$ inducing the homomorphism $\phi_{\ast}$. The space $M_F$ is a closed 3-manifold, so it has a canonical class $[M_F] \in \HH_3(M_F;\Z)$. Define $$\tau(f) = \phi_{\ast}([M_{F}]) \in \HH_3(\Torus^{2g};\Z) \cong \wedge^3 \HHH.$$ The last isomorphism here comes from Exercise \ref{exercise:torus} Summing up, we have constructed a map $\tau : \Torelli_{g,1} \rightarrow \wedge^3 \HHH$. The following exercise is a good test of your understanding of the above construction. \begin{exercise} \label{exercise:tau} Prove that $\tau$ is independent of all the above choices except for the identification of $\pi_1(\Torus^{2g})$ with $\HHH$ (which is fixed). Next, prove that $\tau$ is a homomorphism. \end{exercise} \ParagraphHeading{Effect on generators.} The following lemma calculates $\tau$ on the generators for $\Torelli_{g,1}$. Recall that if $S$ is a genus $h$ surface with at most $1$ boundary component, then a {\em symplectic basis} for $\HH_1(S;\Z) \cong \Z^{2h}$ is a basis $\{a_1,b_1,\ldots,a_h,b_h\}$ for $\HH_1(S;\Z)$ such that $$i_a(a_i,b_j) = \delta_{ij} \quad \text{and} \quad i_a(a_i,a_j) = i_a(b_i,b_j) = 0$$ for all $1 \leq i,j \leq h$. Here $i_a(\cdot,\cdot)$ is the algebraic intersection form. \begin{lemma} \label{lemma:johnsoncalc} \mbox{} \begin{enumerate} \item Let $T_x \in \Torelli_{g,1}$ be a separating twist. Then $\tau(T_x) = 0$. \item Let $T_{x} T_{x'}^{-1}$ be a bounding pair map on $\Torelli_{g,1}$. Let $S$ be the component of $\Sigma_{g,1}$ cut along $x \cup x'$ that does not contain $\partial \Sigma_{g,1}$, so $S \cong \Sigma_{h,2}$ for some $h < g$. Let $S' \subset S$ be an embedded subsurface such that $S' \cong \Sigma_{h,1}$ and let $\{a_1,b_1,\ldots,a_h,b_h\}$ be a symplectic basis for $\HH_1(S';\Z) \subset \HH_1(\Sigma_{g,1};\Z)$. Then $$\tau(T_x T_{x'}^{-1}) = \pm [x] \wedge (a_1 \wedge b_1 + \cdots + a_h \wedge b_g).$$ \end{enumerate} \end{lemma} \noindent We will discuss the proof of this lemma at the end of this section. Right now, we suggest doing the following two exercises. \begin{exercise} Prove that the formula in Lemma \ref{lemma:johnsoncalc} is independent of the choice of $S'$ and its symplectic basis. \end{exercise} \begin{exercise} Using Lemma \ref{lemma:johnsoncalc}, prove that $\tau$ is surjective. \end{exercise} \ParagraphHeading{Johnson's theorem.} In \cite{JohnsonKg}, Johnson proved the following deep theorem, which is a sort of converse to part 1 of Lemma \ref{lemma:johnsoncalc}. For an alternate proof, see \cite{PutmanJohnsonHomo}. \begin{theorem} \label{theorem:kg} The kernel of $\tau$ is generated by separating twists. \end{theorem} \noindent As we indicated at the end of Lecture 1, this theorem together with Lemma \ref{lemma:septwiststorsion} implies that $\HH_1(\Torelli_{g,1};\Z) \cong W \oplus \wedge^3 \HHH$ for $g \geq 3$, where $W$ consists of $2$-torsion. \ParagraphHeading{The Johnson homomorphism mod $p$.} Set $\HHH_p = \HH_1(\Sigma_{g,1};\Z/p)$. We wish to construct a ``mod $p$'' Johnson homomorphism $$\tau_p : \Mod_{g,n}(p) \rightarrow \HHH_p.$$ The construction goes exactly like the construction of the ordinary Johnson homomorphism. Consider $f \in \Mod_{g,1}(p)$. Regard $f$ as a homeomorphism of $\Sigma_{g,1}$. Glue a disc to the boundary component of $\Sigma_{g,1}$ and extend $f$ over this disc by the identity to obtain a homeomorphism $F$ of $\Sigma_g$. Let $p_0 \in \Sigma_g$ be the center of the glued-in disc, so $F(p_0)=p_0$. Let $M_F$ be the mapping torus of $F$ and let $q_0 = (p_0,0)$ be the basepoint for $M_F$. There is a distinguished element $t \in \pi_1(M_F,q_0)$ which traverses the embedded loop $p_0 \times I / \sim$ in $M_F$ in the positive direction. Fix a standard generating set $S=\{s_1,\ldots,s_{2g}\}$ for $\pi_1(\Sigma_g)$ that satisfies the surface relation $$[s_1,s_2] \cdots [s_{2g-1},s_{2g}] = 1.$$ Since $F(p_0)=p_0$, the map $F$ acts on $\pi_1(\Sigma_g,p_0)$. For $1 \leq i \leq 2g$, let $w_i$ be an expression for $F_{\ast}(s_i)$ in terms of the generating set $S$. We then have a presentation $$\pi_1(M_F,q_0) = \GroupPres{s_1,\ldots,s_{2g},t}{[s_1,s_2] \cdots [s_{2g-1},s_{2g}]=1, t s_i t^{-1} = w_i \text{ for } 1 \leq i \leq 2g}.$$ For $\gamma \in \pi_1(\Sigma_g,p_0)$, let $[\gamma]_p \in \HHH_p$ be the associated element. Since $F \in \Mod_{g}(p)$, we have $[s_i]_p = [w_i]_p$ for $1 \leq i \leq 2g$. This implies that we can define a homomorphism $\phi_{\ast} : \pi_1(M_F,q_0) \rightarrow \HHH_p$ such that $\phi_{\ast}(s_i) = [s_i]_p$ for $1 \leq i \leq 2g$ and such that $\phi_{\ast}(t)=0$. The space $M_{F}$ is clearly a $K(\pi_1(M_{F}),1)$. Let $Z$ be a $K(\HHH_p,1)$. The standard properties of Eilenberg-MacLane spaces show that there is a canonical homotopy class of continuous maps $\phi : M_{F} \rightarrow Z$ inducing the homomorphism $\phi_{\ast}$. Define $$\tau_p'(f) = \phi_{\ast}([M_{F}]) \in \HH_3(Z;\Z) \cong \HH_3(\HHH_p;\Z),$$ where $M_F \in \HH_3(M_F;\Z)$ is the canonical class. It is not true that $\HH_3(\HHH_p;\Z) \cong \wedge^3 \HHH_p$; however, $\HH_3(\HHH_p;\Z)$ does contain $\wedge^3 \HHH_p$ as a direct factor (see, for example, \cite[Theorem V.6.4]{BrownCohomology}). Let $\tau_p : \Mod_{g,1}(p) \rightarrow \wedge^3 \HHH_p$ be the composition of $\tau_p'$ with some (fixed for all time) projection of $\HH_3(\HHH_p;\Z)$ onto $\wedge^3 \HHH_p$. By the same argument used in Exercise \ref{exercise:tau}, the map $\tau_p$ is a well-defined homomorphism. From its construction, it is clear that the following diagram commutes. $$\begin{CD} \Torelli_{g,1} @>{\tau}>> \wedge^3 \HHH \\ @VVV @VVV \\ \Mod_{g,n}(p) @>{\tau_p}>> \wedge^3 \HHH_p \end{CD}$$ Since $\tau$ is surjective for $g \geq 3$, we obtain the following. \begin{theorem} \label{theorem:modpjohnson} The map $\tau_p : \Mod_{g,1}(p) \rightarrow \HHH_p$ is surjective for $g \geq 3$. \end{theorem} \ParagraphHeading{Images of mapping tori.} To prove Lemma \ref{lemma:johnsoncalc}, one would need to be able to calculate the images of fundamental classes of mapping tori under maps to a torus. Instead of giving the details of the proof, we will prove an easier theorem which illustrates how one can do this. For the proof of Lemma \ref{lemma:johnsoncalc}, see \cite{ChurchFarbTorelli}. The result we will prove is as follows. In its statement and proof, we will use the natural identification of the graded-commutative algebra structure on $\HH_{\ast}(\Torus^n;\Z)$ with $\wedge^{\ast} \Z^n$ which was identified in Exercise \ref{exercise:torus}. \begin{theorem} Let $\phi : \Sigma_g \rightarrow \Torus^n$ be a continuous map. Choose a standard basis $\{\alpha_1,\beta_1,\ldots,\alpha_g,\beta_g\}$ for $\pi_1(\Sigma_g)$, so $$\pi_1(\Sigma_g) = \GroupPres{\alpha_1,\beta_1,\ldots,\alpha_g,\beta_g}{[\alpha_1,\beta_1]\cdots[\alpha_g,\beta_g]=1}.$$ Then $$\phi_{\ast}([\Sigma_g]) = \pm \sum_{i=1}^g \phi_{\ast}(\alpha_i) \wedge \phi_{\ast}(\beta_i) \in \wedge^2 \Z^n \cong \HH_2(\Torus^n;\Z).$$ \end{theorem} \begin{proof} Assume first that $g=1$, so $\Sigma_g \cong \Torus^2$. We will identify $\Sigma_g$ with $\Torus^2$. Observe that $\pi_1(\Torus^2) \cong \Z^2$ and $\HH_{\ast}(\Torus^2;\Z) \cong \wedge^{\ast}(\Z^2)$. Moreover, $\HH_2(\Torus^2;\Z) \cong \wedge^2 \Z^2$ is generated by $\alpha_1 \wedge \beta_1$, so $[\Torus^2] = \pm \alpha_1 \wedge \beta_1$. By the naturality of the Pontryagin product we have $$\phi_{\ast}([\Torus^2]) = \pm \phi_{\ast}(\alpha_1) \wedge \phi_{\ast}(\beta_1).$$ \Figure{figure:calcjohnson}{CalcJohnson}{Degenerating our surface to a wedge of tori} Now assume that $g > 1$. We will reduce to the case $g = 1$ by a method that will be familiar to algebraic geometers. Namely, we will ``degenerate'' our surface to a nodal surface and make the computation there. Fix a basepoint $p_0$ for $\Sigma_g$. As shown in Figure \ref{figure:calcjohnson}, let $(X,p_0')$ be the result of collapsing $(\Sigma_g,p_0)$ to the wedge of $g$ tori $X_1,\ldots,X_g$. Choose $X$ such that under the collapse map $\rho : \Sigma_g \rightarrow X$, the curves $\alpha_i$ and $\beta_i$ map to generators for $\pi_1(X_i,p_0') \subset \pi_1(X,p_0')$. \begin{exercise} Prove that the map $\phi_{\ast} : \pi_1(\Sigma_g,p_0) \rightarrow \pi_1(\Torus^n)$ factors as $$\pi_1(\Sigma_g,p_0) \stackrel{\rho_{\ast}}{\longrightarrow} \pi_1(X,p_0') \stackrel{\phi'_{\ast}}{\longrightarrow} \pi_1(\Torus^n)$$ for some homomorphism $\phi_{\ast}' : \pi_1(X,p_0') \rightarrow \pi_1(\Torus^n)$. Hint : $\pi_1(\Torus^n)$ is abelian. \end{exercise} \noindent Since the spaces $\Sigma_g$ and $X$ and $\Torus^n$ are all Eilenberg-MacLane spaces, there exists some continuous map $\phi' : X \rightarrow \Torus^n$ inducing $\phi'_{\ast}$ such that $\phi$ is homotopic to $\phi' \circ \rho$. Observe that $\rho_{\ast}([\Sigma_g]) = [X_1] + \cdots + [X_g]$. Moreover, $\rho_{\ast}(\alpha_i)$ and $\rho_{\ast}(\beta_i)$ are a standard basis for $\pi_1(X_i) \subset \pi_1(X)$. By the case $g=1$, we have $$(\phi'|_{X_i})_{\ast}([X_i]) = \pm \phi_{\ast}'(\rho_{\ast}(\alpha_i)) \wedge \phi_{\ast}'(\rho_{\ast}(\beta_i) = \pm \phi_{\ast}(\alpha_i) \wedge \phi_{\ast}(\beta_i).$$ All the $\pm$ signs are identical. Adding everything up, we get the desired result. \end{proof} \section*{Appendix to lecture 2 : the original construction of the Johnson homomorphism} The original construction of the Johnson homomorphism gave a homomorphism $$\tau : \Torelli_{g,1} \rightarrow \Hom(\HHH,\wedge^2 \HHH).$$ Johnson then calculated $\tau$ on generators and showed that its image was isomorphic to $\wedge^3 \HHH$. In this appendix, we will guide you through Johnson's original construction. If $G$ is a group, then let $\gamma_k(G)$ be the $k^{\text{th}}$ term in its {\em lower central series}, which is the inductively defined sequence of groups $$\gamma_0(G) = G \quad \text{and} \quad \gamma_{k+1}(G) = [\gamma_k(G),G].$$ Set $\pi = \pi_1(\Sigma_{g,1},\ast)$, where $\ast \in \partial \Sigma_{g,1}$. We will need the following fact. We will state it as an exercise, but you should only attempt it if you know some of the basics of group cohomology (otherwise, treat it as a black box). \begin{exercise} Prove that $\gamma_1(\pi) / \gamma_2(\pi) \cong \wedge^2 \HHH$. Hint : Apply the 5-term exact sequence in group homology to the short exact sequence $$1 \longrightarrow \gamma_1(\pi) \longrightarrow \pi \longrightarrow \HHH \longrightarrow 1.$$ You will need the fact that $\HH_2(\HHH;\Z) \cong \wedge^2 \HHH$. \end{exercise} Now consider the short exact sequence $$1 \longrightarrow \gamma_1(\pi) \longrightarrow \pi \longrightarrow \HHH \longrightarrow 1.$$ To simplify this, we mod out by $\gamma_2(\pi) < \gamma_1(\pi)$. By the above exercise, we get a short exact sequence \begin{equation} \label{eqn:fundexact} 1 \longrightarrow \wedge^2 \HHH \longrightarrow \Gamma \longrightarrow \HHH \longrightarrow 1, \end{equation} where $\Gamma = \pi / \gamma_2(\pi)$. \begin{exercise} The subgroup $\HHH < \Gamma$ is central. \end{exercise} The mapping class group acts on $\pi$ (this is where we use the fact that the basepoint is on the boundary component, so it is fixed by the mapping class group). This action preserves $\gamma_k(\pi)$ for all $\pi$ (indeed, all automorphisms of $\pi$ do this!). We thus get an action of $\Mod_{g,1}$ on $\Gamma$ which preserves $\wedge^2 \HHH < \Gamma$. Restrict this action to $\Torelli_{g,1}$. The group $\Torelli_{g,1}$ acts trivially on $\HHH$. \begin{exercise} The action of $\Torelli_{g,1}$ on $\wedge^2 \HHH < \Gamma$ is trivial. This looks obvious, but you have to trace through the above definitions to see that the action of $\Mod_{g,1}$ on $\wedge^2 \HHH < \Gamma$ is what you think it is. \end{exercise} Fix $f \in \Torelli_{g,1}$. For $x \in \Gamma$, observe that $f(x)$ and $x$ project to the same element of $\HHH$, so $x (f(x))^{-1} \in \wedge^2 \HHH$. Define a set map $J_f' : \Gamma \rightarrow \wedge^2 \HHH$ by $J_f'(x) = x (f(x))^{-1}$. \begin{exercise} The map $J_f'$ factors through a set map $J_f : \HHH \rightarrow \wedge^2 \HHH$. \end{exercise} \begin{exercise} The set map $J_f$ is a homomorphism. \end{exercise} We can thus define a set map $$\tau : \Torelli_{g,1} \rightarrow \Hom(\HHH,\wedge^2 \HHH)$$ by $\tau(f) = J_f$. \begin{exercise} $\tau$ is a homomorphism. \end{exercise} \section*{Lecture 3 : The abelianization of $\Mod_{g,n}(p)$} \addtocounter{section}{1} \setcounter{theorem}{0} \setcounter{figure}{0} In this lecture, we calculate $\HH_1(\Mod_{g,n}(p);\Z)$. To simplify our exposition, we will do the following. \begin{itemize} \item We will only consider the case of $\Mod_{g,1}(p)$. The case of $\Mod_g(p)$ can be dealt with in a similar way, but there are a few added complications. \item We will only consider the case where $p$ is odd. This greatly simplifies both the statements of the results and their proofs. \end{itemize} \ParagraphHeading{Main theorem.} Our main theorem is as follows. \begin{theorem} \label{theorem:modpabel} Fix $g,p \geq 3$ such that $p$ is odd. Set $\HHH_p = \HH_1(\Sigma_{g,1};\Z/p)$. There is then a short exact sequence $$0 \longrightarrow \wedge^3 \HHH_p \longrightarrow \HH_1(\Mod_{g,1}(p);\Z) \longrightarrow \HH_1(\Sp_{2g}(\Z,p);\Z) \longrightarrow 0.$$ \end{theorem} \begin{remarks} \mbox{} \begin{enumerate} \item Theorem \ref{theorem:modpabel} was proven independently by Putman \cite{PutmanModPAbel} and Sato \cite{SatoModPAbel}. At the same time, Perron \cite{PerronModPAbel} calculated $\HH_1(\Mod_{g,1}(p);\Z)$ up to a $2$-torsion ambiguity. \item Sato \cite{SatoModPAbel} also calculated $\HH_1(\Mod_{g,1}(2);\Z)$. The answer is more complicated than that given in Theorem \ref{theorem:modpabel}. This was later extended by Putman \cite{PutmanPicardGroupLevel} to $\HH_1(\Mod_{g,1}(p);\Z)$ for $p$ not divisible by $4$. The case where $p$ is divisible by $4$ is still open -- see the introduction of \cite{PutmanPicardGroupLevel} for a discussion. \item It was originally proven by Hain \cite{HainTorelli} that $\HH_1(\Mod_{g,1}(p);\Z)$ is finite for $g \geq 3$. \item At the end of this lecture, we will calculate $\HH_1(\Sp_{2g}(\Z,p);\Z)$. \end{enumerate} \end{remarks} \ParagraphHeading{Beginning of the proof.} Of course, Theorem \ref{theorem:modpabel} is derived from the short exact sequence $$1 \longrightarrow \Torelli_{g,1} \longrightarrow \Mod_{g,1}(p) \longrightarrow \Sp_{2g}(\Z,p) \longrightarrow 1.$$ We will need the following exercise. \begin{exercise} \label{exercise:short} Consider a short exact sequence $$1 \longrightarrow G_1 \longrightarrow G_2 \longrightarrow G_3 \longrightarrow 1$$ of groups. There is then a short exact sequence $$0 \longrightarrow V \longrightarrow \HH_1(G_2;\Z) \longrightarrow \HH_1(G_3;\Z) \longrightarrow 0,$$ where $V$ is the image of $\HH_1(G_1;\Z)$ in $\HH_1(G_2;\Z)$. \end{exercise} \noindent Applying Exercise \ref{exercise:short}, we see that it is enough to prove that the image of $\HH_1(\Torelli_{g,1};\Z)$ in $\HH_1(\Mod_{g,1}(p);\Z)$ is $\wedge^3 \HHH_p$. The key to this is the following lemma. \begin{lemma} \label{lemma:killl} For $g \geq 3$, let $v \in \HH_1(\Mod_{g,1}(p);\Z)$ be in the image of $\HH_1(\Torelli_{g,1};\Z)$. Then $p \cdot v = 0$. \end{lemma} Before proving Lemma \ref{lemma:killl}, let use it to finish the proof of Theorem \ref{theorem:modpabel}. Theorem \ref{theorem:torelliabel} says that $$\HH_1(\Torelli_{g,1};\Z) \cong W \oplus \wedge^3 \HHH,$$ where $W$ consists of $2$-torsion. Lemma \ref{lemma:killl} and the fact that $p$ is odd imply that the inclusion map $\HH_1(\Torelli_{g,1};\Z) \rightarrow \HH_1(\Mod_{g,1}(p);\Z)$ factors through $$(W / p \cdot W) \oplus (\wedge^3 \HHH)/(p \cdot \wedge^3 \HHH) = \wedge^3 \HHH_p.$$ Theorem \ref{theorem:modpjohnson} says that the image of $\HH_1(\Torelli_{g,1};\Z)$ in $\HH_1(\Mod_{g,1}(p);\Z)$ contains $\wedge^3 \HHH_p$, so we conclude that the image equals $\HHH_p$, as desired. \ParagraphHeading{The crossed lantern relation.} To prove Lemma \ref{lemma:killl}, we will need the following relation the mapping class group, which is known as the {\em crossed lantern relation}. \Figure{figure:crossedlantern}{CrossedLantern}{The crossed lantern relation $(T_{y_1} T_{y_2}^{-1})(T_{x_1} T_{x_2}^{-1}) = T_{z_1} T_{z_2}^{-1}$.} \begin{lemma} \label{lemma:crossedlantern} Let $x_i$ and $y_i$ and $z_i$ be the curves in Figure \ref{figure:crossedlantern}. Then $$(T_{y_1} T_{y_2}^{-1})(T_{x_1} T_{x_2}^{-1}) = T_{z_1} T_{z_2}^{-1}.$$ \end{lemma} \begin{proof} This rests on two key facts: $$T_{x_2}(y_i) = z_i \quad \text{and} \quad T_{y_2} T_{y_1}^{-1}(x_2) = x_1.$$ Repeatedly using the fact that $f T_{c} f^{-1} = T_{f(c)}$ for any simple closed curve $c$ and any mapping class $f$ (which can be proven exactly like Exercise \ref{exercise:conjugation}), we calculate \begin{align*} T_{x_1} T_{x_2}^{-1} &= \big((T_{y_2} T_{y_1}^{-1}) T_{x_2} (T_{y_1} T_{y_2}^{-1})\big) T_{x_2}^{-1} \\ &= (T_{y_2} T_{y_1}^{-1}) \big(T_{x_2} (T_{y_1} T_{y_2}^{-1}) T_{x_2}^{-1})\big) \\ &= T_{y_2} T_{y_1}^{-1} T_{z_1} T_{z_2}^{-1}. \end{align*} Rearranging this gives the desired relation. \end{proof} \begin{exercise} Give an alternate derivation of the crossed lantern relation by the same technique we used to construct the lantern relation in Lecture 1. Hint : The lantern relation comes from a relation $x y z = 1$ between simple closed curve in a surface group. Try to write down a different relation of the form $x y = z$ between simple closed curves. If you get stuck, see \cite[\S 3.1.3]{PutmanInfinite}. \end{exercise} \ParagraphHeading{The proof of Lemma \ref{lemma:killl}.} We are now in a position to prove Lemma \ref{lemma:killl}. For $f \in \Mod_{g,1}(p)$, let $[f]$ be the corresponding element of $\HH_1(\Mod_{g,1}(p);\Z)$. Theorem \ref{theorem:torelligen} says that $\Torelli_{g,1}$ is generated by bounding pair maps. Letting $T_{x_1} T_{x_2}^{-1}$ be a bounding pair map, it is thus enough to show that $p \cdot [T_{x_1} T_{x_2}^{-1}] = 0$. There exists an embedded subsurface $S$ of $\Sigma_{g,1}$ containing $\{x_1,x_2\}$ such that $S \cong \Sigma_{1,2}$ and such that the curves $\{x_1,x_2\}$ are embedded in $S$ as depicted in Figure \ref{figure:crossedlantern}. Let $\{y_1,y_2\}$ and $\{z_1,z_2\}$ be the curves in $S$ depicted in Figure \ref{figure:crossedlantern}. We thus have a crossed lantern relation \begin{equation} \label{eqn:crossedlantern} (T_{y_1} T_{y_2}^{-1}) (T_{x_1} T_{x_2}^{-1}) = T_{z_1} T_{z_2}^{-1} \end{equation} Also, $z_i = T_{x_2}(y_i)$ for $i=1,2$. The key observation is that for all $k \geq 0$, conjugating \eqref{eqn:crossedlantern} by $T_{x_2}^k$ results in another crossed lantern relation $$(T_{T_{x_2}^k(y_1)} T_{T_{x_2}^k(y_2)}^{-1}) (T_{x_1} T_{x_2}^{-1}) = (T_{T_{x_2}^{k+1}(y_1)} T_{T_{x_2}^{n+1}(y_2)}^{-1}).$$ Since $T_{x_2}^p \in \Mod_{g,1}(p)$, we conclude that $[T_{y_1} T_{y_2}^{-1}]$ is equal in $\HH_1(\Mod_{g,1}(p);\Z)$ to \begin{align*} [T_{x_2}^p (T_{y_1} T_{y_2}^{-1}) T_{x_2}^{-p}] &= [(T_{T_{x_2}^p(y_1)} T_{T_{x_2}^p(y_2)}^{-1})]\\ &= [T_{x_1} T_{x_2}^{-1}] + [(T_{T_{x_2}^{p-1}(y_1)} T_{T_{x_2}^{p-1}(y_2)}^{-1})]\\ &= 2[T_{x_1}T_{x_2}^{-1}] + [(T_{T_{x_2}^{p-2}(y_1)} T_{T_{x_2}^{p-2}(y_2)}^{-1})]\\ &\hspace{5.5pt}\vdots\\ &= p\cdot[T_{x_1}T_{x_2}^{-1}] + [T_{y_1} T_{y_2}^{-1}], \end{align*} so $p\cdot[T_{x_1}T_{x_2}^{-1}] = 0$, as desired. \ParagraphHeading{The abelianization of $\Sp_{2g}(\Z,p)$.} To complete our description of the abelianization of $\Mod_{g,1}(p)$, we need a description of $\HH_1(\Sp_{2g}(\Z,p);\Z)$. We first need some notation. Denote the $n \times n$ identity matrix by $\One_n$ and the $n \times n$ zero matrix by $\Zero_n$. Let $\Omega_g$ be the $2g \times 2g$ matrix $\MatTwoTwo{\Zero_g}{\One_g}{-\One_g}{\Zero_g}$. By definition, for a commutative ring $R$ the group $\Sp_{2g}(R)$ consists of all $2g \times 2g$ matrices $M$ with entries in $R$ such that $M^t \Omega_g M = \Omega_g$. Define $\SpLie_{2g}(R)$ to be the additive group of $2g \times 2g$ matrices $A$ with entries in $R$ such that $A^t \Omega_g + \Omega_g A = 0$. We then have the following theorem. \begin{theorem} \label{theorem:sppabel} Fix $g,p \geq 3$ such that $p$ is odd. Then $$\HH_1(\Sp_{2g}(\Z,p);\Z) \cong \SpLie_{2g}(\Z/p) \quad \text{and} \quad [\Sp_{2g}(\Z,p),\Sp_{2g}(\Z,p)] = \Sp_{2g}(\Z,p^2).$$ \end{theorem} \begin{remarks} \mbox{} \begin{enumerate} \item Theorem \ref{theorem:sppabel} was proved independently by Perron \cite{PerronModPAbel}, Putman \cite{PutmanModPAbel}, and Sato \cite{SatoModPAbel}. All three papers were inspired by a basic result of Lee and Szczarba \cite{LeeSzczarba} which calculates the abelianizations of congruence subgroups of $\SL_n(\Z)$. See Exercise \ref{exercise:leeszczarba} for a discussion of Lee and Szczarba's work. \item For $p$ even, Sato \cite{SatoModPAbel} proved that there is a short exact sequence $$0 \longrightarrow \HHH_2 \longrightarrow \HH_1(\Sp_{2g}(\Z,p);\Z) \longrightarrow \SpLie_{2g}(\Z/p) \longrightarrow 0.$$ Here $\HHH_2 = \HH_1(\Sigma_{g,1};\Z/2)$. \end{enumerate} \end{remarks} \ParagraphHeading{The map from $\Sp_{2g}(\Z,p)$ to $\SpLie_{2g}(\Z/p)$.} To prove Theorem \ref{theorem:sppabel}, we first construct a surjective homomorphism $$\psi : \Sp_{2g}(\Z,p) \longrightarrow \SpLie_{2g}(\Z/p).$$ Consider $M \in \Sp_{2g}(\Z,p)$. By definition, we can write $M = \One_{2g} + p \cdot A$ for some matrix $A$. Define $\psi(M) = A$ modulo $p$. We have $\psi(M) \in \SpLie_{2g}(\Z/p)$; indeed, by definition we have $\Omega_g = M^t \Omega_g M$, so \begin{align*} \Omega_g &= (\One_{2g} + p \cdot A)^t \Omega_g (\One_{2g} + p \cdot A) \\ &= \Omega_g + p \cdot (A^t \Omega_g + \Omega_g A) + p^2 A^{t} \Omega_g A. \end{align*} This implies that modulo $p$, we have $A^t \Omega_g + \Omega_g A = 0$. Next, we prove that $\psi$ is a homomorphism. Consider $M,N \in \Sp_{2g}(\Z,p)$ with $M = \One_{2g} + p A$ and $N = \One_{2g} + p B$. We then have $M N = \One_{2g} + p(A+B) + p^2 A B$, so modulo $p$ we have $\psi(MN) = A + B$. It remains to show that $\psi$ is surjective. If $p$ is prime, then this is easy. Indeed, $\psi$ is equivariant with respect to the conjugation actions of $\Sp_{2g}(\Z)$ on $\Sp_{2g}(\Z,p)$ and $\SpLie_{2g}(\Z/p)$. The latter factors through $\Sp_{2g}(\Z/p)$, so the image of $\psi$ is a $\Sp_{2g}(\Z/p)$-subrepresentation of $\SpLie_{2g}(\Z/p)$. However, it is well-known that $\SpLie_{2g}(\Z/p)$ is an irreducible $\Sp_{2g}(\Z/p)$-representation, so since $\psi$ is not trivial it must be surjective (we remark that this uses the fact that $p$ is odd -- the abelian group $\SpLie_{2g}(\Z/2)$ is {\em not} an irreducible representation of $\Sp_{2g}(\Z/2)$). We leave the general case as an exercise. \begin{exercise} Prove that $\psi$ is surjective by constructing elements of $\Sp_{2g}(\Z,p)$ that map to generators of $\SpLie_{2g}(\Z/p)$. \end{exercise} \ParagraphHeading{Finishing the proof of Theorem \ref{theorem:sppabel}.} Observe now that $\Ker(\psi) = \Sp_{2g}(\Z,p^2)$. Since the target of $\psi$ is abelian, this implies that $$[\Sp_{2g}(\Z,p),\Sp_{2g}(\Z,p)] \subset \Sp_{2g}(\Z,p^2).$$ Theorem \ref{theorem:sppabel} thus follows from the following lemma. \begin{lemma} \label{lemma:expressascommutators} If $g,p \geq 3$ and $p$ is odd, then $\Sp_{2g}(\Z,p^2) \subset [\Sp_{2g}(\Z,p),\Sp_{2g}(\Z,p)]$. \end{lemma} \noindent The proof of Lemma \ref{lemma:expressascommutators} follows from a direct matrix calculation. Rather than giving the details (which can be found in \cite[Lemma 2.4]{PutmanModPAbel} or \cite[Proposition 10.1]{SatoModPAbel}), we will give an exercise which outlines a proof of a somewhat easier result for $\SL_n(\Z)$. The proof of Lemma \ref{lemma:expressascommutators} follows the same basic pattern, though the details are more complicated. \begin{exercise} \label{exercise:leeszczarba} Let $\SL_n(\Z,p)$ be the kernel of the natural map $\SL_n(\Z) \rightarrow \SL_n(\Z/p)$. In this exercise, you will prove a theorem of Lee and Szczarba \cite{LeeSzczarba} which says that for $n \geq 3$, we have $\HH_1(\SL_n(\Z,p);\Z) \cong \SLLie_n(\Z/p)$. Here $\SLLie_n(\Z/p)$ is the abelian group of $n \times n$ matrices over $\Z/p$ with trace $0$. We remark that in this exercise, we do not need to assume that $p$ is odd. \begin{enumerate} \item Construct a surjective homomorphism $\SL_n(\Z,p) \rightarrow \SLLie_n(\Z/p)$ whose kernel is $\SL_n(\Z/p^2)$. Conclude that $[\SL_n(\Z,p),\SL_n(\Z,p)] \subset \SL_n(\Z,p^2)$. \item Prove that $\SL_n(\Z/p^2) \subset [\SL_n(\Z,p),\SL_n(\Z,p)]$. You will need the following theorem of Bass, Milnor, and Serre \cite{BassMilnorSerre}. For $1 \leq i,j \leq n$ such that $i \neq j$, let $e_{ij}$ be the {\em elementary matrix} which is obtained from the $n \times n$ identity matrix by changing the entry at position $(i,j)$ to $1$. Bass, Milnor, and Serre proved that for $n \geq 3$ and $q \geq 2$, the group $\SL_n(\Z,q)$ is normally generated (as a subgroup of $\SL_n(\Z)$) by the set $\Set{$e_{ij}^q$}{$1 \leq i,j \leq n$, $i \neq j$}$. \end{enumerate} \end{exercise} \section*{Lecture 4 : The second rational homology group of $\Mod_g(p)$} \addtocounter{section}{1} \setcounter{theorem}{0} \setcounter{figure}{0} In this final lecture, we turn to the second homology group of $\Mod_g(p)$. To set the stage, let us first recall what happens for the first homology group. Powell \cite{Powell} proved that $\HH_1(\Mod_g;\Z) = 0$ for $g \geq 3$. See \cite[\S 5.1]{FarbMargalitPrimer} for an easier proof (due to Harer). Hain \cite{HainTorelli} later proved that $\HH_1(\Mod_g(p);\Q)=0$ for $g \geq 3$. In other words, over $\Q$ the first homology group of $\Mod_g$ does not change when you pass to the finite-index subgroup $\Mod_g(p)$. Our proof of Theorem \ref{theorem:modpabel} in Lecture 3 is essentially an elaboration of Hain's proof. Harer \cite{HarerH2} proved that $\HH_2(\Mod_g;\Z) \cong \Z$ for $g \geq 4$. The following theorem says that a similar result holds for $\Mod_g(p)$ as long as we work over $\Q$. \begin{theorem}[{Putman, \cite{PutmanSecondHomologyLevel}}] \label{theorem:h2mod} For $g \geq 5$ and $p \geq 2$, we have $\HH_2(\Mod_g(p);\Q) \cong \Q$. \end{theorem} \noindent The first part of this lecture is devoted to motivating why one might expect such a theorem to hold. Next, we show that Theorem \ref{theorem:h2mod} is equivalent to a weak form of homological stability for $\Mod_g(p)$. Finally, we discuss how to prove this stability result. \ParagraphHeading{The transfer map.} What happens to the homology of a group when you pass to a finite-index subgroup? The following standard lemma says that over $\Q$, the homology can get larger but it cannot get smaller. \begin{lemma} \label{lemma:transfer} If $G$ is a finite-index subgroup of $\Gamma$, then the map $\HH_k(G;\Q) \rightarrow \HH_k(\Gamma;\Q)$ is surjective for all $k \geq 1$. \end{lemma} \begin{example} If $\Gamma$ is a free abelian group and $G$ is a finite-index subgroup of $\Gamma$, then there exists some $\ell \geq 1$ such that $\ell \cdot \Gamma < G < \Gamma$. Consequently, we have $G \cong \Gamma$ and it is easy to see that the map $\HH_k(G;\Q) \rightarrow \HH_k(\Gamma;\Q)$ is an isomorphism for all $k \geq 1$. \end{example} \begin{example} If $\Gamma$ is a free group of rank $n$ and $G$ is a proper finite-index subgroup of $\Gamma$, then $G$ is a free group of rank strictly greater than $n$. Consequently, the rank of $\HH_1(G;\Q)$ is strictly greater than the rank of $\HH_1(\Gamma;\Q) \cong \Q^n$. \end{example} \begin{remark} Lemma \ref{lemma:transfer} is false over $\Z$. Indeed, if $\Gamma$ is a finite group, then $G = 1$ is a finite-index subgroup of $\Gamma$. The homology groups of $G$ are all trivial, but $\Gamma$ can certainly have nontrivial homology groups. However, Lemma \ref{lemma:transfer} does imply that the homology groups of $\Gamma$ are all torsion. \end{remark} \begin{proof}[{Proof of Lemma \ref{lemma:transfer}}] Let $n = [\Gamma:G]$. The key to the proof is the {\em transfer map}. Letting $R$ be a commutative ring and letting $i_{\ast} : \HH_k(G;R) \rightarrow \HH_k(\Gamma;R)$ be the map induced by the inclusion $i : G \hookrightarrow \Gamma$, the transfer map is a homomorphism $\tau_k : \HH_k(\Gamma;R) \rightarrow \HH_k(G;R)$ satisfying \begin{equation} \label{eqn:transfer} i_{\ast}(\tau_k(x)) = n x \quad \text{for all $x \in \HH_k(\Gamma;R)$}. \end{equation} The existence of $\tau_k$ immediately implies the lemma. Indeed, if $R = \Q$, then $\frac{1}{n} \tau_k$ is a right-inverse to $i_{\ast}$, so $i_{\ast}$ is surjective. The transfer map is {\em not} induced by a homomorphism $\Gamma \rightarrow G$; rather, it is constructed at the level of chains. The construction goes as follows. Let $B\Gamma$ be a $K(\Gamma,1)$ and let $\rho : BG \rightarrow B\Gamma$ be the cover corresponding to $G < \Gamma$. Thus $BG$ is a $K(G,1)$ and $\rho$ is a degree $n$ cover. Define a map $\hat{\tau}_k : C_k(B\Gamma;R) \rightarrow C_k(BG;R)$ as follows. If $\sigma^k \in C_k(B\Gamma;R)$ is a singular $k$-simplex, then there are exactly $n$ singular $k$-simplices $\tilde{\sigma}^k_1, \ldots, \tilde{\sigma}^k_n$ in $BG$ satisfying $\rho_{\ast}(\tilde{\sigma}^k_i) = \sigma^k$. Define $$\hat{\tau}_k(\sigma^k) = \tilde{\sigma}^k_1 + \cdots + \tilde{\sigma}^k_n.$$ \begin{exercise} The maps $\hat{\tau}_k$ commute with the boundary operators, so we have a commutative diagram $$\begin{CD} \cdots @>{\partial}>> C_k(\Gamma;R) @>{\partial}>> C_{k-1}(\Gamma;R) @>{\partial}>> \cdots\\ @. @VV{\hat{\tau}_k}V @VV{\hat{\tau}_{k-1}}V @.\\ \cdots @>{\partial}>> C_k(G;R) @>{\partial}>> C_{k-1}(G;R) @>{\partial}>> \cdots \end{CD}$$ \end{exercise} \noindent By this exercise, the maps $\hat{\tau}_k$ induce maps $\tau_k : \HH_k(\Gamma;R) \rightarrow \HH_k(G;R)$ in homology. Clearly the maps $\tau_k$ satisfy \eqref{eqn:transfer}, so we are done. \end{proof} \ParagraphHeading{Borel stability.} Lemma \ref{lemma:transfer} implies that $\HH_2(\Mod_g(p);\Q) \cong \Q^m$ for some $m \geq 1$. Why should we expect that $m=1$? There is a very fruitful analogy between the mapping class group and lattices in Lie groups. The Borel stability theorem says that for the classical sequences of arithmetic lattices, passing to finite-index subgroups does not change their rational homology in a stable range. For example, denote $\SL_n(\Z)$ by $\Gamma_n$ and define $$\Gamma_n(p) = \Ker(\SL_n(\Z) \rightarrow \SL_n(\Z/p))$$ for $p \geq 2$. Here the map $\SL_n(\Z) \rightarrow \SL_n(\Z/p)$ comes from reducing all the entries in the matrices modulo $p$. Borel's theorem then takes the following form. \begin{theorem}[{Borel, \cite{BorelStability}}] \label{theorem:borelstab} For $k \geq 1$, there exists some $N_k \geq 1$ such that if $n \geq N_k$ and $p \geq 2$, then $\HH_k(\Gamma_n(p);\Q) \cong \HH_k(\Gamma_n;\Q)$. \end{theorem} \noindent One can view Theorem \ref{theorem:h2mod} as an analogue (for $k=2$) of Theorem \ref{theorem:borelstab} for the mapping class group. As we mentioned, Hain \cite{HainTorelli} proved a similar theorem for $k=1$. It would be very interesting to extend this to $k \geq 3$. \ParagraphHeading{Reduction to stability.} We now discuss the proof of Theorem \ref{theorem:h2mod}. In \cite{HarerStability}, Harer proved that the homology groups of $\Mod_g$ satisfy homological stability. See \cite{IvanovStability} or \cite{WahlStability} for more readable proofs of this. One special case of Harer's theorem is as follows. Choose a subsurface $S$ of $\Sigma_g$ such that $S \cong \Sigma_{g-1,1}$ (see Figure \ref{figure:stability}). Extending mapping classes on $S$ by the identity to $\Sigma_g$, we get a homomorphism $\Mod_{g-1,1} \rightarrow \Mod_g$. Harer proved that for each $k$, there exists some $N_k$ such that if $g \geq N_k$, then the induced map $\HH_k(\Mod_{g-1,1};\Z) \rightarrow \HH_k(\Mod_g;\Z)$ is an isomorphism. \Figure{figure:stability}{Stability}{$S$ is a subsurface of $\Sigma_g$ satisfying $S \cong \Sigma_{g-1,1}$ and $\gamma$ is a simple closed nonseparating curve contained in $\Sigma_g \setminus S$} The following result asserts that something similar happens for $\Mod_{g}(p)$ for $k=2$. \begin{lemma} \label{lemma:stability1} For $g \geq 5$ and $p \geq 2$, let $i : \Mod_{g-1,1}(p) \rightarrow \Mod_g(p)$ be the restriction of the map $\Mod_{g-1,1} \rightarrow \Mod_g$ described above. Then the induced map $i_{\ast} : \HH_2(\Mod_{g-1,1}(p);\Q) \rightarrow \HH_2(\Mod_g(p);\Q)$ is surjective. \end{lemma} \noindent Observe that Theorem \ref{theorem:h2mod} implies Lemma \ref{lemma:stability1}. Indeed, if Theorem \ref{theorem:h2mod} is true, then we have a commutative diagram of the form $$\begin{CD} \HH_2(\Mod_{g-1,1}(p);\Q) @>>> \HH_2(\Mod_g(p);\Q)\\ @VVV @VV{\cong}V \\ \HH_2(\Mod_{g-1,1};\Q) @>>> \HH_2(\Mod_g;\Q) \end{CD}$$ By Lemma \ref{lemma:transfer}, the map $\HH_2(\Mod_{g-1,1}(p);\Q) \rightarrow \HH_2(\Mod_{g-1,1};\Q)$ is a surjection, and Harer's stability theorem say that the map $\HH_2(\Mod_{g-1,1};\Q) \rightarrow \HH_2(\Mod_g;\Q)$ is an isomorphism as long as $g \geq 5$. We conclude that the map $\HH_2(\Mod_{g-1,1}(p);\Q) \rightarrow \HH_2(\Mod_g(p);\Q)$ is a surjection, as desired. Somewhat surprisingly, Lemma \ref{lemma:stability1} also implies Theorem \ref{theorem:h2mod}. \begin{lemma} \label{lemma:stabilityimpliesiso} Lemma \ref{lemma:stability1} implies Theorem \ref{theorem:h2mod}. \end{lemma} \noindent Before proving Lemma \ref{lemma:stabilityimpliesiso}, we need some more abstract nonsense. Let $G$ be a finite-index normal subgroup of $\Gamma$. The conjugation action of $\Gamma$ on $G$ induces an action of $\Gamma$ on $\HH_k(G;\Q)$. Recall that if $M$ is a vector space upon which $\Gamma$ acts, then the {\em coinvariants} of the action, denoted $M_{\Gamma}$, is the quotient $M/I$ with $I$ the subspace spanned by $\Set{$x-g(x)$}{$x \in M$, $g \in \Gamma$}$. In other words, $M_{\Gamma}$ is the largest quotient of $M$ upon which $\Gamma$ acts trivially. Since $\Gamma$ acts trivially on its own homology groups $\HH_k(\Gamma;\Q)$, the map $\HH_k(G;\Q) \rightarrow \HH_k(\Gamma;\Q)$ factors through $(\HH_k(G;\Q))_{\Gamma}$. We then have the following exercise. \begin{exercise} \label{exercise:supertransfer} $\HH_k(\Gamma;\Q) \cong (\HH_k(G;\Q))_{\Gamma}$. Hint : carefully study the proof of Lemma \ref{lemma:transfer}. Alternatively, this can be proven using the Hochschild-Serre spectral sequence of the extension $$1 \longrightarrow G \longrightarrow \Gamma \longrightarrow \Gamma/G \longrightarrow 1.$$ \end{exercise} \begin{proof}[{Proof of Lemma \ref{lemma:stabilityimpliesiso}}] Assume that Lemma \ref{lemma:stability1} holds. Since $\HH_2(\Mod_g;\Q) \cong \Q$, to prove that $\HH_2(\Mod_g(p);\Q) \cong \Q$, it is enough to prove that $\HH_2(\Mod_g(p);\Q) \cong \HH_2(\Mod_g;\Q)$. By Exercise \ref{exercise:supertransfer}, this is equivalent to showing that the action of $\Mod_g$ on $\HH_2(\Mod_g(p);\Q)$ is trivial. This can be checked on a generating set. The group $\Mod_g$ is generated by the set of Dehn twists about nonseparating simple closed curves. Consider such a Dehn twist $T_{\gamma}$. As is shown in Figure \ref{figure:stability}, we can find a subsurface $S$ of $\Sigma_g$ such that $S \cong \Sigma_{g-1,1}$ and such that $\gamma \subset \Sigma_g \setminus S$. Let $i : \Mod_{g-1,1}(p) \rightarrow \Mod_g(p)$ be the map induced by the subsurface inclusion $S \hookrightarrow \Sigma_g$. Clearly $T_{\gamma}$ commutes with $\Image(i)$, so $T_{\gamma}$ acts trivially on $i_{\ast}(\HH_2(\Mod_{g-1,1}(p);\Q))$. Lemma \ref{lemma:stability1} implies that $i_{\ast}(\HH_2(\Mod_{g-1,1}(p);\Q)) = \HH_2(\Mod_g(p);\Q)$, so we conclude that $T_{\gamma}$ acts trivially on $\HH_2(\Mod_g(p);\Q)$, as desired. \end{proof} \ParagraphHeading{A weaker result suffices.} In fact, we do not need the full strength of Lemma \ref{lemma:stability1} to prove Theorem \ref{theorem:h2mod}. If $\gamma$ is the isotopy class of a nonseparating simple closed curve on $\Sigma_g$, then denote by $(\Mod_g(p))_{\gamma}$ the subgroup of $\Mod_g(p)$ which stabilizes $\gamma$. Examining the proof of Lemma \ref{lemma:stabilityimpliesiso}, it is clear that the following lemma also implies Theorem \ref{theorem:h2mod}. \begin{lemma} \label{lemma:stability2} For $g \geq 5$ and $p \geq 2$, let $\gamma$ be a nonseparating simple closed curve on $\Sigma_{g}$. Then the map $\HH_2((\Mod_g(p))_{\gamma};\Q) \rightarrow \HH_2(\Mod_g(p);\Q)$ is surjective. \end{lemma} \ParagraphHeading{The homological stability machine.} Homological stability theorems are known for many different sequences of groups, and there is now a standard procedure for proving them. This procedure goes back to Quillen. See \cite{HatcherWahl} for an excellent discussion of how this machine applies to many different sequences of groups and \cite{WahlStability} for a detailed exposition of how to apply it to the mapping class group. The input for this machine consists of a sequence of highly connected spaces upon which the groups in question act. For $\Mod_g$, Harer proved homological stability by studying the action on the following space. \begin{definition} The {\em nonseparating curve complex}, denoted $\CNosep_g$, is the simplicial complex whose $k$-simplices are sets $\{\gamma_0,\ldots,\gamma_k\}$ of isotopy classes of simple closed curves on $\Sigma_g$ that can be realized such that the $\gamma_i$ are all disjoint and $\Sigma_g \setminus (\gamma_0 \cup \cdots \cup \gamma_k)$ is connected. \end{definition} One's first impulse, of course, is to apply this machine to $\Mod_g(p)$ to try to prove Lemma \ref{lemma:stability2}. Unfortunately, it does not quite work. The problem is that if it did work, then it would prove that $\HH_k(\Mod_g(p);\Z)$ is stable, and this is false even for $k=1$ (see Theorem \ref{theorem:modpabel}). The machine does, however, give the following result. \begin{lemma} \label{lemma:weakstability} For $g \geq 5$ and $p \geq 2$, the map $$\bigoplus_{\gamma \in (\CNosep_g)^{(0)}} \HH_2((\Mod_g(p))_{\gamma};\Q) \longrightarrow \HH_2(\Mod_g(p);\Q)$$ is surjective. \end{lemma} \ParagraphHeading{Rescuing the machine.} To deduce Lemma \ref{lemma:stability2} from Lemma \ref{lemma:weakstability}, we need to prove that if $g \geq 5$ and $p \geq 2$, then for any two nonseparating simple closed curves $\gamma$ and $\gamma'$ on $\Sigma_g$, the images of $\HH_2((\Mod_g(p))_{\gamma};\Q)$ and $\HH_2((\Mod_g(p))_{\gamma'};\Q)$ in $\HH_2(\Mod_g(p);\Q)$ are the same. A standard result about surfaces called the ``change of coordinates principle'' (see \cite{FarbMargalitPrimer}) says that there exists some $\phi \in \Mod_g$ such that $\phi(\gamma) = \gamma'$. This implies that $$\phi (\Mod_g(p))_{\gamma} \phi^{-1} = (\Mod_g(p))_{\gamma'}.$$ Thus the action of $\Mod_g$ on $\HH_2(\Mod_g(p);\Q)$ takes the image of $\HH_2((\Mod_g(p))_{\gamma};\Q)$ to the image of $\HH_2((\Mod_g(p))_{\gamma'};\Q)$. We conclude that Lemma \ref{lemma:weakstability} would following from the following lemma. \begin{lemma} \label{lemma:trivialaction} For $g \geq 5$ and $p \geq 2$, if $\gamma$ is a nonseparating simple closed curve on $\Sigma_g$, then $\Mod_g$ acts trivially on the image of $\HH_2((\Mod_g(p))_{\gamma};\Q)$ in $\HH_2(\Mod_g(p);\Q)$. \end{lemma} The proof of Lemma \ref{lemma:trivialaction} is quite complicated (see \cite{PutmanSecondHomologyLevel}). Rather than give it here, we will prove an analogous result about $\Gamma_n = \SL_n(\Z)$. Recall that $$\Gamma_n(p) = \Ker(\SL_n(\Z) \longrightarrow \SL_n(\Z/p)).$$ There is a natural map $\Gamma_{n-1} \rightarrow \Gamma_{n}$ which takes a matrix $A \in \Gamma_{n-1}$ to the matrix $\MatTwoTwo{1}{0}{0}{A} \in \Gamma_{n}$. In \cite{CharneyCongruence}, Charney proves the following theorem. \begin{theorem}[{Charney, \cite[Proposition 5.5]{CharneyCongruence}}] \label{theorem:charney} For $n \geq 2$ and $k,p \geq 1$, the group $\Gamma_n$ acts trivially on the image of $\HH_k(\Gamma_{n-1}(p);\Q)$ in $\HH_k(\Gamma_n(p);\Q)$. \end{theorem} \noindent Though she does not say it in her paper, it is not hard to combine this with the standard homological stability machinery to give a new proof of the Borel stability theorem for $\SL_n(\Z)$. After the proof, we will comment on the relation between the proofs of Theorem \ref{theorem:charney} and Lemma \ref{lemma:trivialaction}. \begin{proof}[{Proof of Theorem \ref{theorem:charney}}] For $1 \leq i,j \leq n$ such that $i \neq j$, let $e_{ij}$ be the {\em elementary matrix} which is obtained from the $n \times n$ identity matrix by changing the entry at position $(i,j)$ to $1$. The group $\Gamma_n$ is generated by the set $\Set{$e_{ij}$}{$1 \leq i,j \leq n$, $i \neq j$}$. However, we do not need this entire set. \begin{exercise} For $2 \leq i,j \leq n$ such that $i \neq j$, prove that $[e_{i1},e_{1j}] = e_{ij}$. \end{exercise} \noindent Consequently, we only need to prove that $e_{1j}$ and $e_{i1}$ act trivially for $2 \leq i,j \leq n$. We will give the details for $e_{1j}$. The other case is similar. Consider the subgroups $$G = \Set{$\left( \begin{array}{c|c} 1 & c_2 \ \cdots\ c_n \\ \hline 0 & \multirow{3}{*}{\raisebox{-5mm}{\scalebox{2}{$A$}}} \\ \raisebox{2mm}{\vdots} & \\ 0 & \\ \end{array}\right)$} {$A \in \Gamma_{n-1}(p)$, $c_2,\dots,c_n \in p\Z$}$$ and $$\hat{G} = \Set{$\left( \begin{array}{c|c} 1 & c_2 \ \cdots\ c_n \\ \hline 0 & \multirow{3}{*}{\raisebox{-5mm}{\scalebox{2}{$A$}}} \\ \raisebox{2mm}{\vdots} & \\ 0 & \\ \end{array}\right)$} {$A \in \Gamma_{n-1}(p)$, $c_2,\dots,c_n \in \Z$}$$ of $\Gamma_n$. The conjugation action of $e_{1j}$ on $\Gamma_n$ takes $G$ and $\hat{G}$ to themselves. We have $\Gamma_{n-1}(p) \subset G \subset \Gamma_n(p)$, so it is enough to show that $e_{1j}$ acts trivially on $\HH_k(G;\Q)$. Since $e_{1j} \in \hat{G}$, the action of $e_{1j}$ on $\HH_k(\hat{G};\Q)$ is trivial. It is therefore enough to show that the inclusion $G \hookrightarrow \hat{G}$ induces an isomorphism $\HH_k(G;\Q) \cong \HH_k(\hat{G};\Q)$. Define $$K = \Set{$\left( \begin{array}{c|c} 1 & c_2 \ \cdots\ c_n \\ \hline 0 & \multirow{3}{*}{\raisebox{-5mm}{\scalebox{2}{$1$}}} \\ \raisebox{2mm}{\vdots} & \\ 0 & \\ \end{array}\right)$} {$c_2,\dots,c_n \in p\Z$} \subset G.$$ Observe that $K \cong \Z^{n-1}$. We have a short exact sequence $$\begin{CD} 1 @>>> K @>>> G @>>> \Gamma_{n-1}(p) @>>> 1. \end{CD}$$ Similarly, setting $$\hat{K} = \Set{$\left( \begin{array}{c|c} 1 & c_2 \ \cdots\ c_n \\ \hline 0 & \multirow{3}{*}{\raisebox{-5mm}{\scalebox{2}{$1$}}} \\ \raisebox{2mm}{\vdots} & \\ 0 & \\ \end{array}\right)$} {$c_2,\dots,c_n \in \Z$} \subset \hat{G}$$ we have a short exact sequence $$\begin{CD} 1 @>>> \hat{K} @>>> \hat{G} @>>> \Gamma_{n-1}(p) @>>> 1. \end{CD}$$ These short exact sequences fit into a commutative diagram of the form $$\begin{CD} 1 @>>> K @>>> G @>>> \Gamma_{n-1}(p) @>>> 1 \\ @. @VVV @VVV @VV{=}V @.\\ 1 @>>> \hat{K} @>>> \hat{G} @>>> \Gamma_{n-1}(p) @>>> 1. \end{CD}$$ By Lemma \ref{lemma:grouphomology5} (a sort of $5$-lemma for group homology), it is enough to show that the inclusion map $K \hookrightarrow \hat{K}$ induces an isomorphism $\HH_k(K;\Q) \cong \HH_k(\hat{K};\Q)$. However, we have $\hat{K} \cong \Z^{n-1}$ and $K = p \cdot \hat{K}$, so this is immediate. We remark that this final step is the only place we use the fact that we are working over $\Q$. \end{proof} \begin{lemma}[{$5$-lemma for group homology}] \label{lemma:grouphomology5} Fix a commutative ring $R$. Consider a commutative diagram $$\begin{CD} 1 @>>> A_1 @>>> B_1 @>>> C @>>> 1 \\ @. @VVV @VVV @VV{=}V @.\\ 1 @>>> A_2 @>>> B_2 @>>> C @>>> 1 \end{CD}$$ of short exact sequences of groups. Assume that the map $A_1 \rightarrow A_2$ induces an isomorphism $\HH_k(A_1;R) \cong \HH_k(A_2;R)$ for all $k$. Then the map $B_1 \rightarrow B_2$ induces an isomorphism $\HH_k(B_1;R) \cong \HH_k(B_2;R)$ for all $k$. \end{lemma} \begin{proof} Associated each of our short exact sequences is a Hochschild-Serre spectral sequence in group homology. The assumptions in the lemma imply that the induced map between these spectral sequences is an isomorphism on the $E^2$-page. This implies that it converges an an isomorphism on the $E^{\infty}$-pages, so we obtain isomorphisms $\HH_k(B_1;R) \cong \HH_k(B_2;R)$ for all $k$. \end{proof} \ParagraphHeading{Relation between Theorem \ref{theorem:charney} and Lemma \ref{lemma:trivialaction}.} The proof of Lemma \ref{lemma:trivialaction} in \cite{PutmanSecondHomologyLevel} is not nearly as short as the proof above of Theorem \ref{theorem:charney}. However, they share some features. The key to the proof of Theorem \ref{theorem:charney} above is the short exact sequence $$1 \longrightarrow K \longrightarrow G \longrightarrow \Gamma_{n-1}(p) \longrightarrow 1.$$ A similar role in the proof of Lemma \ref{lemma:trivialaction} is played by various analogues for $\Mod_g(p)$ of the Birman exact sequence which was discussed in Lecture 1. However, the kernels of these Birman exact sequences are not free abelian, so passing to finite-index subgroups actually changes their isomorphism types. This necessitates making a number of rather intricate twisted group cohomology computations. \ParagraphHeading{Recap.} The above proof that $\HH_2(\Mod_g(p);\Q) \cong \Q$ was a little involved, so let us recap the main steps. \begin{enumerate} \item We first showed that $\HH_2(\Mod_g(p);\Q) \cong \Q$ if and only if $\HH_2(\Mod_g(p);\Q)$ satisfies a weak form of homological stability (see Lemma \ref{lemma:stabilityimpliesiso}). \item We then attempted to use the standard homological stability machinery to prove that $\HH_2(\Mod_g(p);\Q)$ stabilizes. This failed, but it showed that $\HH_2(\Mod_g(p);\Q)$ is ``concentrated'' on stabilizers of simple closed curves (see Lemma \ref{lemma:weakstability}). \item The proof is reduced to showing that all stabilizers of simple closed curves give the same ``chunk'' of $\HH_2(\Mod_g(p);\Q)$. \end{enumerate} It seems reasonable to conjecture that $\HH_k(\Mod_g(p);\Q) \cong \HH_k(\Mod_g;\Q)$ in a stable range for all $k$. The first two steps above can be easily generalized to $\HH_k(\Mod_g(p);\Q)$ for $k \geq 3$. The real difficulty is the third step. It would be very interesting to extend it to the higher homology groups.

https://arxiv.org/abs/1705.09204

On the generalized principal eigenvalue of quasilinear operators

The notions of generalized principal eigenvalue for linear second order elliptic operators in general domains introduced by Berestycki et al. \cite{BNV,BR0,BR3} have become a very useful and important tool in analysis of partial differential equations. In this paper, we extend these notions for quasilinear operator of the form $$\CK_V[u]:=-\Delta_p u +Vu^{p-1},\quad\quad u \geq0.$$ This operator is a natural generalization of self-adjoint linear operators. If $Ø$ is a smooth bounded domain, we already proved in \cite{NV} that the generalized principal eigenvalue coincides with the (classical) first eigenvalue of $\CK_V$. Here we investigate the relation between three types of the generalized principal eigenvalue for quasilinear operator on general smooth domain (possibly unbounded), which plays an important role in the investigation of their asymptotic properties. These results form the basis for the study of the simplicity of the generalized principal eigenvalues, the maximum principle and the spectrum of $\CK_V$. We further discuss applications of the notions by providing some examples.

\section{Introduction and Main Results} The principal eigenvalue is a basic notion associated with elliptic operators and plays a crucial role in the analysis of partial differential equation, especially in the study of semilinear elliptic problems. The principal eigenvalue for quasilinear operators is also the subject of intensive research since not only it is a natural extension of that of linear operators but also it allows to bring into light new phenomena which stem from the interesting structure of quasilinear operators. In this paper, we investigate the \textit{ generalized principal eigenvalue} of the operator \bel{KV}\CK_V[u]:=-\Delta_p u + V u^{p-1},\quad u \geq 0, \end{equation} in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi \subset\mathbb{R}^N$ (possibly unbounded), where $\Gd_p u =\text{div} (|\nabla u|^{p-2}\nabla u)$ with $p>1$ and $V \in L_{\text{\rm loc}}^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$, $\inf_{\Omega}V>-\infty$. If $\Omega$ is a $C^{1,\nu}$ ($0<\nu<1$) bounded domain and $V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$, it is well-known that the variational problem \begin{equation} \label{varprob} \gl_V^{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}:=\inf_{\phi \in W_0^{1,p}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \setminus \{0\}}\frac{\int_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}(|\nabla \phi} \def\vgf{\varphi} \def\gh{\eta|^p + V|\phi} \def\vgf{\varphi} \def\gh{\eta|^p)dx}{\int_{\Omega}|\phi|^pdx} \end{equation} admits a unique (up to multiplicative constants) positive minimizer $\vgf$ (see, e.g., \cite{DKN}, \cite[Lemma 3]{GS}). Moreover, $\vgf \in C^{1,\theta} \def\vge{\varepsilon}$ ($0<\theta} \def\vge{\varepsilon<1$) and it is a positive solution of the quasilinear eigenvalue problem \bel{eigenprob} \left\{ \begin{aligned} \CK_V [\vgf] &= \gl_V^{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}\vgf^{p-1} \qquad &&\text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi \\ \phantom{-,,} \vgf &= 0 &&\text{on } \partial \Omega} \def\Gx{\Xi} \def\Gy{\Psi. \end{aligned} \right. \end{equation} Here $\gl_V^{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}$ and $\vgf$ are called respectively the \emph{principal eigenvalue} and \emph{eigenfunction} of $\CK_V$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$. Note that since $C_c^1(\Omega)$ is dense in $W_0^{1,p}(\Omega)$ with respect to $W^{1,p}$ norm, the infimum in (\ref{lambda}) can be taken over $C_c^1(\Omega)$. Problem \eqref{eigenprob} has received much attention in the literature because it has various applications, most of which arise from problems in fluid dynamics, where the p-Laplacian operator with $p\neq 2$ is employed to study non-Newtonian fluids ($p > 2$ for dilatant fluids and $p < 2$ for pseudoplastic fluids). This kind of problem has also been used to develop noise reduction and edge detection techniques in image processing (see \cite{CLMC}), where the degenerate diffusion term enables to smoothen the image without destroying the edges. When $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a general (possibly unbounded) domain, we introduced a notion of generalized principal eigenvalue of $\CK_V$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ \cite{NV} \begin{definition} \label{lambda} (i) The quantity \bel{gpe0} \gl(\CK_V,\Omega):=\sup \{\lambda\in\mathbb{R} |\, \, \exists\psi \in W_{\text{\rm loc}}^{1,p}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi),\psi>0, \CK_V[\psi] \geq \lambda \psi^{p-1} ~ \text{{ \textit{in the weak sense}} in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi \} \end{equation} is called a {\em generalized principal eigenvalue} of $\CK_V$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$. Here, the inequality holds \textit{in the weak sense in $\Omega$} means \begin{equation}\label{weaksense} \int_{\Omega}|\nabla\psi|^{p-2}\nabla\psi\cdot\nabla\phi \, dx+\int_{\Omega}V\psi^{p-1}\phi\, dx\geq \lambda\int_{\Omega}\psi^{p-1}\phi \,dx\quad\quad\forall\phi\in C^\infty_c(\Omega). \end{equation} The functions $\psi$ in \eqref{gpe0} are called {\em admissible test functions} for $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. \smallskip (ii) We say that $\gl \in \BBR$ is an eigenvalue of $\CK_V$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ if there exists a positive weak solution $u \in W_{loc}^{1,p}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ of \bel{eigenRN} \CK_V[u]=\gl u^{p-1} \quad \text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi. \end{equation} Such a solution $u$ is called an eigenfunction of $\CK_V$ associated with $\gl$. Denote by $\CE(\Omega)$ the set of all eigenvalues of $\CK_V$ in $\Omega$. \end{definition} An important feature of the notion of generalized principal eigenvalue is that if $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a smooth bounded domain, $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ coincides with the principal eigenvalue $\gl_V^\Omega} \def\Gx{\Xi} \def\Gy{\Psi$, while if $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is unbounded $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ is well defined and can be expressed by a variational formula. This type of eigenvalue is of purely mathematical interest since it is an effective tool in the study of many problems. Indeed, the role of $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ is clearly described in the analysis of equation \bel{nonlinearR} \CK_V[u] + b \,g(u)=0 \quad \text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi \end{equation} where $0 \leq b \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ and $t \mapsto g(t)/t^{p-1}$ is increasing. We refer the reader to \cite{CDG, DuGu} for the case when $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is bounded and to \cite{NV} for the case $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$. In particular, in \cite{NV}, under the assumption on the asymptotic behavior of $V$ near infinity $$ \liminf_{|x| \to \infty}|x|^q V(x)>0 \quad \text{for some } q \in [0,p],\, p>1, $$ we have proved that: \begin{theorem} \label{JFA} (See \cite[Theorem 1.3 and Theorem 1.44]{NV}) \noindent {\sc I. Existence and Uniqueness.} If $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)<0$ then there exists a unique decaying solution of \eqref{nonlinearR}. Moreover, (i)\, If $q \in [0,p)$ then the unique solution decays exponentially (ii)\, If $q=p$ then the solution decays polynomially. \smallskip \noindent {\sc II. Nonexistence.} If $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \geq 0$ then there exists no decaying solution of \eqref{nonlinearR}. \end{theorem} \noindent We emphasize that when $p \geq 2$, the existence, uniqueness and nonexistence results hold in a much larger class of functions, including bounded functions. For more details, we refer the reader to \cite{NV}. It is noteworthy that the notion of generalized principal eigenvalue in Definition \ref{lambda} is closely related to \textit{the best constant in the Hardy-type inequality} which was introduced by Pinchover et al. to establish optimal Hardy-type inequalities (see \cite{DFP,DePi}). This notion is also used to study the structure of positive solution homogeneous equation $\CK_V[u]=0$ in unbounded domains (see, e.g., \cite{FO,Murata, Pinchover}). Moreover, it is directly related to the characterization of the Liouville type result and the maximum principle \cite{Pinchover1, BDPR, GS, BNV,QS,BR3}. Therefore, the investigation of the principal eigenvalue is a crucial ingredient to deal with many fundamental questions in the theory of partial differential equations. The aim of the present paper is to bring the eigentheory for quasilinear operators closer to the level of the well-studied linear case (see \cite{BNV,BR0,BR3}) by establishing qualitative properties, the simplicity, the spectrum of $\CK_V$ and the maximum principle. To this end, in the spirit of the papers \cite{BNV,BR0,BR3}, we introduce other notions of the generalized principal eigenvalue as follows \begin{definition}\label{gpe} Let $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ be a (possibly unbounded) domain in $\mathbb{R}^N$. Define $$ \begin{aligned} &\gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi):=\inf \{\lambda\in\mathbb{R} |\, \, \exists\psi \in W_{0}^{1,p}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi),\psi>0, \CK_V[\psi] \leq \lambda \psi^{p-1} ~ \text{{ \textit{in the weak sense}} in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi\},\\ &\gl''(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi):=\sup \{\lambda\in\mathbb{R} |\, \, \exists\psi \in C_{\text{\rm loc}}^{1}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi), \inf_{\Omega}\psi>0, \CK_V[\psi] \geq \lambda \psi^{p-1} ~ \text{{ \textit{in the weak sense}} in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi \}, \end{aligned} $$ where the inequalities are understood in the weak sense in $\Omega$ as in \eqref{weaksense}. The functions $\psi$ in \eqref{gpe0} are called {\em admissible test functions}. \end{definition} Note that since $\inf_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}V>-\infty$, it follows that $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)>-\infty$ and $\gl''(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)>-\infty$. It may occur that the set of admissible test functions in the definition of $\gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ is empty. In such a case, we set $\gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)=+\infty$. The main difference between our notions of the \textit{generalized principal eigenvalues} and those introduced by Berestycki et al. \cite{BNV,BR0,BR3} is that in Definitions \ref{lambda} and \ref{gpe} admissible test functions $\psi$ are only required to be in $W^{1,p}$ or in $C_{loc}^1$ and all the inequalities are only required to hold \textit{in the weak sense in} $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ while admissible test functions in the definitions of the generalized principal eigenvalue in \cite{BNV,BR0,BR3} belong to $W^{2,N}_{loc}$ and the inequalities are understood \textit{almost everywhere in} $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$. Moreover, our definition of $\lambda'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ differs from the one defined by Berestycki and Rossi for second order linear operators in \cite{BR3} in the sense that we impose admissible test functions $\psi$ to be in $W^{1,p}_0(\Omega)$ instead of requiring them to be in $W_{loc}^{2,N}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ and to satisfy $\lim_{x\to\zeta}\psi(x)=0$ for every $\zeta \in \partial \Omega} \def\Gx{\Xi} \def\Gy{\Psi$. Despite these differences, our notions fit well into the framework of quasilinear operators and allow to obtain the properties of the generalized principal eigenvalue. We emphasize that, many new ideas have been developed in this paper to overcome the fundamental difficulties stemming from the nonlinearity of p-Laplacian since most of the techniques used in \cite{BNV,BR0,BR3} fail to apply in this framework, especially to obtain the relation of the notions of the generalized principal eigenvalue, the simplicity and the maximum principle. Observe, by Definitions \ref{lambda} and \ref{gpe}, that $\gl''(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \leq \gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. Our first result reveals a more profound relation between three notions of generalized principal eigenvalues. \begin{theorem}\label{equivalence} Let $\Omega\subset\mathbb{R}^N$. Assume $0 \not \equiv V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. \noindent (i)\, If $\Omega$ is a smooth bounded domain then $$\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)=\gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) =\gl''(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi).$$ \noindent (ii)\, If $\Omega=\mathbb{R}^N$ then $$\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) = \gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \geq \gl''(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi).$$ \noindent (iii)\, If $\Omega=\mathbb{R}^N$ and $V$ is radially symmetric then $$\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) = \gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) =\gl''(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi).$$ \end{theorem} The following result describes the effect of the diffusion coefficient on the generalized principal eigenvalue. \begin{theorem} \label{lim1} Let $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a (possibly unbounded) domain in $\BBR^N$ and $V \in L_{loc}^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. For $\alpha} \def\gb{\beta} \def\gg{\gamma>0$, denote \bel{La} \CL_\alpha[\phi]:=-\alpha\Delta_p\phi+V\phi^{p-1},\quad\phi\geq 0.\end{equation} Then the following properties hold. \noindent (i)\, The mapping $\alpha} \def\gb{\beta} \def\gg{\gamma \mapsto \gl(\CL_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ is concave, nondecreasing and $$\lim_{\alpha\to0}\lambda(\CL_\alpha,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)=\inf_\Omega} \def\Gx{\Xi} \def\Gy{\Psi V.$$ \noindent (ii)\, If $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a smooth bounded domain then $$\lim_{\alpha\to+\infty}\lambda(\CL_\alpha,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)=\infty.$$ \noindent (iii)\, If $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\mathbb{R}^N$ then \bel{limsup} \limsup_{\alpha\to+\infty}\lambda(\CL_\alpha,\BBR^N)\leq\limsup_{{|x|\to+\infty}} V(x), \end{equation} \bel{liminf} \liminf_{\alpha\to+\infty}\lambda(\CL_\alpha,\BBR^N)\geq\liminf_{{|x|\to+\infty}} V(x). \end{equation} \end{theorem} Theorem \eqref{lim1} yields different interesting phenomena of the effect of diffusion coefficient on the generalized principal eigenvalue between bounded and unbounded domains. From (i) and (ii) and Theorem \ref{JFA}, one sees that if $\Omega$ is bounded then for any bounded potential $V$ with negative infimum, equation \eqref{nonlinearR} admits a positive solution for $\alpha$ small and does not admit any positive solution for $\alpha$ large. The phenomenon is strikingly different when $\Omega=\mathbb{R}^N$. Assume there exists $\lim_{|x|\to\infty}V(x)=\ell$. From \eqref{limsup}) and \eqref{liminf}, if $\ell<0$, then $$\lim_{\alpha\to+\infty}\lambda(\CL_\alpha,\BBR^N)=\ell<0,$$ and Theorem \ref{JFA} implies that equation (\ref{nonlinearR}) admits a unique positive solution for $\alpha$ large while if $\ell>0$ equation (\ref{nonlinearR}) admits no positive solution for $\alpha$ near $+\infty$. Thanks to Theorems \ref{equivalence} and \ref{lim1}, we obtain a result about the effect of the potential on the generalized principal eigenvalues. If $\{V_\alpha} \def\gb{\beta} \def\gg{\gamma\}$ is a sequence of functions in $L_{loc}^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ we denote \bel{Ka} \CK_\alpha} \def\gb{\beta} \def\gg{\gamma[\phi]:=-\Delta_p\phi+V_\alpha} \def\gb{\beta} \def\gg{\gamma\,\phi^{p-1}, \quad \phi} \def\vgf{\varphi} \def\gh{\eta \geq 0. \end{equation} \begin{theorem}\label{lim2} Let $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ be a (possibly unbounded) domain and $V \in L_{loc}^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. \noindent (i)\, Denote $V_\alpha(x)=V(\alpha x)$ for $\alpha} \def\gb{\beta} \def\gg{\gamma>0$. If $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$ and $V(0)=\inf_{\mathbb{R}^N}V$ then \bel{lubu} \lim_{\alpha\to 0}\lambda(\CK_\alpha,\BBR^N)=\inf_{\BBR^N}V \end{equation} and if, in addition, $\sup_{\mathbb{R}^N}V=\liminf_{|x|\to\infty}V(x)$ then \bel{lubu1} \lim_{\alpha\to +\infty}\lambda(\CK_\alpha,\BBR^N)=\sup_{\mathbb{R}^N}V. \end{equation} \noindent (ii)\, Denote $V_\alpha(x)=\alpha V(x)$ for $\alpha} \def\gb{\beta} \def\gg{\gamma>0$. If $V$ is upper semi-continuous then \bel{upper} \lim_{\alpha\to+\infty}\frac{\lambda(\CK_\alpha,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)}{\alpha}=\inf_{\Omega}V.\end{equation} \noindent (iii)\, Denote $V_\alpha(x)=-\alpha V(x)$ for $\alpha} \def\gb{\beta} \def\gg{\gamma>0$. If $V$ is lower semi-continuous then \bel{lower} \lim_{\alpha\to+\infty}\frac{\lambda(\CK_\alpha,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)}{\alpha}=\sup_{\Omega}V.\end{equation} \end{theorem} An example of a potential $V$ which inspires the study of the limits of $\{\gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)\}$ is given by $$ V(x)=\left\{ \begin{aligned} &-e^{-\frac{|x|^2}{1-|x|^2}}+\frac{1}{2} \quad &&\text{if } |x|<1\\ &\frac{1}{2} &&\text{if } |x|\geq1. \end{aligned} \right. $$ For such $V$, we see that $$\min_{\mathbb{R}^N}V(x)=V(0)=-\frac{1}{2} \quad \text{and} \quad \max_{\mathbb{R}^N}V(x)=\lim_{|x| \to +\infty}V(x)=\frac{1}{2}.$$ If we put $V_\alpha(x)=V(\alpha x)$ then $V_\alpha$ is continuous and increasing with respect to $\alpha$. By Theorem \ref{lim2} (i), for $\alpha$ small enough, $\lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)<0$ and $\lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)>0$ for $\alpha$ large enough. Hence in view of Theorem \ref{JFA} (as $q=0$), there exists a threshold value $\alpha^\star$ such that equation \eqref{nonlinearR} admits a unique positive solution if and only if $\alpha<\alpha^\star$. If we put $V_\alpha(x)=\alpha V(x)$ then by Theorem \ref{lim2} (ii) and (iii), equation \eqref{nonlinearR} admits a unique positive solution as $\alpha$ near $+\infty$ and admits no positive solution as $\alpha$ near $-\infty$. Next we discuss the simplicity of $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. For this purpose, we introduce the notion of \textit{solutions of minimal growth at infinity} in the spirit of \cite[Definition 8.2]{BR3}. \begin{definition} \label{defminimal} Assume $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a unbounded domain and $V \in L_{loc}^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. We say that a positive weak solution $u\in W^{1,p}_{loc}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ of \bel{eqKV} \CK_V[u]=0\qquad \text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi, \end{equation} is a solution of \eqref{eqKV} of minimal growth at infinity if for any $\rho>0$ and any positive function $v\in C^{1}_{loc}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi \setminus B_\rho)$ satisfying $\CK_V[v]\geq0$ in the weak sense in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi \setminus B_\rho$, there exist $\rho'>\rho$ and $k>0$ such that $u\leq k v$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi \setminus B_{\rho'}$. \end{definition} In the sequel, we treat the case $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$. We know that if $u$ is a positive eigenfunction of $\CK_V$ in $\BBR^N$ associated to $\gl(\CK_V,\BBR^N)$ then $u$ is a positive weak solution of \begin{equation}\label{eigen} \CK_V[u]=\lambda(\CK_V,\mathbb{R}^N)u^{p-1}\quad \text{in } \BBR^N. \end{equation} \begin{theorem} \label{simplicity1} Assume $V \in L^\infty(\BBR^N)$ satisfies \bel{condV} \limsup_{|x| \to \infty}|x|^{p-1}|V(x)-\gl(\CK_V,\BBR^N)|<\infty. \end{equation} If there exists a positive eigenfunction $u \in W_{loc}^{1,p}(\BBR^N)$ of $\CK_V$ associated to $\gl(\CK_V,\BBR^N)$ such that \begin{align} &u \text{ is a solution of minimal growth solution of } \eqref{eigen}, \\ &\nabla u(x) \ne 0 \quad \forall x \in \BBR^N, \label{nablau}\\ &\liminf_{|x| \to \infty}\frac{|x||\nabla u(x)|}{u(x)}>0, \label{nablauu} \end{align} then $\gl(\CK_V,\BBR^N)$ is simple, i.e. if $v \in W_{loc}^{1,p}(\BBR^N)$ is a positive eigenfunction of $\CK_V$ in $\BBR^N$ associated with $\gl(\CK_V,\BBR^N)$ then $v=\ell u$ in $\BBR^N$ for some $\ell>0$. \end{theorem} Let us remark that the proof of Theorem \ref{simplicity1} is mainly based on the strong comparison principle \cite[Theorem 3.2]{FrPi} and the scaling technique. Assumption \eqref{nablau} is needed to make use of the strong comparison principle and is imposed in many papers, for instance in \cite[Theorem 3.2]{FrPi}, \cite[Theorem 3.4]{BBV} and in a series of celebrated papers by Pucci and Serrin \cite[Lemma 5]{PuSe1}, \cite[Theorem 10.1]{PuSe2}, \cite[Theorem 4]{PSZ}. Assumption \eqref{condV} means that $V$ is not allowed to be far away from $\gl(\CK_V,\BBR^N)$ as $|x|$ near infinity. This assumption and \eqref{nablauu} are employed in the scaling process in order to treat the case when the graph of $u$ and the graph of $\ell v$ are tangent at infinity where $\ell$ is some positive constant. When $p \geq 2$, we obtain the simplicity of $\gl(\CK_V,\BBR^N)$ without making use of the notion of solution of minimal growth at infinity. Unlike assumption \eqref{condV}, in this case, we impose the condition that $V$ is sufficiently far from $\gl(\CK_V,\BBR^N)$ as $|x|$ near $\infty$. This enables us to employ Proposition \ref{class1} in order to deduce that any eigenfunction associated to $\gl(\CK_V,\BBR^N)$ decays exponentially. Consequently, by adapting the classical method \cite{DKN,DrRa}, we obtain the simplicity. \begin{theorem} \label{simplicity2} Assume $p \geq 2$ and $V\in L^\infty(\mathbb{R}^N)$. There exists $\mu>0$ such that if \begin{equation}\label{decay} \lambda(\CK_V,\mathbb{R}^N)<\liminf_{|x|\to\infty}V(x)-\mu \end{equation} then $\lambda(\CK_V,\mathbb{R}^N)$ is simple. Moreover, the unique (up to multiplicative constants) eigenfunction associated with $\lambda(\CK_V,\mathbb{R}^N)$ decays exponentially. \end{theorem} It is well known that if $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a smooth bounded domain then $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ is isolated (see \cite{GS}), i.e. $\CE(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)=\{ \gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \}$. This property no longer holds if $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is unbounded. This is reflected in the following result. \begin{theorem} \label{spectrum} Assume $V \in L^\infty(\BBR^N)$. Then \bel{spec} \CE(\BBR^N)=(-\infty,\gl(\CK_V,\BBR^N)]. \end{equation} \end{theorem} It is noteworthy that Theorem \ref{spectrum} still holds true if $\BBR^N$ is replaced by a smooth unbounded domain. It can be obtained by using an analogue argument as in the proof of Theorem \ref{spectrum}. However, we state the result for the case $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$ in order to simplify the proof and to streamline the exposition. We end the section by providing a criterion in terms of $\gl''(\CK_V,\BBR^N)$ to characterize the maximum principle in $\mathbb{R}^N$. \begin{theorem} \label{MP} (Weak maximum principle) Assume $\gl''(\CK_V,\mathbb{R}^N) >0$. If a function $u\in C^1_{loc}(\mathbb{R}^N)$ satisfies \begin{equation}\nonumber \left\{ \begin{aligned} & \CK_V[u] \leq 0 \quad \text{in the weak sense in } \BBR^N, \\ & \sup_{\BBR^N}u<\infty, \quad \limsup_{|x| \to \infty}u(x) \leq 0, \\ & |\nabla u(x)| \neq 0,\quad\forall x \in \BBR^N, \end{aligned} \right. \end{equation} then $u \leq 0$ in $\BBR^N$. \end{theorem} As a consequence of Theorem \ref{equivalence} (iii) and Theorem \ref{MP}, we obtain \begin{corollary} Assume $0 \not \equiv V \in L^\infty(\BBR^N)$ is radially symmetric. Then the weak maximum principle holds if $\gl(\CK_V,\mathbb{R}^N) >0$. \end{corollary} The paper is organized as follows. In Section 2, we prove Theorem \ref{equivalence}. The proof of Theorem \ref{lim1} and Theorem \ref{lim2} are presented in Section 3. Finally, in Section 4, we demonstrate Theorems \ref{simplicity1}, \ref{simplicity2}, \ref{spectrum} and \ref{MP}. \medskip \noindent \textbf{Notation.} Throughout the paper, $B_r$ denotes the ball of center $0$ and radius $r>0$. Unless otherwise stated, we assume that $p>1$ and $V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. \section{Equivalence of the generalized principal eigenvalues} This section is devoted to the study of the relation between the notions $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$, $\gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ and $\gl''(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. Let us first recall the following result, which is proved in \cite{NV} \begin{theorem}\label{main0} (1)\, Assume $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a $C^{1,\nu}$ ($\nu\in(0,1)$) bounded domain in $\BBR^N$ and $V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. Then $$\gl(\CK_V,\Omega)=\gl_V^\Omega} \def\Gx{\Xi} \def\Gy{\Psi(\Omega} \def\Gx{\Xi} \def\Gy{\Psi).$$ \noindent (2)\, Assume $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a general domain in $\BBR^N$ (possibly unbounded) and $\{\Omega} \def\Gx{\Xi} \def\Gy{\Psi_n\}$ is a smooth exhaustion of $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$. Let $V \in L^\infty_{\text{\rm loc}}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ such that $\inf_{\Omega}V>-\infty$. Then the following properties hold. \medskip (i)\, $\inf_{\Omega}V \leq \gl(\CK_V,\Omega_{n+1})< \gl(\CK_V,\Omega_{n})$ for every $n \in \BBN$. \medskip (ii)\, $\gl(\CK_V,\Omega)=\lim_{n\to\infty}\gl(\CK_V,\Omega_n)$ and there exists a positive weak solution $\vgf \in C^1_{\text{\rm loc}}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ of $$ \CK_V[\vgf]=\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)\vgf^{p-1}\quad\quad\textrm{in $\Omega$.} $$ (iii)\, \bel{lamV} \gl(\CK_V,\Omega)=\inf_{\phi \in C_c^{1}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \setminus\{0\}}\frac{\int_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}(|\nabla \phi} \def\vgf{\varphi} \def\gh{\eta|^p + V|\phi} \def\vgf{\varphi} \def\gh{\eta|^p)dx}{\int_{\Omega}|\phi|^pdx}. \end{equation} \end{theorem} \begin{lemma}\label{la2} Let $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ be a smooth bounded domain or $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$. Assume $0 \not \equiv V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. Then there holds \bel{lub1} \gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \geq \gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi). \end{equation} \end{lemma} \note{Proof} In order to prove \eqref{lub1}, we need to show that $\gl \geq \gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ for any $\gl > \gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. Without lost of generality we assume that $\gl=0$. Since $\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)<0$, by Theorem \ref{main0}, (2.ii), there is a smooth bounded domain $G \Subset \Omega} \def\Gx{\Xi} \def\Gy{\Psi$ such that $\gl(\CK_V,G)<0$ and $V \not \equiv 0$ in $G$. Let $\vgf_V^{G}$ be the first eigenfunction associated to $\gl(\CK_V,G)<0$, normalized by \bel{norm} \max_{G}\vgf_V^G = \min\left \{ 1,-\frac{\gl(\CK_V,G)}{\sup_{G}{|V|}} \right \}. \end{equation} In particular, the function $\vgf_V^G$ satisfies \bel{eigenG} \left\{ \begin{aligned} \CK_V [\vgf_V^G] &= \gl(\CK_V,G) (\vgf_V^G)^{p-1} \qquad &&\text{in } G \\ \vgf_V^G &= 0 &&\text{on } \partial G. \end{aligned} \right. \end{equation} \medskip \noindent \textit{Case 1: $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a smooth bounded domain.} Define $\overline \vgf: = 1$ and $$ \underline \vgf: = \left\{ \begin{aligned} &\vgf_V^G \quad &&\text{in } G, \\ &0 &&\text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi \setminus G. \end{aligned} \right. $$ It can be verified that $\overline \vgf$ and $\underline \vgf$ are respectively weak supersolution and weak subsolution of \bel{lpe} \left\{ \begin{aligned} \CK_V [u] + V^+ u^p &= 0 \qquad &&\text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi \\ u &= 0 &&\text{on } \partial \Omega} \def\Gx{\Xi} \def\Gy{\Psi, \end{aligned} \right. \end{equation} where $V^+=\max\{V,0\}$. By \cite[Theorem 3.1]{LS}, one can find a solution $u$ of \eqref{lpe} such that $\underline \vgf \leq u \leq \overline \vgf$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$. Let $x_0 \in G$ such that $\underline \vgf(x_0)=\max_G \underline \vgf >0$. Then $u(x_0) \geq \underline \vgf(x_0) >0$. By Harnack inequality \cite{Se,Tr}, we deduce that $u>0$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$. \medskip \noindent \textit{Case 2: $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$.} In this case, we can take $G=B_{R_0}$ for some $R_0>0$ large and $\gl(\CK_V,B_{R_0})<0$. Denote by $\vgf_V^{B_{R_0}}$ the eigenfunction associated to $\gl(\CK_V,B_{R_0})$. Hence \eqref{norm} and \eqref{eigenG} hold with $G$ replaced by $B_{R_0}$. Fix $m>0$ and set $$a_{R_0,m}(x):=\left\{\begin{aligned} &-V(x) \quad &&x \in B_{R_0}\\ &-\max\{V(x), m\}&& x \in B_{R_0}^c. \end{aligned} \right.$$ It is easy to see that $$ \limsup_{|x| \to \infty}a_{R_0,m}<-m $$ and that there exists $s_0 > 1$ independent of $R_0$ and $m$ such that $$ -a_{R_0,m}s_0^{p-1} + V^+s_0^p \geq 0.$$ Consider the equation \bel{lomo} -\Delta_p u - a_{R_0,m} u^{p-1} + V^+ u^{p} = 0 \quad\quad \text{in }\mathbb{R}^N. \end{equation} By \cite[Proposition 4.1]{NV}, we deduce that the following function $$ \overline v(x):=s_0 \chi_{B_{R_0}}(x) + Ce^{-\theta} \def\vge{\varepsilon|x|}\chi_{B_{R_0}^c}(x),\quad x \in \BBR^N,$$ with $\theta} \def\vge{\varepsilon=\theta} \def\vge{\varepsilon(p,m)$ and $C=C(s_0,m,p)$, is a decaying supersolution of \eqref{lomo}. For $R>R_0$, define $$ \underline u_R: = \left\{ \begin{aligned} &\vgf_V^{B_{R_0}} \quad &&\text{in } B_{R_0}, \\ &0 &&\text{in } B_R \setminus B_{R_0}. \end{aligned} \right. $$ We can see that $\underline u_R$ is a subsolution of \eqref{lomo}. Indeed, in $B_{R_0}$ \bel{lomosub} \begin{aligned} -\Delta_p \underline u_R - a_{R_0,m} \underline u_R^{p-1} + V^+ \underline u_R^p &= -\Delta_p \vgf_V^{B_{R_0}} + V (\vgf_V^{B_{R_0}})^{p-1} + V^+ (\vgf_V^{B_{R_0}})^p \\ &= \gl_V^{B_{R_0}} (\vgf_V^{B_{R_0}})^{p-1} + V^+(\vgf_V^{B_{R_0}})^p \\ &\leq (\vgf_V^{B_{R_0}})^{p-1}(\gl_V^{B_{R_0}} + V^+\vgf_V^{B_{R_0}}) \\ &\leq 0. \end{aligned} \end{equation} Here the last inequality follows from the normalization of $\vgf_V^{B_{R_0}}$. It is easy to see that $\underline u_R$ and $\overline v$ are respectively sub and supersolution of \bel{lpR} \left\{ \begin{aligned} -\Gd_p u -a_{R_0,m}u^{p-1} + V_+ u^p &= 0 \qquad &&\text{in } B_R \\ u &= 0 &&\text{on } \partial B_R \end{aligned} \right. \end{equation} such that $\underline u_R \leq \overline v$ in $B_R$. Therefore, by sub-supersolutions theorem, we deduce the existence of a solution $u_R$ of \eqref{lpR} in $B_R$ such that $\underline u_R \leq u_R \leq \overline v$ in $B_R$. By regularity result for quasilinear elliptic equations, up to a subsequence, $\{ u_R \}$ converges in $C_{\text{loc}}^1(\BBR^N)$, as $R \to \infty$, to a weak solution $u^*$ of \eqref{lomo} in $\BBR^N$. Moreover, $u^*(x_0) \geq \vgf_V^{B_{R_0}}(x_0)>0$ and $0 \leq u^* \leq \overline v$ a.e. in $\BBR^N$. By Harnack inequality we infer that $u^*>0$ in $\BBR^N$. Since $\overline v$ decays exponentially, so does $u^*$. By local regularity for quasilinear elliptic equations (see \cite{Tol} and \cite[Theorem 3.1]{La}) and Harnack inequality \cite{Tr}, there exists constants $c_1,c_2$ depending on $s_0$, $N$, $p$, $\norm{V}_{L^\infty(\BBR^N)}$, $m$ such that for every $x \in \BBR^N$ $$ \sup_{B_1(x)}|\nabla u^*| \leq c_1 \sup_{B_2(x)} u^* \leq c_2 \inf_{B_2(x)} u^*. $$ It follows that $u^* \in W_0^{1,p}(\BBR^N)$. Since $a_{R_0,m} \leq -V$, we deduce that $$ \CK_V[u^*] = (V+a_{R_0,m})(u^*)^{p-1} - V^+ (u^*)^p \leq 0.$$ By choosing $\psi=u^*$ in the definition of $\gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ we deduce that $\gl'(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \leq 0$. This completes the proof. \\${}$ \hfill $\square$ \begin{lemma} \label{la3} Assume $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a general domain in $\BBR^N$ and $V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. Then there holds \bel{lamlam} \lambda(\CK_V,\Omega)\leq \lambda'(\CK_V,\Omega). \end{equation} \end{lemma} \note{Proof} Take $\lambda\in\mathbb{R}$ such that there exists an admissible test function $\psi\in W^{1,p}_{0}(\Omega)$ satisfying $\psi>0$ in $\Omega$ and $$ \CK_V[\psi]\leq\lambda\psi^{p-1} \quad \text{in the weak sense in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi.$$ we will prove that $\lambda(\CK_V,\Omega)\leq\lambda$. Since $\psi\in W_0^{1,p}(\Omega)$ and $C_c^1(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ is dense in $W_0^{1,p}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$, there exists a sequence $\{\psi_n\} \subset C_c^1(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ converging to $\psi$ in $W^{1,p}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. Since $\psi>0$, we infer that $\psi_n>0$ for $n$ large enough. By \eqref{lamV}, we have $$ \lambda(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)\leq\frac{\int_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}(|\nabla \psi_n|^{p}+V \psi_n^p)dx}{\int_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}\psi_n^pdx}. $$ Letting $n\to \infty$ yields $$ \lambda(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)\leq\frac{\int_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}(|\nabla \psi|^{p}+V \psi^p)dx}{\int_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}\psi^pdx}= \lambda. $$ This completes the proof. \\${}$ \hfill $\square$ \begin{proposition} \label{equal1} Let $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ be a smooth bounded domain or $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$. Assume $0 \not \equiv V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. There holds \bel{lambda01}\lambda(\CK_V,\Omega)=\lambda'(\CK_V,\Omega) \end{equation} \end{proposition} \note{Proof} Equality \eqref{lambda01} follows directly from Lemma \ref{la2} and Lemma \ref{la3}. \\${}$ \hfill $\square$ \begin{lemma} \label{cutoff} Let $\gl \in \BBR$ and $V \in L_{loc}^\infty(\BBR^N)$. If $0 \leq u \in C_{loc}^1(\BBR^N)$ satisfies $\CK_V u \leq \gl u^{p-1}$ in the weak sense in $\BBR^N$ then \bel{estc} (\gl(\CK_V,\BBR^N)-\gl)\int_{\BBR^N} u^p \psi^p dx \leq C(p)\int_{\BBR^N} u^p |\nabla \psi|^p dx \quad \forall \psi \in C_c^1(\BBR^N), \, \psi \geq 0.\end{equation} \end{lemma} \begin{proof}From the assumption, we have \bel{testfor} \int_{\BBR^N}(|\nabla u|^{p-2}\nabla u \nabla \phi} \def\vgf{\varphi} \def\gh{\eta + V u^p \phi} \def\vgf{\varphi} \def\gh{\eta)dx \leq \gl \int_{\BBR^N}u^{p-1}\phi} \def\vgf{\varphi} \def\gh{\eta \,dx \quad \forall \phi} \def\vgf{\varphi} \def\gh{\eta \in C_c^1(\BBR^N). \end{equation} By choosing $\phi} \def\vgf{\varphi} \def\gh{\eta=u\psi^p$ as a test function in \eqref{testfor}, we get $$ \int_{\BBR^N}(|\nabla u|^p \psi^p + pu\psi^{p-1}|\nabla u|^{p-2}\nabla u \nabla \psi + V u^p \psi^p)dx \leq \int_{\BBR^N}Vu^{p}\psi^p \,dx. $$ It follows that $$ \int_{\BBR^N}(|\nabla u|^p \psi^p + V u^p \psi^p)dx \leq p\int_{\BBR^N}u\psi^{p-1}|\nabla u|^{p-1}|\nabla \psi|\, dx + \gl \int_{\BBR^N}u^{p}\psi^p \,dx. $$ By Young's inequality, we deduce that $$ \int_{\BBR^N}(|\nabla u|^p \psi^p + V u^p \psi^p)dx \leq C(p)\Big(\int_{\BBR^N}u^p |\nabla \psi|^p\, dx + \gl \int_{\BBR^N}u^{p}\psi^p \,dx\Big). $$ This, together with the following estimate $$ |\nabla(u\psi)|^p dx \leq C(p)\Big( |\nabla u|^p \psi^p + u^p |\nabla \psi|^p \Big) $$ leads to $$ \int_{\BBR^N}(|\nabla (u\psi)|^p + V (u\psi)^p)dx \leq C(p)\Big(\int_{\BBR^N}u^p |\nabla \psi|^p\, dx + \gl \int_{\BBR^N}(u\psi)^p \,dx\Big). $$ This, joint with Theorem \ref{main0}, implies \eqref{estc}. \end{proof} \begin{proposition} \label{equal2} (i) If $\Omega$ is a smooth bounded domain and $V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ then \bel{equali1} \lambda''(\CK_V,\Omega)=\lambda(\CK_V,\Omega).\end{equation} \noindent (ii) Assume $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ is a smooth bounded domain or $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$. If $0 \not \equiv V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ then \bel{equali2} \lambda''(\CK_V,\Omega)\leq \lambda'(\CK_V,\Omega) .\end{equation} \noindent (iii) If $\Omega} \def\Gx{\Xi} \def\Gy{\Psi=\BBR^N$ and $V \in L_{loc}^\infty(\BBR^N)$ is radially symmetric then \bel{equali3} \gl(\CK_V,\BBR^N)=\gl''(\CK_V,\BBR^N). \end{equation} \end{proposition} \note{Proof} i) Obviously, $\lambda''(\CK_V,\Omega)\leq\lambda(\CK_V,\Omega).$ Let us prove that $\lambda''(\CK_V,\Omega)\geq\lambda(\CK_V,\Omega).$ It suffice to show that $$\lambda \leq \lambda''(\CK_V,\Omega)$$ for any $\lambda<\lambda(\CK_V,\Omega)$. Let $\CO$ be a smooth bounded neighborhood of $\partial\Omega$ then $\Omega} \def\Gx{\Xi} \def\Gy{\Psi \cup \CO$ is a smooth bounded domain. Consider an extension of $V$ to $\Omega} \def\Gx{\Xi} \def\Gy{\Psi \cup \CO$, still denoted by $V$, such that $V \in L^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi \cup \CO)$. Let $\{\Omega} \def\Gx{\Xi} \def\Gy{\Psi_n\}$ be a decreasing sequence of smooth bounded domains such that $$\overline{\Omega} \def\Gx{\Xi} \def\Gy{\Psi_1\setminus\Omega} \def\Gx{\Xi} \def\Gy{\Psi}\subset \CO,\quad \ol\Omega} \def\Gx{\Xi} \def\Gy{\Psi\subset\Omega} \def\Gx{\Xi} \def\Gy{\Psi_n \quad \text{for every }n,\quad\quad\bigcap_n\ol{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}_n=\ol{\Omega} \def\Gx{\Xi} \def\Gy{\Psi},$$ and $\Omega} \def\Gx{\Xi} \def\Gy{\Psi_n$ has the same smoothness as $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ for every $n$. Let $(\lambda_n,\phi_n)$ be the principal eigenvalue and the corresponding eigenfunction of $\CK_V$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi_n$ with normalization $\phi_n(x_0)=1$ for some fixed reference point $x_0 \in \Omega} \def\Gx{\Xi} \def\Gy{\Psi$. We claim that $$\lim_{n\to\infty}\lambda_n=\lambda(\CK_V,\Omega).$$ Assume for the moment that the claim holds true. Since $\gl<\gl(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$, there exists $N_0>0$ large enough such that $\gl_n >\gl$ for every $n>N_0$. Since $\overline\Omega\subset\Omega_n$, it follows that $\inf_{\overline\Omega} \def\Gx{\Xi} \def\Gy{\Psi} \phi_n>0$. Choosing $\psi=\phi_n|_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}$ in the definition of $\lambda''(\CK_V,\Omega)$ in Definition \ref{gpe}, we obtain that $\lambda_n\leq\lambda''(\CK_V,\Omega)$ and thus $\lambda \leq\lambda''(\CK_V,\Omega)$. It remains to prove the claim. By Theorem \ref{main0} (2i), we deduce that $\{ \gl_n \}$ is strictly decreasing and therefore there exists $\gl^*=\lim_{n \to \infty}\gl_n$. By the local regularity of weak solutions of degenerate elliptic equations \cite{Di}, up to a subsequence, $\phi_n$ converges in $C^1_{loc}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ to a function $\phi^*$ which is a weak solution of $$ \CK_V [\phi^*]= \lambda^* (\phi^*)^{p-1} \quad \text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi,$$ Since $\Omega} \def\Gx{\Xi} \def\Gy{\Psi_n$ has the same smoothness as $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$, by the $C^1$ regularity up to the boundary \cite{Lieberman}, we obtain $\phi^*=0$ on $\partial\Omega$. Note that $\phi^*(x_0)=1$, by the Harnack inequality \cite{Tr}, we get $\phi^*>0$ in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$ and hence it is a Dirichlet principal eigenfunction of $\CK_V$ in $\Omega$. This implies $\lambda^*=\lambda(\CK_V,\Omega)$. \medskip \noindent (ii) By Proposition \ref{equal1}, $\lambda'(\CK_V,\Omega)=\lambda(\CK_V,\Omega)$. This, together with the fact that $\lambda''(\CK_V,\Omega) \leq\lambda(\CK_V,\Omega)$, implies inequality \eqref{equali2}. \medskip \noindent (iii) It is easy to see that $\gl''(\CK_V,\BBR^N) \leq \gl(\CK_V,\BBR^N)$. It remains to show that $\gl''(\CK_V,\BBR^N) \geq \gl(\CK_V,\BBR^N)$. To this end, take arbitrary $\gl < \gl(\CK_V,\BBR^N)$ and we will demonstrate that $\gl \leq \gl''(\CK_V,\BBR^N)$. Without loss of the generality, we assume that $\gl=0$. This follows that $\gl(\CK_V,\BBR^N)>0$. For any $n>0$, put $\gl_n=\gl(\CK_V,B_n)$. By Theorem \ref{main0}, $\gl_n \downarrow \gl(\CK_V,\BBR^N)$ as $n \to \infty$, hence $\gl_n \geq \gl(\CK_V,\BBR^N)>0$. For any $n \in \BBN$, let $f_n \in C^\infty(B_n)$ be nonnegative, radially symmetric and not identically equal to zero in $B_n$ with $\text{\rm supp} f_n \in B_n \setminus B_{n-1}$. By \cite[Theorem 2 (v)]{GS}, there exists a unique nonnegative weak solution $w_n \in W_0^{1,p}(B_n)$ of $$ \left\{ \begin{aligned} \CK_V[w_n] &= f_n \quad &&\text{in } B_n \\ w_n &= 0 &&\text{on } \partial B_n. \end{aligned} \right. $$ Since $V$ and $f_n$ are both radially symmetric, the uniqueness implies that $w_n$ is radially symmetric too. Moreover, by the strong maximum principle \cite[Theorem 2 (ii)]{GS}, we obtain that $w_n$ is positive in $B_n$. Put $$ \vgf_n(x):=\frac{w_n(x)}{w_n(0)} $$ then $\vgf_n$ is radially symmetric and $\vgf_n(0)=1$. By the Harnack inequality \cite{Tr} and regularity results for quasilinear elliptic equations \cite{Di}, up to a subsequence, the sequence $\{\vgf_n\}$ converges in $C_{loc}^1(\BBR^N)$ to a function $\vgf$ which is a nonnegative, radially symmetric weak solution of $\CK_{V}[\vgf]=0$ in $\BBR^N$. Since $\vgf(0)=1$, in light of the Harnack inequality, we deduce that $\vgf$ is positive in $\BBR^N$. Next, for $n \in \BBN$, let $\psi_n \in C^\infty(\BBR^N)$ such that $0 \leq \psi_n \leq 1$, $\psi_n=1$ in $B_{n-1}$, $\psi_n=0$ in $B_n$ and $|\nabla \psi_n| \leq C$ where $C$ is a constant independent of $n$. Applying Lemma \ref{cutoff}, we get $$ \gl(\CK_V,\BBR^N)\int_{B_{n-1}}\vgf^p dx \leq c\int_{B_n \setminus B_{n-1}}\vgf^p dx \quad \forall n \in \BBN $$ where $c$ is independent of $n$. Thus there exist constants $C>0$ and $a>1$ such that $$ \int_{B_n \setminus B_{n-1}}\vgf^p dx \geq Ca^n \quad \forall n \in \BBN. $$ Therefore, for each $n$ one can find $x_n \in B_n \setminus B_{n-1}$ such that $\vgf(x_n) \geq Ca^{\frac{n}{p}}$ with another constant $C$ independent of $n$. Since $\vgf$ is radially symmetric, by employing Harnack inequality, we derive that $\vgf$ has exponential growth. In particular, $\inf_{\BBR^N}\vgf>0$. By choosing $\psi=\vgf$ in the definition of $ \gl''(\CK_V,\BBR^N)$ in Definition \ref{gpe}, we derive that $\gl''(\CK_V,\BBR^N) \geq 0$. Thus $\gl(\CK_V,\BBR^N) \leq \gl''(\CK_V,\BBR^N)$. This completes the proof. \\${}$ \hfill $\square$ \medskip \noindent \textbf{Proof of Theorem \ref{equivalence}.} Statements (i)-(iii) follow directly from Proposition \ref{equal1} and Proposition \eqref{equal2}. \\${}$ \hfill $\square$ \section{Qualitative properties of the principal eigenvalue} \noindent \textbf{Proof of Theorem \ref{lim1}.} \noindent (i)\; The fact that $\lambda(\CL_\alpha,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ is concave and nondecreasing is directly deduced by its variational characterization $$\lambda(\CL_\alpha,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)=\inf_{\phi \in C_c^{1}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \setminus\{0\}}\frac{\int_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi}(\alpha|\nabla \phi} \def\vgf{\varphi} \def\gh{\eta|^p + V|\phi} \def\vgf{\varphi} \def\gh{\eta|^p)dx}{\int_{\Omega}|\phi|^pdx}.$$ We see that $$\mathcal{L}_\alpha[\phi]=\alpha} \def\gb{\beta} \def\gg{\gamma\left(-\Delta_p\phi+\frac{1}{\alpha}V(x)\phi^{p-1}\right) = \alpha} \def\gb{\beta} \def\gg{\gamma \CK_{\frac{1}{\alpha} \def\gb{\beta} \def\gg{\gamma}V}[\phi} \def\vgf{\varphi} \def\gh{\eta],$$ it follows that $$\gl(\CL_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)=\alpha} \def\gb{\beta} \def\gg{\gamma \gl(\CK_{\frac{1}{\alpha} \def\gb{\beta} \def\gg{\gamma}V},\BBR^N).$$ By Theorem \ref{lim2} (ii), one has $$\lim_{\alpha\to0}\alpha} \def\gb{\beta} \def\gg{\gamma \gl(\CK_{\frac{1}{\alpha} \def\gb{\beta} \def\gg{\gamma}V},\BBR^N)=\inf_{\Omega}V.$$ Therefore $\lim_{\alpha} \def\gb{\beta} \def\gg{\gamma \to 0}\gl(\CL_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)=\inf_{\Omega}V$. \medskip \noindent (ii)\; Let $\phi_\alpha>0$ be the eigenfunction associated with $\lambda(\CL_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ with normalization $\|\phi_\alpha\|_{L^p(\Omega)}=1$. We have \bel{eigena} \left\{ \begin{aligned} \CL_\alpha [\phi_\alpha] &= \lambda(\CL_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)\phi_\alpha^{p-1} \qquad &&\text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi \\ \phi_\alpha &= 0 &&\text{on } \partial \Omega} \def\Gx{\Xi} \def\Gy{\Psi. \end{aligned} \right. \end{equation} By the variational characterization of $\lambda(\CL_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$, we have $$\lambda(\CL_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)=\alpha\int_{\Omega}|\nabla\phi_\alpha|^pdx +\int_{\Omega}V\phi_\alpha^pdx \geq \alpha\int_{\Omega}|\nabla\phi_\alpha|^pdx+\inf_{\Omega}V. $$ By Poincar\'e inequality, there exists a constant $C=C(N,p,\Omega)>0$ such that $$\int_{\Omega}|\nabla\phi_\alpha|^pdx \geq C \int_{\Omega}|\phi_\alpha|^pdx.$$ Combining the above estimates yields $$\lambda(\CL_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)\geq C\alpha+\inf_{\Omega}V, $$ which implies statement (ii). \medskip \noindent (iii)\; We first prove \eqref{limsup}. It suffices to show that $\lambda(\CL_\alpha,\BBR^N) \leq \eta$ for any $\eta$ satisfying \begin{equation} \eta >\limsup_{|x|\to\infty}V(x).\label{11.1.2} \end{equation} Take $\gl$ satisfying \eqref{11.1.2}, there exists $R$ large enough such that $$\inf_{|x|\geq R}(\eta-V(x))>0,$$ One sees that $$\lambda(-\alpha\Delta_p -(\eta-V),B_R^c(0))<0.$$ Indeed, by the variational formula $$ \begin{aligned} \lambda(-\alpha\Delta_p -(\eta-V(x)),B_R^c(0))&=\inf_{\su{\phi\in C_c^{1}(B_R^c(0))}{\phi\neq 0}}\frac{\alpha\int_{B_R^c(0)}|\nabla\phi|^pdx -\int_{B_R^c(0)}(\eta-V(x))|\phi|^pdx}{\int_{B_R^c(0)}|\phi|^pdx}\nonumber\\ &\leq\inf_{\su{\phi\in C_c^{1}(B_R^c(0))}{\phi\neq 0}}\frac{\alpha\int_{B_R^c(0)}|\nabla\phi|^pdx}{\int_{B_R^c(0)}|\phi|^pdx}-\inf_{B_R^c(0)}(\eta-V(x))\nonumber\\ &=-\inf_{B_R^c(0)}(\eta-V(x))<0,\nonumber \end{aligned} $$ where we use the fact that \begin{equation} \inf_{\su{\phi\in C_c^{1}(B_R^c(0))}{\phi\neq 0}}\frac{\int_{B_R^c(0)}|\nabla\phi|^pdx}{\int_{B_R^c(0)}|\phi|^pdx}=0.\label{16.3.1} \end{equation} Let us prove (\ref{16.3.1}). For $r>R$, it is easily seen that \begin{equation} \inf_{\su{\phi\in C_c^{1}(B_R^c(0))}{\phi\neq 0}}\frac{\int_{B_R^c(0)}|\nabla\phi|^pdx}{\int_{B_R^c(0)}|\phi|^pdx}=\lim_{r\to\infty}\lambda_{r},\label{16.3.2} \end{equation} where $$\lambda_{r}=\inf_{\su{\phi\in C_c^{1}(B_{r}(0)\setminus B_R(0))}{\phi\neq 0}}\frac{\int_{B_{r}(0)\setminus B_R(0)}|\nabla\phi|^pdx}{\int_{B_{r}(0)\setminus B_R(0)}|\phi|^pdx}.$$ Take $x_0\in B_{r}(0)\setminus B_R(0)$ such that $|x_0|=\frac{r+R}{2}$. It is obvious that $$B_{\frac{r-R}{4}}(x_0)\subset B_{r}(0)\setminus B_R(0),$$ and hence \begin{equation} 0\leq\lambda_{r}\leq \inf_{\su{\phi\in C_c^{1}(B_{\frac{r-R}{4}}(x_0))}{\phi\neq 0}}\frac{\int_{B_{\frac{r-R}{4}}(x_0))}|\nabla\phi|^pdx}{\int_{B_{\frac{r-R}{4}}(x_0))}|\phi|^pdx}.\label{16.3.3} \end{equation} By \cite[Theorem 4.3]{Mao}, $$\inf_{\su{\phi\in C_c^{1}(B_{\frac{r-R}{4}}(x_0))}{\phi\neq 0}}\frac{\int_{B_{\frac{r-R}{4}}(x_0))}|\nabla\phi|^pdx}{\int_{B_{\frac{r-R}{4}}(x_0))}|\phi|^pdx}\leq C(N,p,r,R),$$ where $C(N,p,r,R)$ is independent of $x_0$ and \begin{equation} C(N,p,r,R):=\left\lbrace\begin{aligned} &\frac{p^{\frac{4^pp^2-p}{p-N}}}{n^{\frac{np-p}{p-N}}(p-1)^{p-1}(r-R)^p}\quad &&\text{if } N\neq p\\ &\frac{4^N N^Ne^{N-1}}{(N-1)^{N-1}(r-R)^N} &&\text{if } N=p. \end{aligned}\right. \end{equation} By letting $r\to\infty$ in (\ref{16.3.3}), one gets $$\lim_{r\to\infty}\lambda_{r}=0.$$ By Theorem \ref{main0} (ii), we have $$\lambda(-\alpha\Delta_p +(V-\eta),\mathbb{R}^N)\leq\lambda(-\alpha\Delta_p -(\eta-V),B_R^c(x_0))<0.$$ This implies that $\lambda(\CL_\alpha,\BBR^N)< \eta$ and hence we get \eqref{limsup}. \medskip We next prove \eqref{liminf}. Put $$ \begin{aligned} \psi(r)&:=(e^{\gb r +1} + e^{-\gb r-1})^{-\gg} \quad r \geq 0, \\ \tilde \psi(x)&:= \psi(|x|) \quad x \in \BBR^N, \end{aligned} $$ where $\gb>0$ and $\gg>0$ will be made precise later. Then we have \bel{formulation} \Gd_p \tilde \psi(x) = (|\psi_r|^{p-2}\psi_r)_r + \frac{N-1}{r}|\psi_r|^{p-2}\psi_r \quad r=|x|>0 \end{equation} where $\psi_r$ denotes the first derivative of $\psi$. We next compute the left hand-side in \eqref{formulation}. It is easy to see that $$ \begin{aligned} \psi_r(r)&=-\gb\gg(e^{\gb r+1}+e^{-\gb r-1})^{-\gg-1}(e^{\gb r+1}-e^{-\gb r-1}), \\ |\psi_r|^{p-2}\psi_r &= - \gb^{p-1}\gg^{p-1}(e^{\gb r+1}+e^{-\gb r-1})^{-(\gg+1)(p-1)}(e^{\gb r+1}-e^{-\gb r-1})^{p-1}, \\ (|\psi_r|^{p-2}\psi_r)_r &= (p-1)\gb^p \gg^{p-1} g(r)^{p-2}[(\gg+1)g(r)^2-1]\psi^{p-1} \end{aligned} $$ where $$ g(r)=\frac{e^{2(\gb r+1)}-1}{e^{2 (\gb r+1)} +1}. $$ Since $\psi_r(r)<0$ for every $r>0$, it follows that $\Gd_p \tilde \psi \leq (|\psi_r|^{p-2}\psi_r)_r$, namely $$ \Gd_p \tilde \psi \leq (p-1)\gb^p \gg^{p-1} g(r)^{p-2}[(\gg+1)g(r)^2-1]\psi^{p-1}, \quad r=|x|. $$ For $\gb>0$, since $g$ is increasing with respect to $r \in (0,\infty)$, it follows that $$ \frac{e^2-1}{e^2+1}=g(0) < g(r) < \lim_{r \to \infty}g(r)=1. $$ Therefore \bel{form2} \Gd_p \tilde \psi \leq c_p(p-1)\gb^p \gg^{p-1}[(\gg+1)g(r)^2-1]\psi^{p-1}, \quad c_p= \Big( \frac{e^2-1}{e^2+1} \Big)^{p-2}+1. \end{equation} Take arbitrarily $\vge>0$. Then there exists $R>0$ large enough such that $$ V(x) \geq \liminf_{|x| \to \infty}V - \vge \quad \forall x \in B_R^c.$$ It follows that $$ -\alpha} \def\gb{\beta} \def\gg{\gamma \Gd_p \tilde \psi + V \tilde \psi^{p-1} \geq \Big[-c_p(p-1)\alpha} \def\gb{\beta} \def\gg{\gamma\gb^p \gg^{p} + (\liminf_{|x| \to \infty}V-\vge)\Big] \tilde \psi^{p-1} \quad \text{in } B_R^c. $$ Therefore, if we can choose $\alpha} \def\gb{\beta} \def\gg{\gamma$, $\gb$ and $\gg$ such that \bel{cd1} \alpha} \def\gb{\beta} \def\gg{\gamma\gb^p \gg^{p} \to 0 \quad \text{as } \alpha} \def\gb{\beta} \def\gg{\gamma \to \infty, \end{equation} then there exists $\alpha} \def\gb{\beta} \def\gg{\gamma_1$ large enough such that for any $\alpha} \def\gb{\beta} \def\gg{\gamma \geq \alpha} \def\gb{\beta} \def\gg{\gamma_1$, \bel{BRc} -\alpha} \def\gb{\beta} \def\gg{\gamma \Gd_p \tilde \psi + V \tilde \psi^{p-1} \geq (\liminf_{|x| \to \infty}V-2\vge) \tilde \psi^{p-1} \quad \text{in } B_R^c. \end{equation} We next choose $\gb>0$ small enough such that $g(r)^2<\frac{2}{3}$ for every $r \in (0,R)$. Then by \eqref{form2}, $$ \Gd_p \tilde \psi(x) \leq \frac{1}{3}c_p(p-1)\gb^p \gg^{p-1} (2\gg-1)\psi(|x|)^{p-1}, \quad x \in B_R.$$ This implies $$ -\alpha} \def\gb{\beta} \def\gg{\gamma\Gd_p \tilde \psi + V \tilde \psi^{p-1} \geq \Big[ -\frac{1}{3}c_p(p-1)\alpha} \def\gb{\beta} \def\gg{\gamma \gb^p \gg^{p-1} (2\gg-1) + V \Big]\tilde \psi^{p-1} \quad \text{in } B_R. $$ If we can choose $\alpha} \def\gb{\beta} \def\gg{\gamma$, $\gb$ and $\gg$ such that \bel{cd2} \alpha} \def\gb{\beta} \def\gg{\gamma \gb^p \gg^{p-1} \to +\infty \quad \text{as } \alpha} \def\gb{\beta} \def\gg{\gamma \to \infty, \end{equation} then there exists $\alpha} \def\gb{\beta} \def\gg{\gamma_2>0$ large enough such that for any $\alpha} \def\gb{\beta} \def\gg{\gamma \geq \alpha} \def\gb{\beta} \def\gg{\gamma_2$, \bel{BR} -\alpha} \def\gb{\beta} \def\gg{\gamma \Gd_p \tilde \psi + V \tilde \psi^{p-1} \geq (\liminf_{|x|\to \infty}V(x) - 2\vge)\tilde \psi^{p-1} \quad \text{in } B_R. \end{equation} By combining \eqref{BRc} and \eqref{BR} we deduce that \bel{RN} -\alpha} \def\gb{\beta} \def\gg{\gamma \Gd_p \tilde \psi + V \tilde \psi^{p-1} \geq (\liminf_{|x|\to \infty}V(x) - 2\vge)\tilde \psi^{p-1} \quad \text{in } \BBR^N \end{equation} provided that $\alpha} \def\gb{\beta} \def\gg{\gamma \geq \max\{\alpha} \def\gb{\beta} \def\gg{\gamma_1,\alpha} \def\gb{\beta} \def\gg{\gamma_2\}$, $\gb$ and $\gg$ small and \eqref{cd1} and \eqref{cd2} hold. We will choose $\gb$ and $\gg$ in the form \bel{form} \gb=\alpha} \def\gb{\beta} \def\gg{\gamma^{-s_1} \quad \text{and} \quad \gg=\alpha} \def\gb{\beta} \def\gg{\gamma^{-s_2}, \quad s_1>0,s_2>0. \end{equation} Then \eqref{cd1} and \eqref{cd2} become \bel{cds1} \begin{aligned} \alpha} \def\gb{\beta} \def\gg{\gamma^{1-ps_1 -ps_2} &\to 0 \quad \text{as } \alpha} \def\gb{\beta} \def\gg{\gamma \to \infty, \\ \alpha} \def\gb{\beta} \def\gg{\gamma^{1-ps_1 -(p-1)s_2} &\to \infty \quad \text{as } \alpha} \def\gb{\beta} \def\gg{\gamma \to \infty. \end{aligned} \end{equation} We now choose $s_1=\frac{1}{2p}$ and $s_2=\frac{1}{2p-1}$ then \eqref{cds1} holds and thus we get \eqref{RN}. This implies $$ \lambda(\CL_\alpha,\BBR^N) \geq (\liminf_{|x| \to \infty}V - 2\vge). $$ Since $\vge>0$ is arbitrary, we derive $$ \liminf_{\alpha} \def\gb{\beta} \def\gg{\gamma \to \infty}\lambda(\CL_\alpha,\BBR^N) \geq \liminf_{|x| \to \infty}V. $$ \\${}$ \hfill $\square$ Before proving Theorem \ref{lim2}, we first need the following result. \begin{proposition} \label{proplim} Assume $\{V_\alpha} \def\gb{\beta} \def\gg{\gamma\}$ is a sequence of functions in $L_{loc}^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. For $\alpha} \def\gb{\beta} \def\gg{\gamma>0$, denote \bel{Ka} \CK_\alpha} \def\gb{\beta} \def\gg{\gamma[\phi]:=-\Delta_p\phi+V_\alpha} \def\gb{\beta} \def\gg{\gamma\,\phi^{p-1}, \quad \phi} \def\vgf{\varphi} \def\gh{\eta \geq 0. \end{equation} Assume $V \in L_{loc}^\infty(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ and $V_n\rightharpoonup V$ in $L^1_{loc}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ as $\alpha} \def\gb{\beta} \def\gg{\gamma \to \infty$. Then $$\limsup_{\alpha} \def\gb{\beta} \def\gg{\gamma \to\infty}\lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)\leq \lambda(\CK_V,\Omega} \def\Gx{\Xi} \def\Gy{\Psi).$$ \end{proposition} \note{Proof} Put $\ol\lambda:=\limsup_{\alpha} \def\gb{\beta} \def\gg{\gamma\to\infty}\gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. Obviously, $\ol\lambda\in(-\infty,\infty)$. Therefore there exists a subsequence, still denoted by the same notation, such that $\gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi) \to \ol\lambda$ as $\alpha} \def\gb{\beta} \def\gg{\gamma \to \infty$. We denote by $\phi_\alpha} \def\gb{\beta} \def\gg{\gamma$ the generalized principal eigenfunction associated to $\gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$ with normalization $\phi_\alpha} \def\gb{\beta} \def\gg{\gamma(0)=1$. Since $\{V_\alpha} \def\gb{\beta} \def\gg{\gamma\}$ is locally uniformly bounded in $\Omega} \def\Gx{\Xi} \def\Gy{\Psi$, by the Harnack inequality we deduce that $\{\phi} \def\vgf{\varphi} \def\gh{\eta_\alpha} \def\gb{\beta} \def\gg{\gamma\}$ is locally uniformly bounded in $C^1_{loc}(\Omega} \def\Gx{\Xi} \def\Gy{\Psi)$. By local regularity results for quasilinear elliptic equations (see \cite{Di}) and a standard argument, we deduce that, up to a subsequence, $\{\phi_\alpha} \def\gb{\beta} \def\gg{\gamma\}$ converges in $C^1_{loc}(\Omega)$ to a nonnegative function $\ol \phi$ which is a weak solution of $$ - \Gd _p\bar \phi} \def\vgf{\varphi} \def\gh{\eta + V \bar \phi} \def\vgf{\varphi} \def\gh{\eta^{p-1} = \bar \gl \,\bar \phi} \def\vgf{\varphi} \def\gh{\eta^{p-1} \quad \text{in } \Omega} \def\Gx{\Xi} \def\Gy{\Psi $$ and satisfies $\bar \phi(0)=1$. By the strong maximum principle, one has $\ol\phi>0$ in $\Omega$. Therefore, by choosing $\psi=\bar \phi} \def\vgf{\varphi} \def\gh{\eta$ in the definition of $ \lambda(\CK_V,\Omega)$ in \eqref{lambda}, we deduce that $ \lambda(\CK_V,\Omega)\geq \ol\lambda$. \\${}$ \hfill $\square$ \medskip We are led to \noindent \textbf{Proof of Theorem \ref{lim2}.} \noindent (i)\, Observe that $\{V_\alpha} \def\gb{\beta} \def\gg{\gamma\}$ is locally uniformaly bounded in $\BBR^N$ and $V_\alpha \to V(0)$ in $L_{loc}^1(\mathbb{R}^N)$ as $\alpha \to 0$. Define $$ {\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C} [\phi} \def\vgf{\varphi} \def\gh{\eta]:= -\Delta_p \phi} \def\vgf{\varphi} \def\gh{\eta + V(0) \phi} \def\vgf{\varphi} \def\gh{\eta^{p-1} $$ and denote by $\lambda({\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C},\mathbb{R}^N)$ the generalized principal eigenvalue of ${\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C}$ in $\mathbb{R}^N$. By Proposition \ref{proplim}, we get \begin{equation} \label{limsup1} \limsup_{\alpha \to 0}\lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N) \leq \lambda({\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C},\mathbb{R}^N). \end{equation} On the other hand, by Theorem \ref{main0} (i), for any $\alpha>0$, \begin{equation} \label{limsup2} \lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N) \geq \inf_{\mathbb{R}^N}V_\alpha = V(0). \end{equation} Hence from Theorem \ref{equivalence}, \eqref{limsup1} and \eqref{limsup2} we get \begin{equation} \label{limsup3} V(0) \leq \liminf_{\alpha \to 0} \lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N) \leq \limsup_{\alpha \to 0} \lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N) \leq \lambda({\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C},\mathbb{R}^N)=\lambda'({\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C},\mathbb{R}^N). \end{equation} We next prove that \begin{equation} \label{limsup4} \lambda'({\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C},\mathbb{R}^N) \leq V(0). \end{equation} To this end, from the Definition \ref{gpe}, it is sufficient to show that there exists $0<\psi \in W^{1,p}(\mathbb{R}^N)$ such that ${\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C} [\psi] \leq V(0)\psi^{p-1}$ in the weak sense in $\mathbb{R}^N$. Define \bel{psi} \psi(x):=\chi_{B_1}(x) + e^{\frac{N-1}{p-1}(1-|x|)}\chi_{B_1^c}(x) \end{equation} then $\psi$ is the desired function. Indeed, we see that $\Delta_p \psi =0$ in $B_1$ and $$ \Delta_p \psi(x) = \frac{(N-1)^p}{(p-1)^{p-1}}e^{(N-1)(|x|-1)}\Big( 1 - \frac{1}{|x|} \Big) \geq 0 \quad \text{in } B_1^c. $$ It follows that ${\mathcal A}} \def\CB{{\mathcal B}} \def\CC{{\mathcal C} [\psi] \leq V(0)\psi^{p-1}$ in the weak sense in $\mathbb{R}^N$. By combining \eqref{limsup3} and \eqref{limsup4} we obtain \eqref{lubu}. Next, let us prove \eqref{lubu1}. By a similar argument as above and by Proposition \ref{proplim}, we obtain $$\limsup_{\alpha\to\infty}\gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)\leq \lambda(\mathcal{B},\mathbb{R}^N),$$ where $$\mathcal{B}[\phi} \def\vgf{\varphi} \def\gh{\eta]:= -\Delta_p \phi} \def\vgf{\varphi} \def\gh{\eta +\ol V(x) \phi} \def\vgf{\varphi} \def\gh{\eta^{p-1}\quad\quad\textrm{and}\quad\quad \ol V(x):=\left\{ \begin{aligned} &\sup_{\mathbb{R}^N}V(x)\quad &&\text{if } x\neq 0\\ &\inf_{\mathbb{R}^N}V(x)&&\text{if } x= 0. \end{aligned} \right.$$ Thanks to variational characterization of $\lambda(\mathcal{B},\mathbb{R}^N\setminus\{0\})$, one has \begin{eqnarray} \lambda(\mathcal{B},\mathbb{R}^N\setminus\{0\})&=&\inf_{\su{\phi \in C^1_c(\mathbb{R}^N\setminus\{0\}) \setminus \{0\}}{\|\phi\|_{L^p(\mathbb{R}^N)}=1}}\int_{\mathbb{R}^N\setminus\{0\}}(|\nabla \phi} \def\vgf{\varphi} \def\gh{\eta|^p + \ol V|\phi} \def\vgf{\varphi} \def\gh{\eta|^p)dx\nonumber\\ &=&\inf_{\su{\phi \in C^1_c(\mathbb{R}^N\setminus\{0\}) \setminus \{0\}}{\|\phi\|_{L^p(\mathbb{R}^N)}=1}}\int_{\mathbb{R}^N\setminus\{0\}}(|\nabla \phi} \def\vgf{\varphi} \def\gh{\eta|^p + \sup_{\mathbb{R}^N}V|\phi} \def\vgf{\varphi} \def\gh{\eta|^p)dx\nonumber\\ &=&\inf_{\su{\phi \in C^1_c(\mathbb{R}^N) \setminus \{0\}}{\|\phi\|_{L^p(\mathbb{R}^N)}=1}}\int_{\mathbb{R}^N}(|\nabla \phi} \def\vgf{\varphi} \def\gh{\eta|^p + \sup_{\mathbb{R}^N}V|\phi} \def\vgf{\varphi} \def\gh{\eta|^p)dx\nonumber\\ &=&\sup_{\mathbb{R}^N}V.\nonumber \end{eqnarray} Hence $$\limsup_{\alpha\to\infty}\gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)\leq \lambda(\mathcal{B},\mathbb{R}^N)\leq \lambda(\mathcal{B},\mathbb{R}^N\setminus\{0\})= \sup_{\mathbb{R}^N}V.$$ Let us show that $$\liminf_{\alpha\to\infty}\lambda_\alpha\geq \sup_{\mathbb{R}^N} V.$$ By the variational characterization of $\lambda_\alpha$, we have \begin{eqnarray} \gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)&=&\inf_{\su{\phi \in C^1_c(\mathbb{R}^N) \setminus \{0\}}{\|\phi\|_{L^p(\mathbb{R}^N)}=1}}\int_{\mathbb{R}^N}(|\nabla \phi} \def\vgf{\varphi} \def\gh{\eta|^p + V(\alpha x)|\phi} \def\vgf{\varphi} \def\gh{\eta|^p)dx\nonumber\\ &=&\inf_{\su{\phi_\alpha \in C^1_c(\mathbb{R}^N) \setminus \{0\}}{\|\phi_\alpha\|_{L^p(\mathbb{R}^N)}=1}}\int_{\mathbb{R}^N}(\alpha^p|\nabla \phi} \def\vgf{\varphi} \def\gh{\eta_\alpha|^p +V(x)|\phi} \def\vgf{\varphi} \def\gh{\eta_\alpha|^p)dx\nonumber\\ &=&\lambda(-\alpha^p\Delta_p +V,\mathbb{R}^N),\nonumber \end{eqnarray} where $\phi_\alpha(x)=\phi(x/\alpha)$. Applying (\ref{liminf}), we derive $$\liminf_{\alpha\to\infty}\gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)\geq \liminf_{|x|\to\infty}V(x)= \sup_{\mathbb{R}^N}V(x).$$ This concludes the proof. \medskip \noindent (ii)\, By Theorem \ref{main0}, 2i), $\lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)\geq \alpha\inf_{\Omega} V$. Therefore, to prove the statement (ii), it is sufficient to show that \begin{equation} \limsup_{\alpha\to\infty}\frac{\lambda(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)}{\alpha}\leq \inf_{\Omega} \def\Gx{\Xi} \def\Gy{\Psi} V.\label{10.1.1} \end{equation} Since $V$ is upper semi-continuous, for any $\epsilon>0$, there exists a ball $B\subset\Omega$ such that $$ V(x)<\inf_{\Omega} V+\epsilon \quad \forall x \in B.$$ Let $\lambda_B$ and $\phi_B$ be the Dirichlet principal eigenvalue and a corresponding eigenfunction of the operator of $-\Delta_p$ in $B$. For $\alpha>\lambda_B/\epsilon$, we have $$ \begin{aligned} \CK_\alpha[\phi} \def\vgf{\varphi} \def\gh{\eta_B] -\alpha(\inf_\Omega V+2\epsilon)\phi_B^{p-1}&=(\lambda_B+\alpha(V(x)-\inf_\Omega V-2\epsilon))\phi_B^{p-1}\\ &<(\lambda_B-\alpha\epsilon)\phi_B^{p-1}<0\nonumber \end{aligned} $$ By taking $\psi=\phi_B$ in the definition of $\lambda'(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,B)$ in Definition \ref{gpe}, we get $$\lambda'(\CK_\alpha,B)\leq \alpha(\inf_\Omega V+2\epsilon).$$ Since $\lambda'(\CK_\alpha,B)=\lambda(\CK_\alpha,B)\geq \gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)$, it follows that $$\frac{\gl(\CK_\alpha} \def\gb{\beta} \def\gg{\gamma,\BBR^N)}{\alpha}\leq \inf_\Omega V+2\epsilon. $$ We achieve (\ref{10.1.1}) due to the arbitrariness of $\epsilon$. \medskip \noindent (iii)\; This can be proved by using a similar argument as above and thus we omit the detail.\\${}$ \hfill $\square$ \section{Simplicity of the generalized principal eigenvalue and maximum principle} \noindent \textbf{Proof of Theorem \ref{simplicity1}.} Let $v \in W_{loc}^{1,p}(\BBR^N)$ be a positive weak solution of \eqref{eigen} in $\BBR^N$. \noindent \textit{Step 1: We show that} $$ {\mathbb D}:=\{ d>0: u \leq d v \quad \text{in } \BBR^N \} \ne \emptyset. $$ Indeed, suppose by contradiction that ${\mathbb D} = \emptyset$. Then for each $n \in \BBN$, there exists $x_n$ such that $u(x_n) \geq n v(x_n)>0$. We consider two cases: (i) up to a subsequence, $x_n \to x^* \in \BBR^N$, (ii) up to a subsequence, $|x_n| \to \infty$. In case (i), by passing to the limit, $u(x^*)=\infty$, which is a contradiction. In case (ii), since $u$ is a solution of \eqref{eigen} of minimal growth at infinity and $v$ is a solution of \eqref{eigen} then there exists $R^*>0$ and $k>0$ such that $u \leq k v$ in $B_{R^*}^c$. Then we choose $n_0>2k$ large enough such that $|x_n|>R^*$ for every $n \geq n_0$. Therefore $$ u(x_n) \geq n v(x_n) > 2k v(x_n) \geq 2u(x_n) \quad \forall n \geq n_0, $$ which is a contradiction. \smallskip \noindent \textit{Step 2: Scaling process.} Put $ \ell:=\inf \mathbb D} \def\BBE {\mathbb E} \def\BBF {\mathbb F$ then $u \leq \ell v$ in $\BBR^N$. We consider two cases. \textit{Case 1:} There exists $\tilde x \in \BBR^N$ such that $u(\tilde x)=\ell v(\tilde x)$. Put $w:=\ell v - u$ then $w \geq 0$ in $\BBR^N$ and $w(\tilde x)=0$. Put $$W:=V-\lambda(\CK_V,\mathbb{R}^N).$$ We have \bel{trans} \begin{aligned} 0 &= -\Delta_p (\ell v)+W (\ell v)^{p-1} - (-\Delta_p u+Wu^{p-1}) \\ &= -(\Delta_p (\ell v) - \Delta_p u) + W [(\ell v)^{p-1} - u^{p-1}] \\ &= -\sum_{i,j}\partial_{i}(a_{ij}(x)\partial_j w) + b(x)w \end{aligned} \end{equation} where $$ \begin{aligned} a_{ij}:&=|t_i \nabla (\ell v) + (1-t_i)\nabla u|^{p-4}\Big[ \gd_{ij}|t_i \nabla (\ell v) + (1-t_i)\nabla u|^{2} \\ & + (p-2)(t_i \partial_i (\ell v) + (1-t_i)\partial_i u)(t_i \partial_j (\ell v) + (1-t_i)\partial_j u)\Big] \end{aligned} $$ with some $t_i \in (0,1)$ and $$ b(x):=\left\{ \begin{aligned} &W(x)\frac{(\ell v(x))^{p-1}-u(x)^{p-1}}{\ell v(x)-u(x)} \quad &&\text{if } \ell v(x) \neq u(x)\\ &0 &&\text{if } \ell v(x) = u(x). \end{aligned} \right. $$ We see that $\nabla (\ell v(\tilde x))=\nabla u(\tilde x) \ne 0$, hence $$ a_{ij}(\tilde x)=|\nabla u(\tilde x)|^{p-4}\Big[ \gd_{ij}|\nabla u(\tilde x)|^2 + (p-2)\partial_i u (\tilde x) \partial_j u(\tilde x) \Big]. $$ Therefore the matrix $(a_{ij}(\tilde x))$ is positive definite. Consequently, $(a_{ij})$ is also positive definite in ball $B_\gd(\tilde x)$ for some small $\gd>0$. We see that $b$ is bounded in $B_\gd(\tilde x)$. From \eqref{trans}, it follows that $$ -\sum_{i,j}\partial_{i}(a_{ij}\partial_j w) + b^+ w \geq 0 \quad \text{in } B_{\gd}(\tilde x).$$ By the strong maximum principle for linear equations with principal part of divergence form \cite[Theorem 8.19]{GTbook}, we deduce that $w=0$ in $B_\gd(\tilde x)$. In light of the strong comparison principle \cite[Theorem 3.2]{FrPi}, $w=0$ in $\BBR^N$, hence $u=\ell v$ in $\BBR^N$. \textit{Case 2: $u(x)<\ell v(x)$ for every $x \in \BBR^N$ and there exists a sequence $\{x_n\}$ such that $|x_n| \to \infty$ as $n \to \infty$} and \bel{liml} \lim_{n \to \infty}\frac{u(x_n)}{v(x_n)}=\ell. \end{equation} Put $r_n:=|x_n|$ and $M_n:=\sup_{\partial B_{r_n}(0)}v(x)$. Set $u_n(x):=M_n^{-1}u(r_n x)$, $v_n(x):=M_n^{-1}v(r_n x)$ and $W_n(x)=r_n^{p-1} W(r_n x)$. Then $u_n$ and $v_n$ are solutions of $$ - \Gd_p u + W_n u^{p-1} = 0 \quad \text{in } \BBR^N. $$ By \eqref{condV}, there exists $C>0$ and $R_0>0$ such that $$ W(x) \leq C |x|^{-(p-1)} \quad \forall x \in B_{R_0}^c. $$ This implies that, for $n$ large enough (such that $r_n >2R_0$), $$ |W_n(x)|=r_n^{p-1}|W(r_n x)| \leq Cr_n^{p-1}|r_n x|^{-(p-1)}=C|x|^{-(p-1)} \leq C' \quad \text{in } B_{\frac{1}{2}}^c.$$ Since $\sup_{\partial B_1}v_n=1$, by Harnack inequality, for any $\rho>1$ there exists a positive constant $C_{\rho}$ independent of $n$ such that $v_n \leq C_\rho$ in $B_{\rho} \setminus B_{\frac{3}{4}}$ for every $n$. It follows that $u_n \leq C_\rho\ell$ in $B_{\rho} \setminus B_{\frac{3}{4}}$ for every $n$. Note that, up to a subsequence, $\{W_n\}$ converges to a function $\tilde W \in L^\infty( B_{\frac{3}{4}}^c)$ in weak-star topology of $L^\infty$. Therefore, by regularity results for quasilinear elliptic equations and a standard argument, we deduce that, up to a subsequence, $\{u_n\}$ and $\{v_n\}$ converge in the $C_{loc}^1(B_{\frac{3}{4}}^c)$ topology to functions $\tilde u$ and $\tilde v$ respectively which are weak solutions of equation $$ - \Gd_p u + \tilde W u^{p-1} = 0 \quad \text{in } B_{\frac{3}{4}}^c. $$ Put $y_n:=r_n^{-1}x_n$ then $|y_n|=1$ for every $n$. Hence, up to a subsequence, $y_n \to \tilde y \in \partial B_1$ and $\gk_n \to \tilde \gk \leq \ell$. Therefore if $\nabla \tilde u(\tilde y) \ne 0$ then we can use the strong comparison principle as in Case 1 to deduce that $\tilde u=\ell \tilde v$ in $B_{\gd}(\tilde y)$ for some $\gd>0$ small. Consequently, due to the strong comparison principle \cite[Theorem 3.2]{FrPi}, $\tilde u = \ell \tilde v$ in $B_{\frac{3}{4}}^c$. Now we prove that $\nabla \tilde u(\tilde y) \ne 0$. Indeed, from \eqref{liml}, we have $\tilde u(\tilde y)=\ell \tilde v(\tilde y)$. Since $\{v_n\}$ converges to $\tilde v$ in $C_{loc}^1(\BBR^N)$. Therefore, there exists $n_1$ large enough such that for every $n \geq n_1$, \bel{al1} |\nabla \tilde v(x)|>\frac{1}{2}|\nabla v_{n}(x)| \quad \forall x \in B_2(0). \end{equation} From \eqref{nablauu}, we deduce that there exist $\vge>0$ and $n_2>0$ such that for every $n \geq n_2$, \bel{al2} \frac{r_n |\nabla v(r_n \tilde y)|}{v(r_n \tilde y)}> \vge. \end{equation} Next fix $n_0>\max\{ n_1,n_2 \}$. By applying Harnack inequality for $v_{n_0}$, we derive that there exists a positive constant $C=C(N,p,\norm{V}_{L^\infty(\BBR^N)},r_{n_0})$ such that \bel{al3} M_{n_0} \leq C \inf_{\partial B_{r_{n_0}}(0)}v_{n_0} \leq C v(r_{n_0}\tilde y). \end{equation} Combining \eqref{al2} and \eqref{al3} yields \bel{al4} |\nabla v_{n_0}(\tilde y)|=\frac{r_{n_0} |\nabla v(r_{n_0} \tilde y)|}{M_{n_0}}=\frac{r_{n_0} |\nabla v(r_{n_0} \tilde y)|}{v(r_{n_0}\tilde y)}\frac{v(r_{n_0}\tilde y)}{M_{n_0}}>\frac{\vge}{C}>0. \end{equation} From \eqref{al1} and \eqref{al4}, we get $ |\nabla \tilde v(x)|>\frac{\vge}{2C}>0$. Therefore $$|\nabla \tilde u(\tilde y)|=\ell |\nabla \tilde v(\tilde y)|=\frac{\ell \vge}{C}>0.$$ \noindent \textit{Step 3: End of proof.} Put $$\gk_n:=\inf_{x \in \partial B_{r_n}(0)}\frac{u(x)}{v(x)}$$ then $\gk_n \leq \ell$. Therefore, up to a subsequence, $\gk_n \to \gk \leq \ell$. From Step 2, we deduce that $\gk=\ell$. Consequently, for every $\ge>0$ there exists $n_\ge$ such that for any $n \geq n_\ge$, $$ (\ell-\ge)v(x) \leq u(x) \leq \ell v(x) \quad \forall x \in \partial B_{r_n}(0). $$ By the weak comparison principle \cite[Theorem 3.1]{FrPi}, $(\ell -\ge)v \leq u$ in $B_{r_n}(0)$. Letting $n \to \infty$ and $\ge \to 0$ implies $\ell v \leq u$ in $\BBR^N$. Thus $u=\ell v$ in $\BBR^N$. \\${}$ \hfill $\square$ \begin{proposition} \label{class1} Assume $p\geq 2$ and $V \in L^\infty(\BBR^N)$ such that \bel{kcond} \lambda(\CK_V,\BBR^N) < \liminf_{|x| \to \infty}V(x) - \mu \end{equation} for some $\mu>0$. Let $u$ be a positive weak solution of \eqref{eigen}. If \bel{c1}\limsup_{|x|\to\infty}\frac{u(x)}{e^{\tilde{\omega}|x|}}<\infty \quad \text{with} \quad \tilde{\omega}:=\left(\frac{\mu}{N(p-1)}\right)^{\frac{1}{p}}.\end{equation} Then \bel{c2'} \lim_{|x|\to\infty}e^{\tilde{\omega}|x|^{}}u(x)=0.\end{equation} \end{proposition} \note{Proof} Since $u$ satisfies \eqref{eigen}, by \eqref{kcond}, for any $\varepsilon>0$ there exists $R=R(\varepsilon)$ such that $$-\Gd_p u + (\mu+\varepsilon)u^{p-1} \leq 0 \quad \text{in the weak sense in } B_R^c.$$ Let $\CL_\vge[\phi]:=-\Delta_p\phi+(\mu+\varepsilon)\phi^{p-1}$. It is easy to see that $\CL_\vge [u]\leq 0$ in the weak sense in $B_R^c$. For any $\rho>0$, set $$ w^1_\rho(x):=e^{(R+\rho)^{}(\tau-\omega)}e^{\omega|x|^{}}, \qquad w^2_\rho(x):= e^{R^{}(\tau+\omega)}e^{-\omega|x|^{}},$$ $$ w_\rho:=w^1_\rho+w^2_\rho$$ where $\omega,\tau$ and $R$ will be chosen later. Observe that, in $\BBR^N \setminus \{0\}$, \begin{equation} \Delta_p w_\rho=(p-2)|\nabla w_\rho|^{p-4}\left\langle D^2w_\rho \, \nabla w_\rho,\nabla w_\rho\right\rangle+|\nabla w_\rho|^{p-2}\Delta w_\rho. \end{equation} By Cauchy-Schwarz inequality, one has \begin{equation} \label{prop1.1} \Delta_p w_\rho \leq((p-2)N\max_{ij}|\partial_{ij} w_\rho|+|\Delta w_\rho|)|\nabla w_\rho|^{p-2}. \end{equation} Next, we look for an upper bound for the right-hand side of \eqref{prop1.1}. Direct computation yields, for every $x \neq 0$, \begin{equation}\nonumber \nabla w_\rho=\omega\frac{x}{|x|}\, w_\rho^1-\omega\frac{x}{|x|}\, w_\rho^2, \end{equation} thus \begin{equation}\label{NV1_expo1} |\nabla w_\rho|^{p-2}\leq \omega^{p-2}w_\rho^{p-2}. \end{equation} For every $1 \leq i,j\leq N$ and $x \neq 0$, \begin{eqnarray} \nonumber\partial_{ij}w_\rho&=&\omega^2\frac{x_ix_j}{|x|^{2}}w_\rho^1-\omega\frac{x_ix_j}{|x|^{3}}w_\rho^1+\delta(i-j)\omega\frac{w_\rho^1}{|x|^{}}\nonumber + \omega^2\frac{x_ix_j}{|x|^{2}}w_\rho^2 \omega\frac{x_ix_j}{|x|^{3}}w_\rho^2-\delta(i-j)\omega\frac{w_\rho^2}{|x|^{}} \end{eqnarray} where $\delta$ is the Dirac function. Since $|x_ix_j|\leq |x|^2$, it follows that \bel{esti1} \begin{aligned} |\partial_{ij}w_\rho|&\leq \omega^2 w_\rho +\omega\frac{w_\rho}{|x|}+\omega\frac{w_\rho}{|x|}= (\gw^2|x|^{}+2\omega)\frac{w_\rho}{|x|} \end{aligned} \end{equation} and hence \bel{esti2} |\Delta w_\rho|\leq N(\gw^2|x|^{}+2\omega)\frac{w_\rho}{|x|}. \end{equation} Combining \eqref{prop1.1}-\eqref{esti2}, we have $$ \Delta_pw_\rho \leq N(p-1)\omega^{p-1}(\gw|x|+2)\frac{w_\rho^{p-1}}{|x|}. \nonumber $$ Put $A:=2N(p-1)\omega^{p-1}.$ As $|x|\geq R$, one gets \bel{sup1} \CL_\vge[w_\rho]\geq w^{p-1}_\rho\left[-N(p-1)\omega^p- \frac{A}{|x|}+\mu+\varepsilon\right]. \end{equation} One can choose $R$ and $\gw$ such that the right-hand side of \eqref{sup1} is nonnegative. Indeed, since $|x|^{-1}\to0$ as $|x|\to\infty$, there exists $R(\vge)$ such that, for every $R>R(\vge)$, $A|x|^{-1}\leq \varepsilon/2\quad \forall x \in B^c_{R}.$ Take $$ \omega:=\left(\frac{2\mu+\varepsilon}{2N(p-1)}\right)^{\frac{1}{p}}, $$ we obtain $\CL_\ge[w_\rho] \geq 0$ in $B_{R+\rho} \setminus B_\rho$. We next show that $w_\rho$ dominates $u$ on $\partial B_{R+\rho} \cup \partial B_\rho$. Indeed, by \eqref{c1}, one can finds $C>0$ such that $u(x)\leq Ce^{\tilde{\omega}|x|^{}}$ in $\mathbb{R}^N$. Therefore, we can take $\tau$ arbitrarily in $(\tilde{\omega},\omega)$ and $R$ sufficiently large such that for any $\rho>0$, one has \begin{equation} \left\lbrace\begin{array}{ll} w_\rho(x)\geq e^{ R^{}\tau}\geq C e^{ R^{}\tilde{\omega}}\geq u(x),& \textrm{as $|x|=R$}\\ w_\rho(x)\geq e^{ (R+\rho)^{}\tau}\geq C e^{ (R+\rho)^{}\tilde{\omega}}\geq u(x),&\textrm{as $|x|=R+\rho$}.\nonumber \end{array}\right. \end{equation} Fix such $\omega,\tau$ and $R$. Applying the weak comparison principle \cite{GS}, we obtain $$u(x)\leq w_\rho(x)=e^{(R+\rho)^{}(\tau-\omega)}e^{\omega|x|^{}}+e^{R^{}(\tau+\omega)}e^{-\omega|x|^{}}\quad\quad\textrm{in $B_{R+\rho}\setminus B_R$}.$$ Sending $\rho\to\infty$ yields $$u(x)\leq e^{R^{}(\tau+\omega)}e^{-\omega|x|^{}}\quad\quad\textrm{in $\mathbb{R}^N\setminus B_R$}.$$ The fact $\omega>\tilde{\omega}$ confirms the proof.\\${}$ \hfill $\square$ \medskip \begin{lemma} \label{exgrowth1} Assume $V \in L^\infty(\BBR^N)$. There exist positive constants $C$ and $\gb$ depending on $N,p,\| V \|_{L^\infty(\BBR^N)}$ such that if $u \in W_{loc}^{1,p}(\BBR^N)$ is a positive weak solution of \eqref{eqKV} in $\BBR^N$ then \bel{exgrowth} u(x) \leq C e^{\gb |x|}u(0) \quad \forall x \in \BBR^N. \end{equation} \end{lemma} \begin{proof} By Harnack inequality \cite{Se} (see Theorem 5 for $p<N$, Theorem 6 for $p=N$ and Theorem 9 for $p>N$), we deduce that there exists a positive constant $C$ depending on $N,p,\| V \|_{L^\infty(\BBR^N)}$ such that \bel{Har1} u(x) \leq C u(0) \quad \forall x \in B_1(0). \end{equation} Take $x_0 \in \partial B_1(0)$. We claim that for any nonnegative $m$, there holds \bel{ith} u(x) \leq C^{m+1}u(0) \quad \forall x \in B_1(m x_0). \end{equation} We will prove \eqref{ith} by induction on $m$. Obviously, \eqref{ith} holds true for $m=0$. Suppose that \eqref{ith} is valid for some positive integer $m$, we will show that \eqref{ith} also holds true for $m+1$. Indeed, observe that since $V \in L^\infty(\BBR^N)$, \eqref{Har1} still holds if we replace $u$ by $u(\cdot + y)$ for every $y \in \BBR^N$. In particular, with $|x_0|=1$, by replacing $u$ by $u(\cdot + (m+1)x_0)$, we get $$ u(x+(m+1)x_0) \leq Cu((m+1)x_0) \quad \forall x \in B_1(0). $$ By changing the variable and \eqref{ith}, we get $$ u(x) \leq Cu((m+1)x_0) \leq C^{m+2}u(0) \quad \forall x \in B_1((m+1)x_0). $$ Thus we have proved \eqref{ith}. By \eqref{ith} we deduce that $$ u(x) \leq C^{|x|+2}u(0) \quad \forall x \in B_1(m x_0). $$ This implies \eqref{exgrowth}. \end{proof} \noindent \textbf{Proof of Theorem \ref{simplicity2}.} \noindent \textit{Step 1: Exponential decay.} We prove that there exists $\mu>0$ such that if \eqref{decay} holds then every positive eigenfunction associated with $\lambda(\CK_V,\mathbb{R}^N)$ decays exponentially. Indeed, let $\phi} \def\vgf{\varphi} \def\gh{\eta \in W^{1,p}_{loc}(\mathbb{R}^N)$ be a positive eigenfunction associated with $\lambda(\CK_V,\mathbb{R}^N)$ then $\phi} \def\vgf{\varphi} \def\gh{\eta$ is a weak solution of \eqref{eigen}. By Lemma \ref{exgrowth1}, there exist a constant $C_\phi} \def\vgf{\varphi} \def\gh{\eta>1$ depending on $\phi} \def\vgf{\varphi} \def\gh{\eta, N,p,\| V \|_{L^\infty(\BBR^N)}$ and a constant $\gb>0$ depending on $N,p,\| V \|_{L^\infty(\BBR^N)}$ such that $$\phi} \def\vgf{\varphi} \def\gh{\eta(x)\leq C_\phi} \def\vgf{\varphi} \def\gh{\eta\,e^{\beta|x|} \quad \forall x \in \BBR^N.$$ Put $\mu:=N(p-1)\gb^p $. If \eqref{decay} holds then by invoking Proposition \ref{class1}, we deduce that $$ \phi} \def\vgf{\varphi} \def\gh{\eta(x) \leq C'_\phi} \def\vgf{\varphi} \def\gh{\eta e^{-\gb |x|} \quad \forall x \in \BBR^N$$ where $C'_\phi} \def\vgf{\varphi} \def\gh{\eta$ is a positive constant depending on $\phi} \def\vgf{\varphi} \def\gh{\eta,N,p$ and $\| V \|_{L^\infty(\BBR^N)}$. \medskip \noindent \textit{Step 2: Simplicity.} Let $\phi} \def\vgf{\varphi} \def\gh{\eta,\vgf \in W^{1,p}_{loc}(\mathbb{R}^N)$ be two positive eigenfunctions associated with $\gl(\CK_V,\BBR^N)$. We will prove that, under condition \eqref{decay}, there exist a constant $k$ such that $\phi} \def\vgf{\varphi} \def\gh{\eta=k \vgf$ in $\mathbb{R}^N$. Indeed, if \eqref{decay} holds then by Step 1 $u$ and $v$ are decay exponentially. In light of the regularity for quasilinear elliptic equations \cite{Lieberman} and the Harnack inequality, we deduce that $\phi} \def\vgf{\varphi} \def\gh{\eta, \vgf\in W^{1,p}(\mathbb{R}^N)$. Therefore, without lost of generality, we can normalize $u$ and $v$ so that $\norm{\phi} \def\vgf{\varphi} \def\gh{\eta}_{L^p(\BBR^N)}=\norm{\vgf}_{L^p(\BBR^N)}=1$. Set $$\vartheta:=\zeta^{1/p} \quad \text{with} \quad \zeta:=\frac{\phi} \def\vgf{\varphi} \def\gh{\eta^p+\vgf^p}{2},$$ then $\| \vartheta \|_{L^p(\mathbb{R}^N)}=1$. Set $$\theta:=\frac{\phi} \def\vgf{\varphi} \def\gh{\eta^p}{\phi} \def\vgf{\varphi} \def\gh{\eta^p+\vgf^p}\in(0,1).$$ One has $$\nabla \vartheta=\zeta^{-1+\frac{1}{p}}\left(\frac{\phi^{p-1}\nabla\varphi+\varphi^{p-1}\nabla\phi}{2}\right).$$ The convexity of the map $s\mapsto |s|^{p}$ yields \begin{eqnarray*} |\nabla \vartheta|^p&=&\zeta^{1-p}\left|\frac{\phi^{p-1}\nabla\phi+\varphi^{p-1}\nabla\varphi}{2}\right|^p \\ &=&\frac{\zeta^{}}{2^p}\left|\theta(x)\frac{\nabla\phi}{\phi}+(1-\theta(x))\frac{\nabla\varphi}{\varphi}\right|^p\nonumber\\ &\leq&\zeta \left(\theta(x)\left|\frac{\nabla\phi}{\phi}\right|^p+(1-\theta(x))\left|\frac{\nabla\varphi}{\varphi}\right|^p\right) \nonumber \\ &=&\frac{|\nabla u|^p+|\nabla v|^p}{2}.\nonumber \end{eqnarray*} The equality holds if and only if $\frac{\nabla\phi}{\phi}=\frac{\nabla \varphi}{\varphi}$ in $\mathbb{R}^N$. Hence, $$\int_{\mathbb{R}^N}|\nabla \vartheta|^pdx\leq\frac{1}{2}\left(\int_{\mathbb{R}^N}|\nabla \phi|^pdx+\int_{\mathbb{R}^N}|\nabla \varphi|^pdx\right).$$ It follows that $$ \mathcal{F}(\vartheta)\leq \frac{1}{2}(\mathcal{F}(\phi)+\mathcal{F}(\varphi)), $$ where $$\mathcal{F}(w):=\int_{\mathbb{R}^N}(|\nabla w|^p+V|w|^p)dx\qquad w\in W^{1,p}(\mathbb{R}^N).$$ Since both $\phi,\varphi\in W^{1,p}(\mathbb{R}^N)$ satisfy \eqref{eigen}, we deduce \begin{equation}\label{decay2} \mathcal{F}(\vartheta)\leq \frac{1}{2}(\mathcal{F}(\phi)+\mathcal{F}(\varphi))=\gl(\CK_V,\BBR^N). \end{equation} Let $\{\vartheta_n\} \subset C_c^1(\BBR^N)$ be a sequence converging to $\vartheta$ in $W^{1,p}(\BBR^N)$. Since $\vartheta>0$, we infer that $\vartheta_n>0$ for $n$ large enough. For large $n$, by Theorem \ref{main0} iii), we have $$ \lambda(\CK_V,\BBR^N)\leq\frac{\int_{\BBR^N}(|\nabla \vartheta_n|^{p}+V \vartheta_n^p)dx}{\int_{\BBR^N}\vartheta_n^p\,dx}. $$ Letting $n\to \infty$ yields \begin{equation} \label{decay3} \lambda(\CK_V,\BBR^N)\leq\frac{\int_{\BBR^N}(|\nabla \vartheta|^{p}+V \vartheta^p)dx}{\int_{\BBR^N}\vartheta^p\,dx} = \CF(\vartheta). \end{equation} By \eqref{decay2} and \eqref{decay3}, we get $$ \CF(\vartheta)=\gl(\CK_V,\BBR^N). $$ The above equality holds if and only if $\frac{\nabla\phi}{\phi}=\frac{\nabla \varphi}{\varphi}$ in $\mathbb{R}^N$, which implies $\nabla \left(\frac{\phi}{\varphi}\right)=0$ in $\BBR^N$. Therefore there exists $k>0$ such that $\phi=k\varphi$ in $\mathbb{R}^N$. This concludes the proof. \\${}$ \hfill $\square$ \medskip \noindent \textbf{Proof of Thereom \ref{spectrum}.} From the definition of $\CE(\BBR^N)$ and the definition of $\gl(\CK_V,\BBR^N)$ in \eqref{lambda}, we deduce that $\CE(\BBR^N) \subset (-\infty,\gl(\CK_V,\BBR^N) ]$. We next prove the reverse inclusion $(-\infty,\gl(\CK_V,\BBR^N) ] \subset \CE(\BBR^N)$. By Theorem \ref{main0} 2.(ii), $\gl(\CK_V,\BBR^N) \in \CE(\BBR^N)$. It remains to show that $\gl \in \CE(\BBR^N)$ for every $\gl < \gl(\CK_V,\BBR^N)$. Indeed, take $\gl < \gl(\CK_V,\BBR^N)$. For any $n \in \BBN$, let $f_n \in C^\infty(B_n)$ be nonnegative and not identically equal to zero in $B_n$ with $\text{\rm supp} f_n \in B_n \setminus B_{n-1}$. Since $\gl(\CK_V,B_n)>\gl(\CK_V,\BBR^N)>\gl$, it follows that $\gl(\CK_{V-\gl},B_n)>0$. By \cite[Theorem 2 (v)]{GS}, there exists a unique nonnegative weak solution $u_n \in W_0^{1,p}(B_n)$ of $$ \left\{ \begin{aligned} \CK_{V-\gl}[u_n] &= f_n \quad &&\text{in } B_n \\ u &= 0 &&\text{on } \partial B_n. \end{aligned} \right. $$ By the strong maximum principle \cite[Theorem 2 (ii)]{GS}, we obtain that $u_n$ is positive in $B_n$. Put $$ v_n(x):=\frac{u_n(x)}{u_n(0)} $$ then $v_n(0)=1$. By the Harnack inequality \cite{Tr} and regularity results for quasilinear elliptic equations \cite{Di}, up to a subsequence, the sequence $\{v_n\}$ converges in $C_{loc}^1(\BBR^N)$ to a function $v$ which is a nonnegative weak solution of $\CK_{V-\gl}[v]=0$ in $\BBR^N$. Since $v(0)=1$, by the Harnack inequality, we deduce that $v$ is positive in $\BBR^N$. Therefore $v$ is a positive weak solution of \eqref{eigenRN} in $\BBR^N$. It follows that $\gl \in \CE(\BBR^N)$. Finally $(-\infty,\gl(\CK_V,\BBR^N) ] \subset \CE(\BBR^N)$. This completes the proof. \\${}$ \hfill $\square$ \medskip \noindent \textbf{Proof of Theorem \ref{MP}.} Suppose by contradiction that $u$ is positive somewhere in $\mathbb{R}^N$. Since $\gl''(\CK_V,\mathbb{R}^N) >0$, there exist a function $\phi\in C^{1}_{loc}(\mathbb{R}^N)$ and positive number $\lambda$ and $\beta$ such that $\inf_{\mathbb{R}^N}\phi\geq \beta>0$ and $$\CK_V [\phi]\geq \lambda\phi^{p-1}\quad\quad\text{in the weak sense in } \mathbb{R}^N.$$ Since $u$ is continuous, $\sup_{\mathbb{R}^N}u<\infty$ and $\limsup_{|x|\to\infty}u(x)\leq0$, we can find a positive constant $\gamma>0$ and $x_0\in\mathbb{R}^N$ such that $$0<\gamma=\frac{u(x_0)}{\phi(x_0)}=\max_{\mathbb{R}^N}\frac{u(x)}{\phi(x)}<\infty.$$ It follows that there exists $r>0$ such that $u(x)>0$ in $\ol B_r(x_0)$, $u(x)\leq \gamma\phi(x)$, $\forall x\in \ol B_r(x_0)$ and $u(x_0)=\gamma\phi(x_0)$. Therefore, by strong comparison principle \cite[Theorem 3.2]{FrPi}, we get $\gamma\phi=u$ in $B_r(x_0)$. Therefore, for every $0<\psi \in C_c^1(B_r(x_0))$, \bel{contra1} \int_{B_r(x_0)}\CK_V[\gg \phi} \def\vgf{\varphi} \def\gh{\eta]\psi \,dx= \int_{B_r(x_0)}\CK_V[u]\psi \,dx. \end{equation} On the other hand, we have \begin{equation} \label{mp1} \CK_V[\gamma\phi]\geq \lambda\gamma^{p-1}\beta^{p-1}>0\geq \CK_V[u] \quad\text{ in the weak sense in }\in \BBR^N. \end{equation} This implies \bel{contra2} \int_{B_r(x_0)}\CK_V[\gg \phi} \def\vgf{\varphi} \def\gh{\eta]\psi \,dx> \int_{B_r(x_0)}\CK_V[u]\psi \,dx, \end{equation} which contradicts with \eqref{contra1}. Thus $u$ must be nonpositive. \\${}$ \hfill $\square$ \medskip \bigskip \noindent{\bf Acknowledgements.} P-T. Nguyen is supported by Fondecyt Grant 3160207.

https://arxiv.org/abs/1901.08834

Uniform existence of the IDS on lattices and groups

We present a general framework for thermodynamic limits and its applications to a variety of models. In particular we will identify criteria such that the limits are uniform in a parameter. All results are illustrated with the example of eigenvalue counting functions converging to the integrated density of states. In this case, the convergence is uniform in the energy.

\section{Introduction} \label{sec:intro} The thermodynamic limit, i.\,e.\ taking averages over larger and larger volumes, performs the transition from microscopic to macroscopic models. In the context of this paper, thermodynamic limits are used to define the integrated density of states (IDS) of discrete Schr\"odinger operators, in particular, random ones. The IDS of random Schr\"odinger operators has been studied in the mathematical literature at least since 1971, see e.\,g.\ \cite{Pastur-71,Shubin-79,KirschM-07,Veselic-08}. The IDS measures the number of quantum states per unit volume below a threshold energy~$E$ and is thus a function of~$E$. The rigorous implementation of this notion involves a thermodynamic limit, see \cref{sec:defsandmodels}. Classical results about the existence of this macroscopic volume limit apply to fixed energy, which is to say that the limit is considered pointwise. More recently, there has been increased interest to study stronger forms of convergence, for instance uniformly in the energy parameter. This has been studied in various geometric and stochastic contexts. For instance, there are works devoted to Hamiltonians associated to quasicrystals, \cite{LenzS-03a,LenzS-06}, to graphings \cite{Elek-07} and percolations graphs \cite{Veselic-05b,Veselic-06,AntunovicV-08b-short,SamavatSV-14}, as well as abstracts frameworks covering general classes of examples \cite{LenzV-09}, where this is certainly a non-exhaustive list of references. We will review here in particular the results obtained in \cite{LenzMV-08}, \cite{LenzSV-10,LenzSV-12}, \cite{PogorzelskiS-16}, \cite{SchumacherSV-17}, and \cite{SchumacherSV-18}. They have all in common that their findings can be formulated as Banach-valued ergodic theorems. The convergence results in this work are not restriced to the IDS of random Schr\"odinger operators, but provide a general framework for thermodynamic limits with respect to a Banach space topology. To this end, we introduce almost additive fields, which, after normalization with volume, converge as the volume exhausts the physical space. The microscopic structure, like the values of the potential of the Schr\"odinger operator and/or whether a percolation site is open or closed, are encoded in a \emph{coloring}. In applications, the coloring is often random, for the lack of detailed knowledge about the microscopic properties of the material. For the thermodynamic limit to exist, one needs a certain homogeneity of the coloring. In our case, finite portions of the coloring, called \emph{patterns}, should occur with a certain frequency. If there are only finitely many \emph{colors}, i.\,e.\ values of the coloring, this condition is natural and allows the existence of the thermodynamic limit to be proven in any Banach space. For random Schr\"odinger operators, one chooses the Banach space of right continuous bounded functions with $\infty$-norm. Unfortunately, typical potentials have infinitely many values. In that setting, we model the colorings as random fields with some independence. This allows to employ a multivariate version of the theorem of Glivenko--Cantelli. The paper is structured as follows. In \cref{sec:defsandmodels}, we introduce some specific random Schr\"odinger operators and their IDS in more detail. They serve as examples for the abstract theorems as well as illustrations of the meaning of our assumptions. We then present a series of theorems on thermodynamic limits with increasing complexity. \Cref{sec:finiteA} contains the oldest of the presented results, which was in some sense the motivation for further developments. It considers fields over~$\Z^d$ obeying the finiteness condition~\eqref{eq:Afinite} on the set of colors. Since the results are easier to motivate and less technical than later ones, they serve well as a point of departure. In \cref{sec:amenable}, we introduce briefly finitely generated amenable groups and show how the main result generalizes to this setting. We first consider amenable groups which satisfy an additional tiling condition. To remove this tiling condition, we outline quasi tilings and discuss the result on amenable groups, still keeping the finiteness condition. The finiteness condition~\eqref{eq:Afinite} is finally tackled and removed in \cref{sec:infiniteA}. For clarity, we first deal with the euclidean lattice~$\Z^d$ and only after that exhibit the most general formulation on finitely generated amenable groups without finiteness condition. An outlook, a list of symbols, and a list of references conclude the paper. \section{Physical models} \label{sec:defsandmodels} In this section, we introduce example systems to motivate the abstract results. For the moment we will stick to the simple geometry of~$\Z^d$ but use a notation that later generalizes to general geometries on amenable groups in a straightforward way. \subsection{The Anderson model on \texorpdfstring{$\Z^d$}{Zd}} \label{sec:Anderson_Zd} The Anderson model, introduced by P. W. Anderson in \cite{Anderson-58}, is a prototypical random Schr\"odinger operator. To define it, let us introduce some notation. As physical space we choose the group $G:=\Z^d$. To make neighborhood relations explicit, we introduce the \emph{Cayley graph} of~$G$. The Cayley graph of~$G$ has~$G$ itself as vertex set, and two vertices $v,w\in G$ are connected by an undirected edge, if $\norm[1]{v-w}=1$, where $\norm[1]\argmt$ is the usual norm on $\Z^d$ when viewed as subset of $\ell^1\{1,\dotsc,d\}$. In this case, we write $v\sim w$. Thus, the considered geometry is nothing but the usual $d$-dimensional lattice. However, to stay consistent with the notation required in later parts of the article, we already use the notion of a Cayley graph. The general definition of Cayley graphs is introduced in \cref{sec:amenable}. The group~$G$ acts transitively on its Cayley graph by translation: \begin{equation*} \tilde\tau\colon G\times G\to G\textq, \tilde\tau_gv:=(g,v):=v-g\text. \end{equation*} This group action lifts to a unitary group action on the square summable functions on the vertices of the Cayley graph, $\ell^2(G)$, which we denote by \begin{equation*} U_g\from\ell^2(G)\to\ell^2(G)\textq, (U_g\varphi)(v):=\varphi(\tilde\tau_gv)\text. \end{equation*} The \emph{Laplace operator} $\Laplace\from\ell^2(G)\to\ell^2(G)$, given by \begin{equation}\label{eq:laplaceop} (\Laplace\varphi)(v) :=\sum_{w\sim v}\bigl(\varphi(w)-\varphi(v)\bigr)\text, \end{equation} mimics the sum of the second derivatives. The operator~$-\Laplace$ is bounded, self-adjoint, positive semi-definite and serves as the quantum mechanical observable of the kinetic energy. Its spectrum is $\sigma(-\Laplace)=\Icc0{2d}$, as one can see with Fourier analysis. Note that the Laplace operator is equivariant with respect to $G$, i.\,e., for all $g\in G$, we have $\Laplace\circ U_g=U_g\circ\Laplace$. To build a Schr\"odinger operator, we need a multiplication operator~$V$ on $\ell^2(G)$, which plays the role of the observable of potential energy. At this stage, the randomness enters. To this end, we choose a (measurable) set~$\cA\subseteq\R$, which we call the set of colors, equip it with the trace topology inherited from~$\R$ and its Borel $\sigma$-algebra~$\Borel(\cA)$, and choose a probability measure~$\Prob_0$ on $(\cA,\Borel(\cA))$, i.\,e.\ the colors. The probability space for the random potential is $(\Omega,\cB,\Prob) :=(\cA,\Borel(\cA),\Prob_0)^{\tensor G} :=(\cA^G,\Borel(\cA)^{\tensor G},\Prob_0^{\tensor G})$. The \emph{random potential} $V\from\Omega\times G\to\cA$ returns for~$v\in G$ the $v$-th coordinate of~$\omega\in\Omega$: \begin{equation}\label{eq:randompotential} V_\omega(v):=V(\omega,v):=\omega_v:=\omega(v)\text, \end{equation} so that the random variables $\Omega\ni\omega\mapsto V_\omega(v)$, $v\in G$, are independent and identically distributed. For each $\omega\in\Omega$, the random potential~$V_\omega$ operates on $\ell^2(G)$ by multiplication. If~$\cA\in\Borel(\R)$ is bounded, the multiplication operator~$V_\omega$ is bounded and self-adjoint. Analogous to above, the group action of~$G$ on the Cayley graph lifts to an ergodic group action $\tau_g\from\Omega\to\Omega$ on~$\Omega$: \begin{equation*} (\tau_g\omega)_v:=\omega_{\tilde\tau_gv}\text, \end{equation*} and the random potential is equivariant, meaning $V_{\tau_g\omega}\circ U_g=U_g\circ V_\omega$ for all $g\in G$. For each $\omega\in\Omega$, the Schr\"odinger operator \begin{equation} \label{eq:schroedinger} H_\omega:=-\Laplace+V_\omega\from\ell^2(G)\to\ell^2(G) \end{equation} is well-defined, bounded, and self-adjoint. The operator family $(H_\omega)_{\omega\in\Omega}$ is the famous \emph{Anderson Hamiltonian} and is equivariant, too: $H_{\tau_g\omega}\circ U_g=U_g\circ H_\omega$ for all $g\in G$. This equivariance shows that the spectrum $\sigma(H_\omega)$ is a shift invariant quantity. Since the group action of~$G$ on~$\Omega$ is ergodic, the spectrum of $H_\omega$ is almost surely constant: there is a closed set $\Sigma\subseteq\R$ such that \begin{equation*} \Sigma=\sigma(H_\omega) \end{equation*} for $\Prob$-almost all $\omega\in\Omega$. In quantum mechanics, the spectrum is interpreted as the set of possible energies of the particle described by~$H_\omega$. The fact that it is deterministic and does not depend on the microscopic structure of the material makes the interpretation as a homogeneous material possible. More details can be found for example in \cite{FigotinP-92,Kirsch-08}. The main example to motivate and to illustrate the results in later chapters is the integrated density of states (IDS), also known as the spectral distribution function, of the Anderson Hamiltonian. Its definition is \begin{equation*} N\from\R\to\R\textq, N(E):=\E[\spr{\delta_0}{\ifu{\Ioc{-\infty}E}(H_\omega)\delta_0}]\text, \end{equation*} where $\delta_v\in\ell^2(G)$ is the Kronecker delta at $v\in G$, i.\,e.\ $\delta_0=\ifu{\{v\}}$, and $\ifu{\Ioc{-\infty}E}(H_\omega)$ is the spectral projection of~$H_\omega$ onto the energy interval $\Ioc{-\infty}E$. The IDS is monotone, bounded by~$1$, and gives rise to a probability measure~$\d N$, called the \emph{density of states measure}. The topological support of the density of states is the almost sure spectrum of~$H_\omega$. In the following, we will describe how to approximate the IDS using only finite matrices. For each $\Lambda\subseteq G$, we write $\ell^2(\Lambda)$ for the subspace of $\varphi\in\ell^2(G)$ with support $\supp\varphi\subseteq\Lambda$. The indicator function of~$\Lambda$, used as multiplication operator on~$\ell^2(G)$, is the self-adjoint orthogonal projection $\ifu\Lambda\from\ell^2(G)\to\ell^2(\Lambda)$. The operator $H_\omega^\Lambda\from\ell^2(\Lambda)\to\ell^2(\Lambda)$ is given by \begin{equation*} H_\omega^\Lambda:=\ifu\Lambda\circ H_\omega\circ(\ifu\Lambda)^*\text. \end{equation*} Now let~$\cF$ be the (countable) set of finite subsets of~$G$ and assume $\Lambda\in\cF$. The representing matrix of~$H_\omega^\Lambda$ contains the matrix elements $\spr{\delta_v}{H_\omega\delta_w}$ with $v,w\in\Lambda$ and is thus the $\Lambda\times\Lambda$ clipping of the representing $G\times G$ matrix of~$H_\omega$. Of course, $H_\omega^\Lambda$ is Hermitian and has real eigenvalues. For $\omega\in\Omega$, $\Lambda\in\cF$, the eigenvalue counting function $n(\Lambda,\omega)\from\R\to\R$, \begin{equation*} n(\Lambda,\omega)(E):=\tr\ifu{\Ioc{-\infty}E}(H_\omega^\Lambda) =\dim\{\varphi\in\ell^2(\Lambda)\mid\spr\varphi{H_\omega^\Lambda\varphi}\le E\} \end{equation*} is continuous from the right and counts the eigenvalues of~$H_\omega^\Lambda$ below the threshold energy~$E$ according to their multiplicity. We will view the family of eigenvalue counting functions as \begin{equation*} n\from\cF\times\Omega\to\B\text, \end{equation*} where~$\B$ is the Banach space of right continuous functions from~$\R$ to~$\R$. Denote by $\Lambda_L:=\Ico0L^d\isect\Z^d\in\cF$ a cube of side length $L\in\N$. The celebrated \emph{Pastur--Shubin formula} states that, for $\Prob$-almost all $\omega\in\Omega$ and all $E\in\R$ where the IDS~$N$ is continuous, we have \begin{equation}\label{eq:PasturShubin} N(E) =\lim_{L\to\infty}\setsize{\Lambda_L}^{-1}n(\Lambda_L,\omega)(E)\text, \end{equation} see e.\,g.\ \cite{SchumacherS-15}. In fact, since for this model in particular the IDS~$N$ is continuous at all energies, a straight forward argument, using also the boundedness and the monotonicity of~$N$, shows that the convergence does not only hold for all~$E\in\R$, but is actually uniform in~$E$. The Pastur--Shubin formula can be viewed as an ergodic theorem, since it states the equality of an ensemble average and a spatial average. \subsection{Quantum percolation models} \label{sec:percolation} \subsubsection{Site and edge percolation} \label{sec:siteedge} We first introduce \emph{site percolation} on the Cayley graph of $G=\Z^d$. Let $\cA:=\{0,1\}$, and as in \cref{sec:Anderson_Zd} let $(\Omega,\cB,\Prob):=(\cA,\Borel(\cA),\Psite)^{\tensor G}$ with a (non-de\-ge\-ne\-rate) Bernoulli probability~$\Psite$ on~$\cA$. In the percolation setting, the randomness does not determine the potential, but configuration space itself. The \emph{site percolation graph} $(\cV_\omega,\cE_\omega)$ corresponding to $\omega\in\Omega$ is induced by the Cayley graph of~$G$ on the vertex set \begin{equation*} \cV_\omega:=\{v\in G\mid\omega_v=1\}\text. \end{equation*} This means that the edge set is \begin{equation*} \cE_\omega:=\{\{v,w\}\subseteq\cV_\omega\mid v\sim w\}\text, \end{equation*} that is, we keep the edges of the Cayley graph which have both end points in~$\cV_\omega$. In \emph{edge percolation}, one does not erase random sites from the Cayley graph but instead random edges. The color set $\cA:=\{0,1\}^d$ is suitable to implement this strategy. Namely, for a (non-degenerate) Bernoulli probability~$\Pedge$ on $\{0,1\}$, we define $(\Omega,\cB,\Prob):=(\cA,\Borel(\cA),\Pedge^{\tensor d})^{\tensor G}$ and define the edge set corresponding to $\omega\in\Omega$ by \begin{equation*} \cE_\omega:=\Union\nolimits_{j=1}^d\bigl\{\{v,v+e_j\}\bigm|(\omega_v)_j=1\bigr\} \end{equation*} where~$e_j$ is the $j$-th standard basis vector of~$\Z^d$. The vertex set of the edge percolation graph contains all vertices to which an edge in~$\cE_\omega$ is attached: \begin{equation}\label{eq:verticesInEdgePercolation} \cV_\omega:=\{v\in G\mid\exists w\in G\colon\{v,w\}\in\cE_\omega\}\text. \end{equation} The Laplace operator on a subgraph $(\cV_\omega,\cE_\omega)$ of the Cayley graph of~$G$ is \begin{equation}\label{eq:laplace} \Laplace_\omega\from\ell^2(\cV_\omega)\to\ell^2(\cV_\omega)\textq, (\Laplace_\omega\varphi)(v):= \sum_{w\in\cV_\omega,\{v,w\}\in\cE_\omega}\bigl(\varphi(w)-\varphi(v)\bigr) \text. \end{equation} Since we want to use the group action of~$G$, we define the Hamiltonians on $\ell^2(G)$ via \begin{equation*} H_\omega\from\ell^2(G)\to\ell^2(G)\textq, (H_\omega\varphi)(v):= \begin{cases} -(\Laplace_\omega\varphi)(v)&\text{if $v\in\cV_\omega$ and}\\ \alpha\varphi(v) &\text{if $v\in G\setminus\cV_\omega$} \end{cases} \end{equation*} with a constant $\alpha\in\Ioo{2d}\infty$. The operator~$H_\omega$ leaves the subspaces $\ell^2(\cV_\omega)$ and $\ell^2(G\setminus\cV_\omega)$ invariant. Since $\Laplace_\omega$ is bounded by~$2d$ and $\alpha>2d$, the spectrum of of~$H_\omega$ is the disjoint union of $\sigma(-\Laplace_\omega)$ and $\{\alpha\}$, and the two components can be studied separately. In fact, the percolation Hamiltonians are equivariant, and again their spectrum is almost surely constant. As in \cref{sec:Anderson_Zd}, we define the IDS $N\from\R\to\R$ and the eigenvalue counting function $n\from\cF\times\Omega\to\B$ of~$H_\omega$. The Pastur--Shubin formula~\eqref{eq:PasturShubin} remains correct $\Prob$-almost surely for all energies $E\in\R$ at which~$N$ is continuous, see for instance \cite{Veselic-05a,Veselic-05b,KirschM-06,Veselic-06,AntunovicV-08b-short}. But in contrast to the Anderson Hamiltonian, the IDS of a percolation Hamiltonian is not continuous. In fact, the set of discontinuities of~$N$ is dense in~$\Sigma$. Note that since the IDS is monotone, so there can be at most countably many discontinuities. \begin{figure} \begin{center} \tikzsetnextfilename{prob4cluster}% \begin{tikzpicture} \draw[fill] (0,0) circle[radius=1mm] -- (1,0) circle[radius=1mm]; \foreach \x/\y in {0/1, -1/0, 0/-1} \draw[dotted] (0,0) -- (\x,\y) circle[radius=1mm]; \foreach \x/\y in {1/1, 2/0, 1/-1} \draw[dotted] (1,0) -- (\x,\y) circle[radius=1mm]; \begin{scope}[xshift=5cm] \draw[fill] (0,0) circle[radius=1mm] -- (1,0) circle[radius=1mm]; \foreach \x/\y in {0/1, -1/0, 0/-1} \draw[dotted] (0,0) -- (\x,\y); \foreach \x/\y in {1/1, 2/0, 1/-1} \draw[dotted] (1,0) -- (\x,\y); \end{scope} \end{tikzpicture} \end{center} \caption{The probability that $0\in G$ is contained in a cluster of size~$2$ is $2d\Psite(1)^2\Psite(0)^{4d-2}$ in site percolation and $2d\Pedge(1)\Pedge(0)^{4d-2}$ in edge percolation. Illustration for $d=2$.} \label{fig:clustersize2} \end{figure}% That the IDS is discontinuous, can be seen as follows. The percolation graph splits into its connected components, the so-called \emph{clusters}. More precisely, the clusters of a graph are the equivalence classes of the minimal equivalence relation for which neighbors are equivalent. The event that $0\in G$ is contained in a finite cluster has positive probability, see \cref{fig:clustersize2}. A finite cluster of size~$s\in\N$ supports~$s$ eigenfunctions of~$H_\omega$. In particular, the constant function with value $s^{-1/2}$ is an $\ell^2$-normalized eigenfunction of~$H_\omega$ with eigenvalue~$0$. On the event that~$0$ is contained in a finite cluster of size~$s$, we have $\spr{\delta_0}{\ifu{\Ioc{-\infty}0}(H_\omega)\delta_0}=s^{-1}$. Thus \begin{equation*} N(0) =\E[\spr{\delta_0}{\ifu{\Ioc{-\infty}0}(H_\omega)\delta_0}] \ge\sum_{s\in\N}s^{-1}\Prob(\text{the cluster of~$0$ has size~$s$}) >0\text, \end{equation*} and since $N(E)=0$ for all $E<0$, we verified that~$N$ is discontinuous. Note that there can be compactly supported eigenfunctions of~$H_\omega$ on infinite clusters. To construct an example, consider a finite symmetric cluster like $\{-e_1,0,e_1\}\subseteq G$. The symmetry that flips the sign of~$e_1$ commutes with the Laplace operator restricted to this cluster. Therefore, there are antisymmetric eigenfunctions like $(-1,0,1)/\sqrt2$, which have to vanish on each fixed point of the symmetry. Now connect the finite symmetric cluster to an infinite cluster with edges only touching the fixed points. For more details see \cite{ChayesCF-85a,Veselic-05b}. \subsubsection{The Anderson model on a percolation graph} \label{sec:AndPerc} We will now combine the random kinetic energy of the percolation Hamiltonian with the random potential energy of the Anderson model. A suitable set of colors is $\cA=\cA_0\times\{0,1\}$ for site and $\cA=\cA_0\times\{0,1\}^d$ for edge percolation, where the bounded set~$\cA_0\subseteq\R$ contains the values of the random potential. Accordingly, the probability space is \begin{equation*} (\Omega,\cB,\Prob):= \begin{cases} (\cA,\Borel(\cA),\Prob_0\tensor\Psite)^{\tensor G}&\text{or}\\ (\cA,\Borel(\cA),\Prob_0\tensor\Pedge^{\tensor d})^{\tensor G}\text. \end{cases} \end{equation*} For each $v\in G$, the random parameter $\omega=(\omega_v)_{v\in G}\in\Omega$ has a coordinate $\omega_v=(\omega'_v,\omega''_v)$ with $\omega'_v\in\cA_0$ and either $\omega''_v\in\{0,1\}$ or $\omega''_v\in\{0,1\}^d$. The random potential~$V\from G\to\cA_0\subseteq\R$ is defined by \begin{equation V_\omega(v):=V(\omega,v):=\omega'_v\text. \end{equation} Analogously to the construction in \cref{sec:siteedge}, we define the site percolation graph as the graph induced by the Caley graph of~$G$ on the vertex set \begin{equation*} \cV_\omega:=\{v\in G\mid\omega''_v=1\}\text. \end{equation*} The edge percolation graph has edge set \begin{equation*} \cE_\omega:=\Union\nolimits_{j=1}^d \bigl\{\{v,v+e_j\}\bigm|(\omega''_v)_j=1\bigr\} \end{equation*} and vertex set given by~\eqref{eq:verticesInEdgePercolation}. The Laplace operator on the percolation graph is given by~\eqref{eq:laplace}. The Hamiltonians $H_\omega\from\ell^2(G)\to\ell^2(G)$ are defined by \begin{equation*} (H_\omega\varphi)(v):= \begin{cases} -(\Laplace_\omega\varphi)(v)+V_\omega(v) &\text{if $v\in\cV_\omega$ and}\\ \alpha\varphi(v) &\text{if $v\in G\setminus\cV_\omega$} \end{cases} \end{equation*} with $\alpha>2d+\sup\cA_0$. Again, the spectrum $\sigma(H_\omega)$ will have a component~$\{\alpha\}$ and a disjoint remainder, which is the part we are most interested in. We have again an equivariant group action, namely for all $g\in G$, \begin{equation*} H_{\tau_g\omega}\circ U_g =U_g\circ H_{\tau_g\omega}\text. \end{equation*} If the potential has a probability density, the randomness will smooth out at least some discontinuities of the IDS, see \cite{Veselic-06}. \section{Ergodic theorems for finite colors on~\texorpdfstring{$\Z^d$}{Zd}} \label{sec:finiteA} In \cite[Theorem~2]{LenzMV-08}, the authors prove a quantitative version of the following statement. \begin{Theorem}\label{thm:LMV-ecf} In either of the settings presented in \cref{sec:defsandmodels}, assume that the set~$\cA$ of colors is finite: \begin{equation}\label{eq:Afinite} \setsize\cA<\infty\text. \end{equation} Then, for $\Prob$-almost all $\omega\in\Omega$, \begin{equation*} \lim_{L\to\infty}\norm[\infty] {\setsize{\Lambda_L}^{-1}n(\Lambda_L,\omega)-N} =0\text, \end{equation*} where $N\from\R\to\R$ is the IDS. \end{Theorem} Condition~\eqref{eq:Afinite} is automatically satisfied, unless a potential with infinitely many values is present. In \cite{LenzMV-08}, the authors do not talk about a probability space with many configurations, but rather fix one configuration and argue on some required properties. This properties turn out to be almost surely satisfied in the examples in \cref{sec:defsandmodels}. Nonetheless, it is illuminating to see a deterministic example, which we present next. \subsection{Visible points} \label{sec:visible} A point of~$\Z^d$ is \emph{visible} from the origin, if there is no other point of~$\Z^d$ on the straight line connecting the origin and the point in question, see \cref{fig:visiblepoints_small,fig:visiblepoints_large}. The set of visible points is thus \begin{equation*} \cV:=\{x\in\Z^d\mid\{tx\mid t\in\Icc01\}\isect\Z^d=\{0,x\}\}\text. \end{equation*} Equivalently, a point $x\in\Z^d$ is visible, if $x=0$ or if the greatest common divisor of its coordinates is~$1$. See \cite{BaakeMP-00} for a systematic exploration of~$\cV$. \begin{figure} \pgfmathsetmacro{\X}{10} \pgfmathsetmacro{\Y}{6} \begin{center} \tikzsetnextfilename{visible1}% \begin{tikzpicture} \foreach \y in {0, ..., \Y} { \foreach \x in {0, ..., \X} { \pgfmathsetmacro{\visible}{ gcd(\x,\y) == 1 } \ifthenelse{\visible = 1} {\draw[draw=gray!50] (0,0) -- (\x,\y); \filldraw[black] (\x,\y) circle[radius=0.5mm];} {\filldraw[black] (\x,\y) circle[radius=0.1mm];} } } \filldraw[black] (0,0) circle[radius=0.5mm]; \end{tikzpicture} \end{center} \caption{The visible points in $\Icc0\X\times\Icc0\Y$ with the line connecting them to the origin} \label{fig:visiblepoints_small} \end{figure}% The indicator function $\ifu\cV\from\Z^d\to\cA:=\{0,1\}$ can serve as a potential for a Schr\"odinger operator: \begin{equation*} \Hvis\from\ell^2(\Z^d)\to\ell^2(\Z^d)\textq, \Hvis:=-\Laplace+\ifu\cV\text. \end{equation*} Analogous to \cref{sec:Anderson_Zd}, we can define the eigenvalue counting function \begin{equation*} n\from\cF\to\B\textq, n(\Lambda)(E):=\tr\ifu{\Ioc{-\infty}E}(\Hvis^\Lambda)\text. \end{equation*} The results in \cite{LenzMV-08} imply that the thermodynamic limit \begin{equation*} \lim_{L\to\infty}\setsize{\Lambda_L}^{-1}n(\Lambda_L)(E) \end{equation*} exists uniformly for all $E\in\R$, but, to the best of our knowledge, there is no Pastur--Shubin formula. \begin{figure} \pgfmathsetmacro{\X}{150} \pgfmathsetmacro{\Y}{100} \begin{center} \pgfmathsetmacro{\scaleforpicture}{12/(\X+1)} \tikzsetnextfilename{visible2}% \begin{tikzpicture}[scale=\scaleforpicture] \filldraw (0,0) circle[radius=0.5mm]; \foreach \y in {0, ..., \Y} { \foreach \x in {0, ..., \X} { \pgfmathsetmacro{\visible}{ gcd(\x,\y) == 1 } \ifthenelse{\visible = 1} {\filldraw (\x,\y) circle[radius=0.5mm];} {} } } \end{tikzpicture} \end{center} \caption{The visible points in $\Icc0\X\times\Icc0\Y$} \label{fig:visiblepoints_large} \end{figure}% \subsection{Patterns and frequencies}\label{sec:patternsfrequencies} To understand the mechanism behind \cref{thm:LMV-ecf} better and to motivate the abstract formulation of the next theorem, let us examine the situation of percolation without potential in more detail. We already indicated in \cref{sec:siteedge} with the example of the eigenvalue~$0$, how discontinuities of~$N$ arise. We now want to understand, how the limit of the eigenvalue counting functions on large boxes obtains discontinuities of the same height as the IDS. Let us focus on finite clusters again. In $n(\Lambda_L,\omega)$, the eigenvalues corresponding to eigenfunctions supported on finite clusters are counted at least as often as a copy of their cluster is contained in~$\Lambda_L$. Because we normalize with the volume of the box, $\setsize{\Lambda_L}^{-1}n(\Lambda_L,\omega)$, the important quantity turns out to be the relative frequency with which a finite cluster occurs in a large box~$\Lambda_L$. Since the translations $(\tau_g)_{g\in G}$ act ergodically on~$\Omega$, the relative frequencies converge to the probability of their respective cluster at a fixed location. But exactly these probabilities cause the discontinuities of the IDS. Let us formalize the counting of copies of clusters, or, in the more general setting of a finite set of colors~$\setsize\cA<\infty$, shifts of patterns in boxes, following \cite{LenzMV-08}. Recall that the set of finite subsets of~$G$ is denoted by~$\cF$. A (finite) \emph{pattern} with \emph{domain} $\Lambda\in\cF$ is a map $P\from\Lambda\to\cA$. Since a pattern $P\in\cA^\Lambda$ can be identified with the set $\{\omega\in\Omega\mid\omega_\Lambda=P\}$, we reuse the notation for the group action of~$G$ on~$\Omega$ for the shifts of patterns: \begin{equation*} \tau_g\from\cA^\Lambda\to\cA^{\tilde\tau_g\Lambda}\textq, (\tau_gP)(v):=P(\tilde\tau_gv)\text. \end{equation*} This $G$-action induces an equivalence relation on the set of all finite patterns. We denote the equivalence class of a pattern $P\from\Lambda\to\cA$ by $[P]_G:=\{\tau_gP\mid g\in G\}$. A coloring $\omega\in\Omega=\cA^G$ and a finite subset $\Lambda\in\cF$ define the pattern $\omega_\Lambda:=\omega|_\Lambda\from\Lambda\to\cA$, $\omega_\Lambda(v):=\omega_v$. Similarly, for $\Lambda\subseteq\Lambda'\in\cF$, a pattern~$P'$ with domain~$\Lambda'$ induces the pattern $P'_\Lambda:=P'|_\Lambda\from\Lambda\to\cA$ via $P'_\Lambda(v):=P'(v)$. Given $\Lambda'\in\cF$, set \begin{equation*} P'|_\cF:= \{P'|_\Lambda\mid\Lambda\in\cF,\Lambda\subseteq\Lambda'\}\text. \end{equation*} This set of all finite patterns induced by $P'\in\cA^{\Lambda'}$ is useful to count how often a copy of a pattern $P\in\cA^\Lambda$ occurs in~$P'$: \begin{equation}\label{eq:patterncount} \sharp_PP':=\setsize{[P]_G\isect P'|_\cF}\text. \end{equation} See Figure \ref{fig:patterncount} for a visualisation of this pattern counting function. \begin{figure} \begin{tikzpicture}[scale=0.6] { {\color{black} { \foreach \y in {0,1,...,8}{ \draw (1,\y) --(9,\y); } \foreach \x in {1,2,...,9}{ \draw (\x,0) --(\x,8); } \filldraw[blue] (8,2) circle (4pt); \filldraw[blue] (7,8) circle (4pt); \filldraw[blue] (6,5) circle (4pt); \filldraw[blue] (7,2) circle (4pt); \filldraw[blue] (6,8) circle (4pt); \filldraw[red] (1,1) circle (4pt); \filldraw[green] (1,2) circle (4pt); \filldraw[green] (1,3) circle (4pt); \filldraw[red] (1,4) circle (4pt); \filldraw[red] (1,5) circle (4pt); \filldraw[red] (1,6) circle (4pt); \filldraw[green] (1,7) circle (4pt); \filldraw[green] (1,8) circle (4pt); \filldraw[red] (2,1) circle (4pt); \filldraw[green] (2,1) circle (4pt); \filldraw[red] (2,2) circle (4pt); \filldraw[green] (2,3) circle (4pt); \filldraw[blue] (2,4) circle (4pt); \filldraw[blue] (2,5) circle (4pt); \filldraw[red] (2,6) circle (4pt); \filldraw[red] (2,7) circle (4pt); \filldraw[green] (2,8) circle (4pt); \filldraw[green] (3,1) circle (4pt); \filldraw[red] (3,2) circle (4pt); \filldraw[red] (3,3) circle (4pt); \filldraw[red] (3,4) circle (4pt); \filldraw[green] (3,5) circle (4pt); \filldraw[red] (3,6) circle (4pt); \filldraw[blue] (3,7) circle (4pt); \filldraw[green] (3,8) circle (4pt); \filldraw[blue] (4,1) circle (4pt); \filldraw[green] (4,2) circle (4pt); \filldraw[green] (4,3) circle (4pt); \filldraw[green] (4,4) circle (4pt); \filldraw[blue] (4,5) circle (4pt); \filldraw[blue] (4,6) circle (4pt); \filldraw[blue] (4,7) circle (4pt); \filldraw[blue] (4,8) circle (4pt); \filldraw[red] (5,1) circle (4pt); \filldraw[blue] (5,2) circle (4pt); \filldraw[red] (5,3) circle (4pt); \filldraw[green] (5,4) circle (4pt); \filldraw[green] (5,5) circle (4pt); \filldraw[blue] (5,6) circle (4pt); \filldraw[blue] (5,7) circle (4pt); \filldraw[blue] (5,8) circle (4pt); \filldraw[red] (6,1) circle (4pt); \filldraw[red] (6,2) circle (4pt); \filldraw[green] (6,3) circle (4pt); \filldraw[red] (6,4) circle (4pt); \filldraw[green] (6,5) circle (4pt); \filldraw[red] (6,6) circle (4pt); \filldraw[red] (6,7) circle (4pt); \filldraw[green] (6,8) circle (4pt); \filldraw[green] (7,1) circle (4pt); \filldraw[blue] (7,2) circle (4pt); \filldraw[red] (7,3) circle (4pt); \filldraw[red] (7,4) circle (4pt); \filldraw[green] (7,5) circle (4pt); \filldraw[red] (7,6) circle (4pt); \filldraw[green] (7,7) circle (4pt); \filldraw[green] (7,8) circle (4pt); \filldraw[green] (8,1) circle (4pt); \filldraw[blue] (8,2) circle (4pt); \filldraw[green] (8,3) circle (4pt); \filldraw[green] (8,4) circle (4pt); \filldraw[blue] (8,5) circle (4pt); \filldraw[blue] (8,6) circle (4pt); \filldraw[red] (8,7) circle (4pt); \filldraw[green] (8,8) circle (4pt); \filldraw[red] (9,1) circle (4pt); \filldraw[blue] (9,2) circle (4pt); \filldraw[red] (9,3) circle (4pt); \filldraw[green] (9,4) circle (4pt); \filldraw[blue] (9,5) circle (4pt); \filldraw[blue] (9,6) circle (4pt); \filldraw[blue] (9,7) circle (4pt); \filldraw[red] (9,8) circle (4pt); \filldraw[blue] (1,0) circle (4pt); \filldraw[blue] (2,0) circle (4pt); \filldraw[red] (3,0) circle (4pt); \filldraw[red] (4,0) circle (4pt); \filldraw[green] (5,0) circle (4pt); \filldraw[blue] (6,0) circle (4pt); \filldraw[green] (7,0) circle (4pt); \filldraw[red] (8,0) circle (4pt); \filldraw[green] (9,0) circle (4pt); } } } \draw[rounded corners=0.1cm] (4.7,4.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \draw[rounded corners=0.1cm] (0.7,2.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \draw[rounded corners=0.1cm] (6.7,0.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \draw[rounded corners=0.1cm] (0.7,7.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \draw[rounded corners=0.1cm] (3.7,3.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \draw[rounded corners=0.1cm] (7.7,3.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \draw[rounded corners=0.1cm] (6.7,7.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \end{tikzpicture} \caption{The figure shows a pattern $P'$ defined on a square $Q$ of side length~$9$. The marked $2\times2$ squares highlight the $7$ copies of a pattern $P$ in $P'$. Thus, we count $\sharp_PP'=7$. Obviously, to obtian a meaningful proportion of the occurrences of $P$ in $P'$, an appropriate normalization term is the size of the domain of $P'$, i.\,e.~$\tfrac{\sharp_PP'}{\setsize Q}=\tfrac{7}{81}$.} \label{fig:patterncount} \end{figure} In this notation and for the probabilistic models from \cref{sec:defsandmodels}, the ergodic theorem for $\Z^d$-actions, see \cite{Keller-98}, states that, for all $\Lambda\in\cF$, $P\in\cA^\Lambda$, and $\Prob$-almost all $\omega\in\Omega$, \begin{equation*} \setsize{\Lambda_L}^{-1}\sharp_P(\omega_{\Lambda_L}) \xto{L\to\infty}\Prob\{\omega\in\Omega\mid\omega_\Lambda=P\}\text. \end{equation*} The existence of the limit $\lim_{L\to\infty}\setsize{\Lambda_L}^{-1}\sharp_P(\cV|_{\Lambda_L})$ for all patterns $P\in\cA^\Lambda$ with $\Lambda\in\cF$ in the setting of \cref{sec:visible} is shown in \cite{BaakeMP-00} with different methods. See \cref{fig:visiblepoints_large} for an optical impression. Another feature to extract from the percolation example is the following. By restricting the Hamiltonian to finite boxes, we modify some clusters so that some part of their boundary is more ``straight''. As a consequence, the clusters with one ``straight'' boundary are over-represented in the sample. Of course, as the boxes grow larger and larger, their surface increases as well. Fortunately, the even faster growth of the volume of the boxes makes the boundary negligible, or more precisely: the proportion of the surface to the volume vanishes in the limit. By this mechanism, the surplus of clusters with artificial straight boundary becomes negligible. To formalize the notions above, let us introduce the \emph{$r$-boundary} of~$\Lambda\in\cF$ for $r>0$ as \begin{equation}\label{eq:rboundary} \partial^r\Lambda:=\{x\in G\setminus\Lambda\mid\dist(x,\Lambda)\le r\} \union\{v\in\Lambda\mid\dist(v,G\setminus\Lambda)\le r\}\text. \end{equation} The distance~$\dist$ is the length of the shortest path in the Cayley graph, or, for $G=\Z^d$, the distance induced by the $1$-norm. A sequence $(Q_j)_{j\in\N}$ of finite sets $Q_j\in\cF$ is called a \emph{F\o{}lner sequence}, if, for all $r>0$, \begin{equation}\label{eq:folner} \lim_{j\to\infty}\frac{\setsize{\partial^rQ_j}}{\setsize{Q_j}} =0\text. \end{equation} Another common name for F\o{}lner sequences is \emph{van Hove sequence}. The finite boxes $\Lambda_L:=\Ico0L^d\isect\Z^d\in\cF$, $L\in\N$, form an example of a F\o{}lner sequence, because $\setsize{\partial^r\Lambda_L} =(L+2\floor r)^d-(L-2\floor r)^d =4\floor r\sum_{k=0}^{d-1}(L-2\floor r)^k(L-2\floor r)^{d-1+k} \le4dr(L+2r)^{d-1}$ is of the order of~$L^{d-1}$, while $\setsize{\Lambda_L}=L^d$. If, for a pattern $P\in\cA^\Lambda$, an $\omega\in\Omega$, and a F\o{}lner sequence $(Q_j)_{j\in\N}$, the limit \begin{equation*} \nu_P:=\lim_{j\to\infty}\frac{\sharp_P(\omega_{Q_j})}{\setsize{Q_j}} \end{equation*} exists, it is called the \emph{frequency of~$P$ along $(Q_j)_{j\in\N}$}. See Figure \ref{fig:patterncount} for an illustration of one element of this sequence. Note that the existence of $\nu_P$, for all patterns $P$ with finite domain, can be shown almost surely for all our examples from \cref{sec:defsandmodels}. The eigenvalue counting functions are in a certain sense local enough to reflect the effect of small boundary compared to the volume. The following notions encapsulate the crucial properties. \begin{Definition} A map $b\from\cF\to\Ico0\infty$ is called \emph{boundary term}, if \begin{itemize}[nosep] \item $b$ is invariant under~$G$: $b(\Lambda)=b(\tilde\tau_g\Lambda)$ for all $g\in G$, \item $\lim_{j\to\infty}\setsize{Q_j}^{-1}b(Q_j)=0$ for all F\o{}lner sequences $(Q_j)_{j\in\N}$, \item $b$ is bounded, i.\,e.\ its \emph{bound} $D_b:=\sup\{\setsize\Lambda^{-1}b(\Lambda)\mid \Lambda\in\cF,\Lambda\ne\emptyset\}<\infty$ is finite, and \item for $\Lambda,\Lambda'\in\cF$ we have $b(\Lambda\union\Lambda')\le b(\Lambda)+b(\Lambda')$, $b(\Lambda\isect\Lambda')\le b(\Lambda)+b(\Lambda')$, and $b(\Lambda\setminus\Lambda')\le b(\Lambda)+b(\Lambda')$. \end{itemize} \end{Definition} \begin{figure} \begin{center} \tikzsetnextfilename{boundaryrules}% \begin{tikzpicture} \draw [very thick] (0,0) -- (0.75,0.5) arc [start angle=30, end angle=335, x radius=2cm, y radius=1cm] -- cycle; \draw (0,0) -- (2,-1) -- (1.5,1) -- cycle; \draw (1,0) arc [ start angle=0 , end angle=360 , x radius=2cm , y radius=1cm ] -- cycle; \node at (-1.3,0) {$\Lambda\setminus\Lambda'$}; \node at (-2.9,.8) {$\Lambda$}; \node at (2,.6) {$\Lambda'$}; \end{tikzpicture} \end{center} \caption{The requirement $b(\Lambda\setminus\Lambda')\le b(\Lambda)+b(\Lambda')$ is motivated by the fact that the boundary of $\Lambda\setminus\Lambda'$ is a subset of the boundary of~$\Lambda$ united with the boundary of~$\Lambda'$.} \label{fig:pacman} \end{figure} The last property is natural to require, as illustrated in \cref{fig:pacman}, but in fact only necessary when used with quasi tilings, see \cref{sec:general}. \begin{Definition}\label{def:almost-additive} Consider a Banach space $(\B,\norm\argmt)$, a field $F\from\cF\to\B$, and a coloring $\omega\in\cA^G$. \begin{itemize} \item The field~$F$ is \emph{almost additive}, if there exists a boundary term~$b$ such that, for all $n\in\N$, disjoint sets $\Lambda_1,\dotsc\Lambda_n\in\cF$, and $\Lambda:=\Union_{k=1}^n\Lambda_k$, it holds true that \begin{equation*} \norm{F(\Lambda)-\sum_{k=1}^nF(\Lambda_k)} \le\sum_{k=1}^nb(\Lambda_k)\text. \end{equation*} \item The field~$F$ is \emph{$\omega$-invariant}, if for all $\Lambda,\Lambda'\in\cF$ such that the patterns $\omega_\Lambda$ and $\omega_{\Lambda'}$ are $G$-equivalent, we have \begin{equation*} F(\Lambda)=F(\Lambda')\text. \end{equation*} \end{itemize} \end{Definition} To an $\omega$-invariant field $F\from\cF\to\B$, the \emph{pattern function} $\tilde F\from\Union_{\Lambda\in\cF}\cA^\Lambda\to\B$ with \begin{equation*} \tilde F(P):= \begin{cases} F(\Lambda)&\text{if there is a set $\Lambda\in\cF$ with $P\in[\omega_\Lambda]_G$, and}\\ 0&\text{otherwise} \end{cases} \end{equation*} is well defined. Every almost additive field and $\omega$-invariant~$f$ is \emph{bounded} and has a \emph{bound} $C_F$ in the following sense: \begin{equation*} C_F:= \sup\{\setsize\Lambda^{-1}\norm{F(\Lambda)}\mid \Lambda\in\cF\setminus\{\emptyset\}\}<\infty\text. \end{equation*} Indeed, since~$\cA$ is finite, \begin{equation*} C_F \le\sup_{\Lambda\ne\emptyset}\frac1{\setsize\Lambda}\sum_{v\in\Lambda} \bigl(\norm{F(\{v\})}+b(\{v\})\bigr) \le\max_{a\in\cA^{\{0\}}}\tilde F(a)+b(\{0\}) <\infty \end{equation*} The eigenvalue counting functions are good examples for these notions. \begin{Proposition}[{\cite[Proposition~2]{LenzMV-08}}] For almost all $\omega\in\Omega$ and with respect to the Banach space $(\B,\norm[\infty]\argmt)$ of right continuous bounded $\R$-valued functions on~$\R$, the eigenvalue counting functions~$n$ of the models given in \cref{sec:defsandmodels} are $\omega$-invariant, almost additive with boundary term $b(\Lambda):=4\setsize{\partial^1\Lambda}$, $D_b\le8d+4$, and bounded with bound~$C_{n(\argmt,\omega)}=1$. \end{Proposition} The following theorem thus applies to the eigenvalue counting functions of \cref{sec:defsandmodels}. We emphasize that the error estimates imply that the IDS is approximated by the eigenvalue counting functions uniformly in the energy. \begin{Theorem}[{\cite[Theorem~1]{LenzMV-08}}]\label{thm:LenzMV} Let $\cA$ be a finite set, $\omega\in\cA^G$ a coloring, $(\B,\norm\argmt)$ an arbitrary Banach space, $(Q_j)_{j\in\N}$ a F\o{}lner sequence, $F\from\cF\to\B$ a bounded, $\omega$-invariant, and almost additive field with bound~$C_F$, pattern function $\tilde F$, boundary term~$b$, and bound~$D_b$ of~$b$. \par Assume that for every finite pattern~$P\from\Lambda\to\cA$, $\Lambda\in\cF$, the frequency $\nu_P:=\lim_{j\to\infty}\setsize{Q_j}^{-1} \sharp_P(\omega|_{Q_j})$ exists. Then, the limits \begin{equation}\label{eq:Fbar-F-Ftilde} \Fbar :=\lim_{j\to\infty}\frac{F(Q_j)}{\setsize{Q_j}} =\lim_{L\to\infty}\sum_{P\in\cA^{\Lambda_L}} \nu_P\frac{\tilde F(P)}{\setsize{\Lambda_L}} \end{equation} exist and are equal. Moreover, for all $j,L\in\N$, the bounds \begin{equation}\label{eq:Fbar-Ftilde} \biggnorm{\Fbar-\sum_{P\in\cA^{\Lambda_L}} \nu_P\frac{\tilde F(P)}{\setsize{\Lambda_L}}} \le\frac{b(\Lambda_L)}{\setsize{\Lambda_L}} \end{equation} and \begin{multline}\label{eq:Ftilde-F} \biggnorm{\frac{F(Q_j)}{\setsize{Q_j}} -\sum_{P\in\cA^{\Lambda_L}} \nu_P\frac{\tilde F(P)}{\setsize{\Lambda_L}}}\\ \le\frac{b(\Lambda_L)}{\setsize{\Lambda_L}} +(C_F+D_b)\frac{\setsize{\partial^LQ_j}}{\setsize{Q_j}} +C_F\sum_{P\in\cA^{\Lambda_L}} \Bigabs{\frac{\sharp_P(\omega|_{Q_j})}{\setsize{Q_j}}-\nu_P} \end{multline} hold true. \end{Theorem} The error estimates are the crucial part of the theorem. We note that there are two length scales, indexed by~$j$ and~$L$, in the approximation of the limiting object~$\Fbar$. They correspond to two stages of approximation in the proof. In the first step,~$\Fbar$ is compared to the weighted average of the contributions of the patterns on~$\Lambda_L$, the weights being the frequencies of the patterns. According to~\eqref{eq:Fbar-Ftilde}, the patterns on~$\Lambda_L$ capture the behavior of the limiting function up to an error of size $\setsize{\Lambda_L}^{-1}b(\Lambda_L)$. We mentioned above, that $(\Lambda_L)_L$ is a F\o{}lner sequence, so by choosing~$L$ large enough, we can make this error term small. \Cref{eq:Ftilde-F} addresses the problem that we used the limiting frequencies in~\eqref{eq:Fbar-F-Ftilde} and~\eqref{eq:Fbar-Ftilde} instead of the actual number of occurrences of patterns in the finite region~$Q_j$. This error bound requires us to choose~$j$ so large that the empirical frequencies of all patterns on~$\Lambda_L$ get close to their actual frequencies. To recapitulate: First, we have to choose~$L$ large enough to make the patterns on~$\Lambda_L$ meaningful for the limit. Then we have to choose~$j$ large enough such that the patterns on~$\Lambda_L$ are actually observed in proportions that are close to the asymptotic frequencies. In the random Schr\"odinger operator settings from \cref{sec:defsandmodels}, the second error is controlled by the randomness. In fact, the ergodic theorem predicts that the relative number of occurrences of a pattern converges to its frequencies almost surely. It follows from the theory of large deviations that the probability of a fixed difference between the two is exponentially small for large samples, that is, for large~$j$. The frequencies $(\nu_P)_{P\in\cA^{\Lambda_L}}$ as well as their empirical counterparts $(\setsize{Q_j}^{-1}\sharp_P(\omega|_{Q_j}))_{P\in\cA^{\Lambda_L}}$ can be interpreted as a probability mass function on~$\cA^{\Lambda_L}$. The $1$-norm of their difference in~\eqref{eq:Ftilde-F} is also known as the \emph{total variation norm}\label{tv_mentioned} of the corresponding probability measures. In the remainder of this note, we present two generalizations which correspond to the two types of errors described above. The main property of the group~$G=\Z^d$ used in \cref{thm:LenzMV} is that it hosts F\o{}lner sequences. That this is actually the crucial property necessary for this proof is made explicit by generalizing the result to amenable groups, which are characterized by the existence of a F\o{}lner sequence, see \cref{sec:amenable}. The second generalization concerns the finiteness of the set of colors~$\cA$. The main obstacle to overcome this restriction is the probabilistic error. The total variation norm of the difference of the distribution and the empirical measure of a random variable does not converge to zero for continuous random variables. In order to allow infinitely many colors, we will be forced to exploit more properties of the eigenvalue counting functions, see \cref{sec:infiniteA}. \section{Ergodic theorems for finite colors on amenable groups}\label{sec:amenable} In this section we discuss how the above ideas generalize to less restricted geometries. In particular, we present Banach space-valued ergodic theorems for Cayley graphs generated by amenable groups. Let us emphasize that the groups considered in this paper will always be finitely generated and therefore countable. An amenable group is by definition a group containing subsets with an arbitrary small ratio between boundary and volume. It is well known, that amenability is equivalent to the existence of a F\o{}lner sequence. It turns out that even though amenability is the natural condition to generalize the geometry of $\Z^d$, there is an additional requirement needed to almost directly implement the $\Z^d$-methods to this setting. Here we are speaking about a so-called tiling condition, namely the condition that there exists a F\o{}lner sequence consisting of monotiles. Obviously, in~$\Z^d$ a sequence of cubes serves as such a sequence, since for each~$j$ the group~$\Z^d$ can be tiled with cubes of side length~$j$. For an arbitrary amenable group it is not known whether such a sequence exists or not. Therefore, this section is structured in a first part discussing the monotile situation and a second more involved part dealing with the general amenable groups using the technique of quasi tilings. We proceed with some definitions which generalize the notion of previous sections. Given a finitely generated group $G$ with a finite and symmetric generating set $S\subseteq G$, i.\,e.\ $G=\langle S\rangle$, $\setsize S<\infty$ and $S=S^{-1}\not\ni \mbox{id}$, the corresponding Cayley graph~$\Gamma$ has vertex set~$G$ and two vertices $v,w\in G$ are connected if and only if $vs=w$ for some $s\in S$. The induced graph distance, sometimes called word metric, is denoted by~$\dist$. A sequence $(Q_j)$ of finite subsets of~$G$ is called a F\o{}lner sequence if $\setsize{\partial^r Q_j}/\setsize{Q_j}\to 0$ as $j\to\infty$ for all $r>0$, see \eqref{eq:folner}. Here the $r$-boundary is given as in \eqref{eq:rboundary}. For an amenable group~$G$, the introduction of the physical models of \cref{sec:Anderson_Zd,sec:percolation} works completely analogous. Due to the fact that for general groups the group action is usually written as a multiplication, the only difference is that the group action $\tilde\tau$ is defined here as \begin{equation}\label{def:tau} \tilde\tau \from G\times G \to G ,\quad (g,v)\mapsto \tilde\tau_gv := vg^{-1}. \end{equation} As in \eqref{eq:schroedinger} the Sch\"odinger operator in the Anderson model is given by \[ H_\omega:=-\Laplace+V_\omega\from\ell^2(G)\to\ell^2(G) \] with Laplace operator $\Delta\from\ell^2 (G)\to\ell^2(G)$ as in~\eqref{eq:laplaceop} and a (random) potential $V\from\Omega\times G\to \cA$ as in~\eqref{eq:randompotential}. As before we consider the probability space $(\Omega,\cB,\Prob) :=(\cA^G,\Borel(A)^{\tensor G},\Prob_0^{\tensor G})$ with a finite set $\cA$. An element $\omega\in\Omega$ is then interpreted as a coloring of the elements of the group using the finite set of colors~$\cA$. Besides this, also the definition of site and edge percolation does not depend on the $\Z^d$-structure and generalizes straightforwardly to the case of amenable groups. Thereby the Anderson model on percolation Cayley graphs over finitely generated amenable groups is well-defined. \subsection{Symmetrical tiling condition} As mentioned before an additional condition is needed in order to implement the methods of~$\Z^d$ to amenable groups. This so-called symmetric tiling condition is formulated as follows. \begin{Definition}\label{def:STamenable} Let $G$ be a group. A subset $\Lambda\subseteq G$ \emph{symmetrically tiles} $G$ if there exists a set $T$ such that \begin{enumerate}[(i)] \item $T=T^{-1}$, \item $\Lambda T = G$, \item $\{\Lambda t \mid t\in T \}$ are pairwise disjoint. \end{enumerate} An amenable $G$ is said to satisfy the \emph{symmetric tiling condition}, if there exists a F\o{}lner sequence $(\Lambda_L)$ such that each $\Lambda_L$, $L\in \N$, symmetrically tiles $G$. In this situation we call $G$ an \emph{ST-amenable group}. \end{Definition} Note that here $\Lambda T:=\{xt \mid x\in \Lambda, t\in T\}$ and $\Lambda t=\{xt\mid x\in \Lambda\}$. Let us briefly remark on this condition. Krieger proved in \cite{Krieger-07} based on work of Weiss \cite{Weiss-01} that an amenable group satisfies the symmetrical tiling condition if it is residually finite. For instance, each group of polynomial volume growth is nilpotent (by Gromov's theorem) and thus residually finite. Since it is also of subexponential growth, it is amenable, too, and hence ST-amenable. An intensively studied and slightly more general condition than the one stated in \cref{def:STamenable} can be obtained when not assuming the symmetry (i). In this situation a set $Q$ satisfying (ii) and (iii) is usually referred to as a \emph{monotile} and a group $G$ containing a F\o{}lner sequence consisting only of monotiles is called \emph{monotileable}. Let us remark that it is still not known if there exists an amenable group which is not monotileable. In the situation of ST-amenable groups one can prove the following: \begin{Theorem}[{\cite{LenzSV-10,LenzSV-12}}]\label{thm:LenzSV} Let $G$ be a finitely generated ST-amenable group. Let~$(Q_j)$ and~$(\Lambda_L)$ be F\o{}lner sequences such that each $\Lambda_L$, $L\in\N$ symmetrically tiles~$G$. Let $\cA$ be finite and $\omega\in \cA^G$ be given. Moreover, let $(\B,\norm\argmt)$ be a Banach space and $F\from\cF\to\B$ a bounded, $\omega$-invariant and almost additive field with bound~$C_F$, pattern function $\tilde F$, boundary term~$b$, and bound~$D_b$ of~$b$. As before we use the notation $\cF=\{A\subseteq G\mid A \text{ finite}\}$. Assume that for every finite pattern~$P\from\Lambda\to\cA$, $\Lambda\in\cF$, the frequency $\nu_P:=\lim_{j\to\infty}\setsize{Q_j}^{-1} \sharp_P(\omega|_{Q_j})$ exists. Then, the limits \begin{equation}\label{eq:Fbar-F-Ftilde.amenable} \Fbar :=\lim_{j\to\infty}\frac{F(Q_j)}{\setsize{Q_j}} =\lim_{L\to\infty}\sum_{P\in\cA^{\Lambda_L}} \nu_P\frac{\tilde F(P)}{\setsize{\Lambda_L}} \end{equation} exist and are equal. Moreover, for all $j,L\in\N$, the bound \begin{multline}\label{eq:Ftilde-F.amenable} \biggnorm{\frac{F(Q_j)}{\setsize{Q_j}} -\sum_{P\in\cA^{\Lambda_L}} \nu_P\frac{\tilde F(P)}{\setsize{\Lambda_L}}}\\ \le\frac{b(\Lambda_L)}{\setsize{\Lambda_L}} +(C_F+D_b)\frac{\setsize{\partial^{\diam(\Lambda_L)}Q_j}}{\setsize{Q_j}} +C_F\sum_{P\in\cA^{\Lambda_L}} \Bigabs{\frac{\sharp_P(\omega|_{Q_j})}{\setsize{Q_j}}-\nu_P} \end{multline} holds true. Here, $\diam(\Lambda_L)=\max\{\dist(x,y)\mid x,y\in \Lambda_L\}$ is the diameter of the finite set~$\Lambda_L$. \end{Theorem} When comparing \cref{thm:LenzMV} and \cref{thm:LenzSV}, it turns out that the transition from $\Z^d$ to ST-amenable groups does not imply a quantitative difference in the strength of the result. In particular, the only difference is that in the $\Z^d$ setting $(\Lambda_L)$ is a sequence of cubes with side length $L$ (which naturally tiles $\Z^d$) and in the ST-amenable group setting $(\Lambda_L)$ is a F\o{}lner sequence assumed to be symmetrically tiling. In the error bound this results in the substitution of the side length $L$ of a cube by the diameter of $\Lambda_L$. In the following we give a sufficient condition for the existence of the frequencies in the situation of a random coloring. Here we need the notion of a \emph{tempered} F\o{}lner sequence, which is a F\o{}lner $(Q_j)$ sequence with the additional property that there is some $C>0$ such that \[ \Biggl|\bigcup_{k<n}Q_k^{-1}Q_j\Biggr| \leq C \setsize{Q_j} \] for all $j\in\N$. Each F\o{}lner sequence has a tempered subsequence, see \cite{Lindenstrauss-01}. \begin{Theorem} Let~$G$ be a finitely generated amenable group, $\cA$ some finite set and let~$\mu$ be a probability measure on $(\Omega,\cB) =(\cA^G,\Borel(A)^{\tensor G})$. We assume that the action $\tilde\tau$ given via~\eqref{def:tau} of~$G$ on~$\Omega$ is measure preserving and ergodic w.\,r.\,t.~$\mu$. Then, for any tempered F\o{}lner sequence $(Q_j)$, there exists a event $\tilde\Omega$ of full measure, such that the limit \begin{equation*} \lim_{j\to\infty}\frac{\sharp_P(\omega|_{Q_j})}{\abs{Q_j}} \end{equation*} exists for all patterns $P\in\bigcup_{Q\in\cF}\cA^{Q}$ and all $\omega\in\tilde\Omega$. Moreover, the limit is deterministic in the sense that it is independent of the specific choice of~$\omega\in\tilde\Omega$. \end{Theorem} The above result is a direct consequence (see for instance \cite{Schwarzenberger-08}) of Lindenstrauss ergodic theorem \cite{Lindenstrauss-01}. We omit the precise definitions of measure preserving and ergodic action. However, we want to emphasize that in the particular case where $\mu=\Prob_0^{\tensor G}$ is a product measure, the assumptions on~$\tilde\tau$ are met. Thus, in the percolation setting of \cref{sec:percolation} one obtains that almost all configurations satisfy the assumption of well-defined frequencies. \subsection{General amenable groups}\label{sec:general} As outlined before, it is still not known if each amenable group satisfies the symmetrical tiling condition. Roughly speaking, in the previous sections this tiling condition is the crucial tool to mediate between the two F\o{}lner sequence $(\Lambda_L)$ and $(Q_j)$. Thus, when considering amenable groups without an additional tiling assumption the situation is far more challenging. A way to overcome this lack is to apply the theory of $\epsilon$-quasi tilings developed by Ornstein and Weiss \cite{OrnsteinW-87} in 1987, see also \cite{PogorzelskiS-16} for the quantitative estimates used is the present setting. The key idea here is to soften the condition of a \emph{perfect} tiling with copies of \emph{one} set taken from a F\o{}lner sequence, and rather \begin{itemize} \item use finitely many different sets of the F\o{}lner sequence, and \item allow imperfectness in the tiling (in sense of small \emph{overlaps} and \emph{uncovered areas}). \end{itemize} More precisely we use the following definition: \begin{Definition}\label{def:quasitiling} Let $G$ be a finitely generated group, $Q\subseteq G$ a finite set and $\epsilon>0$. We say that $K_1,\dots,K_N\subseteq G$ with center sets $T_1,\dots, T_N$, short $(K_i,T_i)_{i=1}^N$, are an \emph{$\epsilon$-quasi tiling} of~$Q$ if \begin{enumerate}[(i)] \item the sets $K_iT_i$, $i\in\{1,\dotsc,N\}$ are pairwise disjoint and subsets of $Q$; \item $\setsize{Q\setminus \bigcup_{i=1}^N K_iT_i} \leq 2\epsilon \setsize{Q}$; \item there are subsets $\mathring{K_i}\subseteq K_i$, $i\in\{1,\dotsc,N\}$, such that \begin{itemize} \item for each $i$ the sets $\mathring{K_i} t$, $t\in T_i$ are pairwise disjoint \item $\setsize{K_i\setminus \mathring{K_i}}\leq \epsilon \setsize{K_i}$ \end{itemize} \end{enumerate} \end{Definition} For technical reasons, the sequence $(\Lambda_L)$ which will provide the F\o{}lner sets $K_i$ to quasi tile the group is assumed to be \emph{nested}, i.\,e.\ for each $L\in\N$ we have $\id \in \Lambda_L\subseteq \Lambda_{L+1}$. Note that, starting from an arbitrary F\o{}lner sequence, one can construct a nested F\o{}lner sequence by translating elements of an appropriate subsequence. In the following it turns out to be convenient to define for given $\epsilon\in(0,1)$ and $i\in\N_0$ the numbers $N(\epsilon)$ and $\eta_i(\epsilon)$ by \begin{equation}\label{eq:def:N,eta} N(\epsilon):=\biggceil{\frac{\ln(\epsilon)}{\ln(1-\epsilon)}} \qtextq{and} \eta_i(\epsilon):=\epsilon(1-\epsilon)^{N(\epsilon)-i} \text. \end{equation} As usual we use the Gau\ss{}ian bracket notation $\ceil b:=\inf\{z\in\Z\mid z\ge b\}$. The following theorem shows that~$N(\epsilon)$ is the number of required shapes~$K_i$ in order to $\epsilon$-quasi tile a set~$Q$. Moreover, for fixed~$i$ the~$\eta_i(\epsilon)$ can be interpreted as the ratio of the points covered by copies of~$K_i$ in the $\epsilon$-quasi tiling. \begin{Theorem}\label{thm:STP} Let~$G$ be a finitely generated amenable group, $(\Lambda_L)$ a nested F{\o}lner sequence, and $\epsilon\in\Ioo0{0.1}$. Then there is a finite and strictly increasing selection of sets $K_i\in\{\Lambda_L\mid L\in\N\}$, $i\in\{1,\dotsc,N(\epsilon)\}$, with the following property: For each F{\o}lner sequence~$(Q_j)$, there exists $j_0(\epsilon)\in\N$ satisfying that for all $j\geq j_0(\epsilon)$ there exists sets~$T_1^j, \ldots, T_{N(\epsilon)}^j$ such that $(K_i,T_i^j)_{i=1}^{N(\epsilon)}$ is an $\epsilon$-quasi tiling of~$Q_j$. Moreover, for all $j\ge j_0(\ve)$ and all $i\in\{1,\dots,N(\epsilon)\}$, the proportion of~$Q_j$ covered by the tile~$K_i$ satisfies \begin{equation}\label{eq:etdensity} \biggabs{\frac{\setsize{K_iT_i^j}}{\setsize{Q_j}}-\eta_i(\ve)} \le\frac{\ve^2}{N(\ve)}\text. \end{equation} \end{Theorem} The proof of \cref{thm:STP} is to be found in \cite{PogorzelskiS-16}. Although \cref{thm:STP} provides all the elements used to formulate the desired ergodic theorem for general amenable groups (\cref{thm:PSerg}), a far more involved result is applied in the proof. More precisely, in the proof of \cref{thm:PSerg} one does not only need \emph{one} possibility to quasi tile a given set $Q_j$ with $K_1,\dots,K_{N(\epsilon)}$, but one rather needs a bunch of \emph{different} possibilities to quasi tile $Q_j$ with $K_1,\dots,K_{N(\epsilon)}$. These different tilings of $Q_j$ need to be chosen such that for (almost) all $g\in Q_j$ the frequency (over different tilings) that it is covered by one $K_i$ is up to a small error $\eta_i(\epsilon)/|K_i|$. In this sense, this covering result is referred to as uniform $\epsilon$-quasi tiling. We refer to \cite{PogorzelskiS-16} for details. Let us formulate the ergodic theorem based on the $\epsilon$-quasi tiling results. \begin{Theorem}\label{thm:PSerg} Assume: \begin{itemize} \item $G$ is a finitely generated amenable group. \item $\cA$ is a finite set and $\omega\in \cA^G$. \item $(Q_j)$ is a F\o{}lner sequence such that the frequency $\nu_P:=\lim_{j\to\infty}\setsize{Q_j}^{-1} \sharp_P(\omega|_{Q_j})$ exists for every finite pattern~$P\from\Lambda\to\cA$, $\Lambda\in\cF$. \item $(\Lambda_L)$ is a nested F\o{}lner sequence. \item For given $\epsilon\in(0,\tfrac1{10})$ the sets $K_i$, $i\in\{1,\dots,N(\epsilon)\}$ are chosen according to \cref{thm:STP}. \item $(\B,\norm\argmt)$ is a Banach space and $F\from\cF\to\B$ is $\omega$-invariant, and almost additive with bound~$C_F$, pattern function $\tilde F$, boundary term~$b$, and bound~$D_b$ of~$b$. \end{itemize} Then, the limits \begin{equation} \Fbar:= \lim_{j\to\infty}\frac{F(Q_j)}{\abs{Q_j}} = \lim_{\substack{\epsilon \searrow 0 \\ \epsilon<0.1}} \sum_{i=1}^{N(\epsilon)} \eta_i(\epsilon) \sum_{P\in \cA^{K_i}} \nu_P \frac{\tilde F(P)}{\abs{K_i}} \end{equation} exist and are equal. Moreover, for given $\epsilon\in(0,\tfrac1{10})$ there exist some $j(\epsilon),r(\epsilon)\in\N$ such that for all $j\geq j(\epsilon)$ the bound \begin{multline}\label{eq:Ftilde-F.amenable.gen} \biggnorm{ \frac{F(Q_j)}{\abs{Q_j}} - \sum_{i=1}^{N(\epsilon)} \eta_i(\epsilon) \sum_{P\in \cA^{K_i}} \nu_P \frac{\tilde F(P)}{\abs{K_i}}}\\ \le 4 \sum_{i=1}^{N(\epsilon)} \eta_i(\epsilon) \frac{b(K_i )}{|K_i|} +(C_F+4D_b)\frac{\abs{\partial^{r(\epsilon)}Q_j}}{\abs{Q_j}}\sum_{i=1}^{N(\epsilon)} \abs{K_i} \\ + C_F \sum_{i=1}^{N(\epsilon)} \eta_i(\epsilon) \sum_{P \in \cA^{K_i}} \left| \frac{\sharp_P(\omega|_{Q_j})}{|Q_j|} - \nu_P \right| + (11C_F+32D_b) \epsilon \end{multline} holds true. \end{Theorem} When comparing the estimate with the one in \cref{thm:LenzSV}, note that the difference \eqref{eq:Ftilde-F.amenable.gen} gets small if one firstly executes the limit ${j\to\infty}$ and afterwards ${\epsilon\searrow 0}$. Here the limit $\epsilon\searrow0$ corresponds to $L\to\infty$ in the previous setting. For a detailed discussion why the error terms tend to zero we refer to \cite{PogorzelskiS-16}. We confine ourselves to the (rough) statement that the first three terms in the estimate \eqref{eq:Ftilde-F.amenable.gen} correspond in this ordering to three terms in \eqref{eq:Ftilde-F.amenable} or \eqref{eq:Ftilde-F}, respectively. \section{Glivenko--Cantelli type theorems}\label{sec:infiniteA} In this section we will consider the situation that the set of colors~$\cA$ is no longer finite. This means that we are leaving a combinatorial setting and relying on a probabilistic framework instead. This has been already introduced for a number of examples in \cref{sec:Anderson_Zd,sec:percolation}. In particular, we will need some independence and monotonicity assumptions. The monotonicity property will allow us to smooth out and regularize certain quantities which we otherwise do not know how to estimate. The independence assumption gives us a tool to describe the existence of frequencies used in Section \ref{sec:finiteA} in a constructive and quantitative manner. We reformulate the notions of \cref{sec:finiteA} involving fields $F\from\cF\to\B$ with values in an arbitrary Banach space~$\B$ for fields of the form $f\from\cF\times\Omega\to\B$ with a probability space~$\Omega$ and the specific Banach space of right continuous functions with $\sup$-norm. For easy distinction, we denote the latter with lowercase letters. Almost additive is assumed to hold true uniformly on~$\Omega$. The property $\omega$-invariance for fixed $\omega\in\Omega$ is split up in equivariance and locality. \begin{Definition}\label{def:admissible} A field $f\from\cF\times\Omega\to\B$ is \begin{itemize \ite \emph{almost additive}, if there is a boundary term $b\from\cF\to\Ico0\infty$ such that for all $\omega\in\Omega$, pairwise disjoint $\Lambda_1,\dots,\Lambda_n\in\cF$, and $\Lambda:=\bigcup_{i=1}^n\Lambda_i$, we have \begin{equation*} \Bignorm{f(\Lambda,\omega)-\sum_{i=1}^nf(\Lambda_i,\omega)} \le\sum_{i=1}^nb(\Lambda_i)\text. \end{equation*} \ite \emph{equivariant}, if for $\Lambda\in\cF$, $g\in G$ and $\omega\in\Omega$ we have $f(\tilde\tau_g\Lambda,\omega)=f(\Lambda,\tau_g\omega)$. \ite \emph{local}, if for all $\Lambda\in\cF$ and $\omega,\omega'\in\Omega$, $\omega_\Lambda=\omega'_\Lambda$ implies $f(\Lambda,\omega)=f(\Lambda,\omega')$. \ite \emph{bounded}, if $ \sup_{\omega\in\Omega}\norm{f(\{\id\},\omega)} <\infty\text. $ \end{itemize} \end{Definition} \subsection{Glivenko--Cantelli theory} To simplify the motivation, we first consider $G=\Z^d$ and assume that the probability measure~$\Prob=\Prob_0^{\tensor G}$ on~$\Omega$ is a product measure. This implies in particular that, for every local field $f\from\cF\times\Omega\to\B$, the random variables $\omega\mapsto f(\{0\},\tau_g\omega)$, $g\in G$, are independent. Here, we used that for local~$f$, in particular, $f(\{0\},\omega)$ depends only on~$\omega_0$ and not on~$\omega_{G\setminus\{0\}}$. To explain the relation of our methods to Glivenko--Cantelli theory, assume for the moment that $f\from\cF\times\Omega\to\B$ is not only almost additive but \emph{exactly} additive. That means that for a finite subset $\Lambda\in\cF$ of~$G$ and a realization of colors $\omega\in\Omega=\cA^G$, we can split $f(\Lambda,\omega)$ without any errors into a sum over singleton sets \begin{equation*} f(\Lambda,\omega) =\sum_{v\in\Lambda}f(\{v\},\omega)\text. \end{equation*} By equivariance, we can rewrite $f(\{v\},\omega)=f(\{0\},\tau_v^{-1}\omega)$. For the special case $\B=\R$, the law of large numbers allows us to calculate the thermodynamic limit \begin{equation*} \lim_{L\to\infty}\frac{f(\Lambda_L,\omega)}{\setsize{\Lambda_L}} =\lim_{L\to\infty}\frac1{\setsize{\Lambda_L}} \sum_{v\in\Lambda_L}f(\{0\},\tau_v^{-1}\omega) =\E[f(\{0\},\argmt)] \end{equation*} $\Prob$-almost surely. Of course, in view of our examples in \cref{sec:defsandmodels}, we are more interested in the Banach space~$\B$ of right continuous functions from~$\R$ to~$\R$ with $\sup$-norm. To head in this direction it is advantageous to lift the point of view to the empirical probability $\widehat\Prob_\Lambda^\omega :=\setsize\Lambda^{-1}\sum_{v\in\Lambda}\delta_{\omega_v}$ and to write integration as dual pair $\spr\argmt\argmt$. In this notation, the average from above can be written as \begin{equation*} \setsize\Lambda^{-1}f(\Lambda,\omega) =\spr{f(\{0\},\argmt)}{\widehat\Prob_\Lambda^\omega} \text. \end{equation*} Here is a special case, where a theorem from classical probability theory helps. Assume that $\cA\in\Borel(\R)$, and let $f\from\cF\times\Omega\to\B$ count the number of random variables in~$\Lambda$ with value less than a given threshold $E\in\R$: \begin{equation*} f(\Lambda,\omega)(E) :=\sum_{v\in\Lambda}\ifu{\Ico{\omega_v}\infty}(E) =\sum_{v\in\Lambda}\ifu{\Ioc{-\infty}E}(\omega_v) =\setsize\Lambda\spr{\ifu{\Ioc{-\infty}E}}{\widehat\Prob_\Lambda^\omega}\text. \end{equation*} This defines an additive, local, and equivariant field, which can be interpreted as the (not normalized) empirical distribution function of the sample~% $(\omega_v)_{v\in\Lambda}$. The thermodynamic limit in~$\B$, i.\,e.\ uniformly, is $\Prob$-almost surely \begin{equation*} \lim_{L\to\infty}\setsize{\Lambda_L}^{-1}f(\Lambda_L,\omega) =\lim_{L\to\infty} \spr{\ifu{\Ioc{-\infty}\argmt}}{\widehat\Prob_\Lambda^\omega} =\spr{\ifu{\Ioc{-\infty}\argmt}}{\Prob_0} =\Prob(\omega_0\le\argmt) \end{equation*} by the theorem of Glivenko and Cantelli: \begin{Theorem}[\cite{Glivenko-33,Cantelli-33}]\label{thm:GC} Let $V_j$, $j\in\N$, be real valued, independent and identically distributed random variables on $(\Omega,\Borel,\Prob)$ and $\widehat\Prob_n:=\frac1n\sum_{j=1}^n\delta_{V_j}$ the corresponding empirical distribution. Then there exists an event $\Omega_{\mathrm{unif}}$ with probability $\Prob(\Omega_{\mathrm{unif}})=1$ such that for all $\omega\in\Omega_{\mathrm{unif}}$: \begin{equation*} \sup_{E\in\R}\abs{\spr{\ifu{\Ioc{-\infty}E}}{\widehat\Prob_n(\omega)-\Prob}} \xto{n\to\infty}0\text. \end{equation*} \end{Theorem} We learn that the choice of this Banach space~$\B$ means to prove the convergence of~$\widehat\Prob_\Lambda^\omega$ to~$\Prob_0$ with respect to a supremum over appropriate test functions. The route pursued in \cref{thm:LenzMV,thm:LenzSV,thm:PSerg} for finite alphabets corresponds to the estimate \begin{equation*} \abs{\spr g{\widehat\Prob_\Lambda^\omega-\Prob_0}} \le\norm[\infty]g\TVnorm{\widehat\Prob_\Lambda^\omega-\Prob_0} \end{equation*} with $g:=f(\{0\},\argmt)$. But, as the next example shows, for smooth random variables, the difference does not converge to zero in total variation. \begin{Example} Let $\cA=\Icc01$ and~$X_n$, $n\in\N$, be real valued i.\,i.\,d.\ random variables, distributed uniformly on~$\cA$. The empirical distribution $\widehat\Prob_n:=n^{-1}\sum_{j=1}^n\delta_{X_j}$ is an atomic measure on~$\cA$, while the uniform distribution on~$\cA$ is absolutely continuous with respect to Lebesgue measure. The TV-norm of their difference does not vanish for $n\to\infty$: \begin{equation*} \TVnorm{\widehat\Prob_n-\Prob} =\sup_{A\in\Borel(\cA)}\abs{\widehat\Prob_n(A)-\Prob(A)} \ge1\text, \end{equation*} as the set $A:=\{X_n\mid n\in\N\}$ shows. \end{Example} We have to follow a different path. Assume again $\cA\in\Borel(\R)$. Let us abbreviate the difference of the cumulative distribution functions by \begin{equation*} F_\Lambda(E) :=\spr{\ifu{\Ioc{-\infty}E}}{\widehat\Prob_\Lambda^\omega-\Prob_0}\textq, E\in\R\text, \end{equation*} and assume that~$g=f(\{0\},\argmt)$ has bounded variation, or more specifically that it is monotone. Then, we can perform the following partial integration with Riemann-Stieltjes integrals: \begin{align*} \abs{\spr g{\widehat\Prob_\Lambda^\omega-\Prob_0}_{\cA}}& =\Bigabs{\int_\cA g(E)\dd F_\Lambda(E)} =\Bigabs{-\int_\cA F_\Lambda(E)\dd g(E)}\\& \le\norm[\infty]{F_\Lambda}\int_\cA\d\abs g(E) =\TVnorm g\cdot\sup_{E\in\R} \abs{\spr{\ifu{\Ioc{-\infty}E}}{\widehat\Prob_\Lambda^\omega-\Prob_0}_\cA} \text. \end{align*} This calculations generalizes the theorem by Glivenko and Cantelli to bounded monotone functions: For $M>0$, we have \begin{equation*} \sup\{\abs{\spr g{\widehat\Prob_{\Lambda_L}-\Prob_0}} \mid\text{$g\from\R\to\Icc{-M}M$ monotone}\} \xto{L\to\infty}0\text. \end{equation*} In order to deal with fields that are only \emph{almost} additive, we have to treat patterns of all finite sizes and not only singletons. Each pattern corresponds to a multivariate random variable. This means that we require a multivariate version of Glivenko--Cantelli theory. To formulate this, we introduce an multivariate version of the empirical measure: For given (large) set $\Lambda_j$, smaller set $\Lambda_L$, a grid $T_{j,L}$ and a coloring $\omega\in \cA^G$ we define the empirical measure by \[ \widehat \Prob_{j,L}^\omega :\cB(\cA^{\Lambda_j}) \to [0,1],\qquad \widehat\Prob_{j,L}^\omega:=\frac{1}{|T_{j,L}|}\sum_{t\in T_{j,L}} \delta_{(\tau_t\omega)_{\Lambda_L}}. \] Here, the grid $T_{j,L}$ is a set of basepoints to (almost) cover the $\Lambda_j$ with translated versions of $\Lambda_L$ along $T_{j,L}$. An illustration of this it to be found in Figure \ref{fig:emp.measure}. Let us emphasize that the illustration serves well in the $\Z^d$-case or in the ST-amenable case. However, for general amenable groups there is usually no grid $T_{j,L}$ for a (perfect) covering a set with \emph{one} set. In this case one can still use the above definition of the empirical measure, but, as in Section \ref{sec:general}, one needs to implement the technique of $\epsilon$-quasi tilings. Moreover, let use emphasize that counting patterns in the empirical measure along a grid is substantially different from counting patterns in the definition of frequencies in Section \ref{sec:patternsfrequencies}, see the definition of $\sharp_PP'$ in \eqref{eq:patterncount} and compare Figure \ref{fig:patterncount} with Figure \ref{fig:emp.measure}. \begin{figure} \begin{tikzpicture}[scale=0.6] { {\color{black} { \foreach \y in {0,1,...,8}{ \draw (1,\y) --(9,\y); } \foreach \x in {1,2,...,9}{ \draw (\x,0) --(\x,8); } \foreach \x in {2,4,6,8}{ \draw[dotted, thick] (\x+0.5,-0.5) -- (\x+0.5,8.5); \draw[dotted, thick] (0.5,\x-0.5) -- (9.5,\x-0.5); } \filldraw[blue] (8,2) circle (4pt); \filldraw[blue] (7,8) circle (4pt); \filldraw[blue] (6,5) circle (4pt); \filldraw[blue] (7,2) circle (4pt); \filldraw[blue] (6,8) circle (4pt); \filldraw[red] (1,1) circle (4pt); \filldraw[green] (1,2) circle (4pt); \filldraw[green] (1,3) circle (4pt); \filldraw[red] (1,4) circle (4pt); \filldraw[red] (1,5) circle (4pt); \filldraw[red] (1,6) circle (4pt); \filldraw[green] (1,7) circle (4pt); \filldraw[green] (1,8) circle (4pt); \filldraw[red] (2,1) circle (4pt); \filldraw[green] (2,1) circle (4pt); \filldraw[red] (2,2) circle (4pt); \filldraw[green] (2,3) circle (4pt); \filldraw[blue] (2,4) circle (4pt); \filldraw[blue] (2,5) circle (4pt); \filldraw[red] (2,6) circle (4pt); \filldraw[red] (2,7) circle (4pt); \filldraw[green] (2,8) circle (4pt); \filldraw[green] (3,1) circle (4pt); \filldraw[red] (3,2) circle (4pt); \filldraw[red] (3,3) circle (4pt); \filldraw[red] (3,4) circle (4pt); \filldraw[green] (3,5) circle (4pt); \filldraw[red] (3,6) circle (4pt); \filldraw[blue] (3,7) circle (4pt); \filldraw[green] (3,8) circle (4pt); \filldraw[blue] (4,1) circle (4pt); \filldraw[green] (4,2) circle (4pt); \filldraw[green] (4,3) circle (4pt); \filldraw[green] (4,4) circle (4pt); \filldraw[blue] (4,5) circle (4pt); \filldraw[blue] (4,6) circle (4pt); \filldraw[blue] (4,7) circle (4pt); \filldraw[blue] (4,8) circle (4pt); \filldraw[red] (5,1) circle (4pt); \filldraw[blue] (5,2) circle (4pt); \filldraw[red] (5,3) circle (4pt); \filldraw[green] (5,4) circle (4pt); \filldraw[green] (5,5) circle (4pt); \filldraw[blue] (5,6) circle (4pt); \filldraw[blue] (5,7) circle (4pt); \filldraw[blue] (5,8) circle (4pt); \filldraw[red] (6,1) circle (4pt); \filldraw[red] (6,2) circle (4pt); \filldraw[green] (6,3) circle (4pt); \filldraw[red] (6,4) circle (4pt); \filldraw[green] (6,5) circle (4pt); \filldraw[red] (6,6) circle (4pt); \filldraw[red] (6,7) circle (4pt); \filldraw[green] (6,8) circle (4pt); \filldraw[green] (7,1) circle (4pt); \filldraw[blue] (7,2) circle (4pt); \filldraw[red] (7,3) circle (4pt); \filldraw[red] (7,4) circle (4pt); \filldraw[green] (7,5) circle (4pt); \filldraw[red] (7,6) circle (4pt); \filldraw[green] (7,7) circle (4pt); \filldraw[green] (7,8) circle (4pt); \filldraw[green] (8,1) circle (4pt); \filldraw[blue] (8,2) circle (4pt); \filldraw[green] (8,3) circle (4pt); \filldraw[green] (8,4) circle (4pt); \filldraw[blue] (8,5) circle (4pt); \filldraw[blue] (8,6) circle (4pt); \filldraw[red] (8,7) circle (4pt); \filldraw[green] (8,8) circle (4pt); \filldraw[red] (9,1) circle (4pt); \filldraw[blue] (9,2) circle (4pt); \filldraw[red] (9,3) circle (4pt); \filldraw[green] (9,4) circle (4pt); \filldraw[blue] (9,5) circle (4pt); \filldraw[blue] (9,6) circle (4pt); \filldraw[blue] (9,7) circle (4pt); \filldraw[red] (9,8) circle (4pt); \filldraw[blue] (1,0) circle (4pt); \filldraw[blue] (2,0) circle (4pt); \filldraw[red] (3,0) circle (4pt); \filldraw[red] (4,0) circle (4pt); \filldraw[green] (5,0) circle (4pt); \filldraw[blue] (6,0) circle (4pt); \filldraw[green] (7,0) circle (4pt); \filldraw[red] (8,0) circle (4pt); \filldraw[green] (9,0) circle (4pt); } } } \draw[rounded corners=0.1cm] (4.7,4.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \draw[rounded corners=0.1cm] (0.7,2.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \draw[rounded corners=0.1cm] (6.7,0.2)-- ++(0,1.1)-- ++(1.6,0)-- ++(0,-1.6)-- ++(-1.6,0)-- ++(0,0.5); \end{tikzpicture} \caption{The set $\Lambda_j$ is a square of side length~$9$ and $\Lambda_L$ a square of side length~$2$. $\Lambda_j$ is (almost) covered when translating $\Lambda_L$ along the positions of the dashed grid (given by $T_{j,L}$). The marked pattern $P$ is found at $3$ (of $16$ possible) positions along the grid. Thus, the empirical measure of this pattern is $\widehat \Prob_{j,L}^\omega(P)= \widehat \Prob_{9,2}^\omega(P)= \frac{3}{16}$.} \label{fig:emp.measure} \end{figure} In order to apply the multivariate version of Glivenko--Cantelli, we aim to integrate functions mapping from $\cA^\Lambda$ to~$\R$. Such a function $g\from\cA^\Lambda\to\R$ is called monotone, if it is monotone in each coordinate. Besides these generalizations due to higher dimensionality, there is another fundamental difference between univariate and multivariate Glivenko--Cantelli theory: While \Cref{thm:GC} makes no assumptions on the distribution of the random variables, the following example shows that we will have to impose some restrictions on the joint distribution of the coordinates of the random vector. \begin{Example} Let $X_j\sim\Normal(0,1)$, $j\in\N$, be i.\,i.\,d.\ standard normal random variables and $Y_j:=-X_j$. We consider the vectors $(X_j,Y_j)\in\R^2$. Let~$\widehat\Prob_n(\omega):=n^{-1}\sum_{j=1}^n\delta_{(X_j,Y_j)}$ be the empirical distribution of $(X_j,Y_j)_{j=1}^n$ on~$\R^2$. The random test function \begin{equation*} g_\omega:=\ifu{\{(x,y)\in\R^2\mid x+y<0\}} +\ifu{\{(X_j(\omega),Y_j(\omega))\mid j\in\N\}} \from\R^2\to\Icc{-1}1 \end{equation*} is monotone in each coordinate, and we have \begin{equation*} \sup\{\abs{\spr g{\widehat\Prob_n^\omega-\Prob}}\mid\text{$g\from\R^2\to\Icc{-1}1$ monotone}\} \ge\abs{\spr{g_\omega}{\widehat\Prob_n^\omega-\Prob}} =1\text. \end{equation*} \end{Example} The problem arises because the set of discontinuities of the monotone function has positive probability. A correct generalization of \cref{thm:GC} to the multivariate case is as follows. \begin{Theorem}[DeHardt \cite{DeHardt-71}, Wright \cite{Wright-81}]\label{wright:LDP} Let \begin{itemize} \item $V_j$, $j\in\N$, be i.\,i.\,d.\ random variables with values in~$\R^k$ and distribution~$\Prob$, \item $\widehat\Prob_n:=\frac1n\sum_{j=1}^n\delta_{V_j}$ for $n\in\N$ the empirical distribution, and \item $M>0$ and $\cM:=\{g\from\R^k\to\Icc{-M}M\mid\text{$g$ monotone}\}$. \end{itemize} Then, the following are equivalent. \begin{enumerate}[(i)] \item\label{GC-i} For all $J\subseteq\{1,\dotsc,k\}$, $J\ne\emptyset$, strictly monotone $g\from\R^J\to\R$, and $E\in\R$, the continuous part~$\Prob_c^J$ of the marginal~$\Prob^J$ of~$\Prob$ satisfies \begin{equation*} \Prob_c^J\bigl(\partial g^{-1}\bigl(\Ioc{-\infty}E\bigr)\bigr)=0\text. \end{equation*} \item There exists an almost sure event~$\Omega_{\mathrm{unif}}$ on which \begin{equation*} \sup_{g\in\cM}\abs{\spr g{\widehat\Prob_n-\Prob}}\xto{n\to\infty}0\text. \end{equation*} \item For all $\kappa>0$, there are $a_\kappa,b_\kappa>0$ such that for all $n\in\N$, there is an event~$\Omega_{\kappa,n}$ with $\Prob(\Omega_{\kappa,n})\ge1-b_\kappa\exp(-a_\kappa n)$ on which \begin{equation*} \sup_{g\in\cM}\abs{\spr g{\widehat\Prob_n-\Prob}}\le\kappa\text. \end{equation*} \end{enumerate} \end{Theorem} Here $\partial g^{-1}\bigl(\Ioc{-\infty}E\bigr)$ denotes the boundary of the sublevel set $g^{-1}\bigl(\Ioc{-\infty}E$. Condition~\labelcref{GC-i} is trivial in the classical case $k=1$. Also, in any dimension, each product measure~$\Prob$ satisfies condition~\labelcref{GC-i}. In fact, the following theorem holds true. \begin{Theorem}\label{thm:strictlymonotone} Let~$\Prob$ be a probability measure on~$\R^k$ which is absolutely continuous with respect to a product measure $\Tensor_{j=1}^k\mu_j$ on~$\R^k$, where~$\mu_j$, $j\in\{1,\dotsc,k\}$ are measures on~$\R$. Then, condition \cref{wright:LDP}\labelcref{GC-i} is satisfied. \end{Theorem} See \cite[Theorem~5.5]{SchumacherSV-17} for a proof. \subsection{Uniform limits for monotone fields} The theorems which follow for the case of infinitely many colors~$\cA$ all have an additional assumption, namely the monotonicity in the random parameters. This is a natural assumption, indeed: The IDS depends on the potential antitonely in our models. Also, the IDS is monotone in site percolation. \begin{Example} We revisit the Anderson model on a site percolation graph from \cref{sec:AndPerc}, this time with $\cA\subseteq\R$. Fix a bounded set $\cA_0\in\Borel(\R)$ for the values of the potential, and let $\cA:=\cA_0\union\{\alpha\}$ with $\alpha>2d+\sup\cA_0$. The value~$\alpha$ of the potential is interpreted as a closed site in the percolation graph: \begin{equation*} \cV_\omega:=\{v\in G\mid V_\omega(v)\ne\alpha\}\text. \end{equation*} The edges of the percolation graph are as before \begin{equation*} \cE_\omega:=\{\{v,w\}\subseteq\cV_\omega\mid v\sim w\}\text. \end{equation*} The Hamiltonian $H_\omega\from\ell^2(G)\to\ell^2(G)$ is given by \begin{equation*} (H_\omega\varphi)(v):= \begin{cases} -\Laplace_\omega\varphi(v)+V_\omega\varphi(v) &\text{if $v\in\cV_\omega$, and}\\ \alpha\varphi(v) &\text{if $v\in G\setminus\cV_\omega$.} \end{cases} \end{equation*} By the min-max principle, the eigenvalues do not decrease when we increase the potential at a site~$v\in G$. Particularly, when the potential reaches the value~$\alpha$ and the site closes, the eigenfunction experiences de facto a Dirichlet boundary condition on that site, which also at most increases the kinetic energy. \par The eigenvalue counting functions count less eigenvalues below a given threshold, if the eigenvalues increase. Therefore, the eigenvalue counting functions decrease when the potential is raised. The same holds true for the limit, i.\,e.\ the IDS. This is the reason why the IDS in the quantum percolation model with random potential is antitone in the randomness. \end{Example} We first turn to the special case $G=\Z^d$. It is physically most relevant and, since~the group~$\Z^d$ satisfies the tiling property, we do not need to resort to quasi tilings in this case. \begin{Theorem}[\cite{SchumacherSV-17}]\label{thm:infiniteA_Zd} Let~$\cA\in\Borel(\R)$, $\Omega:=\cA^{\Z^d}$, and let $(\Omega,\Borel(\Omega),\Prob)$ be a probability space such that~$\Prob$ satisfies \begin{itemize} \item $\Prob$ is translation invariant with respect to the $\Z^d$-action, \item for all $\Lambda\in\cF$, the marginal~$\Prob_\Lambda$ is absolutely continuous with respect to a product measure on $\R^\Lambda$, and \item for a given $r\ge0$ and all $\Lambda_j\in\cF$, $j\in\N$, with $\min\{\dist(\Lambda_i,\Lambda_j)\mid i\ne j\}>r$, the random variables $\omega\mapsto\omega_{\Lambda_j}$, $j\in\N$, are independent. \end{itemize} Further, let $f\colon\cF\times\Omega\to\B$ be a translation invariant, local, almost additive, monotone, bounded field. Then there exists an event~$\tilde\Omega$ of full probability and a function $\fbar\in\B$ such that for every $\omega\in\tilde\Omega$ we have \begin{equation}\label{eqmain} \lim_{j\to\infty}\biggnorm{ \frac{f(\Lambda_j,\omega)}{\setsize{\Lambda_j}}-\fbar} =0\text, \end{equation} where $\Lambda_j:=\Ico0j\isect\Z^d$ for $j\in\N$. For an estimate on the speed of convergence, denote the bound of~$f$ by~$C_f$, the bound of the boundary term~$b$ of~$f$ by~$D_b$. Then, for every $\kappa>0$ and $L\in\N$, $L>2r$, there are $a,b>0$, depending on~$\kappa$, $L$, and~$C_f$, such that for all $j\in\N$, $j>2L$, there is an event~$\Omega_{\kappa,j}$ with probability $\Prob(\Omega_{\kappa,j})\ge1-b\exp(-a\floor{j/L}^d)$, on which \begin{equation*} \Bignorm{\frac{f(\Lambda_j,\omega)}{\setsize{\Lambda_j}}- \fbar} \le2^{2d+1}\Bigl(\frac{(2C_f+D_b)L^d+D_br^d}{j-2L} +\frac{2(C_f+D_b)r^d+3D_br^d}{L-2r}\Bigr) +\kappa\text. \end{equation*} holds true. \end{Theorem} Of course, there is a version of \cref{thm:infiniteA_Zd} for amenable groups. We follow the strategy in \cref{sec:general} and use quasi tilings to deal with infinitely many colors on amenable groups. This brings new challenges. \Cref{wright:LDP} needs as input i.\,i.\,d.\ samples. But quasi tilings are allowed to overlap (in a relatively small volume), see \cref{def:quasitiling}. This destroys the independence of eigenvalue counting functions associated to overlapping tiles. One is tempted to excise the overlap from some of the tiles. However, this would leave us with an independent but not identically distributed sample. In this situation Glivenko--Cantelli-Theory is difficult to apply. The solution is to independently resample the portions of the quasi tiles which overlap and to account for the error by a volume estimate. The last result we present here treats fields with infinitely many colors on amenable groups. \begin{Theorem}[\cite{SchumacherSV-18}]\label{thm:main} Let~$G$ be a finitely generated amenable group with a F{\o}lner sequence~$(Q_j)_j$. Further, fix~$\cA\in\cB(\R)$, and let $(\Omega=\cA^G,\cB(\Omega),\Prob)$ be a probability space such that~$\Prob$ is translation invariant with respect to~$G$, has finite marginals with density w.\,r.\,t.\ a product measure, and independence at a distance. Further, let~$\cU$ be a set of translation invariant, local, almost additive, monotone, bounded fields $f\from\cF\times\Omega\to\B$ with common bound, i.\,e.\ $C:=\sup\{C_f\mid f\in\cU\}<\infty$, common boundary term~$b\from\cF\to\R$ with bound~$D:=D_b$. \begin{enumerate}[(a), wide] \item Then, there exists an event $\tilde\Omega\in\cB(\Omega)$ such that $\Prob(\tilde\Omega)=1$ and for any $f\in\cU$ there exists a function~$\fbar\in\B$, which does not depend on the specific F{\o}lner sequence~$(Q_j)_j$, with \begin{equation*} \forall\omega\in\tilde\Omega\colon\quad \lim_{j\to\infty} \sup_{f\in\cU}\biggnorm{\frac{f(Q_j,\omega)}{\setsize{Q_j}}-\fbar} =0 \text. \end{equation*} \item Furthermore, for each $\ve\in\Ioo0{1/10}$, there exist $j_0(\ve)\in\N$, independent of~$C$, such that for all $f\in\cU$, there are $a(\ve,C),b(\ve,C)>0$, such that for all $j\in\N$, $j\ge j_0(\ve)$, there is an event $\Omega_{j,\ve,C}\in\Borel(\Omega)$, with the properties \begin{equation*} \Prob(\Omega_{j,\ve,C}) \ge1-b(\ve,C) \exp\bigl(-a(\ve,C)\setsize{Q_j}\bigr) \end{equation*} and \begin{align*} \biggnorm{\frac{f(Q_j,\omega)}{\setsize{Q_j}}-\fbar}& \le(37C_f+47\constb[\cU]+47) \ve \qtext{ for all $\omega\in\Omega_{j,\ve,C}$ and all $f\in\cU$.} \end{align*} \end{enumerate} \end{Theorem} \section{Outlook} We have presented a number of theorems concerning convergence in sup-norm and other Banach-space norms of averaged almost additive fields. More important than the convergence itself are the corresponding quantitative error estimates. They split into two parts of two different origins: The geometric part and the probabilistic part. The probabilistic error measures how far off certain empirical measures are from their theoretical counterparts. This difference can be estimated using large deviations techniques, which is implicit in our use of the \cref{wright:LDP} of DeHardt and Wright. An aspect which is not completely satisfactory is that we are not able to specify the dependence of the positive coefficients~$a_\kappa$ and~$b_\kappa$ on the small parameter $\kappa>0$ and the dimension of the pattern $k\in\N$. For this reason we are not able to choose the two lengths scales which tend both to infinity as functions of each other. Furthermore, the Theorems in \cref{sec:infiniteA} assume certain monotonicity with respect to individual random parameters. While this is sufficient for a wide variety of models in statistical physics, there are examples, e.\,g.\ random hopping Hamiltonians, which do not satisfy this assumption. For this reason it is desirable to relax the monotonicity assumption, or formulate alternative sufficient conditions. These are aims which we will pursue in a forthcoming project. \small \begin{longtabu} to\linewidth { r @{$\colon$} X[l] } \toprule \multicolumn{2}{c}{Table of Notation\label{tableofnotations}}\\ \midrule \endfirsthead \toprule \multicolumn{2}{c}{Table of Notation (continued)}\\ \midrule \endhead \midrule \multicolumn{2}{r}{\footnotesize continues on next page}\\ \endfoot \bottomrule \endlastfoot $G$ & a finitely generated amenable group, or its Cayley graph. \\ $g$ & element of~$G$, when~$G$ is used as group \\ $v,w$ & elements of~$G$, when~$G$ is used as Cayley graph \\ $\tilde\tau_g$ & group action of the group~$G$ on its Cayley graph~$G$ \\ $U_g$ & unitary group action of~$G$ on $\ell^2(G)$ \\ $\Laplace$ & Laplace operator on $\ell^2(G)$, $-\Laplace\ge0$ \\ $\cA$ & set of colors \\ $\Borel(\cA)$ & Borel sets on~$\cA$ \\ $\Prob_0$ & probability measure on $(\cA,\Borel(\cA))$ \\ $\Omega$ & set of colorings of~$G$ \\ $\omega$ & coloring \\ $\Prob$ & probability measure on $(\Omega,\Borel(\Omega))$ \\ $V$ & random potential \\ $\tau_g$ & group action of~$G$ on~$\Omega$ and on patterns \\ $H_\omega$ & random Schr\"odinger operator \\ $\Sigma$ & almost sure spectrum of $H_\omega$ \\ $N$ & integrated density of states (IDS) \\ $\d N$ & density of states measure \\ $\Lambda$ & (finite) subset of~$G$, often domain of pattern \\ $\ell^2(\Lambda)$ & subspace of $\ell^2(G)$ \\ $\ifu\Lambda$ & projection $\ell^2(G)\to\ell^2(\Lambda)$ \\ $H_\omega^\Lambda$ & restriction of~$H_\omega$ to $\ell^2(\Lambda)$ \\ $\cF$ & set of finite subsets of~$G$ \\ $\delta_v$ & Kronecker delta on~$v$ \\ $n(\Lambda,\omega)$ & eigenvalue counting function, not normalized \\ $\B$ & Banach space, often the right continuous $\R\to\R$-functions \\ $\Lambda_L$ & cube of side length~$L$ or small F\o{}lner sequence \\ $\Psite$ & Bernoulli measure for site percolation \\ $\cV_\omega$ & vertices of percolation graph \\ $\cE_\omega$ & edges of percolation graph \\ $\Pedge$ & Bernoulli measure for edge percolation \\ $e_j$ & $j$-th basis vector of $\Z^d$ \\ $\Laplace_\omega$ & Laplace operator on percolation graph \\ $\alpha$ & eigenvalue of~$H_\omega$ on $G\setminus\cV_\omega$ \\ $\cA_0$ & set of values of the random potential on a percolation graph \\ $\cV$ & set of visible points in~$\Z^d$ \\ $\Hvis$ & Schr\"odinger operator with~$\ifu\cV$ as potential \\ $\tilde N$ & limiting function in \cref{thm:LMV-ecf} \\ $P$ & pattern, $P\from\Lambda\to\cA$ \\ $[P]_G$ & equivalence class of patterns with respect to~$G$ \\ $\omega_\Lambda$ & pattern induced by~$\omega$ on~$\Lambda$ \\ $P'|_\Lambda$ & pattern induced by $P'\in\cA^{\Lambda'}$ on~$\Lambda$ \\ $P'|_\cF$ & set of all patterns induced by~$P'$ \\ $\sharp_PP'$ & number of occurrences of~$P$ in~$P'$ \\ $\partial^r\Lambda$ & (two-sided) $r$-boundary of~$\Lambda$ \\ $\dist$ & graph distance on Cayley graph \\ $(Q_j)_j$ & F\o{}lner sequence \\ $\nu_P$ & (asymptotic) frequency of~$P$ \\ $b$ & boundary term \\ $D_b$ & bound of~$b$ \\ $F$ & field (without explicit dependence on~$\omega$) \\ $\tilde F$ & pattern function of~$F$ \\ $C_F$ & bound of field~$F$ \\ $\Fbar$ & limit in \cref{thm:LenzMV} \\ $\widehat\Prob_\Lambda^\omega$ & empirical distribution \end{longtabu} \def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth \lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}}

https://arxiv.org/abs/1501.03983

The Storage-Repair-Bandwidth Trade-off of Exact Repair Linear Regenerating Codes for the Case $d = k = n-1$

In this paper, we consider the setting of exact repair linear regenerating codes. Under this setting, we derive a new outer bound on the storage-repair-bandwidth trade-off for the case when $d = k = n -1$, where $(n, k, d)$ are parameters of the regenerating code, with their usual meaning. Taken together with the achievability result of Tian et. al. [1], we show that the new outer bound derived here completely characterizes the trade-off for the case of exact repair linear regenerating codes, when $d = k = n -1$. The new outer bound is derived by analyzing the dual code of the linear regenerating code.

\section{Introduction}\label{sec:intro} In the regenerating-code framework, a file of size $B$ symbols is encoded into $n\alpha$ symbols and distributed among $n$ nodes in the network, such that each node stores $\alpha$ symbols. These symbols are assumed to be drawn from a finite field $\mathbb{F}_q$. The property of data collection demands that one should be able to recover the entire uncoded file by connecting to any $k$ nodes (see Fig. \ref{fig:intro_regen_framework}) and downloading all the $k\alpha$ coded symbols in them. Further, repair of a single failed node is required to be accomplished by connecting to any $d$ surviving nodes and downloading $\beta \leq \alpha$ symbols from each node. The quantity $d\beta$ is termed as the repair-bandwidth. Two notions of node repair exist, and these are known as functional repair and exact repair. Under functional repair, the code symbols in the replacement node are such that data collection and node repair properties continue to hold. Under exact repair, the contents of the failed and replacement nodes are identical. \begin{figure}[h] \centering \includegraphics[width=5in]{regen_paradigm.pdf} \caption{The Regenerating Code Framework.} \label{fig:intro_regen_framework} \end{figure} A cut-set bound argument based on network-coding was used in \cite{DimGodWuWaiRam} to show that under the framework of functional repair (FR), the file size $B$ is upper bounded by \begin{eqnarray} \label{eq:intro_cut_set_bd} B & \leq & \sum_{i=0}^{k-1} \min\{\alpha,(d-i)\beta\}. \end{eqnarray} For fixed values of parameters $\{B,k,d\}$, there are multiple pairs $(\alpha,\beta)$ that satisfy \eqref{eq:intro_cut_set_bd} with equality. This leads to the storage-repair-bandwidth trade-off shown in Fig.~\ref{fig:intro_trade-off} which is piece-wise linear. Existence of FR regenerating codes that can achieve any point on the storage-repair-bandwidth trade-off was also shown in \cite{DimGodWuWaiRam}. The two extremal points on the trade-off curve are termed as the Minimum Storage Regeneration (MSR) and Minimum Bandwidth Regeneration (MBR) points. At the MSR point, the total storage-overhead is as small as possible, while at the MBR point, the repair-bandwidth is the least. The intermediate points on the curve will be referred to as FR-interior points. \begin{figure}[h] \centering \includegraphics[width=3in]{trade_off.pdf} \caption{The Storage-Overhead Repair-Bandwidth Trade-off for Regenerating Codes for an example set of parameters.} \label{fig:intro_trade-off} \end{figure} Several constructions of MSR and MBR codes, having the property of exact repair (ER) exist in literature. Explicit constructions of ER MSR codes for a class of parameters are presented in \cite{SuhRam,RasShaKum_pm,PapDimCad,ShaRasKumRam_ia,TamWanBru}, whereas the existence of ER MSR codes for all $(n,k,d), \ n > d \geq k$ is shown in \cite{CadJafMalRamSuh}. Explicit ER MBR codes for all $(n,k,d), \ n > d \geq k$ are presented in \cite{RasShaKum_pm}. In \cite{ShaRasKumRam_rbt}, a class of ER MBR codes with $d=(n-1)$ is presented, and these codes are termed as repair-by-transfer MBR codes as they enable node repair without need for any operation other than simple data transfer. \subsection{The Trade-off for the Case of Exact Repair} Following results are known in literature regarding the storage-repair-bandwidth trade-off for the case of ER regenerating codes. \begin{enumerate}[1.] \item The non-existence of ER regenerating codes which operate on the FR-interior points of the trade-off curve (with the possible exception of the line segment from the MSR point to the next deflection point) was shown in \cite{ShaRasKumRam_rbt}. \item The trade-off of ER regenerating codes with parameters $(n=4, k=3,d=3)$ was characterized in \cite{Tia}. Except for a region near the MSR point, the interior points on ER trade-off (to be abbreviated as ER-interior points) for the case $(n=4, k=3,d=3)$ lie strictly away from the FR-interior points. \item In \cite{SasSenKum}, an outer bound on the ER trade-off for any general $(n, k, d), n \geq 4$ was derived, which established that the ER-interior points for any $(n, k, d), n \geq 5$ also lie strictly away from the corresponding FR-interior points (except possibly for a small region near the MSR point). For the case of $(4, 3, 3)$, the bound in \cite{SasSenKum} coincided with the bound in \cite{Tia}. Further, it is also known from the results in \cite{BirenITW} that the bound in \cite{SasSenKum} is optimal when the parameters of the ER regenerating code are given by $(n, k=3, d=n-1)$. However, when $k \geq 4$, the optimality of the outer bound in \cite{SasSenKum} is not known in general. \item Two new outer bounds on the trade-off of ER regenerating codes appear in \cite{Duursma}. These are obtained by extending the techniques of \cite{Tia} and \cite{SasSenKum}. The optimality of these bounds is not known, if we exclude the parameters $(n, k=3, d=n-1)$. \item Constructions of ER regenerating codes which strictly improve upon the space-sharing region of MBR and MSR codes appear in \cite{BirenPVK_canonical}, \cite{ernvall}, \cite{goparaju}. When $k = d = n-1$, the achievable regions presented in all these three works coincide (see Remark $1$, \cite{goparaju}). \end{enumerate} \subsection{Our results} In this paper, we characterize the storage-repair-bandwidth trade-off of $(n, k=n-1, d=n-1), n \geq 5$ ER linear regenerating codes\footnote{The case $n=4$ is already solved in \cite{Tia}, and the case $n < 4$ degenerates to trivial cases.}. This is done by deriving a new upper bound on the file size $B$ of ER linear regenerating codes for the case $k=d=n-1, n \geq 4$. The main result of this paper is stated below. \vspace{0.1in} \begin{thm} \label{thm:new_bound_k_eq_d} Consider an exact repair linear regenerating code, having parameters $(n, k = n-1, d = n-1), (\alpha, \beta), n \geq 4$. Then, the file size $B$ of the code is upper bounded by \begin{eqnarray} \label{eq:bound_rank_G} B & \leq & \left \{ \begin{array}{c} \left \lfloor \frac{r(r-1)n\alpha + n(n-1)\beta}{r^2+r}\right \rfloor, \ \frac{d\beta}{r} \leq \alpha \leq \frac{d\beta}{r-1}, \\ \hspace{1.5in} 2 \leq r \leq n - 2 \\ (n-2)\alpha + \beta, \ \frac{d\beta}{n-1} \leq \alpha \leq \frac{d\beta}{n-2} \end{array} \right. . \end{eqnarray} \end{thm} \vspace{0.1in} The above theorem gives an upper bound on $B$ for the range of $\alpha$ given by $\beta \leq \alpha \leq (n-1)\beta$. Note that when $d=k=n-1$, $\alpha = \beta$ corresponds to the MSR point and $\alpha = (n-1)\beta$ corresponds to the MBR point. In Section \ref{sec:ach}, we will see that the outer bound on storage-repair-bandwidth trade-off corresponding to the bound in Theorem \ref{thm:new_bound_k_eq_d} coincides with the achievability result provided in \cite{BirenPVK_canonical}. Thus, together with this achievability result, the new outer bound completely characterizes the trade-off of ER linear regenerating codes, for the case $k = d = n-1$, $n \geq 5$. \vspace{0.1in} \subsubsection{Illustration of Theorem \ref{thm:new_bound_k_eq_d} for the case of $(5, 4, 4)$ codes} If we specialize \eqref{eq:bound_rank_G} for the case $(5, 4, 4)$, we get \begin{eqnarray} \label{eq:bound_544_rankG_ours} B & \leq & \left \{ \begin{array}{c} \left \lfloor \frac{5\alpha + 10\beta}{3} \right \rfloor, \ 2\beta \leq \alpha \leq 4\beta \\ \left \lfloor \frac{15\alpha + 10\beta}{6} \right \rfloor, \ \frac{4\beta}{3} \leq \alpha \leq 2\beta \\ 3\alpha + \beta, \ \beta \leq \alpha \leq \frac{4\beta}{3} \end{array} \right. . \end{eqnarray} In Fig. \ref{fig:544_tradeoff}, we plot the outer bound on the ``normalized" storage-repair-bandwidth trade-off between $\alpha/B$ and $\beta/B$ corresponding to \eqref{eq:bound_544_rankG_ours}. The normalization is done with respect to the file size $B$. \begin{figure}[ht] \centering \includegraphics[width=3in]{tradeoff_544_comparison.pdf} \caption{Comparison of outer bounds on the storage-repair-bandwidth trade-off of ER regenerating codes for the case $(n = 5, k = d = 4)$. The new outer bound plotted in this figure is obtained under the assumption of linear regenerating codes.} \label{fig:544_tradeoff} \end{figure} In this figure, we have also plotted the following other curves: \begin{enumerate}[1.] \item The FR trade-off of $(5, 4, 4)$ regenerating codes. \item The achievability result from \cite{BirenPVK_canonical} for the parameters $n = 5, d = k = 4$. We see that our outer bound on the trade-off coincides with this achievability result, and thus establishes the optimality of the new outer bound. \item The outer bounds on the trade-off obtained in \cite{SasSenKum} and \cite{Duursma} for $(5, 4, 4)$ ER regenerating codes. We see that the new outer bound is tighter than both these other outer bounds, when the latter bounds are restricted to the case of linear regenerating codes. The expression for the file size bound appearing in \cite{SasSenKum}, when restricted to the case of linear codes, is given by (see Example $2$ of \cite{SasSenKum}) \begin{eqnarray} \label{eq:bound_544_rankG_Biren} B & \leq & \left \{ \begin{array}{c} \left \lfloor \frac{7\alpha + 22\beta}{5} \right \rfloor, \ \frac{18}{7}\beta \leq \alpha \leq 4\beta \\ \\ \left \lfloor \frac{7\alpha + 6\beta}{3} \right \rfloor, \ \frac{3\beta}{2} \leq \alpha \leq \frac{18}{7}\beta \\ \\ 3\alpha + \beta, \ \beta \leq \alpha \leq \frac{3\beta}{2} \end{array} \right., \end{eqnarray} and the bound in \cite{Duursma}, when restricted to the case of linear codes, is given by (see Examples $4.3$ and $5.2$ of \cite{Duursma}) \begin{eqnarray} \label{eq:bound_544_rankG_Duursma} B & \leq & \left \{ \begin{array}{c} \left \lfloor \frac{7\alpha + 22\beta}{5} \right \rfloor, \ \frac{23}{7}\beta \leq \alpha \leq 4\beta \\ \\ \left \lfloor \frac{21\alpha + 57\beta}{14} \right \rfloor, \ \frac{19}{7}\beta \leq \alpha \leq \frac{23}{7}\beta \\ \\ \left \lfloor \frac{11\alpha + 19\beta}{6} \right \rfloor, \ \frac{5}{2}\beta \leq \alpha \leq \frac{19}{7}\beta \\ \\ \left \lfloor \frac{13\alpha + 14\beta}{6} \right \rfloor, \ 2\beta \leq \alpha \leq \frac{5}{2}\beta \\ \\ \left \lfloor \frac{7\alpha + 6\beta}{3} \right \rfloor, \ \frac{3}{2}\beta \leq \alpha \leq 2\beta \\ \\ 3\alpha + \beta, \ \beta \leq \alpha \leq \frac{3}{2}\beta \end{array} \right.. \end{eqnarray} \end{enumerate} \subsection{Our Approach, and Some Preliminaries} We make use of the fact that for the case of linear regenerating codes, maximizing the file size $B$ of the regenerating code is equivalent to minimizing the dimension of dual of the linear regenerating code. More formally, let $\mathcal{C}$ denote an $(n, k, d), (\alpha, \beta)$ ER linear regenerating code, having the (vector-symbol) alphabet $\mathbb{F}_q^{\alpha}$. Also, let the $B \times n\alpha$ matrix $G$ (whose entries are drawn from $\mathbb{F}_q$) denote a generator matrix for $\mathcal{C}$. To be precise, $G$ is the generator matrix for the underlying scalar code (say $\mathcal{C}_s$) of length $n\alpha$, where $\mathcal{C}_s$ is obtained by expanding each vector-symbol of $\mathcal{C}$ into $\alpha$ scalar symbols over $\mathbb{F}_q$. Next, consider the $(n\alpha - B) \times n\alpha$ matrix $H$ which forms a parity check matrix of $\mathcal{C}_s$, i.e., $H$ generates the dual code of $\mathcal{C}_s$. In this paper, we will simply say that $H$ corresponds to the dual of the regenerating code $\mathcal{C}$, and also loosely identify the dual of the code $\mathcal{C}_s$ as the dual of the regenerating code $\mathcal{C}$ itself. Since the code is linear, we have $B = \text{rank}(G) = n\alpha - \text{rank}(H)$. Our approach in this paper will be to find a lower bound on $\text{rank}(H)$ and then convert it to an upper bound on $\text{rank}(G)$. Without loss of generality, we assume that the first $\alpha$ columns of $G$ generate the contents of the first node, the second $\alpha$ columns of $G$ generate the contents of the second node, and so on. The first $\alpha$ columns of $H$ will together be referred to as the first thick column of $H$; similarly the second thick column and so on. For any set $S \subseteq [n]=\{1, 2, \ldots, n\}$, we will write $H|_S$ to denote restriction of $H$ to the thick columns indexed by the set $S$. We will make use of the following properties of the matrix $H$ which were established in \cite{Duursma}. \vspace{0.1in} \begin{lem}[Data Collection] \label{lem:H_datacollection} $\text{Rank}\left({H|_S}\right) = (n-k)\alpha$, for any $S \subseteq [n]$ such that $|S| = n-k$. \end{lem} \begin{proof} This is a re-statement of Part $(1)$ of Proposition $2.1$ of \cite{Duursma}, and is equivalent to the data collection property. \end{proof} \vspace{0.1in} \begin{lem}[Exact Repair] \label{lem:H_repair} Assume that $d= n-1$. Then, under the assumption of ER, the row space of $H$ contains a collection of $n\alpha$ vectors which can be arranged as the rows of an $n\alpha \times n\alpha$ matrix $H_{repair}$, as given below: \begin{eqnarray} \label{eq:Hrepair} H_{repair} & = & \left[ \begin{array}{c|c|c|c} I_{\alpha} & A_{1,2} & & A_{1,n} \\ \hline \\ A_{2,1} & I_{\alpha} & & A_{2, n} \\ \hline \\ & & \ \ \ddots \ \ & \\ \hline \\ A_{n,1} & A_{n,2} & & I_{\alpha} \end{array} \right], \end{eqnarray} where $I_{\alpha}$ denotes the identity matrix of size $\alpha$ and $A_{i,j}$ denotes an $\alpha \times \alpha$ matrix such that $\text{rank}\left(A_{i, j}\right) \leq \beta, 1 \leq i, j \leq n, i \neq j$. \end{lem} \begin{proof} This follows from Part $(2)$ of Proposition $2.1$ of \cite{Duursma}, and is equivalent to the exact-repair property for the case $d = n-1$. \end{proof} \vspace{0.1in} \begin{note} \label{rem:dual_properties} Note from Lemmas \ref{lem:H_datacollection} and \ref{lem:H_repair} that for the case of $d=k=n-1$, the matrix $H_{repair}$ by itself defines an $(n, k=n-1, d=n-1)(\alpha, \beta)$ regenerating code. Since $\text{rank}(H) \geq \text{rank}(H_{repair})$, we will assume that $H = H_{repair}$ while we derive a lower bound on the rank of $H$ for the case of $d=k=n-1$. \end{note} \vspace{0.1in} The technical discussion appearing in the rest of the article is divided as follows: \begin{enumerate}[1.] \item In Section \ref{sec:ach}, we quickly review the achievability result from \cite{BirenPVK_canonical}, for the case of $k = d = n-1$. As mentioned before, the optimality of the new bound derived here will follow from this achievability result. \item In Section \ref{sec:dimakis_via_dual}, we will re-derive \eqref{eq:intro_cut_set_bd} for the case of ER linear codes, by calculating a simple lower bound on $\text{rank}(H)$. \item In Sections \ref{sec:433} and \ref{sec:544}, we will refine the proof presented in Section \ref{sec:dimakis_via_dual}, and obtain the proof of Theorem \ref{thm:new_bound_k_eq_d} for the special cases of $(4, 3, 3)$ and $(5,4,4)$ ER linear regenerating codes, respectively. Our proofs for these two special cases will help us illustrate the key ideas that will be involved in the general proof. Note that the trade-off for the case of $(4,3,3)$ (including non-linear codes) has already been solved by Tian et. al. \cite{Tia}. \item The proof for the general $(n, k = n - 1, d = n-1)$ will be subsequently presented in Section \ref{sec:general}. \end{enumerate} \section{Achievable Region for $(n, k = n-1, d = n-1)$} \label{sec:ach} In \cite{BirenPVK_canonical}, the authors give a construction of $(n, k=d, d)$ ER linear regenerating codes, and these are termed as canonical regenerating codes. When specialised to the case $d = n-1$, code constructions are obtained for the following points on the normalized storage vs repair-bandwidth plot: \begin{equation} \label{eq:ach_Biren} \left(\frac{\alpha}{B} , \frac{\beta}{B} \right) = \left(\frac{r}{n(r-1)}, \frac{r}{n(n-1)}\right), 2 \leq r \leq n-1. \end{equation} Note that in \eqref{eq:ach_Biren}, if we put $r = 2$, we get the MBR point, and as $r$ increases, points closer to the MSR point are achieved. It is also proved that the point corresponding to $ r = n-1$ lies on the FR trade-off, on the line-segment whose one end point is the MSR point. An achievable region on the normalized storage vs repair-bandwidth plot, corresponding to \eqref{eq:ach_Biren} is obtained by 1) connecting the adjacent points in \eqref{eq:ach_Biren} by straight line-segments, and 2) drawing a line segment between the MSR point and the point corresponding to $r = n-1$. For example, if we set $n = 5$, the points of deflection on this achievable region are given by (see Fig. \ref{fig:544_tradeoff}.) \begin{eqnarray} \left(\frac{\alpha}{B} , \frac{\beta}{B} \right) & = & \left(\frac{1}{4}, \frac{1}{4}\right) , \ \text{MSR point} \\ \left(\frac{\alpha}{B} , \frac{\beta}{B} \right) & = & \left(\frac{4}{15}, \frac{1}{5}\right) , \ r = 4 \\ \left(\frac{\alpha}{B} , \frac{\beta}{B} \right) & = & \left(\frac{3}{10}, \frac{3}{20}\right) , \ r = 3 \\ \left(\frac{\alpha}{B} , \frac{\beta}{B} \right) & = & \left(\frac{2}{5}, \frac{1}{10}\right) , \ r = 2, \text{MBR point}. \end{eqnarray} Now, to see that the outer bound on the normalized trade-off induced by the new file-size bound in Theorem \ref{thm:new_bound_k_eq_d} is the same as what is achieved by \cite{BirenPVK_canonical}, we note that 1) the equation of the line-segment obtained by connecting the two points $\left(\frac{r}{n(r-1)}, \frac{r}{n(n-1)}\right)$ and $\left(\frac{(r+1)}{n((r+1)-1)}, \frac{(r+1)}{n(n-1)}\right)$, $ 2 \leq r \leq n-2$ is given by \begin{eqnarray} r(r-1)n\left(\frac{\alpha}{B}\right) + n(n-1)\left(\frac{\beta}{B}\right) & = & r^2+r, \end{eqnarray} and 2) the equation of the line segment obtained by joining the MSR point and the point corresponding to $r=n-1$ is given by $(n-2)\left(\frac{\alpha}{B}\right) + \left(\frac{\beta}{B}\right) = 1$. \section{A Derivation of \eqref{eq:intro_cut_set_bd} Based on Dual Code} \label{sec:dimakis_via_dual} In this section, we will present a simple proof of \eqref{eq:intro_cut_set_bd} for ER linear regenerating codes. As we will see, our proof of Theorem \ref{thm:new_bound_k_eq_d}, to be presented later in this document, will be built up on proof of \eqref{eq:intro_cut_set_bd} that is presented here. As before, we assume that $\mathcal{C}$ denotes an $(n, k, d = n-1) (\alpha, \beta)$ linear regenerating code, and the matrix $H$ generates the dual of $\mathcal{C}$. Also, recall that we use the notation $H|_S$ to denote the restriction of $H$ to the thick columns indexed by the set $S$, where $S \subseteq [n]$. The basic idea of the proof is to get a lower bound on the column rank of the matrix $H$. Towards this, define the quantities $\delta_j, 1 \leq j \leq n$ as follows: \begin{eqnarray} \delta_1 & = & \text{rank}(H|_{[1]}), \label{eq:proof_diamkis_aa}\\ \delta_j & = & \text{rank}(H|_{[j]}) - \text{rank}(H|_{[j-1]}), 2 \leq j \leq n, \label{eq:proof_diamkis_0} \end{eqnarray} where we have used the notation $[t] = \{1, 2, \ldots, t\}$ for any positive integer $t$. We claim that \begin{eqnarray} \label{eq:proof_diamkis_1} \delta_{j} = \alpha, \ 1 \leq j \leq n-k, \end{eqnarray} and \begin{eqnarray} \label{eq:proof_diamkis_2} \delta_{j} \geq (\alpha - (j-1)\beta)^+, \ n-k + 1 \leq j \leq n, \end{eqnarray} where the quantity $a^+$ denotes $\max(a, 0)$. Here, \eqref{eq:proof_diamkis_1} follows because we know from Lemma \ref{lem:H_datacollection} that any $n-k$ thick columns of $H$ has rank given by $(n-k)\alpha$. To see why \eqref{eq:proof_diamkis_2} is true, focus on the $j^{\text{th}}$ thick row of $H_{repair}$ (i.e., the rows from $(j-1)\alpha + 1$ to $j\alpha$ of $H_{repair}$) and note that \begin{eqnarray} \delta_j & \geq & \left(\text{rank}(I_{\alpha}) - \sum_{\ell=1}^{j-1}\text{rank}(A_{j,\ell})\right)^+ \label{eq:dimakis_proof_3a} \\ & \geq & \left(\alpha - (j-1)\beta \right)^+, \ n-k + 1 \leq j \leq n, \label{eq:dimakis_proof_3} \end{eqnarray} where \eqref{eq:dimakis_proof_3} follows because, we know from Lemma \ref{lem:H_repair} that $\text{rank}(A_{i,j}) \leq \beta$. Now, the (column) rank of the matrix $H$ can be lower bounded as \begin{eqnarray} \text{rank}(H) & = & \sum_{i=j}^{n}\delta_j \\ & \geq & (n-k)\alpha + \sum_{j=n-k+1}^{n}\left(\alpha - (j-1)\beta \right)^+, \label{eq:dimakis_proof_4} \end{eqnarray} where \eqref{eq:dimakis_proof_4} follows from \eqref{eq:proof_diamkis_1} and \eqref{eq:proof_diamkis_2}. From this, it follows that the file size $B$ of the code $\mathcal{C}$ can be upper bounded as \begin{eqnarray} B & = & n\alpha - \text{rank}(H) \\ & \leq & n\alpha - (n-k)\alpha - \sum_{j=n-k+1}^{n}\left(\alpha - (j-1)\beta \right)^+ \\ & = & \sum_{j=n-k+1}^{n}\left( \alpha - \left(\alpha - (j-1)\beta \right)^+ \right) \\ & = & \sum_{j=n-k+1}^{n}\min(\alpha, (j-1)\beta) \label{eq:dimakis_proof_5} \\ & = & \sum_{j=0}^{k-1}\min(\alpha, (d-j)\beta), \label{eq:dimakis_proof_6} \end{eqnarray} where \eqref{eq:dimakis_proof_5} follows from noting that $\alpha - \left(\alpha - (j-1)\beta \right)^+ = \min(\alpha, (j-1)\beta)$ and \eqref{eq:dimakis_proof_6} follows from our assumption that $d = n-1$. \section{The Trade-off of $(4, 3, 3)$ ER Linear Regenerating Codes Based on Dual Code} \label{sec:433} In this section, we will re-derive the trade-off of $(4, 3, 3)$ linear regenerating codes, which was originally obtained by \cite{Tia}. Our proof here will be built on the proof of $(1)$ which was presented in Section \ref{sec:dimakis_via_dual}. We will prove that, when restricted to the case of ER, it is possible to get a lower bound on $\text{rank}(H)$ that is in general tighter than what is given in \eqref{eq:dimakis_proof_4}. The following theorem specialises Theorem \ref{thm:new_bound_k_eq_d} for the case of $(4, 3, 3)$ and states the result (to be proved here) in terms of $\text{rank}(H)$. \vspace{0.1in} \begin{thm} \label{thm:433} Consider an exact repair linear regenerating code $\mathcal{C}$, having parameters $(n=4, k=3, d=3,), (\alpha, \beta)$. Let the matrix $H$ correspond to the dual of the code $\mathcal{C}$. Then, the rank of the matrix $H$ is lower bounded by \begin{eqnarray} \label{eq:bound_rank_H_433} \text{rank}(H) & \geq & \left \{ \begin{array}{c} \left \lceil \frac{8\alpha - 6\beta}{3} \right \rceil, \ 1.5\beta \leq \alpha \leq 3\beta \\ \\ 2\alpha - \beta, \ \beta \leq \alpha \leq 1.5\beta \end{array} \right. . \end{eqnarray} \end{thm} \vspace{0.1in} Before we prove Theorem \ref{thm:433},we note the following points regarding this theorem. \begin{enumerate}[1.] \item For the case of $(4, 3, 3)$, $\alpha = \beta$ corresponds to the MSR point and $\alpha = 3\beta$ corresponds to the MBR point. \item From \eqref{eq:dimakis_proof_4}, we see that \begin{eqnarray} \label{eq:433_proof_1} \text{rank}(H) & \geq & 2\alpha - \beta, \ \beta \leq \alpha \leq 2\beta. \label{eq:proof_433_en} \end{eqnarray} Thus, to prove Theorem \ref{thm:433}, we only need to prove that \begin{eqnarray} \text{rank}(H) & \geq & \left \lceil \frac{8\alpha - 6\beta}{3} \right \rceil, \ 1.5\beta \leq \alpha \leq 3\beta. \end{eqnarray} In fact, we will simply prove that \begin{eqnarray} \label{eq:433_proof_4} \text{rank}(H) \geq \left \lceil \frac{8\alpha - 6\beta}{3} \right \rceil, \end{eqnarray} without bothering about the range of $\alpha$. Note that given \eqref{eq:proof_433_en}, this suffices to prove Theorem \ref{thm:433}. \item To see that the bound in Theorem \ref{thm:433} is tighter than the bound in \eqref{eq:dimakis_proof_4}, note that \begin{eqnarray} \label{eq:433_proof_2} \left \lceil \frac{8\alpha - 6\beta}{3} \right \rceil & > & 2\alpha - \beta, \ 1.5\beta < \alpha \leq 2\beta \end{eqnarray} and \begin{eqnarray} \label{eq:433_proof_3} \left \lceil \frac{8\alpha - 6\beta}{3} \right \rceil & > & 3\alpha - 3\beta, \ 2\beta \leq \alpha < 3\beta, \end{eqnarray} where $2\alpha - \beta$ and $3\alpha - 3\beta$ respectively denote the bounds obtained in \eqref{eq:dimakis_proof_4} for the cases when $\beta \leq \alpha \leq 2\beta$ and $2\beta \leq \alpha \leq 3\beta$. \end{enumerate} \vspace{0.1in} A comparison of the bounds in \eqref{eq:dimakis_proof_4} and \eqref{eq:bound_rank_H_433} is shown in Fig. \ref{fig:433_rankH_comparison}. Here, we plot the two lower bounds on $\text{rank}(H)$ as a function of $\alpha$, when $\beta$ is fixed as $48$. \begin{figure}[h] \centering \includegraphics[width=3in]{rankH_433_comparison.pdf} \caption{Comparison of the lower bounds on $\text{rank}(H)$ as function of $\alpha$, for the case of $(n = 4, k = 3, d = 3)$ with $\beta = 48$. The dashed and the solid lines correspond to the cases of functional and exact repairs, respectively. See \eqref{eq:dimakis_proof_4} and \eqref{eq:bound_rank_H_433} for the corresponding equations.} \label{fig:433_rankH_comparison} \end{figure} \vspace{0.1in} \subsection{Proof of Theorem \ref{thm:433}} \label{sec:proof_thm_433} We begin with certain notation needed to prove Theorem \ref{thm:433}. For any matrix $B$ (over $\mathbb{F}_q$), we will use $\mathcal{S}(B)$ to denote the column space of $B$. We will write $\rho(B)$ to mean $\text{rank}(B)$, which is also the same as the dimension of the space $\mathcal{S}(B)$. Next, define $H^{(4)} = H_{repair}$. Also let the matrix $H^{(4)}_j$ denote the $j^{\text{th}}$ thick column of $H^{(4)}, 1 \leq j \leq 4$, i.e., $H^{(4)} = [H^{(4)}_1 \ H^{(4)}_2 H^{(4)}_3 \ H^{(4)}_4 ]$. Next, define the matrices $H^{(3)}_j, 2 \leq j \leq 4$, such that the columns of $H^{(3)}_j$ form a basis for the vector space $\mathcal{S}\left(H^{(4)}_j\right) \cap \mathcal{S}\left(H^{(4)}|_{[j-1]}\right)$. Also define the matrix $H^{(3)}$ as \begin{eqnarray} H^{(3)} = [H^{(3)}_2 H^{(3)}_3 \ H^{(3)}_4 ], \end{eqnarray} i.e., $H^{(3)}$ is obtained by stacking the columns of $H^{(3)}_j, 2 \leq j \leq 4$. Notice that the first thick column of $H^{(3)}$ is denoted as $H^{(3)}_2$ instead of $H^{(3)}_1$ (and so on). This has been done intentionally for notational convenience. The basic idea of our proof for the case of $(4, 3, 3)$ comes from the observation that $\rho(H^{(4)}) \geq \rho(H^{(3)})$. We will show here that \eqref{eq:433_proof_4} is a necessary condition for this to be true. Towards this, we will firstly establish some more notation needed for the proof. We will then separately compute (or bound) the ranks of the two matrices $H^{(4)}$ and $H^{(3)}$. Finally, we will show that the comparison of the two ranks yields the desired bound. \vspace{0.1in} \subsubsection{Some Additional Notation } For any two subspaces $W_1$ and $W_2$, we write $W_1 \subseteq W_2$ to mean that $W_1$ is a subspace of $W_2$. Equation \eqref{eq:Hrepair} will be denoted as $H^{(4)} = \left( A^{(4)}_{i,j} , 1 \leq i, j \leq 4 \ \right )$, where $A^{(4)}_{i,i} = I_{\alpha}, 1 \leq i \leq 4$, and when $i \neq j$, we have added a superscript on $A_{i,j}$. Note that in this notation, the $j^{\text{th}}$ thick column of $H^{(4)}$ is given by $H^{(4)}_j = \left( A^{(4)}_{i,j} , 1 \leq i \leq 4 \ \right )$. Also note that \begin{eqnarray} \label{eq:433_proof_0a} \rho\left(H^{(4)}_j\right) & = & \rho\left( A^{(4)}_{j,j}\right), \ 1 \leq j \leq 4. \end{eqnarray} In terms of block sub-matrices, the matrix $H^{(3)}$ will be identified as $H^{(3)} = \left( A^{(3)}_{i,j} , 1 \leq i \leq 4, 2 \leq j \leq 4 \ \right )$, where $A^{(3)}_{i,j}$ is an $\alpha \times \rho(H^{(3)}_j)$ matrix (over $\mathbb{F}_q$) such that \begin{eqnarray} \mathcal{S}\left(A^{(3)}_{i,j} \right) & \subseteq & \mathcal{S}\left(A^{(4)}_{i,j} \right) \ \bigcap \ \sum_{\ell=1}^{j-1}\mathcal{S}\left(A^{(4)}_{i,\ell}\right). \label{eq:433_proof_0b} \end{eqnarray} Note that \eqref{eq:433_proof_0b} is a direct consequence of the definition of the matrix $H^{(3)}$. Here, we would like to clarify that \eqref{eq:433_proof_0b} is not equivalent to the definition of $H^{(3)}$. The definition of $H^{(3)}$ demands additional restrictions on the matrices $\{A_{i,j}^{(3)}\}$ so that the columns of $H^{(3)}_j$ form a basis for the vector space $\mathcal{S}\left(H^{(4)}_j\right) \cap \mathcal{S}\left(H^{(4)}|_{[j-1]}\right), 2 \leq j \leq 4$. Next, observe that the $j^{\text{th}}$ thick column of $H^{(3)}$ can be written in terms of the block sub-matrices as $H^{(3)}_j = \left( A^{(3)}_{i,j}, 1 \leq i \leq 4 \ \right), 2 \leq j \leq 4$. Also, note that \begin{eqnarray} \rho\left(H^{(3)}_j\right) & = & \rho\left( A^{(3)}_{j,j}\right), \ 2 \leq j \leq 4.\label{eq:433_proof_0} \end{eqnarray} To see why \eqref{eq:433_proof_0} is true, firstly observe that the $\rho\left(H^{(3)}_j\right)$ columns of $H^{(3)}_j$ can be extended to a basis for $\mathcal{S}\left(H^{(4)}_j\right)$ by adding exactly $\rho\left(H^{(4)}_j\right) - \rho\left(H^{(3)}_j\right)$ additional columns. This implies that the $\rho\left(H^{(3)}_j\right)$ columns of $A^{(3)}_{j,j}$ can be extended to a basis for $\mathcal{S}\left(A^{(4)}_{j,j}\right)$ by adding at most $\rho\left(H^{(4)}_j\right) - \rho\left(H^{(3)}_j\right)$ additional columns. But then, we know from \eqref{eq:433_proof_0a} that $\rho\left(H^{(4)}_j\right) = \rho\left( A^{(4)}_{j,j}\right)$. Hence, it must indeed be true that $\rho\left(H^{(3)}_j\right) = \rho\left( A^{(3)}_{j,j}\right), \ 2 \leq j \leq 4$. \vspace{0.1in} \subsubsection{$\text{Rank}(H^{(4)})$} Let us define the quantities $\delta_j, 1 \leq j \leq 4$ in the same manner as we did in \eqref{eq:proof_diamkis_aa} and \eqref{eq:proof_diamkis_0} for the proof of the FR-trade-off, i.e., \begin{eqnarray} \delta_1 & = & \rho\left(H^{(4)}_1\right) \ = \ \rho\left(A^{(4)}_{1,1}\right), \label{eq:433_proof_5z}\\ \delta_j & = & \rho\left(H^{(4)}|_{[j]}\right) - \rho\left(H^{(4)}|_{[j-1]}\right), \ 2 \leq j \leq 4, \label{eq:433_proof_5}. \end{eqnarray} From the discussion in Section \ref{sec:dimakis_via_dual} (see \eqref{eq:dimakis_proof_3a}), we know that \begin{eqnarray} \delta_j & \geq & \left(\rho\left(A_{j,j}^{(4)}\right) - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(4)}\right)\right)^+, \ 2 \leq j \leq 4. \label{eq:433_proof_5a} \end{eqnarray} Thus, let us assume that \begin{eqnarray} \delta_j & = & \left(\rho\left(A_{j,j}^{(4)}\right) - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(4)}\right)\right)^+ + \alpha_j, \ 2 \leq j \leq 4 \label{eq:433_proof_6}, \end{eqnarray} where $\{\alpha_j, 2 \leq j \leq 4\}$ are non-negative integers. The rank of the matrix $H^{(4)}$ can now be written as \begin{eqnarray} \rho\left( H^{(4)} \right) & = & \sum_{j=1}^{4}\delta_j \\ & = & \rho\left(A^{(4)}_{1,1}\right) \ + \ \sum_{j=2}^{4} \left\{ \left(\rho\left(A_{j,j}^{(4)}\right) - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(4)}\right)\right)^+ + \alpha_j \right\} \label{eq:433_proof_7}. \end{eqnarray} \vspace{0.1in} \begin{note} Note that in \eqref{eq:433_proof_5a}, we could have written the bound on $\delta_j$ directly in terms of $\alpha, \beta$, like we did in \eqref{eq:dimakis_proof_3}. The reason for not doing so at this stage of the proof, and keeping the bound on $\delta_j$ only in terms of ranks of $\{A_{i,j}^{(4)}\}$ will become evident when we discuss the case of $(5, 4, 4)$ regenerating codes in Section \ref{sec:544}. See Remark \ref{rem:resuse433} as well. \end{note} \vspace{0.1in} \subsubsection{$\text{Rank}(H^{(3)})$} The following lemma obtains a relation between the ranks of the matrices $\{A^{(3)}_{i,j}\}$ and the ranks of the matrices $\{A^{(4)}_{i,j}\}$. This result will be used to obtain a lower bound on the rank of $H^{(3)}$. \vspace{0.1in} \begin{lem} \label{lem:433_intersections} \begin{enumerate}[a)] \item \begin{eqnarray} \rho\left(A^{(3)}_{j,j} \right) & = & \rho\left(A^{(4)}_{j,j} \right) \ - \ \delta_j \label{eq:433_proof_8} \\ & = & \rho\left(A^{(4)}_{j,j} \right) \ - \ \left\{ \left(\rho\left(A_{j,j}^{(4)}\right) - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(4)}\right)\right)^+ + \alpha_j\right\}, \ 2 \leq j \leq 4. \label{eq:433_proof_9} \end{eqnarray} \item \begin{eqnarray} \sum_{\ell=2}^{j-1} \rho\left(A^{(3)}_{j,\ell} \right) & \leq & \sum_{\ell=1}^{j-1}\rho\left(A^{(4)}_{j,\ell} \right) \ - \ \rho\left(A^{(3)}_{j,j} \right), \ 3 \leq j \leq 4. \label{eq:433_proof_10} \end{eqnarray} \end{enumerate} \end{lem} \begin{proof} Let us prove \eqref{eq:433_proof_8} first. Using the definition of $\delta_j$ from \eqref{eq:433_proof_5}, we get \begin{eqnarray} \delta_j & = & \rho\left(H^{(4)}|_{[j]}\right) - \rho\left(H^{(4)}|_{[j-1]}\right) \\ & = & \text{dim}\left(\mathcal{S}\left(H^{(4)}|_{[j-1]} \right) + \mathcal{S}\left(H^{(4)}_j \right)\right) - \text{dim}\left(\mathcal{S}\left(H^{(4)}|_{[j-1]}\right)\right) \label{eq:433_proof_11} \\ & = & \text{dim}\left(\mathcal{S}\left(H^{(4)}_j \right)\right) - \text{dim}\left(\mathcal{S}\left(H^{(4)}|_{[j-1]} \right) \cap \mathcal{S}\left(H^{(4)}_j \right)\right) \label{eq:433_proof_12} \\ & = & \rho\left(H^{(4)}_j \right) - \rho\left(H^{(3)}_j \right) \label{eq:433_proof_13} \\ & = & \rho\left(A^{(4)}_{j,j} \right) - \rho\left(A^{(3)}_{j,j} \right), \label{eq:433_proof_14} \end{eqnarray} where in \eqref{eq:433_proof_12} we have used the fact that for any two subspaces $W_1, W_2$, $\text{dim}(W_1 + W_2) = \text{dim}(W_1) + \text{dim}(W_2) - \text{dim}(W_1 \cap W_2)$. Equation \eqref{eq:433_proof_13} follows from the definition of $H^{(3)}_j$, while \eqref{eq:433_proof_14} follows from \eqref{eq:433_proof_0a} and \eqref{eq:433_proof_0}. This completes the proof of \eqref{eq:433_proof_8}. Equation \eqref{eq:433_proof_9} now follows directly from \eqref{eq:433_proof_6}. We will next prove \eqref{eq:433_proof_10} which is the second claim in the lemma. Towards this, observe from \eqref{eq:433_proof_0b} that we have $\mathcal{S}\left(A^{(3)}_{j,j}\right) \subseteq \sum_{\ell = 1}^{j-1}\mathcal{S}\left(A^{(4)}_{j,\ell}\right)$ and thus, we get that \begin{eqnarray} \rho \left(A^{(3)}_{j,j}\right) & \leq & \text{dim} \left( \sum_{\ell = 1}^{j-1}\mathcal{S}\left(A^{(4)}_{j,\ell}\right) \right). \label{eq:433_proof_15} \end{eqnarray} The right hand side of \eqref{eq:433_proof_15} can be upper bounded as follows: \begin{eqnarray} \text{dim} \left( \sum_{\ell = 1}^{j-1}\mathcal{S}\left(A^{(4)}_{j,\ell}\right) \right) & = & \text{dim} \left( \sum_{\ell = 1}^{j-2}\mathcal{S}\left(A^{(4)}_{j,\ell}\right) + \mathcal{S}\left(A^{(4)}_{j,j-1}\right)\right) \label{eq:433_proof_16} \\ & = & \text{dim} \left( \sum_{\ell = 1}^{j-2}\mathcal{S}\left(A^{(4)}_{j,\ell}\right) \right) + \text{dim} \left( \mathcal{S}\left(A^{(4)}_{j,j-1}\right)\right) - \text{dim} \left( \sum_{\ell = 1}^{j-2}\mathcal{S}\left(A^{(4)}_{j,\ell}\right) \cap \mathcal{S}\left(A^{(4)}_{j,j-1}\right)\right) \label{eq:433_proof_17} \\ & \leq & \text{dim} \left( \sum_{\ell = 1}^{j-2}\mathcal{S}\left(A^{(4)}_{j,\ell}\right) \right) + \text{dim} \left( \mathcal{S}\left(A^{(4)}_{j,j-1}\right)\right) - \text{dim} \left( \mathcal{S}\left(A^{(3)}_{j,j-1}\right)\right), \label{eq:433_proof_18} \\ & = & \text{dim} \left( \sum_{\ell = 1}^{j-2}\mathcal{S}\left(A^{(4)}_{j,\ell}\right) \right) + \rho\left(A^{(4)}_{j,j-1}\right) - \rho\left(A^{(3)}_{j,j-1}\right), \label{eq:433_proof_19} \end{eqnarray} where \eqref{eq:433_proof_18} follows from \eqref{eq:433_proof_0b}. The term $\text{dim} \left( \sum_{\ell = 1}^{j-2}\mathcal{S}\left(A^{(4)}_{j,\ell}\right) \right)$ appearing in \eqref{eq:433_proof_19} can be further upper bounded (for the case when $j=4$. If $j=3$ \eqref{eq:433_proof_19} completes the proof) by following a similar sequence of steps as in \eqref{eq:433_proof_16} - \eqref{eq:433_proof_19}. Combining with \eqref{eq:433_proof_15}, we eventually get that \begin{eqnarray} \rho \left(A^{(3)}_{j,j}\right) & \leq & \sum_{\ell=1}^{j-1}\rho\left(A^{(4)}_{j,\ell} \right) \ - \ \sum_{\ell=2}^{j-1} \rho\left(A^{(3)}_{j,\ell} \right), \ 3 \leq j \leq 4. \end{eqnarray} This completes the proof of \eqref{eq:433_proof_10}. \end{proof} \vspace{0.1in} We will now use the result in Lemma \ref{lem:433_intersections} to get a lower bound on the rank of $H^{(3)}$. The steps that we follow here are similar to those appearing in the calculation of the rank of $H^{(4)}$. Thus, let us define the quantities $\gamma_j, 2 \leq j \leq 4$ such that \begin{eqnarray} \gamma_2 & = & \rho\left(H^{(3)}_2\right), \\ \gamma_j & = & \rho\left(H^{(3)}|_{\{2, \ldots, j \}}\right) - \rho\left(H^{(3)}|_{ \{2, \ldots, j-1 \} }\right), \ 3 \leq j \leq 4. \label{eq:433_proof_11r} \end{eqnarray} The quantity $\gamma_2$ is given by \begin{eqnarray} \gamma_2 & = & \rho\left(H^{(3)}_2\right) \\ & = & \rho\left(A^{(3)}_{2,2}\right) \label{eq:433_proof_11aa} \\ & = & \rho\left(A^{(4)}_{2,2} \right) \ - \ \left\{ \left(\rho\left(A_{2,2}^{(4)}\right) - \rho\left(A_{2,1}^{(4)}\right)\right)^+ + \alpha_2\right\} \label{eq:433_proof_11a}, \end{eqnarray} where \eqref{eq:433_proof_11aa} and \eqref{eq:433_proof_11a} follow from \eqref{eq:433_proof_0} and \eqref{eq:433_proof_9}, respectively. The quantities $\gamma_j, 3 \leq j \leq 4 $ can be lower bounded in the same way we lower bounded $\delta_j, 2 \leq j \leq 4$ in \eqref{eq:433_proof_5a}. Thus, we get that \begin{eqnarray} \gamma_j & \geq & \left(\rho\left(A_{j,j}^{(3)}\right) - \sum_{\ell=2}^{j-1}\rho\left(A_{j,\ell}^{(3)}\right)\right)^+ \\ & \geq & \rho\left(A_{j,j}^{(3)}\right) - \sum_{\ell=2}^{j-1}\rho\left(A_{j,\ell}^{(3)}\right) \\ & \geq & 2\rho\left(A_{j,j}^{(3)}\right) - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(4)}\right) \label{eq:433_proof_12r}\\ & = & 2\left[ \rho\left(A^{(4)}_{j,j} \right) \ - \ \left\{ \left(\rho\left(A_{j,j}^{(4)}\right) - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(4)}\right)\right)^+ + \alpha_j\right\}\right] - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(4)}\right), \ 3 \leq j \leq 4, \label{eq:433_proof_13r} \end{eqnarray} where \eqref{eq:433_proof_12r} and \eqref{eq:433_proof_13r} follow from \eqref{eq:433_proof_10} and \eqref{eq:433_proof_9}, respectively. The rank of the matrix $H^{(3)}$ is now given by \begin{eqnarray} \rho\left(H^{(3)}\right) & = &\sum_{j=2}^{4}\gamma_j, \label{eq:proof_433_rankH3_short} \end{eqnarray} where $\gamma_2$ is given by \eqref{eq:433_proof_11a}, and $\gamma_3, \gamma_4$ are lower bounded as given by \eqref{eq:433_proof_13r}. \vspace{0.1in} \subsubsection{Comparison of the ranks of the matrices $H^{(3)}$ and $H^{(4)}$} We are now in a position to compare the ranks of the matrices $H^{(3)}$ and $H^{(4)}$. Recall from \eqref{eq:433_proof_7} that the rank of the matrix $H^{(4)}$ is given by \begin{eqnarray} \rho\left( H^{(4)} \right) & = & \rho\left(A^{(4)}_{1,1}\right) \ + \ \sum_{j=2}^{4} \left\{ \left(\rho\left(A_{j,j}^{(4)}\right) - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(4)}\right)\right)^+ + \alpha_j \right\}. \label{eq:433_proof_14r} \end{eqnarray} The goal is to obtain a lower bound on $\sum_{j=2}^{4}\alpha_j$ via the comparison $\rho\left( H^{(4)} \right) \geq \rho\left( H^{(3)} \right)$, and then use this lower bound in \eqref{eq:433_proof_14r} to get the desired lower bound on $\rho\left( H^{(4)} \right)$. One can verify that the comparison yields the following lower bound on $\sum_{j=2}^{4}\alpha_j$: \begin{eqnarray} \sum_{j=2}^{4}\alpha_j & \geq & \frac{1}{3}\left\{ -\rho\left(A^{(4)}_{1,1}\right) + \rho\left(A^{(4)}_{2,2}\right) + 2\sum_{j=3}^{4}\rho\left(A^{(4)}_{j,j}\right) - \right. \nonumber \\ & & \left. \left[2\left( \rho\left(A^{(4)}_{2,2}\right) - \rho\left(A^{(4)}_{2,1}\right)\right)^+ + 3\sum_{j=3}^{4}\left( \rho\left(A^{(4)}_{j,j}\right) - \sum_{\ell=1}^{j-1}\rho\left(A^{(4)}_{j,\ell}\right)\right)^+ + \sum_{j=3}^{4} \sum_{\ell=1}^{j-1}\rho\left(A^{(4)}_{j,\ell}\right)\right] \right\} \label{eq:433_proof_15r} \end{eqnarray} An upper bound on the rank of $H^{(4)}$ is now obtained by substituting \eqref{eq:433_proof_15r} back in \eqref{eq:433_proof_14r}. The result is stated formally in the following theorem: \vspace{0.1in} \begin{thm} \label{thm:433_rankH4} \begin{eqnarray} \rho\left( H^{(4)} \right) & \geq & \frac{1}{3}\left\{ 2\sum_{j=1}^{4}\rho\left(A^{(4)}_{j,j}\right) - \sum_{j=2}^{4} \sum_{\ell=1}^{j-1}\rho\left(A^{(4)}_{j,\ell}\right)\right\}. \label{eq:433_proof_16r} \end{eqnarray} \end{thm} Finally, to get the bound in Theorem \ref{thm:433}, we invoke the facts that $\rho\left(A^{(4)}_{j,j}\right) = \alpha, 1 \leq j \leq 4$ and $\rho\left(A^{(4)}_{i,j}\right) \leq \beta, \ 1 \leq i, j \leq 4, i \neq j$. Using these expressions in \eqref{eq:433_proof_16r}, we get that \begin{eqnarray} \rho\left( H^{(4)} \right) & \geq & \frac{1}{3}\left\{ 2\sum_{j=1}^{4}\alpha - \sum_{j=2}^{4} (j-1)\beta\right\} \\ & = & \frac{8\alpha - 6\beta}{3}. \end{eqnarray} The use of the ceil function in Theorem \ref{thm:433} is justified by the fact the ranks are integers. \vspace{0.1in} \begin{note} \label{rem:resuse433} Note that in the preceding discussion, we never used the facts that $\rho\left(A^{(4)}_{j,j}\right) = \alpha, 1 \leq j \leq 4$ and $\rho\left(A^{(4)}_{i,j}\right) \leq \beta, \ 1 \leq i, j \leq 4, i \neq j$ till (including) Theorem \ref{thm:433_rankH4}. The lower bound in Theorem \ref{thm:433_rankH4} holds for any matrix $H^{(4)} = [H^{(4)}_1 \ H^{(4)}_2 \ H^{(4)}_3 \ H^{(4)}_4 ] = \left( A^{(4)}_{i,j} , 1 \leq i, j \leq 4 \ \right )$ having the following properties: \begin{enumerate}[1.] \item For any $j, 1 \leq j \leq 4$, the columns of $H^{(4)}_j$ are all linearly independent, \item $\rho\left( H^{(4)}_j \right) = \rho\left( A^{(4)}_{j,j} \right), \ 1 \leq j \leq 4$. \end{enumerate} As we will see, this fact will enable us to reuse Theorem \ref{thm:433_rankH4} in a certain way, for the case of $(5, 4, 4)$ as well. \end{note} \section{The Trade-off of $(5, 4, 4)$ ER Linear Regenerating Codes} \label{sec:544} In this section, we will prove Theorem \ref{thm:new_bound_k_eq_d} for the case of $(5, 4, 4) (\alpha, \beta)$ ER linear regenerating codes. The proof for the case of $(5, 4, 4)$ is essentially built on top of the proof for the case of $(4, 3, 3)$ and involves one additional idea which was not present (rather, not needed) in the case of $(4, 3, 3)$. As we will see in Section \ref{sec:general}, this extra step is the key to deriving the proof for any general $n$. We will keep the focus of this section on the main steps of the proof without giving detailed calculations, wherever such calculations resemble those for the case of $(4, 3, 3)$. Like in the case of $(4, 3, 3)$, we begin with a re-statement of Theorem \ref{thm:new_bound_k_eq_d} for the case of $(5, 4, 4)$, where the result is stated in terms of $\text{rank}(H)$. \vspace{0.1in} \begin{thm} \label{thm:544} Consider an exact repair linear regenerating code $\mathcal{C}$, having parameters $(n=5, k=4, d=4,), (\alpha, \beta)$. Let the matrix $H$ correspond to the dual of the code $\mathcal{C}$. Then, the rank of the matrix $H$ is lower bounded by \begin{eqnarray} \label{eq:bound_rank_H_544} \text{rank}(H) & \geq & \left \{ \begin{array}{c} \left \lceil \frac{10(\alpha - \beta)}{3} \right \rceil, \ 2\beta \leq \alpha \leq 4\beta \\ \\ \left \lceil \frac{15\alpha - 10\beta}{6} \right \rceil, \ \frac{4}{3}\beta \leq \alpha \leq 2\beta \\ \\ 2\alpha - \beta, \ \beta \leq \alpha \leq \frac{4}{3}\beta \end{array} \right. . \end{eqnarray} \end{thm} \vspace{0.1in} Like in the of case of $(4, 3, 3)$, the fact $\text{rank}(H) \geq 2\alpha - \beta, \beta \leq \alpha \leq 2\beta $ follows from our proof of FR-trade-off, see \eqref{eq:dimakis_proof_4}. Thus, notice that in order to prove Theorem \ref{thm:544}, it suffices to prove the following bounds on $\text{rank}(H)$ individually, without considering any particular range of $\alpha$: \begin{eqnarray} \textit{Bound 1} : \ \ \text{rank}(H) & \geq & \left \lceil \frac{10(\alpha - \beta)}{3} \right \rceil \label{eq:bound1_544} \\ \textit {Bound 2} : \ \ \text{rank}(H) & \geq & \left \lceil \frac{15\alpha - 10\beta}{6} \right \rceil \label{eq:bound2_544}. \end{eqnarray} We will now separately illustrate the main steps involved in the proofs of \eqref{eq:bound1_544} and \eqref{eq:bound2_544}. \vspace{0.1in} \subsection{Proof of \eqref{eq:bound1_544}} \label{sec:proof_bound1_544} The bound in \eqref{eq:bound1_544} can be derived exactly in the same way we derived \eqref{eq:433_proof_4} for the case of $(4, 3, 3)$. Thus, we define the matrices $H^{(5)}$ and $H^{(4)}$ (in the same way we defined the matrices $H^{(4)}$ and $H^{(3)}$ for the case of $(4, 3, 3)$) and make the comparison $\rho\left(H^{(5)}\right) \geq \rho\left(H^{(4)}\right)$. More formally, the matrices are defined as follows : \begin{eqnarray} H^{(5)} & = & H_{repair}, \end{eqnarray} where $H_{repair}$ is as given by \eqref{eq:Hrepair}. Also, let the matrix $H^{(5)}_j$ denote the $j^{\text{th}}$ thick column of $H^{(5)}, 1 \leq j \leq 5$, i.e., $H^{(5)} = [H^{(5)}_1 \ H^{(5)}_2 H^{(5)}_3 \ H^{(5)}_4 \ H^{(5)}_5]$. Next, define the matrices $H^{(4)}_j, 2 \leq j \leq 5$, such that the columns of $H^{(4)}_j$ form a basis for the vector space $\mathcal{S}\left(H^{(5)}_j\right) \cap \mathcal{S}\left(H^{(5)}|_{[j-1]}\right)$. Also define the matrix $H^{(4)}$ as \begin{eqnarray} \label{eq:proof544_a} H^{(4)} = [H^{(4)}_2 H^{(4)}_3 \ H^{(4)}_4 \ H^{(4)}_5]. \end{eqnarray} Both the matrices $H^{(5)}_j$ and $H^{(4)}_j$ are also associated with corresponding block-submatrix representations, i.e., \begin{eqnarray} H^{(5)} & = & \left( A^{(5)}_{i,j} , 1 \leq i, j \leq 5 \ \right ), \\ H^{(4)} & = & \left( A^{(4)}_{i,j} , 1 \leq i \leq 5, 2 \leq j \leq 5 \ \right ). \label{eq:proof544_b} \end{eqnarray} The calculation of $\rho\left(H^{(5)}\right)$ and $\rho\left(H^{(4)}\right)$ follow steps similar to those in \eqref{eq:433_proof_5z}-\eqref{eq:433_proof_7} and \eqref{eq:433_proof_8}-\eqref{eq:proof_433_rankH3_short},respectively. The subsequent comparison of the two ranks (i.e, $\rho(H^{(5)}) \geq \rho(H^{(4)})$) yields the following lower bound on the rank of $H^{(5)}$: \begin{eqnarray} \rho\left( H^{(5)} \right) & \geq & \frac{1}{3}\left\{ 2\sum_{j=1}^{5}\rho\left(A^{(5)}_{j,j}\right) - \sum_{j=2}^{5} \sum_{\ell=1}^{j-1}\rho\left(A^{(5)}_{j,\ell}\right)\right\} \label{eq:proof544_rankH5}. \end{eqnarray} Note that \eqref{eq:proof544_rankH5} is the analogue of the bound in Theorem \ref{thm:433_rankH4} (see \ref{eq:433_proof_16}), where the upper limits of summations have been changed from $4$ to $5$, and $\{A^{(4)}_{i, j}\}$ have been replaced by $\{A^{(5)}_{i,j}\}$. Now, we invoke the facts that $\rho\left(A^{(5)}_{j,j}\right) = \alpha, 1 \leq j \leq 5$ and $\rho\left(A^{(5)}_{i,j}\right) \leq \beta, 1 \leq i, j \leq 5, \ i \neq j$. Thus, we get \begin{eqnarray} \rho\left( H^{(5)} \right) & \geq & \frac{1}{3}\left\{ 2\sum_{j=1}^{5}\alpha - \sum_{j=2}^{5} (j-1)\beta\right\} \\ & = & \frac{10(\alpha - \beta)}{3}. \end{eqnarray} This completes the proof of \eqref{eq:bound1_544}. \vspace{0.1in} \subsection{Proof of \eqref{eq:bound2_544}} \label{sec:proof_bound2_544} This is the part which is new in the case of $(5, 4, 4)$. For proving \eqref{eq:bound2_544}, consider the matrix $H^{(4)}$ as defined in \eqref{eq:proof544_a}, and whose block submatrix representation is as given by \eqref{eq:proof544_b}, i.e., \begin{eqnarray} H^{(4)} & = & [H^{(4)}_2 \ H^{(4)}_3 \ H^{(4)}_4 \ H^{(4)}_5 ] \ = \ \left( A^{(4)}_{i,j} , 1 \leq i \leq 5, 2 \leq j \leq 5 \ \right ), \label{eq:proof544_c} \end{eqnarray} Define the submatrix $\tilde{H}^{(4)}$ of $H^{(4)}$ as follows: \begin{eqnarray} \tilde{H}^{(4)} & = & \left( A^{(4)}_{i,j} , 2 \leq i \leq 5, 2 \leq j \leq 5 \ \right ), \end{eqnarray} i.e., $\tilde{H}^{(4)}$ is formed by excluding the first thick row of $H^{(4)}$. The matrix $\tilde{H}^{(4)}$ has exactly $4$ thick rows and 4 thick columns. Next, observe that the matrix $H^{(4)}$ satisfies the following properties: \begin{enumerate}[1.] \item For all $j, 2 \leq j \leq 5$, the columns of $H^{(4)}_j$ are linearly independent, and \item $\rho\left(H^{(4)}_j\right) = \rho\left(A^{(4)}_{j,j}\right), \ 2 \leq j \leq 5$. \end{enumerate} The first property follows directly from the definition of $H^{(4)}_j$ (since we know that the columns form a basis for a certain subspace), while the second property is the analogue of \eqref{eq:433_proof_0} for the case of $(5, 4, 4)$. Also, note that the above two properties together imply that the matrix $A^{(4)}_{j,j}, 2 \leq j \leq 5$ has full column rank. It then follows that the sub-matrix $\tilde{H}^{(4)}$ satisfies the following properties : \begin{enumerate}[1.] \item $\rho\left( \tilde{H}^{(4)}_j \right) = \rho\left( A^{(4)}_{j,j} \right), \ 2 \leq j \leq 5$. This follows as a result of the fact that the matrix $A^{(4)}_{j,j}$ has full column rank. \item For all $j, 2 \leq j \leq 5$, the columns of $\tilde{H}^{(4)}_j$ are all linearly independent. This follows because 1) the previous statement implies that $\rho\left( \tilde{H}^{(4)}_j \right) = \rho\left( H^{(4)}_j \right)$, and 2) the columns of $H^{(4)}_j$ are linearly independent. \end{enumerate} Remark \ref{rem:resuse433} now implies that Theorem \ref{thm:433_rankH4} can be used to lower bound the rank of the matrix $\tilde{H}^{(4)}$. The only thing that needs to be taken care of is the fact that in Theorem $\ref{thm:433_rankH4}$, the indices of the thick columns range from $1$ to $4$, where as for the matrix $\tilde{H}^{(4)}$, they range from $2$ to $5$. Accounting for this variation in Theorem \ref{thm:433_rankH4}, we get \begin{eqnarray} \rho\left( H^{(4)} \right) \ \geq \ \rho\left( \tilde{H}^{(4)} \right) & \geq & \frac{1}{3}\left\{ 2\sum_{j=2}^{5}\rho\left(A^{(4)}_{j,j}\right) - \sum_{j=3}^{5} \sum_{\ell=2}^{j-1}\rho\left(A^{(4)}_{j,\ell}\right)\right\}. \label{eq:proof544_d} \end{eqnarray} In summary, we have a got a new lower bound on the rank of $H^{(4)}$ and this new bound in \eqref{eq:proof544_d} is, in general, different from what is used to prove \eqref{eq:bound1_544}. We will now show that \eqref{eq:bound2_544} is obtained as necessary condition for satisfying $\rho(H^{(5)}) \geq \rho(H^{(4)})$, where $\rho(H^{(4)})$ is assumed to be lower bounded as in \eqref{eq:proof544_d}. Towards this, note from \eqref{eq:proof544_rankH5} that $\rho(H^{(5)})$ is given by \begin{eqnarray} \rho\left( H^{(5)} \right) & = & \rho\left(A^{(5)}_{1,1}\right) \ + \ \sum_{j=2}^{5} \left\{ \left(\rho\left(A_{j,j}^{(5)}\right) - \sum_{\ell=1}^{j-1}\rho\left(A_{j,\ell}^{(5)}\right)\right)^+ + \alpha_j \right\}. \label{eq:proof544_e} \end{eqnarray} Recall that \eqref{eq:proof544_e} is the same expression for $\rho(H^{(5)})$ that is used in the proof of $\eqref{eq:bound1_544}$. Given the two bounds on $\rho(H^{(5)})$ and $\rho(H^{(4)})$, the remaining sequence of steps that need to be carried out to complete proof of \eqref{eq:bound2_544} are similar to what we do for the proof of \eqref{eq:bound1_544}. These are given as follows: \begin{enumerate}[1.] \item Express the quantities $\{A^{(4)}_{i,j}\}$ appearing in \eqref{eq:proof544_d} in terms of $\{A^{(5)}_{i,j}\}$. This is accomplished via an analogue of Lemma \ref{lem:433_intersections} for the case of $(5, 4, 4)$. \item Next, obtain a lower bound on $\sum_{j=2}^{5}\alpha_j$ by invoking the comparison $\rho(H^{(5)}) \geq \rho(H^{(4)})$, where $\rho(H^{(5)})$ and $\rho(H^{(4)})$ are given by \eqref{eq:proof544_e} and \eqref{eq:proof544_d}, respectively. \item Finally, use the lower bound on $\sum_{j=2}^{5}\alpha_j$ back in \eqref{eq:proof544_e} to get a lower bound on $\rho(H^{(5)})$. \end{enumerate} We will defer the calculations of the above three steps until Section \ref{sec:general}, where we give the full proof of Theorem \ref{thm:new_bound_k_eq_d}. As we will see, the bound on $\rho(H^{(5)})$ that is obtained by after carrying out the above steps is given by \begin{eqnarray} \rho\left( H^{(5)} \right) & \geq & \frac{1}{6}\left\{ 3\sum_{j=1}^{5}\rho\left(A^{(5)}_{j,j}\right) - \sum_{j=2}^{5} \sum_{\ell=1}^{j-1}\rho\left(A^{(5)}_{j,\ell}\right)\right\} \label{eq:proof544_f}. \end{eqnarray} Given \eqref{eq:proof544_f}, we invoke the facts that $\rho\left(A^{(5)}_{j,j}\right) = \alpha, 1 \leq j \leq 5$ and $\rho\left(A^{(5)}_{i,j}\right) \leq \beta, 1 \leq i, j \leq 5, \ i \neq j$. Thus, we get \begin{eqnarray} \rho\left( H^{(5)} \right) & \geq & \frac{1}{6}\left\{ 3\sum_{j=1}^{5}\alpha - \sum_{j=2}^{5} (j-1)\beta\right\} \\ & = & \frac{15\alpha - 10\beta}{6}. \end{eqnarray} This completes the proof of \eqref{eq:bound2_544}. \vspace{0.1in} \begin{note} We would like to mention that the equivalent of Theorem \ref{thm:433_rankH4} for the case of $(5, 4, 4)$ involves putting together the three equations \eqref{eq:proof544_rankH5}, \eqref{eq:proof544_d} and \eqref{eq:proof544_f}. \end{note} \vspace{0.1in} \section{Proof of Theorem \ref{thm:new_bound_k_eq_d} for general $(n, k = n-1, d = n-1)$} \label{sec:general} \vspace{0.1in} In this section, we will prove Theorem \ref{thm:new_bound_k_eq_d} for any general $(n, k = n-1, d = n-1)$. We begin with a re-statement of Theorem \ref{thm:new_bound_k_eq_d} in terms of rank of $H$. \vspace{0.1in} \begin{thm} \label{thm:rankH_new_bound_k_eq_d} Consider an exact repair linear regenerating code $\mathcal{C}$, having parameters $(n, k = n-1, d = n-1), (\alpha, \beta)$. Let the matrix $H$ correspond to the dual of the code $\mathcal{C}$. Then, the rank of the matrix $H$ is lower bounded by \begin{eqnarray} \label{eq:bound_rank_H_gen} \text{rank}(H) & \geq & \left \{ \begin{array}{c} \left \lceil \frac{2rn\alpha - n(n-1)\beta}{r^2+r}\right \rceil, \ \frac{d\beta}{r} \leq \alpha \leq \frac{d\beta}{r-1}, \text{ where } 2 \leq r \leq n - 2 \\ \\ 2\alpha - \beta, \ \frac{d\beta}{n-1} \leq \alpha \leq \frac{d\beta}{n-2} \end{array} \right. . \end{eqnarray} \end{thm} \vspace{0.1in} Like we mentioned in the special cases of $(4, 3, 3)$ and $(5, 4, 4)$, the fact $\text{rank}(H) \geq 2\alpha - \beta, \beta \leq \alpha \leq 2\beta $ follows from our proof of FR-trade-off, see \eqref{eq:dimakis_proof_4}. Thus, in order to prove Theorem \ref{thm:rankH_new_bound_k_eq_d}, it suffices to prove the following bound on $\text{rank}(H)$, without considering any particular range of $\alpha$: \begin{eqnarray} \text{rank}(H) & \geq & \left \lceil \frac{2rn\alpha - n(n-1)\beta}{r^2+r}\right \rceil, 2 \leq r \leq n - 2. \label{eq:general_bound_to_prove} \end{eqnarray} Note that there are in fact $n-3$ bounds which needs to be established for the general case, where each bound corresponds to a value of $r$ in the range $ 2 \leq r \leq n-2$. A quick outline of the proof is provided next. We will consider the matrices $H^{(t)}, 3 \leq t \leq n$, where $H^{(n)} = H_{repair}$ corresponding to the matrix $H$, and where the matrix $H^{(t)}, 3 \leq t \leq n-1$ is defined based on matrix $H^{(t+1)}$. We will also have the relation $\rho(H^{(t)}) \geq \rho(H^{(t-1)}), 3 \leq t \leq n$. The bound in \eqref{eq:general_bound_to_prove} corresponding to the case of a general $r, 2 \leq r \leq n - 2$ will be obtained as necessary condition for satisfying the chain of inequalities given by \begin{eqnarray} \rho(H^{(n)}) & \geq & \rho(H^{(n-1)}) \ \geq \ \rho(H^{(n-2)}) \ \geq \ \ldots \ \geq \rho(H^{(n-(r-2))}) \ \geq \ \rho(H^{(n-(r-1))}). \end{eqnarray} Recall that for the case of $(5, 4, 4)$, we used the relations 1) $\rho(H^{(5)}) \geq \rho(H^{(4)})$ and 2) $\rho(H^{(5)}) \geq \rho(H^{(4)}) \geq \rho(H^{(3)})$ to prove \eqref{eq:bound1_544} and \eqref{eq:bound2_544}, respectively. The rest of the technical discussion in this section is divided as follows: \begin{enumerate}[1.] \item Formal definition of the matrices $H^{(t)}, 3 \leq t \leq n$, and also the associated block sub-matrix representations. \item Establish a generalization of Lemma \ref{lem:433_intersections} which in turn enables us to prove a generalization of Theorem \ref{thm:433_rankH4}. Arguments based on mathematical induction will be used in the proofs of both these generalizations. \item Deriving the bound in \eqref{eq:general_bound_to_prove} based on the generalization of Theorem \ref{thm:433_rankH4}. \end{enumerate} \vspace{0.1in} \subsection{Notation} \subsubsection{The Matrices $\{H^{(t)}\}$} As stated above, we assume that $H^{(n)} = H_{repair}$, where $H_{repair}$ is as defined by Lemma \ref{lem:H_repair}. Also assume that $H^{(n)}_j$ denotes the $j^{\text{th}}$ thick column of $H^{(n)}, 1 \leq j \leq n$, i.e., \begin{eqnarray} \label{eq:proofgen_Hndef} H^{(n)} & = & [H^{(n)}_1 \ H^{(n)}_2 \ \ldots \ H^{(n)}_n]. \end{eqnarray} The matrices $H^{(t)}, 3 \leq t \leq n-1$ are iteratively defined as follows: \vspace{0.1in} \begin{framed} \begin{enumerate}[Step 1.] \item Let $ t=n-1$. \item Define the matrices $H^{(t)}_j, n-t+1 \leq j \leq n$, such that the columns of $H^{(t)}_j$ form a basis for the vector space $\mathcal{S}\left(H^{(t+1)}_j\right) \cap \mathcal{S}\left(H^{(t+1)}|_{\{n-t, n-t+1, \ldots, j-1\}}\right)$. \item Define the matrix $H^{(t)}$ as \begin{eqnarray} \label{eq:proofgen_Htdef} H^{(t)} = [H^{(t)}_{n-t+1} \ H^{(t)}_{n-t+2} \ \ldots \ H^{(t)}_n]. \end{eqnarray} \item If $t \geq 4$, decrement $t$ by $1$ and go back to Step $2$. \end{enumerate} \end{framed} Steps $2$, $3$, $4$ are carried out in that sequence a total of $n-3$ times so that all the matrices $H^{(t)}, 3 \leq t \leq n-1$ get defined. Clearly, the ranks of the matrices $H^{(t)}, 3 \leq t \leq n$ are ordered as \begin{eqnarray} \label{eq:proofgen_rankorderHs} \rho(H^{(n)}) & \geq & \rho(H^{(n-1)}) \ \geq \ \ \ldots \ \geq \rho(H^{(4)}) \ \geq \ \rho(H^{(3)}). \end{eqnarray} \vspace{0.1in} \subsubsection{Block Sub-Matrix Representation of $H^{(t)}$} The block submatrix representation for the matrix $H^{(n)}$ is given by \eqref{eq:Hrepair}. Like we did in the case of $(4, 3, 3)$ and $(5, 4, 4)$, for easiness of notation, \eqref{eq:Hrepair} will be denoted as \begin{eqnarray} H^{(n)} & = & \left( A^{(n)}_{i,j} , 1 \leq i, j \leq n \ \right ), \end{eqnarray} where $A^{(n)}_{i,i} = I_{\alpha}, 1 \leq i \leq n$, and when $i \neq j$, we have added a superscript on $A_{i,j}$. Next, we introduce block sub-matrix representations for the other $n-3$ matrices $H^{(t)}, 3 \leq t \leq n-1$. The matrix $H^{(t)}$ will be identified as \begin{eqnarray} H^{(t)} & = & \left( A^{(t)}_{i,j} , 1 \leq i \leq n, n-t+1 \leq j \leq n \ \right ), \end{eqnarray} where $A^{(t)}_{i,j}$ is an $\alpha \times \rho(H^{(t)}_j)$ matrix (over $\mathbb{F}_q$) such that \begin{eqnarray} \mathcal{S}\left(A^{(t)}_{i,j} \right) & \subseteq & \mathcal{S}\left(A^{(t+1)}_{i,j} \right) \ \bigcap \ \sum_{\ell=n-t}^{j-1}\mathcal{S}\left(A^{(t+1)}_{i,\ell}\right). \label{eq:proofgen_1} \end{eqnarray} Note that \eqref{eq:proofgen_1} is a direct consequence of the definition of the matrix $H^{(t)}$. The following lemma gives additional properties of the matrices $\{A^{(t)}_{i,j}\}$. While the first part of the lemma is a generalization of \eqref{eq:433_proof_0}, the second and third parts together generalize Lemma \ref{lem:433_intersections}. \vspace{0.1in} \begin{lem} \label{lem:nkk_intersections} \begin{enumerate}[a)] \item \begin{eqnarray} \rho\left(H^{(t)}_j\right) & = & \rho\left( A^{(t)}_{j,j}\right), \ 3 \leq t \leq n, \ n-t+1 \leq j \leq n. \label{eq:proofgen_2} \end{eqnarray} \item \begin{eqnarray} \rho\left(A^{(t)}_{j,j} \right) & = & \rho\left(A^{(t+1)}_{j,j} \right) \ - \ \left\{ \rho\left(H^{(t+1)}|_{\{n-t, \ldots, j\}}\right) - \rho\left(H^{(t+1)}|_{\{n-t, \ldots, j-1\}}\right) \right\}, \nonumber \\ & & \hspace{2in} \ 3 \leq t \leq n-1, \ n-t+1 \leq j \leq n. \label{eq:proofgen_3} \end{eqnarray} \item \begin{eqnarray} \sum_{\ell=n-t+1}^{j-1} \rho\left(A^{(t)}_{j,\ell} \right) & \leq & \sum_{\ell=n-t}^{j-1}\rho\left(A^{(t+1)}_{j,\ell} \right) \ - \ \rho\left(A^{(t)}_{j,j} \right), \ 3 \leq t \leq n-1, \ n-t+2 \leq j \leq n. \label{eq:proofgen_4} \end{eqnarray} \end{enumerate} \end{lem} \begin{proof} See Appendix \ref{app:Prooflem} \end{proof} \vspace{0.1in} \vspace{0.1in} \subsection{Generalization of Theorem \ref{thm:433_rankH4}} \begin{thm} \label{thm:nkk_rankviainduction} Consider the matrices $\{H^{(t)}, 4 \leq t \leq n \}$ as defined together by \eqref{eq:proofgen_Hndef} and \eqref{eq:proofgen_Htdef}. Also, consider the associated block sub-matrix representations given by $H^{(t)} = \left( A^{(t)}_{i,j} , 1 \leq i \leq n, n-t+1 \leq j \leq n \ \right ), 4 \leq t \leq n$. Then, for any $s$ such that $1 \leq s \leq n-3$, and any $t$ such that $3+s \leq t \leq n$, the rank of the matrix $H^{(t)}$ is lower bounded by \begin{eqnarray} \rho\left( H^{(t)} \right) & \geq & \frac{2}{(s+1)(s+2)}\left\{ (s+1)\sum_{j=n-t+1}^{n}\rho\left(A^{(t)}_{j,j}\right) - \sum_{j=n-t+2}^{n} \sum_{\ell=n-t+1}^{j-1}\rho\left(A^{(t)}_{j,\ell}\right)\right\}. \label{eq:boundgen_rankH} \end{eqnarray} \end{thm} \begin{proof} A proof based on induction on the parameter $s$ appears in Appendix \ref{app:ProofThm}. The induction starts at $s=1$ and ends at $s = n-3$. \end{proof} \vspace{0.1in} Observe that if we set $n=4$ in Theorem \ref{thm:nkk_rankviainduction}, there is only one $(s, t)$ pair for which we get a bound via the theorem. The pair is given by $(s = 1, t = 4)$, and in this case, the bound in \eqref{eq:boundgen_rankH} is exactly same as what we have in \eqref{eq:433_proof_16r}. For the case when $n=5$, we get bounds for three pairs of $(s, t)$ given by $(s=1, t=5)$, $(s=1, t = 4)$ and $(s = 2, t = 5)$. The bounds obtained from Theorem \ref{thm:nkk_rankviainduction} for these three cases are given by \eqref{eq:proof544_rankH5}, \eqref{eq:proof544_d} and \eqref{eq:proof544_f}, respectively. \subsection{Proof of Theorem \ref{thm:rankH_new_bound_k_eq_d}} We will now give a proof of \eqref{eq:general_bound_to_prove} based on Theorem \ref{thm:nkk_rankviainduction}. Recall from our earlier discussion that proving \eqref{eq:general_bound_to_prove} is sufficient to prove Theorem \ref{thm:rankH_new_bound_k_eq_d}. For proving \eqref{eq:general_bound_to_prove}, we evaluate the bound in \eqref{eq:boundgen_rankH} for the $(n-3)$ pairs given by $(s, t=n), 1 \leq s \leq n-3$. Further, we also invoke the facts that $\rho\left(A^{(n)}_{j,j}\right) = \alpha, 1 \leq j \leq n$ and $\rho\left(A^{(n)}_{i,j}\right) \leq \beta, 1 \leq i, j \leq n, i \neq j$. Thus, we get that \begin{eqnarray} \rho\left( H^{(n)} \right) & \geq & \frac{2}{(s+1)(s+2)}\left\{ (s+1)\sum_{j=1}^{n}\rho\left(A^{(n)}_{j,j}\right) - \sum_{j=2}^{n} \sum_{\ell=1}^{j-1}\rho\left(A^{(n)}_{j,\ell}\right)\right\} \\ & \geq & \frac{2}{(s+1)(s+2)}\left\{ (s+1)\sum_{j=1}^{n}\alpha - \sum_{j=2}^{n} (j-1)\beta\right\} \\ & = & \frac{2(s+1)n\alpha - n(n-1)\beta}{(s+1)(s+2)}, \ 1 \leq s \leq n-3. \label{eq:proofgenz} \end{eqnarray} Finally, note that \eqref{eq:general_bound_to_prove} follows from \eqref{eq:proofgenz} by substituting $r = s+1$. This completes the proof of \eqref{eq:general_bound_to_prove} and thereby, also the proof of Theorem \ref{thm:rankH_new_bound_k_eq_d}. \bibliographystyle{IEEEtran}

https://arxiv.org/abs/1005.2079

Simulations of Weighted Tree Automata

Simulations of weighted tree automata (wta) are considered. It is shown how such simulations can be decomposed into simpler functional and dual functional simulations also called forward and backward simulations. In addition, it is shown in several cases (fields, commutative rings, Noetherian semirings, semiring of natural numbers) that all equivalent wta M and N can be joined by a finite chain of simulations. More precisely, in all mentioned cases there exists a single wta that simulates both M and N. Those results immediately yield decidability of equivalence provided that the semiring is finitely (and effectively) presented.

\section{Introduction} \label{sec:Intro} Weighted tree automata (or equivalently, weighted tree grammars) are widely used in applications such as model checking~\cite{abdjonmahors02} and natural language processing~\cite{knigra05}. They finitely represent mappings, called tree series, that assign a weight (taken from a semiring) to each tree. For example, a probabilistic parser would return a tree series that assigns to each parse tree its likelihood. Consequently, several toolkits~\cite{klamol01,maykni06,cle08} implement weighted tree automata. The notion of simulation that is used in this paper is a generalization of the simulations for unweighted and weighted (finite) string automata of~\cite{bloesi93,esikui01}. The aim is to relate structurally equivalent automata. The results of~\cite[Section~9.7]{bloesi93} and \cite{koz94} show that two unweighted string automata (i.e., potentially nondeterministic string automata over the \textsc{Boolean} semiring) are equivalent if and only if they can be connected by a finite chain of relational simulations, and that in fact \emph{functional} and \emph{dual functional} simulations are sufficient. Simulations for weighted string automata~(wsa) are called \emph{conjugacies} in~\cite{bealomsak05,bealomsak06}, where it is shown that for all fields, many rings including the ring~$\integer$ of integers, and the semiring~$\nat$ of natural numbers, two wsa are equivalent if and only if they can be connected by a finite chain of simulations. It is also shown that even a finite chain of functional (\emph{covering}) and dual functional (\emph{co-covering}) simulations is sufficient. The origin of those results can be traced back to the pioneering work of \textsc{Sch\"utzenberger} in the early 60's, who proved that every wsa over a field is equivalent to a minimal wsa that is simulated by every \emph{trim} equivalent wsa~\cite{berreu84}. Relational simulations of wsa are also studied in~\cite{buc08}, where they are used to reduce the size of wsa. The relationship between functional simulations and the \textsc{Milner}-\textsc{Park} notion of bisimulation~\cite{mil80,par81} is discussed in~\cite{bloesi93,buc08}. In this contribution, we investigate simulations for weighted (finite) tree automata~(wta). \textsc{Sch\"utzenberger}'s minimization method was extended to wta over fields in~\cite{aleboz89,boz91}. In addition, relational and functional simulations for wta are probably first used in~\cite{esi98,esi10b,hogmalmay07d}. Moreover, simulations can be generalized to presentations in algebraic theories~\cite{bloesi93}, which seems to cover all mentioned instances. Here, we extend the results of~\cite{bealomsak05,bealomsak06} to wta. In particular, we show that two wta over a ring, \textsc{Noetherian} semiring, or the semiring~$\nat$ are equivalent if and only if they are connected by a finite chain of simulations. Moreover, we discuss when the simulations can be replaced by functional and dual functional simulations, which are efficiently computable~\cite{hogmalmay07d}. Such results are important because they immediately yield decidability of equivalence provided that the semiring is finitely and effectively presented. \section{Preliminaries} \label{sec:Prelim} The set of nonnegative integers is~$\nat$. For every $k \in \nat$, the set $\{i \in \nat \mid 1 \leq i \leq k\}$ is simply denoted by~$[k]$. We write $\abs A$ for the cardinality of the set~$A$. A \emph{semiring} is an algebraic structure~${\cal A} = (A, \mathord+, \mathord\cdot, 0, 1)$ such that $(A, \mathord+, 0)$ and $(A, \mathord\cdot, 1)$ are monoids, of which the former is commutative, and $\cdot$~distributes both-sided over finite sums (i.e., $a \cdot 0 = 0 = 0 \cdot a$ for every $a \in A$ and $a \cdot (b + c) = ab + ac$ and $(b + c) \cdot a = ba + ca$ for every $a, b, c \in A$). The semiring~$\mathcal A$ is \emph{commutative} if $(A, \mathord\cdot, 1)$ is commutative. It is a \emph{ring} if for every $a \in A$ there exists an \emph{additive inverse}~$-a \in A$ such that $a + (-a) = 0$. The set~$U$ is the set $\{ a \in A \mid \exists b \in A \colon ab = 1 = ba\}$ of \emph{(multiplicative) units}. The semiring~${\cal A}$ is a \emph{semifield} if $U = A \setminus \{0\}$; i.e., for every $a \in A$ there exists a \emph{multiplicative inverse}~$a^{-1} \in A$ such that $aa^{-1} = 1 = a^{-1}a$. A \emph{field} is a semifield that is also a ring. For every $B \subseteq A$ let $\langle B\rangle_{\mathord+} = \{ b_1 + \dotsb + b_n \mid n \in \nat, \seq b1n \in B\}$. If $A = \langle B\rangle_{\mathord+}$, then ${\cal A}$~is \emph{additively generated by~$B$}. Finally, it is \emph{equisubtractive} if for every $a_1, a_2, b_1, b_2 \in A$ with $a_1 + b_1 = a_2 + b_2$ there exist $c_1, c_2, d_1, d_2 \in A$ such that (i)~$a_1 = c_1 + d_1$, (ii)~$b_1 = c_2 + d_2$, (iii)~$a_2 = c_1 + c_2$, and (iv)~$b_2 = d_1 + d_2$. The semiring~${\cal A}$ is \emph{zero-sum free} if $a + b = 0$ implies $0 \in \{a, b\}$ for every $a, b \in A$. Clearly, any nontrivial (i.e., $0 \neq 1$) ring is not zero-sum free. Moreover, ${\cal A}$~is \emph{zero-divisor free} if $a \cdot b = 0$ implies $a = 0 = b$ for every $a, b \in A$. All semifields are trivially zero-divisor free. Finally, the semiring~${\cal A}$ is \emph{positive} if it is zero-sum free and zero-divisor free. An infinitary sum operation~$\mathord{\sum}$ is a family~$(\mathord{\sum_I})_I$ such that $\mathord{\sum_I} \colon A^I \to A$. We generally write $\sum_{i \in I} a_i$ instead of $\sum_I (a_i)_{i \in I}$. The semiring~${\cal A}$ together with the infinitary sum operation~$\mathord{\sum}$ is \emph{complete}~\cite{eil74,hebwei98,gol99,kar04} if \begin{itemize} \item $\sum_{i \in \{j_1, j_2\}} a_i = a_{j_1} + a_{j_2}$ for all $j_1\neq j_2$ and $a_{j_1}, a_{j_2} \in A$, \item $\sum_{i \in I} a_i = \sum_{j \in J} \bigl( \sum_{i \in I_j} a_i \bigr)$ for every index set~$I$, partition $(I_j)_{j \in J}$ of~$I$, and $(a_i)_{i \in I} \in A^I$, and \item $a \cdot \bigl(\sum_{i \in I} a_i \bigr) = \sum_{i \in I} aa_i$ and $\bigl(\sum_{i \in I} a_i \bigr) \cdot a = \sum_{i \in I} a_ia$ for every $a \in A$, index set~$I$, and $(a_i)_{i \in I} \in A^I$. \end{itemize} An ${\cal A}$-\emph{semimodule} is a commutative monoid $(B, \mathord{+}, 0)$ together with an action $\mathord{\cdot} \colon A \times B \to B$, written as juxtaposition, such that for every $a, a' \in A$ and $b, b' \in B$ \begin{itemize} \item $(a + a') b = ab + a'b$ and $a(b + b') = ab + ab'$, \item $0 b = 0 = a 0$, $1 b = b$ and $(a \cdot a')b = a(a'b)$. \end{itemize} The semiring~${\cal A}$ is \textsc{Noetherian} if all subsemimodules of every finitely-generated ${\cal A}$-semimodule are again finitely-generated. In the following, we often identify index sets of the same cardinality. Let $X \in A^{I_1 \times J_1}$ and $Y \in A^{I_2 \times J_2}$ for some finite sets~$I_1, I_2, J_1, J_2$. We use upper-case letters (like~$C$, $D$, $E$, $X$, $Y$) for matrices and the corresponding lower-case letters for their entries. A matrix $X \in A^{I \times J}$ is \emph{relational} if $x_{ij} \in \{0, 1\}$ for every $i \in I$ and $j \in J$. Clearly, a relational matrix defines a relation $\rho_X \subseteq I \times J$ by $(i, j) \in \rho_X$ if and only if $x_{ij} = 1$ (and vice versa). Moreover, we call a relational matrix \emph{functional}, \emph{surjective}, or \emph{injective} if its associated relation has this property. As usual, we denote the \emph{transpose} of a matrix~$X$ by~$X^{\mathrm T}$, and we call $X$~\emph{nondegenerate} if its has no rows or columns of entirely zeroes. A \emph{diagonal} matrix~$X$ is such that $x_{ij} = 0$ for every $i \neq j$. Finally, the matrix~$X$ is invertible if there exists a matrix~$Y$ such that $XY = I = YX$ where $I$~is the unit matrix. The \textsc{Kronecker} product $X \otimes Y \in A^{(I_1 \times I_2) \times (J_1 \times J_2)}$ is such that $(X \otimes Y)_{(i_1, i_2), (j_1, j_2)} = x_{i_1, j_1} y_{i_2, j_2}$ for every $i_1 \in I_1$, $i_2 \in I_2$, $j_1 \in J_1$, and $j_2 \in J_2$. Clearly, the \textsc{Kronecker} product is, in general, not commutative and $(1) \in A^{[1]}$ acts both-sided as neutral element. We let $X^{0,\mathord{\otimes}} = (1)$ and $X^{i+1, \mathord{\otimes}} = X^{i, \mathord{\otimes}} \otimes X$ for every $i \in \nat$. Finally, let us move to trees. A \emph{ranked alphabet} is a finite set~$\Sigma$ together with a mapping~$\mathord{\rk} \colon \Sigma \to \nat$. We often just write~$\Sigma$ for a ranked alphabet and assume that the mapping~$\rk$ is implicit. We write $\Sigma_k = \{\sigma \in \Sigma \mid \rk(\sigma) = k\}$ for the set of all $k$-ary symbols. The set of $\Sigma$-\emph{trees} is the smallest set~$T_\Sigma$ such that $\sigma(\seq t1k) \in T_\Sigma$ for all $\sigma \in \Sigma_k$ and $\seq t1k \in T_\Sigma$. A \emph{tree series} is a mapping $\varphi \colon T_\Sigma \to A$. The set of all such tree series is denoted by~$\series A\Sigma$. For every $\varphi \in \series A\Sigma$ and $t \in T_\Sigma$, we often write~$(\varphi, t)$ instead of~$\varphi(t)$. Let ${\scriptstyle \Box}$ be a distinguished nullary symbol such that ${\scriptstyle \Box} \notin \Sigma$. A $\Sigma$-\emph{context}~$c$ is a tree of~$T_{\Sigma \cup \{{\scriptstyle \Box}\}}$, in which the symbol~${\scriptstyle \Box}$ occurs exactly once. The set of all $\Sigma$-contexts is denoted by~$C_\Sigma$. For every $c \in C_\Sigma$ and $t \in T_\Sigma$, we write~$c[t]$ for the $\Sigma$-tree obtained by replacing the unique occurrence of~${\scriptstyle \Box}$ in~$c$ by~$t$. A \emph{weighted tree automaton (over~${\cal A}$)}, for short: wta, is a system $(\Sigma, Q, \mu, F)$ with \begin{itemize} \item an input ranked alphabet~$\Sigma$, \item a finite set~$Q$ of \emph{states}, \item transitions $\mu = (\mu_k)_{k \in \nat}$ such that $\mu_k \colon \Sigma_k \to A^{Q^k \times Q}$ for every $k \in \nat$, and \item a \emph{final weight} vector $F \in A^Q$. \end{itemize} Next, let us introduce the semantics~$\sem M$ of~$M$. We first define the function $h_\mu \colon T_\Sigma \to A^Q$ for every $\sigma \in \Sigma_k$ and $\seq t1k \in T_\Sigma$ by \[ h_\mu(\sigma(\seq t1k)) = \bigl( h_\mu(t_1) \otimes \dotsm \otimes h_\mu(t_k) \bigr) \cdot \mu_k(\sigma) \enspace, \] where the final product~$\cdot$ is the classical matrix product. Then $(\sem M, t) = h_\mu(t) F$ for every $t \in T_\Sigma$, where the product is the usual inner (dot) product. Let $f \colon A \to \{0, 1\}$ be such that $f(0) = 0$ and $f(a) = 1$ for all $a \in A \setminus \{0\}$. The Boolean wta~$f(M)$ (i.e., essentially an unweighted tree automaton) corresponding to~$M$ is $(\Sigma, Q, \mu', F')$ where \begin{itemize} \item $\mu'_k(\sigma)_{w, q} = f(\mu_k(\sigma)_{w, q})$ for every $\sigma \in \Sigma_k$, $w \in Q^k$, and $q \in Q$, and \item $F'(q) = f(F(q))$ for every $q \in Q$. \end{itemize} The wta~$M$ is \emph{trim} if every state is accessible and co-accessible in~$f(M)$. In other words, the wta~$M$ is trim if $f(M)$ is trim. \section{Simulation} \label{sec:Sim} Simulations of automata were defined in~\cite{bloesi93,esikui01} in order to provide a structural characterization of equivalent automata. We will essentially follow the presentation of~\cite{bealomsak05} here. \begin{definition} \label{df:Conj} Let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$ be wta. Then $M$~\emph{simulates}~$N$ if there exists a matrix~$X \in A^{Q \times P}$ such that \begin{itemize} \item[(i)] $F = XG$, and \item[(ii)] $\mu_k(\sigma) X = X^{k, \mathord{\otimes}} \cdot \nu_k(\sigma)$ for every $\sigma \in \Sigma_k$. \end{itemize} The matrix~$X$ is called \emph{transfer matrix}, and we write $M \stackrel X\to N$ if $M$~simulates~$N$ with transfer matrix~$X$. \end{definition} \begin{figure} \centering \includegraphics{conj.mps} \caption{Illustration of simulation.} \label{fig:Conj} \end{figure} Note that $X^{k, \mathord{\otimes}}_{\word i1k, \word j1k} = \prod_{\ell = 1}^k x_{i_\ell, j_\ell}$. We illustrate Definition~\ref{df:Conj} in Fig.~\ref{fig:Conj}. If $M \stackrel X\to M'$ and $M' \stackrel Y\to N$, then $M \stackrel{XY}\to N$. Thus, simulations define a preorder on wta. \begin{theorem} \label{thm:Equiv} If $M$~simulates~$N$, then $M$~and~$N$ are equivalent. \end{theorem} \begin{proof} Let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$, and let $X \in A^{Q \times P}$ be a transfer matrix. We claim that $h_\mu(t)X = h_\nu(t)$ for every $t \in T_\Sigma$. We prove this by induction on~$t$. Let $t = \sigma(\seq t1k)$ for some $\sigma \in \Sigma_k$ and $\seq t1k \in T_\Sigma$. \begin{align*} &\phantom{{}={}} h_\mu(\sigma(\seq t1k)) X = \bigl( h_\mu(t_1) \otimes \dotsm \otimes h_\mu(t_k) \bigr) \cdot \mu_k(\sigma) X \\ &= \bigl( h_\mu(t_1) \otimes \dotsm \otimes h_\mu(t_k) \bigr) \cdot X^{k, \mathord{\otimes}} \cdot \nu_k(\sigma) = \bigl( h_\mu(t_1)X \otimes \dotsm \otimes h_\mu(t_k)X \bigr) \cdot \nu_k(\sigma) \\ &= \bigl( h_\nu(t_1) \otimes \dotsm \otimes h_\nu(t_k) \bigr) \cdot \nu_k(\sigma) = h_\nu(\sigma(\seq t1k)) \end{align*} With this claim, the statement can now be proved easily. For every $t \in T_\Sigma$ \[ (\sem M, t) = h_\mu(t)F = h_\mu(t)XG = h_\nu(t)G = (\sem N, t) \enspace. \tag*{\qed} \] \end{proof} \begin{lemma} \label{lm:Trim} Let $M$~and~$N$ be trim wta such $M \stackrel X\to N$. If (i)~$X$ is functional or (ii)~${\cal A}$ is positive, then $X$~is nondegenerate. \end{lemma} \begin{proof} Let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$. Moreover, let \[ J = \{ p \in P \mid \forall q \in Q \colon x_{qp} = 0\} \enspace. \] Then $\nu_k(\sigma)_{w, j} = 0$ for every $\sigma \in \Sigma_k$, $w \in (P \setminus J)^k$, and $j \in J$. This is seen as follows. Since $\mu_k(\sigma)X = X^{k, \mathord{\otimes}} \cdot \nu_k(\sigma)$ we obtain \begin{align} \label{eq:Trim1} \sum_{q \in Q} \mu_k(\sigma)_{\word q1k, q} \cdot x_{qj} = 0 &= \sum_{\seq p1k \in P} \Bigl( \prod_{\ell = 1}^k x_{q_\ell, p_\ell} \Bigr) \cdot \nu_k(\sigma)_{\word p1k, j} \end{align} for every $\seq q1k \in Q$ and $j \in J$. If $X$~is functional, then \[ \sum_{\seq p1k \in P} \Bigl( \prod_{\ell = 1}^k x_{q_\ell, p_\ell} \Bigr) \cdot \nu_k(\sigma)_{\word p1k, j} = \nu_k(\sigma)_{\rho_X(q_1) \dotsm \rho_X(q_k), j} = 0 \enspace, \] which proves the claim. On the other hand, if ${\cal A}$~is positive, then \eqref{eq:Trim1}~implies that $\prod_{\ell = 1}^k x_{q_\ell, p_\ell} \cdot \nu_k(\sigma)_{\word p1k, j} = 0$ for every $\seq p1k \in P$. Since for every $p_\ell \notin J$, there exists $q_\ell$ such that $x_{q_\ell, p_\ell} \neq 0$ and $\prod_{\ell = 1}^k x_{q_\ell, p_\ell} \neq 0$ by zero-divisor freeness, we conclude that $\nu_k(\sigma)_{\word p1k, j} = 0$ for every $\seq p1k \in P \setminus J$, which again proves the claim. Consequently, all states of~$J$ are unreachable. Since $N$ is trim, we conclude $J = \emptyset$, and thus, $X$~has no column of zeroes. If $X$~is functional, then it clearly has no row of zeroes. To prove that $X$ has no row of zeroes in the remaining case, let $I = \{ q \in Q \mid \forall p \in P \colon x_{qp} = 0\}$. Then $F_i = 0$ and $\mu_k(\sigma)_{\word q1k, q} = 0$ for every $\sigma \in \Sigma_k$, $q \in Q \setminus I$, $\seq q1k \in Q$, and $i \in I$ such that $q_\ell = i$ for some $\ell \in [k]$. Clearly, $F_i = \sum_{p \in P} x_{ip} G_p = 0$ for every $i \in I$. Moreover, since $\mu_k(\sigma)X = X^{k, \mathord{\otimes}} \cdot \nu_k(\sigma)$ we obtain \begin{align} \label{eq:Trim2} \sum_{q \in Q} \mu_k(\sigma)_{\word q1k, q} \cdot x_{qp} &= \sum_{\seq p1k \in P} \Bigl( \prod_{\ell = 1}^k x_{q_\ell, p_\ell} \Bigr) \cdot \nu_k(\sigma)_{\word p1k, p} = 0 \end{align} for every $\seq q1k \in Q$, $p \in P$, and $i \in I$ such that $q_\ell = i$ for some $\ell \in [k]$. Since ${\cal A}$ is positive, \eqref{eq:Trim2}~implies that $\mu_k(\sigma)_{\word q1k, q} \cdot x_{qp} = 0$ for every $q \in Q$. However, for all $q \in Q \setminus I$, there exists $p \in P$ such that $x_{qp} \neq 0$ because $q \notin I$. Consequently, $\mu_k(\sigma)_{\word q1k, q} = 0$ by zero-divisor freeness, which proves the claim. Thus, all states of~$I$ are unreachable. Since $M$ is trim, we conclude $I = \emptyset$, and thus, $X$~has no row of zeroes. \qed \end{proof} \begin{definition}[{\protect{see~\cite[Def.~1]{hogmalmay07d}}}] \label{df:ForSim} Let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$ be wta. A surjective function~$\rho \colon Q \to P$ is a \emph{forward simulation} from~$M$ to~$N$ if \begin{itemize} \item[(i)] $F_q = G_{\rho(q)}$ for every $q \in Q$, and \item[(ii)] for every $p \in P$, $\sigma \in \Sigma_k$, and $\seq q1k \in Q$ \[ \sum_{q \in Q \colon \rho(q) = p} \mu_k(\sigma)_{\word q1k, q} = \nu_k(\sigma)_{\rho(q_1) \dotsm \rho(q_k), p} \enspace. \] \end{itemize} Finally, we say that \emph{$M$~forward simulates~$N$}, written $M \twoheadrightarrow N$, if there exists a forward simulation from~$M$ to~$N$. \end{definition} \begin{lemma} \label{lm:FSim} Let $M$~and~$N$ be wta such that $N$~is trim. Then $M \twoheadrightarrow N$ if and only if there exists a functional transfer matrix~$X$ such that $M \stackrel X\to N$. \end{lemma} \begin{proof} Let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$. First suppose that $M \stackrel X\to N$ with functional~$X \in A^{Q \times P}$. Then $\rho_X \colon Q \to P$ is a surjective function by Lemma~\ref{lm:Trim}. Conversely, if $M \twoheadrightarrow N$ with the forward simulation $\rho \colon Q \to P$, then $\rho$ induces a surjective functional matrix $X \in A^{Q \times P}$ such that $\rho_X = \rho$. Let $X \in A^{Q \times P}$ be a surjective, functional matrix. It remains to prove that the conditions that (1)~$X$~is a transfer matrix and (2)~$\rho_X$ is a forward simulation are equivalent. We discuss the two items of Definitions \ref{df:Conj}~and~\ref{df:ForSim} separately. \begin{itemize} \item[(i)] $F = XG$ if and only if $F_q = G_{\rho(q)}$ for every $q \in Q$. \item[(ii)] for every $\sigma \in \Sigma_k$, $\seq q1k \in Q$, and $p \in P$ \begin{align*} (\mu_k(\sigma)X)_{\word q1k, p} &= \sum_{q \in Q \colon \rho_X(q) = p} \mu_k(\sigma)_{\word q1k, q} \\ (X^{k, \mathord{\otimes}} \cdot \nu_k(\sigma))_{\word q1k, p} &= \nu_k(\sigma)_{\rho_X(q_1) \dotsm \rho_X(q_k), p} \enspace. \end{align*} \end{itemize} Thus, $X$~is a transfer matrix if and only if $\rho_X$~is a forward simulation, which proves the statement. \qed \end{proof} \begin{definition}[{\protect{see~\cite[Def.~16]{hogmalmay07d}}}] \label{df:BackSim} Let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$ be wta. A surjective function~$\rho \colon Q \to P$ is a \emph{backward simulation} from~$M$ to~$N$ if \begin{itemize} \item[(i)] $\sum_{q \in Q \colon \rho(q) = p} F_q = G_p$ for every $p \in P$, and \item[(ii)] for every $q \in Q$, $\sigma \in \Sigma_k$, and $\seq p1k \in P$ \[ \sum_{\substack{\seq q1k \in Q \\ \rho(q_1) = p_1, \dotsc, \rho(q_k) = p_k}} \mu_k(\sigma)_{\word q1k, q} = \nu_k(\sigma)_{\word p1k, \rho(q)} \enspace. \] \end{itemize} Finally, we say that \emph{$M$~backward simulates~$N$}, written $M \twoheadleftarrow N$, if there exists a backward simulation from~$M$ to~$N$. \end{definition} \begin{lemma} \label{lm:BSim} Let $M$~and~$N$ be wta such that $N$ is trim. Then $M \twoheadleftarrow N$ if and only if there exists a transfer matrix~$X$ such that $X^{\mathrm T}$ is functional and $N \stackrel X\to M$. \end{lemma} \begin{proof} Let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$. First, suppose that $N \stackrel X\to M$ with the transfer matrix~$X \in A^{P \times Q}$ such that $X^{\mathrm T}$~is functional. Let $Y = X^{\mathrm T}$. Then $\rho_Y \colon Q \to P$ is a surjective function by Lemma~\ref{lm:Trim}. Conversely, if $M \twoheadleftarrow N$ with the backward simulation $\rho \colon Q \to P$, then $\rho$~again induces a surjective, functional matrix $X \in A^{Q \times P}$ such that $\rho_X = \rho$. Let $X \in A^{Q \times P}$ be a surjective, functional matrix. It remains to prove that the conditions that (1)~$X^{\mathrm T}$~is a transfer matrix and (2)~$\rho_X$ is a backward simulation are equivalent. We discuss the two items of Definitions \ref{df:Conj}~and~\ref{df:BackSim} separately. \begin{itemize} \item[(i)] $G = X^{\mathrm T}F$ if and only if $G_p = \sum_{q \in Q \colon \rho_X(q) = p} F_q$ for every $p \in P$. \item[(ii)] for every $\sigma \in \Sigma_k$, $\seq p1k \in P$, and $q \in Q$ \begin{align*} (\nu_k(\sigma)X^{\mathrm T})_{\word p1k, q} &= \nu_k(\sigma)_{\word p1k, \rho_X(q)} \\ ((X^{\mathrm T})^{k, \mathord{\otimes}} \cdot \mu_k(\sigma))_{\word p1k, q} &= \sum_{\substack{\seq q1k \in Q \\ \rho_X(q_1) = p_1, \dotsc, \rho_X(q_k) = p_k}} \mu_k(\sigma)_{\word q1k, q} \enspace. \end{align*} \end{itemize} Thus, $X^{\mathrm T}$~is a transfer matrix if and only if $\rho_X$~is a backward simulation, which proves the statement. \qed \end{proof} \begin{lemma} \label{lm:Help2} If $A = \langle U \rangle_{\mathord{+}}$, then for every $X \in A^{Q \times P}$ there exist matrices $C, E, D$ such that \begin{itemize} \item $X = CED$, \item $C^{\mathrm T}$~and~$D$ are functional, and \item $E$~is an invertible diagonal matrix. \end{itemize} If (i)~$X$ is nondegenerate or (ii)~${\cal A}$ has (nontrivial) zero-sums, then $C^{\mathrm T}$~and~$D$ can be chosen to be surjective. \end{lemma} \begin{proof} For every $q \in Q$ and $p \in P$, let $\ell_{qp} \in \nat$ and $u_{qp1}, \dotsc, u_{qp\ell_{qp}} \in U$ be such that $x_{qp} = \sum_{i = 1}^{\ell_{qp}} u_{qpi}$. In addition, let \[ J = \{ (q, i, p) \mid q \in Q, p \in P, i \in [\ell_{qp}] \} \enspace. \] Finally, let $\pi_1 \colon J \to Q$ and $\pi_3 \colon J \to P$ be such that $\pi_1(\langle q, i, p\rangle) = q$ and $\pi_3(\langle q, i, p\rangle) = p$ for every $\langle q, i, p\rangle \in J$. Then we set $C^{\mathrm T}$~and~$D$ to the functional matrices represented by $\pi_1$~and~$\pi_3$, respectively. Together with the diagonal matrix~$E$ such that $e_{\langle q, i, p\rangle, \langle q, i, p\rangle} = u_{qpi}$ for every $\langle q, i, p\rangle \in J$, we obtain $X = CED$. For every $q \in Q$ and $p \in P$ we have \[ \sum_{j_1, j_2 \in J} c_{q, j_1} e_{j_1, j_2} d_{j_2, p} = \sum_{i = 1}^{\ell_{qp}} e_{\langle q, i, p\rangle, \langle q, i, p\rangle} = \sum_{i = 1}^{\ell_{qp}} u_{qpi} = x_{qp} \enspace. \] It is clear that $C^{\mathrm T}$~and~$D$ are functional matrices. Moreover, $E$~is an invertible diagonal matrix because $EE^{-1} = I = E^{-1}E$ where $E^{-1}$ is the matrix obtained from~$E$ by inverting each nonzero element. If $X$~is nondegenerate, then $C^{\mathrm T}$~and~$D$ are surjective. Finally, if there are zero-sums, then for every $q \in Q$ and $p \in P$ there exist $u, v \in U$ such that $x_{qp} = 0 = u + v$, which yields that we can choose $\ell_{qp} > 0$. This completes the proof. \qed \end{proof} \begin{lemma} \label{lm:Sol} Let ${\cal A}$~be equisubtractive. Moreover, let $R \in A^Q$ and $C \in A^P$ be such that $\sum_{q \in Q} r_q = \sum_{p \in P} c_p$. Then there exists a matrix $X \in A^{Q \times P}$ with row sums~$R$ and column sums~$C$; i.e., $\sum_{q \in Q} x_{qp} = c_p$ for every $p \in P$ and $\sum_{p \in P} x_{qp} = r_q$ for every $q \in Q$. \end{lemma} \begin{proof} If $\abs Q \leq 1$ or $\abs P \leq 1$, then the statement is trivially true. Otherwise, select $i \in Q$ and $j \in P$, and let $Q' = Q \setminus \{i\}$ and $P' = P \setminus \{j\}$. By assumption \[ \sum_{q \in Q'} r_q + r_i = \sum_{p \in P'} c_p + c_j \enspace. \] Thus, by equisubtractivity there exist $a, c'_j, r'_i, x_{ij} \in A$ such that \[ \sum_{q \in Q'} r_q = a + c'_j \qquad r_i = r'_i + x_{ij} \qquad \sum_{p \in P'} c_p = a + r'_i \qquad c_j = c'_j + x_{ij} \enspace. \] Continuing the row decomposition, we obtain $Y \in A^{Q'}$ and $R' \in A^{Q'}$ such that $r_q = r'_q + y_q$ for every $q \in Q'$ and $\sum_{q \in Q'} r'_q = a$. In a similar manner we perform column decomposition to obtain $Y' \in A^{P'}$ and $C' \in A^{P'}$ such that $c_p = c'_p + y'_p$ for every $p \in P'$ and $\sum_{p \in P'} c'_p = a$. Thus, by the induction hypothesis, there exists a matrix $X' \in A^{Q' \times P'}$ with row sums~$R'$ and column sums~$C'$ because $\sum_{q \in Q'} r'_q = \sum_{p \in P'} c'_p$. Then the matrix \[ X = \begin{pmatrix} \; & & \; & \\ & X' & & Y \\ & & & \\ & (Y')^{\mathrm T} & & x_{ij} \end{pmatrix} \] obviously has the required row and column sums $R$~and~$C$, respectively. \qed \end{proof} \begin{lemma} \label{lm:3} If $X \in A^{Q \times P}$~is functional (respectively, invertible diagonal), then $X^{k, \mathord{\otimes}}$ is functional (respectively, invertible diagonal) for every $k \in \nat$. \end{lemma} \begin{proof} Trivial. \qed \end{proof} \begin{theorem} \label{thm:2} Let $M$~and~$N$ be wta and ${\cal A}$~be equisubtractive with $A = \langle U \rangle_{\mathord{+}}$. Then $M \stackrel X\to N$ if and only if there exist two wta $M'$~and~$N'$ such that \begin{itemize} \item $M \stackrel C\to M'$ where $C^{\mathrm T}$~is functional, \item $M' \stackrel E\to N'$ where $E$~is an invertible diagonal matrix, and \item $N' \stackrel D\to N$ where $D$~is functional. \end{itemize} If $M$~and~$N$ are trim, then $M' \twoheadleftarrow M$ and $N' \twoheadrightarrow N$. \end{theorem} \begin{proof} Clearly, $M \stackrel C\to M' \stackrel E\to N' \stackrel D\to N$, which proves that $M \stackrel{CED}\longrightarrow N$. For the converse, let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$. Lemma~\ref{lm:Help2} shows that there exist matrices $C, E, D$ such that \begin{itemize} \item $X = CED$, \item $C^{\mathrm T}$~and~$D$ are functional matrices, and \item $E \in A^{I \times I}$~is an invertible diagonal matrix. \end{itemize} Finally, let $\varphi \colon I \to Q$ and $\psi \colon I \to P$ be the functions associated to $C^{\mathrm T}$~and~$D$. It remains to determine the wta $M'$~and~$N'$. We construct $M' = (\Sigma, I, \mu', F')$ and $N' = (\Sigma, I, \nu', G')$ with \begin{itemize} \item $G' = DG$ and \item $F' = EDG$. \end{itemize} Then $CF' = CEDG = XG = F$. Thus, it remains to specify $\mu'_k(\sigma)$~and~$\nu'_k(\sigma)$ for every $\sigma \in \Sigma_k$. To this end, we determine a matrix $Y \in A^{I^k \times I}$ such that \begin{align} \label{eq:2a} C^{k, \mathord{\otimes}} \cdot Y &= \mu_k(\sigma) CE \\ \label{eq:2b} YD &= E^{k, \mathord{\otimes}} \cdot D^{k, \mathord{\otimes}} \cdot \nu_k(\sigma) \enspace. \end{align} Given such a matrix~$Y$, we then let $\mu'_k(\sigma) = YE^{-1}$ and $\nu'_k(\sigma) = (E^{k, \mathord{\otimes}})^{-1} \cdot Y$. Then \begin{align*} \mu_k(\sigma) C &= C^{k, \mathord{\otimes}} \cdot \mu'_k(\sigma) &\quad \mu'_k(\sigma) E &= E^{k, \mathord{\otimes}} \cdot \nu'_k(\sigma) &\quad \nu'_k(\sigma) D &= D^{k, \mathord{\otimes}} \cdot \nu_k(\sigma) \enspace. \end{align*} These equalities are displayed in Fig.~\ref{fig:Squares}. Finally, we need to specify the matrix~$Y$. For every $q \in Q$ and $p \in P$, let $I_q = \varphi^{-1}(q)$ and $J_p = \psi^{-1}(p)$. Obviously, $Y$~can be decomposed into disjoint (not necessarily contiguous) submatrices $Y_{\word q1k, p} \in A^{(I_{q_1} \times \dotsm \times I_{q_k}) \times J_p}$ with $\seq q1k \in Q$ and $p \in P$. Then \eqref{eq:2a}~and~\eqref{eq:2b} hold if and only if for every $\seq q1k \in Q$ and $p \in P$ the following two conditions hold: \begin{enumerate} \item For every $i \in I$ such that $\psi(i) = p$, the sum of the $i$-column of~$Y_{\word q1k, p}$ is $\mu_k(\sigma)_{\word q1k, \varphi(i)} \cdot e_{i,i}$. \item For all $\seq i1k \in I$ such that $\varphi(i_j) = q_j$ for every $j \in [k]$, the sum of the $(\seq i1k)$-row of~$Y_{\word q1k, p}$ is $\prod_{j = 1}^k e_{i_j, i_j} \cdot \nu_k(\sigma)_{\psi(i_1) \dotsm \psi(i_k), p}$. \end{enumerate} Those two conditions are compatible because \begin{align*} &\phantom{{}={}} \sum_{\substack{i \in I \\ \psi(i) = p}} \mu_k(\sigma)_{\word q1k, \varphi(i)} \cdot e_{i,i} = \bigl( \mu_k(\sigma)CED \bigr)_{\word q1k, p} = \bigl( \mu_k(\sigma)X \bigr)_{\word q1k, p} \\ &\stackrel\dagger= \bigl( X^{k, \mathord{\otimes}} \cdot \nu_k(\sigma) \bigr)_{\word q1k, p} = \bigl( C^{k, \mathord{\otimes}} \cdot E^{k, \mathord{\otimes}} \cdot D^{k, \mathord{\otimes}} \cdot \nu_k(\sigma) \bigr)_{\word q1k, p} \\ &= \sum_{\substack{\seq i1k \in I \\ \forall j \in [k] \colon \varphi(i_j) = q_j}} \Bigl( \prod_{j = 1}^k e_{i_j, i_j} \Bigr) \cdot \nu_k(\sigma)_{\psi(i_1) \dotsm \psi(i_k), p} \enspace. \end{align*} Consequently, the row and column sums of the submatrices~$Y_{\word q1k, p}$ are consistent, which yields that we can determine all the submatrices (and thus the whole matrix) by Lemma~\ref{lm:Sol}. If $M$~and~$N$ are trim, then either \begin{itemize} \item[(a)] ${\cal A}$~is zero-sum free (and thus positive because it is additively generated by its units), in which case $X$~is nondegenerate by Lemma~\ref{lm:Trim}, or \item[(b)] ${\cal A}$~has nontrivial zero-sums. \end{itemize} In both cases, Lemma~\ref{lm:Help2} shows that the matrices $C^{\mathrm T}$~and~$D$ are surjective, which yields the additional statement by Lemmata \ref{lm:FSim}~and~\ref{lm:BSim}. \qed \end{proof} \begin{figure} \centering \includegraphics{squares.mps} \caption{Illustration of the relations between the matrices in the proof of Theorem~\protect{\ref{thm:2}}.} \label{fig:Squares} \end{figure} \section{Category of simulations} \label{sec:Category} In this section our aim is to show that several well-known constructions of wta are \emph{functorial}: they may be extended to simulations in a functorial way. Below we will only deal with the sum, \textsc{Hadamard} product, $\sigma_0$-product, and $\sigma_0$-iteration (cf.~\cite{esi10}). Scalar OI-substition, ${}^\dagger$~\cite{bloesi03}, homomorphism, quotient, and top-concatenation~\cite{esi10} may be covered in a similar fashion. Throughout this section, let $\mathcal{A}$~be commutative. Let $M = (\Sigma, Q, \mu, F)$, $M' = (\Sigma, Q', \mu', F')$, and $M'' = (\Sigma, Q'', \mu'', F'')$ be wta. We already remarked that, if $M \stackrel X\to M'$ and $M' \stackrel Y\to M''$, then $M \stackrel{XY}\to M''$. Moreover, $M \stackrel I\to M$ with the unit matrix~$I \in A^{Q \times Q}$. Thus, wta over the alphabet~$\Sigma$ form a category~$\text{\textbf{Sim}}_\Sigma$. In the following, let $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$ be wta such that $Q \cap P = \emptyset$. \begin{definition} The sum $M + N$ of $M$~and~$N$ is the wta $(\Sigma, Q \cup P, \kappa, H)$ where $H = \langle F, G \rangle = \begin{pmatrix} F \\ G \end{pmatrix}$ and \[ \kappa_k(\sigma)_{\word q1k, q} = (\mu_k(\sigma) + \nu_k(\sigma))_{\word q1k, q} = \begin{cases} \mu_k(\sigma)_{\word q1k, q} & \text{if } q, \seq q1k \in Q \\ \nu_k(\sigma)_{\word q1k, q} & \text{if } q, \seq q1k \in P \\ 0 & \text{otherwise.} \end{cases} \] for all $\sigma \in \Sigma_k$ and $q, \seq q1k \in Q \cup P$. \end{definition} It is well-known that $\sem{M + N} = \sem M + \sem N$. Next, we extend the sum construction to simulations. To this end, let $M \stackrel X\to M'$ with $M' = (\Sigma, Q', \mu', F')$, and let $N \stackrel Y\to N'$ with $N' = (\Sigma, P', \nu', G')$. \begin{definition} The sum $X + Y \in A^{(Q \cup P) \times (Q' \cup P')}$ of the transfer matrices $X$~and~$Y$ is \[ X + Y = \begin{pmatrix} X & 0 \\ 0 & Y \end{pmatrix} \enspace. \] \end{definition} \begin{proposition} We have~$(M + N) \stackrel{X + Y}\longrightarrow (M' + N')$. \end{proposition} \begin{proof} We only need to verify the two conditions of Definition~\ref{df:Conj}. For every $\sigma \in \Sigma_k$ we have \begin{align*} &\phantom{{}={}} \bigl(\mu_k(\sigma) + \nu_k(\sigma) \bigr) \cdot (X + Y) = \mu_k(\sigma)X + \nu_k(\sigma)Y \\ &= X^{k, \mathord{\otimes}} \cdot \mu'_k(\sigma) + Y^{k, \mathord{\otimes}} \cdot \mu'_k(\sigma) = (X + Y)^{k, \mathord{\otimes}} \cdot \bigl(\mu'_k(\sigma) + \nu'_k(\sigma) \bigr) \end{align*} and $\langle F, G \rangle = \langle XF', YG' \rangle = (X + Y) \cdot \langle F', G' \rangle$, which completes the proof. \qed \end{proof} \begin{proposition} The function~$+$, which is defined on wta and transfer matrices, is a functor $\text{\textbf{Sim}}_\Sigma^2 \to \text{\textbf{Sim}}_\Sigma$. \end{proposition} \begin{proof} It is a routine matter to verify that identity transfer matrices are preserved and $(X + Y) \cdot (X' + Y') = XX' + YY'$ for all composable transfer matrices~$X, X', Y, Y'$. \qed \end{proof} \begin{definition} Let $\sigma_0$ be a distinguished symbol in $\Sigma_0$. The $\sigma_0$-product $M \cdot_{\sigma_0} N$ of~$M$ with~$N$ is the wta $(\Sigma, Q \cup P, \kappa, H)$ such that \[ H = \langle F, 0\rangle = \begin{pmatrix} F\\ 0 \end{pmatrix} \] and for each $\sigma \in \Sigma_k$ with $\sigma \neq \sigma_0$, \[ \kappa_k(\sigma)_{\word q1k, q} = \begin{cases} \mu_k(\sigma)_{\word q1k, q} & \text{if } q, \seq q1k \in Q \\ \mu_0(\sigma_0)_q \cdot \sum_{p \in P} \nu_k(\sigma)_{\word q1k, p} G_p & \text{if } q \in Q \text{ and } \seq q1k \in P \\ \nu_k(\sigma)_{\word q1k, q} & \text{if } q, \seq q1k \in P \\ 0 & \text{otherwise.} \end{cases} \] Moreover, \[ \kappa_0(\sigma_0)_q = \begin{cases} \mu_0(\sigma_0)_q \cdot \sum_{p \in P} \nu_0(\sigma_0)_p G_p & \text{if } q\in Q \\ \nu_0(\sigma_0)_q & \text{if } q \in P. \end{cases} \] \end{definition} It is known that $\sem{M \cdot_{\sigma_0} N} = \sem M \cdot_{\sigma_0} \sem N$. We extend this construction to simulations. To this end, let $M \stackrel X\to M'$ and $N \stackrel Y\to N'$. Then we define $X \cdot_{\sigma_0} Y = X + Y$. The next proposition can be verified by a routine calculation. \begin{proposition} The function~$\cdot_{\sigma_0}$, which is defined on wta and transfer matrices, is a functor $\text{\textbf{Sim}}_\Sigma^2 \to \text{\textbf{Sim}}_\Sigma$. \end{proposition} \begin{definition} The \textsc{Hadamard} product $M \cdot_{\mathrm H} N$ is the wta $(\Sigma, Q \times P, \kappa, H)$ where $H = F \otimes G$ and $\kappa_k(\sigma) = \mu_k(\sigma) \otimes \nu_k(\sigma)$ for all $\sigma \in \Sigma_k$. \end{definition} We again extend the construction to simulations. If $M \stackrel X\to M'$ and $N \stackrel Y\to N'$, then we define $X \cdot_{\mathrm H} X \otimes Y$. \begin{proposition} The function~$\cdot_{\mathrm H}$, which is defined on wta and transfer matrices, is a functor $\text{\textbf{Sim}}_\Sigma^2 \to \text{\textbf{Sim}}_\Sigma$. \end{proposition} Finally, we deal with iteration. Let $\sigma_0$~be a fixed symbol in~$\Sigma_0$. Here we assume that $\mathcal{A}$~is complete. Thus, $\mathcal{A}$~comes with a star operation $a^* = \sum_{n \in \nat} a^n$ for every $a \in A$. \begin{definition} The $\sigma_0$-iteration $M^{*_{\sigma_0}}$ of~$M$ is the wta $(\Sigma, Q, \kappa, F)$ where \[ \kappa_k(\sigma)_{\word q1k, q} = \mu_k(\sigma)_{\word q1k, q} + \sem{M}(\sigma_0)^* \cdot \sum_{p \in Q} \mu_k(\sigma)_{\word q1k, p} F_p \] for all $\sigma \in \Sigma_k \setminus \{\sigma_0\}$ and $\kappa_0(\sigma_0) = \mu_0(\sigma_0)$. \end{definition} If $M \stackrel X\to M'$, then we define $X^{*_{\sigma_0}} = X$. \begin{proposition} The $\sigma_0$-iteration, which is defined on wta and transfer matrices, is a functor $\text{\textbf{Sim}}_\Sigma \to \text{\textbf{Sim}}_\Sigma$. \end{proposition} \begin{remark} Several subcategories of~$\text{\textbf{Sim}}_\Sigma$ are also of interest, for example the categories formed by the relational or functional simulations and their duals. The above constructions are preserved by these special kinds of simulations. \end{remark} \section{Joint reduction} \label{sec:Joint} Next we will establish equivalence results using the approach called \emph{joint reduction} in~\cite{bealomsak06}. Let $V \subseteq A^I$ be a set of vectors for a finite set~$I$. Then the ${\cal A}$-semimodule generated by~$V$ is denoted by~$\langle V \rangle$. Given two wta $M = (\Sigma, Q, \mu, F)$ and $N = (\Sigma, P, \nu, G)$ with $Q \cap P = \emptyset$, we first compute $M + N = (\Sigma, Q \cup P, \mu', F')$ as defined in Section~\ref{sec:Category}. Now the aim is to compute a finite set~$V \subseteq A^{Q \cup P}$ such that \begin{itemize} \item[(i)] $(v_1 \otimes \dotsm \otimes v_k) \cdot \mu'_k(\sigma) \in \langle V \rangle$ for every $\sigma \in \Sigma_k$ and $\seq v1k \in V$, and \item[(ii)] $v_1F = v_2G$ for every $(v_1, v_2) \in V$ such that $v_1 \in A^Q$ and $v_2 \in A^P$. \end{itemize} With such a finite set~$V$ we can now construct a wta $M' = (\Sigma, V, \nu', G')$ with $G'_v = v F'$ for every $v \in V$ and \[ \sum_{v \in V} \nu'_k(\sigma)_{\word v1k, v} \cdot v = (v_1 \otimes \dotsm \otimes v_k) \cdot \mu'_k(\sigma) \] for every $\sigma \in \Sigma_k$ and $\seq v1k \in V$. It remains to prove that $M'$ simulates~$M + N$. To this end, let $X = (v)_{v \in V}$, where each $v \in V$ is a row vector. Then for every $\sigma \in \Sigma_k$, $\seq v1k \in V$, and $q \in Q \cup P$, we have \begin{align*} &\phantom{{}={}} (\nu'_k(\sigma) X)_{\word v1k, q} = \sum_{v \in V} \nu'_k(\sigma)_{\word v1k, v} \cdot v_q = \Bigl( \sum_{v \in V} \nu'_k(\sigma)_{\word v1k, v} \cdot v \Bigr)_q \\ &= \bigl( (v_1 \otimes \dotsm \otimes v_k) \cdot \mu'_k(\sigma) \bigr)_q = \sum_{\seq q1k \in Q \cup P} (v_1)_{q_1} \cdot \ldots \cdot (v_k)_{q_k} \cdot \mu'_k(\sigma)_{\word q1k, q} \\ &= \bigl( X^{k, \mathord{\otimes}} \cdot \mu'_k(\sigma) \bigr)_{\word v1k, q} \enspace. \end{align*} Moreover, if we let $X_1$~and~$X_2$ be the restrictions of~$X$ to the entries of $Q$~and~$P$, respectively, then we have $\nu'_k(\sigma)X_1 = X_1^{k, \mathord{\otimes}} \cdot \mu_k(\sigma)$ and $\nu'_k(\sigma)X_2 = X_2^{k, \mathord{\otimes}} \cdot \nu_k(\sigma)$. In addition, $G'_v = v F' = \sum_{q \in Q \cup P} v_q F'_q = (XF')_v$ for every $v \in V$, which proves that $M' \stackrel X\to (M + N)$. Since $v_1F = v_2G$ for every $(v_1, v_2) \in V$, we have $G'_{(v_1, v_2)} = (v_1, v_2)F' = v_1F + v_2G = (1+1)v_1F = (1 + 1)v_2G$. Now, let $G''_{(v_1, v_2)} = v_1F = v_2G$ for every $(v_1, v_2) \in V$. Then \begin{align*} G''_v &= v_1F = \sum_{q \in Q} v_q F_q = (X_1F)_v \\ &= v_2G = \sum_{p \in P} v_p G_p = (X_2G)_v \end{align*} for every $v = (v_1, v_2) \in V$. Consequently, $M'' \stackrel{X_1}\to M$ and $M'' \stackrel{X_2}\to N$, where $M'' = (\Sigma, V, \nu', G'')$. This proves the next theorem. \begin{theorem} \label{thm:Joint} Let $M$~and~$N$ be two equivalent wta. If there exists a finite set~$V \subseteq A^{Q \cup P}$ with properties (i)~and~(ii), then there exists a chain of simulations that join $M$~and~$N$. In fact, there exists a single wta that simulates both $M$~and~$N$. \end{theorem} \subsection{Fields} \label{sec:Fields} In this section, let ${\cal A}$~be a field. We first recall some notions from~\cite{aleboz89}. Let $\varphi \in \series A\Sigma$ be a tree series. The \emph{syntactic ideal} of~$\varphi$ is \[ I_\varphi = \{ \psi \in \series A\Sigma \mid \sum_{t \in T_\Sigma} (\psi, t) (\varphi, c[t]) = 0 \text{ for all } c \in C_\Sigma \} \enspace. \] Moreover, let $\mathord\equiv$ be the equivalence relation on~$\series A\Sigma$ such that $\psi_1 \equiv \psi_2$ if and only if $\psi_1 - \psi_2 \in I_\varphi$. The syntactic algebra is $[\series A\Sigma]_\equiv$. By~\cite[Proposition~2]{aleboz89} the tree series~$\varphi$ is recognizable if and only if its syntactic algebra has finite dimension. Now, let $\varphi$~be recognizable, and let $B$~be a basis of its syntactic algebra. Finally, let $M_\varphi$~be the obtained canonical weighted tree automaton, which recognizes~$\varphi$. \begin{theorem}[{\protect\cite[p.~453]{aleboz89}}] \label{thm:ReltoMin} Every trim wta recognizing~$\varphi$ simulates~$M_\varphi$. \end{theorem} Consequently, all equivalent trim wta $M_1$~and~$M_2$ simulate the canonical wta that recognizes~$\sem M$. Using Theorem~\ref{thm:2} we can show that there exist wta~$M'_1$, $M'_2$, $N'_1$, and~$N'_2$ such that \begin{itemize} \item $M_1 \twoheadleftarrow M'_1$, \item $M'_1 \stackrel E\to N'_1$ with an invertible diagonal matrix~$E$, \item $N'_1 \twoheadrightarrow M_\varphi$, \item $N'_2 \twoheadrightarrow M_\varphi$, \item $M'_2 \stackrel{E'}\to N'_2$ with an invertible diagonal matrix~$E'$, and \item $M_2 \twoheadleftarrow M_2$. \end{itemize} This can be illustrated as follows: \[ M_1 \xleftarrow{\text{backward}} M'_1 \xrightarrow{\text{diagonal}} N'_1 \xrightarrow{\text{forward}} M_\varphi \xleftarrow{\text{forward}} N'_2 \xleftarrow{\text{diagonal}} M'_2 \xrightarrow{\text{backward}} M_2 \] \begin{theorem} \label{thm:Field} Every two equivalent trim wta $M$~and~$N$ over the field~${\cal A}$ can be joined by a chain of simulations. Moreover, there exists a minimal wta~$M_{\sem M}$ such that $M$~and~$N$ both simulate~$M_{\sem M}$. \end{theorem} We could have obtained a similar theorem with the help of Theorem~\ref{thm:Joint} because the finite set~$V$ can be obtained as in~\cite{boz91}. The approach in the next section will cover this case. \subsection{\textsc{Noetherian} semirings} \label{sec:Int} Now, let ${\cal A}$ be a \textsc{Noetherian} semiring. We construct the finite set~$V$ as follows. Let $V_0 = \{ \mu'_0(\alpha) \mid \alpha \in \Sigma_0\}$ and \[ V_{i + 1} = V_i \cup \bigl( \{ (v_1 \otimes \dotsm \otimes v_k) \cdot \mu'_k(\sigma) \mid \sigma \in \Sigma_k, \seq v1k \in V_i\} \setminus \langle V_i\rangle \bigr) \] for every $i \in \nat$. Then \[ \{ 0 \} \subseteq \langle V_0 \rangle \subseteq \langle V_1 \rangle \subseteq \dotsb \subseteq \langle V_k \rangle \subseteq \dotsb {} \] is stationary after finitely many steps because ${\cal A}$ is \textsc{Noetherian}. Thus, let $V = V_k$ for some $k \in \nat$ such that $\langle V_k \rangle = \langle V_{k+1} \rangle$. Clearly, $V$~is finite and has property~(i). Trivially, $V \subseteq \{ h_{\mu'}(t) \mid t \in T_\Sigma \}$, so let $v \in V$ be such that $v = \sum_{i \in I} (h_\mu(t_i), h_\nu(t_i))$ for some finite index set~$I$ and $t_i \in T_\Sigma$ for every $i \in I$. Then \begin{align*} \Bigl( \sum_{i \in I} h_\mu(t_i) \Bigr) F = \sum_{i \in I} (\sem M, t_i) = \sum_{i \in I} (\sem N, t_i) = \Bigl( \sum_{i \in I} h_\nu(t_i) \Bigr) G \end{align*} because $\sem M = \sem N$, which proves property~(ii). \begin{theorem} \label{thm:Noeth} Let ${\cal A}$ be a \textsc{Noetherian} semiring. For every two equivalent wta $M$~and~$N$ over~${\cal A}$, there exists a chain of simulations that join $M$~and~$N$. In fact, there exists a single wta that simulates both $M$~and~$N$. \end{theorem} \begin{proof} Follows from Theorem~\ref{thm:Joint}. \end{proof} Since $\integer$~forms a \textsc{Noetherian} ring, we obtain the following corollary. \begin{corollary}[{\protect{of Theorem~\ref{thm:Noeth}}}] \label{cor:Int} For every two equivalent wta $M$~and~$N$ over~$\integer$, there exists a chain of simulations that join $M$~and~$N$. In fact, there exists a single wta that simulates both $M$~and~$N$. \end{corollary} In fact, since $M + N$ uses only finitely many semiring coefficient, it is sufficient that every finitely generated subsemiring of~${\cal A}$ is contained in a \textsc{Noetherian} subsemiring of~${\cal A}$. Since every finitely generated commutative ring is \textsc{Noetherian}~\cite[Cor.~IV.2.4 \& Prop.~X.1.4]{lan84}, we obtain the following corollary. \begin{corollary}[{\protect{of Theorem~\ref{thm:Noeth}}}] \label{cor:Ring} For every two equivalent wta $M$~and~$N$ over the commutative ring~${\cal A}$, there exists a chain of simulations that join $M$~and~$N$. In fact, there exists a single wta that simulates both $M$~and~$N$. \end{corollary} \subsection{Natural numbers} \label{sec:Nat} Finally, let ${\cal A} = \nat$ be the semiring of natural numbers. We compute the finite set~$V \subseteq \nat^{Q \cup P}$ as follows: \begin{enumerate} \item Let $V_0 = \{ \mu'_0(\alpha) \mid \alpha \in \Sigma_0\}$ and $i = 0$. \item For every $v, v' \in V_i$ such that $v \leq v'$, replace~$v'$ by~$v' - v$. \item Set $V_{i + 1} = V_i \cup \bigl( \{ (v_1 \otimes \dotsm \otimes v_k) \cdot \mu'_k(\sigma) \mid \sigma \in \Sigma_k, \seq v1k \in V_i\} \setminus \langle V_i\rangle \bigr)$. \item Until $V_{i + 1} = V_i$, increase~$i$ and repeat step~2. \end{enumerate} Clearly, this algorithm terminates since every vector can only be replaced by a smaller vector in step~2 and step~3 only adds a finite number of vectors, which after the reduction in step~2 are pairwise incomparable. Moreover, property~(i) trivially holds because at termination $V_{i+1} = V_i$ after step~3. Consequently, we only need to prove property~(ii). To this end, we first prove that $V \subseteq \langle \{ h_{\mu'}(t) \mid t \in T_\Sigma \} \rangle_{\mathord{+}, \mathord{-}}$. This is trivially true after step~1 because $\mu'_0(\alpha) = h_{\mu'}(\alpha)$ for every $\alpha \in \Sigma_0$. Clearly, the property is preserved in steps 2~and~3. Finally, property~(ii) can now be proved as follows. Let $v \in V$ be such that $v = \sum_{i \in I_1} (h_\mu(t_i), h_\nu(t_i)) - \sum_{i \in I_2} (h_\mu(t_i), h_\nu(t_i))$ for some finite index sets $I_1$~and~$I_2$ and $t_i \in T_\Sigma$ for every $i \in I_1 \cup I_2$. Then \begin{align*} &\phantom{{}={}} \Bigl( \sum_{i \in I_1} h_\mu(t_i) - \sum_{i \in I_2} h_\mu(t_i) \Bigr) F = \sum_{i \in I_1} h_\mu(t_i)F - \sum_{i \in I_2} h_\mu(t_i)F \\ &= \sum_{i \in I_1} (\sem M, t_i) - \sum_{i \in I_2} (\sem M, t_i) = \sum_{i \in I_1} (\sem N, t_i) - \sum_{i \in I_2} (\sem N, t_i) \\ &= \sum_{i \in I_1} h_\nu(t_i)G - \sum_{i \in I_2} h_\nu(t_i)G = \Bigl( \sum_{i \in I_1} h_\nu(t_i) - \sum_{i \in I_2} h_\nu(t_i) \Bigr) G \end{align*} because $\sem M = \sem N$. \begin{corollary}[{\protect{of Theorem~\ref{thm:Joint}}}] \label{cor:Nat} For every two equivalent wta $M$~and~$N$ over~$\nat$, there exists a chain of simulations that join $M$~and~$N$. In fact, there exists a single wta that simulates both $M$~and~$N$. \end{corollary} For all finitely and effectively presented semirings, Theorems \ref{thm:Field}~and~\ref{thm:Noeth} and Corollaries \ref{cor:Ring}~and~\ref{cor:Nat}, also yield decidability of equivalence for $M$~and~$N$. Essentially, we run the trivial semi-decidability test for inequality and a search for the wta the simulates both $M$~and~$N$ in parallel. We know that either test will eventually return, thus deciding whether $M$~and~$N$ are equivalent. Conversely, if equivalence is undecidable, then simulation cannot capture equivalence~\cite{esimal10}.

https://arxiv.org/abs/hep-th/9503177

Higher Derivatives and Canonical Formalism

A canonical formalism for higher-derivative theories is presented on the basis of Dirac's method for constrained systems. It is shown that this formalism shares a path integral expression with Ostrogradski's canonical formalism.

\section{Introduction} Higher-derivative theories appear in various scenes of physics:${}^{1),2)}$ higher derivative terms occur as quantum corrections; nonlocal theories, e.g. string theories, are essentially higher-derivative the ories; Einstein gravity supplemented by curvature-squared terms has attracted attenti on because of its renormalizability.${}^{3)}$ A canonical formalism for higher-derivative theories was developed by Ostrogra dski about one and a half centuries ago.${}^{4)}$ Though being self-consistent, his formulation looks different from the convent ional canonical formalism. The purpose of the present paper is to interpret the Ostrogradski's formulatio n in the framework of the ordinary constrained canonical formalism. We show that the path integral expression for the ordinary constrained canonic al formalism is the same as that for the Ostrogrdski formalism. In \S 2 we review the Ostrogradski's formulation. In \S 3 we give another canonical formulation for higher-derivative theories b ased on the usual method of Dirac for constrained systems.${}^{5)}$ In \S 4 it is shown that the two formulations give the same path integral form ulae. Section 5 gives summary. \section{Ostrogradski's canonical formalism} We conside a generic Lagrangian which contains up to $N$th derivative of coord inate $q(t)$: \begin{equation} L = L \left( q^{(0)}, q^{(1)}, \cdots , q^{(N)} \right) , \end{equation} where \begin{equation} q^{(i)} \stackrel{\rm d}{\equiv} \frac{{\rm d}^{i}}{{\rm d}t^{i}}q \makebox[1cm]{} ( i = 0, 1, \cdots , N ) . \end{equation} The Euler-Lagrange equation is \begin{equation} \frac{\delta S}{\delta q} \stackrel{\rm d}{\equiv} \sum_{i=0}^{N} \left( - \frac{{\rm d}}{{\rm d}t} \right) ^{i} \frac{\partial L}{\partial q^{(i)}} = 0 . \end{equation} The canonical formalism of Ostrogradski is the following.${}^{4)}$ For $i = 0, \cdots , N-1$ we regard $q^{(i)}$ as independent coordinates $q^{i }$ : \begin{eqnarray} q^{(i)} & \longrightarrow & q^{i} \makebox[1cm]{} ( i = 0, 1, \cdots , N-1 ) , \\ L \left( q^{(0)}, q^{(1)}, \cdots , q^{(N-1)}, q^{(N)} \right) & \longrightarrow & L \left( q^{0}, q^{1}, \cdots , q^{N-1}, \dot{q}^{N-1} \right) . \end{eqnarray} The momentum $p_{N-1}$ conjugate to $q^{N-1}$ is defined as usual by \begin{equation} p_{N-1} \stackrel{\rm d}{\equiv} \frac{\partial L}{\partial \dot{q}^{N-1}} \left( q^{0}, q^{1}, \cdots , q^{N-1}, \dot{q}^{N-1} \right) . \end{equation} Here and hereafter we assume that the Lagrangian is nondegenerate. This means that the relation (6) can be inverted to give $\dot{q}^{N-1}$ as a function of $q^{i} \; ( i = 0, \cdots , N-1 )$ and $p_{N-1}$ : \begin{equation} \dot{q}^{N-1} = \dot{q}^{N-1} \left( q^{0}, q^{1}, \cdots , q^{N-1}, p_{N-1} \right) . \end{equation} The Hamiltonian is {\it defined} by \begin{eqnarray} H_{{\rm O}} & = & H_{{\rm O}} \left( q^{0}, q^{1}, \cdots , q^{N-1}\, ; \, p_{0}, p_{1}, \cdots, p_{N-1} \right) \nonumber \\ & \stackrel{\rm d}{\equiv} & \sum_{i=0}^{N-2} p_{i}q^{i+1} + p_{N-1}\dot{q}^{N-1} - L \left( q^{0}, q^{1}, \cdots , q^{N-1}, \dot{q}^{N-1} \right) . \end{eqnarray} This definition is different from the usual Legendre transformation. Note that for $i = 0, \cdots , N-2$, the momenta $p_{i}$ are {\it not} defined through relations like (6), but introduced just as independent canonical vari ables. The canonical equations of motion are \begin{eqnarray} \dot{q}^{i} & = & \frac{\partial H_{{\rm O}}}{\partial p_{i}} , \\ \dot{p}_{i} & = & \mbox{} \! - \frac{\partial H_{{\rm O}}}{\partial q^{i}} . \end{eqnarray} Eq.(9) with $i = 0, \cdots , N-2$ gives \begin{equation} \dot{q}^{i} = q^{i+1} \makebox[1cm]{} ( i = 0, 1, \cdots , N-2 ). \end{equation} Under the assumption of nondegeneracy, Eq.(9) with $i = N-1$ reproduces the de finition (6). Eq.(10) gives the following equations: \begin{equation} \left\{ \begin{array}{rcl} \dot{p}_{0} & = & \displaystyle \frac{\partial L}{\partial q^{0}} , \vspace{2mm}\\ \dot{p}_{i} & = & \displaystyle \mbox{} \! - p_{i-1} + \frac{\partial L}{\partial q^{i}} \makebox[1cm]{} ( i = 1, \cdots , N-1 ) . \end{array} \right. \end{equation} {}From Eqs.(12), (11) and (6) we regain the Euler-Lagrange equation (3). \section{Constrained canonical formalism} It has been seen that the Ostrogradski formalism gives special treatment to the highest derivative $q^{(N)}$. Is it possible to treat the highest derivative and the lower derivatives on an equal footing? This is the subject of the present section. To treat all the derivatives equally, we introduce Lagrange multipliers $\lambda _{i}\; ( i = 0, \cdots , N-1 )$ and start from the folllowing equivalent Lagrangian: \begin{equation} L^{*} \stackrel{\rm d}{\equiv} L \left( q^{0}, q^{1}, \cdots , q^{N} \right) + \sum_{i=0}^{N-1} \lambda_{i} \left( \dot{q}^{i} - q^{i+1} \right) . \end{equation} The Euler-Lagrange equations \begin{equation} \left\{ \begin{array}{rcl} \displaystyle \frac{\delta L^{*}}{\delta q^{i}} & = & 0 \makebox[1cm]{} ( i = 0, \cdots , N ) , \vspace{2mm}\\ \displaystyle \frac{\delta L^{*}}{\delta \lambda_{i}} & = & 0 \makebox[1cm]{} ( i = 0, \cdots , N-1 ) \end{array} \right. \end{equation} give \begin{equation} \left\{ \begin{array}{l} \displaystyle \sum_{i=0}^{N} \left( - \frac{{\rm d}}{{\rm d}t} \right) ^{i} \frac{\partial L}{\partial q^{i}} = 0 , \vspace{2mm}\\ \displaystyle \dot{q}^{i} = q^{i+1} \makebox[1cm]{} ( i = 0, \cdots , N-1 ) , \end{array} \right. \end{equation} which are equivalent to Eq.(3). Since the Lagrangian $L^{*}$ describes a constrained system, we follow a way o f Dirac.${}^{5)}$ The conjugate momenta defined by \begin{equation} \left\{ \begin{array}{rcl} p_{i} & \stackrel{\rm d}{\equiv} & \displaystyle \frac{\partial L^{*}}{\partial \dot{q}^{i}} \makebox[1cm]{} ( i = 0, \cdots , N ) , \vspace{2mm}\\ \pi ^{i} & \stackrel{\rm d}{\equiv} & \displaystyle \frac{\partial L^{*}}{\partial \dot{\lambda}_{i}} \makebox[1cm]{} ( i = 0, \cdots , N-1 ) \end{array} \right. \end{equation} provide the following primary constraints: \begin{eqnarray} \phi _{i} & \stackrel{\rm d}{\equiv} & p_{i} - \lambda _{i} \: \approx \: 0 \makebox[1cm]{} ( i = 0, \cdots , N-1 ) , \\ \phi _{N} & \stackrel{\rm d}{\equiv} & p_{N} \makebox[1cm]{} \!\!\! \approx \: 0 , \\ \psi ^{i} & \stackrel{\rm d}{\equiv} & \pi ^{i} \makebox[1cm]{} \! \approx \: 0 \makebox[1cm]{} ( i = 0, \cdots , N-1 ) . \end{eqnarray} The consistency of these constraints under their time developments requires a secondary constraint \begin{equation} \psi ^{N} \stackrel{\rm d}{\equiv} \frac{\partial L}{\partial q^{N}} - \lambda _{N-1} \approx 0 . \end{equation} When the system is nondegenerate, all these constraints \begin{equation} \Phi _{\alpha} \stackrel{\rm d}{\equiv} \left( \phi _{0}, \cdots , \phi _{N} \, ; \, \psi ^{0}, \cdots , \psi ^{N} \right) \end{equation} form a set of second-class constraints: the determinant of the Poisson brackets between these constraints \begin{equation} \det \left( \left[ \Phi _{\alpha}, \Phi _{\beta} \right] _{{\rm P}} \right) = \left( \frac{\partial ^{2}L}{\partial q^{N2}} \right) ^{2} \end{equation} is not zero when \begin{equation} \frac{\partial ^{2}L}{\partial q^{N2}} \neq 0 . \end{equation} The Dirac brackets between the canonical variables $\left( q^{i}, \lambda _{i} \, ; \, p_{i}, \pi ^{i} \right)$ are calculated to be \begin{equation} \left\{ \begin{array}{rllll} \left[ q^{i}, p_{i} \right] _{{\rm D}} & = & \left[ q^{i}, \lambda _{i} \right] _{{\rm D}} & = & 1 , \vspace{2mm}\\ \left[ q^{N}, p_{i} \right] _{{\rm D}} & = & \left[ q^{N}, \lambda _{i} \right] _{{\rm D}} & = & \displaystyle - \left( \frac{\partial ^{2}L}{\partial q^{N2}} \right) ^{-1} \frac{\partial ^{2}L}{\partial q^{i} \partial q^{N}} , \vspace{2mm}\\ \left[ q^{N-1}, q^{N} \right] _{{\rm D}} & = & \displaystyle \left( \frac{\partial ^{2}L}{\partial q^{N2}} \right) ^{-1} , & & \vspace{2mm}\\ {\rm The}\ {\rm others} & = & 0 ,& & \end{array} \right. \end{equation} \[ ( i = 0, \cdots , N-1 ) . \] The Hamiltonian is given by \begin{equation} H_{{\rm D}} = - L \left( q^{0}, \cdots , q^{N} \right) + \sum_{i=0}^{N-1} \lambda _{i}q^{i+1} . \end{equation} This is obtained by performing the ordinary Legendre transformation on the Lag rangian $L^{*}$ of Eq.(13) and by regarding all the constraints $\Phi _{\alpha } \approx 0$ as strong equalities $\Phi _{\alpha} = 0$. Eqs.(25) and (24) allow us to obtain the canonical equations of motion. The independent equations are \begin{equation} \left\{ \begin{array}{llll} \dot{q}^{i} & = & q^{i+1} & ( i = 0, \cdots , N-1 ) , \vspace{2mm}\\ \dot{\lambda}_{0} & = & \displaystyle \frac{\partial L}{\partial q^{0}} , & \vspace{2mm}\\ \dot{\lambda}_{i} & = & \displaystyle \frac{\partial L}{\partial q^{i}} - \lambda _{i-1} & ( i = 1, \cdots , N-1 ) , \vspace{2mm}\\ 0 & = & \displaystyle \frac{\partial L}{\partial q^{N}} - \lambda _{N-1} , & \end{array} \right. \end{equation} which are seen to be equivalent to the Euler-Lagrange equations (15). \section{Path integrals} We have presented two canonical formalisms for higher-derivative theories, the Ostrogradski's one and the constrained one. It has been seen that though looking different from each other, they give an e quivalent set of equations of motion. In this section we show that the two formalisms are completely equivalent to e ach other by comparing path integral expressions for them. The Ostrogradski formalism of \S 2 gives the following path integral expressio n \begin{equation} Z_{{\rm O}} = \int \prod_{i=0}^{N-1}\left( {\cal D}q^{i}{\cal D}p_{i} \right) \exp \left\{ i\int {\rm d}t \left[ \sum_{i=0}^{N-1} p_{i}\dot{q}^{i} - H_{{\rm O}} \left( q^{0}, \cdots , q^{N-1}\, ; \, p_{0}, \cdots , p_{N-1} \right) \right] \right\} , \end{equation} where the Hamiltonian $H_{\rm O}$ is given by Eq.(8). Integrations with respect to $p_{i}\; ( i = 0, \cdots , N-2 )$ offer a factor $\prod _{i=0}^{N-2} \delta \left( \dot{q}^{i} - q^{i+1} \right)$. We can further integrate with respect to $q^{i} \; \left( i = 1, \cdots , N-1 \right)$, obtaining \begin{equation} Z_{{\rm O}} = \int {\cal D}q{\cal D}p_{N-1} \exp \left\{ i\int {\rm d}t \left[ p_{N-1}q^{(N)} - \hat{H} \left( q, \dot{q}, \cdots , q^{(N-1)}, p_{N-1} \right) \right] \right\} , \end{equation} where \begin{equation} \hat{H} \stackrel{\rm d}{\equiv} p_{N-1}\dot{q}^{N-1} - L \left( q, \dot{q}, \cdots , q^{(N-1)}, \dot{q}^{N-1} \right) . \end{equation} In Eq.(29), $\dot{q}^{N-1}$ is a function of $q^{(i)} \; \left( i = 0, \cdots , N-1 \right)$ and $p_{N-1}$, which is obtained by replacing $q^{i}$ with $q^{ (i)}$ in Eq.(7). Path integral expression for the case of the constrained canonical formalism o f \S 3 is \begin{eqnarray} Z_{{\rm D}} & = & \int \prod_{i=0}^{N} \left( {\cal D}q^{i}{\cal D}p_{i} \right) \prod_{i=0}^{N-1} \left( {\cal D}\lambda _{i}{\cal D}\pi ^{i} \right) \prod_{i=0}^{N} \left( \delta (\phi _{i})\delta (\psi ^{i}) \right) \left| \frac{\partial ^{2}L}{\partial q^{N2}} \right| \nonumber \\ & & \makebox[5mm]{} \times \exp \left\{ i\int {\rm d}t \left[ \sum_{i=0}^{N} p_{i}\dot{q}^{i} + \sum_{i=0}^{N-1} \pi {i}\dot{\lambda}_{i} - H_{\rm D} \left( q^{0}, \cdots , q^{N} \, ; \, \lambda _{0}, \cdots , \lambda _{N-1} \right) \right] \right\} , \nonumber \\ & & \end{eqnarray} where the constraints $\left( \phi _{i}, \psi ^{i} \right) \; ( i = 0, \cdots , N )$ are given by Eqs.(17) -- (20), and the Hamiltonian $H_{\rm D}$ by Eq.(2 5). Integrations with respect to $p_{N}, \; \pi ^{i} \; ( i = 0, \cdots , N-2 ), \ ; \lambda _{i} \; ( i = 0, \cdots , N-2 ), \; p_{i} \; ( i = 0, \cdots , N-2 ) $ and $q^{i} \; ( i = 1, \cdots , N )$ give \begin{eqnarray} Z_{{\rm D}} & = & \int \prod_{i=0}^{N}{\cal D}q^{i}\prod_{i=0}^{N-1}{\cal D}p_{i} \delta \left( \frac{\partial L}{\partial q^{N}} - p_{N-1} \right) \left| \frac{\partial ^{2}L}{\partial q^{N2}} \right| \nonumber \\ & & \makebox[2cm]{} \times \exp \left\{ i\int {\rm d}t \left[ \sum_{i=0}^{N-1} p_{i}\left( \dot{q}^{i} - q^{i+1} \right) + L \left( q^{0}, \cdots , q^{N} \right) \right] \right\} \nonumber \\ & = & \int \prod_{i=0}^{N}{\cal D}q^{i}{\cal D}p_{N-1} \delta \left( \frac{\partial L}{\partial q^{N}} - p_{N-1} \right) \left| \frac{\partial ^{2}L}{\partial q^{N2}} \right| \prod_{i=0}^{N-2}\delta \left( \dot{q}^{i} - q^{i+1} \right) \nonumber \\ & & \makebox[2cm]{} \times \exp \left\{ i\int {\rm d}t \left[ p_{N-1}\left( \dot{q}^{N-1} - q^{N} \right) + L \left( q^{0}, \cdots , q^{N} \right) \right] \right\} \nonumber \\ & = & \int {\cal D}q{\cal D}p_{N-1} \exp \left\{ i\int {\rm d}t \left[ p_{N-1}\left( q^{(N)} - q^{N} \right) + L \left( q, \dot{q}, \cdots , q^{(N-1)}, q^{N} \right) \right] \right\} . \nonumber \\ & & \end{eqnarray} In the last line of Eq.(31), $q^{N}$ is a function of $q^{(i)} \; ( i = 0, \cdots , N-1 )$ and $p_{N-1}$ defined by inverting the relation \begin{equation} p_{N-1} = \frac{\partial L}{\partial q^{N}} \left( q, \dot{q}, \cdots , q^{(N-1)}, q^{N} \right) . \end{equation} That means $q^{N}$ is nothing but $\dot{q}^{N-1}$ in Eq.(29). Putting $q^{N} = \dot{q}^{N-1}$ in Eq.(31) shows that the path integral $Z_{{\ rm D}}$ is the same as $Z_{{\rm O}}$ given by Eq.(28) \begin{equation} Z_{\rm D} \equiv Z_{\rm O} . \end{equation} \section{Summary} We have presented a canonical formalism for higher-derivative theories based o n the usual method of Dirac for constrained systems. It has been shown that this formalism shares a path integral expression with t he Ostrogradski's one. We thus have laid the foundation of the ordinary canonical formalism for the O strogradski's formulation. \section*{Acknowledgments} The author would like to thank Minoru Hirayama, Shinobu Hosono and Hitoshi Yam akoshi for discussions.

https://arxiv.org/abs/1411.4820

Reducts of the Generic Digraph

The generic digraph $(D,E)$ is the unique countable homogeneous digraph that embeds all finite digraphs. In this paper, we determine the lattice of reducts of $(D,E)$, where a structure $\mathcal{M}$ is a reduct of $(D,E)$ if it has domain $D$ and all its $\emptyset$-definable relations are $\emptyset$-definable relations of $(D,E)$. As $(D,E)$ is $\aleph_0$-categorical, this is equivalent to determining the lattice of closed groups that lie in between Aut$(D,E)$ and Sym$(D)$.

\section{Preliminaries} \subsection{Notational Conventions} Structures are denoted by $\mathcal{M},\mathcal{N}$, and their domains are $M$ and $N$ respectively. Sym$(M)$ is the set of all bijections $M \to M$ and Aut$(\mathcal{M})$ is the set of all automorphisms of $\mathcal{M}$. Given a formula $\phi(x,y)$, we use $\phi^*(x,y)$ to denote the formula $\phi(y,x)$. $S(\mathcal{M})$ denotes the space of types of the theory of $\mathcal{M}$. If $f$ has domain $A$ and $\bar{a} \in A$, then $f(a_1,\ldots,a_n) \defeq (f(a_1),\ldots,f(a_n))$. If $\bar{a}$ and $\bar{b}$ are tuples of the same length $n$ we say $\bar{a}$ and $\bar{b}$ are isomorphic, and write $\bar{a} \cong \bar{b}$, to mean that the function $a_i \mapsto b_i$ for all $i$ such that $1 \leq i \leq n$ is an isomorphism. There will be instances where we do not adhere to strictly correct notational usage, however, the meaning is always clear from the context. For example, we may write `$a \in (a_1,\ldots,a_n)$' instead of `$a=a_i$ for some $i$ such that $1 \leq i \leq n$'. Another example is that we sometimes use $c$ to represent the singleton set $\{ c \}$ containing it. \subsection{The Generic Digraph} \begin{Def} \begin{enumerate}[(i)] \item A directed graph $(V,E)$ consists of a set $V$ and an irreflexive, antisymmetric relation $E \subseteq V^2$. $V$ represents the set of vertices and $E$ represents the set of directed edges, so if $(a,b) \in E$, we visualise it as an edge going out of $a$ and into $b$. We abbreviate `directed graph' by `digraph'. \item By an empty digraph we mean a digraph whose edge set is empty. \item We say that a structure $\mathcal{M}$ is homogeneous if every isomorphism $f:A \to B$, where $A,B$ are finite substructures of $M$, can be extended to an automorphism of $\mathcal{M}$. \item The generic digraph, which we denote by $(D,E)$, is the unique (up to isomorphism) countable homogeneous digraph that embeds all finite digraphs. \item $N(x,y) \subset D^2$ will denote the non-edge relation of $(D,E)$, so $N(x,y) \defeq \neg E(x,y) \wedge \neg E^*(x,y)$. \end{enumerate} \end{Def} The fact that the generic digraph exists and is unique follows from the theory of Fra\"{i}ss\'{e} limits and amalgamation classes, originally described in \cite{fra53}. Details and proofs can be found in \cite{hod97}. The following lemma collects several useful properties of the generic digraph. \begin{lem} \label{genericdigraph} \begin{enumerate}[(i)] \item Th$((D,E))$ is $\aleph_0$-categorical and has quantifier elimination. \item Let $\bar{a},\bar{b} \in D$. If tp$(\bar{a})=$ tp$(\bar{b})$, then there exists an automorphism mapping $\bar{a}$ to $\bar{b}$. \item The generic digraph $(D,E)$ is the unique, up to isomorphism, countable digraph satisfying the following extension property: for all finite pairwise disjoint subsets $U,V,W \subset D$ there exists $x \in D\backslash (U \cup V \cup W)$ such that $(\forall u \in U) E(x,u)$, $(\forall v \in V) E(v,x)$ and $(\forall w \in W) N(x,w)$. \item All countable digraphs can be embedded into the generic digraph. \item Let $A \subseteq D$ and $B=A^c$. Then $(A,E|_A)$ or $(B,E|_B)$ is isomorphic to the generic digraph. \end{enumerate} \end{lem} Remark: Due to the importance of the property in (iii), we give it the name `the extension property'. Remark: As a result of (ii), there is bijective correspondence between $n$-types and orbits of $n$-tuples. Given a type $p(\bar{x})$ you obtain the orbit $\{\bar{x} \in D\:$ tp$(\bar{x})=p \}$, and given an orbit $A \subset D^n$ you obtain the type $p(\bar{a})$, where $\bar{a} \in A$. In this light, and as has become customary in modern model theory, we sometimes blur the distinction between a type and the set of tuples that realise that type. \begin{proof} (i) This is an instance of the more general statement that any countable homogeneous structure in a finite relational language is $\aleph_0$-categorical and has quantifier elimination. See \cite{hod97} for details. (ii) Since we have quantifier elimination, tp$(\bar{a})=$ tp$(\bar{b})$ implies that $\bar{a} \cong \bar{b}$, so by homogeneiety there is an automorphism that maps $\bar{a}$ to $\bar{b}$. (iii) We leave this as an exercise for the reader. To show that two countable digraphs which satisfy the extension property are isomorphic, you use a back-and-forth argument. An explanation and examples of back-and-forth arguments can be found in \cite{hod97}. (iv) This is proved using only the `forth' part of a back-and-forth argument. We sketch the proof. Let $(D',E')$ be a countable digraph, and let $d_1,d_2,d_3,\ldots$ be an enumeration of the elements of $D'$. You then define an embedding of $D'$ into $D$ inductively. The condition that the generic digraph needs to satisfy to ensure that the inductive step works is precisely the extension property. (v) By (iii), it suffices to show that $(A,E_A)$ or $(B,E_B)$ satisfies the extension property. Suppose for contradiction that both fail the extension property. Let $U_1,V_1,W_1 \subset A$ and $U_2,V_2,W_2 \subset B$ witness this failure. Now let $U=U_1 \cup U_2, V=V_1 \cup V_2$ and $W=W_1 \cup W_2$. These are finite pairwise disjoint subsets of $D$. By (i), we know that $D$ satisfies the extension property, so we can find an appropriate witness $x$ in $D$. Now observe that $x$ is also a witness for $U_1,V_1,W_1$ and for $U_2,V_2,W_2$. But this means we have a contradiction, because $x$ must be in $A$ or in $B$. \end{proof} \subsection{Reducts} Let $\mathcal{M}$ be a structure on domain $M$. A relation $P \subseteq M^k$ is $\emptyset$-definable in $\mathcal{M}$ if there exists a formula $\phi(x_1,\ldots,x_k)$ in the language of $\mathcal{M}$ such that $P=\{(x_1,\ldots,x_k) \in M^k: \mathcal{M} \models \phi(x_1,\ldots,x_k)\}$. Let $\mathcal{M}$ and $\mathcal{N}$ be two structures on the same domain $M$. We say that $\mathcal{N}$ is a reduct of $\mathcal{M}$ if for all $k \in \mathbb{N}$ and all relations $P \subset M^k$, if $P$ is $\emptyset$-definable in $\mathcal{N}$ then $P$ is $\emptyset$-definable in $\mathcal{M}$. We say $\mathcal{N}$ is a proper reduct of $\mathcal{M}$ if $\mathcal{N}$ is a reduct of $\mathcal{M}$ and $\mathcal{N} \neq \mathcal{M}$. The question that is answered here is: What are the reducts of the generic digraph? For this question to be meaningful an important caveat is required, which is that if two structures are both reducts of each other - which implies that they are (first-order) interdefinable - we regard them as being equal. This is the reason you will find the phrase `up to interdefinability' used in the literature. For the sake of conciseness, we choose to avoid this phrase with the understanding that we will always consider two reducts that are interdefinable to be equal. An important fact about the reducts of a fixed structure $\mathcal{M}$ is that they form a lattice, where $\mathcal{N} \leq \mathcal{N'}$ if $\mathcal{N}$ is a reduct of $\mathcal{N'}$. The top element is always the original structure $\mathcal{M}$ and the bottom element is the trivial structure $(M,=)$. The meet (respectively join) of two structures $\mathcal{N}$ and $\mathcal{N'}$ will be the structure whose named relations are precisely the $\emptyset$-definable relations that are definable in both (respectively in at least one of) $\mathcal{N}$ and $\mathcal{N'}$. Intuitively, the meet contains the intersection of the information in the two structures, and the join contains the union of the information. In addition to determining what the reducts of the generic digraph are, we also determine how they relate in this lattice. There is a second, closely related notion of a reduct known as a group reduct. We say that $\mathcal{N}$ is a group reduct of $\mathcal{M}$ if Aut$(\mathcal{N}) \geq$ Aut$(\mathcal{M})$. The group reducts of a fixed structure $\mathcal{M}$ form a lattice via the usual inclusion operation; the bottom element is Aut$(\mathcal{M})$ and the top element is always Sym$(D)$. As a consequence of the Engeler--Ryll-Nardzewski--Svenonius theorem (see \cite{hod97}), if $\mathcal{M}$ is $\aleph_0$-categorical then the lattice of reducts is anti-isomorphic to the lattice of group-reducts. In one direction, a reduct $\mathcal{N}$ is mapped to its automorphism group Aut$(\mathcal{N})$. In the other direction, given a group reduct $G$ you let $\mathcal{N}$ be the structure whose $n$-ary relations are the orbits of the action of $G$ on $M^n$ (where for all $g \in G, \bar{x} \in M^n, g\cdot \bar{x}=g(\bar{x})$). In this light, we often use the word `reduct' to refer to either notion, with the meaning being clear from the context. Furthermore, group reducts of $\mathcal{M}$ can be described purely in terms of permutation group theory, without reference to structures. To do this, we need to consider the topological structure of Sym$(M)$. There are two ways of defining the topology. The first is to say that the topology on Sym$(M)$ is the subspace topology of $M^M$, where $M^M$ has the product topology, where $M$ is given the discrete topology. The second (equivalent) way is to say what it means for $F \subseteq$ Sym$(M)$ to be closed: We say that $g \in$ Sym$(M)$ is in the closure of $F$ if for all finite $A \subset M$, there exists $f \in F$ such that $f(a)=g(a)$ for all $a \in A$. Then, $F$ is closed if $F$ is equal to the closure of itself. It is a central fact in permutation group theory that $G \leq$ Sym$(M)$ is closed if and only if there exists a structure with domain $M$ such that $G$ is its automorphism group. Thus, the group reducts of $\mathcal{M}$ are exactly the closed groups $G \leq$ Sym$(M)$ that contain Aut$(\mathcal{M})$. From the above discussion, since $(D,E)$ is $\aleph_0$-categorical, the task of determining its reducts is the same as determining its group reducts, which in turn is the same as determining the closed groups $G$ where Aut$(D,E) \leq G \leq$ Sym$(D)$. \section{Defining the Reducts} There are two ways of defining reducts, corresponding to the two different notions of reducts. On the permutation group theoretic side, you can define a reduct by adding a function $f \in$ Sym$(D)$ to Aut$(D,E)$, then closing under group operations and closing under the topology. By considering the model theoretic view, you first define a relation, $P$ say, and define the reduct to be the automorphism group of $(D,P)$. In view of this, we establish some notation: \begin{enumerate}[(i)] \item Let $G$ be a topological group (e.g. Sym($D$)). For $F \subseteq G$, let $\langle F \rangle$ denote the smallest closed subgroup of $G$ containing $F$. For brevity, when it is clear we are discussing reducts of $(D,E)$, we may abuse notation and write $\langle F \rangle$ to mean $\langle F \cup $ Aut$(D,E) \rangle$. \item Let $G$ be a group. For $F \subseteq G$, let $\gcl(F)$\footnote{where $\gcl$ stands for `group closure'} denote the smallest subgroup of $G$ containing $F$. As above, we may abuse notation where it is clear we are discussing supergroups of Aut($D,E$). \end{enumerate} We begin by showing in the next few lemmas that three particular functions $-,sw$ and $rot$ exist. These functions will give us the three reducts $\langle - \rangle, \langle sw \rangle$ and $\langle rot \rangle$. \begin{lem} There exists $f:D \to D$ such that for all $x,y \in D$, $E(f(x),f(y))$ iff $E(y,x)$. \end{lem} Remark. For the rest of this article, we fix such a function and denote it by $-$. \begin{proof} The idea is to define a structure $(D,E')$ which is isomorphic to $(D,E)$, in such a way that any isomorphism $f:D \to D$ witnessing this fact has the desired property. For this lemma, we let $E'(x,y) = E^*(x,y) \defeq E(y,x)$. We need to show that $(D,E')$ is isomorphic to $(D,E)$. By \autoref{genericdigraph}, it suffices to show that $(D,E')$ satisifes the extension property. So let $U,V,W$ be finite disjoint subsets of $D$. By the definition of $E'$, we need to find $x \in D\backslash (U \cup V \cup W)$ such that $\forall u \in U, E(u,x), \forall v \in V, E(x,v)$ and $\forall w \in W, N(x,w)$. This is simply the extension property for $(D,E)$ with the role of $U$ and $V$ swapped, so we know such an $x$ exists (again by \autoref{genericdigraph}). Thus, $(D,E)$ and $(D,E')$ are isomorphic. Now let $f:D \to D$ be an isomorphism from $(D,E)$ to $(D,E')$ to complete the proof. \end{proof} \begin{lem} \label{swexists} Let $a \in D$. Then there exists $f:D \to D$ such that \[ E(f(x),f(y)) \text{ if and only if } \begin{cases} E(x,y) \text{ and } x,y \neq a, \text{ OR,}\\ E^*(x,y) \text{ and } x=a \vee y=a \end{cases} \] \end{lem} Remark: For the rest of this article, we fix such a function and denote it by $sw$. \begin{proof} As in the previous lemma, the idea is to find an appropriate structure $(D,E')$ isomorphic to $(D,E)$. For this, we define $E'(x,y)$ as follows: \[ E'(x,y) \defeq \begin{cases} E(x,y),& \text{if } x,y \neq a\\ E^*(x,y),& \text{otherwise} \end{cases} \] As before, \autoref{genericdigraph} tells us that we need to establish the extension property for $(D,E')$. Let $U,V,W \subset D$ be finite and pairwise disjoint. This time the proof splits into three cases. Case 1: $a \in U$. Let $U'=U\backslash \{a\}$ and $V'=V \cup \{a\}$. Then the extension property of $(D,E)$ applied to $U',V',W$ gives an appropriate $x$. Case 2: $a \in V$. Let $U'=U \cup \{a\}$ and $V'=V \backslash \{a\}$. Then again the extension property of $(D,E)$ gives us an appropriate $x$. Case 3: $a \in W$ or $a \notin U \cup V \cup W$. Then applying the extension property of $(D,E)$ gives us an appropriate $x$, without needing to modify $U,V$ or $W$. Thus, $(D,E')$ satisfies the extension property, and hence is isomorphic to $(D,E)$. We end by letting $f$ witness this isomorphism. \end{proof} \begin{lem} Let $a \in D$. Then there exists $f:D \to D$ such that \[ E(f(x),f(y)) \text{ if and only if } \begin{cases} x,y \neq a \text{ and } E(x,y)\\ x=a \text{ and } N(x,y)\\ y=a \text{ and } E^*(x,y) \end{cases} \] \end{lem} Remark: For the rest of this article, we fix such a function and denote it by $rot$. Remark: In words, $rot$ sends edges going out of $a$, to edges going into $a$, to non-edges, to edges going out of $a$. \begin{proof} Use the same strategy as for $-$ and $sw$. \end{proof} \begin{Def} \begin{enumerate}[(i)] \item We let $\Gamma = (D,E_{\Gamma})$, where $E_{\Gamma} \defeq E(x,y) \vee E^*(x,y)$. $\Gamma$ is a graph and, as will be proved later, is in fact (isomorphic to) the random graph. \item We let $-_{\Gamma} \in$ Sym$(D)$ be a function which interchanges the sets of edges and non-edges in $\Gamma$. \item Let $a \in D$. We let $sw_{\Gamma} \in$ Sym$(D)$ be a function which interchanges the sets of edges and non-edges adjacent to $a$, and preserves all other edges and non-edges. \end{enumerate} \end{Def} Remark: $(D,E_{\Gamma})$ is inter-definable with $(D,N)$, where $N(x,y)$ says that $xy$ is a non-edge. We now have all the background definitions necessary to state the main theorem: \begin{thm} \label{maintheorem} The reducts of $(D,E)$ are given by the following lattice, which we call $\mathcal{L}$: \begin{center} \begin{tikzpicture}[node distance=1.3cm, auto] \node (D) {Aut$(D,E)$}; \node (sw) [above left of=D] {$\langle sw \rangle$}; \node (glo) [above right of=D] {$\langle - \rangle$}; \node (sup) [above right of=sw] {$\langle sw,- \rangle$}; \node (Ra) [above of=sup, yshift=-2mm] {Aut($\Gamma$)}; \node (sw2) [above left of=Ra] {$\langle sw_{\Gamma} \rangle$}; \node (glo2) [above right of=Ra] {$\langle -_{\Gamma} \rangle$}; \node (sup2) [above right of=sw2] {$\langle sw_{\Gamma},-_{\Gamma} \rangle$}; \node (top) [above of=sup2] {Sym(D)}; \node (rot) [right of=glo] {$\langle rot \rangle$}; \node (rot2)[above of=rot, yshift=8mm] {$\langle -,rot \rangle$}; \draw[-] (D) to node {} (sw); \draw[-] (D) to node {} (glo); \draw[-] (sw) to node {} (sup); \draw[-] (glo) to node {} (sup); \draw[-] (sup) to node {} (Ra); \draw[-] (Ra) to node {} (sw2); \draw[-] (Ra) to node {} (glo2); \draw[-] (sw2) to node {} (sup2); \draw[-] (glo2) to node {} (sup2); \draw[-] (sup2) to node {} (top); \draw[-] (D) to node{}(rot); \draw[-] (rot) to node{}(rot2); \draw[-] (glo) to node{}(rot2); ..\draw[-] (rot2) to node{}(top); \end{tikzpicture} \end{center} \end{thm} This theorem can be split into two main claims. The first is that $\mathcal{L}$ is a sublattice of the reducts of $(D,E)$ (so for example one needs to show that the meets and joins are correct). The second claim is that $\mathcal{L}$ is in fact the whole lattice - that there are no other reducts. The second claim is the more interesting claim, and requires more work to prove. \section{Understanding the reducts} The purpose of this section is twofold. The first is to establish conditions for an unknown reduct $G$ of $(D,E)$ to be equal to or to contain particular elements of $\mathcal{L}$ - these lemmas will be used throughout the article. The second is to provide familiarity with the reducts, without which the article may be more difficult to understand. The first few lemmas will provide a concrete description of the three groups $\langle sw \rangle, \langle - \rangle$ and $\langle rot \rangle$. The way we do this is by comparing how two functions behave, via the following definition. \begin{Def} \label{behaves} Let $f,g: D \to D$ and $A \subset D$. We say \emph{$f$ behaves like $g$ on $A$} if for all finite tuples $\bar{a} \in A$, $f(\bar{a})$ is isomorphic (as a finite digraph) to $g(\bar{a})$. If $A=D$, we simply say \emph{$f$ behaves like $g$}. \end{Def} Example. All automorphisms of $(D,E)$ behave like the identity $id:D \to D$. Conversely, all $f \in Sym(D)$ which behave like $id$ are automorphisms. \textbf{Important Remark.} If $f:D \to D$ is any function and $g \in$ Aut$(D,E)$, then $h \defeq g \circ f$ behaves like $f$. The converse it also true: if $h$ behaves like $f$, then there is $g \in$ Aut$(D,E)$ such that $h= g \circ f$. Before continuing, we note the following useful fact. If a bijection $f$ and its inverse both preserve a definable relation $P$, then the group $\langle$Aut$(D,E) \cup \{f\} \rangle$ also preserves $P$. This follows straightforwardly by unravelling the definitions, and doing this would be a worthwhile exercise for the reader first encountering these notions. We start with the simplest of the three groups, $\langle - \rangle$. \begin{lem} \label{understanding-}Let $f \in Sym(D)$. Then $f \in \langle - \rangle \backslash$ Aut$(D,E) \Leftrightarrow f$ behaves like $-$. \end{lem} \begin{proof} ``$\Leftarrow$''. We need to show that $f \in \langle - \rangle$. Consider the function $g \defeq - \circ f$. It is easy to see that for all tuples $\bar{a} \in D$, $g(\bar{a})$ is isomorphic to $\bar{a}$. This means that $g$ behaves like $id$, so $g \in$ Aut$(D,E)$. Hence, $f = -^{-1} \circ g \in \gcl(-) \subseteq \langle - \rangle$, so we are done. ``$\Rightarrow$''. $-$ and $-^{-1}$ preserve the weakened edge relation $E_w (x,y;a,b) \defeq E(x,y) \leftrightarrow E(a,b)$, so $\langle - \rangle$ must also preserve $E_w$. In addition, the non-edge relation $N(x,y)$ is definable from $E_w$: $N(x,y) \Leftrightarrow \forall a,b (E_w(x,y;a,b) \leftrightarrow E_w(y,x;a,b))$; hence, $\langle - \rangle$ also preserves non-edges. Now suppose $f$ does not behave like $-$ on $D$ - we want to show $f \notin \langle - \rangle \backslash$ Aut$(D,E)$. If $f$ is an automorphism, then we're trivially done, so assume $f \notin$ Aut$(D,E)$. If $f$ does not preserve non-edges, then we are also done by the previous paragraph; so assume $f$ does preserve non-edges. The only possibility that remains is that there are edges $ab,cd \in D$ such that $E(f(a),f(b))$ and $\neg E(f(c),f(d))$. This means that $E_w(a,b;c,d)$ and $\neg E_w(f(a,b;c,d))$, i.e. that $f$ does not preserve $E_w$. Thus, $f \notin \langle - \rangle$, as required. \end{proof} Next we look at $\langle sw \rangle$. To do this we need some notation. For $A \subset D$, we let $sw_A: D \to D$ denote a function which behaves like $id$ on $A$ and $A^c$, and which switches the direction of all edges between $A$ and $A^c$. For example, $sw=sw_a$ for some $a \in D$, and, $sw_{\emptyset}$ is just an automorphism. The fact that $sw_A$ exists for all $A \subseteq D$ follows from the fact that all countable digraphs are embeddable in the generic digraph (\autoref{genericdigraph}). However, $sw_A$ cannot be a bijection for all $A \subset D$. This is because the image of the generic digraph on applying $sw_A$ may not be isomorphic to the digraph. For example, if you let $A=\{x \in D: E(a,x)\}$ where $a$ is some element of $D$, then $sw_A(a)$ will not have any outward edges. However, there are many subsets of $A$ for which $sw_A$ can be a bijection. For example, if $A \subset D$ is finite, one checks that the digraph obtained by switching with respect to a $A$ satisfies the extension property, so it is isomorphic to the generic digraph. A big idea in the next lemma is this: Let $a_1,\ldots,a_n$ be distinct elements of $D$. Then $sw_{a_1} \circ \ldots \circ sw_{a_n}$ behaves like $sw_A$, where $A=\{a_1,\ldots,a_n\}$. The problem with this idea is that, as stated, it is false: this is because the points $a_1,\ldots,a_n$ will not necessarily be fixed by each of the $sw_{a_i}$'s. \emph{Do} however keep this idea in mind, as it provides the intuition for (parts of) the lemma. \begin{lem} \label{understandingsw} \begin{enumerate}[(i)] \item $\gcl(sw) = \{f \in$ Sym$(D): f$ behaves like $sw_A$, for some finite $A \subset D\}.$ \item For all $A \subseteq D$, if $sw_A \in$ Sym($D)$ then $sw_A \in \langle sw \rangle$. \item For all proper non-empty $A \subset D$, if $sw_A \in$ Sym$(D)$ then $\langle sw_A \rangle = \langle sw \rangle$. \item $\langle sw \rangle = \{ f \in$ Sym$(D): f$ behaves like $sw_A$, for some $A \subseteq D\}.$ \end{enumerate} \end{lem} \begin{proof} For all of this proof, let $a \in D$ be the point such that $sw = sw_a$. (i) RHS $\subseteq$ LHS. From the important remark above, in order to show that every $f$ which behaves like $sw_A$ is in $\gcl(sw)$, it suffices to show that $sw_A \in \gcl(sw)$. First, we show that $sw_{a'} \in \gcl(sw)$, for all $a' \in D$. This is easy: let $g \in$ Aut$(D,E)$ map $a'$ to $a$. Then $sw \circ g \in \gcl(sw)$ and $sw \circ g$ behaves like $sw_{a'}$. Thus, again by the important remark, $sw_{a'} \in \gcl(sw)$. Now let $A=\{a_1,\ldots,a_n\} \subset D$. We start by letting $h_1 = sw_{a_1}$. Then let $h_2 = sw_{h_1(a_2)} h_1$ - observe that $h_2$ behaves like $sw_{\{a_1,a_2\}}$. Next let $h_3 = sw_{h_2(a_3)} h_2$ - $h_3$ behaves like $sw_{\{a_1,a_2,a_3\}}$. Continuing, we obtain $h_n$ which behaves like $sw_A$. By construction, $h_n \in \gcl(sw)$ and so by the important remark $sw_A \in \gcl(sw)$, as required. LHS $\subseteq$ RHS: Any $f \in \gcl(sw)$ can be written as $g_n sw^{\epsilon_n} \ldots g_1 sw^{\epsilon_1} g_0$, where the $g_i$ are automorphisms and $\epsilon_i \in \{1,-1\}$. Since $sw^{-1}$ behaves like $sw$, it is equal to $g\circ sw$, for some $g \in$ Aut$(D,E)$, so without loss, $\epsilon_i=1$ for all $i$. We prove by induction on $n$ that $f$ behaves like $sw_A$ for some finite $A$. In the base case, $f =g_0$ which behaves like $sw_{\emptyset}$. So assume that $f' \defeq g_n sw \ldots g_1 sw g_0$ behaves like $sw_A$ for some finite $A$; we consider $f=g_{n+1}sw f'$. Let $a' = f'^{-1}(a)$. If $a' \notin A$, then $f$ behaves like $sw_{A \cup \{a'\}}$. If $a' \in A$, then $f$ behaves like $sw_{A \backslash \{a'\}}.$ In both cases, we have what we want, thus completing the proof. (ii) Let $A \subseteq D$ and $sw_A \in$ Sym$(D)$. We need to show that for all finite tuples $\bar{d} \in D$, there exists $g \in \gcl(sw)$ such that $g(\bar{d})=sw_A(\bar{d})$. Let $A' = A \cap \bar{d}$. $A'$ is finite, so by part (i), $sw_{A'} \in \gcl(sw)$. Now, $sw_{A'}(\bar{d})$ is isomorphic to $sw_A(\bar{d})$, so by homogeneity let $h \in$ Aut$(D,E)$ map $sw_{A'}(\bar{d})$ to $sw_A(\bar{d})$. Letting $g=hsw_{A'}$ finishes the proof. (iii) Part (ii) tells us that $\langle sw_A \rangle \subseteq \langle sw \rangle$. To show the other direction, it suffices to show that $sw \in \langle sw_A \rangle$. So let $A \subset D$ be such that $A$ and $A^c$ are non-empty. By unravelling the definitions, we need to prove the following: For all $a_1,\ldots a_n \in D$, there exist $b_1,\ldots,b_n \in D$ such that $\bar{a} \cong \bar{b}$, and $A \cap \bar{b} = \{b_1\}$ or $\{b_2,\ldots,b_n\}$. If $A$ is finite, we let $b_1$ be any element of $A$ and find the remaining $b_2,\ldots,b_n$ by homogeneity. By the same reasoning, we are done if $A^c$ is finite. Hence, assume that $A$ is infinite and co-infinite. We prove the result by induction on the length $n$ of the tuple $\bar{a}$. The base case $n=1$ is trivial - simply let $b_1$ be any element of $A$. Now let $(a_1,\ldots, a_{n+1})$ be any tuple of length $n+1$. By the inductive hypothesis, we can find $(b_1,\ldots,b_n)$ isomorphic to $(a_1,\ldots,a_n)$ where $A \cap \bar{b} = \{b_1\}$ or $\{b_2,\ldots,b_n\}$. Without loss, we may assume that $A \cap \bar{b} = \{b_1\}$: the argument is symmetric in the other case. If we find $x \in A^c$ such that $(b_1,\ldots,b_n,x) \cong \bar{a}$, then we are done, so from now on assume that $(b_1,\ldots,b_n,x) \cong \bar{a}$ implies $x \in A$. $(*)$ Now consider a tuple $(c_1,\ldots,c_{n+1})$ satisfying the following: \begin{itemize} \item $c_1$ is some element of $A^c \backslash \{b_2,\ldots,b_n\}$. \item $\bar{c} \cong \bar{a}$. \item For each $2 \leq i \leq n+1$, $(b_1,\ldots,b_n,c_i) \cong \bar{a}$ \end{itemize} The first condition can be satisfied as $A^c$ is infinite. The latter two conditions can be satisfied because $(D,E)$ is homogeneous. By $(*), c_2,\ldots,c_{n+1} \in A$. So $(c_1,\ldots,c_{n+1})$ satisfies all the conditions that we want, completing the induction and hence the proof. (iv) By part (ii), we have RHS $\subseteq$ LHS. To prove the other direction, we find a relation $P$ that all functions in $\langle sw \rangle$ preserve, and show that if $f$ does not behave like $sw_A$ for any $A$, then $f$ does not preserve $P$. The relation is: \[ \begin{aligned} P(x,y,z) \defeq &(E(x,y) \wedge E(y,z) \wedge E(x,z))\\ \vee &(E^*(x,y) \wedge E^*(y,z) \wedge E(x,z))\\ \vee &(E^*(x,y) \wedge E(y,z) \wedge E^*(x,z))\\ \vee &(E(x,y) \wedge E^*(y,z) \wedge E^*(x,z)) \end{aligned} \] `Motto': A function preserves $P$ if for all tournaments on three vertices, it switches an even number of edges. To show that $\langle sw \rangle$ preserves $P$, it suffices to show that $sw$ preserves $P$. This is easy to see. First, $sw$ clearly preserves non-edges. Second, given any three vertices which form a tournament, either $sw$ does not switch any of the edges, or, it switches the direction of precisely two edges (and it would be those two edges which are adjacent to $a$). Now let $f \in$ Sym$(D)$ be a function which does not behave like $sw_A$ for any $A \subseteq D$. Define a partition of $D$ into subsets as follows: \begin{itemize} \item Let $A_0 = \{a_0\}$, where $a_0$ is any element of $D$. \item Let $A_1=\{x \in D: x$ is adjacent to $a_0$ and $f$ does not switch this edge$\}$ \item Let $B_1=\{x \in D: x$ is adjacent to $a_0$ and $f$ switches this edge$\}$ \item Let $A_2=\{x \in D:$ there is an edge between $A_1$ and $x$ that is not switched by $f \}$ \item Let $B_2=\{x \in D:$ there is an edge from $A_1$ to $x$ and all edges between $A_1$ and $x$ are switched by $f \}$ \item Let $A_3=\{x \in D:$ there are no edges between $A_1$ and $x$ and there is an edge between $B_1$ and $x$ switched by $f \}$ \item Let $B_3=\{x \in D:$ there are no edges between $A_1$ and $x$ and all edges between $B_1$ and $x$ are not switched by $f \}$. \end{itemize} By construction, these sets are pairwise disjoint. The fact their union equals $D$ follows from the fact that the maximum path length in the generic digraph is two. The idea behind defining these sets is that \emph{if} $f$ behaved like $sw_A$, then this procedure would find $A$ for us ($A$ would be the union of the $A_i$'s or the union of the $B_i$'s). In this light, let $A=A_0 \cup \ldots \cup A_3$ and $B$ be its complement. By assumption, $f$ does not behave like $sw_A$. What is left in the proof is simply a matter of case checking: we look at the possible reasons $f$ could not behave like $sw_A$ and show in each one that $f$ does not preserve $P$. Case 1a: There exists an edge $x,y \in A_1$ that is switched by $f$. Then consider the tournament $(a_0,x,y)$ - $f$ switches exactly one edge, so by the motto $f$ does not preserve $P$. Case 1b: There exists an edge $x,y \in B_1$ that is switched by $f$. Then $f$ switches all three edges of $(a_0,x,y)$, so $f$ does not preserve $P$. Case 1c: There exists an edge $x,y \in B_2$ switched in $f$. Let $z$ be any element of $A_1$. Then $f$ switches one edge in $(x,y,z)$. Case 1d: There exists an edge $x,y \in A_2$ switched by $f$. By definition of $A_2$ there is an $x' \in A_1$ such that $f$ does not switch the edge $x'x$, and there is a corresponding $y'$ for $y$. If $x'=y'$, then we get that $f$ switches one edge in $(x,y,x')$. If $x' \neq y'$, consider the tournament $(x,y,x',y')$. Now consider any element $z \in D$ such that there is an edge between $z$ and all the vertices $x,y,x',y'$. No matter what $f$ does to these edges, we will be able to find a tournament on three vertices on which $f$ switches an odd number of edges. For example, if $f$ switched all the edges between $z$ and $x,y,x',y'$, then look at $(x,y,z)$. \footnote{Note that what happens between between $x$ and $y'$ and between $x'$ and $y$ does not matter.} If there is an edge inside $B_3$ that is switched, then look at any point in $B_1$. If there is an edge inside $A_3$ that is switched, use a similar argument as in Case 1d but using $B_1$ in place of $A_1$. We have now dealt with all edges whose points lie in the same part. Case 2a: There is an edge $xy$ between $A_1$ and $B_1$ not switched by $f$. Then $f$ switches direction of one edge of $(a_0,x,y)$. Case 2b: There is an edge $xy$ between $A_1$ and $A_2$ which is switched by $f$. Let $y' \in A_1$ be such that $yy'$ is an edge not switched by $f$. Then look at $(x,y,x',a_0)$ and use the argument in Case 1d. We have now dealt with all edges containing a point in $A_1$. Case 2c: There is an edge $xy$ between $A_2$ and $A_3$ which is switched. Let $x' \in A_1$ be adjacent to $x$, and let $y' \in B_1$ be adjacent to $y$ such that $yy'$ is switched. Then consider $(a_0,x,y,x',y')$ and continue as in Case 1d. If there is an edge between $A_2$ and $B_1$ that is not switched, use Case 1d. Dealing with an edge between $A_2$ and $B_2$ that is not switched is straightforward. If there is an edge between $A_2$ and $B_3$ that is not switched, then continue as in Case 2c. We have now dealt with all edges containing a point in $A_2$. Case 2d: There is an edge $xy$ between $B_1$ and $B_2$ switched by $f$. Let $z \in A_1$ be a vertex adjacent to $y$. Consider the tournament $(x,y,z,a_0)$, and use the same argument as in Case 1d. Case 2e: There is an edge $xy$ between $B_1$ and $A_3$ not switched by $f$. Use an argument similar to Case 2b. This deals with all the edges containing a point in $B_1$. Case 2f: The case where there is an edge between $A_3$ and $B_3$ which is not switched is straightforward. If there is an edge $xy$ between $A_3$ and $B_2$ that is not switched, let $x' \in B_1$ be such that $xx'$ is an edge that is switched, and $y' \in A_1$ be an edge that is switched. Then look at $(x,y,x',y',a_0)$ and continue as in Case 1d. Case 2g: There is an edge between $B_2$ and $B_3$ which is switched. Continue as in 2f. This completes all the cases, and thus the proof. \end{proof} \textbf{Remark.} The proof of part (iv) also shows that $\langle sw \rangle = \{f \in Sym(D): f$ preserves $P(x,y,z) \}$. Due to the importance of this relation, we give it a definition. \begin{Def} Let $P_{sw}(x,y,z)$ be the 3-ary relation $P$ from the proof above.\end{Def} The next reduct we analyse is $\langle rot \rangle$. The ideas and proofs are analogous to those of $\langle sw \rangle$ so for the sake of conciseness, we will not go into as much detail and may only sketch the idea for some proofs. \textbf{Notation.} For what follows, $A,B,C \subseteq D$ are pairwise disjoint. For the ordered pair $(A,B)$, an \emph{outward edge} is an edge going from $A$ to $B$ and an \emph{inward edge} is one going from $B$ to $A$. We say $f$ behaves like $rot$ between $(A,B)$\footnote{We may also write `between $A$ and $B$'} if $f$ maps outward edges to inward edges to non-edges to outward edges. We let $rot_{A,B,C}$ be a function $D \to D$ which behaves like $id$ on $A,B$ and $C$ and behaves like $rot$ between $(A,B), (B,C)$ and $(C,A)$. If $C=(A \cup B)^c$, we just write $rot_{A,B}$. If in addition $C=\emptyset$, so that $B=A^c$, we just write $rot_A$. Simple observations: $rot=rot_a$ for some $a \in D$. If $f$ behaves like $rot_{A,B,C}$ then $f^2$ and $f^{-1}$ behave like $rot_{C,B,A}$, and $f^3$ behaves like $id$. $rot_{B,C,A}$ and $rot_{C,A,B}$ both behave like $rot_{A,B,C}$. As we did for $sw$, we describe a key idea in the following lemma. Let $a_1,\ldots, a_n,b_1,\ldots,b_m \in D$ be distinct elements. The idea is that $rot_{a_1}^2 \ldots rot_{a_n}^2 rot_{b_1} \ldots rot_{b_m}$ behaves like $rot_{A,B}$ where $A=\{a_1,\ldots a_n\}$ and $B=\{b_1,\ldots,b_n\}$. As before, this is not true as stated because the $a$'s and $b$'s are not fixed points of the functions involved. \begin{lem} \label{understandingrot} \begin{enumerate}[(i)] \item $\gcl(rot) = \{f \in $Sym$(D): f$ behaves like $rot_{A,B}$ where $A,B$ are finite$\}.$ \item For any disjoint $A,B \subseteq D$, if $rot_{A,B} \in$ Sym$(D)$ then $rot_{A,B} \in \langle rot \rangle$. \item Let $A,B$ be proper disjoint subsets of $D$ such that at least one of $A$ or $B$ is non-empty. If $rot_{A,B} \in$ Sym$(D)$, then $\langle rot_{A,B} \rangle = \langle rot \rangle$. \item $\langle rot \rangle = \{f \in $Sym$(D): f$ behaves like $rot_{A,B}$ where $A,B$ are disjoint subsets of $D\} $. \end{enumerate} \end{lem} \begin{proof} For this proof, let $a \in D$ be the point such that $rot=rot_a$. (i) RHS $\subseteq$ LHS. It suffices to show that $rot_{A,B} \in \gcl(rot)$. We start by showing that $rot_{a'} \in \gcl(rot)$ for all $a' \in D$. This is easy: let $g \in$ Aut$(D,E)$ map $a'$ to $a$ then consider $rot \circ g$. For the general case, let $A=\{a_1,\ldots,a_n\}$ and $B=\{b_1,\ldots,b_m\}$. The idea is to rotate twice about each element of $A$ and rotate once about each element of $B$ - we leave the details to the reader. LHS $\subseteq$ RHS. Any $f \in \gcl(rot)$ can be written in the form $g_n rot^{\epsilon_n} \ldots g_1 rot^{\epsilon_1} g_0$ where for all $i, \epsilon_i \in \{1,-1\}$. Since $rot^{-1}$ behaves like $rot^2$, we can assume that $\epsilon_i=1$ for all $i$. We prove by induction on $n$ that there exist finite disjoint $A,B \subset D$ such that $f$ behaves like $rot_{A,B}$. The base case $n=0$ is trivial, so assume that we know $h=g_{n-1} rot \ldots g_1 rot g_0$ behaves like $rot_{A,B}$ for finite $A,B$, and we consider $f=g_n rot\,h$. There are three cases depending on $a' \defeq h^{-1}(a)$. If $a' \notin A \cup B$, then $f$ behaves like $rot_{A,B \cup \{a'\}}.$ If $a' \in B$, then $f$ behaves like $rot_{A \cup \{a'\}, B \backslash \{a'\}}$. Lastly, if $a' \in A$, then $f$ behaves like $rot_{A \backslash \{a'\},B}$. This completes the induction and hence the proof. (ii) This is straightforward - just unravel the definitions and use part (i). (iii) Let $A,B \subseteq D$ be as described in the lemma, and let $C= (A \cup B)^c$. By (ii), we know that LHS $\subseteq$ RHS. To show the other direction, it suffices to show that $rot$ or $rot^{-1} \in \langle rot_{A,B} \rangle$. If one of $A,B$ or $C$ is empty, then we are done by imitating the corresponding argument for $\langle sw \rangle$. So assume $A,B$ and $C$ are all non-empty. Now, if $(B \cup C, E_{B \cup C}$ is isomorphic to the generic graph, then we can ignore $A$ and treat it as if it were empty, so again we can imitate the argument from the switching case to get the result. Hence, assume that $B \cup C$ is not isomorphic to the generic digraph. This means there exist finite, pairwise disjoint $U,V,W \subset B \cup C$ such that if $x \in D$ satisfies $\phi(x) \defeq (\forall u \in U E(u,x)) \wedge (\forall v \in V E^*(v,x)) \wedge (\forall w \in W N(w,x))$, then $x \in A$. Suppose that there exists $c \in C \cap (U \cup V \cup W)$. We will show that for all $(d_1,\ldots,d_n) \in D$, there exists $a_2,\ldots,a_n \in A$ such that $(c,a_2,\ldots,a_n) \cong (d_1,\ldots,d_n)$. By unravelling definitions, it is easy to see that this is sufficient to show that $rot \in \langle rot_{A,B} \rangle$. So, let $(d_1,\ldots,d_n) \in D$. Then let $(a_2,\ldots,a_n) \in D$ be such that $D \models \phi(a_2),\ldots,\phi(a_n)$ and $(c,a_2,\ldots,a_n) \cong (d_1,\ldots,d_n)$. Such $a_i$ exist by the homogeneity of $(D,E)$. Since $\phi(a_i)$ for all $i$, $(a_2,\ldots,a_n)$ has to be in $A$, as required, so $rot \in \langle rot_{A,B} \rangle$. Now suppose that $C \cap (U \cup W \cup V) = \emptyset$, so there must be $b \in B \cap (U \cup V \cup W)$. By repeating the argument above, we can show that $rot^{-1} \in \langle rot_{A,B} \rangle$, so we are done. (iv) From (ii) we have that RHS $\subseteq LHS$. To prove the other direction, we need to identify relations that $\langle rot \rangle$ preserves. These relations correspond to the orbits when you let $\gcl(rot)$ act on $(D,E)$. We describe the orbits diagrammatically: \begin{center} \begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=0.2cm,y=0.2cm] \begin{scope}[decoration={markings,mark=at position 0.5 with {\arrow[scale=0.5]{>}}}] \clip(2.,-9.) rectangle (69.,9.); \draw[postaction={decorate}] (8.,4.) -- (4.,8.); \draw[postaction={decorate}] (4.,4.) -- (4.,8.); \draw[postaction={decorate}] (8.,4.) -- (4.,4.); \draw[postaction={decorate}] (12.,4.) -- (16.,4.); \draw[postaction={decorate}] (16.,4.) -- (12.,8.); \draw[postaction={decorate}] (20.,8.) -- (20.,4.); \draw[postaction={decorate}] (24.,4.) -- (20.,8.); \draw[postaction={decorate}] (28.,8.) -- (28.,4.); \draw[postaction={decorate}] (28.,8.) -- (32.,4.); \draw[postaction={decorate}] (32.,4.) -- (28.,4.); \draw[postaction={decorate}] (34.,4.) -- (34.,8.); \draw[postaction={decorate}] (34.,4.) -- (38.,4.); \draw[postaction={decorate}] (34.,8.) -- (38.,4.); \draw[postaction={decorate}] (42.,8.) -- (46.,4.); \draw[postaction={decorate}] (54.,4.) -- (50.,4.); \draw[postaction={decorate}] (58.,8.) -- (58.,4.); \draw[postaction={decorate}] (62.,4.) -- (58.,4.); \draw[postaction={decorate}] (64.,4.) -- (64.,8.); \draw[postaction={decorate}] (4.,2.) -- (4.,-2.); \draw[postaction={decorate}] (8.,-2.) -- (4.,-2.); \draw[postaction={decorate}] (8.,-2.) -- (4.,2.); \draw[postaction={decorate}] (12.,-2.) -- (12.,2.); \draw[postaction={decorate}] (12.,-2.) -- (16.,-2.); \draw[postaction={decorate}] (16.,-2.) -- (12.,2.); \draw[postaction={decorate}] (24.,-2.) -- (20.,2.); \draw[postaction={decorate}] (32.,-2.) -- (28.,-2.); \draw[postaction={decorate}] (28.,2.) -- (32.,-2.); \draw[postaction={decorate}] (34.,2.) -- (34.,-2.); \draw[postaction={decorate}] (34.,-2.) -- (38.,-2.); \draw[postaction={decorate}] (34.,2.) -- (38.,-2.); \draw[postaction={decorate}] (42.,-2.) -- (42.,2.); \draw[postaction={decorate}] (42.,2.) -- (46.,-2.); \draw[postaction={decorate}] (54.,-2.) -- (50.,-2.); \draw[postaction={decorate}] (50.,-2.) -- (50.,2.); \draw[postaction={decorate}] (58.,-2.) -- (62.,-2.); \draw[postaction={decorate}] (64.,2.) -- (64.,-2.); \draw[postaction={decorate}] (8.,-8.) -- (4.,-4.); \draw[postaction={decorate}] (8.,-8.) -- (4.,-8.); \draw[postaction={decorate}] (16.,-8.) -- (12.,-4.); \draw[postaction={decorate}] (12.,-4.) -- (12.,-8.); \draw[postaction={decorate}] (12.,-8.) -- (16.,-8.); \draw[postaction={decorate}] (20.,-8.) -- (20.,-4.); \draw[postaction={decorate}] (24.,-8.) -- (20.,-4.); \draw[postaction={decorate}] (28.,-8.) -- (28.,-4.); \draw[postaction={decorate}] (28.,-4.) -- (32.,-8.); \draw[postaction={decorate}] (32.,-8.) -- (28.,-8.); \draw[postaction={decorate}] (34.,-4.) -- (38.,-8.); \draw[postaction={decorate}] (34.,-8.) -- (38.,-8.); \draw[postaction={decorate}] (42.,-4.) -- (42.,-8.); \draw[postaction={decorate}] (42.,-4.) -- (46.,-8.); \draw[postaction={decorate}] (50.,-8.) -- (50.,-4.); \draw[postaction={decorate}] (50.,-8.) -- (54.,-8.); \draw[postaction={decorate}] (58.,-4.) -- (58.,-8.); \draw[postaction={decorate}] (62.,-8.) -- (58.,-8.); \begin{scriptsize} \draw [fill=black] (4.,-2.) circle (1.0pt); \draw [fill=black] (4.,2.) circle (1.0pt); \draw [fill=black] (8.,-2.) circle (1.0pt); \draw [fill=black] (4.,-4.) circle (1.0pt); \draw [fill=black] (4.,-8.) circle (1.0pt); \draw [fill=black] (8.,-8.) circle (1.0pt); \draw [fill=black] (4.,4.) circle (1.0pt); \draw [fill=black] (8.,4.) circle (1.0pt); \draw [fill=black] (4.,8.) circle (1.0pt); \draw [fill=black] (12.,-2.) circle (1.0pt); \draw [fill=black] (12.,2.) circle (1.0pt); \draw [fill=black] (16.,-2.) circle (1.0pt); \draw [fill=black] (12.,-4.) circle (1.0pt); \draw [fill=black] (12.,-8.) circle (1.0pt); \draw [fill=black] (16.,-8.) circle (1.0pt); \draw [fill=black] (12.,4.) circle (1.0pt); \draw [fill=black] (16.,4.) circle (1.0pt); \draw [fill=black] (12.,8.) circle (1.0pt); \draw [fill=black] (20.,-2.) circle (1.0pt); \draw [fill=black] (20.,2.) circle (1.0pt); \draw [fill=black] (24.,-2.) circle (1.0pt); \draw [fill=black] (20.,-4.) circle (1.0pt); \draw [fill=black] (20.,-8.) circle (1.0pt); \draw [fill=black] (24.,-8.) circle (1.0pt); \draw [fill=black] (20.,4.) circle (1.0pt); \draw [fill=black] (24.,4.) circle (1.0pt); \draw [fill=black] (20.,8.) circle (1.0pt); \draw [fill=black] (28.,-2.) circle (1.0pt); \draw [fill=black] (28.,2.) circle (1.0pt); \draw [fill=black] (32.,-2.) circle (1.0pt); \draw [fill=black] (28.,-4.) circle (1.0pt); \draw [fill=black] (28.,-8.) circle (1.0pt); \draw [fill=black] (32.,-8.) circle (1.0pt); \draw [fill=black] (28.,4.) circle (1.0pt); \draw [fill=black] (32.,4.) circle (1.0pt); \draw [fill=black] (28.,8.) circle (1.0pt); \draw [fill=black] (34.,-2.) circle (1.0pt); \draw [fill=black] (34.,2.) circle (1.0pt); \draw [fill=black] (38.,-2.) circle (1.0pt); \draw [fill=black] (34.,-4.) circle (1.0pt); \draw [fill=black] (34.,-8.) circle (1.0pt); \draw [fill=black] (38.,-8.) circle (1.0pt); \draw [fill=black] (34.,4.) circle (1.0pt); \draw [fill=black] (38.,4.) circle (1.0pt); \draw [fill=black] (34.,8.) circle (1.0pt); \draw [fill=black] (42.,-2.) circle (1.0pt); \draw [fill=black] (42.,2.) circle (1.0pt); \draw [fill=black] (46.,-2.) circle (1.0pt); \draw [fill=black] (42.,-4.) circle (1.0pt); \draw [fill=black] (42.,-8.) circle (1.0pt); \draw [fill=black] (46.,-8.) circle (1.0pt); \draw [fill=black] (42.,4.) circle (1.0pt); \draw [fill=black] (46.,4.) circle (1.0pt); \draw [fill=black] (42.,8.) circle (1.0pt); \draw [fill=black] (50.,-2.) circle (1.0pt); \draw [fill=black] (50.,2.) circle (1.0pt); \draw [fill=black] (54.,-2.) circle (1.0pt); \draw [fill=black] (50.,-4.) circle (1.0pt); \draw [fill=black] (50.,-8.) circle (1.0pt); \draw [fill=black] (54.,-8.) circle (1.0pt); \draw [fill=black] (50.,4.) circle (1.0pt); \draw [fill=black] (54.,4.) circle (1.0pt); \draw [fill=black] (50.,8.) circle (1.0pt); \draw [fill=black] (58.,-2.) circle (1.0pt); \draw [fill=black] (58.,2.) circle (1.0pt); \draw [fill=black] (62.,-2.) circle (1.0pt); \draw [fill=black] (58.,-4.) circle (1.0pt); \draw [fill=black] (58.,-8.) circle (1.0pt); \draw [fill=black] (62.,-8.) circle (1.0pt); \draw [fill=black] (58.,4.) circle (1.0pt); \draw [fill=black] (62.,4.) circle (1.0pt); \draw [fill=black] (58.,8.) circle (1.0pt); \draw [fill=black] (64.,-2.) circle (1.0pt); \draw [fill=black] (64.,2.) circle (1.0pt); \draw [fill=black] (68.,-2.) circle (1.0pt); \draw [fill=black] (64.,-4.) circle (1.0pt); \draw [fill=black] (64.,-8.) circle (1.0pt); \draw [fill=black] (68.,-8.) circle (1.0pt); \draw [fill=black] (64.,4.) circle (1.0pt); \draw [fill=black] (68.,4.) circle (1.0pt); \draw [fill=black] (64.,8.) circle (1.0pt); \end{scriptsize} \end{scope} \end{tikzpicture} \end{center} This diagram contains all the possible digraphs you can have on a triple in $D$. Each row of the diagram represents one of the orbits and hence, one of the relations that $\langle rot \rangle$ preserves. Let $P_{rot,1}, P_{rot,2}$ and $P_{rot,3}$ be the relations for the top, middle and bottom rows respectively. One feature worth noting is that given any finite triple in $D$, if you change the relation between exactly one pair of its vertices, you change the orbit the triple is in. For example, given a triple with only non-edges (so it is in $P_{rot,3}$), changing exactly one non-edge into an edge results in the triple no longer being in $P_{rot,3}$. Now let $f \in \langle rot \rangle$. We know that $f$ preserves $P_{rot,i}$, $i=1,2,3$. We want to find disjoint $A,B \subseteq D$ such that $f$ behaves like $rot_{A,B}$. We do this as follows. Pick any $a \in D$. Let $A=\{a\} \cup \{x \in D: E(a,x) \wedge E(f(a,x))$ or $E^*(a,x) \wedge E^*(f(a,x))$ or $N(a,x) \wedge N(f(a,x)) \}$. Let $B=\{x \in D: E(a,x) \wedge E^*(f(a,x))$ or $E^*(a,x) \wedge N(f(a,x))$ or $N(a,x) \wedge E(f(a,x)) \}$. We claim that $f$ behaves like $rot_{A,B}$. This amounts to case checking, which we leave to the reader. We provide one case as an example. Case 1. We need to show that $f$ behaves like $id$ on $A$. Suppose not, and let $a_1,a_2 \in A$ witness this fact. Then we have $(a,a_1,a_2)$ such that $f$ only changes what happens between $a_1$ and $a_2$, contradicting that $f$ preserves $P_{rot,i}$. \end{proof} The relations introduced in this proof are important, so we give them a definition. \begin{Def} For $i=1,2,3$, let $P_{rot,i}(x,y,z)$ be the relations defined in the proof of part (iv) of the lemma above. \end{Def} The descriptions of $\langle -, sw \rangle$ and $\langle -, rot \rangle$ are straightforward: \begin{lem} \label{understandingjoinwith-} \begin{enumerate}[(i)] \item $\langle -,sw \rangle = \{f \in$ Sym$(D): f=g$ or $- \circ g$ for some $g \in \langle sw \rangle \}$. \item $\langle -,rot \rangle = \{f \in$ Sym$(D): f=g$ or $- \circ g$ for some $g \in \langle rot \rangle \}$. \end{enumerate} \end{lem} \begin{proof} (i)$\langle -,sw \rangle$ preserves the 6-ary relation $P_{sw,w} \defeq P_{sw}(\bar{x}) \leftrightarrow P_{sw}(\bar{y})$. Now let $f \in \langle -,sw \rangle$. If $f$ preserves $P_{sw}$, then by \autoref{understandingsw} $f \in \langle sw \rangle$. Now suppose that $f$ does not preserve $P_{sw}$. Since $f$ preserves $P_{sw,w}$, we have that $- \circ f$ preserves $P_{sw}$, so $- \circ f=g \in \langle sw \rangle$. Hence, $f=-^{-1}g$. We can replace $-^{-1}$ by $-$ because $-^{-1}=-\circ h$ for some $h \in$ Aut$(D,E)$. (ii)$\langle -,rot \rangle$ preserves the 6-ary relation $P_{rot,w} \defeq (P_{rot,1}(\bar{x}) \wedge P_{rot,1}(\bar{y})) \vee (P_{rot,2}(\bar{x}) \wedge P_{rot,2}(\bar{y}))$. Now let $f \in \langle -,rot \rangle$. If $f$ preserves $P_{rot,1}$, then by \autoref{understandingrot} $f \in \langle rot \rangle$. Now suppose that $f$ does not preserve $P_{rot,1}$. Since $f$ preserves $P_{rot,w}$, we have that $- \circ f$ preserves $P_1$, so $- \circ f=g \in \langle rot \rangle$. Hence, $f=-^{-1}g$. We can replace $-^{-1}$ by $-$ because $-^{-1}=-\circ h$ for some $h \in$ Aut$(D,E)$. \end{proof} The next lemmas will give us conditions on a group $G$ to be equal to Sym$(D)$ or to contain Aut($\Gamma$). \begin{lem} \label{equalsSymD} Let $G \leq$ Sym$(D)$ be a closed supergroup of Aut$(D,E)$. \begin{enumerate}[(i)] \item If $G$ is $n$-transitive for all $n \in \mathbb{N}$, then $G=$ Sym$(D)$. \item If $G$ is $n$-homogeneous for all $n \in \mathbb{N}$, then $G=$ Sym$(D)$. \item Suppose that whenever $A \subset D$ is finite and has edges, there exists $g \in G$ such that $g(A)$ has less edges than in $A$ (i.e. $|\{(x,y) \in A^2: E(g(x),g(y))\}| < |\{(x,y) \in A^2: E(x,y)\}|)$. Then, $G=$ Sym$(D)$. \item Suppose that there exists a finite $A \subset D$ and $g \in G$ such that $g$ behaves like $id$ on $D\backslash A$, $g$ behaves like $id$ between $A$ and $D\backslash A$, and, $g$ deletes at least one edge in $A$. Then, $G=$ Sym$(D)$. \end{enumerate} \end{lem} Remark: $G$ is $n$-transitive if for all pairs of tuples $\bar{x}, \bar{y} \in D^n$, there exists $g \in G$ such that $g(\bar{x})=\bar{y}$. $G$ is $n$-homogeneous if for all subsets $A,B \subset D$ of size $n$, there exists $g \in G$ such that $g(A)=B$. \begin{proof} (i) Let $f \in$ Sym$(D)$. We want to show that $f \in G$. Since $G$ is closed, it suffices to show that for all finite tuples $\bar{a} \in D$, there exists $g \in G$ such that $g$ maps $\bar{a}$ to $\bar{f(a)}$. But $G$ is $n$-transitive for all $n$, so we can always find an appropriate $g$, so we are done. (ii) We will show that $G$ is $n$-transitive for all $n$. Let $\bar{a},\bar{b}$ be tuples of length $n$ in $D$. Let $f \in G$ be such that $f(\bar{a})$ is empty; this is possible as $G$ is $n$-homogeneous. Similarly, let $g \in G$ be such that $g(\bar{b})$ is empty. Now consider the map $h: f(a_i) \mapsto g(b_i)$. This is an isomorphism of digraphs so can be extended to an automorphism $h'$ of Aut$(D,E)$, by homogeneity. But now $g^{-1}h'f \in G$ maps the tuple $\bar{a}$ to the tuple $\bar{b}$, as required. (iii) We will show that $G$ is $n$-homogeneous for all $n$. It suffices to show that for all finite $A \subset D$, we can map $A$ to the empty digraph. We prove this by induction on the number of edges $k$ in $A$. The base case $k=0$ is trivial. Now let $A$ have $k$ edges. By assumption, there is $f \in G$ such that $f(A)$ has $k'<k$ edges. By the inductive hypothesis, there is $g \in G$ such that $g(f(A))$ is the empty digraph, so we are done. (iv) Let $A$ and $g$ be as in the lemma. We will show that for all finite $B \subset D$, if $B$ contains edges then there is $f \in G$ such that $f(B)$ has less edges than $B$ - this suffices by (iii). So let $B \subset D$ be finite. Let $bb'$ be an edge in $B$, and let $aa' \in A$ be an edge that is deleted by $g$. Let $h$ be an automorphism mapping $bb'$ to $aa'$. Then $gh \in G$ and $gh(B)$ contains less edges than in $B$, as required. \end{proof} Before we describe conditions for $G$ to contain Aut$(\Gamma)$, we first establish a fact we have mentioned earlier, which is that $\Gamma$ is indeed the random graph. \begin{lem} $\Gamma$ is isomorphic to the random graph. \end{lem} \begin{proof} Recall that we defined $\Gamma$ to be $(D,E_{\Gamma})$, where $E_{\Gamma}(x,y) \defeq E(x,y) \vee E(y,x)$. To show $\Gamma$ is isomorphic to the random graph, it suffices to show it satisfies the extension property of the random graph. So let $U,W \subset D$ be finite disjoint - we need to find $x \in D \backslash (U \cup W)$ such that $E_{\Gamma}(x,u)$ for all $u \in U$ and $N(x,w)$ for all $w \in W$. Apply the extension property of the digraph (\autoref{genericdigraph}) to $U, \emptyset, W$ to find an appropriate $x$. \end{proof} \begin{lem} \label{containsgamma} Let $G \leq$ Sym$(D)$ be a closed supergroup of Aut$(D,E)$. \begin{enumerate}[(i)] \item Suppose that whenever $a_1,\ldots,a_n, b_1, \ldots, b_n \in D$ satisfy $N(a_i,a_j) \leftrightarrow N(b_i,b_j)$ for all $i,j$, there exists $g \in G$ such that $g(\bar{a})=\bar{b}$. Then $G \geq$ Aut$(\Gamma)$. \item Suppose that for all $A=\{a_1,\ldots,a_n\} \subset D$, there exists $g \in G$ such that for all edges $a_ia_j$ in $A$, $E(g(a_i),g(a_j))$ iff $i<j$. (Intuitively, such a $g$ is switching the edges so they all point in the same direction.) Then, $G \geq$ Aut$(\Gamma)$. \item Suppose that there exists a finite $A \subset D$ and $g \in G$ such that $g$ behaves like $id$ on $D \backslash A$, $g$ behaves like $id$ between $A$ and $D \backslash A$, and, $g$ switches the direction of (at least) one edge in $A$. Then, $G \geq$ Aut$(\Gamma)$. \end{enumerate} \end{lem} \begin{proof} (i) Let $f \in$ Aut$(\Gamma)$ and let $\bar{a} \in D$ be a finite tuple. We need to find $g \in G$ such that $g(\bar{a})=f(\bar{a})$. Since $f \in$ Aut$(\Gamma)$, we have that $N(a_i,a_j) \leftrightarrow N(f(a_i),f(a_j))$ for all $i,j$. Hence, by the assumptions given in the lemma, there exists an appropriate $g \in G$. (ii) Let $\bar{a}$ and $\bar{b} \in D$ satisfy $N(a_i,a_j) \leftrightarrow N(b_i,b_j)$ - we will show that there is $f \in G$ s.t $f(\bar{a})=\bar{b}$. Let $g_1 \in G$ be a function such that for all edges $a_ia_j \in A$, $E(g(a_i),g(a_j))$ iff $i<j$; such a function exists by assumption. Let $g_2 \in G$ be the corresponding function for $\bar{b}$. By construction, $g_1(\bar{a})$ and $g_2(\bar{a})$ are isomorphic so there is an automorphism $h \in$ Aut$(D,E)$ mapping $g_1(\bar{a})$ to $g_2(\bar{a})$. But then $g_2^{-1}hg_1(\bar{a})=\bar{b}$. Hence we are done by part (i). (iii) Let $A$ and $g$ be as stated in the lemma, and let $aa' \in A$ be an edge whose direction is switched by $g$. Claim: Let $\bar{b} \in D$ be finite and let $bb'$ be any edge in $\bar{b}$. Then there exists $f \in G$ such that $f$ switches the direction of $bb'$ and behaves like $id$ everywhere else on $\bar{b}$. This is easy: By homogeneity, there exists $h \in$ Aut$(D,E)$ such that $h(bb')=aa'$ and $h(\bar{b}) \cap A = \{a,a'\}$. Then $f=gh$ is the function we want. Now suppose we have two tuples $\bar{b},\bar{c}$ as in the statement of (i); we want to find a function in $G$ mapping one to the other. We do this by repeatedly using the above claim to switch the edges in $\bar{b}$ until they are all aligned with the edges in $\bar{c}$. \end{proof} \section{$\mathcal{L}$ is a sublattice of the reducts of $(D,E)$} Before we begin please note a convention that we will use for the remainder of the article. There will be proofs where we want to show that we can map a digraph $A$ to a related digraph $B$. Often, the function will be the composition of a sequence of functions $f_1,f_2,\ldots$, where the definition of each one will depend on those defined earlier. For example, suppose we have defined $f_1$ and $f_2$, and $f_3$ is going to be a switching function. The convention is that we will say `Let $f_3$ be $sw_{A'}$' (where $A'$ will be a particular subset of $A$), in place of the strictly correct phrase `Let $f$ be $sw_{f_2f_1(A')}$'. This may seem odd, but it has benefits. First, the proofs will be easier to follow and will better match the underlying intuition behind the argument. Second, with this convention in place, we often avoid needing to name the functions: We can now use phrases like `First switch about the subset $A_1$, then apply $rot$ about the point $a$', whereas without the convention we would have to say `...then apply $rot$ about the point which is the current image of $a$'. \begin{lem} \label{sublattice} \begin{enumerate}[(i)] \item $\langle - \rangle, \langle sw \rangle$ and $\langle rot \rangle$ are proper reducts of Aut$(D,E)$. \item $\langle - \rangle, \langle sw \rangle$ and $\langle rot \rangle$ are not reducts of each other. \item $\langle -, sw \rangle$ is a proper reduct of $\langle - \rangle$ and $\langle sw \rangle$, and is not equal to Sym$(D)$. \item $\Gamma$ is a proper reduct of $\langle -, sw \rangle$ \item $\langle -,rot \rangle$ is a proper reduct of $\langle - \rangle$ and $\langle rot \rangle$, and is not equal to Sym$(D)$. \item The join of $\langle rot \rangle$ and $\langle sw \rangle$ is Sym$(D)$. \item The meet of $\langle sw \rangle$ and $\langle - \rangle$ is Aut$(D)$. \item The meet of $\langle rot \rangle$ and $\langle sw_{\Gamma}, -_{\Gamma} \rangle$ is Aut$(D)$. \item The meet of $\langle -,rot \rangle$ and $\langle sw_{\Gamma}, -_{\Gamma} \rangle$ is $\langle - \rangle$. \end{enumerate} \end{lem} \begin{proof} (i) This is immediate from the definition of the $\langle \cdot \rangle$. (ii) We need to identify for each reduct a relation that it preserves but which the other two do not preserve. For $\langle - \rangle$ the relation is $E_w$, for $\langle sw \rangle$ the relation it preserves is $P_{sw}$ and for $\langle rot \rangle$ we have $P_{rot, 1}$. (iii) By (ii), $\langle -,sw \rangle$ is a proper reduct of $\langle - \rangle$ and $\langle sw \rangle$. It preserves $P_{sw,w}$, so it is not equal to Sym$(D)$. (iv) Both $-$ and $sw$ preserve $N(x,y)$, so $\langle -, sw \rangle \subseteq$ Aut$(D,N)=\Gamma$. $\Gamma$ is a proper reduct because $\langle -, sw \rangle$ preserves $P_{sw,w}$ but $\Gamma$ does not. (v) By (ii), $\langle -,rot \rangle$ is a proper reduct of $\langle - \rangle$ and $\langle rot \rangle$. It preserves $P_{rot,w}$, so it is not equal to Sym$(D)$. (vi) By \autoref{equalsSymD} (iii), it suffices to show that for all finite $A \subset D$ that has at least one edge, we can find $g \in \langle sw, rot \rangle$ such that $g(A)$ has less edges than in $A$. Let $a \in A$ be a point adjacent to at least one edge. Let $A_1 = \{a' \in A: E(a,a')\}, A_2=\{a' \in A: E(a',a)\}$ and $A_3=\{a' \in A: N(a,a')\}$. First, switch about the subset $A_1$ - the result is that now all the edges adjacent to $a$ are edges going into $a$. Now apply $rot_a^2$: the edges between $a$ and $A_1 \cup A_2$ become outward edges, and the non-edges between $a$ and $A_3$ become inward edges. Now apply $sw_{A_1 \cup A_2}$: the outward edges from $a$ to $A_1 \cup A_2$ now become inward edges. Therefore, between $a$ and $A \backslash \{a\}$ we now only have inward edges. Applying $rot_a$ for the last time results in all these edges becoming non-edges. By noting that at every step, the number of edges within $A \backslash \{a\}$ remains the same, we have shown that we can reduce the number of edges in $A$ using functions in $\langle sw,rot \rangle$, which is what was required. (vii) Let $f \in \langle - \rangle \cap \langle sw \rangle$. By \autoref{understandingsw}, $f$ behaves like $sw_A$ for some $A \subseteq D$. $A$ or $A^c$ must contain an edge. Hence, there exists an edge whose direction $f$ does not switch. In particular, $f$ does not behave like $-$. By \autoref{understanding-}, we conclude that $f$ has to be an automorphism of Aut$(D,E)$, as required. (viii) We first establish some notation. We say $f: D \to D$ graph-behaves like $g: D \to D$ if for all $\bar{a} \in D$, $f(\bar{a})$ is isomorphic to $g(\bar{a})$ as \emph{undirected} graphs. We abbreviate `graph-behaves' by `g-behaves'. Let $A \subseteq D$. We say $f: D \to D$ g-behaves like $sw_{\Gamma,A}$ if $f$ g-behaves like $id$ on $A$ and on $A^c$ and if $f$ swaps edges and non-edges between $A$ and $A^c$. By folklore (or by duplicating the arguments in Section 2), $\langle sw_{\Gamma} \rangle = \{f \in$ Sym$(D): f$ g-behaves like $sw_{\Gamma,A}$ for some $A \subseteq D\}$, and $\langle -_{\Gamma}, sw_{\Gamma} \rangle = \{f \in$ Sym$(D): \exists g \in \langle sw_{\Gamma} \rangle$ such that $f=g$ or $f=-_{\Gamma}\circ g \}$. Let $f \in \langle rot \rangle \cap \langle sw_{\Gamma}, -_{\Gamma} \rangle$. By \autoref{understandingrot}, there exists disjoint $A,B \subseteq D$ such that $f$ behaves like $rot_{A,B}$; let $C=(A \cup B)^c$. We split into two cases. Case 1. $f \in \langle sw_{\Gamma} \rangle$, so $f$ g-behaves like $sw_{\Gamma,U}$ for some $U \subseteq D$; let $V=U^c$. To show that $f \in$ Aut$(D,E)$ it suffices to show that two of $A,B$ and $C$ must be empty. Suppose without loss that $A$ is non-empty, so we want to show that $B$ and $C$ are empty. Since $f$ behaves like $id$ on $A$, $A$ must be a subset of $U$ or a subset of $V$. Without loss, suppose $A \subseteq U$. Similarly, $B$ and $C$ must each be a subset of $U$ or $V$. Furthermore, if $B$ is non-empty it cannot be a subset of $U$; this is because $f$ preserves non-edges in $U$ but $f$ does not preserve non-edges between $A$ and $B$. Similarly, if $C$ is non-empty, then $C \subseteq V$. So if both $B$ and $C$ are non-empty, then they must both be subsets of $V$, which is not possible by the same reasoning. Hence, one of $B$ or $C$ must be empty - without loss we may assume that $C$ is empty. Now we have that $B=A^c$ is non-empty, $A \subseteq U$ and $B \subseteq V$. Hence, $A=U$ and $B=V$. By homogeneity of $D$, there must be an outward edge from $A$ to $B$. But now we get a contradiction: $f$ behaving like $rot_{A,B}$ implies that this edge is mapped to an edge, whereas $f$ g-behaving like $sw_{\Gamma,U}$ implies that $f$ maps this edge to a non-edge. Thus, $B$ must also be non-empty, as required. Case 2. $f=-_{\Gamma}\circ g$ for some $g \in \langle sw_{\Gamma} \rangle$. Let $U \subseteq D$ be such that $g$ g-behaves like $sw_{\Gamma,U}$. Now, for any subset $X$ of $D$ of size at least three, $f$ cannot act like the $id$ on $X$. This is because either $|X \cap U| \geq 2$ or $|X \cap U^c| \geq 2$ and we know that $f$ g-behaves like $-_{\Gamma}$ on $U$ and on $U^c$. However, we also know that $f$ behaves like the $id$ on $A,B$ and $C$, and at least one of them has size at least three. Thus, we have a contradiction. (ix) Let $f \in \langle -,rot \rangle \cap \langle sw_{\Gamma}, -_{\Gamma} \rangle$. By \autoref{understandingjoinwith-}, there exists $g \in \langle rot \rangle$ such that $f=g$ or $f=-\circ g$. Since $f \in \langle - \rangle \Leftrightarrow - \circ f \in \langle - \rangle$, without loss we may assume that $f=g$, i.e. that $f \in \langle rot \rangle$. By (viii), it follows that $f \in$ Aut$(D,E)$, so we are done. \end{proof} \section{$\mathcal{L}$ contains all the reducts} The task of showing that $\mathcal{L}$ contains all the reducts is split up into these lemmas: \begin{lem}\label{threeregions} Let $G$ be a reduct of Aut$(D,E)$. Then either $G$ contains Aut$(\Gamma)$, is contained in Aut$(\Gamma)$, or contains $\langle rot \rangle.$ \end{lem} \begin{lem}\label{region1} Let $G$ be a reduct of Aut$(D,E)$ that contains Aut$(\Gamma)$. Then $G=\Gamma, \langle sw_{\Gamma} \rangle,$ $\langle -_{\Gamma} \rangle, \langle sw_{\Gamma}, -_{\Gamma} \rangle$ or Sym$(D)$. \end{lem} \begin{lem}\label{region2} Let $G$ be a reduct of Aut$(D,E)$ that is contained in Aut($\Gamma)$. Then $G=$Aut$(D,E)$, $\langle sw \rangle$, $\langle - \rangle, \langle sw, - \rangle$ or Aut$(\Gamma)$. \end{lem} \begin{lem} \label{region3} Let $G$ be a reduct of Aut$(D,E)$ that contains $\langle rot \rangle$. Then $G=\langle rot \rangle, \langle rot,- \rangle$ or Sym$(D)$. \end{lem} The main tool that will be used to prove these lemmas will be that of canonical functions, as developed by Bodirsky and Pinsker in \cite{bp11} and \cite{bpt13}. However, before delving into the use of canonical functions, the next subsection describes the details that are obtained by other means. \subsection{Using the classification of the reducts of the random graph and of the random tournament} Knowing the reducts of the random graph is evidently necessary for this result, but it is also helpful to know the reducts of the random tournament. We begin by stating these two classifications. \textbf{Notation.} \begin{enumerate}[(i)] \item We let $\mathcal{T}=(T,E_T)$ denote the random tournament. This can be defined as the countable homogeneous tournament which embeds all finite tournaments. \item Let $-_{\mathcal{T}}$ denote a function which switches the direction of all edges in the random tournament. \item Let $sw_{\mathcal{T}}$ denote a function which switches the direction of only those edges that are adjacent to a particular fixed vertex. \end{enumerate} \begin{thm}\label{gammaandt} \begin{enumerate}[(i)] \item (Thomas \cite{tho91}.) The reducts of the random graph are: $\Gamma, \langle sw_{\Gamma} \rangle,$ $\langle -_{\Gamma} \rangle, \langle sw_{\Gamma}, -_{\Gamma} \rangle$ and the full symmetric group. \item (Bennett, \cite{ben97}.) The reducts of the random tournament are: Aut$(T,E_T),\langle sw_{\mathcal{T}} \rangle, \langle -_{\mathcal{T}} \rangle,$ $\langle sw_{\mathcal{T}}, -_{\mathcal{T}} \rangle$ and the full symmetric group Sym($T$). \end{enumerate} \end{thm} We immediately get: \begin{proof}[Proof of \autoref{region1}] This is exactly the statement of \autoref{gammaandt} (i). \end{proof} Knowing the reducts of the random tournament contributes to the proof of \autoref{region2}, via the following construction: \begin{Def} Let $G$ be a reduct of $(D,E)$. We let $T(G)=\{ f \in$ Sym$(T):$ for all finite tuples $\bar{a} \in T,$ there exist $g \in G$ and a tuple $\bar{b} \in D$ such that $\bar{a} \cong \bar{b}$ and $f(\bar{a}) \cong g(\bar{b})\}$. \end{Def} In words, $T(G)$ contains those functions whose behaviour on finite sets can be replicated by functions in $G$. The intuition is that $T(G)$ tells us what $G$ can do to tournaments. The idea behind this concept is as follows: We show that $T(G)$ must be a reduct of $\mathcal{T}$, so by \autoref{gammaandt} $T(G)$ has five different possibilities. Now if we assume that $G$ fixes non-edges, $G$ can only change the direction of edges. From this, one might suspect that $G$ is determined by how it behaves on tournaments, i.e., that $G$ is determined by $T(G)$. \begin{lem} \label{region2a} Let $G$ be a reduct of $(D,E)$. Then $T(G)$ is a reduct of $\mathcal{T}$. \end{lem} \begin{proof} We need to show that $T(G)$ is a closed supergroup of Aut$\mathcal(T)$. This is an easy exercise in unravelling definitions. We demonstrate by showing that $T(G)$ is closed under composition, and leave the remaining conditions to the reader. Let $f,f' \in T(G)$. We want to show that $f'f \in T(G)$, so let $\bar{a} \in T$ be a finite tuple. Since $f \in T(G)$ we can find $g \in G$ and $\bar{b} \in D$ such that $\bar{a} \cong \bar{b}$ and $f(\bar{a}) \cong g(\bar{b})$. Since $f' \in T(G)$, we can find $g' \in G$ such that $f' (f(\bar{a})) \cong g' (g(\bar{b}))$. Then $g'g$ and $\bar{b}$ satisfy $\bar{a} \cong \bar{b}$ and $f'f(\bar{a}) \cong g'g(\bar{b})$, as required. \end{proof} \begin{lem} \label{region2b} Let $G$ be a reduct of $(D,E)$ contained in Aut$(\Gamma)$. Then: \begin{enumerate}[(i)] \item $G=$Aut$(D,E) \Leftrightarrow T(G)=$Aut$(T,E_T)$. \item $G=\langle sw \rangle \Leftrightarrow T(G) = \langle sw_{\mathcal{T}} \rangle$. \item $G=\langle - \rangle \Leftrightarrow T(G) = \langle -_{\mathcal{T}} \rangle $. \item $G=\langle sw,- \rangle \Leftrightarrow T(G) = \langle sw_{\mathcal{T}}, -_{\mathcal{T}} \rangle$. \end{enumerate} \end{lem} \begin{proof} The following claims are used in all four parts of the lemma. textbf{Claim 1.} $T(G)=$ Aut$(T,E_T), \langle sw_{\mathcal{T}} \rangle, \langle -_{\mathcal{T}} \rangle,$ $\langle sw_{\mathcal{T}}, -_{\mathcal{T}} \rangle$ or Sym$(T)$. \emph{Proof of Claim 1.} This follows immediately from \autoref{region2a} and \autoref{gammaandt}. \textbf{Claim 2.} Let $g \in G$ and let $\bar{b} \in D$ be a tournament. Then there exist $f \in T(G)$ and $\bar{a} \in T$ s.t $\bar{a} \cong \bar{b}$ and $f(\bar{a}) \cong g(\bar{b})$. \emph{Proof of Claim 2.} Let $T_1 \subset D$ satisfy: \begin{itemize} \item $\bar{b} \in T_1$ \item $(T_1,E|_{T_1})$ is isomorphic to the random tournament. \item $T_1$ is a maximal tournament in $D$, i.e. for all $x \in D\backslash T_1$, there exists $y \in T_1$ such that $N(x,y)$. \end{itemize} We sketch how one can show such a $T_1$ exists. Start with $(T,E_T)$, and let $D'=T \cup \{x_1,x_2,x_3,\ldots \}$. We want to define an edge relation on $D'$ so that it extends $E_T$, so that it satisfies the digraph extension property (so by \autoref{genericdigraph} we get the generic digraph), and so that $T$ is a maximal tournament in $D'$. The trickiest condition is ensuring the digraph extension property is satisfied: to deal with this, you enumerate all the pairwise disjoint triples $(U,V,W) \subset D'$, and then you define edge relations so that $x_i$ witnesses the extension property for the $i$th triple. Any edges which are not determined by this process are chosen to be non-edges - this ensures $T$ is a maximal tournament in $D'$. By composing with an element of Aut$(D,E)$ if necessary, we can assume that $g(\bar{b}) \in T_1$. Hence, and because elements of Aut$(\Gamma)$ map maximal tournaments to maximal tournaments, $g(T_1)=T_1$. Now, let $\theta: T \to T_1$ witness the fact that $T_1$ is isomorphic to the random tournament. Now let $f=\theta^{-1}g \theta$. It is easy to see that $f$ satisfies the requirements of the claim. (i) ``$\Rightarrow$''. We prove the contrapositive, so suppose $T(G)$ does not equal Aut$(T,E_T)$. Then there exists $f \in T(G)$ which swapS the direction of some edge in $T$. By definition of $T(G)$, that means there is $g \in G$ which swaps the direction of some edge in $D$, which implies that $G \neq$ Aut$(D,E)$. ``$\Leftarrow$''. Suppose $G \neq$ Aut$(D,E)$. Hence, there exists $g \in G$ and an edge $b_1b_2 \in D$ such that $g$ switches the direction of that edge. Hence, by Claim 2, there exists $f \in T(G)$ which switches the direction of an edge, which implies that $T(G) \neq$ Aut$(T,E_T)$. (ii) ``$\Rightarrow$''. By Claim 1, we have five options for $T(\langle sw \rangle)$. By (i), it cannot be Aut$(T,E_T)$. Suppose $T(G)$ contains $\langle -_T \rangle$. Then there exists $f \in T(G)$ and a triangle in $T$ such that $f$ swaps the direction of all three edges of the triangle. This implies that there is $g \in G$ which swaps the direction of all three edges of a triangle in $D$. But no such function exists in $\langle sw \rangle$, so if $T(G) \geq \langle -_T \rangle$, then $G \neq \langle sw \rangle$. Hence, we have that $T(\langle sw \rangle) = \langle sw_T \rangle.$ ``$\Leftarrow$''. Suppose $T(G)=\langle sw_T \rangle$. By Claim 2, if $G$ does not preserve $P_{sw}$, then this can be witnessed in $T(G)$ also. Since $\langle sw_T \rangle$ does preserve $P_{sw}$, we get that $G$ preserves $P_{sw}$. By \autoref{understandingsw}, we get that $G=$ Aut$(D,E)$ or $\langle sw \rangle$. But it cannot be the former option, so $G=\langle sw \rangle$. (iii) Same arguments as for part (ii). (iv) ``$\Rightarrow$''. This is proved similarly to previous cases. ``$\Leftarrow$''. Suppose $T(G)=\langle sw_T,-_T \rangle$. By Claim 2, we get that $G$ preserves $P_{sw,w}$, which implies that $G \leq \langle sw,- \rangle$. In $\langle sw_T,-_T \rangle$, there is a function that does not preserve $sw$. Hence, there is a function $g \in G$ which does not preserve $sw$. Hence, by \autoref{understandingjoinwith-}, $g=- \circ g'$ where $g' \in sw$. Then $g^2$ will be in $\langle sw \rangle \backslash$Aut$(D,E)$. Hence, by \autoref{understandingsw}, $G \geq \langle sw \rangle$. By composing $g$ with an appropriate element of $\langle sw \rangle$, we get that $- \in G$. Hence, we have that $G \geq \langle sw,- \rangle$. Thus, $G=\langle sw,- \rangle$, as required. \end{proof} This lemma almost completes the proof of \autoref{region1}. What is left to prove is that if $\langle -,sw \rangle < G \leq$ Aut$(\Gamma)$, then $G=$ Aut$(\Gamma)$. We believe that this can be proved directly (without the need of canonical functions), but the combinatorics involved were just out of our reach. \subsection{Canonical functions} \begin{Def} Let $\mathcal{M}, \mathcal{N}$ be any structures. Let $f: M \to N$ be any function between the domains of the structures. \begin{enumerate}[(i)] \item The \emph{behaviour} of $f$ is the relation $\{(p,q) \in S(\mathcal{M})\times S(\mathcal{N}): \exists \bar{a} \in M, \bar{b} \in N$ such that tp$(\bar{a})=p$, tp$(\bar{b})=q$ and $f(\bar{a})=\bar{b} \}$. \item If the behaviour of $f$ is a function $S(M) \to S(N)$, then we say $f$ is \emph{canonical}. Rephrased, we say $f$ is canonical if for all $\bar{a}, \bar{a}' \in M$, tp$(\bar{a})=$ tp$(\bar{a}') \Rightarrow$ tp$(f(\bar{a}))=$ tp$(f(\bar{a}'))$. \item If $f$ is canonical, we use the same symbol $f$ to denote its behaviour. \end{enumerate} \end{Def} \textbf{Examples.} \begin{enumerate} \item Any $f \in$ Aut$(D,E)$ is a canonical function, and for all types $p$, $f(p)=p$. \item $-$ is canonical. \item $sw_a$ is not canonical: Let $b,b'$ be vertices such that we have $E(a,b)$ and $E(b',b)$. Then, tp$(a,b)=$ tp$(b',b)$, but tp$(sw(a,b)) \neq$ tp$(sw(b',b))$. Similarly, $rot_a$ is not canonical. \item $sw_a$ and $rot_a$ \emph{are} canonical when we regard them as functions from $(D,E,a) \to (D,E)$. \item Let $f,g$ be canonical functions. Then $f$ behaves like $g$ (in the sense of \autoref{behaves}) if and only if $f$ and $g$ have the same behaviour (in the sense of the definition above). Note that this is not necessarily true if the functions are not canonical. \end{enumerate} The benefit of canonical functions is that they are particularly well-behaved and can be easily manipulated and analysed. The next theorem will be treated as a `black-box' for this article - a proof can be found in \cite{bpt13}. In order to state the theorem, we need to give a couple of definitions. \begin{Def} Let $F \subseteq D^D$. We let $\tmcl(F)$\footnote{where $\tmcl$ stands for `topological monoid closure'} denote the smallest closed monoid in $M$ containing $F$. We may abuse notation and write $\tmcl(F)$ for $\tmcl($Aut$(D,E) \cup F)$. \end{Def} \begin{Def} We let $(D,E,<)$ denote the countable (linearly) ordered homogeneous digraph that embeds all finite ordered digraphs. \end{Def} The theorem that follows is an application of the theorem in \cite{bpt13} to the structure $(D,E,<)$. In order for this to be valid, we need to know that $(D,E,<)$ is a Ramsey structure. The definition of a Ramsey structure can be found in \cite{bpt13}. The fact that $(D,E,<)$ is Ramsey follows from the main theorem of \cite{nr77}. \begin{thm}\label{blackbox} Let $f \in$ Sym$(D)$ and $c_1,\ldots, c_n \in D$ be any elements. Then there exists a function $g:D \to D$ such that \begin{enumerate}[(i)] \item $g \in \tmcl($Aut$(D,E) \cup \{f\})$. \item $g(c_i)=f(c_i)$ for $i=1,\ldots n$. \item When regarded as a function from $(D,E,<,c_1,...c_n)$ to $(D,E)$, $g$ is a canonical function. \end{enumerate} \end{thm} How is this theorem used? We illustrate by sketching how we will complete the proof of \autoref{region2}: $G$ is a closed group such that $\langle -,sw \rangle < G \leq$ Aut$(\Gamma)$. Thus, $G$ does not preserve $P_{sw,w}$; we let $f \in G$ and $c_1,\ldots, c_6 \in D$ witness this fact. We now use \autoref{blackbox} to obtain the canonical $g$ as in the theorem. We then examine the possibilities for $g$'s behaviour, which boils down to some finite combinatorics. Using \autoref{containsgamma} we show that in all the possible behaviours, $G$ must contain Aut$(\Gamma)$. Implicit in this argument is the fact that we care only about the behaviour of the canonical function. Though this is not immediate, it will certainly become clear as we work with these functions. Intuitively, the idea is that two different canonical functions $f,f'$ with the same behaviour provide the same information about $G$. This means that when we analyse the canonical functions, it suffices to analyse the possible behaviours of canonical functions. This task in turn is greatly simplified by the following: \textbf{Important Observation.} The behaviour of a canonical function $f: (D,E,<,c_1,\ldots, c_n) \to (D,E)$ is determined by the restriction of the behaviour to 2-types. This follows from two facts. The first is that $(D,E,<,c_1,\ldots,c_n)$ has quantifier elimination (see \cite{hod97}). The second is that the arity of the named relations is $\leq 2$. These two facts imply that the type of an $n$-tuple $(a_1,a_2,\ldots,a_n)$ is determined by the set of 2-types $\{$tp$(a_i,a_j): 1\leq i < j \leq n\}$; the observation follows easily from this. \subsubsection{Canonical functions from $(D,E,<)$} We start our analysis with the simplest situation, which is when no constants are added. As per the discussion above, it suffices to analyse the possible behaviours restricted to 2-types. To do this, we first need to describe what the possible 2-types of $(D,E,<)$ and $(D,E)$ are. \textbf{Notation.} Let $\phi_1(x,y),\ldots,\phi_n(x,y)$ be formulas. We let $p_{\phi_1,\ldots,\phi_n}(x,y)$ denote the (partial) type determined by the formula $\phi_1(x,y) \wedge \ldots \wedge \phi_n(x,y)$. For example, let $a,b \in (D,E,<)$ be such that $a<b$ and $E(a,b)$. Then $p_{<,E}(x,y)=$ tp$(a,b)$. We will often omit the free variables $x$ and $y$ and write, for example, $p_{<,E}$. With this notation in place, it is easy to state what the 2-types of $(D,E,<)$ and $(D,E)$ are. \begin{itemize} \item There are three 2-types in $(D,E)$: $p_{E},p_{E^*}$ and $p_{N}$. \item There are six 2-types in $(D,E,<)$: $p_{<,E},p_{<,E^*}, p_{<,N}$, $p_{>,E},p_{>,E^*}$ and $p_{>,N}$. \end{itemize} Now, what are the possible behaviours? For each 2-type in $(D,E,<)$, we must choose which 2-type in $(D,E)$ it gets mapped to. This choice is not free: the image of a type $p(x,y)$, say, determines the image of the corresponding type $p^*(x,y) \defeq p(y,x)$. This is the only restriction - it is easy to show that all functions $\{p_{<,E},p_{<,E^*}, p_{<,N}\} \to \{p_E,p_{E^*},p_N\}$ can be realised as the behaviour of some canonical function $f:(D,E,<) \to (D,E)$. (You use the universality of $(D,E)$. Also, remember that we do not require $f$ to be bijective.) The next lemma contains the analysis of these behaviours. \begin{lem} \label{noconstants} Let $G$ be a closed supergroup of Aut$(D,E)$ and let $f \in \tmcl(G)$ be a canonical function from $(D,E,<)$ to $(D,E)$. Then (at least) one of the following is true: \begin{itemize} \item $f$ behaves like $id$. \item $f$ behaves like $-$. \item $G$ contains Aut($\Gamma$). \end{itemize} \end{lem} \begin{proof} We split up the task according to the behaviour of $f$. For some of the cases, we use the following claim: \textbf{Claim.~} When we consider $f$ as a function $(D,E,<) \to (D,E,<)$, we may assume that $f$ preserves the linear order. \emph{Proof of Claim}. Let $f':D \to D$ be a function with the same behaviour as $f$ and which in addition preserves the linear order; we need to show that $f' \in \tmcl(G)$. Let $\bar{a} \in D$. By definition of $f'$, $f(\bar{a}) \cong f'(\bar{a})$ as unordered digraphs. By homogeneity of $(D,E)$, we can find $h \in$ Aut$(D,E)$ such that $h(f(\bar{a}))= f'(\bar{a})$. Since $f \in \tmcl(G)$, there is $g \in G$ such that $g(\bar{a})=f(\bar{a})$. So we have $hg \in G$ and $hg(\bar{a})=f'(\bar{a})$, as required. \underline{Case 1}. $f(p_{<,N})=p_N$. Case 1a. $f(p_{<,E})=p_{E}$ and $f(p_{<,E^*})=p_{E^*}$, in which case $f$ behaves like $id$. Case 1b. $f(p_{<,E})=p_{E^*}$ and $f(p_{<,E^*})=p_{E}$, in which case $f$ behaves like $-$. Case 1c. $f(p_{<,E})=p_{E}$ and $f(p_{<,E^*})=p_{E}$. We will use \autoref{containsgamma} (ii) to show that $G$ contains Aut$(\Gamma)$, so let $\bar{a} \in D$. Then there is $\bar{b} \in D$ such that $\bar{a} \cong \bar{b}$ \emph{as digraphs}, and $b_i < b_j$ for all $i<j$. Observe that for all edges $b_ib_j$ in $\bar{b}$, we have that $E(f(b_i),f(b_j)) \leftrightarrow i<j$. By homogeneity of $(D,E)$, there is $g_1 \in G$ such that $g_1(\bar{a})=\bar{b}$. Since $f \in \tmcl(G)$, there is $g_2 \in G$ such that $g_2(\bar{b})=f(\bar{b})$. But now $g=g_2g_1$ satisfies the assumptions of \autoref{containsgamma}, so we conclude that $G$ contains Aut$(\Gamma)$. The case where $f(p_{<,E})=p_{E^*}$ and $f(p_{<,E^*})=p_{E^*}$ is symmetric to this case. Case 1d. $f(p_{<,E})=p_{N}$ or $f(p_{<,E^*})=p_{N}$. Without loss suppose the first is true, the latter case is symmetric. We will use \autoref{equalsSymD} (iii) and show that $G=$ Sym$(D)$, so in particular $G$ contains Aut$(\Gamma)$. Let $\bar{a} \in D$ contain an edge $a_ia_j$. By homogeneity of $(D,E)$ there is $g_1 \in G$ such that $g_1(\bar{a}) \cong \bar{a}$ and $g_1(a_i)<g_1(a_j)$. Now let $g_2 \in G$ equal $f$ on $g_1(\bar{a})$. Observe that $g_2$ deletes the edge $g_1(a_i)g_1(a_j)$ (and possibly others too) and we also know that $g_2$ preserves edges. Hence, $g_2g_1(\bar{a})$ contains less edges than $\bar{a}$, so by \autoref{equalsSymD}, we are done. \underline{Case 2} $f(p_{<,N})=p_E$. Case 2a. Neither $f(p_{<,E})=p_{N}$ nor $f(p_{<,E^*})=p_{N}$. We will use \autoref{equalsSymD} (iii) and show that $G=$ Sym$(D)$. Let $\bar{a} \in D$ contain an edge $a_ia_j$. By homogeneity of $(D,E)$, map $\bar{a}$ to an isomorphic (as digraphs) tuple $\bar{b}$ where $b_i$, resp. $b_j$, is the least, resp. second least, element of $\bar{b}$. By assumption, $f$ maps $\{b_i,b_j\}$ to an edge but we do not know its direction. This splits into two cases. Subcase (i). Suppose we have $E(f(b_i),f(b_j))$. Now let $\bar{b}'$ be an ordered digraph which is the same as $\bar{b}$ except that $b_i,b_j$ is changed to a non-edge. Observe that $f(\bar{b}') \cong f(\bar{b})$, because we are in the case where $f(p_{<,N})=p_E$. But now we are done: we can find mappings in $G$ to get from $\bar{a}$ to $\bar{b}$ to $f(\bar{b})$ to $f(\bar{b}')$ to $\bar{b}'$ (noting that though $f$ may not be invertible, the function in $G$ which agrees with $f$ on $\bar{b}'$ is invertible), and $\bar{b}'$ has less edges than in $\bar{a}$. Subcase (ii). Suppose we have $E^*(f(b_i,b_j))$. The previous argument does not work as stated, because the edges $f(b_i,b_j)$ and $f(b_i',b_j')$ will not be in the same direction. To fix this, we modify $\bar{b'}$ by swapping $b_i'$ and $b_j'$ with respect to the linear order. Now, $f(b_i,b_j)$ and $f(b_i',b_j')$ will be in the same direction. Furthermore, because we earlier specified that $b_i$ and $b_j$ should be the two least elements of $\bar{b}$, this swapping only affects the type of the pair $b_ib_j$ - all other pairs' types are unaffected. This ensures that we have $f(\bar{b}') \cong f(\bar{b})$. The rest of the proof continues as in the previous case. Case 2b. $f(p_{<,E})=p_{N}$ and $f(p_{<,E^*})=p_{N}$. By using the claim, it is easy to see that $f^2$ is a canonical function in $\tmcl(G)$ where $f^2(p_{<,N})=p_N$, $f(p_{<,E})=p_{E}$ and $f(p_{<,E^*})=p_{E}$. Hence, by Case 1c, $G$ contains Aut$(\Gamma)$. Case 2c. $f(p_{<,E})=p_{N}$ and $f(p_{<,E^*})=p_{E}$. By considering $f^2$, this case is reduced to Case 1d, so $G$ contains Aut$(\Gamma)$. Case 2d. $f(p_{<,E})=p_{N}$ and $f(p_{<,E^*})=p_{E^*}$. We will use \autoref{equalsSymD} (iii). Let $\bar{a} \in D$ contain an edge $E(a_i,a_j)$. By composing with an element of Aut$(D,E)$ if necessary, we may assume that $a_j$ is the least element, and $a_i$ is the second least element. In particular, we have tp$(a_j,a_i)=p_{<,E^*}$. Let $\bar{b}$ be an ordered digraph such that $f(\bar{b})=\bar{a}$. Now let $\bar{b}'$ be an ordered digraph which is the same as $\bar{b}$ except we swap the position of $b_i'$ and $b_j'$ in the linear order. Now, tp($b_j',b_i')=p_{<,E}$. By choosing $b_i,b_j$ to be the least elements, the types of all the other pairs are unaffected, so $f(\bar{b}')$ is the same digraph as $\bar{a}$ but the edge $a_ia_j$ is replaced by a non-edge. Hence, we are done. Case 2e. $f(p_{<,E})=p_{E}$ and $f(p_{<,E^*})=p_{N}$. Considering $f^2$ reduces us to Case 2a. Case 2f. $f(p_{<,E})=p_{E^*}$ and $f(p_{<,E^*})=p_{N}$. Imitate the argument in Case 2d to show that $G=$ Sym$(D)$. \underline{Case 3} $f(p_{<,N})=p_{E^*}$. This is symmetric to Case 2. \end{proof} \subsubsection{Canonical functions from $(D,E,<,\bar{c})$} We now move on to the general situation where we have added constants $\bar{c} \in D$ to the structure. For convenience, we may as well assume that $c_i<c_j$ for all $i<j$. As is the case for $(D,E)$ (see \autoref{genericdigraph}), the $n$-types of $(D,E,<,\bar{c})$ correspond to the orbits of Aut$(D,E,<,\bar{c})$ acting on the set of $n$-tuples of $D$. As a result we use the concepts of types and orbits interchangeably, and we often abuse notation to provide for a smoother presentation. Unlike the situation with no constants, this structure is not 1-transitive, i.e., we have more than one orbit. There are two kinds of orbits. The first is a singleton containing one of the constants, e.g. $\{c_1\}$ is an orbit. The second kind consists of infinite orbits, which are necessarily isomorphic to the generic digraph. An infinite orbit is determined by how its elements are related to the $c_i$, e.g., one of the orbits will be $\{x \in D: x<c_1 \wedge \bigwedge_i E(x,c_i)\}$. In order to describe the 2-types, we extend notation from the previous section. \textbf{Notation} Let $A,B$ be definable subsets of $D$ and let $\phi_1(x,y),\ldots,\phi_n(x,y)$ be formulas. We let $p_{A,B,\phi_1,\ldots,\phi_n}(x,y)$ denote the (partial) type determined by the formula $x \in A \wedge y \in B \wedge \phi_1(x,y) \wedge \ldots \wedge \phi_n(x,y)$. Now let $X$ and $Y$ be orbits, $\phi \in \{<,>\}$ and $\psi \in \{E,E^*,N\}$. Then all the 2-types are of the form $p_{X,Y,\phi, \psi}= \{(a,b) \in D: a \in X, b \in Y, \phi(a,b)$ and $\psi(a,b)\}$.\footnote{This is an example of how we are abusing notation and blurring the distinction between types and orbits.} Our task now is to analyse the possibilities for $f(p_{X,Y,\phi,\psi})$, where $f$ is a canonical function. The analysis is split into cases depending on how the orbits $X$ and $Y$ relate. The first lemma deals with the situation when $X=Y$. \begin{lem} \label{oneorbit} Let $G$ be a closed supergroup of Aut$(D,E)$, let $f \in \tmcl(G)$ be a canonical function from $(D,E,<,\bar{c})$ and let $X$ be an infinite orbit of Aut$(D,E,\bar{c})$. Then (at least) one of the following holds: \begin{itemize} \item $f$ behaves like $id$ on $X$. \item $f$ behaves like $-$ on $X$. \item $G$ contains Aut$(\Gamma)$. \end{itemize} \end{lem} \begin{proof} By noting that $(X,E|_X)$ is isomorphic to $(D,E)$, unravelling the definitions will show that this lemma has exactly the same mathematical content as \autoref{noconstants}. \end{proof} Next, we look at how $f$ can behave between two infinite orbits. To do this analysis, we need to look at how two infinite orbits can relate to each other with respect to the linear order: \textbf{Facts and Notation} There are two ways that two infinite orbits $X$ and $Y$ of Aut($D,E,<,\bar{c})$ can relate to each other with respect to the linear order $<$: \begin{itemize} \item All of the elements of one orbit, $X$ say, are smaller than all of the elements of $Y$. This is abbreviated by $X<Y$ \item $X$ and $Y$ are interdense: $\forall x<x' \in X, \exists y \in Y$ such that $x<y<x'$ and vice versa. \end{itemize} We deal with these two possibilities separately, starting with the case where one orbit is below the other. \begin{lem} \label{twoorbits} Let $G$ be a closed supergroup of Aut$(D,E)$, let $f \in \tmcl(G)$ be a canonical function from $(D,E,<,\bar{c})$ and let $X$ and $Y$ be infinite orbits of Aut$(D,E,\bar{c})$ such that $X<Y$. Then (at least) one of the following holds: \begin{itemize} \item $f$ behaves like $id,sw,rot$ or $rot^{-1}$ between $X$ and $Y$. \item $f$ behaves like $- \circ rot_{(X,Y)}$ on $X \cup Y$. \item $G$ contains Aut($\Gamma$). \end{itemize} \end{lem} \begin{proof} Let $x_0 \in X$ be fixed. We emphasise now an important feature of this proof, which is that our arguments only depend on how $f$ behaves on $\{x_0\} \cup Y$. This is done intentionally so that these arguments can be used unaltered in later lemmas. By \autoref{oneorbit}, we may assume that $f$ behaves like $id$ or $-$ on $X$ and $Y$. As mentioned above, the arguments only concern one point $x_0 \in X$, so it does not matter how $f$ behaves on $X$. However, whether $f$ behaves like $id$ or $-$ on $Y$ can make a difference. Fortunately, for most cases the arguments require very little, if any, adjustment, so we assume $f$ acts like $id$ on $Y$. When required, we will explain how to modify the argument if $f$ acts like $-$ on $Y$. Furthermore, as the arguments are similar to that of \autoref{oneorbit}, the proofs are more sketchy, and we leave the details to the reader. \underline{Case 1} $f(p_{X,Y,N})=p_{N}$.\footnote{Note that because $X<Y$, $p_{X,Y,\psi}=p_{X,Y,<,\psi}$ for any formula $\psi$} Case 1a. $f(p_{X,Y,E})=p_{E}$ and $f(p_{X,Y,E^*})=p_{E^*}$. Then $f$ behaves like $id$ between $X$ and $Y$. Case 1b. $f(p_{X,Y,E})=p_{E^*}$ and $f(p_{X,Y,E^*})=p_{E}$. Then $f$ behaves like $sw$ between $X$ and $Y$. Case 1c. $f(p_{X,Y,E})=p_{E}$ and $f(p_{X,Y,E^*})=p_{E}$. We will use \autoref{containsgamma} (ii) to show that $G$ contains $\Gamma$. Let $\bar{a}=(a_1,\ldots,a_n) \in D$. We want to show that by using elements of $G$, we can switch the direction of the edges of $\bar{a}$ so they are all pointing in the same direction. We do this by induction on $n$. The base case $n=1$ is trivial so let $n>1$. By the inductive hypothesis, we can assume that for $2\leq i,j \leq n$, if $a_ia_j$ is an edge, then $E(i,j) \leftrightarrow i<j$. By homogeneity, map $a_1$ to $x_0$ and the other $a_i$'s into $Y$. Then applying $f$ switches the edges adjacent to $a_1$ so they are all directed out of $a_1$ and furthermore $f$ does not alter any of the other edges. Thus, the resulting digraph has all edges going in the same direction, as required. In the case where $f$ behaves like $-$ on $Y$, after applying the induction hypothesis, you first map all of $\bar{a}$ into $Y$, apply $f$ (so switch \emph{all} the edges' directions), map $a_1$ to $x_0$, and apply $f$ again (so we `unswitch' all the edges in $\{a_2,\ldots,a_n\}$). Case 1d. $f(p_{X,Y,E})=p_{N}$ or $f(p_{X,Y,E^*})=p_{N}$. Given any $\bar{a} \in D$ which contains edges, we use $f$ to delete edges from it, so by \autoref{equalsSymD} (iii), $G=$ Sym$(D)$. See Case 1d of \autoref{oneorbit} for more detail. \underline{Case 2} $f(p_{X,Y,N})=p_{E}$. Case 2a. Neither $f(p_{X,Y,E})=p_{N}$ nor $f(p_{X,Y,E^*})=p_{N}$. $G=$ Sym$(D)$, by using the same argument as in Case 2a of \autoref{oneorbit}. Note that the argument would be unaffected if $f$ behaves like $-$ on $X$. Case 2b. $f(p_{X,Y,E})=p_{N}$ and $f(p_{X,Y,E^*})=p_{N}$. Given any $\bar{a} \in D$ containing edges, we apply $f$ twice in order to delete edges. Thus, $G=$ Sym$(D)$ by part (iii) of \autoref{equalsSymD}. Case 2c. $f(p_{X,Y,E})=p_{N}$ and $f(p_{X,Y,E^*})=p_{E}$. Same as Case 2b. Case 2d. $f(p_{X,Y,E})=p_{N}$ and $f(p_{X,Y,E^*})=p_{E^*}$. Subcase (i) $f$ acts like $id$ on $Y$. We show that $G=$ Sym$(D)$. The idea is the same to that of Case 2d in \autoref{oneorbit}. Let $\bar{b} \in D$ contain an edge $E(b_i,b_j)$. Let $\bar{b}_1$ be obtained by mapping $b_i$ to $x_0$, applying $f$, then mapping $b_j$ to $x_0$ and applying $f$ again. Note that we have $E^*(b_{1,i},b_{1,j})$. Let $\bar{b}_2$ be obtained from $\bar{b}$ in the same way, except we map $b_j$ to $x_0$ first, and then $b_i$ second. In this case, we have $N(b_{2,i},b_{2,j})$. Furthermore, $\bar{b}_1$ and $\bar{b}_2$ are otherwise the same. Now suppose $\bar{a}$ is given. Find a $\bar{b}$ such that its corresponding $\bar{b}_1$ is isomorphic to $\bar{a}$. Then, we can get from $\bar{a}$ to $\bar{b_1}$ to $\bar{b_2}$, i.e., we can delete an edge from $\bar{a}$. Thus, $G=$ Sym$(D)$. Subcase (ii) $f$ acts like $-$ on $Y$. Now, if $f$ behaves like $id$ on $X$, the subcase above shows that $G=$ Sym$(D)$. Hence, we are left with case where $f$ behaves like $id$ on both $X$ and $Y$. But then $f$ behaves like $- \circ rot_{(X,Y)}$ on $X \cup Y$. Case 2e. $f(p_{X,Y,E})=p_{E}$ and $f(p_{X,Y,E^*})=p_{N}$. Considering $f^2$ reduces us to Case 2a. Case 2f. $f(p_{X,Y,E})=p_{E^*}$ and $f(p_{X,Y,E^*})=p_{E}$. Then $f$ behaves like $rot$ between $X$ and $Y$. Note that if $f$ behaves like $-$ on $Y$ in this case, you can show that $G=$ Sym$(D)$ by considering $- \circ f \in \tmcl(G)$. \end{proof} As mentioned at the start of the proof, what was relevant is how $f$ behaved on $\{x_0\} \cup Y$. More specifically, what was sufficient to make these arguments work was the following: For all finite digraphs $\bar{a} \in D$ and all points $a \in \bar{a}$, we can find a copy of $\bar{a}$ in $\{x_0\} \cup Y$ such that $\bar{a} \cap \{x_0\} = \{a\}$. This condition is satisfied in the remaining situations that need to be analysed, so their corresponding results are immediate corollaries of \autoref{twoorbits}. The statement for interdense orbits has to be modified and will perhaps appear confusing. Clarification will be provided after the statement. \begin{cor} Let $G$ be a closed supergroup of Aut$(D,E)$, let $f \in \tmcl(G)$ be a canonical function from $(D,E,<,\bar{c})$ and let $X$ and $Y$ be interdense infinite orbits of Aut$(D,E,\bar{c})$. Further suppose that $G$ does not contain Aut$(\Gamma)$. Then both of the following hold: \begin{itemize} \item $f$ behaves like $id,sw,rot$ or $rot^{-1}$ between increasing tuples from $X$ to $Y$, or, $f$ behaves like $- \circ rot$ on the set of increasing tuples from $X$ to $Y$. \item $f$ behaves like $id,sw,rot$ or $rot^{-1}$ between decreasing tuples from $X$ to $Y$, or, $f$ behaves like $- \circ rot$ on the set of decreasing tuples from $X$ to $Y$. \end{itemize} \end{cor} For example, $f$ behaves like $sw$ between increasing tuples from $X$ to $Y$ means that $f(p_{X,Y,<,E})=p_{E^*}, (p_{X,Y,<,E^*})=p_{E}$ and $(p_{X,Y,<,N})=p_{N}$. Another example is that when we say $f$ behaves like $- \circ rot$ on the set of decreasing tuples from $X$ to $Y$, we mean that $f$ behaves like $-$ on $X$ and $Y$, $f(p_{X,Y,>,E})=p_{N}, (p_{X,Y,>,E^*})=p_{E^*}$ and $(p_{X,Y,>,N})=p_{E}$. The final situation is to look at how $f$ can behave between one of the constants, $c$ say, and the infinite orbits. \begin{Def}: Let $c$ be one of the named constants of $(D,E,<,\bar{c})$ and let $X_1,X_2$ and $X_3$ be infinite orbits. If it is the case that we have outward edges from $c$ to $X_1$, inward edges from $c$ to $X_2$ and non-edges between $c$ and $X_3$, we called the triple $\bar{X}=(X_1,X_2,X_3)$ a $c$-generic triple. \end{Def} The reason for introducing this definition is that there is nothing to be analysed about how $f$ behaves between $c$ and a \emph{single} orbit $X$. It is only useful to ask how $f$ behaves between $c$ and several infinite orbits, in particular, a $c$-generic triple. Note that if $\bar{X}$ is a $c$-generic triple, then $c \cup X_1 \cup X_2 \cup X_3$ is isomorphic to the generic digraph. \begin{cor} Let $G$ be a closed supergroup of Aut$(D,E)$, let $f \in \tmcl(G)$ be a canonical function from $(D,E,<,\bar{c})$, let $c$ be one of the named constants and let $\bar{X}$ be a $c$-generic triple. Then (at least) one of the following holds: \begin{itemize} \item $f$ behaves like $id,sw,rot$ or $rot^{-1}$ between $c$ and $\bigcup \bar{X}$. \item $f$ behaves like $-\circ rot_{(c,\bar{X})}$ on $\{c\} \cup \bigcup \bar{X}$. \item $G$ contains Aut($\Gamma$). \end{itemize} \end{cor} \subsection{Using canonical functions} With this analysis, we are now in a position to prove the remaining lemmas. \begin{proof}[Proof of \autoref{threeregions}] We recall that we want to show that if $G$ is a reduct of $(D,E)$, then $G$ either contains Aut$(\Gamma)$, is contained in Aut$(\Gamma)$, or contains $\langle rot \rangle$. Suppose none of these are true - we will derive a contradiction. $G$ is not contained in Aut$(\Gamma)$, which means that $G$ does not preserve non-edges. Hence, there is $f \in G$ and an edge $c_1c_2 \in D$ such that $f(c_1c_2)$ is a non-edge. We apply \autoref{blackbox} to obtain a canonical $g: (D,E,<,c_1,c_2) \to (D,E)$ which agrees with $f$ on $c_1$ and $c_2$. By \autoref{oneorbit}, for any infinite orbit $X$, $g$ behaves like $id$ or $-$ on it - otherwise, $G$ would contain Aut($\Gamma$), contradicting our assumptions. By \autoref{twoorbits} and its corollaries, we have that $g$ behaves like $id$ or $sw$ between orbits - otherwise, $G$ would contain either Aut($\Gamma$) or $\langle rot \rangle$, contradicting our assumptions. But now we have a function $g \in \tmcl(G)$ which deletes an edge (namely,$c_1c_2$) and maps all non-edges to non-edges. By imitating the proof of part (iv) of \autoref{equalsSymD}, we conclude that $G$ equals Sym$(D)$, contradicting that $G$ does not contain Aut($\Gamma$). \end{proof} \vspace{5mm} \begin{proof}[Proof of \autoref{region2}] By \autoref{region2a} and \autoref{region2b}, it remains to be proved that if $\langle sw,- \rangle < G \leq$ Aut($\Gamma)$ is a closed group, then $G=$ Aut$(\Gamma)$. So let $G$ be such a closed group and suppose, for contradiction, that $G \neq$ Aut$(\Gamma)$. Since $\langle sw,- \rangle < G$, $G$ does not preserve $P_{sw,w}$, there exists $f \in G$ and $\bar{c} \in D$ such that $P_{sw,w}(\bar{c})$ and $\neg P_{sw,w}(f(\bar{c}))$. Now use \autoref{blackbox} to obtain $g \in \tmcl(G)$ which is canonical from $(D,E,<,\bar{c})$ and which agrees with $f$ on $\bar{c}$. As in the previous proof, we use $\autoref{oneorbit}$ to conclude that for any infinite orbit $X$, $g$ behaves like $id$ or $-$ on $X$, and we use $\autoref{twoorbits}$ and its corollaries to conclude that $g$ behaves like $id$ or $sw$ between orbits. \textbf{Claim 1.} $g$ behaves like $id$ on all infinite orbits, or, $g$ behaves like $-$ on all infinite orbits. \emph{Proof of Claim 1.} Suppose not, so there exists infinite orbits $X$ and $Y$ such that $g$ behaves like $id$ on $X$ and like $-$ on $Y$. There are now two cases. The first case is if $g$ behaves like $id$ between $X$ and $Y$. In this case, by imitating the proof of part (iii) of \autoref{containsgamma}, we conclude that $G \geq$ Aut$(\Gamma)$ - contradiction. The second case is if $g$ behaves like $sw$ between $X$ and $Y$. This reduces to the previous case by considering $- \circ g$: $-\circ g$ behaves like $-$ on $X$, like $id$ on $Y$ and like $id$ between $X$ and $Y$. Thus, we always reach a contradiction, proving the claim. In fact, by considering $- \circ g$ if necessary, we may now assume that $g$ behaves like $id$ on all infinite orbits. Enumerate the (finite number of) infinite orbits as $X_1,X_2,X_3,\ldots$. \textbf{Claim 2.} We may assume $g$ behaves like $id$ between all pairs of orbits from $X_1,X_2,X_3$. \emph{Proof of Claim 2.} If $g$ behaves like $id$ between all three orbits, we are done. If $g$ behaves like $sw$ between precisely two of the pairs - without loss $g$ behaves like $sw$ between $X_1$ and $X_2$, and between $X_1$ and $X_3$ - then by switching about $X_1$ we may assume $g$ behaves like $id$, so again we are done. If $g$ behaves like $sw$ between precisely one pair of infinite orbits, then by imitating the proof of \autoref{containsgamma} (iii), we get that $G \geq$ Aut$(\Gamma)$, which is a contradiction. The final possibility is that $g$ behaves like $sw$ between all pairs: this reduces to the third case by switching about $X_1$, so we again get a contradiction. Thus, we have proved the claim. \textbf{Claim 3.} We may assume $g$ behaves like $id$ between all infinite orbits. \emph{Proof of Claim 3.} First consider how $g$ behaves between $X_4$ and the first three $X_i$'s. If $g$ behaves like $id$ between $X_4$ and all the previous $X_i$'s, we move on. If $g$ behaves like $sw$ between $X_4$ and all the previous $X_i$'s, we switch about $X_4$, reducing to the first case. The last case is if, without loss, $g$ behaves like $id$ between $X_4$ and $X_1$ and like $sw$ between $X_4$ and $X_2$. But this is exactly the same as the contradictory case in the proof Claim 2, so this is not possible. Hence, we have shown that $g$ must behave like $id$ between all the pairs in $X_1,\ldots,X_4$. One then moves on to $X_5$ and repeats this argument to show that we may assume $g$ behaves like $id$ between $X_1,\ldots X_5$. Continuing in this fashion proves the claim. \textbf{Claim 4.} We may assume that $g$ behaves like $id$ between $\bar{c}$ and the infinite orbits. Consider $c_1 \in \bar{c}$, and let $\bar{X}$ and $\bar{Y}$ be $c_1$-generic. Suppose that $g$ behaves like $id$, resp. $sw$ between $c$ and $\bigcup \bar{X}$, resp. $\bigcup \bar{Y}$. Then, for any finite digraph $A$ and edge $aa' \in A$, we can map $aa'$ to an edge between $c_1$ and $\bar{Y}$ and map the remaining vertices of $A$ into $\bar{X}$. Then applying $g$ will have the effect of switching precisely the single edge $aa'$. Thus, by imitating the argument of \autoref{containsgamma} (iii), we get that $G \geq$ Aut$(\Gamma)$, a contradiction. Hence, it must be the case $g$ behaves like $id$ between $c_1$ and all the $c_1$-generic triples, or, $g$ behaves like $sw$ between $c_1$ and the triples. But in the latter case, we can apply $sw_{c_1}$ to reduce to the former case. Repeating this for the other $c_i$ will complete the proof of the claim. Observe that all the manipulations we (may have) used on $g$ have been applications of $-$ or $sw$. This ensures that $g(\bar{c} \notin P_{sw,w}$. In particular, there is at least one edge in $\bar{c}$ whose direction $g$ switches. Combining this observation with all the claims tells us that we are in the situation of \autoref{containsgamma} (iii): $g$ behaves like $id$ everywhere except on the finite set $\bar{c}$. Hence, $G \geq$ Aut$(\Gamma)$, giving us a contradiction, thus completing the proof. \end{proof} \begin{proof}[Proof of \autoref{region3}] We recall the statement of the lemma: If $G$ contains $\langle rot \rangle$, then $G$ equals $\langle rot \rangle, \langle -,rot \rangle$ or Sym$(D)$. To prove the statement, it suffices to prove the following: \begin{enumerate}[(i)] \item If $G > \langle rot \rangle$ and $G \not\geq \langle -,rot \rangle$, then $G=$ Sym$(D)$. \item If $G > \langle - ,rot \rangle$, then $G=$ Sym$(D)$. \end{enumerate} Before continuing, recall that $\langle sw,rot \rangle =$ Sym$(D)$. Hence, we may assume in all that follows that $sw \notin G$. (i) Suppose $G > \langle rot \rangle$ and $G \not\geq \langle -,rot \rangle$. The latter assumption implies that $- \notin G$. Then there exists $\bar{c}$ and $f \in G$ which witness the fact that $G$ does not preserve $P_{rot,1}$. Then use \autoref{blackbox} to obtain a canonical $g: (D,E,<,\bar{c}) \to (D,E)$ which agrees with $f$ on $\bar{c}$. By \autoref{oneorbit}, $g$ behaves like $id$ on all infinite orbits, as otherwise $G$ contains $sw$ or $-$. Similarly, by \autoref{twoorbits} and its corollaries, we have that $g$ behaves like $id$ or $rot$ between orbits. We proceed in a similar fashion to the proof of \autoref{region2}. \textbf{Claim 1.} We may assume that $g$ behaves like the $id$ between all infinite orbits. \emph{Proof of Claim 1.} Let $X_1,X_2,\ldots$ enumerate the infinite orbits. If $g$ behaves like $id$ between $X_1$ and $X_2$ then move on. Otherwise, $g$ behaves like $rot$ or $rot^{-1}$ between $X_1$ and $X_2$. Hence, by composing with a rotation about $X_1$ or $X_2$ as appropriate, we can assume $g$ behaves like $id$ between $X_1$ and $X_2$. Now consider $X_3$. Again by composing with a rotation if necessary, we may assume that $g$ behaves like $id$ between $X_1$ and $X_3$. Suppose that $g$ does not behave like $id$ between $X_2$ and $X_3$ - we will show that $G$ must equal Sym$(D)$. Without loss, $g$ behaves like $rot$ between $X_2$ and $X_3$. Given any finite digraph $A$ and an edge $aa'$ in $A$, we can find a copy of $A$ in $D$ such that $aa'$ is an inward edge from $X_2$ to $X_3$ and such that the other vertices of $A$ all lie in $X_1$. Applying $g$ to this copy results in the edge $aa'$ being deleted, with the rest of $A$ being the same. Hence, by \autoref{equalsSymD} (iii), we conclude that $G=$ Sym$(D)$. Therefore, we may assume that $g$ behaves like $id$ between $X_1,X_2$ and $X_3$. We then consider $X_4$. Using an identical argument, we can show that if $g$ does not behave like $id$ between $X_4$ and the other orbits, then $G=$ Sym$(D)$, so we may assume $g$ behaves like $id$. Continuing in this fashion proves the claim. \textbf{Claim 2.} We may assume that $g$ behaves like $id$ between $\bar{c}$ and all the infinite orbits. \emph{Proof of Claim 2.} First work with $c_1$. Let $\bar{X}$ be a $c_1$-generic triple. We know that $g$ behaves like $id$, $rot$ or $rot^{-1}$ between $c$ and $\bar{X}$, so by composing with a rotation if necessary, we may assume that $g$ behaves like $id$. Now consider another $c_1$-generic triple $\bar{Y}$. If $g$ does not behave like $id$ between $c_1$ and $\bar{Y}$, then we use the same argument as in the proof of Claim 1 to show that $G$ equals Sym$(D)$. So we may assume that $g$ behaves like $id$ between $c_1$ and all infinite orbits. Repeating this for $c_2$ and $c_3$ completes the proof of the claim. Because all the possible modifications of $g$ were compositions with rotations, we still have $\neg P_{rot,1}(g(\bar{c}))$. But that means we have $g$ acting like $id$ everywhere except on $\bar{c}$. Hence, $g$ must either switch an edge or delete an edge in $\bar{c}$, but we do not know which. In either case, using \autoref{containsgamma} or \autoref{equalsSymD} as appropriate, we get that $G \geq$ Aut$(\Gamma)$, so $G$ contains $sw$, so $G=$ Sym$(D)$. This completes the proof of (i). (ii) Suppose $G > \langle rot,- \rangle$. Then there exists $\bar{c}$ and $f \in G$ which witness the fact that $G$ does not preserve $P_{rot,w}$. Then use \autoref{blackbox} to obtain a canonical $g: (D,E,<,\bar{c}) \to (D,E)$ which agrees with $f$ on $\bar{c}$. By \autoref{oneorbit}, $g$ behaves like $id$ or $-$ on all infinite orbits, and by \autoref{twoorbits} and its corollaries, we have that $g$ behaves like $id$ or $rot$ between orbits, or like $-\circ rot$ on the union of the two orbits. \textbf{Claim 1'.} We may assume that $g$ acts like $id$ on all infinite orbits or like $-$ on all infinite orbits. \emph{Proof of Claim 1'.} Suppose not. so there are infinite orbits $X$ and $Y$ such that $g$ acts like $id$ on $X$ and $-$ on $Y$. There are two options for how $g$ behaves between $X$ and $Y$, like $id$ or like $rot$. If $g$ behaves like $id$, then we imitate the idea in \autoref{containsgamma} (iii) to show that $G \geq$ Aut$(\Gamma)$, so then $G$ must equal Sym$(D)$. If $g$ behaves like $rot$ between $X$ and $Y$, just compose with $rot_{X}$ to reduce to the first option. This completes the proof of the claim. Now composing with $-$ if necessary, we may assume that $g$ behaves like $id$ on all infinite orbits. What are the possible ways $g$ can behave between an infinite and another orbit? There are three options: like $id$, like $rot$, or, like $-\circ rot$. This last option is possible as it may have been the case that $g$ originally behaved like $rot$ between the two orbits, so after applying $-$ we get that it behaves like $-\circ rot$. But what does it mean to behave like $- \circ rot$ between two orbits? It means that you swap outward edges with non-edges, while preserving inward edges the same. This means that we would be in Case 2di of \autoref{twoorbits}, where we showed that $G$ must then equal Sym$(D)$. Hence, we may assume that $g$ acts like $id$ on all infinite orbits, and that $g$ behaves like $rot$ or $id$ between an infinite orbit and any other orbit. But this is exactly the same situation as in part (i) of this proof, so we just repeat the argument. This completes the proof. \end{proof} \section{Summary and Open Questions} We summarise the structure of the proof of the main theorem, \autoref{maintheorem}, which states that $\mathcal{L}$ is the lattice of the reducts of the generic digraph. The first task is to show that $\mathcal{L}$ is a sublattice of the reducts of the generic digraph, which was done in \autoref{sublattice}. The second task is to show that $\mathcal{L}$ contains all the reducts. By \autoref{threeregions}, which was proved using canonical functions at the start of Section 5.3, the task is split up into three regions of $\mathcal{L}$: The region above Aut$(\Gamma)$, the region below Aut$(\Gamma)$, and the rest. The region above Aut$(\Gamma)$ is immediately dealt with by Thomas' classification of the reducts of $\Gamma$. The proof of the region below Aut$(\Gamma)$, \autoref{region2}, has two parts. The first part is in Section 5.1, where we use the function $T(G)$ and the classification for the random tournament, and the second part is in Section 5.3. The final region, \autoref{region3}, is proved using canonical functions at the end of Section 5.3. We end by stating some problems of interest in this area. There is the obvious task of determining the reducts of your favourite structure(s), but some more specific questions are: \begin{itemize} \item (Thomas' Conjecture): If a structure is homogeneous in a finite relational language, then it only has finitely many reducts. \item Which lattices can be realised as the lattice of reducts of some structure? \item Is there always a maximal closed group between a closed group $G$ and Sym$(M)$ (where $M$ is countable)? \end{itemize} The answer to the first question may be related to a question in structural Ramsey theory: Given a homogeneous structure, can you finitely extend its language so that the structure becomes Ramsey? For example, it may be easier to prove the conjecture is true for Ramsey structures, and this may be sufficient to prove the full conjecture. Alternatively, a counterexample for one question may lead to a counterexample of the other. Another angle on the second question could be to consider whether there is any relationship between structures which have the same lattice of reducts. For example, it is curious that $(\mathbb{Q},<)$, the random graph, the random tournament and the generic partial order have the same 5-element lattice as their lattice of reducts. For clarification of the third question, we say that a closed group $F<$ Sym$(M)$ is maximal if there are no closed groups $F'$ such that $F < F' <$ Sym$(M)$. To find a counterexample to this question, it is sufficient to find an $\aleph_0$-categorical countable structure such that all of its non-trivial reducts have infinitely many reducts. We remark that such a structure without the condition of being $\aleph_0$-categorical is known: $(\mathbb{Z}, \{(x,y): |x-y|=1\})$. The relations definable in this structure are analysed in \cite{ss12}, and it follows that all of its non-trivial reducts have infinitely many reducts. \bibliographystyle{alpha}

https://arxiv.org/abs/q-alg/9512002

On a Universal Invariant of 3-Manifolds

We construct an invariant of 3-manifolds using a modification of the Kontsevich integral and Kirby's calculus. This invariant, as expected in perturbative Chern-Simon theory, takes values in the algebra of oriented 3-valent graphs. This algebra is a Hopf algebra, graded by half the number of vertices in 3-valent graphs. The degree 1 term of the invariant coincides with Casson-Walker-Lescop invariant. The degree $n$ term is constructed out of the universal Vassiliev invariant of links of degree less than or equal to $(l+1)n$ where $l$ is the number of link components.

\section{Modified universal Vassiliev-Kontsevich invariant} We will review a construction of the universal Vassiliev-Kontsevich invariant of framed oriented links, and the definition of modified one given in \cite{LeMurakamiII}. The modified invariant is well behaved under Kirby move II. We also show other properties of the invariant in this section. \subsection{Chord diagrams} A {\it uni-trivalent graph} is a graph every vertex of which is either univalent or trivalent. A uni-trivalent graph is {\it oriented} if at each trivalent vertex a cyclic order of edges is fixed. Let $X$ be a compact oriented 1-dimensional manifold whose components are labeled. A {\it chord diagram} with support $X$ is the manifold $X$ together with an oriented uni-trivalent graph whose univalent vertices are on $X$; and the graph does not have any connected component homeomorphic to a circle. Note that our definition of a chord diagram is more general than that of \cite{BarNatan,LeMurakami}. In Figures components of $X$ are depicted by solid lines, while the graph is depicted by dashed lines. There may be connected components of the dashed graph which do not have univalent vertices. Each chord diagram has a natural topology. Two chord diagrams $D,D'$ on $X$ are regarded as equal if there is a homeomorphism $f:D\to D'$ such that $f|_X$ is a homeomorphism of $X$ which preserves components and orientation. Let ${\cal A}(X)$ be the vector space over ${\Bbb C}$ spanned by chord diagrams with support $X$, subject to the AS, IHX and STU relations shown in Figure \ref{fig.ASIHXSTU}. The {\it degree} of a chord diagram is half the number of vertices of the dashed graph. Since the relations AS, IHX and STU respect the degree, there is a grading on ${\cal A}(X)$ induced by this degree. We denote by $\hat{\cal A}(X)$ the completion of ${\cal A}(X)$ with respect to the degree. Note that each map on ${\cal A}(X)$ defined in the following has the natural extension of it on $\hat{\cal A}(X)$, and that we use the same notation for the corresponding maps on ${\cal A}(X)$ and $\hat{\cal A}(X)$. \begin{figure}[htpb] $$ \psdraw{fig1}{3.6in} $$ \caption{The AS, IHX and STU relations}\label{fig.ASIHXSTU} \end{figure} Suppose $C$ is a component of $X$. Reversing the orientation of $C$, from $X$ we get $X'$. Let $S_{(C)}: {\cal A}(X)\to{\cal A}(X')$ be the linear mapping which transfers every chord diagram $D$ in ${\cal A}(X)$ to $S_{(C)}(D)$ obtained from $D$ by reversing the orientation of $C$ and multiplying by $(-1)^m$, where $m$ is the number of vertices of the dashed graph on the component $C$. Replacing $C$ by $2$ copies of $C$, from $X$ we get $X^{(2,C)}$, with a projection $p:X^{(2,C)}\to X$. If $x$ is a point on $C$ then $p^{-1}(x)$ consists of $2$ points, while if $x$ is a point of other components, then $p^{-1}(x)$ consists of one point. Let $D$ be a chord diagram on $X$, with the dashed graph $G$. Suppose that there are $m$ univalent vertices of $G$ on $C$. Consider all possible new chord diagrams on $X^{(2,C)}$ with the same dashed graph $G$ such that if a univalent vertex of $G$ is attached to a point $x$ on $X$ in $D$, then this vertex is attached to a point in $p^{-1}(x)$ in the new chord diagram. There are $2^m$ such chord diagrams, and their sum is denoted by $\Delta_{C}(D)$. It is easy to check that these linear mappings $S_{(C)}$ and $\Delta_{(C)}$ are well-defined on ${\cal A}(X)$, and naturally extended to the maps on $\hat{\cal A}(X)$. Suppose that $X$ and $X'$ have distinguished components $C$ and $C'$ respectively, and that $X$ consists of loop components only. Let $D\in{\cal A}(X)$ and $D'\in{\cal A}(X')$ be two chord diagrams. {}From each of $C$ and $C'$ we remove a small arc which does not contain any vertices. The remaining part of $C$ is an arc which we glue to $C'$ in the place of the removed arc such that the orientations are compatible. The new chord diagram is called the {\it connected sum of $D$ and $D'$ along the distinguished components}; it does not depend on the locations of the removed arcs, which follows from the STU relation and the fact that all components of $X$ are loops. The proof is the same as in the case $X=X'=S^1$ as in \cite{BarNatan}. We define a co-multiplication $\hat \Delta$ in ${\cal A}(X)$ and $\hat{\cal A}(X)$ as follows. A {\it chord sub-diagram} of a chord diagram $D$ with dashed graph $G$ is any chord diagram obtained from $D$ by removing some connected components of $G$. The {\it complement chord sub-diagram} of a chord sub-diagram $D'$ is the chord sub-diagram obtained by removing components of $G$ which are in $D'$. We define $$\hat \Delta (D)=\sum D'\otimes D''.$$ Here the sum is over all chord sub-diagrams $D'$ of $D$, and $D''$ is the complement of $D'$. This co-multiplication is co-commutative. \subsection{Associator} Let ${\Bbb C}<<A,B>>$ be the algebra over ${\Bbb C}$ of all formal power series in two non-commutative symbols $A,B$. We are going to define an element $\varphi\in {\Bbb C}<<A,B>>$, known as the Drinfeld associator. We put \begin{equation} \zeta(i_1,\ldots,i_k)=\sum_{n_1<\ldots<n_k\in{\Bbb N}}\frac{1} {n_1^{i_1}\cdots n_k^{i_k}}, \end{equation} for natural numbers $i_1,\ldots,i_k$ satisfying $i_k\ge2$. These values, called multiple zeta values, have recently gained much attention among number theorists. In what follows bold letters ${\bold p}, {\bold q}, {\bold r}, {\bold s}$ stand for non-negative multi-indices. For a multi-index ${\mathbf{p}}=(p_1,\cdots,p_k)$, we call $k$ the {\it length} of $\mathbf{p}$. Let ${\mathbf{1}}_k$ be the multi-index consisting of $k$ letters 1. We denote $\sum p_i$ by $|{\mathbf{p}}|$. For two multi-indices $\mathbf{p}$ and $\mathbf{q}$ of the same length $k$, we put $\eta ({\mathbf{p}};{\mathbf{q}})=0$ if one of $p_i,q_i$ is 0, $$ \eta ({\mathbf{p}};{\mathbf{q}}) \ =\ \zeta ({\mathbf{1}}_{p_1-1},q_1+1,{\mathbf{1}}_{p_2-1}, q_2+1,\dots ,{\mathbf{1}}_{p_k-1},q_k+1), $$ otherwise. Further we set two notations by \begin{align*} (A,B)^{({\mathbf{p}},{\mathbf{q}})} &=A^{p_1}B^{q_1}A^{p_2}B^{q_2}\dots A^{p_k}B^{q_k}, \\ \binom{\mathbf{p}}{\mathbf{q}} &=\binom{p_1}{q_1}\binom{p_2}{q_2}\cdots \binom{p_k}{q_k}. \end{align*} Using the above notations we define the {\it associator} $\varphi$ by \[ \varphi(A,B)=1+\sum_{\mathbf{p},\mathbf{q},\mathbf{r},\mathbf{s}\ge 0} (-1)^{|{\bold{r}}|+|{\bold{q}}|}\eta ({\mathbf{p}}+{\mathbf{r}}; {\mathbf{q}}+{\mathbf{s}}) \,\binom{\mathbf{p+r}}{\mathbf{r}} \binom{\mathbf{q+s}}{\mathbf{s}}B^{|{\mathbf{s}}|} (A,B)^{({\mathbf{p}},{\mathbf{q}})}\,A^{|{\mathbf{r}}|}. \] Here the sum is over all multi-indices $\mathbf{p,q,r,s}$ of the same length $k$ where $k=1,2,3,\cdots$. Note that there exists the inverse of $\varphi(A,B)$; we denote it by $\varphi^{-1}(A,B)$. \subsection{The universal Vassiliev-Kontsevich invariant for framed oriented links} In this subsection, we will define the universal Vassiliev-Kontsevich invariant $\hat Z(L)$ of a framed oriented link $L$. Before defining it, we define an invariant $Z({\cal D})$ of a link diagram ${\cal D}$. Suppose ${\cal D}$ is an {oriented $l$-component link diagram} in a plane with a fixed coordinate system. Using horizontal lines one can decompose ${\cal D}$ into elementary tangle diagrams. Here we mean an {\it elementary tangle diagram} is one of the tangle diagrams shown in Figure \ref{elementary}, maybe with reverse orientation on some components. We number the components of the elementary tangle diagrams from left to right. There are $n$ components in $X^{\pm}_{k,n}$, and the crossing is on the $k$-th and $(k+1)$-th components. There are $n-1$ components in $U_{k,n}, V_{k,n}$, and the non-straight component is numbered by $k$. We define $Z({\cal D})\in\hat{\cal A}(\coprod^lS^1)$ as follows. \begin{figure}[htpb] $$ \psdraw{le1}{3.6in} $$ \caption{Elementary tangle diagrams}\label{elementary} \end{figure} Firstly we define $Z(T)$ for each elementary tangle diagram $T$. When $X$ is $n$ vertical straight numbered lines with downward orientations, $\hat{\cal A}(X)$ is denoted by ${\cal P}_n$. All the ${\cal P}_n$ are algebras: the product of two chord diagrams $D_1$ and $D_2$ is obtained by placing $D_1$ on the top of $D_2$. The algebra ${\cal P}_1$ is commutative (see \cite{BarNatan,Kontsevich}). Consider the following element $x_{k,n}$ in ${\cal P}_n$ $$ x_{k,n}^+=\varphi^{-1}(\frac{1}{2\pi\sqrt{-1}}\sum_{i=1}^{k-1}\Omega_{i,k}, \frac{\Omega_{k,k+1}}{2\pi\sqrt{-1}}) \exp(\frac{\Omega_{k,k+1}}{2})\varphi(\frac{1}{2\pi\sqrt{-1}}\sum_{i=1}^{k-1} \Omega_{i,k+1}, \frac{\Omega_{k,k+1}}{2\pi\sqrt{-1}}). $$ Here $\varphi$ is defined in the previous subsection, and $\Omega_{i,j}$ is the chord diagram in ${\cal P}_n$ with the dashed graph being a line connecting the $i$-th and the $j$-th string. Similarly we put \begin{align*} x_{k,n}^- &=\varphi^{-1}(\frac{1}{2\pi\sqrt{-1}}\sum_{i=1}^{k-1}\Omega_{i,k}, \frac{\Omega_{k,k+1}}{2\pi\sqrt{-1}}) \exp(-\frac{\Omega_{k,k+1}}{2})\varphi(\frac{1}{2\pi\sqrt{-1}} \sum_{i=1}^{k-1}\Omega_{i,k+1}, \frac{\Omega_{k,k+1}}{2\pi\sqrt{-1}}), \\ u_{k,n} &=\varphi^{-1}(\frac{1}{2\pi\sqrt{-1}}\sum_{i=1}^{k-1}\Omega_{i,k}, \frac{\Omega_{k,k+1}}{2\pi\sqrt{-1}}), \\ v_{k,n} &=\varphi(\frac{1}{2\pi\sqrt{-1}}\sum_{i=1}^{k-1}\Omega_{i,k}, \frac{\Omega_{k,k+1}}{2\pi\sqrt{-1}}). \end{align*} We define $Z(X_{k,n}^+)$ as the element in $\hat{\cal A}(X_{k,n}^+)$ obtained by placing $X^+_{k,n}$ without any dashed graph on the top of $x_{k,n}^+$. Here we also use the notation $X_{k,n}^+$ for the set of solid lines of $X_{k,n}^+$. Similarly, $Z(X_{k,n}^-)$ is the element in $\hat{\cal A}(X_{k,n}^-)$ obtained by placing $X^-_{k,n}$ on the top of $x_{k,n}^-$. Further $Z(U_{k,n})$ is obtained by placing $U_{k,n}$ on the bottom of $S_{C_{k+1}}(u_{k,n})$, where $C_{k+1}$ is the $(k+1)$-th strings of the support of chord diagrams in ${\cal P}_n$, and $S_{C_k}$ is defined in the previous subsection. Furthermore $Z(V_{k,n})$ is obtained by placing $V_{k,n}$ on the top of $S_{C_k}(v_{k,n})$. We also show pictures of the definition in Figure \ref{fig.definexuv}. \begin{figure}[htpb] $$ \psdraw{add}{3.6in} $$ \caption{The definition of $Z(T)$ for each elementary tangle diagram $T$}\label{fig.definexuv} \end{figure} Secondly we define $Z(T)$ for each elementary tangle diagram $T$ with arbitrary orientation by using $Z(T)=S_C(Z(T'))$ some times, where $T$ is obtained from $T'$ by reversing the orientation of a component $C$. Lastly we define $Z({\cal D}) \in \hat{\cal A}(\coprod^l S^1)$ for an oriented $l$-component link diagram ${\cal D}$. We can decompose ${\cal D}$ into elementary tangle diagrams $T_1,T_2,\dots,T_m$ by horizontal lines, counting from top to bottom. We set $Z({\cal D})=Z(T_1)\times Z(T_2)\times \dots\times Z(T_m)$. Here, for chord diagrams $D_i$ composing $Z(T_i)$ ($i=1,2,\dots,m$), we mean by $D_1\times\dots\times D_m$ the chord diagram on $\coprod^l S^1$ obtained by placing $D_1$ on the top of $D_2$, and placing the union on top of $D_3$, and so on. Note that the supports of $D_1,D_2,\dots,D_m$ can be glued together, and the result consists of $l$ solid loops. It is known that $Z({\cal D})$ is defined uniquely for an oriented link diagram ${\cal D}$, not depending on the decomposition of ${\cal D}$, i.e. more precisely, not depending on an isotopy of the link diagram ${\cal D}$ in the plane which preserves the maximal points of ${\cal D}$. Suppose now that $L$ is a framed oriented link, represented by a link diagram ${\cal D}$ with blackboard framing. Further suppose that the $i$-th components of ${\cal D}$ has $s_i$ maximal points with respect to the height function of the plane. We define an invariant $\hat Z(L)$ by $$ \hat Z(L)=Z({\cal D})\#(\nu^{s_1}\otimes\dots\otimes \nu^{s_l}) \in \hat{\cal A}(\coprod^l S^1), $$ where we put $\nu=Z(U)^{-1}$ for the link diagram $U$ shown in Figure \ref{U}, and we mean by the above formula that $\hat Z(L)$ is obtained from $Z({\cal D})$ by successively taking connected sum with $\nu^{s_i}$ along the $i$-th component. (Note that a notation $\hat Z_f(L)$ was used instead of $\hat Z(L)$ in the previous papers \cite{LeMurakami,LeMurakamiII}.) The invariant $\hat Z(L)$ of a framed link $L$ is well-defined, not depending on the choice of its link diagram ${\cal D}$, see \cite{LeMurakami}. Note that, for the trivial knot $K$ with framing 0, $\hat Z(K)$ is not trivial; in fact, we have $\hat Z(K)=\nu$. \begin{figure}[htpb] $$ \psdraw{le2}{1.2in} $$ \caption{The link diagram $U$}\label{U} \end{figure} The invariant $\hat Z(L)$ is an invariant of framed oriented links such that it contains in itself all Vassiliev invariants of framed oriented links, and it is a generalization of the Kontsevich integral. We call this invariant {\it the universal Vassiliev-Kontsevich invariant}. \subsection{Parallel of framed links} The following was proved in \cite{LeMurakamiII}. \begin{prop}\label{prop.ZandDcommute} Let $C$ be a component of an oriented framed link $L$. \begin{enumerate} \item Let $L^{(2,C)}$ be the link obtained from $L$ by replacing $C$ by two push-offs of $C$ using the frame. Then we have the following formula; \begin{equation}\label{eq.parallel} \hat Z(L^{(2,C)})=\Delta_{(C)}(\hat Z(L)). \end{equation} \item Let $L'$ be obtained from $L$ by reversing the orientation of $C$. Then we have the following formula; \begin{equation}\label{eq.revori} \hat Z(L')=S_{(C)}(\hat Z(L)). \end{equation} \end{enumerate} \end{prop} Now we put $$ \check Z(L)=\hat Z(L)\# (\nu\otimes\dots\otimes \nu) \in \hat{\cal A}(\coprod^l S^1). $$ This formula means that $\check Z(L)$ is obtained from $\hat Z(L)$ by successively taking connected sum with $\nu$ along every component of $L$. It is easy to see that Proposition \ref{prop.ZandDcommute} is also valid if we replace $\hat Z$ by ${\check Z}$ in the statement of the proposition. Suppose that $X$ consists of $n$ components $C_1,\dots,C_n$. We denote by $X\sqcup X$ the disjoint union of two copies of $X$. We define a linear map: $$ p : {\hat{\cal A}}(X\sqcup X) \to {\hat{\cal A}}(X)^{\otimes 2} $$ as follows. If $D$ is a chord diagram having dashed graph connecting the two copies of $X$, then we put $p(D)=0$. Otherwise, $D$ splits into a disjoint union of two chord diagrams $D_1$ and $D_2$ on the first and the second copies of $X$ respectively, and we put $p(D)=D_1 \otimes D_2$. Then by definition we have \begin{equation}\label{co-mul} p \circ \Delta_{(C_1,\dots,C_n)}(D) =\hat{\Delta}(D), \label{x1} \end{equation} where $\Delta_{(C_1,\dots,C_n)}:\hat{\cal A}(X)\to\hat{\cal A}(X\sqcup X)$ is the mapping obtained by successively applying $\Delta_{(C_i)},i=1,\dots,n$. \begin{thm}\label{thm.bunshin} Let $L$ be an oriented framed link. Then the following formula holds; $$ \hat\D({\check Z}(L))={\check Z}(L) \otimes {\check Z}(L). $$ \end{thm} \begin{pf} By the equations (\ref{eq.parallel}) and (\ref{x1}), the left hand side of the required formula is equal to $p({\check Z}(L^{(2)}))$. Identifying $\hat{\cal A}(X) \otimes \hat{\cal A}(X)$ with a subset of $\hat{\cal A}(X \sqcup X)$ naturally, the right hand side of the required formula becomes equal to ${\check Z}(L \circ L)$, where $L \circ L$ is the split union of two copies of $L$, i.e., the disjoint union not winding with each other. Hence it is sufficient to show that $p({\check Z}(L^{(2)}))$ is equal to ${\check Z}(L \circ L)$. Note that we can obtain $L \circ L$ from $L^{(2)}$ by taking crossing changes between the first and the second copies of $L$. In this procedure ${\check Z}(L^{(2)})$ changes to ${\check Z}(L \circ L)$ only in terms which have dashed graphs connecting the two copies of $X$, since the difference of a crossing change is locally given by \begin{align*} x_{k,n}^+ - x_{k,n}^- &= \varphi^{-1}(\ldots) \left( \exp(\frac{\Omega_{k,k+1}}{2}) - \exp(-\frac{\Omega_{k,k+1}}{2}) \right) \varphi(\ldots) \\ &= \varphi^{-1}(\ldots) \left( \Omega_{k,k+1} + \text{higher terms with respect to $\Omega_{k,k+1}$} \right) \varphi(\ldots). \end{align*} Therefore, after the procedure, ${\check Z}(L^{(2)})$ becomes an element of $\hat{\cal A}(X \sqcup X)$ which is different from ${\check Z}(L^{(2)})$ only in terms with dashed graphs connecting the two copies of $X$. Further the element should vanish in such terms, since it is equal to the invariant of a split union. Hence it must be equal to $p({\check Z}(L^{(2)}))$; it is also equal to ${\check Z}(L \circ L)$. \end{pf} \subsection{The change under Kirby move II} We have the following proposition proved in \cite{LeMurakamiII}. By this proposition we can get the change of ${\check Z}(L)$ under Kirby move II. Throughout the present paper we mean by Kirby move II the handle slide move defined by Kirby in \cite{Kirby}; we show a picture of the move in Figure \ref{fig.handleslide}. \begin{figure}[htpb] $$ \psdraw{add2}{3.6in} $$ \caption{Kirby move II (the handle slide move)}\label{fig.handleslide} \end{figure} \begin{prop}[\cite{LeMurakamiII}]\label{prop.KII} Let $L$ be an oriented framed link, and $L'$ a framed link obtained from $L$ by Kirby move II which preserves the orientation. Then ${\check Z}(L')$ can be obtained from ${\check Z}(L)$ by replacing the left picture in Figure \ref{fig.KII} with the right picture. \end{prop} \begin{figure}[htpb] $$ \psdraw{pic1}{3.6in} $$ \caption{The change under Kirby move II}\label{fig.KII} \end{figure} \section{Replacing solid circles by dashed lines} Our aim of this section is to construct the series of the maps $\iota_n$, whose definition is given in the end of this section. The universal Vassiliev-Kontsevich invariant belongs to the space consisting of both of solid and dashed lines. When we consider quantum $({\frak g}, R)$ invariants of links for a Lie algebra $\frak g$ and its representation $R$, the solid and dashed lines corresponds to $R$ and $\frak g$ respectively. Further, we know that no particular representation $R$ is specified in quantum $\frak g$ invariants of 3-manifolds. It implies that our obstruction in constructing invariants of 3-manifolds from the universal Vassiliev-Kontsevich invariant might be the existence of solid lines. We will remove solid lines from the space of chord diagrams by the maps $\iota_n$ when we construct invariants of 3-manifolds in the following section. \subsection{Chord diagrams with finite support} Let $X$ be a finite set consisting of $m$ ordered points named $0,1,2,$ $\cdots,m-1$. We consider chord diagrams with support $X$, that is, oriented uni-trivalent graphs whose $m$ univalent vertices are on the $m$ fixed points respectively. We denote by ${\cal A}(m)$ the vector space over ${\Bbb C}$ spanned by such chord diagrams subject to the AS and IHX relations. Further we denote by $\cAt{m}$ the vector subspace of ${\cal A}(m)$ spanned by connected and simply connected graphs; note that the AS and IHX relations are closed in the subspace. For an element $\tau$ in the symmetric group ${\cal S}_{m-2}$ acting on the set $\{1,2,\cdots,m-2\}$, let $T_{\tau} \in {\cal A}(m)$ be the graph shown in Figure \ref{fig.Ttau}. \begin{figure}[htpb] $$ \psdraw{fig4}{3.6in} $$ \caption{The definition of $T_{\tau}$}\label{fig.Ttau} \end{figure} \begin{lem}\label{lem.basis} We can take the set of $T_{\tau}$ as basis of the space $\cAt{m}$; in particular the space is $(m-2)!$ dimensional. \end{lem} \begin{pf} Let $D$ be any chord diagram in $\cAt{m}$. We put red color on the path connecting the two univalent vertices $0$ and $m-1$. We can deform $D$ into a linear sum of $T_{\tau}$'s by induction on the number of trivalent vertices on the red path as follows. We choose a trivalent vertex next to the red path, and apply the IHX relation regarding the segment connecting the vertex and the red path as the character \lq\lq I'' in \lq\lq IHX''. Then the number of trivalent vertices on the red path increases. Hence we can show that the space $\cAt{m}$ is spanned by the set of $T_{\tau}$. In order to complete the proof of this lemma, it is sufficient to prove that $T_{\tau}$'s are linearly independent. Suppose that $T_{\tau}$ could be expressed as a linear sum of other $T_{\sigma}$'s. We can \lq\lq substitute'' a Lie algebra $sl(m,{\Bbb C})$ to dashed lines (see \cite{BarNatan}) to make a linear map of $\cAt{m}$ to ${\Bbb C}$. For $j<k$ let $E_{jk}$ be the element in $sl(m,{\Bbb C})$ which has $(j,k)$ entry $1$ and the other entries $0$. We have a relation $[E_{ij},E_{jk}]=E_{ik}$. If we substitute $E_{12}$ to the univalent vertex $0$ and $E_{k+1,k+2}$ to the vertex $\tau(k)$ for $k=1,2,\cdots,m-2$, then $T_{\sigma}$ always vanishes unless ${\sigma}=\tau$, though $T_{\tau}$ does not vanish when we substitute the dual of $E_{1m}$ to the vertex $m-1$, where we mean the dual with respect to the Killing form. This is a contradiction, completing the proof. \end{pf} \subsection{Chord diagrams behaving in a similar way as a solid circle} We define $T_m \in \cAt{m}$ by $$ T_m = \sum_{\tau\in{{\cal S}}_{m-2}} \frac{(-1)^{r(\tau)}}{(m-1)\binom{m-2}{r(\tau)}} T_{\tau}, $$ where we denote by $r(\tau)$ the number of $k$ which satisfies $\tau(k)>\tau(k+1)$. In this subsection, our aim is to show Proposition \ref{prop.STUforTm}. We begin with the following lemma, which is a weak form of Proposition \ref{prop.STUforTm}; we can interchange any two adjacent univalent vertices in Proposition \ref{prop.STUforTm} whereas we can do it for particular pairs in Lemma \ref{lem.STUforT}. We will show symmetries of $T_m$ in this subsection to obtain Proposition \ref{prop.STUforTm} from Lemma \ref{lem.STUforT}. \begin{lem}\label{lem.STUforT} If $1 \le k \le m-3$, then the difference between $T_m$ and the chord diagram obtained from $T_m$ by changing two univalent vertices $k$ and $k+1$ can be expressed using $T_{m-1}$ as shown in Figure \ref{fig.STUinTm}. \end{lem} \begin{figure}[htpb] $$ \psdraw{fig6}{3.6in} $$ \caption{A property of $T_m$ similar to the STU relation}\label{fig.STUinTm} \end{figure} \begin{pf} Since we can take the set of $T_{\tau}$ as basis of $\cAt{m}$, we can express both sides of the required formula as a linear sum of $T_{\tau}$'s. It is sufficient to show that the coefficients of $T_{\tau}$ in both sides are equal for each $\tau$. If $|\tau^{-1}(k)-\tau^{-1}(k+1)| \ge 2$, then the coefficient of the left hand side is equal to zero, since $r(\tau)=r((k\ k+1)\circ \tau)$ holds in this case, where we mean by $(k\ k+1)$ the interchange of $k$ and $k+1$. On the other hand, the coefficient of the right hand side is equal to zero, since the right hand side is equal to a linear sum of $T_{\tau}$ for $\tau$ satisfying $|\tau^{-1}(k)-\tau^{-1}(k+1)|= \pm 1$; we can see it by applying the IHX relation in the right hand side. Therefore the coefficients of $T_{\tau}$ in both sides are equal in this case. If $\tau^{-1}(k)-\tau^{-1}(k+1) = -1$, then the coefficient of the left hand side is equal to $t_{m,\tau}-t_{m,(k\ k+1)\circ \tau}$ where we put $$ t_{m,\tau}= \frac{(-1)^{r(\tau)}}{(m-1)\binom{m-2}{r(\tau)}}. $$ In this case we have $r((k\ k+1)\circ \tau)=r(\tau)+1$ by the definition of $r(\cdot)$. Hence we have \begin{align*} t_{m,\tau}-t_{m,(k\ k+1)\circ \tau} &=\frac{(-1)^{r(\tau)}}{(m-1)\binom{m-2}{r(\tau)}} -\frac{(-1)^{r(\tau)+1}}{(m-1)\binom{m-2}{r(\tau)+1}} \\ &=\frac{(-1)^{r(\tau)}(m-r-2)}{(m-1)(m-2)\binom{m-3}{r(\tau)}} -\frac{(-1)^{r(\tau)+1}(r+1)}{(m-1)(m-2)\binom{m-3}{r(\tau)}} =\frac{(-1)^{r(\tau)}}{(m-2)\binom{m-3}{r(\tau)}} \end{align*} On the other hand, the contribution of the right hand side to $T_{\tau}$ comes from $T_{\hat\tau} \in {\cal A}(m-1)$, where we define $\hat\tau \in {\cal S}_{m-1}$ by putting $\hat\tau^{-1}(j)$ to be $\tau^{-1}(j)$ if $j \le k$, $\tau^{-1}(j+1)-1$ if $j > k$. The coefficient of $T_{\hat\tau}$ is equal to $t_{m-1,\hat\tau}$; in this case we have $r(\hat\tau)=r(\tau)$ by the definition of $r(\cdot)$. Therefore the coefficients of both sides are equal, completing this case. If $\tau^{-1}(k)-\tau^{-1}(k+1) = 1$, we can show that the coefficients are equal in a similar way as above, completing the proof. \end{pf} \begin{lem}\label{lem.sym1} The chord diagram $T_m$ is symmetric with respect to mirror image which replaces $0,1,\cdots,m-1$ with $m-1,m-2,\cdots,0$ respectively, see Figure \ref{fig.T2T3T4} for simple cases. To be exact, the inversion replaces $T_m$ to $(-1)^{m-2}T_m$ because it changes the orientations of $m-2$ trivalent vertices. \end{lem} \begin{pf} By the inversion which replaces $0,1,\cdots,m-1$ with $m-1,m-2,\cdots,0$, the chord diagram $T_{\tau}$ moves to $(-1)^{m-2}T_{\tau'}$ where $$ \tau' = \begin{pmatrix} 1 & 2 & \cdots & m-2 \\ m-2 & m-1 & \cdots & 1 \end{pmatrix} \circ \tau \circ \begin{pmatrix} 1 & 2 & \cdots & m-2 \\ m-2 & m-1 & \cdots & 1 \end{pmatrix} $$ and the sign is derived from the number of trivalent vertices; we use the AS relation such times. We can obtain $r(\tau)=r(\tau')$ by definition of $r(\cdot)$. Therefore the inversion maps $T_m$ to $(-1)^{m-2} T_m$, completing the proof. \end{pf} \begin{lem}\label{lem.sym2} The chord diagram $T_m$ is symmetric with respect to mirror image which replaces $0,1,2,\cdots,m-2,m-1$ with $m-2,m-3,m-4,\cdots,0,m-1$ respectively. To be exact, the inversion replaces $T_m$ to $(-1)^{m-2}T_m$ because of the orientations of trivalent vertices. \end{lem} \begin{pf} Consider the linear map $i$ of $\cAt{m}$ to $\cAt{m+1}$ which maps a chord diagram $D$ to the diagram $D$ added a small branch near the univalent vertex $m-1$ as shown in Figure \ref{fig.addvertex}. The map $i$ is an injection; we can see it by taking basis of $\cAt{m}$ and $\cAt{m+1}$ in the way how we choose the red path connecting $0$ and $m-1$ and connecting $0$ and $m$ respectively as in the proof of Lemma \ref{lem.basis}. \begin{figure}[htpb] $$ \psdraw{pic2}{3.6in} $$ \caption{The map of $\cAt{m}$ to $\cAt{m+1}$}\label{fig.addvertex} \end{figure} We denote by ${\cal S}'_{m-1}$ the symmetric group acting on the set $\{0,1,\cdots,m-2\}$. For an element ${\sigma} \in {\cal S}'_{m-1}$, let $S_{\sigma} \in \cAt{m+1}$ be the chord diagram shown in Figure \ref{fig.Ssigma}. By Lemma \ref{lem.basis} we can take the set of $S_{\sigma}$ as basis of $\cAt{m+1}$. \begin{figure}[htpb] $$ \psdraw{pic3}{3.6in} $$ \caption{The definition of $S_{\sigma}$}\label{fig.Ssigma} \end{figure} We can express $i(T_m)$ as in Lemma \ref{lem.TtoS} below. In order to complete the proof of the present lemma, it is sufficient to show that $i(T_m)$ is symmetric with respect to the inversion replacing $0,1,\cdots,m-2$ with $m-2,m-1,\cdots,0$, since the map $i$ is an injection. The inversion maps $S_{\sigma}$ to $S_{{\sigma}'}$ where $$ {\sigma}' = \begin{pmatrix} 0 & 1 & \cdots & m-2 \\ m-2 & m-1 & \cdots & 0 \end{pmatrix} \circ {\sigma}, $$ note that we have no change of sign in this case because we fix the ends $m-1$ and $m$ of the \lq\lq red path''. We can obtain $r({\sigma}')=m-2-r({\sigma})$ by the definition of $r(\cdot)$. By Lemma \ref{lem.TtoS} below, the inversion maps $i(T_m)$ to $(-1)^{m-2} i(T_m)$, completing the proof. \end{pf} \begin{lem}\label{lem.TtoS} $$ i(T_m) = \sum_{{\sigma}\in{{\cal S}'}_{m-1}} \frac{(-1)^{r({\sigma})}}{(m-1)\binom{m-2}{r({\sigma})}} S_{\sigma}. $$ \end{lem} \begin{pf} We put $S_{m+1}=i(T_m)$. Since the set of $S_{\sigma}$ is basis of $\cAt{m+1}$, we can put $S_{m+1}=\sum_{\sigma} s_{\sigma} S_{\sigma}$ with some scalars $s_{\sigma}$. In the following of this proof we will show the following formula \begin{equation}\label{eq.snodef} s_{\sigma} = \frac{(-1)^{r({\sigma})}}{(m-1)\binom{m-2}{r({\sigma})}} \end{equation} in two steps by induction on $m$. \noindent {\bf Step 1.}\quad If ${\sigma}$ is a cyclic permutation $(0,1,2,\cdots,k)$ for an integer $k$, we can obtain (\ref{eq.snodef}) by definition of $i(T_m)$ as follows. We consider that what $T_{\tau}$ contributes to such $S_{\sigma}$ after changing basis of $\cAt{m}$ to that of $\cAt{m+1}$; we expand the left picture in Figure \ref{fig.expandT} using the IHX relation to obtain the right picture. We use the IHX relation replacing \lq\lq I'' with the difference of \lq\lq H'' and \lq\lq X''. Noting that the univalent vertex $0$ interchanges $k$ times with another vertex, we must use \lq\lq X'' $k$ times in the expansion. Hence the possibilities of $\tau$ are as follow; \begin{align*} \tau &= \left( {\begin{matrix} 1 &2 &\cdots &l_k-1 &l_k &l_k+1 &\cdots \\ k+1 &k+2 &\cdots &l_k+k-1 &k &l_k+k &\cdots \end{matrix}\quad \begin{matrix} l_1-1 &l_1 &l_1+1 &\cdots &m-2 \\ l_1 &1 &l_1+1 &\cdots &m-2 \end{matrix}} \right) \\ &\text{or } \left( {\begin{matrix} 1 &2 &\cdots &l_{k-1}-1 &l_{k-1} &l_{k-1}+1 &\cdots \\ k &k+1 &\cdots &l_{k-1}+k-2 &k-1 &l_{k-1}+k-1 &\cdots \end{matrix}\quad \begin{matrix} m-2 \\ m-2 \end{matrix}} \right) \end{align*} In the first type we can freely choose $\{ l_k,l_{k-1},\cdots,l_1 \}$ from $\{2,3,\cdots,m-2 \}$. Hence there are $\binom{m-3}{k}$ possibilities of $\tau$, and $r(\tau)=k$ holds in this type. Similarly we have $\binom{m-3}{k-1}$ possibilities of $\tau$, which satisfy $r(\tau)=k-1$ in the second type. Therefore we have $$ s_{\sigma} = (-1)^k \binom{m-3}{k} \frac{(-1)^k}{(m-1)\binom{m-2}{k}} + (-1)^k \binom{m-3}{k-1} \frac{(-1)^{k-1}}{(m-1)\binom{m-2}{k-1}} $$ where the term $(-1)^k$ is derived from the number of usage of \lq\lq X''. This formula satisfies (\ref{eq.snodef}), completing Step 1. \begin{figure}[htpb] $$ \psdraw{pic4}{3.6in} $$ \caption{Expanding $T_{\tau}$}\label{fig.expandT} \end{figure} \noindent {\bf Step 2.}\quad In this step we will show that if (\ref{eq.snodef}) holds for ${\sigma}$ then it also holds for $(k\ k+1)\circ {\sigma}$ for any $k = 1,2,\cdots,m-3$, where we mean by $(k\ k+1)$ the interchange of $k$ and $k+1$. We have the formula shown in Figure \ref{fig.STUforS}, where the first and third equalities in the figure are derived from the definition of $S_\star$ and the second equality is derived from Lemma \ref{lem.STUforT}. \begin{figure}[htpb] $$ \psdraw{pic5}{3.6in} $$ \caption{$S_\star$ satisfies a relation similar to the STU relation} \label{fig.STUforS} \end{figure} Hence $S_{m+1}$ satisfies a relation similar to the STU relation; this means $$ s_{\sigma} - s_{(k\ k+1) \circ {\sigma}} = \begin{cases} s_{\hat{\sigma}} \quad & \text{ if } {\sigma}^{-1}(k)-{\sigma}^{-1}(k+1)=-1 \\ -s_{\hat{\sigma}} \quad & \text{ if } {\sigma}^{-1}(k)-{\sigma}^{-1}(k+1)=1 \\ 0 \quad & \text{ if } |{\sigma}^{-1}(k)-{\sigma}^{-1}(k+1)| \ge 2 \end{cases} $$ where we define $\hat{\sigma}$ by putting $\hat{\sigma}^{-1}(j)$ to be ${\sigma}^{-1}(j)$ if $j \le k$, ${\sigma}^{-1}(j+1)-1$ of $j>k$; note that we can obtain the chord diagram $S_{\hat{\sigma}}$ from $S_{\sigma}$ by gluing two adjacent dashed edges who have univalent vertices $k$ and $k+1$ respectively, and we can define $\hat{\sigma}$ only when ${\sigma}^{-1}(k)-{\sigma}^{-1}(k+1)=\pm1$. We note that we can use (\ref{eq.snodef}) for $\hat{\sigma}$ by hypothesis of induction. Since we can check that the above formula satisfies (\ref{eq.snodef}), we obtain the required claim of Step 2. Since we can obtain any ${\sigma} \in {\cal S}_{m-1}$ from some permutation $(0,1,2,\cdots,k)$ by composing interchanges $(j\ j+1)$, we can show (\ref{eq.snodef}) for any ${\sigma}$, completing the proof. \end{pf} By Lemmas \ref{lem.sym1} and \ref{lem.sym2}, we immediately obtain the following proposition. \begin{prop}\label{prop.sym} The chord diagram $T_m$ is symmetric with respect to the action of dihedral group of order $2m$; we show simple cases in Figure \ref{fig.T2T3T4}. Note that the sign of $T_m$ possibly changes by the action as in the statements of Lemmas \ref{lem.sym1} and \ref{lem.sym2}. \end{prop} \begin{figure}[htpb] $$ \psdraw{fig5}{3.6in} $$ \caption{Simple cases of $T_m$}\label{fig.T2T3T4} \end{figure} Noting the symmetry of $T_m$ in Proposition \ref{prop.sym}, we can obtain the following proposition from Lemma \ref{lem.STUforT}. \begin{prop}\label{prop.STUforTm} The difference between $T_m$ and the chord diagram obtained by changing any adjacent two univalent vertices is equal to $T_{m-1}$ with one extra trivalent vertex as shown in Figure \ref{fig.STUinTm}. \end{prop} \subsection{Replacing solid circles with dashed lines} Let $n$ be a positive integer. For an integer $m$ with $m \ge 2n$, we define $T^n_m \in {\cal A}(m)$ as follows. We divide $m$ into $n$ integers as; $$ m=m_1+m_2+\cdots+m_n, \qquad m_1 \ge m_2 \ge \cdots \ge m_n \ge 2. $$ Let $T^n_m$ be the sum of all configurations of disjoint union of $T_{m_i}$'s where we take the configurations preserving cyclic order of univalent vertices of each $T_{m_i}$. We show the picture of $T^2_5$ in Figure \ref{fig.T25}. If $m$ is less than $2n$, then we put $T^n_m$ to be $0$. \begin{figure}[htpb] $$ \psdraw{fig7}{3.6in} $$ \caption{The picture of $T^2_5$}\label{fig.T25} \end{figure} \begin{prop}\label{prop.STUforTnm} For any positive integer $n$, the difference between $T^n_m$ and the chord diagram obtained by changing any adjacent two univalent vertices is equal to $T^n_{m-1}$ with one extra trivalent vertex. Namely $T^n_m$ also satisfies the same formula as in Figure \ref{fig.STUinTm} for $T_m$. \end{prop} \begin{pf} For each term in $T^n_m$, we consider the difference in the left hand side. If the two adjacent vertices are on different connected components of $T^n_m$, then the difference vanishes. If they are on the same connected component, then the difference is equal to a term in $T^n_{m-1}$ in the right hand side by Proposition \ref{prop.STUforTm}. Hence we obtain the required formula. \end{pf} Using the above proposition, we can define a linear map $$ \iota_n:{\cal A}(\coprod^l S^1) \longrightarrow{\cal A}(\phi) $$ by putting $\iota_n(\text{a solid circle with $m$ dashed univalent vertices})$ to be $T^n_m$ as shown in Figure \ref{fig.iota}. We also denote by the same notation $\iota_n$ the naturally extended map of $\hat{\cal A}(\coprod^l S^1)$ to $\hat{\cal A}(\phi)$. \begin{figure}[htpb] $$ \psdraw{fig8}{3.6in} $$ \caption{The definition of $\iota_n$}\label{fig.iota} \end{figure} \section{A series of invariants of 3-manifolds} In this section we will construct a series of topological invariants $\Omega_n(M)$ of a 3-manifold $M$. We show the invariance of ${\check Z}(L)$ under orientation change and Kirby move II using the equivalence relation $P_{\star}$ defined below. Since the relation $P_{n+1}$ vanish in low degrees in the image of $\iota_n$, we will obtain a series of invariants in low degrees of ${\cal A}(\phi)$ through the maps $\iota_n$. \subsection{Invariance under orientation change and Kirby move II} Let ${\overset{\circ}{\cal A}}(X)$ be the space of chord diagrams including dashed trivial circles. We denote by $\hoA{X}$ its completion. For each positive integers $n$, we define an equivalence relation $P_n$ in ${\overset{\circ}{\cal A}}(X)$ and $\hoA{X}$ as follows. Let $P_1$ be the equivalence relation such that any chord diagram with non-empty dashed lines is equivalent to zero. Let $P_2$ be the equivalence relation shown in Figure \ref{fig.P2P3}; the left hand side of the first formula in Figure \ref{fig.P2P3} is the sum over all pairings of $4$ points. Similarly we define the equivalence relation $P_n$ such that the sum over all pairings of $2n$ points is equivalent to zero. \begin{figure}[htpb] $$ \psdraw{fig9}{3.6in} $$ \caption{The relations $P_2$ and $P_3$}\label{fig.P2P3} \end{figure} Before proving the invariance of ${\check Z}(L)$ in Proposition \ref{prop.invariantbyP}, we prepare the following lemma. \begin{lem}\label{lem.decreaselegs} Let $D$ be a chord diagram on $X$, and $C$ a component of $X$. \begin{enumerate} \item Then, with the equivalence relation $P_{n+1}$, the chord diagram $D$ is equivalent to a linear sum of chord diagrams each of which has at most $2n$ univalent vertices on $C$. \item Further, the chord diagram $D$ becomes equivalent to a linear sum of chord diagrams each of which has either $n$ isolated dashed chords on $C$ or at most $2n-1$ univalent vertices on $C$. Here we mean by an isolated chord a dashed arc with no trivalent vertices and two adjacent univalent vertices on $C$. \end{enumerate} \end{lem} \begin{pf} We will see the case $n=3$ before the general case. It is sufficient to show that, if the diagram $D$ has $k$ univalent vertices on $C$ with $k>4$, then it is equivalent to a linear sum of chord diagrams each of which has less than $k$ univalent vertices; we will call them lower terms. We use the relation $P_3$ as in Figure \ref{fig.useP3}, where the second equality is derived from the STU relation; we can use the STU relation to interchange two univalent vertices modulo a lower term. \begin{figure}[htpb] $$ \psdraw{pic6}{3.6in} $$ \caption{Using $P_3$ to make isolated chords}\label{fig.useP3} \end{figure} Then we can replace $D$ with a chord diagram with an isolated chord, where we mean by an isolated chord a trivial dashed arc with adjacent univalent vertices on $X$. Iterating this procedure, we can replace $D$ with a chord diagram with at least two isolated chords. Then we can use the relation $P_3$ again as in Figure \ref{fig.useP3again}, and we can replace the diagram with lower terms, completing the proof of (1) for the case $n=3$. \begin{figure}[htpb] $$ \psdraw{pic7}{3.6in} $$ \caption{Using $P_3$ to decrease univalent vertices}\label{fig.useP3again} \end{figure} In order to prove (2), it is sufficient to show that any chord diagram with $4$ univalent vertices on $C$ is equivalent to a chord diagram with two isolated chords on $C$ modulo lower terms. We use the relation in Figure \ref{fig.useP3} again, to make one isolated chord. We further use the relation $P_3$ as in Figure \ref{fig.useP3further}. Then we can replace the diagram with a diagram with two isolated chords on $C$, completing the proof of (2) for the case $n=3$. \begin{figure}[htpb] $$ \psdraw{add1}{3.6in} $$ \caption{Using $P_3$ to make two isolated chords}\label{fig.useP3further} \end{figure} In a general case for proving (1), it is sufficient to show that, if the diagram $D$ has $m$ univalent vertices on $C$ with $m>2n$, then it is equivalent to a linear sum of chord diagrams with at most $m-1$ univalent vertices. We use the relation $P_{n+1}$ as in Figure \ref{fig.useP} for $0 \le k \le n$; we use it for $k=n$ at the beginning to make an isolated chord, and use it for $k=2,3,\cdots$ to increase isolated chords, and finally use it for $k=0$ to replace the diagram with lower terms; note that we can do it assuming $m>2n$. This completes the proof of (1). In order to prove (2), it is sufficient to show that any chord diagram with $2n$ univalent vertices on $C$ is equivalent to a diagram with $n$ isolated chords on $C$ modulo lower terms. We use the relation in Figure \ref{fig.useP} as above to make $n-1$ isolated chords on $C$. We further use the relation for $k=2$, and we obtain a chord diagram with $n$ isolated chords on $C$. This completes the proof of (2). \end{pf} \begin{figure}[htpb] $$ \psdraw{pic8}{3.6in} $$ \caption{Using $P_{n+1}$ to decrease univalent vertices; this picture is for even $k$}\label{fig.useP} \end{figure} \begin{rem} We can uniquely characterize $T_m^n$ as an element of ${\cal A}(m)/P_{n+1}$ by the following four conditions. \begin{enumerate} \item It satisfies $T_m^n=0$ if $m<2n$. \item It is invariant under cyclic permutation of the $m$ external univalent vertices. \item It satisfies a relation similar to the STU relation; in other words, satisfies Proposition \ref{prop.STUforTnm}. \item It contains neither dashed cycle nor dashed component; that is, it is equal to a linear sum of chord diagrams each of which is a disjoint union of simply connected chord diagrams having external vertices. \end{enumerate} Outline of a proof of the fact is as follows. By Lemma \ref{lem.decreaselegs} a solid circle with $m$ dashed chords is equivalent (modulo $P_{n+1}$ and lower terms) to the disjoint union of a solid circle with $n$ isolated dashed chords and a linear sum of dashed trivalent graphs. We can obtain an alternative definition of $\iota_n$; we define the inverse map $\iota_n^{-1}$ by removing the solid circle with $n$ dashed isolated chords from the above disjoint union. The image of solid circle with $m$ short dashed chords through $\iota^{-1}_n$ is equal to $T_m^n$ by definition of $\iota_n$, and we can check the image satisfies the above four conditions. \end{rem} By Lemma \ref{lem.decreaselegs} we can replace a solid circle with a solid circle with $2n$ univalent vertices modulo lower terms. In the following proposition we show the invariance of ${\check Z}(L)$ by reducing the proof to the above particular solid circle ignoring the lower terms. \begin{prop}\label{prop.invariantbyP} Let $L$ be any oriented framed link of $l$ components, $n$ any positive integer. \begin{enumerate} \item The equivalence class $[{\check Z}(L)]$ including ${\check Z}(L)$ in $\hoA{\coprod^l S^1}/L_{<2n}, P_{n+1}, O_n$ does not depend on orientation of $L$, where we denote by $L_{<2n}$ the equivalence relation such that any chord diagram including a solid circle with less than $2n$ dashed univalent vertices is equivalent to zero, and $O_n$ the equivalence relation such that a trivial dashed circle is equivalent to $-2n$. \item The above equivalence class is invariant under Kirby move II. \end{enumerate} \end{prop} \begin{pf} We will show the proposition for each chord diagram $D$ composing ${\check Z}(L)$. Let $L'$ be the framed link obtained from $L$ by changing the orientation of a component $C$; we denote by $C'$ the corresponding solid circle in the chord diagram in ${\check Z}(L)$. The change of ${\check Z}(L)$ to ${\check Z}(L')$ is given in (\ref{eq.revori}); the diagram in ${\check Z}(L')$ obtained by changing $D$ is equal to $S_{(C)}(D)$. We will show that $S_{(C)}(D)$ is equivalent to $D$. If $D$ has less than $2n$ dashed univalent vertices on $C'$, then both of $D$ and $S_{(C)}(D)$ vanish by the relation $L_{<2n}$. Hence they are equivalent. If $D$ has $2n$ univalent vertices on $C'$, then by (\ref{eq.revori}) we can obtain $S_{(C)}(D)$ from $D$ by changing the order of univalent vertices on $C'$. By the relation $L_{<2n}$ and the STU relation, their equivalence classes are invariant under changing the order of univalent vertices on $C'$ in this case. Hence the equivalence classes are equivalent. If $D$ has $k$ univalent vertices on $C'$ for $k>2n$, then we can reduce this case to the case $k=2n$ as follows. By Lemma \ref{lem.decreaselegs} we can replace $D$ with a chord diagram with less univalent vertices on $C$. The relations we used in the proof are only the STU relation and the relation $P_{n+1}$. Namely, we have a series of a linear sum of chord diagrams; $D=D_0,D_1,D_2,\cdots D_N$, where $D_i$ and $D_{i+1}$ are related by either of the STU relation or the relation $P_{n+1}$. Note that, if $D_i$ and $D_{i+1}$ are related by the STU relation, then $S_{(C)}(D_i)$ and $S_{(C)}(D_{i+1})$ are also related by the relation, because $S_{(C)}$ and the STU relation commute as shown in Figure \ref{fig.SandSTU}; in fact, it guarantees that the map $S_{(C)}$ is well defined. Further note that the relation $P_{n+1}$ and the STU relation commute, because the relation $P_{n+1}$ is independent of solid chords. Hence $S_{(C)}(D_i)$ and $S_{(C)}(D_{i+1})$ are related by the STU relation or the relation $P_{n+1}$. Therefore, if $D_{i+1}$ is equivalent to $S_{(C)}(D_{i+1})$, then if $D_i$ is equivalent to $S_{(C)}(D_i)$; this implies the required reduction, completing this case. This completes the proof of (1). \begin{figure}[htpb] $$ \psdraw{pic9}{3.6in} $$ \caption{The map $S_{(C)}$ and the STU relation commute}\label{fig.SandSTU} \end{figure} Let $L''$ be a framed link obtained from $L$ by Kirby move II taking handle sliding of some component over a component $C$. The change of ${\check Z}(L)$ to ${\check Z}(L'')$ is given by Proposition \ref{prop.KII}; the diagram in ${\check Z}(L'')$ obtained by changing $D$ is equal to $\Delta_{(C)}(D)$. We will show that $\Delta_{(C)}(D)$ is equivalent to $D$. If there are at most $2n-1$ univalent vertices on $C'$, then both of $D$ and $\Delta_{(C)}(D)$ vanish by the relation $L_{<2n}$. Hence they are equivalent. If there are $2n$ univalent vertices on $C'$, then by Proposition \ref{prop.KII} $\Delta_{(C)}(D)$ is equal to a sum of $D$ and the other terms. Further there are less than $2n$ univalent vertices on $C'$ for each of the other terms. Hence their equivalent classes vanish by the relation $L_{<2n}$. Therefore $D$ and $\Delta_{(C)}(D)$ are equivalent. If there are $k$ univalent vertices on $C'$ for $k>2n$, then we can reduce this case to the case $k-1$ as in the above proof of (1). Instead of the commutation of $S_{(C)}$ and the STU relation, we need the commutation of $\Delta_{(C)}$ and the STU relation in the present case; we show it in Figure \ref{fig.DandSTU}. \begin{figure}[htpb] $$ \psdraw{pic10}{3.6in} $$ \caption{The map $\Delta_{(C)}$ and the STU relation commute}\label{fig.DandSTU} \end{figure} This completes the proof of Proposition \ref{prop.invariantbyP}. \end{pf} \subsection{Moving ${\check Z}(L)$ into a set with algebra structure} In order to show the invariance under Kirby move I, we move ${\check Z}(L)$ into a quotient space of ${\cal A}(\phi)$, in which there is an algebra structure with respect to the disjoint union of chord diagrams. Before moving, we prepare the following lemma, which guarantees that we need not consider $P_{n+1}$ in low degrees of ${\cal A}(\phi)$. \begin{lem}\label{lem.Pvanishes} The identity map induces an isomorphism of the quotient space ${\cal A}(\phi)/D_{>n}$ to the quotient space ${\overset{\circ}{\cal A}}(\phi)/D_{>n},P_{n+1},O_n$. \end{lem} \begin{pf} We can remove the trivial dashed component in ${\overset{\circ}{\cal A}}(\phi)$ by $O_n$. Hence it is sufficient to show the claim that, if an element of ${\overset{\circ}{\cal A}}(\phi)$ including $P_{n+1}$, then either it vanishes or its degree is greater than $n$, where we also denote by $P_{n+1}$ the formula defining the relation $P_{n+1}$. Note that $P_{n+1}$ is symmetric with respect to the action of ${\cal S}_{2n}$ acting on the set of $2n$ ends of $P_{n+1}$. We will show the claim for any outside of $P_{n+1}$. We will show that the degree must be more than $n$ unless the element vanishes. We can make an injection of the set of $2n+2$ ends of $P_{n+1}$ to the set of trivalent vertices in the outside as follows. Choose an end of $P_{n+1}$. If the end is not connected to any other end in the outside, we associate one of trivalent vertices in the connected component of the end in the outside. Otherwise we can find a path connecting the end to one of the other ends in the outside. If there are no trivalent vertices on the path, the element vanishes with the relation $O_n$ as shown in Figure \ref{fig.Pvanish}. \begin{figure}[htpb] $$ \psdraw{pic11}{3.6in} $$ \caption{The relation $P_{n+1}$ vanishes}\label{fig.Pvanish} \end{figure} If there are one trivalent vertices on the path, the element vanishes using the symmetry of $P_{n+1}$ and the AS relation as shown in Figure \ref{fig.Pvanishagain}. \begin{figure}[htpb] $$ \psdraw{pic12}{3.6in} $$ \caption{The relation $P_{n+1}$ vanishes again}\label{fig.Pvanishagain} \end{figure} Otherwise there must be at least two trivalent vertices on the path. Then we associate the nearest trivalent vertex to the end, to obtain an injection of the set of $2n+2$ ends of $P_{n+1}$ to the set of trivalent vertices. Hence we showed that the number of trivalent vertices is greater than $2n$; this implies that the degree is greater than $n$, completing the proof. \end{pf} \begin{prop}\label{prop.invariant} \begin{enumerate} \item The equivalence class $[\iota_n({\check Z}(L))] \in {\cal A}(\phi)/D_{>n}$ does not depend on orientation of $L$. \item Further, the above equivalence class does not change under Kirby move II. \end{enumerate} \end{prop} \begin{pf} By the map $\iota_n$ any chord diagram with less than $2n$ univalent vertices on a solid circle of the diagram vanishes. Hence we have the following map; $$ \hoA{\coprod^l S^1}/L_{<2n},P_{n+1},O_n \overset{\iota_n}{\longrightarrow} \hoA{\phi}/P_{n+1},O_n \overset{\text{proj}}{\longrightarrow} {\overset{\circ}{\cal A}}(\phi)/D_{>n},P_{n+1},O_n $$ where the first map is the map induced by $\iota_n$ which we also denote by $\iota_n$, and the second map is the projection. By Proposition \ref{prop.invariantbyP} the equivalence class $[{\check Z}(L)]$ in the first set $\hoA{\coprod^l S^1}/L_{<2n},P_{n+1},O_n$ is invariant under both of orientation change and Kirby move II. Hence it is also true for the equivalence class $[\iota_n {\check Z}(L)]$ in the third set ${\overset{\circ}{\cal A}}(\phi)/D_{>n},P_{n+1},O_n$, which is naturally isomorphic to the set ${\cal A}(\phi)/D_{>n}$ by Lemma \ref{lem.Pvanishes}. This completes the proof. \end{pf} Using the map $\hat\Delta$ we will show a proof of the following lemma in the following section. \begin{lem}\label{lem.invertible} Let $U_+$ (resp. $U_-$) be the trivial knot with $+1$ (resp. $-1$) framing. Then $[\iota_n({\check Z}(U_\pm))]$ is invertible in ${\cal A}(\phi)/D_{>n}$, in which we define an algebra structure such that the disjoint union of two chord diagrams is the product of the diagrams. \end{lem} Using the above proposition and lemma, we obtain the following theorem. \begin{thm} Let $L$ be an oriented framed link, and $M$ the 3-manifold obtained by Dehn surgery on $S^3$ along $L$. Then the equivalence class $$ [\iota_n({\check Z}(U_+)]^{-{\sigma}_+} [\iota_n({\check Z}(U_-)]^{-{\sigma}_-} [\iota_n({\check Z}(L))] \in {\cal A}(\phi)/D_{>n} $$ is a topological invariant of $M$ for any positive integer $n$, where we denote by ${\sigma}_+$ (resp. ${\sigma}_-$) the number of positive (resp. negative) eigenvalues of the linking matrix of $L$. \end{thm} \begin{pf} We obtain invariance under both of orientation change of $L$ and Kirby move II by Proposition \ref{prop.invariant}. Note that we can apply Kirby move II in any way though we prove the proposition for Kirby move II preserving the orientation of $L$, because we also showed the invariance under orientation change. We also obtain invariance under Kirby move I, since the change of $[\iota_n {\check Z}(L)]$ under the move cancels with the change of ${\sigma}_{\pm}$. \end{pf} \begin{defn} We denote the above invariant by $\Omega_n(M)$. \end{defn} \section{A universal quantum invariant of 3-manifolds} In this section we unify the series $\Omega_n(M)$ into a invariant $\Omega(M)$. We show that the series $\Omega_n(M)$ satisfies a property, and that by the property the series has the same information (modulo the order of the first homology group) as $\Omega(M)$ has. We further show that the invariant $\Omega(M)$ satisfies a property derived from the above property of the series, and that there exists the logarithm of $\Omega(M)$ by the property; we denote by the logarithm a universal quantum invariant of 3-manifolds. \subsection{A group-like property of the series $\Omega_n(M)$} We denote by $\hat\D_{n_1,n_2}$ the map ${\cal A}(\phi)/D_{>n_1+n_2}\rightarrow {\cal A}(\phi)/D_{>n_1} \otimes {\cal A}(\phi)/D_{>n_2}$ naturally induced by $\hat\D$; note that this map is well defined since the degree is preserved by $\hat\D$, where we regard the sum of degrees as the degree in the tensor product. Using Theorem \ref{thm.bunshin} we obtain the following proposition. \begin{prop}\label{prop.bunshin} $$ \hat\D_{n_1,n_2}(\Omega_{n_1+n_2}(M)) = \Omega_{n_1}(M) \otimes \Omega_{n_2}(M). $$ \end{prop} \begin{pf} Noting that $\hat\D$ is an algebra homomorphism in ${\cal A}(\phi)$, this proposition is a direct conclusion of Theorem \ref{thm.bunshin} and the following lemma. \end{pf} \begin{lem}\label{lem.Dandiota} Let $n$, $n_1$ and $n_2$ be positive integers satisfying $n=n_1+n_2$. Then the following diagram is commutative; \begin{equation}\label{eq.Dandiota} \begin{CD} {\hoA{\coprod^l S^1}/L_{<2n},PO_n} @>{\iota_n}>> \hoA{\phi}/PO_n \\ @V{\hat\D_{n_1,n_2}}VV @VV{\hat\D}V \\ \hoA{\coprod^l S^1}/L_{<2n_1},PO_{n_1} \otimes \hoA{\coprod^l S^1}/L_{<2n_2},PO_{n_2} @>{\iota_{n_1} \otimes \iota_{n_2}}>> \hoA{\phi}/PO_{n_1} \otimes \hoA{\phi}/PO_{n_2} \end{CD} \end{equation} where we mean by $PO_k$ the equivalence relation generated by $P_{k+1}$ and $O_k$. \end{lem} \begin{pf} By Lemma \ref{lem.decreaselegs} we can assume that an element in $\hoA{\coprod^l S^1}/L_{<2n},PO_n$ is a disjoint union of dashed trivalent graphs and $l$ copies of a solid circle with $n$ dashed isolated chords; in fact, the space consists of linear sums of such elements. Since $\iota_{\star}$ is trivial for dashed trivalent graphs, (\ref{eq.Dandiota}) is commutative for them. Hence it is sufficient to show that (\ref{eq.Dandiota}) is commutative for a solid circle with $n$ dashed isolated chords; we put it to be $D_n$. The image of $D_n$ by the map $\iota_n$ is equal to the chord diagram obtained by attaching $n$ isolated chords to $T_{2n}^n$, by the definition of $\iota_n$. By the same argument in Figure \ref{fig.Pvanish}, $T_{2n}^n$ with one dashed isolated chords is equal to $T_{2n-2}^{n-1}$ times a sum of $2n-2$ and a dashed circle; the sum is equal to $-2$ with the relation $O_n$. Repeating this argument, we can show that the image is equal to $(-2)(-4)\cdots (-2n) = (-2)^n n!$. Therefore the clockwise image of $D_n$ in the diagram is equal to $(-2)^n n!$. The image of $D_n$ by the map $\hat\D$ is equal to the sum of $\binom{n}{k} D_k \otimes D_{n-k}$ by the definition of $\hat\D$. Note that it vanishes with $L_{<2n_1}$ or $L_{<2n_2}$ unless $k=n_1$. Hence the image by $\hat\D_{n_1,n_2}$ is equal to $\binom{n}{n_1} D_{n_1} \otimes D_{n_2}$. By the same argument as above, the image of it by the map $\iota_{n_1} \otimes \iota_{n_2}$ is equal to $\binom{n}{n_1} (-2)^{n_1} n_1! (-2)^{n_2} n_2! = (-2)^n n!$ using $n=n_1+n_2$; this coincides the above value. \end{pf} \begin{pf*}{Proof of Lemma \ref{lem.invertible}} Applying Theorem \ref{thm.bunshin} to ${\check Z}(U_{\pm})$, $n-1$ times, we have the formula $\hat\D^{(n-1)}({\check Z}(U_{\pm}))=({\check Z}(U_{\pm}))^{\otimes n}$, where we define the map $\hat\D^{(k)} : \hat{\cal A}(X) \to \hat{\cal A}(X)^{\otimes (k+1)}$ by $\hat\D^{(k)}= (\hat\D \otimes 1)\circ \hat\D^{(k-1)}$ recursively. Further, putting $X=\phi$, the map $\hat\D^{(k)}$ naturally induces the map of ${\cal A}(\phi)/D_{>k+1}$ to $({\cal A}(\phi)/D_{>1})^{\otimes (k+1)}$; we denote it by $\hat\D^{(k)}_{1,1,\cdots,1}$. Applying Lemma \ref{lem.Dandiota} to the above formula $n-1$ times, we obtain the first equality of the following formula; $$ \hat\D^{(n-1)}_{1,1,\cdots,1}([\iota_n({\check Z}(U_{\pm}))]) =(\iota_1({\check Z}(U_{\pm})))^{\otimes n} = (\mp 1 + \frac{\theta}{16})^{\otimes n} \in ({\cal A}(\phi)/D_{>1})^{\otimes n}. $$ Here the second equality is derived from Lemma \ref{lem.iota1U} below. Therefore the constant term of $[\iota_n({\check Z}(U_{\pm}))]$ does not vanish, which means that it is invertible. \end{pf*} We have the following lemma proved in \cite{LMMOII}. \begin{lem}[\cite{LMMOII}]\label{lem.iota1U} The following formula holds; $$ [\iota_1({\check Z}(U_{\pm}))]= \mp 1 + \frac{\theta}{16} \in {\cal A}(\phi)/D_{>1} $$ where we mean by $\theta$ the dashed trivalent graph consisting of three edges and two vertices, as the Greek character $\theta$. \end{lem} \begin{rem} In fact, the above values of ${\check Z}(U_{\pm})$ are $-2$ times the values in \cite{LMMOII}. The difference occurs from the definition of the map $\iota_1$; it was defined by simply removing $\Theta$ components in \cite{LMMOII}, where we mean by $\Theta$ a solid circle with one dashed line as the Greek letter $\Theta$. On the other hand, we define $\iota_1$ here in the way how we replace $\Theta$ with the trivial dashed circle, which is equivalent to $-2$. \end{rem} \subsection{A power series invariant $\Omega(M)$} We define a map $\varepsilon : {\cal A}(\phi) \to {\Bbb C}$ to be the projection to the degree $0$ part of a linear sum of chord diagrams; recall that we regard the empty diagram as $1$ whose degree is $0$. By the definitions of $\hat\D_{1,n-1}$ and $\varepsilon$, we immediately obtain the following lemma. \begin{lem}\label{lem.contraction} The following formula holds; $$ (\varepsilon \otimes 1)\circ \hat\D_{1,n-1} = p_{n,n-1} $$ where we mean by $1$ the identity map and we denote by $p_{n,n-1}$ the projection of ${\cal A}(\phi)/D_{>n}$ to ${\cal A}(\phi)/D_{>n-1}$. \end{lem} \begin{lem}\label{lem.unifyOmega} Let $(\Omega_1,\Omega_2,\cdots)$ be a series of $\Omega_n \in {\cal A}(\phi)/D_{>n}$ which satisfies \begin{equation}\label{eq.split} \hat\D_{n_1,n_2}(\Omega_{n_1+n_2})=\Omega_{n_1} \otimes \Omega_{n_2} \end{equation} for any positive integers $n_1$ and $n_2$. \begin{enumerate} \item Then the formula $\Omega_n^{(d)} = m^{n-d} \Omega_d^{(d)}$ holds for $d<n$, where we denote by $\alpha^{(d)}$ the degree $d$ part of $\alpha$ and we put $m$ to be $\Omega_1^{(0)}$. \item Further, for $\Omega \in \hat{\cal A}(\phi)$ defined to be $1+ \sum_{n=1}^{\infty} \Omega_n^{(n)}$, the formula $\hat\D(\Omega)=\Omega \otimes \Omega$ holds; recall that we denote by $\hat{\cal A}(\phi)$ the completion of ${\cal A}(\phi)$ with respect to the degree, that is, $\hat{\cal A}(\phi)$ consists of infinite linear sums of chord diagrams. We also denote by $\hat\D$ the natural extension of $\hat\D$ to $\hat{\cal A}(\phi)$. \end{enumerate} \end{lem} \begin{pf} We apply the map in Lemma \ref{lem.contraction} to $\Omega_n$. Then we have the left hand side as; $$ (\varepsilon \otimes 1)\circ \hat\D_{1,n-1} (\Omega_n) = (\varepsilon \otimes 1) (\Omega_1 \otimes \Omega_{n-1}) = m \Omega_{n-1}. $$ Hence we obtain (1). We put $\Omega_0$ to be $1$, then the series $(\Omega_0,\Omega_1,\cdots)$ satisfies (\ref{eq.split}) for any non-negative integers $n_1$ and $n_2$. It is sufficient to show the required formula of (2) in each degree $n$. Hence we will show the formula $\hat\D(\Omega_n^{(n)}) = \sum_{k_1+k_2=n} \Omega^{k_1} \otimes \Omega^{k_2}$. Though this is a formula in $\hat{\cal A}(\phi)\otimes\hat{\cal A}(\phi)$, it consists only of degree $n$ part. Therefore it suffices to show it for the image of it by the map $p_{\infty,n_1} \otimes p_{\infty,n_2}$ for each pair $n_1$ and $n_2$ with $n_1+n_2=n$, where we denote by $p_{\infty,k}$ the projection of $\hat{\cal A}(\phi)$ to ${\cal A}(\phi)/D_{>k}$. The image becomes $\hat\D_{n_1,n_2}(\Omega_n^{(n)})= \Omega_{n_1}^{(n_1)} \otimes \Omega_{n_2}^{(n_2)}$ which is a special case of (\ref{eq.split}). This completes the proof. \end{pf} We have the following lemma by results in \cite{LMMO} where only the invariant $\Omega_1(M)$ was discussed. \begin{lem}\label{lem.degreezero} The degree zero part of $\Omega_1(M)$ is equal to $|H_1(M,{\Bbb Z})|$ if $M$ is a rational homology 3-sphere, $0$ otherwise. Here we mean by $|\cdot|$ the order of the set. \end{lem} We know that we can apply Lemma \ref{lem.unifyOmega} to our series of invariants $\Omega_n(M)$ by Proposition \ref{prop.bunshin}, to reduce the series to $\Omega \in \hat{\cal A}(\phi)$ and a scalar $m$. In our case the scalar $m$ becomes either the order of the first homology group or zero by Lemma \ref{lem.degreezero}. Then we obtain the following definition and proposition from Lemma \ref{lem.unifyOmega} (2). \begin{defn} We define a topological invariant $\Omega(M)$ of a 3-manifold $M$ by $\Omega(M) = 1 + \sum_{n=1}^{\infty} \Omega_n(M)^{(n)} \in \hat{\cal A}(\phi)$. \end{defn} \begin{prop}\label{prop.Omegasplit} The invariant $\Omega(M)$ satisfies $\hat\D(\Omega(M))=\Omega(M) \otimes \Omega(M)$. \end{prop} \subsection{Logarithm of $\Omega(M)$} We denote by ${\cal A}(\phi)_{\rm conn}$ the vector subspace of ${\cal A}(\phi)$ spanned by the set of connected non-empty dashed trivalent graphs with oriented trivalent vertices subject to AS and IHX relations. We put $\hat{\cal A}(\phi)_{\rm conn}$ to be the completion of it with respect to the degree, which becomes a vector subspace of $\hat{\cal A}(\phi)$. It is well known as a property of Hopf algebra, see for example \cite{Hopfalgebra}, that a non-zero element $\Omega \in \hat{\cal A}(\phi)$ is group-like if and only if there exists a primitive element $\omega \in \hat{\cal A}(\phi)$ satisfying $\Omega = \exp(\omega) = 1+\omega+(1/2)\omega^2+\cdots$, where we call $\alpha$ {\it group-like} if it satisfies $\hat\D(\alpha)= \alpha \otimes \alpha$, and call $\alpha$ {\it primitive} if it satisfies $\hat\D(\alpha)=\alpha \otimes 1 + 1 \otimes \alpha$. In our case $\alpha \in \hat{\cal A}(\phi)$ is primitive if and only if $\alpha$ belongs to $\hat{\cal A}(\phi)_{\rm conn}$. Note that $\omega$ is uniquely determined for given $\Omega$, since a primitive element always has a positive degree. Since $\Omega(M)$ is group-like by Proposition \ref{prop.Omegasplit}, we have the following definition. \begin{defn} We define $\omega(M) \in \hat{\cal A}(\phi)_{\rm conn}$ by $\exp(\omega(M))=\Omega(M)$. We call $\omega(M)$ a universal quantum invariant. \end{defn} \section{Properties of the universal quantum invariant $\omega(M)$} We will see some properties of $\omega(M)$ in this section. \subsection{Formulas for connected sum and opposite orientation} \begin{prop} Let $M$ be the connected sum of two closed 3-manifolds $M_1$ and $M_2$, then $\omega(M)$ is given by $$ \omega(M) = \sum_{d=1}^{\infty} \left( m_2^{d} \omega(M_1)^{(d)} + m_1^{d} \omega(M_2)^{(d)} \right), $$ where we put $m_i$ ($i=1,2$) to be the order of $H_1(M_i,{\Bbb Z})$ if $M_i$ is a rational homology 3-sphere, $0$ otherwise. \end{prop} \begin{pf} Let $L$ be the framed link obtained by taking split union of two framed links $L_1$ and $L_2$. By definition of ${\check Z}(L)$ we have ${\check Z}(L)$ as the disjoint union of ${\check Z}(L_1)$ and ${\check Z}(L_2)$. Note that, if $M_1$ and $M_2$ are obtained by Dehn surgery along $L_1$ and $L_2$ respectively, then $M$ is obtained from $L$. By definition of $\Omega_n(M)$ we have $\Omega_n(M)=\Omega_n(M_1)\Omega_n(M_2)$, recall that we define multiplication by disjoint union of chord diagrams. By Lemmas \ref{lem.unifyOmega} and \ref{lem.degreezero} we have \begin{align*} \Omega_n(M)^{(n)} &= \sum_{d_1+d_2=n} \Omega_n(M_1)^{(d_1)}\Omega_n(M_2)^{(d_2)} \\ &= \sum_{d_1+d_2=n} m_1^{d_2}\Omega_{d_1}(M_1)^{(d_1)} m_2^{d_1}\Omega_{d_2}(M_1)^{(d_2)}. \end{align*} Hence we have $$ \Omega(M) = \sum_{d_1,d_2=0}^{\infty} m_1^{d_2} \Omega(M_1)^{(d_1)} m_2^{d_1} \Omega(M_2)^{(d_2)}. $$ If both of $M_1$ and $M_2$ are rational homology 3-spheres, the above formula implies that $\sum_d \Omega(M_i)^{(d)}/m_i^d$ is multiplicative with respect to connected sum. Hence $\sum_d \omega(M_i)^{(d)}/m_i^d$ is additive, and we obtain the required formula. If either of $M_1$ and $M_2$, say $M_1$, is not a rational homology 3-sphere, then we have $m_1=0$. Hence $\Omega(M)=\sum_d \Omega(M_1)^{(d)} m_2^d$ by the above formula putting $d_2=0$, where we regard $0^0$ as $1$. Therefore we obtain the required formula, completing the proof. \end{pf} \begin{prop} Let $-M$ be a 3-manifold $M$ with opposite orientation. Then the following formula holds; $$ \omega(-M) = \sum_{d=1}^{\infty} (-1)^d \omega(M)^{(d)}. $$ \end{prop} \begin{pf} We define a map $\hat S:{\cal A}(X) \to {\cal A}(X)$ by putting $\hat S(D)=(-1)^d D$ for a chord diagram $D$ where $d$ is the degree of $D$. For the mirror image $\bar L$ of a framed link $L$, we have ${\check Z}(\bar L) = \hat S({\check Z}(L))$; we can show the formula by checking it for each elementary tangles. Since $-M$ is obtained by Dehn surgery along $\bar L$, we have $\Omega_n(-M)=\hat S(\Omega_n(M))$, $\Omega(-M)=\hat S(\Omega(M))$ and $\omega(-M)=\hat S(\omega(M))$ by their definitions. The last formula is the required one. \end{pf} \subsection{The first term in $\omega(M)$} By results in \cite{LMMOII} we have the following proposition. \begin{prop} Let $M$ be a oriented closed 3-manifold. Then the coefficient of dashed $\theta$-curve in $\omega(M)$ is equal to $(-1)^{b_1(M)+1}3 \tilde\lambda (M)$ where $b_1(M)$ is the first Betti number of $M$ and we denote $\tilde\lambda(M)$ twice Lescop's generalization \cite{Lescop} of the Casson-Walker invariant (Walker's normalization, which is twice Casson's normalization) $\lambda(M)$ \cite{AkbulutMcCarthy,Walker} satisfying $\tilde\lambda(M)=|H_1(M;{\Bbb Z})|\lambda(M)$ if $M$ is a rational homology 3-sphere. \end{prop} \begin{pf} In \cite{LMMOII} it is shown that the degree $1$ part of $\Omega_1(M)$ is equal to $(-1)^{b_1(M)+1}3 \tilde\lambda (M)$. The proposition immediately follows from that. \end{pf} \begin{rem} We can \lq\lq substitute'' a Lie algebra into dashed lines, see \cite{BarNatanII}. When we substitute $sl_N$, $so_N$ and $sp_N$ to dashed $\theta$-curve, we obtain values $2N(N^2-1)$, $N(N-1)(N-2)/2$ and $2N(N+1)(2N+1)$ respectively. When we substitute $sl_2$ into the first term (i.e. dashed $\theta$-curve and its coefficient) of $\omega(M)$, we have the value $12$ times the coefficient. Hence, if one expects the existence of \lq\lq $G$ Casson invariant $\lambda^G(M)$'' and expects that it should recover from the first term in $\omega(M)$ by substituting the Lie algebra of $G$ into dashed lines, it must satisfy \begin{align*} \lambda^{SU(N)}(M) &= \frac{N(N^2-1)}{6} \lambda(M), \\ \lambda^{SO(N)}(M) &= \frac{N(N-1)(N-2)}{24} \lambda(M), \\ \lambda^{Sp(N)}(M) &= \frac{N(N+1)(2N+1)}{6} \lambda(M). \end{align*} In \cite{Ohtsuki} we have the same formula for $\lambda^{SU(N)}(M)$ of each lens space $M$, which is obtained expanding quantum $PSU(N)$ invariant of lens spaces obtained in \cite{Takata} into a power series in $q-1$. \end{rem}

https://arxiv.org/abs/1410.2841

A convex solution to Psiaki's first joint attitude and spin-rate estimation problem

We consider the problem of jointly estimating the attitude and spin-rate of a spinning spacecraft. Psiaki (J. Astronautical Sci., 57(1-2):73--92, 2009) has formulated a family of optimization problems that generalize the classical least-squares attitude estimation problem, known as Wahba's problem, to the case of a spinning spacecraft. If the rotation axis is fixed and known, but the spin-rate is unknown (such as for nutation-damped spin-stabilized spacecraft) we show that Psiaki's problem can be reformulated exactly as a type of tractable convex optimization problem called a semidefinite optimization problem. This reformulation allows us to globally solve the problem using standard numerical routines for semidefinite optimization. It also provides a natural semidefinite relaxation-based approach to more complicated variations on the problem.

\section{Introduction} \label{sec:intro} Spacecraft attitude estimation is a fundamental problem, arising, for instance, as a natural subproblem whenever attitude control is required. Since spacecraft dynamics are non-linear, a typical and successful approach to attitude estimation is to employ variants of the Extended Kalman Filter (EKF)~\cite{lefferts1982kalman}. As with any method based on linearization of non-linear dynamics, EKF-based approaches can fail to converge given poor initial estimates, and can become unstable in the presence of large disturbances~\cite{psiaki2000attitude}. Many truly non-linear attitude estimation methods have also been proposed (see~\cite{crassidis2007survey} for a survey). An important example is the static least-squares attitude estimation problem known as Wahba's problem~\cite{wahba1965least}. In Wahba's problem we are simultaneously given a batch of vector measurements (from sun sensors, star trackers, etc.) in the body frame and corresponding reference directions in an inertial frame. The aim is to find the rotation matrix (i.e.\ direction cosine matrix) that minimizes the sum of the squared errors between the transformed reference directions and the observed vector measurements. Wahba's problem, as stated, applies most naturally to a static spacecraft. Nevertheless, it has also found use as a subroutine in various recursive estimation algorithms including those that estimate the full dynamical state of the spacecraft (see, e.g., \cite{psiaki2000attitude,gebre2000gyro}). Recently Psiaki has posed a number of generalizations of Wahba's problem to the case of a spinning spacecraft~\cite{psiaki2009generalized}. These problems aim to simultaneously estimate the initial attitude and spin-rate (or, more generally, initial angular momentum) of the spacecraft from vector measurements, without the need for gyroscope measurements. These generalizations are particularly suited to spin-stabilized spacecraft without gyroscopes. We describe Wahba's problem and Psiaki's generalizations formally in Section~\ref{sec:jas}. In this paper we focus on the simplest of Psiaki's generalizations of Wahba's problem. We refer to this problem as \emph{Psiaki's first problem}. In this problem we assume the spacecraft is spinning at a constant unknown angular velocity around a known (stable) inertia axis. This setting is relevant for nutation-damped spin-stabilized spacecraft~\cite{psiaki2009generalized}. The aim is to estimate the initial attitude and the unknown spin-rate given a sequence of noisy vector measurements obtained at certain sampling instants, together with corresponding reference directions. Wahba's problem arises as the special case where the spin-rate is zero. \subsection{Main contribution} \label{sec:contribtions} Our main contribution is to show that, when the sampling period is constant, Psiaki's first problem can be reformulated exactly as a semidefinite optimization problem (see Theorem~\ref{thm:mainsdp}). Semidefinite optimization problems (described in Section~\ref{sec:semidefinite}) are a family of convex optimization problems that generalize linear programming and can be solved globally with provable efficiency guarantees using standard software. Reformulating Psiaki's first problem as a semidefinite optimization problem means that it, like Wahba's problem, can be solved efficiently and globally, to high precision, using numerical methods. A description of Psiaki's first problem as the solution to a semidefinite optimization problem allows us to do more than just solve the original problem as stated. It also allows us to take a semidefinite relaxation-based approach to many variants on Psiaki's problem. We illustrate this in Section~\ref{sec:variations} by considering the example of a version of Psiaki's first problem where explicit bounds on the measurement errors are incorporated into the formulation. \subsection{Organization of the paper} \label{sec:outline} The remainder of the paper is organized as follows. In Section~\ref{sec:notation} we summarize notation not defined elsewhere in the paper. In Section~\ref{sec:jas} we first describe Psiaki's generalizations of Wahba's problem for spinning spacecraft. We then show how to write Psiaki's first problem as an instance of a family of problems we call \emph{\text{trigonometric Wahba problems}{}} (see~\eqref{eq:genform1}). We conclude the section with a summary of prior work on Psiaki's problems. In Section~\ref{sec:semidefinite} we briefly describe semidefinite optimization problems in general before presenting our semidefinite optimization-based reformulation of \text{trigonometric Wahba problems}{}, and in particular of Psiaki's first problem. We defer the proofs to the Appendix. In Section~\ref{sec:variations} we describe a variant on Psiaki's first problem that incorporates additional bounds on the measurement noise (if they are available) and show how to extend our semidefinite optimization-based reformulation of Psiaki's first problem to a semidefinite relaxation of this variant. We also describe the results of a simple numerical experiment comparing Psiaki's first problem and this variant. In Section~\ref{sec:directions} we discuss possible future research related to the work in this paper. \subsection{Notation} \label{sec:notation} We briefly summarize notation used throughout the body of the paper. Additional notation that is used only in the Appendix is introduced separately there. \paragraph{Spaces} Denote by $\mathbb{R}^{n\times n}$ the space of $n\times n$ real matrices. If $X\in \mathbb{R}^{n\times n}$ let $X^T$ be its transpose. Let $\mathcal{S}^n$ be the space of $n\times n$ symmetric matrices (i.e.\ matrices for which $X = X^T$). Let $\mathcal{S}_+^n$ denote the set of $n\times n$ symmetric positive semidefinite matrices (i.e.\ $X\in \mathcal{S}_+^n$ if and only if $u^TXu\geq 0$ for all $u\in \mathbb{R}^n$). If $X\in \mathcal{S}_+^n$ we write $X \succeq 0$ when the dimension is clear from the context. \paragraph{Inner products} If $x,y\in \mathbb{R}^n$ then $\langle x,y\rangle = \sum_{i=1}^{n}x_i y_i$. If $x\in \mathbb{R}^n$ then define $\|x\| = \langle x,x\rangle^{1/2} = \left(\sum_{i=1}^{n}x_i^2\right)^{1/2}$ to be the usual Euclidean norm. If $X,Y\in \mathbb{R}^{n\times n}$ then define an inner product by $\langle X,Y\rangle = \textup{tr}(X^TY) = \sum_{i,j=1}^{n}X_{ij}Y_{ij}$. \paragraph{Convexity} Given a subset $S\subset \mathbb{R}^n$ then \[ \conv(S) = \left\{\sum_{i}\lambda_i x_i:\; \sum_{i}\lambda_i = 1, \;\;x_i\in S,\; \lambda_i \geq 0,\;\text{for all $i$}\right\}\] is the set of all convex combinations of elements of $S$. From the point of view of optimization, if $c\in \mathbb{R}^n$ and $S$ is compact then \[ \max_{x\in S} \langle c,x\rangle = \max_{x\in \conv(S)} \langle c,x\rangle\] so the optimal cost is the same whether we optimize the linear functional defined by $c$ over $S$ or over its convex hull~\cite[Theorem 32.2]{rockafellar1997convex}. \paragraph{Block matrices} If $T_0,T_1,\ldots,T_{N}$ are $d\times d$ matrices with $T_0$ being symmetric, define the corresponding $d(N+1)\times d(N+1)$ symmetric block Toeplitz matrix by \begin{equation} \label{eq:blktoep} \textup{Toeplitz}(T_0,T_1,\ldots,T_N) = \begin{bmatrix} T_0 & T_1 & T_2 & \cdots & T_N\\ T_1^T & T_0 & T_1 & \ddots & \vdots\\ T_2^T & T_1^T & \ddots & \ddots & \vdots\\ \vdots & \ddots & \ddots & \ddots & T_1\\ T_N^T & \cdots & \cdots & T_1^T & T_0\end{bmatrix}. \end{equation} Similarly if $S_1,S_2,\ldots,S_{2N+1}$ are symmetric $d\times d$, matrices define the corresponding $d(N+1)\times d(N+1)$ block Hankel matrix by \begin{equation} \label{eq:blkhank} \textup{Hankel}(S_1,S_2,\ldots,S_{2N+1}) = \begin{bmatrix} S_1 & S_2 & \cdots & S_{N} & S_{N+1}\\ S_2 & & & S_{N+1} & S_{N+2}\\ \vdots & & \iddots & & \vdots\\ S_{N} & S_{N+1} & & & S_{2N}\\ S_{N+1} & S_{N+2} & \cdots & S_{2N} & S_{2N+1}\end{bmatrix}. \end{equation} \paragraph{Unit quaternion parameterization of rotations} We make extensive use of the quadratic parameterization of $SO(3)$, the set of rotation (or direction-cosine) matrices, by unit quaternions, denoted by $\H$. Throughout we think of the unit quaternions geometrically as the unit sphere in $\mathbb{R}^4$ i.e.\ $\H = \{q\in \mathbb{R}^4:\|q\|=1\}$. We only ever work with a unit quaternion $q\in\H$ via the positive semidefinite matrix $qq^T$, avoiding the sign ambiguity that would arise if we were to try to work directly with variables $q\in \H$. It is enough only to consider $qq^T$ because any element of $SO(3)$ can be expressed as $\mathcal{A}(qq^T)$ where $q\in \H$ and $\mathcal{A}:\mathcal{S}^4\rightarrow \mathbb{R}^{3\times 3}$ is the linear map defined (following the convention in~\cite{crassidis2007survey}) by \begin{equation} \label{eq:amap} \mathcal{A}(Z) := \begin{bmatrix} Z_{11} - Z_{22} - Z_{33} + Z_{44} & 2Z_{12} + 2Z_{34} & 2Z_{13} - 2Z_{24}\\2Z_{12} - 2Z_{34} & -Z_{11} + Z_{22} - Z_{33} + Z_{44} & 2Z_{23}+2Z_{14}\\ 2Z_{13} + 2Z_{24} & 2Z_{23} - 2Z_{14} & -Z_{11} - Z_{22} + Z_{33} + Z_{44}\end{bmatrix}. \end{equation} The adjoint of $\mathcal{A}$ (with respect to the inner product on matrices) is $\mathcal{A}^*:\mathbb{R}^{3\times 3}\rightarrow \mathcal{S}^4$ defined by \begin{equation} \label{eq:amapstar} \mathcal{A}^*(Y) := \begin{bmatrix} Y_{11}-Y_{22}-Y_{33} & Y_{12} + Y_{21} & Y_{13}+Y_{31} & Y_{23}-Y_{32}\\ Y_{12}+Y_{21} & -Y_{11}+Y_{22} - Y_{33} & Y_{23}+Y_{32}& -Y_{13} + Y_{31}\\ Y_{13}+Y_{31} & Y_{23} + Y_{32} & -Y_{11} - Y_{22} + Y_{33} & Y_{12}-Y_{21}\\ Y_{23} - Y_{32} & -Y_{13}+Y_{31} & Y_{12}-Y_{21} & Y_{11} + Y_{22} + Y_{33}\end{bmatrix}. \end{equation} In other words for any $Z\in \mathcal{S}^4$ and any $Y\in \mathbb{R}^{3\times 3}$, we have the identity \begin{equation} \label{eq:adjoint} \langle \mathcal{A}(Z),Y\rangle = \langle Z,\mathcal{A}^*(Y)\rangle. \end{equation} \section{Psiaki's generalizations of Wahba's problem for spinning spacecraft} \label{sec:jas} In this section we describe Wahba's problem~\cite{wahba1965least} and Psiaki's generalizations to the case of a spinning spacecraft~\cite{psiaki2009generalized}. For reasons discussed in Section~\ref{sec:psiaki-form} we subsequently focus on the simplest of Psiaki's problems: jointly estimating the attitude and spin-rate of a spacecraft spinning around a stable inertia axis at a constant unknown rate. In this case we show how to reformulate the resulting optimization problem in the general form \begin{equation} \label{eq:genform1}\max_{\substack{Q\in SO(3)\\\omega\in [-\pi,\pi)}} \langle A_0,Q\rangle + \sum_{n=1}^{N}\left[ \langle A_n,\cos(\omega n)Q\rangle + \langle B_n,\sin(\omega n)Q\rangle\right] \end{equation} for appropriate collections of $3\times 3$ matrices $(A_n)_{n=0}^{N}$ and $(B_n)_{n=1}^{N}$. Throughout, we call problems in the form~\eqref{eq:genform1} \emph{\text{trigonometric Wahba problems}{}}. In Section~\ref{sec:semidefinite} to follow, we show how to reformulate \text{trigonometric Wahba problems}{} as semidefinite optimization problems. \subsection{Wahba's problem} \label{sec:wahba} We briefly describe Wahba's least squares attitude estimation problem posed in~\cite{wahba1965least} with solutions published in~\cite{farrell1966least}. \paragraph{Vector measurements} Suppose we are given a batch of noisy unit vector measurements $y_0,y_1,\ldots,y_N$ in the body frame (obtained from star trackers, sun sensors, magnetometers, etc.) of corresponding unit reference directions $x_0,x_1,\ldots,x_N$ in the inertial frame. \paragraph{Least squares objective} Wahba's problem is to find the rotation matrix $Q\in SO(3)$ that transforms the reference directions to best fit the measured vector measurements in the weighted least squares sense by solving \begin{equation} \label{eq:wahba0} \min_{Q\in SO(3)}\; \sum_{n=0}^{N}\frac{\kappa_n}{2}\|y_n - Qx_n\|^2 \end{equation} where $\kappa_0,\kappa_1,\ldots,\kappa_N$ are non-negative scalar weights that one would take to be larger for measurements with smaller noise variance. Since $\|Qx\|^2 = \|x\|^2$ for all $x\in \mathbb{R}^3$ we can expand the squares and see that this optimization problem is equivalent to \begin{equation} \label{eq:wahba} \max_{Q\in SO(3)}\; \langle \sum_{n=0}^{N}\kappa_n y_nx_n^T,Q\rangle \end{equation} where we have dropped an additive constant of $\sum_{n=0}^{N}\frac{\kappa_n}{2}(\|y_n\|^2+\|x_n\|^2)$. \subsection{Psiaki's generalizations} \label{sec:psiaki-form} We now describe Psiaki's generalizations of Wahba's problem, and show how Wahba's problem arises as a special case. \paragraph{Rigid body (Euler) equations} Let $Q(t_0)\in SO(3)$ denote the initial attitude of the spacecraft, $\Omega(t_0)\in \mathbb{R}^3$ the initial body angular velocity, and $I_1\geq I_2\geq I_3$ the principal moments of inertia. Assuming the spacecraft undergoes torque-free motion about its centre of mass then for $t\geq t_0$ the attitude $Q(t)$ and the body angular velocity $\Omega(t) := \begin{bmatrix} \omega_1(t) & \omega_2(t) & \omega_3(t)\end{bmatrix}^T$ satisfy the rigid body equations: \begin{equation} \label{eq:ode} \begin{matrix} I_1\;\dot{\omega}_1(t) = (I_2-I_3)\,\omega_2(t)\,\omega_3(t)\\ I_2\;\dot{\omega}_2(t) = (I_3-I_1)\,\omega_3(t)\,\omega_1(t)\\ I_3\;\dot{\omega}_3(t) = (I_1-I_2)\,\omega_1(t)\,\omega_2(t)\\ \end{matrix}\quad\text{and}\quad \dot{Q}(t) = \begin{bmatrix} 0 & -\omega_3(t) & \phantom{-}\omega_2(t)\\ \phantom{-}\omega_3(t) & 0 & -\omega_1(t)\\-\omega_2(t) & \phantom{-}\omega_1(t) & 0\end{bmatrix}Q(t). \end{equation} Note that for every $t \geq t_0$ and every $\Omega(t_0)$ we have that $Q(t) = \Phi(t-t_0;\Omega(t_0))Q(t_0)$ for some map $\Phi$ taking values in $SO(3)$. In particular $Q(t)$ is always linear in the initial attitude $Q(t_0)$. \paragraph{Vector measurements} Let $t_0,t_1,\ldots,t_N$ be a finite set of sampling instants. Assume, at sample instant $t_n$, that we are given a noisy unit vector measurement $y_n$ in the spacecraft body frame of a corresponding reference directions $x_n$ in the inertial frame. \paragraph{Least squares objective} Following Wahba's least-squares-based objective, Psiaki suggests solving the following weighted least-squares problem to estimate the initial attitude and body angular velocity of the spacecraft, given only the vector measurements $(y_n)_{n=0}^{N}$ and the reference directions $(x_n)_{n=0}^{N}$: \begin{equation} \label{eq:psiaki-lsq} \min_{Q(t_0),\Omega(t_0)}\sum_{n=0}^{N}\frac{\kappa_n}{2}\|y_n - Q(t_n)x_n\|_2^2\end{equation} subject to $Q(t)$ satisfying~\eqref{eq:ode} with initial conditions $Q(t_0)$ and $\Omega(t_0)$. Just as for Wahba's problem, the $\kappa_n$ are non-negative scalars. \paragraph{Dependence on $\Omega(t_0)$} In general, the dependence of $Q(t)$ on the initial body angular velocity $\Omega(t_0)$ is quite complicated. The relationship between $Q(t)$ and $\Omega(t_0)$ simplifies under additional assumptions on $\Omega(t_0)$ and the inertia tensor of the spacecraft. We now summarize these simplified problems and name them for later reference. \begin{description} \item[Wahba's problem] If $\Omega(t_0) = 0$, then $Q(t) = Q(t_0)$ for all $t\geq t_0$ and so the spacecraft is stationary. Adding this as a constraint we recover Wahba's original formulation~\eqref{eq:wahba0}. \item[Psiaki's first problem] Suppose $\Omega(t_0)$ is aligned with the major inertia axis, and (without loss of generality) this is the first axis direction in body coordinates. Then $\Omega(t_0)= \begin{bmatrix}\omega & 0 & 0\end{bmatrix}^T$ and so the dynamical constraints~\eqref{eq:ode} reduce to \begin{equation} \label{eq:spineq} Q(t) = \begin{bmatrix}1 & 0 & 0\\0 & \cos(\omega t) & -\sin(\omega t)\\0 & \sin(\omega t) & \cos(\omega t) \end{bmatrix}Q(t_0) \end{equation} where $\omega$ is the spin-rate (in rad/second). In this case the spacecraft is spinning with an \emph{unknown} constant angular velocity $\omega$ around a \emph{known} axis (fixed in body coordinates). Minimizing the least-squares objective~\eqref{eq:psiaki-lsq} subject to the constraints~\eqref{eq:spineq} is the first generalization of Wahba's problem posed in~\cite{psiaki2009generalized}, and is relevant for a nutation damped spin-stabilized spacecraft. \item[Psiaki's second problem] If $\Omega(t_0)$ is unconstrained and no additional assumptions are made about the moments of inertia of the spacecraft, we obtain the second generalization of Wahba's problem posed in~\cite{psiaki2009generalized}. In this setting the dependence of $Q(t)$ on $\Omega(t_0)$ is more complicated. This case is discussed further in~\cite{psiaki2012numerical} (see Section~\ref{sec:related} to follow). \end{description} In each case, Psiaki's formulations involve solving non-convex optimization problems of the form in~\eqref{eq:psiaki-lsq} subject to dynamical constraints. \paragraph{Focus of the paper} For the remainder of the paper we focus on Psiaki's first problem, because in this case $Q(t)$ only depends on the initial body angular velocity through $\cos(\omega t)$ and $\sin(\omega t)$. In addition to focusing on Psiaki's first problem, we also assume that the sampling instants $t_0,t_1,\ldots,t_N$ are equally spaced. As such we assume there is some $\tau$ such that $t_n = n\tau$ for $n=0,1,\ldots,N$. This paper does not address Psiaki's more general second problem, where the dependence of $Q(t)$ on $\Omega(t_0)$ is significantly more complicated. It would be very interesting if the techniques we develop can be extended to this more general situation. \paragraph{Aliasing} Since we only observe $\omega$ via vector measurements at time instants that are integer multiples of $\tau$, from the data alone we cannot distinguish between spin rates at different integer multiples of $2\pi/\tau$ due to aliasing. Hence we assume that $\omega\in [-\pi/\tau,\pi/\tau)$ so that it is possible to determine the unknown spin-rate from the data. (We could, alternatively, fix some $a$ rad/second and assume $\omega \in [a,a+2\pi/\tau)$.) In a Bayesian formulation of the problem, we could interpret this as encoding prior information on the spin rate. \paragraph{Reformulation} We now reformulate Psiaki's first problem as a \text{trigonometric Wahba problem}{}. Since $\|Q(t)x_n\|^2 = \|x_n\|^2$ for all $t$ and $n$, observe that with $t_n = n\tau$ the optimization problem~\eqref{eq:psiaki-lsq} can be rewritten as \begin{align} \label{eq:inter} \min_{\substack{Q(0)\in SO(3)\\\omega\in [-\pi/\tau,\pi/\tau)}}&\; \sum_{n=0}^{N}\frac{\kappa_n}{2}[\|y_n\|^2 - 2\langle y_n,Q(n\tau)x_n\rangle + \|x_n\|^2]\\ \text{s.t.}\quad& Q(n\tau) = \begin{bmatrix} 1 & 0 & 0\\0 & \cos(n\tau\omega) & -\sin(n\tau\omega)\\ 0 & \sin(n\tau\omega) & \cos(n\tau\omega)\end{bmatrix}Q(0). \label{eq:spin2} \end{align} Putting $\omega' = \tau\omega$, we see that this is equivalent, as an optimization problem, to \begin{equation} \label{eq:psiaki1} \max_{\substack{Q\in SO(3)\\\omega'\in [-\pi,\pi)}} \langle A_0,Q\rangle + \sum_{n=1}^{N}\left[\langle A_n,\cos(n\omega')Q\rangle + \langle B_n,\sin(n\omega')Q\rangle \right] \end{equation} where \begin{equation} \label{eq:A0first} A_0 = \kappa_0\,y_0 x_0^T + \begin{bmatrix} 1 & 0 & 0\\0 & 0 & 0\\0 & 0 & 0\end{bmatrix}\left(\sum_{n=1}^{N}\kappa_n\,y_n x_n^T\right) \end{equation} and for $n=1,2,\ldots,N$, \begin{equation} \label{eq:AnBn} A_n = \kappa_n\begin{bmatrix} 0 & 0 & 0\\0 & 1 & 0\\0 & 0 & 1\end{bmatrix}y_n x_n^T\quad\text{and}\quad B_n = \kappa_n\begin{bmatrix} 0 & 0 & 0\\0 & 0 & 1\\0 & -1 & 0\end{bmatrix}y_n x_n^T. \end{equation} We have now expressed Psiaki's first problem in the general form described in~\eqref{eq:genform1}. \subsection{Prior work and alternative solution methods for Psiaki's problems} \label{sec:related} In this section we summarize previous approaches to Psiaki's generalizations of Wahba's problem for spinning spacecraft. We then briefly discuss a simple discretization-based approach, implicit in the work of Psiaki and Hinks~\cite{psiaki2012numerical}, for solving Psiaki's problems globally. Psiaki's original paper~\cite{psiaki2009generalized} describes a method to globally solve Psiaki's first problem when two noise-free vector measurements (sampled at distinct times) are used. In this situation the problem reduces to finding all the solutions of the corresponding non-linear equations satisfied by the initial attitude and spin-rate. This method seems quite sensitive to measurement noise, and is unable to exploit additional measurements to mitigate the effects of noise. (The advantages of incorporating multiple measurements are demonstrated in Section~\ref{sec:expts}.) In subsequent work~\cite{hinks2011solution} Hinks and Psiaki describe an approach to Psiaki's second problem under the assumption that the spacecraft is axially symmetric and exactly three noise-free vector measurements are used. In this case it is again possible to find an initial body angular velocity $\Omega(t_0)$ and an initial attitude that are consistent with the measurements by solving a set of non-linear equations. They suggest different formulations of these equations, and apply Newton's method (with possibly many different initializations) to obtain a solution to the equations. Again this approach is likely to be useful only when there is very little noise. In later work~\cite{psiaki2012numerical} Psiaki and Hinks describe a method to find local optima of Psiaki's first and second problems (with no additional assumptions) by a novel alternating optimization scheme. The main idea is that for fixed $\Omega(t_0)$, each point $Q(t_0),Q(t_1),\ldots,Q(t_N)$ on the trajectory is linear in $Q(t_0)$. Hence if we can compute the trajectory $(Q(t_n))_{n=0}^{N}$ for fixed $\Omega(t_0)$ we can minimize the objective function of~\eqref{eq:psiaki-lsq} over $Q(t_0)$ for fixed $\Omega(t_0)$ by solving an instance of Wahba's problem. To obtain the trajectory $(Q(t_n))_{n=0}^{N}$ for fixed $\Omega(t_0)$, Psiaki and Hinks suggest numerically solving the rigid body equations. For the other part of the alternating optimization scheme, they employ a trust-region method to locally optimize over $\Omega(t_0)$ for fixed $Q(t_0)$. As presented this problem only finds local optima for $\Omega(t_0)$ and $Q(t_0)$. Nevertheless this method makes very few assumptions, and can incorporate many measurements and so should behave well in the presence of measurement noise. A simpler, but much more naive, strategy would be to discretize the space of $\Omega(t_0)$, solve (in parallel) the corresponding instance of Wahba's problem for each value of $\Omega(t_0)$, then output the pair $(\Omega(t_0),Q(t_0))$ with the smallest cost. This is a reasonable strategy for Psiaki's first problem since aliasing issues mean there is always an optimal $\omega$ in the interval $[-\pi/\tau,\pi/\tau)$. A clear downside of this discretization approach when compared with the semidefinite optimization-based methods we describe in Section~\ref{sec:semidefinite} is that it is expensive to obtain global solutions of high accuracy. Furthermore, the semidefinite optimization-based formulation easily extends to give semidefinite optimization-based formulations for more general problems (see Section~\ref{sec:variations}) where the subproblems for fixed $\Omega(t_0)$ do not reduce to instances of Wahba's problem. \section{Semidefinite optimization reformulations} \label{sec:semidefinite} The main aim of this section is to describe how to reformulate \text{trigonometric Wahba problems}{}, and hence Psiaki's first problem (which is a special case), as semidefinite optimization problems. Before doing so, we briefly explain what semidefinite optimization problems are, and what we mean by a semidefinite reformulation of an optimization problem. We illustrate this in Section~\ref{sec:wahba-sdp} by giving a semidefinite reformulation of Wahba's problem that can be thought of as a more flexible description of the $q$-method~\cite{keat1977analysis}. In Section~\ref{sec:psiaki-sdp} we give a semidefinite reformulation of \text{trigonometric Wahba problems}{}, before giving, in Section~\ref{sec:code}, pseudocode illustrating how to implement the semidefinite optimization problems we formulate using generic semidefinite optimization solvers. \subsection{Semidefinite optimization} \label{sec:sdp} Semidefinite optimization problems are convex optimization problems of the form \[ \max_{x}\; \langle c,x\rangle \quad\text{s.t.}\quad A_0 + \sum_{i=1}^{n}A_ix_i \succeq 0\] where $x\in \mathbb{R}^n$ is a vector of decision variables, $c\in \mathbb{R}^n$ represents a linear cost functional, the matrices $A_0,A_1,\ldots,A_n$ are symmetric $m\times m$ matrices. Recall that $X \succeq 0$ means that the symmetric matrix $X$ is positive semidefinite. An expression of the form \[ A(x) = A_0 + \sum_{i=1}^{n} A_ix_i \succeq 0\] is often called a \emph{linear matrix inequality} because it is linear in the decision variable $x$. Semidefinite optimization problems can be solved to any desired accuracy in time polynomial in $n$ and $m$ using standard software based on interior point methods~\cite{vandenberghe1996semidefinite}. The semidefinite optimization problems that arise in this paper have additional structure that could be exploited to obtain even more efficient algorithms (see Section~\ref{sec:conclusion} for further discussion of this point). For much more information about semidefinite optimization, including duality theory, numerical algorithms, and applications, see for example~\cite{vandenberghe1996semidefinite}. \paragraph{Semidefinite reformulations} Many different optimization problems arising in a variety of contexts, including some optimization problems for which the natural formulation is not convex, can be reformulated as semidefinite optimization problems. Given an optimization problem, by a \emph{semidefinite reformulation} we mean a semidefinite optimization problem such that \begin{enumerate} \item the optimal value of the semidefinite optimization problem and the original optimization problem are the same; \item there is an efficient procedure to take an optimal solution to the semidefinite optimization problem and produce an optimal solution to the original optimization problem. \end{enumerate} \subsection{Wahba's problem} \label{sec:wahba-sdp} We illustrate the basic idea of semidefinite reformulations with the example of solving Wahba's problem. We note that there are much better ways to solve Wahba's problem. The advantage of the semidefinite reformulation is that it can be extended to more complicated situations, such as Psiaki's first problem. The reformulation presented in this section appears (in a more general context) in~\cite{sanyal2011orbitopes} and is generalized to the analogous problem where $SO(3)$ is replaced with $SO(n)$ for any $n\geq 2$ in~\cite{saunderson2014semidefinite}. (See also~\cite{forbes2014linear} where a semidefinite \emph{relaxation} of Wahba's problem is described, as well as conditions under which it is exact.) Wahba's problem fits into the general form~\eqref{eq:genform1} where $A_0 = \sum_{n=0}^{N}\kappa_n y(n\tau)x(n\tau)^T$ and all the other terms vanish. Using the quaternion parameterization of $SO(3)$, Wahba's problem can be expressed as \begin{equation} \label{eq:wahba-original} \max_{Q\in SO(3)}\langle A_0,Q\rangle = \max_{q\in \H} \langle A_0,\mathcal{A}(qq^T)\rangle = \max_{q\in \H} \langle \mathcal{A}^*(A_0),qq^T\rangle. \end{equation} We now explain how to reformulate~\eqref{eq:wahba-original} as a semidefinite optimization problem following a general pattern that we use again in Section~\ref{sec:psiaki-sdp}. \begin{enumerate} \item Rewrite the problem as the optimization of a \emph{linear} functional over some set. In this case \[ \max_{Z} \langle \mathcal{A}^*(A_0),Z\rangle \quad\text{s.t.}\quad Z\in \{qq^T: q\in \H\}.\] \item Replace the constraint set with the \emph{convex hull} of the constraint set. In this case \[ \max_{Z} \langle \mathcal{A}^*(A_0),Z\rangle \quad\text{s.t.}\quad Z \in \conv\{qq^T: q\in \H\}.\] This optimization problem has the same optimal value as the original non-convex problem because the cost function is linear (see Section~\ref{sec:notation}). \item Describe the convex hull of the constraint set as the feasible region of a semidefinite optimization problem (if possible). In this case such a description is well known~(see, e.g.,~\cite[Theorem 3]{overton1992sum}) and given by \[ \conv\{qq^T: q\in \H\} = \{Z\in \mathcal{S}^n: Z \succeq 0,\; \textup{tr}(Z) = 1\}.\] (This holds because if $Z\succeq 0$ and $\textup{tr}(Z) =1$ then any eigendecomposition $Z = \sum_{i=1}^{n}\lambda_i q_iq_i^T$ expresses $Z$ as a convex combination of matrices of the form $qq^T$ with $\|q\|=1$.) \end{enumerate} The resulting semidefinite reformulation of Wahba's problem is \begin{equation} \label{eq:wahbasdp} \max_{Z} \;\langle \mathcal{A}^*(A_0),Z\rangle \quad\text{s.t.}\quad\textup{tr}(Z)=1,\;Z\succeq 0. \end{equation} \paragraph{Extracting an optimal point} Let $Q$ be an optimal solution of Wahba's problem~\eqref{eq:wahba-original}, and suppose $q$ is a corresponding unit quaternion, so that $Q = \mathcal{A}(qq^T)$. Then the positive semidefinite matrix $Z = qq^T$ is an optimum for the semidefinite reformulation of Wahba's problem~\eqref{eq:wahbasdp}. All the optima of the semidefinite reformulation of Wahba's problem are convex combinations of points of the form $qq^T$ where $\mathcal{A}(qq^T)$ is optimal for the original formulation of Wahba's problem. Under mild assumptions (such as having access to at least two generic vector measurements) Wahba's problem has a unique solution $Q^\star = \mathcal{A}(qq^T)$. Whenever Wahba's problem has a unique solution it follows that the semidefinite reformulation also has a unique solution $Z^\star = qq^T$ and we can recover the solution to Wahba's problem from the solution of the semidefinite relaxation by taking $\mathcal{A}(Z^\star)$. \paragraph{Relationship with the $q$-method} The value of the semidefinite optimization problem~\eqref{eq:wahbasdp} is the largest eigenvalue of the Davenport matrix $\mathcal{A}^*(A_0)$. This can already be seen from~\eqref{eq:wahba-original} and the fact that $\max_{q\in \H}\langle \mathcal{A}^*(A_0),qq^T\rangle = \max_{q\in \H}q^T\mathcal{A}^*(A_0)q = \lambda_{\textup{max}}(\mathcal{A}^*(A_0))$. If $q$ is an eigenvector corresponding to the largest eigenvalue of $\mathcal{A}^*(A_0)$ then $Z = qq^T$ is an optimal solution of the semidefinite reformulation~\eqref{eq:wahbasdp}. As such, our reformulation is closely related to the $q$-method for solving Wahba's problem problem~\cite{keat1977analysis}. \paragraph{Discussion} Note that the transformations in the first and second steps above are merely formal and can be applied to essentially any optimization problem. The third step is non-trivial. In general it is not well understood which sets $S$ have the property that $\conv(S)$ can be described as the feasible region of a semidefinite optimization problem---this is an area of active research (see, for example,~\cite{blekherman2013semidefinite}). One view of this paper is that it shows how to express the convex hulls of the non-convex constraint sets appearing in certain joint spin-rate and attitude estimation problems as the feasible regions of semidefinite optimization problems. \subsection{Trigonometric Wahba problems} \label{sec:psiaki-sdp} We now show how to give semidefinite reformulations of \text{trigonometric Wahba problems}{} (defined in~\eqref{eq:genform1}). By specializing to the case where $(A_n)_{n=0}^{N}$ and $(B_n)_{n=1}^{N}$ are given by~\eqref{eq:A0first} and~\eqref{eq:AnBn}, we obtain a semidefinite reformulation of Psiaki's first problem. As in the case of Wahba's problem we use the parameterization of $SO(3)$ in terms of unit quaternions to rewrite \text{trigonometric Wahba problems}{} as \begin{equation} \label{eq:genquat} \max_{\substack{q\in \H\\\omega\in [-\pi,\pi)}} \langle \mathcal{A}^*(A_0),qq^T\rangle + \sum_{n=1}^{N} \left[\langle \mathcal{A}^*(A_n),\cos(n\omega)qq^T\rangle + \langle \mathcal{A}^*(B_n),\sin(n\omega)qq^T\rangle\right]. \end{equation} We can view this problem as the maximization of a linear functional over the set \begin{equation} \label{eq:moment} \mathcal{M}_N:=\{(qq^T,qq^T\cos(\omega),qq^T\sin(\omega),\ldots,qq^T\cos(N\omega),qq^T\sin(N\omega))\in (\mathcal{S}^4)^{2N+1}: q\in \H, \omega\in [-\pi,\pi)\}. \end{equation} As such the convexified version of~\eqref{eq:genquat} is the following optimization problem where the decision variables are the $2N+1$ symmetric matrices $X_0,X_1,Y_1,\ldots,X_N,Y_N$: \begin{multline} \label{eq:genconv} \max_{(X_n)_{n=0}^{N},(Y_n)_{n=1}^{N}}\; \langle \mathcal{A}^*(A_0),X_0\rangle + \sum_{n=1}^{N}\left[\langle \mathcal{A}^*(A_n),X_n\rangle + \langle \mathcal{A}^*(B_n),Y_n\rangle\right]\\ \textup{subject to} \quad (X_0,X_1,Y_1,\ldots,X_N,Y_N)\in \conv(\mathcal{M}_N). \end{multline} This problem is certainly convex, and has the same optimal value as~\eqref{eq:genform1} and~\eqref{eq:genquat}. It may not be immediately clear that the constraint set $\conv(\mathcal{M}_N)$ has a succinct representation in terms of the feasible region of a semidefinite optimization problem. In fact $\conv(\mathcal{M}_N)$ does have such a representation, and we now turn our attention to describing it. \paragraph{A linear matrix inequality description of $\conv(\mathcal{M}_N)$} We now describe $\conv(\mathcal{M}_N)$ in terms of a linear matrix inequality, making use of the block matrix notation defined in Section~\ref{sec:notation}. We establish the correctness of this description in the Appendix, by combining standard results with a novel symmetry reduction argument. \begin{proposition} \label{prop:symm} \begin{multline} \label{eq:symm} \conv(\mathcal{M}_N) = \{(X_0,X_1,Y_1,\ldots,X_N,Y_N)\in (\mathcal{S}^4)^{2N+1}: \;\textup{tr}(X_0) = 1, \\ \textup{Toeplitz}(X_0,X_1,\ldots,X_N) + \textup{Hankel}(Y_N,Y_{N-1},\ldots,Y_1,0,-Y_1,\ldots,-Y_{N-1},-Y_N) \succeq 0\}. \end{multline} \end{proposition} \begin{proof} We provide a proof in the Appendix. \end{proof} \paragraph{Semidefinite reformulation in the general case} Now that we have a semidefinite description of $\conv(\mathcal{M}_N)$, we can give a semidefinite reformulation for all \text{trigonometric Wahba problems}{}. The following theorem explicitly describes this reformulation, which is obtained by replacing $\conv(\mathcal{M}_N)$ in~\eqref{eq:genconv} with its semidefinite description from Proposition~\ref{prop:symm}. \begin{thm} \label{thm:mainsdp} Let $A_0,A_1,\ldots,A_N,B_1,\ldots,B_N\in \mathbb{R}^{3\times 3}$. Then the \text{trigonometric Wahba problem}{} \begin{equation} \label{eq:twthm}\max_{\substack{Q\in SO(3)\\\omega\in [-\pi,\pi)}} \langle A_0,Q\rangle + \sum_{n=1}^{N}\left[ \langle A_n,\cos(\omega n)Q\rangle + \langle B_n,\sin(\omega n)Q\rangle\right] \end{equation} and the semidefinite optimization problem \begin{align} \max_{(X_n)_{n=0}^{N},(Y_n)_{n=1}^{N}}\;&\langle \mathcal{A}^*(A_0),X_0\rangle + \sum_{n=1}^{N}[\langle \mathcal{A}^*(A_n),X_n\rangle + \langle \mathcal{A}^*(B_n),Y_n\rangle]\label{eq:mainsdp}\\ \textup{s.t.} &\;\; \textup{Toeplitz}(X_0,X_1,\ldots,X_N) + \textup{Hankel}(Y_N,Y_{N-1},\ldots,Y_1,0,-Y_1,\ldots,-Y_{N-1},-Y_N) \succeq 0\nonumber\\ & \;\; \textup{tr}(X_0) = 1\nonumber \end{align} have the same optimal value. The set of optimal points of the semidefinite reformulation is \begin{align} \conv&\,\{(qq^T,qq^T\cos(\omega),qq^T\sin(\omega),\ldots,qq^T\cos(N\omega),qq^T\sin(N\omega)):\label{eq:sols}\\ &\qquad\qquad\qquad\qquad\qquad\qquad\qquad \text{$(\omega,\mathcal{A}(qq^T))$ is an optimal point for~\eqref{eq:twthm}}\}\nonumber. \end{align} \end{thm} \paragraph{Extracting an optimal solution} If $N\geq 2$ we expect a generic \text{trigonometric Wahba problem}{} to have a unique optimal point~$(\omega^\star,Q^\star)$~\cite{psiaki2009generalized}. In that case the semidefinite reformulation~\eqref{eq:mainsdp} has a unique optimal point denoted $(X_0^\star,X_1^\star,Y_1^\star,\ldots,X_N^\star,Y_N^\star)$ from which we can recover $(\omega^\star,Q^\star)$ via \begin{equation} \label{eq:relation} Q^\star = \mathcal{A}(X_0^\star),\quad\cos(\omega^\star) = \textup{tr}(X_1^\star)\quad\text{and}\quad \sin(\omega^\star) = \textup{tr}(Y_1^\star). \end{equation} \subsection{Pseudocode} \label{sec:code} In this section we describe code to implement our semidefinite optimization-based formulations~\eqref{eq:mainsdp} of \text{trigonometric Wahba problems}{}. Our motivation for doing this is to show that it is quite straightforward to use standard numerical routines to solve the semidefinite optimization problems that appears in this paper. The code is expressed in a parsing language called YALMIP~\cite{yalmip} that runs under MATLAB. Internally, YALMIP reformulates the human-readable description of the optimization problem we specify into a standard format, then calls a numerical solver for semidefinite optimization problems (we used MOSEK~\cite{andersen2000mosek} version 7 for these experiments) to solve the optimization problem. In what follows, we assume we have functions \begin{description} \item[\underline{\small\texttt{A\_map}}] implementing the linear map $\mathcal{A}$ taking a $4\times 4$ symmetric matrix and returning a $3\times 3$ matrix according to~\eqref{eq:amap}; \item[\underline{\small{\texttt{block\_toeplitz}}}] implementing the linear map $(X_0,X_1,\ldots,X_N)\mapsto \textup{Toeplitz}(X_0,X_1,\ldots,X_N)$ taking a $4\times 4 \times (N+1)$ array and returning a $4(N+1)\times 4(N+1)$ matrix according to~\eqref{eq:blktoep}; \item[\underline{\small{\texttt{block\_hankel}}}] implementing the linear map $(Y_1,Y_2,\ldots,Y_N) \mapsto \textup{Hankel}(-Y_N,\ldots,-Y_1,0,Y_1,\ldots,Y_N)$ taking a $4\times 4 \times N$ array and returning a $4(N+1)\times 4(N+1)$ matrix according to~\eqref{eq:blkhank}. \end{description} We declare variables in YALMIP using the {\small\texttt{sdpvar}} command. \begin{align*} \mbox{\small\texttt{1:}}\qquad& \mbox{\small\texttt{X = sdpvar(4,4,N+1,'symmetric');}}\\ \mbox{\small\texttt{2:}}\qquad& \mbox{\small\texttt{Y = sdpvar(4,4,N,'symmetric');}}\\ \intertext{For example {\small\texttt{Y}} is a $4\times 4 \times N$ array of variables with each slice {\small\texttt{Y(:,:,n)}} being a symmetric matrix. We specify constraints by constructing an array of constraints expressed in a very natural way. We express the two constraints in~\eqref{eq:mainsdp} by} \mbox{\small\texttt{3:}}\qquad& \mbox{\small\texttt{K = [trace(X(:,:,1))==1, \underline{block\_toeplitz}(X) + \underline{block\_hankel}(Y) >= 0];}}\\ \intertext{where we have indexed from 1 following MATLAB's conventions. Note that in YALMIP this latter inequality is automatically interpreted in the positive semidefinite sense since the matrix on the left hand side is structurally symmetric.} \intertext{Suppose the variables {\small\texttt{A}} and {\small\texttt{B}} are respectively $4\times 4 \times (N+1)$ and $4\times 4 \times N$ arrays with {\small\texttt{A(:,:,n+1)}} being $\mathcal{A}^*(A_{n})$ and {\small\texttt{B(:,:,n)}} being $\mathcal{A}^*(B_n)$. Then we can solve the semidefinite optimization problem~\eqref{eq:mainsdp} with the single line} \mbox{\small\texttt{4:}}\qquad&\mbox{\small\texttt{solvesdp(K,-(A(:)'*X(:)\;+\;B(:)'*Y(:)));}}\\ \intertext{which calls a numerical solver with the constraint set {\small\texttt{K}} and the cost function {\small\texttt{-(A(:)'*X(:) + B(:)'*Y(:))}} (with the minus sign because minimization is the default). Assuming that there is a unique solution to the non-convex problem we can extract the optimal rotation matrix $Q$ and optimal $\omega$ with} \mbox{\small\texttt{5:}}\qquad& \mbox{\small\texttt{Q\_opt = \underline{A\_map}(double(X(:,:,1)));}}\\ \mbox{\small\texttt{6:}}\qquad&\mbox{\small\texttt{omega\_opt = atan2(trace(double(Y(:,:,1))),trace(double(X(:,:,2))));}} \end{align*} \section{Variations} \label{sec:variations} In Section~\ref{sec:semidefinite} we formulated Psiaki's first problem as a semidefinite optimization problem by showing how to express the convex hull of $\mathcal{M}_{N}$ in terms of linear matrix inequalities. This description of $\mathcal{M}_{N}$ also allows us to take a semidefinite optimization-based approach to many variations on Psiaki's first problem. In this section we illustrate the possibilities in this direction with one simple example--- a variant on Psiaki's problem where we assume the measurement errors are bounded, and incorporate this additional information into the formulation. \subsection{Psiaki's first problem with bounded measurement errors} \label{sec:psiakivar} Using the notation from Section~\ref{sec:jas}, suppose we know that the error between the measured direction $y_n$ and the true direction $Q(n \tau)x_n$ is bounded in each coordinate, satisfying \begin{equation} \label{eq:box} -\epsilon \leq y_n - Q(n\tau)x_n \leq \epsilon. \end{equation} Here $\epsilon = \begin{bmatrix} \epsilon_1 & \epsilon_2 &\epsilon_3\end{bmatrix}^T$ is a vector of positive constants that are not necessarily equal, and the inequalities in~\eqref{eq:box} are to be interpreted element-wise. Adding these constraints to the formulation of Psiaki's first problem we obtain the following variant: \begin{align} \min_{\substack{Q(0)\in SO(3)\\\omega\in [-\pi/\tau,\pi/\tau)}} & \sum_{n=0}^{N}\frac{\kappa_n}{2}\|y_n - Q(n\tau)x_n\|_2^2\label{eq:psiakivar-lsq}\\ \text{s.t.} & \quad -\epsilon \leq y_n - Q(n\tau)x_n \leq \epsilon \quad\text{for $n=0,1,\ldots,N$}.\nonumber \end{align} Here, as in Section~\ref{sec:jas}, $Q(n\tau)$ is related to $Q(0)$ via~\eqref{eq:spin2} and so putting $\omega' = \omega \tau$, \begin{align*} Q(n\tau) & = \begin{bmatrix} 1 & 0 & 0\\0 & \cos(n\omega') & -\sin(n\omega')\\0 & \sin(n \omega') & \cos(n \omega')\end{bmatrix}Q\\ & = \left[\begin{smallmatrix} 1 & 0 & 0\\0 & 0 & 0\\0 & 0 & 0\end{smallmatrix}\right]\mathcal{A}(qq^T) + \left[\begin{smallmatrix} 0 & 0 & 0\\0 & 1 & 0\\0 & 0 & 1\end{smallmatrix}\right]\mathcal{A}(qq^T\cos(n\omega')) + \left[\begin{smallmatrix} 0 & 0 & 0\\0 & 0 & -1\\0 & 1 & 0\end{smallmatrix}\right]\mathcal{A}(qq^T\sin(n\omega')). \end{align*} Since the objective function of~\eqref{eq:psiakivar-lsq} is identical to the objective function of Psiaki's first problem~\eqref{eq:psiaki-lsq}, using the notation and manipulations of Sections~\ref{sec:jas} and~\ref{sec:semidefinite} we can rewrite the variant of Psiaki's first problem as \begin{align} \max_{\substack{q\in \H\\\omega' \in [-\pi,\pi)}} & \langle \mathcal{A}^*(A_0),qq^T\rangle + \sum_{n=1}^{N}[\langle \mathcal{A}^*(A_n),qq^T\cos(n\omega')\rangle + \langle \mathcal{A}^*(B_n),qq^T\sin(n\omega')\rangle]\\ \label{eq:psiakivar-linear}\\ \textup{s.t.}-\epsilon \leq& y_n - \left(\left[\begin{smallmatrix} 1 & 0 & 0\\0 & 0 & 0\\0 & 0 & 0\end{smallmatrix}\right]\!\mathcal{A}(qq^T) + \left[\begin{smallmatrix} 0 & 0 & 0\\0 & 1 & 0\\0 & 0 & 1\end{smallmatrix}\right]\!\mathcal{A}(qq^T\cos(n\omega')) + \left[\begin{smallmatrix} 0 & 0 & 0\\0 & 0 & -1\\0 & 1 & 0\end{smallmatrix}\right]\!\mathcal{A}(qq^T\sin(n\omega'))\right)x_n\leq \epsilon\label{eq:varconst}\\ &\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\text{for $n=0,1,\ldots,N$.}\nonumber\end{align} Observe that we have rewritten the problem as the maximization of a linear functional over the constraint set defined by \begin{align*} S & = \bigg\{(qq^T,qq^T\cos(\omega),qq^T\sin(\omega),\ldots,qq^T\cos(N\omega),qq^T\sin(N\omega)):q\in \H, \omega \in [-\pi,\pi), \\ &\quad-\epsilon \leq y_n - \left(\left[\begin{smallmatrix} 1 & 0 & 0\\0 & 0 & 0\\0 & 0 & 0\end{smallmatrix}\right]\mathcal{A}(qq^T) + \left[\begin{smallmatrix} 0 & 0 & 0\\0 & 1 & 0\\0 & 0 & 1\end{smallmatrix}\right]\mathcal{A}(qq^T\cos(n\omega)) + \left[\begin{smallmatrix} 0 & 0 & 0\\0 & 0 & -1\\0 & 1 & 0\end{smallmatrix}\right]\mathcal{A}(qq^T\sin(n\omega))\right)x_n\leq \epsilon\\ &\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\text{for $n=0,1,\ldots,N$}\bigg\}. \end{align*} This set $S$ is the intersection of $\mathcal{M}_{N}$ with the additional constraints~\eqref{eq:varconst} coming from incorporating the knowledge that the measurement errors satisfy the explicit deterministic bounds described in~\eqref{eq:box}. \subsection{A semidefinite relaxation} \label{sec:varsdp} Recall from Section~\ref{sec:semidefinite} that if we could exactly describe $\conv(S)$ in terms of linear matrix inequalities that are not too large, we could obtain a semidefinite reformulation of this problem that can be solved efficiently. Unfortunately we do not know of such a concise description of $\conv(S)$, and conjecture that no such concise description exists for all choices of the $x_n$ and $y_n$. Instead, a natural general approach is to construct a \emph{semidefinite relaxation} of $\conv(S)$. By this we mean a convex set $C$ such that \begin{enumerate} \item $C \supseteq \conv(S)\supseteq S$ and \item $C$ has a simple description in terms of linear matrix inequalities. \end{enumerate} One choice would be to take $C$ to be the convex set \begin{align*} C &= \bigg\{(X_0,X_1,Y_1,\ldots,X_N,Y_N)\in \conv\,\mathcal{M}_N:\\ & -\epsilon \leq y_n - \left(\left[\begin{smallmatrix} 1 & 0 & 0\\0 & 0 & 0\\0 & 0 & 0\end{smallmatrix}\right]\mathcal{A}(X_0) + \left[\begin{smallmatrix} 0 & 0 & 0\\0 & 1 & 0\\0 & 0 & 1\end{smallmatrix}\right]\mathcal{A}(X_n) + \left[\begin{smallmatrix} 0 & 0 & 0\\0 & 0 & -1\\0 & 1 & 0\end{smallmatrix}\right]\mathcal{A}(Y_n)\right)x_n\leq \epsilon \quad\text{for $n=0,1,\ldots,N$}\bigg\}. \end{align*} One can check that while $C$ is, in general, strictly larger than $\conv(S)$, it can be expressed using linear matrix inequalities (since $C$ is obtained by adding linear inequalities to $\conv(\mathcal{M}_N)$ which has a linear matrix inequality description from Proposition~\ref{prop:symm}). By optimizing over $C$ rather than $S$ we obtain the following \emph{semidefinite relaxation} of the optimization problem \begin{align} \max_{(X_n)_{n=0}^{N},(Y_n)_{n=1}^{N}}&\;\langle \mathcal{A}^*(A_0),X_0\rangle + \sum_{n=0}^{N}[\langle \mathcal{A}^*(A_n),X_n\rangle + \langle \mathcal{A}^*(B_n),Y_n\rangle] \label{eq:varsdp}\\ \quad\text{s.t.}&\quad (X_0,X_1,Y_1,\ldots,X_N,Y_N)\in C.\nonumber \end{align} When we solve this semidefinite relaxation, if the solution $(X_0^\star,X_1^\star,Y_1^\star,\ldots,X_N^\star,Y_N^\star)\in C$, returned by the solver, is actually in $S$, then it is a solution of the original non-convex problem~\eqref{eq:psiakivar-linear} we are trying to solve. In this case it is typical to say that the semidefinite relaxation is \emph{exact} for this instance. If $(X_0^\star,X_1^\star,Y_1^\star,\ldots,X_N^\star,Y_N^\star) \notin S$, we have not solved the non-convex problem, but can still conclude that the value of the objective function at this point is an upper bound on the optimal value of the original non-convex maximization problem~\eqref{eq:psiakivar-linear}. Such a bound can be used, for example, to assess the quality (in terms of the objective function) of any feasible point obtained, for instance, by a local optimization method. \subsection{Numerical experiments} \label{sec:expts} In this section we describe the results of two simple numerical experiments to illustrate solving Psiaki's first problem using semidefinite optimization, as well as solving the semidefinite relaxation of the variant on Psiaki's problem discussed in Sections~\ref{sec:psiakivar} and~\ref{sec:varsdp}. For all experiments we use the same parameters as in Psiaki's truth-model simulation in~\cite{psiaki2009generalized}---the true spin period is $45.32$ seconds (so the true spin-rate is $\omega = 0.1386$ radians per second), the sampling period is $\tau = 7.7611$ seconds per sample, and the initial attitude is $Q(0) = I$. The attitude dynamics are described by~\eqref{eq:spineq}. \paragraph{Solving Psiaki's first problem} In the first experiment we solve Psiaki's first problem using the semidefinite reformulation~\eqref{eq:mainsdp}. In particular we repeat the following experiment $T=1000$ times: \begin{enumerate} \item Sample reference directions $x_0,x_1,\ldots,x_{10}$ uniformly on the sphere. \item For $n=0,1,2\ldots,N$, sample measurements $y_n$ uniformly distributed on the intersection of the unit sphere and the region \begin{equation} \label{eq:boxexpt} -\epsilon \leq y_n - Q(t_n)x_n \leq \epsilon \end{equation} where $\epsilon = \begin{bmatrix}0.5 & 0.5 & 0.05\end{bmatrix}^T$ (by sampling uniformly on the sphere and rejecting those samples not in the box-shaped region). This corresponds to measurements that are very accurate along one axis, but quite inaccurate in other directions. \item For $N=2,3,\ldots,10$, use the reference directions $x_0,x_1,\ldots,x_N$ and measurements $y_0,y_1,\ldots,y_N$ and solve the semidefinite optimization reformulation of Psiaki's first problem~\eqref{eq:mainsdp}. \end{enumerate} We note that although the data are generated from a model where the measurement errors satisfy the explicit bounds~\eqref{eq:boxexpt}, we do not exploit this in our solution method. Also, to get a sense of the measurement errors introduced, we note that under the noise model we have adopted, the angle between $y_n$ and $Q(t_n)x_n$ over all samples was at most $41.1$ degrees, and on average $16.8$ degrees. The average angular error (in degrees) between the estimate of the initial attitude and the true initial attitude is indicated by cross-shaped markers in Figure~\ref{fig:Qopterr}. Given an estimate $\hat{Q}$ of the true initial attitude $Q(0) = I$, the angular error $\theta$ satisfies $\textup{tr}(\hat{Q}^TQ(0)) = 2\cos(\theta)+1$. Hence we compute the angular error via $|\cos^{-1}[\textup{tr}(\hat{Q}^TQ(0))-1)/2]|$. The corresponding average error in the spin-rate estimate $\hat{\omega}$ is computed by taking the mean of $|\hat{\omega} - \omega|$ over all trials and is indicated by cross-shaped markers in Figure~\ref{fig:omegaerr}. It is clear that as more vector measurements are used (i.e.\ as $N$ increases) the estimates improve, justifying using more than the minimum number of measurements required for the optimization problem to have a unique optimum. \paragraph{Solving the variant with bounded measurement errors} In the second experiment we use exactly the same data as for the first experiment. This time, instead of solving the semidefinite reformulation of Psiaki's first problem, we solve the semidefinite relaxation~\eqref{eq:varsdp} of the variant of Psiaki's first problem with bounded measurement errors. This estimation method explicitly makes use of the fact that the measurement errors satisfy~\eqref{eq:boxexpt}. \begin{figure}[h] \begin{center} \subfigure[Angular errors in the initial attitude]{\label{fig:Qopterr} \includegraphics[trim = 52mm 172mm 90mm 40mm, clip]{Qopterr} } \subfigure[Error in the spin-rate estimate]{\label{fig:omegaerr} \includegraphics[trim = 52mm 172mm 90mm 40mm, clip]{omegaerr} } \end{center} \caption{\label{fig:exptresults}Results from the experiment described in Section~\ref{sec:expts}. Figure~\ref{fig:Qopterr} shows the average (over $1000$ random trials) error in estimating the initial attitude using $N+1$ measurements by solving Psiaki's first problem (cross-shaped markers) and the semidefinite relaxation of the bounded error variant described in Section~\ref{sec:varsdp} (dot-shaped markers). Similarly Figure~\ref{fig:omegaerr} shows the average error in estimating the spin-rate for the same experiment.} \end{figure} As discussed in Section~\ref{sec:varsdp}, and unlike the case for the semidefinite reformulation of Psiaki's first problem itself, when solving the semidefinite relaxation of the bounded error variant of Psiaki's first problem there is no guarantee that the relaxation will be exact. In other words, we do not know, in advance, whether the solution to the semidefinite optimization problem corresponds to a solution of the original non-convex problem~\eqref{eq:psiakivar-linear}. On the other hand, after solving the semidefinite optimization problem we can determine whether or not this is the case by checking whether the solution is feasible for the original problem. The number of trials for which the semidefinite relaxation was exact is listed in Table~\ref{tab:exactness} for each $N=2,3,\ldots,10$. For this experiment, the average angular error in the initial attitude estimates and the average error in the spin-rate estimates are indicated by dot-shaped markers in Figures~\ref{fig:Qopterr} and~\ref{fig:omegaerr}. If the semidefinite relaxation was exact we compute the errors in these quantities in the same way as for the previous experiment. If the semidefinite relaxation was not exact, to be as conservative as possible we take the error to be the maximum possible value: $180$ degrees for the initial attitude error and $\pi$ radians/second for the spin-rate error. The results in Figures~\ref{fig:Qopterr} and~\ref{fig:omegaerr} show that by incorporating explicit bounds on the measurement errors (if known) into the semidefinite optimization framework, significantly better estimates of the initial attitude can be obtained. Furthermore Table~\ref{tab:exactness} indicates that the semidefinite relaxation was indeed exact on many of our random trials, suggesting that the semidefinite relaxation approach may be well suited to tackling at least this variant on Psiaki's first problem, and perhaps others. \begin{table} \begin{center} \begin{tabular}{c|ccccccccc} $N$ & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10\\\hline $T_\textup{exact}$ (/1000) & 842 & 816 & 867 & 918 & 948 & 958 & 965 & 969& 973 \end{tabular} \end{center} \caption{\label{tab:exactness}$T_\textup{exact}$ is the number of random experiments (as described in Section~\ref{sec:expts}) for which the semidefinite relaxation for the variant on Psiaki's first problem~\eqref{eq:varsdp} was exact. Here $N+1$ is the number of vector measurements used.} \end{table} \section{Future directions} \label{sec:directions} We briefly comment on possible future research directions based on the work in the present paper. \subsection{Numerical algorithms} The semidefinite reformulation of Psiaki's first problem~\eqref{eq:mainsdp} has a very specific structure. This structure could be exploited to develop numerical algorithms for its solution (as well as the solution for variants on the problem) that are much faster than the generic interior-point algorithms we used for our experiments. Indeed semidefinite optimization problems with a similar structure arise in problems related to the Kalman-Yakubovich-Popov (KYP) lemma in robust control and in that context numerous specialized algorithms have been developed for their solution (see, e.g.,~\cite{liu2007low,genin2003optimization,kao2007new}). Furthermore, great gains can be made by producing optimized low-level code for a particular family of convex optimization problems. An excellent example of this is the code-generation software CVXGEN which focuses on linear and convex quadratic programs~\cite{mattingley2012cvxgen}. \subsection{Further variants} \label{sec:fvariations} A semidefinite reformulation of a problem is particularly useful because it can be combined in many ways with other semidefinite optimization primitives to yield more problems that can be solved in the semidefinite optimization framework. In Section~\ref{sec:variations} we discussed a variation on Psiaki's first problem that had bounds on the angular noise. In this case it was straightforward to formulate a semidefinite relaxation using our semidefinite reformulation of Psiaki's first problem. Another natural variation that could be approached this way would be to obtain semidefinite relaxations of Psiaki's first problem that are robust to uncertainty in certain model parameters. A similar idea has been carried out in detail for Wahba's problem by Ahmed et al.~\cite{ahmed2012semidefinite}. They extended the semidefinite formulation of Wahba's problem~\cite{sanyal2011orbitopes} to a variant that is robust to uncertainty in certain parameters, such as the reference directions. As suggested by Ahmed at al.~this could be useful when using magnetometer measurements together with a low-order magnetic field model. \subsection{Psiaki's second problem} It would be interesting to try to take a similar approach to the one taken in the present paper to related problems, such as Psiaki's second problem. To do so we would need to give a semidefinite description (or perhaps a relaxation) of \begin{equation} \label{eq:convP2} \conv_{Q(t_0)\in SO(3),\Omega(t_0)}\, \{(Q(t_0),Q(t_1),\ldots,Q(t_N)): Q(t) = \Phi(t-t_0;\Omega(t_0))Q(t_0)\quad\text{for all $t\geq t_0$}\}. \end{equation} Given this, we note that the objective function~\eqref{eq:psiaki-lsq} can be rewritten as the maximization of a linear functional over the convex hull described in~\eqref{eq:convP2}. A more modest goal along similar lines might be to discretize the differential equation~\eqref{eq:ode} and try to compute the convex hull (over all initial conditions $Q(t_0),\Omega(t_0)$) of an appropriately subsampled trajectory of the associated difference equation for the attitude variables. This approach of convexifying a problem based on discretized dynamics would, in a sense, be a convex analogue of the methods proposed for Psiaki's second problem in~\cite{psiaki2012numerical}. \section{Conclusion} \label{sec:conclusion} We have shown how Psiaki's generalization of Wahba's problem to the case of a spacecraft spinning around a fixed axis at an unknown rate can be exactly reformulated as a semidefinite optimization problem. Such convex optimization problems can be solved globally using standard methods for semidefinite optimization. As suggested by Psiaki when formulating his generalizations of Wahba's problem~\cite{psiaki2009generalized}, our solutions to these generalizations of Wahba's problem could be used to initialize standard extended Kalman filter-based methods for attitude estimation. Furthermore, we have illustrated how to use our reformulation of Psiaki's first problem to construct semidefinite relaxations of a more complicated variant on the problem. Our numerical experiments with a variant that includes explicit bounds on the measurement errors suggest that incorporating additional information into the formulation can improve the estimation errors. Our results also suggest the semidefinite relaxation approach we propose, although not exact in general, often computes solutions to the original non-convex variations of Psiaki's problem that we ultimately are aiming to solve.

https://arxiv.org/abs/1311.1487

Approximation pathologies for certain continued fractions

We will provide a family of continued fractions for which there is no correspondence between dynamic and approximation pairs, leading to an anomaly in their corresponding Spaces of Jager Pairs.

\section{Introduction and preliminaries}{} Associated with every continued fraction expansion is its Space of Jager Pairs, used in assessing the approximation quality for its convergents \cite{JK}. In this paper, we will examine the one parameter family of continued fractions \[ [a_1,a_2,...]_k := \frac{k}{k+a_1+\frac{k}{k+a_2+ ...}}, \hspace{1pc} a_n \in \IZ_{\ge 0}, \hspace{1pc} 0 < k \in \mathbb{R},\] introduced in \cite{HM1}, was also studied extensively in \cite{Avi1, Avi2}. Letting $x_0 :=[a_1,a_2,...]_k$ leads to the definitions of the convergents $\frac{p_0}{q_0} = \frac{0}{1}$ and \[ \frac{p_n}{q_n} := [a_1,a_2,...,a_n]_k = \frac{k}{k+a_1+\frac{k}{k+a_2+ ... + \frac{k}{k+a_n}}}, \hspace{.1pc} n\ge1.\] Define the future and past of $x_0$ at time $n \ge 1$ to be $x_n := [a_{n+1},a_{n+2},...]_k \in (0,1)$ and $y_n := -k-a_n - [a_{n-1},a_{n-2},...,a_1]_k \in (-\infty, -k]$ (we take $y_1 = -k-a_1$) and call the pairs $(x_n,y_n)$ the dynamic pairs of $x_0$ at time $n$. The pair of consecutive terms $(\theta_{n-1}(x_0),\theta_n(x_0))$ in the sequence of approximation coefficients $\{\theta_n\}_0^\infty := \left\{\abs{x_0 - \frac{p_n}{q_n}}q_n^2\right\}_0^\infty$ is called the approximation pair of $x_0$ at time $n$. There is a correspondence between these pairs, which was established by Haas and Molnar in \cite[Theorem 4]{HM1}, who proved that $(\theta_{n-1},\theta_n)$ is the image of $(x_n,y_n)$ under the bijective map \begin{equation}\label{Psi} \Psi_k(x,y) := \left(\frac{1}{x-y}, -\frac{x{y}}{k(x-y)}\right). \end{equation} We will prove that there is no such bijection when $0<k<1$ and, consequently, that the formula for the Space of Jager Pairs for these continued fraction expansions does not follow suit with the $k\ge 1$ cases. \section{The space of Jager Pairs when k is less than one}{} \noindent For all $k \in (0,\infty)$ and $a \in \IZ_{\ge 0}$ define the region $P_{(k,a)} := (0,1) \times (-k-a-1, -k-a]$ and $P_{(k,a)}^\# := \Psi_k(P_{(k,a)})$. The continuity of the map $\Psi$ provides us with the partition $\Gamma_k = \Psi_k(\bigcup{P_a}) = \bigcup{\Psi(P_a)} = \bigcup{P_a^\#}$. To find $P_a^\#$, we will use the following propositions: \begin{proposition}{\cite[Proposition 3.1]{Avi1}}\label{P_a}\\ If $\Psi$ is injective on $P_{(k,a)}$, then $P_{(k,a)}^\#$ is the quadrangle with vertices $\left(\frac{1}{k+a},0\right), \hspace{.1pc} \left(\frac{1}{k+a+1},\frac{k+a}{k(k+a+1)}\right),\\ \hspace{.1pc} \left(\frac{1}{k+a+2},\frac{k+a+1}{k(k+a+2)}\right)$ and $\left(\frac{1}{k+a+2},0\right)$. \end{proposition} \begin{proposition}\label{Psi_fold} $\Psi$ is injective on the region $\{(x,y) \in \mathbb{R}^2 : x + y < 0\}$ and is invariant under the reflection about the line $x+y=0$ \end{proposition} \begin{proof} The equality $\Psi(x, y) = \big(\frac{1}{x-y}, -\frac{xy}{k(x-y)}\big) = \Psi(-y,-x)$ proves the first statement. Let $(x_1,y_1)$ and $(x_2,y_2)$ be points which are on or below the line $x + y = 0$, so that $x_1 + y_1 \le 0, \hspace{.1pc} x_2 + y_2 \le 0$ and let $u_1,v_1,u_2,v_2 \in \mathbb{R}$ be such that $(u_1,v_1) = \Psi(x_1,y_1)= \Psi(x_2,y_2) = (u_2,v_2)$. Then the definition of $\Psi$ \eqref{Psi} implies that $\frac{1}{x_1-y_1} = u_1 = u_2 = \frac{1}{x_2-y_2}$, hence \begin{equation}\label{x-y_G} x_1 - y_1 = x_2 - y_2 \ne 0. \end{equation} Also \[-\dfrac{u_1}{k}x_1y_1 = v_1 = v_2 = -\dfrac{u_2}{k}x_2y_2 = -\dfrac{u_1}{k}x_2y_2\] and $u_1,u_2 > 0$ imply that $x_1y_1 = x_2y_2$, so that \[(x_1 +y_1)^2 = (x_1 - y_1)^2 + 4x_1y_1 = (x_2 - y_2)^2 + 4x_2y_2 = (x_2 + y_2)^2.\] Since $x_1 + y_1 \le 0$ and $x_2 + y_2 \le 0$, this last equation proves $x_1 + y_1 = x_2 + y_2$, which in tandem with condition \eqref{x-y_G}, proves $x_1=x_2$ and $y_1 = y_2$, hence $\Psi$ is injective on or below the line $x+y=0$. \end{proof} \noindent These propositions proves that when $k \ge 1$, the Space of Jager Pairs \[\Gamma_k := \{(\theta_{n-1}(x_0),\theta_n(x_0))\}_{x_0 \in (0,1)}\] is the convex quadrangle with vertices $\left(0,0\right), \hspace{.1pc} \left(\frac{1}{k},0\right),\hspace{.1pc}\left(\frac{1}{k+1},\frac{1}{k+1}\right)$ and $\left(0,\frac{1}{k+1}\right)$. \begin{center} \includegraphics[scale=.5]{gauss_larger_than_one.pdf}\\ $\Gamma_k \text{ when $k\ge 1$}$\\ \end{center} \noindent When $0<k<1$, $\Psi$ is injective on $P_{(k,a)}$ precisely when $a>0$. In order to finish the characterization of the space of approximation coefficients for these cases, we prove: \begin{theorem}\label{P_0_k<1} When $0<k<1$, $P_0^\#$ is the intersection of the unbounded regions $u{k}+v < 1, \hspace{.1pc} v > 0, \hspace{.1pc} (k+1)^2u+k{v} > k+1, \hspace{.1pc} u+k{v} < 1$ and $4k{u}{v} \le 1$. \end{theorem} \begin{proof} Let $(u,v) := \Psi(x,-k-a)$. From the definition \eqref{Psi} of $\Psi$, we obtain \begin{equation}\label{u=f(x,y)} u = \dfrac{1}{x-y} = \dfrac{1}{x+k+a} \end{equation} and $v = -\frac{x{y}}{x-y}$ so that \begin{equation}\label{v=f(u)} v = -\dfrac{x{y}}{k(x-y)} = \dfrac{u}{k}x(-y) = \dfrac{u}{k}\bigg(\dfrac{1}{u}- (a+k)\bigg)(a+k) = \dfrac{a+k}{k}\big(1-(a+k)u\big). \end{equation} After restricting $P_0$ to its part on or below $x+y=0$, we are left with the region whose boundary is the 5-gon with vertices $(k,-k), (1,-1), (1,-k), (1,-k-1), (0,-k-1)$ and $(0,-k)$, which includes the part of its perimeter, which is the open line segment connecting the first vertex to the second vertex. We will use these formulas to evaluate the image of each open line segment in this 5-gon under $\Psi$. \begin{enumerate} \item When $y=-k$, we obtain that $u$ ranges between $\frac{1}{k}$ and $\frac{1}{2k}$ as $x$ ranges between $0$ and $k$. Plugging $a=0$ to formula \eqref{v=f(u)} yields that the line segment $(0,k) \times \{-k\}$ in the $x{y}$ plane maps to the open segment in the line $uk+v=1$ between the points $\left(\frac{1}{2k},\frac{1}{2}\right)$ and $\left(\frac{1}{k},0\right)$ on the $u{v}$ plane\\ \item Plugging $x=a=0$ yields that the line segment $\{0\} \times (-k-1,-k)$ in the $x{y}$ plane maps to $\left(\frac{1}{k+1},\frac{1}{k}\right)\times\{0\}$ on the $u{v}$ plane. \item The line segment $(0,1) \times \{-(k+1)\}$ in the $x{y}$ plane maps to the open segment of the line $(k+1)^2u+k{v}=k+1$ between $\left(\frac{1}{k+1},0\right)$ and $\left(\frac{1}{k+2},\frac{k+1}{k(k+2)}\right)$ on the $u{v}$ plane. \item Plugging $x=1$ and $a=0$ yields that the line segment $\{1\} \times \left(-1,-(k+1)\right)$ in the $x{y}$ plane maps to the open segment of the line $u+k{v}=1$ between the points $\left(\frac{1}{k+2},\frac{k+1}{k(k+2)}\right)$ and $\left(\frac{1}{2},\frac{1}{2k}\right)$ on the $u{v}$ plane. \item Finally, if $y=-x$ then $u = \frac{1}{x-y}= \frac{1}{2x}$ hence $x = \frac{1}{2u}$ and $v= -\frac{u}{k}x{y} = \frac{1}{4k{u}}$. Thus the line $x+y=0$ in the $x{y}$ plane is mapped under $\Psi$ to the hyperbola $4k{u}v=1$ in the $u{v}$ plane. The points $(1,-1)$ and $(k,-k)$ map to $\big(\frac{1}{2},\frac{1}{2k}\big)$ and $\big(\frac{1}{2k}, \frac{1}{2}\big)$ under $\Psi$, so that the open segment of the line $x+y=0$ from $(1,-1)$ to $(k,-k)$ maps to the open arc on this hyperbola from $\big(\frac{1}{2},\frac{1}{2k}\big)$ to $\big(\frac{1}{2k}, \frac{1}{2}\big)$. \end{enumerate} \noindent Since $\Psi$ is a continuous bijection on and below the line $x+y=0$, it maps the interior and boundary of this 5-gon bijectively into the interior and boundary of the region in the $u{v}$ plane whose boundary we have just determined and which coincides with the hypothesis. Using proposition \ref{Psi_fold}, we know that the part of $P_0$ which is above the line $x+y=0$ has the same image under $\Psi$ as its reflection about this line. Since this reflection is also contained in $P_0$, we conclude that $P_0$ is mapped in its entirety onto this region, thus concluding the result. \end{proof} \begin{corollary} When $0<k<1$, the Space of Jager Pairs is the region in the $u{v}$ plane which is the union of the quadrangle vertices $\left(0,0\right), \hspace{.1pc} \left(\frac{1}{k},0\right),\hspace{.1pc}\left(\frac{1}{k+1},\frac{1}{k+1}\right)$ and $\left(0,\frac{1}{k+1}\right)$ and the part of the hyperbola $4k{u}{v}=1$ in the $u{v}$ plane between $u=\frac{1}{2}$ and $u=\frac{1}{2k}$. \end{corollary} \begin{center} \includegraphics[scale=.5]{gauss_less_than_one.pdf}\\ $\Gamma_k \text{ when $0 < k < 1$}$ \end{center}

https://arxiv.org/abs/1407.7216

PTAS for Minimax Approval Voting

We consider Approval Voting systems where each voter decides on a subset to candidates he/she approves. We focus on the optimization problem of finding the committee of fixed size k minimizing the maximal Hamming distance from a vote. In this paper we give a PTAS for this problem and hence resolve the open question raised by Carragianis et al. [AAAI'10]. The result is obtained by adapting the techniques developed by Li et al. [JACM'02] originally used for the less constrained Closest String problem. The technique relies on extracting information and structural properties of constant size subsets of votes.

\section{Introduction} Approval Voting systems are widely considered~\cite{brams_av_book} as an alternative to traditional elections, where each voter may select and support at most some small number of candidates. In Approval Voting each voter decides about every single candidate if he approves the candidate or does not approve him/her. A result is obtained by applying a predefined election rule to the set of collected votes. In this paper we study the problem of implementing an appropriate election rule and focus on the Minimax objective \cite{brams}: we minimize the biggest dissatisfaction over voters. The resulting optimization problem is denoted $MAV$, and it is to select a committee composed of exactly $k$ candidates, and minimizing the maximal symmetric difference between the committee and the set of approved candidates by a single voter. Using the string terminology, votes are encoded as strings, and the goal is to find a string encoding a committee minimizing the maximal Hamming distance to an input string. Unlike in the related Closest String problem, in $MAV$ there is also a constraint: the selected committee must be of fixed size $k$, and hence in the string terminology there must be exactly $k$ ones in the string. \subsection{Related work and our results} Many different objective functions have been proposed and studied in the context of selecting the committee based on the set of votes collected in an Approval Voting system~\cite{computations_av,brams_av_book}. Clearly, optimizing the sum of Hamming distances to all votes is an easy task and can be done by simply selecting the $k$ candidates approved by the largest number of voters. By contrast, Minimax Approval Voting was shown by LeGrand~\cite{nphard} to be NP-hard. LeGrand et al.~\cite{kcompletion} obtained $3$-approximation by a very simple $k$-completion algorithm. Next, Carragianis et al.~\cite{markakis} gave the currently best $2$-approximation algorithm. The algorithm was obtained by rounding a fractional solution to the natural LP relaxation of the problem, and obtained approximation ratio essentially matches the integrality gap of the LP. In this paper we give a PTAS for the Minimax Approval Voting problem. Our work is based on the PTAS for Closest String \cite{ptascs}, which is a similar problem to $MAV$ but there we do not have the restriction on the number of 1's in the result. Technically, our contribution is the method of handling the number of 1's in the output. We also believe that our presentation is somewhat more intuitive. Approval Voting systems are also analyzed in respect of manipulability, see e.g., \cite{computations_av} or \cite{markakis}. In particular, \cite{markakis} proved that each strategy-proof algorithm for $MAV$ must have approximation ratio at least $2-\frac{2}{k+1}$, which implies that our PTAS cannot be strategy-proof. \subsection{Definitions} We will use the following notation:\\ $n$ -- number of voters,\\ $m$ -- number of candidates,\\ $s_i \in \{0,1\}^m$ -- a vote of voter $i$,\\ $s_i[j]=1$ if voter $i$ approves candidate $j$,\\ $s_i[j]=0$ if voter $i$ does not approve candidate $j$,\\ $S=\{s_1,s_2,\dotsc,s_n\}$ -- the set of collected votes,\\ $s^{(1)}=\big|\{j:s[j]=1\}\big|$ -- the number of 1's in $s$.\\ For $x,y \in [\,0,1]^m$ we define a distance $d(x,y) = \sum_{j=1}^m \big|x[j]-y[j]\big| = \lVert x-y \rVert _{1}$.\\ For $x,y \in \{0,1\}^m$, $d(x,y)$ is called the Hamming distance. \begin{definition}\label{def_mavk} \[ OPT=\min_{\substack{x \in \{0,1\}^m\\x^{(1)}=k}} \: \max_{i \in \{1,2,\dots,n\}} d(x,s_i)\] Let $s_{OPT}$ be an optimal solution, i.e., $\max_{i \in \{1,2,\dots,n\}} d(s_{OPT},s_i)=OPT$. \end{definition} WLOG we assume that $n>k$. If not, we copy the first string $k-n+1$ times. \subsection{The main idea behind our algorithm}\label{name_consensus} The general idea behind our PTAS is to find a small enough subset $X$ of votes that is a ``good representation'' of the whole set of votes $S$. Then the candidates are partitioned into those for which voters in $X$ agree and the rest of candidates. For the ``consensus candidates'' we fix our decision to the decision induced by votes in $X$ (additionally correcting the number of selected candidates in the ``consensus'' set). Next, we consider the optimization problem of finding a proper subset of the remaining candidates to join the committee. The key insight is that there exists a small enough subset $X$ such that the induced decision for the ``consensus candidates'' will not be a big mistake. \subsection{Organization of the paper} First, in Section~\ref{sec:info_subsets} we formalize the information we may extract from subset of votes, and introduce a measure of inaccuracy of such a subset. Next, in Section~\ref{sec:existence_of_subset} we prove the existence of a small subset of votes with stable inaccuracy. In Section~\ref{sec:aux_prob} we show that the optimization problem of deciding the part of the committee not induced by the subset of votes can be approximated with only a small additional loss in the objective function. Finally, in Section~\ref{sec:alg} we give an algorithm considering all subsets of a fixed size and show that, in the iteration when the algorithm happens to consider a subset with stable inaccuracy, it will produce a $(1+\epsilon)$-approximate solution to $MAV$. \section{Extracting information from subsets} \label{sec:info_subsets} We consider subsets of votes and analyze the information they carry. We measure the inaccuracy of this information with respect to the set of all votes. We show that there exists a small subset with stable inaccuracy, i.e., the drop of inaccuracy after including one more vote is small. Let us define an inaccuracy function $ina:2^S \mapsto \mathbb{R}_{\geqslant 0}$ that measures the inaccuracy if we will consider subset $Y \subseteq S$ instead of $S$. The smaller the $ina(Y)$ is the better the common parts of strings in $Y$ represent $s_{OPT}$. \begin{definition}\label{def_ina} For all $Y \subseteq S, Y \neq \emptyset$ we define functions $t(Y) \in \{0,1\}^m$ and $ina(Y) \in \mathbb{R}_{\geqslant 0}$ as follows: \[\big(t(Y)\big)[j] = \begin{cases} 0 & \text{if } \forall_{y \in Y} \quad y[j]=0\\ 1 & \text{if } \forall_{y \in Y} \quad y[j]=1\\ s_{OPT}[j] & \text{otherwise,} \end{cases}\] \[ ina(Y) = d(t(Y),s_{OPT}).\] \end{definition} Intuitively $t(Y)$ is the optimal solution $s_{OPT}$ changed at positions where all strings from $Y$ agree. Also we define the pattern of a subset of votes. \begin{definition}\label{def_p} For all $Y \subseteq S, Y \neq \emptyset$ we define pattern $p(Y) \in \{0,1,*\}^m$ as: \[ \big(p(Y)\big)[j]= \begin{cases} 0 & \text{if } \forall_{y \in Y} \quad y[j]=0\\ 1 & \text{if } \forall_{y \in Y} \quad y[j]=1\\ * & \text{otherwise.} \end{cases} \] \end{definition} It represents positions that all strings in $Y$ agree. ``$*$'' encodes a mismatch. Note that (from Definitions \ref{def_ina} and \ref{def_p}) $t(Y)$ is an optimal solution $s_{OPT}$ overwritten by a pattern $p_r$ on no-star positions: \[\big(t(Y)\big)[j] = \begin{cases} s_{OPT}[j] & \text{if } \big(p(Y)\big)[j] = * \\ \big(p(Y)\big)[j] & \text{otherwise.} \end{cases} \] The inaccuracy function has the following properties: \begin{lemma}\label{decreasing_ina} $\forall_{s_{i_1} \in S}$, for all sequences $\{s_{i_1}\}=Y_1 \subseteq Y_2 \subseteq \dots \subseteq Y_n = S$ we have \[OPT \geqslant ina(Y_1) \geqslant ina(Y_2) \geqslant \dots \geqslant ina(Y_n) = 0\] \end{lemma} \begin{proof} It is easy to see that \[ina(Y_1) \stackrel{\text{def.}}{=} d\big(t(Y_1),s_{OPT}\big)=d\big(t(\{s_{i_1}\}),s_{OPT}\big)=d\big(s_{i_1},s_{OPT}\big) \leqslant OPT,\] \[ina(Y_n) = ina(S) = d(s_{OPT},s_{OPT}) = 0.\] Still we need to prove $ina(Y_i) \geqslant ina(Y_{i+1})$. Pattern $p(Y_{i+1})$ is built on strings from $Y_i \subseteq Y_{i+1}$ and strings from $Y_{i+1} \setminus Y_i$. So $p(Y_{i+1})$ has at least as many $*$ as $p(Y_i)$ has. Therefore $t(Y_{i+1})$ has at least as many positions as $t(Y_i)$ has that agree with optimal solution $s_{OPT}$, so $d\big(t(Y_i),s_{OPT}\big) \geqslant d\big(t(Y_{i+1}),s_{OPT}\big)$. Using definition of the inaccuracy function (Definition \ref{def_ina}) we prove the lemma. $\hfill\blacksquare$ \end{proof} Intuitively $ina(Y)-ina(Y \cup \{y\})$ is the decrease of the inaccuracy from adding element $y$ to set $Y$. We will show that, when adding one more element $y$ to sets $Y,Z$ such that $Y \subseteq Z$, the inaccuracy decrease more in a case of adding $y$ to the smaller set $Y$ than adding $y$ to the bigger set $Z$. \begin{lemma}\label{supermodular_ina} If we artificially extend the $ina(\cdot)$ function for the empty set:\\ $ina(\emptyset)=2\cdot OPT$, then function $ina(\cdot)$ is supermodular\footnote{according to \cite{schrijver}, $f:2^S\mapsto \mathbb{R}$ is supermodular iff\\ $\forall_{Y,Z \subseteq S} \quad f(Y)+f(Z)\leqslant f(Y \cup Z)+f(Y\cap Z)$ which is equivalent with $\forall_{Y \subseteq Z \subseteq S} \quad\forall_{s\in S} \quad f(Z)-f(Z\cup\{s\})\leqslant f(Y)-f(Y \cup\{s\})$.}, i.e., \begin{equation}\label{supermodular_ina_ieq} \forall_{Y \subseteq Z \subseteq S} \quad\forall_{s \in S} \quad ina(Z)-ina(Z \cup \{s\}) \leqslant ina(Y)-ina(Y \cup \{s\}) \end{equation} \end{lemma} \begin{proof} Let fix $Y,Z$ and $s$ such that $Y \subseteq Z \subseteq S$ and $s \in S$. \paragraph{\textbf{Case 1:}} $Z = \emptyset$: Then also $Y = \emptyset$, and inequality (\ref{supermodular_ina_ieq}) holds obviously. \paragraph{\textbf{Case 2:}} $Z \neq \emptyset, Y=\emptyset$: We have: \[ina(Z)-ina(Z \cup \{s\}) \leqslant OPT = \] \begin{equation}\label{supermodular_ina_c2} = 2\cdot OPT-OPT \leqslant ina(\emptyset)-ina(\{s\}) = ina(Y)-ina(Y \cup \{s\}), \end{equation} because we use respectively: Lemma \ref{decreasing_ina} and the fact that $Z$ has at least one element; definition of $ina(\cdot)$ for empty set and upperbound for $ina(\cdot)$ function; assumption that $Y = \emptyset$. \paragraph{\textbf{Case 3:}} $Z \neq \emptyset, Y \neq \emptyset$: From definition of $ina(\cdot)$ we have: \[ ina(Z)-ina(Z \cup \{s\}) = d\big(t(Z),s_{OPT}\big)-d\big(t(Z \cup \{s\}),s_{OPT}\big) = \] counting a difference by considering two cases for value of $s_{OPT}$ we obtain \[ = \Big| \big\{ j: s_{OPT}[j]=1 \wedge t(Z \cup \{s\})[j]=1 \wedge t(Z)[j]=0 \big\}\Big| + \] \[ +\;\, \Big| \big\{ j: s_{OPT}[j]=0 \wedge t(Z \cup \{s\})[j]=0 \wedge t(Z)[j]=1 \big\}\Big| = \] using definition of function $t(\cdot)$: \[ = \Big| \big\{ j: s_{OPT}[j]=1 \wedge s[j]=1 \wedge \;\forall_{z \in Z} \; z[j]=0 \big\}\Big| + \] \[ +\;\, \Big| \big\{ j: s_{OPT}[j]=0 \wedge s[j]=0 \wedge \;\forall_{z \in Z} \; z[j]=1 \big\}\Big| \leqslant \] taking an universal quantifier over a smaller subset we obtain: \[ \leqslant \Big| \big\{ j: s_{OPT}[j]=1 \wedge s[j]=1 \wedge \;\forall_{y \in Y} \; y[j]=0 \big\}\Big| + \] \[ + \;\,\Big| \big\{ j: s_{OPT}[j]=0 \wedge s[j]=0 \wedge \;\forall_{y \in Y} \; y[j]=1 \big\}\Big| = \] reversing all previous transformations finally we obtain: \[ = ina(Y)-ina(Y \cup \{s\}). \]$\hfill\blacksquare$ \end{proof} \section{Existence of a stable subset} \label{sec:existence_of_subset} \begin{lemma}\label{lem_exists_x} For any fixed $R\in\mathbb{N}_{\geqslant 1}$ there exists a subset $X \subseteq S, |X| = R$ such that \begin{equation}\label{lem_x_exists_ina} \forall_{s \in S \setminus X} \quad ina(X) - ina(X \cup \{s\}) \leqslant \frac{OPT}{R}. \end{equation} We say such $X$ is $\frac{OPT}{R}$-stable. \end{lemma} It means that there exists such a subset of votes $X$ that adding one more vote into $X$ the inaccuracy decreases by at most $\frac{OPT}{R}$. \begin{proof} First, we construct $S_{\underline{r}}$ satisfying (\ref{lem_x_exists_ina}) with at most $R$ elements. Let us construct a sequence of subsets $S_1 \subset S_2 \subset \dotsc \subset S_n = S, |S_i|=i$. We take $S_1=\{s_{i_1}\}$, where $s_{i_1}$ is any element of $S$ and for $r \in \{2,3,\dots,n\}$ we take $S_{r}=S_{r-1}\cup\{s_{i_r}\}$ where $s_{i_r}$ is such a vote that after adding it the~inaccuracy function decreases the most, i.e., \begin{equation}\label{thm_sir} s_{i_r} = \argmax_{s \in S\setminus S_{r-1}} \big(ina(S_{r-1})-ina(S_{r-1} \cup \{s\})\big). \end{equation} \begin{figure}\label{fig:ina} \centering \includegraphics[scale=1.15]{ina2} \vspace{-10pt} \caption{The $ina(\cdot)$ function for the sequence of subsets $S_1 \subset S_2 \subset \dotsc \subset S_n = S$.} \end{figure} We have \[ \min_{r \in \{1,2,\dots,R\}} \quad ina(S_r)-ina(S_{r+1}) \leqslant \frac{1}{R} \left(\sum_{r=1}^R ina(S_r)-ina(S_{r+1})\right) = \] \begin{equation}\label{thm_min_r} = \frac{1}{R} \big( ina(S_1)-ina(S_{R+1}) \big) \leqslant \frac{OPT}{R}, \end{equation} because (from Lemma \ref{decreasing_ina}) we know that $ina(S_1) \leqslant OPT$ and $ina(S_{R+1}) \geqslant 0$. Let $\underline{r}$ be a minimizer for the left-hand side of (\ref{thm_min_r}), then (by the choice of $s_{i_{\underline{r}}}$ in (\ref{thm_sir})) we have: \begin{equation}\label{thm_max_diff_ina} \max_{s \in S\setminus S_{\underline{r}}} \big(ina(S_{\underline{r}})-ina(S_{\underline{r}} \cup \{s\})\big)\leqslant \frac{OPT}{R}, \end{equation} thus $S_{\underline{r}}$ satisfies (\ref{lem_x_exists_ina}), see Figure 1. If $S_{\underline{r}}$ has less elements than $R$ we can extend $S_{\underline{r}}$ to an $R$-elements subset $X$ by adding any elements of $S$. It follows from the supermodularity of $ina(\cdot)$. From Lemma \ref{supermodular_ina} we have: \[\forall_{s \in S \setminus S_{\underline{r}}} \quad ina(X)-ina(X \cup \{s\}) \leqslant ina(S_{\underline{r}})-ina(S_{\underline{r}} \cup \{s\}), \] and hence also: \begin{equation}\label{thm_max_sx_x} \max_{s \in S \setminus S_{\underline{r}}} \big(ina(X)-ina(X \cup \{s\})\big) \leqslant \max_{s \in S \setminus S_{\underline{r}}} \big(ina(S_{\underline{r}})-ina(S_{\underline{r}} \cup \{s\})\big). \end{equation} Finally, taking (\ref{thm_max_diff_ina}) and (\ref{thm_max_sx_x}) we obtain: \[\max_{s \in S \setminus X} \big(ina(X)-ina(X \cup \{s\})\big) \leqslant \frac{OPT}{R}.\]$\hfill\blacksquare$ \end{proof} Of course we cannot construct such a subset efficiently if we do not know $s_{OPT}$. How to find a proper subset $X$? For constructing our PTAS we will fix $R \in \mathbb{N}_{\geqslant 1}$ and consider all subsets $Y \subseteq S$ with cardinality $R$. There is less than $n^R \in Poly(n)$ such subsets. For clarity, we will use $Y \subseteq S$ in arguments valid for all subsets considered by the algorithm, and $X \subseteq S$ for a $\frac{OPT}{R}$-stable subset of votes. For a fixed $Y \subseteq S, Y \neq \emptyset$, WLOG we reorder candidates in such a way that $p(Y)$ is a lexicographically smallest permutation: \[p(Y)=**\dotsc *00\dotsc 011\dotsc 1.\] The first part (from the left) is called ``star positions'' or ``star part''. The remaining part is called ``no-star part''. We define $p^{(*)}(Y)$ as the number of $*$ in $p(Y)$ and we denote it $\beta$: \[ \beta = p^{(*)}(Y) = \Big|\left\{ j:\big(p(Y)\big)[j]=* \right\}\Big|.\] In our PTAS we essentially fix the ``no-star part'' of the answer to the pattern $p(Y)$ and optimize over the choices for the ``star part'' of the outcome. If the~number of stars or number of 1's on star positions of $s_{OPT}$ is small enough, then there is only $Poly(m,n)$ possible solutions and we can consider all of them. Let us analyze the size of the ``star part''. \begin{lemma}\label{size_p_star_y} For all $Y \subseteq S$ we have \[\beta = p^{(*)}(Y) \leqslant |Y| \cdot OPT\] \end{lemma} \begin{proof} Consider an arbitrary $Y=\{y_1,y_2,\dots,y_{|Y|}\}$. We can construct $Y$ in the~following 3 phases: \begin{enumerate} \item[1.] $Y := \{s_{OPT}\}$ \item[2.] for $i=1$ to $|Y|$ do \item[] \quad $Y := Y \cup \{y_i\}$ \item[3.] $Y := Y \setminus \{s_{OPT}\}$ \end{enumerate} After that we obtain set Y. Let us calculate how many stars $p(Y)$ has. In Phase~1 there are no stars. In each step in Phase~2 we add at most $OPT$ stars, because $\forall_{i\in \{1,2,\dots,|Y|\}} \quad d(y_i,s_{OPT}) \leqslant OPT$. In Phase 3 we can at most decrease the~number of stars. So $\beta \leqslant |Y| \cdot OPT$.$\hfill\blacksquare$ \end{proof} Note that for $X$ from Lemma \ref{lem_exists_x} we have \begin{equation}\label{size_p_star} p^{(*)}(X) \leqslant |X| \cdot OPT = R \cdot OPT. \end{equation} Let us now introduce some more notation. Assuming $Y \subseteq S$ and hence also $\beta = p^{(*)}(Y)$ are fixed, we will use the following notation to denote the ``star part'' and the ``no-star part'' of a string $x \in \{0,1\}^m$: \[x'=x[1]\cdot x[2]\cdot \dotsc\cdot x[\beta],\] \[x''=x[\beta+1]\cdot x[\beta+2]\cdot\dotsc\cdot x[m],\] where``$\cdot$'' is a concatenation of strings (letters). So we divide $x$ into two parts: $x=x'\cdot x''$. Let us now define a $k$-completion of $x \in \{0,1\}^m$ (definition from \cite{kcompletion}) to be a $y \in \{0,1\}^m$ such that $y^{(1)}=k$ and $d(y,x)$ is the minimum possible Hamming distance between $x$ and any vector with $k$ of 1's. To obtain a $k$-completion we only add or only delete a proper number of 1's. To be more specific in this paper we assume the $k$-completion is always obtained by changing bits at positions with the smallest possible index\footnote{Any other deterministic rule would work for us just as well.}. In the following lemma we will show that for the pattern from a stable subset $X$ we can change the number of 1's in the ``no-star part'' to the properly guessed number of 1's loosing only twice the stability constant. \begin{lemma}\label{lemma_kbiscompletion} If $X \subseteq S$ is $(\epsilon_1 \cdot OPT)$-stable, $z''$ is a $k''$-completion of $\big(p(X)\big)''$, where $k''=(s_{OPT}'')^{(1)}$, then \begin{equation}\label{lemma_kbis_completion} \forall_{i \in \{1,2,\cdots,n\}} \quad d(s_{OPT}'\cdot z'',s_i) \leqslant (1+2\epsilon_1) \cdot OPT \end{equation} \end{lemma} \begin{proof} WLOG there is insufficient number of 1's in no-star part of pattern $p(X)$, i.e., $k'' \geqslant \big((p(X))''\big)^{(1)}$. The other case is symmetric. Let us fix $s_i \in S$ and consider all combinations of values in strings $\big(p(X)\big)''$, $z''$, $s_i''$, $s_{OPT}''$ at the same position $j$. $\alpha_a \in \mathbb{N}$, for $a\in \{1,2,\cdots,12\}$, counts the number of positions $j$ with combination $a$, see Table 1. \begin{table}[h]\label{table_cases} \centering \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|} \hline & \multicolumn{12}{c|}{combinations} \\ \hline \quad index of a combination \quad & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10& 11& 12\\ \hline $(p(X))''[j]$ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 \\ $z''[j]$ & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\ $s_i''[j]$ & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 \\ $s_{OPT}''[j]$ & 0 & 0 & 1 & 1 & 0 & 0 & 1 & 1 & 0 & 0 & 1 & 1 \\ \hline number of occurrences & $\alpha_1$ & $\alpha_2$ & $\alpha_3$ & $\alpha_4$ & $\alpha_5$ & $\alpha_6$ & $\alpha_7$ & $\alpha_8$ & $\alpha_9$ & $\alpha_{10}$ & $\alpha_{11}$ & $\alpha_{12}$ \\ \hline $d(z''[j],s_i''[j])$ & 0 & 1 & 0 & 1 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 \\ \hline $d(s''_{OPT}[j],s_i''[j])$ & 0 & 1 & 1 & 0 & 0 & 1 & 1 & 0 & 0 & 1 & 1 & 0 \\ \hline \end{tabular} \vspace{5pt} \caption{Combinations of values in strings $(p(X))'', z'', s_i'', s_{OPT}''$. There is only 12 combinations (no $2^4=16$), because by the assumption $k'' \geqslant ((p(X))'')^{(1)}$ we never change from 1 in $((p(X))'')^{(1)}$ to 0 in $z''$.} \vspace{-15pt} \end{table} We have: \[ d(z'',s_i'') = \big|\{j: z''[j] \neq s_i''[j] \}\big| = \] we consider two cases for value of $s_{OPT}$ at position $j$: \[ = \big|\{j: z''[j] \neq s_i''[j] \wedge (z''[j] = s_{OPT} \vee z''[j] \neq s_{OPT})\}\big| = \] we divide it into two components: \begin{alignat*}{4} = &\big|\{j: s_{OPT} = && \; z''[j] \neq s_i''[j] &&\}\big| + \\ + &\big|\{j: && \; z''[j] \neq s_i''[j] = s_{OPT}) &&\}\big| = \end{alignat*} we use case counts from Table 1 to count positions in both components: \[ = ( \underbrace{\alpha_2 + \alpha_7 + \alpha_{11}}_\text{first component} + \underbrace{\alpha_3 + \alpha_6 + \alpha_{10} - \alpha_3 - \alpha_6 - \alpha_{10}}_{=0}) + \underbrace{(\alpha_4 + \alpha_5 + \alpha_9)}_\text{second component} = \] and we use the definition of the Hamming distance: \begin{equation}\label{d_z_bis_s_i_bis} = \big(d(s_{OPT}'',s_i'') - \alpha_3 - \alpha_6 - \alpha_{10}\big) + (\alpha_4 + \alpha_5 + \alpha_9). \end{equation} Since $(z'')^{(1)} = k'' = \big(s_{OPT}''\big)^{(1)}$, \[ \sum_{k=5}^{12} \alpha_k = \alpha_3 + \alpha_4 + \alpha_7 + \alpha_8 + \alpha_{11} + \alpha_{12} \] \begin{equation}\label{alpha5} \alpha_5 = \alpha_3 + \alpha_4 - \alpha_6 - \alpha_9 - \alpha_{10}. \end{equation} Also \begin{equation}\label{epsilon1opt_stable} \alpha_4 + \alpha_8 + \alpha_9 \leqslant \epsilon_1 \cdot OPT, \end{equation} because $X$ is $\epsilon_1 \cdot OPT$-stable. Now we are ready to prove equation (\ref{lemma_kbis_completion}). \[ d(s_{OPT}'\cdot z'',s_i) \stackrel{\rm def.}{=} d(s_{OPT}',s_i') + d(z'',s_i'') \stackrel{(\ref{d_z_bis_s_i_bis})}{=} \] \[ \stackrel{(\ref{d_z_bis_s_i_bis})}{=} d(s_{OPT}',s_i') + d(s_{OPT}'',s_i'') - \alpha_3 - \alpha_6 - \alpha_{10} + \alpha_4 + \alpha_5 + \alpha_9 \stackrel{(\ref{alpha5})}{=} \] \[ \stackrel{(\ref{alpha5})}{=} \underbrace{d(s_{OPT},s_i)}_{\leqslant OPT} + 2(\hspace{-10pt} \underbrace{\alpha_4}_{\stackrel{(\ref{epsilon1opt_stable})}{\leqslant} \epsilon_1 \cdot OPT} \hspace{-15pt} - \alpha_6 - \alpha_{10}) \stackrel{(\ref{epsilon1opt_stable})}{\leqslant} (1+2\epsilon_1) \cdot OPT. \] $\hfill\blacksquare$ \end{proof} \section{An auxiliary optimization problem} \label{sec:aux_prob} In this section we will consider the optimization problem obtained after guessing the number of 1's in the two parts and fixing the ``no-star part'' of the outcome. It has variables for all the positions of the ``star part'' and constraints for all the original votes $s_i \in S$. Let us define the optimization problem in terms of the integer program $IP_{(\ref{ip_q})-(\ref{ip_p01})} (Y,k')$ by (\ref{ip_q})-(\ref{ip_p01}): \begin{equation}\label{ip_q} \min q \end{equation} \begin{equation}\label{ip_no1} (s')^{(1)} = k' \end{equation} \begin{equation}\label{ip_dist} \forall_{i \in \{1,2,\dots,n\}} \quad d(s',s_i') \leqslant q-d(s_{ALG}'',s_i'') \end{equation} \begin{equation}\label{ip_qg0} q \geqslant 0 \end{equation} \begin{equation}\label{ip_p01} \forall_{j \in \{1,2,\dots,\beta\}} \quad s'[j] \in \{0,1\} \end{equation} where $Y \subseteq S, k=k'+k''$, and $s_{ALG}''$ is the $k''$-completion of $(p(Y))''$. Recall that $\beta=p^{(*)}(Y)$ and $(p(Y))''$ is the ``no-star part'' of the~pattern $p(Y)$. In the LP relaxation (\ref{ip_p01}) is replaced with: \begin{equation}\label{lp_p01} \forall_{j \in \{1,2,\dots,\beta\}} \quad s'[j] \in [0,1] \end{equation} Constraints (\ref{ip_q})-(\ref{ip_qg0}),(\ref{lp_p01}) are linear because \[(s')^{(1)}=\sum_{j=1}^\beta s'[j],\] \[d(s',s_i') = \sum_{j=1}^\beta \Big(\chi(s_i'[j]=0)\cdot s'[j] + \chi(s_i'[j]=1)\cdot(1-s'[j])\Big)\] are linear functions of $s'[j]$, where $j \in \{1,2,\dots,\beta\}$. \begin{lemma}\label{lem_aprox_ip} $\forall_{R \in \mathbb{N}_{\geqslant 1}, Y \subseteq S, |Y| \leqslant R, k' \in \mathbb{N}, \epsilon_2 > 0}$ we can find $(1+2\epsilon_2)$-approximation solution for $IP_{(\ref{ip_q})-(\ref{ip_p01})} (Y,k')$ by solving the $LP$ and considering at most \[ (3n)^{\frac{3R \ln(2)}{(\epsilon_2)^2}} + m^{\frac{3R^2\ln(6)}{(\epsilon_2)^2}} \quad\text{cases.} \] \end{lemma} \begin{proof} Let us fix constants $\epsilon_2 \in (0,\frac{1}{2})$ (for $\epsilon_2 \geqslant \frac{1}{2}$ we could use $2$-approximation from \cite{markakis}). Consider three cases: \paragraph{\textbf{Case 1:}} $\beta \leqslant \frac{3 R\ln(3n)}{(\epsilon_2)^2}$\\ There is $2^\beta$ possibilities for $s'$.\\ \[2^\beta \leqslant 2^{\frac{3R\ln(3n)}{(\epsilon_2)^2}} = e^{\ln(3n) \frac{3R \ln(2)}{(\epsilon_2)^2}} = (3n)^{\frac{3R \ln(2)}{(\epsilon_2)^2}} \in Poly(n),\] because $\epsilon_2$ and $R$ are fixed constants. So we will check (in polynomial time) all possibilities for $s'$ and we will find optimal solution for the integer program. \paragraph{\textbf{Case 2:}} $k'\leqslant \frac{3R^2\ln(6)}{(\epsilon_2)^2}$\\ There is $Poly(m)$ possibilities for $s'$ because we can upperbound the number of possibilities of setting 1's into $\beta$ positions by: \[ {\beta \choose k'} \leqslant \beta^{k'} \leqslant \beta^{\frac{3R^2\ln(6)}{(\epsilon_2)^2}} \leqslant m^{\frac{3R^2\ln(6)}{(\epsilon_2)^2}} \in Poly(m), \] because $\epsilon_2$ and $R$ are fixed constants. \paragraph{\textbf{Case 3:}} $\;\beta > \frac{3R\ln(3n)}{(\epsilon_2)^2} \;\wedge\; k' > \frac{3R^2\ln(6)}{(\epsilon_2)^2}$\\ We denote an optimal solution of the~$IP_{(\ref{ip_q})-(\ref{ip_p01})} (Y,k')$ by $\big((s')^{IP},q^{IP}\big)$. Let us use LP relaxation and denote an optimal solution of the LP by $\big((s')^{LP},q^{LP}\big)$. Obviously we have $q^{LP} \leqslant q^{IP}$. We can solve the LP in polynomial time but we may obtain a fractional solution. We want to round it independently. We will use a randomized rounding defined by distributions on each position $j \in \{1,2,\dots,\beta\}$: \begin{equation}\label{pse1} P\big(s'[j]=1\big)= (s')^{LP}[j], \quad P\big(s'[j]=0\big)= 1-(s')^{LP}[j]. \end{equation} We can estimate the expected value of a distance to such a random solution $s'$: \[ \forall_{i \in \{1,2,\cdots,n\}} \quad \mathbb{E}\big[ d(s',s_i') \big] \stackrel{\text{def.}}{=} \mathbb{E}\left[ \sum_{j=1}^\beta \Big|s'[j]-s_i'[j]\Big| \right] = \] \begin{alignat*}{6} = \mathbb{E}\Bigg[ &\sum_{j=1}^\beta& \Big( \quad &\chi(s_i'[j]=0)\cdot s'[j] \quad &+ \quad &\chi(s_i'[j]=1)\cdot (1-s'[j]) &&\Big) \Bigg] \stackrel{\text{lin. of }\mathbb{E}}{=} \\ \stackrel{\text{lin. of }\mathbb{E}}{=} &\sum_{j=1}^\beta& \Big( \quad &\chi(s_i'[j]=0)\cdot \mathbb{E}\big[s'[j]\big]\quad &+ \quad &\chi(s_i'[j]=1)\cdot \mathbb{E}\big[1-s'[j]\big] &&\Big) \stackrel{(\ref{pse1})}{=} \\ \stackrel{(\ref{pse1})}{=} &\sum_{j=1}^\beta& \Big( \quad &\chi(s_i'[j]=0)\cdot (s')^{LP}[j]\quad &+ \quad &\chi(s_i'[j]=1)\cdot \big(1-(s')^{LP}[j]\big) &&\Big) \stackrel{\text{def.}}{=} \end{alignat*} \begin{equation}\label{exp_d_p_si} \stackrel{\text{def.}}{=} d\big((s')^{LP},s_i'\big) \stackrel{(\ref{ip_dist})}{\leqslant} q^{LP}-d(s_{ALG}'',s_i''). \end{equation} $d(s',s_i')$ is a sum of $\beta$ independent 0-1 variables. For $\epsilon' \in (0,1)$ using Chernoff's bound \cite{motvani} we have: \[ P\Big(d(s',s_i') \geqslant (1+\epsilon') \cdot\mathbb{E}\big[d(s',s_i')\big] \Big) \leqslant \exp\left( -\frac{1}{3}(\epsilon')^2 \cdot \mathbb{E}\big[d(s',s_i')\big] \right). \] If we take $\epsilon' = \frac{\epsilon_2 \cdot q^{IP}}{\mathbb{E}[d(s',s_i')]}$ then we obtain: \[ \exp\left( -\frac{1}{3} \cdot \frac{(\epsilon_2)^2 \cdot (q^{IP})^2}{\mathbb{E}\big[d(s',s_i')\big]} \right) \geqslant P\Big(d(s',s_i') \geqslant \mathbb{E}\big[d(s',s_i')\big] + \epsilon_2 \cdot q^{IP} \Big) \stackrel{(\ref{exp_d_p_si})}{\geqslant} \] \begin{equation}\label{expgpbb} \stackrel{(\ref{exp_d_p_si})}{\geqslant} P\Big(d(s',s_i') \geqslant q^{LP}-d(s_{ALG}'',s_i'') + \epsilon_2 \cdot q^{IP} \Big). \end{equation} We want to know an upperbound for the probability that we make an error greater than $\epsilon_2 \cdot q^{IP}$ for at least one vote: \[ P\Big(\exists_{i \in \{1,2,\dots,n\}}: d(s',s_i') \geqslant q^{LP}-d(s_{ALG}'',s_i'') + \epsilon_2 \cdot q^{IP} \Big) \stackrel{(\ref{expgpbb})}{\leqslant} \] \begin{equation}\label{pbb_error_all} \stackrel{(\ref{expgpbb})}{\leqslant} n \cdot \exp\left( -\frac{1}{3} \cdot \frac{(\epsilon_2)^2 \cdot (q^{IP})^2}{\mathbb{E}\big[d(s',s_i')\big]} \right) \leqslant n \cdot \exp\left( -\frac{1}{3} (\epsilon_2)^2 \cdot q^{IP} \right), \end{equation} where the last inequality is because of: \[ \mathbb{E}\big[ d(s',s_i') \big] \stackrel{(\ref{exp_d_p_si})}{\leqslant} q^{LP}-d(s_{ALG}'',s_i'') \leqslant q^{IP}. \] We want to further upperbound the probability in (\ref{pbb_error_all}). From the assumption about $\beta$ and from Lemma~\ref{size_p_star_y} we have: \[ \frac{3R\ln(3n)}{(\epsilon_2)^2} < \beta \stackrel{\rm{Lem. \ref{size_p_star_y}}}{\leqslant} |Y| \cdot OPT \leqslant R \cdot OPT \leqslant R \cdot q^{IP}, \quad \text{equivalently}\] \begin{equation}\label{1deltalnexp} \frac{1}{3} > n \cdot \exp\left( -\frac{1}{3} (\epsilon_2)^2 \cdot q^{IP} \right). \end{equation} So, finally we have: \begin{equation}\label{pexistslqd} P\Big(\exists_{i \in \{1,2,\dots,n\}}: d(s',s_i') \geqslant q^{LP}-d(s_{ALG}'',s_i'') + \epsilon_2 \cdot q^{IP} \Big) \stackrel{(\ref{pbb_error_all}),(\ref{1deltalnexp})}{<} \frac{1}{3}. \end{equation} So with probability at least $\frac{2}{3}$ we obtain: \[ \forall_{i\in\{1,2,\dots,n\}} \quad d(s' \cdot s_{ALG}'', s_i) = d(s', s_i') + d(s_{ALG}'', s_i'') \stackrel{(\ref{pexistslqd})}{<} \] \begin{equation}\label{eps2_approx_ip} \stackrel{(\ref{pexistslqd})}{<} q^{LP}-d(s_{ALG}'', s_i'') +\epsilon_2\cdot q^{IP} + d(s_{ALG}'', s_i'') \leqslant (1+\epsilon_2)\cdot q^{IP}. \end{equation} We can also obtain a wrong number o 1's. The solution $s_{ALG}'$ for that is to take the $k'$-completion of $s'$. We will show that the additional error for such operation is not so big. Expected number of 1's in $s'$ is equal $k'$: \[ \mathbb{E}\big[ (s')^{(1)} \big] \stackrel{\rm{def.}}{=} \mathbb{E}\left[ \sum_{j=1}^\beta s'[j] \right] \stackrel{\text{lin. of }\mathbb{E}}{=} \sum_{j=1}^\beta (s')^{LP}[j] \stackrel{\rm{def.}}{=} \big((s')^{LP}\big)^{(1)} \stackrel{(\ref{ip_no1})}{=} k'. \] We want to know how much we lose taking the $k'$-completion. Similar as before, $(s')^{(1)} = \sum_{j=1}^\beta s'[j]$ is a sum of $\beta$ independent 0-1 variables. For $\epsilon'' \in (0,1)$ using Chernoff's bound \cite{motvani} we have: \[ P\left( (s')^{(1)} \geqslant (1+\epsilon'')\cdot k' \right) \leqslant \exp\left( -\frac{1}{3}(\epsilon'')^2 \cdot k' \right), \] \[ P\left( (s')^{(1)} \leqslant (1-\epsilon'')\cdot k' \right) \leqslant \exp\left( -\frac{1}{2}(\epsilon'')^2 \cdot k' \right). \] Taking both inequalities together, $\epsilon'' = \frac{\epsilon_2}{R}$ and using assumption $k' > \frac{3R^2\ln(6)}{(\epsilon_2)^2}$ we have: \[ P\left( \left|(s')^{(1)}-k'\right| \geqslant \epsilon''\cdot k' \right) \leqslant 2\cdot \exp\left( -\frac{1}{3}(\epsilon'')^2 \cdot k' \right) \leqslant \] \[ \leqslant 2\cdot \exp\left( -\frac{1}{3}\frac{(\epsilon_2)^2}{R^2} \cdot k' \right) < \frac{1}{3}. \] So with probability at least $\frac{2}{3}$ the error from taking the $k'$-completion is not greater than $ \epsilon'' \cdot k' = \frac{\epsilon_2}{R} \cdot k' \leqslant \frac{\epsilon_2}{R} \cdot \beta \stackrel{\rm{Lem. \ref{size_p_star_y}}}{\leqslant} \frac{\epsilon_2}{R} \cdot |Y| \cdot OPT \leqslant \epsilon_2 \cdot OPT \leqslant \epsilon_2 \cdot q^{IP}$. Combining the above with (\ref{eps2_approx_ip}) we obtain a $(1+2\epsilon_2)$-approximate solution with probability at least $\frac{1}{3}$. We may derandomize the algorithm analogously to how it was done in the PTAS for the Closest String problem \cite{ptascs}. For more on derandomization techniques see~\cite{derandomization}. $\hfill\blacksquare$ \end{proof} \section{Algorithm and its complexity analysis} \label{sec:alg} Now we are ready to combine the ideas into a single algorithm. \begin{algorithm} \caption{ALG(R)} \label{alg_mav} \begin{algorithmic}[1] \REQUIRE $S=\{s_1,s_2,\dots,s_n\} \in (\{0,1\}^m)^n, 0 \leqslant k \leqslant m,R \in \mathbb{N}_{\geqslant 1}$ \ENSURE $s_{ALG} \in \{0,1\}^m$ \FOR{each $R$-element subset $Y=\{s_{i_1},s_{i_2},\dots,s_{i_R}\} \subseteq S$}\label{forRsubset} \FOR{each division $k$ into two parts $k=k'+k''$} \STATE{$s''_{ALG} \leftarrow k''$-completion of $(p(Y))''$ \\(if not possible, then skip this inner iteration)} \STATE{$s'_{ALG} \leftarrow$ approximation solution of $IP_{(\ref{ip_q})-(\ref{ip_p01})} (Y,k')$ using Lemma \ref{lem_aprox_ip}\\ (if $LP_{(\ref{ip_q})-(\ref{ip_qg0}),(\ref{lp_p01})}(Y,k')$ infeasible, then skip this inner iteration)} \label{approx_opt_problem} \STATE{evaluate $s'_{ALG}\cdot s''_{ALG}$ by computing $ \max_{i\in \{1,2,\dots,n\}} d(s_i,s'_{ALG}\cdot s''_{ALG})$} \ENDFOR \ENDFOR\label{forRsubsetend} \STATE{$s_{ALG} \leftarrow$ the best solution from a loop in lines \ref{forRsubset}-\ref{forRsubsetend}} \end{algorithmic} \end{algorithm} It remains to argue that for a large enough parameter $R$ the above algorithm will at some point consider a subset of votes $X$ that leads to an accurate enough approximation of the Minimax objective function of our problem. \begin{theorem} $\forall_{\epsilon \in (0,1)}$ we may compute a $(1+\epsilon)$-approximate solution to Minimax Approval Voting in $O\big(Poly(n,m)\big)$ time. \end{theorem} \begin{proof} Let $\epsilon_0 = \frac{\epsilon}{3} < \frac{1}{3}$. By Lemma \ref{lem_exists_x}, there exists an $\frac{\epsilon_0 \cdot OPT}{2}$-stable set of votes $X\subseteq S$ of cardinality $|X|=R=\lceil \frac{2}{\epsilon_0} \rceil$. Consider algorithm ALG(R). In one iteration it will consider $X$ and $k',k''$ such that $(s_{OPT}')^{(1)}=k'$. Recall that $s_{ALG}''$ is the specific $k''$-completion of $\big(p(X)\big)''$. By Lemma \ref{lemma_kbiscompletion} we have: \[ d(s'_{OPT} \cdot s_{ALG}'',s_i) \leqslant (1+\epsilon_0)\cdot OPT,\] hence $\big(s'=s_{OPT}',q=(1+\epsilon_0)\cdot OPT \big)$ is a feasible solution to $IP_{(\ref{ip_q}-\ref{ip_p01})} (X,k')$ and the optimal value of $IP_{(\ref{ip_q}-\ref{ip_p01})} (X,k')$ is at most $(1+\epsilon_0)\cdot OPT$. By Lemma \ref{lem_aprox_ip} with $\epsilon_2 = \frac{\epsilon_0}{2}$ we find a $(1+\epsilon_0)$-approximate solution $\big( s_{ALG}',q_{ALG} \big)$ to $IP_{(\ref{ip_q}-\ref{ip_p01})} (X,k')$. So we have: \[q_{ALG} \leqslant (1+\epsilon_0) \cdot (1+\epsilon_0) \cdot OPT \stackrel{\epsilon_0<1}{\leqslant} (1+3\epsilon_0) \cdot OPT = (1+\epsilon) \cdot OPT.\] It remains to observe, that $s_{ALG}=s_{ALG}' \cdot s_{ALG}''$ is a solution to $MAV$ of cost $q_{ALG} \leqslant (1+\epsilon) \cdot OPT$. The algorithm examined $O(n^R)=O\big(n^{\lceil \frac{6}{\epsilon} \rceil} \big) \in O(Poly(n))$ subsets $Y$, $O(m)$ choices of $k'$ and each time considered \[ O\left( (3n)^{\frac{108 \cdot \lceil \frac{6}{\epsilon} \rceil \cdot \ln(2) }{\epsilon^2}} + m^{\frac{108 \cdot {\lceil \frac{6}{\epsilon} \rceil}^2 \cdot \ln(6)}{\epsilon^2}} \right) \in O(Poly(n,m)) \;\, \text{cases.}\] $\hfill\blacksquare$ \end{proof} \section{Concluding remarks} We showed the existence of a PTAS for Minimax Approval Voting by considering all subsets of a fixed size $R$. If not the discovered supermodularity for the~inaccuracy function $ina(\cdot)$, we would simply consider all subsets of size at most $R$. Although the supermodularity was not essential for our result, it shows that larger subsets of votes are generally more stable (in the sense of definition in Lemma \ref{lem_exists_x}). It seems to suggest that an algorithm considering a smaller number of larger subsets of votes would potentially be more efficient in practice. Perhaps the most interesting open question is whether by randomly sampling a number of subsets of votes to examine, one could obtain a more practical FPRAS for the~problem. Another interesting direction is the optimization of the Minimax objective function subject to a restriction that the voting system must be incentive compatible. According to~\cite{markakis} the best possible approximation ratio in this setting is between $2-\frac{2}{k+1}$ and $3-\frac{2}{k+1}$, and a natural challenge is to narrow this gap. Finally, we know the complexity of the two extreme objectives, i.e., Minimax and Minisum. The latter is easily optimized by selecting the $k$ most often approved candidates. The optimization problem for intermediate objectives such as optimizing the sum of squares of the Hamming distances remains unexplored, and it would be interesting to learn which objective functions are more difficult to approximate than Minimax in the context of Approval Voting systems. \section*{Acknowledgments} We want to thank Katarzyna Staniewicz for many helpful proofreading comments. Also we want to thank reviewers for their valuable suggestions. Krzysztof Sornat was supported by local grant 2139/M/II/14.

https://arxiv.org/abs/1901.07834

Algorithms for the computation of the matrix logarithm based on the double exponential formula

We consider the computation of the matrix logarithm by using numerical quadrature. The efficiency of numerical quadrature depends on the integrand and the choice of quadrature formula. The Gauss--Legendre quadrature has been conventionally employed; however, the convergence could be slow for ill-conditioned matrices. This effect may stem from the rapid change of the integrand values. To avoid such situations, we focus on the double exponential formula, which has been developed to address integrands with endpoint singularity. In order to utilize the double exponential formula, we must determine a suitable finite integration interval, which provides the required accuracy and efficiency. In this paper, we present a method for selecting a suitable finite interval based on an error analysis as well as two algorithms, and one of these algorithms addresses error control.

\section{Introduction} A logarithm of $A\in\mathbb{R}^{n\times n}$ is defined as any matrix $X$ such that \begin{align}\label{eq:def_mat_log} \exp(X)=A,\end{align} where $\exp(X):=I+X+\frac{1}{2!}X^2+\frac{1}{3!}X^3+\cdots$ \cite[p.~269]{higham2008functions}. If all eigenvalues of $A$ lie in the set $\mathbb{C}\setminus (-\infty,0]$, there is a unique logarithm of $A$ whose eigenvalues all lie in the strip $\{z\in\mathbb{C}:-\pi<\mathrm{Im}(z)< \pi\}$ \cite[Thm. 1.31]{higham2008functions}. This logarithm is called the principal logarithm of $A$, and denoted by $\log(A)$. Throughout this paper, we assume that all eigenvalues of $A$ lie in the set $\mathbb{C}\setminus (-\infty,0]$, and we consider the principal logarithm of $A$. The matrix logarithm is utilized in many fields of research, such as quantum mechanics \cite{cosmas2007classical}, quantum chemistry \cite{hesselmann2010random}, biomolecular dynamics \cite{horenko2008likelihood}, buckling simulation \cite{schenk2009modeling}, and machine learning \cite{han2015large,huang2017riemannian,don2017scalable,fitzsimons2017entropic}. The computational methods include the inverse scaling and squaring (ISS) algorithm \cite{al2012improved}, an algorithm based on the arithmetic-geometric mean (AGM) iteration \cite{cardoso2016matrix}, and numerical quadrature. In this paper, we focus on numerical quadrature, which employs the following integral representation (see e.g. \cite[Thm. 11.1]{higham2008functions}): \begin{align}\label{eq:1-01} \log(A) = (A-I) \int_0^1 \bigl[t(A-I) + I\bigr]^{-1} \,\mathrm{d} t. \end{align} First, numerical quadrature can make use of the sparseness of $A$ if $A$ is sparse. It is useful when computing the multiplication of the matrix logarithm and a vector $\log(A)\bm{b}~(\bm{b}\in \mathbb{R}^n)$, which appears in applications such as \cite{han2015large,don2017scalable,fitzsimons2017entropic}\footnote{ In the three applications, the computation of $\log(\det A)$ is performed based on some connections between $\log(A)\bm{b}$ and $\log(\det A)$. For more details, see, e.g., \cite{hutchinson1990stochastic}. }, without computing and storing dense matrices. Conversely, the ISS algorithm and the algorithm based on the AGM iteration include the computation of the matrix square root, which means that the calculation involves dense matrices even if $A$ is sparse. The second reason is that numerical quadrature is potentially more favorable for parallel computers because of independent computation of the integrand on each abscissa. Because the integrand in \eqref{eq:1-01} includes matrix inversion, the computational cost of numerical quadrature depends on the number of evaluations of the integrand. Although numerical quadrature is suitable for parallelization, the quadrature formula should be selected carefully to reduce the computational cost and save computational resources. The method conventionally used to compute \eqref{eq:1-01} is the Gauss--Legendre (GL) quadrature. If the spectral radius of $A-I$ is smaller than 1, the GL quadrature, which can be regarded as a rational approximation of $\log(A)$, coincides with the Pad\'{e} approximants of $\log(A)$ at $I$ \cite[Thm. 4.3]{dieci1996computational}. Therefore, it is natural to use the GL quadrature to reduce the number of abscissas when $A$ is close to $I$. However, the convergence of the GL quadrature becomes slow when $A$ is not close to $I$. For example, the convergence in our experiments became slow when $A$ was ill-conditioned which may be explained by rapid changes in the integrand value when it is closer to the endpoint of the interval. In this paper, we consider the double exponential (DE) formula \cite{takahashi1974double}, which can be used to compute integrals with singularities at one (or both) of the endpoints. For this reason, the DE formula may be useful in scenarios in which the GL quadrature does not perform well. However, when using the DE formula, a finite interval needs to be selected because the integrand in \eqref{eq:1-01} is transformed into a corresponding function on the infinite interval. This selection is important, i.e., if the finite interval is too narrow, the accuracy of the computational result becomes low, but if it is too wide, the convergence of the DE formula becomes slow. By performing an error analysis, we provide a method of selecting the appropriate finite interval, as well as two algorithms for the computation of $\log(A)$ based on the $m$-point DE formula. The remainder of this paper is organized as follows: in Section 2, we present an error analysis and propose two algorithms; in Section 3, we show the results of numerical experiments; in Section 4, we conclude the study. \textbf{Notation:} Unless otherwise stated, $\|\cdot\|$ denotes a consistent matrix norm, e.g., the $p$-norm or the Frobenius norm, and $\|\cdot\|_2$ and $\|\cdot\|_F$ denote the 2-norm and the Frobenius norm, respectively. The inverse functions of $\sinh$ and $\tanh$ are referred to as $\mathrm{asinh}$ and $\mathrm{atanh}$, respectively. \section{Algorithms for the computation of $\log(A)$ based on the DE formula} In this section we propose a method of selecting a finite interval for the DE formula by estimating the interval truncation error and present two algorithms. Before considering the truncation error, let us apply the following transformations to \eqref{eq:1-01}. By substituting $u=2t-1$ in \eqref{eq:1-01}, we obtain \begin{align}\label{eq:2-01} \log(A) = (A-I)\int_{-1}^1 \left[(1+u)(A-I)+2I\right]^{-1} \,\mathrm{d} u. \end{align} Then, by applying the DE transformation $u = \tanh(\sinh(x))$, it follows that \begin{align}\label{eq:2-02} \log(A) = (A-I)\int_{-\infty}^\infty F_{\mathrm{DE}}(x) \,\mathrm{d} x, \end{align} where \begin{align} F_{\mathrm{DE}}(x) := \cosh(x)\mathrm{sech}^2(\sinh(x))\left[ (1+\tanh(\sinh(x)))(A-I)+2I \right]^{-1}. \end{align} The matrix $(1+\tanh(\sinh(x)))(A-I)+2I$ in the integrand $F_{\mathrm{DE}}$ is nonsingular for any $x\in$ $\mathbb{R}$. In Subsection 2.1, we derive an upper bound on the error between the integral in \eqref{eq:2-02} and the same integral defined in the finite interval $[l,r]$, \begin{align}\label{eq:2-51} \left\lVert \log(A) - (A-I)\int_l^r F_{\mathrm{DE}}(x) \,\mathrm{d} x \right\rVert. \end{align} In Subsection 2.2, we propose a method of selecting the interval $[l,r]$ so that the relative truncation error is small than or approximately equal to a given tolerance $\epsilon$. Our algorithms are described in Subsection 2.3. \subsection{Estimation of the error from the interval truncation} The error that stems from the interval truncation \eqref{eq:2-51} can be rewritten as \begin{align}\label{eq:2-03} \left\lVert \log(A) - (A-I)\int_l^r F_{\mathrm{DE}}(x) \,\mathrm{d} x \right\rVert = \left\lVert (A-I)\left[ \int_{-\infty}^lF_{\mathrm{DE}}(x)\,\mathrm{d} x +\int_r^\infty F_{\mathrm{DE}}(x)\,\mathrm{d} x \right] \right\rVert. \end{align} Using the triangle inequality in the right-hand side (RHS) of \eqref{eq:2-03}, it holds that \begin{align}\label{eq:2-04} & \left\lVert (A-I)\left[ \int_{-\infty}^lF_{\mathrm{DE}}(x)\,\mathrm{d} x +\int_r^\infty F_{\mathrm{DE}}(x)\,\mathrm{d} x \right] \right\rVert \\ &\quad \le \left\lVert (A-I)\int_{-\infty}^lF_{\mathrm{DE}}(x)\,\mathrm{d} x \right\rVert + \left\lVert (A-I)\int_r^\infty F_{\mathrm{DE}}(x)\,\mathrm{d} x \right\rVert . \end{align} By estimating the RHS of \eqref{eq:2-04}, we obtain an upper bound on \eqref{eq:2-51}. Initially, we focus on the first term which is on the RHS of \eqref{eq:2-04}. To avoid cumbersome notation, instead of $F_{\mathrm{DE}}$, which including hyperbolic functions, we consider the integrand of \eqref{eq:1-01}. By applying the transformation $t=[\tanh(\sinh(x))+1]/2$, we have \begin{align} \label{eq:2-05} \left\lVert (A-I)\int_{-\infty}^lF_{\mathrm{DE}}(x)\,\mathrm{d} x \right\rVert = \left\lVert (A-I)\int_0^a F(t) \,\mathrm{d} t \right\rVert, \end{align} where, \begin{align} F(t):=[t(A-I)+I]^{-1},\quad a:=\frac{\tanh(\sinh(l))+1}{2}. \end{align} The following lemma shows an upper bound on the RHS of \eqref{eq:2-05}, if $a$ is small enough to warrant the use of the Neumann series expansion of $F(t)$. \begin{lem}\label{thrm:error_l} Suppose that $A\ne I$. Then, for $a\in \bigl(0, 1/(2\|A-I\|)\bigr]$, we have \begin{align}\label{eq:2-06} \left\lVert (A-I)\int_0^a F(t) \,\mathrm{d} t \right\rVert \le \frac{3a}{2} \|A-I\|. \end{align} \end{lem} \begin{pf} For all $t\in[0,a]$ where $a\in \bigl(0, 1/(2\|A-I\|)\bigr]$, it holds that \begin{align} \left\lVert t(I-A) \right\rVert \le \frac{1}{2}~ (<1). \end{align} By applying the Neumann series expansion to $F(t)$ we get the following: \begin{align} F(t) = [t(A-I)+I]^{-1} = \sum_{k=0}^\infty t^k(I-A)^k. \end{align} Therefore, the integral of $F$ can be rewritten as follows: \begin{align} (A-I)\int_0^a F(t) \,\mathrm{d} t &= (A-I)\int_0^a \left[\sum_{k=0}^\infty t^k(I-A)^k\right] \,\mathrm{d} t\label{eq:2-50}\\ &= (A-I)\sum_{k=0}^\infty\frac{a^{k+1}}{k+1}(I-A)^k\\ &= a(A-I) + a(A-I)\sum_{k=1}^\infty\left[\frac{a^k}{k+1}(I-A)^k\right]. \end{align} By using triangle inequality and consistency of the norm we get the following: \begin{align} \left\lVert (A-I)\int_0^a F(t) \,\mathrm{d} t \right\rVert &\le a \left\lVert A-I \right\rVert + a \left\lVert A-I \right\rVert \sum_{k=1}^\infty\left[ \frac{1}{k+1} \left\lVert a(A-I) \right\rVert ^k \right]\\ &\le a \left\lVert A-I \right\rVert + a \left\lVert A-I \right\rVert \sum_{k=1}^\infty\left[ \frac{1}{2} \left(\frac{1}{2}\right)^k \right]\\ &= \frac{3a}{2} \left\lVert A-I \right\rVert . \end{align} \hfill $\square$ \end{pf} The calculation to estimate the second term on the RHS of \eqref{eq:2-04} is similar to that of the first term. By applying the transformation $t=[\tanh(\sinh(x))+1]/2$, we get the following: \begin{align}\label{eq:2-08} \left\lVert (A-I)\int_r^\infty F_{\mathrm{DE}}(x) \,\mathrm{d} x \right\rVert = \left\lVert (A-I)\int_b^1 F(t) \,\mathrm{d} t \right\rVert , \end{align} where $b=[\tanh(\sinh(r))+1]/2$. The following lemma shows an upper bound on \eqref{eq:2-08}, if $b$ is close enough to 1 to warrant the use of the Neumann series expansion of $F(t)$. \begin{lem}\label{thrm:error_r} For $b \in \bigl[2 \|A^{-1}\|/(2\|A^{-1}\|+1), 1\bigr)$, we have \begin{align}\label{eq:2-10} \left\lVert (A-I)\int_b^1 F(t) \,\mathrm{d} t \right\rVert \le \left(-\log(b)+\frac{1-b}{2b}\right) \|A-I\|\|A^{-1}\|. \end{align} \end{lem} \begin{pf} The outline of this proof is similar to that of Lemma \ref{thrm:error_l}. For all $t\in[b,1]$ where $b \in \bigl[2 \|A^{-1}\|/(2\|A^{-1}\|+1), 1\bigr)$, it holds that \begin{align} \left\lVert \frac{t-1}{t}A^{-1} \right\rVert \le \frac{1}{2}~(<1). \end{align} By applying the Neumann series expansion to $F(t)$, we get \begin{align} F(t) &= [t(A-I)+I]^{-1}\\ &= \frac{1}{t}A^{-1}\sum_{k=0}^\infty\left(\frac{t-1}{t}\right)^kA^{-k}. \end{align} Therefore, the integral of $F$ can be rewritten as follows: \begin{align} (A-I)\int_b^1 F(t) \,\mathrm{d} t &= (A-I)\int_b^1 \left[\frac{1}{t}A^{-1}\sum_{k=0}^\infty\left(\frac{t-1}{t}\right)^kA^{-k}\right] \,\mathrm{d} t\\ &= (A-I)A^{-1}\left[ -\log(b)I+\sum_{k=1}^\infty A^{-k}\int_b^1 \frac{1}{t} \left(\frac{t-1}{t}\right)^k \,\mathrm{d} t \right]. \end{align} By using the triangle inequality and consistency of the norm, it follows \begin{align} \label{eq:2-07} \left\lVert (A-I)\int_b^1 F(t) \,\mathrm{d} t \right\rVert &\le \left\lVert A-I \right\rVert \left\lVert A^{-1} \right\rVert \left\{ |-\log(b)| + \sum_{k=1}^\infty\left[ \left\lVert A^{-1} \right\rVert ^k \int_b^1 \frac{1}{t} \left\lvert\frac{t-1}{t}\right\rvert^k \,\mathrm{d} t \right] \right\}. \end{align} In \eqref{eq:2-07}, it holds that $|-\log(b)|=-\log(b)$ because $b\in\bigl[2 \|A^{-1}\|/(2\|A^{-1}\|+1), 1\bigr)$, and $|(t-1)/t| = (1-t)/t$ because $t\in[b,1]$. For the second term in the bracket on the RHS of \eqref{eq:2-07}, we have: \begin{align} \sum_{k=1}^\infty\left[ \left\lVert A^{-1} \right\rVert^k \int_b^1 \frac{1}{t} \left(\frac{1-t}{t}\right)^k \,\mathrm{d} t\right] &\le\sum_{k=1}^\infty\left[ \left\lVert A^{-1} \right\rVert^k \int_b^1 \frac{1}{b} \left(\frac{1-t}{b}\right)^k \,\mathrm{d} t\right]\\ &= \frac{1-b}{b}\sum_{k=1}^\infty\left[\frac{1}{k+1} \left\lVert \frac{1-b}{b}A^{-1} \right\rVert ^k\right]\\ &\le \frac{1-b}{b} \sum_{k=1}^\infty\left[\frac{1}{2}\left(\frac{1}{2}\right)^k\right]\\ & = \frac{1-b}{2b}. \label{eq:2-09} \end{align} By substituting \eqref{eq:2-09} in \eqref{eq:2-07}, we obtain the inequality \eqref{eq:2-10}. \hfill $\square$ \end{pf} In the final part of this Subsection, we estimate the upper bound on \eqref{eq:2-51}. \begin{prop}\label{thrm:2-01} Suppose that $A\ne I$. For a given interval $[l,r]$, let $a=[\tanh(\sinh(l))+1]/2$ and $b=[\tanh(\sinh(r))+1]/2$. Then, if $a \le 1/(2\|A-I\|)$ and $b \ge 2\|A^{-1}\|/(2\|A^{-1}\|+1)$, it holds that \begin{align}\label{eq:2-111} \left\lVert \log(A) - (A-I)\int_l^r F_{\mathrm{DE}}(x) \,\mathrm{d} x \right\rVert \le \frac{3}{2}a\|A-I\| + \left(-\log(b)+\frac{1-b}{2b}\right)\|A-I\|\|A^{-1}\|. \end{align} \end{prop} \begin{pf} By combining \eqref{eq:2-03} and \eqref{eq:2-04}, and as well as substituting \eqref{eq:2-06} and \eqref{eq:2-10} in \eqref{eq:2-04}, we get \eqref{eq:2-111}. \hfill $\square$ \end{pf} \subsection{Setting the integration interval} To develop algorithms for computing $\log(A)$ based on the DE formula, we need to determine the appropriate finite integration interval, $[l,r]$ in advance. The finite interval should be ideally set so that the relative error is guaranteed to be smaller than or equal to a given tolerance, $\epsilon >0$, i.e., \begin{align}\label{eq:2-131} \frac{ \left\lVert \log(A) - (A-I)\int_l^r F_{\mathrm{DE}}(x) \,\mathrm{d} x \right\rVert }{ \|\log(A)\| } \le \epsilon. \end{align} To accomplish this, a lower bound on $\|\log (A)\|$ must be estimated. The following lemma shows a lower bound in terms of the spectral radius of $A$. \begin{lem}\label{thrm:log_lower_bound} Let $\rho(\cdot)$ be the spectral radius, i.e., the largest absolute value of eigenvalues. Then, the following two inequalities hold: \begin{align} &\|\log(A)\| \ge |\log(\rho(A))|, \label{eq:2-21}\\ &\|\log(A)\| \ge |\log(\rho(A^{-1}))|. \label{eq:2-22} \end{align} \end{lem} \begin{pf} By using the consistency of the norm and the fact any eigenvalue of $\log(A)$ is equal to $\log(\lambda)$, for some $\lambda$ being an eigenvalue of $A$, we have the lower bound in \eqref{eq:2-21} as $\|\log(A)\|\ge\rho(\log(A))\ge|\log(\rho(A))|$. The lower bound in \eqref{eq:2-22} can be obtained in a similar way. \hfill $\square$ \end{pf} In the following proposition, we show how to set a finite interval such that the relative truncation error in 2-norm is smaller than or equal to a given tolerance, $\epsilon>0$. \begin{prop}\label{thrm:settings_of_lr} Suppose that $A\ne I$. Let $\theta$ be a lower bound on $\|\log(A)\|_2$, the tolerance $\epsilon>0$ satisfy \begin{align}\label{eq:2-151} \epsilon < \frac{ 3\|A-I\|_{2}\|A^{-1}\|_{2} }{ \theta(1+\|A^{-1}\|_{2}) }, \end{align} and $s$ be a real positive solution of the equation \begin{align}\label{eq:2-133} \frac{1}{\theta}\left(-\log(s)+\frac{1-s}{2s}\right)\|A-I\|_2\|A^{-1}\|_2 = \frac{\epsilon}{2}. \end{align} If \begin{align} l := \mathrm{asinh}(\mathrm{atanh}(2a - 1)),\quad r := \mathrm{asinh}(\mathrm{atanh}(2b - 1)), \end{align} where \begin{align}\label{eq:2-132} a:= \min\left\{\frac{\theta\epsilon}{3\|A-I\|_2},\frac{1}{2\|A-I\|_2}\right\}, \quad b:=\max\left\{s,\frac{2\|A^{-1}\|_2}{2\|A^{-1}\|_2+1}\right\}, \end{align} then, it holds that \begin{align}\label{eq:2-30} \frac{ \left\lVert \log(A) - (A-I)\int_l^r F_{\mathrm{DE}}(x) \,\mathrm{d} x \right\rVert_2 }{ \|\log(A)\|_2 } \le \epsilon. \end{align} \end{prop} \begin{pf} From the definition of $a$ and $b$, it is true that $a\le 1/(2\|A-I\|_2)$ and $b\ge 2\|A^{-1}\|_2/(2\|A^{-1}\|_2+1)$. In addition, from \begin{align} b-a &\ge \frac{2\|A^{-1}\|_2}{2\|A^{-1}\|_2+1} - \frac{\theta\epsilon}{3\|A-I\|_2} > \frac{2\|A^{-1}\|_2}{2\|A^{-1}\|_2+1} - \frac{\|A^{-1}\|_2}{1+\|A^{-1}\|_2}\\ &=\frac{2\|A^{-1}\|_2(1+\|A^{-1}\|_2) -(2\|A^{-1}\|_2+1)\|A^{-1}\|_2 }{(2\|A^{-1}\|_2+1)(1+\|A^{-1}\|_2)}\\ &=\frac{\|A^{-1}\|_2}{(2\|A^{-1}\|_2+1)(1+\|A^{-1}\|_2)} > 0, \end{align} it follows that $a<b$, and $l<r$. Therefore, we can choose a finite interval $[l,r]$ that satisfies the assumptions of Proposition \ref{thrm:2-01}. By dividing the inequality \eqref{eq:2-111} by $\|\log(A)\|_2\ge \theta$, it follows: \begin{align}\label{eq:2-31} \frac{ \left\lVert \log(A) - (A-I)\int_l^r F_{\mathrm{DE}}(x) \,\mathrm{d} x \right\rVert _2}{\|\log(A)\|_2} \le \frac{3a}{2\theta}\|A-I\|_2+\frac{1}{\theta}\left(-\log(b)+\frac{1-b}{2b}\right)\|A-I\|_2\|A^{-1}\|_2. \end{align} From the definition of $a$ and $b$, it holds that $a\le \theta\epsilon/(3\|A-I\|_2)$ and $b\ge s$. Therefore, \begin{align}\label{eq:2-32} \frac{3a}{2\theta}\|A-I\|_2 \le \frac{\epsilon}{2},\quad \frac{1}{\theta}\left(-\log(b)+\frac{1-b}{2b}\right)\|A-I\|_2\|A^{-1}\|_2 \le \frac{\epsilon}{2}. \end{align} We obtain \eqref{eq:2-30} by substituting \eqref{eq:2-32} in \eqref{eq:2-31}. \hfill $\square$ \end{pf} \subsection{Algorithms} In this subsection we establish two algorithms based on the results in subsections 2.1 and 2.2. One of the algorithms is designed to compute $\log(A)$ using the $m$-point DE formula on a finite interval with an interval truncation error smaller than or approximately equal to a given tolerance, $\epsilon>0$. The other algorithm is an adaptive quadrature algorithm designed to compute $\log(A)$ by automatically adding abscissas until the error is smaller than or approximately equal to a given tolerance $\zeta>0$. If the tolerance $\epsilon$ given in Proposition \ref{thrm:settings_of_lr} is sufficiently small, a linear approximation to the nonlinear equation \eqref{eq:2-133} can be used to determine an appropriate interval. We describe our calculation in detail below. Suppose that $\epsilon$ is sufficiently small and the solution $s$ of \eqref{eq:2-133} is approximately equal to 1. Then, because $3(1-s)/2$ is the first-order Taylor approximation to $-\log(s)+(1-s)/2s$ at $s =1$, the solution $s$ can be approximated by using the solution $\tilde{s}$ of the following equation: \begin{align}\label{eq:2-140} \frac{1}{\theta} \frac{3(1-\tilde{s})}{2} \|A-I\|_2\|A^{-1}\|_2 = \frac{\epsilon}{2}. \end{align} The solution of \eqref{eq:2-140} is given by \begin{align} \tilde{s} = 1-\frac{\theta\epsilon}{3\|A-I\|_2\|A^{-1}\|_2}. \end{align} Under the assumptions of Proposition \ref{thrm:settings_of_lr}, by choosing $\tilde{b}$ as \begin{align} \tilde{b} = \max \left\{ \tilde{s},~ \frac{2\|A^{-1}\|_2}{2\|A^{-1}\|_2+1} \right\} \end{align} instead of \eqref{eq:2-132}, and setting $\tilde{r}=\mathrm{asinh}(\mathrm{atanh}(2\tilde{b}-1))$, the interval truncation will be smaller than or approximately equal to $\epsilon$. The summary of the first algorithm, which computes $\log(A)$ using the $m$-point DE formula whose interval truncation error is smaller than or approximately equal to $\epsilon$ is as shown in Algorithm \ref{alg:1}. \begin{algorithm}[htbp] \caption{Computation of $\log(A)$ based on the DE formula.} \begin{algorithmic}[1] \Statex \textbf{Input:} $A\in\mathbb{R}^{n\times n}$, $m\in\mathbb{N}$, $\epsilon>0$ a tolerance for the interval truncation error \Statex \textbf{Output:} $X\approx \log(A)$ \State Set $F_{\mathrm{DE}}(x)=\cosh(x)\mathrm{sech}^2(\sinh(x))\left[ (1+\tanh(\sinh(x)))(A-I)+2I \right]^{-1}$ \State Compute $\|A-I\|_2$, $\|A^{-1}\|_2$, $\rho(A)$ \State $\theta = |\log(\rho(A))|$ \State $\displaystyle\epsilon_{\mathrm{max}} = \frac{3}{\theta} \frac{\|A-I\|_2\|A^{-1}\|_2}{1+\|A^{-1}\|_2}$ \If {$\displaystyle \epsilon \ge \epsilon_{\mathrm{max}}$} \State $\displaystyle \epsilon \gets \epsilon_{\mathrm{max}}/2$ \EndIf \State $\displaystyle a = \min \left\{ \frac{\theta\epsilon}{3\|A-I\|_2}, \frac{1}{2\|A-I\|_2} \right\}$ \State $\displaystyle b = \max \left\{ 1-\frac{\theta\epsilon}{3\|A-I\|_2\|A^{-1}\|_2}, \frac{2\|A^{-1}\|_2}{2\|A^{-1}\|_2+1} \right\}$ \State $l = \mathrm{asinh}(\mathrm{atanh}(2a - 1))$ \State $r = \mathrm{asinh}(\mathrm{atanh}(2b - 1))$ \State $h = (r-l)/(m-1)$ \State $\displaystyle T = \frac{h}{2}(F_{\mathrm{DE}}(l)+F_{\mathrm{DE}}(r)) + h\sum_{i=1}^{m-2}F_{\mathrm{DE}}(l+ih)$ \State $X = (A-I)T$ \end{algorithmic} \label{alg:1} \end{algorithm} When $\epsilon$ is sufficiently small, an accurate computation of $\|I-A\|_2$, $\|A^{-1}\|_2$ and $\rho(A)$ at Step 2 of Algorithm \ref{alg:1} may not be required because the errors that stem from of these values have little effect on the accuracy of $\log(A)$. We give more detail in the following paragraph. Assume that $\epsilon$ in Algorithm \ref{alg:1} is sufficiently small, and that $a$ and $b$ in Step 9 are chosen as \begin{align} a = \frac{\theta\epsilon}{3\|A-I\|_2},\quad b = 1-\frac{\theta\epsilon}{3\|A-I\|_2\|A^{-1}\|_2}, \end{align} where $\theta$ is a lower bound on $\|\log(A)\|_2$. Let $\Delta_1$ and $\Delta_2$ be the relative errors of $\theta/\|A-I\|_2$ and $1/\|A^{-1}\|_2$ respectively. Then, the computational result of $a$ is equal to \begin{align} \frac{\epsilon}{3}\frac{\theta}{\|A-I\|_2}(1+\Delta_1) = \frac{\theta}{3\|A-I\|_2} \epsilon(1+\Delta_1), \end{align} and that of $b$ is equal to \begin{align} 1-\frac{\epsilon}{3}\frac{\theta}{\|A-I\|_2}(1+\Delta_1) \frac{1}{\|A^{-1}\|_2}(1+\Delta_2) = 1-\frac{\theta}{3\|A-I\|_2\|A^{-1}\|_2}\epsilon(1+\Delta_1+\Delta_2+\Delta_1\Delta_2). \end{align} Therefore, the upper bound on the truncation error $\epsilon$ computed by considering the relative errors $\Delta_1,~\Delta_2$ is almost equal to the upper bound on the truncation error when the tolerance is set as $\epsilon(1+\Delta_1+\Delta_2+\Delta_1\Delta_2)$. For example, when $\Delta_1,\Delta_2=10^{-2}$, $\epsilon(1+\Delta_1+\Delta_2+\Delta_1\Delta_2)\approx 1.02\epsilon$, which means that the upper bound on the truncation error changes by approximately 2\%. If $\epsilon$ is sufficiently small, the effect of these errors will be negligible. At Step 3, a lower bound on $\|\log(A)\|_2$ is computed based on \eqref{eq:2-21}. By setting $\theta=\max\{|\log(\rho(A))|,$ $|\log(\rho(A^{-1}))|\}$, a tighter lower bound can be obtained. In particular, when $A$ is positive definite, $|\log(\rho(A^{-1}))|$ can be obtained without additional computation because $\rho(A^{-1})=\|A^{-1}\|_2$ is already computed in Step 2. The computational cost of Algorithm 1 for dense $A$ is $(2m + 2) n^3 + \mathcal{O}(n^2)$. When $A$ is sparse, evaluating $\log(A) \bm{b}$ using Algorithm 1 has computational cost $mc_{\mathrm{abscissa}} + c_{\mathrm{mul}} + c_{\mathrm{param}}$, where $c_{\mathrm{abscissa}}$ is the computational cost of computing $F_{\mathrm{DE}}(x)\bm{b}$, $c_{\mathrm{mul}}$ is the computational cost of a matrix-vector multiplication, and $c_{\mathrm{param}}$ is the computational cost of computing parameters $\rho(A)$, $\|A-I\|_2$, and $\|A^{-1}\|_2$. If the parameters are computed approximately and $F_{\mathrm{DE}}(x)\bm{b}$ is computed accurately, then $c_{\mathrm{param}}$ will be smaller than $c_{\mathrm{abscissa}}$. Therefore, the computational cost will largely depend on $m$. Once an appropriate finite interval is obtained, the accuracy of the DE formula can be improved with the following procedure. Let $m$ be the number of abscissas, $h=(r-l)/(m-1)$ be the mesh size, and $T(h)$ be the trapezoidal rule for the mesh size $h$: \begin{align} T(h) := \frac{h}{2} \left(F_{\mathrm{DE}}(l) + F_{\mathrm{DE}}(r)\right) + h \sum_{i=1}^{m-2} F_{\mathrm{DE}}(l+ih). \end{align} Then, $T(h/2)$ can be computed using $T(h)$: \begin{align} T\left(\frac{h}{2}\right) = \frac{1}{2}T(h) + \frac{h}{2}\sum_{i=1}^{m-1} F_{\mathrm{DE}}\left(l+(2i-1)\frac{h}{2}\right). \end{align} In addition, we can apply the following estimation of the trapezoidal error for a sufficiently small value of $h$ using \cite[Eq. (4.3)]{cardoso2012computation}: \begin{align}\label{eq:2-40} \left\lVert \int_l^r F_{\mathrm{DE}}(x) - T\left(\frac{h}{2}\right) \right\rVert \approx \frac{1}{3} \left\lVert T(h) - T\left(\frac{h}{2}\right) \right\rVert. \end{align} Our adaptive quadrature algorithm which is based on \eqref{eq:2-40} is presented as Algorithm \ref{alg:2}. \begin{algorithm} \caption{Computation of $\log(A)$ by adaptive quadrature based on the DE formula.} \begin{algorithmic}[1] \Statex \textbf{Input:} $A\in\mathbb{R}^{n\times n}$, $m_0\in\mathbb{N}$, $\epsilon>0$ a tolerance for the interval truncation error, $\zeta>0$ a tolerance for the trapezoidal truncation error. \Statex \textbf{Output:} $X\approx \log(A)$ \State Set $F_{\mathrm{DE}}(x)=\cosh(x)\mathrm{sech}^2(\sinh(x))\left[ (1+\tanh(\sinh(x)))(A-I)+2I \right]^{-1}$ \State Computing $l,~r,~\theta$ using steps 2 to 11 of Algorithm 1 \State $h_0 = (r-l) / (m_0-1)$ \State $T_0 = \frac{h_0}{2}F_{\mathrm{DE}}(l) + \frac{h_0}{2}F_{\mathrm{DE}}(r) + h_0\sum_{i=1}^{m_0-2} F_{\mathrm{DE}}(l+ih)$ \For{$k=0,1,2,\ldots$ until convergence} \State $h_{k+1}=h_k/2$ \State $T_{k+1}=\frac{1}{2}T_k + h_{k+1}\sum_{i=1}^{m_k-1} F_{\mathrm{DE}}(l+(2i-1)h_{k+1})$ \State $m_{k+1}=2m_k-1$ \If {$\frac{1}{3}\|T_{k+1}-T_k\|/\theta \le \zeta$} \State $T=T_{k+1}$ \State \textbf{break} \EndIf \EndFor \State $X = (A-I)T$ \end{algorithmic} \label{alg:2} \end{algorithm} The computational cost of Algorithm \ref{alg:2} for dense $A$ is $(2m_{k+1}+2)n^3 + \mathcal{O}(n^2)=[2^{k+1}(m_0-1)+4]n^3 + \mathcal{O}(n^2)$, and the computational cost of $\log(A)\bm{b}$ with sparse $A$ is $[2^k(m_0-1)+1]c_{\mathrm{abscissa}} + c_{\mathrm{mul}} + c_{\mathrm{param}}$. \section{Numerical experiments} The numerical experiments were carried out using Julia 1.0 on a Core-i7 (3.4GHz) CPU with 16GB RAM. We used the IEEE double precision arithmetic. We computed abscissas and weights in the GL quadrature with \texttt{QuadGK.jl} (https://github.com/JuliaMath/QuadGK.jl). \subsection{Experiment 1: Checking the convergence} In this experiment, we checked the convergence of the GL quadrature and the DE formula. Our test matrices are presented in Table \ref{tab:test_matrices}. \begin{table}[htbp] \caption{Test matrices. The condition number of $A$ is denoted by $\kappa_2(A)=\|A\|_2\|A^{-1}\|_2$.} \centering \begin{tabular}{lrll} \hline Matrix & \multicolumn{1}{l}{Size} & $\kappa_2(A)$ & Structure \\ \hline \texttt{SPD 1} & 50 & $1.0\times 10^1$ & SPD\\ \texttt{SPD 2} & 50 & $1.0\times 10^4$ & SPD\\ \texttt{SPD 3} & 50 & $1.0\times 10^7$ & SPD\\ \texttt{parter} \cite{zhang2016matrix} & 10 & $2.4\times 10^{0}$ & Nonsymmetric \\ \texttt{frank} \cite{zhang2016matrix} & 10 & $2.9\times 10^{7}$ & Nonsymmetric \\ \texttt{vand} \cite{zhang2016matrix} & 10 & $3.1\times 10^{12}$ & Nonsymmetric \\ \texttt{bcsstk02} \cite{davis2011university} & 66 & $4.3\times 10^{3}$ & SPD \\ \texttt{bcsstk03} \cite{davis2011university} & 112 & $6.8\times 10^{6}$ & SPD \\ \texttt{ck104} \cite{davis2011university} & 104 & $5.5\times 10^{3}$ & Nonsymmetric \\ \hline \end{tabular} \label{tab:test_matrices} \end{table} We generated the first three matrices in Table \ref{tab:test_matrices} by using the following procedure: \begin{enumerate} \item We generated an orthogonal matrix $Q$ by QR decomposition of a random $50\times 50$ matrix. \item We generated a diagonal matrix whose diagonal elements were from the geometric sequence: $\{d_i\}_{i=1,\ldots,50}$ where $d_1=\kappa^{-1/2}$ and $d_{50}=\kappa^{1/2}$ for $\kappa=10^1$. \item $A = Q D Q^\top$. \item We repeated Step 2 and Step 3 by setting $\kappa$ as $10^4$ and $10^7$. \end{enumerate} The experimental procedure is as follows: \begin{enumerate} \item We scaled the test matrices as $\tilde{A} = (10/\rho(A)) A$ because some matrices had values that were too large to use in computation. \item We computed the reference $\log(\tilde{A})$ with the arbitrary precision arithmetic and rounded the result to double precision. We implemented the ISS algorithm \cite[Alg. 11.10]{higham2008functions} with the \texttt{BigFloat} type of Julia. \item We computed $\log(\tilde{A})$ using Algorithm \ref{alg:1}, where $\epsilon=2^{-53}\approx 1.1\times 10^{-16}$. If the test matrix was symmetric positive definite, we set $\theta=\max\{|\log(\rho(\tilde{A}))|,|\log(\rho(\tilde{A}^{-1}))|\}$ as stated in Subsection 2.3. We computed $\rho(\tilde{A})$ using the \texttt{eigvals} function of Julia, which computes all eigenvalues of $\tilde{A}$. Similarly, we computed $\|I-\tilde{A}\|_2$ and $\|\tilde{A}^{-1}\|_2$ using the \texttt{svdvals} function, which computes all singular values of $\tilde{A}$ \footnote{ If $A$ is large, using \texttt{eigvals} and \texttt{svdvals} may be inefficient. Instead, for Julia, a package \texttt{Arapack.jl} (https://github.com/JuliaLinearAlgebra/Arpack.jl) is available, which can compute a part of eigenvalues and singular values based on the Lanczos (or the Arnoldi) process with the desired accuracy. We present some numerical results, for which $\rho(\tilde{A})$, $\|\tilde{A}-I\|_2$, and $\|\tilde{A}^{-1}\|_2$ are computed with low accuracy, as shown in Appendix A. }. \item We computed $\log(\tilde{A})$ by applying the GL quadrature to \eqref{eq:2-01}. \end{enumerate} \begin{figure}[htbp]\centering \includegraphics[width=0.95\linewidth]{fig1.pdf} \caption{Convergence histories of the DE formula (Algorithm \ref{alg:1}) and the GL quadrature. The vertical axes show the relative error, $\|\log(\tilde{A})-X\|_F/\|\log(\tilde{A})\|_F$, and the horizontal axes show the number of abscissas, $m$.} \label{fig:1} \end{figure} Figure \ref{fig:1} shows the convergence histories of the DE formula and the GL quadrature for each matrix. Several observations can be made: \begin{itemize} \item The accuracy of the DE formula is almost the same as that of the GL quadrature, and the accuracies of the DE formula and the GL quadrature depend on the condition number of test matrices. \item For well-conditioned matrices, such as \texttt{SPD 1} and \texttt{parter} matrix, the GL quadrature converged faster than the DE formula. Conversely, for the ill-conditioned matrices, such as \texttt{SPD 3} and \texttt{vand} matrix, the DE formula converged faster than the GL quadrature. \end{itemize} The above observations suggest that Algorithm \ref{alg:1} selects an appropriate interval and the DE formula is suitable for ill-conditioned matrices. \subsection{Experiment 2: Checking Algorithm 2} In this experiment, we check the performance of Algorithm \ref{alg:2} by using the same matrices that were used in Experiment 1 (see Subsection 3.1). We compared Algorithm \ref{alg:2} with Algorithm \ref{alg:3}, which is based on the GL quadrature (see Appendix B). We conducted the experiment using the following procedure: \begin{enumerate} \item We computed $\log(\tilde{A})$ by using Algorithm \ref{alg:2}. We set $m_0=16$ and $\zeta=\epsilon\in\{10^{-8},10^{-11}\}$. In Step 2 of Algorithm \ref{alg:2}, which calls Algorithm 1, we computed the spectral radius and the 2-norm of matrices using \texttt{eigvals} and \texttt{svdvals} functions, as is done in Experiment 1. We stopped the computation once the number of integrand evaluations reached 1921. \item We computed $\log(\tilde{A})$ by using Algorithm \ref{alg:3}. We set $m_0=16$ and $\zeta\in\{10^{-8},10^{-11}\}$. In Step 2 of Algorithm \ref{alg:3}, the lower bound $\theta$ was computed using \eqref{eq:2-21} and \eqref{eq:2-22} in the same way as was done for the DE formula. If the number of integrand evaluations was more than 2032, we stopped the computation. \end{enumerate} \begin{table}[htbp] \centering \caption{Comparison between Algorithm \ref{alg:2} (based on the DE formula) and Algorithm \ref{alg:3} (based on the GL quadrature), in terms of the number of integrand evaluations (in bold) and the relative error when the algorithm stopped (in parentheses). The notation ``\textbf{--} (--)'' means that the algorithms did not stop before the number of integrand evaluations reached the limit.} \begin{tabular}{l|ll|ll}\hline & \multicolumn{2}{l|}{Algorithm \ref{alg:2} (DE)} & \multicolumn{2}{l}{Algorithm \ref{alg:3} (GL)} \\ $\zeta$& $10^{-8}$ & $10^{-11}$ & $10^{-8}$ & $10^{-11}$ \\\hline \texttt{SPD 1} & \textbf{61} ($2.2 \times 10^{-9}$) & \textbf{61} ($2.7 \times 10^{-12}$) & \textbf{48} ($4.6 \times 10^{-16}$) & \textbf{112} ($5.7 \times 10^{-16}$) \\ \texttt{SPD 2} & \textbf{121} ($6.7 \times 10^{-10}$) & \textbf{241} ($6.4 \times 10^{-13}$) & \textbf{1008} ($1.8 \times 10^{-15}$) & \textbf{1008} ($1.8 \times 10^{-15}$) \\ \texttt{SPD 3} & \textbf{241} ($3.0 \times 10^{-10}$) & \textbf{481} ($4.9 \times 10^{-13}$) & \textbf{--} (--) & \textbf{--} (--) \\ \texttt{parter} & \textbf{61} ($2.6 \times 10^{-9}$) & \textbf{121} ($2.3 \times 10^{-12}$) & \textbf{112} ($3.3 \times 10^{-16}$) & \textbf{112} ($3.3 \times 10^{-16}$) \\ \texttt{frank} & \textbf{481} ($1.0 \times 10^{-12}$) & \textbf{1921} ($2.1 \times 10^{-13}$) & \textbf{496} ($1.5 \times 10^{-11}$) & \textbf{--} (--) \\ \texttt{vand}& \textbf{--} (--) & \textbf{--} (--) & \textbf{--} (--) & \textbf{--} (--) \\ \texttt{bcsstk02} & \textbf{121} ($2.8 \times 10^{-9}$) & \textbf{121} ($3.1 \times 10^{-12}$) & \textbf{496} ($1.7 \times 10^{-15}$) & \textbf{1008} ($1.0 \times 10^{-15}$) \\ \texttt{bcsstk03} & \textbf{241} ($1.4 \times 10^{-9}$) & \textbf{241} ($1.5 \times 10^{-12}$) & \textbf{--} (--) & \textbf{--} (--) \\ \texttt{ck104} & \textbf{121} ($6.7 \times 10^{-10}$) & \textbf{121} ($7.4 \times 10^{-13}$) & \textbf{496} ($2.2 \times 10^{-15}$) & \textbf{496} ($2.2 \times 10^{-15}$) \\ \hline \end{tabular} \label{tab:result2} \end{table} The number of integrand evaluations and the corresponding relative error when the two algorithms stopped are shown in Table \ref{tab:result2}. Several observations can be made: \begin{itemize} \item Algorithm \ref{alg:2} successfully computed the logarithm with the desired accuracy within 2000 integrand evaluations for all test matrices, except \texttt{vand} matrix, whereas Algorithm \ref{alg:3} could not succeed for three matrices. For \texttt{vand} matrix, the stopping criterion $\zeta$ may be too strict because the convergence history of \texttt{vand} matrix (as shown in Figure \ref{fig:1}) hardly reach the value of $10^{-8}$. Our future studies will focus on the method for selecting a suitable stopping criterion. \item Even if the condition number of $A$ is small as is the case for \texttt{SPD 1} and \texttt{parter} matrix, the number of integrand evaluations of Algorithm \ref{alg:2} could be smaller than that of Algorithm \ref{alg:3} because Algorithm \ref{alg:2} can reuse all previous results when improving accuracy. \end{itemize} These observations show that Algorithm \ref{alg:2} can be a practical choice for the computation of the matrix logarithm by numerical quadrature. \section{Conclusion} In this paper, we focused on the DE formula as a new choice for the numerical quadrature formula of $\log(A)$. In order to utilize the DE formula, we proposed a method for selecting an appropriate finite interval based on error analysis, and we proposed two algorithms for practical computation. We carried out two numerical experiments. The first experiment showed that the DE formula converged faster than the GL quadrature for ill-conditioned matrices. The second experiment demonstrated that the proposed adaptive quadrature algorithm worked well with appropriate stopping criteria. Our future work will focus on three problems. The first one is the analyses of the convergence rate for the DE formula and the GL formula, the second one is a method of selecting appropriate stopping criteria, and the third one is the verification of the practical performance of the presented algorithms, when applied to large sparse matrices from current research problems.

https://arxiv.org/abs/0711.1761

The low-dimensional structures formed by tricategories

We form tricategories and the homomorphisms between them into a bicategory, whose 2-cells are certain degenerate tritransformations. We then enrich this bicategory into an example of a three-dimensional structure called a locally cubical bicategory, this being a bicategory enriched in the monoidal 2-category of pseudo double categories. Finally, we show that every sufficiently well-behaved locally cubical bicategory gives rise to a tricategory, and thereby deduce the existence of a tricategory of tricategories.

\section{A bicategory of tricategories}\label{bicat-tricat} We begin by describing the lowest-dimensional structure into which tricategories and their homomorphisms can form themselves. At first, one might think that this would be a category; but unfortunately, composition of trihomomorphisms fails to be associative on the nose, as it requires one to compose 1-cells in a hom-bicategory, which is itself not an associative operation. Consequently, the best we can hope for is a \emph{bicategory} of tricategories, which we will denote by $\cat{Tricat}_2$. The simplest such bicategory would have trihomomorphisms as its 1-cells and \emph{blips} as its 2-cells. According to \cite{GPS}, blips are very degenerate tritransformations which can only exist between two trihomomorphisms $F, G \colon \S \to {\mathcal T}$ which agree on 0-, 1-,~2-\ and 3-cells. Though one might think that this forces $F$ and $G$ to be the same, they can in fact differ with respect to certain pieces of coherence data, and a ``blip'' is the means by which one measures these differences. However, if we are going to form a bicategory of tricategories, it may as well be the most general possible one; and so we will consider more general sorts of both 1-~and 2-cells. Let us begin by looking at the 1-cells. \begin{Defn} Let $\S$ and ${\mathcal T}$ be tricategories. A \defn{lax homomorphism} $F \colon \S \to {\mathcal T}$ is a lax morphism of tricategories in the sense of \cite{GPS}, all of whose coherence 3-cells are invertible. Hence $F$ consists in: \begin{itemize} \item A function $F \colon \ob \S \to \ob {\mathcal T}$; \item Homomorphisms of bicategories $F_{A, B} \colon \S(A, B) \to {\mathcal T}(FA, FB)$; \item 2-cells $\iota_A \colon I_{FA} \to FI_A$; \item 2-cells $\chi_{f, g} \colon Fg.Ff \Rightarrow F(gf)$, pseudo-natural in $f$ and $g$; \item Invertible modifications $\omega$, $\delta$ and $\gamma$ witnessing the coherence of the data for $\iota$ and $\chi$, \end{itemize} all subject to the axioms for a morphism of tricategories as found in \cite{GPS}. \end{Defn} The notion of lax homomorphism is a sensible one from many angles. We can compose lax homomorphisms just as we would compose homomorphisms of tricategories. If we are given a pair of monoidal bicategories \cite{monoidalbicats} which we view as one-object tricategories, then the lax homomorphisms between them are the natural bicategorical generalisation of a lax monoidal functor (the \emph{weak monoidal homomorphisms}, in the terminology of \cite{monoidalbicats}). Lax homomorphisms from $1$ into a tricategory ${\mathcal T}$ classify \emph{pseudomonads} in ${\mathcal T}$~--~that is, monads whose associativity and unit laws have been weakened to hold up to coherent isomorphism, and in a similar vein we may use lax homomorphisms to give a succinct definition of an \emph{enriched bicategory} in the sense of \cite{Carmody, Lack}~--~that is of a bicategory ``enriched in a tricategory'', which is a one-dimension-higher version of a category enriched in a bicategory, which is in turn a generalisation of the familiar notion of a category enriched in a monoidal category. We shall see a little more of enriched bicategories in Section \ref{Sec:locallydouble}. Let us now discuss the 2-cells of $\cat{Tricat}_2$. The most informative precedent is the corresponding notion one dimension down: the \emph{icons} of \cite{LP}. As we mentioned in the Introduction, these are certain degenerate oplax transformations which arise between homomorphisms of bicategories which agree on 0-cells. To be precise, given two such homomorphisms $F, G \colon {\mathcal B} \to {\mathcal C}$, an \defn{icon} $\alpha \colon F \ensuremath{\Rightarrow} G$ is given by specifying for each 1-cell $f \colon A \to B$ of ${\mathcal B}$, a 2-cell $\alpha_f \colon Ff \ensuremath{\Rightarrow} Gf$ of ${\mathcal C}$ such that: \begin{itemize*} \item For each 2-cell $\sigma \colon f \ensuremath{\Rightarrow} g$ of ${\mathcal B}$, the following diagram commutes: \[\cd{ Ff \ar@2[r]^{\alpha_f} \ar@2[d]_{F \sigma} & Gf \ar@2[d]^{G \sigma} \\ Fg \ar@2[r]_{\alpha_g} & Gg\text; }\] \item For each object $A \in {\mathcal B}$, the following diagram commutes: \[\cd{ {\mathrm{id}_{FA}} \ar@2[r]^{\cong} \ar@{=}[d] & F\mathrm{id}_A \ar@2[d]^{\alpha_{\mathrm{id}_A}} \\ {\mathrm{id}_{GA}} \ar@2[r]_{\cong} & G\mathrm{id}_A\text; }\] \item For each pair of composable 1-cells $f \colon A \to B$, $g \colon B \to C$ in ${\mathcal B}$, the following diagram commutes: \[\cd{ Fg . Ff \ar@2[r]^{\cong} \ar@2[d]_{\alpha_{g} . \alpha_{f}} & F(gf) \ar@2[d]^{\alpha_{gf}} \\ Gg . Gf \ar@2[r]_{\cong} & G(gf)\text, }\] \end{itemize*} where the arrows labelled with $\cong$ are expressing the pseudo-functoriality of $F$ and $G$. Now, icons are precisely those oplax natural transformations whose components are all identities (hence the name: \emph{i}dentity \emph{c}omponent \emph{o}plax \emph{n}atural transformation), though the above description suppresses any mention of these identity components. Similarly, the 2-cells of $\cat{Tricat}_2$ we are about to describe can be seen as some sort of degenerate oplax tritransformation, with the degenerate data suppressed.\footnote{Though in order to make this precise, we would first have to define what an ``oplax tritransformation'' is.} So, let $F$, $G \colon \S \to {\mathcal T}$ be lax homomorphisms; then a 2-cell $\alpha \colon F \Rightarrow G$ of $\cat{Tricat}_2$ exists only if $F$ and $G$ agree on objects and 1-cells of $\S$, and is then given by the following data: \begin{enumerate}[(TD1)] \item For each pair of objects $A, B \in \S$, an icon \[\alpha_{A, B} \colon F_{A, B} \Rightarrow G_{A, B} \colon \S(A, B) \to {\mathcal T}(FA, FB)\] (and in particular, for each 2-cell $\theta \colon f \Rightarrow g$ of $\S$, a 3-cell of ${\mathcal T}$: \[\cd{ Ff \ar@2[r]^{F \theta} \ar@{=}[d] \dthreecell{dr}{\alpha_{\theta}}& Fg \ar@{=}[d] \\ Gf \ar@2[r]_{G \theta} & Gg }\quad\text{);}\] \item For each object $A$ of $\S$, a 3-cell of ${\mathcal T}$: \[ \cd{ I_{FA} \ar@2[r]^{\iota^F_A} \ar@{=}[d] \dthreecell{dr}{M^{\alpha}_A} & FI_A \ar@{=}[d] \\ I_{GA} \ar@2[r]_{\iota^G_A} & GI_A\text; }\] \item For each pair of composable 1-cells $f \colon A \to B$, $g \colon B \to C$ of $\S$, a 3-cell of ${\mathcal T}$: \[ \cd{ Fg . Ff \ar@2[r]^{\chi^F_{f, g}} \ar@{=}[d] \dthreecell{dr}{\Pi^{\alpha}_{f, g}} & F(g f) \ar@{=}[d] \\ Gg . Gf \ar@2[r]_{\chi^G_{f, g}} & G(g f)\text; }\] \end{enumerate} subject to the following axioms: \begin{enumerate}[(T{A}1)] \item For each pair of 2-cells $\theta \colon f \Rightarrow g \colon B \to C$ and $\theta' \colon f' \Rightarrow g' \colon A \to B$ of $\S$, the following pasting equality holds: \[ \hphantom{=} \quad \cd[@[email protected]]{ & Ff . Ff' \ar@2[r]^{\chi^F} \ar@{=}[dl] \dlthreecell{d}{\Pi^{\alpha}} & F(ff') \ar@{=}[dl] \ar@2[dr]^{F(\theta \theta')} \dlthreecell{dd}{\alpha_{\theta \theta'}} \\ Gf . Gf' \ar@2[r]^{\chi^G} \ar@2[dr]_{G\theta . G\theta'} & G(ff') \ar@2[dr]^{G(\theta \theta')} \drthreecell{d}{\chi^G} & & F(g g') \ar@{=}[dl] \\ & Gg . Gg' \ar@2[r]_{\chi^G} & G(gg')\text. }\]\[ = \quad \cd[@[email protected]]{ & Ff . Ff' \ar@2[r]^{\chi^F} \ar@{=}[dl] \ar@2[dr]_{F\theta . F\theta'} \dlthreecell{dd}{\alpha_\theta . \alpha_{\theta'}} & F(ff') \ar@2[dr]^{F(\theta \theta')} \drthreecell{d}{\chi^F} \\ Gf . Gf' \ar@2[dr]_{G\theta . G\theta'} & & Fg . Fg' \ar@2[r]^{\chi^F} \ar@{=}[dl] \dlthreecell{d}{\Pi^\alpha} & F(gg')\text; \ar@{=}[dl] \\ & Gg . Gg' \ar@2[r]_{\chi^G} & G(gg') }\] \item For each 1-cell $f \colon A \to B$ of $\S$, the following pasting equality holds: \[ \hphantom{=} \quad \cd[@!C@!R@R+1em]{ & FI_{B}.Ff \ar@2[r]^{\chi^F} \dthreecell{ddr}{\gamma^F} & F(I_B.f) \ar@2[dr]^{F\l} \\ I_{FB}.Ff \ar@2[dr]^{\l} \ar@2[ur]^{\iota^F.1} \ar@{=}[d] & & & Ff \ar@{=}[d] \\ I_{GB}.Gf \ar@2[dr]_{\l} \twoeq{r} & Ff \ar@{=}[r] \ar@{=}[d] \twoeq{dr} & Ff \ar@{=}[ur] \ar@{=}[d] \twoeq{r} & Gf \\ & Gf \ar@{=}[r] & Gf \ar@{=}[ur] }\] \[ = \quad \cd[@!C@!R@R+1em]{ & FI_B . Ff \ar@2[r]^{\chi^F} \ar@{=}[d] \dthreecell{dr}{\Pi^\alpha} & F(I_B . f) \ar@2[dr]^{F\l} \ar@{=}[d] \\ I_{FB}.Ff \ar@2[ur]^{\iota^F.1} \ar@{=}[d] \ar@{}[r] \ar@3?(0.5)+/u 0.2cm/;?(0.5)+/d 0.2cm/_(1.5)*+[l]{\scriptstyle \ \ M^\alpha.1} & GI_B . Gf \ar@2[r]_{\chi^G} \dthreecell{ddr}{\gamma^G} & G(I_B . f) \ar@2[dr]_{G \l} \dthreecell{r}{\alpha_\l} & Ff \ar@{=}[d] \\ I_{GB}.Gf \ar@2[dr]_{\l} \ar@2[ur]_{\iota^G.1} & & & Gf\text. \\ & Gf \ar@{=}[r] & Gf \ar@{=}[ur] }\] \item For each 1-cell $f \colon A \to B$ of $\S$, the following pasting equality holds: \[ \hphantom{=} \quad \cd[@!C@!R@R+1em]{ & Ff.FI_{B} \ar@2[r]^{\chi^F} \dthreecell{ddr}{\delta^F} & F(f.I_B) \ar@2[dr]^{F\r} \\ Ff.I_{FB} \ar@2[dr]^{\r} \ar@2[ur]^{1.\iota^F} \ar@{=}[d] & & & Ff \ar@{=}[d] \\ Gf.I_{GB} \ar@2[dr]_{\r} \twoeq{r} & Ff \ar@{=}[r] \ar@{=}[d] \twoeq{dr} & Ff \ar@{=}[ur] \ar@{=}[d] \twoeq{r} & Gf \\ & Gf \ar@{=}[r] & Gf \ar@{=}[ur] }\] \[ = \quad \cd[@!C@!R@R+1em]{ & Ff.FI_B \ar@2[r]^{\chi^F} \ar@{=}[d] \dthreecell{dr}{\Pi^\alpha} & F(f.I_B) \ar@2[dr]^{F\r} \ar@{=}[d] \\ Ff.I_{FB} \ar@2[ur]^{1.\iota^F} \ar@{=}[d] \ar@{}[r] \ar@3?(0.5)+/u 0.2cm/;?(0.5)+/d 0.2cm/_(1.5)*+[l]{\scriptstyle \ \ 1.M^\alpha} & Gf.GI_B \ar@2[r]_{\chi^G} \dthreecell{ddr}{\delta^G} & G(f.I_B) \ar@2[dr]_{G \r} \dthreecell{r}{\alpha_\r} & Ff \ar@{=}[d] \\ Gf.I_{GB} \ar@2[dr]_{\r} \ar@2[ur]_{1.\iota^G} & & & Gf\text. \\ & Gf \ar@{=}[r] & Gf \ar@{=}[ur] }\] \item For each triple $f, g, h$ of composable 1-cells of $\S$, the following pasting equality holds: \[ \hphantom{=} \quad \cd[@!C@!R@R+1em]{ & F(hg). Ff \ar@2[r]^{\chi^F} \dthreecell{ddr}{\omega^F} & F((hg)f) \ar@2[dr]^{F\a} \\ (Fh.Fg).Ff \ar@2[dr]^{\a} \ar@2[ur]^{\chi^F.1} \ar@{=}[d] & & & F(h(gf)) \ar@{=}[d] \\ (Gh.Gg).Gf \ar@2[dr]_{\a} \twoeq{r} & Fh.(Fg. Ff) \ar@2[r]^{1. \chi^F} \ar@{=}[d] \dthreecell{dr}{1.\Pi^\alpha} & Fh. F(gf) \ar@2[ur]^{\chi^F} \ar@{=}[d] \dthreecell{r}{\Pi^\alpha} & G(h(gf)) \\ & Gh.(Gg.Gf) \ar@2[r]_{1.\chi^G} & Gh.G(gf) \ar@2[ur]_{\chi^G} }\] \[ = \quad \cd[@!C@!R@R+1em]{ & F(hg). Ff \ar@2[r]^{\chi^F} \ar@{=}[d] \dthreecell{dr}{\Pi^\alpha} & F((hg)f) \ar@2[dr]^{F\a} \ar@{=}[d] \\ (Fh.Fg).Ff \ar@2[ur]^{\chi^F.1} \ar@{=}[d] \ar@{}[r] \ar@3?(0.5)+/u 0.2cm/;?(0.5)+/d 0.2cm/_(1.5)*+[l]{\scriptstyle \ \ \Pi^\alpha.1} & G(hg).Gf \ar@2[r]_{\chi^G} \dthreecell{ddr}{\omega^G} & G((hg)f) \ar@2[dr]_{G \a} \dthreecell{r}{\alpha_\a} & F(h(gf)) \ar@{=}[d] \\ (Gh.Gg).Gf \ar@2[dr]_{\a} \ar@2[ur]_{\chi^G.1} & & & G(h(gf))\text. \\ & Gh.(Gg.Gf) \ar@2[r]_{1.\chi^G} & Gh.G(gf) \ar@2[ur]_{\chi^G} }\] \end{enumerate} When we reach Section \ref{Sec:locallydouble}, we will see that (TD3) and (TA1) express that the maps $\Pi^\alpha$ are the components of an $(\ob \S)^3$-indexed family of ``cubical modifications'' \[ \cd{ F({\mathord{\text{--}}}) \otimes F(?) \ar@2[r]^{\chi^F} \ar@2[d]_{\alpha_{({\mathord{\text{--}}})}\otimes \alpha_{(?)}} \dthreecell{dr}{\Pi^\alpha_{A, B, C}} & F\big(({\mathord{\text{--}}}) \otimes (?)\big) \ar@2[d]^{\alpha_{(({\mathord{\text{--}}}) \otimes (?))}} \\ G({\mathord{\text{--}}})\otimes G(?) \ar@2[r]_{\chi^G} & G\big(({\mathord{\text{--}}}) \otimes (?)\big)\text, } \] where a ``cubical modification'' lives inside a square, bounded horizontally by pseudo-natural transformations and vertically by icons, and can be thought of as a modification between the oplax transformations obtained by composing up the two routes around this square. Now, in order to make this collection of 0-, 1-~and 2-cells into a bicategory, we have to give additional \emph{data}~--~vertical composition of 2-cells, horizontal composition of 1-\ and 2-cells and associativity and unitality constraints~--~subject to additional \emph{axioms}~--~the category axioms for vertical composition, the middle-four interchange axiom and the pentagon and triangle axioms for the associativity and unit constraints. We start with the vertical structure: the identity 2-cell $\mathrm{id}_F \colon F \Rightarrow F$ in $\cat{Tricat}_2$ we take to be given by the following data: \[(\mathrm{id}_F)_{A, B} = \mathrm{id}_{F_{A, B}}\text, \quad M^{\mathrm{id}_F}_A = \mathrm{id}_{\iota^F_A} \quad \text{and} \quad \Pi^{\mathrm{id}_F}_{f, g} = \mathrm{id}_{\chi^F_{f, g}} \text.\] Each of the axioms (TA1)--(TA4) now expresses that something is equal to itself pasted together with some identity 3-cells, which is clear enough. Next, given 2-cells $\alpha \colon F \Rightarrow G$ and $\beta \colon G \Rightarrow H$ in $\cat{Tricat}_2$, we take $\beta\alpha \colon F \Rightarrow H$ to be given by the following data: \[(\beta \alpha)_{A, B} = \beta_{A, B} . \alpha_{A, B}\text, \quad M^{\beta\alpha}_A = M^{\beta}_A . M^{\alpha}_A \quad \text{and} \quad \Pi^{\beta\alpha}_{f, g} = \Pi^{\beta}_{f, g} . \Pi^{\alpha}_{f, g}\text.\] Each of the axioms (TA1)--(TA4) for this data follow from juxtaposing the corresponding axioms for $\alpha$ and $\beta$ in a very straightforward manner. Moreover, because vertical composition of 3-cells in a tricategory is strictly associative and unital, so is the vertical composition of 2-cells in $\cat{Tricat}_2$. We turn now to the horizontal structure. Horizontal identities and composition for 1-cells are the identities and composition for lax homomorphisms as detailed in \cite{Lack}; whilst given 2-cells $\alpha \colon F \Rightarrow F' \colon \S \to {\mathcal T}$ and $\beta \colon G \Rightarrow G' \colon {\mathcal T} \to {\mathcal U}$, their horizontal composite $\beta \ast \alpha \colon GF \Rightarrow G'F' \colon \S \to {\mathcal U}$ is given by: \begin{enumerate}[(TD1)] \item $(\beta \ast \alpha)_{A, B} := \beta_{A, B} \ast \alpha_{A, B}$, where $\ast$ on the right-hand side is the horizontal composite of the underlying icons in the 2-category $\cat{Bicat}_2$ of the Introduction. In particular, given a 2-cell $\theta \colon f \Rightarrow g$ of $\S$, we have \[(\beta \ast \alpha)_\theta \quad = \quad \cd{ GFf \ar@2[r]^{GF \theta} \ar@{=}[d] \dthreecell{dr}{\beta_{F\theta}}& GFg \ar@{=}[d] \\ G'Ff \ar@2[r]^{G'F \theta} \ar@{=}[d] \dthreecell{dr}{G'\alpha_{\theta}}& G'Fg \ar@{=}[d] \\ G'F'f \ar@2[r]_{G'F' \theta} & G'F'g\text{;} }\] \item \[ M^{\beta \ast \alpha}_A \quad := \quad \cd{ I_{GFA} \ar@2[r]^{\iota^G} \ar@{=}[d] \dthreecell{dr}{M^{\beta}_{FA}} & GI_{FA} \ar@{=}[d] \ar@2[r]^{G\iota^F} \dthreecell{dr}{\beta_{\iota^F}} & GFI_A \ar@{=}[d] \\ I_{G'FA} \ar@2[r]^{\iota^{G'}} \ar@{=}[d] \twoeq{dr} & G'I_{FA} \ar@{=}[d] \ar@2[r]^{G'\iota^F} \dthreecell{dr}{G'M^{\alpha}_A} & G'FI_A \ar@{=}[d] \\ I_{G'F'A} \ar@2[r]_{\iota^{G'}} & G'I_{F'A} \ar@2[r]_{G'\iota^{F'}} & G'F'I_A\text,} \] \item \[ \Pi^{\beta \ast \alpha}_{f, g} \quad := \quad \cd{ GFg . GFf \ar@2[r]^{\chi^G} \ar@{=}[d] \dthreecell{dr}{\Pi^{\beta}_{Ff, Fg}} & G(Fg . Ff) \ar@2[r]^{G\chi^F} \ar@{=}[d] \dthreecell{dr}{\beta_{\chi^F}} & GF(gf) \ar@{=}[d] \\ G'Fg . G'Ff \ar@2[r]^{\chi^{G'}} \ar@{=}[d] \twoeq{dr} & G'(Fg . Ff) \ar@2[r]^{G'\chi^F} \ar@{=}[d] \dthreecell{dr}{G'\Pi^\alpha_{f, g}} & G'F(gf) \ar@{=}[d] \\ G'F'g . G'F'f \ar@2[r]_{\chi^{G'}} & G'(F'g . F'f) \ar@2[r]_{G'\chi^{F'}} & G'F'(gf)\text. }\] \end{enumerate} We must check that these data satisfy (TA1)--(TA4)\label{checkaxioms}. If we view the pasting equalities in these axioms as equating two ways round a cube or a hexagonal prism, then this verification is a matter of taking a suitable collection of such cubes and prisms for $\beta$ and $\alpha$ and sticking them together in the right way. When realised in two dimensions, this amounts to displaying a succession of equalities of rather large pasting diagrams. We leave the task of reconstructing these to the reader. Let us consider now the middle-four interchange axiom. Asking for this be satisfied amounts to checking that the other obvious way of defining $\beta \ast \alpha$~--~via $GF'$ rather than $G'F$~--~gives the same answer; and this follows quickly from the middle-four interchange law in the hom-bicategories of ${\mathcal U}$, and the first icon axiom for $\beta$. It remains to give the associativity and unit constraints $a$, $l$ and $r$ for $\cat{Tricat}_2$. For the left unit constraint $l$, consider a lax homomorphism $F \colon \S \to {\mathcal T}$, and write $F'$ for the composite $\mathrm{id}_{\mathcal T}.F \colon \S \to {\mathcal T}$. Now, $F'$ agrees with $F$ on 0-cells and on hom-bicategories, but differs in the remaining coherence data; indeed, we have \begin{align*} \iota^{F'}_{A} &= \cd{I_{FA} \ar@2[r]^-{\mathrm{id}_{I_{FA}}} & I_{FA} \ar@2[r]^-{\iota^F_{A}} & FI_A}\\ \text{and} \quad \chi^{F'}_{f, g} &= \cd{Fg. Ff \ar@2[r]^-{\mathrm{id}_{Fg.Ff}} & Fg. Ff \ar@2[r]^-{\chi^F_{f, g}} & F(gf).} \end{align*} Thus we define a 2-cell $l_F \colon \mathrm{id}_{\mathcal T} . F \Rightarrow F$ in $\cat{Tricat}_2$ as follows: \begin{enumerate*} \item[(TD1)] $(l_F)_{A, B} = \mathrm{id}_{F_{A, B}} \colon {F_{A, B}} \Rightarrow {F_{A, B}}$; \item[(TD2)] $M^{l_F}_A$ is the unit isomorphism $\iota^F_A.(\mathrm{id}_{I_{FA}}) \Rrightarrow \iota^F_A$ in the bicategory ${\mathcal T}(FA, FA)$; \item[(TD3)] $\Pi^{l_F}_{f, g}$ is the unit isomorphism $\chi^F_{f, g}.(\mathrm{id}_{Fg.Ff}) \Rrightarrow \chi^F_{f, g}$ in the bicategory ${\mathcal T}(FA, FC)$. \end{enumerate*} Now each axiom (TA1)--(TA4) is a tautology which describes how we obtained $\chi^{F'}$, $\delta^{F'}$, $\gamma^{F'}$ and $\omega^{F'}$ from the corresponding data for $F$. The definition of $r$ is dual to that of $l$, so we pass over it and onto the associativity constraint $a$. Consider three lax homomorphisms $F \colon {\mathcal R} \to \S$, $G \colon \S \to {\mathcal T}$ and $H \colon {\mathcal T} \to {\mathcal U}$ and the two composites $(HG)F$ and $H(GF) \colon {\mathcal R} \to {\mathcal U}$. As above, these agree on 0-cells and on hom-bicategories (and so we write their common value simply as $HGF$) but differ with respect to coherence data. This time we have: \begin{align*} \iota^{(HG)F} &= HG\iota^F.(H\iota^G.\iota^H)\text, & \iota^{H(GF)} &= (HG\iota^F.H\iota^G).\iota^H\text,\\ \chi^{(HG)F} &= HG\chi^F.(H\chi^G.\chi^H)\qquad\text{and} & \chi^{H(GF)} &= (HG\chi^F.H\chi^G).\chi^H\text, \end{align*} where we omit the subscripts for clarity. Thus we take $a_{F, G, H} \colon (HG)F \Rightarrow H(GF)$ in $\cat{Tricat}_2$ to be: \begin{enumerate}[(TD1)] \item $(a_{F,G,H})_{A, B} = \mathrm{id}_{(HGF)_{A, B}} \colon (HGF)_{A, B} \Rightarrow (HGF)_{A, B}$; \item $M^{a_{F,G,H}}_A$ is the associativity isomorphism \[HG\iota^F_A.(H\iota^G_{FA}.\iota^H_{GFA}) \Rrightarrow (HG\iota^F_A.H\iota^G_{FA}).\iota^H_{GFA}\] in the bicategory ${\mathcal U}(HGFA, HGFA)$; \item $\Pi^{a_{F,G,H}}_{f, g}$ is the associativity isomorphism \[HG\chi^F_{f,g}.(H\chi^G_{Ff,Fg}.\chi^H_{GFf,GFg}) \Rrightarrow (HG\chi^F_{f,g}.H\chi^G_{Ff,Fg}).\chi^H_{GFf,GFg}\] in the bicategory ${\mathcal U}(HGFA, HGFC)$. \end{enumerate} We must now verify axioms (TA1)--(TA4) for these data. Observe first that the 3-cell data $\chi$, $\gamma$, $\delta$ and $\omega$ for $H(GF)$ and for $(HG)F$ are, in fact, obtained as different bracketings of the same pasting diagram; for example, both $\omega^{H(GF)}_{f, g, h}$ and $\omega^{(HG)F}_{f, g, h}$ are obtained as pastings of the diagram on the following page. So by the pasting theorem for bicategories, we can obtain the 3-cell data $\chi$, $\gamma$, $\delta$ and $\omega$ for $H(GF)$ from that for $(HG)F$ by pasting with suitable associativity isomorphisms in the appropriate hom-bicategory of ${\mathcal U}$; and this is precisely what axioms (TA1)--(TA4) say. It remains to check the naturality of $l$, $r$ and $a$, and the pentagon and triangle identities. For the naturality of $l$, we must show that for any 2-cell $\alpha \colon F \Rightarrow G$ of $\cat{Tricat}_2$, the following diagram commutes: \[\cd{ {\mathrm{id}_{\mathcal T}} . F \ar@2[r]^{l_F} \ar@2[d]_{\mathrm{id}_{\mathcal T} . \alpha} & F \ar@2[d]^\alpha \\ {\mathrm{id}_{\mathcal T}} . G \ar@2[r]_{l_G} & G\text.} \] We easily verify that the left-hand 2-cell $\alpha' = \mathrm{id}_{\mathcal T} . \alpha$ has components $\alpha'_\theta = \alpha_\theta$, $M^{\alpha'}_A = M^{\alpha}_A.(\mathrm{id}_{I_{FA}})$ and $\Pi^{\alpha'}_{f, g} = \Pi^{\alpha}_{f, g} . (\mathrm{id}_{Ff. Fg})$; therefore the naturality of $l$ is a consequence of the naturality of the left unit constraints in the hom-bicategories of ${\mathcal T}$; and dually for $r$. For the naturality of $a$, we must show that the following diagram commutes in $\cat{Tricat}_2$ for all suitable 2-cells $\alpha$, $\beta$ and $\epsilon$: \[\cd{ (HG)F \ar@2[r]^{a_{F,G,H}} \ar@2[d]_{(\alpha\beta)\epsilon} & H(GF) \ar@2[d]^{\alpha(\beta\epsilon)} \\ (H'G')F' \ar@2[r]_{a_{F',G',H'}} & H'(G'F')\text,} \] for which we must show that (TD1)--(TD3) agree for the two ways around this square. For (TD1) this is trivial; so consider (TD2). For both $(\alpha \beta)\epsilon$ and $\alpha(\beta\epsilon)$, we obtain this datum by pasting together the same $3 \times 3$ diagram of 3-cells; the only difference being the manner in which we bracket together the boundary of this diagram. Thus the commutativity of the above square with respect to (TD2) is a further instance of the pasting theorem for bicategories. (TD3) is obtained in a similar manner. Finally, it is not hard to verify that the pentagon and triangle identities for $a$, $l$ and $r$ follow from instances of the pentagon and triangle identities in the hom-bicategories of the target tricategory. This completes the definition of the bicategory $\cat{Tricat}_2$. \begin{landscape} { \tiny \[\mkern-110mu\cd[@!0@+8em]{ & & & \sh{l}{3em} HGF(hg).HGFf \ar@2[r]^-{\chi^H} \dlthreecell{d}{(\chi^H)^{-1}} & \sh{l}{1em} H(GF(hg).GFf) \ar@2[r]^-{H\chi^G} \dlthreecell{d}{\overline{H(\chi^G)^{-1}}} & \sh{r}{1em} HG(F(hg).Ff) \ar@2[r]^-{HG\chi^F} \dthreecell{ddr}{\overline{HG\omega^F}} & \sh{r}{3em} HGF((hg)f) \ar@2[dr]^{HGF\a} \\ & & \sh{l}{3em} HG(Fh.Fg).HGFf \ar@2[ur]^{HG\chi^F.1} \ar@2[r]^-{\chi^H} \dlthreecell{d}{(\chi^H)^{-1}} & \sh{l}{1em} H(G(Fh.Fg). GFf) \ar@2[r]^-{H\chi^G} \dthreecell{ddr}{\overline{H\omega^G}} \ar@2[ur]_{H(G\chi^F.1)} & \sh{r}{1em} HG((Fh.Fg).Ff) \ar@2[dr]^{HG\a} \ar@2[ur]_{HG(\chi^F.1)} & & & \sh{l}{1.5em} HGF(h(gf)) \\ & \sh{l}{3em} H(GFh.GFg). HGFf \ar@2[r]^{\chi^H} \dthreecell{ddr}{\omega^H} \ar@2[ur]^{H\chi^G.1} & \sh{r}{3em} H((GFh.GFg).GFf) \ar@2[dr]^{H\a} \ar@2[ur]_{H(\chi^G.1)} & & & \sh{l}{3em} HG(Fh.(Fg.Ff)) \ar@2[r]_{HG(1.\chi^F)} \dlthreecell{d}{\overline{H\chi^G}} & \sh{r}{2em} HG(Fh.F(gf)) \ar@2[ur]_{HG\chi^F} \\ (HGFh.HGFg).HGFf \ar@2[dr]^{\a} \ar@2[ur]^{\chi^H.1} & & & \sh{l}{3em} H(GFh.(GFg.GFf)) \ar@2[r]^-{H(1. \chi^G)} \dlthreecell{d}{\chi^H} & \sh{r}{1em} H(GFh. G(Fg.Ff)) \ar@2[ur]_{H\chi^G} \ar@2[r]^-{H(1.G\chi^F)} \dlthreecell{d}{\chi^H} & \sh{r}{3em} H(GFh. GF(gf)) \ar@2[ur]_{H\chi^G} \\ & \sh{l}{4em} HGFh.(HGFg. HGFf) \ar@2[r]_-{1.\chi^H} & \sh{l}{1.5em} HGFh. H(GFg.GFf) \ar@2[ur]_{\chi^H} \ar@2[r]_-{1.H\chi^G} & \sh{r}{1em} HGFh. HG(Fg.Ff) \ar@2[ur]_{\chi^H} \ar@2[r]_-{1.HG\chi^F} & \sh{r}{3.5em} HGFh. HGF(gf) \ar@2[ur]_{\chi^H} }\] } \end{landscape} \endinput \section{Introduction} A major impetus behind many developments in 2-dimensional category theory has been the observation that, just as the fundamental concepts of set theory are categorical in nature, so the fundamental concepts of category theory are 2-categorical in nature. In other words, if one wishes to study categories ``in the small''~--~as mathematical entities in their own right rather than as universes of discourse~--~then a profitable way of doing this is by studying the 2-categorical properties of $\cat{Cat}$, the 2-category of all categories.\footnote{Here, and elsewhere, we will adopt a common-sense attitude to set-theoretic issues, assuming a sufficient supply of Grothendieck universes and leaving it to the reader to qualify entities with suitable constraints on their size.} Once one moves from the study of categories to the study of (possibly weak) $n$-categories, it is very natural to generalise the above maxim, and to assert that \emph{the fundamental concepts of $n$-category theory are $(n+1)$-categorical in nature.} This is a profitable thing to do: for example, consider the coherence theorem for bicategories \cite{MLP}, which in its simplest form states that \[\text{\emph{Every bicategory is biequivalent to a 2-category.}}\] \textit{A priori}, this is merely a statement about individual bicategories; but we may also read it as a statement about the tricategory $\cat{Bicat}$ of all bicategories and cells between them, since ``biequivalent'' is shorthand for ``internally biequivalent in the tricategory $\cat{Bicat}$.''\footnote{In practice, we will often use the more local definition of biequivalence, which says that ${\mathcal B}$ is biequivalent to ${\mathcal B}'$ if there exists a homomorphism $F \colon {\mathcal B} \to {\mathcal B}'$ which is surjective on objects up-to-equivalence, and which is locally an equivalence of categories; but as long as we assume the axiom of choice, the difference between the two definitions is merely one of presentation.} Thus another way of stating the above would be to say that the 2-categories are biequivalence-dense in $\cat{Bicat}$. This maxim permeates almost all research in higher-dimensional category theory, and so we draw attention to here, not in order to point out where we might use it, but rather where we might \emph{not} use it. For instance, consider once again the coherence theorem for bicategories. We may restate it slightly more tightly as: \begin{gather*}\text{\emph{Every bicategory is biequivalent to a 2-category}}\\\text{\emph{via an identity-on-objects biequivalence.}}\end{gather*} The restriction to identity-on-objects biequivalences affords us an interesting simplification, since, as pointed out in \cite{LP}, we can express such a biequivalence as a mere \emph{equivalence} in a suitable 2-category, which we denote by $\cat{Bicat}_2$. The 0-cells of $\cat{Bicat}_2$ are the bicategories; the 1-cells are the homomorphisms between them; and the 2-cells are \emph{icons}, of which we will see more in the next section: for now let us just say that they are special oplax natural transformations, which exist only between homomorphisms which agree on 0-cells. With the help of this 2-category $\cat{Bicat}_2$, the coherence theorem for bicategories can be made into a 2-categorical, rather than tricategorical, statement: namely, that the 2-categories are equivalence-dense in $\cat{Bicat}_2$ (cf. Theorem 5.4 of \cite{LP}). This is a somewhat tighter result; moreover, the 2-category $\cat{Bicat}_2$ is much simpler to work with than the tricategory $\cat{Bicat}$. Thus we should revise our general maxim, and acknowledge that \emph{some} of the fundamental concepts of $n$-category theory may be expressible using fewer than $(n+1)$ dimensions. Consequently, when we study $n$-categories, it may be useful to form them not only into an $(n+1)$-category, but also into suitable lower-dimensional structures, and it is the purpose of this paper to do so in the case where $n=3$. We will construct both a bicategory of tricategories and a tricategory of tricategories: in both cases, the 2-cells are suitably scaled-up analogues of the bicategorical icons mentioned above. However, we will see that there is some information which the mere bicategory carries which the tricategory cannot, and thus we are led to consider a new kind of low-dimensional structure, which we call a \emph{locally double bicategory}, that is rich enough to encode all the information from both the bicategory and tricategory of tricategories which we define. \textbf{Notation.}\ \ \ We follow \cite{bicats} and \cite{review} where it concerns 2-\ and bicategories: so in particular, our oplax natural transformations $\alpha \colon F \Rightarrow G$ have 2-cell components $\alpha_f \colon \alpha_B . Ff \Rightarrow Gf . \alpha_A$. We will tend to use either juxtaposition or the connective ``$.$'' to denote composition, relying on context to sort out precisely which sort of composition is intended. When it comes to tricategories, our primary references are \cite{GPS} and \cite{nicktricats}, but with a preference for the ``algebraic'' presentation of the latter: though we will not use this algebraicity in any essential way. We will also make use of \emph{pasting diagrams} of 2-cells inside bicategories. Such diagrams are only well-defined up to a choice of bracketing of their boundary, and so we assume such a choice to have been made wherever necessary. Occasionally we will need to use similar pasting diagrams of 2-cells in a tricategory, and the same caveat holds, only more so: here, the diagram is only well-defined up to a choice of order in which the pasting should be performed; and again, we assume such a choice to have been made. We adopt one further convention regarding pasting diagrams. Suppose we are given a 2-cell $\alpha \colon h(gf) \Rightarrow h'(g'f')$ in a bicategory ${\mathcal B}$, thus: \[ \cd{ & X \ar[r]^g & Y \ar[dr]^h \\ W \ar[ur]^f \ar[dr]_{f'} \dtwocell{rrr}{\alpha} & & & Z\text, \\ & X' \ar[r]_{g'} & Y' \ar[ur]_{h'}} \] together with a homomorphism of bicategories $F \colon {\mathcal B} \to {\mathcal C}$. Applying $F$ to $\alpha$ yields a 2-cell $F(h(gf)) \Rightarrow F(h'(g'f'))$ of ${\mathcal C}$, but frequently, we will be more interested in the 2-cell \[ \cd{ & FX \ar[r]^{Fg} & FY \ar[dr]^{Fh} \\ FW \ar[ur]^{Ff} \ar[dr]_{Ff'} \dtwocell{rrr}{} & & & FZ\text; \\ & FX' \ar[r]_{Fg'} & FY' \ar[ur]_{Fh'}} \] obtained by pasting $F\alpha$ with suitable coherence constraints for the homomorphism $F$: and we will consistently denote the 2-cell obtained in this way by $\overline{F\alpha}$. \endinput \section{A locally double bicategory of tricategories}\label{Sec:local} We now return to our study of tricategories with the goal of forming them into a locally double bicategory $\mathfrak{Tricat}_3$. The 0-, 1-\ and 2-cells of the bicategory $\cat{Tricat}_2$ will provide us with the 0- and 1-cells and the vertical 2-cells of $\mathfrak{Tricat}_3$, so it remains only to introduce the horizontal 2-cells and the 3-cells, which we will call \emph{pseudo-icons} and \emph{pseudo-icon modifications} respectively. In fact, following our philosophy of being as lax as possible, we begin by considering the more general \emph{oplax icons}: \begin{Defn} Let there be given lax homomorphisms of tricategories $F, G \colon \S \to {\mathcal T}$; then an \defn{oplax icon} $\alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$ may exist only if $F$ and $G$ agree on objects whereupon it consists of the following data: \begin{enumerate}[({I}D1)] \item For each $A$ and $B$ in $\S$, an oplax natural transformation \[\alpha_{A, B} \colon F_{A, B} \Rightarrow G_{A, B} \colon \S(A, B) \to {\mathcal T}(FA, FB)\] (and in particular, for each 1-cell $f \colon A \to B$ of $\S$, we have a 2-cell $\alpha_f \colon Ff \Rightarrow Gf$ of ${\mathcal T}$, and for each 2-cell $\theta \colon f \Rightarrow g$ of $\S$, a 3-cell \[\cd{ Ff \ar@2[r]^{F \theta} \ar@2[d]_{\alpha_f} \dthreecell{dr}{\alpha_{\theta}}& Fg \ar@2[d]^{\alpha_g} \\ Gf \ar@2[r]_{G \theta} & Gg\quad \text{);} }\] \item For each object $A$ of $\S$, a 3-cell of ${\mathcal T}$: \[ \cd{ I_{FA} \ar@2[r]^{\iota^F_A} \ar@{=}[d] \dthreecell{dr}{M^{\alpha}_A} & FI_A \ar@2[d]^{\alpha_{I_A}} \\ I_{GA} \ar@2[r]_{\iota^G_A} & GI_A\text; }\] \item For each $A, B$ and $C$ in $\S$, a modification \[ \cd{ F({\mathord{\text{--}}})\otimes F(?) \ar@2[r]^{\chi^F} \ar@2[d]_{\alpha_{({\mathord{\text{--}}})}\otimes \alpha_{(?)}} \dthreecell{dr}{\Pi_{A, B, C}^\alpha} & F\big(({\mathord{\text{--}}}) \otimes (?)\big) \ar@2[d]^{\alpha_{(({\mathord{\text{--}}}) \otimes (?))}} \\ G({\mathord{\text{--}}})\otimes G(?) \ar@2[r]_{\chi^G} & G\big(({\mathord{\text{--}}}) \otimes (?)\big)\text, } \] where, for instance, $F({\mathord{\text{--}}}) \otimes F(?)$ represents the homomorphism \[\S(B, C) \times \S(A, B) \xrightarrow{F \times F} {\mathcal T}(FB, FC) \times {\mathcal T}(FA, FB) \xrightarrow{\otimes} {\mathcal T}(FA, FC)\] (and in particular, for each pair of composable 1-cells $f \colon A \to B$, $g \colon B \to C$ of $\S$, we have a 3-cell of ${\mathcal T}$: \[ \cd{ Fg . Ff \ar@2[r]^{\chi^F_{f, g}} \ar@2[d]_{\alpha_g . \alpha_f} \dthreecell{dr}{\Pi^{\alpha}_{f, g}} & F(g f) \ar@2[d]^{\alpha_{gf}} \\ Gg . Gf \ar@2[r]_{\chi^G_{f, g}} & G(g f) }\quad\text{).}\] \end{enumerate} These data are subject to the following axioms: \begin{enumerate}[({IA}1)] \item For each 1-cell $f \colon A \to B$ of $\S$, the following pasting equality holds: \[ \hphantom{=} \quad \cd[@!C@!R@R+1em]{ & FI_{B}.Ff \ar@2[r]^{\chi^F} \dthreecell{ddr}{\gamma^F} & F(I_B.f) \ar@2[dr]^{F\l} \\ I_{FB}.Ff \ar@2[dr]^{\l} \ar@2[ur]^{\iota^F.1} \ar@2[d]_{1 . \alpha_f} & & & Ff \ar@2[d]^{\alpha_f} \\ I_{GB}.Gf \ar@2[dr]_{\l} \twocong{r} & Ff \ar@{=}[r] \ar@2[d]^{\alpha_f} \twoeq{dr} & Ff \ar@{=}[ur] \ar@2[d]^{\alpha_f} \twoeq{r} & Gf \\ & Gf \ar@{=}[r] & Gf \ar@{=}[ur] }\] \[ = \quad \cd[@!C@!R@R+1em]{ & FI_B . Ff \ar@2[r]^{\chi^F} \ar@2[d]|{\alpha_{I_B} . \alpha_f} \dthreecell{dr}{\Pi^\alpha} & F(I_B . f) \ar@2[dr]^{F\l} \ar@2[d]|{\alpha_{I_B . f}} \\ I_{FB}.Ff \ar@2[ur]^{\iota^F.1} \ar@2[d]_{1 . \alpha_f} \ar@{}[r] \ar@3?(0.5)+/u 0.2cm/;?(0.5)+/d 0.2cm/_(1.5)*+[l]{\scriptstyle \ \ \overline{M^\alpha.1}} & GI_B . Gf \ar@2[r]_{\chi^G} \dthreecell{ddr}{\gamma^G} & G(I_B . f) \ar@2[dr]_{G \l} \dthreecell{r}{\alpha_\l} & Ff \ar@2[d]^{\alpha_f} \\ I_{GB}.Gf \ar@2[dr]_{\l} \ar@2[ur]_{\iota^G.1} & & & Gf\text. \\ & Gf \ar@{=}[r] & Gf \ar@{=}[ur] }\] \item For each 1-cell $f \colon A \to B$ of $\S$, the following pasting equality holds: \[ \hphantom{=} \quad \cd[@!C@!R@R+1em]{ & Ff.FI_{B} \ar@2[r]^{\chi^F} \dthreecell{ddr}{\delta^F} & F(f.I_B) \ar@2[dr]^{F\r} \\ Ff.I_{FB} \ar@2[dr]^{\r} \ar@2[ur]^{1.\iota^F} \ar@2[d]_{\alpha_f . 1} & & & Ff \ar@2[d]^{\alpha_f} \\ Gf.I_{GB} \ar@2[dr]_{\r} \twocong{r} & Ff \ar@{=}[r] \ar@2[d]^{\alpha_f} \twoeq{dr} & Ff \ar@{=}[ur] \ar@2[d]^{\alpha_f} \twoeq{r} & Gf \\ & Gf \ar@{=}[r] & Gf \ar@{=}[ur] }\] \[ = \quad \cd[@!C@!R@R+1em]{ & Ff.FI_B \ar@2[r]^{\chi^F} \ar@2[d]|{\alpha_f . \alpha_{I_B}} \dthreecell{dr}{\Pi^\alpha} & F(f.I_B) \ar@2[dr]^{F\r} \ar@2[d]|{\alpha_{f . I_B}} \\ Ff.I_{FB} \ar@2[ur]^{1.\iota^F} \ar@2[d]_{\alpha_f . 1} \ar@{}[r] \ar@3?(0.5)+/u 0.2cm/;?(0.5)+/d 0.2cm/_(1.5)*+[l]{\scriptstyle \ \ \overline{1.M^\alpha}} & Gf.GI_B \ar@2[r]_{\chi^G} \dthreecell{ddr}{\delta^G} & G(f.I_B) \ar@2[dr]_{G \r} \dthreecell{r}{\alpha_\r} & Ff \ar@2[d]^{\alpha_f} \\ Gf.I_{GB} \ar@2[dr]_{\r} \ar@2[ur]_{1.\iota^G} & & & Gf\text. \\ & Gf \ar@{=}[r] & Gf \ar@{=}[ur] }\] \item For each triple $f, g, h$ of composable 1-cells of $\S$, the following pasting equality holds: \[ \hphantom{=} \quad \cd[@!C@!R@R+1em]{ & F(hg). Ff \ar@2[r]^{\chi^F} \dthreecell{ddr}{\omega^F} & F((hg)f) \ar@2[dr]^{F\a} \\ (Fh.Fg).Ff \ar@2[dr]^{\a} \ar@2[ur]^{\chi^F.1} \ar@2[d]_{(\alpha_h . \alpha_g) . \alpha_f} & & & F(h(gf)) \ar@2[d]^{\alpha_{h(gf)}} \\ (Gh.Gg).Gf \ar@2[dr]_{\a} \twocong{r} & Fh.(Fg. Ff) \ar@2[r]^{1. \chi^F} \ar@2[d]|{\alpha_h . (\alpha_g . \alpha_f)} \dthreecell{dr}{\overline {1.\Pi^\alpha}} & Fh. F(gf) \ar@2[ur]^{\chi^F} \ar@2[d]^{\alpha_h . \alpha_{gf}} \dthreecell{r}{\Pi^\alpha} & G(h(gf)) \\ & Gh.(Gg.Gf) \ar@2[r]_{1.\chi^G} & Gh.G(gf) \ar@2[ur]_{\chi^G} }\] \[ = \quad \cd[@!C@!R@R+1em]{ & F(hg). Ff \ar@2[r]^{\chi^F} \ar@2[d]^{\alpha_{hg} . \alpha_f} \dthreecell{dr}{\Pi^\alpha} & F((hg)f) \ar@2[dr]^{F\a} \ar@2[d]^{\alpha_{(hg)f}} \\ (Fh.Fg).Ff \ar@2[ur]^{\chi^F.1} \ar@2[d]_{(\alpha_h . \alpha_g) . \alpha_f} \ar@{}[r] \ar@3?(0.5)+/u 0.2cm/;?(0.5)+/d 0.2cm/_(1.5)*+[l]{\scriptstyle \ \ \overline{\Pi^\alpha.1}} & G(hg).Gf \ar@2[r]_{\chi^G} \dthreecell{ddr}{\omega^G} & G((hg)f) \ar@2[dr]_{G \a} \dthreecell{r}{\alpha_\a} & F(h(gf)) \ar@2[d]^{\alpha_{h(gf)}} \\ (Gh.Gg).Gf \ar@2[dr]_{\a} \ar@2[ur]_{\chi^G.1} & & & G(h(gf))\text. \\ & Gh.(Gg.Gf) \ar@2[r]_{1.\chi^G} & Gh.G(gf) \ar@2[ur]_{\chi^G} }\] \end{enumerate} \end{Defn} Now, just as with oplax transformations between homomorphisms of bicategories, these oplax icons fail to admit a well-defined composition along 0-cell boundaries, and in order to remedy this we must restrict our attention to a suitable subclass. In fact, there are two such subclasses: \begin{Defn} Let $F, G \colon \S \to {\mathcal T}$ be lax homomorphisms. Then: \begin{itemize*} \item An \defn{ico-icon} $\alpha \colon F \Rightarrow G$ is an oplax icon $\alpha$ for which each 2-cell $\alpha_f \colon Ff \Rightarrow Gf$ is an identity (it is an \emph{i}dentity \emph{c}omponents \emph{o}plax icon); \item A \defn{pseudo-icon} $\alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$ is an oplax icon $\alpha$ for which each 3-cell $\alpha_\theta$, $M^\alpha_A$ and $\Pi^\alpha_{f, g}$ is invertible. \end{itemize*} \end{Defn} \noindent The ico-icons we have already met: \begin{Prop}\label{embed} Let $F, G \colon \S \to {\mathcal T}$ be lax homomorphisms. Then there is a bijection between the set of 2-cells $\alpha \colon F \Rightarrow G$ of $\cat{Tricat}_2$ and the set of oplax icons $\alpha \colon F \Rightarrow G$ for which each component $\alpha_f \colon Ff \Rightarrow Gf$ is an identity 2-cell. \end{Prop} On the other hand, a pseudo-icon is precisely what one gets by applying a naive dimension-raising to the notion of bicategorical icon: namely, inserting an invertible 3-cell into every diagram of 2-cells which previously commuted on the nose, and then subjecting all of this to an appropriate collection of axioms. Now, these two subclasses of the oplax icons~--~the ico-icons and the pseudo-icons~--~will provide the respective vertical and horizontal 2-cells of $\mathfrak{Tricat}_3$. It remains only to introduce the 3-cells: \begin{Defn} Let there be given pseudo-icons $\alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$ and $\beta \colon F' \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G'$, and ico-icons $\gamma \colon F \Rightarrow F'$ and $\delta \colon G \Rightarrow G'$. A \defn{pseudo-icon modification} \[\cd{ F \ar@2[d]_{\gamma} \ar@2[r]|-{\object@{|}}^{\alpha} \dthreecell{dr}{\Gamma}& G \ar@2[d]^{\delta} \\ F' \ar@2[r]|-{\object@{|}}_{\beta} & G'}\] now consists in the following datum: \begin{enumerate}[(MD1)] \item For each $A, B$ in $\S$, a cubical modification (cf. Definition \ref{cubbicat}) \[\cd{ F_{A, B} \ar@2[d]_{\gamma_{A, B}} \ar@2[r]|-{\object@{|}}^{\alpha_{A, B}} \dthreecell{dr}{\Gamma_{A, B}}& G_{A, B} \ar@2[d]^{\delta_{A, B}} \\ F'_{A, B} \ar@2[r]|-{\object@{|}}_{\beta_{A, B}} & G'_{A, B}}\] (so in particular, for each 1-cell $f \colon A \to B$ of $\S$, we have a 3-cell $\Gamma_f \colon \alpha_f \Rrightarrow \beta_f$ of ${\mathcal T}$); \end{enumerate} subject to the following axioms: \begin{enumerate}[(M{A}1)] \item For each object $A$ of $\S$, the following pasting equality holds: \[\cd[@u@!@+1em]{ I_{F'A} \ar@{=}[r] \ar@{=}[d] \ar@2[dr]_{\iota^{F'}_A} & I_{FA} \ar@2[dr]^{\iota^F_A} \dthreecell[0.45]{d}{M^{\gamma}_A} \\ I_{G'A} \ar@2[dr]_{\iota^{G'}_A} \dlthreecell{r}{M^{\beta}_A} & F'I_A \ar@2[d]^{\beta_{I_A}} \ar@{=}[r] \dthreecell{dr}{\Gamma_{I_A}} & FI_A \ar@2[d]^{\alpha_{I_A}}\\ & G'I_A \ar@{=}[r] & GI_A }\quad = \quad \cd[@u@!@+1em]{ I_{F'A} \ar@{=}[r] \ar@{=}[d] & I_{FA} \ar@{=}[d] \ar@2[dr]^{\iota^F_A} \\ I_{G'A} \ar@{=}[r] \ar@2[dr]_{\iota^{G'}_A} & I_{GA} \ar@2[dr]_{\iota^G_A} \dlthreecell{r}{M^{\alpha}_A} \dthreecell[0.45]{d}{M^{\delta}_A} & FI_A \ar@2[d]^{\alpha_{I_A}}\\ & G'I_A \ar@{=}[r] & GI_A\text; } \] \item For each pair of composable 1-cells $f \colon A \to B$, $g \colon B \to C$ of $\S$, the following pasting equality holds: \[\cd[@u@!@+1.5em@R-1em]{ F'g . F'f \ar@{=}[r] \ar@2[d]_{\beta_g . \beta_f} \ar@2[dr]_{\chi^{F'}_{f, g}} & Fg . Ff \ar@2[dr]^{\chi^F_{f, g}} \dthreecell{d}{\Pi^{\gamma}_{f, g}} \\ G'g . G'f \ar@2[dr]_{\chi^{G'}_{f, g}} \dlthreecell{r}{\Pi^{\beta}_{f, g}} & F'(gf) \ar@2[d]^{\beta_{gf}} \ar@{=}[r] \dthreecell{dr}{\Gamma_{gf}} & F(gf) \ar@2[d]^{\alpha_{gf}}\\ & G'(gf) \ar@{=}[r] & G(gf) }\quad = \quad \cd[@u@!@+1.5em@R-1em]{ F'g . F'f \ar@{=}[r] \ar@2[d]_{\beta_g . \beta_f} \dthreecell{dr}{\Gamma_g . \Gamma_f} & Fg . Ff \ar@2[d]^{\alpha_g . \alpha_f} \ar@2[dr]^{\chi^F_{f, g}} \\ G'g . G'f \ar@{=}[r] \ar@2[dr]_{\chi^{G'}_{f, g}} & Gg . Gf \ar@2[dr]_{\chi^G_{f, g}} \dlthreecell{r}{\Pi^{\alpha}_{f, g}} \dthreecell[0.45]{d}{\Pi^{\delta}_{f, g}} & F(gf) \ar@2[d]^{\alpha_{gf}}\\ & G'(gf) \ar@{=}[r] & G(gf)\text. } \] \end{enumerate} \end{Defn} We are now ready to start constructing $\mathfrak{Tricat}_3$. We begin with the local structure: \begin{Prop}\label{localstruct} Let $\S$ and ${\mathcal T}$ be tricategories. Then the lax homomorphisms, ico-icons, pseudo-icons and pseudo-icon modifications from $\S$ to ${\mathcal T}$ form a weak double category $\mathfrak{Icon}(\S, {\mathcal T})$. \end{Prop} \begin{proof} Underlying each lax homomorphism, ico-icon, pseudo-icon or pseudo-icon modification is an indexed family of homomorphisms of bicategories, bicategorical icons, pseudo-natural transformations, or cubical modifications, respectively: thus our approach will be to lift the compositional structure from the double categories $\mathfrak{Hom}(\mathfrak C, \mathfrak D)$ as defined on page \pageref{hom}. We begin with the vertical structure of $\mathfrak{Icon}(\S, {\mathcal T})$. We have already seen in Section \ref{bicat-tricat} that the lax homomorphisms and ico-icons from $\S$ to ${\mathcal T}$ form a category; we must show the same is true of the oplax icons and the icon modifications between them. So given an oplax icon $\alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$, the unit icon modification $\mathrm{id}_\alpha \colon \alpha \Rrightarrow \alpha$ is given by the identity family of modifications \[(\mathrm{id}_\alpha)_{A, B} = \mathrm{id}_{\alpha_{A, B}} \colon \alpha_{A, B} \Rrightarrow \alpha_{A, B}\text;\] the axioms (MA1) and (MA2) are clear, since every occurence of $\Gamma$ reduces to an identity 3-cell. Next, given icon modifications $\Gamma \colon \alpha \Rrightarrow \beta$ and $\Delta \colon \beta \Rrightarrow \gamma$, their vertical composite $\Delta\Gamma \colon \alpha \Rrightarrow \gamma$ is given by composing their underlying families of modifications: \[(\Delta\Gamma)_{A, B} = \Delta_{A, B}. \Gamma_{A, B} \colon \alpha_{A, B} \Rrightarrow \gamma_{A, B}\text.\] Now the axioms (MA1) and (MA2) follow from an application of the corresponding axiom for $\Delta$ followed by the corresponding axiom for $\Gamma$. Associativity and unitality of this composition follow from that of vertical composition of modifications. We next describe the horizontal structure of $\mathfrak{Icon}(\S, {\mathcal T})$, beginning with the identities functor $\DI{({\mathord{\text{--}}})}$. Given a lax homomorphism $F \colon \S \to {\mathcal T}$, the identity pseudo-icon $\DI F \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} F$ has (ID1) given by the family: \[(\DI F)_{A, B} = \mathrm{id}_{F_{A, B}} \colon F_{A, B} \Rightarrow F_{A, B}\text;\] whilst $M^{\DI F}_A$ and $\Pi^{\DI F}_{A, B, C}$ are given by unnamed coherence isomorphisms in the hom-bicategories of ${\mathcal T}$, whilst for an ico-icon $\alpha \colon F \Rightarrow G$, the identity icon modification $\DI \alpha \colon \DI F \Rrightarrow \DI G$ has (MD1) given by the identity family of 3-cells \[\mathrm{id}_{\mathrm{id}_{Ff}} \colon \mathrm{id}_{Ff} \Rrightarrow \mathrm{id}_{Gf}\text.\] Each of the axioms (IA1)--(IA3) for $\DI F$ and (MA1)--(MA2) for $\DI \alpha$ now asserts that some 3-cell is equal to itself when pasted with such unnamed coherence cells, and this follows from coherence for bicategories. The functoriality of $\DI{({\mathord{\text{--}}})}$ is immediate. We now describe the horizontal composition functor for this double category. Given pseudo-icons $\alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$ and $\beta \colon G \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} H$, we define an pseudo-icon $\beta \alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} H$ as follows: \begin{enumerate}[({ID}1)] \item $(\beta\alpha)_{A, B} = \beta_{A, B}.\alpha_{A, B} \colon F_{A, B} \Rightarrow H_{A, B}$; \item $M^{\beta\alpha}_A$ is the pasting: \[ \cd[@C+1em]{ I_{FA} \ar@2[r]^{\iota^F_A} \ar@{=}[d] \dthreecell{dr}{M^{\alpha}_A} & FI_A \ar@2[d]^{\alpha_{I_A}} \\ I_{GA} \ar@2[r]|{\iota^G_A} \ar@{=}[d] \dthreecell{dr}{M^{\beta}_A} & GI_A \ar@2[d]^{\beta_{I_A}} \\ I_{HA} \ar@2[r]_{\iota^H_A} & HI_A\text; }\] \item $\Pi^\alpha_{A, B, C}$ is the pasting: \[ \cd{ F({\mathord{\text{--}}}) \otimes F(?) \ar@2[r]^{\chi^F} \ar@2[d]|{\alpha_{({\mathord{\text{--}}})}\otimes \alpha_{(?)}} \dthreecell{dr}{\Pi_{A, B, C}^\alpha} \ar@/_5em/@2[dd]_{(\beta\alpha)_{({\mathord{\text{--}}})} \otimes (\beta\alpha)_{(?)}} & F\big(({\mathord{\text{--}}}) \otimes (?)\big) \ar@2[d]^{\alpha_{(({\mathord{\text{--}}}) \otimes (?))}} \\ G({\mathord{\text{--}}})\otimes G(?) \twocong[-0.4]{r} \ar@2[r]^{\chi^G} \ar@2[d]|{\beta_{({\mathord{\text{--}}})}\otimes \beta_{(?)}} \dthreecell{dr}{\Pi_{A, B, C}^\beta} & G\big(({\mathord{\text{--}}}) \otimes (?)\big) \ar@2[d]^{\beta_{(({\mathord{\text{--}}}) \otimes (?))}} \\ H({\mathord{\text{--}}})\otimes H(?) \ar@2[r]_{\chi^H} & H\big(({\mathord{\text{--}}}) \otimes (?)\big)\text. } \] \end{enumerate} Showing that these data satisfy axioms (IA1)--(IA3) is almost as simple as placing the corresponding diagrams for $\beta$ and $\alpha$ alongside each other; the only slight complication arises from the presence of the barred 3-cells, which require us to prove some auxiliary results. For instance, in order to prove (IA1) we must first show that: \begin{multline*} \cd{ I_{FB} . Ff \ar@2[r]^{\iota^F .1} \ar@2[d]|{1 . \alpha_f} \dthreecell{dr}{\overline{M^\alpha . 1}} \ar@/_5em/@2[dd]_{1 . (\beta\alpha)_f} & FI_B . Ff \ar@2[d]^{\alpha_{I_B} . \alpha_f} \ar@/^5em/@2[dd]^{(\beta\alpha)_{I_B} . (\beta\alpha)_f}\\ I_{GB} . Gf \twocong[-0.4]{r} \ar@2[r]^{\iota^G .1} \ar@2[d]|{1 . \beta_f} \dthreecell{dr}{\overline{M^\beta . 1}} & GI_B . Gf \twocong[-0.4]{l} \ar@2[d]^{\beta_{I_B} . \beta_f} \\ I_{HB} . Hf \ar@2[r]_{\iota^H . 1} & HI_B . Hf } \\ \quad = \quad \cd[@+1em]{ I_{FB} . Ff \ar@2[r]^{\iota^F .1} \ar@2[d]_{1 . (\beta\alpha)_f} \dthreecell{dr}{\overline{M^{\beta\alpha} . 1}} & FI_B . Ff \ar@2[d]^{(\beta\alpha)_{I_B} . (\beta\alpha)_f} \\ I_{HB} . Hf \ar@2[r]_{\iota^H .1} & HI_B . Hf\text; } \end{multline*} holds; and similarly for (IA2) and (IA3). These derivations are straightforward bicategorical manipulations and left to the reader. Next we consider the horizontal composition of 2-cells in $\mathfrak{Icon}(\S, {\mathcal T})$. Next, given icon modifications \[\cd{ F \ar@2[d]_{\sigma} \ar@2[r]|-{\object@{|}}^{\alpha} \dthreecell{dr}{\Gamma}& G\ar@2[r]|-{\object@{|}}^{\beta} \dthreecell{dr}{\Delta} \ar@2[d]^{\tau} & H \ar@2[d]^{\upsilon} \\ F' \ar@2[r]|-{\object@{|}}_{\alpha'} & G' \ar@2[r]|-{\object@{|}}_{\beta'} & H'}\] we define the icon modification $\Delta \ast \Gamma \colon \beta \alpha \Rrightarrow \beta' \alpha' \colon F \Rightarrow H$ by taking \[(\Delta \ast \Gamma)_{A, B} = \Delta_{A, B} \ast \Gamma_{A, B} \colon \beta_{A, B} \alpha_{A, B} \Rrightarrow \beta'_{A, B}. \alpha'_{A, B}\text,\] where $\ast$ on the right-hand side is horizontal composition of 2-cells in the weak double category $\mathfrak{Hom}(\S(A, B), {\mathcal T}(FA, FB))$. In particular, for any 1-cell $f$ of $\S$, the 3-cell $(\Delta \ast \Gamma)_f$ is given by the pasting \[\cd[@u]{ Ff \ar@2[d]_{\alpha'_f} \ar@{=}[r] \dthreecell{dr}{\Gamma_f} & Ff \ar@2[d]^{\alpha_f}\\ Gf \ar@2[d]_{\beta'_f} \ar@{=}[r] \dthreecell{dr}{\Delta_f} & Gf \ar@2[d]^{\beta_f}\\ Hf \ar@{=}[r] & Hf\text, }\] and thus (MA1) and (MA2) for $\Delta \ast \Gamma$ follow by placing the corresponding axioms for $\Gamma$ and $\Delta$ beside each other, together with some very simple manipulation with unnamed coherence cells. Finally, note that the middle-four interchange law will hold in $\mathfrak{Icon}(\S, {\mathcal T})$ since it does in each double category $\mathfrak{Hom}(\S(A, B), {\mathcal T}(FA, FB))$. It remains only to give the unitality and associativity constraints $\a$, $\l$ and $\r$ for our double category. So let there be given pseudo-icons $\alpha \colon F \Rightarrow G$, $\beta \colon G \Rightarrow H$ and $\gamma \colon H \Rightarrow K$. Then: \begin{itemize*} \item The associativity constraint $\a_{\alpha, \beta, \gamma} \colon (\gamma\beta)\alpha \Rrightarrow \gamma(\beta\alpha)$ has component modification $(\a_{\alpha, \beta, \gamma})_{A, B}$ given by the associativity constraint $\a_{\alpha_{A, B}, \beta_{A, B}, \gamma_{A, B}}$ in the double category $\mathfrak{Hom}(\S(A, B), {\mathcal T}(FA, FB))$; \item The left unitality constraint $\l_\alpha \colon \mathrm{id}_G . \alpha \Rrightarrow \alpha$ has component modification $(\l_\alpha)_{A, B}$ given by the left unitality constraint $\l_{\alpha_{A, B}}$ in $\mathfrak{Hom}(\S(A, B), {\mathcal T}(FA, FB))$; \item The right unitality constraint $\r_\alpha \colon \alpha . \mathrm{id}_G \Rrightarrow \alpha$ has component modification $(\r_\alpha)_{A, B}$ given by the right unitality constraint $\r_{\alpha_{A, B}}$ in $\mathfrak{Hom}(\S(A, B), {\mathcal T}(FA, FB))$. \end{itemize*} The naturality of these constraints in $\alpha$, $\beta$ and $\gamma$ is inherited from the hom-double categories $\mathfrak{Hom}(\S(A, B), {\mathcal T}(FA, FB))$; and that these data satisfy the axioms (MA1) and (MA2) is also straightforward. In the case of $\a_{\alpha, \beta, \gamma}$, for example, we see that $M^{\gamma( \beta\alpha)}_A$ and $\Pi^{\gamma(\beta\alpha)}_{f, g}$ can be obtained from $M^{(\gamma \beta) \alpha}_A$ and $\Pi^{(\gamma\beta)\alpha}_{f, g}$ by pasting with unnamed coherence isomorphisms; but the components of $\a_{\alpha, \beta, \gamma}$ are built from the selfsame coherence isomorphisms, and so the result follows from the coherence theorem for bicategories. \end{proof} In order for the double categories $\mathfrak{Icon}(\S, {\mathcal T})$ to provide homs for the locally double bicategory $\mathfrak{Tricat}_3$, we must define top level composition and identity homomorphisms. The identity homomorphism \[\elt{I_{\mathcal T}} \colon 1 \to \mathfrak{Icon}({\mathcal T}, {\mathcal T})\] is straightforward; it sends the unique object of 1 to the identity lax homomorphism $\mathrm{id}_{\mathcal T} \colon {\mathcal T} \to {\mathcal T}$, the unique vertical 1-cell to the identity ico-icon on $\mathrm{id}_T$; the unique horizontal 1-cell to the identity pseudo-icon on $\mathrm{id}_{\mathcal T}$; and the unique 2-cell to the identity icon modification on this. The coherence data for this homomorphism arises from unitality constraints in $\mathfrak{Icon}({\mathcal T}, {\mathcal T})$, and so the homomorphism axioms follow from coherence for bicategories. Let us now consider the composition homomorphism \[\otimes \colon \mathfrak{Icon}({\mathcal T}, {\mathcal U}) \times \mathfrak{Icon}(\S, {\mathcal T}) \to \mathfrak{Icon}(\S, {\mathcal U})\text.\] The general approach will be similar to that adopted in the proof of Proposition \ref{localstruct}. There, we defined the compositional structure on $\mathfrak{Icon}(\S, {\mathcal T})$ by lifting it from the weak double categories $\mathfrak{Hom}(\S(A, B), {\mathcal T}(FA, FB))$: here, we will define $\otimes$ by lifting the composition homomorphisms \begin{multline*} \mathfrak{Hom}({\mathcal T}(FA, FB), {\mathcal U}(GFA, GFB)) \times \mathfrak{Hom}(\S(A, B), {\mathcal T}(FA, FB)) \\ \to \mathfrak{Hom}(\S(A, B), {\mathcal U}(GFA, GFB)) \end{multline*} which we obtain from the locally double bicategory $\mathfrak{Bicat}$ of Proposition \ref{bicategories-as-dblcatbicategories}. In detail, $\otimes$ is given as follows. On objects and vertical 1-cells, it is given by the composition law for $\cat{Tricat}_2$. On horizontal 1-cells, we consider pseudo-icons $\alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} F' \colon \S \to {\mathcal T}$ and $\beta \colon G \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G' \colon {\mathcal T} \to {\mathcal U}$, for which the composite pseudo-icon $\beta \otimes \alpha \colon GF \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G'F'$ is given as follows: \begin{enumerate}[({ID}1)] \item $(\beta \otimes \alpha)_{A, B} = \beta_{FA, FB} \otimes \alpha_{A, B}$, where $\otimes$ on the right-hand side is one of the two canonical choices for composition of horizontal composition of pseudo-natural transformations; for concreteness let us take \[(\beta \otimes \alpha)_f = \cd{GFf \ar@2[r]^{\beta_{Ff}} & G'Ff \ar@2[r]^{G'\alpha_f} & G'F'f}\] and \[(\beta \otimes \alpha)_\theta \quad = \quad \cd{ GFf \ar@2[r]^{GF \theta} \ar@2[d]_{\beta_{Ff}} \dthreecell{dr}{\beta_{F\theta}}& GFg \ar@2[d]^{\beta_{Fg}} \\ G'Ff \ar@2[r]^{G'F \theta} \ar@2[d]_{G'\alpha_f} \dthreecell{dr}{G'\alpha_{\theta}}& G'Fg \ar@2[d]^{G'\alpha_g} \\ G'F'f \ar@2[r]_{G'F' \theta} & G'F'g\text{;} }\] \item $M^{\beta \otimes \alpha}_A$ is the following 3-cell: \[\cd{ I_{GFA} \ar@2[r]^{\iota^G} \ar@{=}[d] \dthreecell{dr}{M^{\beta}_{FA}} & GI_{FA} \ar@2[d]^{\beta_{I_{FA}}} \ar@2[r]^{G\iota^F} \dthreecell{dr}{\beta_{\iota^F}} & GFI_A \ar@2[d]^{\beta_{FI_A}} \\ I_{G'FA} \ar@2[r]^{\iota^{G'}} \ar@{=}[d] \twoeq{dr} & G'I_{FA} \ar@{=}[d] \ar@2[r]^{G'\iota^F} \dthreecell{dr}{\overline{G'M^{\alpha}_A}} & G'FI_A \ar@2[d]^{G'\alpha_{I_A}} \\ I_{G'F'A} \ar@2[r]_{\iota^{G'}} & G'I_{F'A} \ar@2[r]_{G'\iota^{F'}} & G'F'I_A\text;} \] \item $\Pi^{\beta \otimes \alpha}_{A, B, C}$ is the pseudo-natural transformation with the following components: \[ \cd[@C+1em]{ GFg . GFf \ar@/_5em/@2[dd]_{(\beta \otimes \alpha)_g . (\beta \otimes \alpha)_f} \ar@2[r]^{\chi^G} \ar@2[d]|{\beta_{Fg} . \beta_{Ff}} \dthreecell{dr}{\Pi^{\beta}_{Ff, Fg}} & G(Fg . Ff) \ar@2[r]^{G\chi^F} \ar@2[d]^{\beta_{Fg . Ff}} \dthreecell{dr}{\beta_{\chi^F}} & GF(gf) \ar@2[d]^{\beta_{F(gf)}} \\ G'Fg . G'Ff \twocong[-0.4]{r} \ar@2[r]^{\chi^{G'}} \ar@2[d]|{G'\alpha_g . G'\alpha_f} \twocong{dr} & G'(Fg . Ff) \ar@2[r]^{G'\chi^F} \ar@2[d]|{G'(\alpha_g . \alpha_f)} \dthreecell{dr}{\overline{G'\Pi^\alpha_{f, g}}} & G'F(gf) \ar@2[d]^{G'\alpha_{gf}} \\ G'F'g . G'F'f \ar@2[r]_{\chi^{G'}} & G'(F'g . F'f) \ar@2[r]_{G'\chi^{F'}} & G'F'(gf)\text. }\] \end{enumerate} The proof that these data satisfy axioms (IA1)--(IA3) consists once again in building large cubes or hexagonal prisms from smaller ones, together with some simple manipulation with unnamed coherence cells: and once again, we leave this task to the reader. Finally we must give the action of $\otimes$ on 2-cells. Given two such: \[ \cd{ F \ar@2[d]_{\sigma} \ar@2[r]|-{\object@{|}}^{\alpha} \dthreecell{dr}{\Gamma}& F' \ar@2[d]^{\sigma'} \\ H \ar@2[r]|-{\object@{|}}_{\gamma} & H' } \quad \text{and} \quad \cd{ G \ar@2[d]_{\tau} \ar@2[r]|-{\object@{|}}^{\beta} \dthreecell{dr}{\Delta}& G' \ar@2[d]^{\tau'} \\ K \ar@2[r]|-{\object@{|}}_{\delta} & K' } \] in $\mathfrak{Icon}(\S, {\mathcal T})$ and $\mathfrak{Icon}({\mathcal T}, {\mathcal U})$ respectively, we define their composite $\Delta \otimes \Gamma \colon \beta \otimes \alpha \Rrightarrow \delta \otimes \gamma$ to be given by horizontally composing the underlying families of cubical modifications in the locally double bicategory $\mathfrak{Bicat}$: \[(\Delta \otimes \Gamma)_{A, B} = \Gamma_{FA, FB} \otimes \Delta_{A, B} \colon \beta_{FA, FB} \otimes \alpha_{A, B} \Rrightarrow \delta_{FA, FB} \otimes \gamma_{A, B}\text.\] So in particular, for any 1-cell $f \colon A \to B$ of $\S$, we have $(\Delta \otimes \Gamma)_f$ given by the following pasting \[\cd{ GFf \ar@2[r]^{\beta_{Ff}} \ar@{=}[d] \dthreecell{dr}{\Delta_{Ff}} & G'Ff \ar@2[r]^{G'\alpha_{f}} \ar@{=}[d] \dthreecell{dr}{\tau'_{\alpha_f}} & G'F'f \ar@{=}[d] \\ KFf \ar@2[r]^{\delta_{Ff}} \ar@{=}[d] \twoeq{dr} & K'Ff \ar@2[r]^{K'\alpha_{f}} \ar@{=}[d] \dthreecell{dr}{K'\Gamma_f} & K'F'f \ar@{=}[d] \\ KHf \ar@2[r]_{\delta_{Hf}} & K'Hf \ar@2[r]_{K'\gamma_{f}} & K'H'f\text.}\] Proving axioms (MA1) and (MA2) for this data amounts to constructing a further succession of pasting equalities which traverse the interior of a $2 \times 2 \times 2$ cube, using: \begin{itemize} \item the corresponding axioms (MA1) or (MA2) for $\Delta$ and $\Gamma$, \item the cubical modification axioms for the components of $\Delta$, \item the icon axioms for the components of $\tau'$, \item the pseudo-natural transformation axioms for the components of $\delta$, \item and some further calculus with unnamed coherence isomorphisms. \end{itemize} We leave this task to the reader. Functoriality of this composition with respect to vertical composition is inherited from that of horizontal composition of modifications in $\mathfrak{Bicat}$. It remains to exhibit the pseudo-functoriality constraints for $\otimes$; so let there be given lax homomorphisms and pseudo-icons \[\cd{ \S \ar@/^2em/[rr]^{F} \ar[rr]|{F'} \ar@/_2em/[rr]_{F''} & {}\dtwocell[0.36]{d}{\gamma} \dtwocell[-0.36]{d}{\alpha} & {\mathcal T} \ar@/^2em/[rr]^{G} \ar[rr]|{G'} \ar@/_2em/[rr]_{G''} & {}\dtwocell[0.36]{d}{\delta} \dtwocell[-0.36]{d}{\beta} & {\mathcal U}\text. \\ & & & & }\]\vskip-0.5\baselineskip\noindent We must exhibit invertible globular icon modifications \[i_{(G, F)} \colon \mathrm{id}_{GF} \Rrightarrow \mathrm{id}_G \otimes \mathrm{id}_F \colon GF \Rightarrow GF \quad \text{and} \quad m_{(\beta, \alpha), (\delta, \gamma)} \colon (\delta \otimes \gamma)(\beta \otimes \alpha) \Rrightarrow (\delta \beta) \otimes (\gamma \alpha)\text;\] whose respective $(A, B)$-components we take to be the invertible modifications witnessing pseudo-functoriality of horizontal composition in the following diagram of homomorphisms and pseudo-natural transformations: \[\cd{ \S(A, B) \ar@/^2em/[rr]^{F_{A, B}} \ar[rr]|{F'_{A, B}} \ar@/_2em/[rr]_{F''_{A, B}} & {}\dtwocell[0.36]{d}{\gamma_{A, B}} \dtwocell[-0.36]{d}{\alpha_{A, B}} & {\mathcal T}(FA, FB) \ar@/^2em/[rr]^{G_{FA, FB}} \ar[rr]|{G'_{FA, FB}} \ar@/_2em/[rr]_{G''_{FA, FB}} & {}\dtwocell[0.36]{d}{\delta_{FA, FB}} \dtwocell[-0.36]{d}{\beta_{FA, FB}} & \sh{r}{0.75em} {{\mathcal U}(GFA, GFB)\text.} \\ & & & & }\] We must check that these data satisfy axioms (MA1) and (MA2). The proof is straightforward manipulation using the pseudo-naturality axioms for $\delta$ and the modification axioms for $\Pi^\delta$. Finally, the naturality of the maps $m_{(\beta, \alpha), (\delta, \gamma)}$ in all variables follows componentwise; as do the coherence axioms which $m$ and $i$ must satisfy. \begin{Thm} The double categories $\mathfrak{Icon}(\S, {\mathcal T})$ together with the identity and composition homomorphisms defined above provide data for a locally double bicategory $\mathfrak{Tricat}_3$ with 0-cells being tricategories; 1-cells, lax homomorphisms; vertical 2-cells, ico-icons; horizontal 2-cells, pseudo-icons; and 3-cells, icon modifications. \end{Thm} \begin{proof} All that remains is to give the associativity and unitality constraints $a$, $l$ and $r$ for $\mathfrak{Tricat}_3$, and check the triangle and pentagon axioms. At the level of 1-cells and vertical 2-cells, these are the corresponding constraints from $\cat{Tricat}_2$; whilst at the level of horizontal 2-cells and 3-cells, suppose we are given trihomomorphisms and pseudo-icons \[\cd{ {{\mathcal R}} \ar@/^1em/[r]^{F} \ar@/_1em/[r]_{F'} \dtwocell{r}{\alpha} & {\S} \ar@/^1em/[r]^{G} \ar@/_1em/[r]_{G'} \dtwocell{r}{\beta} & {{\mathcal T}} \ar@/^1em/[r]^{H} \ar@/_1em/[r]_{H'} \dtwocell{r}{\gamma} & {{\mathcal U}\text.} }\] Then we must give an icon modification \[ \cd[@+0.5em]{ (HG)F \ar@2[d]_{a_{F, G, H}} \ar@2[r]|-{\object@{|}}^{(\gamma \otimes \beta) \otimes \alpha} \dthreecell{dr}{a_{\alpha, \beta, \gamma}}& (H'G')F' \ar@2[d]^{a_{F', G', H'}} \\ H(GF) \ar@2[r]|-{\object@{|}}_{\gamma \otimes (\beta \otimes \alpha)} & H'(G'F') }\] where $a_{F, G, H}$ and $a_{F', G', H'}$ are the corresponding constraints from $\cat{Tricat}_2$. So we set the $(A, B)$th component of this icon modification to be the cubical modification which provides the associativity constraint for the composition \[\cd[@C+2.3em]{ {{\mathcal R}(A, B)} \ar@/^1em/[r]^{F_{A, B}} \ar@/_1em/[r]_{F'_{A, B}} \dtwocell[0.45]{r}{\alpha_{A, B}} & {\S(FA, FB)} \ar@/^1em/[r]^{G_{FA, FB}} \ar@/_1em/[r]_{G'_{FA, FB}} \dtwocell[0.37]{r}{\beta_{FA, FB}} & {{\mathcal T}(GFA, GFB)} \ar@/^1em/[r]^{H_{GFA, GFB}} \ar@/_1em/[r]_{H'_{GFA, GFB}} \dtwocell[0.32]{r}{\gamma_{GFA, GFB}} & {{\mathcal U}(HGFA, HGFB)} }\] in the locally double bicategory $\mathfrak{Bicat}$. We must check that these satisfy the data satisfy the axioms for an icon modification; let us do only (MA2), since (MA1) follows identically. We first observe that the 3-cells $\Pi^{(\gamma \otimes \beta) \otimes \alpha}_A$ and $\Pi^{\gamma \otimes (\beta \otimes \alpha)}_A$ are obtained by pasting together what is essentially the same $3 \times 3$ diagram of 3-cells, and some trivial calculus with unnamed coherence cells shows that they are precisely the same diagram, modulo rewriting of the boundary, so that the latter 3-cell may be obtained from the former by pasting with unnamed coherence cells. But this is precisely the content of axiom (MA2). Finally, we must check that these icon modifications $a_{\alpha, \beta, \gamma}$ are natural in in $\alpha$, $\beta$ and $\gamma$, and satisfy the pentagon and triangle equalities. Each of these follows componentwise from the corresponding facts in $\mathfrak{Bicat}$. \end{proof} \section{From locally double bicategories to tricategories}\label{Sec:fromltt} As we have seen, the structure which captures most precisely the compositional calculus of tricategorical icons is a locally double bicategory. Nonetheless, there may be occasions when we are prepared to sacrifice some of this precision in order to work with a more familiar structure like a tricategory: and so we would like to extract a description of a tricategory $\cat{Tricat}_3$ from our description of $\mathfrak{Tricat}_3$. Happily, there are general principles at work which allow us to do just that. In this Section, we will show how every sufficiently-well behaved locally double bicategory $\mathfrak C$ gives rise to a tricategory. This tricategory will have the same 0-\ and 1-cells as $\mathfrak C$; as 2-cells, the horizontal 2-cells of $\mathfrak C$; and as 3-cells, the globular 3-cells of $\mathfrak C$. The main point of interest is the construction of the tricategorical associativity constraints, which are to be given by \emph{horizontal} 2-cells of $\mathfrak C$. Since the associativity constraints in $\mathfrak C$ are given by \emph{vertical} 2-cells, we will need some kind of linkage between the two types of 2-cell in order to proceed. \begin{Defn} A weak double category $\mathfrak C$ is said to be \defn{fibrant} if the functor $(s, t) \colon {\mathcal C}_1 \to {\mathcal C}_0 \times {\mathcal C}_0$ is an isofibration. \end{Defn} \noindent Recall here that a functor $F \colon {\mathcal A} \to {\mathcal B}$ between categories is an \emph{isofibration} if whenever we have a object $a \in {\mathcal A}$ and isomorphism $\phi \colon Fa \to b$ in ${\mathcal B}$, there exists an object $c \in {\mathcal A}$ and isomorphism $\theta \colon a \to c$ such that $Fc = b$ and $F\theta = \phi$. Thus a weak double category $\mathfrak C$ is fibrant just when every diagram like (a) below has a filler like (b) for which the 2-cell $\theta$ is invertible as an arrow of ${\mathcal C}_1$: \[\text{(a)} \quad \cd{ x \ar[d]_{f} & x' \ar[d]^{g} \\ y \ar[r]|-{\object@{|}}_{k} & y'} \qquad \rightsquigarrow \qquad \text{(b)} \quad \cd{ x \ar[d]_{f} \ar[r]|-{\object@{|}}^{h} \dtwocell{dr}{\theta}& x' \ar[d]^{g} \\ y \ar[r]|-{\object@{|}}_{k} & y'\text.}\] Thus fibrancy is precisely the property which \cite{GP1} refers to as \emph{horizontal invariance}. We may reformulate this property in various useful ways, and since detailed accounts of this process may be found in \cite{tomf} or \cite{GP2}, we restrict ourselves here to recording those equivalent formulations which will be useful to us. For the first, we consider the case of the above filling condition where $g$ and $k$ are both identities: given a vertical map $f \colon x \to y$ of $\mathfrak C$, it asserts the existence of a horizontal 1-cell $\overline f$ and a 2-cell $\epsilon_f$ fitting into the diagram: \[\cd{ x \ar[d]_{f} \ar[r]|-{\object@{|}}^{\overline f} \dtwocell{dr}{\epsilon_f}& y \ar[d]^{\mathrm{id}_y} \\ y \ar[r]|-{\object@{|}}_{\DI y} & y\text.}\] From this, we may define a further 2-cell $\eta_f$ as the composite \[\cd{ x \ar[d]_{\mathrm{id}_x} \ar[r]|-{\object@{|}}^{\DI x} \dtwocell{dr}{\eta_f} & x \ar[d]^{f} \\ x \ar[r]|-{\object@{|}}_{\overline f} & y} \quad := \quad \cd{ x \ar[d]_{f} \ar[r]|-{\object@{|}}^{\DI x} \dtwocell{dr}{\DI f} & x \ar[d]^{f} \\ y \ar[d]_{f^{-1}} \ar[r]|-{\object@{|}}^{g} \dtwocell{dr}{\epsilon_f^{-1}} & y \ar[d]^{\mathrm{id}_y} \\ x \ar[r]|-{\object@{|}}_{\overline f} & y\text.}\] Now the pair $(\eta_f, \epsilon_f)$ satisfy the triangle identities: \[\epsilon_f . \eta_f = \DI f \colon \DI x \Rightarrow \DI y \quad \text{and} \quad \epsilon_f \ast \eta_f = (l^{-1}r)_{\overline f} \colon \overline f . \DI x \Rightarrow \DI y . \overline f\text,\] and so, in the terminology of \cite{}, $f$ and $\overline f$ are \emph{orthogonal companions}; which gives us the ``only if'' direction of: \begin{Prop} A weak double category $\mathfrak C$ is fibrant iff every vertical isomorphism has an orthogonal companion. \end{Prop} \noindent For the ``if'' direction, suppose that we are given a diagram like (a); then we can complete it to a diagram like (b) by taking $h$ to be $\overline{g^{-1}} . (k . \overline f)$ and $\theta$ to be the 2-cell: \[\overline{g^{-1}} . (k . \overline f) \xRightarrow{(\DI g .\epsilon_{g^{-1}}) \ast (\mathrm{id}_k \ast \epsilon_f)} \DI{y'} . (k . \DI y) \xRightarrow{\DI{y'} \ast r_k} \DI{y'} . k \xRightarrow{l_k} k\text.\] Thus each of $\mathfrak{Cat}$, $\mathfrak{Rng}$ and $\mathfrak{Span}({\mathcal C})$ is a fibrant double category: for $\mathfrak{Cat}$, the horizontal companion of a functor $F \colon {\mathcal C} \to D$ is the profunctor $\overline F(\mathord {\mathord{\text{--}}}, \mathord ?) = {\mathcal D}(\mathord {\mathord{\text{--}}}, F\mathord ?)$; for $\mathfrak{Rng}$, the companion of a homomorphism $f \colon R \to S$ is $S$ itself, viewed as a left $S$-, right $R$-module; and in $\mathfrak{Span}({\mathcal C})$, the companion of a morphism $f \colon C \to D$ is the span $C \stackrel{id}{\leftarrow} C \stackrel{f}{\rightarrow} D$. Observe that in all of these examples, it is arbitrary vertical morphisms, and not just the isomorphisms, which have companions: such weak double categories are essentially the \emph{pro-arrow equipments} of \cite{wood1, wood2}. A more detailed analysis of this correspondence may be found in Appendix C of \cite{mike}. \begin{Prop} Let $\mathfrak C$ be a fibrant double category equipped with a choice of orthogonal companion for every vertical isomorphism. Then the assignation $f \mapsto \overline f$ underlies an identity-on-objects homomorphism of bicategories \[(\overline{\ \ \mathstrut})\colon V^{\text{iso}}(\mathfrak C) \to H(\mathfrak C)\text,\] where $V^{\text{iso}}(\mathfrak C)$ is the category of objects and vertical isomorphisms in $\mathfrak C$. Moreover, if we are given vertical isomorphisms $f \colon w \to y$ and $g \colon x \to z$ in $\mathfrak C$, then pasting with $\eta_f$ and $\epsilon_g$ induces a bijection between the set of 2-cells of the form (c) and the set of 2-cells of the form (d): \[(c) \quad \cd{ x \ar[d]_{f} \ar[r]|-{\object@{|}}^{h} \dtwocell{dr}{\alpha} & x' \ar[d]^{g} \\ y \ar[r]|-{\object@{|}}_{k} & y'} \qquad \text{and} \qquad \text{(d)} \quad \cd{ x \ar[d]_{\mathrm{id}_x} \ar[r]|-{\object@{|}}^{\overline g . h} \dtwocell{dr}{\overline \alpha} & y' \ar[d]^{\mathrm{id}_{y'}} \\ x \ar[r]|-{\object@{|}}_{k . \overline f} & y'\text;}\] and $\overline \alpha$ is invertible as an arrow of ${\mathcal C}_1$ just when $\alpha$ is. Furthermore, these bijections satisfy four evident axioms expressing their functoriality with respect to vertical and horizontal composition of 2-cells.\end{Prop} The proof is straightforward manipulation, and it is not hard to prove a converse~--~namely, that from a homomorphism of bicategories $(\overline{\ \ \mathstrut})\colon V^{\text{iso}}(\mathfrak C) \to H(\mathfrak C)$ and a bijective assignation $\alpha \mapsto \overline \alpha$ on 2-cells satisfying the four functoriality axioms, one may define a choice of orthogonal companion for every vertical isomorphism. A detailed proof may be found in the pages leading up to Theorem 3.28 of \cite{tomf}. \begin{Defn} The 2-category $\cat{DblCat}_f$ has objects being fibrant weak double categories equipped with a choice of orthogonal companions; as 1-cells, the homomorphisms between the underlying double categories; and as 2-cells, the vertical transformations between them. \end{Defn} \noindent One may reasonably ask why we do not require the 1-cells $F \colon \mathfrak C \to \mathfrak D$ of $\cat{DblCat}_f$ to respect the choices of orthogonal companions in $\mathfrak C$ and $\mathfrak D$. The reason is that, in fact, \emph{any} homomorphism between objects of $\cat{DblCat}_f$ will automatically respect these choices in a unique way. To make this explicit, let us say that a homomorphism $F \colon \mathfrak C \to \mathfrak D$ between objects of $\cat{DblCat}_f$ is a \defn{fibrant homomorphism} if, for every invertible vertical 1-cell $f \colon x \to y$ of $\mathfrak C$, there is given an invertible globular 2-cell \[\mu_f \colon F(\overline f) \Rightarrow \overline{Ff} \colon Fx \ensuremath{\relbar\joinrel\mapstochar\joinrel\rightarrow} Fy\] of $\mathfrak D$, subject to three axioms. The first two equate, respectively, the two possible 2-cells in $\mathfrak D$ from $\DI {Fx}$ to $\overline{F(\mathrm{id}_x)}$; and from $F(\overline g) \, . \, F (\overline f)$ to $\overline{F(gf)}$. The third axiom concerns a 2-cell $\alpha$ of the type (c) above, and equates the two globular 2-cells \[ \cd[@-1em]{ Fx \ar@/_4em/[dd]|-{\object@{|}}_{\overline {Ff}} \ar[dd]|-{\object@{|}}_{F(\overline f)} \ar[rr]|-{\object@{|}}^{Fh} \dtwocell{ddrr}{F(\overline \alpha)} && x' \ar[dd]|-{\object@{|}}^{F(\overline g)} \\ {}\ltwocell[-0.9]{r}{\mu_f} & \\ Fy \ar[rr]|-{\object@{|}}_{Fk} && Fy'} \quad \text{and} \quad \cd[@-1em]{ Fx \ar[dd]|-{\object@{|}}_{\overline {Ff}} \ar[rr]|-{\object@{|}}^{Fh} \dtwocell{ddrr}{\overline {F\alpha}} && x' \ar[dd]|-{\object@{|}}^{\overline {Fg}} \ar@/^4em/[dd]|-{\object@{|}}^{F(\overline g)} \\ & & {}\ltwocell[-0.9]{l}{\mu_f} \\ Fy \ar[rr]|-{\object@{|}}_{Fk} && Fy'\text.} \] Now, given any homomorphism $F \colon \mathfrak C \to \mathfrak D$ between objects of $\cat{DblCat}_f$, we may make it into a fibrant homomorphism as follows. Given an invertible vertical arrow $f \colon x \to y$ of $\mathfrak C$, we can consider the globular 2-cell \[\overline{F\epsilon_f} \colon \overline{\mathrm{id}_{Fy}} . F \overline f \Rightarrow F\DI y . \overline{Ff}\] of $\mathfrak D$; and since both $\overline{\mathrm{id}_{Fy}}$ and $F\DI y$ are isomorphic to $\DI {Fy}$, we obtain from this a globular 2-cell $\mu_f \colon F \overline f \Rightarrow \overline{Ff}$, which is easily checked to satisfy the three axioms. And in fact, this is the only possible structure of fibrant homomorphism on $F$: for given an arbitrary such structure, applying the third axiom to the 2-cells $\epsilon_f$ in $\mathfrak C$ shows that the maps $\mu_f$ must, in fact, coincide with those defined above. A similar argument applies to the 2-cells of $\cat{DblCat}_f$. A conceptual explanation of why this should be the case is that $\cat{DblCat}_f$ is, in some sense, the 2-category of algebras for a particularly simple kind of 2-dimensional monad on $\cat{DblCat}$, the kind which \cite{propertylike} calls \emph{pseudo-idempotent}: and such monads have the property that the forgetful functor from the 2-category of algebras and algebra pseudomorphisms to the underlying base 2-category is 2-fully faithful. The qualifier ``in some sense'' covers a slight wrinkle in this explanation: namely, that the 2-monad which gives rise to $\cat{DblCat}_f$ lives not on $\cat{DblCat}$ but on $\cat{DblCat}_\text{str}$, the 2-category of weak double categories and \emph{strict} homomorphisms between them, so that making this argument rigourous would require a little more work. \begin{Defn} We will say that a locally double bicategory is \defn{locally fibrant} just when each of its hom-double categories is fibrant. \end{Defn} In particular, a monoidal double category is locally fibrant just when its underlying weak double category is fibrant, so that all of our examples of monoidal double categories are locally fibrant. The locally double bicategory $\mathfrak{DblCat}$ is easily seen \emph{not} to be locally fibrant; on the other hand, we may show that, for weak double categories $\mathfrak C$ and $\mathfrak D$, if $\mathfrak D$ is fibrant then so is $\mathfrak{Hom}(\mathfrak C, \mathfrak D)$. It follows that the locally double bicategory $\mathfrak{DblCat}_f$, of fibrant double categories and all cells between them, is itself locally fibrant; and since any bicategory is trivially fibrant, that the locally double bicategory $\mathfrak{Bicat}$ is too. We will now show that every locally fibrant locally double bicategory gives rise to a tricategory. We begin with a technical result: \begin{Prop}\label{trihomom} Let us write $\cat{DblCat}_g$ for the maximal sub-2-category of $\cat{DblCat}_f$ with only invertible 2-cells. Then the functor of mere categories ${\mathbb H} \colon \cat{DblCat} \to \cat{Bicat}$ can be extended to a trihomomorphism \[{\mathbb H} \colon \cat{DblCat}_g \to \cat{Bicat}\text.\] \end{Prop} \begin{proof} First we define ${\mathbb H}$ on cells. This is already done for 0-\ and 1-cells, and since $\cat{DblCat}_g$ has no non-trivial 3-cells, it remains only to define it on 2-cells. So let there be given an invertible vertical transformation $\alpha \colon F \Rightarrow G \colon \mathfrak C \to \mathfrak D$. We define a pseudo-natural transformation ${\mathbb H}\alpha \colon {\mathbb H} F \Rightarrow {\mathbb H} G$ by taking \[({\mathbb H} \alpha)_x = \overline{\alpha_x} \colon Fx \to Gx\text, \qquad \text{and} \qquad ({\mathbb H} \alpha)_f = \cd{ Fx \ar[d]_{\overline{\alpha_x}} \ar[r]^{Ff} \dtwocell{dr}{\overline{\alpha_f}}& Fy \ar[d]^{\overline{\alpha_y}} \\ Gx \ar[r]_{Gf} & Gy\text.}\] The transformation axioms for ${\mathbb H}\alpha$ follow straightforwardly from the vertical transformation axioms for $\alpha$ and the functoriality of $(\overline{\ \ \mathstrut})$ with respect to 2-cell composition. Next we ensure that ${\mathbb H}$ is locally a homomorphism of bicategories, which entails giving modifications $i_F \colon \mathrm{id}_{{\mathbb H} F} \Rrightarrow {\mathbb H}(\mathrm{id}_F)$ and $m_{\alpha, \beta} \colon {\mathbb H} \beta . {\mathbb H} \alpha\Rrightarrow {\mathbb H}(\beta . \alpha)$. These will have 2-cell components \[(i_F)_x \colon \mathrm{id}_{Fx} \Rightarrow \overline{\mathrm{id}_{Fx}} \qquad \text{and} \qquad (m_{\alpha, \beta})_x \colon \overline{\beta_x} \, . \, \overline{\alpha_x} \Rightarrow \overline{\beta_x \alpha_x}\] in ${\mathbb H} \mathfrak D$ given by the pseudo-functoriality constraints for $(\overline{\ \ \mathstrut})$. The coherence axioms for these data therefore follow pointwise. Next, we must give adjoint pseudo-natural equivalences \[{\b \chi}_{\mathfrak C, \mathfrak D, \mathfrak E} \colon {\mathbb H}({\mathord{\text{--}}}) \otimes {\mathbb H}(?) \Rightarrow {\mathbb H}({\mathord{\text{--}}} \otimes ?) \colon \cat{DblCat}_g(\mathfrak C, \mathfrak D) \times \cat{DblCat}_g(\mathfrak B, \mathfrak C) \to \cat{Bicat}({\mathbb H} \mathfrak B, {\mathbb H} \mathfrak D)\text.\] Observe that the homomorphisms ${\mathbb H}({\mathord{\text{--}}}) \otimes {\mathbb H}(?)$ and ${\mathbb H}(\mathord {\mathord{\text{--}}} \otimes \mathord ?)$ agree on objects, and thus we may consider icons between them: in particular, any \emph{invertible} icon between them will give rise to an adjoint pseudo-natural equivalence, and so to give $\chi$ it suffices to give invertible icons $\chi \colon {\mathbb H}({\mathord{\text{--}}}) \otimes {\mathbb H}(?) \Rightarrow {\mathbb H}(\mathord {\mathord{\text{--}}} \otimes \mathord ?)$. So consider a pair of horizontally composable 2-cells \[\cd{ {\mathfrak B} \ar@/^1em/[r]^{F} \ar@/_1em/[r]_{F'} \dtwocell{r}{\alpha} & {\mathfrak C} \ar@/^1em/[r]^{G} \ar@/_1em/[r]_{G'} \dtwocell{r}{\beta} & {\mathfrak D} }\]in $\cat{DblCat}_g$: we must give a modification $\chi_{\alpha, \beta} \colon {\mathbb H} \beta \ast {\mathbb H} \alpha \Rrightarrow {\mathbb H}(\beta \ast \alpha)$. Now, these two pseudo-natural transformations have respective $x$-components given by \begin{align*} ({\mathbb H} \beta \ast {\mathbb H} \alpha)_x & = GFx \xrightarrow{\overline{\beta_{Fx}}} G'Fx \xrightarrow{G'\overline{\alpha_x}} G'F'x\\ \text{and\ \ } {\mathbb H}(\beta \ast \alpha)_x & = GFx \xrightarrow{\overline{G'{\alpha_x} . \beta_{Fx}}} G'F'x\text,\end{align*} and so we take $(\chi_{\alpha, \beta})_x$ to be the 2-cell \[G'\overline{\alpha_x} \, . \, \overline{\beta_{Fx}} \xRightarrow{\mu_{\alpha_x} . 1} \overline{G'\alpha_x} \, . \, \overline{\beta_{Fx}} \xRightarrow{\ \cong\ } \overline{G'\alpha_x . \beta_{Fx}}\] of ${\mathbb H} \mathfrak D$. The modification axioms for $\chi_{\alpha, \beta}$ follow from the third fibrant homomorphism axiom and the functoriality axioms for $(\overline{\ \ \mathstrut})$ with respect to 2-cell composition. We must verify that these components $\chi_{\alpha, \beta}$ satisfy the three axioms making $\chi$ into an icon. The first is vacuous, whilst the second and third follow by a diagram chase using the axioms for a fibrant homomorphism. We argue entirely analogously in order to give the adjoint equivalences $\iota \colon I_{{\mathbb H} x} \Rightarrow {\mathbb H}(I_x)$. Next we must give invertible modifications $\omega$, $\delta$ and $\gamma$. In the case of $\omega$, for instance, this involves giving invertible modifications \[\cd[@!C@!R@R+1em]{ & ({\mathbb H}({\mathord{\text{--}}})\otimes{\mathbb H}(?))\otimes{\mathbb H}(\mathord \ast) \ar@2[dl]_{\chi \otimes 1} \ar@2[dr]^{\a} \\ {\mathbb H}(\mathord {\mathord{\text{--}}} \otimes \mathord ?)\otimes{\mathbb H}(\mathord\ast) \ar@2[d]_{\chi} \rthreecell{drr}{\omega} & & {\mathbb H}({\mathord{\text{--}}}) \otimes ({\mathbb H}(?) \otimes {\mathbb H}(\mathord \ast)) \ar@2[d]^{1 \otimes \chi} \\ {\mathbb H}((\mathord {\mathord{\text{--}}} \otimes \mathord ?) \otimes \mathord \ast) \ar@2[dr]_{{\mathbb H}\a} & & {\mathbb H}({\mathord{\text{--}}}) \otimes {\mathbb H}(\mathord ? \otimes \mathord \ast) \ar@2[dl]^{\chi} \\ & {\mathbb H}({\mathord{\text{--}}} \otimes (\mathord ? \otimes \mathord \ast))\text. }\] To do this, observe first that every pseudo-natural transformation bounding this diagram may also be viewed as an icon. We already know this for $\chi$ and hence also for $1 \otimes \chi$ and $\chi \otimes 1$; and it is so for $\a$ and ${\mathbb H} \a$ since composition of 1-cells in both $\cat{DblCat}_g$ and $\cat{Bicat}$ is \emph{strictly} associative. If we now compose all the 2-arrows in this diagram \emph{qua} icons, we obtain two further icons $\sigma, \tau \colon ({\mathbb H}({\mathord{\text{--}}})\otimes{\mathbb H}(?))\otimes{\mathbb H}(\mathord \ast) \Rightarrow {\mathbb H}({\mathord{\text{--}}} \otimes (\mathord ? \otimes \mathord \ast))$: and a long but straightforward diagram chase with the fibrant homomorphism axioms shows that these two icons are, in fact, equal. On the other hand, if we compose the two sides \emph{qua} pseudo-natural transformations, then the pseudo-naturals that we get will not necessarily be icons, but they will, at least, be \emph{isomorphic} to icons, namely the icons $\sigma$ and $\tau$ respectively. Thus we take $\omega$ to be the composite of the invertible modification from the left-hand side of this diagram to $\sigma = \tau$ and the invertible modification from $\tau$ to the right-hand side. We proceed similarly for the invertible modifications $\delta$ and $\gamma$. The final thing to check are the two trihomomorphism axioms, equating certain pastings of 3-cells in $\cat{Bicat}$. But all the 3-cells in question are either coherence 3-cells of $\cat{Bicat}$; or component 3-cells of $\omega$, $\delta$ and $\gamma$. But these latter 3-cells are in turn built from coherence 3-cells of $\cat{Bicat}$ and coherence 3-cells for the local homomorphisms $(\overline{\ \ \mathstrut})$. The result thus follows by coherence for tricategories and bicategorical coherence for functors. \end{proof} \begin{Thm}\label{locallydoubletotri} Let $\mathfrak C$ be a locally fibrant locally double bicategory with chosen companions in each hom. Then there is a tricategory ${\mathcal T}$ with the same objects as $\mathfrak C$, and \[{\mathcal T}(A, B) = {\mathbb H}\big(\mathfrak C(A, B)\big)\text.\] \end{Thm} \begin{proof} We begin by observing that both $\cat{DblCat}_g$ and $\cat{Bicat}$ come equipped with finite product structure; and that the trihomomorphism ${\mathbb H}$ preserves the cartesian product of $j$-cells for $j = 0, 1, 2, 3$. Now, the top-level composition and identity functors for ${\mathcal T}$ are given by applying ${\mathbb H}$ to the corresponding data (LDD3) and (LDD4) for $\mathfrak C$: \[1 = {\mathbb H} 1 \xrightarrow{{\mathbb H} \elt{I_A}} {\mathbb H}\big(\mathfrak C(A, A)) = {\mathcal T}(A, A)\] and \[{\mathcal T}(B, C) \times {\mathcal T}(A, B) = {\mathbb H}\big(\mathfrak C(B, C) \times \mathfrak C(A, B)\big) \xrightarrow{{\mathbb H} \mathord \otimes} {\mathbb H}\big(\mathfrak C(A, C)\big) = {\mathcal T}(A, C)\text.\] To obtain the pseudo-natural adjoint equivalences $\a$, $\l$ and $\r$ witnessing the associativity and unitality of this composition, we apply ${\mathbb H}$ to the corresponding data (LDD5) and (LDD6) for $\mathfrak C$. Since each of $a$, $l$ and $r$ is an adjoint equivalence (in fact, an isomorphism) in the relevant hom of $\cat{DblCat}_g$, the same will obtain for their images in $\cat{Bicat}$; and because ${\mathbb H}$ strictly preserves both cartesian products and composition of 1-cells, these adjoint equivalences will have the correct sources and targets. Next we must give the invertible modifications $\pi$, $\mu$, $\lambda$ and $\rho$. To obtain $\pi$, for example, we begin by applying ${\mathbb H}$ to the axiom (LDA2) for $\mathfrak C$. This yields an equality of 2-cells in $\cat{Bicat}$; however, these 2-cells are not of the right form to be the source and target of $\pi$. In order to make them so, we may adjust by coherence 3-cells in $\cat{Bicat}$ whose existence is guaranteed by the coherence theorem for trihomomorphisms. Consequently, we may take $\pi$ to be given by the composite of these coherence 3-cells; and similarly for $\mu$, $\lambda$ and $\rho$. Finally, we must check the three tricategory axioms. These are normally stated in a ``local'' form, asserting the equality of certain pastings of 3-cells in the relevant hom-bicategories; but in this situation, it will be more appropriate to consider them in their ``global'' form. Each such axiom amounts to giving a diagram of 2-\ and 3-cells in $\cat{Bicat}$, whose vertices are pasting diagrams built from copies of the 2-cells $a$, $l$ and $r$, and whose arrows are 3-cells between those 2-cells, built from copies of $\pi$, $\mu$, $\lambda$ and $\rho$; and asserting that the two ways around this diagram coincide. To show this, we consider the corresponding diagram for $\mathfrak C$. This is a diagram of 2-\ and 3-cells in $\cat{DblCat_g}$, which since $\cat{DblCat}_g$ has only identity 3-cells, must commute. Hence by applying ${\mathbb H}$ we obtain a commutative diagram in $\cat{Bicat}$, which, unfortunately, has both the wrong vertices and the wrong arrows. Nonetheless, by the coherence theorem for functors, each ``wrong'' vertex admits an isomorphism 3-cell to the ``right'' vertex; and in such a way that composing these isomorphism 3-cells with the ``wrong'' arrows yields the ``right'' arrows. \end{proof} Special cases of this theorem give us new proofs of some existing results. Restricting to the one-object case, we have the result that \emph{any fibrant monoidal double category gives rise to a monoidal bicategory}; this statement and a sketch proof appear as Theorem B.4 of \cite{mike}. In particular, we obtain elegant proofs that the bicategories of rings and bimodules, of categories and profunctors, and of spans internal to a cartesian category ${\mathcal C}$ are all monoidal bicategories.\footnote{For the last two of these examples, the machinery of \cite{cb2} provides another elegant proof of this fact, and in fact goes further, showing that the monoidal bicategories in question are \emph{symmetric} monoidal bicategories.} Finally, applying this theorem to the fibrant locally double bicategory $\mathfrak{Bicat}$, we deduce the existence of a tricategory of bicategories $\cat{Bicat}$. Again, the result is not new, but the proof is, showing how the tricategory structure on $\cat{Bicat}$ may be induced from a piece of canonical, universally determined structure, namely the biclosed structure on $\cat{DblCat}$. \section{A tricategory of tricategories} We now wish to apply Theorem \ref{locallydoubletotri} to the locally double bicategory $\mathfrak{Tricat}_3$, and in order to do this, we must first prove its homs to be fibrant. \begin{Prop} Each weak double category $\mathfrak{Icon}(\S, {\mathcal T})$ is fibrant. \end{Prop} \begin{proof} Suppose we are given an invertible ico-icon $\alpha \colon F \Rightarrow G \colon \S \to {\mathcal T}$. We must provide a pseudo-icon $\overline \alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$ and an invertible icon modification \[\cd{ F \ar@2[d]_{\alpha} \ar@2[r]|-{\object@{|}}^{\overline \alpha} \dthreecell{dr}{\epsilon_\alpha}& G \ar@2[d]^{\mathrm{id}_G} \\ G \ar@2[r]|-{\object@{|}}_{\DI G} & G. }\] Now by Proposition \ref{embed}, there is a bijection between the ico-icons $F \Rightarrow G$ and the oplax icons $F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$ with identity 2-cell components: for which the invertible ico-icons on the one side correspond to the pseudo-icons on the other. Thus we take $\overline \alpha$ to be the pseudo-icon corresponding to $\alpha$ under this bijection. To give the icon modification $\epsilon_\alpha$, we must give 3-cells $(\epsilon_\alpha)_f \colon {\overline \alpha}_f \Rrightarrow (\DI G)_f$ of ${\mathcal T}$, forming the components of an $\ob \S \times \ob \S$-indexed family of cubical modifications, and satisfying axioms (MA1) and (MA2). Since we have ${\overline \alpha}_f = (\DI G)_f = \mathrm{id}_{Ff} \colon Ff \Rightarrow Ff$, we take $(\epsilon_\alpha)_f = \mathrm{id}_{\mathrm{id}_f}$. The cubical modification axioms and axioms (MA1) and (MA2) now follow by coherence for bicategories. \end{proof} Thus $\mathfrak{Tricat}_3$ is locally fibrant, and so we may apply Theorem \ref{locallydoubletotri} to it, and deduce: \begin{Thm} There is a tricategory $\cat{Tricat}_3$ with objects being tricategories; 1-cells, lax homomorphisms; 2-cells, pseudo-icons; and 3-cells, globular icon modifications. \end{Thm} \endinput \section{Locally double bicategories}\label{Sec:locallydouble} We would now like to take the bicategory of tricategories $\cat{Tricat}_2$ and extend it to a tricategory of tricategories $\cat{Tricat}_3$. This will have the same 0-cells and 1-cells as $\cat{Tricat}_2$, but will have 2-cells with one less level of degeneracy, which consequently admit a notion of 3-cell between them. However, there is a problem with this. From the perspective of $\cat{Tricat}_2$, the composition of lax homomorphisms is associative up-to-isomorphism; but when we move to $\cat{Tricat}_3$, this same composition becomes associative only up-to-\emph{equivalence}, even though the constraint 2-cells are the same in both cases. The reason is that 2-cell composition in $\cat{Tricat}_3$ is less strict than it is $\cat{Tricat}_2$, and so the isomorphisms which witness associativity in $\cat{Tricat}_2$ become mere equivalences in $\cat{Tricat}_3$. To resolve this problem, we will describe a richer structure than a tricategory into which tricategories organize themselves. This structure also has 0-, 1-, 2-~and 3-cells, but the 2-cells now come in two varieties, allowing us to capture both the doubly-degenerate 2-cells of $\cat{Tricat}_2$ with their strictly associative composition and the singly-degenerate ones of $\cat{Tricat}_3$ with their up-to-isomorphism composition. Moreover, the associativity and unit constraints of this structure are given by 2-cells of the first, rather than the second, type, and thus are of a bicategorical, up-to-isomorphism, rather than a tricategorical, up-to-equivalence, kind. What we have, then, is not a tricategory of tricategories, but a \emph{locally double bicategory} of them. A locally double bicategory may be described very succinctly as a ``bicategory weakly enriched in weak double categories''; and in order to expand upon this we must first recall some relevant concepts pertaining to weak double categories. The concept of \emph{strict} double category is due to Ehresmann. It is an example of the notion of \emph{double model} for an essentially-algebraic theory, this being a model of the theory in its own category of ($\cat{Set}$-based) models. Thus a double category~--~which is a double model of the theory of categories~--~is a category object in $\cat{Cat}$. Now, the theory of categories is somewhat special, since its category of ($\cat{Set}$-based) models may be enriched to a 2-category, so that, as well as \emph{strict} category objects in $\cat{Cat}$, we may also consider \emph{pseudo} category objects: and these are the weak double categories which we will be interested in. \begin{Defn} A \defn{weak double category} $\mathfrak C$ is given by specifying a collection of \emph{objects} $x, y, z, \dots$, a collection of \emph{vertical 1-cells} between objects, which we write as $a \colon x \to y$, a collection of \emph{horizontal 1-cells} between objects, which we write as $f \colon x \ensuremath{\relbar\joinrel\mapstochar\joinrel\rightarrow} y$, and a collection of \emph{2-cells}, each of which is bounded by a square of horizontal and vertical arrows, and which we write as: \[\cd{ x \ar[d]_{a} \ar[r]|-{\object@{|}}^{f} \dtwocell{dr}{\alpha}& w \ar[d]^{b} \\ y \ar[r]|-{\object@{|}}_{g} & z\text,}\] or sometimes simply as $\alpha \colon f \Rightarrow g$. Moreover, we must give: \begin{itemize*} \item Identities and composition for vertical 1-cells, $\mathrm{id}_x \colon x \to x$ and $(a, b) \mapsto ab$, making the objects and vertical arrows into a category ${\mathcal C}_0$; \item Vertical identities and composition for 2-cells, $\mathrm{id}_{f} \colon f \Rightarrow f$ and $(\beta, \alpha) \mapsto \beta \alpha$: \[\cd{ x \ar[d]_{\mathrm{id}_x} \ar[r]|-{\object@{|}}^{f} \dtwocell{dr}{\mathrm{id}_{f}}& y \ar[d]^{\mathrm{id}_y} \\ x \ar[r]|-{\object@{|}}_{f} & y} \quad \text; \quad \cd{ u \ar[d]_{a} \ar[r]|-{\object@{|}}^{f} \dtwocell{dr}{\alpha}& x \ar[d]^{b} \\ v \ar[r]|-{\object@{|}}_{g} \ar[d]_{c} \dtwocell{dr}{\beta}& y \ar[d]^{d} \\ w \ar[r]|-{\object@{|}}_{h} & z} \mapsto \cd{u \ar[d]_{ca} \ar[r]|-{\object@{|}}^{f} \dtwocell{dr}{\beta\alpha}& x \ar[d]^{db} \\ w \ar[r]|-{\object@{|}}_{h} & z} \] making the horizontal arrows and 2-cells into a category ${\mathcal C}_1$ for which ``vertical source'' and ``vertical target'' become functors $s, t \colon {\mathcal C}_1 \to {\mathcal C}_0$; \item Identities and composition for horizontal 1-cells, $\DI x \colon x \ensuremath{\relbar\joinrel\mapstochar\joinrel\rightarrow} x$ and $(g, f) \mapsto gf$; \item Horizontal identities and composition for 2-cells, $\DI a \colon \DI x \Rightarrow \DI y$ and $(\beta, \alpha) \mapsto \beta \ast \alpha$: \[\cd{ x \ar[d]_{a} \ar[r]|-{\object@{|}}^{\DI x} \dtwocell{dr}{\DI a}& x \ar[d]^{a} \\ y \ar[r]|-{\object@{|}}_{\DI y} & y} \quad \text; \quad \cd{ u \ar[d]_{a} \ar[r]|-{\object@{|}}^{f} \dtwocell{dr}{\alpha}& v \ar[d]^{b} \ar[r]|-{\object@{|}}^{g} \dtwocell{dr}{\beta} & w \ar[d]^{c} \\ x \ar[r]|-{\object@{|}}_{h} & y \ar[r]|-{\object@{|}}_{k} & z} \mapsto \cd{u \ar[d]_{a} \ar[r]|-{\object@{|}}^{gf} \dtwocell{dr}{\beta \ast \alpha}& w \ar[d]^{c} \\ x \ar[r]|-{\object@{|}}_{kh} & z\text,} \] satisfying functoriality constraints: firstly, $\DI{({\mathord{\text{--}}})}$ is a functor ${\mathcal C}_0 \to {\mathcal C}_1$, which says that $\DI{\mathrm{id}_x} = \mathrm{id}_{\DI x}$ and $\DI{ab} = \DI a.\DI b$ and secondly, horizontal composition is a functor $\mathord{\ast} \colon {\mathcal C}_1 \mathbin{{}_s \times_t} {\mathcal C}_1 \to {\mathcal C}_1$ which says that $\mathrm{id}_{g} \ast \mathrm{id}_{f} = \mathrm{id}_{gf}$ and $(\delta \ast \gamma).(\beta \ast \alpha) = (\delta \beta) \ast (\gamma \alpha)$. \item Horizontal unitality and associativity constraints given by 2-cells \[\cd{ x \ar[d]_{\mathrm{id}_x} \ar[r]|-{\object@{|}}^{\DI y.f} \dtwocell{dr}{\l_{f}}& y \ar[d]^{\mathrm{id}_y} \\ x \ar[r]|-{\object@{|}}_{f} & y}\text, \quad \cd{ x \ar[d]_{\mathrm{id}_x} \ar[r]|-{\object@{|}}^{{f}.\DI x} \dtwocell{dr}{\r_{f}}& y \ar[d]^{\mathrm{id}_y} \\ x \ar[r]|-{\object@{|}}_{f} & y}\text, \quad \text{and} \quad \cd[@C+1.5em]{ x \ar[d]_{\mathrm{id}_x} \ar[r]|-{\object@{|}}^{h(gf)} \dtwocell{dr}{\a_{f,g,h}}& z \ar[d]^{\mathrm{id}_z} \\ x \ar[r]|-{\object@{|}}_{(hg)f} & z\text,} \] natural in $f$, $g$ and $h$, and invertible as arrows of ${\mathcal C}_1$. These 2-cells must obey two laws: the pentagon law, which equates the two routes from $k(h(gf))$ to $((kh)g)f$, and the triangle law, which equates the two routes from $g.(\DI y.f)$ to $gf$. \end{itemize*} \end{Defn} A more comprehensive reference on weak double categories and related matters is \cite{GP1}: though be aware that we interchange its usage of the terms ``horizontal'' and ``vertical'' to give a better fit with the usual 2-categorical terminology. Some simple examples of weak double categories are $\mathfrak{Cat}$, the weak double category of ``categories, functors, profunctors and transformations'', $\mathfrak{Rng}$, the weak double category of ``rings, ring homomorphisms, bimodules and skew-linear maps'', and the weak double category $\mathfrak{Span}({\mathcal C})$ of ``objects, morphisms, spans and span morphisms'' in a category with pullbacks ${\mathcal C}$. These are typical examples of weak double categories, in that they have notions of \emph{homomorphism} and \emph{bimodule} as their respective vertical and horizontal 1-cells. Any bicategory ${\mathcal B}$ gives us a weak double category ${\mathbb U}({\mathcal B})$ with only identity vertical 1-cells, whilst any weak double category $\mathfrak C$ gives a bicategory ${\mathbb H}(\mathfrak C)$ upon throwing away the non-identity vertical 1-cells, and all the 2-cells except for the \defn{globular 2-cells}, whose vertical source and target are identity arrows. Just as in the theory of bicategories, the appropriate notion of morphism between weak double categories only preserves horizontal composition up to comparison 2-cells, the most important case being the \emph{homomorphisms}, for which these 2-cells are invertible. We can define a homomorphism between small weak double categories in terms of a pseudomorphism of pseudocategory objects, but just as easy is to give the elementary description: \begin{Defn} A \defn{homomorphism of weak double categories} $F \colon \mathfrak C \to \mathfrak D$ is given by assignations on objects, 1-cells and 2-cells which preserve source and target and are functorial with respect to vertical composition of 1-\ and 2-cells, together with comparison 2-cells \[\cd{ {Fx} \ar[d]_{\mathrm{id}_{F x}} \ar[r]|-{\object@{|}}^{\DI{Fx}} \dtwocell{dr}{\mathfrak m_{x}}& {F x} \ar[d]^{\mathrm{id}_{F x}} \\ Fx \ar[r]|-{\object@{|}}_{F\DI x} & Fx} \quad \text{and} \quad \cd{ {F x} \ar[d]_{\mathrm{id}_{F x}} \ar[r]|-{\object@{|}}^{F g. F f} \dtwocell{dr}{\mathfrak m_{f, g}}& {F z} \ar[d]^{\mathrm{id}_{F z}} \\ Fx \ar[r]|-{\object@{|}}_{F (g f)} & Fz} \] which are invertible as arrows of ${\mathcal D}_1$, and natural in $x$, respectively $g$ and $f$. Moreover, we require the commutativity of three familiar diagrams, which equate, respectively, the two possible ways of going from $Ff.\DI{Fx}$ to $Ff$, from $\DI{Fy}. Ff$ to $Ff$, and from $Fh.(Fg.Ff)$ to $F((hg)f)$. \end{Defn} With the obvious notion of composition and identities, we obtain a category $\cat{DblCat}$ of (possibly large) weak double categories and homomorphisms between them. If we write $\cat{Bicat}$ for the category of bicategories and homomorphisms, then the assignations ${\mathcal B} \mapsto {\mathbb U}({\mathcal B})$ and $\mathfrak C \mapsto {\mathbb H}(\mathfrak C)$ described above extend to a pair of adjoint functors ${\mathbb U} \dashv {\mathbb H} \colon \cat{DblCat} \to \cat{Bicat}$, for which the composite ${\mathbb H}{\mathbb U}$ is the identity; we can thus view $\cat{Bicat}$ as a coreflective subcategory of $\cat{DblCat}$. Now, $\cat{DblCat}$ is in fact the underlying ordinary category of a 2-category whose 2-cells are the so-called \emph{vertical transformations}. We can understand these 2-cells by observing that there is a 2-monad (or even a $\cat{Cat}$-operad) on the 2-category $\cat{CatGph} := [\,\bullet \rightrightarrows \bullet, \, \cat{Cat}]$ whose strict algebras are small weak double categories, and whose algebra pseudomorphisms are the homomorphisms between them; now the corresponding \emph{algebra 2-cells} are precisely the vertical transformations. Spelling this out, we have: \begin{Defn} A \defn{vertical transformation} $\alpha \colon F \Rightarrow G$ between homomorphisms of weak double categories $F, G \colon \mathfrak C \to \mathfrak D$ is given by specifying, for each object $x \in \mathfrak C$, a vertical 1-cell $\alpha_x \colon Fx \to Gx$ of $\mathfrak D$ and for each horizontal 1-cell $f \colon x \ensuremath{\relbar\joinrel\mapstochar\joinrel\rightarrow} y$ in $\mathfrak {\mathcal C}$ a 2-cell \[\cd{ Fx \ar[d]_{\alpha_x} \ar[r]|-{\object@{|}}^{Ff} \dtwocell{dr}{\alpha_f}& Fy \ar[d]^{\alpha_y} \\ Gx \ar[r]|-{\object@{|}}_{Gf} & Gy}\] of $\mathfrak D$, such that the $\alpha_x$'s are natural in morphisms of ${\mathcal D}_0$, the $\alpha_f$'s are natural in morphisms of ${\mathcal D}_1$, and the following diagrams commute: \[\cd{ \DI{Fx} \ar@2[r]^{\mathfrak m^F_x} \ar@2[d]_{\DI{\alpha_x}} & F\DI x \ar@2[d]^{\alpha_{\DI x}} \\ \DI{Gx} \ar@2[r]_-{\mathfrak m^G_x} & G\DI x } \quad \text{and} \quad \cd{ Fg.Ff \ar@2[r]^{\mathfrak m^F_{g, f}} \ar@2[d]_{\alpha_g \ast \alpha_f} & F(gf) \ar@2[d]^{\alpha_{gf}} \\ Gg.Gf \ar@2[r]_-{\mathfrak m^G_{g, f}} & G(gf)\text.} \] \end{Defn} If we restrict our attention to the bicategories inside $\cat{DblCat}$, then the vertical transformation between homomorphisms are precisely the \emph{icons} of section 1; however, the reader should carefully note that the coreflection of $\cat{DblCat}$ into $\cat{Bicat}$ does \emph{not} enrich to a two-dimensional coreflection, since there is no way of coreflecting a general vertical transformation between homomorphisms of weak double categories into an icon between the corresponding homomorphisms of bicategories. The question of what we can say about this situation is one we return to in Section \ref{Sec:fromltt}. It follows from the algebraic description of $\cat{DblCat}$ that it admits a wide class of 2-dimensional limits, of which we will only be concerned with finite products. That $\cat{DblCat}$ admits these, makes it, of course, into a symmetric monoidal category, but the 2-dimensional aspect of these products give us a \emph{symmetric monoidal 2-category}: that is, a symmetric monoidal category whose tensor product is a 2-functor and whose coherence natural transformations are 2-natural transformations. What we now wish to describe is how we can use this monoidal 2-category $\cat{DblCat}$ as a suitable base for \emph{enrichment}. For any monoidal category ${\mathcal V}$, we have the well-known notion of a \emph{category enriched in ${\mathcal V}$}, which instead of having hom-sets between 0-cells, has hom-objects drawn from ${\mathcal V}$, with the corresponding composition being expressed by morphisms of ${\mathcal V}$ obeying laws assuring associativity and unitality. Slightly less well-known is the two-dimensional generalisation of this notion: if instead of a monoidal category ${\mathcal V}$, we begin with a \emph{monoidal bicategory} ${\mathcal W}$ in the sense of \cite{monoidalbicats}, we obtain a notion of \emph{bicategory enriched in ${\mathcal W}$} or \emph{${\mathcal W}$-bicategory}\footnote{Warning! This is emphatically \emph{not} the same as the notion of ``category enriched in a bicategory'' studied in \cite{Enrichedbicats}: this latter is really just an example of a category enriched in a monoidal category where the monoidal category happens to be spread out over many objects.}, which instead of hom-categories between 0-cells has hom-objects drawn from ${\mathcal W}$; the corresponding composition is once again expressed by morphisms of ${\mathcal W}$, which are now required to be associative and unital only up to coherent 2-cells. Just as a (locally small) category can be viewed as a category enriched in $\cat{Set}$, so a (locally small) bicategory can be viewed as a bicategory enriched in $\cat{Cat}$. Other straightforward examples are obtained by taking ${\mathcal W} = \cat{{\mathcal V}}\text-\cat{Cat}$ for some monoidal category ${\mathcal V}$, for which a ${\mathcal W}$-bicategory has sets of 0-\ and 1-\ cells as usual, but now a ${\mathcal V}$-object of 2-cells between any parallel pair of 1-cells; by taking ${\mathcal W} = \cat{Mod}$, the bicategory of categories and profunctors, for which a ${\mathcal W}$-bicategory is a \emph{probicategory} in the sense of Day \cite{Biclosed}; and by taking ${\mathcal W}$ to be an ordinary monoidal category, viewed as a locally discrete monoidal bicategory, whereupon ${\mathcal W}$-bicategories reduce to categories enriched in ${\mathcal W}$. An account of the general theory of enriched bicategories can be found in \cite{Lack}, but we will need sufficiently little of it that our account can be considered to be self-contained. Indeed, we will only consider in detail the case where ${\mathcal W}$ is the monoidal 2-category $\cat{DblCat}$ described above. \begin{Defn} A \defn{locally double bicategory} or $\cat{DblCat}$-\defn{bicategory} $\mathfrak B$ is given by the following data: \begin{enumerate}[(LDD1)] \item A collection $\ob {\mathfrak B}$ of objects; \item For every pair $A, B \in \ob {\mathfrak B}$, a weak double category ${\mathfrak B}(A, B)$; \item For every $A \in \ob {\mathfrak B}$, a unit homomorphism \[\elt{I_x} \colon 1 \to {\mathfrak B}(A, A)\text;\] \item For every triple $A, B, C \in \ob {\mathfrak B}$, a composition homomorphism \[\otimes \colon {\mathfrak B}(B, C) \times {\mathfrak B}(A, B) \to {\mathfrak B}(A, C)\text;\] \item For every pair $A, B \in \ob {\mathfrak B}$, invertible vertical transformations \[\cd[@C+1em]{ {\mathfrak B}(A, B) \times {\mathfrak B}(A, A) \ar[dr]_{\otimes} & {\mathfrak B}(A, B) \rtwocell[0.36]{dl}{r} \ltwocello[0.36]{dr}{l} \ar[r]^-{\elt{I_B} \times 1} \ar[l]_-{1 \times \elt{I_A}} \ar[d]_{1} & {\mathfrak B}(B, B) \times {\mathfrak B}(A, B)\text; \ar[dl]^{\otimes} \\ & {\mathfrak B}(A, B) & {} }\] \item For every quadruple $A, B, C, D \in \ob {\mathfrak B}$, an invertible vertical transformation \[\cd[@C+1em]{ {\mathfrak B}(C, D) \times {\mathfrak B}(B, C) \times {\mathfrak B}(A, B) \ar[d]_{\otimes \times 1} \ar[r]^-{1 \times \otimes} \dtwocell{dr}{a} & {\mathfrak B}(C, D) \times {\mathfrak B}(A, C) \ar[d]^{\otimes}\\ {\mathfrak B}(B, D) \times {\mathfrak B}(A, B) \ar[r]_-{\otimes} & {\mathfrak B}(A, D)\text.}\] \end{enumerate} This data is subject to the following two axioms: \begin{enumerate}[({LDA}1)] \item For each triple of objects $A, B, C$ of ${\mathfrak B}$, the following pasting equality holds: \[\cd[@+1em@C+1em]{ {\mathfrak B}^2 \ar[r]^{1 \times \elt{I_B} \times 1} \ar[d]_{1} \dtwocell{dr}{1 \times l} & {\mathfrak B}^3 \ar[r]^{1 \times \otimes} \ar[d]^{1 \times \otimes} & {\mathfrak B}^2 \ar[d]^{\otimes}\\ {\mathfrak B}^2 \ar[r]_1 & {\mathfrak B}^2 \ar[r]_{\otimes} & {\mathfrak B} } \quad = \quad \cd[@+1em@C+1em]{ {\mathfrak B}^2 \ar[r]^{1 \times \elt{I_B} \times 1} \ar[d]_{1} \dtwocell{dr}{r \times 1} & {\mathfrak B}^3 \ar[r]^{1 \times \otimes} \ar[d]^{\otimes \times 1} \dtwocell{dr}{a} & {\mathfrak B}^2 \ar[d]^{\otimes}\\ {\mathfrak B}^2 \ar[r]_1 & {\mathfrak B}^2 \ar[r]_{\otimes} & {\mathfrak B}\text, }\] where $\mathfrak B^2$ and $\mathfrak B^3$ abbreviate the appropriate products of hom-double categories; \item For each quintuple of objects $A, B, C, D, E$ of ${\mathfrak B}$, the following pasting equality holds: \[ \cd[@!@+1em]{ {\mathfrak B}^4 \ar[r]^{1 \times 1 \times \otimes} \ar[d]_{\otimes\times 1\times 1} \twocong{dr} & {\mathfrak B}^3 \ar[dr]^{1 \times \otimes} \ar[d]^{\otimes \times 1}\\ {\mathfrak B}^3 \ar[r]_{1 \times \otimes} \ar[dr]_{\otimes \times 1} & {\mathfrak B}^2 \ar[dr]^{\otimes} \drtwocell{d}{a} \dtwocell{r}{a} & {\mathfrak B}^2 \ar[d]^{\otimes} \\ & {\mathfrak B}^2 \ar[r]_{\otimes} & {\mathfrak B} } \quad = \quad \cd[@!@+1em]{ {\mathfrak B}^4 \ar[r]^{1 \times 1 \times \otimes} \ar[dr]|{1 \times \otimes \times 1} \ar[d]_{\otimes\times 1\times 1} & {\mathfrak B}^3 \ar[dr]^{1 \times \otimes} \drtwocell{d}{1 \times a}\\ {\mathfrak B}^3 \ar[dr]_{\otimes \times 1} \drtwocell[0.4]{r}{a \times 1} & {\mathfrak B}^3 \dtwocell{dr}{a} \ar[d]^{\otimes \times 1} \ar[r]_{1 \times \otimes} & {\mathfrak B}^2 \ar[d]^{\otimes} \\ & {\mathfrak B}^2 \ar[r]_{\otimes} & {\mathfrak B}\text, } \] where we observe the same convention regarding $\mathfrak B^4$, $\mathfrak B^3$ and $\mathfrak B^2$. \end{enumerate} \end{Defn} It may be helpful to extract a description of the various sorts of composition that a $\cat{DblCat}$-bicategory possesses. The 0-cells, 1-cells and vertical 2-cells form an ordinary bicategory. Next come the the horizontal 2-cells, which can be composed with each other along either a 1-cell boundary or a 0-cell boundary, with both compositions being associative up to an invertible 3-cell; moreover, the corresponding ``middle four interchange'' law only holds up to an invertible 3-cell. Finally, the 3-cells themselves can be composed with each other along the two different types of 2-cell boundary and along 0-cell boundaries; and these operations are strictly associative modulo the associativity of the boundaries. A \emph{one-object} locally double bicategory amounts to a \defn{monoidal double category} \cite{doubleclubs,GP2,mike}~--~that is, a weak double category with an up-to-isomorphism tensor product on it. In particular, any double category with \emph{finite products} in the appropriate double categorical sense\footnote{By which we mean a \emph{pseudo-functorial choice of double products} in the sense of \cite{GP1}. Such weak double categories are slightly stricter versions of the \emph{cartesian bicategories} of \cite{cb2}.} becomes a monoidal double category under the cartesian tensor. The double categories $\mathfrak{Cat}$, $\mathfrak{Span}({\mathcal C})$ (where ${\mathcal C}$ is a category with finite limits) and $\mathfrak{Rng}$ are all monoidal in this way: though in the case of $\mathfrak{Rng}$, one would more usually be interested in the other natural monoidal structure which is derived from the usual tensor product on the category of rings. For a non-degenerate example of a locally double bicategory, we turn to $\cat{DblCat}$ itself. As demonstrated in \cite{doubleclubs}, we may define an internal hom 2-functor \[\mathfrak{Hom}(\, {\mathord{\text{--}}} , ?) \colon \cat{DblCat}^\mathrm{op} \times \cat{DblCat} \to \cat{DblCat}\label{hom}\] for which $\mathfrak{Hom}(\mathfrak C, \mathfrak D)$ is the following double category. Its objects are homomorphisms $\mathfrak C \to \mathfrak D$, and its vertical 1-cells $\alpha \colon F \Rightarrow G$ are the vertical transformations between them. Its horizontal 1-cells $\alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$ are the \emph{horizontal pseudo-natural transformations}, whose components at an object $x \in \mathfrak C$ are given by horizontal 1-cells $\alpha_x \colon Fx \ensuremath{\relbar\joinrel\mapstochar\joinrel\rightarrow} Gx$ of $\mathfrak D$, satisfying naturality conditions like those for a pseudo-natural transformation between homomorphisms of bicategories; and indeed, in the case that $\mathfrak C$ and $\mathfrak D$ are themselves bicategories the two notions coincide. Finally, the 2-cells of $\mathfrak{Hom}(\mathfrak C, \mathfrak D)$ are the \emph{cubical modifications}, which are bounded by two horizontal and two vertical transformations and whose basic data consists of giving, for each object of the source, a 2-cell of the target bounded by the components of these transformations. When we say that $\mathfrak{Hom}$ acts as an internal hom, we are affirming a universal property: namely, that for each $\mathfrak C$ the 2-functor $({\mathord{\text{--}}}) \times \mathfrak C \colon \cat{DblCat} \to \cat{DblCat}$ is left biadjoint to $\mathfrak{Hom}(\, \mathfrak C, {\mathord{\text{--}}})$, so that what we have is a \emph{biclosed} monoidal bicategory in the sense of \cite{monoidalbicats}. Now, in \cite{Lack}, it is demonstrated that, just as any closed monoidal category can be viewed as a category enriched over itself, so can any biclosed monoidal bicategory be viewed as a bicategory enriched over itself, with the hom-objects being given by the biclosed structure. Applying this result to the 2-category $\cat{DblCat}$, we obtain a locally double bicategory $\mathfrak{DblCat}$, with 0-cells being the weak double categories; 1-cells, the homomorphisms; vertical 2-cells, the vertical transformations; horizontal 2-cells, the horizontal pseudo-natural transformations; and 3-cells the cubical modifications. Moreover, we may restrict our attention to the \emph{bicategories} inside $\mathfrak{DblCat}$ to obtain: \begin{Cor}\label{bicategories-as-dblcatbicategories} There is a locally double bicategory $\mathfrak{Bicat}$ which has as 0-cells, bicategories; as 1-cells, homomorphisms; as vertical 2-cells, icons; as horizontal 2-cells, pseudo-natural transformations; and as 3-cells, cubical modifications. \end{Cor} \noindent These particular cubical modifications are the same ones that we mentioned in Section \ref{bicat-tricat}, and we record their definition for future use: \begin{Defn}\label{cubbicat} Let $F, G, H, K \colon {\mathcal B} \to {\mathcal C}$ be homomorphisms of bicategories; let $\alpha \colon F \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} G$ and $\beta \colon H \ensuremath{\Relbar\joinrel\Mapstochar\joinrel\Rightarrow} K$ be pseudo-natural transformations; and let $\gamma \colon F \Rightarrow H$ and $\delta \colon G \Rightarrow K$ be icons. Then a \defn{cubical modification} \[\cd{ F \ar@2[d]_{\gamma} \ar@2[r]|-{\object@{|}}^{\alpha} \dthreecell{dr}{\Gamma}& G \ar@2[d]^{\delta} \\ H \ar@2[r]|-{\object@{|}}_{\beta} & K}\] is given by specifying, for every object $A \in {\mathcal B}$, a 2-cell $\Gamma_A \colon \alpha_A \Rightarrow \beta_A$, such that for every 1-cell $f \colon A \to B$ of ${\mathcal B}$, the following pasting equality holds: \[ \cd[@!]{ & FB \ar[r]^{\alpha_B} \rtwocell{d}{\alpha_f} \ar@{<-}[dl]_{Ff} & GB \ar@{=}[d] \ar@{<-}[dl]|{Gf} \\ FA \ar[r]^{\alpha_A} \ar@{=}[d] \dtwocell{dr}{\Gamma_A} & GA \ar@{=}[d] \dtwocell{r}{\delta_f} & KB \ar@{<-}[dl]^{Kf} \\ HA \ar[r]_{\beta_A} & KA\text; } \quad = \quad \cd[@!]{ & FB \ar[r]^{\alpha_B} \ar@{=}[d] \dtwocell{dr}{\Gamma_B} \ar@{<-}[dl]_{Ff} & GB \ar@{=}[d] \\ FA \ar@{=}[d] \dtwocell{r}{\gamma_f} & HB \ar[r]_{\beta_B} \ar@{<-}[dl]|{Hf} \rtwocell{d}{\beta_f} & KB \ar@{<-}[dl]^{Kf} \\ HA \ar[r]_{\beta_A} & KA\text. }\] \end{Defn} If we set $\gamma$ and $\delta$ to be identity icons in the above definition, we recapture the standard notion of a modification between pseudo-natural tramsformations, and thus the locally double bicategory $\mathfrak{Bicat}$ is rich enough to encode all of the structure of the tricategory of bicategories, but is able to do so using coherence whose complexity does not rise above the bicategorical level. Pleasing as this is, we should note that not every tricategory can be reduced to a locally double bicategory in this way; for example, given a bicategory ${\mathcal B}$ with bipullbacks, we may form the tricategory $\cat{Span}({\mathcal B})$ of spans in ${\mathcal B}$. In this tricategory, 1-cell composition is given by bipullback, and so is only determined up-to-equivalence, rather than up-to-isomorphism; so evidently, it will be inexpressible as a locally double bicategory. \begin{Rk} There are two canonical ways of forming a tricategory of bicategories, corresponding to the two canonical ways of composing a pair of strong transformations along a 0-cell boundary: however, Proposition \ref{bicategories-as-dblcatbicategories} exhibited a single canonical locally double bicategory of bicategories. The discrepancy is resolved if we observe that to obtain this $\cat{DblCat}$-bicategory we must fix a choice of biclosed structure on $\cat{DblCat}$, and that there are two canonical ways of doing this, depending on how we choose the counit maps $\mathfrak{Hom}(\mathfrak B, \mathfrak C) \times \mathfrak B \to \mathfrak C$ for the biadjunctions in question. \end{Rk} \endinput

https://arxiv.org/abs/2201.07967

A remark on Ehresmann's Fibration Theorem

This note records that in the setting of complex varieties, the cohomological consequence of Ehresmann's fibration theorem holds without the smooth assumption on the base or the total space.

\subsection*{Conventions} A `sheaf' means a `sheaf of vector spaces over some fixed field', and `variety' = `separated reduced scheme of finite type over $\mathrm{Spec}(\CC)$'. Sheaves on varieties are with respect to the complex analytic site. A proper map of topological spaces is a separated and universally closed map. J-L.\ Verdier asserts the following without the locally connected hypothesis \cite[Lemme 2.2.2]{V}. I was unable to understand his proof without this assumption. \begin{lemma}Let $p\colon X \to Y$ be a proper surjective map of topological spaces. Assume $X$ is locally connected. Let $\mathcal{F}$ be a sheaf on $Y$ with finite dimensional stalks. If $p^*\mathcal{F}$ is a local system, then so is $\mathcal{F}$. \end{lemma} \begin{proof} Let $y\in Y$. The stalk $\mathcal{F}_y$ is finite dimensional, so there exist sections $s_1, \ldots, s_n$, of $\mathcal{F}$ over some open neighborhood of $y$, which restrict to a basis of $\mathcal{F}_y$. Since our problem is local, we may assume this neighborhood is all of $Y$. Let $\mathcal{G}$ be the constant sheaf on $Y$ with stalk $\mathrm{span}\{s_1,\ldots,s_n\}$. Then the evident map $u\colon \mathcal{G} \to \mathcal{F}$ induces an isomorphism $\mathcal{G}_y \mapright{\sim}\mathcal{F}_y$. Consequently, $p^*u$ induces isomorphisms: \[ (p^*\mathcal{G})_x \mapright{\sim} (p^*\mathcal{F})_x \quad \mbox{for all $x\in p^{-1}(y)$}.\] For a locally connected space, the set of points at which a morphism of local systems induces an isomorphism on stalks defines an open set. Hence, the set $V\subset X$ of points at which $p^*u$ induces isomorphisms on stalks is open. As $p$ is proper, $U = Y - f(X-V)$ is an open neighborhood of $y$. As $p$ is surjective, $u$ yields an isomorphism $\mathcal{G}|_U \mapright{\sim} \mathcal{F}|_U$. \end{proof} \begin{prop} Let $f\colon Z \to Y$ be a smooth and proper morphism of varieties. Let $\mathcal{L}$ be a local system on $Z$ with finite dimensional stalks. Then the sheaves $R^qf_*\mathcal{L}$ are local systems. \end{prop} \begin{proof} Resolution of singularities (the version in \cite{BP} suffices), the Lemma and proper base change reduce us to the situation where $Z$ and $Y$ are smooth. Here the usual form of Ehresmann's Theorem applies. \end{proof}

https://arxiv.org/abs/1412.2837

Boundedness of the Images of Period Maps

We prove a conjecture of Griffiths on simultaneous normalization of all periods which asserts that the image of the lifted period map on the universal cover lies in a bounded domain in complex Euclidean space.

\section*{Introduction} First we introduce the notations in this paper. Following the terminology of Griffiths, all algebraic varieties are assumed to be smooth over $\mathbb{C}$ and morphisms between algebraic varieties to be rational and holomorphic. In this paper, we will consider the variation of Hodge structure arising from geometry. This means that we will consider an {algebraic family of polarized algebraic varieties}, which is a proper morphism between algebraic varieties $f : \mathfrak{X} \to S$ with the following properties \begin{itemize} \item[(1)] the varieties $\mathfrak{X}$ and $S$ are smooth and connected, and the morphism $f$ is non-degenerate, i.e. the tangent map $df$ is of maximal rank at each point of $\mathfrak{X}$; \item[(2)] $\mathfrak{X}\subseteq \mathbb{P}^N$ is quasi-projective with restricted polarization $\mathcal{L}$ over $\mathfrak{X}$; \item[(3)] each fiber $X_s=f^{-1}(s),\ s\in S$, is smooth, connected, and projective with the polarization $L_s=\mathcal{L}|_{X_s}$. \end{itemize} In general, $S$ is not compact, but it admits a smooth compactification due to Hironaka, i.e. we can embed $S$ as a Zariski open subset of a complete smooth algebraic variety $\bar{S}$ such that $\bar{S}-S$ is divisor with simple normal crossings. That is to say that $\bar{S}-S$ is locally defined by the equation $z_1\cdots z_k=0$ in suitable local coordinates $\{z_1,\cdots, z_n\}$ with $k\le \mbox{dim}_\mathbb{C} S=n$. Since each fiber $X=X_s$ is projective with polarization $L=\mathcal{L}|_{X}\in H^2(X,\mathbb{Z})$, we can define the primitive cohomology group by $$H^n_{pr}(X,\mathbb{F})=\mbox{ker}(L : H^n(X,\mathbb{F})\to H^{n+2}(X,\mathbb{F})),$$ where $n=\mbox{dim}_\mathbb{C} X$ and $\mathbb{F}$ can be $\mathbb{Z}$, $\mathbb{Q}$ and $\mathbb{C}$. From Hodge theory we have the Hodge decomposition of the primitive cohomology group $$H_{pr}^n(X,{\mathbb{C}})=H_{pr}^{n,0}(X)\oplus H_{pr}^{n-1,1}(X)\oplus\cdots\oplus H_{pr}^{0,n}(X),$$ where $H_{pr}^{k,n-k}(X)=H^{k,n-k}(X)\cap H_{pr}^n(X,{\mathbb{C}})$. The Poincar\'e bilinear form $Q$ on $H_{pr}^n(X,{\mathbb{C}})$ is defined by \begin{equation*} Q(u,v)=(-1)^{\frac{n(n-1)}{2}}\int_X u\wedge v \end{equation*} for any $d$-closed primitive $n$-forms $u,v$ on $X$. Then $Q$ is nondegenerate and satisfies the Hodge-Riemann relations \begin{eqnarray} Q\left ( H_{pr}^{n-k,k}(X), H_{pr}^{n-l,l}(X)\right )=0\text{ unless }k+l=n;\label{HR1}\\ \left (\sqrt{-1}\right )^{2k-n}Q\left ( v,\bar v\right )>0\text{ for }v\in H_{pr}^{k,n-k}(X)\setminus\{0\}. \label{HR2} \end{eqnarray} Let $h^{k,n-k}=\mbox{dim}_\mathbb{C} H_{pr}^{k,n-k}(X)$ and $f^k=\sum_{i=k}^nh^{i,n-i}$. Define the decreasing filtration $H_{pr}^n(X,{\mathbb{C}})=F^0\supset\cdots\supset F^n=0$ by taking $F^k=F^k(X)=H_{pr}^{n,0}(X)\oplus\cdots\oplus H_{pr}^{k,n-k}(X)$. We know that \begin{align} & \dim_{\mathbb{C}} F^k=f^k, \label{periodcondition} \\ & H^n_{pr}(X,{\mathbb{C}})=F^{k}\oplus \bar{F^{n-k+1}},\text{ and }H_{pr}^{k,n-k}(X)=F^k\cap\bar{F^{n-k}}.\nonumber \end{align} In terms of Hodge filtrations, the Hodge-Riemann relations \eqref{HR1} and \eqref{HR2} are \begin{align} & Q\left ( F^k,F^{n-k+1}\right )=0;\label{HR1'}\\ & Q\left ( Cv,\bar v\right )>0\text{ if }v\ne 0,\label{HR2'} \end{align} where $C$ is the Weil operator given by $Cv=\left (\sqrt{-1}\right )^{2k-n}v$ for $v\in H_{pr}^{k,n-k}(X)$. The period domain $D$ for polarized Hodge structures is the space of all such Hodge filtrations \begin{equation*} D=\left \{ F^n\subset\cdots\subset F^0=H_{pr}^n(X,{\mathbb{C}})\mid \eqref{periodcondition}, \eqref{HR1'} \text{ and } \eqref{HR2'} \text{ hold} \right \}. \end{equation*} The compact dual $\check D$ of $D$ is \begin{equation*} \check D=\left \{ F^n\subset\cdots\subset F^0=H_{pr}^n(X,{\mathbb{C}})\mid \eqref{periodcondition} \text{ and } \eqref{HR1'} \text{ hold} \right \}. \end{equation*} One can prove that $\check D$ is an algebraic manifold and the period domain $D\subseteq \check D$ is an open submanifold. See Theorem 4.3 in \cite{Griffiths1} for a complete proof and Proposition 8.2 in \cite{Griffiths3} for an alternative proof. From the definition of period domain we naturally get the Hodge bundles on $\check{D}$ by associating to each point in $\check{D}$ the vector spaces $\{F^k\}_{k=0}^n$ in the Hodge filtration of that point. We will denote the Hodge bundles by $\mathscr{F}^k \to \check{D}$ with $\mathscr{F}^k|_{p}=F^k_{p}$ as the fiber for any $p\in \check{D}$ and each $0\leq k\leq n$. For the family $f : \mathfrak{X} \to S$ and some fixed point $s_0\in S$, the period map is defined as a morphism $\Phi : S\to D/\Gamma$ by \begin{equation}\label{perioddefinition} s \mapsto \tau^{[\gamma_s]}(F^n_s\subseteq\cdots\subseteq F^0_s)\in D, \end{equation} where $F_s^k=F^k(X_s)$ and $\tau^{[\gamma_s]}$ is an isomorphism between $\mathbb{C}-$vector spaces $$\tau^{[\gamma_s]}:\, H^n(X_s,\mathbb{C})\to H^n(X_{s_0},\mathbb{C}),$$ which depends only on the homotopy class $[\gamma_s]$ of the curve $\gamma_s$ between $s$ and $s_0$. Then the period map is well-defined with respect to the monodromy representation $\rho : \pi_1(S)\to \Gamma \subseteq \text{Aut}(H_{\mathbb{Z}},Q)$. It is well-known that the period map has the following properties: \begin{itemize} \item[(i)] locally liftable; \item[(ii)] holomorphic: $\partial F^i_s/\partial \bar{s}\subseteq F^i_s$, $0\le i\le n$; \item[(iii)] Griffiths transversality: $\partial F^i_s/\partial s\subseteq F^{i-1}_s$, $1\le i\le n$. \end{itemize} Thanks to (i) we can lift the period map to $\tilde{\Phi} : \mathcal{T} \to D$ by taking the universal cover $\mathcal{T}$ of $S$ such that the diagram \begin{equation}\label{periodlifting} \xymatrix{ \mathcal{T} \ar[r]^-{\tilde{\Phi}} \ar[d]^-{\pi} & D\ar[d]^-{\pi}\\ S \ar[r]^-{\Phi} & D/\Gamma } \end{equation} is commutative. Without loss of generality we will assume that $\Gamma$ is torsion free, and therefore $D/\Gamma$ is smooth. Otherwise, we first choose $\Gamma$ as the whole $\text{Aut}(H_{\mathbb{Z}},Q)$, and then take a torsion free subgroup of $\Gamma$ and proceed on a finite cover of $S$. For the details, one can see the proof of Lemma IV-A, page 705-706 in \cite{Sommese}. In his paper \cite{Griffiths4}, Griffiths raised the following conjecture as Conjecture 10.1 in Section 10, which is now the main theorem of our paper. \begin{theorem}\label{conj} {\em(Griffiths Conjecture)} Given $f:\, \mathfrak{X} \to S$, there exists a simultaneous normalization of all the periods $\Phi(X_s)$ ($s\in S$). More precisely, the image $\tilde{\Phi}(\mathcal{T})$ lies in a bounded domain in a complex Euclidean space. \end{theorem} The main idea of our proof is Riemann extension theorem which asserts that \begin{itemize} \item Suppose that $M$ is a complex manifold and $V\subseteq M$ is an analytic subvariety with $\mbox{codim}_\mathbb{C} V\ge 1$. Then any holomorphic function $f$ defined on $M\setminus V$, which is locally bounded near $V$, can be extended uniquely to a global holomorphic function $\tilde{f}$ on $M$ such that $\tilde{f}|_{M\setminus V}=f$. \end{itemize} Based on this main idea, our proof can be divided into the following steps: \begin{itemize} \item[Step 1 :] Find an analytic subvariety of $\mbox{codim}_\mathbb{C} \ge 1$. \end{itemize} To explain the detail we need to review some results of period domain from Lie theory. We fix a point $p$ in $\mathcal{T}$ and its image $o=\tilde{\Phi}(p)$ as the reference points or base points. If we define the complex Lie group as \begin{align*} G_{\mathbb{C}}=\{ g\in GL(H_{\mathbb{C}})|~ Q(gu, gv)=Q(u, v) \text{ for all } u, v\in H_{\mathbb{C}}\}, \end{align*} then $G_{\mathbb{C}}$ acts transitively on $\check{D}$ with stabilizer $B=\{g\in G_\mathbb{C}| gF_p^k=F_p^k,\ 0\le k\le n\}.$ Let $G_{\mathbb{R}}\subseteq G_{\mathbb{C}}$ be the real subgroup which maps $H_{\mathbb{R}}$ to $H_{\mathbb{R}}$, then we can realize $D$ as $D=G_\mathbb{R}/V$ with $V=B\cap G_{\mathbb{R}}$. On the Lie algebra $\mathfrak{g}$ of $G_{\mathbb{C}}$ we can define a Hodge structure of weight zero by \begin{align*} \mathfrak{g}=\bigoplus_{k\in \mathbb{Z}} \mathfrak{g}^{k, -k}\quad\text{with}\quad\mathfrak{g}^{k, -k}=\{X\in \mathfrak{g}|XH^{r, n-r}_p\subseteq H^{r+k, n-r-k}_p,\ 0\le r \le n-k \}. \end{align*} By definition the Lie algebra of $B$ is $\mathfrak{b}=\bigoplus_{k\geq 0} \mathfrak{g}^{k, -k}$ and the holomorphic tangent space of $\check{D}$ at the base point $o$ is naturally isomorphic to $\mathfrak{g}/\mathfrak{b}\simeq \oplus_{k\geq 1}\mathfrak{g}^{-k,k}\triangleq \mathfrak{n}_+$. We denote the unipotent group to be $N_+=\exp(\mathfrak{n}_+)$. Since $N_+\cap B=\{\text{Id}\}$, we can identify the unipotent group $N_+\subseteq G_\mathbb{C}$ with its orbit $N_+(o)\subseteq \check{D}$ so that $N_+\subseteq \check{D}$ is meaningful. With this we define $\check{\mathcal{T}}=(\tilde{\Phi})^{-1}(N_+\cap D)$ and show that $\mathcal{T}\setminus \check{\mathcal{T}}$ is an analytic subvariety of $\mathcal{T}$ with $\mbox{codim}_\mathbb{C} (\mathcal{T}\setminus \check{\mathcal{T}})\ge 1$. \begin{itemize} \item[Step 2 :] Show that $\tilde{\Phi}|_{\check{\mathcal{T}}} : \check{\mathcal{T}}\to N_+\cap D$ is bounded. \end{itemize} Now we study in detail the structure of the Lie algebras involved. Suppose that the Weil operator $\theta: \,\mathfrak{g}\rightarrow \mathfrak{g}$ is defined by $$\theta(X)=(-1)^k X,\text{ for } X\in \mathfrak{g}^{k,-k}.$$ Then we can decompose Lie algebra $\mathfrak{g}$ into the union of eigenspaces of the Weil operator as $$\mathfrak{g}=\mathfrak{k}\oplus\mathfrak{p},$$ where $\mathfrak{k}$ and $\mathfrak{p}$ correspond to the eigenvalues $+1$ and $-1$ respectively. Let $\mathfrak{g}_0\subseteq \mathfrak{g}$ be the Lie algebra of $G_{\mathbb{R}}$ and $\mathfrak{k}_0=\mathfrak{k}\cap \mathfrak{g}_0$, $\mathfrak{p}_0=\mathfrak{p}\cap \mathfrak{g}_0$, then we get the decomposition of $\mathfrak{g}_0$ as $$\mathfrak{g}_0=\mathfrak{k}_0\oplus \mathfrak{p}_0.$$ In fact, Schmid showed that this decomposition is Cartan decomposition, i.e. $\mathfrak{g}_c\triangleq \mathfrak{k}_0\oplus \sqrt{-1}\mathfrak{p}_0$ is a compact real semi-simple Lie algebra. Next we choose a Cartan subalgebra $\mathfrak{h}_0$ of $\mathfrak{g}_0$ such that $\mathfrak{h}_0\subseteq \mathfrak{v}_0\subseteq \mathfrak{k}_0$ and $\mathfrak{h}_0$ is also a Cartan subalgebra of $\mathfrak{k}_0$, where $\mathfrak{v}_0=\mathfrak{g}_0\cap \mathfrak{g}^{0,0}$ is the Lie algebra of $V$. In Lie theory, root system plays a central role in the structures of Lie algebras. With the root system, one has the decomposition of the Lie algebra as $$\mathfrak{g}=\mathfrak{h}\oplus \sum_{\varphi\in\Delta}\mathfrak{g}^{\varphi}.$$ In our case, we decompose the root system $\Delta$ into the union of $\Delta_{\mathfrak{k}}$ and $\Delta_{\mathfrak{p}}$ due to the corresponding space $\mathfrak{g}^{\varphi}\subseteq \mathfrak{k} \text{ or } \mathfrak{p}$. Then we introduce the notion of strongly orthogonal which says that two different roots $\varphi, \psi\in \Delta$ are strongly orthogonal if and only if $\varphi\pm\psi\notin\Delta\cup \{0\}$. The following two properties are key to our proof. \begin{itemize} \item[(1)] There exists a set of strongly orthogonal noncompact positive roots $$\Lambda=\{\varphi_1, \cdots, \varphi_r\}\subseteq\Delta^+_{\mathfrak{p}}$$ such that \begin{align*} \mathfrak{A}_0=\sum_{i=1}^r\mathbb{R}\left(e_{{\varphi_i}}+e_{{-\varphi_i}}\right) \end{align*} is a maximal abelian subspace in $\mathfrak{A}_0$. \item[(2)] If $\mathfrak{A}_0'$ is any maximal abelian subspace of $\mathfrak{p}_0$, then there exists an element $k\in K$, a maximal compact subgroup of $G$, such that ${\mathrm{Ad}}(k)\cdot \mathfrak{A}_0=\mathfrak{A}'_0$. See Section \ref{Further} for the definition of $K$. Moreover, we have $$\mathfrak{p}_{0}=\bigcup_{k\in K}\mathrm{Ad}(k)\cdot\mathfrak{A}_0$$ \end{itemize} Let $\mathfrak a \subseteq \mathfrak{n}_+$ be the abelian subalgebra of $\mathfrak{n}_+$ determined by the period map $\tilde{\Phi}$. Let $A\triangleq \exp(\mathfrak a)\subseteq N_+$ which is isomorphic to a complex Euclidean subspace, and $P : N_+\cap D \to A\cap D$ be the induced projection map. The restricted period map $\tilde{\Phi} : \, \check{\mathcal{T}}\to N_+\cap D$ composed with the projection map $P$ gives a holomorphic map $\Psi :\, \check{\mathcal{T}} \to A\cap D$ by $\Psi=P \circ \tilde{\Phi}$. In Lemma \ref{abounded} we prove that $\Psi :\, \check{\mathcal{T}} \to A\cap D$ is bounded with respect to the Euclidean metric on $A\subseteq N_+$. In Theorem \ref{locallybounded}, we prove the boundedness of $\tilde{\Phi}(\check{\mathcal{T}})$ in $N_+$ by the finiteness of the map $P|_{\check{\mathcal{T}}}$, where we essentially use the Griffiths transversality of the extended period map, which is introduced in Section \ref{Lie}. \begin{itemize} \item[Step 3 :] By Step 1 and Step 2 and the Riemann Extension Theorem, we can finish the proof of Theorem \ref{conj}. \end{itemize} We remark that without further assumptions, we can only prove that the image $\tilde{\Phi}(\mathcal{T})$ lies in $\mathbb{C}^N$ as a bounded subvariety. In many cases, however, we can embed $\tilde{\Phi}(\mathcal{T})$ in $\mathbb{C}^N$ as a complex sub-manifold, even as bounded open domain. We will come back to this in our future work. \vspace{0.1cm} $\mathbf{Acknowledgement}$ We are very grateful to Professors Vincent Koziarz, Julien Maubon and Azniv Kasparian for their interest and useful comments. \section{Period domains from the viewpoint of Lie theory}\label{Lie} In this section we review the definitions and basic properties of period domains from Lie theory point of views. We consider the nilpotent Lie subalgebra $\mathfrak{n}_+$ and define the corresponding unipotent group to be $N_+=\exp(\mathfrak{n}_+)$. Since $N_+\cap B=\{\text{Id}\}$, we can identify the unipotent group $N_+\subseteq G_\mathbb{C}$ with its orbit $N_+(o)\subseteq \check{D}$. Then we define $\check {\mathcal{T}}=\tilde{\Phi}^{-1}(N_+\cap D)$ and show that $\mathcal{T}\backslash\check{\mathcal{T}}$ is an analytic subvariety of $\mathcal{T}$ with $\text{codim}_{\mathbb{C}}(\mathcal{T}\backslash\check{\mathcal{T}})\geq 1$. \\ First we fix a point $p$ in $\mathcal{T}$ and its image $o=\tilde{\Phi}(p)$ as the reference points or base points. Let us introduce the notion of adapted basis for the given Hodge decomposition or Hodge filtration. For the fixed point $p\in \mathcal{T}$ and $f^k=\dim F^k_p$ for any $0\leq k\leq n$, we call a basis $$\xi=\left\{ \xi_0, \cdots, \xi_{f^{n}-1},\xi_{f^{n}}, \cdots ,\xi_{f^{n-1}-1} \cdots, \xi_{f^{k+1}}, \cdots, \xi_{f^k-1}, \cdots, \xi_{f^{1}},\cdots , \xi_{f^{0}-1} \right\}$$ of $H^n_{pr}(X _p, \mathbb{C})$ an f{adapted basis for the given Hodge decomposition} $$H^n_{pr}(X _p, {\mathbb{C}})=H^{n, 0}_p\oplus H^{n-1, 1}_p\oplus\cdots \oplus H^{1, n-1}_p\oplus H^{0, n}_p, $$ if it satisfies $ H^{k, n-k}_p=\text{Span}_{\mathbb{C}}\left\{\xi_{f^{k+1}}, \cdots, \xi_{f^k-1}\right\}$ with $h^{k,n-k}=f^k-f^{k+1}$. We call a basis \begin{align*} \zeta=\left\{ \zeta_0, \cdots, \zeta_{f^{n}-1},\zeta_{f^{n}}, \cdots ,\zeta_{f^{n-1}-1} \cdots, \zeta_{f^{k+1}}, \cdots, \zeta_{f^k-1}, \cdots,\zeta_{f^{1}},\cdots , \zeta_{f^{0}-1} \right\} \end{align*} of $H^n_{pr}(X _p, {\mathbb{C}})$ an {adapted basis for the given filtration} \begin{align*} F^{n}_p\subseteq F^{n-1}_p\subseteq\cdots\subseteq F^0_p \end{align*} if it satisfies $F^{k}_p=\text{Span}_{\mathbb{C}}\{\zeta_0, \cdots, \zeta_{f^k-1}\}$ with $\text{dim}_{\mathbb{C}}F^{k}=f^k$. For the convenience of notations, we set $f^{n+1}=0$ and $m=f^0$. The blocks of an $m\times m$ matrix $T$ are set as follows: for each $0\leq \alpha, \beta\leq n$, the $(\alpha, \beta)$-th block $T^{\alpha, \beta}$ is \begin{align}\label{block} T^{\alpha, \beta}=\left(T_{ij}\right)_{f^{-\alpha+n+1}\leq i \leq f^{-\alpha+n}-1, \ f^{-\beta+n+1}\leq j\leq f^{-\beta+n}-1}, \end{align} where $T_{ij}$ is the entries of the matrix $T$. In particular, $T =(T^{\alpha,\beta})$ is called a {block lower triangular matrix} if $T^{\alpha,\beta}=0$ whenever $\alpha<\beta$. Let $H_{\mathbb{F}}=H^n_{pr}(X, \mathbb{F})$, where $\mathbb{F}$ can be chosen as $\mathbb{Z}$, $\mathbb{R}$, $\mathbb{C}$. We define the complex Lie group \begin{align*} G_{\mathbb{C}}=\{ g\in GL(H_{\mathbb{C}})|~ Q(gu, gv)=Q(u, v) \text{ for all } u, v\in H_{\mathbb{C}}\}, \end{align*} and the real one \begin{align*} G_{\mathbb{R}}=\{ g\in GL(H_{\mathbb{R}})|~ Q(gu, gv)=Q(u, v) \text{ for all } u, v\in H_{\mathbb{R}}\}. \end{align*} Griffiths in \cite{Griffiths1} showed that $G_{\mathbb{C}}$ acts on $\check{D}$ transitively and so does $G_{\mathbb{R}}$ on $D$. The stabilizer of $G_{\mathbb{C}}$ on $\check{D}$ at the fixed point $o$ is $$B=\{g\in G_\mathbb{C}| gF_p^k=F_p^k,\ 0\le k\le n\},$$ and the one of $G_{\mathbb{R}}$ on $D$ is $V=B\cap G_\mathbb{R}$. Thus we can realize $\check{D}$ as $$\check{D}=G_\mathbb{C}/B,\text{ and }D=G_\mathbb{R}/V$$ so that $\check{D}$ is an algebraic manifold and $D\subseteq \check{D}$ is an open complex submanifold. One can find a complete proof of this result in Theorem 4.3 of \cite{Griffiths1}, or Proposition 8.2 in \cite{Griffiths3} for an alternative proof. The Lie algebra $\mathfrak{g}$ of the complex Lie group $G_{\mathbb{C}}$ is \begin{align*} \mathfrak{g}&=\{X\in \text{End}(H_\mathbb{C})|~ Q(Xu, v)+Q(u, Xv)=0, \text{ for all } u, v\in H_\mathbb{C}\}, \end{align*} and the real subalgebra $$\mathfrak{g}_0=\{X\in \mathfrak{g}|~ XH_{\mathbb{R}}\subseteq H_\mathbb{R}\}$$ is the Lie algebra of $G_\mathbb{R}$. Note that $\mathfrak{g}$ is a simple complex Lie algebra and contains $\mathfrak{g}_0$ as a real form, i.e. $\mathfrak{g}=\mathfrak{g}_0\oplus \sqrt{-1}\mathfrak{g}_0$. On the linear space $\text{Hom}(H_\mathbb{C},H_\mathbb{C})$ we can give a Hodge structure of weight zero by \begin{align*} \mathfrak{g}=\bigoplus_{k\in \mathbb{Z}} \mathfrak{g}^{k, -k}\quad\text{with}\quad\mathfrak{g}^{k, -k}=\{X\in \mathfrak{g}|XH^{r, n-r}_p\subseteq H^{r+k, n-r-k}_p,\ 0\le r \le n-k \}. \end{align*} By definition of $B$ the Lie algebra $\mathfrak{b}$ of $B$ has the form $\mathfrak{b}=\bigoplus_{k\geq 0} \mathfrak{g}^{k, -k}$. Then the Lie algebra $\mathfrak{v}_0$ of $V$ is $$\mathfrak{v}_0=\mathfrak{g}_0\cap \mathfrak{b}=\mathfrak{g}_0\cap \mathfrak{b}\cap\bar{\mathfrak{b}}=\mathfrak{g}_0\cap \mathfrak{g}^{0, 0}.$$ With the above isomorphisms, the holomorphic tangent space of $\check{D}$ at the base point is naturally isomorphic to $\mathfrak{g}/\mathfrak{b}$. Let us consider the nilpotent Lie subalgebra $\mathfrak{n}_+:=\oplus_{k\geq 1}\mathfrak{g}^{-k,k}$. Then one gets the holomorphic isomorphism $\mathfrak{g}/\mathfrak{b}\cong \mathfrak{n}_+$. We denote the unipotent group to be $$N_+=\exp(\mathfrak{n}_+).$$ As $\text{Ad}(g)(\mathfrak{g}^{k, -k})$ is in $\bigoplus_{i\geq k}\mathfrak{g}^{i, -i} \text{ for each } g\in B,$ the subspace $\mathfrak{b}\oplus \mathfrak{g}^{-1, 1}/\mathfrak{b}\subseteq \mathfrak{g}/\mathfrak{b}$ defines an Ad$(B)$-invariant subspace. By left translation via $G_{\mathbb{C}}$, $\mathfrak{b}\oplus\mathfrak{g}^{-1,1}/\mathfrak{b}$ gives rise to a $G_{\mathbb{C}}$-invariant holomorphic subbundle of the holomorphic tangent bundle. It will be denoted by $T^{1,0}_{h}\check{D}$, and will be referred to as the horizontal tangent subbundle. One can check that this construction does not depend on the choice of the base point. The horizontal tangent subbundle, restricted to $D$, determines a subbundle $T_{h}^{1, 0}D$ of the holomorphic tangent bundle $T^{1, 0}D$ of $D$. The $G_{\mathbb{C}}$-invariace of $T^{1, 0}_{h}\check{D}$ implies the $G_{\mathbb{R}}$-invariance of $T^{1, 0}_{h}D$. Note that the horizontal tangent subbundle $T_{h}^{1, 0}D$ can also be constructed as the associated bundle of the principle bundle $V\to G_\mathbb{R} \to D$ with the adjoint representation of $V$ on the space $\mathfrak{b}\oplus\mathfrak{g}^{-1,1}/\mathfrak{b}$. As another interpretation of the horizontal bundle in terms of the Hodge bundles $\mathscr{F}^{k}\to \check{D}$, $0\le k \le n$, one has \begin{align}\label{horizontal} T^{1, 0}_{h}\check{D}\simeq T^{1, 0}\check{D}\cap \bigoplus_{k=1}^{n}\text{Hom}(\mathscr{F}^{k}/\mathscr{F}^{k+1}, \mathscr{F}^{k-1}/\mathscr{F}^{k}). \end{align} In \cite{schmid1}, Schmid call a holomorphic mapping $\Psi: \,M\rightarrow \check{D}$ of a complex manifold $M$ into $\check{D}$ {horizontal} if the tangent map $d\Psi:\, T^{1,0}M \to T^{1,0}\check{D}$ takes values in $T^{1,0}_h\check{D}$. The period map $\tilde{\Phi}: \, \mathcal{T}\rightarrow D$ is horizontal due to Griffiths transversality. \begin{remark}\label{N+inD} We remark that elements in $N_+$ can be realized as nonsingular block lower triangular matrices with identity blocks in the diagonal; elements in $B$ can be realized as nonsingular block upper triangular matrices. If $c, c'\in N_+$ such that $cB=c'B$ in $\check{D}$, then $c'^{-1}c\in N_+\cap B=\{I \}$, i.e. $c=c'$. This means that the matrix representation in $N_+$ of the unipotent orbit $N_+(o)$ is unique. Therefore with a fixed base point $o\in \check{D}$, we can identify $N_+$ with its unipotent orbit $N_+(o)$ in $\check{D}$ by identifying an element $c\in N_+$ with $[c]=cB$ in $\check{D}$. Then $N_+\subseteq\check{D}$ is meaningful. In particular, when the base point $o$ is in $D$, we have $N_+\cap D\subseteq D$. \end{remark} Now we define \begin{align*} \check{\mathcal{T}}=\tilde{\Phi}^{-1}(N_+\cap D). \end{align*} Then we show that $\mathcal{T}\backslash\check{\mathcal{T}}$ is an analytic subvariety of $\mathcal{T}$ with $\text{codim}_{\mathbb{C}}(\mathcal{T}\backslash\check{\mathcal{T}})\geq 1$. \begin{lemma}\label{transversal}Let $p\in\mathcal{T}$ be the reference point with $\tilde{\Phi}(p)=\{F^n_p\subseteq F^{n-1}_p\subseteq \cdots\subseteq F^0_p\}.$ Let $q\in \mathcal{T}$ be any point with $\tilde{\Phi}(q)=\{F^n_q\subseteq F^{n-1}_q\subseteq \cdots\subseteq F^0_q\}$, then $\tilde{\Phi}(q)\in N_+$ if and only if $F^{k}_q$ is isomorphic to $F^k_p$ for all $0\leq k\leq n$. \end{lemma} \begin{proof}For any $q\in \mathcal{T}$, we choose an arbitrary adapted basis $\{\zeta_0, \cdots, \zeta_{m-1}\}$ for the given Hodge filtration $\{F^{n}_q\subseteq F^{n-1}_q\subseteq\cdots\subseteq F^0_q\}$. We fix $\{\eta_0, \cdots, \eta_{m-1}\}$ as the adapted basis for the Hodge filtration $\{F^{n}_p \subseteq F^{n-1}_p\subseteq \cdots\subseteq F^0_p\}$ at the base point $p$. Let $[A^{i,j}(q)]_{0\leq i,j\leq n}$ be the transition matrix between the basis $\{\eta_0,\cdots, \eta_{m-1}\}$ and $\{\zeta_0, \cdots, \zeta_{m-1}\}$ for the same vector space $H_\mathbb{C}$, where $A^{i,j}(q)$ are the corresponding blocks. Then $\tilde{\Phi}(q)\in N_+=N_+B/B\subseteq \check{D}$ if and only if its matrix representation $[A^{i,j}(q)]_{0\leq i,j\leq n}$ can be decomposed as $L(q)\cdot U(q)$, where $L(q)$ is a nonsingular block lower triangular matrix with identities in the diagonal blocks, and $U(q)$ is a nonsingular block upper triangular matrix. By basic linear algebra, we know that $[A^{i,j}(q)]$ has such decomposition if and only if $\det[A^{i,j}(q)]_{0\leq i, j\leq k}\neq 0$ for any $0\leq k\leq n$. In particular, we know that $[A(q)^{i,j}]_{0\leq i,j\leq k}$ is the transition map between the bases of $F^k_p$ and $F^k_q$. Therefore, $\det([A(q)^{i,j}]_{0\leq i,j\leq k})\neq 0$ if and only if $F^k_q$ is isomorphic to $F^k_p$. \end{proof} \begin{proposition}\label{codimension} The subset $\check{\mathcal{T}}$ is an open complex submanifold in $\mathcal{T}$, and $\mathcal{T}\backslash \check{\mathcal{T}}$ is an analytic subvariety of $\mathcal{T}$ with $\text{codim}_{\mathbb{C}}(\mathcal{T}\backslash \check{\mathcal{T}})\geq 1$. \end{proposition} \begin{proof} From Lemma \ref{transversal}, one can see that $\check{D}\setminus N_+\subseteq \check{D}$ is defined as an analytic subvariety by equations \begin{align*} \{q\in \check{D} : \det ((A^{i,j}(q))_{0\leq i,j\leq k})=0\text{ for some } 0\leq k\leq n\}. \end{align*} Therefore $N_+$ is dense in $\check{D}$, and that $\check{D}\setminus N_+$ is an analytic subvariety, which is closed in $\check{D}$ and with $\text{codim}_{\mathbb{C}}(\check{D}\backslash N_+)\geq 1$. We consider the period map $\tilde{\Phi}:\,\mathcal{T}\to \check{D}$ as a holomorphic map to $\check{D}$, then $\mathcal{T}\setminus \check{\mathcal{T}}=\tilde{\Phi}^{-1}(\check{D}\setminus N_+)$ is the preimage of $\check{D}\setminus N_+$ of the holomorphic map $\tilde{\Phi}$. Therefore $\mathcal{T}\setminus \check{\mathcal{T}}$ is also an analytic subvariety and a closed set in $\mathcal{T}$. Because $\mathcal{T}$ is smooth and connected, $\mathcal{T}$ is irreducible. If $\dim_\mathbb{C}(\mathcal{T}\setminus \check{\mathcal{T}})=\dim_\mathbb{C}\mathcal{T}$, then $\mathcal{T}\setminus \check{\mathcal{T}}=\mathcal{T}$ and $\check{\mathcal{T}}=\emptyset$, but this contradicts to the fact that the reference point $p$ is in $\check{\mathcal{T}}$. Thus we conclude that $\dim_\mathbb{C}(\mathcal{T}\setminus \check{\mathcal{T}})<\dim_\mathbb{C}\mathcal{T}$, and consequently $\text{codim}_{\mathbb{C}}(\mathcal{T}\backslash \check{\mathcal{T}})\geq 1$. \end{proof} In the introduction, we have assumed that $S$ admits a compactification $\bar{S}$ such that $\bar{S}$ is smooth and complete and $\bar{S}\setminus S$ is a divisor with simple normal crossings. Let $S'\supseteq S$ be the subset of $\bar{S}$ to which the period map $\Phi:\, S\to D/\Gamma$ extends continuously and $\Phi' : \, S'\to D/\Gamma$ be the extended map. Then one has the commutative diagram \begin{equation*} \xymatrix{ S \ar@(ur,ul)[r]+<16mm,4mm>^-{\Phi}\ar[r]^-{i} &S' \ar[r]^-{\Phi'} & D/\Gamma. } \end{equation*} with $i :\, S\to S'$ the inclusion map. \begin{lemma}\label{openness1} $S'$ is an open complex submanifold of $\bar{S}$ and the complex codimension of $\bar{S}\setminus S'$ is at least one. \end{lemma} \begin{proof} Here by abusing notation, in the following discussion we still denote by $\Gamma$ the discrete subgroup in $\text{Aut}(H_{\mathbb{Z}},Q)$ containing the monodromy group which is the image of $\pi_1(S)$. As a standard procedure, after going to a finite cover we may assume $\Gamma$ is torsion free, or even neat. See Theorem 3.6 in \cite{Milne}. We will give two proofs of the lemma. First we can use Theorem 9.6 of Griffiths in \cite{Griffiths3}, see also Corollary 13.4.6 in \cite{CMP}, to get the Zariski open submanifold $S''$ of $\bar{S}$, where $S''$ contains all points of finite monodromy in $\bar{S}$, and hence of trivial monodromy in $\bar{S}$, since $\Gamma$ is torsion free. Then the extension map $\Phi'':\, S''\to D/\Gamma$ is a proper holomorphic map. For the related discussion, see also page 705-706 in \cite{Sommese}. Since $S''$ is Zariski open in $\bar{S}$, we only need to prove that $S'=S''$. In fact, by the definition of $S'$, we see that $S''\subseteq S'$. Conversely, for any point $q\in S'$ with image $u=\Phi'(q)\in D/\Gamma$, we can choose the points $q_{k}\in S$, $k=1,2,\cdots$ such that $q_{k}\longrightarrow q$ with images $u_{k}=\Phi(q_{k})\longrightarrow u$ as $k\longrightarrow \infty$. Since $\Phi'':\, S''\to D/\Gamma$ is proper, the sequence $\{q_{k}\}_{k=1}^{\infty}\subset (\Phi'')^{-1}(\{u_{k}\}_{k=1}^{\infty})$ has a limit point $q$ in $S''$, that is to say $q\in S''$ and $S'\subseteq S''$. From this we see that $S'=S''$ is an open complex submanifold of $\bar{S}$ with $\text{codim}_{\mathbb{C}}(\bar{S}\setminus S')\ge 1$. For the second proof, note that $\bar{S}$ is smooth and $S\subseteq \bar{S}$ is Zariski open, we only need to show that $S'$ is open in $\bar{S}$. To prove this we use the compactification space $\bar{D/\Gamma}$. There are several natural notions of the compactification space $\bar{D/\Gamma}$, see \cite{KU}, \cite[Page 2]{Green14}, \cite[Page 29, 30]{Green14}, in which it is proved that the period map has continuous, even holomorphic extension. We can choose any one of them together with the continuous extension of the period map as $$\bar{\Phi} :\, \bar{S}\to \bar{D/\Gamma}.$$ By the definition of $S'$, $S'= \bar{\Phi}^{-1}(D/\Gamma)$. Since $D/\Gamma$ is open and dense in the compacification $\overline{D/\Gamma}$, $S'$ is therefore an open and dense submanifold of $\bar{S}$. \end{proof} Let $\mathcal{T}'$ be the universal cover of $S'$ with the universal covering map $\pi':\, \mathcal{T}'\to S'$. We then have the following commutative diagram \begin{equation}\label{maindiagram} \xymatrix{\mathcal{T} \ar[r]^{i_{\mathcal{T}}}\ar[d]^{\pi}&\mathcal{T}'\ar[d]^{\pi'}\ar[r]^{\tilde{\Phi}'}&D\ar[d]^{\pi_D}\\ S\ar[r]^{i}&S'\ar[r]^{\Phi'}&D/\Gamma, } \end{equation} where $i_\mathcal{T}$ is the lifting of $i\circ \pi$ with respect to the covering map $\pi': \, \mathcal{T}'\to S'$ and $\tilde{\Phi}'$ is the lifting of $\Phi'\circ \pi'$ with respect to the covering map $\pi_D:\, D\to D/\Gamma$. Then $\tilde{\Phi}'$ is continuous. There are different choices of $i_\mathcal{T}$ and $\tilde{\Phi}'$, but Lemma A.1 in the Appendix shows that we can choose $i_\mathcal{T}$ and $\tilde{\Phi}'$ such that $\tilde{\Phi}=\tilde{\Phi}'\circ i_\mathcal{T}$. Let $\mathcal{T}_0\subseteq \mathcal{T}'$ be defined by $\mathcal{T}_0=i_\mathcal{T}(\mathcal{T})$. \begin{lemma}\label{openness2} $\mathcal{T}_0=\pi'^{-1}(S)$. \end{lemma} \begin{proof} From the diagram (\ref{maindiagram}), we see that $\pi'(\mathcal{T}_0)=\pi'(i_\mathcal{T}(\mathcal{T}))=i(\pi(\mathcal{T}))=S$, hence $\mathcal{T}_0 \subseteq \pi'^{-1}(S)$. Conversely, for any $q \in \pi'^{-1}(S)$, we need to prove that $q\in \mathcal{T}_0$. Let $p=\pi'(q)\in S$. If there exists $r\in \pi^{-1}(p)$ such that $i_\mathcal{T}(r)=q$, then we are done. Otherwise, we can draw a curve $\gamma$ from $i_\mathcal{T}(r)$ to $q$ for some $r\in \pi^{-1}(p)$, as $\mathcal{T}'$ is connected and thus path connected. Then we get a circle $\Gamma=\pi'(\gamma)$ in $S'$. But Lemma A.2 in the Appendix implies that we can choose the circle $\Gamma$ contained in $S$. Note that $p\in \Gamma$. Since $\pi : \, \mathcal{T} \to S$ is covering map, we can lift $\Gamma$ to a unique curve $\tilde{\gamma}$ from $r$ to some $r'\in \pi^{-1}(p)$. Notice that both $\gamma$ and $i_\mathcal{T}(\tilde{\gamma})$ map to $\Gamma$ via the the covering map $\pi' :\, \mathcal{T}' \to S'$, that is $\gamma$ and $i_\mathcal{T}(\tilde{\gamma})$ are both the lifts of $\Gamma$ starting from the same point $i_\mathcal{T}(r)$. By the uniqueness of homotopy lifting, $i_\mathcal{T}(r')=q$, i.e. $q\in i_\mathcal{T}(\mathcal{T})=\mathcal{T}_0$. \end{proof} Lemma \ref{openness2} implies that $\mathcal{T}_0$ is an open complex submanifold of $\mathcal{T}'$ and $\text{codim}_\mathbb{C}(\mathcal{T}'\setminus \mathcal{T}_0)\ge 1$. Since $\tilde{\Phi}=\tilde{\Phi}'\circ i_\mathcal{T}$ is holomorphic, $\tilde{\Phi}'|_{\mathcal{T}_0} : \, \mathcal{T}_0\to D$ is also holomorphic. Since $\tilde{\Phi}' : \mathcal{T}' \to D$ is continuous, $\tilde{\Phi}'$ is locally bounded around $\mathcal{T}'\setminus \mathcal{T}_0$, then by applying Riemann extension theorem we have that $\tilde{\Phi}' :\, \mathcal{T}' \to D$ is holomorphic. \begin{lemma}\label{extendedtransversality} The extended holomorphic map $\tilde{\Phi}' :\, \mathcal{T}' \to D$ satisfies the Griffiths transversality. \end{lemma} \begin{proof} Let $\text{T}^{1,0}_hD$ be the horizontal subbundle. Since $\tilde{\Phi}' : \mathcal{T}' \to D$ is a holomorphic map, the tangent map $$\tilde{\Phi}'_* : \, \text{T}^{1,0}\mathcal{T}' \to \text{T}^{1,0}D$$ is at least continuous. We only need to show that the image of $\tilde{\Phi}'_*$ is contained in the horizontal tangent bundle $\text{T}^{1,0}_hD$. Since $\text{T}^{1,0}_hD$ is closed in $\text{T}^{1,0}D$, $(\tilde{\Phi}'_*)^{-1}(\text{T}^{1,0}_hD)$ is closed in $\text{T}^{1,0}\mathcal{T}'$. But $\tilde{\Phi}'|_{\mathcal{T}_0}$ satisfies the Griffiths transversality, i.e. $(\tilde{\Phi}'_*)^{-1}(\text{T}^{1,0}_hD)$ contains $\text{T}^{1,0}\mathcal{T}_0$, which is open in $\text{T}^{1,0}\mathcal{T}'$. Hence $(\tilde{\Phi}'_*)^{-1}(\text{T}^{1,0}_hD)$ contains the closure of $\text{T}^{1,0}\mathcal{T}_0$, which is $\text{T}^{1,0}\mathcal{T}'$. \end{proof} \section{Further results of period domains from Lie theory}\label{Further} In this section, we study the structure of the Lie algebra $\mathfrak{g}$ by considering the root system, which is fundamental to our proof of Theorem \ref{locallybounded} about the boundeness of the image of the restricted period map. Many results are from \cite{GS} and \cite{schmid1} to which we refer the reader for detailed proofs. \\ Now we define the Weil operator $\theta: \,\mathfrak{g}\rightarrow \mathfrak{g}$ by $$\theta(X)=(-1)^k X,\text{ for } X\in \mathfrak{g}^{k,-k}.$$ Then $\theta$ is an involutive automorphism of ${\mathfrak g}$, and defined over $\mathbb{R}$. Let $\mathfrak{k}$ and $\mathfrak{p}$ be the $(+1)$ and $(-1)$ eigenspaces of $\theta$ respectively. Considering the types, we have $$\mathfrak{k}=\oplus_{k \text{ even }}\mathfrak{g}^{k,-k}, \ \mathfrak{p}=\oplus_{k \text{ odd }}\mathfrak{g}^{k,-k}.$$ Set \begin{align*} \mathfrak{k}_0=\mathfrak{k}\cap \mathfrak{g}_0, \quad \mathfrak{p}_0=\mathfrak{p}\cap \mathfrak{g}_0. \end{align*} Then we have the decompositions \begin{align*} \mathfrak{g}=\mathfrak{k}\oplus\mathfrak{p}, \quad \mathfrak{g}_0=\mathfrak{k}_0\oplus \mathfrak{p}_0 \end{align*} with the property that \begin{align*} [\mathfrak{k}, \,\mathfrak{k}]\subseteq \mathfrak{k}, \quad [\mathfrak{p},\,\mathfrak{p}]\subseteq\mathfrak{k}, \quad [\mathfrak{k}, \,\mathfrak{p}]\subseteq \mathfrak{p}. \end{align*} Let $\mathfrak{g}_c=\mathfrak{k}_0\oplus \sqrt{-1}\mathfrak{p}_0$. Then $\mathfrak{g}_c$ is also a real form of $\mathfrak{g}$. Let us denote the complex conjugation of $\mathfrak{g}$ with respect to the real form $\mathfrak{g}_c$ by $\tau_c$, and the complex conjugation of $\mathfrak{g}$ with respect to the real form $\mathfrak{g}_0$ by $\tau_0$. Recall that on the complex Linear space $H_\mathbb{C}$ we can define an Hermitian inner product $(\cdot,\cdot)$ induced by the Poincar\'e bilinear form $Q$ as \begin{equation}\label{Hermitian} (u,v)=Q(Cu,\bar{v}) \ \ u,v\in H_\mathbb{C}, \end{equation} where $C$ is the Weil operator on $H_\mathbb{C}$ and defined over $\mathbb{R}$. Thus $C$ can be considered as an element in $G_\mathbb{R}$, whose adjoint action on $\mathfrak{g}$ is just $\theta$. For any $Z=X+\sqrt{-1} Y\in {\mathfrak g}_c$, where $X\in \mathfrak{k}_0$ and $Y\in \mathfrak{p}_0$, we have that $\forall\, u,v \in H_\mathbb{C}$ \begin{eqnarray} (Z\cdot u, v)&=&Q(C((X+\sqrt{-1} Y)\cdot u),\bar{v}) \label{H1} \\ &=&Q((X-\sqrt{-1} Y)\cdot Cu,\bar{v})\nonumber \\ &=&-Q(Cu,(X-\sqrt{-1} Y)\cdot \bar{v})\nonumber \\ &=&-Q(Cu, \bar{(X+\sqrt{-1} Y)\cdot v})\nonumber \\ &=&-(u,Z\cdot v).\nonumber \end{eqnarray} Thus ${\mathfrak g}_c$ is the intersection of ${\mathfrak g}$ with the algebra of all skew Hermitian transforms with respect to the Hermitian inner product $(\cdot,\cdot)$. Using this result, Schmid in \cite{schmid1} proved that: \begin{itemize} \item ${\mathfrak g}_c$ is a compact real form of ${\mathfrak g}$, and the Killing form $B$ $$B(X,Y)=\text{Trace}(\text{ad}X\circ \text{ad}Y), \ \ X,Y\in{\mathfrak g}$$ restricts to a negative definite bilinear form $B|_{{\mathfrak g}_c}$ on ${\mathfrak g}_c$. Moreover, one has an Hermitian inner product $-B(\theta\ \cdot,\bar{\cdot})$ on ${\mathfrak g}$, making ${\mathfrak g}$ an Hermitian complex linear space. \end{itemize} Following this result, we have that $\mathfrak{g}_0=\mathfrak{k}_0\oplus \mathfrak{p}_0$ is Cartan decomposition. Now we define the Lie subgroup $G_c\subseteq G_\mathbb{C}$ corresponding to ${\mathfrak g}_c\subseteq{\mathfrak g}$. Then (\ref{H1}) implies that $G_c$ contains the elements in $G_\mathbb{C}$ which preserve the Hermitian inner product $(\cdot,\cdot)$, i.e. $G_c$ is the unitary subgroup of $G_\mathbb{C}$. Thus $G_c$ is compact. As noted by Schmid, a compact real form in a connected complex semisimple Lie group is always connected and is its own normalizer, which implies that $G_c$ is also connected. The intersection $$K=G_c\cap G_{\mathbb{R}}$$ is a compact subgroup of $G_{\mathbb{R}}$ with Lie algebra $\mathfrak{g}_c\cap {\mathfrak g}_0=\mathfrak{k}_0$. In pages 278-279 of \cite{schmid1}, Schmid showed that: \begin{itemize} \item $K$ is a maximal compact subgroup of $G_{\mathbb{R}}$ and it meets every connected component of $G_{\mathbb{R}}$. \item $G_c\cap B=V$, which implies $V\subseteq K$ and their Lie algebras $\mathfrak{v}_0\subseteq \mathfrak{k}_0$. \end{itemize} In \cite{GS}, Griffiths and Schmid observed that: \begin{itemize} \item There exists a Cartan subalgebra $\mathfrak{h}_0$ of $\mathfrak{g}_0$ such that $\mathfrak{h}_0\subseteq \mathfrak{v}_0\subseteq \mathfrak{k}_0$ and $\mathfrak{h}_0$ is also a Cartan subalgebra of $\mathfrak{k}_0$; \end{itemize} Denote $\mathfrak{h}$ to be the complexification of $\mathfrak{h}_0$. Then $\mathfrak{h}$ is a Cartan subalgebra of $\mathfrak{g}$ such that $\mathfrak{h}\subseteq \mathfrak{v}\subseteq \mathfrak{k}$. Now we review the root systems we need. Write $\mathfrak{h}_0^*=\text{Hom}(\mathfrak{h}_0, \mathbb{R})$ and $\mathfrak{h}^*_{_\mathbb{R}}=\sqrt{-1}\mathfrak{h}^*_0. $ Then $\mathfrak{h}^*_{_\mathbb{R}}$ can be identified with $\mathfrak{h}_{_\mathbb{R}}:=\sqrt{-1}\mathfrak{h}_0$ by the restricted Killing form $B|_{\mathfrak{h}_{\mathbb{R}}}$ on $\mathfrak{h}_{\mathbb{R}}$. Let $\rho\in\mathfrak{h}_{{\mathbb{R}}}^*\simeq \mathfrak{h}_{{\mathbb{R}}}$, one can define the following subspace of $\mathfrak{g}$, \begin{align*} \mathfrak{g}^{\rho}=\{x\in \mathfrak{g}| [h, \,x]=\rho(h)x\quad\text{for all } h\in \mathfrak{h}\}. \end{align*} An element $\varphi\in\mathfrak{h}_{{\mathbb{R}}}^*\simeq\mathfrak{h}_{{\mathbb{R}}}$ is called a root of $\mathfrak{g}$ with respect to $\mathfrak{h}$ if $\mathfrak{g}^\varphi\neq \{0\}$. Let $\Delta\subseteq \mathfrak{h}^{*}_{_\mathbb{R}}\simeq\mathfrak{h}_{{\mathbb{R}}}$ denote the space of nonzero $\mathfrak{h}$-roots. Then each root space \begin{align*} \mathfrak{g}^{\varphi}=\{x\in\mathfrak{g}|[h,x]=\varphi(h)x \text{ for all } h\in \mathfrak{h}\} \end{align*} with respect to some $\varphi\in \Delta$ is one-dimensional over $\mathbb{C}$, generated by a root vector $e_{{\varphi}}$. Since the involution $\theta$ is a Lie-algebra automorphism fixing $\mathfrak{k}$, we have $$\theta([h, \theta(e_{{\varphi}})])=[h, e_{{\varphi}}]$$ and hence $$[h, \theta(e_{{\varphi}})]=\theta([h, e_{{\varphi}}])=\varphi(h)\theta(e_{{\varphi}}),$$ for any $h\in \mathfrak{h}$ and $\varphi\in \Delta.$ Thus $\theta(e_{{\varphi}})$ is also a root vector belonging to the root $\varphi$, so $e_{{\varphi}}$ must be an eigenvector of $\theta$. It follows that there is a decomposition of the roots $\Delta$ into the union $\Delta_{\mathfrak{k}}\cup\Delta_{\mathfrak{p}}$ of compact roots and non-compact roots with root spaces $\mathbb{C}e_{{\varphi}}\subseteq \mathfrak{k}$ and $\mathfrak{p}$ respectively. The adjoint representation of $\mathfrak{h}$ on $\mathfrak{g}$ determins a decomposition \begin{align*} \mathfrak{g}=\mathfrak{h}\oplus \sum_{\varphi\in\Delta}\mathfrak{g}^{\varphi}. \end{align*} Let $\{\phi_i: 1\leq i\leq l\}$, $l=\text{rank}(\mathfrak{g})$ be any simple root system of $\Delta$, then we can introduce the notion of positivity in $\Delta$, i.e. we can define $\Delta^+$ to be the subset of $\Delta$ consisting of the roots in $\Delta$ which are positive integral linear combination of the simple roots. Let $\{h_i: 1\leq i\leq l\}$ be the basis of ${\mathfrak h}_\mathbb{R}$ corresponding to the simple roots. That is to say that $\phi_i(h)=B(h_i,h)$ for any $h\in{\mathfrak h}$, where $1\leq i\leq l$ and the Cartan matrix $(B(h_i,h_j))$ is positive definite. Following Serre \cite{Serre}, we can choose a Weyl basis $\{e'_{\alpha} : \alpha\in\Delta \}$ such that \begin{align*} &[{\mathfrak h},{\mathfrak h}]=0;\\ &[h,e'_\alpha]=\alpha (h)e'_\alpha, \ \forall h \in {\mathfrak h};\\ &[e'_\alpha,e'_{-\alpha}]=B(e'_\alpha,e'_{-\alpha})h_\alpha, \ \forall \phi \in \Delta;\\ &[e'_\alpha,e'_{\beta}]=N_{\alpha,\beta}e'_{\alpha+\beta}, \ \alpha +\beta\ne 0, \end{align*} where $N_{\alpha,\beta}=0$ if $\alpha+\beta \ne 0$, $\alpha+\beta \notin \Delta$; $N_{\alpha,\beta}\ne 0$ if $\alpha+\beta \in \Delta$ with relation that $$N_{-\alpha,-\beta}=-N_{\alpha,\beta}\in \mathbb{R}.$$ Now we define the real subspace of ${\mathfrak g}$ as $${\mathfrak g}'_c={\mathfrak h}_0+\sum_{\alpha\in\Delta}\mathbb {R}(e'_\alpha-e'_{-\alpha})+\sum_{\alpha\in\Delta}\mathbb {R}\sqrt{-1}(e'_\alpha+e'_{-\alpha}).$$ In fact, Theorem 6.3 in Chapter III of \cite{Hel} shows that ${\mathfrak g}'_c$ is a compact real form of ${\mathfrak g}$ with complex conjugation $\tau'_c$. Again by Theorem 7.1 in Chapter III of \cite{Hel} and its proof, there exists one-parameter subgroup ${\vartheta}_t$ of automorphisms of ${\mathfrak g}$ such that ${\vartheta}_{\frac{1}{4}}({\mathfrak g}'_c)$ is invariant under $\tau_c$ and ${\vartheta}_1=(\tau_c\tau'_c)^2$. Then the proof of Theorem 7.2 and Corollary 7.3 in Chapter III of \cite{Hel} show that ${\mathfrak g}_c={\vartheta}_\frac{1}{4}({\mathfrak g}'_c)$ and there exist some $X\in{\mathfrak g}$ such that ${\vartheta}_t=\text{exp}(t\cdot \text{ad}X)$. Since $\text{exp}(\text{ad}X)=(\tau_c\tau'_c)^2$ and $\tau_c|_{{\mathfrak h}_0}=\tau'_c|_{{\mathfrak h}_0}=\text{id}$, $X$ must be in ${\mathfrak h}$, hence ${\vartheta}_t|_{{\mathfrak h}_0}=\text{exp}(t\cdot \text{ad}X)|_{{\mathfrak h}_0}=\text{id}$. Therefore we have proved that there exists an automorphism ${\vartheta}={\vartheta}_\frac{1}{4}$ of ${\mathfrak g}$ such that ${\mathfrak g}_c={\vartheta}({\mathfrak g}'_c)$ and ${\vartheta}|_{{\mathfrak h}_0}=\text{id}$. For any $\alpha\in \triangle$, let $e_\alpha\triangleq {\vartheta}_{\frac{1}{4}}(e'_\alpha)$. Since ${\vartheta}_{\frac{1}{4}}$ is an automorphism, $\{e_\alpha : \alpha\in \triangle\}$ is also a Weyl basis. Then under the Weyl basis $\{e_\alpha : \alpha\in \triangle\}$, \begin{eqnarray*} {\mathfrak g}_c&=&{\vartheta}_{\frac{1}{4}}({\mathfrak g}'_c)\\ &=&{\mathfrak h}_0+\sum_{\alpha\in\Delta}\mathbb {R}(e_\alpha-e_{-\alpha})+\sum_{\alpha\in\Delta}\mathbb {R}\sqrt{-1}(e_\alpha+e_{-\alpha})\\ &=&\mathfrak{k}_0+\sqrt{-1}\mathfrak{p}_0. \end{eqnarray*} Therefore, by considering types we have proved the following theorem. \begin{thm}\label{rootdecomp} There exist a basis $\{h_i : 1\le i\le l\}$ of ${\mathfrak h}_\mathbb{R}$ and a Weyl basis $\{e_{{\varphi}} : \varphi\in\Delta\}$ such that \begin{align} &\tau_c(h_i)=\tau_0(h_i)=-h_i, \quad \text{ for any }1\leq i\leq l;\nonumber \\ &\tau_c(e_{{\varphi}})=\tau_0(e_{{\varphi}})=-e_{{-\varphi}}, \quad \text{for any } \varphi\in \Delta_{\mathfrak{k}};\label{bar}\\ &\tau_0(e_{{\varphi}})=-\tau_c(e_{{\varphi}})=e_{{-\varphi}}, \quad\text{for any }\varphi\in \Delta_{\mathfrak{p}},\nonumber \end{align} and \begin{align} \mathfrak{k}_0&=\mathfrak{h}_0+\sum_{\varphi\in \Delta_{\mathfrak{k}}}\mathbb{R}(e_{{\varphi}}-e_{{-\varphi}})+\sum_{\varphi\in\Delta_{\mathfrak{k}}} \mathbb{R}\sqrt{-1}(e_{{\varphi}}+e_{{-\varphi}});\label{rootk}\\ \mathfrak{p}_0&=\sum_{\varphi\in \Delta_{\mathfrak{p}}}\mathbb{R}(e_{{\varphi}}+e_{{-\varphi}})+\sum_{\varphi\in\Delta_{\mathfrak{p}}} \mathbb{R}\sqrt{-1}(e_{{\varphi}}-e_{{-\varphi}}).\label{rootp} \end{align} \end{thm} \begin{lemma}\label{puretype}Let $\Delta$ be the set of $\mathfrak{h}$-roots as above. Then for each root $\varphi\in \Delta$, there is an integer $-n\leq k\leq n$ such that $e_{\varphi}\in \mathfrak{g}^{k,-k}$. In particular, if $e_{{\varphi}}\in \mathfrak{g}^{k, -k}$, then $e_{-\varphi}=-\tau_0(e_{{\varphi}})\in \mathfrak{g}^{-k,k}$ for any $-n\leq k\leq n$. \end{lemma} \begin{proof} Let $\varphi$ be a root, and $e_{\varphi}$ be the generator of the root space $\mathfrak{g}^{\varphi}$, then $e_{\varphi}=\sum_{k=-n}^n e^{-k,k}$, where $e^{-k,k}\in \mathfrak{g}^{-k,k}$. Because $\mathfrak{h}\subseteq \mathfrak{v}\subseteq \mathfrak{g}^{0,0}$, the Lie bracket $[h,e^{-k,k}]\in \mathfrak{g}^{-k,k}$ for each $k$. Then the condition $[h,e_{\varphi}]=\varphi (h)e_{\varphi}$ implies that \begin{align*} \sum_{k=-n}^n[h,e^{-k,k}]=\sum_{k=-n}^n\varphi(h)e^{-k,k}\quad \text{ for each } h\in \mathfrak{h}. \end{align*} By comparing the type, we get \begin{align*} [h,e^{-k,k}]=\varphi(h)e^{-k,k}\quad \text{ for each } h\in \mathfrak{h}. \end{align*} Therefore $e^{-k,k}\in g^{\varphi}$ for each $k$. As $\{e^{-k,k}\}_{k=-n}^n$ forms a linear independent set, but $g^{\varphi}$ is one dimensional, thus there is only one $-n\leq k\leq n$ with $e^{-k,k}\neq 0$. \end{proof} Let us now introduce a lexicographic order (cf. pp.41 in \cite{Xu} or pp.416 in \cite{Sugi}) in the real vector space $\mathfrak{h}_{{\mathbb{R}}}$ as follows: we fix an ordered basis $h_1, \cdots, h_l$ for $\mathfrak{h}_{{\mathbb{R}}}$. Then for any $h=\sum_{i=1}^l\lambda_ih_i\in\mathfrak{h}_{{\mathbb{R}}}$, we call $h>0$ if the first nonzero coefficient is positive, that is, if $\lambda_1=\cdots=\lambda_k=0, \lambda_{k+1}>0$ for some $1\leq k<l$. For any $h, h'\in\mathfrak{h}_{{\mathbb{R}}}$, we say $h>h'$ if $h-h'>0$, $h<h'$ if $h-h'<0$ and $h=h'$ if $h-h'=0$. In particular, let us identify the dual spaces $\mathfrak{h}_{{\mathbb{R}}}^*$ and $\mathfrak{h}_{{\mathbb{R}}}$, thus $\Delta\subseteq \mathfrak{h}_{{\mathbb{R}}}$. Let us choose a maximal linearly independent subset $\{\phi_1, \cdots, \phi_s\}$ of $\Delta_{\mathfrak{p}}$, then a maximal linearly independent subset $\{\phi_{s+1}, \cdots, \phi_{l}\}$ of $\Delta_{\mathfrak{k}}$. Then $\{\phi_1, \cdots, \phi_s, \phi_{s+1}, \cdots, \phi_l\}$ forms a basis for $\mathfrak{h}_{{\mathbb{R}}}^*$ since $\text{Span}_{\mathbb{R}}\Delta=\mathfrak{h}_{{\mathbb{R}}}^*$. Then define the above lexicographic order in $\mathfrak{h}_{{\mathbb{R}}}^*\simeq \mathfrak{h}_{{\mathbb{R}}}$ using the ordered basis $\{\phi_1,\cdots, \phi_l\}$. In this way, we can also define \begin{align*} \Delta^+=\{\varphi>0: \,\varphi\in \Delta\};\quad \Delta_{\mathfrak{p}}^+=\Delta^+\cap \Delta_{\mathfrak{p}}. \end{align*} Similarly we can define $\Delta^-$, $\Delta^-_{\mathfrak{p}}$, $\Delta^{+}_{\mathfrak{k}}$, and $\Delta^-_{\mathfrak{k}}$. \begin{lemma}\label{RootSumNonRoot}Using the above notation, we have \begin{equation}\label{R1} (\Delta_{\mathfrak{k}}+\Delta^{\pm}_{\mathfrak{p}})\cap \Delta\subseteq\Delta_{\mathfrak{p}}^{\pm}; \end{equation} \begin{equation}\label{R2} (\Delta_{\mathfrak{p}}^{+}+\Delta_{\mathfrak{p}}^{+})\cap\Delta=\emptyset;\quad (\Delta_{\mathfrak{p}}^{-}+\Delta_{\mathfrak{p}}^{-})\cap\Delta=\emptyset. \end{equation} If one defines $$\mathfrak{p}^{\pm}=\sum_{\varphi\in\Delta^{\pm}_{\mathfrak{p}}} \mathfrak{g}^{\varphi}\subseteq\mathfrak{p},$$ then $\mathfrak{p}=\mathfrak{p}^+\oplus\mathfrak{p}^-$, $[\mathfrak{p}^+, \,\mathfrak{p}^-] \subseteq\mathfrak{k}$ and \begin{equation}\label{R3} [\mathfrak{k},\, \mathfrak{p}^{\pm}] \subseteq\mathfrak{p}^{\pm}; \end{equation} \begin{equation}\label{R4} [\mathfrak{p}^{+}, \, \mathfrak{p}^{+}]=0, \ [\mathfrak{p}^{-}, \, \mathfrak{p}^{-}]=0. \end{equation} \end{lemma} \begin{proof} Relation (\ref{R1}) follows directly from the definition of lexicographic order, which also implies (\ref{R3}). For (\ref{R2}), let $\phi=\sum_{i>s}a_i\phi_i,\ \psi=\sum_{i>s}b_i\phi_i$ in $\Delta_{\mathfrak{p}}^{+}$, then $\phi+\psi=\sum_{i>s}(a_i+b_i)\phi_i \in \Delta_{\mathfrak{p}}^{+}$ and $e_{\phi},\ e_{\psi},\ e_{\phi+\psi} \in \mathfrak{p}$. Suppose that $\phi+\psi\in \Delta$, then $$0\ne[e_{\phi},\ e_{\psi}]=N_{\phi+\psi}e_{\phi+\psi}\in [\mathfrak{p},\ \mathfrak{p}]\subseteq \mathfrak{k}.$$ Then $0\ne e_{\phi+\psi}\in \mathfrak{k}\cap \mathfrak{p}=\{0\}$, which is a contradiction. Therefore (\ref{R2}) holds, from which (\ref{R4}) follows. \end{proof} \begin{definition}Two different roots $\varphi, \psi\in \Delta$ are said to be strongly orthogonal if and only if $\varphi\pm\psi\notin\Delta\cup \{0\}$, which is denoted by $\varphi \independent\psi$. \end{definition} \begin{lemma}\label{stronglyortho}There exists a set of strongly orthogonal noncompact positive roots $\Lambda=\{\varphi_1, \cdots, \varphi_r\}\subseteq\Delta^+_{\mathfrak{p}}$ such that $0<\varphi_1< \cdots < \varphi_r$ and \begin{align*} \mathfrak{A}_0=\sum_{i=1}^r\mathbb{R}\left(e_{{\varphi_i}}+e_{{-\varphi_i}}\right) \end{align*} is a maximal abelian subspace in $\mathfrak{p}_0$. \end{lemma} \begin{proof} Let $\varphi_1$ be the minimum in $\Delta_\mathfrak{p}^+$, and $\varphi_2$ be the minimal element in $\{\varphi\in\Delta_{\mathfrak{p}}^+:\,\varphi\independent \varphi_1\}$, then we obtain inductively an maximal ordered set of roots $\Lambda=\{\varphi_1,\cdots, \varphi_r\}\subseteq \Delta_\mathfrak{p}^+$, such that for each $1\leq k\leq r$ \begin{align*} \varphi_k=\min\{\phi\in\Delta_{\mathfrak{p}}^+:\,\varphi\independent \varphi_j \text{ for } 1\leq j\leq k-1\}. \end{align*} Because $\varphi_i\independent \varphi_j$ for any $1\leq i<j\leq r$, we have $[e_{{\pm\varphi_i}}, e_{{\pm\varphi_j}}]=0$. Therefore $$\mathfrak{A}_0=\sum_{i=1}^r\mathbb{R}\left(e_{{\varphi_i}}+e_{{-\varphi_i}}\right)$$ is an abelian subspace of $\mathfrak{p}_0$. We claim that $\varphi_1,\cdots, \varphi_r$ are $\mathbb{R}-$linearly independent. In fact, since $\varphi_i\pm \varphi_j\notin \Delta\cup \{0\}$ for $i\ne j$, we conclude that $B(\varphi_i, \varphi_j)=0$, otherwise either $\frac{2B(\varphi_i, \varphi_j)}{B(\varphi_j, \varphi_j)}$ or $\frac{2B(\varphi_j, \varphi_i)}{B(\varphi_i, \varphi_i)}$ equals to $\pm 1$ and we assume that $\frac{2B(\varphi_i, \varphi_j)}{B(\varphi_j, \varphi_j)}=\pm 1$, then the property of root system implies that $$\varphi_i\mp \varphi_j=\varphi_i-\frac{2B(\varphi_i, \varphi_j)}{B(\varphi_j, \varphi_j)}\varphi_j\in \Delta,$$ which is a contradiction. Now we assume that $\sum_{i}k_i\varphi_i=0$ with each $k_i\in\mathbb{R}$. Then $$0=B(\sum_{i}k_i\varphi_i,\sum_{i}k_i\varphi_i)=\sum_{i}k_i^2B(\varphi_i,\varphi_i).$$ As $B|_{{\mathfrak h}_0}$ is negative definite, we conclude that each $k_i=0$. It remains to prove that $\alpha_0$ is maximal. If it is not so, we can find $X\in \mathfrak{p}_0$ with $$X=\sum_{\alpha\in \Delta^+_{\mathfrak{p}}\setminus \Lambda}\lambda_\alpha(e_\alpha+e_{-\alpha})+\sum_{\alpha\in \Delta^+_{\mathfrak{p}}\setminus \Lambda}\mu_\alpha\sqrt{-1}(e_\alpha-e_{-\alpha}), \ \lambda_\alpha,\mu_\alpha\in \mathbb{R}$$ such that $[X, e_{{\varphi_i}}+e_{{-\varphi_i}}]=0$ for $1\leq i\leq r$. Let $c_\alpha=\lambda_\alpha+\sqrt{-1}\mu_\alpha$, then $X=\sum_{\alpha\in \Delta^+_{\mathfrak{p}}\setminus \Lambda}(c_\alpha e_{\alpha}+\bar{c_\alpha}e_{-\alpha})$ and for $1\le i\le r$ \begin{eqnarray} 0&=&[X, e_{{\varphi_i}}+e_{{-\varphi_i}}] \nonumber \\ &=&\sum_{\alpha\in \Delta^+_{\mathfrak{p}}\setminus \Lambda}\big( c_\alpha(N_{\alpha+\varphi_i}e_{\alpha+\varphi_i}+N_{\alpha-\varphi_i}e_{\alpha-\varphi_i})+\bar{c_\alpha}(N_{-\alpha+\varphi_i}e_{-\alpha+\varphi_i}+N_{-\alpha-\varphi_i}e_{-\alpha-\varphi_i})\big ).\label{maxabel} \end{eqnarray} By Lemma \ref{RootSumNonRoot}, for any $1\le i\le r$, $\alpha+\varphi_i\notin \Delta$, hence $N_{\alpha+\varphi_i}=0$. Similarly, $N_{-\alpha-\varphi_i}=0$. Then (\ref{maxabel}) implies \begin{equation}\label{maxabel2} \sum_{\alpha\in \Delta^+_{\mathfrak{p}}\setminus \Lambda}\big( c_\alpha N_{\alpha-\varphi_i}e_{\alpha-\varphi_i}+\bar{c_\alpha}N_{-\alpha+\varphi_i}e_{-\alpha+\varphi_i}\big )=0. \end{equation} For any $\alpha\in \Delta^+_{\mathfrak{p}}\setminus \Lambda$, by the construction of $\Lambda$ there exists an $i$ such that $\alpha-\varphi_i$ lies in $\Delta$, and then $N_{\alpha-\varphi_i}\ne 0$. Equation (\ref{maxabel2}) then implies that $c_\alpha=0$. Hence $X=0$. \end{proof} For further use, we also state a proposition about the maximal abelian subspace of $\mathfrak{p}_0$ as Lemma 6.3 in Chapter V of \cite{Hel}. \begin{proposition}\label{adjoint} Let $\mathfrak{A}_0'$ be an arbitrary maximal abelian subspaces of $\mathfrak{p}_0$, then there exists an element $k\in K$ such that $\emph{Ad}(k)\cdot \mathfrak{A}_0=\mathfrak{A}'_0$. Moreover, we have $$\mathfrak{p}_{0}=\bigcup_{k\in K}\emph{Ad}(k)\cdot\mathfrak{A}_0,$$ where $\emph{Ad}$ denotes the adjoint action of $K$ on $\mathfrak{A}_0$. \end{proposition} \section{Boundedness of the period map}\label{bound} In the previous two sections, we have discussed the basic properties of the period domains. In particular, we have found the analytic subvariety $\mathcal{T}\setminus \check{\mathcal{T}}$ with $\mbox{codim}_\mathbb{C} (\mathcal{T}\setminus \check{\mathcal{T}})\ge 1$. In this section, we will prove the boundedness of the restricted period map $\tilde{\Phi}|_{\check{\mathcal{T}}} :\, \check{\mathcal{T}}\to N_+\cap D$ by using the structure theory of the complex semi-simple Lie algebra ${\mathfrak g}$. \\ Recall that we have fixed the reference points $p\in \mathcal{T}$ and $o=\tilde{\Phi}(p)\in D$. Then $N_+$ can be viewed as a subset in $\check{D}$ by identifying it with its orbit $N_+(o)$ in $\check{D}$. At the base point $\tilde{\Phi}(p)=o\in N_+\cap D$, the tangent space $T_{o}^{1,0}N_+=T_o^{1,0}D\simeq \mathfrak{n}_+$ and the exponential map $\text{exp} :\, \mathfrak{n}_+ \to N_+$ is an isomorphism. Then the Hodge metric on $T_o^{1,0}D$ induces an Euclidean metric on $N_+$ so that $\text{exp} :\, \mathfrak{n}_+ \to N_+$ is an isometry. Also recall that we have defined $\check{\mathcal{T}}=\tilde{\Phi}^{-1}(N_+\cap D)$ and have shown that $\mathcal{T}\backslash\check{\mathcal{T}}$ is an analytic subvariety of $\mathcal{T}$ with $\text{codim}_{\mathbb{C}}(\mathcal{T}\backslash\check{\mathcal{T}})\geq 1$. In this section, we prove that $\tilde{\Phi} :\, \check{\mathcal{T}}\to N_+\cap D$ is bounded in $N_+$ with respect to the Euclidean metric on $N_+$. Let $\mathfrak a \subseteq \mathfrak{n}_+$ be the abelian subalgebra of $\mathfrak{n}_+$ determined by the tangent map of period map $$\tilde{\Phi}_* :\, \text{T}^{1,0}\check{\mathcal{T}} \to \text{T}^{1,0}D.$$ By Griffiths transversality, $\mathfrak a\subseteq {\mathfrak g}^{-1,1}$ is an abelian subspace. Let $$A\triangleq \exp(\mathfrak a)\subseteq N_+$$ and $P : N_+\cap D \to A\cap D$ be the projection map induced by the projection from $N_+$ to its subspace $A$. Then $A$, as a complex Euclidean space, has the induced Euclidean metric from $N_+$. We can consider $A$ as an integral submanifold of the abelian algebra $\mathfrak a$ passing through the base point $o=\Phi(p)$, see page 248 in \cite{CZTG}. For the basic properties of integral manifolds of horizontal distribution, see Chapter 4 of \cite{CZTG}, or \cite{CKT} and \cite{Mayer}. The restricted period map $\tilde{\Phi} : \, \check{\mathcal{T}}\to N_+\cap D$ composed with the projection map $P$ gives a holomorphic map \begin{equation}\label{maptoA} \Psi :\, \check{\mathcal{T}} \to A\cap D, \end{equation} that is $\Psi=P \circ \tilde{\Phi}|_{\check{\mathcal{T}}}$. \begin{lemma}\label{abounded} The image of the holomorphic map $\Psi : \check{\mathcal{T}} \to A\cap D$ is bounded in $A$ with respect to the Euclidean metric on $A\subseteq N_+$. \end{lemma} \begin{proof} We need to show that there exists $0\leq C<\infty$ such that for any $q\in \check{\mathcal{T}}$, $d_{E}({\Psi}({p}), {\Psi}(q))\leq C$, where $d_E$ is the Euclidean distance on $A$. By the definition of $A$, for any $q\in \check{\mathcal{T}}$ there exists a unitary $X^+\in \mathfrak a$ and a real number $T_{0}$ such that $\Psi(q)=\exp(T_{0}X^+)$. Consider $X^-=\tau_0(X^+)\in\tau_0(\mathfrak a)$, then $X=X^++X^-\in \mathfrak a_0\triangleq \mathfrak a+ \tau_0(\mathfrak a)\subseteq \mathfrak{p}_0$. For any $q\in \check{\mathcal{T}}$, there exists $T$ such that $\gamma=\exp(tX): [0, T]\rightarrow G_{\mathbb{R}}$ defines a geodesic from $\gamma(0)=I$ to $\gamma(T)\in G_{\mathbb{R}}$ such that $\gamma(T)\cdot o=\Psi(q)\in A\cap D$, where $\tilde{\Phi}(p)=o$ is the base point. Since $X\in \mathfrak{p}_0$, by Proposition \ref{adjoint}, there exists $k\in K$ such that $ X\in \text{Ad}(k)\cdot\mathfrak{A}_0$. As the adjoint action of $K$ on $\mathfrak{p}_0$ is unitary action and we are considering the length in this proof, we may simply assume that $X\in \mathfrak{A}_0$ up to a unitary transformation. The following proof is an analogue of the proof of the Harish-Chandra embedding theorem for Hermitian symmetric spaces, see page 94 in \cite{Mok}. Let $\Lambda=\{\varphi_1, \cdots, \varphi_r\}\subseteq \Delta^+_{\mathfrak{p}}$ be a set of strongly orthogonal roots given in Lemma \ref{stronglyortho}. We denote $x_{\varphi_i}=e_{\varphi_i}+e_{-\varphi_i}$ and $y_{\varphi_i}=\sqrt{-1}(e_{\varphi_i}-e_{-\varphi_i})$ for any $\varphi_i\in \Lambda$. Then \begin{align*} \mathfrak{A}_0=\mathbb{R} x_{\varphi_{_1}}\oplus\cdots\oplus\mathbb{R}x_{\varphi_{_r}},\quad\text{and}\quad\mathfrak{A}_c=\mathbb{R} y_{\varphi_{_1}}\oplus\cdots\oplus\mathbb{R}y_{\varphi_{_r}}, \end{align*} are maximal abelian spaces in $\mathfrak{p}_0$ and $\sqrt{-1}\mathfrak{p}_0$ respectively. For any $X\in \mathfrak{A}_0$ there exists $\lambda_{i}\in \mathbb{R}$ for $1\leq i\leq r$ such that \begin{align*} X=\lambda_{1}x_{{\varphi_1}}+\lambda_{2}x_{{\varphi_2}}+\cdots+\lambda_{r}x_{{\varphi_r}}. \end{align*} Since $\mathfrak{A}_0$ is commutative, we have \begin{align*} \exp(tX)=\prod_{i=1}^r\exp (t\lambda_{i}x_{{\varphi_i}}). \end{align*} Now for each $\varphi_i\in \Lambda$, we have $\text{Span}_{\mathbb{C}}\{e_{{\varphi_i}}, e_{{-\varphi_i}}, h_{{\varphi_i}}\}\simeq \mathfrak{sl}_2(\mathbb{C}) \text{ with}$ \begin{align*} h_{{\varphi_i}}\mapsto \left[\begin{array}[c]{cc}1&0\\0&-1\end{array}\right], \quad &e_{{\varphi_i}}\mapsto \left[\begin{array}[c]{cc} 0&1\\0&0\end{array}\right], \quad e_{{-\varphi_i}}\mapsto \left[\begin{array}[c]{cc} 0&0\\1&0\end{array}\right]; \end{align*} and $\text{Span}_\mathbb{R}\{x_{{\varphi_i}}, y_{{\varphi_i}}, \sqrt{-1}h_{{\varphi_i}}\}\simeq \mathfrak{sl}_2(\mathbb{R})\text{ with}$ \begin{align*} \sqrt{-1}h_{{\varphi_i}}\mapsto \left[\begin{array}[c]{cc}\sqrt{-1}&0\\0&-\sqrt{-1}\end{array}\right], \quad &x_{{\varphi_i}}\mapsto \left[\begin{array}[c]{cc} 0&1\\1&0\end{array}\right], \quad y_{{\varphi_i}}\mapsto \left[\begin{array}[c]{cc} 0&\sqrt{-1}\\-\sqrt{-1}&0\end{array}\right]. \end{align*} Since $\Lambda=\{\varphi_1, \cdots, \varphi_r\}$ is a set of strongly orthogonal roots, we have that \begin{align*}&\mathfrak{g}_{\mathbb{C}}(\Lambda)=\text{Span}_\mathbb{C}\{e_{{\varphi_i}}, e_{{-\varphi_i}}, h_{{\varphi_i}}\}_{i=1}^r\simeq (\mathfrak{sl}_2(\mathbb{C}))^r,\\\text{and} \quad&\mathfrak{g}_{\mathbb{R}}(\Lambda)=\text{Span}_\mathbb{R}\{x_{{\varphi_i}}, y_{{\varphi_i}}, \sqrt{-1}h_{{\varphi_i}}\}_{i=1}^r\simeq (\mathfrak{sl}_2(\mathbb{R}))^r. \end{align*} In fact, we know that for any $\varphi, \psi\in \Lambda$ with $\varphi\neq\psi$, $[e_{{\pm\varphi}}, \,e_{{\pm\psi}}]=0$ since $\varphi$ is strongly orthogonal to $\psi$; $[h_{\phi},\,h_{{\psi}}]=0$, since $\mathfrak{h}$ is abelian; and $$[h_{{\varphi}},\, e_{{\pm\psi}}]=[[e_{{\varphi}},\,e_{{-\varphi}}],\,e_{{\pm\psi}}]=-[[e_{-\phi},\ e_{\pm\psi}],\ e_\phi]-[[e_{\pm\psi},\ e_\phi],\ e_{-\phi}]=0.$$ Then we denote $G_{\mathbb{C}}(\Lambda)=\exp(\mathfrak{g}_{\mathbb{C}}(\Lambda))\simeq (SL_2(\mathbb{C}))^r$ and $G_{\mathbb{R}}(\Lambda)=\text{exp}(\mathfrak{g}_{\mathbb{R}}(\Lambda))=(SL_2(\mathbb{R}))^r$, which are subgroups of $G_{\mathbb{C}}$ and $G_{\mathbb{R}}$ respectively. With the fixed reference point $o=\tilde{\Phi}(p)$, we denote $D(\Lambda)=G_\mathbb{R}(\Lambda)(o)$ and $S(\Lambda)=G_{\mathbb{C}}(\Lambda)(o)$ to be the corresponding orbits of these two subgroups, respectively. Then we have the following isomorphisms, \begin{align} &D(\Lambda)=G_{\mathbb{R}}(\Lambda)\cdot B/B\simeq G_{\mathbb{R}}(\Lambda)/G_{\mathbb{R}}(\Lambda)\cap V,\label{isomorphism1}\\ &S(\Lambda)\cap (N_+B/B)=(G_{\mathbb{C}}(\Lambda)\cap N_+)\cdot B/B\simeq G_{\mathbb{C}}(\Lambda)\cap N_+. \label{isomorphism2} \end{align} With the above notations, we will show that (i) $D(\Lambda)\subseteq S(\Lambda)\cap (N_+B/B)\subseteq \check{D}$; (ii) $D(\Lambda)$ is bounded inside $S(\Lambda)\cap(N_+B/B)$. By Lemma \ref{puretype}, we know that for each pair of roots $\{e_{{\varphi_i}}, e_{{-\varphi_i}}\}$, there exists a positive integer $k$ such that either $e_{{\varphi_i}}\in \mathfrak{g}^{-k,k}\subseteq \mathfrak{n}_+$ and $e_{{-\varphi_i}}\in \mathfrak{g}^{k,-k}$, or $e_{{\varphi_i}}\in \mathfrak{g}^{k,-k}$ and $e_{{-\varphi_i}}\in \mathfrak{g}^{-k,k}\subseteq\mathfrak{n}_+$. For notation simplicity, for each pair of root vectors $\{e_{{\varphi_i}}, e_{{-\varphi_i}}\}$, we may assume the one in $\mathfrak{g}^{-k,k}\subseteq \mathfrak{n}_+$ to be $e_{{\varphi_i}}$ and denote the one in $\mathfrak{g}^{k,-k}$ by $e_{{-\varphi_i}}$. In this way, one can check that $\{\varphi_1, \cdots, \varphi_r\}$ may not be a set in $\Delta^+_{\mathfrak{p}}$, but it is a set of strongly orthogonal roots in $\Delta_{\mathfrak{p}}$. Therefore, we have the following description of the above groups, \begin{align*} &G_{\mathbb{R}}(\Lambda)=\exp(\mathfrak{g}_{\mathbb{R}}(\Lambda))=\exp(\text{Span}_{\mathbb{R}}\{ x_{{\varphi_1}}, y_{{\varphi_1}},\sqrt{-1}h_{{\varphi_1}}, \cdots, x_{{\varphi_r}}, y_{{\varphi_r}}, \sqrt{-1}h_{{\varphi_r}}\})\\ &G_{\mathbb{R}}(\Lambda)\cap V=\exp(\mathfrak{g}_{\mathbb{R}}(\Lambda)\cap \mathfrak{v}_0)=\exp(\text{Span}_{\mathbb{R}}\{\sqrt{-1}h_{{\varphi_1}}, \cdot,\sqrt{-1} h_{{\varphi_r}}\})\\ &G_{\mathbb{C}}(\Lambda)\cap N_+=\exp(\mathfrak{g}_{\mathbb{C}}(\Lambda)\cap \mathfrak{n}_+)=\exp(\text{Span}_{\mathbb{C}}\{e_{{\varphi_1}}, e_{{\varphi_2}}, \cdots, e_{{\varphi_r}}\}). \end{align*} Thus by the isomorphisms in \eqref{isomorphism1} and \eqref{isomorphism2}, we have \begin{align*} &D(\Lambda)\simeq \prod_{i=1}^r\exp(\text{Span}_{\mathbb{R}}\{x_{{\varphi_i}}, y_{{\varphi_i}}, \sqrt{-1}h_{{\varphi_i}}\})/\exp(\text{Span}_{\mathbb{R}}\{\sqrt{-1}h_{{\varphi_i}}\},\\ &S(\Lambda)\cap (N_+B/B)\simeq\prod_{i=1}^r\exp(\text{Span}_{\mathbb{C}}\{e_{{\varphi_i}}\}). \end{align*} Let us denote $G_{\mathbb{C}}(\varphi_i)=\exp(\text{Span}_{\mathbb{C}}\{e_{{\varphi_i}}, e_{{-\varphi_i}}, h_{{\varphi_i}})\simeq SL_2(\mathbb{C}),$ $S(\varphi_i)=G_{\mathbb{C}}(\varphi_i)(o)$, and $G_{\mathbb{R}}(\varphi_i)=\exp(\text{Span}_{\mathbb{R}}\{x_{{\varphi_i}}, y_{{\varphi_i}}, \sqrt{-1}h_{{\varphi_i}}\})\simeq SL_2(\mathbb{R}),$ $D(\varphi_i)=G_{\mathbb{R}}(\varphi_i)(o)$. Now each point in $S(\varphi_i)\cap (N_+B/B)$ can be represented by \begin{align*} \text{exp}(ze_{{\varphi_i}})=\left[\begin{array}[c]{cc}1&z\\ 0& 1\end{array}\right] \quad \text{for some } z\in \mathbb{C}. \end{align*} Thus $S(\varphi_i)\cap (N_+B/B)\simeq \mathbb{C}$. In order to see $D(\varphi_i)$ in $G_\mathbb{C}/B$, we decompose each point in $D(\varphi_i)$ as follows. Let $z=a+bi$ for some $a,b\in\mathbb{R}$, then \begin{eqnarray*} \exp(ax_{{\varphi_i}}+by_{{\varphi_i}})&=&\left[\begin{array}[c]{cc}\cosh |z|&\frac{z}{|z|}\sinh |z|\\ \frac{\bar{z}}{|z|}\sinh |z|& \cosh |z|\end{array}\right]\\ &=&\left[\begin{array}[c]{cc}1&\frac{z}{|z|}\tanh |z|\\ 0&1\end{array}\right]\left[\begin{array}[c]{cc}(\cosh |z|)^{-1}&0\\0&\cosh |z|\end{array}\right]\left[\begin{array}[c]{cc}1& 0\\ \frac{\bar{z}}{|z|}\tanh |z|&1\end{array}\right]\\ &=&\exp\left[(\frac{z}{|z|}\tanh |z| )e_{{\varphi_i}}\right]\exp\left[-\log (\cosh |z|)h_{{\varphi_i}}\right]\exp\left[(\frac{\bar{z}}{|z|}\tanh |z|) e_{{-\varphi_i}}\right]\\ &\equiv&\exp\left[(\frac{z}{|z|}\tanh |z| )e_{{\varphi_i}}\right] \ (\text{mod }B). \end{eqnarray*} So elements of $D(\varphi_i)$ in $G_\mathbb{C}/B$ can be represented by $\exp[(z/|z|)(\tanh |z|) e_{{\varphi_i}}]$, i.e. \begin{align*} \left[\begin{array}[c]{cc}1& \frac{z}{|z|}\tanh |z|\\0&1\end{array}\right], \end{align*} in which $\frac{z}{|z|}\tanh |z|$ is a point in the unit disc $\mathfrak{D}$ of the complex plane. Therefore the $D(\varphi_i)$ is a unit disc $\mathfrak{D}$ in the complex plane $S(\varphi_i)\cap (N_+B/B)$. Therefore $$D(\Lambda)\simeq \mathfrak{D}^r\quad\text{and}\quad S(\Lambda)\cap N_+\simeq \mathbb{C}^r. $$ So we have obtained both (i) and (ii). As a consequence, we get that for any $q\in \check{\mathcal{T}}$, $\Psi(q)\in D(\Lambda)$. This implies \begin{align*}d_E(\Psi(p), \Psi(q))\leq \sqrt{r} \end{align*} where $d_E$ is the Eulidean distance on $S(\Lambda)\cap (N_+B/B)$. To complete the proof, we only need to show that $S(\Lambda)\cap (N_+B/B)$ is totally geodesic in $N_+B/B$. In fact, the tangent space of $N_+$ at the base point is $\mathfrak{n}_+$ and the tangent space of $S(\Lambda)\cap N_+B/B$ at the base point is $\text{Span}_{\mathbb{C}}\{e_{{\varphi_1}}, e_{{\varphi_2}}, \cdots, e_{{\varphi_r}}\}$. Since $\text{Span}_{\mathbb{C}}\{e_{{\varphi_1}}, e_{{\varphi_2}}, \cdots, e_{{\varphi_r}}\}$ is a sub-Lie algebra of $\mathfrak{n}_+$, the corresponding orbit $S(\Lambda)\cap N_+B/B$ is totally geodesic in $N_+B/B$. Here the basis $\{e_{{\varphi_1}}, e_{{\varphi_2}}, \cdots, e_{{\varphi_r}}\}$ is an orthonormal basis with respect to the pull-back Euclidean metric. \end{proof} As proved in Proposition 3 of Chapter 2 in \cite{Schwartz}, for the period map $\tilde{\Phi}:\, \mathcal{T} \to D$, there is a stratification $\mathcal{T}=\cup_{\nu,k}\mathcal{T}^k_\nu$ such that if $\tilde{\Phi}^k_\nu=\tilde{\Phi}|_{\mathcal{T}^k_\nu}$, the rank of $d\tilde{\Phi}^k_\nu$ is constant on $\mathcal{T}^k_\nu$. We will consider the image of a local neighborhood of $\mathcal{T}$ under the period map $\tilde{\Phi}$ as a horizontal slice, which is given by the union of the image of each $\mathcal{T}^k_\nu$. Note that for the extended period map $\tilde{\Phi}' :\, \mathcal{T}'\to D$, we can define horizontal slices in the same way, since the extended period map $\tilde{\Phi}'$ still satisfies the Griffiths transversality. Another point of view about horizontal slice is from complex analytic foliation. By using the natural Whitney stratification of the singular set of a complex analytic foliation, the following lemma is proved in \cite{MY} as Corollary 2.10. \begin{lemma}\label{singular foliation} Let $E$ be a complex analytic singular foliation on a complex manifold $M$. Then there exists a family $\mathcal{L}$ of complex sub-manifolds of $M$ such that $M=\cup_{L\in \mathcal{L}}L$ is disjoint union and that, for any $L\in \mathcal{L}$ and $p\in L$, we have $E(p)=\text{T}_pL$. \end{lemma} Here $E$ is considered as a coherent subsheaf of the tangent sheaf of $M$, and $E(p)$ denotes the tangent vectors in $E$ at $p\in M$. We refer the reader to \cite{MY} for the details about singular complex analytic foliations. Note that in general the foliation $E$ on the period domain $D$ defined by the tangent map of the period map $\tilde{\Phi}$ is a singular complex analytic foliation, since in this case we have $[E,E]=0$. With the notations above, each $L\in \mathcal{L}$ is identified with $\tilde{\Phi}(\mathcal{T}^k_\nu)$ for some $k$ and $\nu$, and similarly for the images of $\tilde{\Phi}'$. By Lemma \ref{singular foliation}, a neighborhood of any point in the image of the period map is a union of integral submanifolds of a singular foliation. We will also call these integral submanifolds the horizontal slices. One sees that the two descriptions of horizontal slices are equivalent. The projection map $\pi:\,D\to G_\mathbb{R}/K$ is a one-to-one map when restricted to a horizontal slice by the Griffiths tranversality. We will need to use this fact in the following discussion. Now let $\mathfrak{p}_+=\mathfrak{p}/\mathfrak{p}\cap \mathfrak{b}\subseteq \mathfrak{n}_+$ denote the subspace of the holomorphic tangent space of $D$ at the base point $o$, which is viewed as an Euclidean subspace of $\mathfrak{n}_+$, and consider the natural projection $P_+:\, N_+\cap D\to \mathrm{exp}(\mathfrak{p}_+) \cap D$ by viewing $ \mathrm{exp}(\mathfrak{p}_+)$ as an Euclidean subspace of $N_+$. It is easy to see that the natural projection $\pi:\,D\to G_\mathbb{R}/K$, restricted to $\mathrm{exp}(\mathfrak{p}_+) \cap D$ is diffeomorphic to $G_\mathbb{R}/K$. For simplicity we will assume that $A\cap D$ is contained in $\mathrm{exp}(\mathfrak{p}_+) \cap D$. Then there exists a neighborhood $U$ of the base point $o$ in $D$ such that the image of $U\cap \tilde{\Phi}(\check{\mathcal{T}}) $ under the projection $P_+$ lies inside $A\cap D$, since they are both integral submanifolds of the horizontal distribution through the base point given by the abelian algebra $\mathfrak{a}$. From this we see that $P_+( \tilde{\Phi}(\check{\mathcal{T}}))$ must lie inside $A\cap D$ completely, since $\tilde{\Phi}(\check{\mathcal{T}})$ is a connected irreducible complex analytic variety, which is induced from the irreducibility of $\mathcal{T}$. Note that, with the above notations, $P:\, N_+\cap D\to A\cap D$ may be considered as the composite of $P_+$ with the projection from $\mathrm{exp}(\mathrm{p}_+)\cap D$ to its subspace $A\cap D$. \begin{theorem}\label{locallybounded} The image of the restriction of the period map $\tilde{\Phi} :\, \check{\mathcal{T}} \to N_+\cap D$ is bounded in $N_+$ with respect to the Euclidean metric on $N_+$. \end{theorem} \begin{proof} In Lemma \ref{abounded}, we already proved that the image of $\Psi=P\circ \tilde{\Phi}$ is bounded with respect to the Euclidean metric on $A\subseteq N_+$. Now together with the Griffiths transversality, we will deduce the boundedness of the image of $\tilde{\Phi} :\, \check{\mathcal{T}} \to N_+\cap D$ from the boundedness of the image of $\Psi$. Since the projection $\pi:\,D\to G_\mathbb{R}/K$ restricted to $\mathrm{exp}(\mathfrak{p}_+) \cap D$ is a diffeomorphism, any two points in $N_+\cap D$ are mapped to the same point in $\mathrm{exp}(\mathfrak{p}_+) \cap D$ via $P_+$, if and only if they are are mapped to the same point in $G_\mathbb{R}/K$ via $\pi$. Hence $P_+$ maps the fiber $\pi^{-1}(z')\cap N_+$ for $z' \in G_\mathbb{R}/K$ to a point $z\in \exp(\mathfrak{p}_+) \cap D$ with $\pi(z) =z'$. From this and the discussion above Theorem \ref{locallybounded}, we can see that for any $z\in \Psi(\check{\mathcal{T}})$, the restricted map $P|_{\tilde{\Phi}(\check{\mathcal{T}})}$ is just the projection map from the intersection $\tilde{\Phi}(\check{\mathcal{T}})\cap \pi^{-1}(z')$ to $z$, where $z'=\pi(z)\in G_\mathbb{R}/K$. With the above in mind, our proof can be divided into two steps. It is an elementary argument to apply the Griffiths transversality on $\mathcal{T}'$. (i) We claim that there are only finite points in $(P|_{\tilde{\Phi}(\check{\mathcal{T}})})^{-1}(z)$, for any $z\in \Psi(\check{\mathcal{T}})$. Otherwise, we have $\{q_i\}_{i=1}^{\infty}\subseteq \check{\mathcal{T}}$ and $\{y_i=\tilde{\Phi}(q_i)\}_{i=1}^{\infty}\subseteq (P|_{\tilde{\Phi}(\check{\mathcal{T}})})^{-1}(z)$ with limiting point $y_{\infty}\in \pi^{-1}(z')\simeq K/V$, since $K/V$ is compact. We project the points $q_i$ to $q'_i\in S$ via the universal covering map $p:\, \mathcal{T} \to S$. There must be infinite many $q'_i$'s. Otherwise, we have a subsequence $\{q_{j_k}\}$ of $\{q_j\}$ such that $p(q_{j_k})=q'_{i_0}$ for some $i_0$ and $$y_{j_k}=\tilde{\Phi}(q_{j_k})=\gamma_{k}\tilde{\Phi}(q_{j_0})=\gamma_{k}y_{j_0},$$ where $\gamma_k \in \Gamma$ is the monodromy action. Since $\Gamma$ is discrete, the subsequence $\{y_{j_k}\}$ is not convergent, which is a contradiction. Now we project the points $q_i$ on $S$ via the universal covering map $p:\, \mathcal{T} \to S$ and still denote them by $q_i$ without confusion. Then the sequence $\{q_i\}_{i=1}^{\infty}\subseteq S$ has a limiting point $q_\infty$ in $\bar{S}$, where $\bar{S}$ is the compactification of $S$ as defined in the introduction section. By continuity the period map $\Phi :\, S \to D/\Gamma$ can be extended over $q_\infty$ with $\Phi(q_\infty)=\pi_D(y_\infty) \in D/\Gamma$, where $\pi_D :\, D\to D/\Gamma$ is the projection map. Thus $q_\infty$ lies $S'$. Now we can regard the sequence $\{q_i\}_{i=1}^{\infty}$ as a convergent sequence in $S'$ with limiting point $q_\infty \in S'$. We can also choose a sequence $\{\tilde{q}_i\}_{i=1}^{\infty}\subseteq \mathcal{T}'$ with limiting point $\tilde{q}_\infty\in \mathcal{T}'$ such that $\tilde{q}_i$ maps to $q_i$ via the universal covering map $\pi' :\, \mathcal{T}'\to S'$ and $\tilde{\Phi}'(\tilde{q}_i)=y_i \in D$, for $i\ge 1$ and $i=\infty$. Since the extended period map $\tilde{\Phi}' :\, \mathcal{T}' \to D$ still satisfies the Griffiths transversality by Lemma \ref{extendedtransversality}, we can choose a small neighborhood $U$ of $\tilde{q}_\infty$ such that $U$, and thus the points $\tilde{q}_i$ for $i$ sufficiently large are mapped into a horizontal slice, which is a contradiction. Denote $P|_{\tilde{\Phi}'(\mathcal{T}')}$ to be the restricted map $P|_{\tilde{\Phi}'(\mathcal{T}')}:\, N_{+}\cap \tilde{\Phi}'(\mathcal{T}') \to A\cap D$. In fact, a similar argument also implies that there are finite points in $(P|_{\tilde{\Phi}'(\mathcal{T}')})^{-1}(z)$, for any $z\in \Psi(\check{\mathcal{T}})$. Furthermore, we have the following conclusions. (ii) Let $r(z)$ be the cardinality of the fiber $(P|_{\tilde{\Phi}'(\mathcal{T}')})^{-1}(z)$, for any $z\in \Psi(\check{\mathcal{T}})$. We claim that $r(z)$ is locally constant. To be precise, for any $z\in \Psi(\check{\mathcal{T}})$, let $r=r(z)$ and choose points $x_1,\cdots,x_r$ in $N_+\cap D$ such that $P|_{\tilde{\Phi}'(\mathcal{T}')}(x_i)=z$. We can find a horizontal slice $U_i$ around $x_i$ such that \begin{equation}\label{locallyinj} P|_{\tilde{\Phi}'(\mathcal{T}')} :\, U_i\cap \tilde{\Phi}'(\mathcal{T}')\to P|_{\tilde{\Phi}'(\mathcal{T}')}(U_i) \end{equation} is injective, $i=1,\cdots,r$. We choose the balls $B_i$, $i=1,\cdots,r$ small enough in $N_+\cap D$ such that $B_i$'s are mutually disjoint and $B_i\supseteq U_i$. We claim that there exists a small neighborhood $V\ni z$ in $A\cap D$ such that the restricted map \begin{equation}\label{localconstant} P|_{\tilde{\Phi}'(\mathcal{T}')} :\, (P|_{\tilde{\Phi}'(\mathcal{T}')})^{-1}(V)\cap B_i \to V \text{ is injective, for } i=1,\cdots,r. \end{equation} We define the pair of sequences $(z,y)$ for some $i$ as follows \begin{itemize} \item [($*_i$)] $\{z_k\}_{k=1}^\infty$ is a convergent sequence in $A\cap D$ with limiting point $z$. $\{y_k\}_{k=1}^\infty$ is a convergent sequence in $N_+\cap D$ with limiting point $x_i$ such that $P(y_k)=z_k$ and $y_k \in B_i\setminus U_i$, for any $k\ge 1$. \end{itemize} We will prove that the pair of sequences $(z,y)$ as ($*_i$) does not exist for any $i=1,\cdots,r$, which implies that we can find a small neighborhood $V\ni z$ in $A\cap D$ such that $$(P|_{\tilde{\Phi}'(\mathcal{T}')})^{-1}(V)\cap B_i\subseteq U_i.$$ Hence (\ref{localconstant}) holds and $r(z)$ is locally constant. In fact, if for some $i$ there exists a pair of sequences $(z,y)$ as ($*_i$), we can find a sequence $\{q_k\}_{k=1}^\infty$ in $\mathcal{T}'$ with limiting point $q_\infty \in \mathcal{T}'$ by a similar argument as (i) such that $\tilde{\Phi}'(q_k)=y_k$ for any $k\ge 1$ and $\tilde{\Phi}'(q_\infty)=x_i$. Since $\tilde{\Phi}' : \, \mathcal{T}' \to D$ still satisfies the Griffiths transversality due to Lemma \ref{extendedtransversality}, we can choose a small neighborhood $W\ni q_\infty$ in $\mathcal{T}'$ such that $\tilde{\Phi}'(W)\subseteq U_i$. Then for $k$ sufficiently large, $y_k=\tilde{\Phi}'(q_k)\in U_i$, which is a contradiction to ($*_i$). Therefore from (i) and (ii) we deduce that the image $\tilde{\Phi}(\check{\mathcal{T}})\subseteq N_+\cap D$ is also bounded. \end{proof} \begin{remark} The proof of Theorem \ref{locallybounded} makes substantial use of the basic properties of the variation of Hodge structure arising from geometry, such as the integrability of the horizontal distribution and the local structure of the horizontal slices, as well as the quasi-projectivity of the family of polarized manifolds and the existence of the arithmetic discrete subgroup. Theorem \ref{locallybounded} may not be valid for general horizontal holomorphic maps as pointed out by Professors Vincent Koziarz and Julien Maubon. \end{remark} \section{Proof of the Griffiths conjecture} In this section, we prove the Griffiths conjecture by using the boundedness of the restricted period map $\tilde{\Phi}|_{\check{\mathcal{T}}} :\, \check{\mathcal{T}}\to N_+\cap D$ and the Riemann extension theorem. This is our main result. \\ \begin{theorem}\label{main}{\em(Main Theorem)} The image of $\tilde{\Phi} : \mathcal{T}\rightarrow D$ lies in $N_+\cap D$ and is bounded with respect to the Euclidean metric on $N_+$. \end{theorem} \begin{proof} According to Proposition \ref{codimension}, $\mathcal{T} \setminus \check{\mathcal{T}}$ is an analytic subvariety of $\mathcal{T}$ \and the complex codimension of $\mathcal{T} \setminus \check{\mathcal{T}}$ is at least one; by Theorem \ref{locallybounded}, the holomorphic map $\tilde{\Phi} : \check{\mathcal{T}}\rightarrow N_+\cap D$ is bounded in $N_+$ with respect to the Euclidean metric. Thus by the Riemann extension theorem, there exists a holomorphic map $\tilde{\Phi}_\mathcal{T} : \mathcal{T} \to N_+$, such that $\tilde{\Phi}_\mathcal{T}|_{\check{\mathcal{T}}}=\tilde{\Phi}|_{\check{\mathcal{T}}}$. Since as holomorphic maps, $\tilde{\Phi}_\mathcal{T}$ and $\tilde{\Phi}$ agree on the open subset $\check{\mathcal{T}}$, they must be the same on the entire $\mathcal{T}$. Therefore, the image of $\tilde{\Phi}$ is in $N_+\cap D$, and the image is bounded with respect to the Euclidean metric on $N_+.$ As a consequence, we also get $\mathcal{T}=\check{\mathcal{T}}=\tilde{\Phi}^{-1}(N_+\cap D).$ \end{proof} From the proof of Theorem \ref{locallybounded} and Theorem \ref{main}, we also have the following corollary, which improves the boundedness on $\mathcal{T}$ to its completion space $\mathcal{T}'$. \begin{corollary} The image of the extended period map $\tilde{\Phi}' : \mathcal{T}' \rightarrow D$ also lies in $N_+\cap D$ and is bounded with respect to the Euclidean metric on $N_+$. \end{corollary}

https://arxiv.org/abs/1808.01905

Nuisance Parameters Free Changepoint Detection in Non-stationary Series

Detecting abrupt changes in the mean of a time series, so-called changepoints, is important for many applications. However, many procedures rely on the estimation of nuisance parameters (like long-run variance). Under the alternative (a change in mean), estimators might be biased and data-adaptive rules for the choice of tuning parameters might not work as expected. If the data is not stationary, but heteroscedastic, this becomes more challenging. The aim of this paper is to present and investigate two changepoint tests, which involve neither nuisance nor tuning parameters. This is achieved by combing self-normalization and wild bootstrap. We study the asymptotic behavior and show the consistency of the bootstrap under the hypothesis as well as under the alternative, assuming mild conditions on the weak dependence of the time series and allowing the variance to change over time. As a by-product of the proposed tests, a changepoint estimator is introduced and its consistency is proved. The results are illustrated through a simulation study, which demonstrates computational efficiency of the developed methods. The new tests will also be applied to real data examples from finance and hydrology.

\section{Main goals} In the statistical analysis, it is of particular interest to be able to detect systematic changes---so-called \emph{changepoints}---in the underlying structure despite the random fluctuations and to estimate the time of these changes. Under the assumption of finite expectations, changes in the location are typically detected by comparing sample means and the asymptotic distribution can be derived from an~invariance principle for the partial sum process. However, one has to estimate the \emph{long-run variance} to utilize the traditional CUSUM-statistic and this involves some difficulties. For time series, the long-run variance includes the covariances, which have to be estimated and combined with, e.g., kernels. Under the alternative, the estimation of the covariances is biased, so the long-run covariance is typically overestimated, which results in a~loss of power, see~\cite{huvskova2010note}. In many applications, the observations do not seem to be stationary even under the hypothesis of no change in mean, because the amount of fluctuation is not constant. To estimate the time-varying long-run variance is even more difficult. In a~recent article, \cite{gorecki2017change} have followed this approach. The main aim of this paper is to develop tests for the hypothesis of a~constant expectation against the alternative of at most one changepoint that avoid the problems of long-run variance estimation and heteroscedasticity. Our new test statistics will not involve \emph{any nuisance parameters} and will work for \emph{heteroscedastic and dependent} time series under some mild mixing conditions. Additionally, we will give a~consistent estimator for the time of the change. Some authors proposed to use nonparametric resampling methods like bootstrap (e.g., \cite{HK2012} and~\cite{PP2018}) or subsampling (e.g., \cite{betken2017subsampling}) to avoid the estimation of the long-run variance. However, these methods still involve the choice of tuning parameters like bandwidths or block sizes and only work for stationary time series. Other approaches are \emph{ratio statistics} and \emph{self-normalized statistics}, which do not rely on tuning parameters. Ratio tests have been introduced to detect changes in persistence by \cite{Kim2000} and since have been studied for changes in mean~\citep{HHH2008}, for changes in variance~\citep{zhao2011ratio}, for heavy-tailed sequences~\citep{dan2017detection}, for panel date framework~\citep{PP2015}, and for robust $M$-estimators~\citep{PP2018}. Self-normalized test statistics for changepoints were firstly proposed by \cite{shao2010testing} and were generalized to long range dependent time series~\citep{Shao2011}. \cite{betken2016} developed a~robust self-normalized test based on the Wil\-coxon-statistic, \cite{zhang2018unsupervised} proposed a~self-normalized test for multiple changepoints. Our approach is to combine new variants of the self-normalized test statistics with the \emph{wild bootstrap}. The wild bootstrap was proposed by \cite{wu1986jackknife} and is consistent under heteroscedasticity. However, it does not reproduce the dependence of the data. We will show that under our model assumptions, it still gives the correct critical values for the self-normalized test statistics. In this way, we can avoid using the dependent wild bootstrap of \cite{shao2010dependent}, which involves the choice of a~kernel and of a~bandwidth parameter. The paper is organized as follows: In the next section, we will introduce our data model and our new test statistics. In Section~\ref{sec:mainresult}, the technical assumptions are discussed, the bootstrap is introduced and our main theoretical results are presented. We provide a~table with critical values of our test statistics under stationarity in Section~\ref{sec:simu} and investigate the finite sample properties through simulation results. Two data examples from finance and hydrology are provided in Section~\ref{sec:data}. Afterwards, our conclusion follows. The proofs of our theoretical results can be found in the appendix. \section{Stochastic model and methods}\label{sec:model} \subsection{Changepoint model} We tend to study time series with one abrupt change in the mean at an unknown point in time. Let us consider observations $Y_{1,n},\ldots,Y_{n,n}$ obtained at time ordered points. We are interested in testing the null hypothesis of all observations being random variables having equal expectation. Our goal is to test against the alternative of the first $\tau_n$ observations have expectation $\mu$ and the remaining $n-\tau_n$ observations come from distributions with expectation $\mu + \d_n$, where $\d_n\neq 0$. More precisely, our model is \begin{equation}\label{eq:locmodel Y_{n,k}=\mu+\d_n\ind\{k>\tau_n\} + \sigma\left(\frac{k}{n}\right)\varepsilon_k,\quad k=1,\ldots,n, \end{equation} where $\mu$, $\d_n$, and $\tau_n$ are unknown parameters, $\{Y_{n,k}\}_{n=1,k=1}^{\infty,n}$ is a~\emph{triangular array} of random variables, $\{\varepsilon_{n}\}_{n=1}^{\infty}$ is a~sequence of stationary centered disturbances, $\sigma(t)$ is a~non-stochastic variance function, and $\ind\{A\}$ denotes the indicator of set~$A$. The time point~$\tau_n$ is called the \emph{changepoint}. A~similar model was assumed by \cite{gorecki2017change}. We are going to test the null hypothesis that no change occurred against the alternative that a~change occurred at some unknown time point $\tau_n$, i.e., \begin{equation*} \mathcal{H}_0: \tau_n=n\quad\mbox{versus}\quad\mathcal{H}_1:\tau_n<n,\ \d_n \ne 0. \end{equation*} \subsection{Test statistics} The \emph{CUSUM-statistic} is frequently used to detect changes in the mean and it is based on the partial sums $\sum_{i=1}^{k}\left(Y_{n,i}-\bar{Y}_{n,1:n}\right)$, $k=1,\ldots,n-1$ of the centered observations, where \[ \bar{Y}_{n,i:j}=\frac{1}{j-i+1}\sum_{k=i}^{j}Y_{n,k},\quad i\leq j. \] To combine the values of the partial sums for different~$k$ into a~single test statistic, one can use the \emph{supremum-type} CUSUM-statistic $\max_{k=1,\ldots,n-1}\left|\sum_{i=1}^{k}\big(Y_{n,i}-\bar{Y}_{n,1:n}\big)\right|$ or the \emph{integral-type} CUSUM-statistic $\sum_{k=1}^{n-1}\left(\sum_{i=1}^{k}\left(Y_{n,i}-\bar{Y}_{n,1:n}\right)\right)^2$. These test statistics need to be standardized by a~variance of the series. However, it is practically \emph{difficult to find a~variance estimator} with satisfactory properties. Such difficulty can occur in situations with dependent or heteroscedastic random errors. Nonetheless, the variance estimators often do not perform well even in the i.i.d.~case, especially under alternatives~\citep{Antoch1997}. To avoid the estimation of variance parameters, different ratios of such test statistics have been proposed. \cite{HHH2008} divide the supremum-test statistic of the first part of the series by the supremum-test statistic of the second part of the data. \cite{wenhua2016ratio} use a~ratio test based on the integral-type statistic. \cite{shao2010testing} introduced a self-normalized statistic, which uses the supremum-type CUSUM-statistic of the whole data set in the numerator, divided by the sum of two integral-type statistics of the data before~$k$ and after~$k$. Our idea is to use a~self-normalization of the CUSUM-statistic by the same type: We divide the supremum-type statistic by two supremum-type statistics and the integral-type statistic by two integral-type statistics. Our test statistics can be expressed as functionals of the \emph{cumulative sums} \[ V_n(k):=\sum_{i=1}^{k}Y_{n,i}\quad\mbox{and}\quad \widetilde{V}_n(k):=V_n(n)-V_n(k). \] We define the \emph{self-normalized test statistics} as \begin{align}\label{eq:statistic1} \mathscr{Q}(V_n)&:=\max_{1\leq k \leq n}\Bigg|\frac{V_n(k)-k/n V_n(n)}{\max\limits_{1\leq i \leq k}\big|V_n(i)-i/kV_n(k)\big|+\max\limits_{k< i \leq n}\big|\widetilde{V}_n(i)-(n-i)/(n-k)\widetilde{V}_n(k)\big|}\Bigg|\\ &\equiv \max_{1\leq k \leq n}\frac{\left|\sum_{i=1}^{k}\left(Y_{n,i}-\bar{Y}_{n,1:n}\right)\right|}{\max\limits_{1\leq i \leq k}\left|\sum_{j=1}^{i}\left(Y_{n,j}-\bar{Y}_{n,1:k}\right)\right|+\max\limits_{k< i \leq n}\left|\sum_{j=i}^{n}\left(Y_{n,j}-\bar{Y}_{n,(k+1):n}\right)\right|}\notag \end{align} and \begin{align}\label{eq:statistic2} \mathscr{R}(V_n)&:=\sum_{k=1}^n\frac{\left\{V_n(k)-k/n V_n(n)\right\}^2}{\sum_{i=1}^k\big\{V_n(i)-i/kV_n(k)\big\}^2+\sum_{i=k+1}^n\big\{\widetilde{V}_n(i)-(n-i)/(n-k)\widetilde{V}_n(k)\big\}^2}\\ &\equiv \sum_{k=1}^n\frac{\left\{\sum_{i=1}^{k}\left(Y_{n,i}-\bar{Y}_{n,1:n}\right)\right\}^2}{\sum_{i=1}^k\left\{\sum_{j=1}^{i}\left(Y_{n,j}-\bar{Y}_{n,1:k}\right)\right\}^2+\sum_{i=k+1}^n\left\{\sum_{j=i}^{n}\left(Y_{n,j}-\bar{Y}_{n,(k+1):n}\right)\right\}^2}.\notag \end{align} For many changepoint tests, one has to skip, for instance, the first and the last $10\%$ of observations as possible candidates for a~changepoint, see, e.g., \cite{shao2010testing}. Moreover, the amount of trimming can be viewed as an~additional tuning parameter. For our test statistics, we are able to consider all time points $k=1,\ldots,n$. The limit distribution of our statistics is obtained with the help of the continuous mapping theorem, using limit theorems for the partial sum process under weak dependence and heteroscedasticity by~\cite{Cav2005}. \section{Main results}\label{sec:mainresult} \subsection{Assumptions} Prior to deriving asymptotic properties of the test statistic, we summarize the notion of \emph{strong mixing} ($\alpha$-mixing) dependence in more detail, which will be imposed on the model's errors. Suppose that $\{\e_n\}_{n=1}^{\infty}$ is a~sequence of random elements on a~probability space $(\Omega,\mathcal{F},\P)$. For sub-$\sigma$-fields $\mathcal{A},\mathcal{B}\subseteq\mathcal{F}$, we define $\alpha(\mathcal{A},\mathcal{B}):=\sup_{A\in\mathcal{A},B\in\mathcal{B}}\left|\P(A\cap B)-\P(A)\P(B)\right|$. Intuitively, $\alpha(\cdot,\cdot)$ measures the dependence of the events in $\mathcal{B}$ on those in $\mathcal{A}$. There are many ways in which one can to describe weak dependence or, in other words, \emph{asymptotic independence} of random variables, see~\cite{Bradley2005}. Considering a~filtration $\mathcal{F}_m^n:=\sigma\{\e_i\in\mathcal{F},m\leq i\leq n\}$, sequence $\{\e_n\}_{n=1}^{\infty}$ of random variables is said to be \emph{strong mixing} ($\alpha$-mixing) if $\alpha(n):=\sup_{k\in\mathbbm{N}}\alpha(\mathcal{F}_{1}^k,\mathcal{F}_{k+n}^{\infty})\to 0$ as $n\to\infty$. We proceed to the assumptions that are needed for deriving asymptotic properties of the proposed test statistics. For the functional central limit theorem, we need assumptions controlling the dependence and the moments of the underlying errors. \begin{assumpE}\label{assump:UP1} $\left\{\varepsilon_{n}\right\}_{n=1}^{\infty}$ form a~zero-mean strictly stationary $\alpha$-mixing sequence such that $\var\varepsilon_{n}=1$, $\E|\varepsilon_{n}|^p<\infty$ for some $p>2$ with mixing coefficients $\alpha(n)$ satisfying $\sum_{n=1}^{\infty}\{\alpha(n)\}^{2(1/r-1/p)}<\infty$ for some $r\in(2,4]$, $r\leq p$, and, additionally, $\sum_{n=1}^{\infty}n\alpha(n)<\infty$. Furthermore, for the long-run variance, it holds that $0<\lambda:=1+2\sum_{n=1}^{\infty}\E\varepsilon_1\varepsilon_{n+1}<\infty$. \end{assumpE} The mixing properties will be inherited by $Y_{n,k}=\mu+\delta_n\ind\{j>\tau_n\}+\sigma(k/n)\epsilon_k$, but this process can additionally model heteroscedasticity, which is important for many applications. To control variability of the series, we have an~assumption regarding heteroscedasticity. \begin{assumpV}\label{assump:UP2} $\sigma:\,[0,1]\to\mathbb{R}^+$ has finite number of points of discontinuity satisfying a~first-order Lipschitz condition except at points of discontinuity. \end{assumpV} Our tests will be consistent not only for fixed alternatives with $\delta_n\equiv\delta\neq 0$, but also under local alternative, when the size of the change converges to 0. However, it will only consistently detect changes that are not too small compared to the variance of the partial sum process. \begin{assumpC}\label{assump:UP3} $|\delta_n|\sqrt{n}\to\infty$ as $n\to\infty$. \end{assumpC} Henceforth, $\xrightarrow{\mathbbm{P}}$ denotes convergence in probability, $\xrightarrow{\dist}$ convergence in distribution, $\xrightarrow[n\to\infty]{\mathsf{D}[0,1]}$ weak convergence in the Skorokhod space $\text{D}[0,1]$ of c\`adl\`ag functions on $[0,1]$, and $[x]$ denotes the integer part of the real number~$x$. \subsection{Asymptotic distribution of the test statistics} Under the null hypothesis and the technical assumptions from the previous subsection, the test statistics defined in~\eqref{eq:statistic1} and~\eqref{eq:statistic2} converge to \emph{non-degenerate limit distributions} (their quantiles can be found in Subsection~\ref{subsec:crit}). \begin{theorem}[Under the null]\label{theorem:undernull} Under Assumptions~\ref{assump:UP1}, \ref{assump:UP2}, and under the null hypothesis $\mathcal{H}_0$, \begin{equation}\label{eq:limit_dist} \mathscr{Q}(V_n)\xrightarrow{\dist}\mathscr{S}(W_{\eta})\quad\mbox{and}\quad\mathscr{R}(V_n)\xrightarrow{\dist}\mathscr{T}(W_{\eta}),\quad n\to\infty, \end{equation} where~$W_{\eta}(t):=W(\eta(t))$, $\{W(t),0\leq t\leq 1\}$ is a~standard Wiener process, $\eta(t):=\frac{\int_{0}^t\sigma^2(s)\mbox{d} s}{\int_{0}^1\sigma^2(s)\mbox{d} s}$, and the functionals~$\mathscr{S}$ and~$\mathscr{T}$ are defined in~\eqref{functionalS} and~\eqref{functionalT}. \end{theorem} The null hypothesis is rejected at significance level $\alpha$ for large values of $\mathscr{Q}(V_n)$ and $\mathscr{R}(V_n)$. The critical values can be obtained as the $(1-\alpha)$-quantiles of the asymptotic distributions from~\eqref{eq:limit_dist}, if $\eta$ is known. Furthermore, the tests based on these two statistics are consistent, as the test statistics converge to infinity under the alternative, provided that the size of the change does not convergence to~$0$ to fast (see Assumption~\ref{assump:UP3}). \begin{theorem}[Under the alternative]\label{theorem:underalt} Suppose Assumptions~\ref{assump:UP1}, \ref{assump:UP2}, and~\ref{assump:UP3} hold. Under the alternative hypothesis $\mathcal{H}_1$ such that $\tau_n=[n\zeta]$ for some $\zeta\in(0,1)$, \[ \mathscr{Q}(V_n)\xrightarrow{\mathbbm{P}}\infty\quad\mbox{ and }\quad\mathscr{S}(V_n)\xrightarrow{\mathbbm{P}}\infty,\quad n\to\infty. \] \end{theorem} Theorem~\ref{theorem:underalt} says that in presence of the structural change in mean, the test statistics \emph{explode above all bounds}. Hence, the procedures are \emph{consistent} and the asymptotic distributions from Theorem~\ref{theorem:undernull} can be used to construct the tests. Although, explicit forms of those distributions are unknown. Therefore in order to obtain the critical values, we have to use either simulations from the limit distributions or resampling methods. For the simulation purposes, one would need to know or to estimate the nuisance function $\eta(t)$. The resampling techniques will help us to avoid and overcome such an~issue. \subsection{Wild bootstrap} Wild bootstrap \emph{replications} are defined as \[ Y_{n,k}^{\star}:=\left(Y_{n,k}-\bar{Y}_{n,1:n}\right)X_{k},\quad k=1,\ldots,n, \] where $\{X_{n}\}_{n=1}^{\infty}$ is a~sequence of i.i.d.~random variables having \emph{standard normal} $\mathsf{N}(0,1)$ distribution. Moreover, $\{Y_{n,k}\}_{n=1,k=1}^{\infty,n}$ and $\{X_{n}\}_{n=1}^{\infty}$ are also \emph{independent}. The schematic algorithm of the wild bootstrap can be seen as Procedure~\ref{alg:wildboot}. In general, the wild bootstrap replications should be defined in the following way $Y_{n,k}^*:=\bar{Y}_{n,1:n}+\left(Y_{n,k}-\bar{Y}_{n,1:n}\right)X_{k}$, $k=1,\ldots,n$. However, in our case how the test statistics are defined, there would be no impact by adding $\bar{Y}_{n,1:n}$. We define \[ V_n^{\star}(k):=\sum_{i=1}^{k}Y_{n,i}^{\star}\quad\mbox{and}\quad \widetilde{V}_n^{\star}(k):=V_n^{\star}(n)-V_n^{\star}(k). \] \begin{algorithm}[!htb] \caption{Wild bootstrap of the test statistic $\mathscr{Q}(V_n)$ and $\mathscr{R}(V_n)$} \label{alg:wildboot} \begin{algorithmic}[1] \REQUIRE Sequence of observations $Y_{1,n},\ldots,Y_{n,n}$ and number of bootstrap replications~$B$ \ENSURE Bootstrap distributions of $\mathscr{Q}(V_n)$ and $\mathscr{R}(V_n)$, respectively; i.e., the empirical distributions where probability mass $1/B$ concentrates at each of ${}_{(1)}\mathscr{Q}(V_n^{\star}),\ldots,{}_{(B)}\mathscr{Q}(V_n^{\star})$ and ${}_{(1)}\mathscr{R}(V_n^{\star}),\ldots,{}_{(B)}\mathscr{R}(V_n^{\star})$, respectively \FOR[repeat in order to obtain the empirical distributions]{$b=1$ to $B$} \STATE generate random sample $[{}_{(b)}X_{1},\ldots,{}_{(b)}X_{n}]$ from $\mathsf{N}(0,1)$ independently for different $b$'s \STATE calculate ${}_{(b)}Y_{n,k}^{\star}=\left(Y_{n,k}-\bar{Y}_{n,1:n}\right)\times{}_{(b)}X_{k}$ for all $k$'s \STATE calculate ${}_{(b)}V_n^{\star}(k)=\sum_{i=1}^{k}{}_{(b)}Y_{n,i}^{\star}$ and ${}_{(b)}\widetilde{V}_n^{\star}(k)={}_{(b)}V_n^{\star}(n)-{}_{(b)}V_n^{\star}(k)$ for all $k$'s \STATE compute the bootstrap test statistics ${}_{(b)}\mathscr{Q}(V_n^{\star})$ and ${}_{(b)}\mathscr{R}(V_n^{\star})$ \ENDFOR \end{algorithmic} \end{algorithm} The idea behind bootstrapping is to \emph{mimic the original distribution} of the test statistic in some sense with the distribution of the bootstrap test statistic. It is not known and it does not matter whether our observations come form the null hypothesis or the alternative. We are going to prove that $\mathscr{Q}(V_n^{\star})$ and $\mathscr{R}(V_n^{\star})$, respectively, provide asymptotically correct critical values for the test based on $\mathscr{Q}(V_n)$ and $\mathscr{R}(V_n)$, respectively. \begin{theorem}[Wild bootstrap validity]\label{theorem:boot} Suppose that Assumptions~\ref{assump:UP1} and~\ref{assump:UP2} hold. Under the null hypothesis $\mathcal{H}_0$ or under local alternatives $\mathcal{H}_1$ with $\delta_n\rightarrow 0$ as $n\rightarrow 0$, \begin{equation*} \mathscr{Q}(V_n^{\star})\xrightarrow{\dist}\mathscr{S}(W_{\eta})\quad\mbox{and}\quad\mathscr{R}(V_n^{\star})\xrightarrow{\dist}\mathscr{T}(W_{\eta}),\quad n\to\infty \end{equation*} almost surely conditionally on $\{Y_{n,k}\}_{n=1,k=1}^{\infty,n}$. The functionals~$\mathscr{S}$ and~$\mathscr{T}$ are defined in~\eqref{functionalS} and~\eqref{functionalT}, $W_{\eta}(t):=W(\eta(t))$ with~$\eta(t):=\frac{\int_{0}^t\sigma^2(s)\mbox{d} s}{\int_{0}^1\sigma^2(s)\mbox{d} s}$. Under the alternative hypothesis $\mathcal{H}_1$ with $\tau_n=[n\zeta]$ for some $\zeta\in(0,1)$ and having $\delta_n\equiv\delta\neq 0$ fixed, let $\{B(t),0\leq t\leq 1\}$ be a standard Wiener processes independent of~$W$. Then, \begin{equation*} \mathscr{Q}(V_n^{\star})\xrightarrow{\dist}\mathscr{S}\left(W_{\eta}-\frac{\delta}{\varsigma} B_{\zeta}\right)\quad\mbox{and}\quad\mathscr{R}(V_n^{\star})\xrightarrow{\dist}\mathscr{T}\left(W_{\eta}-\frac{\delta}{\varsigma} B_{\zeta}\right),\quad n\to\infty \end{equation*} almost surely conditionally on $\{Y_{n,k}\}_{n=1,k=1}^{\infty,n}$ with $\varsigma^2=\int_{0}^1\sigma^2(t)\mbox{d} t$ and \begin{equation*} B_{\zeta}(t)=\left\{ \begin{array}{ll} (1-\zeta)B(t), & t\leq \zeta;\\ B(\zeta)-\zeta B(t), & t> \zeta. \end{array} \right. \end{equation*} \end{theorem} Theorem~\ref{theorem:boot} assures that the asymptotic distribution of the bootstrap test statistics and the limit distribution of the original test statistics \emph{coincide} under the null hypothesis. Thus, the bootstrap tests approximately keep the same level as the original tests based on the asymptotics from Theorem~\ref{theorem:undernull} even without knowing or estimating the nuisance function~$\eta(t)$. Moreover, the limit distribution of the bootstrap test is not changed under local alternatives, so we avoid the power loss that would be caused by overestimation of the long-run variance under the alternative. Even under fixed alternatives, the distribution of the bootstrap statistics converge to an~almost sure finite limit. In contrast, an~uncorrected kernel estimator for the long-run variance would converge to infinity in this case. Depending on the function $\eta$ and the time of the change represented by~$\zeta$, the quantiles $\mathscr{S}(W_{\eta}-\frac{\delta}{\varsigma} B_{\zeta})$ and $\mathscr{T}(W_{\eta}-\frac{\delta}{\varsigma} B_{\zeta})$, respectively, might be larger or smaller than the quantiles of $\mathscr{S}\left(W_{\eta}\right)$ and $\mathscr{T}\left(W_{\eta}\right)$, respectively, resulting in a~loss or gain of power compared to the use of the asymptotic quantiles from Theorem~\ref{theorem:undernull} (which is only feasible when~$\eta$ is known). Now, the simulated (empirical) distributions of the bootstrap test statistics can be used to calculate the bootstrap critical values, which will be compared to the values of the original test statistics in order to reject the null or not. \subsection{Changepoint estimator} If a~change is \emph{detected}, it is of interest to estimate the time of the change. It is sensible to use \begin{align*} \hat{\tau}_n&:=\operatorname{argmax}_{1\leq k \leq n}\frac{\left|\sum_{i=1}^{k}\left(Y_{n,i}-\bar{Y}_{n,1:n}\right)\right|+\left|\sum_{i=n-k+1}^{n}\left(Y_{n,i}-\bar{Y}_{n,1:n}\right)\right|}{\max\limits_{1\leq i \leq k}\left|\sum_{j=1}^{i}\left(Y_{n,j}-\bar{Y}_{n,1:k}\right)\right|+\max\limits_{k< i \leq n}\left|\sum_{j=i}^{n}\left(Y_{n,j}-\bar{Y}_{n,(k+1):n}\right)\right|}\\ &\equiv \operatorname{argmax}_{1\leq k \leq n}\frac{\big|V_n(k)-k/n V_n(n)\big|+\big|\widetilde{V}_n(n-k)-k/nV_n(n)\big|}{\max\limits_{1\leq i \leq k}\big|V_n(i)-i/kV_n(k)\big|+\max\limits_{k< i \leq n}\big|\widetilde{V}_n(i)-(n-i)/(n-k)\widetilde{V}_n(k)\big|} \end{align*} as a~\emph{changepoint estimator}. Our next theorem shows that under the alternative, the changepoint $\tau_n$ is consistently estimated by the estimator $\hat{\tau}_n$. \begin{theorem}[Estimator's consistency]\label{theorem:estimation} Suppose Assumptions~\ref{assump:UP1}, \ref{assump:UP2}, and~\ref{assump:UP3} hold. Under the alternative hypothesis $\mathcal{H}_1$ such that $\tau_n=[n\zeta]$ for some $\zeta\in(0,1)$, it holds $\hat{\tau}_n/n\xrightarrow{\mathbbm{P}}\zeta$ as $n\to\infty$. \end{theorem} \section{Simulations}\label{sec:simu} \subsection{Asymptotic critical values}\label{subsec:crit} The explicit forms of the limit distributions stated in~\eqref{eq:limit_dist} are not known. The critical values for the simplest case $\eta(t)=t$ may be determined by simulations from the limit distributions $\mathscr{S}(W)$ and $\mathscr{T}(W)$ from Theorem~\ref{theorem:undernull}. Theorem~\ref{theorem:underalt} ensures that we reject the null hypothesis for large values of the test statistics. We have simulated the asymptotic distributions~\eqref{eq:limit_dist} for the stationary case (i.e., a~situation when $\eta(t)=t$) by \emph{discretizing} the standard Wiener process and using the relationship of a~random walk to the standard Wiener process. We considered $1000$ as the number of discretization points within $[0,1]$ interval and the number of simulation runs equals to $100000$. In Table~\ref{tab:crit_val}, we present several critical values for the test statistics~$\mathscr{Q}(V_n)$ and~$\mathscr{R}(V_n)$ under stationarity. \begin{table}[!ht] \caption{Simulated critical values corresponding to the asymptotic distributions of the test statistics~$\mathscr{Q}(V_n)$ and~$\mathscr{R}(V_n)$ under the null hypothesis, where $\eta(t)=t$} \label{tab:crit_val} \begin{center} \begin{tabular}{rrrrrr} \toprule $100(1-\alpha)\%$ & $90\%$ & $95\%$ & $97.5\%$ & $99\%$ & $99.5\%$ \\ \midrule $\mathscr{Q}(V_n)$-based & $1.209008$ & $1.393566$ & $1.571462$ & $1.782524$ & $1.966223$\\ $\mathscr{R}(V_n)$-based & $5.700222$ & $7.165705$ & $8.807070$ & $10.597625$ & $11.755233$\\ \bottomrule \end{tabular} \end{center} \end{table} \subsection{Simulation study} We are interested in the performance of the tests based on the self-normalized test statistics $\mathscr{Q}(V_n)$, $\mathscr{R}(V_n)$ (with $\eta(t)=t$ corresponding to the stationary case) and their wild bootstrap counterparts $\mathscr{Q}(V_n^{\star})$, $\mathscr{R}(V_n^{\star})$ that are completely nuisance parameter free. We focused on the comparison of the \emph{accuracy of critical values} obtained by the wild bootstrap method with the accuracy of critical values obtained by the simulation from the limit distributions. In Figures~\ref{fig:H0} and~\ref{fig:H1}, one may see \emph{size-power plots} for choices of $n\in\{100,400\}$, $\tau_n\in\{n/4,n/2\}$, and $\delta_n\in\{0.5,1.0\}$ considering the test statistics $\mathscr{Q}(V_n)$, $\mathscr{R}(V_n)$, $\mathscr{Q}(V_n^{\star})$, and $\mathscr{R}(V_n^{\star})$ under the null hypothesis and under the alternative. In Figure~\ref{fig:H0}, the empirical rejection frequency under the null hypothesis (actual $\alpha$-errors) is plotted against the theoretical size (theoretical $\alpha$-errors with $\alpha\in\{1\%,5\%,10\%\}$), illustrating the power of the test. The ideal situation under the null hypothesis is depicted by the straight diagonal dotted line. The empirical rejection frequencies ($1-$($\beta$-errors)) under the alternative (with different changepoints and values of the change) are shown in Figure~\ref{fig:H1}. Under the alternative, the desired situation would be a~steep function with values close to~1. For more details on the size-power plots we may refer, e.g., to~\cite{kir2006}. The error terms $\{\sigma(k/n)\varepsilon_k\}_{k=1}^n$ were simulated as two stationary and two non-stationary time series: \begin{itemize} \setlength\itemsep{-.1em} \item \textsf{IID} \ldots~independent and identically distributed random variables; \item \textsf{AR(1)} \ldots~autoregressive (AR) process of order one having a~coefficient of autoregression equal $0.3$; \item \textsf{AR(1)--AR(1)} \ldots~AR process with the coefficient $0.3$, which realizations are multiplied by $\sqrt{2}$ after the first quarter of the time series (deterministic change of volatility); \item \textsf{ARCH(1) inc} \ldots~autoregressive conditional heteroscedasticity (ARCH) process with the second coefficient equal $0.9$, whose realizations are randomly and `increasingly' multiplied (random and in average linearly increasing change of volatility). \end{itemize} The standard normal distribution and the Student $t$-distribution with 3~degrees of freedom are used for generating the innovations of the models' errors. All of the processes are standardized such that they have unit variance at the beginning. The non-stationary ones have changing variance such that their variance is doubled at the end. In the simulations of the rejection rates, we used $5000$ repetitions. When bootstrapping, for each sample we used $2000$ bootstrap samples to compute the bootstrap critical values. \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{FigH0JointBoot.pdf} \caption{Size-power plots for $\mathscr{Q}$ and $\mathscr{R}$ under $\mathcal{H}_0$} \label{fig:H0} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{FigH1JointBoot.pdf} \caption{Size-power plots for $\mathscr{Q}$ and $\mathscr{R}$ under $\mathcal{H}_1$} \label{fig:H1} \end{figure} In all of the subfigures of Figure~\ref{fig:H0} depicting a~situation under the null hypothesis, we may see that comparing the accuracy of $\alpha$-levels (sizes) for different self-normalized test statistics, the integral-type ($\mathscr{R}$-based) method seems to keep the theoretical significance level more firmly than the supremum-type ($\mathscr{Q}$-based) method. The bootstrap approach generally gives critical values that are more accurate than the asymptotic critical values (assuming stationarity, i.e. $\eta(t)=t$), especially for the non-stationary situations. Comparing the case of $\mathsf{N}(0,1)$ innovations with the case of~$t_3$ innovations, the rejection rates under the null tend to be slightly higher for the~$t_3$ distribution. As expected, the accuracy of the critical values tends to be better for larger~$n$. While the $\mathscr{R}$-method performs better under the null, under the alternative method, it has a tendency to have slightly lower power than the $\mathscr{Q}$-method (see Figure~\ref{fig:H1}). In addition, the wild bootstrap technique provides higher power in some situations besides the fact that they are nuisance parameter free. So we strongly recommend to use this bootstrap method. We may also conclude that under~$\mathcal{H}_1$ with larger abrupt change, the power of the test increases. The power decreases when the changepoint is closer to the beginning or the end of the time series. The heavier tails ($t_3$ against $\mathsf{N}(0,1)$) give worse results in general for both test statistics. Moreover, `more dependent' and `more non-stationary' scenarios reveal worsening of the test statistics' performance. Additionally, one can use a~size-power plot with the \emph{adjusted} (empirical) $\alpha$-errors to compare the performance of $\mathscr{Q}(V_n)$ against $\mathscr{R}(V_n)$. The \emph{empirical size-power plots} in Figure~\ref{fig:Adjusted} display the empirical size of the test (i.e., $1-$sensitivity) on the $x$-axis versus the empirical power of the test (i.e., specificity) on the $y$-axis. The ideal shape of the curve is as steep as possible. The empirical size-power plots demonstrate that the self-normalized test statistic $\mathscr{Q}(V_n)$ gives approximately the same empirical powers for the adjusted empirical sizes comparing to the test statistic $\mathscr{R}(V_n)$. This is due to two opposing facts: $\mathscr{R}(V_n)$ keeps the significance level of the test better, but $\mathscr{Q}(V_n)$ gives higher power of the test. The very same conclusion can be made for the bootstrap add-ons: When comparing bootstrapping versus asymptotics, the wild bootstrap method gives slightly higher empirical powers for the adjusted empirical sizes compared to the traditional asymptotics assuming underlying stationarity of the time series' disturbances (i.e., $\eta(t)=t$). \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{FigAdjJointBoot.pdf} \caption{Empirical (adjusted) size-power plots for $\mathscr{Q}$ and $\mathscr{R}$} \label{fig:Adjusted} \end{figure} Furthermore, a~comparison with a~standard and widely used change point detection procedure is provided with emphasis on computational performance. A~classical representative is the \emph{supremum-type cumulative sums (CUSUM)} test statistic \begin{equation}\label{eq:nonR_CUSUM} \mathscr{C}_{M}(V_n):=\frac{1}{\sqrt{\widehat{\sigma}^2_n(M)}}\max_{1\le k\le n-1}\left|\sum_{i=1}^k\left(Y_{n,i}-\bar{Y}_{n,1:n}\right)\right| \end{equation} where $\widehat{\sigma}^2_n(M)$ is a~suitable variance estimator. The null hypothesis is rejected for large values of $\mathscr{C}_{M}(V_n)$. For surveys, we refer to, e.g., \cite{Perron2006}. In order to ensure that a~test statistic is asymptotically distribution-free under the null hypothesis, it is necessary to use a~suitable estimator of variance for the underlying process of random errors. The minimal requirement for $\widehat{\sigma}^2_{n}(M)$ would be consistency under the null hypothesis and boundedness (in probability) under the alternative. Often, the \emph{Bartlett estimator} is used to estimate the variance \begin{equation*} \hat\sigma_n^2(M)=\hat{R}(0)+2\sum_{1\le k\le M}\left(1-\frac{k}{M}\right)\hat{R}(k),\quad M<n, \end{equation*} where $\hat{R}(k)=\frac{1}{n}\sum_{1\le i\le n-k}(Y_{n,i}-\bar{Y}_{n,1:n})(Y_{n,i+k}-\bar{Y}_{n,1:n})$, $0\le k<n$. However, it does not always provide satisfactory results and finding a~proper value of~$M$ may be troublesome. The rate of convergence is small even under the null hypothesis and $\hat\sigma_n^2(M)$ might go to infinity under the alternative~\citep{HHH2008}. Other similar types of estimators can be used instead, for instance Parzen kernels \citep{Andrews1991}, but they still possess the described deficiency. The consistency properties of the above described Bartlett estimator and of its modification are studied in~\cite{Antoch1997}. The authors also describe difficulties of long-run variance estimation when detecting a~change in the mean of a~linear process in more detail. A~simulation study shows that it is not easy to find a~variance estimate that would work well both under null hypothesis and under alternative. Furthermore, such estimators are often very sensitive to the \emph{choice of the window length~$M$}. Based on simulation studies performed by~\cite{Antoch1997}, we decided to use a~rule of thumb of $M=n/10$. Asymptotic distribution of the test statistic~\eqref{eq:nonR_CUSUM} can be found, e.g., in~\cite{CH1997}. The corresponding critical values come from~\cite{Kulperger1990}. Now, we demonstrate the performance of our approaches---the $\mathscr{Q}$ and $\mathscr{R}$ self-normalized test statistics---compared to the traditional CUSUM test statistic (i.e., based on $\mathscr{C}_M$). Empirical sizes under the null hypothesis and empirical powers under the alternative ($\tau_n=n/2$ and $\delta_n=0.5$) of our two detection procedures compared to the standard one are shown in Figure~\ref{fig:H0Compare} and in Figure~\ref{fig:H1Compare}, respectively. The number of repetitions for the simulation of the rejection rates is again set to $5000$. The sample size is chosen as $n=200$ and the window length for the variance estimate is $M=20$. \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{FigH0CompareBoot.pdf} \caption{Size-power plots for $\mathscr{C}_M$, $\mathscr{Q}$, and $\mathscr{R}$ under $\mathcal{H}_0$ (sample size $n=200$)} \label{fig:H0Compare} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{FigH1CompareBoot.pdf} \caption{Size-power plots for $\mathscr{C}_M$, $\mathscr{Q}$, and $\mathscr{R}$ under $\mathcal{H}_1$ (sample size $n=200$, change of $\delta_n=0.5$ at time $\tau_n=n/2$)} \label{fig:H1Compare} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{FigAdjJointCompareBoot.pdf} \caption{Empirical (adjusted) size-power plots for $\mathscr{C}_M$, $\mathscr{Q}$, and $\mathscr{R}$ (sample size $n=200$, change of $\delta_n=0.5$ at time $\tau_n=n/2$)} \label{fig:AdjCompare} \end{figure} To conclude, the CUSUM test statistic yields too small empirical size, see Figure~\ref{fig:H0Compare}. It rejects more often than it should and, moreover, it provides lower power (Figure~\ref{fig:H1Compare}) compared to the self-normalized type test procedures, especially for small significance levels ($5\%$ and $1\%$). There are two possible reasons for that: the classical CUSUM procedure relies on the variance estimate, which can be troublesome, and it requires a~suitable choice of the nuisance parameter. This illustrates that avoiding the nuisance parameter estimation should really be considered as one of advantages of the proposed methods. Besides that, the wild bootstrap performs better compared to the traditional asymptotics, which can be illustrated via adjusted size-power plots in Figure~\ref{fig:AdjCompare}. Furthermore, one can concentrate on a~situation that is far away from a~stationary case. Let us take into consideration a~zero-mean AR(1) sequence (the AR-coefficient is set to $0.3$) of $n=200$ random errors, where the random variables from the first quarter of the series are multiplied by~$10$. This leads to the variance function $\eta(t)=40t/13$ for $t\in[0,1/4)$ and $\eta(t)=10/13+(4t-1)/13$ for $t\in[1/4,1]$. The corresponding alternative $\mathcal{H}_1$ is chosen as $\tau_n=n/2$ and $\d_n=1$. Figure~\ref{fig:Extreme} evidence, on one hand, that the asymptotic approach assuming underlying stationarity yields very unreliable results. On the other hand, the bootstrap method provides reasonable rejection rates even in such non-stationary case. \begin{figure}[!ht] \begin{center} \includegraphics[width=0.49\textwidth]{FigH0CompareExtremeBoot.pdf} \includegraphics[width=0.49\textwidth]{FigH1CompareExtremeBoot.pdf} \includegraphics[width=0.61\textwidth]{FigAdjJointCompareExtremeBoot.pdf} \end{center} \caption{Size-power plots for $\mathscr{Q}$ and $\mathscr{R}$ under $\mathcal{H}_0$ (top-left) and under $\mathcal{H}_1$ (top-right) for a~very heteroscedastic case together with the corresponding empirical (adjusted) size-power plots (bottom)} \label{fig:Extreme} \end{figure} Afterwards, a~simulation experiment is performed to study the \emph{finite sample} properties of the changepoint estimator for an~abrupt change in the mean. In particular, the interest lies in the \emph{empirical distributions} of the proposed estimator visualized via boxplots, see Figure~\ref{fig:Estimator}. The simulation setup is kept the same as described above. \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{FigBoxplot.pdf} \caption{Boxplots of the estimated changepoint $\hat{\tau}_n$} \label{fig:Estimator} \end{figure} It can be concluded that the precision of our changepoint estimator is satisfactory even for relatively short time series regardless of the errors' structure. Furthermore, the disturbances with heavier tails or changing variance yield less precise estimators than stationary innovations with light tail. One may notice that higher precision is obtained when the changepoint is closer to the middle of the time series. It is also clear that the precision of $\hat{\tau}_n$ improves markedly as $\delta_n$ increases. \section{Practical Applications}\label{sec:data} \subsection{Dieselgate} Especially in many time series from finance, (conditional) heteroscedasticity appears frequently. In our first data example, we analyze the daily absolute log returns of the Volkswagen stock prices from January~1, 2015 to November~26, 2015 (VOW.DE, XETRA -- XETRA Delayed Price. Currency in EUR. Open. Downloaded on May 30, 2018 \url{https://finance.yahoo.com/quote/VOW.DE?p=VOW.DE}), which are visualized in Figure~\ref{fig:VW}. \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{VW.pdf} \caption{Absolute log returns of the Volkswagen stock price (January~1, 2015 -- November~26, 2015). The changepoint estimate corresponding to the emissions scandal on September~18, 2015 is depicted by the vertical line} \label{fig:VW} \end{figure} Both our tests as well as their wild bootstrap add-ons reject the null hypothesis of a~constant mean in the absolute log returns (cf.~Table~\ref{tab:VW}), indicating an~increased volatility of the Volkswagen stock price. In contrast, \cite{dehling2015robust} did not find a~significant change using the classical CUSUM-test. In their article, several robust tests detected a~change. There are several large values in this time series, but the reason could be a~period with strongly increased variance. \begin{table}[!ht] \caption{Self-normalized test statistics (asymptotic and bootstrap) together with the corresponding critical values for the Volkswagen stock price data, considering a~significance level of $5\%$} \label{tab:VW} \begin{center} \begin{tabular}{rrrrr} \toprule & $\mathscr{Q}(V_n)$ & $\mathscr{R}(V_n)$ & $\mathscr{Q}(V_n^{\star})$ & $\mathscr{R}(V_n^{\star})$ \\ \midrule Test statistic & $1.546779$ & $10.74026$ & $1.546779$ & $10.74026$\\ Critical value & $1.393566$ & $7.165705$ & $1.388683$ & $8.109334$\\ \bottomrule \end{tabular} \end{center} \end{table} As an estimator for our change, we obtain $\hat{\tau}_n=182$ (depicted by a~vertical line in Figure~\ref{fig:VW}), which corresponds to September 18, 2015. On this day, the United States Environmental Protection Agency issued a notice of violation, which lead to the Volkswagen emissions scandal. Our procedure is capable to detect and, consequently, to estimate the changepoint based on only 10~weeks of daily data after the emissions scandal. \subsection{Elbe river} Our second data example consists of the annual maximum discharge of the river Elbe at Dresden, Germany, in the years~1851 to~2012. The variance seems to be lower in the 20th century compared to second half of the 19th century (see Figure~\ref{fig:Labe}). Therefore, we think that our tests are a~good choice for this data set, as they are not effected by heteroscedasticity. \begin{figure}[!ht] \centering \includegraphics[width=.8\textwidth]{Labe.pdf} \caption{Yearly maximal discharges (1851 -- 2012) of the Elbe river in Dresden, Germany. Year~1901 depicted by the vertical line is the changepoint estimate} \label{fig:Labe} \end{figure} All of our four testing procedures reject the null hypothesis of constant mean of the maximum discharges over the whole observation period, see Table~\ref{tab:Labe}. This data set has been previously analyzed with other methods: \cite{sharipov2016sequential} used a~Cram\'er-von Mises-type test statistic and detected a~change in the marginal distribution. \cite{vogel2017studentized} detected a~shift in location using a~robust test based on the Hodges-Lehmann-estimator, while the usual CUSUM statistic did not lead to rejection of the hypothesis (stationarity). So, our self-normalized tests seem to work more reliable than the ordinary CUSUM-test also in this example. \begin{table}[!ht] \caption{Self-normalized test statistics (asymptotic and bootstrap) together with the corresponding critical values for the annual maximum discharge of the river Elbe at Dresden, Germany, considering a~significance level of $5\%$} \label{tab:Labe} \begin{center} \begin{tabular}{rrrrr} \toprule & $\mathscr{Q}(V_n)$ & $\mathscr{R}(V_n)$ & $\mathscr{Q}(V_n^{\star})$ & $\mathscr{R}(V_n^{\star})$ \\ \midrule Test statistic & $1.481084$ & $7.936363$ & $1.481084$ & $7.936363$\\ Critical value & $1.393566$ & $7.165705$ & $1.476675$ & $7.599609$\\ \bottomrule \end{tabular} \end{center} \end{table} \section{Conclusions}\label{sec:con} We have proposed two tests for changepoints with desirable theoretical properties: The asymptotic size of the tests is guaranteed by a~limit theorem even under heteroscedasticity and dependence, the tests and the related changepoint estimator are consistent. By combining self-normalization and the wild bootstrap, there are neither tuning nor nuisance parameters involved in the whole testing procedure, which makes this framework effortlessly applicable. In our simulations, the tests show reliable performance. Especially the bootstrap test based on the integral-type self-normalized CUSUM-statistic~$\mathscr{R}$ has an~empirical size very close to the nominal level in a~wide range of situations. In the data examples, we have shown that our tests can find changes, which were not detected before using the ordinary CUSUM-tests. Let us note that the test statistic could also be applied for other data generating processes. The limit distribution is derived from the limit distribution of the partial sum process by the continuous mapping theorem. \cite{Shao2011} studied long range dependent process, where the partial sum process converges weakly to a~fractional Brownian motion, and it would be possible to obtain the limit distribution of our new test statistic in the same way. Furthermore in the case of heavy-tailed random variables, the partial sum process might converge weakly to a~stable L\'evy process. While it should be possible to identify the limit distribution of our test statistics with the help of the continuous mapping theorem, we would expect a~loss of power. Another possibility would be robustified tests following ideas of \cite{PP2018}. But this goes beyond the scope of this paper and is a~topic for future research.

https://arxiv.org/abs/1611.00084

The Largest Pure Partial Planes of Order 6 Have Size 25

In this paper, we prove that the largest pure partial plane of order 6 has size 25. At the same time, we classify all pure partial planes of order 6 and size 25 up to isomorphism. Our major approach is computer search. The search space is very large so we use combinatorial arguments to rule out some of the cases. For the remaining cases, we subdivide each search by phases and use multiple checks to reduce search space via symmetry.

\section{Introduction} The problem of the existence of finite projective planes of certain orders has been attracting mathematicians' interest for hundreds of years. However, the problem still remains widely open. People know that finite projective planes of order equal to prime powers exist and no finite projective planes of order not equal to a prime power have been found. Therefore, some mathematicians conjectured that finite projective planes can only have prime power orders. There was some progress made in the past. In 1938, Bose \cite{bose1938application} proved that there is no projective plane of order 6 by relating the existence of a finite projective plane to the existence of a \textit{hyper-Graeco-Latin square}, which is known as \textit{orthogonal Latin squares} in modern terminology. In 1949, Bruck and Ryser \cite{bruck1949nonexistence} proved that if the order $n$ is congruent to 1 or 2 modulo 4 and $n$ cannot be represented as the sum of two perfect squares, then there does not exist any finite projective planes of order $n$. This result is known as Bruck-Ryser theorem. By this famous combinatorial theorem, infinitely many cases are solved but still infinitely many cases are left. Due to Bose's result and the Bruck-Ryser theorem, the smallest unsolved case is order $n=10$. After some progress using binary codes \cite{assmus1970possibility}, Lam, Thiel and Swierczp \cite{lam1989non} proved the nonexistence of finite projective planes of order 10 with the help of super computers and a total of 2 to 3 years of running time. Aside from finding more finite projective planes or proving their nonexistence, there is one more attractive question. We already know that there is no finite projective planes of order 6, but how close can we come to constructing such a plane? In particular, what is the largest pure partial plane (see Definition \ref{Def:ppp}) of order 6 we can construct? In \cite{hering2007partial}, a pure partial plane of order 6 related to icosahedron is constructed and in \cite{prince2009pure}, two pure partial planes of order 6 with 25 lines that extend the dual of the point-line incidence structure of $PG(3,2)$, the three-dimensional projective geometry, are constructed. In \cite{mccarthy1976approximations}, McCarthy et al. proved that there are no pure partial plane of order 6 and size 29 with very long combinatorial arguments. However, the exact maximum has not been given. In this paper, we prove that the maximum size of pure partial planes of order 6 is 25. Or in other words, a pure partial plane of order 6 contains at most 25 lines. \begin{theorem}\label{thm:main} The maximum size of a pure partial plane of order 6 is 25. Furthermore, all pure partial planes of size 25 are listed in Appendix \ref{Sec:allPPP}. \end{theorem} To do this, we use computer search combined with standard combinatorial arguments. In Section \ref{Sec:Prelim}, we define the notion of saturated pure partial planes and introduce some related and useful lemmas. In Section \ref{Sec:Alg}, we give our algorithmic model for computer search and in Section \ref{Sec:Result}, we present our results. We also attach our code for readers to verify. A list of codes that we provide is shown in Appendix \ref{Sec:files}. Finally, the proof of our main theorem comes in Section \ref{Sec:Main}. All the pure partial planes of order 6 and size 25 are provided in Appendix \ref{Sec:allPPP}. \section{Preliminaries}\label{Sec:Prelim} In this paper, a ``point" means an element in a universe and a ``line" means a subset of this universe, or equivalently, a set of ``points". We will consider ``points" and ``lines" only in a set-theoretic view and won't discuss any finite geometry here. \begin{definition} A \textit{finite projective plane (FPP) of order} $n$, or a \textit{projective plane of order} $n$, is a collection of $n^2+n+1$ points and $n^2+n+1$ lines, such that \begin{enumerate} \item[(1)] every line contains $n+1$ points; \item[(2)] every point is on $n+1$ lines; \item[(3)] every two distinct lines intersect at exactly one point; \item[(4)] every any two distinct points lies on exactly one line. \end{enumerate} \end{definition} \begin{definition}\label{Def:ppp} A \textit{pure partial plane (PPP) of order} $n$ \textit{and size} $s$ is a collection of $n^2+n+1$ points and $s$ lines, such that \begin{enumerate} \item[(1)] every line contains $n+1$ points; \item[(2)] every two distinct lines intersect at exactly one point; \end{enumerate} \end{definition} In Definition \ref{Def:ppp}, we say that there are $n^2+n+1$ points. Just for clarity, there can be at most $n^2+n+1$ points since some points may not appear in any of the lines. \begin{definition} We say that a pure partial plane is \textit{saturated} if no lines can be added to it such that it still remains a pure partial plane. We use the abbreviation SPPP for \textit{saturated pure partial plane}. \end{definition} \begin{definition} Two pure partial planes are isomorphic if there exists a bijection of their points and a bijection of their lines such that the point-in-line relation is equivalent under these two bijections. \end{definition} We are only interested in (saturated) pure partial planes up to isomorphism. And for the rest of the paper, we will consider two (saturated) pure partial planes to be the same if they are isomorphic. In other words, we only care about isomorphism classes. For convenience, we make the following definition. \begin{definition} We say that two lines are \textit{compatible} if they intersect at exactly one point and that two sets of lines are \textit{compatible} if every line from one set is compatible with every line from the other set. \end{definition} It is immediate that a finite projective plane is always a saturated pure partial plane of the same order, and is a largest one, in terms of the size. From now on, we will always use $n$ for the order and $s$ for the size. For convenience, we label all points as $0,1,\ldots,n^2+n$ and represent straightforwardly a line as a set of cardinality $n+1$, e.g., $\{0,1,2,3,4,5,6\}$. \begin{lemma}\label{Lem:NoN} For a saturated pure partial plane of order $n$, no points appear in exactly $n$ lines. \end{lemma} \begin{proof} We use proof by induction. Suppose that there is a SPPP with point 0 appearing in $n$ lines. Assume that these $n$ lines are $\{0,1,\ldots,n\},\ \{0,n+1,\ldots,2n\},\ \ldots,\ \{0,n^2-n+1,\ldots,n^2\}$. For any other line $L$, it does not contain point $0$, so by definition, it must intersect $\{in+1,in+2,\ldots,in+n\}$ at exactly one point, for $i=0,1,\ldots,n-1$. Since $L$ contains $n+1$ points in total, we know that $L$ must intersect with $\{n^2+1,n^2+n,\ldots,n^2+n\}$ at exactly one point, too. Line $L$ does not contain 0 so $L$ intersects with $\{0,n^2+1,\ldots,n^2+n\}$ at exactly one point. It is then obvious that we can add a new line $\{0,n^2+1,\ldots,n^2+n\}$ to this collection, contradicting the property of saturation. \end{proof} \begin{lemma}\label{Lem:SumA} For a pure partial plane of order $n$ and size $s$, suppose that there are $a_k$ points that appear in $k$ lines, for $k=0,1,\dots$. Then $$\sum_{k}ka_k=(n+1)s;\qquad\sum_{k}k^2a_k=s^2+ns.$$ \end{lemma} \begin{proof} For each line $i$ in this pure partial plane, with $i=1,\ldots,s$, associate a vector $L_i\in\{0,1\}^{n^2+n+1}$ with it, such that $L_{i,j}=1$ if point $j$ appears in line $i$ and equals 0 otherwise. Let $v=L_1+\cdots+L_s$. The entries of $v$ are just a permutation of $m_0,m_1,\ldots,m_{n^2+n}$ so the $L^1$ norm of $v$ is $\sum ka_k$ and at the same time, it is the sum of the $L^1$ norms of $L_i$'s, giving us $(n+1)s$. At the same time, $v\cdot v=\sum k^2a_k$. By definition, $L_i\cdot L_j=1$ if $i\neq j$ and $L_i\cdot L_j=n+1$ if $i=j$ so $v\cdot v=(n+1)s+(s^2-s)=s^2+ns$, as desired. \end{proof} Lemma \ref{Lem:SumA} is a simple but useful lemma that has appeared in other forms in previous works. For example, \cite{mccarthy1976approximations} mentions essentially the same thing in Section 3 but in a different format. \begin{lemma}\label{Lem:SumC} Suppose that $\{i_1,i_2,\ldots,i_{n+1}\}$ is a line in a pure partial plane of order $n$ and size $s$, and suppose that point $i_k$ appears $c_{i_k}$ times. Then $$c_{i_1}+c_{i_2}+\cdots+c_{i_{n+1}}=s+n.$$ \end{lemma} \begin{proof} All the lines in this pure partial plane are either the line $\{i_1,\ldots,i_{n+1}\}$ or contain exactly one of $i_1,\ldots,i_{n+1}$. Since $i_k$ appears $c_{i_k}$ times, there are $c_{i_k}-1$ lines that contain $i_k$ but not $i_j$ for all $j\neq k$ with $1\leq j\leq n+1$. Therefore, we have $(c_{i_1}-1)+(c_{i_2}-1)+\cdots+(c_{i_{n+1}}-1)+1=s$ and thus $$c_{i_1}+c_{i_2}+\cdots+c_{i_{n+1}}=s+n.$$ \end{proof} \begin{theorem}\label{Thm:even} For any saturated pure partial plane of even order, there exists a point that appear in at least 3 lines. \end{theorem} \begin{proof} Assume the opposite that there exists a SPPP such that the order $n$ is even and every point appears in at most 2 lines. By Lemma \ref{Lem:SumA} and following its notation, we have $a_1+2a_2=(n+1)s$ and $a_1+4a_2=s^2+ns$. It gives $a_1=s(n+2-s)$ and $a_2=\frac{s^2-s}{2}={s\choose2}$. Since $a_1\geq0$, $s\leq n+2$. Notice that the value of $a_2$ is actually obvious because no three lines intersect at one point so the intersection points of different pairs of lines are different. When $n=2$, we can first assume that our first two lines are $\{0,1,2\}$ and $\{0,3,4\}$. Forcing point 0 to appear in only 2 lines, we can determine the next two lines to be $\{1,3,5\}$ and $\{2,4,6\}$, which are unique up to permutation. Now, the size of this pure partial plane is already 4, which equals $n+2$. However, it is not saturated since the line $\{0,5,6\}$ is compatible with it. This yields a contradiction. So we assume that $n\geq4$. Then $$a_1+a_2=-\frac{1}{2}s^2+\frac{2n+3}{2}s=\frac{n^2+3n+2}{2}-(s-n-1)(s-n-2)\leq\frac{n^2+3n+2}{2}\leq n^2$$ as $n\geq4$. So $a_0\geq n+1$, meaning that we have plenty of points to use. Suppose that points $0,1,\ldots,a_2-1$ appear two times and points $n^2,n^2+1,\ldots,n^2+n$ do not appear. Label the lines as $1,2,\ldots,s$. If $s$ is even, say point 0 appears in lines $1,2$, point 1 appears in lines $3,4$, $\ldots$, point $\frac{s}{2}-1$ appears in lines $s-1,s$, then we can add a new line $\{0,1,\ldots,\frac{s}{2}-1,n^2,n^2+1,\ldots,n^2+n-\frac{s}{2}\}$. It is easy to see that this new line intersects with previous lines at exactly one point, as it intersects lines $2k-1,2k$ at point $k-1$ only, for $k=1,\ldots,\frac{s}{2}$. If $s$ is odd, then as $n$ is even, $a_1=s(n+2-s)\neq0$. So we can assume that point $n^2-1$ appears exactly one time and line $s$ contains it. Further, since all pairs of lines intersect at some point, we can assume that point 0 appears in lines $1,2$, point 1 appears in lines $3,4$, $\ldots$, point $\frac{s-3}{2}$ appears in lines $s-2,s-1$. Similarly, a new line, $\{0,1,\ldots,\frac{s-3}{2},n^2-1,n^2,\ldots,n^2+n-\frac{s+1}{2}\}$ can be added. Therefore, this pure partial plane cannot be saturated. \end{proof} \begin{theorem}\label{Thm:odd} For each odd number $n\geq3$, there exists a saturated pure partial plane of order $n$ such that no points appear in more than 2 lines. \end{theorem} \begin{proof} The construction is straightforward. Draw $n+2$ lines in $\mathbb{R}^2$ such that no two are parallel and no three are concurrent. This gives us ${n+2\choose 2}<n^2+n+1$ intersection points and $n+2$ lines, each passing through $n+1$ points. Clearly it is a pure partial plane. If it is not saturated, then we should be able to find a subset of these intersection points, as well as some previously unused points, such that each previously existing line passes through exactly one of them. However, each intersection points appear in exactly 2 lines and each previously unused points appear in exactly 0 lines, while there are $n+2$ previously existing lines. Because $n+2$ is odd, such a set cannot be found. Therefore, this construction indeed provides a SPPP as desired. An example is given in Figure \ref{Fig:sppp3}. \end{proof} \begin{figure}[h!] \centering \includegraphics[scale=0.4]{1.png} \caption{a SPPP of order 3 and size 5} \label{Fig:sppp3} \end{figure} Theorem \ref{Thm:even} and Theorem \ref{Thm:odd} are not related to our main theorem. It is still good to have them in the sense that we want to understand the notion of ``saturation" better. The problem of testing isomorphism between pure partial planes can be reduced to the problem of graph isomorphism. \begin{definition}\label{Def:PLAG} For a pure partial plane of order $n$ and size $s$, define its \textit{point-line-adjacency graph} to be an undirected simple bipartite graph with $n^2+n+1+s$ vertices, representing all points and lines where a vertex representing a line is connected to a vertex representing a point if and only if the line contains the point. \end{definition} Notice that in the definition, we allow some vertices to have degree 0, although this detail is negligible. \begin{theorem}\label{Thm:GraphIso} Two pure partial planes of the same order that are not finite projective planes are isomorphic if and only if their point-line-adjacency graphs are isomorphic. \end{theorem} \begin{proof} One direction is clear. If two pure partial planes are isomorphic, then their point-line-adjacency graphs are isomorphic. The bijection between points and the corresponding bijection between lines give a bijection between vertices in the graphs. So now let $A$ and $B$ be two pure partial planes of the same order with point-line-adjacency graphs $G_A$ and $G_B$, respectively, and assume that $G_A$ is isomorphic to $G_B$. Clearly, $G_A$ and $G_B$ have the same number of vertices so we know that $A$ and $B$ have the same size $s$. If $s=0$, the theorem is correct. Now we assume $s>0$. Then, $G_A$ and $G_B$ must have the same number of vertices of degree 0, so the number of points appearing in $A$ is the same as the number of points appearing in $B$. For the other vertices of degree at least 1, $G_A$ and $G_B$ are bipartite since there are no edges between points and no edges between lines. By definition, since two lines intersect at exact one point, there must be a length 2 path from any vertex representing a line to another vertex representing a line. So all vertices representing lines are in one connected components, and therefore, this connected component contains all vertices representing points that appear at least once. So, the number of connected components in $G_A$ and $G_B$, ignoring the degree 0 vertices, is 1. Now, the isomorphism between $G_A$ and $G_B$ will first pair up their degree 0 vertices. And then, it will pair up one part (of the bipartite graph) of $G_A$ to one part of $G_B$. We only need to make sure that vertices representing points in $G_A$ are not paired up with vertices representing lines in $G_B$. For that to happen, all points in $G_A$ that appear will have to appear $n+1$ times since all lines in $G_B$ contain $n+1$ points. The number of points that appear is $(n+1)\cdot s/(n+1)=s$. Suppose points that appear are labeled as $1,\ldots,s$. For $i=1,\ldots,s$ let $L_i\in\{0,1\}^s$ be such that $L_{i,j}=1$ if line $i$ contains point $j$ and 0 otherwise. Then the dot product $L_i\cdot L_k=1$ if $i\neq k$ and $L_i\cdot L_i=n+1$. Let us compute $x=(L_1+\cdots+L_s)\cdot(L_1+\cdots+L_s)$. First, $L_1+\cdots+L_s=(n+1,n+1,\ldots,n+1)$ so $x=(n+1)^2\cdot s$. At the same time, $x=s\cdot(n+1)+s(s-1)$. Comparing these two, since $s>0$, we easily get $s=n^2+n+1$, which means that $A$ is a finite projective plane, contradicting our assumption. \end{proof} \begin{remark} Theorem \ref{Thm:GraphIso} fails when we consider two finite projective planes that are not self-dual. Examples include Hall planes \cite{hall1943projective}. \end{remark} By Theorem \ref{Thm:GraphIso}, we are able to transform the problem of testing isomorphism between saturated pure partial planes into graph isomorphism. Therefore, we can then use the fastest online code for graph isomorphism, \textbf{nauty} and \textbf{Traces} \cite{McKay201494}, to do so. \section{Programming Model}\label{Sec:Alg} From now on, we will use computer search to find saturated pure partial planes. In this section, we will present a programming model for searching. Our model is highly adjustable with many conditions under specified. In the next section, we will go into details about specific cases and will specify the conditions that are unclear for now. We start with a certain pure partial plane which we call a \textit{starting configuration}. By brute force, we then generate a list of lines for us to choose from, which we call a \textit{starting list}, that are compatible with this starting configuration. Using the starting configuration and this list of lines, we do a depth first search, adding line by line to this starting configuration from the list and removing incompatible lines from the list until the list becomes empty. Whenever we get a saturated pure partial plane (the corresponding list of compatible lines is empty), we check if it is isomorphic to any of the saturated pure partial planes we already have by Theorem \ref{Thm:GraphIso} and \textbf{nauty} and \textbf{Traces} \cite{McKay201494}. If not, we record this saturated pure partial plane. Here is the basic algorithmic model for depth first search (DFS). In the following diagram, there are steps that are not specified since we use different implementations for different purposes, including Step \ref{Stp:End}, Step \ref{Stp:Saturated} and Step \ref{Stp:Check}. We will also explain them below. \begin{algorithm} \caption{Depth First Search for Saturated Pure Partial Planes}\label{Alg:dfs} \begin{algorithmic}[1] \Procedure{Depth-First-Search}{$ppp_0,rl_0$}\Comment{$ppp_0$ is a pure partial plane, $rl_0$ is a list of lines} \If{$ppp_0,rl_0$ satisfy certain terminating properties}\label{Stp:End} \If{$ppp_0$ satisfies certain properties and is not isomorphic to all PPPs recorded already}\label{Stp:Saturated} \State Record $ppp_0$ globally \EndIf \Else \For{each line $L$ in $rl_0$} \State $ppp_1\leftarrow ppp_0+L$\Comment{a new pure partial plane} \State Construct $rl_1$ from $rl_0$ by selecting the lines that intersect with $L$ at exactly one point \If{$ppp_1$ passes all the checks}\label{Stp:Check} \State \textsc{Depth-First-Search}($ppp_1,rl_1$) \EndIf \EndFor \EndIf \EndProcedure \end{algorithmic} \end{algorithm} This paradigm is very straightforward and simple. However, usually the search space is very large so we need methods to cut down some symmetric cases beforehand. In Step \ref{Stp:End}, the \textit{certain terminating properties} is usually implemented as checking if $rl_0$ is empty. We will assume so if not specified. Step \ref{Stp:Check} is the main step in which we cut off symmetric cases. In this step, we will typically check the following properties of $ppp_1$: \begin{itemize} \itemsep0em \item[1.] If point $i$ has appeared in this pure partial plane, then point $i-1$ must also appear in this pure partial plane, for $i=1,2,\ldots,n^2+n$. \item[2.] The line $L$ just added must be lexicographically greater than the last line in $ppp_0$, assuming all points in each line are sorted. \item[3.] The number of times that certain points appear should not exceed certain values. These parameters will be specified in Section \ref{Sec:Result} where we are using this algorithm. \end{itemize} Check 1 above in Step \ref{Stp:Check} is not always useful. When $n=6$, the cases we are dealing with usually have one point that appears 7 times, meaning that in the starting configuration, all points have already appeared. Check 2 above in Step \ref{Stp:Check} can also be implemented in the way that in our DFS step after adding the new line to the pure partial plane, we discard all lines from the list of compatible lines that are lexicographically greater than this new line. Check 3 above in Step \ref{Stp:Check} is the most important one. Typically we will divide the whole case by assuming the number of appearance of certain points. Here, we check that if the number of appearance of such points in $ppp_1$ has already exceeded our assumption. Requiring some properties of $ppp_0$ in Step \ref{Stp:Saturated} usually helps us reduce the number of isomorphism testing. For example, if in our starting configuration, points $3,4,5,6$ are symmetric, then we can require that in $ppp_0$ the number of times that $3,4,5,6$ appear forms a non-decreasing sequence. In this way, some isomorphic cases will be quickly discarded. \section{Search Results}\label{Sec:Result} Our goal is to prove that all possible pure partial planes of order 6 have size at most 25 and to give all pure partial planes of order 6 and size 25. Following our previous notations, let $a_i$ be the number of points that appear exactly $i$ times. We will only consider saturated pure partial planes, in order to use Lemma \ref{Lem:NoN} and get that $a_6=0$, meaning that no points can appear exactly 6 times. Intuitively, if we want our SPPPs to have large sizes, we need to have the points appear as many times as possible. So as an overview, we will search for the cases where $a_7\geq2$ and also the cases where $a_5$ is sufficiently large. In the next section (Section \ref{Sec:Main}), we will give a proof showing that all possible SPPPs with size at least 25 are already covered in our search. And at that point, it will be clear why we discuss these cases. In this section, we will consider five cases specified in each subsection. For each of them, we will use the algorithm given in Section \ref{Sec:Alg} in multiple phases. In each phase, the inputs are some pure partial planes regarded as starting configurations and the outputs are some bigger pure partial planes, that will be used as starting configurations for the next phase. Intuitively, using multiple phases instead of one will reduce search time since some symmetric cases can be cut off when they have not grown very big. Essentially, searching for pure partial planes in multiple phases is like doing breadth first search. Since we do isomorphism testings after each phase, the idea of combining breadth first search into the depth first search backbone can speed up the search. However, we want the number of phases to be small because breadth first search may consume too much space. For convenience, we will assume that $\{0,1,2,3,4,5,6\}$ is the first line in our starting configuration (except the last case). Also, whenever we talk about a particular ``Step", we are referring to our algorithmic model in Section \ref{Sec:Main}. We provide a list of programs for readers to verify (Appendix \ref{Sec:allPPP}). \subsection{At least 3 points appear 7 times} First, 0 appears 7 times. Assume that these 7 lines are $\{0,6k+1,6k+2,\ldots,6k+6\}$ where $k=0,1,\ldots,6$. At this stage, all other points are equivalent under symmetric group so we can safely assume that 1 appears 7 times, too. Let the next 6 lines be $\{1,k+7,k+13,k+19,k+25,k+31,k+37\}$ where $k=0,1,\ldots,5$. It is also clear that the choice of these 6 lines are unique. Now that we have 13 lines in our starting configuration, we need to divide this case. For the third point that appears 7 times, it may be a point that appear in the same line with both 0 and 1, i.e. $2,3,4,5,6$ or other points. We divide this case into two subcases where 2 appears 7 times and where 7 appears 7 times. Notice that in both cases, we can add one more line $\{2,7,14,21,28,35,42\}$ into the collection using symmetry. \subsubsection{Point 2 appears 7 times}\label{Res:2x7} \noindent\textbf{Phase 1} We use our program with stating configuration being this 14-line pure partial plane, the starting list being all lines that start with point 2 and are compatible with the starting configuration. In Step \ref{Stp:End} (described in Section \ref{Sec:Alg}), we simply require that point 2 appears 7 times or equivalently, the size of $ppp_0$ is 19. In Step \ref{Stp:Saturated}, we do nothing and in Step \ref{Stp:Check}, we only do check 2. which checks the lexicographical order. Running the program gives us a total of 12 nonisomorphic pure partial planes of size 19, where 0,1,2 appear 7 times. These starting configurations are shown in file ``case1-1-phase1.txt". \ \noindent\textbf{Phase 2} Then we treat these 12 pure partial planes as starting configurations and run our program again, with the starting list being all lines that are compatible with the starting configuration. The search space is pretty small in this case so we do not actually need a lot of checks. The only check we implemented here is the check of lexicographical order in Step \ref{Stp:Check}. In Step \ref{Stp:End}, we require the list of lines $rl_0$ to be empty. These 12 starting configurations provide 36 nonisomorphic saturated pure partial planes. The results are shown in file ``case1-1-phase2(SPPP).txt". Among these results, the maximum size is 25 and there are 3 SPPPs that achieve 25. \subsubsection{Point 7 appears 7 times} \noindent\textbf{Phase 1} We use our program with starting configuration begin the 14-line pure partial plane described above, the starting list being all lines with point 7 and compatible with the starting configuration. Similarly, in Step \ref{Stp:End}, we require that point 7 must appear exactly 7 times, or equivalently, the size of $ppp_0$ is 18. And in Step \ref{Stp:Check}, we only do check 2. which checks the lexicographical order. The program produces 2 nonisomorphic pure partial planes of size 18 where 0,1,7 appear 7 times. These starting configurations are shown in file ``case1-2-phase1.txt". \ \noindent\textbf{Phase 2} Then we use these 2 pure partial planes as starting configurations to get saturated pure partial planes using our program. In Step \ref{Stp:Check}, we require that points 2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 19, 25, 31, 37 can appear at most 5 times. Otherwise, if one of them appears at least 6 times, by Lemma \ref{Lem:NoN}, it must appear 7 times in the corresponding saturated pure partial planes and we are then back to the previous case where 0,1,2 appear 7 times. We find that there are 30 SPPPs while none of these can achieve size 25. The results are shown in ``case1-2-phase2(SPPP).txt". \subsection{Exactly 2 points appear 7 times}\label{Res:01x7} In this case, we have only one phase. We assume that 0 and 1 appear 7 times and thus have the unique 13-lines starting configuration: $\{0,6k+1,6k+2,\ldots,6k+6\}$ where $k=0,1,\ldots,6$ and $\{1,k+7,k+13,k+19,k+25,k+31,k+37\}$ where $k=0,1,\ldots,5$. Actually, it can be easily seen that if we can add one more line $\{2,7,14,21,28,35,42\}$ while still keeping the uniqueness. We use the program for this 14-line starting configuration and with the starting list being the list of all possible lines that are compatible with the starting configuration. In Step \ref{Stp:Saturated}, we require that $c_3\geq c_4\geq c_5\geq c_6$, where $c_i$ is the number of times that $i$ appears. This requirement is valid because in our starting configuration, points 3,4,5,6 are equivalent under the symmetric group. In Step \ref{Stp:Check}, we check the lexicographical order as usual and we also require that all points except 0,1 must appear at most 5 times, by Lemma \ref{Lem:NoN}. Eventually, we get 2166 SPPPs while the maximum size is 23. The results are shown in ``case2-phase1(SPPP).txt". \subsection{Exactly one point, 0, appears 7 times; 1,2,3 appear 5 times and 4 appears at least 4 times}\label{Res:0x7,123x5} The reason that we do not do the case where exactly one point appears 7 times is largeness of our search space. Therefore, we restrict our attention to the case that one point appears 7 times while a lot of points appear 5 times. As before, $\{0,1,2,3,4,5,6\}$ is a line in our starting configuration. We have only two phases while the second phase has little work to do. \ \noindent\textbf{Phase 1} Let us determine the starting configuration. The first 7 lines are $\{0,6k+1,6k+2,\ldots,6k+6\}$ where $k=0,1,\ldots,6$ and the next 4 lines are $\{1,k+7,k+13,k+19,k+25,k+31,k+37\}$ where $k=0,1,2,3$. The next line, containing one of $2,3$ can also be uniquely added, which we assume to be $\{2,7,14,21,28,35,41\}$. We use our program for this 12-line starting configuration with the starting list being the list of all possible lines that start with one of $2,3,4$ and are compatible with the starting configuration. In Step \ref{Stp:End}, we no longer require $rl_0$ to be empty; instead, we check that if $c_2=c_3=5$ and $c_4=4$, where $c_i$ is the number of times that $i$ appears in $ppp_0$. Equivalently, this is to say that the size of $ppp_0$ is 22 by Lemma \ref{Lem:SumC} used on the first line. In Step \ref{Stp:Check}, we check the lexicographical order as usual and also make sure that no point except 1 can appear more than 5 times. The result from the program is 26 pure partial planes of size 22, presented in ``case4-phase1.txt". \ \noindent\textbf{Phase 2} The second phase is simply extending these pure partial planes to saturation. It turns out that some of them are already saturated and the others can be made saturated by appending one line. We get a 23 pure partial planes of size 22 or 23 in total, presented in ``case4-phase2(SPPP).txt". \subsection{Points 0,1,2,3,4 appear 5 times each}\label{Res:01234x5} Recall that we require $\{0,1,2,3,4,5,6\}$ to be our first line. Importantly, in this case, we do not actually require that no points appear 7 times, but rather, we require that point 0,1,2,3,4 appear exactly 5 times each. We will see from the results that actually no points can appear more than 5 times in all the saturated pure partial planes we get in the end. \ \noindent\textbf{Phase 1} As before, we can determine the first 9 lines uniquely. They are \begin{tabular}{|l|ccccccc|} \hline line 1 & 0&1&2&3&4&5&6\\ \hline line 2 & 0&7&8&9&10&11&12\\ \hline line 3 & 0&13&14&15&16&17&18\\ \hline line 4 & 0&19&20&21&22&23&24\\ \hline line 5 & 0&25&26&27&28&29&30\\ \hline line 6 & 1&7&13&19&25&31&32\\ \hline line 7 & 1&8&14&20&26&33&34\\ \hline line 8 & 1&9&15&21&27&35&36\\ \hline line 9 & 1&10&16&22&28&37&38\\ \hline \end{tabular} \ In this phase, we add two lines that start at point 2 to this 9-line starting configuration. Namely, we use our program with the starting list being all possible lines that include point 2 and are compatible with the 9-line starting configuration shown above. And in Step \ref{Stp:End}, we require the size of $ppp_0$ to be 11. In this way, we get 29 pure partial planes with size 11, used as starting configurations for our next phase. These starting configurations are presented in file ``case4-phase1.txt". \ \noindent\textbf{Phase 2} For each of the starting configuration with size 11 we just obtained, we use the program with the starting list being all possible lines that start at point 2,3,4 and are compatible with the starting configuration with 11 lines. In Step \ref{Stp:Check}, we make sure that 2,3,4 appear at most 5 times. And in Step \ref{Stp:End}, we require point 2,3,4 to appear exactly five times. In this phase, we get 30 pure partial planes with size 21. They are shown in file ``case4-phase2.txt". \ \noindent\textbf{Phase 3} For each of the 21-line starting configurations, we run our program with the starting list being all possible lines that are compatible with the starting configuration. In Step \ref{Stp:Check}, we make sure that no point can appear more than 5 times and in Step \ref{Stp:End}, we require that the list $rl_0$ is empty, meaning that we require the pure partial plane to be saturated. Interestingly, 18 of these 21-line starting configurations are already saturated and the rest of them cannot be made saturated without letting one of point 0,1,2,3,4 appear 7 times. All possible saturated pure partial planes in this case are shown in file ``case4-phase3(SPPP).txt". \subsection{Points 0,...,14 appear exactly 5 times and points 15,...,39 appear exactly 4 times}\label{Res:40} In this case, we focus our attention to the situations where there are 15 points appearing 5 times, 25 points appearing 4 times and 3 points not appearing at all. Also, we require that in each line, three points are from $0,\ldots,14$ and four points are from $15,\ldots,39$. The use of this case will become clear in Section \ref{Sec:Main} where we give the main theorem. By a simple counting formula, we know that if such pure partial plane exists, it must have size 25. For this case, we will need a different isomorphism testing function in order to differentiate between a point that appears 5 times and a point that appears 4 time. To do this, we simply add another vertex to our point-line-adjacency graphs (Definition \ref{Def:PLAG}), connect vertices $0,1,\ldots,14$ to it and use graph isomorphism testing for the new graphs. Notice that this different isomorphism testing function is used solely for this case. The first 5 lines starting at 0 can be uniquely determined. \begin{tabular}{|l|ccccccc|} \hline line 1 & 0&1&2&15&16&17&18\\ \hline line 2 & 0&3&4&19&20&21&22\\ \hline line 3 & 0&5&6&23&24&25&26\\ \hline line 4 & 0&7&8&27&28&29&30\\ \hline line 5 & 0&9&10&31&32&33&34\\ \hline \end{tabular} \ The next line starting at 1 must pair up with two points from $\{3,4,\ldots,14\}$. There are two possibilities: 1,3,5 or 1,3,11. Specifically, the lines are $\{1,3,5,27,31,35,36\}$ and $\{1,3,11,23,27,31,35\}$. \ \noindent\textbf{Phase 1} With these two possible starting configurations, we add 4 lines to them that start with 1. We run our program (separately for these two starting configurations) with all lines that start at 1, contain three points from $\{1,\ldots,14\}$ and four points from $\{15,\ldots,39\}$ and are compatible with the starting configuration. In Step \ref{Stp:End}, we check that $ppp_0$ has size 9. Here, we get a total of 13 pure partial planes of size 9, presented in ``case5-phase1.txt". \ \noindent\textbf{Phase 2} This phase exists because we want to save some running time. We add just 1 compatible line that starts at 2 to the starting configurations. Then we get a total of 620 pure partial planes of size 10, shown in ``case5-phase2.txt". \ \noindent\textbf{Phase 3} We run our program with all lines that are compatible with the starting configuration, contain three points from $\{2,\ldots,14\}$ and four points from $\{15,\ldots,39\}$. In Step \ref{Stp:Check}, we make sure that points $2,\ldots,14$ never appear more than 5 times and points $15,\ldots,39$ never appear more than 4 times. In Step \ref{Stp:End}, we check that if our pure partial plane has size 25. Finally, we get a single pure partial plane of size 25, shown in ``case5-phase3(SPPP).txt". In fact, it must be saturated and we will explain this in next Section. \subsection{Summary}\label{Sec:Summary} For all these 5 cases described above, we find no pure partial planes of size 26 or greater. We find a total of 4 pure partial planes of size 25: three from Section \ref{Res:2x7} and one from Section \ref{Res:40}. We will list all of them in Appendix \ref{Sec:allPPP} for clarity. \section{Main Theorem}\label{Sec:Main} In this section, we restate our main theorem and finish the rest of the proof. \begin{theorem*}[\ref{thm:main}] The maximum size of a pure partial plane of order 6 is 25. Furthermore, all pure partial planes of size 25 are listed in Appendix \ref{Sec:allPPP}. \end{theorem*} \begin{proof} Essentially, we want to show that there are no pure partial planes of size 25 or greater outside our search. To do this, we restrict our attention to saturated pure partial planes instead of pure partial planes in general because we want to use Lemma \ref{Lem:NoN}. Assume that there exists a saturated pure partial plane $A$ of size $s$ with $s\geq25$. Specifically, assume that $A$ is a saturated pure partial plane that is not mentioned in our search in Section \ref{Sec:Result}. Define $a_i$ to be the number of points that appear $i$ times in $A$. Since we have already searched all possible cases for $a_7\geq2$ (Section \ref{Sec:Result}), now we assume that $a_7\leq1$. Use $c_i$ to denote the number of times that point $i$ appears in $A$. In other words, $c_i$ is the number of lines in $A$ that contain point $i$. Lemma \ref{Lem:SumA} gives us the following useful equations, with $n=6$: \begin{align*} 7a_7+5a_5+4a_4+3a_3+2a_2+a_1=&7s,\\ 49a_7+25a_5+16a_4+9a_3+4a_2+a_1=&s^2+6s. \end{align*} Notice that according to Lemma \ref{Lem:NoN}, $a_6=0$ so we ignore this term. In the above two equations, we subtract the second one by 5 times the first one, in order to get rid of $a_5$, which is potentially the largest term. We then divide this equation by 2. Together with the first equation, we have \begin{equation}\label{Eqn:1} 7a_7+5a_5+4a_4+3a_3+2a_2+a_1=7s, \end{equation} \begin{equation}\label{Eqn:11} 2a_4+3a_3+3a_2+2a_1=\displaystyle{\frac{29s-s^2}{2}}+7a_7. \end{equation} \noindent\textbf{Case 1:} $a_7=0$ and $s\geq 26$. For any line $\{i_1,\ldots,i_7\}$ of $A$, according to Lemma \ref{Lem:SumC}, we have $c_{i_1}+\cdots+c_{i_7}=s+6\geq32$. $a_7=0$ means that $c_{i_k}\leq5$ for $k=1,\ldots,7$. Also, according to the search result from Section \ref{Res:01234x5}, we have already covered the cases where there are at least 5 points that appear 5 times in a line. So we know that at most 4 of $c_{i_1},\ldots,c_{i_7}$ can be 5. Then $s+6\leq 5+5+5+5+4+4+4$ so $s\leq 26$. There is only one possibility now: $s=26$, 4 of $c_{i_1},\ldots,c_{i_7}$ equal 5 and the other 3 equal 4. In other words, $a_k>0$ only when $k=4,5$. Equation \eqref{Eqn:1} and \eqref{Eqn:11} become $5a_5+4a_4=182$, $2a_4=39$. It clearly does not have integer solutions. \ \noindent\textbf{Case 2:} $a_7=0$ and $s=25$. For any line $\{i_1,\ldots,i_7\}$ of $A$, according to Lemma \ref{Lem:SumC}, we have $c_{i_1}+\cdots+c_{i_7}=s+6=31$. Similarly as above, since we have already searched for cases where at least 5 points in this line appear a total of 5 times, there are these two possibilities left for $c_{i_1},\ldots,c_{i_7}$: $5,5,5,5,4,4,3$ and $5,5,5,4,4,4,4$. Thus, we must have that $a_1=a_2=0$. Equation \eqref{Eqn:1} and \eqref{Eqn:11} become $5a_5+4a_4+3a_3=175$, $2a_4+3a_3=50$. The second one gives that $a_3$ is a multiple of $2$. We subtract the first equation by two times the second equation and get $5a_5-3a_3=75$ and it gives that $a_3$ is a multiple of 5. Thus, $a_3$ is a multiple of 10. Since a point that appears 3 times must be contained in 3 lines of the form 5,5,5,5,4,4,3, we then know that $3a_3\leq 25$. These arguments give us $a_3=0$. By solving the equations, we get $a_5=15$ and $a_4=25$. Further, every line has the form $5,5,5,4,4,4,4.$ This is exactly the case we considered in Section \ref{Res:40}. \ \noindent\textbf{Case 3:} $a_7=1$. We can assume that 0 appears 7 times and the lines that contain 0 are $\{0,6k+1,6k+2,\ldots,6k+6\}$ where $k=0,1,\ldots,6$. Therefore, we can see that all points have appeared at least once so $a_0=0$. We thus have an additional equation $a_7+a_5+a_4+\cdots+a_1=n^2+n+1=43$. If $a_1\neq0$, without loss of generality, we assume that point 6 appears $1$ time, meaning $c_6=1$. According to Lemma \ref{Lem:SumC}, $c_0+c_1+\cdots+c_6=s+6\geq31$. So $c_1+c_2+c_3+c_4+c_5\geq23$ with $c_i\leq 5$ for $i=1,\ldots,5$. Therefore, either at least four of $c_1,c_2,c_3,c_4,c_5$ have value 5 or at least three of them have value 5 and a fourth one have value at least 4. These situations is covered in our computer search in Section \ref{Res:0x7,123x5}. So we then assume that $a_1=0$. If $a_2\neq0$, assume that point 6 appears in two lines. At least one of these two lines won't contain point 0 since they intersect at point 6 already. Suppose that this line is $6,j_1,j_2,\ldots,j_6$. Then by Lemma \ref{Lem:SumC}, $c_6+c_{j_1}+\cdots+c_{j_6}\geq31$. Since $c_6=2$ and $c_{j_i}\leq5$ for $i=1,\ldots,6$, at least five of $c_{j_1},\ldots,c_{j_6}$ must be 5. This situation is covered in our computer search in Section \ref{Res:01234x5}. So we then assume that $a_2=0$. Simplify Equations \eqref{Eqn:1}, \eqref{Eqn:11} and together with the new equation, we now have \begin{equation}\label{Eqn:2} 5a_5+4a_4+3a_3=7s-7, \end{equation} \begin{equation}\label{Eqn:3} 2a_4+3a_3=\frac{29s-s^2}{2}+7, \end{equation} \begin{equation}\label{Eqn:4} a_5+a_4+a_3=42. \end{equation} Manipulating the equations by $3\cdot\text{Equation }\eqref{Eqn:2}+2\cdot\text{Equation }\eqref{Eqn:3}-15\cdot\text{Equation }\eqref{Eqn:4}$, we get $$a_4=50s-s^2-637=-12-(s-25)^2<0,$$ a clear contradiction. Therefore, there are no saturated pure partial planes of order 6 and size at least 25 that are outside of our search. \end{proof} \section{Appendix}\label{Sec:Appendix} \subsection{All Pure Partial Planes of Order 6 and Size 25}\label{Sec:allPPP} Here is a list of all (saturated) pure partial planes of order 6 and size 25, up to isomorphism. \ \noindent\{\{0, 1, 2, 3, 4, 5, 6\}, \{0, 7, 8, 9, 10, 11, 12\}, \{0, 13, 14, 15, 16, 17, 18\}, \{0, 19, 20, 21, 22, 23, 24\}, \{0, 25, 26, 27, 28, 29, 30\}, \{0, 31, 32, 33, 34, 35, 36\}, \{0, 37, 38, 39, 40, 41, 42\}, \{1, 7, 13, 19, 25, 31, 37\}, \{1, 8, 14, 20, 26, 32, 38\}, \{1, 9, 15, 21, 27, 33, 39\}, \{1, 10, 16, 22, 28, 34, 40\}, \{1, 11, 17, 23, 29, 35, 41\}, \{1, 12, 18, 24, 30, 36, 42\}, \{2, 7, 14, 21, 28, 35, 42\}, \{2, 8, 13, 22, 27, 36, 41\}, \{2, 9, 16, 23, 30, 31, 38\}, \{2, 10, 15, 24, 29, 32, 37\}, \{2, 11, 18, 19, 26, 34, 39\}, \{2, 12, 17, 20, 25, 33, 40\}, \{3, 7, 15, 20, 30, 34, 41\}, \{3, 9, 14, 19, 29, 36, 40\}, \{3, 10, 13, 23, 26, 33, 42\}, \{4, 7, 18, 22, 29, 33, 38\}, \{4, 11, 13, 21, 30, 32, 40\}, \{4, 12, 14, 23, 27, 34, 37\}\}. \ \noindent\{\{0, 1, 2, 3, 4, 5, 6\}, \{0, 7, 8, 9, 10, 11, 12\}, \{0, 13, 14, 15, 16, 17, 18\}, \{0, 19, 20, 21, 22, 23, 24\}, \{0, 25, 26, 27, 28, 29, 30\}, \{0, 31, 32, 33, 34, 35, 36\}, \{0, 37, 38, 39, 40, 41, 42\}, \{1, 7, 13, 19, 25, 31, 37\}, \{1, 8, 14, 20, 26, 32, 38\}, \{1, 9, 15, 21, 27, 33, 39\}, \{1, 10, 16, 22, 28, 34, 40\}, \{1, 11, 17, 23, 29, 35, 41\}, \{1, 12, 18, 24, 30, 36, 42\}, \{2, 7, 14, 21, 28, 35, 42\}, \{2, 8, 13, 22, 27, 36, 41\}, \{2, 9, 16, 23, 30, 31, 38\}, \{2, 10, 15, 24, 29, 32, 37\}, \{2, 11, 18, 19, 26, 34, 39\}, \{2, 12, 17, 20, 25, 33, 40\}, \{3, 7, 15, 20, 30, 34, 41\}, \{3, 9, 14, 19, 29, 36, 40\}, \{3, 10, 13, 23, 26, 33, 42\}, \{4, 7, 18, 23, 27, 32, 40\}, \{4, 11, 14, 22, 30, 33, 37\}, \{4, 12, 13, 21, 29, 34, 38\}\}. \ \noindent\{\{0, 1, 2, 3, 4, 5, 6\}, \{0, 7, 8, 9, 10, 11, 12\}, \{0, 13, 14, 15, 16, 17, 18\}, \{0, 19, 20, 21, 22, 23, 24\}, \{0, 25, 26, 27, 28, 29, 30\}, \{0, 31, 32, 33, 34, 35, 36\}, \{0, 37, 38, 39, 40, 41, 42\}, \{1, 7, 13, 19, 25, 31, 37\}, \{1, 8, 14, 20, 26, 32, 38\}, \{1, 9, 15, 21, 27, 33, 39\}, \{1, 10, 16, 22, 28, 34, 40\}, \{1, 11, 17, 23, 29, 35, 41\}, \{1, 12, 18, 24, 30, 36, 42\}, \{2, 7, 14, 21, 28, 35, 42\}, \{2, 8, 13, 22, 27, 36, 41\}, \{2, 9, 17, 19, 30, 32, 40\}, \{2, 10, 18, 23, 25, 33, 38\}, \{2, 11, 16, 24, 26, 31, 39\}, \{2, 12, 15, 20, 29, 34, 37\}, \{3, 7, 15, 23, 26, 36, 40\}, \{3, 8, 17, 24, 28, 33, 37\}, \{3, 11, 13, 21, 30, 34, 38\}, \{4, 7, 16, 20, 30, 33, 41\}, \{4, 8, 18, 21, 29, 31, 40\}, \{4, 12, 13, 23, 28, 32, 39\}\}. \ \noindent\{\{0, 1, 2, 15, 16, 17, 18\}, \{0, 3, 4, 19, 20, 21, 22\}, \{0, 5, 6, 23, 24, 25, 26\}, \{0, 7, 8, 27, 28, 29, 30\}, \{0, 9, 10, 31, 32, 33, 34\}, \{1, 3, 5, 27, 31, 35, 36\}, \{1, 4, 6, 28, 32, 37, 38\}, \{1, 11, 12, 19, 23, 29, 33\}, \{1, 13, 14, 20, 24, 30, 34\}, \{2, 7, 9, 19, 24, 35, 37\}, \{2, 8, 10, 20, 23, 36, 38\}, \{2, 11, 13, 21, 25, 27, 32\}, \{2, 12, 14, 22, 26, 28, 31\}, \{3, 7, 11, 15, 26, 34, 38\}, \{3, 8, 14, 16, 25, 33, 37\}, \{3, 9, 13, 17, 23, 28, 39\}, \{4, 8, 11, 18, 24, 31, 39\}, \{4, 9, 12, 15, 25, 30, 36\}, \{4, 10, 13, 16, 26, 29, 35\}, \{5, 7, 12, 16, 20, 32, 39\}, \{5, 9, 14, 18, 21, 29, 38\}, \{5, 10, 11, 17, 22, 30, 37\}, \{6, 7, 13, 18, 22, 33, 36\}, \{6, 8, 12, 17, 21, 34, 35\}, \{6, 10, 14, 15, 19, 27, 39\}\}. \subsection{List of Files}\label{Sec:files} Here is a list of all the files that we provide for the project, under the folder ``cases". The case number of each file corresponds directly to the subsection number under Section \ref{Sec:Result} so we won't give redundant reference in the table. For each program, the ``input file" name is already written in the code. Each program will directly print the result that is supposed to be the same as what is written in the ``output file". The run time approximation is rough and serves as an upper bound. We also provide a file \verb|testcases.sh| to automatically test that the output files we provided are correct. The run time of \verb|testcases.sh| is supposed to be the sum of run times listed below. Readers should refer to \verb|README.txt| for more details. \ \begin{tabular}{|c|c|c|c|} \hline file name & input file & output file & run time \\ \hline \verb|case1-1-phase1.cpp| & \verb|case1-phase0.txt| & \verb|case1-1-phase1.txt| & 1 min \\ \hline \verb|case1-1-phase2.cpp| & \verb|case1-1-phase1.txt| & \verb|case1-1-phase2_SPPP.txt| & 1 min \\ \hline \verb|case1-2-phase1.cpp| & \verb|case1-phase0.txt| & \verb|case1-2-phase1.txt| & 1 min \\ \hline \verb|case1-2-phase2.cpp| & \verb|case1-2-phase1.txt| & \verb|case1-2-phase2_SPPP.txt| & 1 min \\ \hline \verb|case2-phase1.cpp| & \verb|case2-phase0.txt| & \verb|case2-phase1_SPPP.txt| & 10 days \\ \hline \verb|case3-phase1.cpp| & \verb|case3-phase0.txt| & \verb|case3-phase1.txt| & 10 hours \\ \hline \verb|case3-phase2.cpp| & \verb|case3-phase1.txt| & \verb|case3-phase2_SPPP.txt| & 1 min \\ \hline \verb|case4-phase1.cpp| & \verb|case4-phase0.txt| & \verb|case4-phase1.txt| & 1 hour \\ \hline \verb|case4-phase2.cpp| & \verb|case4-phase1.txt| & \verb|case4-phase2.txt| & 2 days \\ \hline \verb|case4-phase3.cpp| & \verb|case4-phase2.txt| & \verb|case4-phase3_SPPP.txt| & 2 min \\ \hline \verb|case5-phase1.cpp| & \verb|case5-phase0.txt| & \verb|case5-phase1.txt| & 1 day \\ \hline \verb|case5-phase2.cpp| & \verb|case5-phase1.txt| & \verb|case5-phase2.txt| & 1 min \\ \hline \verb|case5-phase3.cpp| & \verb|case5-phase2.txt| & \verb|case5-phase3_SPPP.txt| & 2 days \\ \hline \end{tabular} \subsection*{Acknowledgements} Thanks to Henry Cohn for supervising this project. \bibliographystyle{plain}

https://arxiv.org/abs/1605.05959

On circular flows: linear stability and damping

In this article we establish linear inviscid damping with optimal decay rates around 2D Taylor-Couette flow and similar monotone flows in an annular domain $B_{r_{2}}(0) \setminus B_{r_{1}}(0) \subset \mathbb{R}^{2}$. Following recent results by Wei, Zhang and Zhao, we establish stability in weighted norms, which allow for a singularity formation at the boundary, and additional provide a description of the blow-up behavior.

\section{Introduction} \label{sec:introduction} In this article we consider the linear stability and long-time asymptotic behavior of circular flows in an annular domain $(x,y) \in B_{r_{2}}(0) \setminus B_{r_{1}}(0)$. Such two-dimensional flows can for example be established experimentally in rotating cylinders, where the rotation is sufficiently slow as to not cause a (three-dimensional) Taylor-Couette instability. In this setting, radial vorticities \begin{align} \begin{split} \omega(x,y)&=\omega(r), \\ v(x,y)&= \p_r \psi e_\theta= \begin{pmatrix} -y \\ x \end{pmatrix} \frac{\psi'(r)}{r}, \\ \psi''(r)+\frac{1}{r}\psi'(r)&=\omega(r), \end{split} \end{align} are stationary solutions of the incompressible 2D Euler equations. Considering a small perturbation to Taylor-Couette flow, \begin{align} \label{eq:1} \frac{\phi'(r)}{r}= A + \frac{B}{r^{2}}, \end{align} we observe in Figure~\ref{fig:TC} that for $B=0$, i.e. constant angular velocity, perturbations are rotated while keeping their shape. However, in the general case when $B \neq 0$, $\frac{\phi'(r)}{r}$ is strictly monotone and the perturbation is sheared in way reminiscent of plane Couette flow, as is depicted in Figure~\ref{fig:Mon}. This mixing behavior underlies the phenomenon of (linear) inviscid damping. \begin{figure}[h] \centering \includegraphics[width=0.4\linewidth,page=1]{pictures.pdf} \includegraphics[width=0.5\linewidth, page=2]{pictures.pdf} \caption{Transport with constant angular velocity. We consider the Taylor-Couette flow $r$ in an annulus. The time $1$ flow-lines are drawn as arrows. A perturbation initially concentrated on a line stays concentrated on a line. On the right this behavior is expressed in polar coordinates.} \label{fig:TC} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.4\linewidth, page=3]{pictures.pdf} \includegraphics[width=0.4\linewidth, page=4]{pictures.pdf} \includegraphics[width=0.4\linewidth, page=5]{pictures.pdf} \includegraphics[width=0.4\linewidth, page=6]{pictures.pdf} \caption{Transport by a \emph{monotone} flow. We consider the Taylor-Couette flow $r + \frac{1}{r}$, which we observe to be mixing. As time tends to infinity this mixing results in weak convergence to an averaged quantity.} \label{fig:Mon} \end{figure} Considering polar coordinates, the linearized Euler equations around these stationary solutions are given by \begin{align} \label{eq:polarLE} \begin{split} \dt f + U(r) \p_{\theta} f &= b(r) \p_{\theta} \phi, \\ (\p_{r}^{2}+\frac{1}{r}\p_{r}+\frac{1}{r^{2}} \p_{\theta}^{2})\phi&= f, \\ \p_{\theta}\phi|_{r=r_{1},r2_{2})}&=0, \\ (t,\theta,r) & \in \R \times \T \times [r_{1},r_{2}], \end{split} \end{align} where $U$ and $b$ are given by \begin{align*} U(r)&=\frac{\phi'(r)}{r}, \\ b(r)&=-\frac{1}{r}\p_{r} (\p_{r}^{2}\phi(r)+\frac{1}{r}\p_{r}\phi(r)), \end{align*} and $b(r)\equiv 0$ if and only if one considers Taylor-Couette flow, $U(r)=A+\frac{B}{r^2}$. As suggested by our notation, these equations share strong similarities with the linearized Euler equations around a shear flow $(U(y),0)$ in a plane finite periodic channel, $\T \times [0,1]$: \begin{align} \begin{split} \dt \omega + U(y) \p_{x} \omega - U''(y) \p_{x}\phi&=0, \\ (\p_{y}^{2}+\p_{x}^{2})\phi &= \omega, \\ \p_{x}\phi|_{y=0,1}&=0 , \\ (t,x,y) &\in \R \times \T \times [0,1]. \end{split} \end{align} Here, various different approaches have been used to study this and related settings. \begin{itemize} \item In \cite{stepin1995nonself}, Stepin studies the asymptotic stability of monotone shear flows using spectral methods. Under the assumption that the associated Rayleigh boundary value problem possesses no eigenvalues, he obtains an asymptotic description of the stream function and non-optimal decay rates. \item In \cite{Euler_stability}, Bouchet and Morita provide heuristic results which suggest that the algebraic decay rates of Couette flow should hold for general monotone flows as well. However, their methods are not rigorous and do not provide sufficient error and stability estimates, especially in higher Sobolev regularity, in order to prove decay with optimal rates. \item In \cite{Zill5} and \cite{Zill3}, the author establishes linear inviscid damping and scattering for monotone shear flows in an infinite and finite periodic channel. In the latter setting, we restrict to perturbations in $H^2 \cap H^{1}_0$ in order to obtain the optimal decay rates. Conversely, in the setting without vanishing Dirichlet boundary values, the sharp stability threshold is shown to be given by $H^{s},s=3/2$ due to asymptotic singularity formation at the boundary. \item In \cite{Zhang2015inviscid}, Wei, Zhang and Zhao follow similar methods as in \cite{stepin1995nonself} and establish linear inviscid damping with optimal decay rates for monotone shear flows under the condition of there being no embedded eigenvalues. In particular, they remove the requirement of vanishing Dirichlet data and note that, using the boundary conditions of the velocity field and Hardy's inequality, one may allow for some blow-up at the boundary and still attain optimal decay rates. \item In a seminal work \cite{bedrossian2013asymptotic}, \cite{bedrossian2015inviscid} Bedrossian and Masmoudi establish nonlinear inviscid damping for Couette flow in an infinite periodic channel. There perturbations are required to be extremely regular, more precisely of Gevrey 2 class, in order to control nonlinear resonances. In particular, due to the singularity formation at the boundary and the associated blow-up of relatively low Sobolev norms, the question of linear inviscid damping for settings with boundary remains open. \item In addition to the inviscid setting, Bedrossian, Germain and Masmoudi also consider Couette flow as a solution of the Navier-Stokes equation in a two and three-dimensional infinite periodic channel. There, in addition to inviscid damping, the interaction between the mixing and viscous behavior yields additional stabilization by enhanced dissipation. Nonlinear inviscid damping is then established in Gevrey regularity \cite{bedrossian2015dynamics} and more recently in Sobolev regularity \cite{Bedrossian2015}, \cite{bedrossian2016sobolev}, where the threshold for stability results depends on $\nu>0$. \item In the circular setting, research has focused on instability results, such as Taylor-Couette instability, bifurcation and turbulence. For an introduction we refer to the book of Chossat and Iooss \cite{chossat2012couette}. \end{itemize} As the main results of this article we prove linear inviscid damping and scattering for a general class of circular flows, satisfying suitable monotonicity and smallness assumptions. In comparison to our previous results, we note the following changes and improvements: \begin{itemize} \item We obtain optimal decay rates also for perturbations without vanishing Dirichlet data. \item We show that $\p_yW$ splits into a bulk part $\Gamma$, which is stable also in unweighted higher Sobolev spaces, and a boundary correction $\beta$, which is stable in a suitably weighted $H^1$ space, but exhibits blow-up in $L^{\infty}$. \item The smallness condition is strongly reduced for results in higher regularity. \item In this circular setting, periodicity in $\theta$ is a natural condition, unlike in the setting of a plane periodic channel. \item We obtain a finer description of the boundary layer in terms of only the Dirichlet boundary values of the initial data. \end{itemize} \subsection{Main results} \label{sec:main-results} Our main results are summarized in the following theorem. \begin{thm}[Linear inviscid damping with optimal decay rates] \label{thm:main1} Let $0<r_1<r_2<\infty$ and let $U:(r_1,r_2)\rightarrow (a,b)$ be bilipschitz and suppose that $h(\cdot)=b(U^{-1}(\cdot)) \in W^{3,\infty}((a,b))$ and that $\|h\|_{W^{1,\infty}}$ is sufficiently small. Then, for any $f_{0}\in H^{-1}_{\theta}H^{2}_{r}$ there exists $v_{\infty}(r)$ such that the solution $f$ of~\eqref{eq:polarLE} satisfies \begin{align} \label{eq:4} \|v(t,\theta,r)- v_{\infty}(r) e_{\theta}\|_{L^{2}} \lesssim <t>^{-1}\|f_{0}\|_{H^{-1}_{\theta}H^{1}_{r}}, \\ \|v(t,\theta,r)e_{r}\|_{L^{2}} \lesssim <t>^{-2} \|f_{0}\|_{H^{-1}_{\theta}H^{2}_{r}}, \end{align} as $t \rightarrow \infty$. There exists $f_{\infty} \in L^{2}_{\theta}H^{1}_{r}$ such that \begin{align*} f(t,\theta-tU(r),r) \rightarrow f_{\infty} \text{ in } L^{2}, \end{align*} and \begin{align} \label{eq:5} \|f(t,\theta-tU(r), r) - f_{\infty}(\theta, r)\|_{L^{2}_{\theta,r}} \lesssim <t>^{-1}\|f_{0}\|_{H^{-1}_{\theta}H^{2}_{r}}. \end{align} \\ Furthermore, $f$ satisfies \begin{align*} \|f(t,\theta-rU(r),r)\|_{H^{-1}H^{1}}+ \|(r-r_{1})(r-r_{2})\frac{d^{2}}{dr^{2}}f(t,\theta-rU(r),r)\|_{H^{-1}H^{1}} \lesssim \|f_{0}\|_{H^{-1}H^{2}}. \end{align*} However, unless $bf|_{r=r_1,r_2}$ is constant, \begin{align*} \sup_{t \geq 0} \|f(t,\theta-rU(r),r)\|_{H^{-1}H^{s}}= \infty, \end{align*} for any $s>3/2$. More precisely, there exists an (explicit) function $\nu(t,\theta,r)$ determined solely by $f_0|_{r=r_1,r_2}$ and $U$ such that \begin{align*} \|\frac{d^2}{dr^2} f(t,\theta-rU(r),r)- \nu \|_{L^2L^2} \leq \|f_0\|_{L^2H^2}, \end{align*} and such that \begin{align*} \|(r-r_1)(r-r_2)\nu\|_{L^2} \leq |f_0|_{r=r_1,r_2}|. \end{align*} \end{thm} \begin{rem} \begin{itemize} \item While $h=b(U^{-1})$ is required to be regular, the smallness assumption is only imposed on the $W^{1,\infty}$ norm. \item This theorem summarizes the main results of Proposition ~\ref{prop:damping} and Theorems~\ref{thm:L2},~\ref{thm:StabilityofGamma} and ~\ref{thm:Stabilityofbeta} in terms of common norms in the variables $t,\theta,r$. In Section~\ref{sec:chang-vari-auxil} we introduce a \emph{scattering formulation}, which is used throughout the article. \item The function $\nu$ is introduced in Section~\ref{sec:it-works}. \item In~\cite{Zhang2015inviscid} it has been observed that, by a use of Hardy's inequality, the second derivative of $W$ can be allowed to form a singularity as $t \rightarrow \infty$ while still attaining the optimal $t^{-2}$ decay rate. Here, we stress that stability in $H^2$ indeed does not hold due to singularity formation at the boundary as $t\rightarrow \infty$, as quantified in $\nu$ and $\beta$ (c.f. Section \ref{sec:Higher stability}). \item As we discuss in Section~\ref{sec:chang-vari-auxil}, our method of proof does not rely on cancellations or conserved quantities. Hence, the results extend to complex-valued $b(r)$ and various modified equations in a straightforward manner. In the case of the linearized Euler equations in a plane finite periodic channel, however, Wei, Zhang and Zhao~\cite{Zhang2015inviscid} have shown, using different methods, that weaker assumptions suffice to obtain damping. \end{itemize} \end{rem} Similarly to~\cite{Zill3} our strategy is to first establish the damping and scattering result, \emph{assuming} stability in higher Sobolev norms. We stress that the damping estimate necessarily loses regularity. Hence, usual Duhamel fixed point iteration approaches or energy methods can not yield stability results. Instead we employ a finer study of the damping mechanism, which allows us to construct a Lyapunov functional using the mode-wise decay to avoid the necessary loss of regularity of uniform damping estimates. The remainder of the article is organized as follows: \begin{itemize} \item In Section~\ref{sec:damping-mixing}, we show that regularity of the vorticity in coordinates moving with the flow can be exchanged for uniform damping estimates and that the problem of linear inviscid damping thus reduces to a stability problem. As motivating examples, we discuss the specific cases of Taylor-Couette flow, a point vortex and of Couette flow in a plane channel, where explicit solutions are available and, in a sense, trivial. \item In Section~\ref{sec:chang-vari-auxil}, we introduce several reductions and changes of variables to arrive at a \emph{scattering formulation} of the linearized Euler equations. Subsequently, we analyze the structure of the equation and establish $L^{2}$ stability. \item Section~\ref{sec:Higher stability} considers higher regularity and singularity formation at the boundary. Compared to~\cite{Zill4}, in addition to considering a circular setting, we introduce a splitting $\p_{y}W=\Gamma+\beta$, where $\Gamma$ is shown to be stable in higher regularity, regardless of Dirichlet boundary data. On the other hand, $\beta$ is determined solely by the underlying circular flow and the Dirichlet boundary data of the initial perturbation and provides an explicit characterization of the boundary layer. Subsequently, we further split $\p_y\beta$ to obtain an explicit characterization of the $H^2$ blow-up in the form of $\nu$ and stability in weighted spaces. Here, we rely on a new approach based on Duhamel's principle and an iterative estimate in order to control the evolution of the weighted quantities. \item The Appendices~\ref{sec:auxil-funct-bound} and~\ref{sec:Duhamel} provide a description of boundary evaluations for elliptic ODEs and a variant of Duhamel's formula adapted to a time-dependent right-hand-side of the equation. \end{itemize} \section{Damping by mixing, the role of regularity and examples} \label{sec:damping-mixing} As in the case of inviscid damping in a plane channel or Landau damping, decay of the velocity/force field and regularity of the solution in a coordinate system moving with the flow are closely linked. More precisely, in this section we show that uniform damping estimates closely correspond to a control of the regularity of \begin{align*} W(t,\theta,r):=f(t,\theta-tU(r),r) \end{align*} with respect to $r$ and that such a control is necessary. The problem of linear inviscid damping with optimal decay rates thus turns out to be a stability problem, studied in Section \ref{sec:Higher stability}, which is the main focus of this article. \\ We consider the linearized Euler equations \begin{align} \label{eq:polarLE_3} \begin{split} \dt f + U(r) \p_{\theta} f &= b(r) \p_{\theta} \phi, \\ (\p_{r}^{2}+\frac{1}{r}\p_{r}+\frac{1}{r^{2}} \p_{\theta}^{2})\phi&= f, \\ \p_{\theta}\phi|_{r=r_{1},r_{2}}&=0, \\ (t,\theta,r) & \in \R \times \T \times [r_{1},r_{2}], \end{split} \end{align} as a perturbation around the transport problem \begin{align} \label{eq:transport_r} \begin{split} \dt g + U(r) \p_{\theta} g &=0, \\ (t,\theta,r) & \in \R \times \T \times [r_{1},r_{2}]. \end{split} \end{align} Based on this view, we measure the deviation of these equations by introducing the \emph{scattered vorticity} \begin{align} W(t,\theta,r):= f(t,\theta-tU(r),r). \end{align} \emph{Assuming} regularity of $W$ uniformly in time, damping results for \eqref{eq:polarLE_3} then reduce to estimates for \eqref{eq:transport_r}. Here, we it has recently been observed by Wei, Zhang and Zhao ~\cite{Zhang2015inviscid} that quadratic decay rates only require control of a weighted $H^{2}$ norm \begin{align} \label{eq:9} \|W\|_{H^{1}}+ \|y (1-y)\p_{y}^{2}W\|_{L^{2}} \end{align} by using a Hardy inequality in the duality estimate. The following two propositions provide damping estimates in terms of regularity of $W$ in the case of a plane channel and a circular domain, respectively. \begin{prop}[Damping by regularity for plane channel \cite{Zhang2015inviscid}, \cite{Zill4}, \cite{Lin-Zeng}] \label{prop:damping_plane} Let $-\infty\leq a<b \leq \infty$ and let $U:(a,b) \rightarrow \R$ be locally $C^1$ and suppose that $U'(y)\neq 0$ for almost every $y\in (a,b)$. Let $W\in H^{-1}_xH^{1}_y(\T \times (a,b))$ with $\int_{\T}W dx=0$ and let \begin{align*} \omega(t,x,y)=W(t,x-tU(y),y). \end{align*} Let further the associated velocity field $v$ be defined by \begin{align*} v_1&=-\p_y\phi,\\ v_2&=\p_x\phi,\\ \Delta \phi &=\omega, \\ \p_x\phi|_{y=a,b}&=0, \\ \phi &\in \dot H^1. \end{align*} Then $v$ satisfies \begin{align} \label{eq:t1plane} \|v(t)\|_{L^2} &\lesssim \min \Big( \left\|W\right\|_{H^{-1}_{x}L^{2}_{y}}, t^{-1}\left\|\frac{W}{U'}\right\|_{H^{-1}_{x}H^{1}_{y}}, \\ & \quad t^{-1} \left(\left\|W|\frac{1}{U'}| + W|\p_y\frac{1}{U'}|\right\|_{H^{-1}_{x}L^{2}_{y}}+ \left\|(y-a)(y-b)\frac{\p_yW}{U'}\right\|_{H^{-1}_{x}L^{2}_{y}}\right)\Big). \end{align} Furthermore, suppose that $\p_{y}^2 W$ exists. Then, $v_2$ additionally satisfies \begin{align} \label{eq:t2plane} \begin{split} \|v_2(t)\|_{L^2} &\leq t^{-2} \Big(\left\| \frac{W}{(U')^2}\right\|_{H^{-1}_{x}H^{1}_{y}} + \left\| W\p_y\frac{1}{(U')^2}\right\|_{H^{-1}_{x}H^{1}_{y}}\\ & \quad + \min\left(\left\|\frac{(y-a)(y-b)}{(U')^2}\p_{y}^2W\right\|_{H^{-1}_{x}L^{2}_{y}},\left\|\frac{\p_{y}^2W}{(U')^2}\right\|_{H^{-1}_{x}L^{2}_{y}}\right)\Big). \end{split} \end{align} \end{prop} \begin{proof}[Proof of Proposition \ref{prop:damping_plane}] We note that, by integration by parts, \begin{align*} \|v\|_{L^2}^2=\|\nabla \phi\|_{L^2}^2= -\iint \phi \omega dx dy. \end{align*} Applying Plancherel's theorem with respect to $x$ and noting that, \begin{align*} \mathcal{F}_x (\omega(t,\cdot,y))(k)= e^{iktU(y)} \hat{W}(t,k,y), \end{align*} this equals \begin{align*} \sum_{k \neq 0}\int \overline{\hat{\phi}}(t,k,y) e^{iktU(y)} \hat{W}(t,k,y). \end{align*} Integrating \begin{align*} e^{iktU}= \frac{1}{iktU'}\p_y e^{iktU(y)} \end{align*} by parts, we further obtain \begin{align*} \sum_{k \neq 0} \frac{1}{t}\int e^{iktU(y)}\p_y \left(\overline{\hat{\phi}} \frac{\hat{W}}{ikU'}\right), \end{align*} which is controlled by \begin{align*} t^{-1} \|\phi\|_{L^2H^1} \left\|\frac{W}{U'}\right\|_{H^{-1}H^1}. \end{align*} The estimate \eqref{eq:t1plane} thus follows by noting that \begin{align*} \|\phi\|_{L^2H^1} \leq \|v\|_{L^2}. \end{align*} \\ In order to prove \eqref{eq:t2plane}, we note that \begin{align*} \Delta v_2= \p_x \omega \end{align*} and define $\psi$ s.t. \begin{align*} \Delta \psi&=v_2, \\ \p_x\psi|_{y=a,b}&=0, \\ \psi &\in \dot H^{1}. \end{align*} Then, using integration by parts, we obtain \begin{align*} \|v\|_{L^2}^2&= \iint \psi \p_x \omega= \sum_{k \neq} \overline{\hat{\psi}} e^{iktU(y)} ik\hat{W} dy \\ &= \frac{1}{t^2} \sum_{k \neq 0} \int e^{iktU(y)}\p_y \left( \frac{1}{U'} \p_y \left(\frac{1}{U'} \overline{\hat{\psi}} \frac{\hat{W}}{k}\right)\right) dy \\ & \quad + \frac{1}{t^2} \sum_{k \neq 0} e^{iktU(y)}\frac{1}{U'} \p_y \left(\frac{1}{U'} \overline{\hat{\psi}} \frac{\hat{W}}{k}\right) \Big|_{y=a}^b \\ \end{align*} The result hence follows by the Cauchy-Schwarz inequality, the trace map and by using the estimates \begin{align*} \|\phi\|_{H^1} &\lesssim \|v_2\|_{L^2}, \\ \|\frac{\phi}{(y-a)(y-b)}\|_{L^2} &\lesssim \|\phi\|_{H^1}, \end{align*} The first estimate here follows by standard elliptic regularity theory, while the second one is given Hardy's inequality, as observed in~\cite{Zhang2015inviscid}. \end{proof} The following proposition adapts these results to the setting of circular flows. \begin{prop}[Damping for circular flows;~\cite{Zhang2015inviscid},~\cite{Zill4},~\cite{Lin-Zeng}] \label{prop:damping} Let $0<r_1<r_2<\infty$ and let $U:(r_1,r_2)\rightarrow \R$ be locally $C^1$ with $U'(r)\neq 0$ for almost every $r \in (r_1,r_2)$. Let $\frac{W(t,\theta,r)}{U'} \in H^{-1}H^1(\T \times (r_1,r_2), r dr d\theta)$ with $\int_\T W d\theta=0$ and let \begin{align*} f(t,\theta,r)=W(t,\theta-tU(r),r). \end{align*} Let further the associated velocity field be by defined by \begin{align} \begin{split} v_{r}(t,\theta,r)&= \frac{1}{r}\p_{\theta} \phi(t,\theta,r), \\ v_{\theta}(t,\theta,r)&= \p_{r} \phi(t,\theta,r), \\ (\p_{r}^{2}+ \frac{1}{r}\p_{r} +\frac{1}{r^{2}}\p_{\theta}^{2})\phi &=f, \\ \p_{\theta}\phi|_{r=r_{1},r_{2}}&=0, \\ v &\in L^2(r dr d\theta). \end{split} \end{align} Then $v$ satisfies \begin{align} \label{eq:t1circular} \|v(t)\|_{L^{2}(r dr d\theta)} &\lesssim \min \Big( \|W\|_{L^{2}(rdr d\theta)}, t^{-1} \|\frac{W(t)}{U'}\|_{H^{-1}_{\theta}H^{1}_{r}( r dr d\theta)}, \\ & \quad t^{-1} \left(\|W(t)|\frac{1}{U'}|+ W r|\p_r\frac{1}{U'}|\|_{H^{-1}_{\theta}L^{2}_{r}( r dr d\theta)} + \left\| \frac{(r-r_1)(r-r_2)\p_rW}{U'}\right\|_{H^{-1}L^2(r dr d\theta)}\right) \Big). \end{align} Furthermore, suppose that $\p_{r}^2W$ exists. Then, $v_r$ additionally satisfies \begin{align} \|v_{r}(t)\|_{L^{2}(r dr \theta)} &\lesssim t^{-2} \Big(\left\| \frac{W}{(U')^2}\right\|_{H^{-1}H^1(r dr d\theta)} + \left\| W \p_r \frac{1}{(U')^2}\right\|_{H^{-1}H^1( rdr d\theta)} \\ & \quad + \min \left( \left\| \frac{\p_{r}^2 W}{(U')^2}\right\|_{H^{-1}L^2( r dr d\theta)}, \left\| \frac{(r-r_1)(r-r_2)\p_{r}^2 W}{(U')^2}\right\|_{H^{-1}L^2( rdr d\theta)} \right)\Big). \end{align} \end{prop} \begin{rem} We note that for any given $0<r_1<r_2<\infty$ we could replace $r dr$ by just $dr$ in the above estimates at the cost of a constant $C(r_1,r_2)$. In this way the result and its proof can be made more similar to the setting of a finite channel. However, the above formulation also allows us to pass to the limits $r_1 \downarrow 0$ and $r_2 \uparrow \infty$. \end{rem} \begin{proof}[Proof of Proposition~\ref{prop:damping}] In order to obtain a more tractable stream function formulation of Euler's equations, in this proof we consider conformal coordinates, i.e. \begin{align*} (r,\theta)&=(e^s,\theta),\\ s \in (\log(r_1),\log(r_2))&=:(s_1,s_2). \end{align*} With respect to these coordinates, the stream function $\psi$ and the velocity field are given by \begin{align*} e^{-2s}(\p_{s}^2+ \p_{\theta}^2)\psi &= \omega, \\ v_r&=e^{-s}\p_\theta \psi, \\ v_\theta&= e^{-s}\p_s \psi. \end{align*} Furthermore, the kinetic energy satisfies \begin{align*} \int |v_r|^2 r dr d\theta = \int e^{-2s} |\p_\theta\psi|^2 e^{2s}ds d\theta &= \int |\p_\theta \psi|^2 ds d\theta, \\ \int |v_\theta|^2 r dr d\theta &= \int |\p_s \psi|^2 ds d\theta, \\ \int |v|^2 dr d\theta &= - \int \psi \omega e^{2s} ds d\theta. \end{align*} Applying a Fourier transform in $\theta$ and using the definition of $W$, we hence obtain \begin{align*} \int |\p_s \psi|^2 + |\p_\theta\psi|^2 ds d\theta = - \sum_{k} \int \overline{\hat{\psi}} e^{2s} e^{iktU(e^s)} \hat{W} ds. \end{align*} Integrating \begin{align*} e^{iktU(e^s)} = e^{-s} \frac{1}{iktU'(e^s)}\p_s e^{iktU(e^s)} \end{align*} by parts, we further compute \begin{align*} \int |\p_s \psi|^2 + |\p_\theta\psi|^2 ds d\theta= \sum_{k \neq 0} \int e^{iktU(e^s)} \p_s \left( \frac{1}{iktU'(e^s)} \overline{\hat{\psi}} e^{s} \hat{W}\right) ds. \end{align*} In order to estimate this integral, we use various different tools: \begin{itemize} \item If $\p_s$ does not fall on $\psi$, we control \begin{align*} \sum_{k \neq 0} \int |\hat{\psi} \frac{1}{k}X| ds \leq \|\psi\|_{L^2(ds d\theta)} \|X\|_{H^{-1}_{\theta}L^{2}_s(ds d\theta)} \end{align*} and use Poincar\'e's inequality to further estimate \begin{align*} \|\psi\|_{L^2(ds d\theta)} \leq C \|\p_\theta\psi\|_{L^2}. \end{align*} \item Alternatively, instead of Poincar\'e's inequality, duality yields an estimate by \begin{align*} \|\p_\theta \psi\|_{L^2(ds d\theta)} \|X\|_{H^{-2}_{\theta}L^{2}_s(ds d\theta)}. \end{align*} \item Since $\psi$ has zero boundary values, we can also use Hardy's inequality to control by \begin{align*} & \quad \|\frac{\psi}{(e^s-e^{s_1})(e^s-e^{s_2})}\|_{L^2(ds d\theta)} \|(e^s-e^{s_1})(e^s-e^{s_2}) X\|_{H^{-1}L^2(ds d\theta)} \\ &\leq \|\p_s \psi\|_{L^2(ds d\theta)} \|(e^s-e^{s_1})(e^s-e^{s_2}) X\|_{H^{-1}L^2(ds d\theta)}. \end{align*} \end{itemize} In the case of a fixed annulus $\T \times (r_1,r_2),$ $ 0<r_1<r_2<\infty,$ the precise choice of estimate is not essential. However, when considering a non-periodic setting, e.g. $\R \times [a,b]$, or a point vortex, i.e. $r_1=0$, or initial data with singularities at the boundary, all these estimates can yield improvements. \\ In order to obtain the quadratic decay estimate for $v_{r}=e^{-s}\p_\theta\psi$, we note that \begin{align*} (\p_{s}^2+\p_{\theta}^2)\p_\theta\psi= e^{2s}\p_{\theta}\omega. \end{align*} Thus, we define a potential $\gamma$ by \begin{align*} (\p_{s}^2 + \p_{\theta}^2)\gamma &= \p_\theta \psi, \\ \gamma|_{s=s_1,s_2}&=0,\\ \nabla \gamma &\in L^2, \end{align*} and compute \begin{align*} \int |v_r|^2 r dr d\theta &= \int |\p_\theta \psi|^2 ds d\theta \\ &= \int \gamma (\p_{s}^2+\p_{\theta}^2)\p_\theta\psi \\ &= \int \gamma e^{2s} \p_\theta \omega ds d\theta = \sum_{k\neq 0} ik \int \overline{\hat{\gamma}} e^{2s}e^{iktU(e^s)} \hat{W} ds. \end{align*} The result hence follows by integrating $e^{iktU(e^s)}$ by parts twice and using the Dirichlet data of $\gamma$ and $\p_\theta \psi$, the trace inequality and a variant of Hardy's inequality. That is, since $\gamma$ has zero Dirichlet boundary values, \begin{align*} \|\frac{\gamma}{(e^s-e^{s_1})(e^s-e^{s_2})}\|_{L^2(ds d\theta)} &\leq \|\frac{\gamma}{(s-s_1)(s-s_2)}\|_{L^2(ds d\theta)} \\ &\lesssim \|\p_s \gamma\|_{L^{2}(ds d\theta)}= \|v_r\|_{L^2(rdr d\theta)}. \end{align*} \end{proof} We stress that these uniform damping estimates necessarily lose regularity, since the associated change of coordinates is a unitary operator. Thus, the operator norm of $f \mapsto v$ considered as a mapping from $L^2$ to $L^2$ does not improve in time. Hence, it is not possible to derive stability of \eqref{eq:polarLE_3} using a common Duhamel-type approach or a fixed point mapping. Instead, in Sections \ref{sec:L2stability} and \ref{sec:Higher stability} we have to make use of finer properties of the dynamics and the mode-wise decay of the principal symbol of the evolution operator. Before that, in the following we discuss some examples for which explicit computations are possible. \subsubsection{Taylor-Couette flow} \label{sec:examples} As an application of the damping results, we discuss some exceptional cases for which $W$ can be trivially computed in terms of the initial datum. \begin{cor}[Couette flow] Let $U(y)=y$ on $\T \times [a,b]$ with $a,b \in [-\infty,\infty]$, then the linearized Euler equations reduce to the free transport equations. Furthermore, if $\omega_0 \in H^{-1}_xH^{2}_y$, then the associated velocity field satisfies \begin{align*} \|v(t)-\langle v|_{t=0} \rangle_x\|_{L^2}&\leq t^{-1}\|\omega_0\|_{H^{-1}H^1}, \\ \|v_2(t)\|_{L^2} &\leq t^{-2}\|\omega_0\|_{H^{-1}H^2}. \end{align*} \end{cor} \begin{cor}[Taylor-Couette flow; Point vortex] \label{cor:pointvortex} Let $A, B \in \R$ and let $0\leq r_1<r_2 \leq\infty$, then the linearized Euler equations around Taylor-Couette flow \begin{align*} U(r)= (Ar + \frac{B}{r})e_{\theta}, \end{align*} are given by \begin{align*} \dt f + (A+\frac{B}{r^2})\p_\theta f &=0, \text{ on } (0,\infty)\times \T \times (r_1,r_2) \\ f|_{t=0}&=f_0 \text{ on } \T \times (r_1,r_2). \end{align*} Furthermore, the associated velocity field $v$ satisfies \begin{align*} \|v\|_{L^2(rdrd\theta)} \leq C t^{-1}B^{-1} \|f_0\|_{H^{-1}H^1((r^7+r^5)drd\theta)}, \\ \|v_r\|_{L^2( rdr d\theta)} \leq C t^{-2}B^{-2} \|f_0\|_{H^{-1}H^2((r^7+r^5)drd\theta)}. \end{align*} Here, the case the case $r_1=0$, $A=0, B \neq 0$ corresponds to a \emph{point vortex}. \end{cor} \begin{proof}[Proof of Corollary \ref{cor:pointvortex}] We note that $(A+\frac{B}{r^2})'=-B \frac{2}{r^3}$. Hence, by direct computation \begin{align*} \|\frac{\omega_0}{U'}\|_{H^{-1}H^1(rdrd\theta)} + \|\frac{\omega_0}{U'}\|_{L^2(r^{-1} dr \theta)} \leq 2B^{-1} (\|\omega_0\|_{H^{-1}L^2((r^7+r^5) dr \theta)} + \|\p_r\omega_0\|_{H^{-1}L^2(r^7 dr \theta)}). \end{align*} \end{proof} \section{Scattering formulation and $L^2$ stability} \label{sec:chang-vari-auxil} As established in Section~\ref{sec:damping-mixing}, the core problem of (linear) inviscid damping consists of establishing a control of higher Sobolev norms of the vorticity moving with the flow: \begin{align} \label{eq:12} W(t,\theta,r):= f(t,\theta-tU(r),r). \end{align} Here, we largely follow a similar approach as in the plane setting considered in~\cite{Zill4}. As key improvements we obtain a less restrictive smallness condition and develop a splitting of $\p_{r}W$ into a well-behaved and more regular part $\Gamma$ and a (relatively) explicit boundary layer $\beta$. This then allows us to deduce damping with optimal decay rates and a detailed stability in suitable weighted Sobolev spaces, such as the ones considered in Proposition~\ref{prop:damping}. In order simplify our analysis, in this section we introduce several changes of variables as well as useful auxiliary functions. \subsection{Scattering formulation} \label{sec:scatt-form} Expressing the linearized Euler equations \begin{align} \label{eq:polarLE_2} \begin{split} \dt f + U(r) \p_{\theta} f &= b(r) \p_{\theta} \phi, \\ (\p_{r}^{2}+\frac{1}{r}\p_{r}+\frac{1}{r^{2}} \p_{\theta}^{2})\phi&= f, \\ \p_{\theta}\phi|_{r=r_{1},r_{2})}&=0, \\ (t,\theta,r) & \in \R \times \T \times [r_{1},r_{2}], \end{split} \end{align} in terms of the \emph{scattered} quantities \begin{align} \begin{split} F(t,\theta,r)&= f(t,\theta-tU(r),r), \\ \Upsilon(t,\theta,r)&= \phi(t,\theta-tU(r),r), \end{split} \end{align} we obtain \begin{align} \begin{split} \dt F &= b(r) \p_{\theta} \Upsilon, \\ ((\p_{r}-tU'(r)\p_{\theta})^{2}+\frac{1}{r}(\p_{r}-tU'(r)\p_{\theta})+\frac{1}{r^{2}} \p_{\theta}^{2})\Upsilon&= F, \\ \p_{\theta}\Upsilon|_{r=r_{1},r2_{2})}&=0, \\ (t,\theta,r) & \in \R \times \T \times [r_{1},r_{2}], \end{split} \end{align} As none of the coefficient functions depend on $\theta$, our system decouples with respect to Fourier modes $k$ in $\theta$. \begin{align} \begin{split} \dt \hat{F} &= b(r) ik \hat{\Upsilon}, \\ ((\p_{r}-iktU'(r))^{2}+\frac{1}{r}(\p_{r}-iktU'(r))-\frac{k^{2}}{r^{2}})\hat{\Upsilon}&= \hat{F}, \\ ik\hat{\Upsilon}|_{r=r_{1},r2_{2})}&=0, \\ (t,k,r) & \in \R \times 2\pi\Z \times [r_{1},r_{2}], \end{split} \end{align} We in particular note that the mode $k=0$, which corresponds to a purely circular flow, is conserved in time. Using the linearity of our equations, in the following we hence without loss of regularity consider $k \in 2\pi (\Z\setminus \{0\})$ as a given parameter. In view of the structure of the differential equation for $\Phi$, it is further advantageous to use that $U$, as a strictly monotone function, is invertible. Introducing a change of coordinates \begin{align} r \mapsto y=U(r). \end{align} as well a denoting \begin{align} \begin{split} h(y) &= \frac{(\omega_{0})'}{r}|_{r=U^{-1}(y)}, \\ g(y)&= U'(r)|_{r=U^{-1}(y)}, \\ W(t,y,k) &= \hat{F}(t,r,k)|_{r=U^{-1}(y)},\\ \Phi(t,y,k) &= \frac{1}{k^{2}}\hat{\Upsilon}(t,r,k)|_{r=U^{-1}(y)}, \end{split} \end{align} our system is then given by the following definition. \begin{defi}[Euler's equations in scattering formulation] \label{defi:SLE} Let $U:[r_1,r_2]\rightarrow \R$ be strictly monotone and let $h(y)=b|_{r=U^{-1}(y)}$ and $g=U'(U^{-1}(y))$. Then \emph{Euler's equations in scattering formulation} are given by \begin{align} \label{eq:SLE} \begin{split} \dt W = \frac{ih(y)}{k}\Phi &=: \frac{ih(y)}{k}L_{t}W, \\ \mathcal{E}_{t}\Phi:= \left(\left(g(y)(\frac{\p_{y}}{k}-it)\right)^{2}+ \frac{g(y)}{kr(y)}(\frac{\p_{y}}{k}-it)-\frac{1}{r^{2}(y)}\right) \Phi &=W, \\ \Phi|_{y=a,b}&=0, \\ (t,k,y) &\in \R \times 2\pi (\Z \setminus \{0\}) \times [a,b], \end{split} \end{align} where $a=\min(U^{-1}(r_{1}), U^{-1}(r_{2}))$, $b=\max(U^{-1}(r_{1}), U^{-1}(r_{2}))$ and $k \in 2 \pi (\Z \setminus \{0\})$. \end{defi} \begin{rem} \begin{itemize} \item Our methods do not rely on the specific form of $h$ or $g$ in terms of $U$. For example, we can allow for $h$ to be an arbitrary \emph{complex valued} $W^{1,\infty}$ function. \item Here the notation $L_{t}W$ is used to stress that the mapping $W \mapsto \Phi$ is a linear operator in $W$. \item As this system decouples with respect to $k$, we will often treat $k\neq 0$ as a fixed given external parameter and with slight abuse of notation use $W(t,y)$ to refer to $W(t,k,y)$ for the given $k$. \end{itemize} \end{rem} \subsection{Shifted elliptic regularity and modified spaces} \label{sec:modif-spac-glid} We note that in this scattering formulation $\mathcal{E}_{t}$ is obtained from an elliptic operator by conjugation with $e^{ikty}$ and hence define suitable replacements of the $H^1$ and $H^{-1}$ energies: \begin{defi}[ $\tilde{H}^{1}_{t}$ and $\tilde{H}^{-1}_{t}$ energies] Let $u \in H^{1}([a,b])$ and let $k \in 2\pi(\Z \setminus \{0\})$ be given, then for every $t \in \R$, we define \begin{align} \label{eq:tildeH1energy} \|u\|_{\tilde{H}^{1}_t}^2:= \|e^{ikty}u\|_{H^1}^2= \|u\|_{L^2}^2 + \|(\frac{\p_y}{k}-it)u\|_{L^2}^2. \end{align} Furthermore, we define a dual quantity in the following way. Let $v \in L^2$ and let $\Psi[v]$ be the unique solution of \begin{align*} (-1+(\frac{\p_y}{k}-it)^2) \Psi[v]&=v, \\ \Psi[v]|_{y=a,b}&=0. \end{align*} Then we define \begin{align*} \|v\|_{\tilde{H}^{-1}_t}:= \|\Psi[v]\|_{\tilde{H}^{1}_t}. \end{align*} \end{defi} \begin{lem}[Duality] \label{lem:dual} Let $W \in L^{2}$ and let $k \in (\Z\setminus\{0\})$ be given. Then \begin{align} \label{eq:13} \|W\|_{\tilde{H}^{-1}_{t}}=\sup \{ \langle W, \alpha \rangle_{L^{2}}: \alpha \in H^{1}_{0}, \|\alpha\|_{\tilde{H}^{1}}\leq 1 \}, \end{align} i.e. $\tilde{H}^{-1}_{t}$ is dual to $\tilde{H}^{1}_{t}$. \end{lem} \begin{proof}[Proof of Lemma~\ref{lem:dual}] Since multiplication by $e^{ikty}$ is a unitary operation and preserves zero Dirichlet boundary values and \begin{align*} \Psi_t[v]= e^{-ikty}\Psi_0[e^{ikty}v], \end{align*} it suffices to consider the case $t=0$, which is given by the usual $H^1$ and $H^{-1}$ norms (where we use $\frac{\p_y}{k}$ instead of $\p_y$). The result then follows using integration by parts: \begin{align} \label{eq:14} -\langle W, \alpha \rangle = \langle (1-\frac{\p_{y}}{k}^{2}) \Psi[W], \alpha \rangle \\ = \langle \Psi[W], \alpha \rangle + \langle \frac{\p_{y}}{k}\Psi[W], \frac{\p_{y}}{k}\alpha \rangle \leq \|W\|_{H^{-1}}\|\alpha\|_{H^{1}}, \end{align} with equality if $\alpha= -\frac{1}{\|\Psi[W]\|_{H^1}}\Psi[W]$. Taking the supremum over all $\alpha$ with $\|\alpha\|_{H^{1}}$ we hence obtain the result. \end{proof} \subsection{Heuristics and obstructions} \label{sec:heuristics} On a \emph{heuristic} level, in order to establish stability in $L^2$, we use that \begin{align*} \frac{d}{dt}\|W(t)\|_{L^2}^2 =2\Re \langle W, \frac{ih}{k}L_t W\rangle \lesssim C(h,k)\|W(t)\|_{\tilde{H}^{-1}_t}^2, \end{align*} and that for fixed functions $u \in L^2$, which \emph{do not depend on time}, \begin{align*} \int_{0}^{\infty} \|u\|_{\tilde{H}^{-1}_t}^2 dt \leq C \|u\|_{L^2}^2, \end{align*} as can be computed from a Fourier characterization. Hence, it seems reasonable to expect that solutions $W(t)$ of~\eqref{eq:SLE} satisfy an estimate of the form \begin{align*} \|W(t)\|_{L^2} \leq \exp(C\|h\|_{L^\infty}|k|^{-1}) \|f_0\|_{L^2}, \end{align*} also for complex valued $h$, which is the case for some explicit model problems (c.f.~\cite{Zill3}). However, we stress that this heuristic is very rough and does not account for several obstructions: \begin{itemize} \item We note that integrability in time in general fails for time-dependent $u \in L^{\infty}_t(L^2)$. For example, choosing \begin{align*} u(t,k,y)=e^{ikty}u_0(k,y), \end{align*} we observe that \begin{align*} \int_{0}^{T} \|u\|_{\tilde{H}^{-1}_t}^2 dt = T \|u_0\|_{H^{-1}}^2, \end{align*} which diverges as $T \rightarrow \infty$ despite $\|u(t,y)\|_{L^2}= \|u_0\|_{L^2}$ being uniformly bounded. \item Since the first estimate does not account for antisymmetric operators in $\frac{d}{dt}W$ it is not sufficient to establish $L^2$ stability. For example, this estimate is satisfied by solutions $u(t,y)$ to \begin{align*} \dt u +iy u &= \Phi, \\ (-1+(\p_y-it)^2)\Phi&=u, \\ \Phi|_{y=a,b}&=0. \end{align*} Considering $v(t,y)=e^{ity}u(t,y)$, we observe that $v$ solves \begin{align*} \dt v &= \phi, \\ (1-\p_{y}^2)\phi &= v, \\ \phi|_{y=a,b}&=0. \end{align*} Hence, choosing $u|_{t=0}$ to be an eigenfunction of $(1-\p_{y}^2)$, we obtain an \emph{exponentially growing} solution. \end{itemize} \subsection{$L^2$ stability} \label{sec:L2stability} As the main result of this section, we adapt the Lyapunov functional approach of~\cite{Zill5} to this circular setting and prove stability of~\eqref{eq:SLE}. In the following we formulate the main ingredients of our approach as a series of Lemmata, which are then used to prove $L^2$ stability in Theorem \ref{thm:L2}. Subsequently, we elaborate on the theorem's statement and assumptions in comparison to existing results and prove the lemmata. Here, the lemmata are formulated in a general way in order to facilitate their use for higher regularity estimates in later sections. \begin{lem} \label{lem:EllipticContribution} Let $L_{t}$ be given by~\eqref{eq:SLE} and let $\kappa \in W^{1,\infty}$. Then, for any $u,v \in L^{2}$ \begin{align} \label{eq:15} | \langle u, \kappa L_{t}v \rangle | \leq (\|\kappa\|_{L^\infty} + \frac{1}{|k|} \|\p_y\kappa\|_{L^\infty}) \|u\|_{\tilde{H}^{-1}_{t}}\|L_{t} v\|_{\tilde{H}^{1}_t} \end{align} \end{lem} \begin{lem} \label{lem:compCC} Let $L_{t}$ be as in~\eqref{eq:SLE}. Then there exists a constant $C=C(a,b,g)$ such that for any $u \in L^{2}$ and any $t\geq 0$ \begin{align} \label{eq:16} \|L_{t}u\|_{\tilde{H}^{1}_{t}} \leq C \|u\|_{\tilde{H}^{-1}_{t}}. \end{align} \end{lem} \begin{lem}[{\cite[Lemma 4.5]{Zill4}}] \label{lem:A} Let $u \in L^2([a,b])$ and let $\sum_{n \in (b-a)\N} u_{n} \sin(ny)$ be its series expansion. Define the symmetric, positive definite, non-increasing operator $A$ by \begin{align} \label{eq:17} \langle u, A u \rangle := \sum_{n} \exp(\arctan(\frac{n}{k}-t)) |u_{n}|^{2}. \end{align} Then $A$ is symmetric, positive definite, non-increasing, $C^{1}$ in time and comparable to the identity, i.e. \begin{align} \label{eq:18} e^{-\pi} \|u\|_{L^{2}} \leq \langle u, A u \rangle \leq e^\pi \|u\|_{L^{2}}, \end{align} for all $u \in L^{2}$. Furthermore, there exists a constant $e^{-\pi}\leq C_{2} \leq e^{\pi}$ and $\delta>0$ such that \begin{align} \label{eq:19} \|u\|_{\tilde{H}^{-1}_{t}}\|Au\|_{\tilde{H}^{-1}_{t}} \leq -C_{2} \langle u, \dot A u \rangle \end{align} \end{lem} Using the preceding lemmata, we can establish $L^{2}$ stability. \begin{thm}[$L^{2}$ stability] \label{thm:L2} Let $A$ and $C_{2}$ be given by Lemma~\ref{lem:A} and let $C$ be as in Lemma~\ref{lem:compCC}. Further suppose that there exists $\delta>0$ such that \begin{align} \label{eq:20} |k|^{-1}(\|h\|_{L^{\infty}}+|k|^{-1}\|\p_y h\|_{L^\infty}) \leq \frac{1}{C}(\frac{1}{C_2}-\delta). \end{align} Then for any solution $W$ to the Euler equations in scattering formulation~\eqref{eq:SLE} the functional \begin{align} \label{eq:21} I(t):= \langle W, A(t) W \rangle. \end{align} is non-increasing and satisfies \begin{align} \label{eq:22} \dt I(t) \leq -\delta \|W(t)\|_{\tilde{H}^{-1}_t}\|A(t)W(t)\|_{\tilde{H}^{-1}_t} \leq 0. \end{align} In particular, this implies \begin{align} \label{eq:23} e^{-\pi}\|W(t)\|_{L^{2}}^{2} \leq I(t) \leq I(0) \leq e^\pi \|\omega_{0}\|_{L^{2}}^{2}. \end{align} \end{thm} \begin{proof}[Proof of Theorem~\ref{thm:L2}] Using Lemma~\ref{lem:EllipticContribution}, we estimate \begin{align} \label{eq:24} \dt I(t) \leq \langle W, \dot A W \rangle + 2 |k|^{-1}(\|h\|_{L^{\infty}}+|k|^{-1}\|\p_y h\|_{L^\infty}) \|AW\|_{\tilde{H}^{-1}_{t}}\|L_tW\|_{\tilde{H}^{1}_{t}}. \end{align} Applying Lemma~\ref{lem:A} and Young's inequality, we further control \begin{align} \label{eq:25} \|AW\|_{\tilde{H}^{-1}_{t}}\|L_tW\|_{\tilde{H}^{-1}_{t}} \leq \frac{C}{2} \|AW\|_{\tilde{H}^{-1}_{t}} \|W\|_{\tilde{H}^{-1}_{t}}. \end{align} The result then follows by an application of Lemma~\ref{lem:A} and noting that, by our smallness assumption, \begin{align*} \dt I(t) \leq \langle W, \dot A W \rangle + |k|^{-1}(\|h\|_{L^{\infty}}+|k|^{-1}\|\p_y h\|_{L^\infty})C\|AW\|_{\tilde{H}^{-1}_{t}}\|W\|_{\tilde{H}^{-1}_{t}} \\ \leq \langle W, \dot A W \rangle + (\frac{1}{C_2}-\delta)\|AW\|_{\tilde{H}^{-1}_{t}} \|W\|_{\tilde{H}^{-1}_{t}} \\ \leq -\delta \|AW\|_{\tilde{H}^{-1}_{t}}\|W\|_{\tilde{H}^{-1}_{t}} \leq 0 \end{align*} \end{proof} Let us briefly remark on this result and its assumptions: \begin{itemize} \item We require a smallness condition on $\frac{ih}{k}L_t$ in order to rule out the obstacles mentioned in Section~\ref{sec:heuristics}. \item Since $h$ is allowed to be complex-valued, we do not rely on conserved quantities or classical stability results such as the ones of Rayleigh, Fjortoft or Arnold. \item In the setting of a plane finite periodic channel, in~\cite{Zhang2015inviscid} Wei, Zhang and Zhao use a spectral approach to establish linear stability and decay with optimal rates for monotone shear flows under the assumption that the strictly monotone shear flow $U(y)$ possesses no embedding eigenvalues. In comparison, our smallness assumption is more restrictive, but extends to related problems such as stability in fractional Sobolev spaces, complex valued functions $h$ and fractional operators $L_t$ in a straightforward way. \item In Section~\ref{sec:Higher stability}, we show that $\p_{y}^2W$ can be split into a very regular, stable part $\Gamma$ and a boundary layer part $\beta$ which develops a singularity at the boundary. Here, $\beta$ is determined solely by the Dirichlet boundary data of the initial datum, $\omega_0$, and allows for a detailed study of the stability properties of the evolution. \end{itemize} It remains to prove Lemmata~\ref{lem:EllipticContribution} and~\ref{lem:compCC}. \begin{proof}[Proof of Lemma~\ref{lem:EllipticContribution}] Let $u,v \in L^2$ and let $\Psi[u]$ be the unique solution of \begin{align} \label{eq:26} (-1+(\p_{y}-ikt)^{2})\Psi[u]&=u, \\ \Psi[u]|_{y=a,b}&=0. \end{align} Then we directly compute \begin{align} \label{eq:27} | \langle u, \kappa L_{t}v \rangle | = | \langle (1-(\frac{\p_{y}}{k}-ikt)^{2})\Psi[u], \kappa L_{t}v \rangle | \\ \leq \|\Psi[u]\|_{L^{2}} \|\kappa L_{t}\|_{L^{2}} + \|(\frac{\p_{y}}{k}-it)\Psi[u]\|_{L^{2}}\|(\frac{\p_{y}}{k}-it) \kappa L_{t}v\|_{L^{2}} \\ \leq \|\Psi[u]\|_{\tilde{H}^{1}_{t}}(\|\kappa\|_{L^\infty} + \frac{1}{|k|} \|\p_y\kappa\|_{L^\infty}) \|L_{t}v\|_{\tilde{H}^{1}_{t}}. \end{align} Here, we used that $L_tv$ by definition satisfies zero Dirichlet boundary conditions and hence no boundary contributions appear when integrating by parts. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:compCC}] We recall that $L_{t}u$ is the solution of \begin{align} \label{eq:28} \mathcal{E}_{t} L_{t} u= \left(\left(g(y)(\frac{\p_{y}}{k}-it)\right)^{2}+ \frac{g(y)}{kr(y)}(\frac{\p_{y}}{k}-it)-\frac{1}{r^{2}(y)}\right) L_{t}u&=0 , \\ L_{t}u|_{y=a,b}&=0, \end{align} and that $g(y)$ and $\frac{g(y)}{r(y)}$ are bounded from below (and above). Hence $\mathcal{E}_{t}$ is a shifted elliptic operator and testing by $L_{t}u$ (or $\frac{1}{g}L_{t}u$) we obtain that \begin{align} \label{eq:29} \|L_{t}u\|_{\tilde{H}^{1}_{t}}^{2} \leq -C \langle L_{t}u, u \rangle, \end{align} for some $C>0$. Applying Lemma~\ref{lem:dual}, we thus obtain \begin{align} \label{eq:30} \|L_{t}u\|_{\tilde{H}^{1}_{t}}^{2} &\leq C \|L_{t}u\|_{\tilde{H}^{1}_{t}}\|\Psi[u]\|_{\tilde{H}^{1}_{t}},\\ \Leftrightarrow \|L_{t}u\|_{\tilde{H}^{1}_{t}} &\leq C \|u\|_{\tilde{H}^{-1}_{t}}. \end{align} \end{proof} Having introduced the basic tools of our approach, in the following section we consider higher stability of $W$, i.e. control of $\p_{y}W$. Here, boundary effects qualitatively change the dynamics and necessitate a modification of the weight $A(t)$. \section{Higher stability and boundary layers} \label{sec:Higher stability} In this section we show that the $L^{2}$ stability result can be extended to higher Sobolev regularity. However, unlike in the setting of an infinite periodic channel, boundary effects can not be neglected and result in the formation of singularities. As the main improvements over our previous work for the plane channel in \cite{Zill4}, we provide an explicit splitting into a more regular good parts and a boundary layer exhibiting blow-up as well as an improved smallness condition. This splitting then also allows to provide a more detailed description of the blow-up also in weighted Sobolev spaces. For this purpose we also introduce a different method of proof. Let thus $W$ be a solution to \eqref{eq:SLE} \begin{align*} \dt W &= \frac{ih}{k}L_{t}W, \\ \mathcal{E}_{t} L_t W &=W, \\ L_t|_{y=a,b}&=0, \\ (t,k,y) &\in \R \times 2\pi (\Z \setminus \{0\}) \times [a,b]. \end{align*} We begin by studying $\p_yW$, which satisfies \begin{align} \label{eq:pyW} \dt \p_yW &= \frac{ih}{k}L_t \p_y W + \frac{ih'}{k}L_tW + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_tW + \frac{ih}{k}H^{(1)}, \\ \mathcal{E}_{t} H^{(1)}&=0 ,\\ H^{(1)}_{y=a,b}&=\p_yL_tW|_{y=a,b}. \end{align} In contrast to the $L^2$ setting (or a setting without boundary such as $\T \times \R$) we hence obtain a correction $H^{(1)}$ due to $\p_y L_tW$ not satisfying zero Dirichlet boundary conditions. As a main result of Appendix~\ref{sec:auxil-funct-bound}, we study the boundary behavior of $\p_yL_t$ (also confer~\cite{Zill4}) and obtain the following description of $H^{(1)}$: \begin{lem} Let $W$ be a solution of~\eqref{eq:SLE} and let $H^{(1)}$ be the unique solution of \begin{align*} \mathcal{E}_{t} H^{(1)}&=0 ,\\ H^{(1)}_{y=a,b}&=\p_yL_tW|_{y=a,b}. \end{align*} Then there exist functions $u_1,u_2,\tilde{u}_1,\tilde{u}_2 \in H^{2}$ (depending on $a,b,k$ and $g$ but not on $t$) and constants $c_1,c_2$ such that \begin{align*} H^{(1)}(t,y)&= c_1\langle W, e^{ikt(y-a)}\tilde{u}_1 \rangle e^{ikt(y-a)}u_1 \\ & \quad + c_2\langle W, e^{ikt(y-b)}\tilde{u}_2 \rangle e^{ikt(y-b)}u_2. \end{align*} Furthermore, for instance for $u_1$ for any $t>0$ \begin{align*} \langle W, e^{ikt(y-a)}\tilde{u}_1 \rangle = \frac{\omega_0(a)}{ikt} - \frac{1}{ikt} \langle W, e^{ikt(y-a)}\p_y\tilde{u}_1 \rangle - \frac{1}{ikt} \langle \p_yW, e^{ikt(y-a)}\tilde{u}_1 \rangle. \end{align*} \end{lem} Based on this characterization of $H^{(1)}$, we introduce a splitting of $\p_yW$ into a function $\beta$ depending only on $\omega_0|_{y=a,b}$ and $\Gamma=\p_yW-\beta$. As we show in Theorem~\ref{thm:StabilityofGamma}, $\Gamma$ is stable also in higher regularity. In contrast, unless $\omega_0|_{y=a,b}$ is trivial, $\beta$ asymptotically develops singularities at the boundary and exhibits blow-up in $H^{s},s>1/2$. If one however considers weighted spaces, it is possible to compensate for these singularities by vanishing weights and hence establish sufficient control for damping with optimal decay rates. \begin{lem} \label{lem:splitting} Let $W$ be a solution of~\eqref{eq:SLE} and let $\Gamma$ be the solution of \begin{align} \label{eq:Gamma} \begin{split} \dt \Gamma &= \frac{ih}{k}L_t \Gamma - \frac{h}{k^2t} \langle \Gamma, e^{ikt(y-a)}\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 - \frac{h}{k^2t} \langle \Gamma, e^{ikt(y-b)}\tilde{u}_2 \rangle e^{ikt(y-b)} u_2 \\ & \quad+ \frac{ih'}{k}L_tW + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_tW -c_1 \frac{h}{k^2t} \langle W, e^{ikt(y-a)}\p_y\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 \\ & \quad -c_2 \frac{h}{k^2t} \langle W, e^{ikt(y-b)}\p_y\tilde{u}_2 \rangle e^{ikt(y-b)} u_2, \\ \Gamma|_{t=0}&=\p_y \omega_0, \end{split} \end{align} and let $\beta$ be the solution of \begin{align} \label{eq:beta} \begin{split} \dt \beta &= \frac{ih}{k}L_t \beta - \frac{h}{k^2t} \langle \beta, e^{ikt(y-a)}\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 - \frac{h}{k^2t} \langle \beta, e^{ikt(y-b)}\tilde{u}_2 \rangle e^{ikt(y-b)} u_2 \\ & \quad -\frac{c_1h\omega_0(a)}{k^2t} e^{ikt(y-a)}u_1 -\frac{c_2h\omega_0(b)}{k^2t} e^{ikt(y-b)}u_2, \\ \beta|_{t=0}&=0. \end{split} \end{align} Then $\p_yW=\Gamma+\beta$. The function $\beta$ is called the \emph{boundary layer}. \end{lem} \begin{thm}[$H^2$ regularity of $\Gamma$] \label{thm:StabilityofGamma} Suppose that $g,h$ satisfy the assumptions of Theorem~\ref{thm:L2}. \begin{enumerate} \item Suppose that additionally $g \in W^{2,\infty}$ and $h \in W^{2,\infty}$. Then there exists a constant $C_1$ such that for all $\omega_0 \in H^1$ and any $t\geq 0$, the solution $\Gamma$ of~\eqref{eq:Gamma} satisfies \begin{align*} \|\Gamma(t)\|_{L^2} \leq C \|\omega_0\|_{H^1}. \end{align*} \item Suppose that additionally $g \in W^{3,\infty}$ and $h \in W^{3,\infty}$, then there exists a second constant $C_2$ such that for any $\omega_0 \in H^2$ and for any $t \geq 0$, \begin{align*} \|\Gamma(t)\|_{H^1} \leq C_2 \|\omega_0\|_{H^2}. \end{align*} \end{enumerate} \end{thm} \begin{thm}[$H^2$ regularity of $\beta$] \label{thm:Stabilityofbeta} Suppose $g,h$ satisfy the assumptions of Theorem~\ref{thm:L2}. \begin{enumerate} \item Then there exists a constant $C_1$ such that for all $t\geq 0$, the solution $\beta$ of~\eqref{eq:beta} satisfies \begin{align*} \|\beta(t)\|_{L^2} \leq C_1 (|\omega_0(a)|+|\omega_0(b)|). \end{align*} \item Suppose that additionally $g,h \in W^{2,\infty}$, then there exists a second constant $C_2$ such that \begin{align*} \|(y-a)(y-b)\p_y\beta(t)\|_{L^2} \leq C_2 (|\omega_0(a)|+|\omega_0(b)|). \end{align*} However, if for instance $|\omega_0(a)|>0$, then \begin{align*} |\beta(t,a)| \gtrsim \log(t) \end{align*} as $t \rightarrow \infty$ (similarly for $b$). In particular, by the Sobolev embedding, we obtain blow-up in $H^{s},s>1/2$. \end{enumerate} \end{thm} \begin{rem} \begin{itemize} \item Combining Theorems~\ref{thm:L2},~\ref{thm:StabilityofGamma} and~\ref{thm:Stabilityofbeta} and Proposition~\ref{prop:damping}, we obtain Theorem~\ref{thm:main1}. \item It is possible to further split $\Gamma$ into functions controlled solely in terms of $\|\omega_0\|_{L^2}$, $\|\p_y \omega_0\|_{L^2}$ and $\|\p_{y}^2 \omega_0\|_{L^2}$, if finer control is desired. \item Like Theorem~\ref{thm:L2}, in addition to these stability results we obtain Lyapunov functionals. As a key difference, these functionals are however in general only decreasing for times $t\geq T>0$. Control up time $T$ is hence provided by a Gronwall-type argument, which determines the constants $C_1,C_2$. \item We stress that we do not require higher norms of $g,h$ to be small but only finite, so that derivatives of the equation are well-defined as mappings in $L^2$. \item When considering a setting without boundary contributions such as $\T \times \R$ or $\T \times \T$, no boundary correction $\beta$ is needed. Thus (a suitable modification of) this result already yields the desired stability for decay with optimal rates. Furthermore, this result generalizes to higher derivatives in a straightforward way, where again only finiteness of higher norms has to be required. \end{itemize} \end{rem} \subsection{Stability of $\Gamma$} \label{sec:h1gamma} As the main result of this subsection we provide a proof of Theorem \ref{thm:StabilityofGamma}. Here, the $L^2$ stability result is self-contained, while the $H^1$ estimate presupposes the $L^2$ stability of $\beta$, which is established in the following subsection. Furthermore, we briefly discuss the implications of Theorem \ref{thm:StabilityofGamma} for settings without boundary and provide an improved stability result for the setting of an infinite plane periodic channel, $\T_L\times \R$. We recall that $\Gamma$ is the solution of \begin{align*} \dt \Gamma &= \frac{ih}{k}L_t \Gamma - \frac{h}{k^2t} \langle \Gamma, e^{ikt(y-a)}\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 - \frac{h}{k^2t} \langle \Gamma, e^{ikt(y-b)}\tilde{u}_2 \rangle e^{ikt(y-b)} u_2 \\ & \quad+ \frac{ih'}{k}L_tW + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_tW -c_1 \frac{h}{k^2t} \langle W, e^{ikt(y-a)}\p_y\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 \\ & \quad -c_2 \frac{h}{k^2t} \langle W, e^{ikt(y-b)}\p_y\tilde{u}_2 \rangle e^{ikt(y-b)} u_2, \\ \Gamma|_{t=0}&=\p_y \omega_0, \end{align*} In addition to the estimates for $L_t$ derived in Section~\ref{sec:L2stability}, we hence need to control contributions of the form \begin{align*} \frac{1}{ikt} \langle \Gamma, e^{ikt(y-a)}\tilde{u}_1 \rangle \langle A \Gamma, e^{ikt(y-b)} u_2 \rangle, \end{align*} which can not be controlled by the previous choice of $A(t)$. Instead, we construct a modified weight $A_1(t)$, which is introduced in the following Lemmata (cf. \cite{Zill5} for a similar construction adapted to fractional Sobolev spaces). \begin{lem} \label{lem:modweight1} Let $u \in H^1$, then for $0<\mu<1/2$ and for every $v=\sum_n v_n e^{iny} \in L^2$ \begin{align*} |\langle v, e^{ikt(y-a)} u \rangle|^{2} \leq C_\mu \|u\|_{H^1}^2 \sum_n <n-kt>^{-2\mu}|v|_{n}^2. \end{align*} \end{lem} \begin{proof}[Proof of Lemma~\ref{lem:modweight1}] By expanding the $L^2$ inner product in a basis, we obtain that \begin{align*} \langle v, e^{ikt(y-a)} u\rangle= \sum_n v_n \langle e^{iny},e^{ikt(y-a)} u \rangle. \end{align*} Integrating by parts and using the trace inequality, we further estimate \begin{align*} |\langle e^{iny} ,e^{ikt(y-a)} u \rangle | \leq <n-kt>^{-1} \|u\|_{H^1}. \end{align*} The result hence follows by an application of the Cauchy-Schwarz inequality: \begin{align*} |\langle v, e^{ikt(y-a)} u\rangle | &\lesssim \sum_n v_n <n-kt>^{-\mu} <n-kt>^{1-\mu} \\ &\leq \|v_n <n-kt>^{-\mu}\|_{l^2} \|<n-kt>^{1-\mu}\|_{l^2} \\ &\leq 2 \|v_n <n-kt>^{-\mu}\|_{l^2} \|<n>^{-(1-\mu)}\|_{l^2} \\ &=: C_\mu \|v_n <n-kt>^{-\mu}\|_{l^2}, \end{align*} where we used that $<n>^{-(1-\mu)} \in l^2$ if $\mu<1/2$. \end{proof} \begin{lem} \label{lem:A1} Let $0<\lambda,\mu <1$ with $\lambda+2\mu>1$ and let $\epsilon>0$ and define the symmetric operator $A_1(t)$ by its action on the basis: \begin{align*} A_1(t): e^{iny} \mapsto \exp\left(\arctan\left(\frac{\eta}{k}-t\right)- \epsilon \int^{t} <\tau>^{-\lambda} <n-k\tau>^{-2\mu} d\tau\right). \end{align*} Then for every $u \in L^2$ and every $t \in \R$ \begin{align*} C \|u\|_{L^2}^2\leq \langle u, A_1(t) u \rangle &\leq C^{-1} \|u\|_{L^2}^2, \\ \langle u, \dot{A}_1(t)u \rangle \leq -C_1\|u\|_{\tilde{H}^{-1}_t}^2 - C\epsilon \sum_n <t>^{-\lambda} <n-kt>^{-2\mu}|u_n|^2 &\leq 0. \end{align*} \end{lem} \begin{proof}[Proof of Lemma \ref{lem:A1}] We note that $<t>^{-\lambda} <n-kt>^{-2\mu} \in L^1(\R)$ and that \begin{align*} - \epsilon \int^{t} <\tau>^{-\lambda} <n-k\tau>^{-2\mu} d\tau \end{align*} is monotonically decreasing. The properties of $A_1(t)$ hence follow by direct computation, where \begin{align*} C=\exp(-\pi - \epsilon \|<\cdot>^{-\lambda} <n-k\cdot>^{-2\mu}\|_{L^1(\R)}). \end{align*} and $C_1$ is determined by $C$ and Lemma~\ref{lem:A}. \end{proof} \begin{lem} \label{lem:com1} Let $g \in W^{2,\infty}$, $g \geq c>0$, then for every $u \in L^2$ and for every $t \geq 0$, \begin{align*} \|L_t[\mathcal{E}_{t}, \p_y]L_tu\|_{\tilde{H}^{1}_{t}} \lesssim \|u\|_{\tilde{H}^{-1}_t}. \end{align*} \end{lem} \begin{proof}[Proof of Lemma \ref{lem:com1}] By Lemma~\ref{lem:compCC}, we obtain that \begin{align*} \|L_t[\mathcal{E}_{t}, \p_y]L_tu\|_{\tilde{H}^{1}_{t}} \lesssim \|[\mathcal{E}_{t}, \p_y]L_tu\|_{\tilde{H}^{-1}_{t}}. \end{align*} We further note that \begin{align*} [\mathcal{E}_{t}, \p_y]= e^{-ikty}[\mathcal{E}_0, \p_y-ikt]e^{ikty}= e^{-ikty}[\mathcal{E}_0, \p_y]e^{ikty}, \end{align*} and that, by direct computation, $[\mathcal{E}_0, \p_y]$ is a second-order operator. Hence, using integration by parts, we further estimate \begin{align*} \|[\mathcal{E}_{t}, \p_y]L_tu\|_{\tilde{H}^{-1}_{t}} \lesssim \|L_tu\|_{\tilde{H}^{1}_t} \lesssim \|u\|_{\tilde{H}^{-1}_t}. \end{align*} \end{proof} Using these results, we can now provide a proof of Theorem \ref{thm:StabilityofGamma} and thus establish $L^2$ stability. \begin{proof}[Proof of Theorem~\ref{thm:StabilityofGamma}, part $(1)$] Fix $0<\lambda,\mu<1$ with $2\mu+\lambda>1$ and let $A_1$ be given by Lemma \ref{lem:A1}, where \begin{align*} 0<\epsilon < \frac{1}{100} \|<n-k\cdot>^{-2\mu}<\cdot>^{-\lambda}\|_{L^1(\R)}^{-1}. \end{align*} Then we define \begin{align} \label{eq:LyapGamma1} I(t):= \langle \Gamma, A_1(t)\Gamma \rangle + C_1 \langle W, A(t)W \rangle, \end{align} where $C_1 \gg 0$ is to be chosen later. We then claim that there exists $T>0$ such that for all initial data and for all $t \geq 0$, $I(t)$ satisfies \begin{align*} \frac{d}{dt}I(t) \leq C t^{-2(1-\mu/2)} \|\omega_0\|_{L^2}^2 \in L^1(\R). \end{align*} Using Gronwall's inequality, we further obtain that \begin{align*} I(T)\leq \exp(CT)I(0), \end{align*} which concludes the proof. \\ It remains to prove the claim. Using Theorem~\ref{thm:L2} and Lemma \ref{lem:A1}, we directly compute \begin{align*} \frac{d}{dt}I(t) &\leq -C\|\Gamma\|_{\tilde{H}^{-1}_t}^2 - C\epsilon \sum_n <t>^{-\lambda} <n-kt>^{-2\mu}|\Gamma_n|^2 \\ & \quad - C_1\delta \|W(t)\|_{\tilde{H}^{-1}_t}^2 + 2 \Re \langle \frac{d}{dt}\Gamma, A_{1}(t)\Gamma \rangle . \end{align*} Using Lemma~\ref{lem:modweight1} and Lemma~\ref{lem:EllipticContribution} and recalling~\eqref{eq:Gamma}, we further estimate \begin{align*} 2 \Re \langle \frac{d}{dt}\Gamma, A_{1}(t)\Gamma \rangle &\leq C(h,k)\|\Gamma\|_{\tilde{H}^{-1}_t} \|A_1\Gamma\|_{\tilde{H}^{-1}_t} + C(h,k,\mu) \frac{1}{t} \left( \sum_n <n-kt>^{-2\mu}|\Gamma_n|^2 \right) \\ & \quad + C(h,h',k) \|A_{1}\Gamma\|_{\tilde{H}^{-1}_t} (\|L_tW\|_{\tilde{H}^{1}_t} + \|L_t [\mathcal{E}_{t},\p_y]L_t\|_{\tilde{H}^{1}_t}) \\ & \quad + C(h,k,g) t^{-1} \|\omega_0\|_{L^2} \sqrt{\sum_n <n-kt>^{-2\mu}|\Gamma_n|^2}. \end{align*} Splitting $t=t^{-(1-\mu)} t^{-\mu}$ and using Young's inequality and Lemmata~\ref{lem:compCC} and~\ref{lem:com1}, we further control \begin{align*} \frac{1}{t} \left( \sum_n <n-kt>^{-2\mu}|\Gamma_n|^2 \right)= t^{-(1-\lambda)}\sum_n t^{-\lambda} <n-kt>^{-2\mu}|\Gamma_n|^2 , \\ \|A_{1}\Gamma\|_{\tilde{H}^{-1}_t} (\|L_tW\|_{\tilde{H}^{1}_t} + \|L_t [\mathcal{E}_{t},\p_y]L_t\|_{\tilde{H}^{1}_t}) \leq \sigma \|A_{1}\Gamma\|_{\tilde{H}^{-1}_t}^{2} + \sigma^{-1} \|W(t)\|_{\tilde{H}^{-1}_t}^2 , \\ t^{-1} \|\omega_0\|_{L^2} \sqrt{\sum_n <n-kt>^{-2\mu}|\Gamma_n|^2} \leq \sigma \langle \Gamma, \dot{A}_1(t) \Gamma \rangle| + \sigma^{-1} t^{-2(1-\mu/2)}\|\omega_0\|_{L^2}^2. \end{align*} Choosing $\sigma$ sufficiently small and letting $T>0$ be sufficiently large and using the smallness assumption of Theorem~\ref{thm:L2}, we observe that \begin{align*} -C\|\Gamma\|_{\tilde{H}^{-1}_t}^2 - C\epsilon \sum_n <t>^{-\lambda} <n-kt>^{-2\mu}|\Gamma_n|^2 + C(h,k)\|\Gamma\|_{\tilde{H}^{-1}_t} \|A_1\Gamma\|_{\tilde{H}^{-1}_t} \\ + (C(h,k,\mu)t^{-(1-\mu)}) \sum_n <t>^{-\lambda} <n-kt>^{-2\mu}|\Gamma_n|^2 \\ + \sigma \|A_{1}\Gamma\|_{\tilde{H}^{-1}_t}^{2} + \sigma \langle \Gamma, \dot{A}_1(t) \Gamma \rangle| \leq 0. \end{align*} Similarly, choosing $C_1$ sufficiently large, we observe that \begin{align*} - C_1\delta \|W(t)\|_{\tilde{H}^{-1}_t}^2 + \sigma^{-1} \|W(t)\|_{\tilde{H}^{-1}_t}^2 \leq 0. \end{align*} Hence, we conclude that for $t \geq T>0$, $I(t)$ satisfies \begin{align*} \frac{d}{dt}I(t) \leq \sigma^{-1} t^{-2(1-\mu/2)}\|\omega_0\|_{L^2}^2. \end{align*} which finishes the proof of the claim and hence of the $L^2$ stability result, $(1)$. \end{proof} Next, we consider the evolution of $\p_y\Gamma$: \begin{align} \label{eq:pyGamma} \dt \p_y\Gamma&= \frac{ih}{k}L_t \p_y \Gamma + \frac{ih'}{k}L_t \Gamma + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_t\Gamma - (\p_yL_t\Gamma)(a) e^{ikt(y-a)}u_1 - (\p_yL_t\Gamma)(b) e^{ikt(y-b)}u_1\\ & \quad+ \p_y\Big( \frac{h}{k^2t} \langle \Gamma, e^{ikt(y-a)}\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 - \frac{h}{k^2t} \langle \Gamma, e^{ikt(y-b)}\tilde{u}_2 \rangle e^{ikt(y-b)} u_2 \\ & \quad-c_1 \frac{h}{k^2t} \langle W, e^{ikt(y-a)}\p_y\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 -c_2 \frac{h}{k^2t} \langle W, e^{ikt(y-b)}\p_y\tilde{u}_2 \rangle e^{ikt(y-b)} u_2 \Big) \\ & \quad + \p_y \left( \frac{ih'}{k}L_tW + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_tW \right), \\ \p_y \Gamma|_{t=0}&=\p_{y}^2 \omega_0 \end{align} Since we here also have to compute $\p_yW=\Gamma+\beta$ in order to control $\|\p_y \Gamma\|_{L^2}$, we require $L^2$ estimates on $\beta$. Before continuing with the proof of Theorem~\ref{thm:StabilityofGamma}, we hence prove the first part of Theorem~\ref{thm:Stabilityofbeta} as well as some further properties of the evolution of $\beta$, which are formulated in the following proposition. \begin{prop} \label{prop:betaL2} Suppose $g,h$ satisfy the assumptions of Theorem~\ref{thm:Stabilityofbeta}. Let $\beta$ be the solution of~\eqref{eq:beta} and let $A_1(t)$ be given by Lemma~\ref{lem:A1}. Then there exists $T>0$ such that for all $t\geq 0$ \begin{align*} I_2(t)=\langle \beta, A_1(t)\beta \rangle \end{align*} satisfies \begin{align*} \frac{d}{dt}I_2(t) \leq \delta \langle \beta, \dot A_1(t)\beta \rangle + C t^{-2(1-\mu/2)}|\omega_0|_{y=a,b}|^2. \end{align*} \end{prop} \begin{proof}[Proof of Proposition~\ref{prop:betaL2}] Using the same weight $A_{1}$, we observe that \begin{align*} \Re \langle A_{1}(t)\beta, \frac{ih}{k}L_t \beta -\frac{h}{k^2t} \langle \beta, e^{ikt(y-a)}\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 - \frac{h}{k^2t} \langle \beta, e^{ikt(y-b)}\tilde{u}_2 \rangle e^{ikt(y-b)} u_2 \rangle \\ \leq (C(h,k)+ C(h,k,g)t^{-(1-\mu)})|\langle \beta, \dot{A}_1(t)\beta \rangle|. \end{align*} Using the smallness assumption and restricting to $t\geq T>0$, this contribution can thus be absorbed by \begin{align*} \langle \beta, \dot{A}_1(t)\beta \rangle \leq 0. \end{align*} Hence, we focus on \begin{align*} \Re \langle A_{1}\beta, \omega_{0}(a)\frac{1}{ikt}e^{ikty}u \rangle \lesssim C_{\lambda} \|\beta_{n}<n-kt>^{-\lambda}\|_{l^{2}} |\omega_{0}|_{y=a,b}| \frac{1}{|kt|}. \end{align*} Using Young's inequality and choosing $\sigma$ sufficiently small, we thus obtain that \begin{align*} \frac{d}{dt}\langle \beta, A_1 \beta \rangle \leq \delta \langle \beta, \dot{A}_1 \beta \rangle + C\sigma^{-1}t^{-2(1-\mu/2)} |\omega_0|_{y=a,b}|^2. \end{align*} The first part of Theorem \ref{thm:Stabilityofbeta} then follows by integrating this inequality and using a Gronwall-type estimate to control the growth up to time $T$. \end{proof} Additionally, we make use of the following estimates for boundary evaluations of $L_t\Gamma, W$ and $\Gamma$, which are obtained as an application of the results of Appendix \ref{sec:auxil-funct-bound}. \begin{lem} \label{lem:bdGamma} Let $g,h,k$ satisfy the assumptions of the second part of Theorem \ref{thm:StabilityofGamma}. Then, \begin{align*} (\p_yL_t\Gamma)(a)=c_1 \langle \Gamma, e^{ikt(y-a)}\tilde{u}_1 \rangle, \\ (\p_yL_t\Gamma)(b)=c_2 \langle \Gamma, e^{ikt(y-b)}\tilde{u}_2 \rangle, \end{align*} and the following estimates hold: \begin{align*} |\langle \Gamma, e^{ikt(y-a)}\tilde{u}_1 \rangle| &\lesssim \frac{C_\mu}{kt}\sqrt{\sum_n |(\p_y\Gamma)|_{n}^2 <n-kt>^{-2\lambda}} + \frac{C}{kt}|\Gamma(a,t)|, \\ |\langle W, e^{ikt(y-a)}\tilde{u}_1 \rangle| &\lesssim \frac{C_\mu}{kt}(\|\Gamma\|_{L^2}+\|\beta(t)\|_{L^2})+ \frac{C}{kt}|\omega_0(a,t)|, \\ |\Gamma(a,t)| &\leq \log(t)(|\omega_0(a)|+\|\omega_0\|_{L^2}). \end{align*} \end{lem} \begin{proof}[Proof of Lemma~\ref{lem:bdGamma}] The evaluations of $\p_yL_t\Gamma$ at the boundary are obtained as an application of Lemma~\ref{lem:H1calc}. The first two estimates follow by integration by parts. In order to show the last estimate, we restrict~\eqref{eq:Gamma} to the boundary and obtain that \begin{align*} |\dt \Gamma(a,t)|\lesssim \frac{1}{kt}(\|\Gamma(t)\|_{L^2}+\|W(t)\|_{L^2}), \end{align*} where we used that $L_t$ enforces zero Dirichlet data. The result hence follows by using Theorem~\ref{thm:L2} and the first part of Theorem ~\ref{thm:StabilityofGamma} to control \begin{align*} \|\Gamma(t)\|_{L^2}+\|W(t)\|_{L^2}\lesssim |\omega_0(a)|+\|\omega_0\|_{L^2}, \end{align*} and then integrating the inequality. \end{proof} \begin{lem} \label{lem:commutator} Let $W$ be the solution of~\eqref{eq:SLE} with initial datum $\omega_0 \in H^1$ and let $\Gamma$ and $\beta$ be as in Lemma~\ref{lem:splitting}. Then, for any $\sigma>0$, \begin{align*} \Re \langle A_1 \p_y\Gamma, \left[\left(\frac{ih'}{k}L_t\cdot + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_t \cdot \right), \p_y \right] W \rangle \leq \sigma |\langle \p_y\Gamma, \dot{A}_1 \p_y \Gamma \rangle| + C\sigma^{-1} \|W\|_{\tilde{H}^{-1}_t}^2. \end{align*} \end{lem} \begin{proof}[Proof of Lemma~\ref{lem:commutator}] The contribution due to $\frac{ih'}{k}L_t$ can be estimated as in Lemma \ref{lem:com1}. In the following we thus focus on the commutator and decompose the commutator into the cases where $\p_y$ falls on $h$, \begin{align*} \frac{ih''}{k}L_tW + \frac{ih'}{k}L_t[\mathcal{E}_{t},\p_y]L_tW, \end{align*} the terms solving an elliptic equation with vanishing Dirichlet data, \begin{align*} \frac{ih'}{k}L_t[\mathcal{E}_{t},\p_y]L_t W + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_t[\mathcal{E}_{t},\p_y]L_tW, \end{align*} and the homogeneous corrections, \begin{align*} \frac{ih'}{k} ((\p_yL_tW)(a,t) e^{ikt(y-a)}u_1+(\p_yL_tW)(b,t) e^{ikt(y-b)}u_2), \\ \frac{ih}{k} ((\p_yL_t[\mathcal{E}_{t},p_y]L_tW)(a,t) e^{ikt(y-a)}u_1+(\p_yL_t[\mathcal{E}_{t},\p_y]L_tW)(b,t) e^{ikt(y-b)}u_2), \end{align*} In the first and second case, we use Lemmata~\ref{lem:compCC} and~\ref{lem:EllipticContribution} to estimate by \begin{align*} \|\p_y\Gamma\|_{\tilde{H}^{-1}_t} \|W\|_{\tilde{H}^{-1}_t}, \end{align*} which is of the desired form by Young's inequality. \\ It hence only remains to consider the homogeneous corrections. Here, we estimate \begin{align*} \Re \Big\langle A_1(t)\p_y \Gamma &, \frac{ih'}{k} ((\p_yL_tW)(a,t) e^{ikt(y-a)}u_1+(\p_yL_tW)(b,t) e^{ikt(y-b)}u_2) \\ &+ \frac{ih}{k} ((\p_yL_t[\mathcal{E}_{t},p_y]L_tW)(a,t) e^{ikt(y-a)}u_1+(\p_yL_t[\mathcal{E}_{t},\p_y]L_tW)(b,t) e^{ikt(y-b)}u_2) \\ &\leq C_\mu \sqrt{\sum_n |\p_y\Gamma_n|^{2}<n-kt>^{-2\mu}} \Big(|\p_yL_tW)(a,t)|+|\p_yL_tW)(b,t)| \\ & \quad +|\p_yL_t[\mathcal{E}_{t},p_y]L_tW)(a,t)| + |\p_yL_t[\mathcal{E}_{t},p_y]L_tW)(b,t)|\Big). \end{align*} We further recall from Section~\ref{sec:auxil-funct-bound} that boundary evaluations can be obtained by testing with suitable homogeneous solution to the adjoint problem. Hence, \begin{align*} |\p_yL_tW)(a,t)|+|\p_yL_tW)(b,t) \lesssim t^{-1} \|\p_yW\|_{L^2}\leq t^{-1}\|\omega_0\|_{H^1}, \\ |\p_yL_t[\mathcal{E}_{t},p_y]L_tW)(a,t)| + |\p_yL_t[\mathcal{E}_{t},p_y]L_tW)(b,t)| \lesssim t^{-1} \|[\mathcal{E}_{t},\p_y]L_tW\|_{H^1}. \end{align*} We can thus conclude the proof, if we can show that \begin{align*} \|[\mathcal{E}_{t},\p_y]L_tW\|_{H^1} \lesssim \|W\|_{H^1}. \end{align*} Expressing $\p_y [\mathcal{E}_{t},\p_y]L_tW= [\mathcal{E}_{t},\p_y]L_t \p_yW +[[\mathcal{E}_{t},\p_y]L_t,\p_y]W$, this estimate follows from elliptic regularity theory for $[\mathcal{E}_{t},\p_y]L_t|_{t=0}$ and using that multiplication by $e^{ikty}$ is an isometry. \end{proof} Building on these results, we can now complete the proof of Theorem \ref{thm:StabilityofGamma}. \begin{proof}[Proof of Theorem~\ref{thm:StabilityofGamma}, part 2] Following a similar strategy as in the previous part, we consider \begin{align*} I_{2}(t):= \langle \p_y \Gamma, A_1(t)\p_y \Gamma \rangle + C_1 \langle \Gamma, A_1(t)\Gamma \rangle + C_2 \langle \beta, A_1(t)\beta \rangle + C_3 \langle W, A(t)W\rangle, \end{align*} where $C_1,C_2,C_3>0$ are to be chosen later. Using the preceding results and strategy, it suffices to study \begin{align*} \Re \langle \p_t \p_y \Gamma, A_1 \p_y \Gamma \rangle. \end{align*} Following the same strategy as in the previous part of the proof and using Lemma~\ref{lem:bdGamma}, we estimate \begin{align*} & \quad \Re \langle A_1 \p_y\Gamma, \frac{ih}{k}L_t \p_y \Gamma + \frac{ih'}{k}L_t \Gamma + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_t\Gamma \\ & \quad- (\p_yL_t\Gamma)(a) e^{ikt(y-a)}u_1 - (\p_yL_t\Gamma)(b) e^{ikt(y-b)}u_1 \rangle \\ &\leq (C+\sigma+ct^{-(1-\mu)}\log(t)) \|\p_y\Gamma\|_{\tilde{H}^{-1}_t}^2+ \sigma^{-1}|\langle \Gamma, \dot A \Gamma \rangle|, \end{align*} which can be absorbed. \\ Furthermore, applying Lemma~\ref{lem:modweight1}, we can control \begin{align*} & \quad\Re \Big\langle A_1 \p_y\Gamma , \p_y\Big( \frac{h}{k^2t} \langle \Gamma, e^{ikt(y-a)}\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 - \frac{h}{k^2t} \langle \Gamma, e^{ikt(y-b)}\tilde{u}_2 \rangle e^{ikt(y-b)} u_2 \\ & \quad -c_1 \frac{h}{k^2t} \langle W, e^{ikt(y-a)}\p_y\tilde{u}_1 \rangle e^{ikt(y-a)} u_1 -c_2 \frac{h}{k^2t} \langle W, e^{ikt(y-b)}\p_y\tilde{u}_2 \rangle e^{ikt(y-b)} u_2 \Big) \Big \rangle \\ &\leq C(\mu,g,h,k) \left( \sum_n |\Gamma_n|^2<n-kt>^{-2\mu} \right)^{1/2} \Big( \left| \langle \Gamma, e^{ikt(y-a)}\tilde{u}_1 \rangle \right| \\ & \quad +\left|\langle \Gamma, e^{ikt(y-b)}\tilde{u}_2 \rangle\right| +\left|\langle W, e^{ikt(y-a)}\p_y\tilde{u}_1 \rangle\right| +\left|\langle W, e^{ikt(y-b)}\p_y\tilde{u}_2 \rangle\right| \Big). \end{align*} Applying the estimates of Lemma~\ref{lem:bdGamma} and using Young's inequality, these contributions can hence again be partially absorbed provided $\sigma$ is sufficiently small and $T>0$ is sufficiently large. The remaining non-absorbed terms can be estimated by \begin{align*} t^{-2(1-\mu/2)} (|\Gamma(a,t)| + |\Gamma(b,t)| + \|\p_y W(t)\|_{L^2} + |\omega_0(a)| + |\omega_0(b)|) \lesssim t^{-2(1-\mu/2)} \|\omega_0\|_{H^1}, \end{align*} where we used Theorem~\ref{thm:L2}, the first part of Theorem \ref{thm:StabilityofGamma} and the Sobolev embedding. \\ It remains to estimate \begin{align*} \Re \left\langle A_1(t)\p_y\Gamma, \p_y \left( \frac{ih'}{k}L_tW + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_tW \right) \right\rangle . \end{align*} Recalling the definition of $\Gamma$ and $\beta$, we express the right function as \begin{align*} \left(\frac{ih'}{k}L_t\cdot + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_t \cdot \right) (\Gamma+\beta) \\ + \left[\left(\frac{ih'}{k}L_t\cdot + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_t \cdot \right), \p_y \right] W. \end{align*} We then estimate \begin{align*} \Re \left\langle A_1(t)\p_y\Gamma, \left(\frac{ih'}{k}L_t\cdot + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_t \cdot \right) (\Gamma+\beta) \right\rangle \lesssim \|\p_y\Gamma\|_{\tilde{H}^{-1}_t} (\|\Gamma\|_{\tilde{H}^{-1}_t} + \|\beta\|_{\tilde{H}^{-1}_t}). \end{align*} Using Young's inequality, the respective terms can then again be controlled, given a suitable choice of $\sigma$. Finally, using Lemma~\ref{lem:commutator}, \begin{align*} \Re \langle A_1 \p_y\Gamma, \left[\left(\frac{ih'}{k}L_t\cdot + \frac{ih}{k}L_t[\mathcal{E}_{t},\p_y]L_t \cdot \right), \p_y \right] W \rangle \\ \leq \sigma |\langle \p_y\Gamma, \dot{A}_1 \p_y \Gamma \rangle| + C\sigma^{-1} \|W\|_{\tilde{H}^{-1}_t}^2, \end{align*} which can again be absorbed and hence concludes the proof. \end{proof} \subsection{Weighted stability of $\p_y\beta$ and boundary blow-up} \label{sec:boundary-blow-up} In this section we consider the evolution of $\p_y\beta$. Since the behavior at both boundary points is similar and separates, we for simplicity of notation consider the case $\omega_0(a)\neq 0$, $\omega_0(b)=0$. The general case can then be obtained by switching $a$ and $b$ and using the linearity of the equation. The function $\beta$ then satisfies~\eqref{eq:beta}: \begin{align} \label{eq:31} \begin{split} \dt \beta -\frac{ih}{k}L_{t}\beta - \frac{h}{k^2t}\langle \beta ,e^{ikt(y-a)}u \rangle e^{ikt(y-a)}u &= \omega_{0}(a)\frac{h}{k^2t}e^{ikt(y-a)}u, \\ \beta|_{t=0}&=0. \end{split} \end{align} We note that, if $\omega_{0}|_{y=a,b}=0$, then $\beta$ identically vanishes. We recall that by Proposition \ref{prop:betaL2} under suitable assumptions on $h,g$ and $k$, $\beta$ is stable in $L^2$. However, stability in $H^{1}$ or, indeed in $H^{s},s>1/2$, does not hold due to the asymptotic formation of singularities at the boundary. \begin{lem}[Boundary blow-up] \label{lem:bdblowup} Suppose that for some $s>0$, \begin{align*} \sup_{t>0} \|\beta(t)\|_{H^{s}} = C <\infty. \end{align*} Then $\beta(a,t)$ satisfies \begin{align*} |\beta(a,t)-h(a)\omega_{0}(a)k^{-2} \log(t)| \leq C_{s}C, \end{align*} as $t \rightarrow \infty$ In particular, if $\omega_{0}(a)\neq 0$, then \begin{align*} \sup_{t}\|\beta(t)\|_{C^{0}}=\infty. \end{align*} Hence, by the Sobolev embedding, in that case, \begin{align*} \sup_{t} \|\beta(t)\|_{H^{s}} \geq \sup_{t} \log(t) = \infty, \end{align*} for any $s>\frac{1}{2}$. \end{lem} \begin{proof} Restricting the evolution by ~\eqref{eq:31} to the boundary, we obtain \begin{align} \label{eq:32} \dt \beta(a) + \frac{h(a)}{k^2t} \langle \beta, e^{ikt(y-a)}u \rangle &= \frac{h(a)\omega_{0}(a)}{k^2t}. \end{align} Let $s>0$ and without loss of generality $s<1/2$, then by direct computation \begin{align*} |\langle \beta, e^{ikt(y-a)}u \rangle | \lesssim C t^{-s}\|\beta\|_{H^s}. \end{align*} Hence, $\beta(a,t)$ satisfies \begin{align*} \dt \beta(a) - \frac{\omega_{0}(a)h(a)}{k^2} \dt \log(t) = t^{-1} \mathcal{O}(t^{-s}) \in L^{1}_{t}. \end{align*} The result hence follows by integrating in time. \end{proof} Letting $s=1$ in the preceding Lemma, we in particular note that in general $H^1$ stability of $\beta$ fails. Following a similar approach as in~\cite{Zill5}, one can further show that $s=1/2$ is indeed critical in the sense that stability holds for $H^{s},s<1/2$. As this is however not sufficient for optimal decay rates in the damping estimate of Section~\ref{sec:damping-mixing}, in the following we prove weighted $H^1$ stability as formulated in Theorem~\ref{thm:Stabilityofbeta}. Here, we use a different method of proof based Duhamel's formula, the details of which can be found in Appendix~\ref{sec:Duhamel}. \subsubsection{Splitting $\p_y\beta$} \label{sec:weight-h1-stab} We recall that $\beta$ solves \begin{align} \begin{split} \dt \beta- \frac{ih}{k}L_{t}\beta - \frac{h}{k^2t} \langle \beta, e^{ikt(y-a)}u \rangle e^{ikt(y-a)} u &= \omega_{0}(a) \frac{h}{k^2t}e^{ikt(y-a)}u, \\ \beta|_{t=0}&=0. \end{split} \end{align} Applying one $y$ derivative to this equation, we obtain \begin{align} \label{eq:pybeta} \begin{split} & \quad \dt \p_{y}\beta - \frac{ih}{k}L_{t}\p_{y}\beta + \frac{h}{k^2t} \langle \p_{y} \beta, e^{ikt(y-a)}u \rangle e^{ikt(y-a)} u \\ &= [\frac{ih}{k}L_{t},\p_{y}]\beta + \frac{h}{k^2t} \langle \beta, e^{ikt(y-a)} \p_{y} u \rangle e^{ikt(y-a)} u + \frac{h}{k^2t} \beta(a,t) e^{ikt(y-a)} u \\ &\quad - \frac{h}{k^2t} \langle \beta, e^{ikt(y-a)}u \rangle e^{ikt(y-a)} u - \omega_{0}(a) \frac{h'}{k^2t}e^{ikt(y-a)}u \\ & \quad+\omega_{0}(a) \frac{h}{k^2t}e^{ikt(y-a)}\p_{y}u + \frac{i\omega_{0}(a)h}{k} e^{ikt(y-a)}u, \end{split} \end{align} where we used that \begin{align*} \frac{h}{k^2t} \langle \beta, e^{ikt(y-a)}u \rangle \p_y(e^{ikt(y-a)})u= \frac{h}{k^2t} (\langle \p_y(\beta u), e^{ikt(y-a)} \rangle - \beta u e^{ikt(y-a)}|_{y=a}^b) e^{ikt(y-a)}u. \end{align*} We note that most terms in~\eqref{eq:pybeta} are very similar to ones in equation~\eqref{eq:pyGamma} satisfied by $\p_{y}\Gamma$, with the exception of \begin{align*} \frac{i\omega_{0}(a)h}{k} e^{ikt(y-a)}u, \end{align*} which is hence identified as the term driving the blow-up. Based on this reasoning the following lemma introduces a splitting of $\p_y \beta$. \begin{lem} \label{lem:splitpybeta} Let $\beta_{I}$ be the solution of \begin{align} \label{eq:33} \begin{split} \dt \beta_{I} - \frac{ih}{k}L_{t}\beta_{I} + \frac{1}{ikt} \langle \beta_{I}, e^{ikty}u \rangle e^{ikty} u &= [\frac{ih}{k}L_{t},\p_{y}]\beta + \frac{h}{k^2t} \langle \beta, e^{ikty} \p_{y} u \rangle e^{ikty} u \\ & \quad + \frac{h}{k^2t} \beta(a,t) e^{ikty} u - \frac{h}{k^2t} \langle \beta, e^{ikty}u \rangle e^{ikty} u \\ & \quad - \omega_{0}(a) \frac{h}{k^2t}e^{ikty}\p_{y}u +\omega_{0}(a) \frac{h}{k^2t}e^{ikty}\p_{y}u, \\ \beta_{I}|_{t=0}&=0, \end{split} \end{align} and let $\beta_{II}$ be the solution of \begin{align*} \dt \beta_{II} - \frac{ih}{k}L_{t}\beta_{II} + \frac{h}{k^2t} \langle \beta_{II}, e^{ikty}u \rangle e^{ikty} u &= \omega_{0}(a) e^{ikty}u, \\ \beta_{V}|_{t=0}&=0. \end{align*} Then $\p_y\beta=\beta_{I}+\beta_{II}$. \end{lem} Following the same strategy as in Section~\ref{sec:h1gamma}, we obtain $L^{2}$ stability of $\beta_{I}$. \begin{prop} \label{prop:beta1} Suppose the assumptions of Theorem~\ref{thm:Stabilityofbeta} are satisfied, then \begin{align*} \|\beta_{I}(t)\|_{L^{2}} \lesssim |\omega_{0}|_{y=a,b}|. \end{align*} \end{prop} \begin{proof} Following the same strategy as in the proof of Theorem ~\ref{thm:StabilityofGamma}, we show that, \begin{align*} \frac{d}{dt}\langle \beta_{I}, A_1(t) \beta_{I} \rangle \leq \langle \beta_{I}, \dot{A}_1(t) \beta_{I} \rangle + C \|\beta_{I}\|_{\tilde{H}^{-1}_t}^2 \\ + (Ct^{-(1-\mu)} + \sigma) \sum_n |(\beta_{I})_n|^2<n-kt>^{-2\lambda}t^{-\mu} \\ + C\sigma^{-1} t^{-2(1-\mu/2)} (|\beta(a,t)|^2 + \|\beta\|_{L^2}^2 + |\omega_0(a)|^2). \end{align*} Hence, restricting to $t \geq T>0$ and choosing $\sigma$ sufficiently small, \begin{align*} \frac{d}{dt}\langle \beta_{I}, A_1(t) \beta_{I} \rangle \leq C\sigma^{-1} t^{-2(1-\mu/2)} (|\beta(a,t)|^2 + \|\beta\|_{L^2}^2 + |\omega_0(a)|^2) \\ \leq C\sigma^{-1} t^{-2(1-\mu/2)}\log(t)^2 |\omega_0(a)|^2, \end{align*} where we used Proposition~\ref{prop:betaL2} and that, by equation~\eqref{eq:beta}, \begin{align*} |\beta(a,t)|\lesssim \int^t \tau^{-1}\|\beta(\tau)\|_{L^2} d\tau \lesssim \log(t) |\omega_0(a)|. \end{align*} \end{proof} For later reference, we note that we have thus also proven the following proposition. \begin{prop} Suppose that $g,h,k$ satisfy the assumptions of the second part of Theorem ~\ref{thm:StabilityofGamma}. Then, for any $\omega_0 \in L^2$, the solution $W$ of~\eqref{eq:SLE} satisfies \begin{align*} \|W(t)\|_{H^1} + \|\p_y^2W(t) - \beta_{II}(t)\|_{L^2} \lesssim \|\omega_0\|_{H^2}, \end{align*} where $\beta_{II}$ is given by Lemma \ref{lem:splitpybeta}. \end{prop} \begin{proof} This result combines Theorems~\ref{thm:L2} and~\ref{thm:StabilityofGamma} and Propositions~\ref{prop:betaL2} and~\ref{prop:beta1}. \end{proof} \subsubsection{Weighted stability of $\beta_{II}$} \label{sec:it-works} In order to complete the proof of Theorem~\ref{thm:Stabilityofbeta}, it only remains to study the stability of \begin{align*} \dt \beta_{II} - \frac{ih}{k}L_{t}\beta_{II} + \frac{h}{k^2t} \langle \beta_{II}, e^{ikty}u \rangle e^{ikty} u &= \frac{ih}{k}\omega_{0}(a) e^{ikty}u, \\ \beta_{II}|_{t=0}&=0. \end{align*} While it would be possible to study this equation directly, we instead build on our previous analysis of \begin{align} \label{eq:34} \dt - \frac{ih}{k}L_{t} \end{align} and introduce an additional boundary layer $\nu$ (c.f. Theorem~\ref{thm:main1}) solving \begin{align} \label{eq:35} (\dt - \frac{ih}{k}L_{t})\nu &= \frac{h}{k}\omega_{0}(a)e^{ikty}, \\ \nu|_{t=0} &=0, \end{align} and also define $\beta_{V}=\beta_{II}-\nu$. Then $\beta_{V}$ solves \begin{align} \label{eq:36} \begin{split} \dt \beta_{V} - \frac{ih}{k}L_{t}\beta_{V} + \frac{\langle \beta_{V}, e^{ikty}\tilde{u} \rangle}{ikt} e^{ikty}u &= \frac{\langle \nu, e^{ikty}\tilde{u} \rangle}{ikt} e^{ikty}u. \\ \beta_{V}|_{t=0}&=0. \end{split} \end{align} \begin{rem} Instead of $\nu$ one might attempt to choose the explicit function \begin{align*} \int^t \frac{ih}{k}\omega_0(a)e^{ik\tau y} d\tau= \frac{ih}{k}\omega_0(a)\frac{e^{ikty}-1}{iky}=:\chi. \end{align*} However, we note that part of this function oscillates like $e^{ikty}$ and that \begin{align*} L_t\chi = e^{ikty}L_0 \frac{h}{k^2y}\omega_0(a) + L_t \frac{h}{k^2y}, \end{align*} where $L_0 \frac{h}{k^2y}\omega_0(a)$ is independent of $t$. Hence, even for a constant function $u$ \begin{align*} \langle u, L_t \chi \rangle \end{align*} would not decay or oscillate rapidly enough to be an integrable perturbation. \end{rem} As the main result of this section we establish the following proposition, which concludes the proof of Theorem~\ref{thm:Stabilityofbeta}. \begin{prop} \label{prop:betaV} Suppose the assumptions of Theorem~\ref{thm:Stabilityofbeta} are satisfied. Then the functions $\beta_V$ and $\nu$ satisfy \begin{align} \label{eq:37} \|\beta_{V}(t)\|_{L^{2}} \lesssim |\omega_{0}(a)|, \\ \|(y-a)(y-b)\nu(t)\|_{L^{2}} \lesssim |\omega_{0}(a)|. \end{align} \end{prop} As the evolution of $\beta_{V}$ depends on $\nu$ via \begin{align} \label{eq:38} \langle \nu, e^{ikt(y-a)}\tilde{u} \rangle \end{align} and as our estimates of $\nu$ rely on properties of the solution operator of~\eqref{eq:34} (and hence $W$), we follow a multi-step approach: \begin{enumerate} \item Using Propositions~\ref{prop:beta1} and~\ref{prop:damping}, we show that~\eqref{eq:38} grows at most like $\sqrt{t}$. \item By direct computation, we show that $\|(y-a)(y-b) \nu\|_{L^{2}}$ grows at most like $\log(t)$. \item This yields a weaker form of Proposition~\ref{prop:betaV} with an estimate by $\sqrt{t} |\omega_{0}(a)|$. \item Combining this estimate with the damping result of Section~\ref{sec:damping-mixing}, the estimate of ~\eqref{eq:38} improves to $\log(t)$ and we obtain a uniform bound of $\|(y-a)(y-b) \nu\|_{L^{2}}$. \item Finally, we establish $L^{2}$ stability of $\beta_{V}$ and thus conclude the proof of Proposition~\ref{prop:betaV}. \end{enumerate} \begin{lem} \label{lem:step1} Assume that the assumptions of Theorem~\ref{thm:Stabilityofbeta} are satisfied. Then~\eqref{eq:38} satisfies \begin{align} \label{eq:39} |\langle e^{ikty} \tilde u , \nu(t)\rangle | \lesssim \sqrt{t}|\omega_{0}(a)| \end{align} as $t \rightarrow \infty$. \end{lem} \begin{lem} \label{lem:step2} Assume that the assumptions of Theorem~\ref{thm:Stabilityofbeta} are satisfied. Then $\nu(t)$ satisfies \begin{align} \label{eq:40} \|(y-a)(y-b)\nu(t)\|_{L^{2}} \lesssim \log(t)|\omega_{0}(a)| \end{align} as $t\rightarrow \infty$. \end{lem} \begin{lem} \label{lem:step3} Assume that the assumptions of Theorem~\ref{thm:Stabilityofbeta} are satisfied. Then, as $t \rightarrow \infty$, $\beta_{V}$ and $\nu$ satisfy \begin{align*} \|\beta_{V}(t)\|_{L^{2}} &\lesssim \sqrt{t}|\omega_{0}(a)|, \\ \|(y-a)(y-b)\nu(t)\|_{L^{2}} &\lesssim \log(t)|\omega_{0}(a)|. \end{align*} In particular, we conclude that the solution opertor \begin{align*} S(t,0): H^{2}(dy) &\rightarrow H^{2}\left( (y-a)(y-b) dy\right), \\ \omega_0 &\mapsto W(t), \end{align*} satisfies \begin{align*} ||| S(t,0)||| \lesssim \sqrt{t}. \end{align*} \end{lem} \begin{lem} \label{lem:step4} Assume that the assumptions of Theorem~\ref{thm:Stabilityofbeta} are satisfied. Then $\nu$ satisfies \begin{align} \label{eq:41} \|(y-a)(y-b)\nu(t)\|_{L^{2}} \lesssim |\omega_{0}(a)| \end{align} as $t\rightarrow \infty$. \end{lem} \begin{lem} \label{lem:step5} Assume that the assumptions of Proposition~\ref{prop:betaV} are satisfied. Then $\beta_{V}$ satisfies \begin{align} \label{eq:42} \|\beta_{V}(t)\|_{L^{2}} \lesssim |\omega_{0}(a)| \end{align} as $t\rightarrow \infty$. \end{lem} In our proof of Lemmata~\ref{lem:step1} to~\ref{lem:step5}, we rely on more detailed, (semi-explicit) characterization of $\nu(t)$ via Duhamel's formula, which is established in Appendix~\ref{sec:Duhamel}. \begin{proof}[Proof of Lemma~\ref{lem:step1}] We directly compute \begin{align} \label{eq:43} \langle e^{ikty}\tilde{u}, \int_{0}^{t} e^{ikty}S(t,\tau) e^{ik\tau y} u d\tau \rangle \\ = \langle \tilde{u}, \int_{0}^{t} e^{ik(t-\tau) y}S(t-\tau,0) u d\tau \rangle . \end{align} Next, we integrate \begin{align} \label{eq:44} e^{ik(t-\tau) y} = \p_{\tau} \frac{e^{ik(t-\tau)y}-1}{iky} \end{align} by parts in $\tau$. Here, we obtain a boundary term \begin{align} \label{eq:45} \langle \tilde{u},\frac{e^{ikty}-1}{iky} S(t,0) u \rangle \end{align} and an integral term \begin{align} \label{eq:46} \langle \tilde{u},\int_{0}^{t}\frac{e^{ik(t-\tau) y}-1}{iky} \p_{\tau} S(t-\tau,0) u d\tau \rangle . \end{align} For~\eqref{eq:45} we apply Hölder's inequality and control by \begin{align} \label{eq:47} \|\tilde{u}\|_{L^{\infty}} \|\frac{e^{ikty}-1}{iky}\|_{L^{1}_{y}} \|S(t,0) u\|_{L^{\infty}} \lesssim \log(t) \|u\|_{H^{1}}. \end{align} In the integral term we use the damping estimate, Proposition~\ref{prop:damping}, to control by \begin{align} \label{eq:48} \int_{0}^{t} \|\tilde{u}\|_{L^{\infty}} \|\frac{e^{ik(t-\tau)y}-1}{iky}\|_{L^{2}} \|\p_{\tau} S(t-\tau,0) u\|_{L^{2}} d\tau \\ \lesssim \int_{0}^{t}\sqrt{|t-\tau|} <t-\tau>^{-1} \|S(t-\tau,0) u\|_{H^{1}} d\tau \\ \lesssim \int <t-\tau>^{-1/2} d\tau \lesssim \sqrt{t}. \end{align} \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:step2}] Using Lemmata~\ref{lem:tDuhamel} and~\ref{lem:shiftS}, we obtain that \begin{align} \label{eq:49} \nu(t)= \int_{0}^{t} e^{ikt(t-\tau)(y-a)}S(t-\tau,0)u d\tau \end{align} Multiplying with $(y-a)$, we use that \begin{align*} -\p_{\tau} \frac{e^{ik(t-\tau)(y-a)}-1}{ik}= (y-a)e^{ik(t-\tau)(y-a)} \end{align*} and hence control \begin{align*} \|(y-a)\nu(t)\|_{L^{2}} & \leq \| \frac{e^{ik(t-\tau)(y-a)}-1}{ik} S(t-\tau,0)u|_{\tau=0}^{t}\|_{L^{2}} \\ &\quad + \int_{0}^{t} \|\frac{e^{ik(t-\tau)(y-a)}-1}{ik} \p_{\tau}S(t-\tau,0)u \|_{L^{2}} d\tau \\ &\lesssim |k|^{-1} \|u\|_{L^{2}} + |k|^{-2} \|h\|_{L^{\infty}} \int_{0}^{t} \|L_{t-\tau} S(t-\tau,0)u\|_{L^{2}} \\ & \lesssim |k|^{-1} \|u\|_{L^{2}} + |k|^{-2} \|h\|_{L^{\infty}} \int_{0}^{t} <t-\tau>^{-1} \|S(t-\tau,0)u\|_{H^{1}} d\tau \\ & \lesssim |k|^{-1} \|u\|_{L^{2}} + |k|^{-2} \|h\|_{L^{\infty}} \|u\|_{H^{1}}\log(t), \end{align*} where we used Proposition~\ref{prop:damping} and Theorem~\ref{thm:StabilityofGamma}. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:step3}] Using our Lyapunov functional approach on $\beta_{V}$, we need to estimate \begin{align} \label{eq:50} \langle A_{1}\beta_{V}, e^{ikty}u\rangle \frac{\langle \nu, e^{ikty}\tilde{u} \rangle}{ikt}. \end{align} By Lemma~\ref{lem:step1}, we control \begin{align*} |\frac{\langle \nu, e^{ikty}\tilde{u} \rangle}{ikt}|\lesssim t^{-1/2}, \end{align*} and using Lemma~\ref{lem:A1}, we estimate. \begin{align*} |\langle A_{1}\beta_{V}, e^{ikty}u\rangle | \leq C_{\lambda} \|(\beta_{V})<n-kt>^{-\lambda}\|_{l^{2}}, \end{align*} where $0<\lambda<\frac{1}{2}$. Hence, using Young's inequality, we can control~\eqref{eq:50} by \begin{align*} \epsilon \|(\beta_{V})_n<n-kt>^{-\lambda}\|_{l^{2}}t^{-1/2} + C(\epsilon,\lambda) t^{-1/2}. \end{align*} Here, for $\epsilon$ sufficiently small, the first term can be absorbed by \begin{align*} \langle \beta_{V}, \dot{A}_{1} \beta_{V} \rangle \end{align*} and in summary we obtain \begin{align*} \dt \langle \beta_{V}, A_{1} \beta_{V} \rangle \leq C(\epsilon,\lambda) t^{-1/2}. \end{align*} Integrating this inequality then yields the result. \end{proof} We remark that already in step 3 we could obtain a better growth bound by optimizing in $\lambda$ and the splitting of $t^{-1/2}$ in Young's inequality. However, since $t^{-1/2}\not \in L^{2}$ this would only yield a non-uniform bound and our multi-step proof only requires a better than linear growth bound. \begin{proof}[Proof of Lemma~\ref{lem:step4}] Following the proof of Lemma~\ref{lem:step2} it suffices to show that \begin{align} \label{eq:51} \int_{0}^{t}\|\p_{\tau}S(t-\tau,0)u\|_{L^{2}} d\tau \lesssim 1, \end{align} uniformly in $t$. Using Hölder's inequality and Proposition~\ref{prop:damping}, we estimate \begin{align*} \|\p_{\tau}S(t-\tau,0)u\|_{L^{2}} &= \|\frac{ih}{k}L_{t-\tau}S(t-\tau,0)u\|_{L^{2}} \leq \|h\|_{L^{\infty}}|k|^{-1} \|L_{t-\tau}S(t-\tau,0)u\|_{L^{2}} \\ &\leq \|h\|_{L^{\infty}}|k|^{-1}<t-\tau>^{-2}(\|(y-a)(y-b)\p_{y}^{2}S(t-\tau,0)u\|_{L^{2}} \\ & \quad +\|S(t-\tau,0)u\|_{H^{1}}) \leq \|h\|_{L^{\infty}}|k|^{-1}<t-\tau>^{-2} |||S(t-\tau,0)||| \|u\|_{H^{2}}, \end{align*} the operator norm of $S(t-\tau,0)$ is given by Lemma~\ref{lem:step3}. Hence, we obtain that \begin{align*} \|\p_{\tau}S(t-\tau,0)u\|_{L^{2}} \lesssim \|h\|_{L^{\infty}}|k|^{-1}\|u\|_{H^{2}}<t-\tau>^{-2} \sqrt{t-\tau}, \end{align*} which is integrable in $\tau$ and thus concludes the proof. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:step5}] We claim that \begin{align} \label{eq:52} |\langle e^{ikty} \tilde u , \nu(t)\rangle | \lesssim \log (t)|\omega_{0}(a)|. \end{align} Following the proof of Lemma~\ref{lem:step3}, this implies that \begin{align} \label{eq:53} |\langle A_{1}\beta_{V}, e^{ikty}u\rangle \frac{\langle \nu, e^{ikty}\tilde{u} \rangle}{ikt} | \leq C \|(\beta_{V})<n-kt>^{-\lambda}\|_{l^{2}} \frac{\log(t)}{t} \\ \leq \epsilon \|(\beta_{V})<n-kt>^{-\lambda}\|_{l^{2}}^{2} t^{-2\mu} + C(\epsilon) \log(t)^{2}t^{-2(1-\mu)}, \end{align} where $C(\epsilon)$ is given by Young's inequality and $0<\mu<1$ is chosen such that $2\lambda+2\mu>1$ and $2(1-\mu)>1$. Choosing $\epsilon$ sufficiently small, we thus obtain \begin{align*} \dt \langle \beta_{V}, A_{1}\beta_{V} \rangle &\leq \langle \beta_{V}, \dot{A}_{1}\beta_{V} \rangle + \epsilon \|(\beta_{V})<n-kt>^{-\lambda}\|_{l^{2}}^{2} t^{-2\mu} + C(\epsilon) \log(t)^{2}t^{-2(1-\mu)} \\ &\leq C(\epsilon) \log(t)^{2}t^{-2(1-\mu)} \in L^{1}_{t}([1,\infty)). \end{align*} Integrating this inequality then yields the desired result. \\ It remains to prove the claim~\eqref{eq:52}. Here, we estimate \begin{align*} |\langle e^{ikty} \tilde{u}, \nu(t) \rangle| \lesssim \log(t) \|u\|_{H^{1}} + \int_{0}^{t}\|\tilde{u}\|_{L^{\infty}} \|\frac{e^{ik(t-\tau)y}-1}{iky}\|_{L^{2}} \|\p_{\tau} S(t-\tau,0) u\|_{L^{2}} d\tau . \end{align*} Using Lemma~\ref{lem:step3} and Proposition~\ref{prop:damping} we control \begin{align*} \|\p_{\tau} S(t-\tau,0) u\|_{L^{2}} \leq <t-\tau>^{-2}|||S(t-\tau,0)||| \|u\|_{H^{2}} \lesssim <t-\tau>^{-3/2} \|u\|_{H^{2}}, \end{align*} and we directly compute that \begin{align*} \|\frac{e^{ik(t-\tau)y}-1}{iky}\|_{L^{2}} \lesssim \sqrt{t-\tau}. \end{align*} Hence, we control \begin{align*} \int_{0}^{t}\|\tilde{u}\|_{L^{\infty}} \|\frac{e^{ik(t-\tau)y}-1}{iky}\|_{L^{2}} \|\p_{\tau} S(t-\tau,0) u\|_{L^{2}} d\tau \\ \lesssim \|\tilde{u}\|_{L^{\infty}} \|u\|_{H^{2}} \int_{0}^{t} <t-\tau>^{-1} d\tau \leq \|\tilde{u}\|_{L^{\infty}} \|u\|_{H^{2}} \log(t), \end{align*} which proves the claim. \end{proof}

https://arxiv.org/abs/1201.5228

Almost exponential maps and integrability results for a class of horizontally regular vector fields

We show a higher order integrability theorem for distributions generated by a family of vector fields under a horizontal regularity assumption on their coefficients. We use as chart a class of almost exponential maps which we discuss in details

\section{Introduction and main results} In this paper we discuss the integrability of distributions defined by families of vector fields under a higher order \emph{horizontal regularity} hypothesis and assuming an involutivity condition of order $s\in \mathbb{N}$. The central tool we exploit is given by a class of \emph{almost exponential maps} which we will analyze in details assuming only low regularity on the coefficients of the vector fields. To start the discussion, fix a family $\H= \{X_1, \dots, X_m\} $ of at least Lipschitz-continuous vector fields. For any $x\in \mathbb{R}^n$ define the Sussmann's \emph{orbit}, or \emph{leaf} \begin{equation}\label{orbitale} \cal{O}_{\cal{H}}^x : = \{ e^{t_1X_{j_1}} \cdots e^{t_p X_{j_p}}x: p\in\mathbb{N}, J:=(j_1, \dots, j_p)\in \{1,\dots,m\}^p, t\in \Omega_{J,x}\}, \end{equation} where for fixed $x\in\mathbb{R}^n$ we denote by $\Omega_{J, x}\subset\mathbb{R}^p$ the open neighborhood of the origin where the map $t\mapsto e^{t_1X_{j_1}} \cdots e^{t_p X_{j_p}}x$ is well defined. We equip the leaf $\cal{O}_\H^x$ with the topology $\t_d$ defined by the Franchi--Lanconelli distance $d$; see \eqref{dicocco}. Our purpose is to describe a regularity class of order $s\ge 2$ and a $s$\emph{-involutivity} assumption that ensure that each orbit $\O_\H$ is a integral manifold of the distribution generated by the family~$\P:=\P_s:= \{Y_1, \dots, Y_q\}$ of all nested commutators of length at most $ s $ constructed from the original family $\H$. To give coordinates on $\O$ we shall use the following \emph{almost exponential maps}. Fix $s\ge 2$ and denote by $\P$ the aforementioned family of commutators. Assign to each $Y_j$ the \emph{length} $\ell_j\le s$, just its order. Then, let \begin{equation}\label{almostino} E_{I,x}(h):= \exp_{\textup{ap}}(h_1 Y_{i_1})\cdots \exp_{\textup{ap}}(h_p Y_{i_p})x, \end{equation} where $I= (i_1, \dots, i_p)$ is a multiindex which fixes $p$ commutators $Y_{i_1},\dots, Y_{i_p}\in\P$, $h\in\mathbb{R}^p$ belongs to a neighborhood of the origin and $p\in\{1, \dots, n\}$ is suitable. See \eqref{appsto} for the definition of the \emph{approximate exponential} $\exp_{\textup{ap}}$. We shall use the maps in~\eqref{almostino} to construct charts, developing a higher order, nonsmooth, quantitative extension of some ideas appearing in a paper by Lobry; see~\cite{Lobry70}; see Theorem~\ref{viola} and Remarks~\ref{noterella} and~\ref{lobrao} below. Here is a description of our regularity class. Let $\H= \{X_1, \dots, X_m\}$ and let $s\ge 2$. Assume that $X_j=:f_j\cdot\nabla\in C^1_\textup{Euc}$ for all $j$ (here and hereafter $C^1_\textup{Euc}$ refers to Euclidean regularity). Assume also that for each $p\le s$ and $j_1, \dots, j_p\in\{1,\dots,m\}$, all derivatives $X_{j_1}^\sharp \cdots X_{j_{p-1}}^\sharp f_{j_p}$ exist and are locally Lipschitz-continuous functions with respect to distance $d$ associated to the vector fields. Here, following \cite{MontanariMorbidelli11a}, we denote by $X^\sharp f$ the Lie derivative along the vector field $X$ of the scalar function $f$. Moreover we require that for any commutator $Y_j=: g_j\cdot\nabla\in\P$, all maps of the form $g_j\circ E_{I,x}$ are continuous for all $p\in\{1,\dots,n\}$, $I= (i_1, \dots, i_p)$ and $x\in\mathbb{R}^n$. \footnote{This condition is widely ensured for instance as soon as we assume that $g_j$ is continuous in the Euclidean topology, or at least in the Sussmann's orbit topology defined on $\O$ by the family $\H$; see \cite{Sussmann}.} Furthermore, we require the following $s$-involutivity condition. For any $X_j\in\H$ and for any $Y_k\in\P$ with maximal length $\ell_k=s$, at any $x\in\Omega$ where the derivative $X_j^\sharp g_k(x)$ exists one can write for suitable $b^i= b^i(x)$ \begin{equation}\label{integreria}\begin{aligned} (\ad_{X_j} Y_{k})_x:= & ( X_j^\sharp g_k(x) - Y_k f_j(x))\cdot\nabla = \sum_{ i=1}^{q} b^i Y_{i,x} \quad\text{with $b^i$ locally bounded.}\end{aligned} \end{equation} The class of vector fields satisfying all those assumptions will be denoted by $\mathcal{A}_s$; see Definition \ref{laint}, where a more precise formulation of this assumption is described. Note that in the smooth case we have $\ad_{X_j} Y_{k} = [X_j, Y_k]$ and ultimately \eqref{integreria} is equivalent to the Hermann condition \cite{Hermann} \begin{equation}\label{ajello} [Y_i, Y_j] = \sum_{1\le k\le q} c_{ij}^k Y_k,\quad\text{with $c_{ij}^k\in L^\infty_\textup{loc}$, } \end{equation} which ensures that any Sussmann's orbit $\O_\cal{P}$ of the family of commutators $\cal{P}$ is a integral manifold of the distribution generated by~$\P$. If furthermore $s=1$, then $\P= \H$ and \eqref{ajello} and \eqref{integreria} are the same. Note that the appearance of operators of the form $\ad_{X_j}Y_k$ is very natural in the framework of our almost exponential maps; see the non-commutative calculus formulas discussed in \cite[Section 3]{MontanariMorbidelli11a}. Here is the statement of our result. \begin{theorem}\label{grossouno} Let $\H= \{X_1, \dots, X_m\}$ be a family of vector fields of class $\mathcal{A}_s$. Then, for any $x_0\in\mathbb{R}^n$, the orbit $\O:= \O^{x_0}_\H$ with the topology $\t_d$ is a $C^1$ immersed submanifold of $\mathbb{R}^n$ with tangent space $T_y\O = P_y$ for all $y\in \O$. \end{theorem} Note that this result does not follow from standard ones, because the commutators $Y_j$ are not assumed to be $C^1$ in the Euclidean sense. In Example \ref{esempietto} we exhibit a family of vector fields where our theorem apply, but classical results do not. See also Remark~\ref{finalremark} for some further comments. Furthermore, let us mention that if $s=1$, i.e.~$\H=\P$, then Theorem \ref{grossouno} is a consequence of the Frobenius Theorem for singular $C^1$ distributions (it is well known to experts that in such case one can prove that orbits are even $C^2$ smooth). Note that if $s=1$, in \cite{MontanariMorbidelli11b} we proved a singular Frobenius-type theorem assuming only Lipschitz-continuity of the involved vector fields, generalizing part of Rampazzo's results \cite{Rampazzo} to singular distributions; in fact, in \cite{MontanariMorbidelli11b}, orbits are~$C^{1,1}$. On a technical level, the main tool we discuss is the approximate exponential~$E_{I,x}$ in~\eqref{almostino}. Introduce the notation $ p_x:= \dim P_x := \dim \Span\{Y_{1}(x),\dots, Y_{q}(x)\}$ for all $x\in\mathbb{R}^n$. Fix $x$, take $p:= p_x$ commutators $Y_{i_1}, \dots, Y_{i_p}$, which are linearly independent at $x$ and construct the map $E$, defined in \eqref{almostino}. Then, under the hypotheses of Theorem \ref{grossouno}, we shall show that if the family $\H$ satisfies condition $\mathcal{A}_s$, then $E$ is a $C^1_\textup{Euc}$, full rank map in a neighborhood of the origin $0\in \mathbb{R}^p$, whose derivative enjoys the following remarkable expansion \begin{equation}\label{polacco} \begin{aligned}E_*(\partial_{h_k}) = Y_{i_k}(E(h)) +\sum_{\ell_j=\ell_{i_k}+1}^{s} a^j_k(h) Y_{j }(E(h)) + \sum_{i=1}^q \omega_k^i(x,h) Y_{i}(E(h)). \end{aligned}\end{equation} The functions $a_k^j$ and $\omega_k^i$ have a very precise rate of convergence to $0$, as $h\to 0$ which will be specified in \eqref{sogliola} and \eqref{merluzzo}. Note that an expansion of $E_*(\partial_{h_k})$ can be obtained either with the Campbell--Hausdorff formula in the smooth case (see \cite{Morbidelli98} or \cite{VaropoulosSaloffCosteCoulhon}), or in nonsmooth situations with the techniques of \cite{MM}. However, the expansions in the mentioned papers contain some remainders appearing either as formal series, or in integral form. Here we are able to express such reminders via the pointwise terms $\omega_k^j$, improving all previous results. Note also that we are improving the mentioned papers both from a regularity standpoint and because here we do not assume the H\"ormander condition. At the authors' knowledge, expansion \eqref{polacco} with precise estimates on $a_k^j$ and $\omega_k^i$ is new even in the smooth case. As a final remark, observe that Theorem \ref{deduciamo} contains an explicit detailed proof of the fact that the map $E$ is $C^1$ smooth, avoiding any use of the Campbell--Hausdorff formula. Note that, even if the vector fields are smooth, such maps are not much more than $C^1$; see Remark \ref{tredodici}-(ii). The useful information one can extract from \eqref{polacco} is that $ E_*(\partial_{h_k}) \in P_{E(h)}$ (note that we are interested to situations where the inclusion $P_{E(h)}\subset\mathbb{R}^n$ is strict); see Theorem~\ref{deduciamo} for a precise statement. Observe that, if $O\subset\mathbb{R}^p$ is a small open set containing the origin, then $E(O)$ is a $C^1$ submanifold of $\mathbb{R}^n$ and \eqref{polacco} shows that $ T_{E(h)}E(O) \subseteq P_{E(h)}$ for all $h$. This is the starting point to prove that $\O_\H^x$ is a integral manifold of the distribution generated by~$\P$. Another fact we need to prove is that the dimension of $P_y:= \Span\{Y_j(y): 1\le j\le q\}$ is constant if $y$ belongs to a fixed orbit $ \O_\H^x$. This is obtained by means of a nonsmooth quantitative curvilinear version of the original Hermann's argument inspired to the work of Nagel, Stein and Wainger~\cite{NagelSteinWainger} and Street~\cite{Street}. To conclude this introduction, we give some references and motivations to study our almost exponential maps~$E$. Such maps appear in \cite{NagelSteinWainger}, and were used by the authors to show equivalence between different control distances; see also \cite{VaropoulosSaloffCosteCoulhon}. More recently they have revealed to be a useful tool to study Poincar\'e inequalities (see \cite{LM}), subelliptic Sobolev spaces (see \cite{Danielli91,Morbidelli98,Coulhon01,MM}), and geometric theory of Carnot--Carath\'eodory spaces (see \cite{MontiMorbidelli02,FerrariFranchi03,Vittone10}). Finally, note that the precise expansion \eqref{polacco} will be a fundamental tool in the companion paper \cite{MontanariMorbidelli11c}, where we shall prove a Poincar\'e inequality on orbits for a family of vector fields satisfying an integrability condition. \section{Preliminaries} \label{preliminarmente} \paragraph{Vector fields and the control distance.} Consider a family of vector fields $\H=\{ X_1, \dots, X_m\}$ and assume that $X_j\in C^1_\textup{Euc}(\mathbb{R}^n)$ for all $j$. Here and later $C^1_\textup{Euc}$ means $C^1$ in the Euclidean sense. Write $X_j = :f_j\cdot\nabla$, where $f_j\colon\mathbb{R}^n\to \mathbb{R}^n$. The vector field $X_j$, evaluated at a point $x\in \mathbb{R}^n$, will be denoted by $X_{j,x}$ or $ X_j(x)$. All the vector fields in this paper are always defined on the whole space $\mathbb{R}^n$. Define the Franchi--Lanconelli distance \cite{FL} \begin{equation}\label{dicocco} \begin{aligned} d(x,y)&:= \inf \Big\{r>0: \text{$y=e^{t_1 Z_1}\cdots e^{t_\mu Z_\mu} x$ for some $\mu\in \mathbb{N}$} \\& \qquad\qquad \qquad \text{where $\sum \abs{t_j}\le 1$ with $Z_j\in r\cal{H} $} \Big\}. \end{aligned} \end{equation} Here and hereafter we let $r\H:= \{r X_1,\dots, r X_m\}$ and $\pm r\H:= \{\pm r X_1,\dots, \pm r X_m\}$. The topology associated with $d$ will be denoted with $\t_d$. We denote instead by $d_{\textup{cc}}$ the standard \emph{Carnot--Carath\'eodory} or \emph{control} distance (see Feffermann--Phong \cite{FeffermanPhong81} and Nagel--Stein--Wainger~\cite{NagelSteinWainger}). In the present paper we shall make a prevalent use of the distance~$d$. It is well known that $\t_d$ is (possibly strictly) stronger than the topology $\t_\textup{Euc}|_\O$ received by $\cal{O}$ from $\mathbb{R}^n$. See \cite[Chapter 3]{Berhanu} and \cite[Example 5.5]{Agrachev}. In view of the mentioned examples, we need to use the broad definition of submanifold; see \cite{Chevalley,KobayashiNomizu}. Below, if $\Sigma\subset\mathbb{R}^n$, we denote by $\t_\textup{Euc}|_\Sigma$ the induced topology. \begin{definition} [Immersed submanifold] Let $\Sigma\subset\mathbb{R}^n$ and let $\t\supseteq\t_\textup{Euc}|_\Sigma$ be a topology on $\Sigma$. We say that $\Sigma$ is a $C^k$ submanifold if $\Sigma $ is connected and for all $x\in \Sigma$ there is $\Omega\in\tau$, open neighborhood of $x$ such that $\Omega$ is a $C^k$ graph. If moreover $\t=\t_\textup{Euc}|_\Sigma$ then we say that $\Sigma$ is an \emph{embedded} submanifold. \end{definition} \paragraph{Horizontal regularity classes.} Here we define our notion of horizontal regularity in terms of the distance $d$. Note that we \emph{do not} use the control distance $d_{\textup{cc}}$. \begin{definition}\label{montgomery} Let $\H :=\{ X_1, ,\dots,X_m \}$ be a family of vector fields, $X_j\in C^1_\textup{Euc}$. Let $d$ be their distance \eqref{dicocco} Let $g:\mathbb{R}^n\to \mathbb{R}$. We say that $g$ is \emph{$d$-continuous}, and we write $g\in C^0_\H(\mathbb{R}^n)$, if for all $x\in \mathbb{R}^n$, we have $ g(y)\to g(x)$, as $d(y,x)\to 0$. We say that $g:\mathbb{R}^n\to \mathbb{R}$ is \emph{$\H$-Lipschitz} or \emph{$d$-Lipschitz} in $A\subset\mathbb{R}^n$ if \begin{equation*} \Lip_\H(g; A):= \sup_{ {x,y\in A,\; x\neq y}} \frac{|g(x)- g(y)|}{d(x,y)}<\infty. \end{equation*} We say that $g\in C_\H^1(\mathbb{R}^n)$ if the derivative $X_j^\sharp g(x): = \lim_{t\to 0}(f(e^{tX_j}x) - f(x))/t$ is a $d$-continuous function for any $j=1, \dots, m$. We say that $g\in C^k_{\cal{H}}(\mathbb{R}^n)$ if all the derivatives $X_{j_1}^\sharp\cdots X_{j_p}^\sharp g$ are $d$-continuous for $p\le k$ and $j_1, \dots, j_p\in\{1,\dots,m\}$. If all the derivatives $X_{j_1}^\sharp\cdots X_{j_k}^\sharp g$ are $d$-Lipschitz on each $\Omega$ bounded set in the Euclidean metric, then we say that $g\in C^{k,1}_{\cal{H},\textup{loc}}(\mathbb{R}^n)$. Finally, denote the usual Euclidean Lipschitz constant of $g$ on $A\subset\mathbb{R}^n$ by $\Lip_{\textup{Euc}}(g; A)$. \end{definition} We will usually deal with vector fields which are of class at least $C^1_\textup{Euc}\cap C^{s-1,1}_{\H,\textup{loc}}$, where $s\ge 1$ is a suitable integer. In this case it turns out that commutators up to the order $s$ can be defined; see Definition \ref{ffff}. In the companion paper \cite{MontanariMorbidelli11a} we study several issues related with this definition. \paragraph{Definitions of commutator.} Our purpose now is to show that, given a family $\H$ of vector fields with $X_j\in C^{s-1,1}_{\H,\textup{loc}} \cap C^1_\textup{Euc}$, then commutators can be defined up to length $s$. For any $\ell\in \mathbb{N}$, denote by $\mathcal W_\ell: = \{ w_1\cdots w_\ell: w_j\in \{1,\dots, m\}\}$ the words of length $\abs{w}:=\ell$ in the alphabet $1,2,\dots,m$. Let also $\mathfrak{S}_\ell $ be the group of permutations of $\ell$ letters. Then for all $\ell\ge 1$, there are functions $\pi_\ell:\mathfrak{S}_\ell\to \{-1,0,1\}$ such that \begin{equation} \label{commototo} [A_{w_1},[A_{w_2},\dots [A_{w_{\ell-1}},A_{w_\ell}]]\dots ]= \sum_{\sigma\in \mathfrak{S}_\ell}\pi_\ell(\sigma) A_{\sigma_1(w)} A_{\sigma_2(w)}\cdots A_{\sigma_\ell(w)}, \end{equation} for all $A_1, \dots, A_m:V\to V$ linear operators on a vector space $V$. See \cite{MontanariMorbidelli11a} for a more formal definition and an in-depth discussion. We are now ready to define commutators for vector fields in our regularity classes. \begin{definition} [Definitions of commutator]\label{ffff} Given a family $\H=\{X_1,\dots, X_m\}$ of vector fields of class $C^{s-1,1}_{\H,\textup{loc}}\cap C^1_\textup{Euc} $, define for any function $\psi\in C^1_\H$ the operator \(X_j^\sharp\psi(x) := \cal{L}_{X_j}\psi(x),\) the Lie derivative. Let also $ X_j\psi(x) := f_j(x) \cdot \nabla \psi(x)$ where $\psi\in C^1_\textup{Euc}$. Moreover, let \begin{equation*} \begin{aligned} f_w &: = \sum_{\sigma\in \mathfrak{S}_\ell } {\pi}_\ell(\sigma)\big( X_{\sigma_1(w)}\cdots X_{\sigma_{\ell -1}(w)} f_{\sigma_\ell(w)} \big)\quad\text{for all $w$ with $\abs{w}\le s$,} \\ X_w \psi & :=[X_{w_1}, , \dots, [X_{w_{\ell-1}}, X_{w_\ell}]] \psi : = f_w \cdot\nabla \psi\quad\text{for all $\psi\in C^1_\textup{Euc}\quad \abs{w}\le s$,} \\ X_w^\sharp \psi& : = \sum_{\sigma\in \mathfrak{S}_\ell } {\pi}_\ell(\sigma) X_{\sigma_1(w)}^\sharp\cdots X_{\sigma_{\ell-1}(w)}^\sharp X_{\sigma_\ell(w)}^\sharp\psi \quad\text{for all $\psi\in C^{\ell}_{\cal{H}}$\quad $\abs{w}\le s-1$. } \end{aligned} \end{equation*} Finally, for any $j\in\{1,\dots, m\}$ and $w$ with $1\le \abs{w}\le s$, let \begin{equation}\label{addio} \ad_{X_j} X_w \psi : = (X_j^\sharp f_w - f_w\cdot\nabla f_j )\cdot\nabla \psi =(X_j^\sharp f_w - X_w f_j) \cdot\nabla\psi \quad\text{for all $\psi\in C^1_\textup{Euc}$.} \end{equation} \end{definition} Non-nested commutators are precisely defined in \cite{MontanariMorbidelli11a}. \begin{remark} \begin{itemize*}\item Let $Z\in \pm \H $. If $\abs{w}\le s-1$, then there are no problems in defining $\ad_Z X_w$. More precisely, in \cite{MontanariMorbidelli11a} we show that $\ad_Z X_w =[ Z, X_w]$. If instead $\abs{w}=s$, then the function $t\mapsto f_w(e^{t Z}x)$ is Euclidean Lipschitz. In particular it is differentiable for a.e.~$t$. In other words, for any fixed $x\in \mathbb{R}^n$, the limit $ \frac{d}{dt} f_w(e^{t Z}x) = : Z^\sharp f_w(e^{tZ}x )$ exists for a.e.~$t$ close to $0$. Therefore the pointwise derivative $Z^\sharp f_w(y)$ exists for almost all $y\in \mathbb{R}^n$ and ultimately $\ad_Z X_w$ is defined almost everywhere. \item Both our definitions of commutator, $X_w$ and $X_w^\sharp$ are well posed from an algebraic point of view, i.e.~they satisfy antisymmetry and the Jacobi identity; see \cite{MontanariMorbidelli11a}. \item In \cite{MontanariMorbidelli11a} we will also recognize that the first order operator $X_w$ agrees with $X_w^\sharp$ against functions $\psi\in C^{s-1,1}_{\cal{H},\textup{loc}} \cap C^1_\textup{Euc}$ as soon as $|w|\le s-1$. \end{itemize*} \end{remark} \paragraph{The integrability class $\mathcal{A}_s$.} \begin{definition}[Vector fields of class $\mathcal{A}_s$]\label{laint} Let $\H= \{X_1, \dots, X_m\} $ be a family in the regularity class $C^1_\textup{Euc}\cap C^{s-1,1}_{\H,\textup{loc}} $. We say that the family $\cal{H}$ belongs to the class $\mathcal{A}_s$ if, fixed an open bounded set $\Omega\subset \mathbb{R}^n$, there is $C_0>1$ such that the following holds: for any $Z\in \pm\H$, for any word $w$ with $\abs{w}=s$, for each $x\in \Omega$ and for a.e.~$t\in [-C_0^{-1}, C_0^{-1}]$, there are coefficients $b^v\in \mathbb{R}$ such that \begin{align}\label{arte} \ad_{Z} X_w (e^{tZ}x) & = \sum_{1\le \abs u\le s}b^u X_{u}(e^{tZ}x)\quad\text{with} \\ \label{artefatto} \abs{b^u } & \le C_0 \qquad\text{for all $u$ with $1\le \abs{u}\le s$; } \end{align} finally assume that if $1\le \abs{w}\le s$, for all $p\in\{1,\dots,n\}$, for any $I\in\mathcal{I}(p,q)$, $x\in \mathbb{R}^n$, we have at any $h^*$ where $E_{I,x}$ is defined \begin{equation}\label{sussmann} f_w(E_{I,x}(h)) \longrightarrow f_w(E_{I,x}(h^*)) \quad\text{as $h\to h^*$.} \end{equation} \end{definition} \begin{remark} \begin{itemize*} \item Assumption \eqref{sussmann} will be used only once, in \eqref{sosso}, but it is essential in order to ensure that the almost exponential maps we define later are actually $C^1_\textup{Euc}$ smooth. It is easy to check that assumption \eqref{sussmann} is satisfied as soon as $f_w: (\O_\H,\t_\H)\to \mathbb{R}$ is continuous, where $\t_\H$ denotes the Sussmann's orbit topology defined by the family $\H$, see \cite{Sussmann}. Note that at this stage assumption \eqref{sussmann} is not ensured by the $d$-Lipschitz continuity of $f_w$. \item Conditions \eqref{arte} and \eqref{artefatto} scale nicely. Namely, letting for all $r\le 1$, $\widetilde Z= r Z$, $\widetilde X_w = r^{\abs{w}}X_w$ with $\abs{w}=s$, we have \begin{equation}\begin{aligned}\label{artusi} \ad_{\widetilde Z} \widetilde X_w (x) = \sum_{1\le \abs u\le s} \widetilde b^u \widetilde X_{u}(x) \quad\text{ where $\abs{\widetilde b^u} \le C_0 r\le C_0$ for all $u$.} \end{aligned}\end{equation} \item Let $\H $ be a family of vector fields in the class $C^1_\textup{Euc}\cap C^{s-1,1}_{\H,\textup{loc}}$ satisfying the H\"ormander bracket-generating condition of step $s$ and assume that each $f_w $ with $\abs{w}\le s$ is continuous in the Euclidean sense. Then $\H$ satisfies $\mathcal{A}_s$. The constant $C_0$ in \eqref{artefatto} depends also on a positive lower bound on $\inf_\Omega\abs{\Lambda_n(x,1)}$, see \eqref{panacea}. This case is discussed in \cite[Section~4]{MontanariMorbidelli11a}. \item The pathological vector fields $X_1= \partial_{x_1} $ and $X_2= e^{-1/{x_1}^2}\partial_{x_2}, $ in spite of their $C^\infty$ smoothness, do not satisfy \eqref{artefatto} for any $s\in \mathbb{N}$. \end{itemize*} \end{remark} Let $\Omega_0\subset\mathbb{R}^n$ be a fixed open set, bounded in the Euclidean metric. Given a family $\H$ of vector fields of class $C^1_\textup{Euc}\cap C^{s-1,1}_{\H,\textup{loc}}$, introduce the constant \begin{equation} \label{lipo} \begin{aligned} L_0 :& = \sum_{j_1,\dots, j_s=1}^m\Big\{ \sup_{\Omega_0 } \Big( |f_{j_1}| + |\nabla f_{j_1}| + \sum_{p\le s} |X_{j_1}^\sharp \cdots X_{j_{p-1}}^\sharp f_{j_p}|\Big) \\ & \qquad \qquad \qquad +\Lip_\H(X_{j_1}^\sharp\cdots X_{j_{s-1}}^\sharp f_{j_s}; \Omega_0)\Big\}. \end{aligned} \end{equation} We shall always choose points $x\in\Omega\Subset\Omega_0$ and we fix a constant $t_0>0$ small enough to ensure that \begin{equation}\label{hello} e^{\t_1 Z_1}\cdots e^{\t_N Z_N}x\in \Omega_0 \quad \text{if $x\in \Omega$, $Z_j\in \H$, $\abs{\t_j}\le t_0$ and $N\le N_0 $,} \end{equation} where $N_0$ is a suitable constant which depends on the data $n, m$ and $s$. \begin{proposition}[measurability] \label{misod} Let $\H$ be a family of class $\mathcal{A}_s$. Let $\abs{w} =s$ and let $Z\in \pm\H$, Then for any $x\in \Omega$ we can write \begin{equation} \ad_{Z} X_w(e^{t Z}x) = \sum_{1\le\abs{v}\le s} b^v(t) X_v(e^{t Z}x) \qquad \text{for a.e. $t\in (-t_0, t_0)$,} \end{equation} where the functions $t\mapsto b^v(t)$ are \emph{measurable} and for a.e.~$t$ we have $\abs{b^v(t)}\le C_0$, where $C_0$ denotes the constant in \eqref{artefatto}. \end{proposition} \begin{proof} The statement can be proved arguing as in \cite[Proposition~4.1]{MontanariMorbidelli11a}. \end{proof} \paragraph{Wedge products and $\eta$-maximality conditions.} Following \cite{Street}, denote by $\cal{P} := \{ Y_1, \dots, Y_q\} = \{ X_w: 1\le \abs{w} \le s\}$ the family of commutators of length at most $s$. Let $ \ell_j\le s$ be the length of $Y_j$ and write $Y_j=:g_j\cdot\nabla$. % Define for any $p,\mu\in \mathbb{N}$, with $1\le p\le \mu$, $\mathcal{I}(p,\mu) := \{I=(i_1, \dots, i_p): 1\le i_1<i_2< \cdots<i_p\le \mu\} $. For each $x\in\mathbb{R}^n$ define $ p_x:= \dim \Span \{ Y_{j,x} : 1\le j\le q\}.$ Obviousely, $p_x\le \min\{n,q\}$. Then for any $p\in \{1,\dots,\min\{n,q\}\}$, let \begin{equation*} Y_{I,x} : = Y_{i_1,x}\wedge\cdots\wedge Y_{i_p,x}\in {\textstyle\bigwedge}_p T_x\mathbb{R}^n\sim {\textstyle\bigwedge}_p\mathbb{R}^n \quad\text{for all $I\in \mathcal{I}(p,q)$,} \end{equation*} and, for all $K\in \mathcal{I}(p,n)$ and $ I\in \mathcal{I}(p,q)$ \begin{equation}\label{girocollo} \begin{aligned} Y_I^K(x)& : = dx^K(Y_{i_1}, \dots, Y_{i_p}) (x) : = \det(g_{i_\a}^{k_\b})_{\a,\b=1,\dots, p}. \end{aligned} \end{equation} Here we let $dx^K:=dx^{k_1}\wedge \cdots \wedge d x^{k_p}$ for any $K=(k_1,\dots, k_p)\in \mathcal{I}(p,n)$. The family $e_K:= e_{k_1}\wedge\cdots\wedge e_{k_p}$, where $K\in\mathcal{I}(p,n)$, gives an othonormal basis of $\bigwedge_p\mathbb{R}^n$, i.e. $\langle e_K, e_H\rangle = \delta_{K,H}$ for all $K,H$. Then we have the orthogonal decomposition $ Y_I(x) =\sum_{K}Y_J^K(x) e_K\in {\bigwedge}_p \mathbb{R}^n $, so that the number \[ |Y_I(x)| : =\bigl(\sum_{K\in \mathcal{I}(p, n)}Y_I^K(x)^2\bigr)^{1/2} = \abs{Y_{i_1}(x)\wedge\cdots\wedge Y_{i_p}(x)} \] gives the $p$-dimensional volume of the parallelepiped generated by $Y_{i_1}(x), \dots, Y_{i_p}(x)$. Let $I= (i_1, \dots, i_p)\in \mathcal{I}(p, q)$ such that $\abs{Y_I}\neq 0$. Consider the linear system $\sum_{k=1}^p\xi^k Y_{i_k}= W$, for some $W\in\Span\{Y_{i_1}, \dots, Y_{i_p}\}$. The Cramer's rule gives the unique solution \begin{equation} \label{cromo} \xi^k = \frac{\langle Y_I, \iota^k(W) Y_I\rangle}{\abs{Y_I}^2}\quad\text{for each $k=1,\dots, p$,} \end{equation} where we let $\iota^k_W Y_I:= \iota^k(W) Y_I:= Y_{(i_1,\dots, i_{k-1})}\wedge W\wedge Y_{(i_{k+1},\dots,i_p)}.$ Let $r>0$. Given $J\in \mathcal{I}(p,q)$, let $\ell(J):= \ell_{j_1}+ \cdots + \ell_{j_p}$. Introduce the vector-valued function \begin{equation}\label{panacea} \begin{aligned} \Lambda_p(x,r)& := \bigl( Y_J^K (x)r^{\ell(J)} \bigr)_{{J\in \mathcal{I}(p, q)}, K\in \mathcal{I}(p, n)} =: \bigl( \widetilde Y_J^K (x) \bigr)_{{J\in \mathcal{I}(p, q),\, K\in\mathcal{I}(p, n)}}, \end{aligned} \end{equation} where we adopt the tilde notation $\widetilde Y_k = r^{\ell_k}Y_k $ and its obvious generalization for wedge products. Note that $\abs{\Lambda_p(x,r)}^2 = \sum_{I\in \mathcal{I}(p, q)} r^{2\ell(I)}\abs{Y_I(x)}^2$. \begin{definition} [$\eta$-maximality] Let $x\in\mathbb{R}^n$, let $I\in \mathcal{I}(p_x,q)$ and $\eta\in (0,1)$. We say that $(I,x,r)$ is $\eta$-maximal if $\abs{Y_I (x)} r^{\ell (I)} >\eta \displaystyle \max_{J\in \mathcal{I}(p_x, q)} \abs{Y_J(x)}r^{\ell(J)}. $ \end{definition} Note that, if $(I,x,r)$ is a candidate to be $\eta$-maximal with $I\in \mathcal{I}(p,q) $, then by definition it \emph{must} be $p=p_x=\dim\Span\{Y_j(x):1\le j\le q\}$. \paragraph{Approximate exponentials of commutators.} \label{capuozzo} Let $w_1, \dots, w_\ell\in \{1,\dots,m\}$. Given $\t>0$, we define, as in \cite{NagelSteinWainger,Morbidelli98} and \cite{MM}, \begin{equation}\label{navetta} \begin{aligned} C_\t( X_{w_1})& := \exp(\t X_{w_1}), \\ C_\t( X_{w_1}, X_{w_2})& :=\exp(-\t X_{w_2})\exp(-\t X_{w_1})\exp(\t X_{w_2})\exp(\t X_{w_1}), \\&\vdots \\C_\t( X_{w_1}, \dots, X_{w_\ell})& :=C_\t( X_{w_2}, \dots, X_{w_\ell})^{-1}\exp(-\t X_{w_1}) C_\t( X_{w_2}, \dots, X_{w_\ell})\exp(\t X_{w_1}). \end{aligned} \end{equation} Then let \begin{equation}\label{appsto} \espo_{\textup{ap}}^{tX_{w_1 w_2\dots w_\ell}} := \exp_{\textup{ap}}(t X_{{w_1 w_2\dots w_\ell}}):= \left\{\begin{aligned} & C_{t^{1/\ell}}(X_{w_1}, \dots, X_{w_\ell} ), \quad &\text{ if $t\geq 0$,} \\ &C_{|t|^{1/\ell}}(X_{w_1}, \dots, X_{w_\ell} )^{-1}, \quad &\text{ if $t<0$.} \end{aligned}\right. \end{equation} By standard ODE theory, there is $t_0$ depending on $\ell, \Omega$, $\Omega_0$, $ \sup\abs{ f_j } $ and $\sup\abs{\nabla f_j} $ such that $\exp_*(t X_{{w_1 w_2\dots w_\ell}})x\in\Omega_0$ for any $x\in \Omega$ and $|t|\le t_0$. Define, given $I= (i_1,\dots,i_p)\in\{1,\dots,q\}^p $, $x\in \Omega$ and $h\in\mathbb{R}^p$, with $|h|\le C^{-1}$ \begin{equation} \begin{aligned} \label{hhh}E_{I,x}(h)& :=\exp_{\textup{ap}}(h_1 Y_{i_1})\cdots \exp_{\textup{ap}}(h_p Y_{i_p})(x) \\ \bigl\|h\bigr\|_I & : =\max_{j=1,\dots,p}|h_j|^{1/\ell_{i_j} }\quad \text{and}\quad Q_I(r): =\{h\in\mathbb{R}^p:\norm{h}_I < r\}. \end{aligned}\end{equation} \paragraph{Gronwall's inequality.} We shall refer several times to the following standard fact: for all $a\ge 0$, $b>0$, $T>0$ and $f$ continuous on $[0,T]$, \begin{equation}\label{grammatica} 0\le f(t)\le a t+ b\int_0^t f(\t) d\t \quad\forall \; t\in [0, T]\quad\Rightarrow \quad f(t)\le \frac{a}{b} (e^{b t}-1)\quad\forall\, t\in[0,T]. \end{equation} \section{Approximate exponentials and regularity of \texorpdfstring{$A_s$}{As} orbits} \label{kobayashi} Let $\H = \{X_1, \dots, X_m\}$ be a family of $\mathcal{A}_s$ vector fields in $\mathbb{R}^n$. The main purpose of this section is to prove that any $\H$-orbit $\cal{O}_\H$ with the topology $\t_d$ generated by the distance~$d$ is a $C^1$ integral manifold of the distribution generated by $\P$. Recall our usual notation $ \cal{P} : = \{ Y_j: 1\le j\le q\}$, $ P_x :=\Span\{Y_{j,x}: 1\le j\le q\}$ and $p_x:= \dim P_x.$ \subsection{Geometric properties of orbits}\label{fiveone} In this subsection we look at the properties of orbits $\O_{\H}$ for vector fields of class $\mathcal{A}_s$. First we study how the geometric determinants $\widetilde Y_J^K$ change along a given orbit $\O_\H$. The argument we use is known, see for instance \cite{TaoWright03,MM} and especially \cite{Street}. However, we need to address some issues which appear due to our low regularity assumptions. Ultimately, we will show that the positive integer $p_x$ is constant as $x\in\O_\H$. Below we shall use the following notation: given $r>0$, we let $\widetilde Y_j = r^{\ell_j} Y_j = : \widetilde g_j\cdot\nabla$ and $\widetilde Z = r Z$, if $Z\in \pm \cal{H}$. Let also $\widetilde Y_J^K:=r^{\ell(J)}Y_J^K$, where the notation for $Y_J^K$ has been introduced in \eqref{girocollo}. \begin{lemma} \label{lili} Let $\H$ be a family of vector fields of class $\mathcal{A}_s$. Let $ p\in\{1,\dots, q \wedge n\}$. Let $x\in \Omega$ and $r_0>0$ so that $B_d(x,r_0)\subset\Omega_0$. Let $J\in \mathcal{I}(p , q)$, $K\in \mathcal{I}(p, n)$, $r\in (0, r_0]$ and $\widetilde Z\in \pm r\H$. Then the function $[-1,1]\ni t\mapsto \widetilde Y_J^K(e^{t \widetilde Z}x)$ is Lipschitz continuous and there is $C>1$ depending on $C_0$ and $L_0$ in \eqref{artefatto} and \eqref{lipo} such that \begin{equation*} \Bigl|\frac{d}{dt}\widetilde Y_J^K (e^{t \widetilde Z}x) \Bigr|\le C\abs{\Lambda_p(e^{t \widetilde Z}x, r)} \quad\text{ for a.e. $t\in (-1, 1)$}. \end{equation*} \end{lemma} \begin{proof} Denote $\gamma_t:=e^{t \widetilde Z}x$ and let $t, \t\in (-1, 1)$. Then \begin{equation*} \begin{aligned} |\widetilde Y_J^K(\gamma_\t) -\widetilde Y_J^K(\gamma_t)| & = \Big| \sum_{1\le \a\le p} dx^{K}(\dots, \widetilde Y_{j_{\a+1}}(\gamma_t), \widetilde Y_{j_\a}(\gamma_\t)- \widetilde Y_{j_\a}(\gamma_t), \widetilde Y_{j_{\a+1}}(\gamma_t), \dots )\Big| \\&\le C \abs{\t- t}, \end{aligned} \end{equation*} where $C$ depends on $L_0$ in \eqref{lipo}. Then $t\mapsto \widetilde Y_J^K(\gamma_t)$ belongs to $\Lip_\textup{Euc}(-1, 1)$. The estimate for the Lipschitz constant here is quite rough and it can be refined through a computation of the derivative. Indeed, we claim that for a.e. $t\in (-1, 1)$ we have \begin{equation}\begin{aligned}\label{maino} \frac{d}{dt} \widetilde Y_J^K(\gamma_t) & = \sum_{\substack{1\le\a\le p\\ \ell_{j_\a}\le s-1}} dx^K(\dots, \widetilde Y_{j_{\a-1}}, [{\widetilde Z}, \widetilde Y_{j_\a}], \widetilde Y_{j_{\a+1}}, \dots, \widetilde Y_{j_p})(\gamma_t) \\& +\sum_{\substack{ 1\le \a\le p\\ \ell_{j_\a}= s}} \sum_{1\le \b\le q} b_{\a }^\b(\gamma_t) dx^K(\dots, \widetilde Y_{j_{\a-1}}, \widetilde Y_\b, \widetilde Y_{j_{\a+1}}, \dots, \widetilde Y_{j_p})(\gamma_t) \\&+ \sum_{1\le \gamma\le n}\sum_{1\le \b\le p} \partial_\gamma \widetilde f^{k_\b} dx^{(k_1,\dots,k_{\b-1},\gamma,k_{\b+1},\dots, k_p)} ( \widetilde Y_{j_1},\dots,\widetilde Y_{j_p})(\gamma_t) \\&=: (A)+(B)+(C), \end{aligned} \end{equation} where we wrote $\widetilde Z= \widetilde f\cdot\nabla\in C^1_\textup{Euc}$ and $b_\a^\b$ are measurable functions with $\abs{b_{\a }^\b}\le C_0$. To prove \eqref{maino}, observe that, if $\ell(Y_{j_\a})\le s-1$, then $t\mapsto \widetilde Y_{j_\a}(\gamma_t)$ is $C^1_\textup{Euc}(-1, 1)$ and \begin{equation*} \begin{aligned} \lim_{\t\to t}\frac{ \widetilde Y_{j_\a}(\gamma_\t)- \widetilde Y_{j_\a}(\gamma_t)}{\t- t} & = {\widetilde Z}^\sharp \widetilde g_{j_\a}(\gamma_t) \cdot\nabla = [{\widetilde Z} , \widetilde Y_{j_\a}](\gamma_t ) +\widetilde Y_{j_\a} \widetilde f(\gamma_t)\cdot\nabla \quad \text{for all $t\in [-1, 1]$. } \end{aligned} \end{equation*} Note that here we used \cite[Theorem~3.1]{MontanariMorbidelli11a} to claim that $\ad_{\widetilde Z}\widetilde Y_{j_\a} = [\widetilde Z, \widetilde Y_{j_\a}]$. If instead $\ell(Y_{j_\a})= s$, then for almost any $t$ we have \begin{equation}\label{gigos} \begin{aligned} \lim_{\t\to t} \frac{\widetilde Y_{j_\a}(\gamma_\t)- \widetilde Y_{j_\a}(\gamma_t)}{\t- t} & = {\widetilde Z}^\sharp \widetilde g_{j_\a}(\gamma_t) \cdot\nabla = \ad_{{\widetilde Z}}\widetilde Y_{j_\a}(\gamma_t )+ \widetilde Y_{j_\a} \widetilde f (\gamma_t)\cdot\nabla \\& =\sum_{\b=1}^q b_\a^\b(t) \widetilde Y_\b(\gamma_t) + \widetilde Y_{j_\a}\widetilde f(\gamma_t)\cdot\nabla. \end{aligned} \end{equation} In the first equality we used the definition of $\ad$. Here $\widetilde Y_{j_\a}\widetilde f:= \widetilde g_{j_\a}\cdot\nabla \widetilde f$, is well defined. In the second line we used Proposition \ref{misod}. The term $\widetilde Y_{j_\a}\widetilde f$, in view of Lemma \ref{ollo} gives the third line of \eqref{maino}. Next we estimate each line of \eqref{maino}, starting with $(A)$. \begin{equation*} \begin{aligned} |(A)|& \le \big|dx^K(\dots, \widetilde Y_{j_{\a-1}}(\gamma_t), [{\widetilde Z}, \widetilde Y_{j_\a}](\gamma_t), \widetilde Y_{j_{\a+1}}(\gamma_t), \dots ) \big| \le C \abs{\Lambda_p(\gamma_t, r)}, \end{aligned} \end{equation*} for all $t\in[-1, 1]$. Estimate is correct even if $\Lambda_p(\gamma_t, r)=0$. To estimate $(B)$, recall that $|b_{\a }^\b|\le C$. Then, for all $t\in[-1,1]$, \begin{equation*} \begin{aligned} |(B)| &\le \sum_{1\le\a\le p} \sum_{1\le \b\le q} \big| dx^K(\dots,\widetilde Y_{j_{\a-1}}, \widetilde Y_\b, \widetilde Y_{j_{\a+1}}, \dots)\big| \le C\abs{\Lambda_p(\gamma_t,r)}. \end{aligned} \end{equation*} Finally the estimate of $(C)$ is easy and takes the form \begin{align*} \abs{(C)}& \le \sup_{B_d(x, r)}|\nabla \widetilde f|\max_{K\in \mathcal{I}(p,n)}|\widetilde Y_J^K(\gamma_t)| \le C\abs{\Lambda_p(\gamma_t, r)}\quad \text{if $\abs{t}\le 1$.}\qedhere \end{align*} \end{proof} The previous lemma immediately implies the following proposition. \begin{proposition} \label{valida} Let $\H$ be a family in the regularity class $\mathcal{A}_s$. Let $x\in\Omega$, let $r\le r_0$, where $r_0$ is small enough so that $B_d(x, r_0)\subset\Omega_0$. Let $\gamma(t):=\gamma_t$ be a piecewise integral curve of $\pm r\H$ with $\gamma(0)=x$. Let $p\in\{1,\dots, q\wedge n\}$. Then we have \begin{equation}\label{spray} \big|\Lambda_p(\gamma(t),r)- \Lambda_p(x,r)\big|\le \abs{\Lambda_p(x,r)}\,(e^{Ct}-1)\quad\text{for all $t\in[0,1]$}. \end{equation} In particular, if $p= p_x$ and $(I, x, r)$ is $\eta$-maximal, then \begin{equation}\label{damettere} \abs {\widetilde Y_J(\gamma(t))- \widetilde Y_J(x)} \le \frac{C t}{\eta}\abs{\widetilde Y_I(x)} \quad\text{for all $J\in \mathcal{I}(p,q)\quad t\in[0,1]$.} \end{equation} Finally, if $x,y$ belong to the same orbit, then $p_x = p_{y}$. \end{proposition} \begin{remark}\label{osservo} As a consequence of the proposition and of the Cramer's rule \eqref{cromo}, if $(I,x,r)$ is $\eta$-maximal, then $(I, y, r)$ is $ C^{-1}\eta$-maximal for all $y\in B_d(x, C^{-1}\eta r)$ and we may write for all such $y$ and for any $j\in\{1,\dots,q\}$ \begin{equation} \label{giorgetto} \widetilde Y_{j,y} = \sum_{k=1}^p \frac{b_j^k }{\eta}\widetilde Y_{i_k, y}, \end{equation} where $\abs{b_j^k}\le C$. \end{remark} \begin{remark}\label{optima} Proposition \ref{valida} shows that the oscillation of determinants $\Lambda_p$ on a ball is controlled in terms of the value of $\Lambda_p$ at the center of the ball. It is not true that the oscillation of a single vector field on a ball can be controlled by its value at the center of the ball. For instance, we can take the vector fields $X=\partial_x$ and $Y=y\partial_y + x\partial_x$. Look at the ball $B((0,y), r)$, where $0< y \ll r$. Note that $(r,y) $ belongs to such ball, but the oscillation $\abs{Y(0,y)- Y(r,y)} \sim r$ can not be controlled with the value $\abs{Y(0,y)} = \abs{y}$. \end{remark} \begin{proof}[Proof of Proposition \ref{valida}] (See \cite{TaoWright03,MM,Street}). Let $p\in \{1,\dots,q\wedge n\}$. By Lemma~\ref{lili}, the map $t\mapsto \Lambda_p(\gamma_t,r) $ is Lipschitz. Moreover, we have for a.e. $t\in[0, 1]$, \begin{equation*} \begin{aligned} \Big|\frac{d}{d t}\Lambda_p(\gamma_t,r)\Big| &= \Big|\Big(\frac{d}{d t}\widetilde Y_J^K(\gamma_t)\Big)_{\substack{J\in \mathcal{I}(p,q)\\K\in\mathcal{I}(p,n)}}\Big| \le C \abs{\Lambda_p(\gamma_t,r)}, \end{aligned} \end{equation*} by Lemma \ref{lili}. Then the Gronwall's inequality \eqref{grammatica} provides immediately the required estimate \eqref{spray}. Note that this implies that if $\Lambda_p(x,r)=0$, then $\Lambda_p(\gamma_t,r)=0$ for all $t\in[0, 1]$. Estimate \eqref{damettere} follows immediately. Let now $x $ and $y$ be a couple of points on the same leaf $\cal{O}_\H$. Let $1\le p\le q\wedge n$ and let $I\subset\mathbb{R}$ be an interval. Let $I= [a,b]$ and take $\gamma:I\to \mathbb{R}$ a piecewise integral curve of the vector fields $X_j$ with $\gamma(a)= x$ and $\gamma(b)=y$. Let $A_p:=\{t\in I:\abs{\Lambda_p(\gamma(t))}=0\}$. Note that $A_p$ is closed, because it is the zero set of the continuous function $I \ni t\mapsto \abs{\Lambda_p(\gamma(t))}\in\mathbb{R}$. The set $A_p$ is also open by estimate \eqref{spray}. Therefore, either $A_p=\varnothing$ or $A_p= I$ and the proof is concluded. \end{proof} The fact we are going to establish in the following theorem will have a key role in Subsection~\ref{pastorius}, when we shall study our almost exponential maps $E$. See Remark~\ref{noterella} below. \begin{theorem} \label{viola} Let $\H$ be a family of vector fields of class $\mathcal{A}_s$. Let $(I, x, r)$ be $\eta$-maximal where $x\in \Omega$, $r\le r_0$, $I\in \mathcal{I}(p_x, q)$ and $\eta\in (0,1)$. Denote $\widetilde U_j:= r^{\ell_{i_j}} Y_{i_j}$ for $j=1,\dots, p:= p_x$ and $\widetilde Z:= r Z\in \pm r\H $. Then there is $C>0$ depending on $L_0$ and $C_0$ in \eqref{lipo} and \eqref{artefatto} so that \begin{equation}\label{dentro} e^{-t\widetilde Z}_*(\widetilde U_{j, e^{t\widetilde Z}x})\in P_x\quad\text{for all $t$ with $\abs{t}\le C^{-1}\eta$}. \end{equation} Moreover, if we write, for a given test function $\psi\in C^1_\textup{Euc}(\mathbb{R}^n)$, \begin{equation}\label{75} \widetilde U_j(\psi e^{-t \widetilde Z})(e^{t \widetilde Z}x) = : \sum_{k=1}^p \big(\delta_j^k + \theta_j^k(t) \big) \widetilde U_k\psi(x), \end{equation} then we have \begin{equation}\label{zast} |\theta_j^k(t)| \le \frac{C |t|}{\eta} \quad\text{for all $j,k=1,\dots,p$\qquad $ \abs{t} \le C^{-1}\eta$}. \end{equation} Finally, for any commutator $\widetilde Y_h := \widetilde g_h\cdot\nabla$, where $h\in \{1,\dots,q\}$, we have at any $t\in (-C^{-1}\eta , C^{-1}\eta)$ \begin{equation} \label{909} \widetilde Y_h(\psi e^{-t\widetilde Z})(e^{t\widetilde Z}x) = \sum_{k=1}^p \frac{b_h^k(t)}{\eta}\widetilde U_k\psi(x), \end{equation} where $ |b_h^k(t)|\le C$ if $\abs{t}\le C^{-1}\eta$. \end{theorem} \begin{remark}\label{noterella} The geometric interpretation of \eqref{dentro} tells that $e^{-t \widetilde Z}_*P_{e^{t\widetilde Z}x} = P_x$, i.e. the tangent map of the $C^1$ diffeomorphism $e^{-t\widetilde Z}$ maps the (candidate) tangent bundle $\cup_x P_x $ to the orbit $\cal{O}$ to itself (we say ``candidate'' because we do not know yet that $\O$ is a manifold). Theorem \ref{viola} has an important consequence. Namely, in in Theorem \ref{dascalare}, it will enable us to show that integral remainders have in fact a pointwise form. Ultimately, we will apply such property in Theorem \ref{deduciamo} to show that $E_*(\partial_{h_k})\in P_{E(h)}$. \end{remark} \begin{remark}\label{lobrao} The proof below is inspired to an argument due to Lobry; see \cite[Lemma~1.2.1]{Lobry70}. Here we generalize such argument to a higher order, nonsmooth situation and we get more quantitative estimates. See also \cite{Lobry76} and the related discussion by Balan \cite{Balan}; see finally the paper~\cite{Pelletier}, for an up-to-date bibliography on the subject. Note that Lobry's idea is also used in \cite[Lemma 5.15]{Agrachev}. \end{remark} \begin{proof}[Proof of Theorem \ref{viola}] Without loss of generality, we can work with positive values of $t$. First, we differentiate the left-hand side of \eqref{75}. If $\ell_{i_j}\le s-1$, then we use \cite[Theorem~2.6-(a) and Theorem~3.1-(ii)]{MontanariMorbidelli11a} which give \begin{equation}\label{popo} \begin{aligned} \frac{d}{dt}\widetilde U_j(\psi e^{-t \widetilde Z})(e^{t \widetilde Z} x)& =[\widetilde Z, \widetilde U_j](\psi e^{-t \widetilde Z})(e^{t\widetilde Z}x) = \sum_{k=1}^p \frac{b_j^k(t)}{\eta} \widetilde U_k(\psi e^{-t\widetilde Z}) (e^{t\widetilde Z}x), \end{aligned} \end{equation} provided that $0<t\le C^{-1}\eta $. Here $\abs{b_j^k(t)}\le C$. In last equality we used \eqref{giorgetto} with $\widetilde Y_h = [\widetilde Z, \widetilde U_j]$. If instead $\ell_{i_j}= s$, then we need first \cite[Theorem 2.6-(b)]{MontanariMorbidelli11a}, then \eqref{sussmann} and Proposition \ref{misod} in the present paper. This gives for a.e. $t\in[0, C^{-1}\eta]$ \begin{equation}\label{popo2} \begin{aligned} \frac{d}{dt}\widetilde U_j (\psi e^{-t \widetilde Z})(e^{t\widetilde Z} x)& = \sum_{1\le h\le q } b_j^h(t) \widetilde Y_h (\psi e^{-t \widetilde Z})(e^{t\widetilde Z}x) \quad \text{by \eqref{giorgetto}} \\&= \sum_{1\le h\le q}\sum_{1\le k\le p} b_j^h(t) b_h^k(t)\frac{ 1}{\eta} \widetilde U_k (\psi e^{-t \widetilde Z})(e^{t \widetilde Z}x) \\&=:\sum_{1\le k\le p} \frac{b_j^k(t)}{\eta} \widetilde U_k(\psi e^{-t \widetilde Z})(e^{t \widetilde Z}x) \end{aligned} \end{equation} provided that $0<t\le C^{-1}\eta $. In this formula $b_j^h$, $b_h^k$ and $b_j^k$ denote measurable functions, bounded in term of the admissible constants $C_0$ and $ L_0$. By elementary ODE theory, for any fixed $\psi$, the functions $t\mapsto \widetilde U_j(\psi e^{-t\widetilde Z})(e^{t\widetilde Z}x) $ with $j=1,\dots,p$ are uniquely determined by their value $\widetilde U_j\psi(x)$ at $t=0$. Moreover, if we denote by $(a_j^k(t))\in \mathbb{R}^{p\times p}$ the solution of the Cauchy problem \begin{equation}\label{odofusi} \dot a(t)= \frac{b(t)}{\eta} a(t) \quad\text{with}\quad a(0)= I_p\in\mathbb{R}^{p\times p}, \end{equation} then we can write \begin{equation} e_*^{-t \widetilde Z}(\widetilde U_{j, e^{t\widetilde Z}x})\equiv \widetilde U_j(\psi e^{-t \widetilde Z})(e^{t\widetilde Z}x) = \sum_{k=1}^p a_j^k(t) \widetilde U_k\psi(x). \end{equation} Then we have proved \eqref{dentro}. The Cramer's rule \eqref{cromo} confirms that the coefficients $a_j^k(t)$ are unique for each $t$. To estimate the functions $\theta_j^k:= a_j^k(t)- \delta_j^k $, where $a_j^k$ satisfy \eqref{odofusi}, it suffices to use estimate $|b_{j}^k(t)|\le C$ if $0\le t \le C^{-1}\eta $. The Gronwall inequality \eqref{grammatica} gives \( |a_j^k(t) - \delta_j^k|\le C |t|/\eta \) for all $j,k=1,\dots, p$ and $ 0< t\le C^{-1}\eta .$ Therefore \eqref{zast} follows. To obtain the proof of \eqref{909} it suffices to repeat the computation in \eqref{popo} starting from $\widetilde Y_h$ instead of $\widetilde U_j$. This ends the proof. \end{proof} Under the hypotheses of Theorem \ref{viola}, iterating the argument, we get for all $x\in \Omega$, $\mu\le N_0$ (see \eqref{hello}), $j\in\{1,\dots,p\}$ and $Z_1,\dots, Z_\mu\in\H$, \begin{equation}\label{jappo} \widetilde U_j(\psi e^{-t_1 \widetilde Z_1}\cdots e^{-t_\mu \widetilde Z_\mu})(e^{t_\mu \widetilde Z_\mu}\cdots e^{t_1 \widetilde Z_1}x) = \sum_{1\le k\le p} (\d_j^k +\theta_j^k(t))\widetilde U_k\psi(x) \end{equation} where $\abs{\theta(t)}\le C \abs{t}/\eta$, as soon as $\sum_{j=1}^\mu\abs{t_j}\le C^{-1}\eta$. Moreover, for each $h\in \{1,\dots,q\}$, we get, if $x\in \Omega $, for the same values of $(t_1,\dots, t_\mu)$ and for almost all $\t\in (- C^{-1}\eta, C^{-1}\eta)$, \begin{equation*} \begin{aligned} \frac{d}{d\t}\widetilde Y_h& (\psi e^{-t_1 \widetilde Z_1} \cdots e^{-t_\mu \widetilde Z_\mu}e^{-\t \widetilde X}) ( e^{\t \widetilde X}e^{t_\mu \widetilde Z_\mu}\cdots e^{t_1 \widetilde Z_1}x) \\&= \ad_{\widetilde X}\widetilde{Y}_h(\psi e^{-t_1 \widetilde Z_1}\cdots e^{-t_\mu \widetilde Z_\mu} e^{-\t \widetilde X}) (e^{\t \widetilde X}e^{t_\mu \widetilde Z_\mu }\cdots e^{t_1 \widetilde Z_1}x) =\sum_{k=1}^p \frac{b_k(x,t,\tau )}{\eta}\widetilde U_k\psi(x), \end{aligned} \end{equation*} where $\abs{b_k(x,t,\t)}\le C$ for a.e.~$\t$. Here $X\in \H$. If we do not care about maximality and choose $r= 1$, we get, for any fixed $(t_1, \dots, t_\mu)$ with $\sum_j\abs{t_j}\le C^{-1}$ and for almost all $\t $ with $\abs{\t}\le C^{-1}$, \begin{equation} \label{jpn4} \begin{aligned} \frac{d}{d\t} Y_h & (\psi e^{-t_1 Z_1}\cdots e^{-t_\mu Z_\mu}e^{-\t X}) (e^{\t X}e^{t_\mu Z_\mu}\cdots e^{t_1 Z_1}x) \\ & = \ad_{ X} {Y_h} (\psi e^{-t_1 Z_1}\cdots e^{-t_\mu Z_\mu} e^{-\t X}) (e^{\t X}e^{t_\mu Z_\mu}\cdots e^{t_1 Z_1}x) \\& =\sum_{1\le j\le q} b_j(x,t,\tau) Y_j\psi(x), \end{aligned} \end{equation} where $\abs{b_j (x,t,\t)}\le C$ for a.e.~$\t$. Here again $x\in \Omega$ and $\psi\in C^1_\textup{Euc}$ is a test function. Formula \eqref{jpn4} will be referred to later. \subsection{Derivatives of almost exponential maps and regularity of orbits} \label{pastorius} In this subsection we get several information on the derivatives of the approximate exponentials $E_{I,x,r}$ associated with a family $\H$ of $\mathcal{A}_s$ vector fields and we show that each orbit $\cal{O}$ with topology $\t_d$ is a $C^1$ immersed submanifold of $\mathbb{R}^n$ with $T_y\O= P_y$ for all $y\in\O$. We will tacitly but heavily rely on the results of \cite[Section~3]{MontanariMorbidelli11a}, namely on formulae \begin{equation}\label{usbus} \ad_{X_{v_1}}\cdots \ad_{X_{v_k}}X_w = X_{vw}\quad\text{for all $v,w$ such that $\abs{v}+ \abs{w} = k+\abs{w}\le s$} \end{equation} These formulae have a key role. In the proof of Theorem \ref{dascalare} below, we shall follow the arguments of \cite[Theorems 3.4 and 3.5]{MM}, modifying everywhere the remainders $O_{s+1}$ in \cite{MM} with our remainders defined in \cite{MontanariMorbidelli11a}. This will give us a formula with integral remainder, see \eqref{grall}. Then, using the results of Subsection~\ref{fiveone}, we shall show that such integral remainder can be specified in a pointwise form. \begin{theorem}\label{dascalare} Let $1\le \abs{w}=:\ell\le s$, take $x\in\Omega$ and $ t\in [0, t_0]$, where $t_0$ is small enough to ensure that $C_t x\in \Omega_0$ for all $t\in[0, t_0]$. Let $C_t=C_t(X_{w_1}, \dots, X_{w_\ell})$ be the map defined in \eqref{navetta}. Fix a test function $\psi\in C^1_\textup{Euc}(\mathbb{R}^n)$. Then we have \begin{equation*} \frac{d}{dt}\psi (C_t x )= \ell t^{\ell-1} X_w \psi(C_tx) +\sum_{|v|=\ell+1}^s a_v t^{|v|-1} X_v\psi(C_tx) +t^s\sum_{\abs{u}=1}^s b_u(x,t)X_u\psi(C_t x), \end{equation*} and \begin{equation*} \begin{aligned} \frac{d}{dt}\psi(C_t^{-1} x) = & - \ell t^{\ell-1} X_w \psi(C_t^{-1}x) +\sum_{|v|=\ell+1}^s \overline a_v t^{|v|-1} X_v \psi(C_t^{-1}x) \\& + t^s \sum_{\abs{u}=1}^s \overline b_u(x,t)X_u\psi (C_t^{-1}x). \end{aligned} \end{equation*} Both the sums on $v$ are empty if $\abs{w}=s$. Otherwise, we have the cancellations $\sum_{\abs{v}=\ell+1}(a_v+\overline a_v)f_v(x)=0$ for all $x\in \Omega$. The (real) coefficients $b_u$ and $\overline b_u$ are bounded in terms of the constants $L_0$ and $C_0$ in \eqref{lipo} and \eqref{artefatto}. \end{theorem} \begin{remark} As already observed, the theorem just stated improves \cite[Theorem~3.5]{MM}, both because we relax regularity assumptions and because we devise a pointwise form of the remainders. In particular, choosing as $\psi$ the identity function, we see that the remainder belongs to the subspace $P_{C_tx}=\Span\{Y_{j,C_t x}: j=1,\dots, q\}$ which can be a strict subspace of $\mathbb{R}^n$. \end{remark} \begin{proof}[Proof of Theorem~\ref{dascalare}] We prove the statement for $t>0$. By \cite[Theorem~3.5]{MM}, we know that \begin{equation}\label{grall} \frac{d}{dt}\psi (C_t x )= \ell t^{\ell-1} X_w \psi(C_tx) +\sum_{|v|=\ell+1}^s a_v t^{|v|-1} X_v\psi(C_tx) +O_{s+1}(t^s,\psi, C_tx), \end{equation} where the numbers $a_v$ are suitable algebraic coefficients. Note that formula \eqref{grall} in \cite{MM} is proved for smooth vectro fields. Using \eqref{usbus} and changing everywhere the remainders in \cite{MM} with the remainders introduced in \cite[Subsection~2.1]{MontanariMorbidelli11a}, one can check that all computations fit to our setting. Therefore, we only need to deal with the integral remainders introduced and discussed in~\cite{MontanariMorbidelli11a}. Concerning such remainders, recall that \begin{equation* \begin{aligned} O_{s+1}(t^s, \psi,C_t x)& =\text{(sum of terms like)} \int_0^t \omega(t, \t)\frac{d}{d\t} X_v(\psi\phi^{-1}e^{-\t Z})(e^{\t Z}\phi C_t x) d\t \end{aligned} \end{equation*} where $\abs{v}=s$, $\phi = e^{t Z_1 }\cdots e^{t Z_\nu}$ and $Z, Z_j\in \pm \H $. Next, by \eqref{jpn4}, we may write for a.e.~$\t$ \begin{align*} \frac{d}{d\t} X_v(\psi\phi^{-1}e^{-\t Z})(e^{\t Z}\phi C_t x) & = \sum_{1\le \abs{u}\le s} b_u(x,t,\tau)X_u\psi(C_t x), \end{align*} where for any $t,x$ the functions $\t\mapsto b_u(x,t,\tau)$ are measurable and satisfy $\abs{ b_u(t,\tau,x)}\le C$ for a.e.~$\t$. Therefore we get \begin{equation*} \sum_{1\le \abs{u}\le s}\int_0^t \omega(t, \t) b_u(x,t,\tau) d\t \; X_u \psi( C_t x)=: t^s \sum_{1\le \abs{u}\le s} b_u(x,t) X_u\psi(C_t x), \end{equation*} where $\abs{b_u(x,t)}\le C$ for all $x\in\Omega$ and $\abs{t}\le t_0$. This ends the proof. \end{proof} Our purpose now is to study the maps \begin{equation}\label{fivetwenty} E(h):= E_{I,x,r}(h) := \exp_{\textup{ap}}(h_1 \widetilde Y_{i_1})\cdots \exp_{\textup{ap}}(h_p \widetilde Y_{i_p})= \espo_{\textup{ap}}^{h_1 \widetilde U_1}\cdots \espo_{\textup{ap}}^{h_p \widetilde U_p}x \end{equation} where $1\le p\le q$, $I\in \mathcal{I}(p,q)$, $\widetilde U_k:= \widetilde Y_{i_k}$ and $d_k:= \ell_{i_k}$. We always take $x\in \Omega$ and $h$ sufficiently close to the origin so that $E(h)\in \Omega_0$, see \eqref{hello}. Some elementary properties of $E$ are contained in the following lemma. Without loss of generality we choose $r=1$ and $I= (1,\dots, p)$. \begin{lemma}\label{torrette} The map $ h \mapsto \espo_{\textup{ap}}^{h_1 Y_1}\cdots\espo_{\textup{ap}}^{h_p Y_p}x=:E_{I,x}(h)$ satisfies for $x,x^*\in \Omega$ and $h, h^*\in B_\textup{Euc}(C^{-1})$ \begin{equation}\label{galileo} \abs{E_{I,x}(h)- E_{I, x^*}(h^*)} \le C\big(\bigl\|h-h^*\bigr\|_I+\abs{x-x^*}\big). \end{equation} Moreover, for any $w$ with $1\le \abs{w}\le s$, the function $F_{X_w}\colon[-C^{-1},C^{-1}]\times \Omega\to \mathbb{R}^{n\times n}$, defined as $F_{X_w}(t,x ):= \nabla _x \espo_{\textup{ap}}^{t X_w}(x),$ is continuous. \end{lemma} \begin{proof} Observe first that, since each $Z\in \pm \H$ is $C^1_\textup{Euc}$, by the Gronwall inequality we have \begin{equation}\label{baio} \abs{e^{\t Z}y - e^{ \t_0 Z}y_0} \le C\big(\abs{y- y_0} + \abs{\t-\t_0}\big)\quad\text{for all $y,y_0\in \Omega\quad \abs{\t},\abs{\t_0}\le C^{-1}$. } \end{equation} Next, assume first that $t\ge t^*\ge 0$. Write $\espo_{\textup{ap}}^{t X_w}x= e^{\t Z_1}\cdots e^{\t Z_\nu}x$, where $Z_1, \dots, Z_\nu\in \pm\H$ are suitable, see \eqref{appsto}, and $\t = t^{1/\ell}$, with $\ell:=\abs{w}$. Then iterating \eqref{baio} we get \begin{equation*} \begin{aligned} \bigl|\espo_{\textup{ap}}^{t X_w}x - \espo_{\textup{ap}}^{t^* X_w}x^*\bigr| &= \bigl|e^{\t Z_1}\cdots e^{\t Z_\nu}x - e^{\t^* Z_1}\cdots e^{\t^* Z_\nu}x^*\bigr| \le C\bigl(\abs{x- x^*} + \abs{t - t^*}^{1/\ell} \bigr). \end{aligned} \end{equation*} If instead $t>0>t^*$, then we get \begin{align*} \bigl|\espo_{\textup{ap}}^{t X_w}x - \espo_{\textup{ap}}^{t^* X_w}x^*\bigr| &\le \abs{\espo_{\textup{ap}}^{t X_w}x - x} + \abs{ x ^*- \espo_{\textup{ap}}^{t^* X_w}x^*}+ \abs{x- x^*} \\&\le C\bigl(\abs{t}^{1/\ell} + \abs{t^*}^{1/\ell} + \abs{x- x^*}\bigr) \le C\bigl( \abs{t-t^*}^{1/\ell} + \abs{x- x^*}\bigr). \end{align*} This shows \eqref{galileo} for $p=1$. Iterating one gets the general case. Next we prove existence and continuity of the derivative $F_{X_w}$. Assume first that $t\ge 0$ and decompose $\espo_{\textup{ap}}^{t X_w} x= e^{t^{1/\ell }Z_1}\cdots e^{t^{1/\ell} Z_\nu}x$, where $\ell= \abs{w}$ and $Z_1, \dots,Z_\nu \in \pm\H$ are suitable. Euclidean regularity of the vector fields $Z_j$ implies that the functions $(\t,y)\mapsto F_{Z_j}(\t, y):= \nabla_y e^{\t Z_j}y$ are continuous if $y\in \Omega$ and $\abs{\tau}$ is small. Therefore, the chain rule gives \begin{align*} F_{X_w}(t,x)&= \nabla_x \espo_{\textup{ap}}^{t X_w} (x) \\& = F_{Z_1}(t^{1/\ell}, e^{t^{1/\ell}Z_2}\cdots e^{t^{1/\ell }Z_\nu}x) F_{Z_2}(t^{1/\ell}, e^{t^{1/\ell}Z_3}\cdots (x)) \cdots F_{Z_\nu}(t^{1/\ell} ,x). \end{align*} Thus $F_{X_w}\bigr|_{[0, C^{-1}]\times\Omega}$ is continuous. Note that $F_{X_w}(0,x)= I_n$ for all $x$. An analogous argument shows that $F_{X_w}\bigr|_{[-C^{-1},0]\times\Omega}$ is continuous and concludes the proof. \end{proof} At this point we may deduce the following result. See \eqref{fivetwenty} for notation on the map~$E$. \begin{theorem}\label{deduciamo} Let $\H$ be an $\mathcal{A}_s$ family. Let $x\in \Omega$ and let $r\in (0,r_0)$. Fix $p\in\{1,\dots,q\}$ and $I\in \mathcal{I}(p, q)$. Then the function $E_{I,x,r}$ is $C^1$ smooth on $B_\textup{Euc}(C^{-1})$. Moreover, for all $h\in B_\textup{Euc}(C^{-1})$ and for any $k\in\{1,\dots,p\}$ we have $E_*(\partial_{h_k})\in P_{E(h)}$ and we can write \begin{equation}\label{nhb} E_*(\partial_{h_k}) = \widetilde U_{k,E(h)} +\sum_{\ell_j=d_k+1}^{s} a^j_k(h) \widetilde Y_{j, E(h)} + \sum_{i=1}^q \omega_k^i(x,h)\widetilde Y_{i,E(h)}, \end{equation} where, for some $C>1$ depending on $L_0$ and $C_0$ in \eqref{lipo} and \eqref{artefatto}, we have \begin{align}\label{sogliola} \abs{a_k^j (h)}& \le C \bigl\|h \bigr\|_I^{\ell_j-d_k}\quad\text{for all $h\in B_\textup{Euc}(C^{-1})$} \\ \label{merluzzo} \abs{\omega_i(x,h)} &\le C \bigl\|h \bigr\|_I^{s+1- d_k} \quad\text{for all $h\in B_\textup{Euc}(C^{-1})\quad x\in\Omega$}. \end{align} \end{theorem} \begin{proof} For notational simplicity we delete everywhere the tilde. In fact, the statement holds uniformly in $r\in (0, r_0)$, where $r_0$ depends on the already mentioned constants $L_0$ and $C_0$. \step{Step 1.} We first prove the theorem for $p=1$. Using the definition of $\exp_{\textup{ap}}$ and Theorem \ref{dascalare}, we easily obtain by a change of variable that for any commutator $Y$ of length $\ell\in\{1,\dots,s\}$ and for all $\psi\in C^1_\textup{Euc}$, \begin{equation}\begin{aligned} \label{daeo} \frac{d}{dh}\psi (\espo_{\textup{ap}}^{h Y }(x))& = Y\psi(\espo_{\textup{ap}}^{h Y}(x))+ \sum_{\ell_j=\ell+1}^s \a_j(h) Y_k\psi(\espo_{\textup{ap}}^{h Y}x) \\& \qquad + \abs{h}^{(s+1-\ell)/\ell} \sum_{i=1}^q b_i(x,h) Y_i\psi(\espo_{\textup{ap}}^{h Y}x), \end{aligned}\end{equation} for all $x\in K$ and $0<\abs{h}\le C^{-1} $, where the sum is empty if $\ell=s$. If $\ell< s$, then $\a_j(h)= \ell^{-1} a_j h^{(\ell_j-\ell)/\ell}$ if $h>0$, while $\a_j(h)=- \ell^{-1}\overline a_j h^{(\ell_j-\ell)/\ell}$ if $h<0$. The functions $a_j$ come from the statement of Theorem \ref{dascalare}. The functions $b_i(x,h)$ can be discontinuous, if we pass from $h>0$ to $h<0$, but we have estimate $\abs{b_i(x,h)}\le C$ uniformly in $x,h$. To complete Step 1, we need to show that the function $h\mapsto\frac{d}{dh} \espo_{\textup{ap}}^{h Y}z$ is continuous for all fixed $z\in \Omega$. Continuity at any $h\neq 0$ (say $h>0$) follows immediately from the decomposition $\espo_{\textup{ap}}^{h Y} = e^{h^{1/\ell} Z_1}\cdots e^{h^{1/\ell} Z_\nu} $, where $Z_j\in\pm\H$. We show now continuity at $h=0$. Formula \eqref{daeo} gives $ \Bigl| \frac{\partial}{\partial h}\espo_{\textup{ap}}^{h Y}z - g(\espo_{\textup{ap}}^{h Y}z) \Bigr|\le C\abs{h}^{1/\ell}$ (recall notation $Y=:g\cdot\nabla$). Therefore, using the l'H\^opital's rule, we get \begin{equation*} \begin{aligned} \frac{d}{d h}\espo_{\textup{ap}}^{h Y}z\bigr|_{h=0} & := \lim_{h\to 0}\frac{\espo_{\textup{ap}}^{h Y}z - z }{h} =\lim_{h\to 0} g(\espo_{\textup{ap}}^{h Y}z)+ O(\abs{h}^{1/\ell}) = g(z), \end{aligned} \end{equation*} where we need the $d$-continuity of $ g$. This shows existence of the derivative at $h=0$. To see continuity, just let $h\to 0$ in \eqref{daeo}. \step{Step 2.} By induction on $p$, we show that $E$ is $C^1$ smooth. Assume that $(h_1, \dots, h_{p-1})\mapsto \espo_{\textup{ap}}^{h_1 U_1}\cdots \espo_{\textup{ap}}^{h_{p-1}U_{p-1}}(x)$ is $C^1$ for all choice of $U_1, \dots, U_{p-1}$. We need to show that $(h_1, \dots, h_{p})\mapsto \espo_{\textup{ap}}^{h_1 U_1}\cdots \espo_{\textup{ap}}^{h_{p}U_{p}}(x)$ is $C^1$ smooth. Let $U_1, \dots, U_p\in \P$. First of all we show that the map $(h_1, \dots, h_p)\mapsto E_*(\partial_{h_1})$ is continuous. If $h_1\neq 0$, say $h_1>0$, then we decompose for suitable $Z_1, \dots, Z_\mu\in \H$, \begin{equation*} \espo_{\textup{ap}}^{h_1 U_1} \cdots \espo_{\textup{ap}}^{h_p U_p}x = e^{ h_1^{1/d_1}Z_1} \cdots e^{ {h_1}^{1/d_1}Z_\mu} \espo_{\textup{ap}}^{h_2 U_2}\cdots\espo_{\textup{ap}}^{h_p U_p}x. \end{equation*} Note that by standard ODE theory, the map $(\t_1, \dots,\t_\mu,z )\mapsto e^{\t_1 Z_1}\cdots e^{\t_\mu Z_\mu}z$ is $C^1$. Therefore, by means of Lemma \ref{torrette}, we have existence and continuity of $\partial_1 E(h)= E_*(\partial_{h_1})$ at any point of the form $h= (h_1, h_2, \dots, h_p)$ with $h_1\neq 0$. To discuss the case $h_1=0$, recall that formula \eqref{daeo} gives \begin{equation*} \begin{aligned} & \Bigl| \frac{\partial}{\partial h_1}\espo_{\textup{ap}}^{h_1 U_1}\cdots\espo_{\textup{ap}}^{h_p U_p}x -U_1(\espo_{\textup{ap}}^{h_1 U_1}\cdots\espo_{\textup{ap}}^{h_p U_p}x)\Bigr| \le C\abs{h_1}^{1/d_1}. \end{aligned} \end{equation*} Therefore, using de~l'H\^opital's rule, for all $h=(0,h_2, \dots, h_p)=:(0,\widehat h_1)$, we get \begin{equation*} \begin{aligned} \partial_1E(0,\widehat h_1) &:=\lim_{h_1\to 0}\frac{\espo_{\textup{ap}}^{h_1 U_1}\espo_{\textup{ap}}^{h_2 U_2}\cdots\espo_{\textup{ap}}^{h_p U_p}x-\espo_{\textup{ap}}^{h_2 U_2}\cdots\espo_{\textup{ap}}^{h_p U_p}x}{h_1} \\&=\lim_{h_1\to 0}U_1(\espo_{\textup{ap}}^{h_1 U_1}\espo_{\textup{ap}}^{h_2 U_2}\cdots\espo_{\textup{ap}}^{h_p U_p}x)+ O(\abs{h_1}^{1/d_1}) = U_1(E(0, \widehat h_1)), \end{aligned} \end{equation*} where we need the $d$-continuity of $U_1$. This shows existence of $\partial_{1}E(0, \widehat h_1)$. To show continuity of $\partial_{h_1}E$ at $h^* = (0, \widehat h_1^*)\in B_\textup{Euc}(C^{-1})$, write by expansion \eqref{daeo} \begin{equation} \label{sosso} \begin{aligned} \bigl| \partial_1E(h_1, \widehat h_1) & - \partial_1E(0,\widehat h_1^*)\bigr| \\ &= \Bigl|U_1(E(h_1, \widehat h_1))+\sum_{d_1 +1\le \ell_j\le s} \a_j(h_1)Y_j(E(h_1, \widehat h_1)) \\&\quad + \abs{h_1}^{(s+1-d_1)/d_1} \sum_{1\le i\le q} b_i Y_i(E(h_1, \widehat h_1)) - U_1(E(0, \widehat h_1^*))\Bigr| \\&\le C\abs{h_1}^{1/d_1}+ \abs{U_1(E(h_1, \widehat h_1)) -U_1(E(0, \widehat h_1^*))}\to 0, \end{aligned} \end{equation} as $(h_1, \widehat h_1)\to (0, \widehat h_1^*)$, here we used assumption \eqref{sussmann} for $ U_1$. To conclude \emph{Step 2}, we show the continuity of $\partial_{h_k} E$ for all $2\le k\le p$. Write by the chain rule \begin{equation}\label{cksh} \begin{aligned} \frac{\partial}{\partial h_k} E(h) & = F_{U_1}(h_1, \espo_{\textup{ap}}^{h_2 U_2}\cdots(x))\cdots F_{U_{k-1}}(h_{k-1}, \espo_{\textup{ap}}^{h_k U_k}\cdots(x))\frac{\partial}{\partial h_k}\espo_{\textup{ap}}^{h_k U_k}\cdots(x). \end{aligned} \end{equation} This ends the proof, because the right-hand side depends continuosly on $h_1, \dots, h_p$, by Lemma \ref{torrette} and the first part of \emph{Step 2}. \step{Step 3.} We show expansion \eqref{nhb} and estimates \eqref{sogliola} and \eqref{merluzzo} for any $p$ and for all $k=1,\dots, p$. Let $U_k = Y_{i_k}$, $d_k:= \ell_{i_k}$ and $E_{ \langle j,k\rangle }(x):= \espo_{\textup{ap}}^{h_j U_j } \cdots \espo_{\textup{ap}}^{h_k U_k}(x)$ for all $1\le j\le k\le p$. We agree that $E_{\langle j,j-1\rangle } $ denotes the identity function. Observe that the function $z\mapsto E_{ \langle j,k\rangle }(z) $ is a $C^1$ diffeomorphism for any fixed $h_j, h_{j+1}, \dots, h_k$. Then, for $k\in\{1,\dots,p\}$, we may use \eqref{daeo} and we get \begin{equation}\label{baleno} \begin{aligned} E_*(\partial_{h_k}) &= U_k E_{\langle 1,k-1\rangle }(E_{\langle k,p\rangle}(x)) +\sum_{\ell_j = d_k +1}^s\a_j(h_k) Y_{j} E_{\langle 1,k-1\rangle }(E_{\langle k, p \rangle}(x)) \\&\quad+ \abs{h_k}^{(s+1-d_k)/d_k} \sum_{i=1}^q b_i Y_i E_{\langle 1,k-1\rangle}(E_{\langle k, p \rangle } (x)), \end{aligned} \end{equation} where $b_i$ denote bounded functions and $\abs{\alpha_j(h_k)}\le C\abs{h_k}^{(\ell_j - d_k)/d_k}$. To get formula \eqref{nhb}, it suffices to use a rough expansion of each term as follows. Write for $\lambda\in\{1,\dots,p\}$ and $h_\lambda >0$, $\espo_{\textup{ap}}^{h_\lambda U_\lambda} = e^{- h_\lambda^{1/d_\lambda} Z_1}\cdots e^{-h_\lambda^{1/d_\lambda} Z_\nu}$, for suitable $Z_i\in \pm \H$. Then for all $j\in \{1,\dots,q\}$ write \begin{align*} Y_j(\psi \espo_{\textup{ap}}^{h_\lambda U_\lambda})(z) &= Y_j(\psi e^{-{h_\lambda}^{1/d_\lambda} Z_1}\cdots e^{-{h_\lambda}^{1/d_\lambda} Z_\nu})(z) \\&= Y_j\psi (\espo_{\textup{ap}}^{h_\lambda U_\lambda} z)+ \sum_{\abs{\a}=1 }^{s-\ell_j}\ad_{ Z_\nu}^{\a_\nu}\cdots\ad_{ Z_1}^{\a_1} Y_j \psi (\espo_{\textup{ap}}^{h_\lambda U_\lambda} z)\frac{h_\lambda^{\abs{\a}/d_\lambda}}{\a!} \\&\qquad + O_{s+1}(\abs{h_\lambda}^{(s+1-\ell_j)/d_\lambda}, \psi, \espo_{\textup{ap}}^{h_\lambda U_\lambda}z ) \\&= Y_j\psi (\espo_{\textup{ap}}^{h_\lambda U_\lambda} z) + \sum_{\ell_i= \ell_j + 1}^s c_i \abs{h_\lambda}^{(\ell_i - \ell_j)/d_\lambda} Y_i\psi (\espo_{\textup{ap}}^{h_\lambda U_\lambda}x) \\&\qquad \qquad +\abs{h_\lambda}^{(s+1-\ell_j)/d_\lambda}\sum_{i=1}^q b_i Y_i\psi (\espo_{\textup{ap}}^{h_\lambda U_\lambda}x), \end{align*} where we use the pointwise form of the remainder, see the proof of Theorem \ref{dascalare}. Here $c_i$ are constants, while $b_i$ are bounded functions. The proof of \eqref{nhb} follows from \eqref{baleno} via a repeated application of this expansion. If $h_\lambda<0$, then the terms $c_i$ and $b_i$ may change, but the argument gives the same conclusion. The proof of the theorem is concluded \end{proof} \begin{remark}\label{tredodici} $\,$\begin{itemize*} \item[(i)] Let $X_w$ be a commutator of length $\abs{w}\le s$. Define the function $ H(t,x):=\frac{d}{dt} \espo_{\textup{ap}}^{t X_w}(x).$ Under our assumptions $\mathcal{A}_s$ we may claim that $ H(t,x)$ exists for all $(t,x)$. However, we can not expect that the function $(t,x)\mapsto H(t,x)$ is continuous in $(-t_0, t_0)\times\Omega$. Indeed, in order to show the continuity of $H$ at a point $(0, \widetilde x)$, because \begin{equation*} \begin{aligned} \abs{H(t,x)- H(0,\widetilde x)}&\le\abs{H(t,x)- H(0,x)}+\abs{H(0,x)- H(0, \widetilde x)}. \\&=\bigl|\frac{d}{dt}\espo_{\textup{ap}}^{t X_w}x - f_w(x)\bigr|+\abs{f_w(x)- f_w(\widetilde x)}. \end{aligned} \end{equation*} The first term can be made small uniformly in $x$, if $\abs{t}$ is small. In order to make the second term small, we can use only assumption \eqref{sussmann}, which does not ensure any continuity if $x$ and $ \widetilde x$ belong to different orbits. \item[(ii)] Under our assumptions, we cannot expect that maps $h\mapsto E_{I,x}(h)$ are more than $C^1$. Indeed, the term $F_{U_1}(h_1, \espo_{\textup{ap}}^{h_2 U_2} \cdots \espo_{\textup{ap}}^{h_p U_p}x)$ in \eqref{cksh} depends continuously on $h_2, \dots, h_p$, if $\H$ is a $C^1$ family (recall that $F_{U_1}(h,x):=\nabla \espo_{\textup{ap}}^{h U_1}(\xi)$ is only continuous in $\xi$). An inspection of the proof above shows that if $\H$ is a $C^2$ family and $\mathcal{A}_s$ holds, then $E_{I,x}\in C^{1, 1/s}_\textup{loc}$, but this regularity cannot be improved, even if $X_j\in C^{\infty}$ or $ C^\omega$; see \cite[Example~5.7]{MM}. \end{itemize*} \end{remark} Now we can easily prove the regularity of orbits, along the lines of the proof in \cite{Agrachev}. \begin{theorem}[Regularity of $\mathcal{A}_s$ orbits]\label{frr} Let $\H$ be a system of $\mathcal{A}_s$ vector fields. Then each orbit $\O$ with the topology $\t_d$ is a connected $C^1$ smooth immersed submanifold of $\mathbb{R}^n$ satisfying $T_x\O= P_x:=\Span\{X_w(x):1\le \abs{w}\le s\}$ for all $x\in \O$. \end{theorem} \begin{proof} Let $x_0\in \mathbb{R}^n$ and let $\O:= \O_\H^{x_0}$ be its $\H$-orbit. We know from Remark \ref{osservo} that $\dim P_x = \dim P_{x_0} =:p$ is constant in $\O$. For each $x\in \O$ choose $I\in \mathcal{I}(p, q)$ such that $\abs{Y_I(x)} \neq 0 $. By Theorem \ref{deduciamo} and by the implicit function theorem, we may claim that for a suitable $ O_{I,x}\subset\mathbb{R}^p$, open neighborhood of the origin, the map $E_{I,x}:O_{I,x}\to\mathbb{R}^n$ is a $C^1$ full-rank map which parametrizes a $C^1$ smooth, $p$-dimensional embedded submanifold $E_{I,x}(O_{I,x})\subset \mathbb{R}^n$. Note also that $E_{I,x}(O_{I,x})\subset\O$ and, by Theorem~\ref{deduciamo}, $T_{E_{I,x}(h)} E_{I,x}(O_{I,x}) = P_{E_{I,x}(h)}$, for all $h\in O_{I,x}$. Let \[ \begin{aligned} \mathcal{U} &:=\{ E_{I, x}(O): x\in \O, I\in \mathcal{I}(p,q),\abs{Y_I(x)}\neq 0 \\ &\qquad \qquad \text{and $O\subset O_{I, x}$ is a open neighborhood of the origin$\}.$} \end{aligned}\] We claim that the family $\mathcal{U}$ can be used as a base for a topology $\t(\cal{U})$ on $\O$. To see that, we need to show that if the intersection of the $p$-dimensional submanifolds $E_{I,x}(O)$ and $ E_{I',x'}(O')$ is nonempty, then it contains a small manifold of the form $E_{I'',x''}(O'')$, if $O''$ is a sufficiently small neighborhood of the origin. Let $\Sigma:=E_{I,x}(O)$ and $\Sigma'= E_{I',x'}(O')$ and let $x''\in\Sigma\cap\Sigma'$. Recall that both $\Sigma$ and $\Sigma'$ are embedded $C^1$ submanifolds of $\mathbb{R}^n$. Let $I''\in \mathcal{I}(p,q)$ be such that $\abs{Y_{I''}(x'')}\neq0$. Let $O''\subset\mathbb{R}^p$ be a small open neighborhood of the origin. For any $h\in O''$, the point $E_{I'',x''} (h) $ can be written as $ e^{\t_1Z_1}\cdots e^{\t_\nu Z_\nu}x$ where $Z_j\in \pm\H$ and $\sum_j\abs{\t_j}\le C\|h\|_I$. By a repeated application of Bony's theorem \cite[Theorem 2.1]{Bony69}, it follows that $E(h)\in\Sigma$, provided that $h$ is sufficiently close to the origin. The same argument applies to $\Sigma'$. Thus we have proved that $\mathcal{U}$ can be used as a topology base. A similar argument shows that any submanifold of the form $E_{I, x}(O)\in\mathcal{U}$ contains a small ball $B_d(x, \sigma)$. Therefore $\t_d$ is stronger than $\t(\mathcal{U})$. The fact that $\t(\mathcal{U})$ is stronger that $\t_d$ follows easily from estimate $d(E_{I,x}(h),x)\le C\|h\|_I.$ Finally, since all paths of the form $t\mapsto e^{tZ}x\in(\O,\t(\mathcal{U}))=(\O,\t_d)$ are continuous, the orbit is connected. The $C^1$ differential structure on $\O$ is given by the family maps $E_{I,x}\bigr|_{O}$ where $x\in \O$, $I\in\mathcal{I}(p_x, q)$ is such that $\abs{Y_I(x)}\neq 0$ and $O\subset O_{I,x}$ is an open neighborhood of the origin. \end{proof} \begin{example} \label{esempietto} Let us consider in $\mathbb{R}^3$ the family $\H=\{X_1, X_2, X_3\}$: \[X_1= a(t)\partial_x\quad X_2 = x a(t)\partial_y \quad\text{and} \quad X_3 = t\partial_t,\] where the function $a$ satisfies $a(t) = 1+t^3 \sin\big(\frac 1t\big)$, if $0<\abs{t}<1$, $a(0)=0$, $a\in C^\infty(\mathbb{R}\setminus\{0\})$ and $\inf_\mathbb{R} a>0$. Note that $X_j\in C^1_\textup{Euc}(\mathbb{R}^3)$ and \[ [X_1, X_2]= a(t)^2 \partial_y,\quad [X_1, X_3] = - t a'(t) \partial_{x}\quad\text{and}\quad [X_2, X_3] = -t a'(t) x\partial_{y}. \] If $0< \abs{t}<1$, then \[ \frac{d}{dt} (ta'(t)) = \frac{d}{dt}\Big(3t^3 \sin \frac 1t - t^2\cos \frac 1t\Big) =9t^2 \sin\frac 1t - 5t\cos\frac 1t - \sin\frac{1}{t} \] is discontinuous at $t=0$. Therefore $X_{13}$ and $X_{23}\notin C^1_\textup{Euc}$ and the $C^1$ singular Frobenius theorem does not apply to the family $\P=\{ X_1, X_2, X_3, [X_1, X_2], [X_1, X_3], [X_2, X_3]\}$. However, we claim that the family $\H$ belongs to our class $\mathcal{A}_2$. To show this claim, we first prove that $X_j\in C^{1,1}_{\H,\textup{loc}}$ . To see that, it suffices to show that $X_3^\sharp X_3^\sharp a \in C^0_\H$. But, if $0<\abs{t}<1$, we have \begin{equation}\label{continuous} X_3^\sharp X_3^\sharp a(t)= t\partial_t(ta'(t)) = 9t^3 \sin\frac 1t - 5 t^2\cos\frac 1 t - t\sin\frac 1t, \end{equation} which is a continuous function up to $t=0$ (note that, since $ X_3^\sharp a(0)=0$, we have $X_3^\sharp X_3^\sharp a(0) = \lim_{t\to 0}t^{-1}(X_3^\sharp a (e^{t X_3}(0)) - X_3^\sharp a(0) )= 0$). Since $X_{12},X_{13}$ and $X_{23}\in C^0_\textup{Euc}$, condition \eqref{sussmann} is fulfilled. Finally, we have to check the $2$-involutivity, i.e.~that for all $i,j,k$ we can write $\ad_{X_i}X_{jk}= \sum_{\abs{w}\le 2}b^w X_w$ with $b^w$ locally bounded. A computation shows that the nonzero terms are the following (we work with $0<\abs{t}<1$) \begin{align*} -\ad_{X_1}X_{23}&= \ad_{X_2}X_{13} =\frac 12 \ad_{X_3}X_{12}= ta(t)a'(t)\partial_y = \frac{ta'(t)}{a(t)}X_{12} \\ \ad_{X_3}X_{13} & =-t\partial_t(ta'(t))\partial_x =\frac{-t\partial_t(ta'(t))}{a(t)}X_1 \\ \ad_{X_3} X_{23}&= -xt\partial_t(ta'(t))\partial_y = \frac{-t\partial_t(ta'(t))}{a(t)}X_2. \end{align*} Since $\inf_{\mathbb{R}} a>0$, one can see with the help of \eqref{continuous} that both the coefficients $ta'(t)/a(t)$ and $-t\partial_t(ta'(t))/a(t)$ are locally bounded. Thus, hypothesis $\mathcal{A}_2$ is fulfilled and our main theorem applies. Note finally that it is very easy to see that there are three orbits of the family $\H$. Namely, $\O_1:= \{(x,y,t):t>0\}$, $\O_2= \{t=0\}$ and $\O_3 = \{t<0\}$ and they are integral manifolds of the distribution generated by the family $\P.$ \end{example} \begin{remark} \label{finalremark} A natural question concerns sharpness of the $C^1$ regularity of $\O_\H$. It is reasonable to guess that $C^1$ regularity is not sharp. Actually, we do not have any example of vector fields of class $\mathcal{A}_s$ where the integral manifolds $\O_\H$ are less than $C^2$. However, under our assumptions, maps $E_{I,x }$ cannot provide more than $C^1$ regularity, see Remark~\ref{tredodici}-(ii). A related issue concerns the regularity of the orbit $\O_\H$ of a generic family of $C^1$ (or even Lipschitz-continuous) vector fields which do not satisfy any involutivity assumptions. This would require a careful discussion of a nonsmooth version of Sussmann's orbit theorem. We plan to discuss such questions in a future study. \end{remark}

https://arxiv.org/abs/2106.08900

Random feature neural networks learn Black-Scholes type PDEs without curse of dimensionality

This article investigates the use of random feature neural networks for learning Kolmogorov partial (integro-)differential equations associated to Black-Scholes and more general exponential Lévy models. Random feature neural networks are single-hidden-layer feedforward neural networks in which only the output weights are trainable. This makes training particularly simple, but (a priori) reduces expressivity. Interestingly, this is not the case for Black-Scholes type PDEs, as we show here. We derive bounds for the prediction error of random neural networks for learning sufficiently non-degenerate Black-Scholes type models. A full error analysis is provided and it is shown that the derived bounds do not suffer from the curse of dimensionality. We also investigate an application of these results to basket options and validate the bounds numerically.These results prove that neural networks are able to \textit{learn} solutions to Black-Scholes type PDEs without the curse of dimensionality. In addition, this provides an example of a relevant learning problem in which random feature neural networks are provably efficient.

\section{Introduction} \label{sec:Intro} A fundamental problem in science and engineering is to infer an unknown input-output relation from data. In recent years (artificial) neural networks have become an important tool to address such problems in complex, high-dimensional situations. Neural networks have shown a strikingly efficient computational performance in an enormous range of applications and impressive progress has also been made regarding the theoretical and mathematical foundations of neural network-based methods. In many situations additional a priori information about the unknown input-output relation is available and the problem amounts to learning the solution of a partial differential equation (PDE) or, for instance in a financial context, an expectation of a stochastic process. Examples of applications of neural networks in this area can be found e.g.\ in \hbox{\cite{Han2018}}, \cite{EHanJentzen2017CMStat}, \cite{Sirignano2018}, \cite{ComePhamWarin2020}, \cite{Buehler2018}, \cite{Cuchiero2019}. We refer to the surveys \cite{Ruf2020}, \cite{Beck2020}, \cite{Germain2021} for an overview of the numerous recent applications of neural network-based learning in the context of PDEs, stochastic processes and finance. There has also been important progress regarding the theoretical and mathematical foundations of neural network-based methods in this area, see again the surveys mentioned above for an overview. Many of these recent mathematical results prove that deep neural networks are able to approximate solutions to various classes of PDEs without the curse of dimensionality, see, for instance, \cite{EGJS18_787}, \cite{HornungJentzen2018}, \cite{HJKNvW2020}, \cite{ReisingerZhang2019}, \cite{Laakmann2020}, \cite{KutyniokPeterseb2019}, \cite{GS21}. In some articles then a learning problem is considered and such approximation error bounds are combined with generalization error bounds in order to prove that the empirical risk-minimizing deep neural network is capable of overcoming the curse of dimensionality for learning solutions to certain PDEs, see e.g.\ \cite{BernerGrohsJentzen2018}, \cite{CarmonaLauriere_DL_periodic}. In practice, the neural network that minimizes the empirical risk needs to be calculated approximately, which is typically achieved using a variant of the stochastic gradient descent algorithm. This introduces a further error component, the optimization error, which has remained challenging to analyze mathematically for general neural networks. As a consequence, in the context of PDEs there have been no results in the literature so far which address all three error components and explain mathematically the success of neural networks at \textit{learning} solutions to high-dimensional PDEs. In this work such an explanation is provided by proving that neural networks are capable of learning solutions to certain PDEs without the curse of dimensionality. This is achieved by considering neural networks in which only certain weights are trainable and the remaining parameters are generated randomly, as we will now describe in more detail. We investigate the capabilities of random (feature) neural networks \cite{HCS2006}, \cite{RahimiRecht2008a}, \cite{RahimiRecht2008} as a learning method in the context of certain Kolmogorov PDEs. Random neural networks are feedforward neural networks with a single hidden layer and the property that the parameters of the hidden layer are randomly initialized and then fixed. Hence, only the parameters of the output layer can be trained. The non-convex optimization problem that needs to be solved in order to train a standard neural network reduces to a convex optimization problem here. This simplifies both training in practice and theoretical analysis. On the other hand, allowing only parts of the parameters to be trained reduces the approximation capabilities and so, at least a priori, it is not clear if random neural networks still have any of the powerful approximation properties of general deep neural networks. In several other contexts these questions have been addressed and learning (or prediction/test) error bounds for random features or random neural networks have been proved (see for instance \cite{RahimiRecht2008a}, \cite{RudiRosasco2017}, \cite{CRR2018}, \cite{MM19}, \cite{MMM21} and the references therein), but not in the context of PDEs. This is precisely the subject of this article. We investigate these questions for the problem of learning an unknown function from a class of Kolmogorov PDEs, which include the Black-Scholes PDE as a special case. These partial (integro-)differential equations, which are also referred to as \textit{(non-local) PDEs}, arise for instance in the context of option pricing in exponential L\'evy models, see e.g.\ \cite{Cont2004}, \cite{EberKall19} and the references therein. The main results of this article prove that, indeed, random neural networks are capable of learning non-degenerate Black-Scholes type PDEs without the curse of dimensionality. We provide a full error analysis, i.e., bounds on the approximation error, the generalization (or estimation) error and the optimization error. For each of these error components we obtain polynomial convergence rates which do not depend on the dimension $d$ of the underlying PDE and constants which grow at most polynomially in $d$. Thus, the article contributes to the literature in several aspects. Firstly, it provides for the first time a neural network-based algorithm for learning Kolmogorov PDEs for which a full error analysis (covering all three error components) is available and which does not suffer from the curse of dimensionality. The solution to the PDE can be learnt on a full hypercube from observational data even without knowing the parameters of the PDE. Secondly, it provides an example of a practically relevant learning problem in which random features are provably efficient. Finally, the techniques developed in the article may also be helpful for the theoretical analysis of more general neural network-based learning methods in future works. Neural networks with randomly sampled weights already appear in Barron's work \cite{Barron1992}, \cite{Barron1993}. The random sampling-based dimension-independent convergence rates obtained there were also extended to the larger class of ``generalized Barron functions'' in \cite{EMaWWu2020}, \cite{EMaWu2019}, \cite{EWojtowytsch2020}, see also \citet[Section~4.2]{BGKP2021}. For further related results and extensions we refer, for instance, to \cite{BarronKlusowski2018}, \cite{SiegelXu2020}, \cite{CPV2020} and the references therein. In all these results, the random sampling procedure is an intermediate step to establish the existence of neural network weights and obtain approximation bounds. This does not yield a constructive sampling procedure in general, since the random sampling distribution depends on the unknown target function. In contrast, in the random features approach \cite{RahimiRecht2008}, \cite{RahimiRecht2008a} considered here the distribution from which the random weights are sampled is chosen a priori and does not depend on the target function. Numerical methods for partial (integro-)differential equations associated to univariate and certain multivariate exponential L\'evy models were developed, e.g., in \cite{ContVolt2005}, \cite{FRS07}, \cite{MvS04_373}, \cite{Hilber2009}. In a high-dimensional setting, when the parameters of the PDE are known and the solution of the PDE needs to be evaluated at a single point, then Monte Carlo methods are able to approximate the solution of the PDE without the curse of dimensionality. In contrast, here we consider a more challenging situation, which includes both the problem of evaluating the solution of the PDE on a full hypercube $[-M,M]^d$ by a numerical method and the problem of learning the solution of the PDE from observed values. In the latter situation, in particular, the true parameters of the PDE are unknown. The remainder of the article is structured as follows. Section~\ref{sec:RandomNN} introduces random neural networks and provides a general approximation result. In Section~\ref{sec:Approx} we build on this result to provide random neural network approximation bounds for a class of convolutional functions and then specialize to the case of partial (integro-)differential equations or (non-local) PDEs associated to exponential L\'evy models. Section~\ref{sec:Learning} introduces the learning problem, provides error bounds for different learning methods (regression, constrained regression and stochastic gradient descent) and develops an application to basket option pricing. These results are then applied in Section~\ref{sec:Kolmogorov} to prove that random neural networks are capable of learning Black-Scholes type PDEs without the curse of dimensionality. The paper concludes with a numerical experiment to validate the obtained bounds. \subsection{Notation} In most parts of the article we will consider the dimension $d \in \mathbb{N}$ as fixed, but we will work out explicitly the dependence of all constants on $d$. In Sections~\ref{subsec:ApproxLevy} and \ref{sec:Kolmogorov} we will consider a family of models indexed by $d \in \mathbb{N}$ and thus $d$ appears explicitly in the notation there. Throughout, $\|\cdot\|$ denotes the Euclidean norm on $\mathbb{R}^d$ or $\mathbb{R}^N$ (the appropriate space will always be clear from the context). For $M>0$ we denote the Euclidean ball by $B_M(0) = \{x \in \mathbb{R}^d \,|\, \|x\|\leq M\}$. All random variables are defined on a probability space $(\Omega,\mathcal{F},\P)$ and we write $\|\cdot\|_{L^\infty(\P)} = \|\cdot\|_{L^\infty(\Omega,\mathcal{F},\P)}$ for the $L^\infty$-norm on $(\Omega,\mathcal{F},\P)$. For $x \in \mathbb{R}^d$ we use the notation $\exp(x) =(\exp(x_1),\ldots,\exp(x_d))$. \section{Random neural networks: preliminary results} \label{sec:RandomNN} In this section we recall the definition of random (feature) neural networks and provide a general approximation result. Such networks will be used to learn an unknown target function. A random neural network is a feedforward neural network with one hidden layer and randomly generated hidden weights. More specifically, let $N \in \mathbb{N}$, let $B_1,\ldots,B_N$ be i.i.d.\ random variables, let $A_1,\ldots,A_N$ be i.i.d.\ $\mathbb{R}^d$-valued random vectors, assume that $A=(A_1,\ldots,A_N)$ and $B=(B_1,\ldots,B_N)$ are independent and for an $\mathbb{R}^N$-valued random vector $W$ consider the (random) function \begin{equation} \label{eq:RandomNN} H^{A,B}_W(x):= \sum_{i=1}^N W_i \varrho(A_i \cdot x + B_i), \quad x \in \mathbb{R}^d, \end{equation} where $\varrho \colon \mathbb{R} \to \mathbb{R}$ is a fixed activation function. Throughout the article we will consider random neural networks with the ReLU activation function given by $\varrho(z)=\max(z,0)$ for $z \in \mathbb{R}$. The random variables $A$ and $B$ will be referred to as the (random) hidden weights of the neural network and $W$ as the vector of output weights. To approximate an unknown function $H \colon \mathbb{R}^d \to \mathbb{R}$ the (random) hidden weights $A,B$ are considered as fixed and only the output vector $W$ can be trained. Thus, the goal is to find $W$ such that the expected uniform approximation error $\mathbb{E}[ \|H^{A,B}_W - H \|_{L^\infty([-M,M]^d)}]$ is small. Approximation properties of such random neural networks have been studied for instance in \cite{HCS2006}, \cite{RahimiRecht2008a}, \cite{RudiRosasco2017} and most recently in \cite{RC12}. Theorem~\ref{thm:RC12Linfty} below is a novel approximation result for sufficiently regular functions, which will be crucial for the results in Section~\ref{sec:Approx}. The result and parts of the proof of Theorem~\ref{thm:RC12Linfty} are similar to \citet[Theorem~1]{RC12}; however in \citet[Theorem~1]{RC12} a more general Hilbert space setting and more general sampling distributions are considered. In contrast, Theorem~\ref{thm:RC12Linfty} works under stronger hypotheses and employs Rademacher complexity-based techniques to obtain a \textit{uniform error bound} instead of an $L^2$-error bound. More specifically, in what follows we make the following assumptions on the distribution of the hidden weights of the random neural network \eqref{eq:RandomNN}: \begin{itemize} \item the distribution of $A_1$ has a strictly positive Lebesgue-density $\pi_{\text{w}}$ on $\mathbb{R}^d$ and \item the distribution of $B_1$ has a strictly positive Lebesgue-density $\pi_{\text{b}}$ on $\mathbb{R}$. \end{itemize} In this situation, the following random neural network approximation result holds. \begin{theorem} \label{thm:RC12Linfty} Let $H \colon \mathbb{R}^d \to \mathbb{R}$, let $M >0$ and assume there exists $G \colon \mathbb{R}^d \to \mathbb{C}$ such that \begin{equation} \label{eq:Hrepresentation2} H(x) = \int_{\mathbb{R}^d} e^{i x \cdot \xi} G(\xi) d\xi \end{equation} for all $x \in [-M,M]^d$. Suppose that \begin{equation} \label{eq:BarronCondLInfty} \int_{\mathbb{R}^d} \max(1,\|\xi\|^2) |G(\xi)| d\xi < \infty, \end{equation} $\bar{F}(r) := 2\int_{-r}^0 \frac{1}{\pi_{\text{b}}(s)} ds \in (-\infty,\infty)$ for all $r \in \mathbb{R}$ and \begin{equation} \label{eq:IfiniteLinfty} I = \max(16,M^2) \int_{\mathbb{R}^d} [ \bar{F}(M\|\xi\|_1)\|\xi\|_1^2 + ( \bar{F}(1)-\bar{F}(-1)) \max(1,\|\xi\|^2)] \frac{(|G(\xi)|+|G(-\xi)|)^2}{ \pi_{\text{w}}(\xi)} d \xi \end{equation} is finite. Then there exists an $\mathbb{R}^N$-valued, $\sigma(A,B)$-measurable random vector $W$ such that \begin{equation} \label{eq:LInftyerror} \mathbb{E}\left[\sup_{x \in [-M,M]^d} |H^{A,B}_W(x) - H(x) | \right] \leq \frac{4 (M \sqrt{d}+1) \sqrt{I}}{\sqrt{N}}. \end{equation} Moreover, $\|W_i\|_{L^\infty(\P)} \leq \frac{1}{N} \sup_{(u,\xi) \in \mathbb{R}\times \mathbb{R}^d} (\mathbbm{1}_{[-M\|\xi\|_1,0]}(u)+4\mathbbm{1}_{[-1,1]}(u)) \frac{|G(\xi)|+|G(-\xi)|}{\pi_{\text{b}}(u) \pi_{\text{w}}(\xi)}$ for $i=1,\ldots,N$. \end{theorem} \begin{remark} The proof of Theorem~\ref{thm:RC12Linfty} is based on several ingredients: firstly, \eqref{eq:Hrepresentation2} is used to derive an integral representation for $H$ (see \eqref{eq:auxEq37}). This representation is related to the Radon-wavelet integral representation (as used in \cite{MaiorovMeir2000}) and representations in \cite{Barron1992}, \cite{Barron1993}, \cite{KlusowskiBarron2018}. Secondly, the output weights $W$ are selected based on an ``importance sampling procedure'' (see \eqref{eq:auxEq39}). This matches the distribution of the random weights (which is chosen a priori and does not depend on $H$) with the function $\alpha$ in the integral representation for $H$ (see \eqref{eq:auxEq37}). Thirdly, Rademacher complexity-based techniques (\cite{Bartlett2003}, \cite{Boucheron2013}, \cite{Ledoux2013}) are employed to bound the $L^\infty$-error between the random neural network and the target function $H$ on the hypercube $[-M,M]^d$. The first two ingredients were also used in the proof of \citet[Theorem~1]{RC12}. \end{remark} \begin{proof} First, let us point out that for any $\mathbb{R}^N$-valued random vector $W$ the mapping $(\omega,x) \mapsto H^{A(\omega),B(\omega)}_{W(\omega)}(x) = \sum_{i=1}^N W_i(\omega) \varrho(A_i(\omega) \cdot x + B_i(\omega))$ is $\mathcal{F} \otimes \mathcal{B}(\mathbb{R}^d)$-measurable by \citet[Lemma~4.51]{aliprantis:border:infinite}. We now proceed in two steps. The first step consists in deriving an integral representation of $H$ based on \eqref{eq:Hrepresentation2}, as in \cite{RC12}. From this integral representation we construct the output weights $W$ based on an importance sampling procedure. The second step then uses Rademacher complexities to estimate the expected $L^\infty$-error. \textit{Step 1:} By considering separately the cases $r>0$ and $r<0$, one obtains for any $r \in \mathbb{R}$ the identity \begin{equation} \label{eq:ExpIdentity} e^{ir} - i r - 1 = - \int_{0}^\infty (r-u)^+ e^{i u} + (-r-u)^+ e^{-i u} d u. \end{equation} Inserting $r=\xi \cdot x$, multiplying by $G(\xi)$, integrating over $\xi \in \mathbb{R}^d$, employing the representation \eqref{eq:Hrepresentation2} and using Fubini's theorem (which can be applied due to \eqref{eq:BarronCondLInfty}) hence yields for any $x \in [-M,M]^d$ that \begin{equation} \label{eq:auxEq35} \begin{aligned} H(x) - x \cdot \nabla H(0) - H(0) & = \int_{\mathbb{R}^d} e^{i x \cdot \xi } G(\xi) - i x \cdot \xi G(\xi) - G(\xi) d \xi \\ & = - \int_{0}^\infty \int_{\mathbb{R}^d} [ (x \cdot \xi-u)^+ e^{i u} + (-x \cdot \xi-u)^+ e^{-i u}] G(\xi) d \xi d u . \end{aligned} \end{equation} Changing variables in the integral and using that for $x \in [-M,M]^d$ and $u \leq - M \|\xi\|_1$ we have $(x \cdot \xi + u)^+ =0$ then shows for all $x \in [-M,M]^d$ that \begin{equation} \label{eq:auxEq38} \begin{aligned} H(x) - x \cdot \nabla H(0) - H(0) & = - \int_{\mathbb{R}^d} \int_{- M \|\xi\|_1}^0 (x \cdot \xi+u)^+ [e^{-i u} G(\xi) + e^{i u} G(-\xi)] d u d \xi . \end{aligned} \end{equation} From the fact that $H(0)$ and $\nabla H(0) $ are elements of $\mathbb{R}$ one obtains $\int_{\mathbb{R}^d} \mathrm{Im}[G](\xi) d \xi = 0$ and $ \int_{\mathbb{R}^d} \xi \mathrm{Re}[G](\xi) d \xi = 0$ and hence we can represent \begin{equation} \label{eq:auxEq36} \begin{aligned} x \cdot \nabla H(0) + H(0) & = \int_{\mathbb{R}^d} - (x \cdot \xi) \mathrm{Im}[G](\xi) d \xi + \int_{\mathbb{R}^d} \mathrm{Re}[G](\xi) d \xi \\ & = \int_{\mathbb{R}^d} \int_0^1 [(x \cdot \xi +u)^+ - (-x \cdot \xi -u)^+ ](2\mathrm{Re}[G](\xi)-\mathrm{Im}[G](\xi)) du d \xi. \end{aligned} \end{equation} Combining \eqref{eq:auxEq38} and \eqref{eq:auxEq36} we obtain for all $x \in [-M,M]^d$ \begin{equation} \label{eq:auxEq37} \begin{aligned} H(x) & = \int_{\mathbb{R}^d} \int_{- \infty}^\infty (x \cdot \xi+u)^+ \alpha(\xi,u) d u d \xi \end{aligned} \end{equation} with \[ \alpha(\xi,u) = -\mathbbm{1}_{(-M \|\xi\|_1,0]}(u)\mathrm{Re}[e^{-i u} G(\xi) + e^{i u} G(-\xi)] + \mathbbm{1}_{[0,1]}(u)\tilde{g}(\xi)-\mathbbm{1}_{[-1,0]}(u)\tilde{g}(-\xi) \] for $\tilde{g}(\xi) = 2\mathrm{Re}[G](\xi)-\mathrm{Im}[G](\xi)$. Define for $(\xi,u) \in \mathbb{R}^d \times \mathbb{R}$ the function \[ f(\xi,u) = \frac{\alpha(\xi,u)}{\pi_{\text{w}}(\xi) \pi_{\text{b}}(u)} \] and choose the random vector $W=(W_1,\ldots,W_N)$ as \begin{equation} \label{eq:auxEq39} W_i =\frac{1}{N} f(A_i,B_i), \quad i=1,\ldots,N. \end{equation} The estimate \begin{equation}\label{eq:auxEq46} |f(\xi,u)| \leq (\mathbbm{1}_{(-M \|\xi\|_1,0]}(u) + 4\mathbbm{1}_{[-1,1]}(u)) \frac{|G(\xi)|+|G(-\xi)|}{\pi_{\text{w}}(\xi) \pi_{\text{b}}(u)} \end{equation} then proves the claimed bound on $\|W_i\|_{L^\infty(\P)}$ for $i=1,\ldots,N$. \textit{Step 2:} We now use the representation \eqref{eq:auxEq37} to prove \eqref{eq:LInftyerror} for the choice of $W$ made in \eqref{eq:auxEq39}. To this end, first notice that for any $x \in [-M,M]^d$ we have by the choice of $f$ and by the integral representation \eqref{eq:auxEq37} \[ \mathbb{E}[f(A_i,B_i) \varrho(A_i \cdot x + B_i)] = \int_{\mathbb{R}^d} \int_\mathbb{R} \varrho(x \cdot \xi + u) \alpha(\xi,u) d u d \xi = H(x). \] Therefore, letting $U_{i,x} = f(A_i,B_i) \varrho(A_i \cdot x + B_i)$ for $i=1,\ldots,N$ and $x \in [-M,M]^d$, we have \begin{equation} \label{eq:auxEq40} \begin{aligned} \mathbb{E}\left[\sup_{x \in [-M,M]^d} |H^{A,B}_W(x) - H(x) | \right] = \mathbb{E}\left[\sup_{x \in [-M,M]^d} \left|\frac{1}{N}\sum_{i=1}^N \left(U_{i,x} - \mathbb{E}[U_{i,x}]\right) \right| \right]. \end{aligned} \end{equation} Let $\varepsilon_1,\ldots,\varepsilon_N$ by i.i.d.\ Rademacher random variables which are independent of $A$ and $B$. Symmetrization (see e.g.\ \cite{Boucheron2013}) and \eqref{eq:auxEq40} then yields \begin{equation} \label{eq:auxEq41} \begin{aligned} \mathbb{E}\left[\sup_{x \in [-M,M]^d} |H^{A,B}_W(x) - H(x) | \right] \leq 2 \mathbb{E}\left[\sup_{x \in [-M,M]^d} \left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i U_{i,x} \right| \right]. \end{aligned} \end{equation} For $a=(a_1,\ldots,a_N) \in \mathbb{R}^{d}\times \cdots \times \mathbb{R}^d$, $b \in \mathbb{R}^N$ we let $T_{a,b} = \{(f(a_i,b_i)[ a_i \cdot x + b_i])_{i=1,\ldots,N} \,|\, x \in [-M,M]^d\}$. Then $T_{a,b} \subset \mathbb{R}^N$ is bounded, $\varrho$ is a contraction and hence independence and \citet[Theorem~4.12]{Ledoux2013} yield \begin{equation} \label{eq:auxEq42} \begin{aligned} \mathbb{E} & \left[\sup_{x \in [-M,M]^d} \left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i U_{i,x} \right| \right] \\ & = \mathbb{E} \left[ \left. \mathbb{E} \left[\sup_{t \in T_{a,b}} \left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i \varrho(t_i) \right|\right] \right\rvert_{(a,b)=(A,B)}\right] \\ & \leq 2 \mathbb{E} \left[ \left. \mathbb{E} \left[\sup_{t \in T_{a,b}} \left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i t_i \right|\right] \right\rvert_{(a,b)=(A,B)}\right] \\ & = 2 \mathbb{E} \left[ \left. \mathbb{E} \left[\sup_{x \in [-M,M]^d} \left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i f(a_i,b_i)[ a_i \cdot x + b_i] \right|\right] \right\rvert_{(a,b)=(A,B)}\right]. \end{aligned} \end{equation} Now for each $a,b$ we use Jensen's inequality and the fact that $\mathbb{E}[\varepsilon_i \varepsilon_j] = \delta_{ij}$ to estimate \begin{equation} \label{eq:auxEq43} \begin{aligned} \mathbb{E} & \left[\sup_{x \in [-M,M]^d} \left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i f(a_i,b_i)[ a_i \cdot x + b_i] \right|\right] \\ & \leq \mathbb{E} \left[ M \sqrt{d} \left\|\frac{1}{N}\sum_{i=1}^N \varepsilon_i f(a_i,b_i) a_i \right\|+ \left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i f(a_i,b_i) b_i \right|\right] \\ & \leq M \sqrt{d} \mathbb{E} \left[ \left\|\frac{1}{N}\sum_{i=1}^N \varepsilon_i f(a_i,b_i) a_i \right\|^2 \right]^{1/2}+ \mathbb{E} \left[\left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i f(a_i,b_i) b_i \right|^2\right]^{1/2} \\ & = \frac{M \sqrt{d}}{N} \left( \sum_{i=1}^N \left\| f(a_i,b_i) a_i \right\|^2 \right)^{1/2}+ \frac{1}{N} \left(\sum_{i=1}^N f(a_i,b_i)^2 b_i^2 \right)^{1/2}. \end{aligned} \end{equation} Inserting this in \eqref{eq:auxEq42} and using first Jensen's inequality and subsequently the fact that $(A_1,B_1),\ldots,(A_N,B_N)$ are identically distributed yields \begin{equation} \label{eq:auxEq44} \begin{aligned} \mathbb{E} & \left[\sup_{x \in [-M,M]^d} \left|\frac{1}{N}\sum_{i=1}^N \varepsilon_i U_{i,x} \right| \right] \\ & \leq \frac{2M \sqrt{d}}{N} \mathbb{E} \left[ \left( \sum_{i=1}^N \left\| f(A_i,B_i) A_i \right\|^2 \right)^{1/2} \right] + \frac{2}{N} \mathbb{E} \left[ \left(\sum_{i=1}^N f(A_i,B_i)^2 B_i^2 \right)^{1/2} \right] \\ & \leq \frac{2 M \sqrt{d}}{\sqrt{N}} \mathbb{E} \left[ \left\| f(A_1,B_1) A_1 \right\|^2 \right]^{1/2} + \frac{2}{\sqrt{N}} \mathbb{E} \left[ f(A_1,B_1)^2 B_1^2 \right]^{1/2}. \end{aligned} \end{equation} From the bound \eqref{eq:auxEq46} we obtain \begin{equation} \label{eq:auxEq45} \begin{aligned} & \mathbb{E} \left[ f(A_1,B_1)^2 \max(\| A_1 \|^2,B_1^2) \right] \\ & \leq \int_{\mathbb{R}^d} \int_\mathbb{R} \left[(\mathbbm{1}_{(-M \|\xi\|_1,0]}(u) + 4\mathbbm{1}_{[-1,1]}(u)) \frac{|G(\xi)|+|G(-\xi)|}{\pi_{\text{w}}(\xi) \pi_{\text{b}}(u)}\right]^2 \max(\|\xi\|^2,u^2) \pi_{\text{w}}(\xi) \pi_{\text{b}}(u) du d \xi \\ & \leq 2 \int_{\mathbb{R}^d} \int_\mathbb{R} (\mathbbm{1}_{(-M \|\xi\|_1,0]}(u) + 16\mathbbm{1}_{[-1,1]}(u)) \frac{(|G(\xi)|+|G(-\xi)|)^2}{\pi_{\text{w}}(\xi) \pi_{\text{b}}(u)} \max(\|\xi\|^2,u^2) du d \xi \\ & \leq \max(M^2,1) \int_{\mathbb{R}^d} \bar{F}(M \|\xi\|_1)\frac{(|G(\xi)|+|G(-\xi)|)^2}{\pi_{\text{w}}(\xi) } \|\xi\|_1^2 d \xi \\ & \quad +16 \int_{\mathbb{R}^d} (\bar{F}(1) - \bar{F}(-1)) \frac{(|G(\xi)|+|G(-\xi)|)^2}{\pi_{\text{w}}(\xi) } \max(\|\xi\|^2,1) d \xi \\ & \leq I. \end{aligned} \end{equation} Combining this with \eqref{eq:auxEq41} and \eqref{eq:auxEq44} yields \begin{equation} \label{eq:auxEq47} \begin{aligned} \mathbb{E} \left[\sup_{x \in [-M,M]^d} |H^{A,B}_W(x) - H(x) | \right] & \leq \frac{4 (M \sqrt{d}+1)}{\sqrt{N}} \mathbb{E} \left[ f(A_1,B_1)^2 \max(\| A_1 \|^2,B_1^2) \right]^{1/2} \\ & \leq \frac{4 (M \sqrt{d}+1) \sqrt{I}}{\sqrt{N}}, \end{aligned} \end{equation} which completes the proof. \end{proof} \begin{remark} With some additional work the weight distributions in Theorem~\ref{thm:RC12Linfty} could also be allowed to have compact support as in \citet[Theorem~1]{RC12}. However, in the results below (for instance in Theorem~\ref{thm:ApproxError}) such weight distributions would require much more restrictive assumptions on the unknown function $H$ and thus we do not pursue this direction here. \end{remark} \begin{corollary} \label{cor:RC12L2version} Assume that the hypotheses of Theorem~\ref{thm:RC12Linfty} are satisfied. Then the random vector $W$ from Theorem~\ref{thm:RC12Linfty} also satisfies that for any probability measure $\mu$ on $(\mathbb{R}^d,\mathcal{B}(\mathbb{R}^d))$ which is supported in $[-M,M]^d$ we have that \begin{equation} \label{eq:L2error5} \mathbb{E}\left[ \|H^{A,B}_W - H \|_{L^2(\mathbb{R}^d,\mu)}^2 \right]^{1/2} \leq \frac{(\sqrt{d} M+1)\sqrt{I}}{\sqrt{N}}. \end{equation} \end{corollary} \begin{proof} Using the same notation as in the proof of Theorem~\ref{thm:RC12Linfty}, we obtain from the proof of Theorem~\ref{thm:RC12Linfty} and by Tonelli's theorem and independence that \[ \begin{aligned} \mathbb{E}\left[ \|H^{A,B}_W - H \|_{L^2(\mathbb{R}^d,\mu)}^2 \right] & = \mathbb{E}\left[\int_{\mathbb{R}^d} \left|\frac{1}{N}\sum_{i=1}^N U_{i,x} - \mathbb{E}[U_{i,x}] \right|^2 \mu(dx) \right] \\ & = \int_{\mathbb{R}^d} \frac{1}{N^2}\sum_{i=1}^N \mathbb{E}\left[ \left| U_{i,x} - \mathbb{E}[U_{i,x}] \right|^2 \right] \mu(dx) \\ & \leq \int_{\mathbb{R}^d} \frac{1}{N} \mathbb{E}\left[ \left| f(A_1,B_1) \varrho(A_1 \cdot x + B_1) \right|^2 \right] \mu(dx) \\ & \leq (\sqrt{d} M + 1)^2 \frac{\mathbb{E}\left[ \left| f(A_1,B_1) \max(\|A_1\|,|B_1|) \right|^2 \right] }{N} . \end{aligned} \] Therefore, \eqref{eq:auxEq45} yields the claimed bound. \end{proof} \section{Random neural network approximation bounds} \label{sec:Approx} In this section we use random neural networks to approximate functions with a convolutional structure. In Section~\ref{subsec:ApproxGeneral} we derive approximation error bounds with explicit dependence on the dimension $d$. These results are then applied in Section~\ref{subsec:ApproxLevy} in the context of exponential L\'evy models, which include the Black-Scholes model as a special case. \subsection{Bounds for convolutional functions} \label{subsec:ApproxGeneral} Consider a function $H \colon \mathbb{R}^d \to \mathbb{R}$ given by $H(x) = \mathbb{E}[\Phi(x+V)]$ for an $\mathbb{R}^d$-valued random vector $V$ and a function $\Phi\colon \mathbb{R}^d \to \mathbb{R}$. Assume that the characteristic function of $V$ satisfies the following bound: there exists $C>0$ such that \begin{equation} \label{eq:charFctAss} |\mathbb{E}[e^{i \xi \cdot V}]|\leq \exp(-C\|\xi\|^2) \quad \text{ for all } \xi \in \mathbb{R}^d. \end{equation} Examples of functions $H$ of this type include expectations (respectively option prices) and associated solutions to PDEs in (exponential) L\'evy models with non-degenerate Gaussian component, see Section~\ref{subsec:ApproxLevy} below. We now approximate $H$ by a random neural network $H^{A,B}_W$ and analyze the approximation error. As above the randomly generated hidden weights $A,B$ are not trainable and the goal is to find $W$ such that the expected uniform approximation error $\mathbb{E}[ \|H^{A,B}_W - H \|_{L^\infty([-M,M]^d)}]$ is small. The output weight vector $W$ may be chosen depending on $A,B$, i.e., it is a $\sigma(A,B)$-measurable random variable. Our goal is to obtain approximation error bounds in which the dependence on the dimension $d$ is fully explicit. To achieve this we need more specific assumptions on the distributions from which the hidden weights $A$ and $B$ are drawn. Recall that $\pi_{\text{b}}$ denotes the Lebesgue-density of $B_1$ and $\pi_{\text{w}}$ denotes the density of $A_1$. We will assume below that $\pi_{\text{w}}$ is the density of a multivariate $t$-distribution $t_{\nu}(0,\mathbbm{1}_d)$ for some $\nu > 1$ and that $\pi_{\text{b}}$ has \textit{at most polynomial decay}, that is, there exists a polynomial $p_{\text{b}} \colon \mathbb{R} \to (0,\infty)$ such that \begin{equation} \label{eq:polyTails1D} 1 \leq p_{\text{b}}(z) \pi_{\text{b}}(z) \quad \text{ for all } z \in \mathbb{R}. \end{equation} This hypothesis is satisfied, for instance, by Student's $t$-distribution. These assumptions allow us to obtain explicit control of the normalizing constant of the weight distribution $\pi_{\text{w}}$. \begin{theorem} \label{thm:ApproxError} Let $C>\frac{1}{2^{3/2} \pi}$ and let $\nu>1$. Suppose $A_1 \sim t_{\nu}(0,\mathbbm{1}_d)$ and $B_1$ has density $\pi_{\text{b}}$ satisfying \eqref{eq:polyTails1D}. Then there exist $k \in \mathbb{N}$ and an absolute constant $C_{\text{app}}>0$ such that for any $H \colon \mathbb{R}^d \to \mathbb{R}$ of the form $H(x) = \mathbb{E}[\Phi(x+V)]$ with $\Phi \in L^1(\mathbb{R}^d)$ and $V$ satisfying \eqref{eq:charFctAss} the following random neural network approximation result holds: there exists an $\mathbb{R}^N$-valued, $\sigma(A,B)$-measurable random vector $W$ such that \begin{equation} \label{eq:L2error2} \mathbb{E}\left[\sup_{x \in [-M,M]^d} |H^{A,B}_W(x) - H(x) | \right] \leq \frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}}. \end{equation} The constant $k$ only depends on $\pi_{\text{b}}$ and the constant $C_{\text{app}}$ depends on $\nu, \pi_{\text{b}}, C, M$, but it does not depend on $d$, $N$ or $H$. Moreover, \begin{equation} \label{eq:Wbound} \|W_i\|_{L^\infty(\P)} \leq \frac{C_{\text{wgt}}\|\Phi\|_{L^1(\mathbb{R}^d)}(\nu+d)^{2k+\frac{1}{2}}}{N} \end{equation} for $i=1,\ldots,N$, where the constant $C_{\text{wgt}}>0$ depends on $\nu, \pi_{\text{b}}, C, M$, but it does not depend on $d$, $N$ or $H$. \end{theorem} \begin{remark} \label{rmk:L2error} In addition to the uniform bound in \eqref{eq:L2error2} the proof of Theorem~\ref{thm:ApproxError} also shows that for any probability measure $\mu$ on $(\mathbb{R}^d,\mathcal{B}(\mathbb{R}^d))$ supported in $[-M,M]^d$ we have \begin{equation} \label{eq:L2error6} \mathbb{E}\left[ \|H^{A,B}_W - H \|_{L^2(\mathbb{R}^d,\mu)}^2 \right]^{1/2} \leq \frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}}. \end{equation} This follows directly by using Corollary~\ref{cor:RC12L2version} instead of Theorem~\ref{thm:RC12Linfty} in \eqref{eq:auxEq48} below. \end{remark} \begin{remark} Hypothesis \eqref{eq:charFctAss} in Theorem~\ref{thm:ApproxError} is employed in the proof in order to guarantee that the constant in the error bound does not grow exponentially in the dimension $d$. In low-dimensional situations this behaviour may not be required and hence \eqref{eq:charFctAss} could be replaced by the weaker hypothesis $|\mathbb{E}[e^{i \xi \cdot V}]|\leq \exp(-C\|\xi\|^\alpha)$ for some $C>0$, $\alpha >0$ or even by the assumption that $|\mathbb{E}[e^{i \xi \cdot V }]|\leq C (1+\|\xi\|)^{-\beta}$ for some $C>0$ and sufficiently large $\beta >0$ (depending on $\nu$ and $\pi_{\text{b}}$). In this situation the error bound \eqref{eq:L2error2} is still valid with a different constant $C_{\text{app}}$ and an additional factor which is potentially exponential in $d$. \end{remark} \begin{proof} Let $H \colon \mathbb{R}^d \to \mathbb{R}$ be of the form $H(x) = \mathbb{E}[\Phi(x+V)]$ with $\Phi \in L^1(\mathbb{R}^d)$ and $V$ satisfying \eqref{eq:charFctAss}. We verify that $H$ satisfies the hypotheses of Theorem~\ref{thm:RC12Linfty} and derive a bound for the constant $I$ in \eqref{eq:IfiniteLinfty} with the claimed properties. For $f \in L^1(\mathbb{R}^d)$ we denote by $\hat{f}$ the Fourier transform of $f$ given for all $\xi \in \mathbb{R}^d$ by $\hat{f}(\xi) = (2\pi)^{-\frac{d}{2}}\int_{\mathbb{R}^d} e^{-i x \cdot \xi} f(x) d x$ . By \eqref{eq:charFctAss} and \citet[Proposition~2.5(xii)]{Sato1999} the random variable $-V$ has a bounded Lebesgue-density $p_{-V}$. Thus, we can write $H(x) = \int_{\mathbb{R}^d} \Phi(x-y) p_{-V}(y) d y = (\Phi * p_{-V})(x)$. The convolution theorem (see for instance \citet[Theorem~X.9.16]{Amann2009}) hence shows that $\hat{H}(\xi) = (2\pi)^{\frac{d}{2}} \hat{\Phi}(\xi)\widehat{p_{-V}}(\xi)$. Combining this with $\widehat{p_{-V}}(\xi) = (2\pi)^{-\frac{d}{2}}\int_{\mathbb{R}^d} e^{-i x \cdot \xi} p_{-V}(x) d x = (2\pi)^{-\frac{d}{2}} \mathbb{E}[e^{-i\xi \cdot (-V)}]$ and \eqref{eq:charFctAss} we obtain that $\hat{H}$ is integrable. The Fourier inversion theorem (see for instance \citet[Theorem~X.9.12]{Amann2009}) therefore yields for all $x \in \mathbb{R}^d$ that \begin{equation} \label{eq:HFourier} H(x) = (2 \pi)^{-\frac{d}{2}}\int_{\mathbb{R}^d} e^{i \xi \cdot x} \hat{H}(\xi) d \xi = (2\pi)^{-\frac{d}{2}} \int_{\mathbb{R}^d} e^{i \xi \cdot x} \hat{\Phi}(\xi) \mathbb{E}[e^{i\xi \cdot V}] d \xi. \end{equation} Hence, the representation \eqref{eq:Hrepresentation2} holds for all $x \in \mathbb{R}^d$ with $G(\xi) = (2\pi)^{-\frac{d}{2}} \hat{\Phi}(\xi) \mathbb{E}[e^{i\xi \cdot V}]$, $\xi \in \mathbb{R}^d$. Condition \eqref{eq:BarronCondLInfty} is satisfied, since \eqref{eq:charFctAss} implies \[ \int_{\mathbb{R}^d} \max(1,\|\xi\|^2) |G(\xi)| d\xi \leq (2\pi)^{-{d}} \|\Phi\|_{L^1(\mathbb{R}^d)} \int_{\mathbb{R}^d} \max(1,\|\xi\|^2) \exp(-C\|\xi\|^2) d\xi < \infty. \] Denote by $k \in \mathbb{N}$ the degree of $p_{\text{b}}$, then there exist $a_0,\ldots,a_k \in \mathbb{R}$ such that $p_{\text{b}}(s) = \sum_{l=0}^k a_l s^l$ for all $s \in \mathbb{R}$. Then $|p_{\text{b}}(s)| \leq (k+1) \max_l\{|a_l|\} \max(1,|s|^k) \leq C_{\text{b}} (1+s^{2k})$ for all $s \in \mathbb{R}$, where $ C_{\text{b}} = (k+1) \max_l\{|a_l|\}$. Consequently, we may use \eqref{eq:polyTails1D} to estimate for any $r\geq 0$ \[ \bar{F}(r) = 2\int_{-r}^0 \frac{1}{\pi_{\text{b}}(s)} ds \leq 2\int_{-r}^0 p_{\text{b}}(s) ds \leq 2 C_{\text{b}} (r+\frac{r^{2k+1}}{2k+1})< \infty \] and for $r<0$ analogously $|\bar{F}(r)| = 2\int_{0}^{-r} \frac{1}{\pi_{\text{b}}(s)} ds \leq - 2 C_{\text{b}} (r+\frac{r^{2k+1}}{2k+1}) < \infty$. Therefore, $\bar{F}(1)-\bar{F}(-1) \leq 8 C_{\text{b}} $ and we can now use the comparison $\|\cdot\|_1 \leq \sqrt{d} \|\cdot\| $ on $\mathbb{R}^d$ to estimate the constant $I$ in \eqref{eq:IfiniteLinfty} as \[ \begin{aligned} I & \leq 2 C_{\text{b}} c_{M,1} \int_{\mathbb{R}^d} \left[ M\|\xi\|_1^3+\frac{\|\xi\|_1^2(M\|\xi\|_1)^{2k+1}}{2k+1} + 4 \max(1,\|\xi\|^2)\right] \frac{(|G(\xi)|+|G(-\xi)|)^2}{ \pi_{\text{w}}(\xi)} d \xi \\ & \leq 2 C_{\text{b}} c_{M,2} d^{k+\frac{3}{2}} \int_{\mathbb{R}^d} \left[\|\xi\|^3+\|\xi\|^{2k+3} + \max(1,\|\xi\|^2)\right] \frac{(|G(\xi)|+|G(-\xi)|)^2}{ \pi_{\text{w}}(\xi)} d \xi \\ & \leq 6 C_{\text{b}} c_{M,2} d^{k+\frac{3}{2}} \int_{\mathbb{R}^d} \max(1,\|\xi\|^{2k+3}) \frac{(\hat{\Phi}(\xi) \mathbb{E}[e^{i\xi \cdot V}]+\hat{\Phi}(-\xi) \mathbb{E}[e^{-i\xi \cdot V}])^2}{(2\pi)^{d} \pi_{\text{w}}(\xi)} d \xi \\ & \leq C_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2 \int_{\mathbb{R}^d} \max(1,\|\xi\|^{2k+3}) \frac{\exp(-2C\|\xi\|^2)}{(2\pi)^{2d} \pi_{\text{w}}(\xi)} d \xi \end{aligned} \] with $c_{M,1} = \max(M^2,16)$, $c_{M,2}=c_{M,1} \max(M^{2k+1},4)$, $C_I = 24 C_{\text{b}} c_{M,2}$. We now insert the density $\pi_{\text{w}}(x) = \frac{\Gamma((\nu+d)/2)}{\Gamma(\nu/2) \nu^{d/2} \pi^{d/2}} (1 + \nu^{-1} \|x\|^2)^{-(\nu+d)/2}$ and use the estimate $(\nu + \|x\|^2)^{p} \leq 2^{p-1}(\nu^{p}+ \|x\|^{2p})$ for $p \geq 1$ to obtain \begin{equation} \label{eq:auxEq1} \begin{aligned} I & \leq C_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2 \int_{\mathbb{R}^d} \max(1,\|\xi\|^{2k+3}) (\nu + \|\xi\|^2)^{(\nu+d)/2} \frac{\exp(-2C\|\xi\|^2) \Gamma(\frac{\nu}{2}) \nu^{-\frac{\nu}{2}} \pi^{d/2}}{(2\pi)^{2d} \Gamma(\frac{\nu+d}{2})} d \xi \\ & \leq C_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2 \Gamma\left(\frac{\nu}{2}\right) \nu^{-\nu/2} \int_{\mathbb{R}^d} (\nu + \|\xi\|^2)^{(2k+3+\nu+d)/2} \frac{\exp(-2C\|\xi\|^2) \pi^{d/2}}{(2\pi)^{2d} \Gamma(\frac{\nu+d}{2})} d \xi \\ & \leq \frac{\tilde{C}_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2}{(2\pi)^{3d/2} \Gamma(\frac{\nu+d}{2})} \int_{\mathbb{R}^d} (\nu^{(\tilde{\nu}+d)/2} + \|\xi\|^{\tilde{\nu}+d}) \exp(-2C\|\xi\|^2) d \xi \end{aligned} \end{equation} with $\tilde{\nu} = 2k+3+\nu$, $\tilde{C}_I = 2^{(\tilde{\nu}/2)-1} C_I \Gamma(\frac{\nu}{2}) \nu^{-\nu/2}$. Denote by $X_C$ a random variable with a $\mathcal{N}(0,\frac{1}{4C}\mathbbm{1}_d)$-distribution. Then the last line in \eqref{eq:auxEq1} can be rewritten in terms of $X_C$, yielding \begin{equation} \label{eq:auxEq2} \begin{aligned} I & \leq \frac{\tilde{C}_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2}{(2\pi)^{d} \Gamma(\frac{\nu+d}{2})(2^d C^{d/2})} \int_{\mathbb{R}^d} (\nu^{(\tilde{\nu}+d)/2} + \|\xi\|^{\tilde{\nu}+d}) \frac{\exp(-2C\|\xi\|^2)}{(2\pi)^{d/2}(4 C)^{-d/2}} d \xi \\ & = \frac{\tilde{C}_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2}{(2\pi)^{d} \Gamma(\frac{\nu+d}{2})(2^d C^{d/2})} \left[ \nu^{(\tilde{\nu}+d)/2} + \mathbb{E}[ \|X_C\|^{\tilde{\nu}+d}] \right]. \end{aligned} \end{equation} On the other hand, $\mathbb{E}[ \|X_C\|^{\tilde{\nu}+d}] = \mathbb{E}[ \|(2\sqrt{C})^{-1}Z\|^{\tilde{\nu}+d}] = (2\sqrt{C})^{-(\tilde{\nu}+d)} \mathbb{E}[ (\|Z\|^2)^{\frac{\tilde{\nu}+d}{2}}]$ where $Z$ is a $d$-dimensional standard normal random vector. Hence, $\|Z\|^2$ has a $\chi^2(d)$-distribution and thus \[ \begin{aligned} \mathbb{E}[ (\|Z\|^2)^{(\tilde{\nu}+d)/2}] & = 2^{-d/2} [\Gamma(d/2)]^{-1}\int_0^\infty x^{\frac{\tilde{\nu}+2d}{2}-1} e^{-x/2} dx = \frac{2^{(\tilde{\nu}+2d)/2} \Gamma((\tilde{\nu}+2d)/2)}{2^{d/2} \Gamma(d/2)} \\ & = \frac{2^{(\tilde{\nu}+d)/2} \Gamma((\tilde{\nu}+2d)/2)}{ \Gamma(d/2)}. \end{aligned} \] Combining this with \eqref{eq:auxEq2} and the upper and lower bounds for the gamma function (see, e.g., \citet[Lemma~2.4]{Gonon2019}) we obtain \begin{align} \label{eq:auxEq3} I & \leq \frac{\tilde{C}_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2}{(2\pi)^{d} \Gamma(\frac{\nu+d}{2})(2^d C^{d/2})} \left[ \nu^{(\tilde{\nu}+d)/2} + (2\sqrt{C})^{-(\tilde{\nu}+d)} \frac{2^{(\tilde{\nu}+d)/2} \Gamma((\tilde{\nu}+2d)/2)}{ \Gamma(d/2)} \right] \\ \notag & \leq \frac{\tilde{C}_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2}{(2\pi)^{d} (2^d C^{d/2})} (\frac{\nu+d}{4 \pi})^{\frac{1}{2}} (\frac{2e}{\nu+d})^{\frac{\nu+d}{2}} \left[ \nu^{\frac{\tilde{\nu}+d}{2}} + \frac{e}{(2C)^{\frac{\tilde{\nu}+d}{2}}} (\frac{2e}{d})^{\frac{d}{2}} (\frac{d}{\tilde{\nu}+2d})^{\frac{1}{2}} (\frac{\tilde{\nu}+2d}{2e})^{\frac{\tilde{\nu}+2d}{2}} \right] \\ \notag & \leq \frac{\tilde{C}_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2\sqrt{\nu+d} }{(16 \pi^2 C)^{\frac{\tilde{\nu}+d}{2}}} (\frac{2e}{\nu+d})^{\frac{\tilde{\nu}+d-2k-3}{2}} (4 \pi \sqrt{C})^{\tilde{\nu}} \left[ \nu^{\frac{\tilde{\nu}+d}{2}} + \frac{(2e)^{\frac{d}{2}}}{(2C)^{\frac{\tilde{\nu}+d}{2}}d^{\frac{d}{2}}} (\frac{\tilde{\nu}+2d}{2e})^{\frac{\tilde{\nu}+2d}{2}} \right] \\ \notag & \leq \bar{C}_I d^{k+\frac{3}{2}} \|\Phi\|_{L^1(\mathbb{R}^d)}^2 (\nu+d)^{k+2} \left[ (\frac{e\nu}{8\pi^2C(\nu+d)})^{\frac{\tilde{\nu}+d}{2}} + (\frac{\tilde{\nu}+2d}{32 \pi^2 C^2(\nu+d)})^{\frac{\tilde{\nu}+d}{2}} (\frac{\tilde{\nu}+2d}{d})^{\frac{d}{2}} \right] \\ \notag & \leq \bar{C}_I \|\Phi\|_{L^1(\mathbb{R}^d)}^2 (\nu+d)^{2k+\frac{7}{2}} \left[ (\frac{e\nu}{8\pi^2C(\nu+d)})^{\frac{\tilde{\nu}+d}{2}} + (\frac{\tilde{\nu}+2d}{32 \pi^2 C^2(\nu+d)})^{\frac{\tilde{\nu}}{2}} (\frac{(\tilde{\nu}+2d)^2}{32 \pi^2 C^2(\nu+d)d})^{\frac{d}{2}} \right] \end{align} with $\bar{C}_I = \tilde{C}_I (2e)^{-\frac{2k+3}{2}} (16 \pi^2 C)^{\frac{\tilde{\nu}}{2}}$. Now clearly \begin{equation} \label{eq:C1} C_1 := \sup_{m \in \mathbb{N}} \left\{ \left(\frac{e\nu}{8\pi^2C(\nu+m)}\right)^{\frac{\tilde{\nu}+m}{2}} \right\} < \infty \end{equation} and \begin{equation} \label{eq:C2} \begin{aligned} C_2 : & = \sup_{m \in \mathbb{N}} \left\{ \left(\frac{(\tilde{\nu}+2m)^2}{32 \pi^2 C^2(\nu+m)m}\right)^{\frac{m}{2}} \right\} = \sup_{m \in \mathbb{N}} \left\{ \left(\frac{(\frac{\tilde{\nu}}{m}+2)^2}{32 \pi^2 C^2(\frac{\nu}{m}+1)}\right)^{\frac{m}{2}} \right\} \\ & \leq \sup_{m \in \mathbb{N}} \left\{ \left(\frac{(\frac{\tilde{\nu}}{m}+2)^2}{32 \pi^2 C^2}\right)^{\frac{m}{2}} \right\} < \infty, \end{aligned} \end{equation} because $C^2 > \frac{1}{8 \pi^2}$ and hence $m_0:=\frac{\tilde{\nu}}{2(\sqrt{8}\pi C-1)} >0$ and $(\frac{\tilde{\nu}}{m}+2)^2 < 32 \pi^2 C^2$ for all $m \in \mathbb{N}$ with $m > m_0$. Combining this with \eqref{eq:auxEq3} we have therefore proved that \begin{equation} \label{eq:auxEq55} \begin{aligned} I & \leq \bar{C}_I \|\Phi\|_{L^1(\mathbb{R}^d)}^2 (\nu+d)^{2k+\frac{7}{2}} \left[ C_1 + \left(\frac{\tilde{\nu}+2d}{32 \pi^2 C^2(\nu+d)}\right)^{\frac{\tilde{\nu}}{2}} C_2 \right] \\ & \leq \bar{C}_I \|\Phi\|_{L^1(\mathbb{R}^d)}^2 (\nu+d)^{2k+\frac{7}{2}} \left[ C_1 + \left(\frac{\tilde{\nu}+2}{32 \pi^2 C^2}\right)^{\frac{\tilde{\nu}}{2}} C_2 \right] \\ & = C_3 \|\Phi\|_{L^1(\mathbb{R}^d)}^2 (\nu+d)^{2k+\frac{7}{2}} \end{aligned} \end{equation} with $C_3 = \bar{C}_I(C_1 + (\frac{\tilde{\nu}+2}{32 \pi^2 C^2})^{\frac{\tilde{\nu}}{2}} C_2) $. Altogether, the hypotheses of Theorem~\ref{thm:RC12Linfty} are satisfied and hence there exists an $\mathbb{R}^N$-valued, $\sigma(A,B)$-measurable random vector $W$ such that the error bound \eqref{eq:LInftyerror} holds. Inserting \eqref{eq:auxEq55} yields \begin{equation} \label{eq:auxEq48} \begin{aligned} \mathbb{E}\left[\sup_{x \in [-M,M]^d} |H^{A,B}_W(x) - H(x) | \right] & \leq \frac{4 (M+1) \sqrt{d} \sqrt{I}}{\sqrt{N}} \\ & \leq \frac{4 (M+1)\sqrt{C_3} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}} . \end{aligned} \end{equation} Hence, the $L^\infty$-error estimate \eqref{eq:L2error2} follows with \[ C_{\text{app}} =4 (M+1)\sqrt{C_3} =4 (M+1) (\bar{C}_I)^{1/2} \left[ C_1 + \left(\frac{\tilde{\nu}+2}{32 \pi^2 C^2}\right)^{\frac{\tilde{\nu}}{2}} C_2 \right]^{1/2} \] where we recall that $\bar{C}_I = 24 C_{\text{b}} \max(M^2,16) \max(M^{2k+1},4) 2^{(\tilde{\nu}/2)-1} \Gamma(\frac{\nu}{2}) \nu^{-\nu/2} (2e)^{-\frac{2k+3}{2}}$ $ (16 \pi^2 C)^{\frac{\tilde{\nu}}{2}}$, $\tilde{\nu} = 2k+3+\nu$, the constants $ C_{\text{b}}$ and $k$ only depend on $p_{\text{b}}$ and $C_1$, $C_2$ are given by \eqref{eq:C1} and \eqref{eq:C2}, respectively. To prove the upper bound on $W$ we insert the bound from Theorem~\ref{thm:RC12Linfty}, then \eqref{eq:charFctAss} and \eqref{eq:polyTails1D} can be used to estimate similarly as before for $i=1,\ldots,N$ \[\begin{aligned} \|W_i\|_{L^\infty(\P)} & \leq \frac{1}{N} \sup_{(u,\xi) \in \mathbb{R}\times \mathbb{R}^d} (\mathbbm{1}_{[-M\|\xi\|_1,0]}(u)+4\mathbbm{1}_{[-1,1]}(u)) \frac{|G(\xi)|+|G(-\xi)|}{\pi_{\text{b}}(u) \pi_{\text{w}}(\xi)} \\ & \leq \frac{5}{N} \sup_{\xi \in \mathbb{R}^d} \frac{2(2\pi)^{-d} \|\Phi\|_{L^1(\mathbb{R}^d)} \exp(-C\|\xi\|^2) C_{\text{b}} (1+\max(1,M\|\xi\|_1)^{2k})}{\pi_{\text{w}}(\xi)} \\ & \leq \frac{\tilde{C}_{\text{wgt}} d^k \|\Phi\|_{L^1(\mathbb{R}^d)}}{N} \sup_{\xi \in \mathbb{R}^d} \frac{ \exp(-C\|\xi\|^2) (2+M^{2k}\|\xi\|^{2k}) (\nu + \|\xi\|^2)^{(\nu+d)/2}}{2^{\frac{d}{2}} \Gamma((\nu+d)/2) (2\pi)^{\frac{d}{2}}} \\ & \leq \frac{\tilde{C}_{\text{wgt}} d^k \|\Phi\|_{L^1(\mathbb{R}^d)}\max(2,M^{2k})}{N} \max_{r \geq 0} \frac{ \exp(-Cr) (\nu + r)^{(\nu+d+2k)/2}}{2^{\frac{d}{2}} \Gamma((\nu+d)/2) (2\pi)^{\frac{d}{2}}} \\ & = \frac{\tilde{C}_{\text{wgt}} d^k \|\Phi\|_{L^1(\mathbb{R}^d)}\max(2,M^{2k})}{N} e^{\nu C} \frac{ ((\nu+d+2k)/(2Ce))^{(\nu+d+2k)/2}}{2^{\frac{d}{2}} \Gamma((\nu+d)/2)(2\pi)^{\frac{d}{2}}} \\ & \leq \frac{\bar{C}_{\text{wgt}}\|\Phi\|_{L^1(\mathbb{R}^d)}(\nu+d)^{2k+\frac{1}{2}}}{N} \left(\frac{\nu+d+2k}{4 \pi C(\nu+d)}\right)^{(\nu+d+2k)/2} \end{aligned} \] with $\tilde{C}_{\text{wgt}} =10C_{\text{b}}\Gamma(\nu/2) \nu^{-\nu/2}$, $\bar{C}_{\text{wgt}}= \tilde{C}_{\text{wgt}} \max(2,M^{2k}) e^{\nu C} (2e)^{-k} (4\pi)^{(\nu+2k-1)/2} $. In the last two steps we used that the maximum is attained for $\nu + r = (\nu+d+2k)/(2C)$ and we applied the lower bound for the gamma function as in \eqref{eq:auxEq3}. The hypothesis $8 \pi^2 C^2> 1$ implies $4 \pi C >1$ and therefore \[ C_4 := \sup_{m \in \mathbb{N}} \left\lbrace \left(\frac{\nu+m+2k}{4\pi C(\nu+m)}\right)^{(\nu+m+2k)/2} \right\rbrace < \infty \] by a similar reasoning as used to argue that $C_2$ in \eqref{eq:C2} is finite. Hence, the bound on $\|W_i\|_{L^\infty(\P)}$ follows with $C_{\text{wgt}} = \bar{C}_{\text{wgt}} C_4$. This completes the proof. \end{proof} We now show that an analogous approximation result holds when the assumption $\Phi \in L^1(\mathbb{R}^d)$ is replaced by the assumptions that $\Phi$ satisfies a Lipschitz-condition and $V$ admits certain moments. Here we call $\psi \colon \mathbb{R}^d \to \mathbb{R}^d$ increasing if for any $x,y \in \mathbb{R}^d$ with $x_i\leq y_i$ for all $i=1,\ldots,d$ it holds that $\psi_i(x) \leq \psi_i(y)$ for all $i=1,\ldots,d$. Furthermore, we denote $\mathbf{1}=(1,\ldots,1) \in \mathbb{R}^d$. \begin{proposition}\label{prop:RNNapprox} Let $C, C_{\text{Lip}} >0$ with $C^2 > \frac{1}{8 \pi^2}$ and let $\nu>1$. Suppose $A_1 \sim t_{\nu}(0,\mathbbm{1}_d)$ and $B_1$ has density $\pi_{\text{b}}$ satisfying \eqref{eq:polyTails1D}. Let $\psi \colon \mathbb{R}^d \to \mathbb{R}^d$ be increasing and measurable. Let $H \colon \mathbb{R}^d \to \mathbb{R}$ be of the form $H(x) = \mathbb{E}[\Phi(x+V)]$ for $\Phi$ satisfying \begin{equation} \label{eq:Psi-Lipschitz} |\Phi(x)-\Phi(y)| \leq C_{\text{Lip}} \|\psi(x)-\psi(y)\|, \quad x,y \in \mathbb{R}^d \end{equation} and $V$ satisfying \eqref{eq:charFctAss}, $\mathbb{E}[\|\psi(M\mathbf{1}+V)\|^2]< \infty$. Then for any $R>0$ there exists an $\mathbb{R}^N$-valued, $\sigma(A,B)$-measurable random vector $W$ such that \begin{equation} \label{eq:L2error3} \mathbb{E}\left[\sup_{x \in B_M(0)} |H^{A,B}_W(x) - H(x) | \right] \leq \frac{C_{\text{app}} \mathcal{I}(R) (\nu+d)^{k+3}}{\sqrt{N}} + C_{\text{mom}} \P(\|V\|> R)^{1/2}, \end{equation} where $\mathcal{I}(R) = \int_{\mathbb{R}^d} |\Phi(x)| \mathbbm{1}_{\{\|x\|\leq M + R\}} dx $, $C_{\text{mom}} = C_{\text{Lip}} (\mathbb{E}[\|\psi(M\mathbf{1}+V)\|^2]^{1/2} + \|\psi(0)\|)+|\Phi(0)| $ and $k \in \mathbb{N}$, $C_{\text{app}}>0$ are as in Theorem~\ref{thm:ApproxError}. \end{proposition} \begin{proof} Let $R>0$ and denote $\Phi^R(x)= \Phi(x) \mathbbm{1}_{\{\|x\|\leq M + R\}}$. Set $\bar{H}^R(x) = \mathbb{E}[\Phi^R(x+V)] $. Then for $x \in B_M(0)$ we estimate \[ \begin{aligned} |& \bar{H}^R(x) -H(x)| \\ & \leq \mathbb{E}[|\Phi(x+V)\mathbbm{1}_{\{\|x+V\|\leq M + R\}}-\Phi(x+V) |] \\ & = \mathbb{E}[\mathbbm{1}_{\{\|x+V\|> M + R\}}|\Phi(x+V)|] \\ & \leq C_{\text{Lip}}\mathbb{E}[\mathbbm{1}_{\{\|x+V\|> M + R\}}\|\psi(x+V)-\psi(0)\|] + |\Phi(0)|\P(\|x+V\|> M + R) \\ & \leq C_{\text{Lip}}\mathbb{E}[\mathbbm{1}_{\{\|x+V\|> M + R\}}\|\psi(M\mathbf{1}+V)\|] + (C_{\text{Lip}}\|\psi(0)\|+|\Phi(0)|)\P(\|x+V\|> M + R) \\ & \leq C_{\text{Lip}}\mathbb{E}[\mathbbm{1}_{\{\|V\|> R\}}\|\psi(M\mathbf{1}+V)\|] + (C_{\text{Lip}}\|\psi(0)\|+|\Phi(0)|)\P(\|V\|> R) \\ & \leq C_{\text{Lip}}\P(\|V\|> R)^{1/2}\mathbb{E}[\|\psi(M\mathbf{1}+V)\|^2]^{1/2} + (C_{\text{Lip}}\|\psi(0)\|+|\Phi(0)|)\P(\|V\|> R)^{1/2}. \end{aligned} \] The truncated function $\Phi^R$ is integrable and \[ \|\Phi^R\|_{L^1(\mathbb{R}^d)} = \int_{\mathbb{R}^d} |\Phi(x)| \mathbbm{1}_{\{\|x\|\leq M + R\}} dx = \mathcal{I}(R). \] Therefore, the result follows from Theorem~\ref{thm:ApproxError} and the triangle inequality. \end{proof} \begin{remark} Let us now explain how Proposition~\ref{prop:RNNapprox} could be applied. In the case of exponential L\'evy models we would choose $\Phi(x)=\varphi(\exp(x))$ for $\varphi \colon \mathbb{R}^d \to \mathbb{R}$. Hence, if $\varphi$ is $C_{\text{Lip}}$-Lipschitz-continuous, then \eqref{eq:Psi-Lipschitz} is satisfied with $\psi(x)=(\exp(x_1),\ldots,\exp(x_d))$ for $x \in \mathbb{R}^d$. Consequently, if we choose $R = \frac{1}{\alpha}\log(N)$ for some $\alpha>1$, then \[ \begin{aligned} \mathcal{I}(R) & = \int_{\mathbb{R}^d} |\Phi(x)| \mathbbm{1}_{\{\|x\|\leq M + R\}} dx \leq \int_{\mathbb{R}^d} c (1+\|\exp(x)\|) \mathbbm{1}_{\{\|x\|\leq M + R\}} dx \\ & \leq \int_{\mathbb{R}^d} c (1+d^{\frac{1}{2}} \exp(M+R)) \mathbbm{1}_{\{\|x\|\leq M + R\}} dx \\ & = c (1+d^{\frac{1}{2}} \exp( M+R)) \mathrm{Vol}(B_{M+R}(0)) \\ & \leq c (1+d^{\frac{1}{2}} e^{M} N^{\frac{1}{\alpha}}) (d\pi)^{-1/2}\left(\frac{2\pi e}{d}\right)^{d/2}\left(M+\frac{\log(N)}{\alpha}\right)^d \\ & \leq \tilde{c} N^{\frac{1}{\alpha}} \end{aligned} \] with $c = \max(C_{\text{Lip}},|\varphi(0)|)$, $\tilde{c} = 2 c \max(1,e^{M}) \pi^{-1/2}$ and where the last step holds if the number of nodes satisfies the condition $N \leq \exp(\alpha[d^{1/2}(2\pi e)^{-1/2}-M])$ (which is, however, exponential in $d$). Furthermore, \[\begin{aligned} C_{\text{mom}} & = C_{\text{Lip}} \mathbb{E}[\|\exp(M\mathbf{1}+V)\|^2]^{1/2} + C_{\text{Lip}}\|\mathbf{1}\|+|\varphi(\mathbf{1})| \\ & \leq C_{\text{Lip}} e^M \left( \sum_{i=1}^d \mathbb{E}[\exp(2V_i)]\right)^{1/2} + d^{1/2} C_{\text{Lip}}+c(1+d^{1/2}) \end{aligned} \] is finite under exponential moment hypotheses on $V$. Therefore, from \eqref{eq:L2error3} and Markov's inequality we obtain \begin{equation} \label{eq:auxEq5} \begin{aligned} \mathbb{E}\left[\sup_{x \in B_M(0)} |H^{A,B}_W(x) - H(x) | \right] & \leq \frac{ \tilde{c} C_{\text{app}} (\nu+d)^{k+3} + C_{\text{mom}} \mathbb{E}[\exp(\alpha\|V\|)]^{1/2} }{N^{\frac{1}{2}-\frac{1}{\alpha}}}. \end{aligned} \end{equation} \end{remark} \subsection{Bounds for non-degenerate L\'evy models} \label{subsec:ApproxLevy} In this section we apply Theorem~\ref{thm:ApproxError} to prove that random neural networks are capable of overcoming the curse of dimensionality in the numerical approximation of solutions to partial (integro-)differential equations (also referred to as (non-local) PDEs) associated to exponential L\'evy models with a non-degenerate Gaussian component. This includes the Black-Scholes PDE as a special case. We refer to \cite{Cont2004}, \cite{EberKall19} for background on exponential L\'evy models and their applications in financial modelling and, e.g., to \cite{Sato1999} for an extensive treatment of L\'evy processes. For each $d \in \mathbb{N}$ we consider a payoff function $\varphi_d \colon (0,\infty)^d \to \mathbb{R}$ and a L\'evy process $L^d$ with characteristic triplet $(\Sigma^d,\gamma^d,\nu^d_\mathrm{L})$ satisfying $ \nu^d_\mathrm{L}(\{y \in \mathbb{R}^d \, | \, \|y\|>R\}) = 0$ for some $R>1$. We define the shifted drift vector $\tilde{\gamma}^d$ given by $\tilde{\gamma}^d_i = \gamma_i^d + \frac{1}{2} \Sigma_{i,i}^d + \int_{\mathbb{R}^d} (e^{y_i}-1- y_i \mathbbm{1}_{\{\|y\|\leq 1\}}) \nu^d_\mathrm{L}(d y)$ for $i=1,\ldots,d$. We now consider the partial (integro-)differential equation \begin{equation} \label{eq:PIDEs} \begin{array}{rl} \partial_t u_d(t,s) & = \frac{1}{2} \sum_{k,l=1}^d s_k s_l \Sigma^d_{k,l} \partial_{s_k} \partial_{s_l} u_d(t,s) + \sum_{i=1}^d s_i \tilde{\gamma}^d_i \partial_{s_i} u_d(t,s) \\ & \quad + \int_{\mathbb{R}^d} \left[u_d(t,s e^y)-u_d(t,s)-\sum_{i=1}^d (e^{y_i}-1) s_i \partial_{s_i} u_d(t,s) \right] \nu^d_\mathrm{L}(d y) , \\ u_d(0,s) &= \varphi_d(s) \end{array} \end{equation} for $s \in (0,\infty)^d, t > 0$, where we write $s\exp(x) =(s_1\exp(x_1),\ldots,s_d\exp(x_d))$ for $s,x \in \mathbb{R}^d$. The (non-local) PDE \eqref{eq:PIDEs} is the Kolmogorov PDE for the exponential L\'evy model associated to $L^d$. By \citet[Theorem~25.17]{Sato1999} the exponential L\'evy process $(\exp({L^d_t}))_{t \geq 0}$ is a martingale if (and only if) $\tilde{\gamma}^d = 0$. In this case, $u_d(T,s)$ is the price at time $0$ of an option with payoff $\varphi_d$ at maturity $T$ when price of the underlying at time $0$ is $s$. Furthermore, if the jump-measure vanishes ($\nu^d_\mathrm{L}=0$), then \eqref{eq:PIDEs} is the Black-Scholes PDE. We now show that $u_d(T,\cdot)$ can be approximated by random neural networks without the curse of dimensionality. To achieve this, the weights of the random neural networks are generated as follows: let $\nu>1$ and for each $d \in \mathbb{N}$ let $A^d_1,A_2^d,\ldots$ by i.i.d.\ $t_{\nu}(0,\mathbbm{1}_d)$-distributed $\mathbb{R}^d$-valued random vectors independent of the i.i.d.\ random variables $B_1,B_2,\ldots$ which have a strictly positive Lebesgue-density $\pi_{\text{b}}$ of at most polynomial decay (see \eqref{eq:polyTails1D}). For $N \in \mathbb{N}$ we write $A^{d,N}=(A^d_1,\ldots,A^d_N)$ and $B^N = (B_1,\ldots,B_N)$. Theorem~\ref{thm:LevyApprox} complements the results in \cite{HornungJentzen2018}, \cite{GS20_925}. \begin{theorem}\label{thm:LevyApprox} Let $p\geq0$, $c,C, M,T>0$. For each $d \in \mathbb{N}$ assume the payoff function satisfies $\varphi_d \circ \exp \in L^1(\mathbb{R}^d)$ and $\|\varphi_d \circ \exp \|_{L^1(\mathbb{R}^d)} \leq c d^p$, the characteristic triplet $(\Sigma^d,\gamma^d,\nu^d_\mathrm{L})$ of the L\'evy process $L^d$ satisfies for all $\xi \in \mathbb{R}^d$ \begin{equation} \label{eq:Ccond} \frac{1}{2} \xi \cdot \Sigma^d \xi \geq C \| \xi\|^2, \end{equation} assume $C T > \frac{1}{2^{3/2} \pi}$ and suppose $u_d \in C^{1,2}((0,T] \times (0,\infty)^d) \cap C([0,T]\times (0,\infty)^d)$ is an at most polynomially growing solution to the PDE \eqref{eq:PIDEs}. Then there exist constants $C_0,\mathfrak{p}>0$ such that for any $d,N \in \mathbb{N}$ there exists an $\mathbb{R}^N$-valued, $\sigma(A^{d,N},B^{N})$-measurable random vector $W^{d,N}$ such that the random neural network \begin{equation} \bar{H}_{d,N}(x) := H^{A^{d,N},B^N}_{W^{d,N}}(x)= \sum_{i=1}^N W_{i}^{d,N} \varrho(A_i^d \cdot x + B_i), \quad x \in \mathbb{R}^d, \end{equation} satisfies the approximation bound \begin{equation} \label{eq:L2error4} \mathbb{E}\left[ \sup_{x \in [-M,M]^d} |\bar{H}_{d,N}(x) - u_d(T,\exp(x))| \right] \leq \frac{C_0 d^{\mathfrak{p}}}{\sqrt{N}}. \end{equation} \end{theorem} \begin{proof} Let $d, N \in \mathbb{N}$, $\Phi(x) = \varphi_d(\exp(x))$ and $H(x)= u_d(T,\exp(x))$ for $x \in \mathbb{R}^d$. Then Proposition~\ref{prop:FeynmanKac} below shows that $H(x)= \mathbb{E}[\varphi_d(\exp(x+{L^d_T}))]= \mathbb{E}[\Phi(x+L_T^d)]$. Furthermore, by the L\'evy-Khintchine representation (see for instance \citet[Theorem~8.1]{Sato1999} or \citet[Theorem~1.2.14 and Theorem~1.3.3]{Applebaum2009}) we have $\mathbb{E}[e^{i \xi \cdot L^d_T}] = \exp(T \eta(\xi))$ with \begin{equation}\label{eq:LevySymbol} \eta(\xi) = i \xi \cdot \gamma^d -\frac{1}{2} \xi \cdot \Sigma^d \xi + \int_{\mathbb{R}^d \setminus \{0\}} \left[e^{i \xi \cdot y}-1- i \xi \cdot y \mathbbm{1}_{\{\|y\|\leq 1\}} \right] \nu^d_\mathrm{L}(d y)\;, \quad \xi \in \mathbb{R}^d. \end{equation} In particular, \[ \mathrm{Re} \, \eta(\xi) = -\frac{1}{2} \xi \cdot \Sigma^d \xi + \int_{\mathbb{R}^d \setminus \{0\}} \left[\cos(\xi \cdot y)-1 \right] \nu^d_\mathrm{L}(d y) \leq -\frac{1}{2} \xi \cdot \Sigma^d \xi, \] since the integrability property $\int_{\mathbb{R}^d} (\|y\|^2 \wedge 1)\nu^d_\mathrm{L}(dy) < \infty$ guarantees that $y \mapsto \cos(\xi \cdot y)-1$ and $y \mapsto \sin(\xi \cdot y)-\xi \cdot y \mathbbm{1}_{\{\|y\|\leq 1\}}$ are indeed $\nu^d_\mathrm{L}$-integrable for any $\xi \in \mathbb{R}^d$. This and \eqref{eq:Ccond} show that for all $\xi \in \mathbb{R}^d$ \begin{equation} \label{eq:auxEq8} |\mathbb{E}[e^{i \xi \cdot L^d_T}]| = e^{T \mathrm{Re} \, \eta(\xi)} \leq \exp(-CT\|\xi\|^2). \end{equation} Theorem~\ref{thm:ApproxError} hence shows that there exist $C_{\text{app}}>0$, $k \in \mathbb{N}$ and an $\mathbb{R}^N$-valued, $\sigma(A^{d,N},B^{N})$-measurable random vector $W^{d,N}$ such that the random neural network $ \bar{H}_{d,N} = H^{A^{d,N},B^N}_{W^{d,N}}$ satisfies \begin{equation} \label{eq:L2errorAux} \mathbb{E}\left[\sup_{x \in [-M,M]^d} |\bar{H}_{d,N}(x) - H(x) | \right] \leq \frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}}. \end{equation} Thus, we obtain \[ \begin{aligned} \mathbb{E}\left[\sup_{x \in [-M,M]^d} |\bar{H}_{d,N}(x) - u_d(T,\exp(x))|\right] & \leq \frac{C_{\text{app}} c d^p (\nu+1)^{k+3} d^{k+3}}{\sqrt{N}} \\ & = \frac{C_0 d^{\mathfrak{p}}}{\sqrt{N}} \end{aligned} \] with $C_0 = (\nu+1)^{k+3} C_{\text{app}} c$ and $\mathfrak{p} = p + k+3$. This proves \eqref{eq:L2error4} and the statement, since $C_{\text{app}}$ in Theorem~\ref{thm:ApproxError} does not depend on $d$ or $N$ and hence the constants $C_0, \mathfrak{p}$ are the same for all $d,N \in \mathbb{N}$. \end{proof} \begin{remark}\label{rmk:stochastic} Theorem~\ref{thm:LevyApprox} also holds if we directly assume $u_d(T,\exp(x)) = \mathbb{E}[\varphi_d(\exp(x+L_T^d))]$ instead of considering the PDE \eqref{eq:PIDEs}. For instance in the context of mathematical finance many quantities of interest (such as option prices or ``greeks'') are defined in terms of such expectations. In particular, in this situation the hypothesis $\varphi_d \in C((0,\infty)^d,\mathbb{R})$ is not required (in Theorem~\ref{thm:LevyApprox} this hypothesis is implicit in the assumption $u_d \in C^{1,2}((0,T] \times (0,\infty)^d) \cap C([0,T]\times (0,\infty)^d)$). The integrability hypothesis $\varphi_d \circ \exp \in L^1(\mathbb{R}^d)$ is more restrictive, but currently it can not be avoided in the proof of Theorem~\ref{thm:ApproxError}. The hypothesis is satisfied e.g.\ for butterfly or binary options. More general payoffs can be incorporated by truncation (which is often possible without affecting the price significantly) or potentially by employing Fourier representations as in \cite{Carr1999OptionVU} instead of \eqref{eq:HFourier}. \end{remark} \begin{remark} The assumption $ \nu^d_\mathrm{L}(\{y \in \mathbb{R}^d \, | \, \|y\|>R\}) = 0$ for some $R>1$ is only required to obtain a ``Feynman-Kac representation'' from the results of \cite{BBP1997} (see Proposition~\ref{prop:FeynmanKac} below). This assumption on $\nu^d_\mathrm{L} $ can be weakened to $\int_{\{\|y\|>1\}} e^{y_i} \nu^d_\mathrm{L}(dy) < \infty$ for $i=1,\ldots,d$ for instance in the situation of Remark~\ref{rmk:stochastic} when we directly assume a stochastic representation for $u_d$. Alternatively, instead of assuming $ \nu^d_\mathrm{L}(\{y \in \mathbb{R}^d \, | \, \|y\|>R\}) = 0$ for some $R>1$ we could impose that $ \nu^d_\mathrm{L}$ is a finite measure and \eqref{eq:Ccond} holds. Then we may apply \citet[Proposition~5.3]{Pham1998} instead of \cite{BBP1997} in the proof of Proposition~\ref{prop:FeynmanKac} below and also obtain the representation $u_d(t,s)=\mathbb{E}[\varphi_d(s\exp({L^d_t}))]$. \end{remark} The proof of Theorem~\ref{thm:LevyApprox} employs the ``Feynman-Kac representation'' from Proposition~\ref{prop:FeynmanKac} below. Proposition~\ref{prop:FeynmanKac} is essentially a consequence of the results from \cite{BBP1997}. For the readers' convenience we provide a proof of Proposition~\ref{prop:FeynmanKac} and make explicit how it can be obtained from \cite{BBP1997}. Related results and further references can be found, for instance, in \citet{Pham1998}, \cite{ContVolt2005}, \citet[Proposition~3.3]{ContVolt2006}, \cite{GlauClassLevy2016}. \begin{proposition}\label{prop:FeynmanKac} Suppose $u_d \in C^{1,2}((0,T] \times (0,\infty)^d) \cap C([0,T]\times (0,\infty)^d)$ is an at most polynomially growing solution to the PDE \eqref{eq:PIDEs} and $\varphi_d$ is bounded. Then for all $(t,s) \in [0,T]\times (0,\infty)^d$ it holds that $u_d(t,s)=\mathbb{E}[\varphi_d(s\exp({L^d_t}))]$. \end{proposition} \begin{proof} Let $\Phi_d(x) = \varphi_d(\exp(x))$ and $v_d(t,x)=u_d(T-t,\exp(x))$. Firstly, the assumptions on $u_d$ imply that $v_d \in C^{1,2}([0,T) \times \mathbb{R}^d) \cap C([0,T]\times \mathbb{R}^d)$ and a straightforward calculation shows that $v_d$ satisfies the (non-local) PDE \begin{equation} \label{eq:PIDEx} \begin{array}{rl} - \partial_t v_d(t,x) & = \frac{1}{2} \sum_{k,l=1}^d \Sigma^d_{k,l} \partial_{x_k} \partial_{x_l} v_d(t,x) + \sum_{i=1}^d \left(\gamma^d_i+\int_{\mathbb{R}^d} y_i \mathbbm{1}_{\{\|y\|> 1 \}} \nu^d_\mathrm{L}(d y) \right) \partial_{x_i} v_d(t,x) \\ & \quad + \int_{\mathbb{R}^d} \left[v_d(t,x+y)-v_d(t,x)-\sum_{i=1}^d y_i \partial_{x_i} v_d(t,x) \right] \nu^d_\mathrm{L}(d y) , \\ v_d(T,x) &= \Phi_d(x) \end{array} \end{equation} for $x \in \mathbb{R}^d, t \in [0,T)$. Set $\hat{\gamma}^d = (\gamma^d+\int_{\mathbb{R}^d} y \mathbbm{1}_{\{\|y\|> 1 \}} \nu^d_\mathrm{L}(d y))$ and for $\phi \in C^2(\mathbb{R}^d)$ write \begin{equation} \label{eq:auxEq7} \begin{aligned} \mathcal{A}\phi(x) & = \frac{1}{2}\mathrm{Trace}(\Sigma^d D^2_x \phi(x)) + [D_x \phi(x)]\hat{\gamma}^d \\ \mathcal{K}\phi(x) & = \int_{\mathbb{R}^d} (\phi(x+y)-\phi(x)- [D_x\phi(x)] y) \nu^d_\mathrm{L}(d y). \end{aligned} \end{equation} Now if $\phi \in C^2([0,T]\times \mathbb{R}^d)$ and $(t_0,x_0) \in [0,T) \times \mathbb{R}^d$ is a global maximum point of $v_d-\phi$, then $D_{(t,x)}(v_d-\phi)(t_0,x_0) = 0$ and $D^2_{x}(v_d-\phi)(t_0,x_0) \leq 0$. Thus, \eqref{eq:PIDEx} implies \begin{equation} \label{eq:auxEq6} \begin{aligned} - & \partial_t \phi(t_0,x_0) - \mathcal{A}\phi(t_0,x_0) - \mathcal{K}\phi(t_0,x_0) \\ & = \mathcal{A}(v_d-\phi)(t_0,x_0) + \mathcal{K}(v_d-\phi)(t_0,x_0) \\ & = \frac{1}{2}\mathrm{Trace}(\sqrt{\Sigma^d} D^2_{x}(v_d-\phi)(t_0,x_0)\sqrt{\Sigma^d}) + \int_{\mathbb{R}^d} (v_d-\phi)(t_0,x_0+y)-(v_d-\phi)(t_0,x_0) \nu^d_\mathrm{L}(d y) \\ & \leq 0. \end{aligned} \end{equation} This and \citet[Lemma~3.3]{BBP1997} show that $v_d$ is a viscosity subsolution of \eqref{eq:PIDEx} in the sense of \cite{BBP1997}. Similarly, one argues that $v_d$ is also a viscosity supersolution to \eqref{eq:PIDEx}. \citet[Theorem~3.5]{BBP1997} hence shows that for all $(t,x)\in [0,T]\times \mathbb{R}^d$ we have $v_d(t,x) = \mathbb{E}[\Phi_d(X_T^{t,x})]$ (see also the proof of \citet[Corollary~5.4]{GS21}) where $(X^{t,x}_r)_{r \geq t}$ is the unique solution to $X^{t,x}_t = x$, \[ \begin{aligned} d X^{t,x}_r & = \hat{\gamma}^d dr + \sqrt{\Sigma^d} W^d_r + \int_{\mathbb{R}^d \setminus \{0\}} z \tilde{N}^d(dt,dz) \\ & = \gamma^d dr + \sqrt{\Sigma^d} W^d_r + \int_{\mathbb{R}^d \setminus \{0\}} z \mathbbm{1}_{\{\|z\|\leq 1\}} \tilde{N}^d(dr,dz) + \int_{\mathbb{R}^d \setminus \{0\}} z \mathbbm{1}_{\{\|z\|> 1\}} N^d(dr,dz) \end{aligned} \] where $N^d$ is a Poisson random measure on $\mathbb{R}_+ \times (\mathbb{R}^d \setminus \{0\})$ with intensity $\nu^d_\mathrm{L}$, $W^d$ is an independent $d$-dimensional standard Brownian motion and $\tilde{N}^d(dt,dz) = N^d(dt,dz) - dt \nu^d_\mathrm{L}(dz)$. Note that the assumption $ \nu^d_\mathrm{L}(\{y \in \mathbb{R}^d \, | \, \|y\|>R\}) = 0$ for some $R>1$ guarantees that the function $\beta$ in \citet[Theorem~3.5]{BBP1997} can be chosen so that it satisfies the required boundedness hypothesis. Hence, by the L\'evy-It\^o-decomposition (see for instance \citet[Theorem~19.2]{Sato1999} or \citet[Theorem~2.4.16]{Applebaum2009}) we obtain that $X^{t,x}_T$ has the same distribution as $x+L_{T-t}^d$. Thus, we have proved the representation $ v_d(t,x) = \mathbb{E}[\Phi_d(x+L_{T-t}^d)] $ and therefore for all $x \in \mathbb{R}^d$, with $s = \exp(x)$, \[ u_d(t,s)=v_d(T-t,x)=\mathbb{E}[\varphi_d(\exp(x+L_t^d))]=\mathbb{E}[\varphi_d(s\exp({L^d_t}))]. \] \end{proof} \section{Learning by random neural networks} \label{sec:Learning} In this section we use random neural networks $H^{A,B}_W$ to learn functions of the type considered in Section~\ref{subsec:ApproxGeneral}. In Section~\ref{subsec:learningProblem} we formulate the considered learning problem. In Sections~\ref{subsec:Regression}, \ref{subsec:RidgeRegression}, \ref{subsec:SGD} we then provide bounds on the prediction error that arises when $W$ is learnt by means of regression, constrained regression and stochastic gradient descent, respectively. In Sections~\ref{sec:Options} we will then apply these results to obtain prediction error bounds for random neural networks applied to learning option prices in certain non-degenerate models. \subsection{Formulation of the learning problem} \label{subsec:learningProblem} Let $n \in \mathbb{N}$ and suppose that we are given i.i.d.\ $\mathbb{R}^d\times \mathbb{R}$-valued random variables $(X_1,Y_1),\ldots,$ $(X_n,Y_n)$ (the data) which are independent of $(A,B)$. Let $H \colon \mathbb{R}^d \to \mathbb{R}$ be the target function (which we will assume to be of the form specified in Section~\ref{subsec:ApproxGeneral}) and suppose that \begin{equation}\label{eq:regressionFunction} H(x) = \mathbb{E}[Y_1 | X_1=x], \end{equation} for $(\P \circ (X_1)^{-1})$-a.e.\ $x \in \mathbb{R}^d$, that is, $H$ is the regression function. This encompasses two important situations: \begin{itemize} \item \textit{Learning $H$ from noisy observations}: We observe the unknown function $H$ (the solution to a PDE or market prices of options) at $n$ data points up to some additive noise. Thus, in this situation we suppose $Y_i = H(X_i) + \varepsilon_i$, $i=1,\ldots,n$, for $\varepsilon_1,\ldots,\varepsilon_n$ i.i.d.\ random variables which are independent of $(X_1,\ldots,X_n)$ and satisfy $\mathbb{E}[\varepsilon_1] = 0$. \item \textit{Solving PDEs by learning}: Solving linear Kolmogorov PDEs with affine coefficients has been formulated as a learning problem in \cite{BernerGrohsJentzen2018}. The setting considered here also covers this type of learning problem. \end{itemize} The target function $H$ is considered unknown and is to be learnt from the data $D_n=((X_1,Y_1),\ldots,(X_n,Y_n))$ using random neural networks. To do this, we recall that $H(X_1) = \mathbb{E}[Y_1|X_1]$ minimizes \begin{equation}\label{eq:risk} \mathcal{R}(f) = \mathbb{E}[(f(X_1)-Y_1)^2] \end{equation} among all measurable functions $f \colon \mathbb{R}^d \to \mathbb{R}$. Thus, to learn $H(x) =\mathbb{E}[Y_1|X_1=x]$ from the data one aims at finding a minimizer of \begin{equation}\label{eq:Empiricalrisk} \mathcal{R}_n(f) = \frac{1}{n}\sum_{i=1}^n (f(X_i)-Y_i)^2. \end{equation} $\mathcal{R}_n(f)$ is the empirical version of \eqref{eq:risk}. In the situation considered here we know from Section~\ref{sec:Approx} that $H$ can be approximated well by random neural networks and so we learn $H$ by minimizing $\mathcal{R}_n(\cdot)$ only over this class of functions, i.e.\ by minimizing $\mathcal{R}_n(H^{A,B}_W)$ over neural networks $H^{A,B}_W$ with random weights $(A,B)$ and trainable $W$ (see Section~\ref{sec:RandomNN}). This leads to the optimization problem \begin{equation} \label{eq:ERM} \widehat{W} = \arg \min_{W \in \mathcal{W}} \left\lbrace \frac{1}{n} \sum_{i=1}^n (H^{A,B}_W(X_i) - Y_i)^2 \right\rbrace \end{equation} for a suitable set $\mathcal{W}$ of $\mathbb{R}^N$-valued, $\sigma(A,B,D_n)$-measurable random vectors. The measurability requirement incorporates the fact that $A,B$ are generated randomly and then fixed and hence the trainable weights may depend on $A,B$. Having solved \eqref{eq:ERM}, the learning algorithm then returns the (random) function \[ H^{A,B}_{\widehat{W}}(x) = \sum_{i=1}^N \widehat{W}_i \varrho(A_i \cdot x + B_i), \quad x \in \mathbb{R}^d \] as our approximation for $H$. To evaluate the learning performance of the random features regression algorithm we need to bound the (squared) learning error (or prediction error) \begin{equation} \label{eq:testError} \mathbb{E}[|H(\bar{X}) - H^{A,B}_{\widehat{W}}(\bar{X}) |^2], \end{equation} where $(\bar{X},\bar{Y})$ has the same distribution as $(X_1,Y_1)$ and is independent of $(A,B,D_n)$. \subsection{Regression} \label{subsec:Regression} Consider first the case $\mathcal{W} = \{W \colon \Omega \to \mathbb{R}^N \, | \, W \text{ is } \sigma(A,B,D_n)\text{-measurable}\}$. In this case computing \eqref{eq:ERM} amounts to a simple least squares optimization. Hence $\widehat{W}$ can be calculated explicitly by solving \begin{equation}\label{eq:OLSSol} (\mathbf{X}^\top \mathbf{X}) \widehat{W} = \mathbf{X}^\top \mathbf{Y} \end{equation} where $\mathbf{X}$ is the $n \times N$-random matrix with entries $\mathbf{X}_{i j} = \varrho(A_j \cdot X_i + B_j)$ and $\mathbf{Y}$ is the $n$-dimensional random vector with $\mathbf{Y}_i = Y_i$ for $i=1,\ldots,n$, $j=1,\ldots,N$. Thus, there is no additional ``optimization error'' component in this case and we can directly bound the prediction error \eqref{eq:testError} by combining the approximation error estimates from Section~\ref{sec:Approx} with a result from \citet{DistributionFreeTheory}. The trained neural network $ H^{A,B}_{\widehat{W}}$ will be capped at a level $L>0$ by applying the truncation $T_L \colon \mathbb{R} \to \mathbb{R}$, $T_L(u) = \max(\min(u,L),-L)$. \begin{theorem}\label{thm:TrainingErrRegression} Let $C>\frac{1}{2^{3/2} \pi}$ and let $\nu>1$. Suppose $A_1 \sim t_{\nu}(0,\mathbbm{1}_d)$ and $B_1$ has density $\pi_{\text{b}}$ satisfying \eqref{eq:polyTails1D}. Suppose $H \colon \mathbb{R}^d \to \mathbb{R}$ is of the form $H(x) = \mathbb{E}[\Phi(x+V)]$ with $\Phi \in L^1(\mathbb{R}^d)$ and $V$ satisfying \eqref{eq:charFctAss}. Assume that $\|X_1\|_\infty \leq M$, $\P$-a.s. Let $L>0$ and assume $\sigma^2=\sup_{x \in \mathbb{R}^d} \mathbb{E}[(Y_1-H(X_1))^2|X_1=x] < \infty$ and $|H(x)| \leq L$ for all $x \in \mathbb{R}^d$. Then there exist $k \in \mathbb{N}$ and $\tilde{C}_{\text{app}}>0$ such that \begin{equation} \label{eq:fullError} \begin{aligned} & \mathbb{E}[|H(\bar{X}) - T_L(H^{A,B}_{\widehat{W}}(\bar{X})) |^2]^{1/2} \\ & \leq \tilde{C}_{\text{app}} \max(\sigma,L) \frac{(\log(n)+1)^{1/2}\sqrt{N}}{\sqrt{n}} + \frac{\tilde{C}_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}}. \end{aligned} \end{equation} The constant $k$ only depends on $\pi_{\text{b}}$ and the constant $\tilde{C}_{\text{app}}$ depends on $\nu, \pi_{\text{b}}, C, M$, but it does not depend on $d$, $n$ or $N$. \end{theorem} \begin{remark} Theorem~\ref{thm:TrainingErrRegression} bounds the square-root of the prediction error by ${\mathcal O}(\frac{\log(n)^{1/2}\sqrt{N}}{\sqrt{n}} + \frac{1}{\sqrt{N}})$. This matches, up to constants, the error bound obtained in the seminal work \cite{Barron1994ApproximationAE} for general ``Barron functions''. In \cite{Barron1994ApproximationAE} all parameters of the network are trainable and the neural network estimator is defined via empirical risk minimization over a constrained parameter set. However, the optimization error, which arises when the neural network estimator is calculated based e.g.\ on the stochastic gradient descent algorithm, is not addressed in \cite{Barron1994ApproximationAE}. In contrast, in our situation the class of considered functions is smaller, but the neural network estimator can be directly calculated by solving the linear system \eqref{eq:OLSSol}. Hence, the bound in Theorem~\ref{thm:TrainingErrRegression} captures \textit{the full training error}. \end{remark} \begin{proof} Firstly, for fixed $a \in (\mathbb{R}^d)^N$, $b \in \mathbb{R}^N$ we consider the function class $\mathcal{F}_{a,b} = \{H_W^{a,b} \, | \, W \in \mathbb{R}^N \}$, $\mathcal{F}_{a,b}(D_n)=\{H_W^{a,b} \, | \, W \colon \Omega \to \mathbb{R}^N \text{ is $\sigma(D_n)$-measurable} \}$ (in \cite{DistributionFreeTheory} the same symbol is used for these two sets) and let $\hat{f}_{a,b} = \arg \min_{f \in \mathcal{F}_{a,b}(D_n)} \mathcal{R}_n(f)$. Then $\mathcal{F}_{a,b}$ is an $N$-dimensional vector space and hence \citet[Theorem~11.3]{DistributionFreeTheory} implies that \begin{equation}\label{eq:auxEq9} \begin{aligned} \mathbb{E}& \left[\int_{\mathbb{R}^d} |T_L(\hat{f}_{a,b}(x)) - H(x)|^2 \mu_X(d x)\right] \\ & \leq c \max(\sigma^2,L^2) \frac{(\log(n)+1)N}{n} + 8 \inf_{f \in \mathcal{F}_{a,b}} \int_{\mathbb{R}^d}|f(x)-H(x)|^2 \mu_X(dx), \end{aligned} \end{equation} where $\mu_X$ is the law of $X_1$ under $\P$ and $c = 8+2304[\log(9)+4\log(12e)+1]$. For any $a \in (\mathbb{R}^d)^N$, $b \in \mathbb{R}^N$ the minimization problem for $\hat{f}_{a,b}$ can be solved explicitly and we obtain $\hat{f}_{a,b} = H^{a,b}_{\hat{w}_{a,b}}$, where $\hat{w}_{a,b}$ is a solution to the linear system \eqref{eq:OLSSol} with $A,B$ fixed to $a,b$. A solution always exists (see for instance \citet[Chapter~4.8.1]{Stoer2002}) and, e.g.\ by choosing the solution given in terms of the pseudo-inverse matrix as $\widehat{W} = (\mathbf{X}^\top \mathbf{X})^{\dagger} \mathbf{X}^\top \mathbf{Y}$, it is possible to write $ \widehat{W} = F(A,B,D_n)$ for a measurable function $F \colon (\mathbb{R}^d)^N\times \mathbb{R}^N\times (\mathbb{R}^d\times\mathbb{R})^n \to \mathbb{R}^N$ and select $\hat{w}_{a,b}$ in such a way that $\hat{w}_{a,b} = F(a,b,D_n)$. Using independence we thus obtain from \eqref{eq:auxEq9} \begin{equation}\label{eq:auxEq10} \begin{aligned} \mathbb{E}& \left[|T_L(H^{A,B}_{\widehat{W}}(\bar{X})) - H(\bar{X})|^2 | A,B \right] = \left. \mathbb{E} [|T_L(H^{a,b}_{\hat{w}_{a,b}}(\bar{X})) - H(\bar{X})|^2 ]\right\rvert_{(a,b)=(A,B)} \\ & \leq c \max(\sigma^2,L^2) \frac{(\log(n)+1)N}{n} + 8 \left. \left(\inf_{W \in \mathbb{R}^N} \mathbb{E}[|H_W^{a,b}(\bar{X})-H(\bar{X})|^2 ]\right)\right\rvert_{(a,b)=(A,B)} \\ & \leq c \max(\sigma^2,L^2) \frac{(\log(n)+1)N}{n} + 8 \mathbb{E}[|H_{W^*}^{A,B}(\bar{X})-H(\bar{X})|^2 |A,B ], \end{aligned} \end{equation} where $W^*$ denotes the random vector from Theorem~\ref{thm:ApproxError}. We may therefore take expectations in \eqref{eq:auxEq10}, use $\|\bar{X}\|_\infty \leq M$ and insert the bound from Theorem~\ref{thm:ApproxError} (c.f.\ also Remark~\ref{rmk:L2error}) to deduce \eqref{eq:fullError} with $\tilde{C}_{\text{app}} = \max(\sqrt{c},\sqrt{8} C_{\text{app}})$. \end{proof} \subsection{Constrained regression} \label{subsec:RidgeRegression} In the next result we consider a constrained regression estimator, i.e., $\widehat{W}$ in \eqref{eq:ERM} is calculated with a smaller set of potential weights $\mathcal{W}$. This leads to a different bound than in Theorem~\ref{thm:TrainingErrRegression}, but for instance for $N=\sqrt{n}$ the same rate is achieved. Set $\mathcal{W}_\lambda = \{W \colon \Omega \to \mathbb{R}^N \, | \, W \text{ is } \sigma(A,B,D_n)\text{-measurable}, \|W\| \leq \lambda \text{ $\P$-a.s.} \}$. Computing \eqref{eq:ERM} now corresponds to a constrained regression problem \begin{equation} \label{eq:ConstrainedRegression} \widehat{W}_\lambda = \arg \min_{W \in \mathcal{W}_\lambda} \left\lbrace \frac{1}{n} \sum_{i=1}^n (H^{A,B}_W(X_i) - Y_i)^2 \right\rbrace. \end{equation} The solution to \eqref{eq:ConstrainedRegression} is given explicitly as follows: $\widehat{W}_\lambda$ coincides with the solution $\widehat{W}$ to the unconstrained problem \eqref{eq:OLSSol} with minimal norm in case $\widehat{W}$ satisfies $\|\widehat{W}\| \leq \lambda$. Otherwise $\widehat{W}_\lambda$ is given explicitly as \begin{equation}\label{eq:ConstrainedSol} \widehat{W}_\lambda = (\mathbf{X}^\top \mathbf{X} + \mathbbm{1} \Lambda)^{-1} \mathbf{X}^\top \mathbf{Y} \end{equation} with $\Lambda$ a non-negative $\sigma(A,B,D_n)$-measurable random variable\footnote{This means that once the data and the random weights have been sampled/observed (i.e.\ conditionally on these) $\Lambda$ is just a constant.} such that $\|\widehat{W}_\lambda\| =\lambda$. The two cases can be summarized by setting $\Lambda = 0$ in the first case and interpreting the inverse in \eqref{eq:ConstrainedSol} as a pseudo-inverse, then $\widehat{W}_\lambda$ is given by \eqref{eq:ConstrainedSol} in both cases. We now provide a bound on the prediction error for random neural networks with parameters learned according to \eqref{eq:ConstrainedRegression}. \begin{theorem}\label{thm:TrainingErrConstrRegression} Let $C>\frac{1}{2^{3/2} \pi}$ and let $\nu>2$. Suppose $A_1 \sim t_{\nu}(0,\mathbbm{1}_d)$ and $B_1$ has density $\pi_{\text{b}}$ satisfying \eqref{eq:polyTails1D}. Suppose $H \colon \mathbb{R}^d \to \mathbb{R}$ is of the form $H(x) = \mathbb{E}[\Phi(x+V)]$ with $\Phi \in L^1(\mathbb{R}^d)$ and $V$ satisfying \eqref{eq:charFctAss}. Assume that $\|X_1\|_{\infty} \leq M$, $\P$-a.s.\ and $\mathbb{E}[|Y_1|^4]< \infty$. Let $k \in \mathbb{N}$ and $C_{\text{app}},C_{\text{wgt}}>0$ be as in Theorem~\ref{thm:ApproxError}. Let $\lambda >0$ satisfy $\frac{C_{\text{wgt}}\|\Phi\|_{L^1(\mathbb{R}^d)}(\nu+d)^{2k+\frac{1}{2}}}{\sqrt{N}} \leq \lambda \leq \frac{C_{\text{lam}} d^{p} }{\sqrt{N}}$ for some $p\geq 0$, $C_{\text{lam}}>0$ not depending on $n,N,d$. Then there exists $C_{\text{est}}>0$ such that \begin{equation} \label{eq:fullError2} \begin{aligned} & \mathbb{E}[|H(\bar{X}) - H^{A,B}_{\widehat{W}_\lambda}(\bar{X}) |^2]^{1/2} \leq \frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}} + \frac{C_{\text{est}} d^{p+1} }{n^{\frac{1}{4}}}. \end{aligned} \end{equation} The constant $C_{\text{est}}$ depends on $\nu, \pi_{\text{b}}, C_{\text{lam}}, M, \mathbb{E}[Y_1^4]$, but it does not depend on $d$, $n$ or $N$. \end{theorem} \begin{remark} Theorem~\ref{thm:TrainingErrConstrRegression} shows that the prediction error is of order ${\mathcal O}(\frac{1}{N}+\frac{1}{\sqrt{n}})$. Thus, the error bound decays more quickly than the bound ${\mathcal O}(\frac{1}{\sqrt{N}}+\frac{1}{\sqrt{n}})$ that was obtained in the seminal work \cite{RahimiRecht2008}, where high-probability bounds were obtained for random neural networks trained by constrained regression in a classification setting ($\P(Y_i \in \{1,-1\})=1$). The reason for this faster rate is that we use the mean-square loss here. This allows to write $|\mathcal{R}(H)-\mathcal{R}(\tilde{H})| = \mathbb{E}[|H(\bar{X}) - \tilde{H}(\bar{X}) |^2]$ due to \eqref{eq:regressionFunction}. For $L$-Lipschitz loss functions the bound $\mathcal{R}(H)-\mathcal{R}(\tilde{H}) \leq L \mathbb{E}[|H(\bar{X}) - H^{A,B}_{\widehat{W}_\lambda}(\bar{X}) |^2]^{1/2}$ can be deduced (see \citet[Lemma~2]{RahimiRecht2008}), which leads to an approximation error of order $1/\sqrt{N}$ instead of $1/N$. Thus, we are concerned here with a slightly different setting, but our proof of the ``estimation error'' (or generalization error) component is based on similar arguments as the proof in \cite{RahimiRecht2008}. \end{remark} \begin{proof} Firstly, \eqref{eq:regressionFunction} and independence imply \begin{equation} \label{eq:auxEq12} \begin{aligned} \mathbb{E}[H(\bar{X}) H^{A,B}_{\widehat{W}_\lambda}(\bar{X}) ] & = \mathbb{E}[\left. \mathbb{E}[\mathbb{E}[\bar{Y}|\bar{X}] H^{a,b}_{w}(\bar{X})] \right\rvert_{(a,b,w)=(A,B,\widehat{W}_\lambda)} ] \\ & = \mathbb{E}[\left. \mathbb{E}[\bar{Y} H^{a,b}_{w}(\bar{X})] \right\rvert_{(a,b,w)=(A,B,\widehat{W}_\lambda)} ] \\ & = \mathbb{E}[\bar{Y} H^{A,B}_{\widehat{W}_\lambda}(\bar{X})] \end{aligned} \end{equation} and analogously $\mathbb{E}[H(\bar{X}) H^{A,B}_{W}(\bar{X})] = \mathbb{E}[\bar{Y} H^{A,B}_{W}(\bar{X})]$ for any $W \in \mathcal{W}_\lambda$. Thus, we calculate \begin{equation} \label{eq:auxEq11} \begin{aligned} & \mathbb{E}[|H(\bar{X}) - H^{A,B}_{\widehat{W}_\lambda}(\bar{X}) |^2] \\ & \quad = \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W}(\bar{X}) |^2] + \mathbb{E}[|H^{A,B}_{\widehat{W}_\lambda}(\bar{X})- \bar{Y} |^2] - \mathbb{E}[|H^{A,B}_{W}(\bar{X})- \bar{Y} |^2] \\ & \quad = \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W}(\bar{X}) |^2] + \mathbb{E}[\mathcal{R}(H^{A,B}_{\widehat{W}_\lambda}) - \mathcal{R}(H^{A,B}_{W})] \\ & \quad \leq \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W}(\bar{X}) |^2] + \mathbb{E}[\mathcal{R}(H^{A,B}_{\widehat{W}_\lambda}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) + \mathcal{R}_n(H^{A,B}_{W})- \mathcal{R}(H^{A,B}_{W})], \end{aligned} \end{equation} where we used \eqref{eq:ConstrainedRegression} and $W \in \mathcal{W}_\lambda$ in the last step. Consider the first term in the right hand side of \eqref{eq:auxEq11}. Theorem~\ref{thm:ApproxError} (c.f.\ also Remark~\ref{rmk:L2error}) guarantees that there exists an $\mathbb{R}^N$-valued, $\sigma(A,B)$-measurable random vector $W^*$ such that \begin{equation} \label{eq:auxEq13} \begin{aligned} \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W^*}(\bar{X}) |^2]^{1/2} \leq \frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}}, \end{aligned} \end{equation} where we used that $\|\bar{X}\|_{L^\infty(\P)} \leq M$. Furthermore, \eqref{eq:Wbound} shows that $\P$-a.s.\ the weight vector satisfies $\|W^*\| \leq \sqrt{N} \max_{i=1}^N \|W_i^*\|_{L^\infty(\P)} \leq \frac{C_{\text{wgt}}\|\Phi\|_{L^1(\mathbb{R}^d)}(\nu+d)^{2k+\frac{1}{2}}}{\sqrt{N}} \leq \lambda$. Hence, it follows that $W^* \in \mathcal{W}_\lambda$ and so the decomposition \eqref{eq:auxEq11} can be applied with $W=W^*$. For the second term in the right hand side of \eqref{eq:auxEq11} we let $\widehat{W}^{a,b}_\lambda$ denote the solution to \eqref{eq:ConstrainedRegression} for $(A,B)$ fixed to $(a,b)$. The random variable $\Lambda$ can be written as $\Lambda = F(A,B,D_n)$ for a measurable function $F \colon (\mathbb{R}^d)^N\times \mathbb{R}^N\times (\mathbb{R}^d\times\mathbb{R})^n \to [0,\infty)$ (in fact, $F(a,b,d_n) = \inf \{t \geq 0 \, | \, f_{a,b,d_n}(t) \leq \lambda\} $ for the strictly decreasing function $f_{a,b,d_n}(t)= \|(\mathbf{X}_{a,b,d_n}^\top \mathbf{X}_{a,b,d_n} + \mathbbm{1} t)^{-1} \mathbf{X}_{a,b,d_n}^\top \mathbf{Y}_{a,b,d_n}\|$, where $\mathbf{X}_{a,b,d_n}, \mathbf{Y}_{a,b,d_n}$ are $\mathbf{X}, \mathbf{Y}$ with $(A,B,D_n)$ fixed to $(a,b,d_n)$). Then from the formula \eqref{eq:ConstrainedSol} it is clear that $\widehat{W}_\lambda = G(A,B,D_n)$ for a measurable function $G$ and $\widehat{W}^{a,b}_\lambda = G(a,b,D_n)$. Furthermore, we write $(a,b) \mapsto W^{a,b}$ for the measurable function with $W^{A,B} = W$ (which exists, since $W$ is $\sigma(A,B)$-measurable) and $\mathcal{W}_\lambda^0 = \{w \in \mathbb{R}^N \, | \, \|w\| \leq \lambda \}$. Then by independence \begin{equation} \label{eq:auxEq14} \begin{aligned} & \mathbb{E}[\mathcal{R}(H^{A,B}_{\widehat{W}_\lambda}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) + \mathcal{R}_n(H^{A,B}_{W})- \mathcal{R}(H^{A,B}_{W})] \\ & = \mathbb{E}[\left.\mathbb{E}[\mathcal{R}(H^{a,b}_{\widehat{W}^{a,b}_\lambda}) - \mathcal{R}_n(H^{a,b}_{\widehat{W}^{a,b}_\lambda}) + \mathcal{R}_n(H^{a,b}_{W^{a,b}})- \mathcal{R}(H^{a,b}_{W^{a,b}})]\right\rvert_{(a,b)=(A,B)}] \\ & \leq 2 \mathbb{E}\left[ \left.\mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left| \mathcal{R}(H^{a,b}_{w}) - \mathcal{R}_n(H^{a,b}_{w}) \right| \right] \right\rvert_{(a,b)=(A,B)}\right]. \end{aligned} \end{equation} We now fix $(a,b)$, consider for $i=1,\ldots,n$, $w \in \mathcal{W}_\lambda^0$ the random variables $U_{w,i}^{a,b} = (H^{a,b}_{w}(X_i)-Y_i)^2$ and let $\varepsilon_1,\ldots,\varepsilon_n$ denote i.i.d.\ Rademacher random variables independent of all other random variables. Employing symmetrization (see for instance \cite[Lemma~11.4]{Boucheron2013}) we obtain \begin{equation} \label{eq:auxEq15} \begin{aligned} \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left| \mathcal{R}(H^{a,b}_{w}) - \mathcal{R}_n(H^{a,b}_{w}) \right| \right] & \leq 2 \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{1}{n} \sum_{i=1}^n \varepsilon_i U_{w,i}^{a,b} \right| \right]. \end{aligned} \end{equation} \iffalse We now fix $(a,b)$ and employ a standard symmetrization argument (see for instance \cite[Lemma~11.4]{Boucheron2013}). Consider for $i=1,\ldots,n$, $w \in \mathcal{W}_\lambda^0$ the random variables $U_{w,i}^{a,b} = (H^{a,b}_{w}(X_i)-Y_i)^2$ and denote by $\mathcal{U}$ the $\sigma$-algebra generated by these. Denote by $\tilde{U}_{w,i}^{a,b}, i\in\{1,\ldots,n\},w \in \mathcal{W}_\lambda^0$, an independent copy of $U_{w,i}^{a,b},i\in\{1,\ldots,n\},w \in \mathcal{W}_\lambda^0$ and let $\varepsilon_1,\ldots,\varepsilon_n$ denote i.i.d.\ Rademacher random variables independent of all other random variables. Then we obtain by Jensen's inequality \begin{equation} \label{eq:auxEq15} \begin{aligned} \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left| \mathcal{R}(H^{a,b}_{w}) - \mathcal{R}_n(H^{a,b}_{w}) \right| \right] & = \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{1}{n} \sum_{i=1}^n U_{w,i}^{a,b} - \mathbb{E}[\tilde{U}_{w,i}^{a,b}|\mathcal{U}] \right| \right] \\ & \leq \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{1}{n} \sum_{i=1}^n U_{w,i}^{a,b} - \tilde{U}_{w,i}^{a,b} \right| \right] \\ & = \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{1}{n} \sum_{i=1}^n \varepsilon_i (U_{w,i}^{a,b} - \tilde{U}_{w,i}^{a,b}) \right| \right] \\ & \leq 2 \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{1}{n} \sum_{i=1}^n \varepsilon_i U_{w,i}^{a,b} \right| \right]. \end{aligned} \end{equation} \fi In the next step we denote by $\mathbf{X}^i$ the vector with components $\mathbf{X}^i_j = \varrho(a_j \cdot X_i + b_j)$, $j=1,\ldots,N$ and rewrite $H^{a,b}_{w}(X_i)=w \cdot \mathbf{X}^i$. Then we use the triangle inequality, Jensen's inequality and independence to estimate \begin{equation} \label{eq:auxEq16} \begin{aligned} \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{1}{n} \sum_{i=1}^n \varepsilon_i U_{w,i}^{a,b} \right| \right] & \leq \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{1}{n} \sum_{i=1}^n \varepsilon_i H^{a,b}_{w}(X_i)^2 \right| \right] + \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{2}{n} \sum_{i=1}^n \varepsilon_i H^{a,b}_{w}(X_i) Y_i\right| \right] \\ & \quad \quad + \frac{1}{n} \mathbb{E}\left[ \left| \sum_{i=1}^n \varepsilon_i Y_i^2 \right| \right] \\ & \leq \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|w^\top \left(\frac{1}{n} \sum_{i=1}^n \varepsilon_i \mathbf{X}^i [\mathbf{X}^i]^\top \right) w \right| \right] \\ & \quad \quad + \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{2}{n} w^\top \sum_{i=1}^n \varepsilon_i \mathbf{X}^i Y_i\right| \right] + \frac{1}{n} \mathbb{E}\left[ \left| \sum_{i=1}^n \varepsilon_i Y_i^2 \right|^2 \right]^{1/2} \\ & \leq \frac{\lambda^2}{n} \mathbb{E}\left[ \left\| \sum_{i=1}^n \varepsilon_i \mathbf{X}^i [\mathbf{X}^i]^\top \right\|_{F}^2 \right]^{1/2} + \frac{2 \lambda}{n} \mathbb{E}\left[ \left\| \sum_{i=1}^n \varepsilon_i \mathbf{X}^i Y_i\right\|^2 \right]^{1/2} \\ & \quad \quad + \frac{1}{n} \left( \sum_{i=1}^n \mathbb{E}[Y_i^4] \right)^{1/2}, \end{aligned} \end{equation} where $\|\cdot\|_F$ is the Frobenius norm on $\mathbb{R}^{N \times N}$. Denoting by $\langle \cdot,\cdot \rangle_F $ the Frobenius (matrix) inner product on $\mathbb{R}^{N \times N}$ and using independence and $\mathbb{E}[\varepsilon_i \varepsilon_j] = \delta_{ij}$ we obtain \[ \mathbb{E}\left[ \left\| \sum_{i=1}^n \varepsilon_i \mathbf{X}^i [\mathbf{X}^i]^\top \right\|_{F}^2 \right] = \mathbb{E}\left[ \sum_{i,j=1}^n \varepsilon_i \varepsilon_j \langle \mathbf{X}^i [\mathbf{X}^i]^\top, \mathbf{X}^j [\mathbf{X}^j]^\top \rangle \right] = n \mathbb{E}\left[ \left\| \mathbf{X}^1 [\mathbf{X}^1]^\top \right\|_{F}^2 \right]. \] Employing an analogous argument for the second term in the right hand side of \eqref{eq:auxEq16} (now with the standard inner product on $\mathbb{R}^N$) yields \begin{equation} \label{eq:auxEq17} \begin{aligned} \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left|\frac{1}{n} \sum_{i=1}^n \varepsilon_i U_{w,i}^{a,b} \right| \right] & \leq \frac{\lambda^2}{\sqrt{n}} \mathbb{E}\left[ \left\| \mathbf{X}^1 [\mathbf{X}^1]^\top \right\|_{F}^2 \right]^{1/2} + \frac{2 \lambda}{\sqrt{n}} \mathbb{E}\left[ \left\| \mathbf{X}^1 Y_1\right\|^2 \right]^{1/2} \\ & \quad \quad + \frac{1}{\sqrt{n}} \mathbb{E}[Y_1^4]^{1/2}. \end{aligned} \end{equation} Using $\left\| \mathbf{X}^1 [\mathbf{X}^1]^\top \right\|_{F}^2 = \sum_{k,l=1}^N [\mathbf{X}^1_k]^2 [\mathbf{X}^1_l]^2 = \|\mathbf{X}^1\|^4$ and inserting the bound \eqref{eq:auxEq17} in \eqref{eq:auxEq15} we obtain \begin{equation} \label{eq:auxEq18} \begin{aligned} \mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left| \mathcal{R}(H^{a,b}_{w}) - \mathcal{R}_n(H^{a,b}_{w}) \right| \right] & \leq \frac{2\lambda^2}{\sqrt{n}} \mathbb{E}\left[ \left\| \mathbf{X}^1 \right\|^4 \right]^{1/2} + \frac{4 \lambda}{\sqrt{n}} \mathbb{E}\left[ \left\| \mathbf{X}^1 \right\|^2 Y_1^2 \right]^{1/2} \\ & \quad \quad + \frac{2}{\sqrt{n}} \mathbb{E}[Y_1^4]^{1/2}. \end{aligned} \end{equation} Employing the bound \begin{equation} \label{eq:auxEq56} \left\| \mathbf{X}^1 \right\|^2 = \sum_{j=1}^N [\varrho(a_j \cdot X_1 + b_j)]^2 \leq 2 \sum_{j=1}^N \|a_j\|^2 \| X_1 \|^2 + |b_j|^2 \end{equation} we estimate using the Minkowski integral inequality and the triangle inequality \[ \begin{aligned} \mathbb{E}\left[ \left\| \mathbf{X}^1 \right\|^4 \right]^{1/2} & \leq \mathbb{E}\left[ \left(2 \sum_{j=1}^N \|a_j\|^2 \| X_1 \|^2 + |b_j|^2 \right)^2 \right]^{1/2} \leq 2 \sum_{j=1}^N \mathbb{E}\left[ \left( \|a_j\|^2 \| X_1 \|^2 + |b_j|^2 \right)^2 \right]^{1/2} \\ & \leq 2 \sum_{j=1}^N \|a_j\|^2\mathbb{E}[\| X_1 \|^4]^{1/2} + |b_j|^2. \end{aligned} \] The second term in the right hand side of \eqref{eq:auxEq18} can be bounded similarly with \eqref{eq:auxEq56}. Inserting this and \eqref{eq:auxEq18} in \eqref{eq:auxEq14} yields \begin{equation} \label{eq:auxEq19} \begin{aligned} & \mathbb{E}[\mathcal{R}(H^{A,B}_{\widehat{W}_\lambda}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) + \mathcal{R}_n(H^{A,B}_{W})- \mathcal{R}(H^{A,B}_{W})] \\ & \leq 2 \mathbb{E}\left[ \frac{4\lambda^2}{\sqrt{n}} \left(\sum_{j=1}^N \|A_j\|^2\mathbb{E}[\| X_1 \|^4]^{1/2} + |B_j|^2\right) \right] \\ & \quad +2\mathbb{E}\left[ \frac{2^{2+\frac{1}{2}} \lambda}{\sqrt{n}} \left( \sum_{j=1}^N \|A_j\|^2 \mathbb{E}[Y_1^2\| X_1 \|^2] + |B_j|^2 \mathbb{E}[Y_1^2] \right)^{1/2}\right] + \frac{4}{\sqrt{n}} \mathbb{E}[Y_1^4]^{1/2} \\ & \leq \frac{8\lambda^2N}{\sqrt{n}} (\mathbb{E}[\|A_1\|^2]\mathbb{E}[\| X_1 \|^4]^{1/2} + \mathbb{E}[|B_1|^2]) \\ & \quad +\frac{2^{3+\frac{1}{2}} \lambda \sqrt{N}}{\sqrt{n}}\left(\mathbb{E}[\|A_1\|^2] \mathbb{E}[Y_1^2\| X_1 \|^2] + \mathbb{E}[|B_1|^2] \mathbb{E}[Y_1^2]\right)^{1/2} + \frac{4}{\sqrt{n}} \mathbb{E}[Y_1^4]^{1/2}. \end{aligned} \end{equation} Recall that $A_1$ has a multivariate $t$-distribution $t_{\nu}(0,\mathbbm{1}_d)$, hence $A_1 \,{\buildrel d \over =}\, Z/\sqrt{U/\nu}$ where $Z \sim \mathcal{N}(0,\mathbbm{1}_d)$ and $U \sim \chi^2(\nu)$ are independent. Thus, $\mathbb{E}[\|A_1\|^2]=\mathbb{E}[\|Z\|^2] \mathbb{E}[\nu/U] = \nu d /(\nu-2)$. Using that $\|X_1\|_\infty \leq M$ and $\lambda \leq \frac{C_{\text{lam}} d^{p} }{\sqrt{N}}$ we may thus deduce from \eqref{eq:auxEq19} that \begin{equation} \label{eq:auxEq20} \begin{aligned} & \mathbb{E}[\mathcal{R}(H^{A,B}_{\widehat{W}_\lambda}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) + \mathcal{R}_n(H^{A,B}_{W})- \mathcal{R}(H^{A,B}_{W})] \\ & \leq \frac{C_{\text{est}}^2 d^{2p+2} }{\sqrt{n}} \end{aligned} \end{equation} with $C_{\text{est}}^2 = 8C_{\text{lam}}^2 (\frac{\nu}{\nu-2} M^2 + \mathbb{E}[|B_1|^2]) +2^{3+\frac{1}{2}} C_{\text{lam}}(\frac{\nu}{\nu-2}M^2\mathbb{E}[Y_1^2] + \mathbb{E}[|B_1|^2] \mathbb{E}[Y_1^2])^{1/2} + 4 \mathbb{E}[Y_1^4]^{1/2}$ not depending on $d$, $n$ or $N$. Combining \eqref{eq:auxEq20} with \eqref{eq:auxEq11} and \eqref{eq:auxEq13} we obtain \begin{equation} \begin{aligned} & \mathbb{E}[|H(\bar{X}) - H^{A,B}_{\widehat{W}_\lambda}(\bar{X}) |^2] \leq \left(\frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}}\right)^{2} + \frac{C_{\text{est}}^2 d^{2p+2} }{\sqrt{n}}. \end{aligned} \end{equation} \end{proof} \subsection{Stochastic gradient descent} \label{subsec:SGD} For the most common choices of $\mathcal{W}$ the solution to the optimization problem \eqref{eq:ERM} can be obtained by solving the system of linear equations \eqref{eq:OLSSol} or \eqref{eq:ConstrainedSol}, respectively. There may nevertheless be situations in which one is interested in solving \eqref{eq:ERM} using a stochastic gradient descent method (e.g.\ when comparing the performance of different learning methods in an experiment). Therefore, we will briefly discuss optimization of \eqref{eq:ConstrainedRegression} by stochastic gradient descent here and combine our error bound in Theorem~\ref{thm:TrainingErrConstrRegression} with the stochastic gradient descent optimization error bound from \cite{pmlr-v28-shamir13}. To this end, let $\mathcal{V} = \{w \in \mathbb{R}^N \,|\, \|w\|\leq \lambda \}$ denote the set within which we look for an optimizer, let $\Pi_{\mathcal{V}} \colon \mathbb{R}^N \to \mathcal{V}$ be the orthogonal projection onto $\mathcal{V}$, for $i=1,\ldots,n$ write $\mathbf{X}^i$ for the $\mathbb{R}^N$-valued random vector with components $\mathbf{X}^i_j = \varrho(A_j \cdot X_i + B_j)$, $j=1,\ldots,N$, let $\mathcal{T} \in \{2,3,\ldots\}$ denote the number of stochastic gradient descent iterations, let $\mathfrak{B} \in \{1,\ldots,n\}$ denote the batch size and let $J=\{J_{i,t}\}_{(i,t)\in \{1,\ldots,\mathfrak{B}\} \times \{1,\ldots,\mathcal{T}\}}$ denote i.i.d.\ random variables each having a uniform distribution on $\{1,\ldots,n\}$ and independent of $(A,B,D_n,\bar{X},\bar{Y})$. Then, starting with $W_1 =0$, we iteratively compute \begin{equation}\label{eq:SGD} W_{t+1} = \Pi_{\mathcal{V}}\left( W_t - \frac{2\eta_t}{\mathfrak{B}} \sum_{i=1}^{\mathfrak{B}} \mathbf{X}^{J_{i,t}} (W_t \cdot \mathbf{X}^{J_{i,t}} - Y_{J_{i,t}}) \right) , \quad t=1,\ldots,\mathcal{T}-1, \end{equation} where $\eta_t = \eta_0 t^{-1/2}$ for $t=1,\ldots,\mathcal{T}-1$. The parameter vector $W_\mathcal{T}$ is then used for the random neural network, i.e., $ H^{A,B}_{W_\mathcal{T}}$ is the learned function approximating $H$. The next proposition provides a bound on the prediction error. \begin{proposition} \label{prop:SGDtrained} Let $C>\frac{1}{2^{3/2} \pi}$, $\eta_0>0$ and $\nu>4$. Suppose $A_1 \sim t_{\nu}(0,\mathbbm{1}_d)$ and $B_1$ has density $\pi_{\text{b}}$ satisfying \eqref{eq:polyTails1D}. Suppose $H \colon \mathbb{R}^d \to \mathbb{R}$ is of the form $H(x) = \mathbb{E}[\Phi(x+V)]$ with $\Phi \in L^1(\mathbb{R}^d)$ and $V$ satisfying \eqref{eq:charFctAss}. Assume that $\|X_1\|_\infty\leq M$, $\P$-a.s.\ and $\mathbb{E}[|Y_1|^4]< \infty$. Let $\eta_t = \eta_0 t^{-1/2}$ for $t=1,\ldots,\mathcal{T}-1$ and $\lambda \in \frac{1}{\sqrt{N}}[C_{\text{wgt}}\|\Phi\|_{L^1(\mathbb{R}^d)}(\nu+d)^{2k+\frac{1}{2}},C_{\text{lam}}d^p] $ with $k \in \mathbb{N}$, $C_{\text{wgt}}>0$ as in Theorem~\ref{thm:ApproxError} and $p \geq 0$, $C_{\text{lam}} >0$ not depending on $n,N,d$ or $\mathcal{T}$. Then there exist $C_{\text{app}},C_{\text{est}},C_{\text{opt}}>0$ such that \begin{equation} \label{eq:fullError3} \begin{aligned} \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W_\mathcal{T}}(\bar{X}) |^2]^{1/2} & \leq \frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}} + \frac{C_{\text{est}} d^{p+1} }{n^{\frac{1}{4}}} \\ & \quad \quad + \frac{C_{\text{opt}} d^{p+2} N (2+\log(\mathcal{T}))^{\frac{1}{2}}}{\mathcal{T}^{\frac{1}{4}}}. \end{aligned} \end{equation} The constant $k$ only depends on $\pi_{\text{b}}$ and the constants $C_{\text{app}},C_{\text{est}},C_{\text{opt}}$ depend on $\nu, \pi_{\text{b}}, C, M$, $\mathbb{E}[Y_1^4], \eta_0, C_{\text{lam}}$, but they do not depend on $d$, $n$, $N$ or $\mathcal{T}$. \end{proposition} \begin{remark} The first two terms in the error bound in \eqref{eq:fullError3} are as in the bound \eqref{eq:fullError2} in Theorem~\ref{thm:TrainingErrConstrRegression}, whereas the last term in \eqref{eq:fullError3} is due to the stochastic gradient descent optimization. The rate of convergence to $0$ of this last error term as a function of $\mathcal{T}$ could be further improved, e.g., by using a more refined optimization scheme (based on averaging) than \eqref{eq:SGD}, see for instance \cite{pmlr-v28-shamir13}. However, for our purposes the bound in Proposition~\ref{prop:SGDtrained} suffices as this bound already proves that the overall error does not suffer from the curse of dimensionality. \end{remark} \begin{proof} Let $C_{\text{app}}>0$ be as in Theorem~\ref{thm:ApproxError}, let $W$ be the $\mathbb{R}^N$-valued, $\sigma(A,B)$-measurable random vector satisfying \eqref{eq:L2error2} (see Theorem~\ref{thm:ApproxError}) and let $C_{\text{est}}>0$ be as in Theorem~\ref{thm:TrainingErrConstrRegression}. By independence and \eqref{eq:regressionFunction} we obtain (as in \eqref{eq:auxEq12}-\eqref{eq:auxEq11} in the proof of Theorem~\ref{thm:TrainingErrConstrRegression}) \begin{equation} \label{eq:auxEq21} \begin{aligned} & \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W_\mathcal{T}}(\bar{X}) |^2] \\ & \quad = \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W}(\bar{X}) |^2] + \mathbb{E}[\mathcal{R}(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}(H^{A,B}_{W})] \\ & \quad \leq \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W}(\bar{X}) |^2] \\ & \quad \quad + \mathbb{E}[\mathcal{R}(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}_n(H^{A,B}_{W_\mathcal{T}}) +\mathcal{R}_n(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) + \mathcal{R}_n(H^{A,B}_{W})- \mathcal{R}(H^{A,B}_{W})], \end{aligned} \end{equation} where we used \eqref{eq:ConstrainedRegression} and $W \in \mathcal{W}_\lambda$ (as established in the proof of Theorem~\ref{thm:TrainingErrConstrRegression}) in the last step. The first expectation in the right hand side of \eqref{eq:auxEq21} has been bounded in \eqref{eq:auxEq13} in the proof of Theorem~\ref{thm:TrainingErrConstrRegression}. For the second expectation we may proceed analogously as in \eqref{eq:auxEq14}: we use the same notation as in \eqref{eq:auxEq14} and, in addition, write $W^{a,b}_\mathcal{T}$ for the output of the stochastic gradient descent algorithm with $(A,B)$ fixed to $(a,b)$. Then independence yields \begin{equation} \label{eq:auxEq22} \begin{aligned} \mathbb{E}[&\mathcal{R}(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}_n(H^{A,B}_{W_\mathcal{T}}) + \mathcal{R}_n(H^{A,B}_{W})- \mathcal{R}(H^{A,B}_{W})] \\ & = \mathbb{E}[\left.\mathbb{E}[\mathcal{R}(H^{a,b}_{W^{a,b}_\mathcal{T}}) - \mathcal{R}_n(H^{a,b}_{W^{a,b}_\mathcal{T}}) + \mathcal{R}_n(H^{a,b}_{W^{a,b}})- \mathcal{R}(H^{a,b}_{W^{a,b}})]\right\rvert_{(a,b)=(A,B)}] \\ & \leq 2 \mathbb{E}\left[ \left.\mathbb{E}\left[ \sup_{w \in \mathcal{W}_\lambda^0} \left| \mathcal{R}(H^{a,b}_{w}) - \mathcal{R}_n(H^{a,b}_{w}) \right| \right] \right\rvert_{(a,b)=(A,B)}\right]. \end{aligned} \end{equation} Now we can compare \eqref{eq:auxEq21} and \eqref{eq:auxEq22} to \eqref{eq:auxEq11} and \eqref{eq:auxEq14} in the proof of Theorem~\ref{thm:TrainingErrConstrRegression}. We see that the decomposition \eqref{eq:auxEq21} yields the same error terms as in Theorem~\ref{thm:TrainingErrConstrRegression} plus the additional term $ \mathbb{E}[\mathcal{R}_n(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) ]$. Therefore, Theorem~\ref{thm:TrainingErrConstrRegression} shows that \begin{equation} \label{eq:auxEq23} \begin{aligned} \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W_\mathcal{T}}(\bar{X}) |^2]^{1/2} & \leq \frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}} + \frac{C_{\text{est}} d^{p+1} }{n^{\frac{1}{4}}} \\ & \quad + \mathbb{E}[\mathcal{R}_n(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) ]^{1/2}. \end{aligned} \end{equation} We now analyze the last term. Write $W_\mathcal{T}^{a,b,d_n}$ for the output of the stochastic gradient descent algorithm and $\widehat{W}_\lambda^{a,b,d_n}$ for the solution to \eqref{eq:ConstrainedRegression} when $(A,B,D_n)=(a,b,d_n)$. From the updating scheme it is clear that there exists a measurable function $F$ such that $W_\mathcal{T} = F(A,B,D_n,J) = W_\mathcal{T}^{A,B,D_n}$. Furthermore (as argued in the proof of Theorem~\ref{thm:TrainingErrConstrRegression}), $\widehat{W}_\lambda^{a,b,d_n} = G(a,b,d_n)$ for a measurable function $G$ and $\widehat{W}_\lambda^{A,B,D_n} =\widehat{W}_\lambda $. Thus, we may use independence to write \begin{equation} \label{eq:auxEq24} \begin{aligned} \mathbb{E}[\mathcal{R}_n(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) ] = \mathbb{E}[\left.\mathbb{E}[\mathcal{R}_n^{d_n}(H^{a,b}_{W_\mathcal{T}^{a,b,d_n}}) - \mathcal{R}_n^{d_n}(H^{a,b}_{\widehat{W}_\lambda^{a,b,d_n}}) ]\right\rvert_{(a,b,d_n)=(A,B,D_n)}], \end{aligned} \end{equation} where $\mathcal{R}_n^{d_n}(f) = \frac{1}{n} \sum_{i=1}^n (f(x_i) - y_i)^2$ for ${d_n}=((x_1,y_1),\ldots,(x_n,y_n))$. Consider $(a,b,d_n) \in (\mathbb{R}^d)^N\times \mathbb{R}^N\times ([-M,M]^d\times\mathbb{R})^n$ as fixed now and write $\mathbf{x}^i$ for the vector with $\mathbf{x}^i_j = \varrho(a_j \cdot x_i + b_j)$, $j=1,\ldots,N$. Let $F \colon \mathcal{V} \to \mathbb{R}$, $F(w) = \frac{1}{n} \sum_{i=1}^n (w\cdot \mathbf{x}^i - y_i)^2$. Then $H^{a,b}_w(x_i) = w \cdot \mathbf{x}^i$, $\mathcal{R}_n^{d_n}(H^{a,b}_w) = F(w)$ and hence $\hat{w}:=\widehat{W}_\lambda^{a,b,d_n}$ is a (global) minimizer of $F$ in $\mathcal{V}$. Write $w_t := W_t^{a,b,d_n}$ and recall \begin{equation}\label{eq:SGDconditional} w_{t+1} = \Pi_{\mathcal{V}}\left( w_t - \eta_t \hat{g}_t \right) , \quad t=1,\ldots,\mathcal{T}-1 \end{equation} with $\hat{g}_t = \frac{2}{\mathfrak{B}} \sum_{i=1}^{\mathfrak{B}} \mathbf{x}^{J_{i,t}} (w_t \cdot \mathbf{x}^{J_{i,t}} - y_{J_{i,t}})$. Independence implies $\mathbb{E}[\hat{g}_t |w_t] = \frac{2}{\mathfrak{B}} \sum_{i=1}^{\mathfrak{B}} \mathbb{E}[\mathbf{x}^{J_{i,t}} (w \cdot \mathbf{x}^{J_{i,t}} - y_{J_{i,t}})]\rvert_{w=w_t} = \frac{2}{n} \sum_{j=1}^n \mathbf{x}^{j} (w_t \cdot \mathbf{x}^{j} - y_{j}) = \nabla F(w_t)$. Furthermore, $F$ is convex and the Minkowski integral inequality and independence yield \begin{equation} \label{eq:auxEq25} \begin{aligned} \mathbb{E}[\|\hat{g}_t\|^2] & \leq 4 \mathbb{E}\left[\left(\frac{1}{\mathfrak{B}} \sum_{i=1}^{\mathfrak{B}} \|\mathbf{x}^{J_{i,t}}\|(|w_t \cdot \mathbf{x}^{J_{i,t}}| +| y_{J_{i,t}}|)\right)^2\right] \\ & \leq 4 \left( \frac{1}{\mathfrak{B}} \sum_{i=1}^{\mathfrak{B}} \left( \mathbb{E}\left[ \|\mathbf{x}^{J_{i,t}}\|^2(|w_t \cdot \mathbf{x}^{J_{i,t}}| +| y_{J_{i,t}}|)^2\right] \right)^{1/2} \right)^2 \\ & \leq \frac{8}{n} \sum_{j=1}^n \|\mathbf{x}^{j}\|^2 (\mathbb{E}[\|w_t\|^2] \|\mathbf{x}^{j}\|^2 +| y_{j}|^2) \\ & \leq \frac{16}{n} \sum_{i=1}^n \left(\sum_{j=1}^N \|a_j\|^2 \| x_i \|^2 + |b_j|^2\right) (\lambda^2 \|\mathbf{x}^{i}\|^2 +| y_{i}|^2) \\ & \leq 32 \left(1+M^2 d \|a\|_F^2 + \|b\|^2\right)^2 (\lambda^2 +\frac{1}{n} \sum_{i=1}^n| y_{i}|^2), \end{aligned} \end{equation} where in the last two inequalities we used the estimate $\left\| \mathbf{x}^i \right\|^2 = \sum_{j=1}^N [\varrho(a_j \cdot x_i + b_j)]^2 \leq 2 \sum_{j=1}^N \|a_j\|^2 \| x_i \|^2 + |b_j|^2 $. \citet[Theorem~2]{pmlr-v28-shamir13} hence implies that \[ \mathbb{E}[F(w_\mathcal{T})-F(\hat{w})] \leq \left(\frac{4 \lambda^2}{\eta_0} + \eta_0 32 \left(1+d M^2 \|a\|_F^2 + \|b\|^2\right)^2 (\lambda^2 +\frac{1}{n} \sum_{i=1}^n| y_{i}|^2)\right) \frac{2+\log(\mathcal{T})}{\sqrt{\mathcal{T}}}. \] Inserting this in \eqref{eq:auxEq24} and using independence yields \begin{equation} \label{eq:auxEq26} \begin{aligned} \mathbb{E}[& \mathcal{R}_n(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) ] \\ & \leq \mathbb{E} \left[\frac{4 \lambda^2}{\eta_0} + 32 \eta_0 \left(1+ d M^2 \|A\|_F^2 + \|B\|^2\right)^2 (\lambda^2 +\frac{1}{n} \sum_{i=1}^n| Y_{i}|^2)\right] \frac{2+\log(\mathcal{T})}{\sqrt{\mathcal{T}}} \\ & \leq \left(\frac{4 \lambda^2}{\eta_0} + 96\eta_0 (1+d^2 M^4\mathbb{E}[ \|A\|_F^4] + \mathbb{E}[\|B\|^4]) (\lambda^2 +\mathbb{E}[| Y_{1}|^2])\right) \frac{2+\log(\mathcal{T})}{\sqrt{\mathcal{T}}}. \end{aligned} \end{equation} Employing Minkowski's integral inequality we estimate \begin{equation} \label{eq:auxEq27} \begin{aligned} d^2M^4\mathbb{E}[ \|A\|_F^4] + \mathbb{E}[\|B\|^4] & \leq d^2M^4\left(\sum_{j=1}^N \mathbb{E}[\|A_j\|^4]^{1/2} \right)^2 + \left(\sum_{j=1}^N \mathbb{E}[|B_j|^4]^{1/2}\right)^2 \\ & = N^2(d^2 M^4 \mathbb{E}[\|A_1\|^4] +\mathbb{E}[|B_1|^4] ). \end{aligned} \end{equation} Recall that $A_1 \,{\buildrel d \over =}\, Z/\sqrt{U/\nu}$, where $Z \sim \mathcal{N}(0,\mathbbm{1}_d)$ and $U \sim \chi^2(\nu)$ are independent. Therefore $\mathbb{E}[\|A_1\|^4]=\mathbb{E}[\|Z\|^4] \mathbb{E}[\nu^2/U^2]$ and one obtains analogously to \eqref{eq:auxEq27} the estimate $ \mathbb{E}[\|Z\|^4]\leq d^2 \mathbb{E}[Z_1^4]$. Inserting this into \eqref{eq:auxEq27} and \eqref{eq:auxEq26} and estimating $\lambda \leq \frac{C_{\text{lam}} d^{p} }{\sqrt{N}}$ yields \begin{equation} \label{eq:auxEq28} \begin{aligned} \mathbb{E}[& \mathcal{R}_n(H^{A,B}_{W_\mathcal{T}}) - \mathcal{R}_n(H^{A,B}_{\widehat{W}_\lambda}) ] \\ & \leq \frac{C_{\text{opt}}^2}{4} \left((1+N^2)\frac{ d^{2p+4} }{N} + (1+N^2)d^4\right) \frac{2+\log(\mathcal{T})}{\sqrt{\mathcal{T}}} \\ & \leq C_{\text{opt}}^2 d^{2p+4} N^2 \frac{2+\log(\mathcal{T})}{\sqrt{\mathcal{T}}} \end{aligned} \end{equation} with $C_{\text{opt}}^2 = 4 \max(\frac{4}{\eta_0},96\eta_0)\max(2,3 M^4 \nu^2/[(\nu-2)(\nu-4)] +\mathbb{E}[|B_1|^4]) \max(C_{\text{lam}}^2,\mathbb{E}[| Y_{1}|^2]) $ and where we used $\mathbb{E}[Z_1^4]=3$, $\mathbb{E}[U^{-2}] = 1/[(\nu-2)(\nu-4)]$. Combining this with \eqref{eq:auxEq23} yields \eqref{eq:fullError3}, as claimed. \end{proof} \subsection{Application to basket option pricing} \label{sec:Options} As a first application of the results derived in Sections~\ref{subsec:Regression}--\ref{subsec:SGD} we consider the problem of learning prices of basket put options in certain ``non-degenerate'' models. Suppose that $y_i$ is the market price of a put option with strike $K_i>0$ written on a basket of $m$ assets. Assume that, up to some additive noise, these market prices are ``generated'' from an unknown, non-degenerate stochastic model. This means that we assume \[y_i = \mathbb{E}\left[\max\left(K_i-\sum_{i=1}^m w_i S_{T,i},0\right)\right] + \varepsilon_i, \quad i=1,\ldots,n, \] where $\varepsilon_1,\ldots,\varepsilon_n$ are i.i.d.\ random variables, $S_T = (S_{T,1},\ldots,S_{T,m})$ is a $[0,\infty)^m$-valued random vector and $w_1,\ldots,w_m \in [0,\infty)$ are non-negative weights. Assume that $\mathbb{E}[\varepsilon_1]=0$, $\mathbb{E}[\varepsilon_1^4] < \infty$ and $\{\varepsilon_i\}_{i=1,\ldots,n}$ are independent of $(A,B,S_T)$. We think of $S_T$ as the value at time $T$ of a price process $S$ (for which $\P$ is a martingale measure). The goal is to learn the pricing function $H(K):= \mathbb{E}\left[\max\left(K-\sum_{i=1}^m w_i S_{T,i},0\right)\right]$ from the observed market prices $y_1,\ldots,y_n$. This fits into the framework introduced above (see Section~\ref{subsec:learningProblem}) if we let $M=\max_{i=1,\ldots,n} K_i $ and consider $K_1,\ldots,K_n$ as the observed realizations of the $n$ i.i.d.\ random variables $X_1,\ldots,X_n$ so that also $y_i = H(K_i) + \varepsilon_i$ is the realization of $Y_i = H(X_i) + \varepsilon_i$. We assume that $X_1$ is distributed uniformly on $[0,M]$ and $\{X_i\}_{i=1,\ldots,n}$ are independent of $\{\varepsilon_i\}_{i=1,\ldots,n}, (A,B)$. Then the option pricing function $H$ is indeed the regression function \eqref{eq:regressionFunction} and we obtain the following corollary. Recall that $\bar{X}$ has the same distribution as $X_1$ and is independent of $(A,B,D_n)$. \begin{corollary} \label{cor:appl} Let $\nu>4$, $C>\frac{1}{2^{3/2} \pi}$, $\eta_0>0$ and $\bar{c} > 0 $ be constants which do not depend on $n, N$ or $\mathcal{T}$. Suppose $A_1 \sim t_{\nu}(0,1)$ and $B_1$ has density $\pi_{\text{b}}$ satisfying \eqref{eq:polyTails1D}. Assume that the $[0,\infty)^m$-valued random vector $S_T$ satisfies $|\mathbb{E}[e^{-i \xi w \cdot S_T }]|\leq \exp(-C|\xi|^2)$ for all $\xi \in \mathbb{R}$. Then there exists $C_0>0$ such that the prediction error bound \begin{align} \label{eq:OptionPriceError} & \mathbb{E}[|H(\bar{X}) - T_M(H^{A,B}_{\widehat{W}}(\bar{X})) |^2]^{1/2} \leq C_0 \left( \frac{(\log(n)+1)^{1/2}\sqrt{N}}{\sqrt{n}} + \frac{1}{\sqrt{N}}\right) \end{align} holds and there exist $C_1,C_2,\underline{c} >0$ such that for any $\lambda \in \frac{1}{\sqrt{N}}[\underline{c},\bar{c}]$ the prediction error bounds \begin{align} \label{eq:OptionPriceError2} & \mathbb{E}[|H(\bar{X}) - H^{A,B}_{\widehat{W}_\lambda}(\bar{X}) |^2]^{1/2} \leq C_1 \left( \frac{1}{\sqrt{N}} + \frac{1 }{n^{\frac{1}{4}}} \right), \\ \label{eq:OptionPriceError3} & \mathbb{E}[|H(\bar{X}) - H^{A,B}_{W_\mathcal{T}}(\bar{X}) |^2]^{1/2} \leq C_2\left(\frac{1}{\sqrt{N}} + \frac{1}{n^{\frac{1}{4}}} + \frac{ N (2+\log(\mathcal{T}))^{\frac{1}{2}}}{\mathcal{T}^{\frac{1}{4}}}\right) \end{align} hold. The constants $C_0,C_1,C_2,\underline{c}$ do not depend on $n$, $N$ or $\mathcal{T}$. \end{corollary} \begin{remark} The proof of Corollary~\ref{cor:appl} shows that $\underline{c}$ does not depend on $\bar{c} $. Hence, by choosing $\bar{c} > \underline{c}$ it can always be guaranteed that $[\underline{c},\bar{c}]$ is not empty. \end{remark} \begin{remark} The hypothesis $|\mathbb{E}[e^{-i \xi w \cdot S_T }]|\leq \exp(-C|\xi|^2)$ is inherited from Theorem~\ref{thm:ApproxError}. In Theorem~\ref{thm:ApproxError} this hypothesis guarantees that the constants do not grow exponentially in the dimension $d$. In the situation here $d=1$ and so this hypothesis could be relaxed considerably: it could be replaced by the assumption $|\mathbb{E}[e^{-i \xi w \cdot S_T }]|\leq \exp(-C|\xi|^\alpha)$ for some $C>0$, $\alpha >0$ or even by the assumption that $|\mathbb{E}[e^{-i \xi w \cdot S_T }]|\leq C (1+|\xi|)^{-\beta}$ for some $C>0$ and sufficiently large $\beta >0$ (depending on $\nu$ and $\pi_{\text{b}}$). \end{remark} \begin{proof} Firstly, by assumption we have $|X_1|\leq M$, $\P$-a.s.\ and $H \colon \mathbb{R} \to \mathbb{R}$ satisfies for $K \in [0,M]$ that \[ H(K) =\mathbb{E}\left[\max\left(K-w \cdot S_{T},0\right)\right] = \mathbb{E}[\Phi(K+V)] \] with $V = - w \cdot S_T$ and $\Phi(y) = y \mathbbm{1}_{[0,M]}(y)$. Hence, $H(\bar{X}) = \tilde{H}(\bar{X})$ $\P$-a.s.\ with $\tilde{H}(x) = \mathbb{E}[\Phi(x+V)]$ for $x \in \mathbb{R}$. Furthermore, $\Phi \in L^1(\mathbb{R})$, $V$ satisfies \eqref{eq:charFctAss}, $\sigma^2=\sup_{x \in \mathbb{R}} \mathbb{E}[(Y_1-H(X_1))^2|X_1=x] = \mathbb{E}[\varepsilon_1^2]< \infty$ and $|\tilde{H}(x)| \leq M$ for all $x \in \mathbb{R}$. Thus, the hypotheses of Theorem~\ref{thm:TrainingErrRegression} with $L=M$ are satisfied and so, using $H(\bar{X}) = \tilde{H}(\bar{X})$ $\P$-a.s., we obtain that there exist $k \in \mathbb{N}$ and $\tilde{C}_{\text{app}}>0$ such that the prediction error bound \eqref{eq:fullError} holds. Hence \eqref{eq:OptionPriceError} follows with $C_0 = \max(\tilde{C}_{\text{app}} \max(\sigma,M), \tilde{C}_{\text{app}} \|\Phi\|_{L^1(\mathbb{R})} (\nu+1)^{k+3})$. Next we prove \eqref{eq:OptionPriceError2}. To this end, notice $\mathbb{E}[|Y_1|^4] \leq 8(\mathbb{E}[|H(X_1)|^4] + \mathbb{E}[|\varepsilon_1|^4]) \leq 8(M^4 + \mathbb{E}[|\varepsilon_1|^4]) < \infty$ and let $C_{\text{app}},C_{\text{wgt}}>0$ be as in Theorem~\ref{thm:ApproxError}. Then Theorem~\ref{thm:TrainingErrConstrRegression} proves that for any $\lambda >0$ satisfying $\frac{C_{\text{wgt}}\|\Phi\|_{L^1(\mathbb{R})}(\nu+1)^{2k+\frac{1}{2}}}{\sqrt{N}} \leq \lambda \leq \frac{\bar{c}} {\sqrt{N}}$ there exists a constant $C_{\text{est}}>0$ such that the prediction error bound \eqref{eq:fullError2} holds. The proof actually shows that the same constant can be chosen for all $\lambda$ in the specified range. Thus, \eqref{eq:OptionPriceError2} follows with $C_1 = \max(C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R})} (\nu+1)^{k+3},C_{\text{est}})$ and $\underline{c}= C_{\text{wgt}}\|\Phi\|_{L^1(\mathbb{R})}(\nu+1)^{2k+\frac{1}{2}}$. Furthermore, Proposition~\ref{prop:SGDtrained} proves that there exists $C_{\text{opt}}>0$ such that \eqref{eq:fullError3} holds. Setting $C_2 = \max(C_1,C_{\text{opt}}) $ we obtain \eqref{eq:OptionPriceError3}. In these results we proved that the constants $C_{\text{app}},C_{\text{est}},C_{\text{opt}}$ depend on $\nu, \pi_{\text{b}}, C, M$, $\mathbb{E}[Y_1^4]$, $\eta_0$, $\bar{c}$, but they do not depend on $n$, $N$ or $\mathcal{T}$, hence it follows that $C_0,C_1,C_2,\underline{c}$ do not depend on $n$, $N$ or $\mathcal{T}$. \end{proof} \section{Learning Black-Scholes type PDEs} \label{sec:Kolmogorov} In this section we apply the results from Section~\ref{sec:Learning} to prove that random neural networks are capable of learning Black-Scholes type partial (integro-)differential equations (also referred to as (non-local) PDEs) without the curse of dimensionality. More specifically, we consider the problem of learning solutions to Kolmogorov PDEs associated to exponential L\'evy-processes, which includes the Black-Scholes PDE as a special case. The learning methods used to tackle this problem are random neural networks trained by (constrained) regression or stochastic gradient descent. By combining the results from Theorems~\ref{thm:LevyApprox}, \ref{thm:TrainingErrRegression}, \ref{thm:TrainingErrConstrRegression} and from Proposition \ref{prop:SGDtrained} we obtain bounds on the prediction error. The dependence on the dimension $d$ in these bounds is explicit and at most polynomial, whereas the bounds decay at polynomial rate in the number of samples $n$ and the network size $N$ (and the number of stochastic gradient descent iterations $\mathcal{T}$). Hence, the number of samples, hidden nodes of the network and gradient steps required to achieve a prescribed prediction accuracy $\varepsilon >0$ grows at most polynomially in $d$ and $\varepsilon^{-1}$. This means that random neural networks are capable of learning solutions to such Kolmogorov PDEs without the curse of dimensionality. For the reader's convenience we introduce in Section~\ref{subsec:learningProblemLevy} in detail again all the objects relevant to the discussion. Section~\ref{subsec:learningResultsLevy} then contains the prediction error bounds for Black-Scholes type PDEs. We conclude in Section~\ref{sec:numerics} with a numerical experiment. \subsection{Formulation of the learning problem for PDEs} \label{subsec:learningProblemLevy} We again put ourselves in the situation studied in Section~\ref{subsec:ApproxLevy} and consider for each $d \in \mathbb{N}$ the partial (integro-)differential equation \begin{equation} \label{eq:PIDEs2} \begin{array}{rl} \partial_t u_d(t,s) & = \frac{1}{2} \sum_{k,l=1}^d s_k s_l \Sigma^d_{k,l} \partial_{s_k} \partial_{s_l} u_d(t,s) + \sum_{i=1}^d s_i \tilde{\gamma}^d_i \partial_{s_i} u_d(t,s) \\ & \quad + \int_{\mathbb{R}^d} \left[u_d(t,s e^y)-u_d(t,s)-\sum_{i=1}^d (e^{y_i}-1) s_i \partial_{s_i} u_d(t,s) \right] \nu^d_\mathrm{L}(d y) , \\ u_d(0,s) &= \varphi_d(s) \end{array} \end{equation} for $s \in (0,\infty)^d, t > 0$, where $\varphi_d \colon (0,\infty)^d \to \mathbb{R}$ is a ``payoff'' function and $(\Sigma^d,\gamma^d,\nu^d_\mathrm{L})$ is the characteristic triplet of a L\'evy process $L^d$, we write $\tilde{\gamma}^d_i = \gamma_i^d + \frac{1}{2} \Sigma_{i,i}^d + \int_{\mathbb{R}^d} (e^{y_i}-1- y_i \mathbbm{1}_{\{\|y\|\leq 1\}}) \nu^d_\mathrm{L}(d y)$, $i=1,\ldots,d$, for the shifted drift vector and we assume $ \nu^d_\mathrm{L}(\{y \in \mathbb{R}^d \, | \, \|y\|>R\}) = 0$ for some $R>1$. Furthermore, we recall the notation $s\exp(x) =(s_1\exp(x_1),\ldots,s_d\exp(x_d))$ for $s,x \in \mathbb{R}^d$. The (non-local) PDE \eqref{eq:PIDEs2} is the Kolmogorov PDE for the exponential L\'evy model associated to $L^d$, see Section~\ref{subsec:ApproxLevy} for further interpretation and a discussion on the relation to option pricing and the assumption on $\nu^d_\mathrm{L}$. If $\nu^d_\mathrm{L}=0$, then \eqref{eq:PIDEs2} is the Black-Scholes PDE. Let $T>0$ and suppose we are given i.i.d.\ $\mathbb{R}^d\times \mathbb{R}$-valued random variables $(X_1^d,Y_1^d),(X_2^d,Y_2^d)$, $\ldots$ with the property that \begin{equation}\label{eq:regressionFunction2} u_d(T,\exp(x)) = \mathbb{E}[Y_1^d | X_1^d=x], \end{equation} for $(\P \circ (X_1^d)^{-1})$-a.e.\ $x \in \mathbb{R}^d$, that is, $u_d(T,\exp(\cdot))$ is the regression function. We are interested in learning $u_d(T,\cdot)$ on the set $\mathcal{D}^d = \{\exp(x) \,|\, x \in [-M,M]^d \} \subset (0,\infty)^d$. This encompasses two particularly relevant situations. \begin{example} Suppose that the solution $u_d(T,\cdot)$ of the PDE can be observed at $n$ points $\exp(X_1^d),\ldots,\exp(X_n^d)$. The observations are not perfect, but perturbed by some additive noise. The goal is to learn the solution of the PDE on the entire set $\mathcal{D}^d$ from these noisy observations. This situation is captured in our setting with $Y_i^d = u_d(T,\exp(X_i^d)) + \varepsilon_i^d$ for $i=1,\ldots,n$, where $\varepsilon_1^d,\ldots,\varepsilon_n^d$ are i.i.d.\ random variables independent of $X_1^d,\ldots,X^d_n$. \end{example} \begin{example} A different situation of interest arises when neural networks are employed as a \textit{solution method} for the PDE \eqref{eq:PIDEs2} in the way proposed in \cite{BernerGrohsJentzen2018} for a related setting. Let $X_1^d,\ldots,X_n^d$ be i.i.d.\ random variables uniformly distributed on $[-M,M]^d$ and independent of $L^d$ and let $Y_i^d = \varphi_d(\exp(X_i^d + L^d_T))$ for $i=1,\ldots,n$. Then one may show using the Feynman-Kac formula (see Proposition~\ref{prop:FeynmanKac}) that \[ u_d(T,\exp(x)) = \mathbb{E}[\varphi_d(\exp(x+L^d_T))] = \mathbb{E}[\varphi_d(\exp(X_1^d +L^d_T))|X_1^d=x] = \mathbb{E}[Y_1^d|X_1^d=x] \] for $(\P \circ (X_1^d)^{-1})$-a.e.\ $x \in \mathbb{R}^d$ and hence $u_d(T,\exp(\cdot))$ is indeed the regression function \eqref{eq:regressionFunction2}. Thus, in this situation we have formulated the problem of solving the PDE \eqref{eq:PIDEs2} on $\mathcal{D}^d$ as a statistical learning problem with data points $(X_i,\varphi_d(\exp(X_i^d + L^d_T))) $, $i=1,\ldots,n$. \end{example} In order to learn the unknown function $u_d(T,\cdot)$ from the data $D_n^d=((X_1^d,Y_1^d),\ldots,(X_n^d,Y_n^d))$ we employ a random neural network. Recall from Section~\ref{sec:RandomNN} that a random neural network is a single-hidden-layer feedforward neural network in which the hidden weights are randomly generated and then considered fixed and only the output-layer weight vector can be trained. The weights of the random neural networks are generated as follows: let $\nu>4$, for each $d \in \mathbb{N}$ let $A^d_1,A_2^d,\ldots$ be i.i.d.\ $\mathbb{R}^d$-valued random vectors and let $B_1,B_2,\ldots$ be i.i.d.\ random variables. Assume that $A^d_1$ is $t_{\nu}(0,\mathbbm{1}_d)$-distributed and $B_1$ has a strictly positive Lebesgue-density $\pi_{\text{b}}$ of at most polynomial decay (see \eqref{eq:polyTails1D}). For each $d,n \in \mathbb{N}$ we assume that $\{A^d_i\}_{i \in \mathbb{N}}$, $\{B_i\}_{i \in \mathbb{N}}$ and $D_n^d$ are independent. For $d, N \in \mathbb{N}$ we write $A^{d,N}=(A^d_1,\ldots,A^d_N)$ and $B^N = (B_1,\ldots,B_N)$. If $N$ hidden nodes are used, the random neural network employed for learning is then given by \begin{equation}\label{eq:RandomNNLevy} H^{A^{d,N},B^N}_{W}(x)= \sum_{i=1}^N W_{i} \varrho(A_i^d \cdot x + B_i), \quad x \in \mathbb{R}^d, \end{equation} where $W$ is an $\mathbb{R}^N$-valued, $\sigma(A^{d,N},B^{N},D_n^d)$-measurable random vector which needs to be chosen. The (squared) learning error (or prediction error) is given by \begin{equation} \label{eq:testErrorLevy} \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - H^{A^{d,N},B^N}_{W}(\bar{X}^d) |^2], \end{equation} where $(\bar{X}^d,\bar{Y}^d)$ has the same distribution as $(X_1^d,Y_1^d)$ and is independent of $\{(A^d_i,B_i)\}_{i \in \mathbb{N}}$ and $D_n^d$. Learning $u_d(T,\cdot)$ by $H^{A^{d,N},B^N}_{W}$ then amounts to selecting an $\mathbb{R}^N$-valued (random) vector $W$ that minimizes the prediction error. $W$ may be chosen depending on the random weights $A^{d,N},B^N$ and the data $D_n^d = ((X_1^d,Y_1^d),\ldots,(X_n^d,Y_n^d))$. We consider three choices: \begin{itemize} \item $W$ is chosen as $\widehat{W}^{d,N,n}$, where \begin{equation} \label{eq:ERMLevy} \widehat{W}^{d,N,n} = \arg \min_{W \in \mathcal{W}^{d,N,n}} \left\lbrace \frac{1}{n} \sum_{i=1}^n (H^{A^{d,N},B^N}_{W}(X_i^d) - Y_i^d)^2 \right\rbrace \end{equation} for $\mathcal{W}^{d,N,n} = \{W \colon \Omega \to \mathbb{R}^N \, | \, W \text{ is } \sigma(A^{d,N},B^N,D_n^d)\text{-measurable}\}$. Note that $\widehat{W}^{d,N,n}$ can be calculated explicitly by solving a system of linear equations (see Section~\ref{subsec:Regression}). \item $W$ is chosen as $\widehat{W}_\lambda^{d,N,n}$, where \begin{equation} \label{eq:ConstrainedRegressionLevy} \widehat{W}_\lambda^{d,N,n} = \arg \min_{W \in \mathcal{W}_\lambda^{d,N,n}} \left\lbrace \frac{1}{n} \sum_{i=1}^n (H^{A^{d,N},B^N}_{W}(X_i^d) - Y_i^d)^2 \right\rbrace \end{equation} for $\mathcal{W}_\lambda^{d,N,n} = \{W \in \mathcal{W}^{d,N,n} \,|\, \|W\| \leq \lambda \text{ $\P$-a.s.} \}$. Recall that $\widehat{W}_\lambda^{d,N,n}$ can be calculated explicitly by solving a system of linear equations (see Section~\ref{subsec:RidgeRegression}). \item $W$ is chosen as $W_{\mathcal{T}}^{d,N,n}$, where $W_{\mathcal{T}}^{d,N,n}$ is computed using the stochastic gradient descent algorithm as introduced in Section~\ref{subsec:SGD}. \end{itemize} \begin{remark} As pointed out above, training of random neural networks can be performed by solving a system of linear equations (see \eqref{eq:OLSSol} in Section~\ref{subsec:Regression} and \eqref{eq:ConstrainedSol} in Section~\ref{subsec:RidgeRegression}). There may nevertheless be situations in which one is interested in training a random neural network using a stochastic gradient descent method (e.g.\ a performance comparison in an experiment). This is the reason why we also analyze optimization by stochastic gradient descent here. \end{remark} \subsection{Learning error bounds}\label{subsec:learningResultsLevy} With these preparations (see Section~\ref{subsec:learningProblemLevy}) we now use the results from Sections~\ref{sec:Approx} and \ref{sec:Learning} to prove that $u_d(T,\cdot)$ can be learnt using random neural networks without the curse of dimensionality. \begin{corollary}\label{cor:LevyLearning} Let $p\geq0$, $c, L, M, \eta_0>0$, $C > \frac{1}{2^{3/2} T \pi}$. Assume that for each $d \in \mathbb{N}$ the payoff function satisfies $\varphi_d \circ \exp \in L^1(\mathbb{R}^d)$ and $\|\varphi_d \circ \exp \|_{L^1(\mathbb{R}^d)} \leq c d^p$, the characteristic triplet $(\Sigma^d,\gamma^d,\nu^d_\mathrm{L})$ of the L\'evy process $L^d$ satisfies for all $\xi \in \mathbb{R}^d$ \begin{equation} \label{eq:Ccond2} \frac{1}{2} \xi \cdot \Sigma^d \xi \geq C \| \xi\|^2, \end{equation} assume that $\|X_1^d\|_\infty \leq M$, $\P$-a.s.\ and suppose $u_d \in C^{1,2}((0,T] \times (0,\infty)^d) \cap C([0,T]\times (0,\infty)^d)$ is an at most polynomially growing solution to the PDE \eqref{eq:PIDEs2}. \begin{itemize} \item[(i)] Assume for all $d \in \mathbb{N}$ that $\sigma_d^2=\sup_{x \in \mathbb{R}^d} \mathbb{E}[(Y_1^d-u_d(T,\exp(X_1^d)))^2|X_1^d=x] \leq c d^p$ and $|u_d(T,s)| \leq L$ for all $s \in (0,\infty)^d$. Then there exist constants $C_0,\mathfrak{p}>0$ such that for any $d,N,n \in \mathbb{N}$ the prediction error of random neural network regression satisfies \begin{equation} \label{eq:fullError1Levy} \begin{aligned} & \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - T_L(H^{A^{d,N},B^N}_{\widehat{W}^{d,N,n}}(\bar{X}^d)) |^2]^{1/2} \leq C_0 d^{\mathfrak{p}} \left( \frac{ (\log(n)+1)^{1/2}\sqrt{N}}{\sqrt{n}} + \frac{1}{\sqrt{N}} \right). \end{aligned} \end{equation} \item[(ii)] Assume for all $d \in \mathbb{N}$ that $\mathbb{E}[|Y_1^d|^4]\leq c d^p$. Then there exist $\underline{p},\underline{c} >0 $ such that for any $\overline{p} > \underline{p}$, $\overline{c} > \underline{c}$ there exist $ C_0,\mathfrak{p}>0$ such that for any $d, N,n \in \mathbb{N}$ the random neural network trained by constrained regression with parameter $\lambda \in \frac{1}{\sqrt{N}} [\underline{c} d^{\underline{p}}, \overline{c} d^{\overline{p}}]$ satisfies \begin{equation} \label{eq:fullError2Levy} \begin{aligned} & \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - H^{A^{d,N},B^N}_{\widehat{W}^{d,N,n}_\lambda}(\bar{X}^d) |^2]^{1/2} \leq C_0 d^{\mathfrak{p}} \left( \frac{1}{\sqrt{N}} + \frac{1} {n^{\frac{1}{4}}}\right). \end{aligned} \end{equation} \item[(iii)] Consider the same situation as in (ii). Then, in addition, there exist constants $C_1,\mathfrak{q}>0$ such that for any $d, N, n, \mathcal{T} \in \mathbb{N}$ the random neural network trained by stochastic gradient descent for $\mathcal{T}$ steps with learning rate $\eta_t = \eta_0 t^{-1/2}$ for $t=1,\ldots,\mathcal{T}-1$ and with $\lambda $ as in (ii) satisfies \begin{equation} \label{eq:fullError3Levy} \begin{aligned} \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - H^{A^{d,N},B^N}_{W_{\mathcal{T}}^{d,N,n}}(\bar{X}^d) |^2]^{1/2} & \leq C_1 d^{\mathfrak{q}} \left( \frac{1}{\sqrt{N}} + \frac{1} {n^{\frac{1}{4}}} + \frac{N (2+\log(\mathcal{T}))^{\frac{1}{2}}}{\mathcal{T}^{\frac{1}{4}}} \right) . \end{aligned} \end{equation} \end{itemize} \end{corollary} \begin{remark} Each of these statements can be translated directly into a statement on the number of samples and hidden nodes required to guarantee a prescribed learning error of precision at most $\varepsilon > 0$. For instance, in the case of regression (corresponding to the bound \eqref{eq:fullError1Levy}) we see that there exist constants $\tilde{C}_0,\tilde{\mathfrak{p}}>0$ such that for all $d \in \mathbb{N}$, $\varepsilon >0$ at most $N\leq \tilde{C}_0 d^{\tilde{\mathfrak{p}}} \varepsilon^{-2}$ weights and $n\leq \tilde{C}_0 d^{\tilde{\mathfrak{p}}} \varepsilon^{-8}$ samples suffice to guarantee \begin{equation} \label{eq:fullError1LevyTranslated} \begin{aligned} & \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - T_L(H^{A^{d,N},B^N}_{\widehat{W}^{d,N,n}}(\bar{X}^d)) |^2]^{1/2} \leq \varepsilon. \end{aligned} \end{equation} This follows from \eqref{eq:fullError1Levy} by choosing $N=4C_0^2 d^{2\mathfrak{p}}\varepsilon^{-2}$, $n = 16c^2 C_0^4 d^{4\mathfrak{p}}\varepsilon^{-4} N^2$ and $\tilde{C}_0=\max(4C_0^2,256c^2 C_0^8),\tilde{\mathfrak{p}} = 8\mathfrak{p}$ where $c$ is a constant such that $\log(m)+1 \leq c \sqrt{m}$ for all $m \in \mathbb{N}$. \end{remark} \begin{proof} For fixed $d\in \mathbb{N}$ let $\Phi(x) = \varphi_d(\exp(x))$ and $H(x)= u_d(T,\exp(x))$ for $x \in \mathbb{R}^d$. Then Proposition~\ref{prop:FeynmanKac} shows that $H(x)= \mathbb{E}[\Phi(x+L_T^d)]$ and, as argued in the proof of Theorem~\ref{thm:LevyApprox}, the characteristic function of $L_T^d$ satisfies the bound \eqref{eq:auxEq8}. \textit{Proof of (i):} Theorem~\ref{thm:TrainingErrRegression} hence implies that there exist $k \in \mathbb{N}$ and $\tilde{C}_{\text{app}}>0$ such that \begin{equation} \label{eq:auxEq29} \begin{aligned} & \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - T_L(H^{A^{d,N},B^N}_{\widehat{W}^{d,N,n}}(\bar{X}^d)) |^2]^{1/2} \\ & \leq \tilde{C}_{\text{app}} \max(\sigma_d^2,L) \frac{(\log(n)+1)^{1/2}\sqrt{N}}{\sqrt{n}} + \frac{\tilde{C}_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}} \\ & \leq C_0 d^{\mathfrak{p}} \left( \frac{ (\log(n)+1)^{1/2}\sqrt{N}}{\sqrt{n}} + \frac{1}{\sqrt{N}} \right) \end{aligned} \end{equation} with $C_0 = \tilde{C}_{\text{app}} \max(\max(c,L),c(2 \nu)^{k+3})$ and $\mathfrak{p} = p+ k+3$. This proves (i), since $k$ and $\tilde{C}_{\text{app}}$ in Theorem~\ref{thm:TrainingErrRegression} do not depend on $d$, $n$ or $N$. \textit{Proof of (ii):} Let $k \in \mathbb{N}$ and $C_{\text{app}},C_{\text{wgt}}>0$ be as in Theorem~\ref{thm:ApproxError}, choose $\underline{p} = 2 k + \frac{1}{2} + p$, $\underline{c} =C_{\text{wgt}}c(2\nu)^{2k+\frac{1}{2}} $ and let $\overline{p} > \underline{p}$, $\overline{c} > \underline{c}$. Then $\lambda \in \frac{1}{\sqrt{N}} [\underline{c} d^{\underline{p}}, \overline{c} d^{\overline{p}}]$ satisfies $ \frac{1}{\sqrt{N}} C_{\text{wgt}}\|\Phi\|_{L^1(\mathbb{R}^d)}(\nu+d)^{2k+\frac{1}{2}} \leq \lambda $ and hence Theorem~\ref{thm:TrainingErrConstrRegression} shows that there exists $C_{\text{est}}>0$ such that \begin{equation} \label{eq:auxEq32} \begin{aligned} & \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - H^{A^{d,N},B^N}_{\widehat{W}^{d,N,n}_\lambda}(\bar{X}^d) |^2]^{1/2} \leq \frac{C_{\text{app}} \|\Phi\|_{L^1(\mathbb{R}^d)} (\nu+d)^{k+3}}{\sqrt{N}} + \frac{C_{\text{est}} d^{\overline{p}+1} }{n^{\frac{1}{4}}}. \end{aligned} \end{equation} From the proof of Theorem~\ref{thm:TrainingErrConstrRegression} (with $C_{\text{lam}} = \overline{c}$ here) the constant $C_{\text{est}}$ is given by \[ \begin{aligned} C_{\text{est}}^2 & = 8\overline{c}^2 (\frac{\nu M^2} {\nu-2}+ \mathbb{E}[|B_1|^2]) +2^{3+\frac{1}{2}} \overline{c}(\frac{\nu M^2} {\nu-2}\mathbb{E}[(Y_1^d)^2] + \mathbb{E}[|B_1|^2] \mathbb{E}[(Y_1^d)^2])^{1/2} + 4 \mathbb{E}[(Y_1^d)^4]^{1/2} \end{aligned} \] and hence $C_{\text{est}} \leq d^{\frac{p}{4}} \tilde{C}_{\text{est}}$ with $\tilde{C}_{\text{est}}^2 = 8\overline{c}^2 (\frac{\nu M^2} {\nu-2}+ \mathbb{E}[|B_1|^2]) +2^{3+\frac{1}{2}} \overline{c} c^{\frac{1}{4}}(\frac{\nu M^2} {\nu-2} + \mathbb{E}[|B_1|^2])^{1/2} + 4 c^{\frac{1}{2}}$. Thus, \eqref{eq:auxEq32} yields \begin{equation} \label{eq:auxEq33} \begin{aligned} & \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - H^{A^{d,N},B^N}_{\widehat{W}^{d,N,n}_\lambda}(\bar{X}^d) |^2]^{1/2} \leq C_0 d^{\mathfrak{p}} \left( \frac{1}{\sqrt{N}} + \frac{1 }{n^{\frac{1}{4}}}\right) \end{aligned} \end{equation} with $C_0 = \max(C_{\text{app}} c (2\nu)^{k+3} ,\tilde{C}_{\text{est}})$ and $\mathfrak{p} = \max(p+k+3,\overline{p}+1+\frac{p}{4})$. As shown in the above results (and visible from the explicit expressions available for these constants) neither $k \in \mathbb{N}$ nor the constants $C_{\text{app}},C_{\text{wgt}}>0$ depend on $d$, $n$ or $N$. Hence, the constants $C_0, \mathfrak{p}$ do not depend on $d,N,n$ or $\lambda$. This proves (ii). \textit{Proof of (iii):} Let $k \in \mathbb{N}$, $C_{\text{app}},C_{\text{wgt}}$, $\underline{p} $, $\underline{c},C_0,\mathfrak{p} >0$ be as in the proof of (ii) and let $\overline{p} > \underline{p}$, $\overline{c} > \underline{c}$. Applying Proposition~\ref{prop:SGDtrained} with $C_{\text{lam}} = \overline{c}$ and using the estimate provided in the proof of (ii) (see \eqref{eq:auxEq32} and \eqref{eq:auxEq33}) for the first two terms in \eqref{eq:fullError3} we obtain that there exists $C_{\text{opt}}>0$ such that \begin{equation} \label{eq:auxEq34} \begin{aligned} \mathbb{E}[|u_d(T,\exp(\bar{X}^d)) - H^{A^{d,N},B^N}_{W_{\mathcal{T}}^{d,N,n}}(\bar{X}^d) |^2]^{1/2} & \leq C_0 d^{\mathfrak{p}} \left( \frac{1}{N^{\frac{1}{2}}} + \frac{1 }{n^{\frac{1}{4}}}\right) + \frac{C_{\text{opt}} d^{\overline{p}+2} N (2+\log(\mathcal{T}))^{\frac{1}{2}}}{\mathcal{T}^{\frac{1}{4}}}. \end{aligned} \end{equation} The constant $C_{\text{opt}}$ was given explicitly in the proof and we deduce that $C_{\text{opt}} \leq d^{p/4} \tilde{C}_{\text{opt}} $ with $\tilde{C}_{\text{opt}}^2 = 4 \max(\frac{4}{\eta_0},96\eta_0)\max(2,3M^4 \nu^2/[(\nu-2)(\nu-4)] +\mathbb{E}[|B_1|^4]) \max(\overline{c}^2,c^{1/2}) $. Combining this with \eqref{eq:auxEq34} proves \eqref{eq:fullError3Levy} with $C_1 = \max(C_0,\tilde{C}_{\text{opt}})$, $\mathfrak{q} = \max(\mathfrak{p}, \overline{p}+2+ \frac{p}{4})$. By the same reasoning as above $C_1, \mathfrak{q}$ do not depend on $d,N,n,\mathcal{T}$ or $\lambda$. This proves (iii). \end{proof} \subsection{Numerical example} \label{sec:numerics} In this section we consider a numerical example in which the solution $u_d(T,\cdot)$ to \eqref{eq:PIDEs2} is learnt from noisy observations. We fix $d$, $T$ and generate $n$ training data points $(X_1^d,Y_1^d),\ldots,(X_n^d,Y_n^d)$ for our experiment. The goal is then to learn $u_d(T,\cdot)$ based only on these data points, i.e.\ without using any knowledge about the underlying PDE or its parameters. This is achieved by employing neural networks with randomly generated hidden weights, as explained in detail in Section~\ref{subsec:learningProblemLevy}. For the unknown PDE we choose the pricing PDE for a max-call option in a $d$-dimensional Black-Scholes model with equal correlations among the assets. Thus, we fix $d=50$, choose $\varphi_d(s) = \max(\max(s_1,\ldots,s_d)-K,0)$ as initial value for the PDE and let $\Sigma^d$ be given for $i,j=1,\ldots,d$ by $\Sigma^d_{i,j} = \sigma^2 \rho $ for $i \neq j$ and $\Sigma^d_{i,i} = \sigma^2 $. Furthermore, $\tilde{\gamma}^d = 0$, $ \nu^d_\mathrm{L} = 0$ and the parameter values are chosen as $\sigma=0.2$, $\rho = 0.2$, $T=1$. The strike $K$ is chosen as $K=1$ (which corresponds to expressing prices in units of the ``actual'' strike). From the solution $u_d(T,\cdot)$ with $K=1$ on $\mathcal{D}^d$ one can also directly obtain the solution $\tilde{u}_d(T,\cdot)$ for other values of $K$ (e.g.\ $K=100$) on the set $\{K \exp(x) \,|\, x \in [-M,M]^d \}$ by using $\tilde{u}_d(T,s) = K u_d(T,s/K)$. For our experiment we now select $M=1$ and generate the $i$-th data point as follows: we randomly uniformly sample $X_i^d$ on $[-1,1]^d$ and then use a Monte Carlo simulation with $5\cdot 10^6$ sample paths to calculate an approximate value of $u_d(T,\exp(X_i^d))$. $Y_i^d$ is then defined as this approximate value and corresponds to a noisy observation of $u_d(T,\cdot)$ at $\exp(X_i^d)$. By using this procedure for $i=1,\ldots, n$ we generate $n= 5\cdot 10^6$ data points (the training data). The goal is now to learn the solution $u_d(T,\cdot)$ to \eqref{eq:PIDEs2} based only on these (noisy) observations. To achieve this we use random neural networks as described in Section~\ref{subsec:learningProblemLevy}. We consider different choices for the number of hidden nodes $N$. For the weight distributions we choose $A_1^d \sim t_5(0,\mathbbm{1}_d)$ and let $B_1$ have a Student's $t$-distribution with $2$ degrees of freedom (i.e. $\nu=5$ and $\pi_{\text{b}}$ is the density of a $t$-distribution with $2$ degrees of freedom). Unconstrained regression is employed to fit the output weights (see \eqref{eq:ERMLevy}), resulting in an output weight vector $\widehat{W}^{d,N,n}$ and a random neural network approximation $H^{A^{d,N},B^N}_{\widehat{W}^{d,N,n}}$ (see \eqref{eq:RandomNNLevy}) to $u_d(T,\exp(\cdot))$. Then we generate $n_{\mathrm{test}} = 5 \cdot 10^5$ test samples $(\bar{X}^d_1,\bar{Y}^d_1),\ldots,(\bar{X}^d_{n_\mathrm{test}},\bar{Y}^d_{n_\mathrm{test}})$ according to the same procedure that we used for the training data above. Based on these training data points we calculate the squared error $\hat{e}^2 = \frac{1}{n_{\mathrm{test}}} \sum_{i=1}^{n_{\mathrm{test}}} (\bar{Y}^d_i - H^{A^{d,N},B^N}_{\widehat{W}^{d,N,n}}(\bar{X}^d_i))^2$. The error $\hat{e}$ is an estimate of the prediction error (see \eqref{eq:testErrorLevy}, \eqref{eq:fullError1Levy} and recall that the Monte Carlo price $\bar{Y}^d_i$ is an unbiased estimate of $u_d(T,\exp(\bar{X}^d_i))$). Figure~\ref{plot} displays $\hat{e} = \hat{e}(N)$ for different choices of the number of hidden nodes, namely, $N \in \{1\} \cup \{10,20,\ldots,190\}$. The figure also displays the function $x \mapsto \frac{e_0}{\sqrt{x}}$, where $e_0$ is chosen as $\hat{e}(1)$. The theoretical results from Corollary~\ref{cor:LevyLearning} show that, for $n$ large, the theoretical prediction error decays at least as $1/\sqrt{N}$ when $N$ increases. The numerical results here reproduce this behaviour for the estimated prediction error $\hat{e}(N)$. This can be seen from Figure~\ref{plot}, where the estimated error $\hat{e}(N)$ matches closely the function $x \mapsto \frac{e_0}{\sqrt{x}}$. \begin{figure}[h] \centering \includegraphics[width=0.95\textwidth]{plot.pdf} \caption{Plot of the estimated learning error committed when a random neural network with $N$ hidden nodes is used to learn a $50$-dimensional Black-Scholes PDE from observations. The dots show the estimated learning error $\hat{e}(N)$ for different values of $N$, the line shows the decay implied by the theoretical results $\frac{e_0}{\sqrt{N}}$ (with $e_0$ chosen as $\hat{e}(1)$).} \label{plot} \end{figure} This numerical experiment also indicates that the integrability and smoothness assumptions in Corollary~\ref{cor:LevyLearning} can potentially be relaxed. More specifically, the payoff $\varphi_d$ considered in the example here does not satisfy the hypothesis $\varphi_d \circ \exp \in L^1(\mathbb{R}^d)$ and for the chosen parameters the matrix $\Sigma^d$ does not satisfy \eqref{eq:Ccond2}, since the smallest eigenvalue of $\frac{1}{2}\Sigma^d$ is smaller than $\frac{1}{2^{3/2} T \pi}$ and hence any eigenvector $\xi$ of $\frac{1}{2}\Sigma^d$ corresponding to this eigenvalue satisfies $\frac{1}{2} \xi \cdot \Sigma^d \xi < C \| \xi\|^2$ for any $C > \frac{1}{2^{3/2} T \pi}$. Nevertheless, the numerical results suggest that Corollary~\ref{cor:LevyLearning}(i) is still valid in this situation. While Theorem~\ref{thm:RC12Linfty} may be used to establish the $N^{-1/2}$-decay in $N$ also without the hypotheses $\varphi_d \circ \exp \in L^1(\mathbb{R}^d)$ and \eqref{eq:Ccond2}, these hypotheses were needed in the proof of Theorem~\ref{thm:ApproxError} (and propagate to Corollary~\ref{cor:LevyLearning}) in order to guarantee that the constant in the error bound does not grow exponentially in $d$. The numerical experiment and the choice $d=50$ indicates non-exponential constants also here and hence it may be possible to relax these assumptions by taking a different approach than the one that was used in the proof of Theorem~\ref{thm:ApproxError}. {\small \bibliographystyle{abbrvnat}

https://arxiv.org/abs/2112.00126

Martingale product estimators for sensitivity analysis in computational statistical physics

We introduce a new class of estimators for the linear response of steady states of stochastic dynamics. We generalize the likelihood ratio approach and formulate the linear response as a product of two martingales, hence the name "martingale product estimators". We present a systematic derivation of the martingale product estimator, and show how to construct such estimator so its bias is consistent with the weak order of the numerical scheme that approximates the underlying stochastic differential equation. Motivated by the estimation of transport properties in molecular systems, we present a rigorous numerical analysis of the bias and variance for these new estimators in the case of Langevin dynamics. We prove that the variance is uniformly bounded in time and derive a specific form of the estimator for second-order splitting schemes for Langevin dynamics. For comparison, we also study the bias and variance of a Green-Kubo estimator, motivated, in part, by its variance growing linearly in time. Presented analysis shows that the new martingale product estimators, having uniformly bounded variance in time, offer a competitive alternative to the traditional Green-Kubo estimator. We compare on illustrative numerical tests the new estimators with results obtained by the Green-Kubo method.

\section{Introduction} We consider the linear response at infinite time ($t\to\infty$) to a perturbation in the drift vector field of an ergodic stochastic dynamics defined by a system of stochastic differential equations \begin{equation}\label{eq:dynamics} dy_t^\eta = (b(y_t^\eta)+\eta F(y_t^\eta)) dt + \sigma(y_t^\eta) dW_t\,{\,,} \end{equation} where~$\eta \in \mathbb{R}$ is a small parameter. Under the assumption that both the unperturbed ($\eta=0$) and perturbed system are ergodic with respect to probability measures $\pi$ and $\pi^\eta$, respectively, the central object of the numerical analysis we present is the estimation of the quantity \begin{equation}\label{eq:linresponse} \left.\frac{d}{d\eta}\right|_{\eta=0}\pi^{\eta}(f) \equiv \lim_{\eta \to 0} \frac{1}{\eta}\left[\int f(y) \,\pi^\eta(dy) - \int f(y)\,\pi(dy)\right]\,, \end{equation} for a suitable class of functions $f$, termed observables. The well posedness of this limit is guaranteed under certain assumptions on the dynamics, see for instance~\cite{HM10,lelievre2016partial}. Our motivation for developing an efficient numerical estimation of~\eqref{eq:linresponse} is the computation of transport coefficients in molecular systems at thermal equilibrium in the canonical ensemble. Accurately estimating transport properties such as the mobility, the shear stress or the thermal conductivity, presents computational challenges in spite of algorithmic developments in computational methods for statistical mechanics of fluids and gasses, see, \emph{e.g.}, \cite{evans-moriss,tuckerman2010statistical} for physical background, \cite[Section~5]{lelievre2016partial} for mathematical background and analysis, and \cite{leimkuhler2016computation} for elements of numerical analysis. We present in Section~\ref{sec:general-continuous-setting} a strategy for deriving a new type of estimators for the problem \eqref{eq:linresponse} under the dynamics~\eqref{eq:dynamics}. A systematic but formal derivation is followed by rigorous numerical analysis and results in Sections~\ref{sec:MP_for_Langevin} and~\ref{sec:LR-vs-GK}, focusing specifically on the Langevin dynamics \begin{equation}\label{eq:Langevinintro} \left\{ \begin{aligned} dq^\eta_t &= M^{-1} p^\eta_t \, ,\\ dp^\eta_t &= \Big(-\nabla V(q^\eta_t)+\eta F(q^\eta_t,p^\eta_t)\Big)dt -\gamma M^{-1} p^\eta_t dt + \sqrt{\frac{2\gamma}{\beta}} dW_t\,, \end{aligned} \right. \end{equation} which describes particles with positions~$q$ and momenta~$p$ at a fixed inverse temperature $\beta$. The particle interaction is defined by the potential $V(q)$ and $M$ denotes the mass tensor. The specific form of the magnitude~$\sqrt{2\gamma\beta^{-1}} dW_t$ of the standard Wiener process~$W_t$ ensures that the Boltzmann--Gibbs probability measure $\pi(dq\,dp) = Z^{-1} \mathrm{e}^{-\beta(p^T M^{-1} p/2 + V(q))} \,dq\,dp$ is invariant for the unperturbed system at~$\eta=0$. The perturbing field $F$ is {\it not} assumed to be the gradient of a potential, hence the corresponding invariant measure $\pi^\eta$ of the perturbed dynamics is unknown in general, and corresponds to what is called a {\it non-equilibrium steady state}. As we shall see in Section~\ref{sec:general-continuous-setting} the proposed estimators do not require simulations of the perturbed dynamics. Instead the estimator is based on certain ergodic averages on trajectories of the {\it unperturbed dynamics} ($\eta=0$). Let us however emphasize that, although we consider as a reference dynamics a Langevin dynamics where the force derives from a gradient, our results can be straightforwardly extended to the situation when~$-\nabla V(q)$ is replaced by a generic (non-gradient) force field, in which case the reference probability measure~$\pi$ is not known. Standard methods to estimate transport coefficients are reviewed in \cite[Section~5]{lelievre2016partial}. There are essentially two approaches. The first one is to numerically approximate the linear response, which is however computationally quite expensive since averages with respect to~$\pi^\eta$ have to be computed by time averages over very long times in order for the statistical error to scale as~$\eta^{-2}$, so that the limit in~\eqref{eq:linresponse} is sufficiently well approximated. The second popular option is to rely on Green--Kubo formulas, which rewrite the linear response as some integrated correlation function for the equilibrium dynamics. Properly estimating this time integral is however challenging~\cite{dSOG17,EMB17} because the correlation function is very difficult to reliably estimate for long times since it is a small quantity plagued by a large relative statistical error. For these two standard approaches, the numerical analysis of the bias arising from the timestep discretization of the underlying SDE can be read in~\cite{leimkuhler2016computation,lelievre2016partial}. An alternative to the GK estimator relies on the so-called likelihood ratio method, which is based on the Girsanov theorem. More precisely, the estimator which is suggested in this approach is constructed by linearizing the Girsanov weight considered on a finite time interval, and passing to the infinite time limit. It can be shown that this indeed provides a consistent estimator of the linear response~\eqref{eq:linresponse}. This approach was employed in~\cite{glynn2019likelihood} for discrete time Markov chains and in~\cite{wang2019steady} for continuous time Markov chains, resulting in likelihood ratio type estimators with a variance uniformly bounded in time. The derived low-variance modification of the likelihood ratio (LR) method was further extended to the linear response of the dynamics described by stochastic differential equations in~\cite{plechac2019convergence}. However, the approach developed in \cite{plechac2019convergence} did not provide a systematic way to derive estimators of order higher than~1 when discretizing the underlying dynamics. Our aim in this work is to generalize the likelihood ratio approach and derive a new class of estimators that we call {\it martingale product} estimators (MP). The MP estimators reformulate the linear response as an expectation of the product of two martingales. The analysis of the proposed approach also provides a systematic way to obtain estimators whose time-step bias is consistent with the weak order of the underlying stochastic dynamics. In particular, we present estimators that lead to biases of order~2 in the time-step for weak second order discretizations of Langevin dynamics~\eqref{eq:Langevinintro} based on splitting schemes~\cite{leimkuhler2013rational}. Furthermore, we prove that the variance of the MP estimators is uniformly bounded in time, as for CLR estimators, and therefore provide a competitive alternative to Green-Kubo estimators for which the variance grows linearly in time. We more precisely compare the new MP estimators and the Green-Kubo method in Section~\ref{sec:LR-vs-GK} from a mathematical viewpoint, and in Section~\ref{sec:numerics} from a numerical perspective on toy examples. Let us however immediately emphasize that it is not our aim in this study to obtain strong conclusions on the actual relative numerical performances of these two methods. The manuscript is organized as follows. We start by presenting in Section~\ref{sec:general-continuous-setting} the heuristic strategy for constructing MP estimators for general stochastic dynamics. In order to make this strategy precise and rigorous for Langevin dynamics, we review some properties of this dynamics and its discretization by splitting schemes in Section~\ref{sec:LD-application}. We can then perform the numerical analysis of the bias and variance of MP estimators for Langevin dynamics in Section~\ref{sec:MP_for_Langevin} (with some technical results postponed to Appendices~\ref{app:useful_estimates} and~\ref{app:secondorder}). For comparison, we also study the bias and variance of the GK estimators in Section~\ref{sec:LR-vs-GK}. Finally, several numerical benchmarks are provided to demonstrate the consistent between our theoretical results and the numerical simulations in Section~\ref{sec:numerics}. \section{Formal construction of martingale product estimators} \label{sec:general-continuous-setting} We present in this section a general strategy for constructing the new MP estimators for the linear response of stationary measures of stochastic dynamics. We first recall in Section~\ref{sec:LR_stoch_dyn} the mathematical framework for the linear response of stationary measures of general diffusion processes. In particular, we express the response coefficients in terms of solutions to Poisson equations and their computation in terms of an auxiliary martingale~\cite{plechac2019convergence}. We then present in Section~\ref{subsec:formal-derivation} a general form of the numerical estimators we consider when the underlying continuous diffusion process is discretized in time. Finally, in order to reduce the bias induced by the time discretization on the estimation of the linear response, the key point in our approach is to construct discrete auxiliary martingales tailored to the weak order of the numerical scheme at hand, as formally discussed in Section~\ref{sec:construction_discrete_martingale} for schemes of both first and second weak order. \subsection{Linear response of stochastic dynamics} \label{sec:LR_stoch_dyn} \subsubsection{Reference dynamics.} Given a filtered probability space $(\Omega, \mathcal{F}, \mathcal{F}_t, \bP)$, we denote by $W_t = (W_{1,t},\dots, W_{k,t})^{\T}$ a $k$-dimensional $\{\mathcal{F}_t\}_{t>0}$-adapted standard Wiener process. We study certain perturbations of a $d$-dimensional time homogeneous stochastic differential equation (SDE) on the state space~$\mY$ (for example, $\mY = \mathbb{R}^d$ or~$\mathbb{T}^d$ with $\mathbb{T}=\mathbb{R}\backslash \mathbb{Z}$ the one-dimensional torus): \begin{equation}\label{eqn:SDE} dy_t = b(y_t)\, dt + \sigma(y_t) \, dW_t{\,,} \end{equation} where $b: \mY \to \mathbb{R}^d$ is a smooth vector field and $\sigma: \mY \to \mathbb{R}^{d \times k}$ is a matrix valued function. It is implicitly assumed that $b$ and $\sigma$ satisfy all required conditions so that the above SDE is well posed~\cite{oksendal2013stochastic}. We consider the case of degenerate noise, \emph{i.e.} $k < d$, but under the assumption that the matrix field $\sigma(y) = \left(\sigma_{ij}(y)\right)_{1 \leq i \leq d, 1 \leq j \leq k}$ has full rank~$k$ for all~$y \in \mY$. Equation~\eqref{eqn:SDE} can be written component-wise as \[ dy_{i,t} = b_i(y_t)\, dt + \sum_{j=1}^k \sigma_{ij}(y_t) \, dW_{j,t}{\,,} \qquad i=1,\dots,d {\,.} \] We assume that the dynamics~\eqref{eqn:SDE} is ergodic and that its unique invariant measure~$\pi(dy)$ has a smooth positive density~$\rho(y)$ with respect to the Lebesgue measure (see~\cite{kliemann1987recurrence, bellet2006ergodic} for sufficient conditions). To write the expression of the generator associated with~\eqref{eqn:SDE}, we introduce the $d\times d$ diffusion matrix~$S = \sigma \sigma^{\T}$, and denote by~$\nabla^2 f$ the Hessian matrix with entries $\partial^2_{y_i,y_j} f$ for $1 \leq i,j \leq d$. The symbol ``$:$'' represents the Frobenius inner product of $d\times d$ matrices, \emph{i.e.}, $S:\nabla^2 f = \sum_{1 \leq i,j \leq d} S_{ij} \partial^2_{y_i,y_j} f$. We emphasize that the diffusion matrix $S(y)$ can be degenerate. With this notation the generator of~\eqref{eqn:SDE} acts on smooth test functions~$f$ as \begin{equation} \mathcal{L}f = b^{\T} \nabla f + \frac{1}{2} S : \nabla^2 f {\,.} % \end{equation} {Here and in the sequel, we adopt the following notational convention: all vectors are understood as column vectors, the gradient of a scalar function $f$ is represented by a column vector $\nabla f$ and thus $b^T\nabla f = (\nabla f)^\T b = \sum_i b_i \partial_{y_i} f $. The notation $\nabla F$ for the gradient of a vector field $F$ represents a matrix with entries $(\nabla F)_{ij} = \partial_{y_i} F_j$, i.e., the $j$th column of~$\nabla F$ is the gradient of the individual component~$F_j$ (understood as a column vector).} \subsubsection{Perturbations of the reference dynamics.} We next consider perturbations of the reference dynamics~\eqref{eqn:SDE} obtained by adding to the drift field~$b$ a vector field $F: \mY \to \mathbb{R}^d$, with a forcing magnitude~$\eta \in \mathbb{R}$. The perturbed dynamics then reads \begin{equation} \label{eqn:perturbed-SDE} dy_t^{\eta} = (b(y_t^{\eta}) + \eta F(y_t^{\eta}))\, dt + \sigma(y_t^{\eta})\, dW_t{\,.} \end{equation} Its generator can be written as \[ \mL^{\eta} = \mL + \eta \LOPERTIL{\,,} \qquad \LOPERTIL = F^{\T} \nabla{\,.} \] As for the reference dynamics~\eqref{eqn:SDE}, we assume that the perturbed dynamics~\eqref{eqn:perturbed-SDE} is ergodic and that its unique invariant measure~$\pi^\eta(dy)$ has a positive smooth density~$\rho^\eta(y)$ with respect to the Lebesgue measure. The linear response of a given observable of interest~$f$ is then defined as (provided the limit exists) \begin{equation} \label{eqn:lin-response} \mathrm{Lin} (f) = \lim_{\eta \to 0}\frac{1}{\eta}(\pi^{\eta}(f) - \pi(f)) = \lim_{\eta \to 0}\frac{1}{\eta}\int_{\mY} f(y) (\rho^{\eta}(y) - \rho(y)) \, dy{\,,} \end{equation} where for the ease of notation $\pi = \pi^0$ and $\rho=\rho^0$ denote the reference measure and reference density, respectively. We emphasize that the linear response depends both on the forcing~$F$ and the reference dynamics. We present one possible framework for justifying the well-posedness of the limit~\eqref{eqn:lin-response}, relying on linear response theory (see for instance~\cite{HM10}). The reference functional space is the Hilbert space \[ L^2(\pi) = \left\{f : \mY \to \mathbb{R}~\text{measurable}~\left|~ \int_{\mY} |f|^2 \, d\pi < \infty\right.\right\}{\,.} \] Introducing the projection operator \begin{equation} \label{eqn:projection-operator} \Pi f = f - \pi(f), \end{equation} we further define the projected subspace \[ L_0^2(\pi) = \Pi L^2(\pi) = \left\{ f \in L^2(\pi) ~\middle|~ \pi( f )= 0 \right\}, \] which consists of all $L^2(\pi)$ functions with zero mean with respect to the reference invariant measure~$\pi$. We denote by $\mL^*$ the adjoint of~$\mL$ on~$L^2(\pi)$, \emph{i.e.}, for any $C^{\infty}$ compactly supported test functions~$f, g$, \[ \int_{\mY} (\mL f) g \,d\pi = \int_{\mY} f (\mL^* g) \,d\pi{\,.} \] We next assume that the operator~$-\mL$ is invertible on~$L^2_0(\pi)$, and that it stabilizes some dense space of smooth functions which, together with their derivatives, satisfy some growth conditions (see for instance Theorem~\ref{thm:L-stability} below for Langevin dynamics). When~$\pi$ admits moments of sufficiently high order, it can then be shown (see the discussion in~\cite[Remark~5.5]{lelievre2016partial}) that the linear response~\eqref{eqn:lin-response} is well-defined and can be rewritten, for $f \in L^2(\pi)$ given, as \begin{equation} \label{eqn:lin-response-2} \mathrm{Lin} (f) = -\int_{\mY} \LOPERTIL \mL^{-1} \Pi f \, d\pi {\,.} \end{equation} The above expression serves as the starting point for various reformulations of the linear response that lead to different numerical strategies, as made precise below. These reformulations can be seen as alternatives to the straightforward method which consists in approximating~\eqref{eqn:lin-response} by estimating averages with respect to the invariant measure as time averages over one long realization of the dynamics under consideration and computing, for a given small value of~$\eta$, the estimator \[ \frac{1}{\eta} \left(\frac1t \int_0^t f(y_s^\eta)\, ds - \frac1t \int_0^t f(y_s) \, ds\right)\,. \] This approach corresponds to the so-called Nonequilibrium Molecular Dynamics approach in computational statistical physics~\cite{CKS05,evans-moriss}. When a central limit theorem applies to the time averages under consideration, it is easily seen that the latter estimator has a variance which scales as~$1/(\eta^2 t)$, so that the integration time~$t$ has to be quite large since small values of~$\eta$ are necessary to ensure that the response remains in the linear regime. This indeed motivates turning to alternative expressions of the linear response. \subsubsection{Green--Kubo method.} One popular reformulation of~\eqref{eqn:lin-response-2} is based on the operator identity \begin{equation}\label{eqn:L-inverse} -\mL^{-1} = \int_0^{\infty} \mathrm{e}^{t \mL} \, dt{\,,} \end{equation} which can be given a meaning when~$\mathrm{e}^{t \mL}$ is a bounded operator over~$L^2_0(\pi)$, with an operator norm which decays sufficiently fast with respect to~$t$. This allows us to rewrite the linear response as an integrated correlation function, the famous Green--Kubo formula (see for instance~\cite{leimkuhler2016computation, lelievre2016partial}): for $f \in L^2_0(\pi)$, \begin{equation} \label{eqn:GK} \mathrm{Lin} (f) = \int_{\mY} \left(-\mL^{-1} f\right)\left(\LOPERTIL^* \mathbf{1}\right) d\pi = \int_0^{\infty} \bE^{\pi}\left[ f(y_t) \varphi(y_0) \right] dt, \end{equation} where $\varphi = \LOPERTIL^* \mathbf{1}$, and the expectation $\bE^{\pi}$ is taken over the initial conditions distributed as $y_0 \sim \pi$ and over all realizations of the reference dynamics~\eqref{eqn:SDE}. We discuss more precisely the numerical efficiency of the Green-Kubo (GK) estimator based on~\eqref{eqn:GK} in Section~\ref{sec:LR-vs-GK} (in the context of the Langevin dynamics, but our analysis can easily be extended to other diffusion processes). \subsubsection{Likelihood ratio method.} A different approach to estimating~\eqref{eqn:lin-response-2} is based on the so-called likelihood ratio formula, obtained by re-weighting the realizations of the reference dynamics with a Girsanov weight to account for the external perturbation, and formally passing first to the small forcing limit, and then to the infinite time limit. As shown in \cite{glynn2019likelihood, plechac2019convergence, wang2019steady}, this leads for $f \in L^2_0(\pi)$ to \begin{equation} \label{eqn:LR} \mathrm{Lin} (f) = \lim_{t \to \infty} \bE^{\mu_0}\left[\left(\frac{1}{t}\int_0^t f(y_s) \, ds\right) z_t \right], \qquad z_t = \int_0^t u(y_s)^{\T} \, dW_s, \end{equation} where the function~$u:\mY \to \mathbb{R}^k$ satisfies \begin{equation} \label{eq:su=F} \sigma u = F, \end{equation} and the expectation $\bE^{\mu_0}$ is taken over initial conditions from an arbitrary initial distribution $\mu_0$, \emph{i.e.}, $y_0 \sim \mu_0$, and over all realizations of the reference dynamics~\eqref{eqn:SDE}. It is clear that~$z_t$ is a zero mean $\mathcal{F}_t$-local martingale. The main aim of this work is to propose appropriate discrete counterparts of~$z_t$, which we call {\it an auxiliary martingale process}. Another aim is to compare the estimators~\eqref{eqn:GK} and~\eqref{eqn:LR}. \subsection{Formal strategy for constructing martingale product estimators} \label{subsec:formal-derivation} Before proceeding with the rigorous analysis of the numerical methods that we propose for Langevin dynamics, we outline in this section the heuristic derivation of estimators motivated by the likelihood ratio method~\eqref{eqn:LR}. We leave the key question of the construction of the discrete counterparts to the auxiliary martingale process~$z_t$ to Section~\ref{sec:construction_discrete_martingale}. However, we warn the reader that these discrete counterparts are {\it not} based on a direct discretization of the process $z_t$, but on some discrete martingale fluctuation identity -- hence the name {\it martingale product estimators} for the estimation of the linear response. The formal arguments presented in this section are made rigorous for Langevin dynamics in Section~\ref{sec:MP_for_Langevin}. We consider numerical schemes associated with a fixed timestep~$\dt>0$, written as \begin{equation} \label{eqn:numerstep} y^{n+1} = y^n + \Phi_\dt(y^n;G^n){\,,} \end{equation} for a given increment function~$\Phi_\dt$ and random increments~$G^n$ (typically standard Gaussian random variables). We assume that this Markov chain admits a unique invariant probability measure~$\pi_\dt$. Furthermore, the chain is exponentially ergodic provided that the discrete semigroup induced by the Markov chain satisfies both a Lyapunov condition and a minorization condition~\cite{hairer2011yet, meyn2012markov,DMPS18}. The basic idea of our approach is to find a zero mean $\mathcal{F}^n$-martingale $\{z^n\}_{n\geq 0}$ re-weighting the centered ergodic average so that \begin{equation}\label{eqn:desired-estimator} \mathrm{Lin} (f) = \lim_{\dt \to 0}\lim_{N \to \infty}\bE_{\dt}\left\{\left(\frac{1}{N}\sum_{n=0}^{N-1} f(y^n) - \pi_{\dt}(f)\right) z^N\right\}{\,,} \end{equation} where the expectation $\bE_{\dt}$ is taken over an arbitrary initial distribution~$\mu_0$ for~$y^0$ and over all realizations of the Markov chain~\eqref{eqn:numerstep}. The order of the limits on the right hand side of the above equality is unimportant. We comment on the fact that we present the method for the ``perfect'' centering~$\pi_{\dt}(f)$, whereas in practice one should rather center the estimator with the empirical mean of $\{f(y^n)\}_{0 \leq n \leq N-1}$ (see the discussion in~\cite{plechac2019convergence}). This issue does not impact the construction of the method. In view of~\eqref{eqn:desired-estimator}, the discovery of a MP estimator boils down to identifying the corresponding discrete auxiliary martingale $\{z^n\}_{n\geq 0}$ and analyzing asymptotic expansions in~$\dt$ and~$1/N$ of the error \begin{equation} \label{eqn:error} \bE_{\dt}\left\{\left(\frac{1}{N}\sum_{n=0}^{N-1} f(y^n) - \pi_{\dt}(f)\right) z^N\right\} - \mathrm{Lin}(f){\,.} \end{equation} We recall below how to quantify time-step biases on the computation of averages along one realization, and then proceed to the analysis of the time-step bias in the product of such averages with discrete martingales as in~\eqref{eqn:error}. Before we do so, we first rewrite the linear response using the solution to the Poisson equation associated with the continuous time process \begin{equation} \label{eqn:cont-Poisson-eqn-general} - \mL \widehat{f} = f - \pi(f){\,.} \end{equation} The linear response~\eqref{eqn:lin-response-2} can then be rewritten as \begin{equation} \label{eq:Lin_f_Poisson} \mathrm{Lin}(f) = \int_\mY F^\T \nabla \widehat{f} \, d\pi\,. \end{equation} \subsubsection{Quantifying errors on averages along one realization.} The discrete evolution operator~$P_\dt$ associated with~\eqref{eqn:numerstep} is \begin{equation} \label{eqn:discsemigroup} \left(P_{\dt} f\right)(y) = \bE_{\dt}\{f(y^{n+1}) \, | \, y^n = y\} = \bE_G\{f(y + \Phi_h(y;G))\}{\,.} \end{equation} Using Taylor expansions, the action of the operator~$P_h$ can usually be written as an expansion in powers of~$\dt$: for any $C^{\infty}$ test function $f$, \begin{equation} \label{eq:expansion_Pdt} P_{\dt}f = f + \dt {\widetilde{\mathcal{A}}}_1 f + \ldots + \dt^{\alpha} {\widetilde{\mathcal{A}}}_{\alpha}f + \dt^{\alpha+1} \mathcal{R}_{\alpha, \dt}f {\,,} \end{equation} where~${\widetilde{\mathcal{A}}}_i$ and~$\mathcal{R}_{\alpha, \dt}$ are operators which depend on the underlying numerical scheme $\Phi_h$ (the operators~${\widetilde{\mathcal{A}}}_i$ being linear differential operators with smooth coefficients). Since the semigroup of the continuous time process~\eqref{eqn:SDE} has the expansion \[ \mathrm{e}^{\dt \mL} f = f + \dt \mL f + \frac{\dt^2}{2} \mL^2 f + \dots + \frac{\dt^{\alpha}}{\alpha !}\mL^\alpha f + \BIGO(\dt^{\alpha+1}) {\,,} \] it is easily seen that ${\widetilde{\mathcal{A}}}_1 = \mL$ corresponds to weak first order schemes ($\alpha=1$), while the additional condition ${\widetilde{\mathcal{A}}}_2 = \mL^2/2$ characterizes weak second order schemes $(\alpha=2$); see, for instance, \cite{KP92,MT04} for an introduction to the numerical analysis of discretization schemes for diffusion processes. Moreover, it can be shown that an ergodic scheme of weak order~$\alpha$ has an invariant probability measure which is correct at least at order~$\alpha$, in the following sense: for any smooth function~$f$ satisfying some growth conditions, there exists $C_f \in \mathbb{R}_+$ such that \begin{equation} \label{eq:error_inv_meas_dt} \left|\pi_\dt(f) - \pi(f)\right| \leq C_f \dt^{\alpha}. \end{equation} We can now recall the approximation result for pathwise averages of~$\pi(f)$ using a numerical scheme of weak order~$\alpha$ (see for instance~\cite{mattingly2010convergence}): \begin{equation} \label{eqn:ergodicaverage} \left|\bE_{\dt}\left\{ \frac{1}{N}\sum_{n=0}^{N-1} f(y^n) \right\} - \pi(f)\right| \leq C\left(\dt^{\alpha} + \frac{1}{N\dt}\right){\,.} \end{equation} The proof of this result is based on a decomposition of the pathwise average into a sum of discrete martingale increments. To make this decomposition precise, since it will be of paramount importance for the numerical analysis we present, we introduce the operator \begin{equation} \label{eqn:discrete_generator} \LOPERH = \frac{1}{\dt}(P_\dt - I){\,.} \end{equation} By definition, \begin{equation} \label{eqn:obs_increment} f(y^{n+1}) - f(y^n) = \dt \LOPERH(f)(y^n) + {\Delta \mathcal{M}_{n}}(f){\,,} \qquad {\Delta \mathcal{M}_{n}}(f) = f(y^{n+1}) - \left(P_\dt f\right)(y^n){\,,} \end{equation} with ${\Delta \mathcal{M}_{n}}(f)$ the increment (\emph{i.e.}, martingale difference) of a discrete martingale. We next define the discrete Poisson equation \begin{equation} \label{eqn:disc-Poisson-eqn-general} -\LOPERH \widehat{f}_{\dt} = \frac{1}{\dt}(I - P_{\dt}) \widehat{f}_{\dt} = f - \pi_{\dt}(f){\,.} \end{equation} This equation is well-posed when~$P_\dt^n$ decays sufficiently fast with~$n$ in a suitable functional space, typically spaces of measurable functions with a maximal growth at infinity dictated by a Lyapunov function, in which case (see for instance~\cite{meyn2012markov}) \[ (I - P_{\dt})^{-1} = \sum_{n=0}^{\infty} P_{\dt}^n {\,.} \] Using the solution~$\widehat{f}_{\dt}$ of~\eqref{eqn:disc-Poisson-eqn-general}, we can rewrite pathwise averages as \begin{equation}\label{eqn:estimator-idea1} \begin{split} \frac{1}{N}\sum_{n=0}^{N-1} f(y^n) - \pi_{\dt}(f) & = -\frac{1}{N}\sum_{n=0}^{N-1} \LOPERH \widehat{f}_{\dt}(y^n) = \frac{1}{N\dt} \sum_{n=0}^{N-1} \left[{\Delta \mathcal{M}_{n}}(\widehat{f}_{\dt})+\widehat{f}_{\dt}(y^n)-\widehat{f}_{\dt}(y^{n+1})\right]\\ &= \frac{1}{Nh} \left[ \widehat{f}_{\dt}\left(y^0\right) - \widehat{f}_{\dt}\left(y^N\right)\right] + \frac{1}{Nh}\sum_{n=0}^{N-1} \DMNFH {\,.} \end{split} \end{equation} The error estimate~\eqref{eqn:ergodicaverage} then follows by taking the expectation of both sides of the previous equality (provided~$\bE_{\dt}\{\widehat{f}_{\dt}(y^n)\}$ can be bounded uniformly in~$n \geq 0$), and recalling~\eqref{eq:error_inv_meas_dt}. \subsubsection{Quantifying errors in~\eqref{eqn:error}.} Following the derivation leading to~\eqref{eqn:estimator-idea1}, we start by rewriting the first term of~\eqref{eqn:error} as \begin{equation} \label{eqn:estimator-idea} \begin{split} \bE_{\dt}\left\{\left(\frac{1}{N}\sum_{n=0}^{N-1} f(y^n) - \pi_{\dt}(f) \right) z^N\right\} & = \frac{1}{N\dt}\bE_{\dt}\left\{\sum_{n=0}^{N-1} \DMNFH z^N\right\} \\ & \ \ + \frac{1}{N\dt}\bE_{\dt}\left\{ \left(\widehat{f}_{\dt}(y^0) - \widehat{f}_{\dt}(y^N)\right) z^N\right\}\,. \end{split} \end{equation} At this point, it is tempting to expand ${\Delta \mathcal{M}_{n}}(\widehat{f}_{\dt})$ in powers of~$\dt$ using a Taylor expansion of~$\widehat{f}_{\dt}(y^{n+1})=\widehat{f}_{\dt}(y^{n}+\Phi_\dt(y^n;G^n))$ at $y^n$. However, such an expansion requires sufficient regularity of the solution~$\widehat{f}_\dt$ to the discrete Poisson equation~\eqref{eqn:disc-Poisson-eqn-general}. A control of even the first derivatives of the solution to the discrete Poisson equation is not guaranteed in general since the natural functional spaces for the well posedness of the discrete Poisson equation are (weighted) spaces of bounded measurable functions. However, there exists a sufficiently regular function~$\widetilde{f}_{\dt}$ which approximates the solution~$\widehat{f}_\dt$ to the discrete Poisson equation to an arbitrary order in powers of~$\dt$; and also happens to approximate the solution to the continuous time Poisson equation~\eqref{eqn:cont-Poisson-eqn-general}. Indeed, it turns out that, by a technical result whose proof is deferred to Section~\ref{subsec:approx-Poisson-equation} for Langevin dynamics (see also~\cite{plechac2019convergence} for non-degenerate diffusions on a torus), we obtain the following: there exists a function~$\widetilde{f}_{\dt}$, and operators~$\mathcal{Q}_1$ and~$\mathcal{Q}_2$ such that \begin{equation} \label{eqn:approx-operator-preview} \widetilde{f}_{\dt} - \widehat{f} = \left\{ \begin{array}{ll} -\dt \mathcal{Q}_1(f - \pi(f)), & \alpha = 1;\\[3pt] \dfrac{\dt}{2}(f - \pi(f)) -\dt^2 \mathcal{Q}_2(f - \pi(f)), & \alpha =2{\,,} \end{array} \right. \end{equation} where $\alpha=1$ or~$2$ refers to the order of the weak approximation order for the numerical discretization~\eqref{eqn:numerstep} at hand. We next replace $\widehat{f}_\dt$ by~$\widetilde{f}_{\dt}$ in the sum on the right hand side of~\eqref{eqn:estimator-idea} (with a remainder which can be as small as needed, but we restrict ourselves to~$\BIGO(\dt^{\alpha})$ since this will be sufficient for our analysis), thus obtaining \begin{equation} \label{eqn:general-MP-estimate} \begin{aligned} \bE_{\dt}\left\{\left(\frac{1}{N}\sum_{n=0}^{N-1} f(y^n) - \pi_{\dt}(f) \right) z^N\right\} & = \frac{1}{N\dt}\bE_{\dt}\left\{\sum_{n=0}^{N-1} \DMNFT z^N\right\} \\ & \quad + \frac{1}{N\dt}\bE_{\dt}\left\{ \left(\widehat{f}_\dt(y^0) - \widehat{f}_\dt(y^N)\right) z^N\right\} + \BIGO(\dt^{\alpha}){\,.} \end{aligned} \end{equation} Ensuring that the remainder term in the previous estimate is indeed of order~$\dt^\alpha$ requires some work; see for instance~\eqref{eqn:1st-order-estimate}-\eqref{eqn:1st-order-estimate-1}. Under mild conditions, the second term on the right hand side of~\eqref{eqn:general-MP-estimate} can be controlled by~$\BIGO((N\dt)^{-1/2})$ uniformly with respect to the timestep~$\dt$. This term therefore contributes to the finite integration time error (it is larger than for standard time averages, compare with the term~$\BIGO((N\dt)^{-1})$ in~\eqref{eqn:ergodicaverage}), and hence the dominating discretization error is completely determined by the first term on the right hand side of~\eqref{eqn:general-MP-estimate}. In view of~\eqref{eq:Lin_f_Poisson}, our goal is therefore to construct the auxiliary discrete martingale~$z^N$ by exploring the expansion in~$\dt$ of the error term \[ \frac{1}{N\dt}\bE_{\dt}\left\{\sum_{n=0}^{N-1} \DMNFT z^N\right\} - \int_\mY F^\T \nabla \widehat{f} \,d\pi{\,.} \] \subsection{Construction of the discrete auxiliary martingale} \label{sec:construction_discrete_martingale} As mentioned above, rather than starting with the continuous auxiliary martingale~$\{z_t\}_{t\geq 0}$, we construct a discrete time auxiliary martingale~$\{z^n\}_{n\geq 0}$ which allows to obtain a consistent approximation of~$\mathrm{Lin}(f)$ in the limit~$\dt\to 0$. We consider a specific form for the auxiliary martingale, namely \begin{equation} \label{eq:def_z_n} z^{n+1} - z^n = {\Delta \mathcal{M}_{n}}(g) \equiv g(y^{n+1}) - \left(P_{\dt} g\right)(y^n), \qquad z^0 = 0{\,,} \end{equation} where~$g$ is a sufficiently regular function chosen so that \begin{equation} \label{eq:estimator_Lin_f} \mathrm{Lin}_{\dt,N}(f) \equiv \frac{1}{N\dt}\bE_{\dt}\left\{\sum_{n=0}^{N-1} \DMNFT \sum_{m=0}^{N-1} {\Delta \mathcal{M}_{m}}(g)\right\} \end{equation} approximates $\mathrm{Lin}(f)$ to the desired order~$\BIGO(\dt^\alpha)$. Clearly, $\{z^n\}_{n\geq 0}$ defined in~\eqref{eq:def_z_n} is a zero mean martingale adapted to~$\mathcal{F}^n$. In fact, since ${\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt})$ and~${\Delta \mathcal{M}_{m}}(g)$ are increments of mean zero which are conditionally independent for $m\ne n$, \[ \frac{1}{N\dt}\bE_{\dt}\left\{\sum_{n=0}^{N-1} \DMNFT \sum_{m=0}^{N-1} {\Delta \mathcal{M}_{m}}(g)\right\} = \frac{1}{N\dt} \sum_{n=0}^{N-1} \bE_{\dt}\left\{\DMNFT {\Delta \mathcal{M}_{n}}(g)\right\}. \] The expression on the right hand side of the previous equality motivates the name ``martingale product'' of our estimator. The next step is to obtain an expansion in powers of~$\dt$ of the right hand side of the previous equality. To simplify the notation and the computation of these expansions, we introduce the ``carr\'e du champ operator'' associated with the generator~$\mL$: for two smooth funtions~$f,g$, \begin{equation} \label{eqn:GammaDef} \Gamma(f,g) = \mL(fg) - f\mL(g) - g\mL(f) = \nabla f^\T S \nabla g{\,.} \end{equation} We also consider its discrete counterpart (recall the definition~\eqref{eqn:discrete_generator} for~$\LOPERH$): \begin{equation} \label{eqn:GammahDef} \Gamma_\dt(f,g) = \LOPERH(fg) - f\LOPERH(g) - g\LOPERH(f) - \dt\LOPERH(f)\LOPERH(g){\,.} \end{equation} In view of the definition~\eqref{eqn:GammahDef}, the following crucial identity holds \begin{equation} \label{eq:crucial_identity} \bE_{G^n}\left\{\DMNFT {\Delta \mathcal{M}_{n}}(g)\right\} = \left[P_{\dt}(\widetilde{f}_{\dt} g)\right](y^n) - (P_{\dt}\widetilde{f}_{\dt})(y^n) (P_{\dt}g)(y^n) = \dt \Gamma_\dt(\widetilde{f}_{\dt},g)(y^n){\,,} \end{equation} where $\bE_{G^n}$ denotes the expectation with respect to the random variable~$G^n$ in~\eqref{eqn:numerstep}. Therefore, for a numerical scheme of weak order~$\alpha$, \begin{equation} \label{eqn:expansionDMDN} \mathrm{Lin}_{\dt,N}(f) = \frac{1}{N} \sum_{n=0}^{N-1}\bE_\dt\left\{\Gamma_h(\widetilde{f}_{\dt},g)(y^n)\right\} = \int_\mY \Gamma_h(\widetilde{f}_{\dt},g) \, d\pi + \BIGO(\dt^\alpha) + \BIGO\left(\frac{1}{N\dt}\right){\,,} \end{equation} where the second equality follows from applying~\eqref{eqn:ergodicaverage} to the observable $\Gamma_h(\widetilde{f}_{\dt}, g)$. In order to expand $\mathrm{Lin}_{\dt,N}(f)$ in powers of~$\dt$, it suffices to expand the integral in the last equality of~\eqref{eqn:expansionDMDN}. We consider successively schemes of weak order~$\alpha = 1$ or~2, using the following expansions: \begin{eqnarray} \mbox{$\alpha=1$:} & \LOPERH = \mL + \BIGO(h){\,,} &\Gamma_\dt(f,g) = \Gamma(f,g) + \BIGO(h){\,,} \label{eqn:order1}\\ \mbox{$\alpha=2$:} & \displaystyle \LOPERH = \mL + \frac{h}{2} \mL^2 + \BIGO(h^2){\,,} &\Gamma_\dt(f,g) = \Gamma(f,g) + h \Lambda(f,g) + \BIGO(h^2){\,,} \label{eqn:order2} \end{eqnarray} where \begin{equation} \label{eqn:Lambda-operator} \Lambda(f, g) = \frac{1}{2}\Big(\mL^2(fg) - f \mL^2 g - g\mL^2 f - 2 (\mL f)( \mL g)\Big){\,.} \end{equation} Note that \begin{equation} \label{eq:int_Lambda_pi} \begin{aligned} \int_\mY \Lambda(f, g) \, d\pi & = -\int_\mY \Big(\frac{\mL+\mL^*}{2}\Big)f \mL g + \mL f \Big(\frac{\mL+\mL^*}{2}\Big) g\, d\pi \\ & = \frac12 \int_\mY \nabla f^\T S \nabla(\mL g) + \nabla g^\T S \nabla(\mL f) \, d\pi{\,.} \end{aligned} \end{equation} \subsubsection{Weak first order schemes ($\alpha=1$).} \label{sec:first_order_general} For a weak first order discretization scheme, \eqref{eqn:approx-operator-preview}, \eqref{eqn:GammaDef}, \eqref{eqn:expansionDMDN} and~\eqref{eqn:order1} lead to \begin{equation} \label{eqn:carre-du-champ-1st-order} \mathrm{Lin}_{\dt,N}(f) = \int_\mY \Gamma(\widehat{f}, g) \, d\pi + \BIGO(\dt) + \BIGO\left(\frac{1}{N\dt}\right) = \int_\mY \nabla \widehat{f}^\T S \nabla g \, d\pi + \BIGO(\dt) + \BIGO\left(\frac{1}{N\dt}\right){\,.} \end{equation} This coincides at dominant order with the linear response as given by~\eqref{eq:Lin_f_Poisson} provided the following condition holds: for all smooth functions~$\varphi$ with compact support, \begin{equation} \label{eqn:constraint-g} \int_\mY \nabla \varphi^\T S \nabla g \, d\pi = \int_\mY \nabla \varphi^\T F \, d\pi\,. \end{equation} The latter condition can be seen as the Euler--Lagrange equation associated with the minimization problem \[ \min_{\phi \in H^1(\pi)} \left\{ \int_\mY \left|\sigma^\T\nabla \phi-u\right|^2 d\pi \right\}, \] where~$u$ satisfies~\eqref{eq:su=F}, \emph{i.e.}, $\sigma u = F$, and $H^1(\pi) = \{\phi\in L^2(\pi) ~|~ \partial_{y_i} \phi \in L^2(\pi), \forall i = 1, \ldots, d\} $. This minimization problem always admits a solution when~$\sigma$ is bounded (since it can be seen as the orthogonal projection of~$u$ onto the closed subspace~$\{\sigma^\T \nabla \phi ~| \ \phi \in H^1(\pi)\}$ of~$L^2(\pi)$), but the corresponding minimizer~$g$ may not be unique, even if the function~$g$ is required to have mean zero with respect to~$\pi$ (in order to remove the trivial obstruction to uniqueness provided by adding constants to a minimizer). Alternatively, \eqref{eqn:constraint-g} can be interpreted as a weak formulation of the following Poisson-type equation for~$g$: \begin{equation} \label{eq:PDE_g} \nabla^* (S \nabla g) = \nabla^* F \equiv \sum_{i=1}^d \partial_{y_i}^* F_i, \end{equation} with $\partial_{y_i}^* = -\partial_{y_i} - \partial_{y_i} (\ln \rho)$ the adjoint of~$\partial_{y_i}$ with respect to the canonical scalar product on~$L^2(\pi)$. The well-posedness of the equation defining~$g$ is discussed in Section~\ref{sec:LD-application} for Langevin dynamics. When~\eqref{eqn:constraint-g} holds, we can then consider (at least formally) the following first order MP estimator \begin{equation} \label{eqn:general-first-order-MP} \mathcal{E}_{\dt, N}^{\mathrm{MP1}} = \frac{1}{N}\sum_{n=0}^{N-1}\left( f(y^n) -\pi_{\dt}(f)\right) \left( g(y^{n+1}) - (P_{\dt} g)(y^n) \right){\,.} \end{equation} To obtain an actual expression of~$z^n$ which can be implemented in practice, the final step is to Taylor expand $g(y^{n+1})$ in (fractional) powers of~$\dt$ around~$y^n$, and to replace~$P_\dt$ in~$(P_{\dt} g)(y^n)$ with~\eqref{eq:expansion_Pdt}. \begin{example}[Euler-Maruyama scheme] {\rm Consider a non-singular noise ($\sigma(y)$ is invertible for all~$y \in \mY$), and the Euler-Maruyama discretization \[ y^{n+1} = y^n + h\, b(y^n) + h^{1/2}\, \sigma(y^n) G^n{\,,} \] where $G^n$ are independent and identically distributed standard random Gaussian variables with identity covariance. A simple computation then shows that the martingale increment appearing in the definition~\eqref{eq:def_z_n} can be expanded as~${\Delta \mathcal{M}_{n}}(g) = h^{1/2} \nabla g(y^n)^\T\sigma(y^n) G^n +\mathcal{O}_P(\dt^{3/2})$, ~where $\mathcal{O}_P(\dt^\alpha)$ denotes random terms of order~$h^\alpha$ (in the sense that $\bE\{|\mathcal{O}_P(\dt^\alpha)|\} = \BIGO(\dt^\alpha)$). The condition~\eqref{eqn:constraint-g} suggests to replace~$\sigma^\T\nabla g$ with~$u=\sigma^{-1}F$, which leads to the following definition for the discrete auxiliary martingale: \[ z^{n+1} = z^n + h^{1/2} \left(G^n\right)^\T \sigma^{-1}(y^n) F(y^n){\,.} \] The discrete martingale~$z^n$ is in this case a consistent approximation of the likelihood process~$z_t$, and the associated first order estimator is \[ \mathcal{E}_{\dt, N}^{\mathrm{MP1}} = \frac{1}{N}\sum_{n=0}^{N-1}\left( f(y^n) -\pi_{\dt}(f)\right) \left( \dt^{1/2} \left(G^n\right)^\T \sigma^{-1}(y^n) F(y^n) \right){\,.} \] This corresponds to the first order estimator already introduced in~\cite{plechac2019convergence}. } \end{example} \subsubsection{Weak second order schemes $\alpha=2$.} The interest of our approach becomes more apparent with a second order discretization scheme as it allows us to systematically discover the corresponding second order MP estimator. In order to correct the discretization error of order~$\dt$, we consider an additional martingale of order~$\mathcal{O}_P(\dt^{3/2})$, which can be seen as a correction to~\eqref{eq:def_z_n}, so that \[ z^{n+1} - z^n = {\Delta \mathcal{M}_{n}}(g+\dt\WTG ) {\,,} \] where the leading order function~$g$ is determined by~\eqref{eqn:constraint-g} as for first order schemes. From~\eqref{eqn:approx-operator-preview}, \eqref{eqn:expansionDMDN},~\eqref{eqn:order2} and the bilinearity of the operator $\Gamma$, \[ \begin{aligned} \mathrm{Lin}_{\dt,N}(f) & = \int_\mY \Gamma_\dt\left(\widehat{f}+\frac{\dt}{2}f, g+\dt\WTG \right) d\pi + \BIGO(\dt^2) + \BIGO\left(\frac{1}{N\dt}\right) \\ & = \int_\mY \Gamma\left(\widehat{f},g\right) d\pi + \dt \int_\mY \frac12 \Gamma(f,g) + \Gamma\left(\widehat{f},\WTG \right) + \Lambda\left(\widehat{f},g\right) d\pi+ \BIGO(\dt^2) + \BIGO\left(\frac{1}{N\dt}\right){\,.} \end{aligned} \] As for first order schemes, this coincides at leading order with the linear response as given by~\eqref{eq:Lin_f_Poisson} provided~\eqref{eqn:constraint-g} holds. The bias of order~$\dt$ vanishes provided~$\WTG $ is chosen such that \[ \begin{aligned} \int_\mY \Gamma\left(\widehat{f},\WTG \right) d\pi & = - \int_\mY \frac12 \Gamma(f,g) + \Lambda\left(\widehat{f},g\right) d\pi \\ & = - \frac12 \int_\mY \nabla f^\T S \nabla g \, d\pi - \frac12 \int_\mY \nabla (\mL \widehat{f})^\T S \nabla g \, d\pi - \frac12\int_\mY \nabla \widehat{f}^\T S \nabla (\mL g) \, d\pi \\ & = - \frac12 \int_\mY \nabla \widehat{f}^\T S \nabla (\mL g) \, d\pi = -\frac12\int_\mY \Gamma(\widehat{f},\mL g) \, d\pi {\,,} \end{aligned} \] where we used~\eqref{eq:int_Lambda_pi} to obtain the second equality and ~\eqref{eqn:cont-Poisson-eqn-general} to obtain the third equality. This calculation suggests to choose \[ \WTG = -\frac12 \mL g\,. \] This leads to the following second order MP estimator: \begin{equation} \label{eqn:general-second-order-MP} \mathcal{E}_{\dt, N}^{\mathrm{MP}2} = \frac{1}{N}\sum_{n=0}^{N-1}\left[ f(y^n) -\pi_{\dt}(f)\right] \left[ \left(g-\frac{\dt}{2}\mL g\right)(y^{n+1}) - P_{\dt} \left(g-\frac{\dt}{2}\mL g\right)(y^n) \right]{\,,} \end{equation} where $g$ satisfies~\eqref{eqn:constraint-g}. As discussed after~\eqref{eqn:general-first-order-MP}, one needs to Taylor expand $g(y^{n+1})$ in (fractional) powers of~$\dt$ around~$y^n$, and to replace~$P_\dt$ in~$(P_{\dt} g)(y^n)$ with~\eqref{eq:expansion_Pdt} in order to obtain a computable expression for $z^n$. This answers in any case the question left open in~\cite{plechac2019convergence} about constructing second order schemes for general non degenerate diffusion processes in spaces of dimension higher than or equal to two. We emphasize that the derivation of the MP estimators~\eqref{eqn:general-first-order-MP} and~\eqref{eqn:general-second-order-MP} presented in this section is heuristic since we have not provided any uniform controls for the error terms. In order to carry out rigorous analysis in the next sections we consider the particular case of Langevin dynamics with numerical schemes based on operator splitting methods~\cite{leimkuhler2013rational}. \section{Splitting schemes for Langevin dynamics} \label{sec:LD-application} In this section, we present Langevin dynamics and define the class of perturbations we consider (see Section~\ref{sec:Langevin_transport}). We next recall in Section~\ref{sec:splitting_Lang} splitting schemes to discretize Langevin dynamics, as well as some of their properties. We conclude in Section~\ref{subsec:approx-Poisson-equation} by providing a construction of the smooth approximations to discrete Poisson equations introduced in~\eqref{eqn:approx-operator-preview}. \subsection{Langevin dynamics and transport coefficients} \label{sec:Langevin_transport} Langevin dynamics is a diffusion process which describes the evolution of $\mathcal{N}$ particles in a space of physical dimension~$d$, at fixed temperature. We denote by $q = (q_{1}, \cdots, q_{\mathcal{N}}) \in \mathcal{D}$ the positions of the particles (typically, $\mathcal{D} = \mathbb{T}^D$ with $\mathbb{T} = \mathbb{R}\backslash \mathbb{Z}$ and~$D=\mathcal{N} d$, or~$\mathcal{D}=\mathbb{R}^D$), and by~$p = (p_{1}, \cdots, p_{\mathcal{N}})\in \mathbb{R}^{D}$ their momenta. The phase space for the elements~$y=(q,p)$ is therefore $\mY = \mathcal{D} \times \mathbb{R}^{D}$. \subsubsection{Reference Langevin dynamics.} Langevin dynamics is defined by the system of stochastic differential equations \begin{equation} \label{eqn:ULD} \left\{ \begin{aligned} dq_t &= M^{-1}p_t\,dt\,, \\ dp_t &= -\nabla V(q_t)\, dt - \gamma M^{-1} p_t \,dt + \displaystyle\sqrt{\frac{2\gamma}{\beta}} dW_t\,, \end{aligned} \right. \end{equation} where~$\beta$ is (proportional to) the inverse temperature, $V:\mathcal{D} \to \mathbb{R}$ is the potential function, the mass matrix $M$ is a symmetric definite positive matrix such as $M= \mathrm{diag}(m_1 \mathrm{Id}_d, \ldots, m_N \mathrm{Id}_d)$ with~$m_i > 0$ for $1 \leq i \leq \mathcal{N}$, $\beta$ is proportional to the inverse temperature, $\gamma>0$ is the friction coefficient, and $W_t$ is a standard $D$-dimensional Brownian motion. We consider the following running assumption. \begin{assumption} \label{ass:V_smooth_and_domain_compact} The potential $V$ belongs to~$C^{\infty}(\mathcal{D})$ and the position space is compact: $\mathcal{D} = \mathbb{T}^D$. \end{assumption} Periodic boundary conditions for the positions are natural in molecular dynamics simulations of condensed matter systems~\cite{FrenkelSmit,tuckerman2010statistical}, so that the assumption $\mathcal{D} = \mathbb{T}^D$ is physically relevant. The main benefit of this assumption is that it allows us to simplify various mathematical arguments, although many of them could be extended to unbounded position spaces. The existence and uniqueness of strong solutions to~\eqref{eqn:ULD} is standard when the position space is compact~\cite{bellet2006ergodic}. Moreover, the unique invariant measure of~\eqref{eqn:ULD} is the Boltzmann--Gibbs probability measure \[ \pi(dq\, dp) = Z^{-1} \mathrm{e}^{-\beta H(q, p)}\, dq\,dp\,,\; \qquad Z = \int_\mY \mathrm{e}^{-\beta H}\,, \;\;\qquad H(q, p) = V(q) + \frac{1}{2}p^{\T} M^{-1} p\,, \] where $H$ denotes the Hamiltonian of the system. \subsubsection{Poisson equation associated with the reference Langevin dynamics.} In order to control remainder terms we restrict the analysis to a subclass class of smooth observables. Since the position space~$\mathcal{D}$ is compact, it is sufficient to control the growth of functions and their derivatives with respect to the momentum variable. Specifically, as in~\cite{talay2002stochastic,kopec2015weak}, we consider scale functions of polynomial form \[ \mKS(q, p) = 1 + |p|^{2s}\,, \;\;\;\qquad s = 1, 2, \ldots{\,,} \] and define \[ B_{\mKS}^{\infty}(\mY) = \left\{ f \mathrm{~measurable} \ \middle| \ \left\|f\right\|_{B_{\mKS}^{\infty}} = \sup_{y \in \mY} \left|\frac{f(y)}{\mKS(y)}\right| < \infty \right\}\,. \] The set of observables we consider is then composed of smooth functions which, together with their derivatives, grow at most polynomially: \[ \mS = \left\{f \in C^{\infty}(\mY) \left| ~\forall k \in \mathbb{N}^{2D}, \ \exists s \in \mathbb{N} ~\text{such that}~ \partial^k f \in B_{\mKS}^{\infty}(\mY) \right. \right\}{\,,} \] where $\partial^k f = \partial_{y_1}^{k_1}\dots \partial_{y_{2D}}^{k_{2D}} f$ for $k = (k_1,\dots,k_{2D}) \in \mathbb{N}^{2D}$. Clearly, the invariant measure $\pi$ integrates all functions~$\mKS$ for $s \in \mathbb{N}$, and hence $\mS$ is a dense subspace of~$L^2(\pi)$. Finally, we define \[ \mS_0 = \Pi \mS = \mS \cap L^2_0(\pi)\,. \] An important result for our analysis is the following regularity result concerning the existence and uniqueness of solutions to the continuous time Poisson equation \begin{equation} \label{eqn:cont-Poisson-eqn} -\mL_{\gamma} \widehat{f} = f - \pi(f){\,,} \end{equation} where $\mathcal{L}_{\gamma}$ is the generator of the Langevin dynamics: \begin{equation} \mathcal{L}_{\gamma} = A + B + \gamma C {\,,} \label{eqn:Lgamma} \end{equation} with \begin{equation} A = p^\T M^{-1} \nabla_q\,, \qquad B = -\nabla V(q)^{\T} \nabla_p\,, \qquad C = - p^\T M^{-1} \nabla_p + \frac{1}{\beta} \Delta_p\,. \label{eqn:ABC} \end{equation} The invertibility of~$\mL_{\gamma}$ considered as an operator on~$L_0^2(\pi)$ can be obtained by results of hypocoercivity (relying on the exponential convergence of the semigroup as provided by~\cite{Herau06,villani2009hypocoercivity,DMS15} for instance, see the introduction of~\cite{bernard2020hypocoercivity} for an extensive review) or by a direct analysis based on Schur complements~\cite{bernard2020hypocoercivity}. In fact, the solution to~\eqref{eqn:cont-Poisson-eqn} belongs in fact to~$\mS_0$ when~$f \in \mS$, as made precise in the following result proved in~\cite{kopec2015weak, talay2002stochastic}. \begin{theorem} \label{thm:L-stability} The space $\mS_0$ is stable under $\mL_{\gamma}^{-1}$: % for all $f \in \mS$, there is a unique $\widehat{f} \in \mS_0$ % satisfying the Poisson equation~\eqref{eqn:cont-Poisson-eqn} associated with $\mL_{\gamma}$. % \end{theorem} The regularity result of Theorem~\ref{thm:L-stability} is of fundamental importance for justifying the consistency of MP estimators, since it allows us to replace solutions to the Poisson equation associated with discretizations of the stochastic differential equation~\eqref{eqn:ULD} by smooth functions, up to a small error term (see Section~\ref{subsec:approx-Poisson-equation} below). \subsubsection{Perturbations of the Langevin dynamics.} Various dynamical properties of Langevin dynamics can be of interest, in particular transport coefficients such as the mobility, thermal conductivity or shear viscosity; see for instance~\cite{evans-moriss,tuckerman2010statistical} as well as~\cite{hairer2008ballistic,Lefevere,JS12,BK21} for some representative mathematical studies dealing with these transport coefficients. As discussed in Section~\ref{sec:LR_stoch_dyn}, we view transport coefficients as being obtained from the linear response of steady state averages of Langevin dynamics perturbed by an external forcing. More precisely, we consider a smooth external forcing $F: \mathcal{D} \to \mathbb{R}^{D}$, and the following non-equilibrium Langevin dynamics for~$\eta \in \mathbb{R}$: \begin{equation} \label{eqn:ULD-perturbed} \left\{ \begin{aligned} dq_t^\eta &= M^{-1}p_t^\eta\,dt {\,,}\\ dp_t^\eta &= \left(-\nabla V(q_t^\eta) + \eta F(q_t^\eta)\right)\, dt - \gamma M^{-1} p_t^\eta \,dt + \displaystyle\sqrt{\frac{2\gamma}{\beta}} dW_t {\,.} \end{aligned} \right. \end{equation} Note that the forcing~$F(q)$ is a function of the positions only, and is in general not the gradient of a smooth periodic function. For instance, mobility can be computed with~$F(q)$ constant (as made precise in~\cite[Section~5]{lelievre2016partial}), while the shear viscosity can be recovered by considering a force whose magnitude in one direction depends only on the component of the position in another direction (the so-called sinusoidal transverse field method~\cite{GMS73,JS12}). It can be shown that~\eqref{eqn:ULD-perturbed} admits a unique strong solution. Moreover, this SDE has a unique invariant probability measure~$\pi^{\eta}$, using standard arguments (a minorization condition is obtained as in~\cite{MSH02} and we can then use the results of~\cite{hairer2011yet} since a Lyapunov condition holds as well). In addition, the invariant measure has a smooth density with respect to the Lebesgue measure by hypoellipticity~\cite{hormander1967hypoelliptic}. \begin{remark} { It would be possible to consider an external forcing which depends both the positions~$q$ and momenta~$p$. In order to guarantee the existence and uniqueness of strong solutions to~\eqref{eqn:ULD-perturbed}, at least for~$|\eta|$ small enough, this would require conditions on~$F$ such that~$\mKS$ are Lyapunov functions. This is for instance the case when there exist $a$, $b \in \mathbb{R}$ such that \[ p^\T F(q,p) \leq a |p|^2 + b\,. \] However, we refrain from considering such a more general setting since the applications we have in mind consider a position-dependent forcing only. } \end{remark} The generator of the perturbed Langevin dynamics~\eqref{eqn:ULD-perturbed} can be written as \[ \mL_{\gamma}^{\eta} = \mL_{\gamma} + \eta\LOPERTIL\,, \] with $\mathcal{L}_{\gamma}$ defined in~\eqref{eqn:Lgamma}, and the perturbation operator \begin{equation} \LOPERTIL = F^\T \nabla_p {\,.} \label{eqn:Ltilde} \end{equation} Following the argument in~\cite[Remark~5.5]{lelievre2016partial}, it can be shown that the linear response of an observable~$f \in \mS$ is well defined and can be written as \begin{equation*} \mathrm{Lin}(f) = -\int_{\mY} \LOPERTIL \mL_{\gamma}^{-1} \Pi f \,d\pi {\,.} \end{equation*} In view of the Poisson equation~\eqref{eqn:cont-Poisson-eqn} associated with the generator~\eqref{eqn:Lgamma}, we end up with the reformulation \begin{equation} \label{eqn:Langevin-lin-res-reformulation} \mathrm{Lin}(f) = \int_{\mY} F^\T \nabla_p \widehat{f}\, d\pi {\,,} \end{equation} that serves as the starting point for deriving MP estimators. \subsection{Splitting schemes for Langevin dynamics and their properties} \label{sec:splitting_Lang} We briefly describe in this section splitting schemes to discretize the Langevin dynamics~\eqref{eqn:ULD} (see~\cite{leimkuhler2013rational, leimkuhler2016computation} for a full exposition) and then recall some of their properties, in particular concerning their invariant measures and their convergence to stationarity. \subsubsection{Splitting schemes.} The splitting schemes we consider for Langevin dynamics are based on the decomposition \eqref{eqn:Lgamma} of the generator $\mathcal{L}_{\gamma}$. First order splitting schemes with a timestep~$\dt$ are obtained by a Lie-Trotter approximation of the continuous time evolution operator~$\mathrm{e}^{\dt\mL_{\gamma}}$ by the discrete time evolution operator \[ P_{\dt}^{Z, Y, X} = \mathrm{e}^{\dt Z} \mathrm{e}^{\dt Y} \mathrm{e}^{\dt X}, \] where $(Z, Y, X)$ is one of the six possible permutations of the elementary operators~$(A, B, \gamma C)$ introduced in~\eqref{eqn:ABC}. For example, the scheme associated with~$P_{\dt}^{B, A, \gamma C}$ reads \begin{equation} \label{eqn:BAC-scheme} \left\{ \begin{aligned} p^{n+\frac{1}{2}} &= p^n - \dt \nabla V(q^n) {\,,} \\ q^{n+1} &= q^n + \dt M^{-1} p^{n + \frac{1}{2}} {\,,} \\ p^{n+1} &= \alpha_{\dt} p^{n+\frac{1}{2}} + \sqrt{\frac{1-\alpha_{\dt}^2}{\beta} M} G^n, \end{aligned} \right. \end{equation} where $\alpha_{\dt} = \exp(-\gamma M^{-1} \dt)$ and $(G^n)_{n \geq 0}$ are independent and identically distributed standard $D$-dimensional Gaussian random vectors. Similarly, second order schemes approximate the continuous time evolution operator~$\mathrm{e}^{\dt\mL_{\gamma}}$ through a Strang splitting \[ P_{\dt}^{Z, Y, X, Y, Z} = \mathrm{e}^{\dt Z/2} \mathrm{e}^{\dt Y/2} \mathrm{e}^{\dt X} \mathrm{e}^{\dt Y/2} \mathrm{e}^{\dt Z/2}. \] For instance, the second order scheme associated with~$P_{\dt}^{B, A, \gamma C, A, B}$ reads \begin{equation} \label{eqn:BACAB} \left\{ \begin{aligned} p^{n+\frac{1}{3}} &= p^n - \frac{\dt}{2} \nabla V(q^n) {\,,} \\ q^{n+\frac{1}{2}} &= q^n + \frac{\dt}{2} M^{-1} p^{n + \frac{1}{3}} {\,,} \\ p^{n+\frac{2}{3}} &= \alpha_{\dt} p^{n+\frac{1}{3}} + \sqrt{\frac{1-\alpha_{\dt}^2}{\beta} M} G^n{\,,} \\ q^{n+1} &= q^{n+\frac{1}{2}} + \frac{\dt}{2} M^{-1} p^{n+\frac{2}{3}} {\,,} \\ p^{n+1} &= p^{n+\frac{2}{3}} - \frac{\dt}{2} \nabla V(q^{n+1}) {\,,} \end{aligned} \right. \end{equation} where $(G^n)_{n \geq 0}$ are independent and identically distributed standard $D$-dimensional Gaussian random vectors; while the second order scheme associated with~$P_{\dt}^{\gamma C,B, A, B,\gamma C}$ reads \begin{equation} \label{eqn:CBABCB} \left\{ \begin{aligned} p^{n+\frac{1}{4}} &= \alpha_{\dt/2} p^{n} + \sqrt{\frac{1-\alpha_{\dt/2}^2}{\beta} M} G_1^n{\,,} \\ p^{n+\frac{1}{2}} &= p^{n+\frac14} - \frac{\dt}{2} \nabla V(q^n) {\,,} \\ q^{n+1} &= q^n + \dt M^{-1} p^{n + \frac{1}{2}} {\,,} \\ p^{n+\frac{3}{4}} &= p^{n+\frac12} - \frac{\dt}{2} \nabla V(q^{n+1}) {\,,} \\ p^{n+1} &= \alpha_{\dt/2} p^{n+\frac34} + \sqrt{\frac{1-\alpha_{\dt/2}^2}{\beta} M} G_2^n{\,,} \\ \end{aligned} \right. \end{equation} where $(G_1^n)_{n \geq 0}$ and $(G_2^n)_{n \geq 0}$ are two independent families of independent and identically distributed standard $D$-dimensional Gaussian random vectors. We recall in the next subsections various ergodicity results and error estimates for the schemes under consideration. \subsubsection{Ergodicity results.} One preliminary result used to prove the ergodicity of the numerical schemes is that the functions~$\mKS$ are Lyapunov functions for the numerical schemes under consideration (see~\cite[Lemma~2.7]{leimkuhler2016computation}): for any~$s^* \in \mathbb{N}$, there exists constants $\lambda>0$ and $K \in \mathbb{R}_+$ and a timestep~$\dt^*>0$ such that, for all $s = 1, 2, \ldots, s^*$ and $0 < \dt \leq \dt^*$, \[ \forall n \in \mathbb{N}, \qquad (P_{\dt}^n \mKS)(y) \leq \mathrm{e}^{-\lambda n \dt} \mKS(y) + K {\,.} \] In particular, there exists~$C \in \mathbb{R}_+$ such that \begin{equation} \label{eqn:K-evolution-estimate} \forall n \in \mathbb{N}, \qquad (P_{\dt}^n \mKS)(y) \leq C \mKS(y) {\,.} \end{equation} With this result at hand, the following theorem regarding the ergodicity of the associated Markov chain $\{y^n\}_{n\geq 0}$ is obtained from the results in~\cite{leimkuhler2016computation,EDMS21}. \begin{theorem} \label{thm:ergodicity-splitting} Consider $s^* \geq 1$. For any $0 < \gamma < \infty$, there exists $\dt^* > 0$ such that, for any $0 < \dt \leq \dt^*$, the Markov chain associated with a first or second order splitting scheme has a unique invariant probability measure $\pi_{\gamma, \dt}$, which integrates the scale functions~$\mKS$ for all $s = 1, 2, \ldots, s^*$ uniformly in $\dt \in (0,\dt^*]$: \begin{equation} \label{eqn:integrate-scale-function} \forall s \in \{1, 2, \ldots, s^*\}, \qquad \sup_{0 < \dt \leq \dt^*}\int_{\mY} \mKS \, d\pi_{\gamma, \dt} < \infty. \end{equation} Moreover, there exist constants $\lambda = \lambda(s^*, \gamma) > 0$ and $K = K(s^*, \gamma) > 0$ such that, for any observable~$f \in B_{\mKS}^{\infty}(\mY)$, \begin{equation*} \forall n \in \mathbb{N}, \qquad \left\| P_{\dt}^n f - \int_{\mY} f \,d\pi_{\gamma, \dt}\right\|_{B^\infty_{\mKS}} \leq K \mathrm{e}^{-\lambda n \dt} \|f\|_{B_{\mKS}^{\infty}} {\,.} \end{equation*} \end{theorem} An important consequence of the above ergodicity result is the uniform control of the solution to the discrete time Poisson equation \begin{equation} \label{eqn:disc-Poisson-eqn} - \LOPERH\widehat{f}_{\dt} = f - \pi_{\gamma,\dt}(f){\,,} \end{equation} where $\LOPERH$ is defined in~\eqref{eqn:discrete_generator}. Upon defining the subspace of~$B_{\mKS}^{\infty}(\mY)$ of functions with average~0 with respect to the probability measure~$\pi_{\gamma, \dt}$: \[ B_{\mKS,\dt}^{\infty}(\mY) = \left\{f \in B_{\mKS}^{\infty}(\mY) ~\left|~ \int_{\mY} f \, d\pi_{\gamma, \dt} = 0\right. \right\}{\,,} \] we can recall the following result (see~\cite{EDMS21} and~\cite[Corollary~2.10]{leimkuhler2016computation}). \begin{corollary} \label{cor:bounded-disc-Poission-solution} Consider an integer $s^* \geq 1$ and $\gamma\in (0,\infty)$. There exists $\dt^* > 0$ such that, for all $ s \in \{ 1, \ldots, s^*\}$, \begin{equation} \label{eqn:bounded-disc-resolvent} \sup_{0<\dt \leq \dt^*}\left\|\widehat{f}_{\dt}\right\|_{B_{\mKS}^{\infty}} \leq \frac{2K}{\lambda}\left\|f \right\|_{B_{\mKS}^{\infty}} {\,,} \end{equation} with the same constants $K,\lambda$ as in Theorem~\ref{thm:ergodicity-splitting}. \end{corollary} \subsubsection{Error estimates on average properties.} In the remainder of this work, we assume that the initial configuration $y^0 = (q^0, p^0)$ follows a distribution~$\mu_0$ which integrates all scale functions~$\mKS$, namely \[ \forall s \geq 1, \qquad \int_{\mY} \mKS \,d\mu_0 < \infty{\,.} \] To simplify the notation, we denote by $\bE_{\dt} = \bE_{\dt}^{\mu_0}$ the expectation taken with respect to all initial conditions $y^0 \sim \mu_0$ and for all realizations of the Markov chain $\{y^n\}_{n\geq 0}$ starting from~$y^0$. The following result quantifies two sources of bias in the estimation of averages with respect to the target measure~$\pi$ by averages over one realization of the Markov chain: (i) a systematic bias arising from the use of a finite timestep~$\dt$; (ii) a truncation bias from the finite integration time. The former bias has been carefully studied in~\cite{leimkuhler2016computation} for Langevin dynamics, following the framework initiated by~\cite{TT90,talay2002stochastic}; while the latter bias is a direct consequence of Theorem~\ref{thm:ergodicity-splitting} (see also~\cite{mattingly2010convergence}). \begin{theorem}[Bias of splitting schemes] \label{thm:bias-splitting} Fix~$\gamma \in (0,\infty)$ and an observable~$f \in \mS$, and consider either any first order ($\alpha = 1$) or second order ($\alpha =2$) splitting scheme. There exist $\dt^* > 0$ and $C_f > 0$ (which both depend on~$f$) such that, for any $0 < \dt \leq \dt^*$ and integer $N \geq 1$, \begin{equation} \label{eqn:bias-splitting} \left|\frac{1}{N}\sum_{n=0}^{N-1} \bE_{\dt}\{f(y^n)\} - \int_{\mY} f \,d\pi \right| \leq C_f \left( \frac{1}{N\dt} + \dt^{\alpha} \right){\,.} \end{equation} \end{theorem} \begin{proof} Denote by~$s \in \mathbb{N}$ an integer such that $f \in B^\infty_{\mKS}(\mY)$, and consider the timestep~$\dt^*$ as given by Theorem~\ref{thm:ergodicity-splitting}. It is a direct consequence of Theorem~$2.13$ ($\alpha = 1$) and Theorem~$2.16$ ($\alpha = 2$) in~\cite{leimkuhler2016computation} that there exists a constant $C_f \in \mathbb{R}_+$ such that, for all $\dt \in (0,\dt^*]$, \begin{equation} \label{eq:error_ergodic_avg_splitting} \left|\int_{\mY} f \, d\pi_{\gamma, \dt} - \int_{\mY} f \, d\pi \right| \leq C_f \dt^{\alpha} {\,.} \end{equation} On the other hand, by Theorem~\ref{thm:ergodicity-splitting}, \begin{equation*} \begin{aligned} & \left|\frac{1}{N}\sum_{n=0}^{N-1} \bE_{\dt}\{f(y^n)\} - \int_{\mY} f \,d\pi_{\gamma, \dt} \right| \leq \frac{1}{N}\sum_{n=0}^{N-1} \int_\mY \left|P_\dt^n f - \int_{\mY} f \,d\pi_{\gamma, \dt}\right| d\mu_0 \\ & \qquad \leq \frac{K}{N} \mu_0(\mKS)\|f\|_{B_{\mKS}^{\infty}}\sum_{n=0}^{N-1} \mathrm{e}^{-\lambda n \dt} = \frac{K(1-\mathrm{e}^{-\lambda N\dt})}{N(1-\mathrm{e}^{\lambda \dt})} \mu_0(\mKS)\|f\|_{B_{\mKS}^{\infty}} \leq \frac{2K}{\lambda N\dt} \mu_0(\mKS) \|f\|_{B_{\mKS}^{\infty}} {\,,} \end{aligned} \end{equation*} upon possibly reducing~$\dt^*$. The error estimate~\eqref{eqn:bias-splitting} then immediately follows from the triangle inequality, with $C = \max\{C_f, 2K \mu_0(\mKS)\|f\|_{B_{\mKS}^{\infty}}/\lambda\}$. \end{proof} \subsection{Approximate solutions to Poisson equations} \label{subsec:approx-Poisson-equation} The formal derivation of MP estimators in Section~\ref{sec:general-continuous-setting} crucially relies on the fact that solutions to the discrete Poisson equation~\eqref{eqn:disc-Poisson-eqn}, which belong to functional spaces such as~$B^\infty_{\mKS}(\mY)$ when~$f \in \mathcal{S}$, can in fact be approximated by smooth functions, as given by~\eqref{eqn:approx-Poisson-solution} below. As in~\cite[Section~4.3]{leimkuhler2016computation} (see also~\cite{lelievre2016partial,plechac2019convergence}), we rely on approximate inverse operators to this end. We briefly recall in this section how these operators are constructed, and provide some technical estimates on expectations involving approximate solutions to Poisson equations. For the splitting schemes considered in Section~\ref{sec:splitting_Lang}, the evolution operator~$P_\dt$ can be expanded in powers of~$\dt$ as follows, for a given function $f\in \mathcal{S}$: \begin{equation} \label{eq:expansion_P_dt} P_{\dt}f = f + \dt\mathcal{A}_1f + \ldots + \dt^{\alpha+1}\mathcal{A}_{\alpha+1}f + \dt^{\alpha+2} r_{\alpha,\dt,f}{\,,} \end{equation} with $\alpha = 1$ for first order and $\alpha = 2$ for second order schemes. The operators~$\mathcal{A}_i$ can be systematically identified by the Baker--Campbell--Hausdorff formula~\cite[Section~III.4.2]{hairer2006geometric}, while the functions~$r_{\alpha, \dt,f}$ are in~$\mathcal{S}$ and there exists~$s \in \mathbb{N}$, $\dt^*>0$ and~$K\in \mathbb{R}_+$ such that~$\|r_{\alpha, \dt,f}\|_{B^\infty_{\mKS}} \leq K<\infty$ for~$\dt \in (0,\dt^*]$. We next note that the discrete time Poisson equation~\eqref{eqn:disc-Poisson-eqn} can be rewritten as \begin{equation} \label{eqn:discrete-Poisson-equation-projected} -\LOPERHPI \widehat{f}_{\dt} = f - \pi(f){\,,} \end{equation} with (recalling the definition~\eqref{eqn:projection-operator} for~$\Pi$) \[ \LOPERHPI = \Pi\LOPERH\Pi \equiv \Pi\left[\frac{1}{\dt}(P_h - I)\right]\Pi{\,.} \] In order to find an approximation to $\widehat{f}_{\dt}$, we introduce~${\widetilde{\mathcal{B}}_{\dt}} = {\widetilde{\mathcal{A}}}_2 + \dt {\widetilde{\mathcal{A}}}_3 + \ldots + \dt^{\alpha-1} {\widetilde{\mathcal{A}}}_{\alpha+1}$, where ${\widetilde{\mathcal{A}}}_j = \Pi \mathcal{A}_j \Pi$ for $j = 1, 2, \ldots, \alpha+1$ are operators mapping~$\mS_0$ to itself, so that $\LOPERHPI f ={\widetilde{\mathcal{A}}}_1 f + \dt {\widetilde{\mathcal{B}}_{\dt}} f + \dt^{\alpha+2} r_{\alpha,\dt,f}$. We next truncate the formal series expansion of the inverse \[ \left({\widetilde{\mathcal{A}}}_1 + \dt {\widetilde{\mathcal{B}}_{\dt}}\right)^{-1} = {\widetilde{\mathcal{A}}}_1^{-1} - \dt {\widetilde{\mathcal{A}}}_1^{-1}{\widetilde{\mathcal{B}}_{\dt}} {\widetilde{\mathcal{A}}}_1^{-1} + \dt^2 {\widetilde{\mathcal{A}}}_1^{-1}{\widetilde{\mathcal{B}}_{\dt}}{\widetilde{\mathcal{A}}}_1^{-1}{\widetilde{\mathcal{B}}_{\dt}}{\widetilde{\mathcal{A}}}_1^{-1} + \ldots{\,.} \] up to terms involving at most~$\alpha$ instances of~${\widetilde{\mathcal{B}}_{\dt}}$. This motivates introducing \[ \widetilde{Q}_{\dt} = {\widetilde{\mathcal{A}}}_1^{-1}\sum_{j=0}^{\alpha} (-1)^j \dt^j \left({\widetilde{\mathcal{B}}_{\dt}}{\widetilde{\mathcal{A}}}_1^{-1}\right)^j{\,,} \] which satisfies the following identity on~$\mathcal{S}_0$: \[ \LOPERHPI \widetilde{Q}_{\dt} f = \left({\widetilde{\mathcal{A}}}_1 + \dt {\widetilde{\mathcal{B}}_{\dt}}\right)\widetilde{Q}_{\dt} f + \dt^{\alpha+2} r_{\alpha,\dt,\widetilde{Q}_{\dt} f} = \Pi f + (-1)^{\alpha} \dt^{\alpha+1} \left({\widetilde{\mathcal{B}}_{\dt}}{\widetilde{\mathcal{A}}}_1^{-1}\right)^{\alpha+1} f + \dt^{\alpha+2} \widetilde{r}_{\alpha,\dt,f} {\,.} \] Note that the functions~$\widetilde{r}_{\alpha, \dt, f}$ are in~$\mathcal{S}$ and there exists~$s \in \mathbb{N}$, $\dt^*>0$ and~$K \in \mathbb{R}_+$ such that~$\|\widetilde{r}_{\alpha, \dt, f}\|_{B^\infty_{\mKS}} \leq K<\infty$ for any~$\dt \in (0,\dt^*]$. This can be seen from the fact that $\widetilde{r}_{\alpha, \dt,f} = r_{\alpha, \dt, \widetilde{Q}_{\dt} f}$ is a finite sum of terms of the form~$\dt^j r_{\alpha, \dt, g_j}$ with~$g_j \in \mS$. We finally define the {\it approximate inverse operator} $Q_{\dt}$ by substituting the formula for~${\widetilde{\mathcal{B}}_{\dt}}$ into $\widetilde{Q}_{\dt} $ and keeping only the terms of order at most~$\dt^{\alpha}$, \emph{i.e.} \[ Q_{\dt} \triangleq \mathcal{Q}_0 + \dt \mathcal{Q}_1 + \ldots + \dt^{\alpha-1}\mathcal{Q}_{\alpha-1} + \dt^{\alpha} \mathcal{Q}_{\alpha}{\,,} \] where the operators~$\mathcal{Q}_{j}$ for $0 \leq j \leq \alpha$ are operators mapping~$\mathcal{S}_0$ to itself. Therefore, $Q_{\dt}$ also maps~$\mS_0$ to itself. \begin{remark} \label{rmk:second_order_Q1} To derive MP estimators in Section~\ref{sec:MP_for_Langevin} (in particular, the second order MP estimator in Section~\ref{subsec:2nd-order-MP}), we only need the explicit expressions of $\mathcal{Q}_0 = {\widetilde{\mathcal{A}}}_1^{-1}$ and $\mathcal{Q}_1 = -{\widetilde{\mathcal{A}}}_1^{-1} {\widetilde{\mathcal{A}}}_2 {\widetilde{\mathcal{A}}}_1^{-1}$. For schemes of weak order~2, it holds~$\mathcal{A}_1 = \mL_{\gamma}$ and $\mathcal{A}_2 = \mL_{\gamma}^2/2$, so that~$\mathcal{Q}_1 = -\Pi/2$. \end{remark} We are now in a position to define the {\it approximate solution to the discrete time Poisson equation} \begin{equation} \label{eqn:approx-Poisson-solution} \widetilde{f}_{\dt} = -Q_{\dt}(f - \pi(f)){\,.} \end{equation} We emphasize that the function~$\widetilde{f}_{\dt}$ indeed belongs to~$\mathcal{S}_0$ when~$f \in \mS$, while the solution~$\widehat{f}_{\dt}$ to the Poisson equation~\eqref{eqn:disc-Poisson-eqn-general} does not. Applying the operator $-\LOPERHPI$ to both sides of~\eqref{eqn:approx-Poisson-solution} leads to \begin{equation} \label{eqn:phi} % -\LOPERHPI \widetilde{f}_{\dt} = f - \pi(f) + \dt^{\alpha+1} \phi_{\alpha, \dt, f} \end{equation} for some function $\phi_{\alpha, \dt, f} \in \mathcal{S}_0 $, and there exists~$s \in \mathbb{N}$, $\dt^*>0$ and~$K\in \mathbb{R}_+$ such that \begin{equation} \label{eqn:control-of-phi} \forall \dt \in (0,\dt^*], \qquad \|\phi_{\alpha, \dt,f}\|_{B^\infty_{\mKS}} \leq K {\,.} \end{equation} Moreover, in view of~\eqref{eqn:K-evolution-estimate}, for any $a^* \in \mathbb{N}$, upon possibly increasing~$s$ and~$K$ and decreasing~$\dt^*$, \begin{equation} \label{eq:second-moment-bound-widetilde} \forall a \in \{1, \ldots, a^*\},\qquad \forall \dt \in (0,\dt^*], \qquad \forall n \geq 0, \qquad \bE_{\dt}\left\{ \left|\widetilde{f}_{\dt}(y^n)\right|^{2a} \right\} \leq K \left\|f\right\|_{B_{\mKS}^{\infty}}^{2a} {\,.} \end{equation} Finally, from the definitions~\eqref{eqn:cont-Poisson-eqn} and~\eqref{eqn:approx-Poisson-solution}, we can compare~$\widetilde{f}_{\dt}$ with~$\widehat{f}$ when $\mathcal{A}_1 = \mL_{\gamma}$ (\emph{i.e.} the numerical scheme is at least of weak order~1): \begin{equation} \label{eqn:diff-discrete-continuous-Poisson-pointwise} \widetilde{f}_{\dt}-\widehat{f} = -\left[\dt \mathcal{Q}_1 + \dt^2 \mathcal{Q}_2 + \ldots + \dt^{\alpha-1}\mathcal{Q}_{p-1} + \dt^{\alpha} \mathcal{Q}_{\alpha}\right](f - \pi(f)){\,.} \end{equation} In view of Remark~\ref{rmk:second_order_Q1}, this corresponds to the equalities~\eqref{eqn:approx-operator-preview} in the general argument of Section~\ref{sec:general-continuous-setting}. \section{Martingale product estimators for Langevin dynamics} \label{sec:MP_for_Langevin} We are now able to derive the MP estimators for the linear response of Langevin dynamics, by following the general strategy presented in Section~\ref{sec:general-continuous-setting}. We first discuss in Section~\ref{sec:choice_g_Lang} the choice of the function~$g$ appearing in the MP estimator, which allows us to study the consistency of both the first and second order MP estimators in Sections~\ref{subsec:1st-order-MP} and~\ref{subsec:2nd-order-MP}, respectively. The boundedness of the variance of MP estimators is proved in Section~\ref{subsec:MP-var-analysis}. In all this section, we rewrite the elementary evolutions associated with first or second order splittings as \[ y^{n+1} = y^n + \Phi_{\dt}(y^n; G^n), \qquad \Phi_{\dt}(y; G) = \begin{pmatrix} \Phi_{\dt}^q(y; G) \\ \Phi_{\dt}^p(y; G) \end{pmatrix} {\,,} \] where the expression of the increment function~$\Phi_{\dt}$ depends on the specific scheme under consideration, and $(G^n)_{n \geq 0}$ are independent and identically distributed standard Gaussian random vectors. It may be the case that two independent Gaussian vectors $G^{n,1},G^{n,2}$ are needed per timestep for second order splittings such as~$\mathrm{e}^{\dt C/2} \mathrm{e}^{\dt A/2} \mathrm{e}^{\dt B} \mathrm{e}^{\dt A/2} \mathrm{e}^{\dt C/2}$, in which case~$G^n=(G^{n,1},G^{n,2})$. \subsection{Choice of the function~$g$ in the discrete auxiliary martingale} \label{sec:choice_g_Lang} Given our specific choice for the external forcing $F(q)$, the equation~\eqref{eqn:constraint-g} on the function $g$ which appears in the MP estimator is easily solved by choosing \begin{equation} \label{eq:solutio_g_MP_Langevin} g(q,p) = \frac{\beta}{2\gamma} p^\T F(q). \end{equation} The solution is not unique\footnote{We do not make use of this flexibility here, but an interesting question is whether one of the possible solutions leads to a final estimator with minimal variance.} since any function of the~$q$ variable only can be added to~$g$. The important point here is that we have an explicit solution which can be used to construct the numerical scheme. \begin{remark}\label{rem:gfun} For more general forcings depending both on momenta and positions, i.e., forces $F(q,p)$, a solution to~\eqref{eqn:constraint-g} is given by \begin{equation}\label{eqn:gfun} g(q,p) = \frac{\beta}{2\gamma} \left(\nabla_p^* \nabla_p \right)^{-1} \nabla_p^* F(q,p)\,, \end{equation} where $\nabla_p^*$ is defined in~\eqref{eq:PDE_g}. This function is well-defined $q$ by $q$, since $\nabla_p^* F$ has zero average with respect to the marginal measure~$\kappa(dp)$ of~$\pi$ in the $p$-variable (\emph{i.e.}, $\kappa$ is the Gaussian measure in~$p$ with mean zero and covariance~$M\beta$), and $\nabla_p^* \nabla_p$ is coercive on the subspace of functions in~$L^2(\kappa)$ with mean zero with respect to~$\kappa$ (thanks to the Poincar\'e inequality satisfied by~$\kappa$). On the other hand, the expression for~$g$ is not explicit in general. This prevents explicitly constructing the discrete martingale by expansions in powers of~$\dt$ of~${\Delta \mathcal{M}_{n}}(g)$. When the expression of~$g$ is explicit (which is the case when $F(q,p)$ depends polynomially in~$p$, for example), the methodology presented in this section can be adapted in a straightforward way, see Appendix~\ref{app:secondorder} for more details. \end{remark} \subsection{First order martingale product estimator} \label{subsec:1st-order-MP} A simple computation shows that, for all the first order splitting schemes, \begin{equation} \label{eq:expansion_first_order_splitting} \Phi_{\dt}^p(y^n; G^n) = \dt^{1/2}\sqrt{\frac{2\gamma}{\beta}} G^n + \Delta_{\dt}(y^n; G^n), \end{equation} where the remainder term~$\Delta_{\dt}(y^n; G^n)$ and~$\Phi_\dt^q$ are of order~$\dt$, in the following sense: for any~$a^* \in \mathbb{N}$, there exists~$C_* \in \mathbb{R}_+$ and~$\dt^* >0$ such that \begin{equation} \label{eq:remainders_first_order_splitting} \forall a \in \{1,\ldots,a^*\}, \quad \forall \dt \in (0,\dt^*], \qquad \bE_{\dt}\{ |\Delta_{\dt}(y^n;G^n)|^a\} + \bE_{\dt}\{ |\Phi_\dt^q(y^n;G^n)|^a\} \leq C_* \dt^{a}. \end{equation} As heuristically derived in Section~\ref{sec:first_order_general}, the first order MP estimator based on an arbitrary first order splitting scheme is \begin{equation} \label{eqn:first-order-MP-estimator} \mathcal{E}_{\dt, N}^{\mathrm{MP}1}(f) =\frac{1}{N} \sum_{n=0}^{N-1}\left(f(y^n) - \pi_{\gamma, \dt}(f)\right) z^N{\,,} \end{equation} where \begin{equation} \label{eqn:1st-order-weight} z^N = \dt^{1/2} \sqrt{\frac{\beta}{2\gamma}} \sum_{n=0}^{N-1} F(q^n)^{\T} G^n \end{equation} is the auxiliary discrete martingale, obtained from expanding $g(y^{n+1}) - (P_h g)(y^n)$ up to order $\sqrt{h}$ in~\eqref{eqn:general-first-order-MP} and replacing~$g$ by its expression~\eqref{eq:solutio_g_MP_Langevin}. Note that the weight process remains the same for all the first order splittings considered in Section~\ref{sec:splitting_Lang}. The first order MP estimator coincides with the CLR estimator (see~\cite{plechac2019convergence} for the overdamped Langevin setting). The benefit of the proposed martingale product derivation is that it provides a systematic approach for discovering higher order MP schemes, as discussed in Section~\ref{subsec:2nd-order-MP}. The reason why we carefully study first order estimators is that the proof of their consistency paves the way for the proof of consistency of second order estimators. The main result concerning the consistency of first order MP estimators is the following. \begin{theorem}[Bias of first order MP estimators] \label{thm:1st-order-bias} Consider any of the first order splitting schemes introduced in Section~\ref{sec:splitting_Lang}, and fix an observable~$f \in \mS$. Then there exist $\dt^* > 0$ and $C \in \mathbb{R}_+$ such that, for any $0<\dt \leq \dt^*$ and $N \geq 1$ satisfying~$N\dt \geq 1$, \begin{equation}\label{eqn:1st-order-MP-bias} \left|\bE_{\dt}\left\{ \mathcal{E}_{\dt, N}^{\mathrm{MP}1}(f)\right\} - \mathrm{Lin} (f)\right| \leq C\left(\dt + \frac{1}{\sqrt{N\dt}}\right){\,.} \end{equation} \end{theorem} \begin{proof} Throughout the proof, we denote by~$C$ a positive constant that may vary line by line (but does not depend on~$N$ and~$\dt$) and by~$s$ a sufficiently large positive integer. We also repeatedly use the following bound, directly obtained from~\eqref{eqn:K-evolution-estimate}: for given functions~$\varphi_1,\varphi_2 \in \mathcal{S}$ (so that $\varphi_1 \varphi_2 \in \mS$), there exist~$K \in \mathbb{R}_+$ and~$\dt > 0$ such that \begin{equation} \label{eq:bound_sum_expectation_functions} \forall \dt \in (0,\dt^*], \quad \forall N \geq 1, \qquad \frac{1}{N} \left| \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \varphi_1(y^n) \varphi_2(y^n) \right\}\right| \leq K. \end{equation} We start by rewriting the MP estimator in a form more amenable for our estimates. Since $\bE_{\dt}\{z^N\} = 0$, one can easily verify that \[ \bE_{\dt}\left\{ \left(f(y^n) - \pi_{\gamma, \dt}(f)\right) z^N \right\} = \bE_{\dt}\left\{ \left(f(y^n) - \pi(f)\right) z^N \right\}. \] In view of~\eqref{eqn:phi} (with~$\alpha=1$), it holds \begin{equation} \label{eqn:1st-order-estimate} \begin{aligned} & \bE_{\dt}\left\{ \mathcal{E}_{\dt, N}^{\mathrm{MP}1}(f)\right\} = \frac{1}{N\dt} \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \left(\Pi (I - P_{\dt}) \Pi \widetilde{f}_{\dt}(y^n)\right) z^N \right\} - \frac{\dt^2}{N} \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \phi_{1, \dt, f}(y^n) z^N \right\} \\ & \qquad = \frac{1}{N\dt} \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \left( \widetilde{f}_{\dt}(y^{n+1}) - (P_{\dt} \widetilde{f}_{\dt})(y^n) \right)z^N \right\} + \frac{1}{N\dt} \bE_{\dt}\left\{ \left( \widetilde{f}_{\dt}(y^0) - \widetilde{f}_{\dt}(y^N) \right) z^N \right\}\\ & \qquad \ \ \ - \frac{\dt^2}{N} \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \phi_{1, \dt, f}(y^n) z^N \right\}{\,.} \end{aligned} \end{equation} Let us first show that the last two terms in the last right hand side vanish as~$\dt \to 0$. Since the functions~$\phi_{1, \dt, f}$ belong to some space~$B_{\mKS}^{\infty}(\mY)$ for~$\dt$ sufficiently small by~\eqref{eqn:control-of-phi}, the estimates in Proposition~\ref{prop:elementary-term} immediately imply that \begin{equation} \label{eqn:1st-order-estimate-1} \frac{\dt^2}{N} \left| \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \phi_{1, \dt, f}(y^n) z^N \right\} \right| \leq C\dt^{3/2} \end{equation} for $\dt$ sufficiently small. For the second term in the last right hand side of~\eqref{eqn:1st-order-estimate}, we use~\eqref{eq:bound_sum_expectation_functions} and the assumption that each component of $F$ is a smooth function to write \begin{equation} \label{eq:bound_zN^2} \bE_{\dt}\left\{\left(z^N\right)^2\right\} = \frac{\beta\dt}{2\gamma}\sum_{n=0}^{N-1} \bE_\dt\left\{ \left|F(q^n)\right|^2 \right\}\leq C N\dt {\,.} \end{equation} A Cauchy--Schwarz inequality together with~\eqref{eq:second-moment-bound-widetilde} then leads to \begin{equation}\label{eqn:1st-order-estimate-2} \frac{1}{N\dt} \left| \bE_{\dt}\left\{ \left( \widetilde{f}_{\dt}(y^0) - \widetilde{f}_{\dt}(y^N) \right) z^N \right\} \right| \leq \frac{C}{\sqrt{N\dt}} \left\|f \right\|_{B_{\mKS}^{\infty}}, \end{equation} which contributes to the finite integration time error~$C /\sqrt{N\dt}$ in the bias. It only remains to estimate the difference between the first term on the right hand side of~\eqref{eqn:1st-order-estimate} and~$\mathrm{Lin}(f)$. Since the martingale increments~$z^{m+1}-z^m$ and~${\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt})$ are conditionally independent for $m \neq n$ (recalling the notation in~\eqref{eqn:obs_increment}), \[ \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \left( \widetilde{f}_{\dt}(y^{n+1}) - (P_{\dt} \widetilde{f}_{\dt})(y^n) \right) z^N \right\} = \sum_{n=0}^{N-1} \bE_{\dt}\left\{ {\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt})(z^{n+1} - z^n) \right\}. \] Now, by a Taylor expansion and in view of~\eqref{eq:remainders_first_order_splitting}, \begin{equation} \label{eq:Taylor_WTF} {\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt}) = \sqrt{\frac{2\gamma\dt}{\beta}} \left(G^n\right)^\T \nabla_p \widetilde{f}_{\dt}(y^n) + \dt \psi(y^n;G^n) + \dt^{3/2}\Upsilon_\dt(y^n;G^n) {\,,} \end{equation} with $\bE_\dt(\psi (y^n;G^n) G^n) = 0$ and terms~$\psi(y^n;G^n),\Upsilon_\dt(y^n;G^n)$ of order~1 in the following sense: for any~$a \geq 1$, there exist~$C \in \mathbb{R}_+$ and~$\dt^*>0$ such that, for any~$0 \leq n \leq N$ and~$\dt \in (0,\dt^*]$, \begin{equation} \label{eq:moment_bounds_psi_Upsilon} \bE_{\dt}\left\{ |\psi (y^n;G^n)|^a \right\} \leq C, \qquad \bE_{\dt}\left\{ |\Upsilon_\dt (y^n;G^n)|^a \right\} \leq C. \end{equation} A Cauchy--Schwarz inequality then gives \[ \left|\sum_{n=0}^{N-1} \bE_{\dt}\left\{ \Upsilon_\dt (y^n;G^n)(z^{n+1} - z^n) \right\}\right| \leq \left(\sum_{n=0}^{N-1} \bE_{\dt}\left\{ \Upsilon_\dt (y^n;G^n)^2 \right\}\right)^{1/2}\left( \sum_{n=0}^{N-1} \bE_{\dt}\left\{ (z^{n+1}-z^n)^2\right\}\right)^{1/2}, \] so that, since the second sum on the right hand side of the above inequality is bounded by~$CN\dt$, \begin{equation} \label{eq:1st-order-bound_remainder_martingale} \frac{1}{N\dt} \left|\sum_{n=0}^{N-1} \bE_{\dt}\left\{ \left[ \dt \psi(y^n;G^n)+\dt^{3/2} \Upsilon_\dt (y^n;G^n)\right](z^{n+1} - z^n) \right\}\right| \leq C \dt. \end{equation} Moreover, by~\eqref{eqn:diff-discrete-continuous-Poisson-pointwise} with~$\alpha=1$, \[ \begin{aligned} & \frac{1}{N\dt} \sqrt{\frac{2\gamma\dt}{\beta}} \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \left(G^n\right)^\T \nabla_p \widetilde{f}_{\dt}(y^n) (z^{n+1} - z^n) \right\} = \frac{1}{N}\sum_{n=0}^{N-1} \bE_{\dt}\left\{ F(y^n)^\T \nabla_p \widetilde{f}_{\dt}(y^n) \right\} \\ & \qquad = \frac{1}{N}\sum_{n=0}^{N-1} \bE_{\dt}\left\{ F(y^n)^\T \nabla_p \widehat{f}(y^n) \right\} - \frac{\dt}{N}\sum_{n=0}^{N-1} \bE_{\dt}\left\{ F(y^n)^\T \nabla_p \mathcal{Q}_1 \Pi f(y^n) \right\} {\,.} \end{aligned} \] The second sum in the previous equality is uniformly bounded by~$C\dt$ in view of~\eqref{eq:bound_sum_expectation_functions}, while we use Theorem~\ref{thm:bias-splitting} for the first one: \begin{equation} \label{eq:1st-order_estimates_dominant_term} \left| \frac{1}{N}\sum_{n=0}^{N-1} \bE_{\dt}\left\{ F(y^n)^\T \nabla_p \widehat{f}(y^n) \right\} - \int_{\mY} F^\T \nabla_p\widehat{f} \,d\pi \right| \leq C \left( \dt + \frac{1}{N\dt} \right). \end{equation} The integral on the left-hand side of the previous equality is equal to~$\mathrm{Lin}(f)$ by the reformulation~\eqref{eqn:Langevin-lin-res-reformulation} for the linear response. The desired bias estimate finally follows by combining~\eqref{eqn:1st-order-estimate} and the estimates provided by~\eqref{eqn:1st-order-estimate-1}, \eqref{eqn:1st-order-estimate-2}, \eqref{eq:1st-order-bound_remainder_martingale} and~\eqref{eq:1st-order_estimates_dominant_term}. \end{proof} \subsection{Second order martingale product estimator} \label{subsec:2nd-order-MP} We present in this section second order MP estimators based on second order splittings. For first order MP estimators, the auxiliary discrete martingales~$\{z^n\}_{n \geq 0}$ are identical for different first order splittings. However, this is not the case for second order MP estimators. We choose to first present in Section~\ref{sec:general_2nd_order} a general result on second order MP estimators, following the strategy outlined in Section~\ref{sec:construction_discrete_martingale}. The so-obtained estimators are however not explicit, which is why we make them precise in Section~\ref{sec:specific_2nd_order} for the various splitting schemes we consider. \subsubsection{General second order estimators.} \label{sec:general_2nd_order} The second order MP estimator obtained from~\eqref{eqn:general-second-order-MP} is \begin{equation} \mathcal{E}_{\dt, N}^{\mathrm{MP}2} = \frac{1}{N}\sum_{n=0}^{N-1}\left[ f(y^n) -\pi_{\gamma,\dt}(f)\right] {\Delta \mathcal{M}_{n}}\left(g-\frac{\dt}{2}\mL g\right)\,, \label{eq:MP2_estimator} \end{equation} where we recall \[ {\Delta \mathcal{M}_{n}}\left(g-\frac{\dt}{2}\mL g\right) = \left(g-\frac{\dt}{2}\mL g\right)(y^{n+1}) - P_{\dt} \left(g-\frac{\dt}{2}\mL g\right) (y^n){\,.}\notag \] The second order consistency of this estimator is rigorously established by the following result. \begin{theorem}[Bias of second order MP estimators] \label{thm:2nd-order-bias} Consider any of the second order splitting schemes introduced in Section~\ref{sec:LD-application}, and fix an observable $f \in \mS$. Then there exist $\dt^* > 0$ and $C \in \mathbb{R}_+$ such that, for any $0<\dt \leq \dt^*$ and $N \geq 1$, \begin{equation} \label{eqn:2nd-order-MP-bias} \left|\bE_{\dt}\left\{ \mathcal{E}_{\dt, N}^{\mathrm{MP}2}(f)\right\} - \mathrm{Lin} (f)\right| \leq C\left(\dt^2 + \frac{1}{\sqrt{N\dt}}\right){\,.} \end{equation} \end{theorem} \begin{proof} Throughout the proof, we denote by~$C$ a positive constant and by~$s$ a positive integer that is sufficiently large. These two numbers may vary line by line. We introduce the auxiliary discrete martingale \[ z^N = \sum_{n=0}^{N-1} {\Delta \mathcal{M}_{n}}\left(g-\frac{\dt}{2}\mL g\right). \] First note that~\eqref{eq:expansion_first_order_splitting} and~\eqref{eq:remainders_first_order_splitting} hold true for all second order splitting schemes as well. Furthermore, from~\eqref{eqn:K-evolution-estimate} and together with the fact that~$g,\mL g \in \mS$, \[ \bE_\dt\left\{ \left(z^{n+1}-z^n\right)^2 \right\} \leq 2\bE_\dt\left\{ {\Delta \mathcal{M}_{n}}(g)^2 \right\} + \frac{\dt^2}{2}\bE_\dt\left\{ {\Delta \mathcal{M}_{n}}(\mL g)^2 \right\} \leq D \dt{\,,} \] for some constant $D> 0$. Similarly to~\eqref{eqn:1st-order-estimate}, using~\eqref{eqn:phi} with~$\alpha=2$ and the conditional independence of the martingale increments $z^{m+1}-z^m$ and~$\widetilde{f}_{\dt}(y^{n+1}) - (P_{\dt} \widetilde{f}_{\dt})(y^n)$ when $m \neq n$, \begin{equation*} \begin{split} \bE_{\dt}\left\{ \mathcal{E}_{\dt, N}^{\mathrm{MP}2}(f)\right\} & = \frac{1}{N\dt} \sum_{n=0}^{N-1} \bE_{\dt}\left\{ {\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt})(z^{n+1}-z^n) \right\}\\ & \ \ \ \ + \frac{1}{N\dt} \bE_{\dt}\left\{ \left( \widetilde{f}_{\dt}(y^0) - \widetilde{f}_{\dt}(y^n) \right) z^N \right\} - \frac{\dt^3}{N} \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \phi_{2, \dt, f}(y^n) z^N \right\} {\,.} \end{split} \end{equation*} The second and third terms on the right hand side of the above equality can be uniformly controlled as in the proof of Theorem~\ref{thm:1st-order-bias}. Now, in view of~\eqref{eq:crucial_identity}, \[ \bE_{\dt}\left\{ {\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt})(z^{n+1}-z^n) \right\} % = \dt \bE_\dt\left\{ \Gamma_\dt\left(\widetilde{f}_{\dt},g-\frac{\dt}{2}\mL g\right)(y^n)\right\}, \] so that, with~\eqref{eqn:bias-splitting}, \[ \left| \frac{1}{N\dt} \sum_{n=0}^{N-1} \bE_{\dt}\left\{ {\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt})(z^{n+1}-z^n) \right\} - \int_\mY \Gamma_\dt\left(\widetilde{f}_{\dt},g-\frac{\dt}{2}\mL g\right) d\pi \right| \leq C \left(\dt^2 + \frac{1}{N\dt}\right) {\,,} \] where the constant~$C$ depends on~$F$ through~$g$, and on the observable~$f$ under consideration. At this stage, it only remains to show that the difference between~$\mathrm{Lin}(f)$ and the integral on the left hand side of the previous inequality is of order~$\dt^2$. To this end, we first use~\eqref{eq:expansion_P_dt} (which allows us to give a rigorous meaning to~\eqref{eqn:order2}) and~\eqref{eqn:diff-discrete-continuous-Poisson-pointwise} (with~$\mathcal{Q}_1\varphi = -\Pi\varphi/2$ for second order splitting schemes) to write \[ \Gamma_\dt\left(\widetilde{f}_{\dt},g-\frac{\dt}{2}\mL g\right) = \Gamma\left(\widehat{f},g\right) + \dt\left[\Lambda\left(\widehat{f},g\right)- \frac12 \Gamma(\mL \widehat{f},g) - \frac12 \Gamma\left(\widehat{f},\mL g\right) \right] + \dt^2 r_{\dt,f} {\,,} \] with $\|r_{\dt,f}\|_{B^\infty_{\mKS}} \leq K$ for~$\dt$ sufficiently small. The result then follows from~\eqref{eqn:integrate-scale-function} and the fact that $\Gamma\left(\widehat{f},g\right) = F^\T \nabla_p \widehat{f}$, while the order~$\dt$ term in the expansion of~$\Gamma_\dt\left(\widetilde{f}_{\dt},g-\dt\mL g/2\right)$ in powers of~$h$ has average~0 with respect to~$\pi$ in view of~\eqref{eq:int_Lambda_pi}. \end{proof} \begin{remark} A closer inspection of the proof of Theorem~\ref{thm:2nd-order-bias} shows that similar results can be obtained for other discretization schemes as long as ${\widetilde{\mathcal{A}}}_1 = \mL_{\gamma}$ and~${\widetilde{\mathcal{A}}}_2$ is proportional, but not necessarily equal, to~$\mL_{\gamma}^2$; see~\cite{FS17} for examples of such schemes, based on a Barker acceptance/rejection rule together with a second order discretization of overdamped Langevin dynamics. \end{remark} \subsubsection{Application to second order splitting schemes.} \label{sec:specific_2nd_order} We now provide a more explicit expression for the estimator~\eqref{eq:MP2_estimator} depending on the numerical scheme at hand. We illustrate our approach for the second order splitting~$P_{\dt}^{B,A,\gamma C, A, B}$ considered in~\eqref{eqn:BACAB} and then discuss the extension to other second order splittings. The strategy is to expand the martingale increments~${\Delta \mathcal{M}_{n}}(g-\dt\mL g/2)$ up to order~$\dt^{3/2}$ as \begin{equation} \label{eq:expansion_martingale_increment_2nd_order} \begin{aligned} {\Delta \mathcal{M}_{n}}\left(g-\frac{\dt}{2}\mL g\right) = h^{1/2}\sqrt{\frac{\beta}{2\gamma}} F(y^n)^{\T} G^n & + \dt \psi_1(y^n,G^n) + \dt^{3/2}\psi_{3/2}(y^n,G^n) \\ & + \dt^2 \psi_2(y^n,G^n) + \dt^{5/2} \Upsilon_\dt(y^n,G^n), \end{aligned} \end{equation} where the dominant term on the right hand side of course agrees with the discrete martingale increments in~\eqref{eqn:1st-order-weight} for the first order MP estimator. It then suffices to replace the last factor in~\eqref{eq:MP2_estimator} by the first three terms above. The remainder terms involving~$\psi_2$ and~$\Upsilon_\dt$ which arise when quantifying the bias of the estimator can be uniformly controlled by noting that $\bE_\dt(\psi_2(y^n,G^n) G^n) = 0$ and using uniform estimates on moments of~$\psi_2$ and~$\Upsilon_\dt$. We do not explicitly perform these manipulations as they follow very closely estimates used in the proofs of Theorems~\ref{thm:1st-order-bias} and~\ref{thm:2nd-order-bias}. The first step towards an equality such as~\eqref{eq:expansion_martingale_increment_2nd_order} is to expand the increment function of the scheme under consideration up to the order of $\dt^{3/2}$. For~\eqref{eqn:BACAB}, \begin{equation} \label{eq:expansion_BACAB} \begin{split} \Phi_{\dt}^q(y^n; G^n) &= \dt M^{-1} p^n + \dt^{3/2} \sqrt{\frac{\gamma}{2\beta}}M^{-1} G^n + \Delta_{\dt}^q(y^n; G^n),\\ \Phi_{\dt}^p(y^n; G^n) &= h^{1/2}\sqrt{\frac{2\gamma}{\beta}} G^n - \dt (\nabla V(q^n) + \gamma M^{-1} p^n) - \dt^{3/2}\sqrt{\frac{\gamma}{2\beta}}\gamma M^{-1} G^n + \Delta_{\dt}^p(y^n; G^n), \end{split} \end{equation} with remainders terms of order~$\dt^2$: for any $a^* \in \mathbb{N}$, there exists~$C_* \in \mathbb{R}_+$ and~$\dt^* >0$ such that \[ \forall a \in \{1,\ldots,a^*\}, \quad \forall \dt \in (0,\dt^*], \qquad \bE_{\dt}\{ |\Delta_{\dt}^q(y^n;G^n)|^a\} + \bE_{\dt}\{ |\Delta_{\dt}^p(y^n;G^n)|^a\} \leq C_* \dt^{2a} {\,.} \] This leads to the following second order MP estimator for the scheme~\eqref{eqn:BACAB}: \begin{equation} \label{eqn:second-order-MP-estimator} \mathcal{E}_{\dt, N}^{\mathrm{MP}2,\mathrm{BACAB}}(f) =\frac{1}{N} \sum_{n=0}^{N-1}\left(f(y^n) - \pi_{\gamma, \dt}(f)\right) z^N, \end{equation} with the second order weight process { \begin{equation} \label{eqn:2nd-order-weight} z^N = \sqrt{\frac{\beta}{2\gamma}} \sum_{n=0}^{N-1} \dt^{1/2} F(q^n)^\T G^n + \frac{1}{2}\dt^{3/2} (p^n)^\T M^{-1} \nabla_q F(q^n) G^n \,, \end{equation} } obtained from $z^N=\sum_{n=0}^{N-1} {\Delta \mathcal{M}_{n}}(g-\dt\mathcal{L}g/2)$ by truncating the expansion in \eqref{eq:expansion_martingale_increment_2nd_order}. See Appendix~\ref{app:secondorder} for more detailed calculations. We conclude this section by providing the expressions for the auxiliary discrete martingale~$z^N$ associated with other splitting schemes: \begin{itemize} \item[(i)] For the scheme $P_{\dt}^{A, B, \gamma C, B, A}$, the same expansion as~\eqref{eq:expansion_BACAB} holds, so that the second order MP estimator for this scheme coincides with~\eqref{eqn:second-order-MP-estimator}. \item[(ii)] For the scheme $P_{\dt}^{\gamma C, B, A, B, \gamma C}$ (recall~\eqref{eqn:CBABCB}), the counterpart of the expansion~\eqref{eq:expansion_BACAB} reads \begin{equation} \label{eq:expansion_CBABC} \begin{split} \Phi_{\dt}^q(y^n; G^n) &= \dt M^{-1} p^n + \dt^{3/2} \sqrt{\frac{\gamma}{\beta}}M^{-1} G_1^n + \Delta_{\dt}^q(y^n; G^n)\,,\\ \Phi_{\dt}^p(y^n; G^n) &= h^{1/2}\sqrt{\frac{\gamma}{\beta}} (G_1^n+G_2^n) - \dt (\nabla V(q^n) + \gamma M^{-1} p^n) \\ & \;\;\; -\frac{\dt^{3/2}}{4}\sqrt{\frac{\gamma}{\beta}}\gamma M^{-1} (3G_1^n+G_2^n) + \Delta_{\dt}^p(y^n; G^n)\,. \end{split} \end{equation} The expansion is in fact the same for the scheme~$P_{\dt}^{\gamma C, A, B, A, \gamma C}$. A second order MP estimator is then obtained for these two schemes by choosing (see Appendix~\ref{app:secondorder} for more detailed calculations) \begin{equation}\label{eqn:2nd-order-weight-CBABC} \begin{aligned} z^{n+1} - z^n & = \frac{1}{2}\dt^{1/2}\sqrt{\frac{\beta}{\gamma}} F(q^n)^{\T} (G_1^n + G_2^n)\\ & + \frac{\dt^{3/2}}{4} \sqrt{\frac{\beta}{\gamma}} \left[ (p^n)^\T M^{-1} \nabla_q F(q^n) (G_1^n+G_2^n)+ (p^n)^\T \nabla_q F(q^n)^\T M^{-1} (G_1^n-G_2^n) \phantom{\frac12}\right.\\ & \qquad \qquad \qquad \qquad \qquad \qquad\left. - \frac{\gamma}{2} F(q^n)^\T M^{-1} (G_1^n - G_2^n)\right]\,. \end{aligned} \end{equation} \end{itemize} We do not consider the second order splitting schemes associated with the semigroups $P_{\dt}^{B, \gamma C, A, \gamma C, B}$ and $P_{\dt}^{A, \gamma C, B, \gamma C, A}$ since the marginal in the position variable of the invariant measures of these numerical schemes in the overdamped limit are not consistent with the Boltzmann--Gibbs measure with density proportional to~$\mathrm{e}^{-\beta V(q)}$ \cite{leimkuhler2016computation}. \subsection{Variance analysis of martingale product estimators} \label{subsec:MP-var-analysis} The key advantage of the martingale product estimator is that its variance does not grow with respect to the integration time, as stated in the following result. As will be seen below, this is in sharp contrast with estimators based on Green--Kubo formulas for instance (see Theorem~\ref{thm:GK-variance}). \begin{theorem}[Variance of MP estimators] \label{thm:MP-variance} Fix an observable $f \in \mS$, and consider any of the first ($\alpha=1$) or second ($\alpha=2$) order splitting schemes introduced in Section~\ref{sec:LD-application}, and their associated MP estimators constructed in~\eqref{eqn:first-order-MP-estimator}-\eqref{eqn:1st-order-weight} and Section~\ref{subsec:2nd-order-MP} (such as~\eqref{eq:MP2_estimator}), respectively. There exist $\dt^* > 0$ and $C \in \mathbb{R}_+$ such that, for any $\dt\in (0,\dt^*]$ and any integer $N\geq 1$ satisfying~$N\dt \geq 1$, \[ \mathrm{Var}_{\dt}\left(\mathcal{E}_{\dt,N}^{\mathrm{MP}\alpha} \right) \leq C\left(1+\frac{1}{N\dt}\right){\,.} \] \end{theorem} \begin{proof} For the ease of notation, we write the proof for the first order MP estimator~\eqref{eqn:first-order-MP-estimator}-\eqref{eqn:1st-order-weight}, but the estimates are completely similar for second order estimators. The proof closely follows the lines of the proof of~\cite[Theorems~4.3 and~4.6]{plechac2019convergence}. Here as well, $C$ denotes a generic constant which may change from line to line. The index~$s$ of Lyapunov functions can also change from line to line. In view of the discrete time Poisson equation~\eqref{eqn:disc-Poisson-eqn}, we can rewrite $\mathcal{E}_{\dt, N}^{\mathrm{MP}1}$ as \[ \mathcal{E}_{\dt, N}^{\mathrm{MP}1} = \frac{1}{N\dt} \sum_{n=0}^{N-1} \left( \widehat{f}_{\dt}(y^{n+1}) - (P_{\dt} \widehat{f}_{\dt})(y^n) \right)z^N + \frac{1}{N\dt} \left( \widehat{f}_{\dt}(y^0) - \widehat{f}_{\dt}(y^N) \right) z^N {\,,} \] so that, using $\mathrm{Var}_{\dt}\left( \mathcal{E}_{\dt, N}^{\mathrm{MP}1} \right) \leq \bE_{\dt}\left\{ \left( \mathcal{E}_{\dt, N}^{\mathrm{MP}1} \right)^2 \right\}$ and a Cauchy--Schwarz inequality, \begin{equation} \label{eqn:var-estimate} \begin{aligned} \mathrm{Var}_{\dt}\left( \mathcal{E}_{\dt, N}^{\mathrm{MP}1} \right) & \leq \frac{2}{N^2\dt^2} \mathbb{E}_{\dt}\left\{\left(\sum_{n=0}^{N-1}{\Delta \mathcal{M}_{n}}(\widehat{f}_{\dt}) z^N\right)^2 \right\} + \frac{2}{N^2\dt^2}\mathbb{E}_{\dt}\left\{ \left[\left(\widehat{f}_{\dt}(y^N) - \widehat{f}_{\dt}(y^0)\right)z^N \right]^2\right\} {\,.} \end{aligned} \end{equation} We start by estimating the first term on the right hand side of~\eqref{eqn:var-estimate}. Denoting by~$\DN_n = z^{n+1} - z^n$ for notational homogeneity, a simple computation shows that \begin{equation*} \begin{split} & \bE_{\dt}\left\{\left(\sum_{n=0}^{N-1}{\Delta \mathcal{M}_{n}}\!\left(\widehat{f}_{\dt}\right) z^N\right)^2 \right\} = \sum_{n_1, n_2, n_3, n_4=0}^{N-1}\bE_{\dt}\left\{ \DM_{n_1}\!\!\left(\widehat{f}_{\dt}\right)\DM_{n_2}\!\!\left(\widehat{f}_{\dt}\right) \DN_{n_3} \DN_{n_4}\right\}\\ & = \sum_{n_1,n_2=0}^{N-1} \bE_{\dt}\left\{ \DM_{n_1}\!\!\left(\widehat{f}_{\dt}\right)^2 \DN_{n_2}^2 \right\} + 2\!\!\!\!\!\!\!\!\! \sum_{0 \leq n_1 < n_2 \leq N-1}\!\!\!\!\!\!\!\!\!\bE_{\dt}\left\{\DM_{n_1}\!\!\left(\widehat{f}_{\dt}\right)\DM_{n_2}\!\!\left(\widehat{f}_{\dt}\right)\DN_{n_1}\DN_{n_2}\right\}{\,,} \end{split} \end{equation*} where the second equality holds because the underlying martingale increments are conditionally independent between successive steps. We next replace~$\widehat{f}_{\dt}$ by~$\widetilde{f}_{\dt}$, up to some error term, and perform expansions in powers of~$\dt$ of ${\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt})$. More precisely, we obtain from~\eqref{eqn:discrete-Poisson-equation-projected} and~\eqref{eqn:phi} that $-\LOPERHPI (\widetilde{f}_{\dt}-\widehat{f}_{\dt}) = \dt^2 \phi_{1, \dt, f}$ and hence by definition of $\widetilde{\mL}_{\dt}$, \[ \widetilde{f}_{\dt}-\widehat{f}_{\dt} = -h^2 \mL_{\dt}^{-1} (\phi_{1, \dt, f} - \pi_{\gamma, \dt}(\phi_{1, \dt, f})). \] By~\eqref{eqn:control-of-phi} and Corollary~\ref{cor:bounded-disc-Poission-solution}, there exist therefore~$\dt^*>0$, ~$C \in \mathbb{R}_+$ and~$s \in \mathbb{N}$ such that \[ \forall \dt \in (0,\dt^*], \qquad \left|\widetilde{f}_{\dt}-\widehat{f}_{\dt}\right| \leq C \dt^2 \mKS {\,.} \] This implies that, for any~$\dt \in (0,\dt^*]$, \[ \left|{\Delta \mathcal{M}_{n}}\!\left(\widetilde{f}_{\dt}\right)^2-{\Delta \mathcal{M}_{n}}\!\left(\widehat{f}_{\dt}\right)^2\right| = \left|{\Delta \mathcal{M}_{n}}\!\left(\widetilde{f}_{\dt}-\widehat{f}_{\dt}\right)\right| \, \left|{\Delta \mathcal{M}_{n}}\!\left(\widetilde{f}_{\dt}+\widehat{f}_{\dt}\right)\right|\leq C \dt^2 \left[\mKS(y^n)+\mKS(y^{n+1})\right] {\,.} \] so that \begin{equation} \label{eq:bound_difference_WTF_WHF_variance} \begin{aligned} & \left| \bE_{\dt}\left\{ \DM_{n_1}\!\!\left(\widehat{f}_{\dt}\right)^2 \DN_{n_2}^2 \right\} - \bE_{\dt}\left\{ \DM_{n_1}\!\!\left(\widetilde{f}_{\dt}\right)^2 \DN_{n_2}^2 \right\} \right| \\ & \qquad \qquad \leq C\dt^3 \bE_{\dt}\left[ \left(F(q^{n_2})^{\T} G^{n_2} \right)^2\left(\mKS(y^{n_1})+\mKS(y^{n_1+1})\right)\right] \\ & \qquad \qquad \leq C\dt^3 \bE_{\dt}\left[ \left(F(q^{n_2})^{\T} G^{n_2} \right)^4 + \mKS(y^{n_1})^2 +\mKS(y^{n_1+1})^2 \right] \leq C\dt^3. \end{aligned} \end{equation} We therefore obtain that, for $\dt \in (0,\dt^*]$, \[ \left| \sum_{n_1,n_2=0}^{N-1} \bE_{\dt}\left\{ \DM_{n_1}\!\!\left(\widehat{f}_{\dt}\right)^2 \DN_{n_2}^2 \right\} - \sum_{n_1,n_2=0}^{N-1} \bE_{\dt}\left\{ \DM_{n_1}\!\!\left(\widetilde{f}_{\dt}\right)^2 \DN_{n_2}^2 \right\} \right| \leq C N^2\dt^3. \] Moreover, in view of the expansion~\eqref{eq:Taylor_WTF} of~${\Delta \mathcal{M}_{n}}(\widetilde{f}_{\dt})$ in powers of~$\dt$, and the bounds~\eqref{eq:moment_bounds_psi_Upsilon}, a computation similar to~\eqref{eq:bound_difference_WTF_WHF_variance} provides \[ \frac{1}{N^2\dt^2}\sum_{n_1=0}^{N-1}\sum_{n_2=0}^{N-1} \bE_{\dt}\left\{ \DM_{n_1}\!\!\left(\widetilde{f}_{\dt}\right)^2 \DN_{n_2}^2 \right\} \leq C{\,.} \] By summing the last two estimates, we finally obtain that, for $\dt \in (0,\dt^*]$, \[ \frac{1}{N^2\dt^2}\sum_{n_1=0}^{N-1}\sum_{n_2=0}^{N-1} \bE_{\dt}\left\{ \DM_{n_1}\!\!\left(\widehat{f}_{\dt}\right)^2 \DN_{n_2}^2 \right\} \leq C{\,.} \] A similar estimate can be obtained for the other double sum \[ \frac{1}{N^2\dt^2}\sum_{0 \leq n_1 < n_2 \leq N-1}^{N-1}\!\!\!\!\!\!\!\!\!\bE_{\dt}\left\{\DM_{n_1}\!\!\left(\widehat{f}_{\dt}\right)\DM_{n_2}\!\!\left(\widehat{f}_{\dt}\right)\DN_{n_1}\DN_{n_2}\right\} \leq C\dt{\,.} \] so that the first term on the right hand side of~\eqref{eqn:var-estimate} is uniformly bounded by $C$ for $\dt$ sufficiently small. Now, for the second term on the right hand side of~\eqref{eqn:var-estimate}, a Cauchy--Schwarz inequality gives \[ \mathbb{E}_{\dt}\left\{ \left[ \left(\widehat{f}_{\dt}(y^N) - \widehat{f}_{\dt}(y^0)\right)z^N \right]^2\right\} \leq \mathbb{E}_{\dt}\left\{ \left(\widehat{f}_{\dt}(y^N) - \widehat{f}_{\dt}(y^0)\right)^4 \right\}^{1/2} \mathbb{E}_{\dt}\left\{ (z^N)^4 \right\}^{1/2}. \] Note that by Corollary~\ref{cor:bounded-disc-Poission-solution}, $\widehat{f}_{\dt}$ is uniformly bounded in~$B_{\mKS}^{\infty}(\mY)$ for some integer~$s$ and hence we can write (upon possibly increasing~$s$) \[ \mathbb{E}_{\dt}\left\{ \left(\widehat{f}_{\dt}(y^N) - \widehat{f}_{\dt}(y^0)\right)^4 \right\} \leq C \bE_{\dt}\left\{\mKS(y^0) + \mKS(y^N)\right\}, \] which is uniformly bounded by a constant for~$\dt$ sufficiently small in view of~\eqref{eqn:K-evolution-estimate}. The estimation of $\mathbb{E}_{\dt}\left\{ (z_N)^4 \right\}$ is similar to that of the first term on the right hand side of~\eqref{eqn:var-estimate} by noting that \[ \mathbb{E}_{\dt}\left\{ (z^N)^4 \right\} = \sum_{n=0}^{N-1} \bE_{\dt}\left\{ \DN_{n}^4 \right\} + 6\sum_{0 \leq n_1 < n_2 \leq N-1} \bE_{\dt}\left\{ \DN_{n_1}^2 \DN_{n_2}^2 \right\}, \] which is uniformly bounded by $C N^2\dt^2$ for $\dt$ sufficiently small. Therefore, the second term on the right hand side of~\eqref{eqn:var-estimate} is uniformly bounded by $C/(N \dt)$, which concludes the desired variance estimate. \end{proof} \section{Comparison with the Green-Kubo method} \label{sec:LR-vs-GK} The Green-Kubo (GK) formula is one of the traditional methods for computing linear responses and transport coefficients (see for instance~\cite[Chapter~13]{tuckerman2010statistical}). We first recall known results on the bias of GK estimators in Section~\ref{sec:GK_bias}, and then characterize their variance in Section~\ref{sec:variance_GK}. The main conclusion is that the bias of GK estimators is comparable to the timestep bias of MP estimators, but the finite time bias of GK estimators is much smaller than the one from for MP estimators, while on the other hand the variance of GK estimators scales as~$N\dt$, in sharp contrast to the variance of MP estimators, which is uniformly bounded in time. This suggests that MP estimators could be more reliable than GK estimators. \subsection{Green-Kubo estimators} \label{sec:GK_bias} Recall the reformulation~\eqref{eqn:GK} of the linear response as an integrated correlation function: \[ \mathrm{Lin}(f) = \int_0^{\infty} \bE^{\pi}\left\{ f(y_t) \varphi(y_0) \right\}\, dt = \int_0^{\infty} \bE^{\pi}\left\{ \left[f(y_t)-\pi(f)\right] \varphi(y_0) \right\}\, dt, \] where \begin{equation} \label{eq:def_varphi_GK} \varphi = \LOPERTIL^* \mathbf{1} \end{equation} has average~0 with respect to~$\pi$ and~$\bE^{\pi}$ is the expectation taken with respect to all realizations of the Markov process $\{y_t\}_{t \geq 0}$ and with respect to all initial conditions~$y_0$ distributed according to~$\pi$. The timestep bias of GK estimators was made precise in~\cite{leimkuhler2016computation,lelievre2016partial}. The result essentially states that the linear response can be approximated by a Riemann sum, or using a trapezoidal rule for second order schemes. To make this result precise, we introduce the following estimators: \[ \mathcal{E}_{\dt, N}^{\mathrm{GK}1}(f) = \dt \sum_{n=0}^{N-1} \left[f(y^n) - \pi_{\gamma, \dt}(f)\right] \varphi(y^0), \qquad y^0 \sim \pi_{\gamma, \dt}{\,,} \] and \[ \mathcal{E}_{\dt, N}^{\mathrm{GK}2}(f) = \dt \left( \frac12 \left[f(y^0) - \pi_{\gamma, \dt}(f)\right] + \sum_{n=1}^{N-1} \left[f(y^n) - \pi_{\gamma, \dt}(f)\right] \right)\varphi(y^0), \qquad y^0 \sim \pi_{\gamma, \dt} {\,.} \] Note that, as discussed after~\eqref{eqn:desired-estimator}, we consider here as well a ``perfect'' centering based on subtracting~$\pi_{\gamma, \dt}(f)$; but the same discussion as for MP estimators applies here. Let us also mention that the initial distribution for~$y^0$ has to be the invariant probability measure~$\pi_{\gamma, \dt}$ for GK estimators whereas MP estimators do not require this condition. We can now recall the results from~\cite{leimkuhler2016computation,lelievre2016partial}, which correspond to the case $N \to \infty$. We denote by $\bE_{\dt}^{\pi_{\gamma, \dt}}$ the expectation taken with respect to all realizations of the Markov chain $\{y^n\}_{n \geq 0}$ and with respect to all initial conditions~$y^0$ distributed according to~$\pi_{\gamma,\dt}$. \begin{theorem} \label{thm:error-estimates-GK} Fix an observable $f \in \mS$, and consider any of the first ($\alpha=1$) or second ($\alpha=2$) order splitting schemes introduced in Section~\ref{sec:LD-application}. Then there exist~$\dt^* > 0$ and $C \in \mathbb{R}_+$ such that, for any $\dt\in(0,\dt^*]$, \[ \left| \bE_{\dt}^{\pi_{\gamma, \dt}}\left\{\mathcal{E}_{\dt, \infty}^{\mathrm{GK}\alpha}(f)\right\} - \mathrm{Lin}(f) \right| \leq C \dt^\alpha. \] \end{theorem} In practice, the integrated correlation over the infinite time horizon has to be approximated by that over a finite time horizon, which adds another bias, exponentially small in the integration time~$N\dt$ -- and hence asymptotically much smaller than the finite time integration bias of order~$1/\sqrt{N\dt}$ for MP estimators. \begin{corollary}[Bias of GK estimators] \label{cor:GK-bias} Fix an observable $f \in \mS$, and consider any of the first ($\alpha=1$) or second ($\alpha=2$) order splitting schemes introduced in Section~\ref{sec:LD-application}. Then there exist $\lambda,\dt^* > 0$ and $C \in \mathbb{R}_+$ such that, for any $\dt\in(0,\dt^*]$, \[ \forall N \geq 1, \qquad \left| \bE_{\dt}^{\pi_{\gamma, \dt}}\left\{\mathcal{E}_{\dt,N}^{\mathrm{GK}\alpha}(f)\right\} - \mathrm{Lin}(f) \right| \leq C \left( \dt^\alpha + \mathrm{e}^{-\lambda N\dt} \right) {\,.} \] \end{corollary} \begin{proof} We bound the error as \[ \begin{aligned} \left| \bE_{\dt}^{\pi_{\gamma, \dt}}\left\{\mathcal{E}_{\dt,N}^{\mathrm{GK}\alpha}(f)\right\} - \mathrm{Lin}(f) \right| & \leq \left| \bE_{\dt}^{\pi_{\gamma, \dt}}\left\{\mathcal{E}_{\dt,N}^{\mathrm{GK}\alpha}(f)\right\} - \bE_{\dt}^{\pi_{\gamma, \dt}}\left\{\mathcal{E}_{\dt,\infty}^{\mathrm{GK}\alpha}(f)\right\} \right| \\ & \quad + \left| \bE_{\dt}^{\pi_{\gamma, \dt}}\left\{\mathcal{E}_{\dt,\infty}^{\mathrm{GK}\alpha}(f)\right\} - \mathrm{Lin}(f) \right|. \end{aligned} \] The second term on the right hand side of the previous inequality is bounded by~$C\dt^{\alpha}$ in view of Theorem~\ref{thm:error-estimates-GK}. Choosing an integer~$s$ such that $f,\varphi \in B^\infty_{\mKS}(\mY)$, and using the exponential decay estimates of Theorem~\ref{thm:ergodicity-splitting}, we obtain, for some constant~$K \in \mathbb{R}_+$, \[ \begin{split} \left| \bE_{\dt}^{\pi_{\gamma, \dt}}\left\{\mathcal{E}_{\dt,N}^{\mathrm{GK}\alpha}(f)\right\} - \bE_{\dt}^{\pi_{\gamma, \dt}}\left\{\mathcal{E}_{\dt,\infty}^{\mathrm{GK}\alpha}(f)\right\} \right| & \leq \dt \sum_{n = N}^{\infty}\bE_{\dt}^{\pi_{\gamma, \dt}} \left( \left| \left[f(y^n) - \pi_{\gamma, \dt}(f)\right] \varphi(y^0) \right| \right)\\ & \leq K \|f\|_{B^\infty_{\mKS}} \|\varphi\|_{B^\infty_{\mKS}} \pi_{\gamma, \dt}\left(\mKS^2\right) \dt \sum_{n = N}^{\infty} \mathrm{e}^{-\lambda n \dt} \end{split} \] The result follows by noting that $\dt\sum_{n = N}^{\infty}\mathrm{e}^{-\lambda n \dt} \leq 2\lambda^{-1} \mathrm{e}^{-\lambda N \dt}$ for~$\dt$ sufficiently small. \end{proof} Let us conclude this section by emphasizing that all the results presented here require that initial conditions are distributed according to the invariant probability measure~$\pi_{\gamma,\dt}$, in contrast to the estimators constructed in Section~\ref{sec:general-continuous-setting}. \subsection{Variance analysis of Green-Kubo estimators} \label{sec:variance_GK} We can now present the main result of this section, which states that the variance of the Green--Kubo estimator is given at dominant order by the variance of the random variable \[ \mathscr{E}_{\dt, N}^{\mathrm{GK}\alpha}(f) = \varphi(y^0) \sum_{n=0}^{N-1} {\Delta \mathcal{M}_{n}}\left(\widehat{f}_{\dt} \right) {\,,} \] which is expected to scale linearly with the integration time~$Nh$. Recall that~$\varphi$ is given by~\eqref{eq:def_varphi_GK}. The increase of the variance with time can significantly deteriorate the performance of the estimator especially when correlation functions decay slowly. \begin{theorem}[Variance of GK estimators] \label{thm:GK-variance} Fix an observable $f \in \mS$, and consider any of the first ($\alpha=1$) or second ($\alpha=2$) order splitting schemes introduced in Section~\ref{sec:LD-application}. Then there exist $\dt^* > 0$ and $C \in \mathbb{R}_+$ such that, for any $\dt\in(0,\dt^*]$, \[ \forall N \geq 1, \qquad \left| \mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left( \mathcal{E}_{\dt, N}^{\mathrm{GK}\alpha}(f) \right)- \mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left( \mathscr{E}_{\dt, N}^{\mathrm{GK}\alpha}(f) \right) \right| \leq C\left(1+ \sqrt{N\dt}\right). \] As a consequence, there exists a constant~$K \in \mathbb{R}_+$ such that, for any $\dt\in(0,\dt^*]$, the following bound holds when~$N\dt \geq 1$: \[ \mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left( \mathcal{E}_{\dt, N}^{\mathrm{GK}\alpha}(f)\right) \leq K N\dt. \] \end{theorem} \begin{proof} The proof is quite similar to the proof of Theorem~\ref{thm:MP-variance}, which is not surprising since the GK estimator shares some structural similarities with MP estimators (in fact, replacing~$z^N/(N\dt)$ by~$\varphi(y^0)$). We therefore follow the notation introduced in Theorem~\ref{thm:MP-variance}. Note first that \[ \mathcal{E}_{\dt, N}^{\mathrm{GK}\alpha}(f) = \mathscr{E}_{\dt, N}^{\mathrm{GK}\alpha}(f) + \mathscr{R}_{\dt, N}^{\mathrm{GK}\alpha}(f), \qquad \mathscr{R}_{\dt, N}^{\mathrm{GK}\alpha}(f) = \left( \widehat{f}_{\dt}(y^0) - \widehat{f}_{\dt}(y^N) \right) \varphi(y^0) {\,.} \] Therefore, in view of the inequality $|\mathrm{Var}(X+Y)-\mathrm{Var}(X)| \leq \mathrm{Var}(Y) + 2\sqrt{\mathrm{Var}(Y)\mathrm{Var}(X)}$, we obtain \[ \begin{aligned} \left| \mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left( \mathcal{E}_{\dt, N}^{\mathrm{GK}\alpha}(f) \right) - \mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left( \mathscr{E}_{\dt, N}^{\mathrm{GK}\alpha}(f) \right) \right| & \leq \mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left(\mathscr{R}_{\dt, N}^{\mathrm{GK}\alpha}(f)\right) \\ & \quad + 2 \sqrt{\mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left(\mathscr{R}_{\dt, N}^{\mathrm{GK}\alpha}(f)\right)\mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left( \mathscr{E}_{\dt, N}^{\mathrm{GK}\alpha}(f) \right)}. \end{aligned} \] The estimate of the variance of~$\mathscr{E}_{\dt, N}^{\mathrm{GK}\alpha}(f)$ is similar to that of $\sum_{n=0}^{N-1}{\Delta \mathcal{M}_{n}}(\widehat{f}_{\dt}) z^N$ in Theorem~\ref{thm:MP-variance} and can be shown to be bounded by $CN\dt$. Furthermore, a simple computation shows that \[ \mathrm{Var}_{\dt}^{\pi_{\gamma, \dt}}\left(\mathscr{R}_{\dt, N}^{\mathrm{GK}\alpha}(f)\right) \leq \mathbb{E}_{\dt}^{\pi_{\gamma, \dt}}\left(\mathscr{R}_{\dt, N}^{\mathrm{GK}\alpha}(f)^2\right) \leq 2 \int_\mY \left[ \widehat{f}_{\dt}^2 + \left(P_\dt^N \widehat{f}_{\dt}\right)^2 \right] \varphi^2 \, d\pi_{\gamma, \dt}{\,.} \] The right hand side of the latter inequality is uniformly bounded in~$N$ in view of~\eqref{eqn:K-evolution-estimate} since~$\varphi,\widehat{f}_{\dt}$ are both in~$B^\infty_{\mKS}(\mY)$ for some integer~$s \in \mathbb{N}$ (with uniform bounds for~$\widehat{f}_{\dt}$ by Corollary~\ref{cor:bounded-disc-Poission-solution}). % \end{proof} \section{Computational benchmarks} \label{sec:numerics} We demonstrate the accuracy and efficiency of the derived MP estimators by applying them to simple two dimensional Langevin systems. Denoting by~$q =(q_1, q_2)$ and $p = (p_1, p_2)$, we consider the Langevin dynamics~\eqref{eqn:ULD} with $\beta = \gamma = 1$ and $M = \mathrm{Id}$. We study the behavior of the proposed MP estimators in the case of unbounded position space and two different types of perturbing forces (Examples~1 and~2). In Example~3, we present comparison between the MP and Green-Kubo estimators in the case of periodic (compact) position domain. \ifx\IMAJNA\undefined \subsubsection*{Example 1: position dependent perturbation $F(q)$.} \else \medskip\noindent{\it Example 1:} {position dependent perturbation $F(q)$.} \fi In this example we choose a harmonic potential for $q \in\mathbb{R}^2$: \[ V(q,\omega) = \frac{1}{2}\omega |q|^2 = \frac{1}{2}\omega (q_1^2 + q_2^2), \] with~$\omega=1$ in our simulations. Note that the dynamics is defined on $\mathcal{Y}=\R^2\times \R^2$, thus strictly speaking not satisfying our set of assumptions. Recall however that the restriction to a compact position domain is for the ease of presentation, as hinted at after Assumption~\ref{ass:V_smooth_and_domain_compact}. \begin{figure}[ht] \includegraphics[width=0.95\textwidth]{./Figures/convRateQuadExMP12inset.png} \caption{{\it Example 1.} Convergence rates of the bias for the MP estimators, $\mathcal{E}_{\dt, N}^{\mathrm{MP}\alpha}$, illustrating the error estimates~\eqref{eqn:1st-order-MP-bias} and~\eqref{eqn:2nd-order-MP-bias}. The inset depicts the dependence of the bias on the timestep $\dt$. The exact value $\mathrm{Lin}(f)=0.5$ of the sensitivity index is computed from Gaussian integrals.} \label{fig:MPquadconv} \end{figure} The perturbing force $F$ and the observable $f$ are chosen as \[ F_i(q)\equiv\frac{\partial^2 V}{\partial q_i\partial\omega} (q,\omega) = q_i\,{\,,} \qquad f(q) = |q|^2\,. \] Thus we estimate the sensitivity of the second moment of the $q$-marginal for the invariant distribution (a Gaussian distribution in $\R^2$) with respect to the parameter $\omega$. Since the perturbing force is only position dependent, one can apply directly the formula \eqref{eqn:2nd-order-weight}, \emph{i.e.}, the increment of the re-weighting process is \begin{equation} \label{eqn:2nd-order-weight-inc} z^{n+1}-z^n = \sqrt{\frac{\beta}{2\gamma}} \Big[ \dt^{1/2} F(q^n)^{\T} + \frac{1}{2}\dt^{3/2} (p^n)^\T M^{-1} \nabla_q F(q^n)\Big] G^n \,. \end{equation} The example allows us to demonstrate, in a simple setting, the importance of the $\dt^{3/2}$ term in the definition of re-weighting process $z^n$ in \eqref{eqn:2nd-order-weight}. Simulations are run with~$N_{\mathrm{smpl}} = 10^6$ realizations, for dynamics integrated up to a physical time~$T_{\mathrm{fin}} = 10^3$. Figure~\ref{fig:MPquadconv} depicts the convergence of the bias for the sensitivity index $\mathrm{Lin}{f}$ for the 1st order ($\alpha=1$) and 2nd order ($\alpha=2$) MP estimators~$\mathcal{E}_{\dt, N}^{\mathrm{MP}\alpha}$, confirming the error estimates~\eqref{eqn:1st-order-MP-bias} and~\eqref{eqn:2nd-order-MP-bias}. The 2nd order estimator uses the BACAB splitting~\eqref{eqn:BACAB} for the Langevin dynamics. In contrast, an MP estimator (here called 1st order with the 2nd order dynamics) omitting the $\dt^{3/2}$ term in \eqref{eqn:2nd-order-weight-inc} % % % % achieves only a first order rate of convergence. \ifx\IMAJNA\undefined \subsubsection*{Example 2: momentum dependent perturbation $F(p)$.} \else \medskip\noindent{\it Example 2:} {momentum dependent perturbation $F(p)$.} \fi The purpose of this benchmark is to demonstrate the full expression for the 2nd-order increment of the re-weighting process in the MP estimator $\mathcal{E}_{\dt, N}^{\mathrm{MP}\alpha}$. We consider the same potential as in Example 1, and denote by% \[ \pi_\beta(dq\,dp) = \frac{\beta^2 \omega}{4\pi^2} \mathrm{e}^{-\beta |p|^2/2} \mathrm{e}^{-\beta\omega |q|^2/2}\,dq\,dp\, \] the invariant probability measure of the Langevin dynamics~\eqref{eqn:ULD}. We modify the perturbing force as $F(p) = p$, which corresponds to a perturbation of the friction coefficient $\gamma$ in the Langevin equation~\eqref{eqn:ULD}. In view of the fluctuation-dissipation relation $\sigma\sigma^T = 2\beta^{-1} \gamma$, which relates the diffusion matrix $S=\sigma\sigma^T$ with the friction matrix $\gamma$ and the inverse temperature~$\beta$, the perturbation of $\gamma$ results in a perturbation of the average~$\pi_\beta(f)$ with respect to $\beta$. Therefore, for a suitable choice of observables, \emph{e.g.}, $f_1(q) = q_1^2 + q_2^2$ and $f_2(p)=p_1^4 + p_2^4$ used in this example, the sensitivity/linear response $d\pi_\beta(f)/d\beta$ can be computed exactly by evaluating Gaussian integrals. A simple computation based on~\eqref{eqn:gfun} shows that~$g(q,p) = \tfrac{\beta}{2\gamma}|p|^2$, which allows us to obtain an explicit expression for the martingale increment of the 2nd-order MP estimator by substituting the formula for~$g$ into~\eqref{eqn:zinc} in Appendix~\ref{app:secondorder}. The results presented in Figure~\ref{fig:MPbetaconv}, obtained with~$N_{\mathrm{smpl}} = 10^7$ realizations of the dynamics integrated up to a physical time~$T_{\mathrm{fin}} = 10^3$, show that the second order MP estimators allow for passing from a bias of order~$\dt$ to a bias of order~$\dt^2$, and even to~$\dt^4$ for the observable~$f_1$. For the latter case, the bias cannot be decreased below the value~$10^{-3}$ since the statistical error starts to dominate the overall error. \begin{figure}[ht] \includegraphics[width=0.95\textwidth]{./Figures/convRateExMPbeta12inset.png} \caption{{\it Example 2.} Convergence rates of the bias for the MP estimators, $\mathcal{E}_{\dt, N}^{\mathrm{MP}\alpha}$. The exact values $\mathrm{Lin}(f_1)=-2$ and $\mathrm{Lin}(f_2)=-12$ of the sensitivity index with respect to $\beta$ are computed from Gaussian integrals.} \label{fig:MPbetaconv} \end{figure} \ifx\IMAJNA\undefined \subsubsection*{Example 3.} \else \medskip\noindent{\it Example 3.} \fi In the last example we consider a Langevin dynamics on the phase space $\mathcal{Y} = \mathbb{T}^2 \times \R^2$ with a non-conservative constant perturbing force $F \in \R^2$ (which does not derive from the gradient of a smooth periodic potential). More precisely, using the Langevin dynamics \eqref{eqn:ULD}, we sample the invariant distribution for $q =(q_1, q_2) \in \mathcal{D} = (2\pi \mathbb{T})^2$ and $p = (p_1, p_2) \in \mathbb{R}^2$, with the periodic potential \[ V(q) = 2 \cos(2q_1) + \cos(q_2)\,. \] We are interested in estimating the mobility in a direction given by the constant unit vector $F$. The observable $f$ and the corresponding function $\varphi$ in~\eqref{eqn:GK} are \[ f(q, p) = F^{\T} M^{-1} p\,, \;\;\;\;\; \varphi(q, p) = \beta F^{\T} M^{-1} p\,. \] The Green-Kubo formula for the mobility therefore reduces to the following time integral of the velocity autocorrelation function: \[ \mathrm{Lin}(f)= \beta \int_{0}^{\infty} \bE^{\pi}\left\{\left(F^{\T} M^{-1} p_t\right) \left(F^{\T} M^{-1} p_0 \right)\right\} dt{\,.} \] For the simulation results presented here, the perturbation is chosen to be $F = (1, 0)^{\T}$, which corresponds to estimating the mobility in the $q_1$-direction. Figure~\ref{fig:GKMPconv} shows that the predicted convergence rates for the bias of both martingale product ($\mathcal{E}_{\dt, N}^{\mathrm{MP}\alpha}$) and Green-Kubo ($\mathcal{E}_{\dt, N}^{\mathrm{GK}\alpha}$) estimators are indeed of the order predicted by our theoretical analysis. These results were produced with~$N_{\mathrm{smpl}} = 10^7$ realizations of the dynamics integrated up to a physical time~$T_{\mathrm{fin}} = 900$ for~MP estimators, and~$T_{\mathrm{fin}} = 25$ for GK estimators. \begin{figure}[ht] \includegraphics[width=0.95\textwidth]{./Figures/convergence_biasMP-GK12.png} \caption{{\it Example 3.} Convergence rates of the bias for the different estimators. The 2nd order MP estimator is tested with the splitting schemes BACAB (splitting 1) and CBABC (splitting 2) }\label{fig:GKMPconv} \end{figure} An important property of the proposed MP estimators is their bounded variance while the variance of the GK estimators grows in time as demonstrated in the inset of Figure~\ref{fig:GKMPvar}. However, a fair comparison of efficiency should take into account the fact that auto-correlation functions can be decaying fast (as in the example studied here) and thus the GK estimator may not require long trajectories to be integrated, see Corollary~\ref{cor:GK-bias} for the precise statement. On the other hand, initial conditions for the Green-Kubo estimator need to be sampled from the invariant distribution and thus preparation of an initial state may still require a long time integration of the Langevin dynamics or the use of another equilibration algorithm, thus contributing to the overall computational cost of the method. \begin{figure}[ht] \centering \includegraphics[width=0.80\textwidth]{./Figures/MPVarCompare2inset.png} \caption{{\it Example 3.} Comparison variance for the martingale product (MP) and Green-Kubo (GK) estimators.} \label{fig:GKMPvar} \end{figure}

https://arxiv.org/abs/1607.03800

Moser stability for volume forms on noncompact fiber bundles

We prove a stability result for volume forms on fiber bundles with compact base and noncompact fibers. This generalizes the classical results of Moser and Greene--Shiohama, and recent work by the authors.

\section{Appendix: ends} \label{sec:appendix} \subsection{Ends} \label{ssec:end} For any topological space $X$, $(\mathcal{K}(X), {\subset})$ is a directed partially ordered set. Let $\rho^K_L \colon \Conn(X \setminus L) \rightarrow \Conn(X \setminus K)$ sending any $C$ to the connected component of $X \setminus K$ containing $C$, for any $K, L \in \mathcal{K}(X)$ such that $K \subset L$. Let $\mathcal{NE}(X) = \Set{\Connub(X \setminus K)}_{K \in \mathcal{K}(X)}$. As in \cite[Subsection 1.1]{MR2599132}, \begin{equation*} \mathcal{N}(X) = \left\langle \mathcal{NE}(X), \Set{\rho^K_L}_{K, L \in \mathcal{K}(X)}^{K \subset L} \right\rangle \end{equation*} is an inverse system of sets. \begin{definition} \label{def:end} Let $X$ be a topological space. An \notation{end} of $X$ is an element $\epsilon$ in $\mathcal{E}(X)$, the inverse limit of the inverse system $\mathcal{N}(X)$. If $X$ is compact $\mathcal{E}(X)$ is empty. \end{definition} \begin{definition} \label{def:refinement-end} For a topological space $X$, a subcollection $\mathcal{A} \subset \mathcal{K}(X)$ is a \notation{refinement} of $\mathcal{B} \subset \mathcal{K}(X)$ if for any $K \in \mathcal{B}$ there is $L \in \mathcal{A}$ such that $K \subset L$. Being refinements is a preorder among all subcollections of $\mathcal{K}(X)$. If $\mathcal{A}$ is a refinement of $\mathcal{K}(X)$ we say $\mathcal{A}$ is \notation{exhaustive}. \end{definition} \begin{definition} \label{def:finiteness-end} Suppose $X$ is a smooth manifold, and $\omega$ is a volume form on $X$. With respect to $\omega$, we say $\epsilon \in \mathcal{E}(X)$ a \notation{finite end} if there exist $K \in \mathcal{K}(X)$ such that $\epsilon(K)$ has finite volume; we say $\epsilon$ an \notation{infinite end} if for any $K \in \mathcal{K}(X)$, $\epsilon(K)$ has infinite volume. \end{definition} If $\mathcal{A}$ is a subcollection of $\mathcal{K}(X)$, define $\mathcal{NE}(X \pipe \mathcal{A}) = \Set{\Connub(X \setminus K)}_{K \in \mathcal{A}}$. Then we denote by $\mathcal{E}(X \pipe \mathcal{A})$ the inverse limit of the inverse system $\mathcal{N}\Pa{X \pipe \mathcal{A}} = \left\langle \mathcal{NE}(X \pipe \mathcal{A}), \Set{\rho^K_L}_{K, L \in \mathcal{A}}^{K \subset L} \right\rangle$. \begin{proposition} \label{prop:exhaustive-relative-end} If $\mathcal{K}(X)$ has an exhaustive subcollection $\mathcal{A}$, then $\mathcal{E}(X \pipe \mathcal{A}) = \mathcal{E}(X)$, in the sense that the map $\mathcal{E}(X) \rightarrow \mathcal{E}(X \pipe \mathcal{A}), \epsilon \mapsto \Res{\epsilon}_{\mathcal{A}}$, is a bijection. \end{proposition} \begin{proof} Suppose we know $\Res{\epsilon}_{\mathcal{A}} \in \mathcal{E}(X \pipe \mathcal{A})$, we want to recover $\epsilon(K)$ for any $K \in \mathcal{K}(X)$. Since $\mathcal{A}$ is exhaustive, there is $L \in \mathcal{A}$ are such that $K \subset L$. Then let $\epsilon(K) = \rho^K_L \Res{\epsilon}_{\mathcal{A}}(L)$. If there is $L_1 \in \mathcal{A}$ such that $K \subset L_1$, then there is $L_2 \in \mathcal{A}$ that contains $L \cup L_1 \in \mathcal{K}(X)$. So $\epsilon(K) = \rho^K_L \Res{\epsilon}_{\mathcal{A}}(L) = \rho^K_L \rho_{LL_2} \Res{\epsilon}_{\mathcal{A}}(L_2) = \rho_{KL_2} \Res{\epsilon}_{\mathcal{A}}(L_2) = \rho_{KL_1} \Res{\epsilon}_{\mathcal{A}}(L_1)$ is well defined. If $K \in \mathcal{A}$ then $\epsilon(K) = \rho_{KK} \Res{\epsilon}_{\mathcal{A}}(K) = \Res{\epsilon}_{\mathcal{A}}(K)$, so $\Res{\epsilon}_{\mathcal{A}}$ is the restriction of $\epsilon$ on $\mathcal{A}$. Then we check that if $K, K_1 \in \mathcal{K}(X)$ are such that $K \subset K_1$ and $L \in \mathcal{A}$ contains $K \cup K_1$, then $\epsilon(K) = \rho^K_L \Res{\epsilon}_{\mathcal{A}}(L) = \rho_{KK_1} \rho_{K_1L} \Res{\epsilon}_{\mathcal{A}}(L) = \rho_{KK_1} \epsilon(K_1)$. Hence $\epsilon \in \mathcal{E}(X)$. \end{proof} Let $\lb{M}$ be an exhausted bundle with compact base. Let $\mathcal{K}(\lb{M})$ be compact exhausted subbundles of $\lb{M}$. Note that the collection of underlying spaces of bundles in $\mathcal{K}(\lb{M})$ is a refinement of $\mathcal{K}(M)$ by Proposition~\ref{prop:exhaustive-relative-end}, so $\mathcal{E}(\lb{M}) = \mathcal{E}(M)$. Let $\omega, \tau \in \FVform(\lb{M})$. \begin{lemma} \label{lem:refinement-commensurability} Let $\mathcal{B}$ is a refinement of $\mathcal{A}$. If $\omega, \tau$ are smoothly commensurable on any element of $\mathcal{NE}(X \pipe \mathcal{A})$ then they are smoothly commensurable on every element of $\mathcal{NE}(X \pipe \mathcal{B})$. \end{lemma} \begin{proof} For $\lb{K} \in \mathcal{A}$ let $\lb{L} \in \mathcal{B}$ such that $\undsp(\lb{K}) \subset \undsp(\lb{L})$. Let $K = \undsp(\lb{K})$ and $L = \undsp(\lb{L})$. Then any $C \in \Connub(X \setminus K)$ is a filled subspace of $\lb{M}$, so let $\lb{C} = \Res{\lb{M}}_C$. For any $E \in \Connub(C \setminus L)$, let $\lb{E} = \Res{\lb{M}}_{E}$. Let $\alpha \in \Regular(f)$ such that $\alpha \geq \max_L f \geq \max_K f$. Now we have \begin{align*} \rint_{\lb{C}} \omega &= \int_{\Res{(\Rls \lb{C})}_{C_{(-\infty, \alpha]}}} \omega + \int_{\Res{(\Rls \lb{C})}_{C_{(\alpha, +\infty)}}} \omega \\ &= \int_{\Res{(\Rls \lb{C})}_{C_{(-\infty, \alpha]}}} \omega + \sum_{E \in \Connub(C \setminus L)} (\kappa^{\lb{E}}_{\lb{C}})_* \int_{\Res{(\Rls \lb{E})}_{E_{(\alpha, +\infty)}}} \omega \\ &= \int_{\Res{(\Rls \lb{C})}_{C_{(-\infty, \alpha]}}} \omega + \sum_{E \in \Connub(C \setminus L)} (\kappa^{\lb{E}}_{\lb{C}})_* \Pa{\rint_{\lb{E}} \omega - \int_{\Res{(\Rls \lb{E})}_{E_{(-\infty, \alpha]}}} \omega}, \end{align*} if $\omega$ has finite released integral on $\lb{E}$. In case that $\omega$ has infinite released integral on $\lb{E}$, by the second step of above formula $\omega$ also has infinite released integral on $\lb{E}$. Analogous arguments work for $\tau$. Note that $\Res{(\Rls \lb{C})}_{C_{(-\infty, \alpha]}}$ and $\Res{(\Rls \lb{E})}_{E_{(-\infty, \alpha]}}$ are compact exhausted bundles, so the fiber integrals on them are smooth functions. If $\omega, \tau$ are smoothly commensurable on every $\lb{E}$ for $E \in \Connub(C \setminus L)$ then they are smoothly commensurable on $\lb{C}$, which completes the proof. \end{proof} \subsection{Equivalence of theorems} \label{ssec:equivalence-theorems} In this subsection, fix $p \in B$ and $(U, \varphi_U)$ a trivialized chart such that $p \in U$. Let $\mathcal{A} = \Setby{\lb{M}_{(-\infty, \alpha]}}{\alpha \in \Regular(f)} \subset \mathcal{K}(\lb{M})$. Let $\mathcal{B} = \bigcup_{\alpha \in \Regular(f)} \Conn F_\alpha \subset \mathcal{K}(F)$ where $F_\alpha$ is the fiber of $\lb{M}_{(-\infty, \alpha]}$. Note that $\mathcal{A}, \mathcal{B}$ are exhaustive, see the proof of Lemma~\ref{lem:saturating-threshold}. Since $\mathcal{NE}(\lb{M} \pipe \mathcal{A})$ is the set of superlevel subbundles of $\lb{M}$, by Lemma~\ref{lem:refinement-commensurability}, any two fiber volume forms on $\lb{M}$ are smoothly commensurable if and only if they are smoothly commensurable on any element of $\mathcal{NE}(\lb{M})$. Since $\mathcal{A}, \mathcal{B}$ are exhaustive, a similar argument as Lemma~\ref{lem:refinement-commensurability} shows that Theorem~\ref{thm:main-theorem} is equivalent to Theorem~\ref{cor:smooth-family} when $\lb{M}$ is a trivial bundle, and is equivalent to Theorem~\ref{thm:greene-shiohama} when $B$ is a singleton. The following proposition demonstrates the relationship between ends of the underlying space and ends of the fiber. \begin{proposition} \label{prop:bundle=fiber=end} There is a surjection $\gamma \colon \mathcal{E}(F) \to \mathcal{E}(M)$. \end{proposition} \begin{proof} For any $\epsilon' \in \mathcal{E}(F \pipe \mathcal{B})$, define $\epsilon = \tilde \gamma(\epsilon') \in \mathcal{E}(M \pipe \mathcal{A})$ as follows: for any $\alpha \in \Regular(f)$, let $\epsilon(\lb{M}_{(-\infty, \alpha]})$ be the connected components of $\lb{M}_{(\alpha, +\infty)}$ whose fiber contains $\epsilon'(F_\alpha)$ in $F$. To prove the surjectivity of $\tilde \gamma$, we can first choose a strictly increasing sequence $\Set{\alpha_i}_{i \in \mathbb{N}} \subset \Regular(f)$, and let $\mathcal{B}_1 = \bigcup_{i \in \mathbb{N}} \Conn F_{\alpha_i}$. Then for any $\epsilon \in \mathcal{E}(M \pipe \mathcal{A})$, we first construct $\epsilon'_1 \in \mathcal{E}(M \pipe \mathcal{B}_1)$ by induction, then extend it to $\epsilon' \in \mathcal{E}(M \pipe \mathcal{B})$ such that $\tilde \gamma(\epsilon') = \epsilon$. Since $\mathcal{A}, \mathcal{B}$ are exhaustive, by Proposition~\ref{prop:exhaustive-relative-end}, there are bijections $\mathcal{E}(F) \rightarrow \mathcal{E}(F \pipe \mathcal{B})$ and $\mathcal{E}(M) \rightarrow \mathcal{E}(M \pipe \mathcal{A})$, then $\tilde \gamma$ induces a surjection $\gamma \colon \mathcal{E}(F) \to \mathcal{E}(M)$. \end{proof} \section{Combinatorial construction: trees} \label{sec:combinatorics} Consider the following combinatorial notions from~\cite[Definition 9.10]{MR1940513} and \cite[Section 1]{math/0412554} which will be very useful for our purposes. \begin{definition} \label{def:tree} A \notation{tree} is a strict partially ordered set $(\mathcal{T}, \prec)$ with the property that for each $x \in \mathcal{T}$, the set $\tPre(x) = \Setby{y \in \mathcal{T}}{y \prec x}$ of all \notation{predecessors} of $x$ is well ordered by $\prec$. We then define $\tpre(x) = \Setby{y \in \tPre(x)}{\forall z \in \tPre(x), y \not\prec z}$ as the \notation{immediate predecessor}. We write $\mathcal{T}$ for $(\mathcal{T}, \prec)$ when there is no ambiguity. A \notation{branch} in $\mathcal{T}$ is a maximal linearly ordered subset of $\mathcal{T}$. Let $\tbranch(\mathcal{T})$ be the collection of branches of $\mathcal{T}$. Let $\tRoot(\mathcal{T}) = \Setby{x \in T}{\forall y \in T, y \not\prec x} \neq \emptyset$ be the set of roots of $\mathcal{T}$. If $\tRoot(\mathcal{T})$ is a singleton we call $\mathcal{T}$ \notation{rooted} and let $\troot(\mathcal{T})$ the \notation{root} of $\mathcal{T}$ be the only element of $\tRoot(\mathcal{T})$. \end{definition} \begin{definition} \label{def:successor} Let $\tSuc(x) = \Setby{y \in \mathcal{T}}{y \succ x}$ be the set of all \notation{successors} of $x$, then $(\tSuc(x), \prec)$ is a tree with $\tSuc(x) \subset \mathcal{T}$. Let $\tsuc(x) = \tRoot(\tSuc(x))$ be the set of \notation{immediate successors} or \notation{children} of $x$. If for any $x \in \mathcal{T}$, $\tsuc(x)$ is finite, we call $\mathcal{T}$ \notation{locally finite}. Let $\tssuc(x) = \bigcup_{y \in \tsuc(x)} \tsuc(y)$ be the set of \notation{grandchildren} of $x$. Let $\tLeaf(\mathcal{T}) = \Setby{x \in T}{\forall y \in T, x \not\prec y}$ be the set of \notation{pendant vertices} or \notation{leaves} of $\mathcal{T}$. If $\tLeaf(\mathcal{T}) = \emptyset$ we call $\mathcal{T}$ \notation{leafless}. \end{definition} \begin{definition} \label{def:tree-height} The \notation{depth} of $x$ is the ordinal of $\tPre(x)$, which we denote by $\tdepth(x)$. Let $\theight(\mathcal{T}) = \sup\Setby{\tdepth(x) + 1}{x \in \mathcal{T}}$ be the \notation{height} of $\mathcal{T}$. If $\theight(\mathcal{T})$ is the smallest infinite ordinal, then any $x \in \mathcal{T}$ has finite depth. or \notation{parent} of $x$. Then for $l \in \mathbb{N} \cup \Set{0}$, let $\tlevel(l) = \Set{x \in \mathcal{T}}{\tdepth(x) = l}$ be the $l$-th \notation{level} of $\mathcal{T}$. \end{definition} Let $\upomega$ denote the smallest infinite ordinal. Any other infinite ordinal is larger than $\upomega$. If $\theight(\mathcal{T}) = \upomega$, then every node in $\mathcal{T}$ has finite depth, but these depths can be arbitrarily large. We have the following essential construction for the combinatorial part of the proof of Theorem~\ref{thm:main-theorem}. \begin{lemma} \label{lem:bundle-slicing-tree} Let $\lb{M} = (\pi, M, B, F, f, h)$ be a connected exhausted bundle, $\alpha_0 = -\infty$ and $\Set{\alpha_l}_{l \in \mathbb{N}} \subset \Regular(f) \cap f(M)$ be an unbounded strictly increasing sequence. Let $\tlevel(l)$ be the collection of $\lb{A} = \Res{\lb{M}}_A$ where $A$ is any unbounded connected component of $M_{(\alpha_{l-1}, +\infty)}$. Then there is a tree $(\mathcal{T}, \supsetneq)$ of filled subbundles of $\lb{M}$ such that $\mathcal{T} = \coprod_{l \in \mathbb{N} \cup \Set{0}} \tlevel(l)$, $\lb{A} \supsetneq \lb{C}$ if $\lb{C}$ is a filled subbundle of (not equal to) $\lb{A}$. Moreover, $\Pa{\mathcal{T}, {\supsetneq}}$ is a rooted locally finite leafless tree of height $\upomega$, and for each $l \in \mathbb{N} \cup \Set{0}$, $\tlevel(l)$ constructed above is indeed the $l$-th level of $\mathcal{T}$. \end{lemma} \begin{proof} Let $\lb{A}_i \in \tlevel(l_i) \subset \mathcal{T}$ where $l_i \in \mathbb{N} \cup \Set{0}$, for $i = 1, 2$ and $3$. By definition of connected components we have the following: if $\lb{A}_1 \supsetneq \lb{A}_2$, then $l_1 < l_2$; if $\lb{A}_1, \lb{A}_2 \supsetneq \lb{A}_3$ and $l_1 < l_2$, then $\lb{A}_1 \supsetneq \lb{A}_2$. Hence $\Pa{\mathcal{T}, \supsetneq}$ is a tree. The only root of $\mathcal{T}$ is $\lb{M} \in \tlevel(0)$, by induction $\tlevel(l)$ is indeed the $l$-th level of $\mathcal{T}$, which is finite, so $\mathcal{T}$ is locally finite. Hence $\Setby{\tdepth(\lb{A})}{\lb{A} \in \mathcal{T}} = \mathbb{N} \cup \Set{0}$, and $\theight(\mathcal{T}) = \upomega$. \end{proof} For any $A \in \tlevel(l)$, we define \begin{equation*} \Tse_{\mathcal{T}} \lb{A} = \lb{A}_{(-\infty, \alpha_{l+1}]}, \qquad \Shta_{\mathcal{T}} \lb{A} = \lb{A}_{(-\infty, \alpha_{l+2}]}. \end{equation*} Then we have \begin{equation*} \undsp\Pa{\Tse_{\mathcal{T}} \lb{A}} = A \setminus \coprod_{\lb{C} \in \tsuc(\lb{A})} \undsp(\lb{C}), \qquad \undsp\Pa{\Shta_{\mathcal{T}} \lb{A}} = A \setminus \coprod_{\lb{E} \in \tssuc(\lb{A})} \undsp(\lb{E}). \end{equation*} \section{Filtration theorem: combining the topological, combinatorial, and geometric-analytic constructions} \label{sec:filtration-theorem} \subsection{Statement} \label{ssec:statement-filtration-theorem} The objects in the following result are illustrated in Figures~\ref{fig:preparation} and~\ref{fig:tree}. \begin{theorem} \label{thm:filtration-theorem} Let $\lb{M} = (\pi, M, B, F, f, h)$ be an exhausted bundle with compact base and noncompact fiber without boundary. Suppose $\omega, \tau \in \FVform(\lb{M})$ that have equal fiber integral, and are smoothly commensurable. Then there is a tree $\Pa{\mathcal{T}, {\supsetneq}}$ of connected exhausted subbundles of $\lb{M}$ and $\Set{\omega_n}_{n \in \mathbb{N} \cup \Set{0}}, \Set{\tau_n}_{n \in \mathbb{N} \cup \Set{0}} \subset \FVform(\lb{M})$ such that $\omega_0 = \omega, \tau_0 = \tau$ and for any $n \in \mathbb{N}$, \begin{equation} \label{eq:support-in-level} \support(\omega_n - \omega_{n-1}) \cup \support(\tau_n - \tau_{n-1}) \subset \bigcup_{\lb{C} \in \tlevel(2n-2)} \Pa{\undsp(\Shta_\mathcal{T} \lb{C})}^\circ, \end{equation} and for each $\lb{A} \in \tlevel(2n-3)$ (when $n > 1$), $\lb{C} \in \tlevel(2n-2)$, $\lb{E} \in \tlevel(2n-1)$, \begin{align} \rint_{\Tse_\mathcal{T} \lb{M}} \omega_1 &= \rint_{\Tse_\mathcal{T} \lb{M}} \tau_1, &\rint_{\Shta_\mathcal{T} \lb{A}} \omega_n &= \rint_{\Shta_\mathcal{T} \lb{A}} \tau_n \text{~for $n > 1$}; \label{eq:volume-in-Shta-A} \\ \rint_{\Shta_\mathcal{T} \lb{C}} \omega_n &= \rint_{\Shta_\mathcal{T} \lb{C}} \omega_{n-1}, &\rint_{\Shta_\mathcal{T} \lb{C}} \tau_n &= \rint_{\Shta_\mathcal{T} \lb{C}} \tau_{n-1}; \label{eq:volume-in-Shta-C} \\ \rint_{\lb{E}} \omega_n &= \rint_{\lb{E}} \tau_n. & & \label{eq:volume-in-E} \end{align} \end{theorem} \inputfigure{fiber-bundle}{preparation}{ The exhausted bundle $\lb{M} = (\pi, M, B, F, f, h)$ in Theorem~\ref{thm:filtration-theorem}. It can have finitely or infinitely many ends going upwards in the top of the figure. } The abstract tools we have developed so far in the paper allow us to give an inductive proof of Theorem~\ref{thm:filtration-theorem} with a minimum of technical fuss. \inputfigure{tree}{tree}{ The tree $\mathcal{T}$ in Theorem~\ref{thm:filtration-theorem}. } \subsection{Inductive proof of filtration theorem} \label{ssec:proof-filtration-theorem} We aim to find $\alpha_0 = -\infty$ and $\Set{\alpha_l}_{l \in \mathbb{N}} \subset \Regular(f) \cap f(M)$ such that $\mathcal{T}$ is constructed by Lemma~\ref{lem:bundle-slicing-tree}. Although $\mathcal{T}$ is determined by the sequence $\Set{\alpha_l}_{l \in \mathbb{N} \cup \Set{0}}$, the $\tlevel(l)$ only has something to do with $\alpha_l$, so if we have found $\Set{\alpha_l}_{0 \leq l \leq m}$ for some $m \in \mathbb{N} \cup \Set{0}$ then we say $\mathcal{T}$ is constructed up to the $m$-th level, and $\tlevel(l)$ and relating constructions makes sense for any $l$ with $0 \leq l \leq m$. We proceed by induction on $n \in \mathbb{N} \cup \Set{0}$ to find $\alpha_{2n-1}$, $\alpha_{2n}$ and $\omega_n, \tau_n \in \FVform(\lb{M})$ such that $\rint_{\lb{E}} \omega_n = \rint_{\lb{E}} \tau_n$ for any $\lb{E} \in \tlevel(2n-1)$ ($\lb{E} \in \tlevel(0)$ if $n = 0$). \emph{Case $0$.} Set $\alpha_0 = -\infty$, then $\lb{M}_{(-\infty, \alpha_0]} = \emptyset$, and $\tlevel(0) = \Set{\lb{M}}$. Since $\omega$ and $\tau$ has equal fiber integral and $\lb{M}$ has connected fiber, we have \begin{equation} \label{eq:equal-volume-M} \rint_{\lb{M}} \omega_0 = \rint_{\lb{M}} \tau_0. \end{equation} \medskip \emph{Case $(n-1)$ for $n \in \mathbb{N}$.} Assume by induction \begin{equation} \label{eq:equal-volume-A} \rint_{\lb{A}} \omega_{n-1} = \rint_{\lb{A}} \tau_{n-1}. \end{equation} for any $\lb{A} \in \tlevel(2n-3)$ ($\lb{A} \in \tlevel(0)$ when $n = 1$). \medskip \emph{Case $n$ for $n \in \mathbb{N}$.} Let $\alpha_{2n-1} \in \Regular(f)$ such that $\alpha_{2n-1} > \max\Setby{\theta_{\lb{C}}}{\lb{C} \in \tlevel(2n-2)}$. Then $\mathcal{T}$ is constructed up to the $(2n-1)$-th level. Let $\lb{A} \in \tlevel(2n-3)$ (if $n = 1$ let $\lb{A} = \lb{M}$ and replace $\tssuc(\lb{A})$ by $\tsuc(\lb{M})$, $\Shta_\mathcal{T} \lb{A}$ by $\Tse_\mathcal{T} \lb{M}$ throughout this paragraph). The base of $\Rls \lb{A}$ is $B_A$. Let $\tssuc_0(\lb{A})$ (resp. $\tssuc_1(\lb{A})$) be the subcollection of elements in $\tssuc(\lb{A})$ with finite (resp. infinite) volume. For any $\lb{E} \in \tssuc(\lb{A})$, we define $\delta_{\lb{E}} \in \smth(B_E; \mathbb{R})$ as follows: if $\lb{E}$ has finite volume, let \begin{equation*} \delta_{\lb{E}} = \rint_{\lb{E}} \omega_{n-1} - \rint_{\lb{E}} \tau_{n-1}; \end{equation*} if $\lb{E}$ has infinite volume, let \begin{equation*} \delta_{\lb{E}} = \dfrac{1}{\sum\limits_{\lb{G} \in \tssuc_1(\lb{A})} \# \kappa^{\lb{G}}_{\lb{A}}} (\kappa^{\lb{E}}_{\lb{A}})^* \Pa{ \rint_{\Shta_\mathcal{T} \lb{A}} \tau_{n-1} - \rint_{\Shta_\mathcal{T} \lb{A}} \omega_{n-1} - \sum_{\lb{G} \in \tssuc_0(\lb{A})} (\kappa^{\lb{G}}_{\lb{A}})_* \delta_{\lb{G}}}. \end{equation*} Then by \eqref{eq:covering-space-equation} and \eqref{eq:equal-volume-A} we have \begin{equation} \sum_{\lb{E} \in \tssuc(\lb{A})} (\kappa^{\lb{E}}_{\lb{A}})_* \delta_{\lb{E}} = \rint_{\Shta_\mathcal{T} \lb{A}} \tau_{n-1} - \rint_{\Shta_\mathcal{T} \lb{A}} \omega_{n-1}. \end{equation} For any $\lb{C} \in \tsuc(\lb{A})$, let $B_C$ be the base of $\Rls \lb{C}$, and let $u_{\lb{C}} \in \smth(B_C; \mathbb{R})$ be such that \begin{align*} \max\Pa{- \rint_{\Tse_\mathcal{T} \lb{C}} \omega_{n-1}, - \rint_{\Tse_\mathcal{T} \lb{C}} \tau_{n-1} + \sum_{\lb{E} \in \tsuc(\lb{C})} (\kappa^{\lb{E}}_{\lb{C}})_* \delta_{\lb{E}}} < u_{\lb{C}} < \rint_{\lb{C}} \omega_{n-1} - \rint_{\Tse_\mathcal{T} \lb{C}} \omega_{n-1}. \end{align*} Note that if $\lb{C}$ has finite volume, \begin{align*} &\phantom{{}=} \Pa{\rint_{\lb{C}} \omega_{n-1} - \rint_{\Tse_\mathcal{T} \lb{C}} \omega_{n-1}} - \Pa{ -\rint_{\Tse_\mathcal{T} \lb{C}} \tau_{n-1} + \sum_{\lb{E} \in \tsuc(\lb{C})} (\kappa^{\lb{E}}_{\lb{C}})_* \delta_{\lb{E}}} \\ &= \rint_{\lb{C}} \omega_{n-1} + \Pa{\rint_{\Tse_\mathcal{T} \lb{C}} \tau_{n-1} - \rint_{\Tse_\mathcal{T} \lb{C}} \omega_{n-1}} + \sum_{\lb{E} \in \tsuc(\lb{C})} (\kappa^{\lb{E}}_{\lb{C}})_* \Pa{\rint_{\lb{E}} \tau_{n-1} - \rint_{\lb{E}} \omega_{n-1}} \\ &= \rint_{\lb{C}} \tau_{n-1} > 0, \end{align*} so such $u_{\lb{C}}$ exists. Since \begin{equation*} u_{\lb{C}} < \sum_{\lb{E} \in \tsuc(\lb{C})} (\kappa^{\lb{E}}_{\lb{C}})_* \rint_{\lb{E}} \omega_{n-1} = \rint_{\lb{C}} \omega_{n-1} - \rint_{\Tse_\mathcal{T} \lb{C}} \omega_{n-1}, \end{equation*} by Lemma~\ref{lem:approximation-lemma} on the covering space $\coprod_{\lb{E} \in \tsuc(\lb{C})} B_E \to B_C$, we can choose $v_{\lb{E}} \in \smth(B_E; \mathbb{R})$ (where $B_E$ is the base of $\Rls \lb{E}$) such that $v_{\lb{E}} < \rint_{\lb{E}} \omega_{n-1}$ and $\sum_{\lb{E} \in \tsuc(\lb{C})} (\kappa^{\lb{E}}_{\lb{C}})_* v_{\lb{E}} = u_{\lb{C}}$. For any $\lb{E} \in \tsuc(\lb{C})$, if $\lb{E}$ has infinite volume, take $\beta_{\lb{E}} > \theta_{\lb{E}}$ be any regular value of $f$. Otherwise, the following function defined on $B_E \times \mathbb{R}$, \begin{equation} \label{eq:lambdaps} (\cdot, \beta) \mapsto \min\Pa{\int_{(\Rls \lb{E})_{(-\infty, \beta]}} \omega_{n-1}, \int_{(\Rls \lb{E})_{(-\infty, \beta]}} \tau_{n-1} + \delta_{\lb{E}}} - v_{\lb{E}} \end{equation} is continuous in the first variable, is increasing in $\beta$, and converges to $\rint_{\lb{E}} \omega_{n-1} - v_{\lb{E}} > 0$ as $\beta \to +\infty$ pointwise. Since $B_E$ is compact there is $\beta_{\lb{E}} > \alpha_{2n-1}$ such that \eqref{eq:lambdaps} is positive when $\beta = \beta_{\lb{E}}$. Let $\alpha_{2n} = \max_{\lb{E} \in \tlevel(2n-1)} \beta_E$, then $\mathcal{T}$ is constructed up to the $2n$-th level. So $\Tse_\mathcal{T} \lb{E} = \lb{E}_{(-\infty, \alpha_{2n}]}$, then we have $v_{\lb{E}} < \rint_{\Tse_\mathcal{T} \lb{E}} \omega_{n-1}$, and $v_{\lb{E}} - \delta_{\lb{E}} < \rint_{\Tse_\mathcal{T} \lb{E}} \tau_{n-1}$. Since all right hand sides are positive smooth functions, by Lemma~\ref{lem:volume-lemma}, there are $\omega_n, \tau_n \in \FVform(\lb{M})$ such that \begin{align*} \rint_{\Tse_\mathcal{T} \lb{C}} \omega_n &= \rint_{\Tse_\mathcal{T} \lb{C}} \omega_{n-1} + u_{\lb{C}}, &\rint_{\Tse_\mathcal{T} \lb{C}} \tau_n &= \rint_{\Tse_\mathcal{T} \lb{C}} \tau_{n-1} + u_{\lb{C}} - \sum_{\lb{E} \in \tsuc(\lb{C})} (\kappa^{\lb{E}}_{\lb{C}})_* \delta_{\lb{E}}, \\ \rint_{\Tse_\mathcal{T} \lb{E}} \omega_n &= \rint_{\Tse_\mathcal{T} \lb{E}} \omega_{n-1} - v_{\lb{E}} , &\rint_{\Tse_\mathcal{T} \lb{E}} \tau_n &= \rint_{\Tse_\mathcal{T} \lb{E}} \omega_{n-1} - (v_{\lb{E}} - \delta_{\lb{E}}), \end{align*} and \begin{align*} \support(\omega_n - \omega_{n-1}) \cup \support(\tau_n - \tau_{n-1}) &\subset \lb{M}_{(\alpha_{2n-2}, \alpha_{2n})} = \bigcup_{\lb{C} \in \tlevel(2n-2)} \Pa{\undsp(\Shta_\mathcal{T} \lb{C})}^\circ. \end{align*} Then we have \begin{align*} \rint_{\Shta_\mathcal{T} \lb{A}} \omega_n &= \rint_{\Shta_\mathcal{T} \lb{A}} \omega_{n-1} + \sum_{\lb{C} \in \tsuc(\lb{A})} (\kappa^{\lb{C}}_{\lb{A}})_* u_{\lb{C}} \\ &= \rint_{\Shta_\mathcal{T} \lb{A}} \tau_{n-1} - \sum_{\lb{E} \in \tssuc(\lb{A})} (\kappa^{\lb{C}}_{\lb{A}})_* \Pa{(\kappa^{\lb{E}}_{\lb{C}})_* \delta_{\lb{E}} - u_{\lb{C}}} = \rint_{\Shta_\mathcal{T} \lb{A}} \tau_n, \end{align*} and \begin{align*} \rint_{\Shta_\mathcal{T} \lb{C}} \omega_n &= \rint_{\Tse_\mathcal{T} \lb{C}} \omega_n + \sum_{\lb{E} \in \tsuc(\lb{C})} (\kappa^{\lb{E}}_{\lb{C}})_* \rint_{\Tse_\mathcal{T} \lb{E}} \omega_n = \rint_{\Shta_\mathcal{T} \lb{C}} \omega_{n-1}, \\ \rint_{\Shta_\mathcal{T} \lb{C}} \tau_n &= \rint_{\Tse_\mathcal{T} \lb{C}} \tau_n + \sum_{\lb{E} \in \tsuc(\lb{C})} (\kappa^{\lb{E}}_{\lb{C}})_* \rint_{\Tse_\mathcal{T} \lb{E}} \tau_n = \rint_{\Shta_\mathcal{T} \lb{C}} \tau_{n-1}, \end{align*} and \begin{equation*} \begin{split} \rint_{\lb{E}} \omega_n &= \rint_{\Tse_\mathcal{T} \lb{E}} \omega_n + \rint_{\lb{E}} \omega_{n-1} - \rint_{\Tse_\mathcal{T} \lb{E}} \omega_{n-1} \\ &= \rint_{\lb{E}} \omega_{n-1} - v_{\lb{E}} = \rint_{\lb{E}} \tau_{n-1} - (v_{\lb{E}} - \delta_{\lb{E}}) = \rint_{\lb{E}} \tau_n. \end{split} \end{equation*} \section{Geometric-analytic constructions} \label{sec:geometric-analytic} Throughout this section, $\lb{M} = (\pi, M, B, F, f, h)$ is an oriented connected exhausted bundle with compact base and noncompact connected fiber without boundary. \subsection{Fiber forms with compactly supported difference} \label{sec:compact-support} \begin{lemma} \label{lem:fiber-collar-neighborhood} Suppose $\lb{M}$ has compact base. Let $N$ be a compact hypersurface of $M$ through regular points of $f$, on which $\pi$ is a submersion. Then there exists $\varepsilon > 0$ and a diffeomorphism $\Phi \colon N \times (-\varepsilon, \varepsilon) \to V_N$ such that $V_N \subset M$ is a neighborhood of $N$, $\Phi(y, 0) = y$, $\pi(\Phi(y, s)) = \pi(y)$ and $f(\Phi(y, s)) = f(y) + s$ for any $(y, s) \in N \times (-\varepsilon, \varepsilon)$. Moreover, if in addition that $N$ is a filled subspace of $\lb{M}$ then $V_N$ is also a filled subspace of $\lb{M}$. Let $\lb{V} = \Res{\lb{M}}_V$ and $\lb{V}_{\lb{N}} = \Res{\lb{M}}_{V_N}$. If $\lb{N}$ has connected fiber then $\lb{V}_{\lb{N}}$ has connected fiber too. \end{lemma} \begin{proof} Denote by $\mathrm{V}\lb{M} = \ker (\der \pi \colon \mathrm{T}M \to \mathrm{T}B)$ the vertical tangent bundle of $\lb{M}$. Pick an arbitrary Riemannian metric $g$ on $M$. Let $Y \in \Gamma(\mathrm{V}\lb{M})$ such that $\Res{Y}_{\pi^{-1}(p)} = \nabla\Pa{\Res{f}_{\pi^{-1}(p)}}$ is the gradient of $\Res{f}_{\pi^{-1}(p)}$ for any $p \in B$, then $\Res{Y(f)}_{\pi^{-1}(p)} = \Res{Y}_{\pi^{-1}(p)}\Pa{\Res{f}_{\pi^{-1}(p)}} = \abs{\nabla \Pa{\Res{f}_{\pi^{-1}(p)}}}_g^2$. Hence there is a neighborhood $V_N \supset N$ such that $Y(f) > 0$ in $V_N$. Let $X \in \Gamma(\mathrm{V}\lb{M})$ be such that $X(x) = \abs{\nabla \Pa{\Res{f}_{\pi^{-1}(p)}}(x)}_g^{-2} Y(x)$, then $X(f) = 1$ in $V_N$. Take the flow of $X$, $\Phi \colon N \times (-\varepsilon, \varepsilon) \to M , (y, s) \mapsto x$, that is $\Phi(y, 0) = y$ for all $y \in N$ and $\frac{\partial \Phi}{\partial s}(y, s) = X(\Phi(y, s))$ for all $(y, s) \in N \times (-\varepsilon, \varepsilon)$, for $\varepsilon > 0$ small enough such that the image of $\Phi$ is contained in $V_N$. Then we redefine $V_N = \Phi(N \times (-\varepsilon, \varepsilon))$. Since $X$ is vertical and $X(f) = 1$ in $V_N$, we have $\pi(\Phi(y, s)) = \pi(y)$ and $f(\Phi(y, s)) = f(y) + s$ for any $(y, s) \in N \times (-\varepsilon, \varepsilon)$, and $\Phi$ is a diffeomorphism. If $N$ is a filled subspace of $\lb{M}$ then there is $\alpha \in \Regular(f)$ such that $N \subset \Conn{f^{-1}(\alpha)}$, so $V_N$ is the union of some connected components of $f^{-1}((\alpha - \varepsilon, \alpha + \varepsilon))$ and by Lemma~\ref{lem:subbundle-slicing} $V_N$ is a filled subspace of $\lb{M}$. If $\lb{N}$ has connected fiber, then since the fiber of $\lb{V}_{\lb{N}}$ is the image of the fiber of $\lb{N}$ under the flow of $X$ restricted on the fiber, which is connected. \end{proof} \begin{theorem} \label{thm:prelimitive-compact-support} Let $Z$ be an open submanifold of $F$ such that $\overline{Z}$ is a submanifold of $F$ with boundary $\partial Z$. Then for any $q \in \mathbb{N}$ with $1 \leq q \leq \dim F$ there is an operator preserving smooth families of $q$-forms \begin{equation*} I_Z^q \colon \Setby{\xi \in \Omega^q_\cspt(F)}{\support \xi \subset Z, \Res{\xi}_Z \in \der \Omega^{q-1}_\cspt(Z)} \to \Setby{\eta \in \Omega^{q-1}_\cspt(F)}{\support \eta \subset \overline{Z}} \end{equation*} satisfying $d \circ I_Z^q = \identity$. \end{theorem} \begin{proof} By \cite{MR1893604} there is a weighted Hodge-Laplacian $\Delta_\mu \colon \Omega^q_\cspt(Z) \to \Omega^q_\cspt(Z)$ on $Z$ equipped with a specific metric $g$ and measure $\mu$. Its Green operator $G_\mu \colon \Omega^q_\cspt(Z) \to \Omega^q(Z)$ and the weighted codifferential $\delta_\mu \colon \Omega^q_\cspt(Z) \to \Omega^{q-1}_\cspt(Z)$ satisfy the identity $\der \circ \delta_\mu \circ G_\mu \circ \der = \der$. Moreover if $\eta \in G_\mu(\Omega^q_\cspt(Z))$, then it has an extension $\tilde{\eta} \in \Omega^q_\cspt(F)$, such that $\support \tilde{\eta} \subset \overline{Z}$ and $\Res{\tilde{\eta}}_Z = \eta$. For $\xi \in \Omega^q_\cspt(F)$ that is supported in $Z$, if $\Res{\xi}_Z \in \der \Omega^{q-1}_\cspt(Z)$ we define $I_Z^q(\xi)$ as the extension of $(\delta_\mu \circ G_\mu) (\Res{\xi}_Z)$ to $\Omega^{q-1}_\cspt(F)$, see Figure~\ref{fig:inverseextder}. Then we have $\der \circ I_Z^q = \identity$. \end{proof} \inputfigure{inverse-of-d}{inverseextder}{ From compactly supported forms to forms with zero extensions. } \begin{lemma} \label{thm:prelimitive-compact-support-bundle} Let $\lb{W} = (\Res{\pi}_W, W, B, Z, \Res{f}_W, \Res{h}_Z)$ be a filled subbundle of $\lb{M}$ with connected fiber. Suppose $W$ is an open submanifold of $M$ such that $\overline{W}$ is a compact submanifold with boundary $\partial W$. Let $\xi \in \pcspt{\Omega^q_F}(\lb{M})$. If $\support \xi \subset W$ and $\Res{\xi}_{W \cap \pi^{-1}(p)} \in \der \Omega^{q-1}_\cspt(W \cap \pi^{-1}(p))$ for any $p \in B$, then there is an $\eta \in \pcspt{\Omega^{q-1}_F}(\lb{M})$ such that $\xi = \fibder \eta$ and $\support \eta \subset \overline{W}$. \end{lemma} \begin{proof} Let $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ be a local trivialization of $\lb{M}$ with respect to which $W$ is a filled subspace of $\lb{M}$, then we can assume $\phi_i\Pa{W \cap \pi^{-1}(U_i)} = U_i \times Z$ for any $i \in \mathcal{I}$. By Lemma~\ref{lem:compact-exhausted-bundle}, $\overline{Z}$ is compact. Let $\Set{\chi_i}_{i \in \mathcal{I}}$ be a partition of unity subordinated to the open cover $\Set{U_i}_{i \in \mathcal{I}}$ of $B$. We apply Theorem~\ref{thm:prelimitive-compact-support} to $Z$ to get an operator $I_Z^q$, then define \begin{equation*} \eta = \sum_{i \in \mathcal{I}} \phi_i^* I_Z^q \Pa{\phi_i^{-1}}^* ((\chi_i \circ \pi) \cdot \xi). \end{equation*} Since $\der \circ I_Z^q = \identity$, we have $\xi = \fibder \eta$ and $\support \eta \subset \overline{W}$. \end{proof} Figure~\ref{fig:lemma1} illustrates the main point of the following lemma, where the shaded region is the support of $\omega - \tau$. \begin{lemma} \label{lem:lemma1} Let $\lb{V}$ be a filled subbundle of $\lb{M}$ with connected fiber. Suppose $V = \undsp(\lb{V})$ is an open submanifold of $M$ such that $\overline{V}$ is a compact submanifold with boundary $\partial V$. Let $\omega, \tau \in \FVform(\lb{M})$ such that $\support(\omega - \tau) \subset V$ and $\int_{\lb{V}} \omega = \int_{\lb{V}} \tau$. Then there is a fiber diffeomorphism $\varphi \colon M \to M$ such that $\varphi$ is the identity in a neighborhood of $M \setminus V$ and $\varphi^* \omega = \tau$. \end{lemma} \begin{proof} Let $N = \partial V$. Applying Lemma~\ref{lem:fiber-collar-neighborhood} to $N$ there are $\varepsilon > 0$ and $V_N$ a neighborhood of $N$ with the properties stated in the lemma. Since $B$ is compact and $\support(\omega - \tau) \subset V$, we may decrease $\varepsilon$ if necessary so that $\support(\omega - \tau) \subset V \setminus \overline{V_N}$. Let $W = V \setminus \overline{V_N}$. For each $p \in B$, let $W_p = W \cap \pi^{-1}(p)$. Since the map $H_\cspt^{\dim F}(W_p) \to \mathbb{R}, [\omega] \mapsto \int_{W_p} \omega$ is a linear isomorphism, and $\int_{\lb{V}} \omega = \int_{\lb{V}} \tau$, we have $\Res{(\omega_p - \tau_p)}_{W_p} \in \der \Omega_\cspt^{\dim F-1}(W_p)$ for any $p \in B$. Therefore by Lemma~\ref{thm:prelimitive-compact-support-bundle} there exists $\sigma \subset \pcspt{\Omega^{\dim F-1}_F}(\lb{M})$ with $\support \sigma \subset \overline{W}$ such that $\fibder \sigma = \omega - \tau$. Let $\omega_t = (1-t)\omega + t\tau \in \FVform(\lb{M})$ for any $t \in [0, 1]$. Since $\omega_t$ is nowhere vanishing there exists a unique smooth family of vertical vector fields $\Set{X_t}_{t \in [0, 1]} \subset \Gamma(\mathrm{V}\lb{M})$ where each $X_t$ is supported in $\overline{W}$ and such that $\omega_t(X_t, \cdot) = \sigma$. Since $\overline{V}$ is compact, for each $t \in [0,1]$ there exists a fiber diffeomorphism $\varphi_t \colon M \to M$ generated by $X_t$ that is the identity outside of $\overline{W}$. Then $\varphi_t^* \omega_t = \omega$. If $\varphi = \varphi_1^{-1}$ then we have $\varphi^* \omega = \tau$. Since $X_t = 0$ in $M \setminus W$ for $t \in [0, 1]$, $\varphi$ is the identity outside of $\overline{W}$. \end{proof} \subsection{The transfer of volumes} \label{sec:transfer-volume} \begin{lemma} \label{lem:volume-lemma} Let $\omega \in \FVform(\lb{M})$, and let $\lb{K}$ be a connected filled subbundle of $\lb{M}$ whose underlying space $K$ is a compact manifold with or without boundary which has nonempty interior. Let $B_K$ be the base of $\Rls \lb{K}$, and let $w \in \smth(B_K; \mathbb{R})$. Then there exists $\tau \in \FVform(\lb{M})$ such that $\support(\omega - \tau) \subset K^\circ$ and $\rint_{\lb{K}} \tau = w$. \end{lemma} \begin{proof} Let $\xi \in \FVform(\lb{M})$ be such that $\support(\xi - \omega) \subset K^\circ$ and $\rint_{\lb{K}} \xi < w$. Let $\eta \geq 0$ be a fiber top-form on $\lb{M}$ such that $\support \eta \subset K^\circ$ and $\rint_{\lb{K}} \eta > 0$. Let $\Rls (\Res{\pi}_K)$ be the bundle map of $\Rls(\lb{K})$. Then define \begin{equation*} \tau = \xi + \Rls (\Res{\pi}_K)^*\Pa{\frac{w - \rint_{\lb{K}} \xi}{\rint_{\lb{K}} \eta}} \eta. \end{equation*} Note that $\frac{w - \rint_{\lb{K}} \xi}{\rint_{\lb{K}} \eta}$ is a function on $B_K$, and $\Rls (\Res{\pi}_K) \colon K \to B_K$. \end{proof} \begin{lemma} \label{lem:lemma2} Let $\lb{N}$ be a filled subbundle of $\lb{M}$ with connected fiber such that its underlying space $N$ is a hypersurface of $M$ and $\omega, \tau \in \FVform(\lb{M})$. Then there is $V_N$ a neighborhood of $N$, $V_N^+$ and $V_N^-$ are the connected components of $V_N \setminus V$. Moreover, $V_N$ is a filled subspace of $\lb{M}$. Let $\lb{V}_{\lb{N}} = \Res{\lb{M}}_{V_N}$ and $\lb{V}_{\lb{N}}^\pm = \Res{\lb{M}}_{V_N^\pm}$. There is a fiber diffeomorphism $\varphi \colon M \to M$ such that the following hold: $\varphi$ is the identity in a neighborhood of $M \setminus V_N$; $\varphi^*\omega = \tau$ in a neighborhood of $N$; $\int_{\lb{V}_{\lb{N}}^+} \varphi^* \omega = \int_{\lb{V}_{\lb{N}}^+} \omega$; and $\int_{\lb{V}_{\lb{N}}^-} \varphi^* \omega = \int_{\lb{V}_{\lb{N}}^-} \omega$. \end{lemma} \begin{proof} By Lemma~\ref{lem:fiber-collar-neighborhood}, there exists $\lb{V}_{\lb{N}}$ with underlying space $V_N$ as a filled subbundle of $\lb{M}$, $\varepsilon > 0$ and a diffeomorphism $\Phi \colon N \times (-\varepsilon, \varepsilon) \to V_N$ such that $V_N \subset M$ is a neighborhood of $N$, $\Phi(y, 0) = y$, $\pi(\Phi(y, s)) = \pi(y)$ and $f(\Phi(y, s)) = f(y) + s$ for any $(y, s) \in N \times (-\varepsilon, \varepsilon)$, see Figure~\ref{fig:lemma2}. Let $V_N^+ = \Phi(N \times (0, \varepsilon))$ and $V_N^- = \Phi(N \times (-\varepsilon, 0))$. First we consider $\Phi(N \times [0, \varepsilon))$. Since $B$ is compact, there exists $\delta$ with $0 < \delta < \varepsilon / 2$ such that $\int_{\Res{\lb{M}}_{\Phi(N \times (0, \varepsilon-\delta))}} \tau > \int_{\Res{\lb{M}}_{\Phi(N \times (0, \delta))}} \omega$ and $\int_{\Res{\lb{M}}_{\Phi(N \times (0, \varepsilon-\delta))}} \omega > \int_{\Res{\lb{M}}_{\Phi(N \times (0, \delta))}} \tau$. \inputfigure{graph-alpha}{graphalpha}{ The graph of $\zeta(s, t)$. } Let $\zeta \colon (0, \varepsilon) \times (0, 1) \to [0,1]$ be a smooth function with the properties (see Figure~\ref{fig:graphalpha}): \begin{equation*} \begin{cases} \zeta(s, \cdot) = 1, &s \in (0, \delta]; \\ \displaystyle\lim_{t \to 0+} \zeta(s, t) = 0, \frac{\partial \zeta}{\partial t}(s, \cdot) > 0, \displaystyle\lim_{t \to 1-} \zeta(s, t) = 1, &s \in (\delta, \varepsilon - \delta); \\ \zeta(s, \cdot) = 0, &s \in [\varepsilon - \delta, \varepsilon). \end{cases} \end{equation*} Define $\theta \colon (0, 1) \times B \to \mathbb{R}$ by \begin{equation*} \theta(t, \cdot) = \int_{\lb{V}_{\lb{N}}^+} \zeta(s(\cdot), t) \tau - \int_{\lb{V}_{\lb{N}}^+} \zeta(s(\cdot), 1 - t) \omega \end{equation*} where $s = \proj_2 \circ \Phi^{-1} \colon \Phi(N \times (-\varepsilon, \varepsilon)) \to (-\varepsilon, \varepsilon)$. Since $\zeta$ is smooth and $\omega, \tau \in \FVform(M)$ it follows that $\theta$ is smooth. Furthermore \begin{equation*} \frac{\partial \theta}{\partial t}(t, \cdot) = \int_{\lb{V}_{\lb{N}}^+} \frac{\partial \zeta}{\partial t}(s(\cdot), t) \tau + \int_{\lb{V}_{\lb{N}}^+} \frac{\partial \zeta}{\partial t}(s(\cdot), 1 - t) \omega > 0 \end{equation*} for any $t \in (0, 1)$ and $\lim_{t \to 0+} \theta(t, p) < 0 < \lim_{t \to 1-} \theta(t, p)$ for any $p \in B$. Then for every $p \in B$ there is a unique $t = t(p)$ solving $\theta(t(p), p) = 0$. By differentiation rules the derivatives of $t$ of any order can be explicitly given in terms of the derivatives of $\theta$, so $t \colon B \to \mathbb{R}$ is smooth. Define $\lambda(x) = \zeta(s(x), t(\pi(x)))$ and $\mu(x) = \zeta(s(x), 1 - t(\pi(x)))$ in $V_N^+$. The functions $\lambda$ and $\mu$ are smooth in $x$ and satisfy $\int_{\lb{V}_{\lb{N}}^+} \mu \omega = \int_{\lb{V}_{\lb{N}}^+} \lambda \tau$. Analogously we can define $\lambda$ and $\mu$ in $V_N^-$, then let $\lambda = \mu = 1$ on $N$. Notice that $\lambda = \mu = 1$ in $\Phi(N \times [-\delta, \delta])$, so we obtain smooth extensions of $\lambda, \mu$ which we also denote by $\lambda, \mu \colon V_N \to \mathbb{R}$. Hence \begin{align*} \int_{\lb{V}_{\lb{N}}^+} \Pa{(1 - \mu) \omega + \lambda \tau} &= \int_{\lb{V}_{\lb{N}}^+} \omega, \\ \int_{\lb{V}_{\lb{N}}^-} \Pa{(1 - \mu) \omega + \lambda \tau} &= \int_{\lb{V}_{\lb{N}}^-} \omega. \end{align*} By Lemma~\ref{lem:lemma1} applied to $(1 - \mu) \omega + \lambda \tau$ and $\omega$ on $\lb{V}_{\lb{N}}^+$ and $\lb{V}_{\lb{N}}^-$ respectively, combining the results we obtain a fiber diffeomorphism $\varphi \colon M \to M$ such that, $\varphi = \identity$ in $M \setminus \Phi(N \times (\delta - \varepsilon, \varepsilon - \delta))$ and $\varphi^* \omega = (1 - \mu) \omega + \lambda \tau$. \end{proof} \begin{lemma} \label{lem:lemma3} Let $\Set{L_j}_{j \in \mathbb{N}}$ be a cover of $M$ by compact submanifolds with boundary, which have the same dimension as $M$, and whose interiors are pairwise disjoint. Suppose for any $j \in \mathbb{N}$, $L_j$ is a filled subspace, and $\lb{L}_j = \Res{\lb{M}}_{L_j}$ is a filled subbundle of $\lb{M}$ with connected fiber. If $\omega, \tau \in \FVform(\lb{M})$ are such that $\int_{\lb{L}_j} \omega = \int_{\lb{L}_j} \tau$ for each $j \in \mathbb{N}$ then there is a fiber diffeomorphism $\varphi \colon M \to M$ such that $\varphi^* \omega = \tau$. \end{lemma} \begin{proof} By the construction of $\Set{\lb{L}_j}_{j \in \mathbb{N}}$, any three different $L_j$'s for $j \in \mathbb{N}$ do not intersect. Let $\mathcal{C} = \Setby{\Res{\lb{M}}_N}{N \in \Conn{L_j \cap L_k}, j, k \in \mathbb{N}, j \neq k}$. Then $\mathcal{C}$ is a collection of pairwise disjoint filled subbundles of $\lb{M}$ whose underlying spaces are hypersurfaces of $M$. So for each $\lb{N} \in \mathcal{C}$ with underlying space $N$, let $j, k \in \mathbb{N}$ be such that $N \subset L_j \cap L_k$, by Lemma~\ref{lem:fiber-collar-neighborhood}, we obtain $\varepsilon_N > 0$ and a diffeomorphism $\Phi_N \colon V_N \times (-\varepsilon_N, \varepsilon_N) \to V_N$ where $V_N$ is a neighborhood of $N$ and a filled subspace of $\lb{M}$. We require $V_N \subset L_j \cup L_k$. Let $\lb{V}_{\lb{N}} = \Res{\lb{M}}_{V_N}$. We apply Lemma~\ref{lem:lemma2} to $\Rls \lb{V}_{\lb{N}}$ to obtain a fiber diffeomorphism $\varphi_N\colon M \to M$ such that $\varphi_N = \identity$ in a neighborhood of $M \setminus V_N$, $\varphi_N^*\omega = \tau$ in a neighborhood of $N$, and \begin{equation*} \rint_{\lb{V}_{\lb{N}}^+} \varphi_N^* \omega = \rint_{\lb{V}_{\lb{N}}^+} \omega, \qquad \rint_{\lb{V}_{\lb{N}}^-} \varphi_N^* \omega = \rint_{\lb{V}_{\lb{N}}^-} \omega. \end{equation*} Thus in particular, \begin{equation*} \int_{\lb{V}_{\lb{N}}^+} \varphi_N^* \omega = \int_{\lb{V}_{\lb{N}}^+} \omega, \qquad \int_{\lb{V}_{\lb{N}}^-} \varphi_N^* \omega = \int_{\lb{V}_{\lb{N}}^-} \omega. \end{equation*} hence $\int_{\lb{L}_j} \varphi_N^* \omega = \int_{\lb{L}_j} \omega, \int_{\lb{L}_k} \varphi_N^* \omega = \int_{\lb{L}_k} \omega$. If necessary, choose $\varepsilon_N$ small so that $\Set{\overline{V_N}}_{N \in \mathcal{C}}$ is mutually disjoint. Since replacing $\omega$ by $\varphi_N^*\omega$ each time does not change the fiber volume of $\lb{L}_j$ for any $j \in \mathbb{N}$, we compose these $\varphi_N$ for $N = \undsp(\lb{N})$, $\lb{N} \in \mathcal{C}$, as they are the identity away from disjoint open sets, to obtain a fiber diffeomorphism $\varphi' \colon M \to M$ such that $\omega' = \varphi'^* \omega$ is equal to $\tau$ in some neighborhood of $\bigcup_{N \in \mathcal{C}} N$ and $\int_{\lb{L}_j} \omega' = \int_{\lb{L}_j} \omega = \int_{\lb{L}_j} \tau$ for each $j \in \mathbb{N}$. Applying Lemma~\ref{lem:lemma1} to each $\lb{L}_j$ for $j \in \mathbb{N}$ we get a fiber diffeomorphism $\psi_j \colon M \to M$ such that $\tau = \psi_j^* \omega'$ in $L_j$ and $\psi_j = \identity$ in a neighborhood of $M \setminus L_j$. Replacing $\omega'$ by $\psi_j^* \omega'$ each time and composing $\Set{\psi_j}_{j \in \mathbb{N}}$ we obtain a fiber diffeomorphism $\psi' \colon M \to M$ such that $\tau = \psi'^* \omega'$. Let $\varphi = \varphi' \circ \psi'$. \end{proof} \inputfiguretri {lemma-1}{lemma1}{Lemma~\ref{lem:lemma1}.} {lemma-2}{lemma2}{Lemma~\ref{lem:lemma2}.} {lemma-3}{lemma3}{Lemma~\ref{lem:lemma3}.} \subsection{Covering space and approximation lemma} \label{ssec:approximation-lemma} \begin{definition} \label{def:pullback-pushforward} Let $\kappa \colon B' \to B$ be a covering space. We define the \notation{pullback} $\kappa^*$ and the \notation{pushforward} $\kappa_*$ of functions as follows \begin{align*} \kappa^* \colon \cont(B; \mathbb{R}) &\to \cont(B'; \mathbb{R}), &\kappa_* \colon \cont(B'; \mathbb{R}) &\to \cont(B; \mathbb{R}); \\ (\kappa^*u)(p') &= u(\kappa(p')), &(\kappa_*u)(p) &= \sum_{p' \in \kappa^{-1}(p)} u(p'). \end{align*} \end{definition} If $B$ is connected, then for any $u \in \cont(B; \mathbb{R})$, \begin{equation} \label{eq:covering-space-equation} (\kappa_* \kappa^* u)(p) = \sum_{p' \in \kappa^{-1}(p)} u(\kappa(p')) = \#\kappa \cdot u(p). \end{equation} \begin{lemma} \label{lem:approximation-lemma} Let $\kappa \colon B' \to B$ be a covering space with $B'$ compact (so $B$ is compact). Let $a \in \cont(B; \mathbb{R})$, $u \in \smth(B; \mathbb{R})$ such that $u < a$. Then for any $a' \in \cont(B'; \mathbb{R})$ with $\kappa_* a' = a$, there is $u' \in \smth(B'; \mathbb{R})$ such that $u' < a'$ and $\kappa_* u' = u$. \end{lemma} \begin{proof} Without loss of generality we assume $B$ is connected and $u = 0$ otherwise we can deal with each connected component of $B$ one by one and replace $a'$ by $a' - u/\#\kappa$, $u'$ by $u' - u/\#\kappa$. Choose $\varepsilon > 0$ such that $\#\kappa \cdot \varepsilon < \min a$. Define $h' = a' - \varepsilon$, then $\kappa_* h' = a - \#\kappa \cdot \varepsilon > 0$. So $\kappa_* (h')^+ > \kappa_* (h')^- \geq 0$. Since $h'$ is bounded from below we set $c = \max \kappa_* (h')^- > 0$. Define \begin{equation*} w' = \frac{(h')^+}{\kappa^* \kappa_* (h')^+} \kappa^* \kappa_* (h')^- -(h')^-, \end{equation*} then $\kappa_*(w') = 0$. Moreover, \begin{equation*} h' - w' = (h')^+ - \frac{\kappa^* \kappa_* (h')^-}{\kappa^* \kappa_*(h')^+} (h')^+ \geq 0. \end{equation*} By Whitney Approximation Theorem, \cite[Theorem 10.16]{MR2954043}, there is $v' \in \smth(B'; \mathbb{R})$ such that $\abs{v' - w'} < \varepsilon / 2$. Then let $u' = v' - \frac{1}{\# \kappa} \kappa^* \kappa_*(v') \in \smth(B'; \mathbb{R})$. So $\abs{u' - w'} < \varepsilon$, and $\kappa_* u' = 0$ by \eqref{eq:covering-space-equation}, hence $a' - u' > h' - w' \geq 0$ is as required. \end{proof} \section{Introduction} \label{sec:introduction} The main result of this paper concerns fiber volume forms. A \emph{fiber volume form} is a family of volume forms, one for each fiber of a fiber bundle. The bundle here is sliced by an exhaustion function compatible with the bundle structure. \begin{theorem}[See Theorem~\ref{thm:main-theorem} for the complete statement] \label{thm:main-theorem-intro} Two fiber volume forms can be intertwined by a fiber-preserving diffeomorphism if each fiber has the same volume (finite or infinite) with respect to both forms and, the variations of both volumes of each end along fibers are continuous, and their difference is smooth. \end{theorem} Indeed, it is a well known theorem due to Moser~\cite{MR0182927} that if two volume forms $\omega$ and $\tau$ on a compact manifold without boundary satisfy $\int_M \omega = \int_M \tau$ then one can find a diffeomorphism $\varphi \in \mathrm{C}^{\infty}(M, M)$ such that $\varphi^* \omega = \tau$. Later Greene and Shiohama~\cite{MR542888} realized that Moser's theorem also holds even if $M$ is not compact provided that the forms $\omega$ and $\tau$ are \notation{commensurable} in the sense of giving the same (finite or not) type of volume for each \notation{end} $\epsilon$ of $M$. By an end $\epsilon$ of $M$ having \notation{finite volume} with respect to a volume form $\omega$ we mean that there exists a neighborhood of $\epsilon$ in $M$ which has finite volume with respect to $\omega$, otherwise we say that $\epsilon$ has \notation{infinite volume} with respect to $\omega$. The proof in~\cite{MR542888} is more complicated than Moser's proof because the authors have to deal with the behavior at infinity of the forms (this is made rigorous with the notion of ends). Their proof has three stages: first they extend Moser's proof to forms which are compactly supported. Then they chop their noncompact manifold into pieces, and finally a careful analysis of the behavior at the boundaries and interiors, allows them to construct a global diffeomorphism by pasting together the local diffeomorphisms, in effect bypassing any analytic estimates. The Moser and Greene--Shiohama results have applications in geometry and analysis, and also in classical mechanics, where understanding the geometry of volume forms is relevant~\cite{MR2827114, MR2551999}. \inputfigure{torus-bundle}{torusbundle}{ Surface bundle over a torus. $B$ is a $2$-dimensional torus, $F$ is a noncompact surface. In this case, $\pi \colon M \to B$ is also a symplectic fiber bundle. } If one mimics the Greene--Shiohama argument in the case of two \notation{smooth families} of volume forms $\omega_p, \tau_p$, indexed by some compact manifold without boundary which plays the role of parameter space $B$, this produces for each $p$ a diffeomorphism $\varphi_p$ such that $\varphi_p^* \omega_p = \tau_p$, but there is no information given about how $\varphi_p$ changes when $p$ changes in $B$. This is a separate, independent problem, of an analytic nature. Understanding the smooth dependence of this variation is a {sine qua non} condition to understanding the case of a variation of volume forms on a fiber bundle with nontrivial topology $F \hookrightarrow M \xrightarrow{\pi} B$ with noncompact fiber $F$ without boundary, which is our goal in this paper, see Figure~\ref{fig:torusbundle} (if $F$ and $M$ are compact and without boundary this is close to Moser's original theorem, see~\cite[Theorem~2.4]{MR2551999}). This problem is essentially the same as one faces in Hodge theory, where given a smooth family of exact $q$-forms $\alpha_p$, if $M$ itself is compact, the Hodge theorem produces for us a \notation{smooth} family of $(q-1)$-forms $\beta_p$ such that $d\beta_p=\alpha_p$ (note that for each fixed $p$ this statement is evident). But of course Hodge theory on noncompact manifolds is much more complicated than its compact counterpart. To address the analytic problem in the noncompact case we will resort to recent work by Bueler and Brokhorenkov on Hodge theory and cohomology~\cite{MR1893604}. This smooth family problem is a particular instance of the fiber bundle case, more precisely, it coincides with the case of trivial fiber bundles. If the fiber bundle has topology, we incorporate combinatorial and topological techniques into the analytic problem. The noncompactness of the fibers adds a layer of complexity, one which was already present in~\cite{MR542888} (and the general ideas of~\cite{MR542888} have served as inspiration for the induction argument in the end of the paper). In the fiber bundle case we also have to control the variations from fiber to fiber of the forms. In order to do this, we look at the structure of each fiber from a combinatorial angle, using the setting of slicing and trees which in turn also makes the proofs conceptually transparent, and with this view point we are able to prove our results with a minimum of technical fuss. We in this paper consider fiber bundles meeting the requirements that for each end $\epsilon$ of $M$ we require that the fibers in it must either all have finite volume, or all have infinite volume; nonetheless the finiteness of volume can vary with the end. A key ingredient of the proof is the association of certain trees encoding important features of noncompact manifolds, see Figure~\ref{fig:manifoldtree}. Combinatorial techniques have in the past been applied very successfully to study diffeomorphisms and other geometric problems~\cite{MR934253}. \inputfigure{manifold-tree}{manifoldtree}{ The tree associated to a noncompact surface. This association works in any dimension. The nodes correspond to regions above the hypersurfaces immediately below the nodes. } The paper is organized around the proof of the main result Theorem~\ref{thm:main-theorem}, which is a generalization to fiber bundles with noncompact fibers of Moser's classical stability result on volume forms. In Section~\ref{sec:category}, we give a precise description of the category of the type of fiber bundles in question. The statement of the main theorems for fiber bundles and smooth families are in Section~\ref{sec:main-result}. Because there are well differentiated parts that must be combined to prove the theorem, we have split the paper into three parts: topology, combinatorics, and geometry-analysis; in Sections~\ref{sec:topology}--\ref{sec:geometric-analytic} we develop the tools for the proof of the theorem. The Sections~\ref{sec:filtration-theorem}~and~\ref{sec:proof-main-theorem} contain the proof of the theorem, divided into a longer inductive proof (second to last section) which combines the material from these parts, and a shorter section concluding the proof relying primarily in the geometry-analysis section. With the abstract viewpoint of the paper, the main theorem can be stated in quite simple terms. Finally we have an appendix in Section~\ref{sec:appendix} where we discuss the properties of ends of the bundle and the fiber, with which we have alternative statements of the main theorems. \medskip \emph{Acknowledgments}. We are very grateful to Bruce Driver, Alessio Figalli, Rafe Mazzeo, and Alan Weinstein for helpful discussions, and to Ioan Benjenaru and Lei Ni for providing us with useful references. Also, Xiudi Tang thanks Rafe Mazzeo for the warm hospitality and helpful conversations in his visit to Stanford in December 2015. In addition, we are very thankful to Alan Weinstein for insightful and detailed comments on a preliminary version of this paper. The authors are supported by NSF CAREER grant DMS-1518420. The first author also received support from Severo Ochoa Program at ICMAT in Spain, where part of this work was carried out. \section{Main results} \label{sec:main-result} \subsection{Fiber volume forms} \label{ssec:fiber-volume-form} \begin{definition} \label{def:fiber-form} Let $\lb{M} = (\pi, M, B, F, f, h)$ be an oriented filled bundle, and $q \in \mathbb{N}$ such that $0 \leq q \leq \dim F$. A \notation{fiber $q$-form} on $\lb{M}$ is a family $\Set{\omega_p}_{p \in B}$ such that $\omega_p$ is a $q$-form on $\pi^{-1}(p)$ for each $p \in B$, and there exists $\omega \in \Omega^q(M)$ such that $\omega_p = \iota_p^* \omega$ for each $p \in B$. By a slight abuse of notation we denote $\Set{\omega_p}_{p \in B}$ by $\omega$. Let $\Omega^q_F(\lb{M})$ be the space of fiber $q$-forms on $\lb{M}$. The space of compactly supported fiber $q$-forms on $\lb{M}$ is defined analogously and denoted by $\pcspt{\Omega^q_F}(\lb{M})$. A \notation{fiber top-form} on $\lb{M}$ is a fiber $(\dim F)$-form $\omega$ on $\lb{M}$. A \notation{fiber volume form} on $\lb{M}$ is a fiber top-form $\omega$ on $\lb{M}$ such that $\omega_p$ is a volume form on $\pi^{-1}(p)$ for any $p \in B$. We denote by $\FVform(\lb{M})$ be the space of fiber volume forms on $\lb{M}$. \end{definition} \begin{definition} \label{def:fiber-derivative} Let $\lb{M} = (\pi, M, B, F, f, h)$ be an oriented filled bundle. We define the \notation{fiber exterior derivative} as the map $\fibder \colon \Omega^q_F(\lb{M}) \rightarrow \Omega^{q+1}_F(\lb{M}), \eta \mapsto \fibder \eta$, uniquely defined by $(\fibder \eta)_p = \der \eta_p$ for any $p \in B$. \end{definition} \begin{definition} \label{def:fiber-integral} Let $\lb{M} = (\pi, M, B, F, f, h)$ be an oriented filled bundle. Let $\omega$ be a fiber top-form on $\lb{M}$. If for all $p \in B$ the integral of the top-form $\omega_p$, $\int_{\pi^{-1}(p)} \omega_p$, exists in $[-\infty, +\infty]$, we call the map \begin{align*} \int_{\lb{M}} \omega \colon B &\longrightarrow [-\infty, +\infty] \\ p &\longmapsto \int_{\pi^{-1}(p)} \omega_p \end{align*} the \notation{fiber integral of $\omega$ on $\lb{M}$}. \end{definition} \begin{definition} \label{def:released-fiber-integral} Let $\lb{M} = (\pi, M, B, F, f, h)$ be a connected oriented filled bundle. Let $\omega$ be a fiber top-form on $\lb{M}$. If the fiber integral of $\omega$ on $\Rls \lb{M}$ exists then define \begin{align*} \rint_{\lb{M}} \omega = \int_{\Rls{\lb{M}}} \omega \colon B_M &\longrightarrow [-\infty, +\infty] \end{align*} as the \notation{released fiber integral of $\omega$ on $V$}. \end{definition} \begin{remark} If for each $p \in B$, $\omega_p \geq 0$ with respect to the orientation of $\pi^{-1}(p)$ (in this case we call $\omega$ \notation{nonnegative}), then $\int_{\lb{M}} \omega$ is well defined. \end{remark} \begin{definition} \label{def:fiber-diffeomorphism} Let $\lb{M} = (\pi, M, B, F, f, h)$ be a filled bundle. A diffeomorphism $\varphi \colon M \to M$ is a \notation{fiber diffeomorphism} if $\pi \circ \varphi = \varphi \circ \pi$. Let $\omega \in \Omega^q_F(\lb{M})$ for some $q \in \mathbb{N}$ such that $0 \leq q \leq \dim F$, then define the pullback $\varphi^* \omega = \Set{\Res{\varphi}_{\pi^{-1}(p)}^* \omega_p}_{p \in B} \in \Omega^q_F(\lb{M})$. \end{definition} It is worth noting that a fiber diffeomorphism of $\lb{M}$ is not necessarily a filled morphism. \subsection{Main theorem: Moser stability for fiber bundles} \label{ssec:main-theorem} \begin{definition} \label{def:smoothly-commensurable} Let $\lb{A}$ be a filled subbundle of $\lb{M}$, and let $\omega, \tau \in \FVform(\lb{M})$. We say $\omega$ and $\tau$ are \notation{smoothly commensurable on $\lb{A}$} if the released fiber integral of $\omega, \tau$ on $\lb{A}$ are either both continuous with smooth difference, or are both infinite. We say $\omega$ and $\tau$ are \notation{smoothly commensurable} if they are smoothly commensurable on any superlevel subbundle of $\lb{M}$. \end{definition} \begin{theorem} \label{thm:main-theorem} Let $\lb{M} = (\pi, M, B, F, f, h)$ be an oriented connected exhausted bundle with compact base and noncompact connected fiber without boundary. For any smoothly commensurable fiber volume forms $\omega, \tau$ on $\lb{M}$ with equal fiber integral, there exists a fiber diffeomorphism $\varphi \colon M \to M$ such that $\varphi^* \omega = \tau$. \end{theorem} \begin{remark} If we only require $\lb{M}$ to be connected, but may have disconnected fiber, then the condition $\omega, \tau$ having equal fiber integral on $\lb{M}$ should be replaced by they having same released fiber integral. To prove this we apply Theorem~\ref{thm:main-theorem} on $\Rls \lb{M}$. \end{remark} \begin{remark} The geometry of volume preserving diffeomorphisms is much simpler than that of their symplectic counterparts (see \cite{MR728456} and \cite{MR1826128}). \end{remark} \subsection{Corollary: Moser stability for families} \label{ssec:smooth-family-theorem} The goal of this section is to give a smooth family version of the Moser--Greene--Shiohama result \cite[Theorem~1]{MR542888}. \begin{definition} \label{def:smooth-family} Let $M$ be a noncompact connected manifold and let $B$ be a compact manifold. A family of differential $q$-forms $\Set{\omega_p}_{p \in B} \subset \Omega^q(M)$ is \notation{smooth} if the map $\omega \colon B \times M \rightarrow \wedge^q T^*M, (p, x) \mapsto \omega_p(x)$ is smooth. Smooth families of diffeomorphisms are defined in the same way. \end{definition} \begin{corollary} \label{cor:smooth-family} Let $M$ be an oriented connected manifold without boundary, $\mathcal{NE}(M) = \coprod_{K \in \mathcal{K}(M)} \Connub(M \setminus K)$. Let $B$ be a compact manifold. Let $\Set{\omega_p}_{p \in B}$ and $\Set{\tau_p}_{p \in B}$ be smooth families of volume forms on $M$ such that: \begin{itemize} \item for each $p \in B$, $\int_M \omega_p = \int_M \tau_p \leq +\infty$; \item for any $C \in \mathcal{NE}(M)$, either $C$ has infinite volume with respect to both $\Set{\omega_p}_{p \in B}$ and $\Set{\tau_p}_{p \in B}$, or $\Pa{p \mapsto \int_C \omega_p}$ and $\Pa{p \mapsto \int_C \tau_p}$ are continuous and there difference is smooth. \end{itemize} Then there is a smooth family of diffeomorphisms $\Set{\varphi_p \colon M \to M}_{p \in B}$ such that $\varphi^*_p\omega_p=\tau_p$ for each $p \in B$. \end{corollary} \begin{proof} Let $f \colon M \to \mathbb{R}$ be an exhaustion function for $M$, then $\lb{P} = (\proj_1, B \times M, B, M, f \circ \proj_2, f)$ is an exhausted bundle with compact base and connected fiber, and $\omega = \Set{\omega_p}_{p \in B}$, $\tau = \Set{\tau_p}_{p \in B}$ are fiber $q$-form on $\lb{P}$. If $M$ is compact, the only condition is $\int_M \omega_p = \int_M \tau_p$ for any $p \in B$, this is a result of the standard Moser stability theorem. If $M$ is noncompact, note that any superlevel subbundle of $P$ has underlying space in the form $B \times C$, where $C \in \mathcal{NE}(M)$, then Corollary~\ref{cor:smooth-family} is the result of Theorem~\ref{thm:main-theorem}. \end{proof} \begin{remark} If $B = [0, 1]$, a version of Corollary~\ref{cor:smooth-family} was given for continuous families as \cite[Theorem 1]{MR1034272} for the case of manifolds $M$ which are the interior of a compact manifold with boundary. The work relies on a version of Moser's theorem for compact manifolds with boundary due to Banyaga \cite{MR0358649}. \end{remark} If $B$ is a singleton, Theorem~\ref{thm:main-theorem} or Corollary~\ref{cor:smooth-family} recovers Theorem~\ref{thm:greene-shiohama}. \begin{theorem}[Moser--Greene--Shiohama] \label{thm:greene-shiohama} Let $M$ be a noncompact oriented connected manifold without boundary. If $\omega, \tau$ are fiber volume forms on $M$ that with equal volume, and for each end $\epsilon$ of $M$ either has finite volume or infinite volume with respect to $\omega$ and $\tau$, then there exists a diffeomorphism $\varphi \colon M \to M$ such that $\varphi^* \omega = \tau$. \end{theorem} \section{Category of $\ell$-bundles} \subsection{Objects: $\ell$-bundles} \begin{definition} An $\ell$-bundle is a 7-tuple $\lb{M} = (\pi, M, B, F, \mathcal{G}, h, f)$ such that $M, F$ are manifolds, $B$ is a compact manifold, $\pi \colon M \to B$ a smooth map, $h$ is an exhaustion function for $M$ and $f$ iss such that for any point in $B$ there is a neighborhood $U$ such that $\pi^{-1}(U)$ is diffeomorphic to $U \times F$ via the following commutative diagram: \begin{equation*} \xymatrix{ \pi^{-1}(U) \ar[r]^{\phi_i} \ar[d]^{\pi} & U \times F \ar[dl]^{\proj_1} \\ U & }, \end{equation*} Suppose $(\pi, M, B, F)$ is a fiber bundle with fiber $F$ and let $\mathcal{G}$ be a subgroup of the diffeomorphism group $\Diff(F)$ of $F$. For each $i, j \in \mathcal{I}$ such that $U_i \cap U_j \neq \emptyset$ let the \notation{transition map} $\phi_{ij} \colon U_i \cap U_j \to \Diff(F)$ be uniquely defined by the commutative diagram for $p \in U_i \cap U_j$, \begin{equation*} \xymatrixcolsep{3pc} \xymatrix{ \Set{p} \times F \ar[d]^{\proj_2} \ar[r]^{\phi_i \circ \phi_j^{-1}} & \Set{p} \times F \ar[d]^{\proj_2} \\ F \ar[r]^{\phi_{ij}(p)} & F }. \end{equation*} The fiber bundle $(\pi, M, B, F)$ is called a \notation{fiber $\mathcal{G}$-bundle} if the image of $\phi_{ij}$ is in $\mathcal{G}$ for each $i, j \in \mathcal{I}$ such that $U_i \cap U_j \neq \emptyset$. We call $\mathcal{G}$ a \notation{structure group} of $(\pi, M, B, F)$. There exists a unique map $\fatlarge \colon M \to \mathbb{R}$ such that for any $i \in \mathcal{I}$ the following diagram commutes \begin{equation*} \xymatrixcolsep{5pc} \xymatrix{ \pi^{-1}(U) \ar[d]^{\phi_i} \ar[r]^-{\Res{\fatlarge}_{\pi^{-1}(U)}} & \mathbb{R} \\ U \times F \ar[r]^-{\proj_2} & F \ar[u]^{f} }. \end{equation*} Moreover $\fatlarge$ is an exhaustion function for $M$. For any $A \in \Inv(\mathcal{G})$, $A \in \mathcal{K}(F)$ if and only if $\El_\pi(A) \in \mathcal{K}(M)$. The restriction $\Res{\pi}_{\El_\pi(A)} \colon \El_\pi(A) \to B$ is a fiber subbundle of $\pi \colon M \to B$ with fiber $A$ if $A$ is a submanifold of $F$ with or without boundary. \end{definition} \begin{lemma} Let $(\pi, M, B, F, \mathcal{G}, h, f)$ be an $\ell$-bundle. Any local trivialization maps on $U$ differ by diffeos in $\mathcal{G}$. \end{lemma} \begin{proof} \begin{equation*} \xymatrixcolsep{5pc} \xymatrix{ \pi^{-1}(U) \ar[d]^{\phi_U} \ar[r]^-{\Res{\fatlarge}_{\pi^{-1}(U)}} & \mathbb{R} & \pi^{-1}(U) \ar[d]^{\tilde \phi_U} \ar[l]_-{\Res{\fatlarge}_{\pi^{-1}(U)}} \\ U \times F \ar[r]^-{\proj_2} & F \ar[u]^{f} & U \times F \ar[l]_-{\proj_2} }. \end{equation*} \end{proof} \begin{lemma} Given $(\pi, M, B, F, \mathcal{G}, f)$, there is always an $h$ such that $(\pi, M, B, F, \mathcal{G}, h, f)$ is an $\ell$-bundle. \end{lemma} \begin{proof} \begin{equation*} \xymatrixcolsep{5pc} \xymatrix{ \pi^{-1}(U) \ar[d]^{\phi_i} \ar[r]^-{\Res{\fatlarge}_{\pi^{-1}(U)}} & \mathbb{R} \\ U \times F \ar[r]^-{\proj_2} & F \ar[u]^{f} }. \end{equation*} Let $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ be the local trivialization of $\pi$ and $I$ is finite. We take into account that $\pi$ is a fillable fiber $\mathcal{G}$-bundle with respect to $f$. The definition of $\fatlarge$ is given by the commutative diagram, and $\fatlarge$ is well defined since $f \circ \proj_2 \circ \Res{\phi_i^{-1}}_{\pi^{-1}(U_i \cap U_j)} = f \circ \proj_2 \circ \Res{\phi_j^{-1}}_{\pi^{-1}(U_i \cap U_j)}$ for each $i, j \in \mathcal{I}$ with $U_i \cap U_j \neq \emptyset$. Let $A \in \Inv(\mathcal{G})$. If $\El_\pi(A) \in \mathcal{K}(M)$, let $\alpha = \max_{\El_\pi(A)} \fatlarge = \max_A f \in \mathbb{R}$. Then $\El_\pi(A) \cap \pi^{-1}(p) = \El_\pi(A) \cap \phi_i^{-1}\Pa{\Set{p} \times f^{-1}((-\infty, \alpha])} \in \mathcal{K}(\pi^{-1}(p))$ for any $i \in \mathcal{I}$ and $p \in U_i$, so $A = \phi_i\Pa{\El_\pi(A) \cap \pi^{-1}(p)} \in \mathcal{K}(F)$. If $A \in \mathcal{K}(F)$, let $\Set{V_j}_{j \in J}$ be a finite refinement of $\Set{U_i}_{i \in \mathcal{I}}$, namely a finite open cover of $B$ such that for any $j \in J$ there is an $i(j) \in I$ such that $V_j \ssubset U_{i(j)}$. Then we have $\El_\pi(A) = \bigcup_{j \in J} \Pa{\El_\pi(A) \cap \pi^{-1}(\overline{V}_j)} = \bigcup_{j \in J} \phi_{i(j)} \Pa{\overline{V}_j \times A} \in \mathcal{K}(M)$. The fact that $\Res{\pi}_{\El_\pi(A)}$ is a subbundle of $\pi$ is immediate by the definition of $\El_\pi$. Since $f$ is an exhaustion function for $F$, let $\alpha \in \Regular(f)$, then $\fatlarge^{-1}((-\infty, \alpha]) = \El_\pi\Pa{f^{-1}((-\infty, \alpha])}$ is compact. Hence $\fatlarge$ is an exhaustion function for $M$. \end{proof} \subsection{Morphisms: $\ell$-maps} If $E \subset M$ is a submanifold, $\Res{\pi}_{E} \colon E \to B$ is called a \notation{fiber subbundle} of $\pi$ if there is a local trivialization $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ of $\pi$ such that there exists $A \subset F$ a submanifold and for any $i \in \mathcal{I}$, $U_i \times A$ is the image of $\phi_i$ on $\pi^{-1}(U_i) \cap E$. Then $\Pa{\Res{\pi}_E, E, B, A}$ is a fiber bundle. \subsection{The Release functor of $\ell$-bundles} We define a endofunctor $\Rls \colon \lBundle \to \lBundle, \lb{A} \mapsto \Rls\lb{A}$ as follows. Let \begin{equation} \xymatrix{ & \pi_A^{-1}(U) \ar[dl]_{\Rls(\pi_A)} \ar[d]^{\pi_A} \\ U \times \Conn{\pi_A^{-1}(U)} \ar[r]_{p_U} & U }, \end{equation} which is shown in Figure~\ref{fig:coveringspace}. We want to show that $p$ is a covering map and $Rls(\pi_A)$ has connected fibers. Suppose $i \colon C \to A$ is an inclusion then there is a map from $B_C \to B_A$ by the natural $\Conn{\pi_C^{-1}(U)} \to \Conn{\pi_A^{-1}(U)}$. To show Rls is a functor, \begin{equation} \xymatrix{ & \pi^{-1}(U) \ar[dl]_{\Rls(\pi)} \ar[d]^{\pi} & \pi^{-1}(U') \ar[l]^{\varphi} \ar[dr]^{\Rls(\pi')} \ar[d]_{\pi'} & \\ U \times \Conn{\pi^{-1}(U)} \ar[r]_{p_U} & U & U' \ar[l]^{\varphi} & U' \times \Conn{\pi^{-1}(U')} \ar[l]^{p_{U'}} }, \end{equation} \subsection{Proof of filtration theorem} We proceed by induction on $m \in \mathbb{N} \cup \Set{0}$. At this moment, we assume it is the case that all volumes are finite. \emph{Case $m = 0$.} Let's start from $\lb{M}$ and let $\omega_0 = \omega$, $\tau_0 = \tau$. We have $$\int_{\lb{M}} \omega_0 = \int_{\lb{M}} \tau_0.$$ \medskip \emph{Case $(0 \leq n \leq m-1) \implies (n = m)$ for $m \in \mathbb{N}$.} Assume $$\int_{\lb{A}} \omega_{m-1} = \int_{\lb{A}} \tau_{m-1}.$$ Then consider the statement for $n = m$: Let $\lb{A} \in \tlevel(2m-3)$. Let $C \in \tsuc(\lb{A})$ be a connected sub-$\ell$-bundle. Consider $E \in \tsuc(C)$. First we choose $u_{\lb{C}} \colon B_C \to \mathbb{R}$ such that $$0 < u_{\lb{C}} < \min\Set{\int_{\lb{C}} \omega_{m-1}, \int_{\lb{C}} \tau_{m-1}}.$$ Then we pick $v_{\lb{E}} = \frac{1}{\#C}p_{EC}^*u_{\lb{C}} \colon B_E \to \mathbb{R}$ where $\#E$ is the number of sheets of $p_{EC} \colon B_E \to B_C$. The following function defined on $B_E \times \mathbb{R}$, \begin{equation} \label{eq:lambdaps} (p, \beta) \mapsto v_{\lb{E}}(p) - \max\Set{\Pa{\int_{\lb{E}_{[\beta, +\infty)}} \omega_{m-1}}(p), \Pa{\int_{\lb{E}_{[\beta, +\infty)}} \tau_{m-1}}(p)}, \end{equation} is continuous in $p$, is increasing in $\beta$, and converges to $v_{\lb{E}} > 0$ as $\beta \to +\infty$ pointwise for $p \in B_E$. Since $B_E$ is compact there is $\beta_{\lb{E}} \geq \alpha_{\lb{E}}$ such that \eqref{eq:lambdaps} is positive when $\beta = \beta_{\lb{E}}$. Define $\Tse_\mathcal{T} \lb{E} = \lb{E}_{[\alpha_{\lb{E}}, \beta_{\lb{E}}]}$ we have $\int_{\lb{E}_{[\beta_{\lb{E}}, +\infty)}} \omega_{m-1} < v_{\lb{E}}$, and $\int_{\lb{E}_{[\beta_{\lb{E}}, +\infty)}} \tau_{m-1} < v_{\lb{E}}$. Now we require $\omega_m, \tau_m$ by \begin{align*} \int_{\Tse_\mathcal{T} \lb{C}} \omega_m &= \int_{\lb{C}} \omega_{m-1} - u_{\lb{C}}, &\int_{\Tse_\mathcal{T} \lb{C}} \tau_m &= \int_{\lb{C}} \tau_{m-1} - u_{\lb{C}}, \\ \int_{\Tse_\mathcal{T} \lb{E}} \omega_m &= v_{\lb{E}} - \int_{\lb{E}_{[\beta_{\lb{E}}, +\infty)}} \omega_{m-1}, &\int_{\Tse_\mathcal{T} \lb{E}} \tau_m &= v_{\lb{E}} - \int_{\lb{E}_{[\beta_{\lb{E}}, +\infty)}} \tau_{m-1}. \end{align*} and \begin{align*} \support(\omega_m - \omega_{m-1}) &\subset \Shta_\mathcal{T} \lb{C}, &\support(\tau_m - \tau_{m-1}) &\subset \Shta_\mathcal{T} \lb{C}. \end{align*} We define the following functors \begin{align*} p_{EC}^* \colon C(B_C) &\to C(B_E), &(p_{EC}^*u)(x) &= u(p_{EC}x); \\ s_E \colon C(B_E) &\to C(B_C), &(s_Eu)(x) &= \sum_{y \in \pi_E^{-1}(x)} u(y). \end{align*} Then $$(s_E p_{EC}^*u)(x) = \sum_{y \in \pi_E^{-1}(x)} u(p_{EC}(y)) = \#E\,u(x).$$ Then we have $$\sum_{\lb{E} \in \tsuc(\lb{C})} s_E(v_{\lb{E}}) = \sum_{\lb{E} \in \tsuc(\lb{C})} \frac{\#E}{\#C} u_{\lb{C}} = u_{\lb{C}}.$$ $$\int_{\Tse_\mathcal{T} \lb{C}} \omega_m = \int_{\lb{C}} \omega_{m-1} - u_{\lb{C}} = \int_{\lb{C}} \omega_{m-1} - \sum_{\lb{E} \in \tsuc(\lb{C})} s_E(v_{\lb{E}}) =\int_{\Tse_\mathcal{T} \lb{C}} \omega_{m-1}.$$ Similarly, $$\int_{\Tse_\mathcal{T} \lb{C}} \tau_m = \int_{\Tse_\mathcal{T} \lb{C}} \tau_{m-1}.$$ Since by induction $\int_{\lb{A}} \omega_{m-1} = \int_{\lb{A}} \tau_{m-1}$, and by the above construction $\int_{\lb{A}} \omega_m = \int_{\lb{A}} \omega_{m-1}$, $\int_{\lb{A}} \tau_m = \int_{\lb{A}} \tau_{m-1}$, and that $\int_{\lb{E}} \omega_m = \int_{\lb{E}}\tau_m$, we have $\int_{\Shta_\mathcal{T} \lb{A}} \omega_m = \int_{\Shta_\mathcal{T} \lb{A}} \tau_m$. \begin{comment} Let $\tilde B_C = \coprod_{E \in \tsuc(C)} B_E$. \rho_E \colon C(\tilde B_C) &\to C(B_E), &(\rho_Eu)(x) &= u(x); \\ e_E \colon C(B_E) &\to C(\tilde B_C), &(e_Eu)(x) &= \chi_{B_E} u(x); \\ Let $z_E = \fibint{\omega_{m-1}}{E} - \fibint{\tau_{m-1}}{E} \colon B_E \to \mathbb{R}$. We fix $\int_{\Tse_\mathcal{T} \lb{C}} \omega_m$ and $\int_{\Tse_\mathcal{T} \lb{C}} \tau_m$, then $\int^{\tilde B_C}_{\Tse(C)} \omega_m = p_C^* \int_{\Tse_\mathcal{T} \lb{C}} \omega_m$. This gives $\sum_E e_E(u_E) = \int^{\tilde B_C}_{\Tse(C)} \omega_m - \int^{\tilde B_C}_{\Tse(C)} \omega_{m-1}$. If we can find $u_E \colon B_E \to \mathbb{R}$ such that $u_E, u_E - z_E > 0$ and $u_E < \int_{\lb{E}} \omega_{m-1}$. This $\sum_E e_E(z_E^+) < p_C^*\int_{\Tse_\mathcal{T} \lb{C}} (\omega_m - \omega_{m-1}) < \sum_E e_E\int_{\lb{E}} \omega_{m-1}$. \end{comment} \begin{comment} First we choose $w_C \colon B_C \to \mathbb{R}$ such that $$\max\Set{0, \int_{\lb{C}} \omega_{m-1} - \int_{\lb{C}} \tau_{m-1}} < w_C < \int_{\lb{C}} \omega_{m-1},$$ then let $w'_C = w_C - \Pa{\int_{\lb{C}} \omega_{m-1} - \int_{\lb{C}} \tau_{m-1}}$, we have $0 < w'_C < \int_{\lb{C}} \tau_{m-1}$. Then we pick $v_{\lb{E}} \colon B_E \to \mathbb{R}$ such that $v_{\lb{E}} > 0$ and $\sum_E e_E(v_{\lb{E}}) = \int_{\lb{C}} \omega_{m-1} - w_C$. We define $v_{\lb{E}} = \frac{\#E}{\#C}\Res{p_C^*\Pa{\int_{\lb{C}} \omega_{m-1} - w_C}}_{B_E}$. Then we let $\fibint{\omega_m}{\Tse(C)} = w_C$, $\fibint{\tau_m}{\Tse(C)} = w'_C$, $\fibint{\omega_m}{E} = \fibint{\tau_m}{E} = v_{\lb{E}}$. We can find $\Tse(E)$ by requiring $\int_{\lb{E}_{[\beta_{\lb{E}}, +\infty)}} \omega_m < v_{\lb{E}}$. \end{comment} \subsection{Commensurability lemma} \label{ssec:commensurabilitylemma} \begin{lemma} Let $\kappa \colon B' \to B$ be a covering space with $B'$ compact (so $B$ is compact). Let $S' \subset B'$ be open, and $S = B \setminus \kappa(B' \setminus S')$. Then for any $u \in \smth(B; \mathbb{R})$ vanishing in $S$ there is $u' \in \smth(B'; \mathbb{R})$ vanishing in $S'$ such that $s(u') = u$. \end{lemma} \begin{proof} For $0 \leq l \leq \#\kappa$, let \begin{equation*} S_l = \Setby{p \in B}{\# \Pa{\kappa^{-1}(p) \setminus S'} \leq l}. \end{equation*} Then \begin{equation*} S = S_0 \subset S_1 \subset \dotsb \subset S_{\#\kappa} = B. \end{equation*} is a filtration of $B$ by open sets. Let $S'_l = \kappa^{-1}(S_l)$ for $0 \leq l \leq \#\kappa$, then $\kappa(S') = S_{\#\kappa-1}$. Define $\Res{u'}_{S'} = 0$. For any $x \in S'_1 \setminus S'$, define $u'(x) = u(\kappa(x))$. Note that $$ \Pa{S'_1 \setminus S'} \cup S'_0 = \Setby{x \in B'}{\kappa^{-1}(\kappa(x)) \setminus \Set{x} \subset S'}$$ is open, and $u(\kappa(x)) = 0$ when $x \in S'_0$. Then $\Res{u'}_{S'_1 \cup S'}$ is defined smoothly. For any $x \in S'_2 \setminus S'$, We can define $u'$ on $\kappa^{-1}(S_1)$ in a smooth manner. If we extend $u'$ to $w'$ onto $\kappa^{-1}(S_2)$ arbitrarily, then $s(w') - u$ is smooth in $S_2$ and vanishes outside $S_2 \setminus S_1$. Let $\Res{u'}_{S_2} = w' - \frac12(\kappa^*(s(w') - u))$, then $\Res{u'}_{S_2}$ is an extension of $\Res{u'}_{S_1}$ and $s\Res{u'}_{S_2} = u$ We define $u'(x) = 0$ if $x \in S'$ and $u'(x) = u(\kappa(x))$ if $x \in S_1$. \end{proof} \begin{lemma} Let $\kappa \colon B' \to B$ be a covering space with $B'$ compact (so $B$ is compact). Let $a, b \in \cont(B; \peR)$, $u \in \smth(B; \mathbb{R})$ such that $a = b + u$. Then if there are $a', b' \in \cont(B'; \peR)$ with $s(a') = a$, $s(b') = b$, and $a', b'$ are $\smth$-equivalent, then there is $u' \in \smth(B'; \mathbb{R})$ such that $a' = b' + u'$ and $s(u') = u$. \end{lemma} \begin{proof} Without loss of generality we can assume $a' = b'$. Then $a = b$ and we are left to prove: Let $S' = \Set{a' < \infty} \subset B'$ and $S = \Set{a < \infty} \subset B$ then we have $s(B' \setminus S') = B \setminus S$. Then we can define $\Res{u'}_{S'} = \Res{a'}_{S'} - \Res{b'}_{S'}$. On $\Set{\#(\infty) \leq 1}$, which is open, the choice of $u_i$ is unique and smooth. Consider each connected component of $\Set{\#(\infty) \leq 2}$. For those points on a component which have a double cover (open set) \end{proof} \begin{lemma} \label{lem:ApproximationLemma} Let $\kappa \colon B' \to B$ be a covering space with $B'$ compact (so $B$ is compact). Let $b \in \cont(B; \peR)$, $u \in \smth(B; \mathbb{R})$ such that $b + u > 0$. Then for any $b' \in \cont(B'; \peR)$ with $s(b') = b$, there is $u' \in \smth(B'; \mathbb{R})$ such that $b' + u' > 0$ and $s(u') = u$. \end{lemma} \begin{proof} Without loss of generality we assume $s(u') = 0$ otherwise we can replace $b'$ by $b' + u/\#\kappa$ and $u'$ by $u' - u/\#\kappa$, so then we have $u = 0$. Choose $\varepsilon > 0$ such that $\#\kappa \cdot \varepsilon < \min a$. Define $h' = b' - \varepsilon$, then $s(h') = a - \#\kappa \cdot \varepsilon > 0$. Since $h'$ is bounded from below we set $c = \max s((h')^-) > 0$. Define \begin{equation*} w' = (h')^- - \frac{\min(c, (h')^+)}{\kappa^* s\min(c, (h')^+)} \kappa^* s((h')^-) \in \cont(B'; \mathbb{R}), \end{equation*} then $s(w') = 0$. Moreover, \begin{equation*} h' + w' = (h')^+ - \frac{\kappa^* s((h')^-)}{\kappa^* s\min(c, (h')^+)} \min(c, (h')^+) \geq (h')^+ - \min(c, (h')^+) \geq 0, \end{equation*} since $\kappa^* s((h')^-) \leq \min\Pa{c, \kappa^* s((h')^+)} \leq \kappa^* s\min(c, (h')^+)$. By Whitney Approximation Theorem, \cite[Theorem 10.16]{MR2954043}, there is $v' \in \smth(B'; \mathbb{R})$ such that $\abs{v' - w'} < \varepsilon / 2$. Then let $u' = v' - \frac{1}{\# \kappa} \kappa^* s(v') \in \smth(B'; \mathbb{R})$. So $\abs{u' - w'} < \varepsilon$, and $u = s(u') = 0$, hence $b' + u' > h' + w' \geq 0$ is as required. \end{proof} \subsection{Inductive proof of filtration theorem} \label{sec:prooffiltrationtheorem} We proceed by induction on $m \in \mathbb{N} \cup \Set{0}$. At this moment, we assume it is the case that all volumes are finite. \emph{Case $m = 0$.} Let's start from $\lb{M}$ and let $\omega_0 = \omega$, $\tau_0 = \tau$. We have $$\int_{\lb{M}} \omega_0 = \int_{\lb{M}} \tau_0.$$ \medskip \emph{Case $(0 \leq n \leq m-1) \implies (n = m)$ for $m \in \mathbb{N}$.} Assume $$\int_{\lb{A}} \omega_{m-1} = \int_{\lb{A}} \tau_{m-1}.$$ Then consider the statement for $n = m$: Let $\lb{A} \in \tlevel(2m-3)$. Let $C \in \tsuc(\lb{A})$ be a connected sub-$\ell$-bundle. Consider $E \in \tsuc(C)$. Since $\omega, \tau \in \FVform(M)$ are [[smoothly commensurable]], there is $\delta_C \in \smth$ such that $\rint_{\lb{C}} \omega_{m-1} = \rint_{\lb{C}} \tau_{m-1} + \delta_C$. We choose and $a_{\lb{C}} \in \smth(B_C; \mathbb{R})$ such that $$ \max\Set{0, \delta} < a_{\lb{C}} < \int_{\lb{C}} \omega_{m-1},$$ then let $u_{\lb{C}} = \int_{\lb{C}} \omega_{m-1} - a_C \in \cont(B_C; \peR)$. We have $u_{\lb{C}} < \int_{\lb{C}} \omega_{m-1}$ and $u_C < \int_{\lb{C}} \tau_{m-1}$. [[we mean where the larger one is infinity the smaller one can also be infinite.]] Let $\alpha = \max_E\Set{\alpha_{\mathrm{lofty}, E}}$. Since $u_{\lb{C}} = \int_{\lb{C}_{(-\infty, \alpha]}} \omega_{m-1} + \int_{\lb{C}_{[\alpha, +\infty)}} \omega_{m-1} - a_C > 0$, $a_C \in \smth(B_C; \mathbb{R})$, by [[Commensuablitiy lemma]] to the covering space $\nu \colon \coprod B_E \to B_C$, there are $b_E \in \smth(B_E; \mathbb{R})$ such that $\sum_E s_E(b_E) = \int_{\lb{C}_{(-\infty, \alpha]}} \omega_{m-1} - a_C$, and let $v_E = \int_{\lb{E}_{[\alpha, +\infty)}} \omega_{m-1} + b_E > 0$. Let $\delta_E \in \smth$ such that $\rint_{\lb{E}} \omega_{m-1} = \rint_{\lb{E}} \tau_{m-1} + \delta_E$. The following function defined on $B_E \times \mathbb{R}$, \begin{equation} \label{eq:lambdaps} (\cdot, \beta) \mapsto b_E - \int_{(\Rls \lb{E})_{(-\infty, \alpha]}} \omega_{m-1} + \min\Set{\int_{(\Rls \lb{E})_{(-\infty, \beta]}} \omega_{m-1}, \delta_E + \int_{(\Rls \lb{E})_{(-\infty, \beta]}} \tau_{m-1}} \end{equation} is continuous in the first variable, is increasing in $\beta$, and converges to $v_{\lb{E}} > 0$ as $\beta \to +\infty$ pointwise. Since $B_E$ is compact there is $\beta_{\lb{E}} \geq \alpha_{\lb{E}}$ such that \eqref{eq:lambdaps} is positive when $\beta = \beta_{\lb{E}}$. Define $\Tse_\mathcal{T} \lb{E} = \lb{E}_{[\alpha_{\lb{E}}, \beta_{\lb{E}}]}$ we have $\int_{\lb{E}_{[\beta_{\lb{E}}, +\infty)}} \omega_{m-1} < v_{\lb{E}}$, and $\int_{\lb{E}_{[\beta_{\lb{E}}, +\infty)}} \tau_{m-1} < v_{\lb{E}}$. Now by [[Volume lemma]], there are $\omega_m, \tau_m$ such that, \begin{align*} \int_{\Tse_\mathcal{T} \lb{C}} \omega_m &= a_{\lb{C}}, \qquad \int_{\Tse_\mathcal{T} \lb{C}} \tau_m = a_{\lb{C}} - \delta, \\ \int_{\Tse_\mathcal{T} \lb{E}} \omega_m &= b_E + \int_{(\Rls \lb{E})_{[\alpha, \beta_{\lb{E}}]}} \omega_{m-1}, \\ \int_{\Tse_\mathcal{T} \lb{E}} \tau_m &= b_E + \delta_E + \int_{(\Rls \lb{E})_{(- \infty, \beta_{\lb{E}}]}} \tau_{m-1} - \int_{(\Rls \lb{E})_{(-\infty, \alpha]}} \omega_{m-1}. \end{align*} and \begin{align*} \support(\omega_m - \omega_{m-1}) &\subset \Shta_\mathcal{T} \lb{C}, &\support(\tau_m - \tau_{m-1}) &\subset \Shta_\mathcal{T} \lb{C}. \end{align*} We define the following functors \begin{align*} p_{EC}^* \colon C(B_C) &\to C(B_E), &(p_{EC}^*u)(x) &= u(p_{EC}x); \\ s_E \colon C(B_E) &\to C(B_C), &(s_Eu)(x) &= \sum_{y \in \pi_E^{-1}(x)} u(y). \end{align*} Then $$(s_E p_{EC}^*u)(x) = \sum_{y \in \pi_E^{-1}(x)} u(p_{EC}(y)) = \#E\,u(x).$$ Then we have $$\sum_{\lb{E} \in \tsuc(\lb{C})} s_E(v_{\lb{E}}) = \sum_{\lb{E} \in \tsuc(\lb{C})} \frac{\#E}{\#C} u_{\lb{C}} = u_{\lb{C}}.$$ $$\int_{\Shta_\mathcal{T} \lb{C}} \omega_m = a_{\lb{C}} + \sum_{\lb{E} \in \tsuc(\lb{C})} s_E \Pa{b_E + \int_{(\Rls \lb{E})_{[\alpha, \beta_{\lb{E}}]}} \omega_{m-1}} =\int_{\Shta_\mathcal{T} \lb{C}} \omega_{m-1}.$$ Similarly, $$\int_{\Shta_\mathcal{T} \lb{C}} \tau_m = \int_{\Shta_\mathcal{T} \lb{C}} \tau_{m-1}.$$ Since by induction $\int_{\lb{A}} \omega_{m-1} = \int_{\lb{A}} \tau_{m-1}$, and by the above construction $\int_{\lb{A}} \omega_m = \int_{\lb{A}} \omega_{m-1}$, $\int_{\lb{A}} \tau_m = \int_{\lb{A}} \tau_{m-1}$, and that $\int_{\lb{E}} \omega_m = \int_{\lb{E}}\tau_m$, we have $\int_{\Shta_\mathcal{T} \lb{A}} \omega_m = \int_{\Shta_\mathcal{T} \lb{A}} \tau_m$. \begin{comment} \subsection{Archive} Let $(\pi, M, B, F)$ be a fiber $\mathcal{G}$-bundle. For any $A \in \mathcal{B}(F)$ let $\mathcal{G}(A) = \bigcup_{g \in \mathcal{G}} g(A)$. Denote the collection of invariant Borel subsets under $\mathcal{G}$ by \begin{equation*} \Inv(\mathcal{G}) = \Setby{A \in \mathcal{B}(F)}{\mathcal{G}(A) = A} = \Setby{\mathcal{G}(A)}{A \in \mathcal{B}(F)}. \end{equation*} Let $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ be a local trivialization of $\pi$. To each $A \in \Inv(\mathcal{G})$ we can assign a subspace $\El_\pi(A) \in \mathcal{B}(M)$ of $M$ defined by $\phi_i\Pa{\El_\pi(A) \cap \pi^{-1}(U_i)} = U_i \times A$ for any $i \in \mathcal{I}$. \begin{proposition} \label{prop:subbundle} The map $\El_\pi \colon \Inv(\mathcal{G}) \to \mathcal{B}(M)$ is well defined and injective. \end{proposition} \begin{proof} Let $(U_i, \phi_i)$ and $(U_j, \phi_j)$ be two trivialized chart such that $U_i \cap U_j \neq \emptyset$, then $\El_\pi(A) \cap \pi^{-1}(U_i \cap U_j) = \phi_i^{-1} \Pa{(U_i \cap U_j) \times A} = \phi_j^{-1} \Pa{(U_i \cap U_j) \times A}$ for any $A \in \Inv(\mathcal{G})$. So $\El_\pi(A)$ is well defined. The injectivity of $\El_\pi$ is immediate by definition. \end{proof} \begin{definition} \label{def:diffeoactonend} A diffeomorphism $\phi \colon F \to F$ descends to a permutation $\mathcal{E}(\phi) \colon \mathcal{E}(F) \to \mathcal{E}(F), \epsilon \mapsto \Pa{ \mathcal{E}(\phi)(\epsilon) \colon K \mapsto (\phi \circ \epsilon \circ \phi^{-1})(K)}$. Define the subgroups of $\Diff(f)$ given by $\Diff(F, f) = \Setby{\phi \in \Diff(F)}{\phi^*f = f}$, $\Diff_{\mathcal{E}}(F) = \Setby{\phi \in \Diff(F)}{\mathcal{E}(\phi) = \identity_{\mathcal{E}(F)}}$, and \begin{align*} \Diff_{\mathcal{E}}(F, f) &= \Diff(F, f) \cap \Diff_{\mathcal{E}}(F). \end{align*} \end{definition} \begin{lemma} \label{lem:treefiltration} Let $X$ be a connected locally connected locally compact Hausdorff space and let $\Set{K_l}_{l \in \mathbb{N} \cup \Set{0}}$ be an exhaustive filtration of $X$ by compact sets with $K_0 = \emptyset$. Define \begin{equation*} \mathcal{T} = \coprod_{l \in \mathbb{N} \cup \Set{0}} \Connub{X \setminus K_l}. \end{equation*} Then $\Pa{\mathcal{T}, {\supsetneq}}$ is a rooted locally finite leafless tree of height $\upomega$. \end{lemma} \begin{proof} Let $C_i \in \Conn{X \setminus K_{l_i}} \subset \mathcal{T}$ where $l_i \in \mathbb{N} \cup \Set{0}$ and $K_i \in \mathcal{A}$, for $i = 1, 2$ and $3$. By definition of connected components we have the following: if $C_1 \supsetneq C_2$, then $l_1 < l_2$; if $C_1, C_2 \supsetneq C_3$ and $l_1 < l_2$, then $C_1 \supsetneq C_2$. Hence $\Pa{\mathcal{T}, \supsetneq}$ is a tree. The only root of $\mathcal{T}$ is $X \in \Conn{X \setminus K_0}$, and $\tlevel(l) = \Conn{X \setminus K_l}$, which is finite by Lemma~\ref{lem:finitemanycomponent}, so $\mathcal{T}$ is locally finite. Hence $\Set{\tdepth(x) \mid x \in \mathcal{T}} = \mathbb{N} \cup \Set{0}$, and $\theight(\mathcal{T}) = \upomega$. \end{proof} \begin{lemma} \label{lem:treefiltration} Let $X$ be a connected locally connected locally compact Hausdorff space and let $\Set{K_l}_{l \in \mathbb{N} \cup \Set{0}}$ be an exhaustive filtration of $X$ by compact sets with $K_0 = \emptyset$. Define \begin{equation*} \mathcal{T} = \coprod_{l \in \mathbb{N} \cup \Set{0}} \Connub{X \setminus K_l}. \end{equation*} Then $\tbranch(T)$ is bijective to $\mathcal{E}(X)$. \end{lemma} \begin{proof} The statement on $\mathcal{E}(X)$ is the result of Proposition~\ref{prop:exhaustive-relative-end}. \end{proof} \begin{proposition} \label{prop:exhaustive-relative-end} If $\mathcal{B}$ is a refinement of $\mathcal{A}$, then $\mathcal{E}(X \pipe \mathcal{A}) = \mathcal{E}(X)$, in the sense that the map $\mathcal{E}(X \pipe \mathcal{A}) \rightarrow \mathcal{E}(X \pipe \mathcal{B})$ is a surjection. \end{proposition} \begin{proof} Suppose we know $\Res{\epsilon}_{\mathcal{A}} \in \mathcal{E}(X \pipe \mathcal{A})$, we want to recover $\epsilon(K)$ for any $K \in \mathcal{K}(X)$. Since $\mathcal{A}$ is exhaustive, there is $L \in \mathcal{A}$ are such that $K \subset L$. Then let $\epsilon(K) = \rho^K_L \Res{\epsilon}_{\mathcal{A}}(L)$. Let $\epsilon \in \mathcal{E}(X \pipe \mathcal{A})$. For any $\mathcal{B}_1 \subset \mathcal{B}$, if $\epsilon_1 \in \mathcal{E}(X \pipe \mathcal{B}_1)$ such that for any $L \in \mathcal{B}_1$ and $K \in \mathcal{A}$ with $K \subset L$ we have $\rho^K_L \epsilon_1(L) = \epsilon(K)$. Let $L_1 \in \mathcal{B} \setminus \mathcal{B}_1$. If $K \in \mathcal{A}$ such that $K \supset L_1$ then let $\epsilon'(L_1) = \rho^{L_1}_K \epsilon(K)$. Otherwise choose $\epsilon'(L_1)$ arbitrarily. If there is $L_1 \in \mathcal{A}$ such that $K \subset L_1$, then there is $L_2 \in \mathcal{A}$ that contains $L \cup L_1 \in \mathcal{K}(X)$. So $\epsilon(K) = \rho^K_L \Res{\epsilon}_{\mathcal{A}}(L) = \rho^K_L \rho_{LL_2} \Res{\epsilon}_{\mathcal{A}}(L_2) = \rho_{KL_2} \Res{\epsilon}_{\mathcal{A}}(L_2) = \rho_{KL_1} \Res{\epsilon}_{\mathcal{A}}(L_1)$ is well defined. If $K \in \mathcal{A}$ then $\epsilon(K) = \rho_{KK} \Res{\epsilon}_{\mathcal{A}}(K) = \Res{\epsilon}_{\mathcal{A}}(K)$, so $\Res{\epsilon}_{\mathcal{A}}$ is the restriction of $\epsilon$ on $\mathcal{A}$. Then we check that if $K, K_1 \in \mathcal{K}(X)$ are such that $K \subset K_1$ and $L \in \mathcal{A}$ contains $K \cup K_1$, then $\epsilon(K) = \rho^K_L \Res{\epsilon}_{\mathcal{A}}(L) = \rho_{KK_1} \rho_{K_1L} \Res{\epsilon}_{\mathcal{A}}(L) = \rho_{KK_1} \epsilon(K_1)$. Hence $\epsilon \in \mathcal{E}(X)$. \end{proof} \begin{definition} The fiber structure defined a monodromy structure on $\mathcal{E}(F)$. \begin{align*} \pi_1(B) \longrightarrow \mathcal{S}_{\mathcal{E}(F)}. \end{align*} For $[\gamma] \in \pi_1(B)$, the lift of $\gamma$ maps one component of $F \setminus F_\alpha$ to another. For $\beta > \alpha$, and $C \supset C_1$, we want to show $\gamma \cdot C \supset \gamma \cdot C_1$ because of the naturality of $\nu$. \end{definition} \begin{definition} \label{def:filtration} A sequence $\Set{K_l}_{l \in \mathbb{N} \cup \Set{0}} \subset \mathcal{K}(X)$ is an \notation{exhaustive filtration by compact sets} if $K_l \ssubset K_{l+1}$ for each $l \in \mathbb{N} \cup \Set{0}$ and $\bigcup_{l=0}^\infty K_l = X$. \end{definition} One can check that an exhaustive filtration by compact sets is in fact exhaustive. This suggests that an exhaustive subcollection is a good representative of all compact subsets when considering ends. \end{comment} \section{Conclusion: Proof of Main Theorem} \label{sec:proof-main-theorem} We prove Theorem~\ref{thm:main-theorem}. We apply Theorem~\ref{thm:filtration-theorem} to $\lb{M}$ and $\omega, \tau$ and obtain the tree $\mathcal{T}$ of filled subbundles of $\lb{M}$, such that \eqref{eq:support-in-level}--\eqref{eq:volume-in-E} are satisfied. For $n \in \mathbb{N}$ and $\lb{C} \in \tlevel(2n-2)$, applying Lemma~\ref{lem:lemma1} to $\Pa{\Shta_\mathcal{T} \lb{C}}^\circ$, there are fiber diffeomorphisms $\varphi_n, \psi_n \colon M \to M$ such that $\varphi_n ^* \omega_{n-1} = \omega_n$, $\psi_n ^* \tau_{n-1} = \tau_n$, and $\varphi_n = \psi_n = \identity$ outside of $\lb{M}_{(\alpha_{2n-2}, \alpha_{2n})}$. Let \begin{equation} \label{eq:omega-infinity} \begin{aligned} \omega_\infty &= \lim_{n \to \infty} \omega_n, & \tau_\infty &= \lim_{n \to \infty} \tau_n, \\ \varphi_\infty &= \varphi_1 \circ \varphi_2 \circ \cdots, & \psi_\infty &= \psi_1 \circ \psi_2 \circ \cdots. \end{aligned} \end{equation} Since $\Set{\Pa{\undsp(\Shta_\mathcal{T} \lb{C})}^\circ}_{\lb{C} \in \mathcal{T}, 2\divides\tdepth(C)}$ is mutually disjoint, the pointwise limits in \eqref{eq:omega-infinity} will be stable at a finite $n$, so $\omega_\infty, \tau_\infty \in \FVform(\lb{M})$, $\varphi_\infty, \psi_\infty \colon M \to M$ are fiber diffeomorphisms, \begin{equation*} \rint_{\Tse_\mathcal{T} \lb{M}} \omega_\infty = \rint_{\Tse_\mathcal{T} \lb{M}} \tau_\infty, \qquad \rint_{\Shta_\mathcal{T} \lb{A}} \omega_\infty = \rint_{\Shta_\mathcal{T} \lb{A}} \tau_\infty \end{equation*} for each $\lb{A} \in \mathcal{T}$ with odd depth, $\varphi_\infty^* \omega = \omega_\infty$, and $\psi_\infty^* \tau = \tau_\infty$. We have left to show that there is a fiber diffeomorphism $\varphi' \colon M \to M$ such that $\varphi'^* \omega_\infty = \tau_\infty$. Let $\Set{\lb{L}_j}_{j \in \mathbb{N}}$ be $\Set{\Tse_\mathcal{T} \lb{M}} \cup \Set{\overline{\Shta_\mathcal{T} \lb{A}}}_{\lb{A} \in \mathcal{T},2\notdivides\tdepth(A)}$, then this is the result of Lemma~\ref{lem:lemma3}. Finally, \begin{equation*} \varphi = \varphi_\infty \circ \varphi' \circ \psi_\infty^{-1} \colon M \to M \end{equation*} is as required. \section{Category of filled bundles} \label{sec:category} Throughout this paper manifolds are always assumed to be ($\mathrm{C}^\infty$) smooth manifolds with or without boundary except where explicitly stated. Let $X, Y$ be manifolds. The space of $q$-forms on $X$ is denoted by $\Omega^q(X)$, and the space of compactly supported $q$-forms on $X$ is denoted by $\Omega^q_\cspt(X)$. The map $\der \colon \Omega^q(X) \to \Omega^{q+1}(X)$ denotes the exterior derivative. For any smooth map $f \colon X \to Y$ denote by $f^* \colon \Omega^q(Y) \to \Omega^q(X)$ the pullback map of $q$-forms. Let $\cont(X; Y)$ be the space of continuous functions, and $\smth(X; Y)$ the space of smooth functions from $X$ to $Y$. For any topological spaces $X, Y$ we denote by $\proj_1 \colon X \times Y \to X$ and $\proj_2 \colon X \times Y \to Y$ the projection maps corresponding to the product space $X \times Y$. For any map $f \colon X \to Y$ we may use the same notation $f$ for the restriction $\Res{f}_A$ to some subspace $A \subset X$ if $A$ is clear from the context. Denote by $\mathcal{K}(X)$ the collection of compact subspaces of $X$. Let $\identity_X \colon X \to X$ denote the identity map. We may omit the subscript ``$X$'' if it is clear from the context. Let $X$ be a manifold, and let $f \colon X \to \mathbb{R}$ be a smooth function. Let $\Diff(X)$ be the group of diffeomorphisms of $X$, then define $\Diff(X, f) = \Setby{\phi \in \Diff(X)}{f \circ \phi = f}$. Let $\Regular(f)$ be the set of regular values of $f$ (including $\mathbb{R} \setminus f(X)$). \subsection{Objects: filled bundles} \label{ssec:object} \begin{definition} \label{def:filled-bundle} A \notation{filled bundle} is a 6-tuple $\lb{M} = (\pi, M, B, F, f, h)$ with the following properties: \begin{enumerate}[label=\textup{(\roman*)}] \item \label{itm:def-filbund-elements} $B, F$ are smooth manifolds with or without boundary; if $B, F$ both have boundaries, then $M$ is a manifold with corners, otherwise, $M$ is a manifold with or without boundary. $\pi \colon M \to B$ is a smooth map, $f$ is an exhaustion function for $M$ and $h \colon F \to \mathbb{R}$ is a smooth function. We call $\pi$ the \notation{projector}, $M$ the \notation{underlying space}, $B$ the \notation{base}, $F$ the \notation{fiber}, $f$ the \notation{filler}, $h$ the \notation{fiber filler} of $\lb{M}$. Write $\undsp(\lb{M}) = M$; \item \label{itm:def-filbund-bundle} every point $p \in B$ has a neighborhood $U$ such that the following diagram commutes: \begin{equation} \xymatrix{ \pi^{-1}(U) \ar[r]^{\phi_U} \ar[d]_{\pi} & U \times F \ar[dl]^{\proj_1} \\ U & }, \end{equation} where $\phi_U \colon \pi^{-1}(U) \to U \times F$ is a diffeomorphism. We call $U$ a \notation{trivializing neighborhood} of $p$ with respect to $\lb{M}$ and $(U, \phi_U)$ a \notation{trivialized chart} of $\lb{M}$; \item \label{itm:def-filbund-group} there is an open cover $\Set{U_i}_{i \in \mathcal{I}}$ of $B$ by trivializing neighborhoods with corresponding trivialized charts $(U_i, \phi_i)$ such that for each $i, j \in \mathcal{I}$ and $p \in U_i \cap U_j$, the following diagram commutes: \begin{equation*} \xymatrixcolsep{5pc} \xymatrix{ \pi^{-1}(p) \ar[d]^{\pi} \ar[dr]^{\pi} & \\ \Set{p} \times F \ar[d]^{\proj_2} \ar[r]_{\phi_i \circ \phi_j^{-1}} & \Set{p} \times F \ar[d]^{\proj_2} \\ F \ar[r]^{\phi_{ij}(p)} \ar[dr]_{h} & F \ar[d]^{h} \\ & \mathbb{R} }. \end{equation*} We call the atlas $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ a \notation{local trivialization} of $\lb{M}$. The \notation{transition map} $\phi_{ij} \colon U_i \cap U_j \to \Diff(F)$ is uniquely defined by the above diagram; \item \label{itm:def-filbund-filler} for any $i \in \mathcal{I}$ the following diagram commutes: \begin{equation*} \xymatrixcolsep{5pc} \xymatrix{ \pi^{-1}(U_i) \ar[d]^{\phi_i} \ar[r]^-{\Res{f}_{\pi^{-1}(U_i)}} & \mathbb{R} \\ U_i \times F \ar[r]^-{\proj_2} & F \ar[u]_{h} }. \end{equation*} \end{enumerate} \end{definition} Unless otherwise stated, we will always choose trivialized charts of $\lb{M}$ satisfying \ref{itm:def-filbund-group} and local trivializations satisfying \ref{itm:def-filbund-filler}. Note that $\Regular(f) = \Regular(h)$. \begin{definition} \label{def:oriented-connected-filled-bundle} If $\lb{M}$ is a filled bundle with oriented underlying space, base and fiber, we call $\lb{M}$ an \notation{oriented filled bundle}. If $\lb{M}$ is a filled bundle with connected underlying space, we call $\lb{M}$ a \notation{connected filled bundle}. \end{definition} \begin{remark} Noting that the projector is a continuous map, a connected filled bundle always has a connected base. \end{remark} \begin{remark} The notion of manifolds with corners is exclusively used for the case that both the base and the fiber are manifolds with boundary. In this case, the underlying space is locally the product of two manifolds with boundary, so the smoothness on it is well-defined. \end{remark} \begin{remark} A \notation{fiber bundle} is a $4$-tuple $(\pi, M, B, F)$ satisfying \ref{itm:def-filbund-elements} and \ref{itm:def-filbund-bundle}. The \notation{structure group} $\mathcal{G}$ of a fiber bundle $(\pi, M, B, F)$ is a subgroup of $\Diff(F)$ such that there is a local trivialization $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ whose transition maps $\phi_{ij}$ have images in $\mathcal{G}$. By \ref{itm:def-filbund-group} one observes that a filled bundle $\lb{M}$ is a fiber bundle with structure group $\Diff(F, h)$. If $\lb{M}$ has oriented base $B$ and fiber $F$, and its structure group is the group of orientation-preserving diffeomorphisms in $\Diff(F, h)$, then $\lb{M}$ is an oriented filled bundle. \end{remark} \begin{lemma} \label{lem:existance-filler} If the $5$-tuple $(\pi, M, B, F, h)$ satisfies \ref{itm:def-filbund-elements} -- \ref{itm:def-filbund-group}, there exists an $f \colon M \to \mathbb{R}$ such that $\lb{M} = (\pi, M, B, F, f, h)$ is a filled bundle. \end{lemma} \begin{proof} Let $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ be a local trivialization of $(\pi, M, B, F, h)$ satisfying \ref{itm:def-filbund-group}. For each $i \in \mathcal{I}$ we define $f$ on $\pi^{-1}(U_i)$ such that the diagram in \ref{itm:def-filbund-filler} commutes. Then for each $i, j \in \mathcal{I}$ with $U_i \cap U_j \neq \emptyset$, we have $h \circ \proj_2 \circ \Res{\phi_i^{-1}}_{\pi^{-1}(U_i \cap U_j)} = h \circ \proj_2 \circ \Res{\phi_j^{-1}}_{\pi^{-1}(U_i \cap U_j)}$. Hence $f$ is well defined on $M$. \end{proof} \begin{definition} \label{def:exhausted-bundle} Let $X$ be a manifold. An \notation{exhaustion function} $f$ for $X$ is a smooth function $f \colon X \to \mathbb{R}$ such that for any $\alpha \in \mathbb{R}$, $f^{-1}((-\infty, \alpha])$ is compact. By \cite[Proposition 2.28]{MR2954043} the exhaustion functions for $X$ always exist. Let $\lb{M} = (\pi, M, B, F, f, h)$ be a filled bundle. If $f$ is an exhaustion function for $M$ and $h$ is an exhaustion function for $F$, then we call $\lb{M}$ an \notation{exhausted bundle}, $f$ the \notation{exhauster}, and $h$ the \notation{fiber exhauster} of $\lb{M}$. \end{definition} \begin{lemma} \label{lem:compact-exhausted-bundle} A filled bundle with compact base and whose fiber filler is an exhaustion function is an exhausted bundle. An exhausted bundle $\lb{M}$ has compact base and compact fiber if and only if it has compact underlying space. \end{lemma} \begin{proof} Let $\lb{M} = (\pi, M, B, F, f, h)$ be such a filled bundle. Since $B$ is compact, let $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ be a local trivialization of $\lb{M}$ such that $I$ is finite and $\overline{U_i}$ is compact for any $i \in \mathcal{I}$. Now that $h$ is an exhaustion function for $F$, let $\alpha \in \Regular(f)$, then \begin{equation*} f^{-1}((-\infty, \alpha]) = \bigcup_{i \in \mathcal{I}} f^{-1}((-\infty, \alpha]) \cap \pi^{-1}(\overline{U}_i) = \bigcup_{i \in \mathcal{I}} \phi_i \Pa{\overline{U}_i \times h^{-1}((-\infty, \alpha])} \end{equation*} is compact. Hence $f$ is an exhaustion function for $M$ and then $\lb{M}$ is an exhausted bundle. If $\lb{M}$ is exhausted and $B$, $F$ are compact, then we have $M = \bigcup_{i \in \mathcal{I}} \Pa{M \cap \pi^{-1}(\overline{U_i})} = \bigcup_{i \in \mathcal{I}} \phi_i \Pa{\overline{U_i} \times F}$ is compact. If $\lb{M}$ is exhausted and compact, then $B = \pi(M)$ is compact, and $f$ is bounded. So $h$ is bounded, which implies the compactness of $F$. \end{proof} \subsubsection*{Examples of filled and exhausted bundles} In this subsubsection, we give some examples of filled and exhausted bundles, on the latter of which our Theorem~\ref{thm:main-theorem} performs. Recall that in the Greene--Shiohama theorem, we require not only the total volume be equal, but also for each end either has finite or infinite volume with respect to both volume forms. An end $\epsilon$ of a topological space $X$ is an assignment of any compact $K \subset X$ to $\epsilon(K)$ an unbounded connected component of $X \setminus K$ such that if $K \subset L$ are compact subsets of $X$ then $\epsilon(K) \supset \epsilon(L)$. As a generalization of their result, our Theorem~\ref{thm:main-theorem} is backed by the concept of ends both in the statement and the proof. Things are more involved in our situation, since the bundle itself and the fiber are both noncompact manifolds and have ends. In our theorem, we will consider the ends of the bundle. The ends of the bundle is determined by the ends of the fiber and fiber bundle structure. We show later in Section~\ref{sec:appendix} that there is a surjection from the ends of the fiber to the ends of the bundle. \begin{example} \label{ex:product-space} Perhaps the simplest example of filled bundles is the trivial bundle. Let $M$ be an oriented connected manifold without boundary, and $B$ be a manifold. Let $f \colon M \to \mathbb{R}$ be a smooth function. Then $\lb{P} = (\proj_1, B \times M, B, M, f \circ \proj_2, f)$ is a filled bundle, which is an exhausted bundle if and only if $B$ is compact and $f$ is an exhaustion function for $M$. \end{example} \begin{example} \label{ex:vector-bundle} Let $(\pi, M, B, \mathbb{R}^k)$ be a vector bundle with rank $k$ and let $g \colon M \times_B M \to B \times \mathbb{R}$ be a metric on the bundle, that is, for any $b \in B$, $\Res{g}_{\pi^{-1}(b)} \colon \pi^{-1}(b) \times \pi^{-1}(b) \to \mathbb{R}$ is an inner product. Let $f \colon M \to \mathbb{R}, x \mapsto g(x, x)$ and $h \colon \mathbb{R}^k \to \mathbb{R}, y \mapsto \abs{y}^2$. Then $(\pi, M, B, \mathbb{R}^k, f, h)$ is a filled bundle, which is exhausted if and only if $B$ is compact. This bundle has only one end. \end{example} \begin{example} \label{ex:product-compact-bundle} Let $(\pi, N, B, E)$ be a compact fiber bundle and let $F$ be a noncompact manifold with a smooth function $h \colon F \to \mathbb{R}$. Then $(\pi \circ \proj_1, N \times F, B, E \times F, h \circ \proj_2, h \circ \proj_2)$ is a filled bundle which is exhausted if and only if $h$ is an exhaustion function. Note that $N \times F$ has as many ends as $F$ has, which can be an arbitrary number from one to the cardinality of the continuum. \end{example} \begin{example} \label{ex:cylinder-with-holes} Let \begin{equation*} F = \Setby{(x,y,z) \in \mathbb{R}^3}{x^2 + y^2 = 1, y^2 + z^2 > \frac14}, \end{equation*} then $F$ is a noncompact $2$-manifold with $4$ ends. Let $\phi \in \Diff(F)$ be the diffeomorphism given by $\phi(x, y, z) = (x, -y, -z)$, switching two ends $z \to +\infty$ and $z \to -\infty$. Let $h \colon F \to \mathbb{R}, h(x, y, z) = z^2 + \Pa{(y^2+z^2) - \frac14}^{-1}$, then $h$ is an exhaustion function with the property $h \circ \phi = h$. Let $B = \mathbb{S}^1 = (0, 2) / (p \mapsto p+1)$. Then define $M = (0, 2) \times_\varphi F$ where $\varphi \colon (1, 2) \times F \to (0, 1) \times F$ is given by $\varphi(p, y) = (p-1, \phi(y))$. Let $\pi \colon M \to B$ be the map induced by the $\proj_1 \colon (0, 2) \times F \to (0, 2)$, and $f \colon M \to \mathbb{R}$ be the map induced by the $h \circ \proj_2 \colon (0, 2) \times F \to \mathbb{R}$. Then $\lb{M} = (\pi, M, B, F, f, h)$ is an oriented exhausted bundle where $M$ has $3$ ends, since $\phi$ is orientation preserving. \end{example} \begin{example} \label{ex:open-subset-of-euclidean} Let $k \in \mathbb{N}$ and $\mathcal{G}$ be a subgroup of $\mathrm{SO}(n)$. Let $E \subset \mathbb{R}^k$ be a noncompact complete submanifold without boundary, which is invariant under $\mathcal{G}$. Let $u \colon \mathbb{R}^k \to \mathbb{R}$ be a smooth function such that $u \circ \phi = u$ for any $\phi \in \mathcal{G}$. Let \begin{equation*} F = E \cap \Set{u > 0}. \end{equation*} Let $h \colon F \to \mathbb{R}, h(x) = \abs{x}^2 + u(x)^{-1}$, then $h$ is an exhaustion function with the property $h \circ \phi = h$ for any $\phi \in \mathcal{G}$. Let $(\pi, M, B, F)$ be any fiber bundle with structure group $\mathcal{G}$ such that $B$ is compact. Let $f \colon M \to \mathbb{R}$ be defined by Lemma~\ref{lem:existance-filler}. Then $\lb{M} = (\pi, M, B, F, f, h)$ is an oriented exhausted bundle with noncompact fiber. \end{example} \subsection{Morphisms: filled morphisms} \label{ssec:morphism} \begin{definition} Let $\lb{M} = (\pi, M, B, F, f, h), \lb{M'} = (\pi', M', B', F', f', h') \in \Object(\lBundle)$ be two filled bundles. We define $\Morphism(\lb{M'}, \lb{M})$ as follows: $\sfmu \in \Morphism(\lb{M'}, \lb{M})$ is a pair $\sfmu = (\mu, \mu_{B'})$, where $\mu \colon M' \to M$, $\mu_{B'} \colon B' \to B$ are smooth maps such that the following diagram commutes: \begin{equation*} \xymatrixcolsep{3pc} \xymatrix{ & M' \ar[r]^{\pi'} \ar[dl]_{f'} \ar[d]^{\mu} & B' \ar[d]^{\mu_{B'}} \\ \mathbb{R} & M \ar[r]^{\pi} \ar[l]_{f} & B }. \end{equation*} We call $\mu$ the \notation{bundle mapping} and $\mu_{B'}$ the \notation{base mapping} of $\sfmu$. Hence filled morphisms are maps $\lb{M'} \to \lb{M}$ between filled bundles which preserve bundle structures and fillers. The base morphism $\mu_{B'}$ can be derived from $\mu$, moreover, there is a smooth map $\mu_{F'} \colon F' \to F$ induced from $\mu$ by the following approach. Let $p' \in B', p = \mu_{B'}(p') \in B$ and let $(U', \phi'_{U'})$ (resp.\@ $(U, \phi_{U})$) be a trivialized chart of $\lb{M'}$ (resp.\@ $\lb{M}$) where $U'$ is a neighborhood of $p'$ (resp.\@ $U$ is a neighborhood of $p$). Then define $\mu_{F'}$ such that the following diagram commutes: \begin{equation*} \xymatrixcolsep{3pc} \xymatrix{ (\pi')^{-1}(p') \ar[r]^{\phi'_{U'}} \ar[d]^{\mu} & \Set{p'} \times F' \ar[r]^-{\simeq} \ar[d] & F' \ar[d]^{\mu_{F'}} \\ \pi^{-1}(p) \ar[r]^{\phi_U} & \Set{p} \times F \ar[r]^-{\simeq} & F }. \end{equation*} Note that the map $\mu_{F'}$ depends on the choice of $p'$, $\phi_U$ and $\phi'_{U'}$. We call $\mu_{F'}$ a \notation{fiber map} of $\sfmu$. \end{definition} \begin{remark} Let $\lb{M} \in \Object(\lBundle)$, then we define $\lb{id}_{\lb{M}} = (\identity_M, \identity_B) \in \Morphism(\lb{M}, \lb{M})$ as the identity of $\lb{M}$. The composition of filled morphisms is the component-wise composition. A filled morphism is called an \notation{filled isomorphism} if it is invertible. The objects of $\lBundle$ are actually defined as equivalence classes under filled isomorphisms, but we will ignore this difference throughout the paper. \end{remark} \subsection{Subbundles} \label{ssec:subbundle} \begin{definition} \label{def:filled-subbundle} Let $\lb{M} = (\pi, M, B, F, f, h)$ and $\lb{A} = (\Res{\pi}_A, A, B, P, \Res{f}_A, \Res{h}_P)$ be filled bundles and $\sfiota = (\iota, \identity_B) \in \Morphism(\lb{A}, \lb{M})$. If $\iota \colon A \to M$ is an embedding of manifolds (can have corners), then we say $\iota$ is an \notation{embedding} and $\lb{A}$ is a \notation{filled subbundle} of $\lb{M}$. \end{definition} \begin{definition} \label{def:filled-subspace} Let $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ be a local trivialization of $\lb{M}$. We call $A$ a \notation{filled subspace} of $\lb{M}$ (with respect to $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$) if for any $i \in \mathcal{I}$, there is a subset $P_i \subset F$ such that, \begin{equation*} \phi_i (A \cap \pi^{-1}(U_i)) = U_i \times P_i. \end{equation*} Note that for any $i, j \in \mathcal{I}$ with $U_i \cap U_j \neq \emptyset$, there is a diffeomorphism of $F$ sending $P_i$ to $P_j$. By connectedness of $B$, all $P_i$'s are diffeomorphic, and we denote one of the representatives by $P$. Let $\psi_i \colon U_i \times F \to U_i \times F$ be a diffeomorphism which commutes with $\proj_1$ and such that $\psi_i(U_i \times P_i) = U_i \times P$. If $P$ is a submanifold of $F$ then $\lb{A} = (\Res{\pi}_A, A, B, P, \Res{f}_A, \Res{h}_P)$ is a filled subbundle of $\lb{M}$, with $\Set{(U_i, \psi_i \circ \Res{\phi_i}_{\Res{\pi}_A^{-1}(U_i)})}_{i \in \mathcal{I}}$ as a local trivialization. We write $\Res{\lb{M}}_A = \lb{A}$ the \notation{restriction} of $\lb{M}$ on $A$ when $A$ is a filled subspace of $\lb{M}$. If $A \subset M$ is a filled subspace of $\lb{M}$ then its closure of $\overline{A}$ and interior $A^\circ$ in $M$ are filled subspaces of $\lb{M}$. We write $\overline{\lb{A}} = \Res{\lb{M}}_{\overline{A}}$ and $\lb{A}^\circ = \Res{\lb{M}}_{A^\circ}$ for simplicity. \end{definition} \begin{definition} \label{def:category-subbundle} Let $\lb{M} = (\pi, M, B, F, f, h)$ be a filled bundle. Define $\lBundle_{\lb{M}}$ the category of filled subbundles of $\lb{M}$ as the subcategory of $\lBundle$ with subbundles of $\lb{M}$ as objects and embeddings as morphisms. \end{definition} \begin{lemma} \label{lem:subbundle-slicing} Let $m \in \mathbb{N}$. If $A = \bigcup_{k = 1}^m A_k$, where $A_k \in \Conn(f^{-1}(I_k))$ for some interval $I_k$ whose endpoints are regular values of $f$, then $A$ is a filled subspace of $\lb{M}$ with respect to any local trivialization. \end{lemma} \begin{proof} Let $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$ be a local trivialization of $\lb{M}$. Then for any $i \in \mathcal{I}$, $k \in \Set{1, \dotsc, m}$, by continuity of $\phi_i$, $\phi_i(A_k \cap \pi^{-1}(U_i))$ is equal to $U_i \times P_{ik}$ for some $P_{ik}$, which is the disjoint union of some connected components of $h^{-1}(I_k)$. Hence let $P_i = \bigcup_{k = 1}^m P_{ik}$, then $\phi_i(A \cap \pi^{-1}(U_i)) = U_i \times P_i$. So $A$ is a filled subspace of $\lb{M}$ with respect to $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$. \end{proof} \begin{definition} \label{superlevel-subbundle} A \notation{superlevel subbundle} of $\lb{M}$ is the restriction of $\lb{M}$ on an unbounded connected component of $f^{-1}((\alpha, +\infty))$ for some $\alpha \in \Regular(f)$. \end{definition} \subsection{Functor: release} \label{ssec:release-functor} Consider $\Conn \colon \mathbf{Top} \to \mathbf{Set}$ the functor from the category of topological spaces with continuous maps to the category of sets with maps (usually written as $\pi_0$). For any $X \in \Object(\mathbf{Top})$ a topological space, $\Conn X$ is the set of connected components of $X$, and let $\Connub{X} \subset \Conn X$ be the subset of those which are unbounded (has noncompact closure). For any $X, Y \in \Object(\mathbf{Top})$ and $\mu \colon X \to Y$ a continuous map, $\Conn \mu \colon \Conn X \to \Conn Y$ sends any $C \in \Conn X$ to the unique connected component of $Y$ containing $\mu(C)$. Here we are using the fact that the image of connected sets under a continuous map is still connected. \begin{definition} \label{def:category-connected-bundle} The category $\ConnlBundle$ of connected filled bundles is the subcategory of $\lBundle$ which consists of all connected filled bundles with all filled morphisms between them. \end{definition} We define a functor $\Rls \colon \ConnlBundle \to \ConnlBundle$ on the category of connected filled bundles. Let $\lb{M} = (\pi, M, B, F, f, h)$ be a connected filled bundle, then \begin{equation*} \Rls \lb{M} = (\Rls \pi, M, B_M, F_M, f, h_M) \in \Object(\ConnlBundle) \end{equation*} is defined as follows: Set-theoretically, let \begin{equation*} B_M = \coprod_{p \in B} \Conn \pi^{-1}(p), \end{equation*} and let $c_M \colon B_M \to B$ be such that $c_M \Pa{\Conn \pi^{-1}(p)} = \Set{p}$. Let $\Rls \pi \colon M \to B_M$ be such that $\Res{(\Rls \pi)}_{\pi^{-1}(p)} = \Conn \colon \pi^{-1}(p) \to \Conn \pi^{-1}(p)$, that is, sending any point in $\pi^{-1}(p)$ to the connected component of $\pi^{-1}(p)$ it lies in. Let $(U, \phi_U)$ be a locally trivialized chart of $\lb{M}$. Define $\lambda_U \colon \coprod_{p \in U} \Conn \pi^{-1}(p) \to U \times \Conn{\pi^{-1}(U)}$ as the map sending a connected component of $\pi^{-1} (p)$ for each $p \in U$ to $p$ paired with the connected component of $\pi^{-1}(U)$ containing that component of $\pi^{-1} (p)$. Choose the unique topological and smooth structures on $B_M$ with which $\lambda_U$ is a diffeomorphism. \begin{lemma} \label{release-covering-space} The map $c_M \colon B_M \to B$ is a covering map, and $\Rls \pi \colon M \to B_M$ is the map of a fiber bundle with connected fiber. \end{lemma} \begin{proof} For any locally trivialized chart $(U, \phi_U)$ of $\lb{M}$, we have \begin{equation*} \xymatrixcolsep{5pc} \xymatrix{ \pi^{-1}(U) \ar[r]^-{\pi \times \Conn} \ar[dr]^{\Rls \pi} \ar[d]^{\pi} & U \times \Conn{\pi^{-1}(U)} \\ U & \coprod\limits_{p \in U} \Conn \pi^{-1}(p) \ar[u]_{\lambda_U} \ar[l]^-{c_M} }, \end{equation*} so $c_M$ is a covering map, $\Rls \pi$ is smooth and locally trivial, and $(\Rls \pi)^{-1}(p')$ is connected for each $p' \in B_M$. Since $M$ is connected, $B_M = (\Rls \pi)(M)$ is connected. Since $\Rls \pi$ is locally trivial, fibers are locally diffeomorphic, and since $B_M$ is connected, this shows that every fiber are diffeomorphic. Choose any $F_M \in \Conn(F)$. Then $(\Rls \pi, M, B_M, F_M)$ is a filled bundle with connected fiber. \end{proof} Let $h_M = \Res{h}_{F_M}$. Then $\Rls \lb{M} = (\Rls \pi, M, B_M, F_M, f, h_M)$ is a filled bundle with connected fiber. If $\lb{M}$ is exhausted, $\Rls \lb{M}$ is also exhausted. Note that $(\identity_M, c_M)$ is a filled morphism from $\Rls \lb{M}$ to $\lb{M}$. The relation between $\lb{M}$ and $\Rls \lb{M}$ is shown in Figure~\ref{fig:coveringspace}. \inputfigure{covering-space}{coveringspace}{ A connected filled bundle $\lb{M} = (\pi, M, B, F, f, h)$ and its release $\Rls \lb{M} = (\Rls \pi, M, B_M, F_M, f, h_M)$. } Let $\lb{M} = (\pi, M, B, F, f, h)$, $\lb{M'} = (\pi', M', B', F', f', h')$ be connected filled bundles, and let $\sfmu = (\mu, \mu_{B'}) \in \Morphism(\lb{M'}, \lb{M})$ be a filled morphism. We set $\Rls(\sfmu) = (\mu, \nu)$, where $\nu \colon B_{M'} \to B_M$ is defined by the following commuting diagram: \begin{equation*} \xymatrixcolsep{5pc} \xymatrix{ \pi^{-1}(U) \ar@/_5pc/[dd]_{\Rls \pi} \ar[d]^{\pi \times \Conn} & (\pi')^{-1}(U') \ar[l]^{\mu} \ar@/^5pc/[dd]^{\Rls \pi'} \ar[d]_{\pi' \times \Conn} \\ U \times \Conn \pi^{-1}(U) \ar[d]^{\lambda_U} & U' \times \Conn (\pi')^{-1}(U') \ar[d]_{\lambda_{U'}} \ar[l]^{\mu_{B'} \times \Conn \mu} \\ \coprod_{p \in U} \Conn \pi^{-1}(p) \ar[d]^{c_M} & \coprod_{p' \in U'} \Conn (\pi')^{-1}(p') \ar[l]^{\nu} \ar[d]_{c_{M'}} \\ U & U' \ar[l]^{\mu_{B'}} }. \end{equation*} The map $c_M$ is natural in the sense that $c_M \circ \nu = \mu_{B'} \circ c_{M'}$. \subsubsection*{Embedding and covering map} Let $\lb{M} = (\pi, M, B, F, f, h)$ and $\lb{A} = (\Res{\pi}_A, A, B, P, \Res{f}_A, \Res{h}_P)$ be filled bundles and $\sfiota = (\iota, \identity_B) \in \Morphism(\lb{A}, \lb{M})$. If $\sfiota$ is an embedding, let $\Rls{\sfiota} = (\iota, \kappa)$, then $\kappa \colon B_A \to B_M$ is the map induced by the natural map $\Conn{\Res{\pi}_A^{-1}(U)} \to \Conn{\pi^{-1}(U)}$. \begin{lemma} If $B$ is compact and $P$ has finitely many connected components then $\kappa$ is a covering map. \end{lemma} \begin{proof} Since $B_A$ is compact, $\kappa(B_A)$ is compact. But since $\kappa$ is a local diffeomorphism, $\kappa(B_A)$ is open in $B_M$, which means $\kappa$ is surjective, hence a covering map. \end{proof} \begin{definition} Let $\kappa \colon B' \to B$ be a covering space with $B$ connected. Define $\# \kappa$ the \notation{number of sheets} of $\kappa$ as the number of $\kappa^{-1}(p)$ for any $p \in B$ (independent of $p$). \end{definition} \section{Topological construction: saturated slicings} \label{sec:topology} Let $\lb{M} = (\pi, M, B, F, f, h)$ be an exhausted bundle. Fix a basepoint $x_0 \in M$ as a minimum point of $f$ whenever we use saturated slicings. For any $\alpha \in \Regular(f) \cap f(M)$, let $C_\alpha$ be the connected component of $f^{-1}((-\infty, \alpha])$ containing $x_0$. Define $M_{(-\infty, \alpha]}$ as $C_\alpha$ union the unbounded connected components of $M \setminus C_\alpha$, then $M_{(-\infty, \alpha]}$ is compact and connected, see Figure~\ref{fig:Malpha}. Let $M_{(-\infty, \alpha)}$ be the interior of $M_{(-\infty, \alpha]}$. For $\alpha \in \mathbb{R} \setminus f(M)$, let $M_{(-\infty, \alpha)} = M_{(-\infty, \alpha]} = \emptyset$. For any interval $I \subset \mathbb{R}$ whose endpoints are regular values of $f$, we can define $M_I$ from the half-bounded intervals. For instance, if $\alpha, \beta \in \Regular(f)$, define $M_{(\alpha, \beta]} = M_{(-\infty, \beta]} \setminus M_{(-\infty, \alpha]}$. By Lemma~\ref{lem:subbundle-slicing}, $M_I$ is a filled subspace of $\lb{M}$ with respect to any local trivialization. Let $\lb{M}_I = (\Res{\pi}_{M_I}, M_I, B, F_I, \Res{f}_{M_I}, \Res{h}_{F_I})$, which is an object in $\lBundle_{\lb{M}}$, and we call it the \notation{saturated slicing} of $\lb{M}$ by $I$. For any $\lb{A} = (\Res{\pi}_A, A, B, P, \Res{f}_A, \Res{h}_P) \in \lBundle_{\lb{M}}$, if $A$ is a filled subspace of $\lb{M}$ with respect to $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$, then we define $A_I = A \cap M_I$ which is also a filled subspace of $\lb{M}$ with respect to $\Set{(U_i, \phi_i)}_{i \in \mathcal{I}}$. Let $\lb{A}_I = \Res{\lb{M}}_{A_I} \in \lBundle_{\lb{M}}$. \inputfigure{M-alpha}{Malpha}{ The saturated slicing $\lb{M}_{(-\infty, \alpha]}$. } \begin{definition} \label{def:saturated-slicing} Let $\lb{A} \in \lBundle_{\lb{M}}$ be a connected precompact filled subbundle of $\lb{M}$. For any $\alpha \in \Regular(f) \cap f(M))$, let $\sfiota_\alpha \colon \lb{A}_{(-\infty, \alpha]} \to \lb{A}$ be the embedding, and let $\Rls \sfiota_\alpha = (\iota_\alpha, \kappa_\alpha)$. If $\# \kappa_\alpha = 1$ we say $\lb{A}_{(-\infty, \alpha]}$ is a \notation{saturated slicing}. Define the \notation{saturating threshold} $\theta_{\lb{A}}$ for $\lb{A}$ as \begin{equation*} \theta_{\lb{A}} = \inf \Setby{\alpha \in \Regular(f) \cap f(A)}{\# \kappa_\alpha = 1}. \end{equation*} \end{definition} \begin{lemma} \label{lem:connecting} Let $X$ be a locally connected locally compact Hausdorff space. Let $K \in \mathcal{K}(X)$ and let $A, A' \subset X$ be connected and precompact. If $A, A'$ lies in the same connected component $C$ of $X$ then there is $L \in \mathcal{K}(X)$ such that they lie in the same connected component of $L \cap C$. \end{lemma} \begin{proof} For any topological space $X$ and a nonempty connected subset $A$, denote by $\conn(X, A)$ be the unique $C \in \Conn X$ such that $C \supset A$. Denote $\mathcal{K}' = \Setby{L \in \mathcal{K}(X)}{L \supset A \cup A'}$. For any $A \subset X$ that is connected and precompact, let $C = \conn(X, A)$. Since $X$ is locally connected, $C$ is open in $X$ and locally connected. Let $P(A) = \bigcup_{L \in \mathcal{K}'} \conn(L \cap C, A)$. For any $L \in \mathcal{K}'$, if $x \in \conn(L \cap C, A)$, then there is a compact connected neighborhood $\overline{U}_x \subset C$ of $x$. If $L' = L \cup \overline{U}_x \in \mathcal{K}'$, then $\conn(L \cap C, A) \cup \overline{U}_x \subset \conn(L' \cap C, A)$, so $P(A)$ is open. For any precompact connected open set $U \subset C$, we have \begin{equation*} \Pa{ \begin{aligned} U \cap P(A) \neq \emptyset &\implies \exists L \in \mathcal{K}', \quad U \cap \conn(L \cap C, A) \neq \emptyset \\ &\implies U \subset \conn\Pa{(L \cup \overline{U}) \cap C, A} \subset P(A) \end{aligned} }. \end{equation*} Thus $U \setminus P(A) \neq \emptyset$ implies that $U \subset C \setminus P(A)$. Since $C$ has a topology base consisting of connected sets, $P(A)$ is closed in $C$. Hence $P(A)$ is nonempty and clopen in $C$ and $C$ is connected, so $P(A) = C$. For any $A' \subset X$ that is connected and precompact, suppose $\conn(X, A') = C$ and $\conn(L_1 \cap C, A) \cap \conn(L_2 \cap C, A') \neq \emptyset$, $L_1, L_2 \in \mathcal{K}'$. Let $L' = L_1 \cup L_2 \in \mathcal{K}'$, since $P(A) = P(A') = C$ then $\conn(L' \cap C, A) \cap \conn(L' \cap C, A') \supset \conn(L_1 \cap C, A) \cap \conn(L_2 \cap C, A') \neq \emptyset$, which means $\conn(L' \cap C, A) = \conn(L' \cap C, A')$. \end{proof} \begin{lemma} \label{lem:saturating-threshold} For any connected exhausted subbundle $\lb{A}$ of $\lb{M}$, the saturating threshold $\theta_{\lb{A}} < +\infty$. For any $\beta \in \Regular(f)$ such that $\beta > \theta_{\lb{A}}$, $\# \kappa_\beta = 1$. \end{lemma} \begin{proof} Let $\lb{A} = (\Res{\pi}_A, A, B, P, \Res{f}_A, \Res{h}_P)$, $\Rls \lb{A} = (\Rls(\Res{\pi}_A), A, B_A, P_A, f, h_A)$. Fix $\alpha \in \Regular(f) \cap f(A)$ then we consider $\lb{A}_{(-\infty, \alpha]}$. Since the fiber of $\lb{A}_{(-\infty, \alpha]}$ is the interior of a compact manifold with boundary, it can only have finitely many components. Let $P_\alpha$ denote the fiber of $\Res{(\Rls \lb{A})}_{A_{(-\infty, \alpha]}}$. By Lemma~\ref{lem:connecting}, there is $K \in \mathcal{K}(\overline{P_A})$ which is connected and contains $x_0$ and every component of $P_\alpha$, where the closure is taken in $F$. Suppose $\beta \in \Regular(f)$ and $\beta \geq \max_K fh$, then $P_\beta \supset K$ contains every component of $P_\alpha$. Let $\sfiota^\alpha_\beta \colon \Res{(\Rls \lb{A})}_{A_{(-\infty, \alpha]}} \to \Res{(\Rls \lb{A})}_{A_{(-\infty, \beta]}}$ be the embedding, and let $\Rls \sfiota^\alpha_\beta = (\iota^\alpha_\beta, \kappa^\alpha_\beta)$. Then $\kappa^\alpha_\beta \colon P^A_\alpha \to P^A_\beta$ is a covering map. Note that a component of $P_\beta$ contains all components of $P_\alpha$, this means the image of $\kappa^\alpha_\beta$ is a one-fold covering of $P_A$. Let $\sfiota_\beta \colon \Res{(\Rls \lb{A})}_{A_{(-\infty, \beta]}} \to \Rls \lb{A}$ be the embedding, and let $\Rls \sfiota_\beta = (\iota_\beta, \kappa_\beta)$, then that $\kappa_\beta \colon P^A_\beta \to P_A$ is a diffeomorphism. For the same reason, $\# \kappa_{\beta'} = 1$ for any $\beta' \in \Regular(f)$ no less than $\beta$. \end{proof}

https://arxiv.org/abs/1811.02550

Strange Expectations and the Winnie-the-Pooh Problem

Motivated by the study of simultaneous cores, we give three proofs (in varying levels of generality) for the expected norm of a weight in a highest weight representation of a complex simple Lie algebra. First, we argue directly using the polynomial method and the Weyl character formula. Second, we use the combinatorics of semistandard tableaux to obtain the result in type A. Third, and most interestingly, we relate this problem to the "Winnie-the-Pooh problem" regarding orthogonal decompositions of Lie algebras; although this approach offers the most explanatory power, it applies only to Cartan types other than A and C. We conclude with computations of many combinatorial cumulants.

\section{Introduction} The representation theory of complex simple Lie algebras is a classical source of algebraic combinatorics, intimately related to tableaux and plane partitions, symmetric functions and positivity questions, quantum groups and crystals, and the plactic monoid and RSK. Having fixed a Cartan subalgebra $\mathfrak{h}$, the finite-dimensional irreducible representions of a complex simple Lie algebra $\mathfrak{g}$ are completely classified by the dominant weights $\lambda$ in its weight lattice $\Lambda \subset \mathfrak{h}^*$. Using the symmetric bilinear form $\langle \cdot,\cdot \rangle$ on $\mathfrak{h}^*$ induced by the Killing form and writing $\rho$ for the half-sum of the positive roots $\Phi^+$, the Weyl dimension formula asserts that for $\lambda$ dominant, the dimension of the finite-dimensional irreducible representation $V_\lambda$ \begin{equation} \mathrm{dim}(V_\lambda) = \prod_{\alpha \in \Phi^+} \frac{\langle \alpha,\lambda+\rho \rangle}{\langle \alpha,\rho \rangle}. \end{equation} Motivated by the recent interest in statistics on simultaneous cores (see~\Cref{sec:motivation}), the purpose of this paper is to give a formula for the average norm of a weight in a highest weight representation. \begin{theorem} For $\mathfrak{g}$ a complex simple Lie algebra with $V_\lambda$ its finite-dimensional irreducible representation of highest weight $\lambda$, the expected norm of a weight in $V_\lambda$ is \[\Expt{\mu \in V_\lambda}{\langle \mu,\mu\rangle}=\frac{1}{\mathrm{dim}(V_\lambda)} \sum_{\mu \in V_\lambda} \mathrm{dim}(V_\lambda(\mu)) \langle \mu,\mu\rangle = \frac{1}{h+1} \big\langle \lambda, \lambda + 2\rho\big\rangle,\] where $\mathrm{dim}(V_\lambda(\mu))$ is the multiplicity of $\mu$ in $V_\lambda$ and $h$ is the Coxeter number of $\mathfrak{g}$. \label{thm:main_thm} \end{theorem} \Cref{thm:main_thm} is illustrated in~\Cref{fig:example1,fig:example2} for highest weight representations for $\mathfrak{sl}_3$ and for $\mathfrak{sp}_4$. \begin{figure}[htbp] \begin{center} \raisebox{-0.6\height}{\begin{tikzpicture}[scale=1.5] \draw[black,thin] (-1.5,0) -- (1.5,0); \draw[black,thin] (.75,1.299) -- (-.75,-1.299); \draw[black,thin] (-.75,1.299) -- (.75,-1.299); \filldraw[red] (0,1) circle[radius=1pt]; \node at (1,.7) (a) {$(1,0,-1)$}; \filldraw[red] (0,-1) circle[radius=1pt]; \filldraw[red] (.866,.5) circle[radius=1pt]; \filldraw[red] (.866,-.5) circle[radius=1pt]; \filldraw[red] (-.866,.5) circle[radius=1pt]; \filldraw[red] (-.866,-.5) circle[radius=1pt]; \filldraw[blue] (0,0) circle[radius=1pt]; \draw[blue] (0,0) circle[radius=3pt]; \end{tikzpicture}} \end{center} \caption{In $\mathfrak{sl}_3$, the \textcolor{black}{eight} weights in $V_\lambda$ with $\lambda=(1,0,-1)$. The average norm is $\frac{\textcolor{red}{6} \cdot 2 + \textcolor{blue}{2} \cdot 0}{\textcolor{black}{8}} = \frac{3}{2} = \frac{1}{3+1} \langle \lambda,\lambda+2\rho \rangle$.} \label{fig:example1} \end{figure} \begin{figure}[htbp] \begin{center} \raisebox{-0.6\height}{\begin{tikzpicture}[scale=.8] \draw[black,thin] (-2.5,0) -- (2.5,0); \draw[black,thin] (0,-2.5) -- (0,2.5); \draw[black,thin] (-1.77,-1.77) -- (1.77,1.77); \draw[black,thin] (-1.77,1.77) -- (1.77,-1.77); \filldraw[red] (2,1) circle[radius=1pt]; \node at (2.1,1.3) (a) {$(2,1)$}; \node at (1.2,.4) (a) {$(1,0)$}; \filldraw[red] (-2,1) circle[radius=1pt]; \filldraw[red] (-2,-1) circle[radius=1pt]; \filldraw[red] (2,-1) circle[radius=1pt]; \filldraw[red] (1,2) circle[radius=1pt]; \filldraw[red] (1,-2) circle[radius=1pt]; \filldraw[red] (-1,2) circle[radius=1pt]; \filldraw[red] (-1,-2) circle[radius=1pt]; \filldraw[blue] (1,0) circle[radius=1pt]; \filldraw[blue] (0,1) circle[radius=1pt]; \filldraw[blue] (-1,0) circle[radius=1pt]; \filldraw[blue] (0,-1) circle[radius=1pt]; \draw[blue] (0,1) circle[radius=3pt]; \draw[blue] (0,-1) circle[radius=3pt]; \draw[blue] (1,0) circle[radius=3pt]; \draw[blue] (-1,0) circle[radius=3pt]; \end{tikzpicture}} \end{center} \caption{In $\mathfrak{sp}_4$, the \textcolor{black}{sixteen} weights in $V_\lambda$ with $\lambda=(2,1)$. The average norm is $\frac{\textcolor{red}{8} \cdot 5 + \textcolor{blue}{4 \cdot 2} \cdot 1}{\textcolor{black}{16}} = 3 = \frac{1}{4+1} \langle \lambda,\lambda+2\rho \rangle$.} \label{fig:example2} \end{figure} \section{Lie Algebras and their Representation Theory} \label{sec:lie_algebras} Recall that the complex simple Lie algebras are classified by their Dynkin diagrams, illustrated in~\Cref{fig:dynkin}. Fix a complex simple Lie algebra $\mathfrak{g}$ with Cartan subalgebra $\mathfrak{h}$; all Cartan subalgebras of $\mathfrak{g}$ are conjugate. Given a complex representation $V:\mathfrak{g}\to \mathfrak{gl}(V)$, we say that the \defn{weight space} for $\mu \in \mathfrak{h}^*$ is the subspace \[V(\mu) = \{ v \in V : H \cdot v = \mu(H) v \text{ for all } H \in \mathfrak{h}\}.\] The adjoint representation of $\mathfrak{g}$ has non-zero weights called \defn{roots}, and we obtain the \defn{Cartan decomposition} \begin{equation}\mathfrak{g} = \mathfrak{h} \oplus \bigoplus_{\alpha \in \Phi^+} \mathfrak{g}_\alpha \oplus \bigoplus_{\alpha \in \Phi^-} \mathfrak{g}_\alpha,\label{eq:decomp}\end{equation} where $\Phi^+$ and $\Phi^-$ are the positive and negative roots, respectively. A \defn{simple root} is a positive root that cannot be written as the sum of two positive roots, and we write $\widetilde{\alpha}$ for the highest root. \begin{figure}[tbp] \begin{center} \begin{tabular}{cc} \parbox[t]{.3\textwidth}{ \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$A_n$}; \draw[thick] (0 cm,0) -- (4 cm,0); \draw[dotted,thick] (4 cm,0) -- (8 cm,0); \draw[thick] (8 cm,0) -- (10 cm,0); \draw[thick,solid,fill=white] (0cm,0) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,0) circle (.5cm); \draw[thick,solid,fill=white] (8cm,0) circle (.5cm); \draw[thick,solid,fill=white] (10cm,0) circle (.5cm); \end{tikzpicture} \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$B_n$}; \draw[thick] (0 cm,-.2) -- (2 cm,-.2); \draw[thick] (0 cm,.2) -- (2 cm,.2); \draw[thin] (1.3 cm,.6) -- (.8 cm,0); \draw[thin] (1.3 cm,-.6) -- (.8 cm,0); \draw[thick] (2 cm,0) -- (4 cm,0); \draw[dotted,thick] (4 cm,0) -- (8 cm,0); \draw[thick] (8 cm,0) -- (10 cm,0); \draw[thick,solid,fill=white] (0cm,0) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,0) circle (.5cm); \draw[thick,solid,fill=white] (8cm,0) circle (.5cm); \draw[thick,solid,fill=white] (10cm,0) circle (.5cm); \end{tikzpicture} \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$C_n$}; \draw[thick] (0 cm,-.2) -- (2 cm,-.2); \draw[thick] (0 cm,.2) -- (2 cm,.2); \draw[thin] (.7 cm,.6) -- (1.2 cm,0); \draw[thin] (.7 cm,-.6) -- (1.2 cm,0); \draw[thick] (2 cm,0) -- (4 cm,0); \draw[dotted,thick] (4 cm,0) -- (8 cm,0); \draw[thick] (8 cm,0) -- (10 cm,0); \draw[thick,solid,fill=white] (0cm,0) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,0) circle (.5cm); \draw[thick,solid,fill=white] (8cm,0) circle (.5cm); \draw[thick,solid,fill=white] (10cm,0) circle (.5cm); \end{tikzpicture} \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$D_n$}; \draw[thick] (0 cm,-1) -- (2 cm,0); \draw[thick] (0 cm,1) -- (2 cm,0); \draw[thick] (2 cm,0) -- (4 cm,0); \draw[dotted,thick] (4 cm,0) -- (8 cm,0); \draw[thick] (8 cm,0) -- (10 cm,0); \draw[thick,solid,fill=white] (0cm,-1) circle (.5cm); \draw[thick,solid,fill=white] (0cm,1) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,0) circle (.5cm); \draw[thick,solid,fill=white] (8cm,0) circle (.5cm); \draw[thick,solid,fill=white] (10cm,0) circle (.5cm); \end{tikzpicture}} & \parbox[t]{.3\textwidth}{ \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$E_6$}; \draw[thick] (0 cm,0) -- (8 cm,0); \draw[thick] (4 cm,0) -- (4 cm,2); \draw[thick,solid,fill=white] (0cm,0) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,0) circle (.5cm); \draw[thick,solid,fill=white] (6cm,0) circle (.5cm); \draw[thick,solid,fill=white] (8cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,2) circle (.5cm); \end{tikzpicture} \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$E_7$}; \draw[thick] (0 cm,0) -- (10 cm,0); \draw[thick] (6 cm,0) -- (6 cm,2); \draw[thick,solid,fill=white] (0cm,0) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,0) circle (.5cm); \draw[thick,solid,fill=white] (6cm,0) circle (.5cm); \draw[thick,solid,fill=white] (8cm,0) circle (.5cm); \draw[thick,solid,fill=white] (10cm,0) circle (.5cm); \draw[thick,solid,fill=white] (6cm,2) circle (.5cm); \end{tikzpicture} \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$E_8$}; \draw[thick] (0 cm,0) -- (12 cm,0); \draw[thick] (4 cm,0) -- (4 cm,2); \draw[thick,solid,fill=white] (0cm,0) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,0) circle (.5cm); \draw[thick,solid,fill=white] (6cm,0) circle (.5cm); \draw[thick,solid,fill=white] (8cm,0) circle (.5cm); \draw[thick,solid,fill=white] (10cm,0) circle (.5cm); \draw[thick,solid,fill=white] (12cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,2) circle (.5cm); \end{tikzpicture} \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$F_4$}; \draw[thick] (0 cm,0) -- (2 cm,0); \draw[thick] (2 cm,-.2) -- (4 cm,-.2); \draw[thin] (3.3 cm,.6) -- (2.8 cm,0); \draw[thin] (3.3 cm,-.6) -- (2.8 cm,0); \draw[thick] (2 cm,.2) -- (4 cm,.2); \draw[thick] (4 cm,0) -- (6 cm,0); \draw[thick,solid,fill=white] (0cm,0) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \draw[thick,solid,fill=white] (4cm,0) circle (.5cm); \draw[thick,solid,fill=white] (6cm,0) circle (.5cm); \end{tikzpicture} \begin{tikzpicture}[scale=.2] \draw (-1,0) node[anchor=east] {$G_2$}; \draw[thick] (0 cm,-.3) -- (2 cm,-.3); \draw[thick] (0 cm,.3) -- (2 cm,.3); \draw[thick] (0 cm,0) -- (2 cm,0); \draw[thin] (1.3 cm,.7) -- (.8 cm,0); \draw[thin] (1.3 cm,-.7) -- (.8 cm,0); \draw[thick,solid,fill=white] (0cm,0) circle (.5cm); \draw[thick,solid,fill=white] (2cm,0) circle (.5cm); \end{tikzpicture}} \end{tabular} \end{center} \caption{The Dynkin diagrams.} \label{fig:dynkin} \end{figure} The \defn{Killing form} is the nondegenerate symmetric bilinear form defined by \[B(X,Y)=\mathrm{tr}(\mathrm{ad}(X),\mathrm{ad}(Y)).\] We normalize the Killing form so that the norm of a long root is 2, and we will write this normalized form as $\langle \cdot,\cdot \rangle$. We write $\|\alpha\|^2:=\langle \alpha,\alpha \rangle$. Restricting the Killing form to $\mathfrak{h}$ and writing $\mathrm{dim}(\mathfrak{h})=n$ allows us to view weights and roots as points in $\mathbb{R}^n$, and the \defn{Weyl group} of $\mathfrak{g}$ is the reflection group $W$ generated by the reflections perpendicular to the roots $\alpha \in \Phi$. The \defn{coroot} of a root $\alpha$ is $\chk{\alpha}:=\frac{2\alpha}{\langle \alpha,\alpha \rangle}$. Let $\mathcal{A}[\mathfrak{h}^*]$ be the character ring of formal linear combinations of formal exponentials of weights. For a weight $\lambda\in\mathfrak{h}^*$, let \[A_{\lambda}=\sum_{w\in W}\op{sgn}(w)e^{w(\lambda)} \in \mathcal{A}[\mathfrak{h}^*].\] Then $A_{\lambda}$ is alternating with respect to the $W$-action on $\mathcal{A}[\mathfrak{h}^*]$, so that $w(A_{\lambda})=\op{sgn}(w)A_{\lambda},$ and it is also a $W$-alternating function of $\lambda$: $A_{w(\lambda)}=\op{sgn}(w)A_{\lambda}.$ A weight $\lambda$ is called \defn{regular} if $w(\lambda)\neq \lambda$ for all $e \neq w \in W$---in particular, $A_{\lambda}=0$ if $\lambda$ is not regular. It is called \defn{integral} if its inner product with every coroot is integral. The \defn{fundamental weights} are the dual basis to the simple coroots, and a weight is called \defn{dominant} if it is a nonnegative linear combination of the fundamental weights. Let $\rho$ denote the lowest regular dominant integral weight. \begin{theorem}[Weyl character formula] \label{thm:weyl} If $\lambda$ is an integral weight, there exists a unique $f_{\lambda}\in\mathcal{A}[\mathfrak{h}^*]$ with \[A_{\lambda}=f_{\lambda}A_{\rho}.\] Furthermore, if $\lambda$ is dominant, then \[f_{\lambda+\rho}=\sum_{\mu\in\mathfrak{h}^*}\mathrm{dim}(V_\lambda(\mu)) e^\mu\] is the character of the finite dimensional irreducible representation $V_\lambda$, where the multiplicity of $\mu$ in $V_\lambda$ is denoted $\mathrm{dim}(V_\lambda(\mu))$. \end{theorem} \subsection{Casimir Elements and the Universal Enveloping Algebra} \label{sec:casimir} The Harish-Chandra isomorphism is an isomorphism between the center of the universal enveloping algebra of $\mathfrak{g}$, $Z(U(\mathfrak{g}))$, and $W$-invariant polynomials $S(\mathfrak{h})^W$. By the Shephard-Todd-Chevelley theorem, $S(\mathfrak{h})^W$ is a polynomial algebra with $n$ generators, and the degrees $d_1,d_2,\ldots,d_n$ of these generators play an important numerological role: for example, the highest degree is the \defn{Coxeter number} $h$ (the order of a Coxeter element of $W$), the dimension of the Lie algebra is $\mathrm{dim}(\mathfrak{g})=n(h+1)$, the number of reflections in $W$ is $\sum_{i=1}^n (d_i-1)$, and the number of elements in $W$ is $|W|=\prod_{i=1}^n d_i$. See also~\Cref{sec:coxeter_cumulants} for further numerology. We call an element of $Z(U(\mathfrak{g}))$ a \defn{Casimir element}---the Harish-Chandra isomorphism combined with the Shephard-Todd-Chevelley theorem shows that there are $n$ algebraically independent Casimir elements. Special emphasis is given to the Casimir element of degree two, which may be defined as follows: fixing any basis $\{X_i\}_{i=1}^{\mathrm{dim}(\mathfrak{g})}$, we obtain a dual basis $\{X^i\}_{i=1}^{\mathrm{dim}(\mathfrak{g})}$ using the Killing form, and then define \begin{equation}\Omega = \sum_{i=1}^{\mathrm{dim}(\mathfrak{g})} X_iX^i \in Z(U(\mathfrak{g})).\label{eq:casimir}\end{equation} As representations of $\mathfrak{g}$ coincide with modules for its universal enveloping algebra, Schur's lemma implies that since $\Omega$ is in the center of $U(\mathfrak{g})$, it acts as a scalar on any highest weight representation of $\mathfrak{g}$. The following well-known theorem explicitly identifies this scalar. \begin{theorem} Let $\lambda$ be a dominant weight. Then $\Omega$ acts as multiplication by $\langle\lambda,\lambda+2\rho\rangle$ on $V_\lambda$. \label{thm:casimir_action} \end{theorem} \begin{proof} Using the Cartan decomposition of~\Cref{eq:decomp}, write $\Omega = \sum_{1 \leq i \leq n} H_i H^*_i + \sum_{\alpha \in \Phi} E_\alpha E_{-\alpha}$, where we have chosen $\langle E_\alpha,E_{-\alpha}\rangle = 1$ and $[E_\alpha,E_{-\alpha}]=H_\alpha.$ In the representation $V_\lambda$, we compute \begin{align*}\Omega &= \sum_{1 \leq i \leq n} H_i H^*_i + \sum_{\alpha \in \Phi} E_\alpha E_{-\alpha} = \sum_{1 \leq i \leq n} H_i H^*_i + 2\sum_{\alpha \in \Phi^+} E_{-\alpha} E_{\alpha} + \sum_{\alpha \in \Phi^+} [E_\alpha, E_{-\alpha}] \\ &= \sum_{1 \leq i \leq n} H_i H^*_i + 2\sum_{\alpha \in \Phi^+} E_{-\alpha} E_{\alpha} + \sum_{\alpha \in \Phi^+} H_\alpha. \end{align*} Acting on a highest weight vector in $V_\lambda$, the term $ 2\sum_{\alpha \in \Phi^+} E_{-\alpha} E_{\alpha}$ vanishes, leaving only $\sum_{1 \leq i \leq n} \lambda(H_i) \lambda(H^*_i) + \lambda\left(\sum_{\alpha \in \Phi^+} H_\alpha\right)$. The first term is now computed as $\langle \lambda,\lambda \rangle$, while the second term gives $\langle \lambda,2\rho\rangle$ for $\rho$ the half sum of the positive roots. \end{proof} \section{Motivation: Cores and Ehrhart Theory} \label{sec:motivation} In this section we relate a special case of~\Cref{thm:main_thm} (for the first fundamantal weight in type $A$) to the study of simultaneous core partitions. \subsection{Simultaneous Cores and Armstrong's Conjecture} An \defn{$a$-core} is an integer partition with no hook-length of size $a$. The study of simultaneous $(a,b)$-cores---that is, partitions that are both $a$-cores and $b$-cores---is a topic that has recently seen quite a lot of interest from the combinatorics community~\cite{stanley2013catalan,aggarwal2014armstrong}. When $\gcd(a,b)=1$, Anderson proved that the number of $(a,b)$-cores has the simple expression \[\left|{\sf core}(a,b)\right| = \frac{1}{a+b}\binom{a+b}{b}\] by giving a bijection to Dyck paths in an $a \times b$ rectangle~\cite{anderson2002partitions}. It is well-known that the dominant alcoves in the affine symmetric group $\widetilde{S}_a$ are naturally indexed by $a$-cores, and in this language Anderson's result had previously been proven in the generality of affine Weyl groups by both Haiman and Suter~\cite{haiman1994conjectures,suter1998number}. \medskip While investigating the interpretation of $q,t$-statistics and the zeta map using the affine symmetric group~\cite{armstrong2011conjecture,armstrong2014results}, Armstrong was led to conjecture that the expected number of boxes of a simultaneous core (its ``${\sf size}$'') had a beautiful formula. \begin{theorem}[{Amstrong (conjectured), Johnson (proof)~\cite{johnson2015lattice}}] \label{eq:drew_exp} The expected number of boxes of a simultaneous core is given by \[\Expt{\lambda \in {\sf core}(a,b)}{{\sf size}(\lambda)} = \frac{(a-1)(b-1)(a+b+1)}{24}.\] \end{theorem} \begin{example} We compute the expected number of boxes for the five simultaneous $(3,4)$-cores \[ \emptyset,\hspace{1em} {\tiny \yng(1)},\hspace{1em} {\tiny \yng(2)},\hspace{1em} {\tiny \yng(1,1)},\hspace{1em} {\tiny \yng(3,1,1)},\] as \[\frac{1}{5}\left(0+1+2+2+5\right)=2=\frac{(3-1)(4-1)(3+4+1)}{24}.\] \end{example} \Cref{eq:drew_exp} was first proven by Johnson using Ehrhart theory~\cite{johnson2015lattice}. Building on Johnson's approach in \cite{thiel2017strange}, we showed that the statistic ${\sf size}$ could be interpreted as a slight modification of the natural norm on the weight space (see~\Cref{fig:weights_and_reps}), and generalized the result to all simply-laced affine Weyl groups (we now have a generalization to all affine Weyl groups): \begin{align} {\sf size}_b(x):=\frac{h}{2}\left\|x-b\frac{\chk{\rho}}{h}\right\|^2-\frac{h}{2}\left\|\frac{\chk{\rho}}{h}\right\|^2. \label{def:size} \end{align} Briefly, by composing the bijection between $a$-cores and dominant alcoves in $\widetilde{S}_a$, and the natural bijection between dominant alcoves and coroots, one obtains a bijection between simultaneous $(a,b)$-cores and $\chk{Q} \cap b \mathcal{A}$---coroot points inside a $b$-fold dilation of the fundamental alcove in $\widetilde{S}_a$. When $a$ is coprime to $b$, the cyclic symmetry of the affine Dynkin diagram gives rise to an affine isometry that partitions the weights inside $b \mathcal{A}$ into regular orbits, each of which contains a single coroot. It is therefore enough to consider the \emph{coweights} in the dilation of the fundamental alcove $\chk{\Lambda} \cap b \mathcal{A}$, and one may then apply Ehrhart theory to prove (generalizations of) Armstrong's conjecture. \subsection{Simultaneous Cores and Highest Weight Representations} As motivation for~\Cref{thm:main_thm}, we wish to show that the problem of computing the expected number of boxes in a simultaneous $(a,b)$-core is roughly equivalent to computing the expected norm of a weight in a particular highest weight representation. Having already related cores and $\chk{\Lambda} \cap b \mathcal{A}$, we now wish to find a relation to representations. In $\mathfrak{sl}_a$, there is a bijection between coweights inside the $b$-fold dilation of the fundamental alcove $b\mathcal{A}$, and coweights in the highest weight representation $\mathfrak{sl}_a(b\omega_1)$, where $\omega_1$ is the first fundamental weight---indeed, both are counted by the binomial coefficient $\binom{a+b}{b}$. This bijection may be described as follows, and is illustrated in~\Cref{fig:weights_and_reps}. We first center $b\mathcal{A}$ around the origin by sending \[x \mapsto x- \frac{b\chk{\rho}}{h}.\] There is a natural bijection between the coweight and coroot lattice defined by \begin{align*} \chk{\Lambda} &\to \chk{Q}\\ x &\mapsto (1-c)x,\end{align*} where $c=(a,a{-}1,\ldots,1) \in \mathfrak{S}_a$ is a long cycle. This reflects the fact that $(1-c)$ is conjugate to the Cartan matrix, with determinant equal to the index of connection $|\chk{Q}/\chk{\Lambda}|$. \begin{proposition} \label{prop:bij} The composition of these two maps gives the desired bijection \normalfont \begin{align*}\phi: \chk{\Lambda} \cap b\mathcal{A} &\to \mathfrak{sl}_a(b\omega_1) \\ x &\mapsto (1-c)\left(x - \frac{b\chk{\rho}}{h}\right) \end{align*} \end{proposition} \begin{proof} The vertices of the polytope $b\mathcal{A}$ are $\{0\}\cup \{b\omega_i\}_{i=1}^{n-1}$, and $\chk{\Lambda} \cap b\mathcal{A}$ is (by definition) all coweight points inside the convex hull of those vertices. On the other hand, the coweights in $\mathfrak{sl}_a(b\omega_1)$ are exactly those coweights in the convex hull of $\{c^i (b\omega_1)\}_{i=0}^{n-1}$ whose difference from $b\omega_1$ is in the coroot lattice. It therefore suffices to check that the map $(1-c)\left(x - \frac{b\chk{\rho}}{h}\right)$ takes the vertices $\{0\}\cup \{\omega_i\}_{i=1}^{n-1}$ to the vertices $\{c^i (b\omega_1)\}_{i=0}^{n-1}$. But this is a simple computation---writing $e_i$ for the usual basis of $\mathbb{R}^n$ so that $\omega_i = \sum_{j=1}^i e_j - \frac{i}{n}\sum_{j=1}^n e_j$ and $\chk{\rho}=\sum_{i=1}^{n-1} \omega_i$, we check \begin{align*}(1{-}c)\left(b\omega_i - \frac{b\chk{\rho}}{h}\right) &= \frac{b(1{-}c)}{n}\left(\sum_{j=1}^i \left(\frac{n{-}1}{2}-i+j\right)e_j-\sum_{j=i+1}^n \left(\frac{n{+}1}{2}+i-j\right)e_j\right) \\ &= \frac{b}{n} \left((n-1)e_i+\sum_{j\neq i} -e_j\right)=c^i(b \omega_1). \end{align*} \end{proof} \subsection{Equivalence of~\Cref{eq:drew_exp} and~\Cref{thm:main_thm} for $\mathfrak{sl}_a(b\omega_1)$} Under the bijection of~\Cref{prop:bij}, we show that computing the expected ${\sf size}_b$ on $b\mathcal{A}$ (and hence the expected number of boxes in an $(a,b)$-core) is roughly equivalent to computing the expected norm on $\mathfrak{sl}_a(b\omega_1)$. Write $x_b = x-b\frac{\rho}{h}$ and compute: \begin{align*} \Expt{\mu \in \mathfrak{sl}_a(b\omega_1)}{\langle \mu,\mu\rangle} &= \frac{1}{\binom{a+b}{b}}\sum_{\mu \in \mathfrak{sl}_a(b\omega_1)} \| x\|^2 = \frac{1}{\binom{a+b}{b}}\sum_{\mu \in b\mathcal{A}} \left\|(1-c)x_b\right\|^2\\ &=\frac{1}{\binom{a+b}{b}}\sum_{\mu \in b\mathcal{A}}\left( 2\left\|x_b\right\|^2-2\left\langle cx_b,x_b\right\rangle\right)\\ &=\frac{a^2-1}{6a}+\frac{1}{\binom{a+b}{b}}\left(\frac{4}{a}\sum_{\mu \in b\mathcal{A}}{\sf size}_b(x)-2\sum_{\mu \in b\mathcal{A}}\left\langle cx_b,x_b\right\rangle\right) \end{align*} It is slightly surprising, but follows from an Ehrhart-theoretic computation of roughly the same degree of difficulty as the computation of the expectation of ${\sf size}_b$ (the only difficulty arising from a repeated root), that \[\frac{1}{\binom{a+b}{b}} \sum_{\mu \in b\mathcal{A}}\left\langle cx_b,x_b\right\rangle = \frac{(a-5)(a-1)b(a+b)}{12a(a+1)},\] so that \begin{align*} \Expt{\mu \in \mathfrak{sl}_a(b\omega_1)}{\langle \mu,\mu\rangle} &= \frac{a^2-1}{6a} + \frac{4}{a} \frac{(a-1)(b-1)(a+b+1)}{24} - 2 \frac{(a-5)(a-1)b(a+b)}{12a(a+1)} \\ &= \frac{(a-1)b(a+b)}{a(a+1)}. \end{align*} On the other hand---since the representation is multiplicity-free---we could have applied Ehrhart theory directly to compute the expected norm of a weight in $\mathfrak{sl}_a(b\omega_1)$. We conclude that computing $\Expt{\mu \in \mathfrak{sl}_a(b\omega_1)}{\langle \mu,\mu\rangle}$ is roughly equivalent to computing the expectation of size on simultaneous $(a,b)$-cores. \medskip Thus, given the success of studying moments of norms of weights in $b\mathcal{A}$, we found it a reasonable extension to ask for the expected norm of a weight in a highest weight representation. \begin{figure}[htbp] \begin{center} \raisebox{-0.6\height}{ \begin{tikzpicture}[scale=1.3] \draw[->,black,thin] (0,0) -- (0,1.41421); \draw[->,black,thin] (0,0) -- (1.2247,-.70711); \draw[black,thin] (-2.8,0) -- (3.8,0); \draw[black,thin] (1.4,2.4249) -- (-1.9,-3.291); \draw[black,thin] (-1.9,3.291) -- (1.4,-2.4249); \tikzmath{\l = -1.225+.8165*4+.40825*0;}; \tikzmath{\k = -.7071+.7071*0;}; \tikzmath{m = -1.225+.8165*0+.40825*4;}; \tikzmath{\n = -.7071+.7071*4;}; \tikzmath{\o = -1.225+.8165*0+.40825*0;}; \tikzmath{\p = -.7071+.7071*0;}; \draw[gray!60,thin] (\l,\k) -- (m,\n) -- (\o,\p) -- (\l,\k); \tikzmath{\l = 3.266-1.225*0;}; \tikzmath{\k = 1.414*0-.7071*0;}; \tikzmath{m = 3.266-1.225*4;}; \tikzmath{\n = 1.414*4-.7071*4;}; \tikzmath{\o = 3.266-1.225*4;}; \tikzmath{\p = 1.414*0-.7071*4;}; \draw[black,thin] (\l,\k) -- (m,\n) -- (\o,\p) -- (\l,\k); \foreach \i in {0,...,4}{ \tikzmath{\ii=4-\i;}; \foreach \j in {0,...,\ii}{ \tikzmath{\l = -1.225+.8165*\i+.40825*\j;}; \tikzmath{\k = -.7071+.7071*\j;}; \filldraw[color=gray!60, fill=white, very thick] (\l,\k) circle[radius=2pt]; }; }; \foreach \i in {0,...,4}{ \foreach \j in {\i,...,4}{ \tikzmath{\l = 3.266-1.225*\j;}; \tikzmath{\k = 1.414*\i-.7071*\j;}; \filldraw (\l,\k) circle[radius=1.3pt]; }; }; \node at (-1.225,-.83) (a) {$-\rho$}; \node at (.40825,.2357) (b) {$\rho/h$}; \node at (.8,-.2) (c) {$\omega_1$}; \node at (.3,.9) (d) {$\omega_2$}; \node at (3.266,-.2) (a) {$4\omega_1$}; \node[black] at (1.47,-.85) (b) {$\alpha_1$}; \node[black] at (0,1.7) (c) {$\alpha_2$}; \end{tikzpicture}} \end{center} \caption{The weights inside a $4$-fold dilation of the fundamental alcove in $\mathfrak{sl}_3$ are drawn as gray circles inside the gray triangle, while the weights in the representation $V_{4\omega_1}$ are drawn as black disks. The statistic ${\sf size}$ on the weights in the dilation of the fundamental alcove is a quadratic form that is a slight modification of the norm.} \label{fig:weights_and_reps} \end{figure} \section{Proof of~\Cref{thm:main_thm} using the Weyl Character Formula} In this section we use the polynomial method and the Weyl character formula (\Cref{thm:weyl}) to give an elementary, uniform proof of~\Cref{thm:main_thm} in all types. We first prove polynomiality of sums of polynomial functions in the weights in $V_\lambda$. \begin{theorem}\label{avgpoly} Suppose that $\lambda$ is a dominant integral weight of $\mathfrak{h}$ and let $V_{\lambda}$ be the finite dimensional irreducible representation of $\mathfrak{g}$ of highest weight $\lambda$. For a weight $\mu$, let $\mathrm{dim}(V_\lambda(\mu))$ be the multiplicity of $\mu$ in $V_\lambda$. Let $P\in S(\mathfrak{h})^W$ be a $W$-invariant polynomial on $\mathfrak{h}^*$ of degree $d$. Then \[S(P,\lambda):=\frac{1}{\op{dim}(V_\lambda)}\sum_{\mu\in\mathfrak{h}^*}\mathrm{dim}(V_\lambda(\mu))P(\mu)\] is a polynomial in $\lambda$ of at most degree $d$. It is a $W$-invariant polynomial in variables given by $\lambda+\rho$. \end{theorem} \begin{proof} We follow the ideas and notation of the derivation of the Weyl dimension formula from the Weyl character formula in \cite[Section 7.1.2]{goodman2009symmetry}. We write $f_{\lambda+\rho}=\frac{A_{\lambda+\rho}}{A_{\rho}}$. We define the linear differential operator \begin{align*}N:\mathcal{A}[\mathfrak{h}^*]&\rightarrow \mathcal{A}[\mathfrak{h}^*] \\ N(e^{\lambda})&=P(\lambda)e^\lambda,\end{align*} and the linear evaluation map \begin{align*}\epsilon:\mathcal{A}[\mathfrak{h}^*]&\rightarrow\mathbb{C}\\ \epsilon(e^\lambda)&=1.\end{align*} Then we have that \[\frac{\epsilon(N(f_{\lambda+\rho}))}{\epsilon(f_{\lambda+\rho})}=\frac{\sum_{\mu\in\mathfrak{h}^*}\mathrm{dim}(V_\lambda(\mu))P(\mu)}{\sum_{\mu\in\mathfrak{h}^*}\mathrm{dim}(V_\lambda(\mu))}=S(P,\lambda).\] Write $A_\rho=\prod_{\alpha\in\Phi^+}(e^{\alpha/2}-e^{-\alpha/2})$. Now $N$ is a differential operator of degree $d$, so the quotient rule implies that \begin{align*} N(f_\lambda)&=N\left(\frac{A_\lambda}{A_\rho}\right)=N\left(\frac{A_\lambda}{\prod_{\alpha\in\Phi^+}(e^{\alpha/2}-e^{-\alpha/2})}\right)\\ &=\sum \frac{B_{\lambda,\mathbf{m}}}{\prod_{\alpha\in\Phi^+}(e^{\alpha/2}-e^{-\alpha/2})^{m_\alpha+1}}, \end{align* where the sum is over all $\mathbf{m}=(m_\alpha)\in[d]^{\Phi^+}$ such that $\sum_{\alpha\in\Phi^+}m_\alpha\leq d$, and where $B_{\lambda,\mathbf{m}}\in\mathcal{A}[\mathfrak{h}^*]$ has coefficients that are polynomials in $\lambda$ of degree at most $d-\sum_{\alpha\in\Phi^+}m_\alpha$. Using L'H\^{o}pital's rule we see that $\epsilon(N(f_\lambda))$ is a polynomial in $\lambda$ of degree at most $|\Phi^+|+d$. It is alternating in $\lambda$. In particular, $\epsilon(N(f_\lambda))=0$ if $\lambda$ is not regular. So for every $\alpha\in\Phi^+$ the linear factor $\langle\lambda,\alpha\rangle$ divides the polynomial $\epsilon(N(f_\lambda))$. Consider the special case where $P=1$, so that $N$ is the identity and $d=0$. Then this implies that $\epsilon(N(f_\lambda))=\epsilon(f_\lambda)=C\prod_{\alpha\in\Phi^+}\langle\lambda,\alpha\rangle$ for a constant $C$. We have that $f_\rho=e^0$ so that $C=\frac{1}{\prod_{\alpha\in\Phi^+}\langle\rho,\alpha\rangle}$. For any $W$-invariant polynomial $P\in S(\mathfrak{h})^W$, the polynomial $\epsilon(f_\lambda)$ therefore divides the polynomial $\epsilon(N(f_\lambda))$, so that the quotient $\frac{\epsilon(N(f_\lambda))}{\epsilon(f_\lambda)}$ is a polynomial in $\lambda$ of degree at most $d$. It is $W$-invariant, since both $\epsilon(N(f_\lambda))$ and $\epsilon(f_\lambda)$ are alternating in $\lambda$. Thus $S(\lambda,P)=\frac{\epsilon(N(f_{\lambda+\rho}))}{\epsilon(f_{\lambda+\rho})}$ is given by a $W$-invariant polynomial of degree at most $d$ in $\lambda+\rho$. \end{proof} \begin{example} Consider $\mathfrak{g}=\mathfrak{sl}_2$ and $P=\|\cdot\|^2$, where $\|\cdot \|^2$ is given by the $\mathfrak{S}_2$-equivariant polynomial $\|(x,-x)\|^2 = 2x^2$. Since the dominant weights are given by $\lambda = (m,-m)$ for $m \in \mathbb{N}$, and since the weights in $V_\lambda$ are exactly of the form $(m,-m),(m-2,-m+2),\ldots,(-m,m)$ (with no multiplicity), we compute that \begin{align*}S(\|\cdot \|^2,\lambda) = \frac{2}{m+1} \sum_{i=0}^m (m-2i)^2 = \frac{2}{3} m(m+2) &= \frac{1}{3}\left(2(m+1)^2-2\right),\\ &=\frac{1}{h+1}(\|\lambda+\rho\|^2-\|\rho\|^2), \end{align*} which is $\mathfrak{S}_2$-equivariant as a polynomial in $m+1$. \end{example} \begin{proof}[First proof of~\Cref{thm:main_thm}] For this proof, we use the normalization of the Killing form that $\|\widetilde{\alpha} \|^2=\frac{1}{g}$, where $g$ is the dual Coxeter number. By Theorem \ref{avgpoly}, $S(\|\cdot\|,\lambda)$ is a $W$-invariant polynomial of degree at most $2$ in $\lambda+\rho$ so that $S(\|\cdot\|^2,\lambda)=a+b\|\lambda+\rho\|^2$ for some $a,b\in{\mathbb{C}}$. We have that $S(\|\cdot\|^2,0)=0$, so $a=-b\|\rho\|^2$. Furthermore, if $\widetilde{\alpha}\in\Phi^+$ is the highest root, then $V_{\widetilde{\alpha}}$ is the adjoint representation of $\mathfrak{g}$, so we get \[S(\|\cdot\|^2,\widetilde{\alpha})=\frac{1}{n(h+1)}\sum_{\alpha\in\Phi}\|\alpha\|^2=\frac{1}{h+1}\] using $\sum_{\alpha\in\Phi}\|\alpha\|^2=n$ for the Killing form \cite{brown1964remark}. So $\frac{1}{h+1}=b(\|\widetilde{\alpha}+\rho\|^2-\|\rho\|^2)=b$, using that $\|\widetilde{\alpha}+\rho\|^2-\|\rho\|^2$ is the Casimir eigenvalue on the adjoint representation and therefore equals $1$. We conclude that \[S(\|\cdot\|^2,\lambda)=\frac{1}{h+1}(\|\lambda+\rho\|^2-\|\rho\|^2).\qedhere\] \end{proof} \section{A Combinatorial Proof in Type $A$} \label{sec:typea} In this section, we make the polynomiality argument of the previous section more concrete using the combinatorics of the representation theory of $\mathfrak{sl}_n$. Fix $\mathfrak{g}=\mathfrak{sl}_{n}$ and $\omega_i=\sum_{j=1}^i e_i$. In $\mathfrak{sl}_n$, dominant weights of $\mathfrak{h}$ may be parametrized as integer partitions \[\lambda=[\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_n] \vdash m,\] where parts $\lambda_i$ may be equal to zero. Fix a highest weight $\lambda$, and write \begin{align*} \overline{\lambda_i} = \lambda_i - \frac{|\lambda|}{n}, \text{ } \overline{\lambda}=\left[\overline{\lambda_1}\geq \cdots \geq \overline{\lambda_n}\right] \text{, and } \rho = \left[n-1,n-2,\ldots,1,0\right]. \end{align*} With these conventions, weights $\mu$ in the highest weight representation $\mathfrak{sl}_{n}(\lambda)$ may be thought of as certain points in $\mathbb{R}^n$ with positive entries and sum equal to $m$. Combinatorially, the multiplicity of $\mu$ in $\mathfrak{sl}_{n}(\lambda)$ is well-known to be given by the number of semistandard tableaux of shape $\lambda$ on the alphabet $[n]$ with content $\mu$; $m$ is just the number of boxes in the Ferrers shape $\lambda$: \[\mathrm{ch}\left(\mathfrak{sl}_{n,\lambda}\right) = s_\lambda(x_1,\ldots,x_n)= \sum_{\substack{T \text{ semistandard} \\ \text{ of shape }\lambda}} \mathbf{x}^T,\] where $\mathbf{x}^T = \prod_{i=1}^n x_i^{\left|\{i \in T\}\right|}$ and $s_\lambda(x_1,\ldots,x_n)$ is a Schur polynomial. As a simple consequence of this combinatorial description of the character, we have Weyl's ``interlacing'' multiplicity-free formula for the branching of the representation $\mathfrak{sl}_{n,\lambda}$ to $\mathfrak{sl}_{n-1}$:\[\mathfrak{sl}_{n,\lambda} = \bigoplus_{\mu} \mathfrak{sl}_{n-1,\mu}, \text{ where } \lambda_1 \geq \mu_1 \geq \lambda_2 \geq \cdots \geq \mu_{n-1} \geq \lambda_n.\] By symmetry of the Schur function, and since this branching rule exactly peels off the boxes containing the entry $n$ (corresponding to the value of the coordinate $x_n$), one could imagine using this formula to determine the norm by computing \[\sum_{\mu \in V_\lambda} \mathrm{dim}\left(\mathfrak{sl}_{n,\lambda}(\mu)\right) \langle \mu,\mu\rangle = n\sum_{\mu} \mathrm{dim}\left(\mathfrak{sl}_{n-1,\mu}\right)\left(|\lambda|-|\mu|\right)^2.\] We do not follow this approach here, but instead isolate the boxes containing the entry $n$ using the Pieri rule and an inclusion-exclusion argument. \begin{theorem} Let $\mathfrak{g}=\mathfrak{sl}_{n}$. Suppose that $\lambda$ is a dominant weight of $\mathfrak{h}$ and let $\mathfrak{sl}_{n,\lambda}$ be the finite dimensional irreducible representation of $\mathfrak{sl}_n$ of highest weight $\lambda$. Then \[\frac{1}{\mathrm{dim}(\mathfrak{sl}_{n,\lambda})} \sum_{\mu \in \mathfrak{sl}_{n}(\lambda)} \mathrm{dim}(\mathfrak{sl}_{n,\lambda}(\mu)) \langle \overline{\mu},\overline{\mu}\rangle = \frac{1}{h+1} \big\langle \overline{\lambda}, \overline{\lambda + 2\rho}\big\rangle.\] \end{theorem} \begin{proof}[Second proof of~\Cref{thm:main_thm}, valid in type $A$] Some care is needed when we compute the length of $\mu \in \mathfrak{sl}_{n}(\lambda)$---we wish to compute the length of the \emph{normalized} weight $\overline{\mu}$. Of course, there is a simple relationship between the length of $\mu$ and of $\overline{\mu}$: $\langle \overline{\mu},\overline{\mu} \rangle = \langle \mu,\mu\rangle- \frac{m^2}{n},$ where $m=\langle \mu, [1]^n \rangle$ (constant for all $\mu \in \mathfrak{sl}_{n}(\lambda)$). We may therefore compute with unnormalized weights using the relationship \begin{align*}\frac{1}{\mathrm{dim}(\mathfrak{sl}_{n,\lambda})} \sum_{\mu \in \mathfrak{sl}_{n,\lambda}} \mathrm{dim}(\mathfrak{sl}_{n,\lambda}(\mu)) \langle \overline{\mu},\overline{\mu}\rangle =& -\frac{m^2}{n}\\&+\frac{1}{\mathrm{dim}(\mathfrak{sl}_{n,\lambda})} \sum_{\mu \in \mathfrak{sl}_{n,\lambda}} \mathrm{dim}(\mathfrak{sl}_{n,\lambda}(\mu)) \langle \mu,\mu\rangle.\end{align*} Define $n$ new partitions \[\lambda^{(i)} = \left[\lambda_1+1 \geq \lambda_2+1 \geq \cdots \geq \lambda_{i-1}+1 \geq \lambda_{i+1} \geq \cdots \geq \lambda_n\right] \text{ for } 1 \leq i \leq n.\] Using the fact that Schur polynomials in the variables $x_i$ are symmetric, conditioning on which boxes of $\lambda$ contain the entry $n$, and using the Pieri rule allows us to write \begin{multline*}\frac{1}{\mathrm{dim}(\mathfrak{sl}_{n,\lambda})}\sum_{\mu \in \mathfrak{sl}_{n,\lambda}} \mathrm{dim}(\mathfrak{sl}_{n,\lambda}(\mu)) \langle \overline{\mu},\overline{\mu}\rangle= -\frac{m^2}{n}+\\+\overbrace{\frac{n}{\mathrm{dim}(\mathfrak{sl}_{n,\lambda})}}^{\substack{\text{Schur polynomial}\\ \text{symmetry}}} \sum_{j=1}^n \overbrace{(-1)^{j+1}}^{\substack{\text{inclusion-}\\\text{exclusion}}} \Big( \underbrace{s_{\lambda^{(j)}}([1]^{n-1}) \sum_{i=0}^{\lambda_j-(j-1)} h_i([1]^{n-1})}_{\text{Pieri rule; leftover boxes contain $n$}} \underbrace{\left(\lambda_j-(j-1)-i\right)^2}_{\substack{\text{contribution to norm}\\ \text{ of boxes containing $n$}}} \Big), \end{multline*} where the alternating sum reflects an inclusion-exclusion argument that removes the over-count of those partitions that aren't contained in $\lambda$. The point is that the numerator has now been expressed as a polynomial. \begin{example} As in \Cref{fig:example1}, let $\lambda = (2,1)$ and $n=3$. We consider all eight semistandard tableaux of shape $(2,1)$ with entries at most 3: \[ \begin{tabular}{c}\young(13,3)\\[\smallskipamount] \young(23,3)\end{tabular} \hspace{3em} \begin{tabular}{c} \young(11,3) \\[\smallskipamount] \young(12,3) \\[\smallskipamount] \young(22,3) \\[\medskipamount] \young(13,2) \end{tabular} \hspace{3em} \begin{tabular}{c}\young(11,2) \\[\smallskipamount] \young(12,2) \end{tabular}.\] By symmetry of the Schur function, \[\sum_T \sum_{i=1}^3 \left(\text{number of $i$s in T}\right)^2 = 3\sum_T \left(\text{number of $3$s in T}\right)^2,\] and so we should group the tableaux by the number of boxes containing the entry $n=3$. The sum above uses the Pieri rule to do this, expressing these eight tableaux as \begin{center} \begin{tabular}{c}\young(1\bullet,\bullet)\\[\smallskipamount] \young(2\bullet,\bullet) \\[\smallskipamount] \hline $s_1 \cdot h_0 \cdot 2^2$ \end{tabular} \hspace{1.5em}+\hspace{1.5em} \begin{tabular}{c} \young(11,\bullet) \\[\smallskipamount] \young(12,\bullet) \\[\smallskipamount] \young(12,\bullet) \\[\medskipamount] \young(1\bullet,2) \\[\smallskipamount]\hline $s_1 \cdot h_1 \cdot 1^2$ \end{tabular} \hspace{1.5em}+\hspace{1.5em} \begin{tabular}{c}\young(11,2) \\[\smallskipamount] \young(12,2) \\[\smallskipamount] \young(111) \\[\smallskipamount] \young(112) \\[\smallskipamount] \young(122) \\[\smallskipamount] \young(222) \\[\smallskipamount] \hline $s_1 \cdot h_2 \cdot 0^2$ \end{tabular} \hspace{1.5em}-\hspace{1.5em} \begin{tabular}{c} \young(111) \\[\smallskipamount] \young(112) \\[\smallskipamount] \young(122) \\[\smallskipamount] \young(222) \\[\smallskipamount] \hline $s_3 \cdot h_0 \cdot 0^2$ \end{tabular}. \end{center} \end{example} We have the evaluations of the Schur and homogeneous functions at $x_i=1$ \begin{align} s_\lambda ([1]^n) = \prod_{1 \leq i < j \leq n} \frac{\lambda_i-\lambda_j+j-i}{j-i} \text{ and } h_i ([1]^n) = \binom{n+i-1}{i}. \end{align} Dividing by $\mathrm{dim}(\mathfrak{sl}_{n}(\lambda))=s_{\lambda}([1]^n)$, using the formulas above, performing the obvious cancellations, and explicitly evaluating the sums, we obtain \begin{multline}\frac{1}{\mathrm{dim}(\mathfrak{sl}_{n,\lambda})}\sum_{\mu \in \mathfrak{sl}_{n,\lambda}} \mathrm{dim}(\mathfrak{sl}_{n,\lambda}(\mu)) \langle \overline{\mu},\overline{\mu}\rangle=\\ =-\frac{m^2}{n}+n! \sum_{j=1}^n\left(\prod_{\substack{1 \leq i \leq n \\ i\neq j}} \frac{1}{\lambda_j-\lambda_i+i-j}\right) \sum_{i=0}^{\lambda_j-(j-1)} \binom{n+i-2}{i} \left(\lambda_j-(j-1)-i\right)^2\\ =-\frac{m^2}{n}+\frac{n!}{n+1}\sum_{j=1}^n \left(\prod_{\substack{1 \leq i \leq n \\ i\neq j}} \frac{1}{\lambda_j-\lambda_i+i-j}\right) \binom{n+\lambda_j-j}{\lambda_j-j}\left(n+2(\lambda_j-j)+1\right).\label{eq:rhs} \end{multline} On the other hand, \begin{align} \frac{1}{h+1}\langle \overline{\lambda}, \overline{\lambda+2\rho}\rangle &= \frac{1}{n+1}\left\langle \lambda - \frac{m}{n}[1]^n, \lambda- \frac{m}{n}[1]^n+2\rho - (n-1)[1]^n \right \rangle \\ &=-\frac{m^2-mn+n^2}{n(n+1)}+\frac{1}{n+1}\left(\sum_{i=1}^n\lambda_i^2 +2(n-i)\lambda_i\right)\label{eq:lhs} \end{align} Setting~\Cref{eq:lhs,eq:rhs} equal, multiplying by $(n+1)$, pushing the constants to one side, and writing \begin{align*} x_j = \lambda_j - j \text{ and } P(x_j)= \left(n+2x_j+1\right) \prod_{i=1}^n (x_j+i) \end{align*} we must check that \begin{align} 2\left(\sum_{1\leq i \leq j \leq n} x_ix_j\right) + (1 + n)^2 \left(\sum_{i=1}^n x_i\right)+\frac{n (n + 1) (3n^2+5n+4)}{12}\label{eq:lhs2}\\ = \sum_{j=1}^n \left(\prod_{\substack{1 \leq i \leq n \\ i\neq j}} \frac{1}{x_j-x_i}\right) P(x_j).\label{eq:rhs2} \end{align} Treating the $x_i$ as formal variables, \Cref{eq:lhs2} is clearly a symmetric polynomial of degree $2$ in the $x_j$. For any $1 \leq i < j \leq n$, the $i$th and $j$th terms of the sum in \Cref{eq:rhs2} are the only two terms with $x_i-x_j$ in the denominator---and the sum of these terms is multiplied by $P(x_j)-P(x_i)$. This is true for any $i$ and $j$, and so all residues cancel. We conclude that since $P(x_j)$ has degree $n+1$, \Cref{eq:rhs2} is also a symmetric polynomial in the $x_j$ of degree at most $2$. It remains to confirm that these are the same degree $2$ symmetric function Setting $x_{i}=-i$ for $1 \leq i \leq n-2$, every term except the last two vanish in \Cref{eq:rhs2}. These remaining terms simplify to \begin{align*} &\left(\frac{n+2x_n+1}{x_n-x_{n-1}} \frac{\prod_{i=1}^n x_n+i}{\prod_{i=1}^{n-2}x_n+i}\right) +\left( \frac{n+2x_{n-1}+1}{x_{n-1}-x_{n}} \frac{\prod_{i=1}^n x_{n-1}+i}{\prod_{i=1}^{n-2}x_{n-1}+i}\right)\\ &= \frac{\left(n{+}2x_n{+}1\right)(x_n{+}n{-}1)(x_n{+}n)-\left(n{+}2x_{n-1}{+}1\right)(x_{n-1}{+}n-1)(x_{n-1}{+}n)}{x_n{-}x_{n-1}}\\ &=2\left(x_{n-1}^2+x_{n-2}^2\right)+2x_nx_{n-1}+(5n-1)(x_n+x_{n-1})+4n^2-n-1, \end{align*} which proves that the coefficients of $x_i^2$ and $x_ix_j$ in \Cref{eq:rhs2} agree with those in \Cref{eq:lhs2}: \begin{align} &\sum_{j=1}^n \left(\prod_{\substack{1 \leq i \leq n \\ i\neq j}} \frac{1}{x_j-x_i}\right) P(x_j) =2\left(\sum_{1\leq i \leq j \leq n} x_ix_j\right) + C_1 \left(\sum_{i=1}^n x_i\right)+C_0\label{eq:rhs3} .\end{align} We now determine $C_1$. Setting $x_i=-i$ for $1 \leq i \leq n$ makes every term in \Cref{eq:rhs2} vanish; setting $x_i=-i+1$ leaves only the first term, which simplifies to $n(n+1)$. Specializing $x_i$ to these values in \Cref{eq:rhs3} and subtracting, we obtain \begin{align*} &\left(\frac{n(n+1)(n+2)(3n+1)}{12}-\frac{n(n+1)}{2}C_1+C_0\right)\\&-\left(\frac{(n-1)n(n+1)(3n-2)}{12}-\frac{(n-1)n}{2}C_1+C_0\right)=n^3+n^2-nC_1 \end{align*} Equating $n^3+n^2-nC_1 = 0-n(n+1)$, we obtain $C_1=(n+1)^2$. Again setting $x_i=-i$ for $1 \leq i \leq n$ so that \Cref{eq:rhs2} is 0, we finally determine that $C_0=\frac{n (n + 1) (3n^2+5n+4)}{12}$ by computing \[ C_0=2\left(\sum_{1\leq i \leq j \leq n} ij\right)-(n{+}1)^2 \left(\sum_{i=1}^n i\right)=\frac{n(n{+}1)(3n^2{+}5n{+}4)}{12}. \qedhere \] \end{proof} \begin{comment} Write $\prec$ for the dominance order on weights. Assume we know the formula for all $\lambda \prec \lambda_1+\lambda_2$. By definition of the weights that occur in the tensor product of two highest-weight representations and by symmetry, we obtain \begin{align*} \sum_{\mu \in V(\lambda_1) \otimes V(\lambda_2)} n^{\lambda_1 \otimes \lambda_2}_{\mu} |\mu|^2 &= \sum_{\mu_1 \in V(\lambda_1)} n_{\mu_1}^{\lambda_1} \sum_{\mu_2 \in V(\lambda_2)} n_{\mu_2}^{\lambda_2} |\mu_1+\mu_2|^2 \\ &= \sum_{\mu_1 \in V(\lambda_1)} n_{\mu_1}^{\lambda_1} \sum_{\mu_2 \in V(\lambda_2)} n_{\mu_2}^{\lambda_2}\left( |\mu_1|^2+|\mu_2|^2+2\langle \mu_1,\mu_2 \rangle\right)\\ &= \frac{\text{dim}(V(\lambda_1) \otimes V(\lambda_2))}{h+1}\left( \big\langle \lambda_1,\lambda_1+2\rho\big\rangle + \big\langle \lambda_2,\lambda_2+2\rho\big\rangle\right). \end{align*} On the other hand, decomposing the tensor into irreducible components, we have \begin{align*} \sum_{\mu \in V(\lambda_1) \otimes V(\lambda_2)} n^{\lambda_1 \otimes \lambda_2}_{\mu} |\mu|^2&=\sum_{V(\lambda) \in V(\lambda_1) \otimes V(\lambda_2)} \left(\sum_{\mu \in V(\lambda)} n_{\mu}^\lambda |\mu|^2 \right)\\ &= \left(\sum_{\mu \in V(\lambda_1+\lambda_2)} n_{\mu}^{\lambda_1+\lambda_2} |\mu|^2 \right)+\frac{1}{h+1}\sum_{\substack{V(\lambda) \in V(\lambda_1) \otimes V(\lambda_2)\\ \lambda \neq \lambda_1+\lambda_1}} \text{dim}(V(\lambda))\langle \lambda,\lambda+2\rho\rangle. \end{align*} Equating and solving for the sum over $V(\lambda_1+\lambda_2)$, we conclude that \begin{align*}\sum_{\mu \in V(\lambda_1+\lambda_2)} n_{\mu}^{\lambda_1+\lambda_2} |\mu|^2 = \frac{1}{h+1} \Big(&\text{dim}(V(\lambda_1) \otimes V(\lambda_2))\left(\big\langle \lambda_1,\lambda_1+2\rho\big\rangle + \big\langle \lambda_2,\lambda_2+2\rho\big\rangle\right)\\ &- \sum_{\substack{V(\lambda) \in V(\lambda_1) \otimes V(\lambda_2)\\ \lambda \neq \lambda_1+\lambda_1}} \text{dim}(V(\lambda))\big\langle \lambda,\lambda+2\rho\big\rangle \Big).\end{align*} \end{comment} \section{Proof of \Cref{thm:main_thm} using Orthogonal Decompositions In this section, we give a conceptual proof of~\Cref{thm:main_thm} (in types other than $A$ and $C$) using the theory of orthogonal decompositions. Our strategy is to compute the trace of the degree two Casimir element $\Omega$ in two different ways on the representation $V_\lambda$. \subsection{Orthogonal Decompositions of Lie algebras} \label{sec:orthogonal} The usual decomposition of $\mathfrak{g}$ using a fixed Cartan subalgebra $\mathfrak{h}$ and the adjoint representation is given in~\Cref{eq:decomp}. Numerologically, this reflects the identity \[n(h+1)=n+nh=\mathrm{dim}(\mathfrak{h})+|\Phi|.\] But since $\mathrm{dim}(\mathfrak{h})=n$ divides $\mathrm{dim}(\mathfrak{g})=n(h+1)$, we might ask for a \emph{different} decomposition of $\mathfrak{g}$ using a direct sum of $h+1$ Cartan subalgebras: \begin{equation} \mathfrak{g} = \bigoplus_{i=0}^{h} \mathfrak{h}_i, \text{ with } \mathfrak{h}_i \text{ a Cartan subalgebra of }\mathfrak{g} \text{ and } \mathfrak{h}_0=\mathfrak{h}. \end{equation} In fact, such a decomposition is always possible. More difficult is to require that these $h+1$ Cartan subalgebras are pairwise orthogonal with respect to the Killing form; such a decomposition is called an \defn{orthogonal decomposition}. We refer the reader to~\cite{kostrikin1981orthogonal, kostrikin1994orthogonal} for background and references, pausing only to remark that such decompositions have a number of applications, including Thompson's construction of his sporadic simple group from the Lie algebra of type $E_8$~\cite{thompson1976conjugacy} and the construction of mutually unbiased bases for quantum cryptography~\cite{boykin2007mutually}. We note that Kostant used a dual approach to the related numerological problem of trying to uniformly explain the duality between degrees and the heights of roots~\cite{kostant2009principal}. Kostant decomposed $\mathfrak{g}$ into direct sum of $n$ irreducible representations of the principal three dimensional simple subalgebra (a distinguished copy of $\mathfrak{sl}_2$ inside $\mathfrak{g}$), reflecting the identity $n(h+1) = \sum_{i=1}^n (2d_i-1)$. \begin{theorem}[{\cite{kostrikin1994orthogonal}}] A complex simple Lie algebra $\mathfrak{g}$ has an orthogonal decomposition, except possibly if \begin{itemize} \item $\mathfrak{g}=\mathfrak{sl}_n$ for $n$ not a prime power; or if \item $\mathfrak{g}=\mathfrak{sp}_{2n}$ for $n \neq 2^m$. \end{itemize} \label{thm:orthogonal} \end{theorem} Although types $A$ and $C$ are usually the easiest Lie algebras to work with, it is widely believed that these classical Lie algebras \emph{do not} have orthogonal decompositions (outside the cases listed above); this problem is wide open, even for $\mathfrak{sl}_6$. The problem of finding such decompositions was dubbed the \defn{Winnie-the-Pooh problem} in the Russian paper~\cite{kostrikin1981orthogonal}, due to a play on words found in Zahoder's translation of Milne's famous children's book ``Winnie-the-Pooh'' into Russian. Zahoder's play on words can be interpreted as the sequence of Cartan types $A_5$---corresponding to the smallest open case $\mathfrak{sl}_6$---then $A_6, A_7,$ and $A_8$. This play on words apparently has no counterpart in Milne's original text, so when translating~\cite{kostrikin1981orthogonal} into English, Queen also translated Zahoder's verse---while managing to preserve the pun~\cite{kostrikin1994orthogonal}. \begin{comment} { \renewcommand{\thetheorem}{\ref{thm:main_thm}} \begin{theorem} For $\mathfrak{g}$ a complex simple Lie algebra with $V(\lambda)$ its finite-dimensional irreducible representation of highest weight $\lambda$, the expected norm of a weight in $V(\lambda)$ is \[\frac{1}{\mathrm{dim}(V(\lambda))} \sum_{\mu \in V(\lambda)} \mathrm{dim}(V_\lambda(\mu)) \langle \mu,\mu\rangle = \frac{1}{h+1} \big\langle \lambda, \lambda + 2\rho\big\rangle,\] where $n_\lambda^\mu$ is the multiplicity of $\mu$ in $V(\lambda)$ and $h$ is the Coxeter number of $\mathfrak{g}$. \end{theorem} \addtocounter{theorem}{-1} } \end{comment} \begin{proof}[Third proof of~\Cref{thm:main_thm}, valid in types not $A$ or $C$] Suppose $\mathfrak{g}$ has an orthogonal decomposition $\mathfrak{g} = \bigoplus_{i=0}^{h} \mathfrak{h}_i.$ For each $0\leq i\leq h$, pick an orthonormal basis $\{X_{i,1},\ldots,X_{i,n}\}$ of $\mathfrak{h}_i$. Then \[\Big\{X_{i,j} : 0 \leq i \leq h\text{ and } 1 \leq j \leq n \Big\}\] is an orthonormal basis of $\mathfrak{g}$, so we may write the degree two Casimir element $\Omega$ as \[\Omega = \sum_{\substack{0 \leq i \leq h\\ 1 \leq j \leq n }} X_{i,j}^2.\] We compute the trace of $\Omega$ on $V_\lambda$ in two different ways. On the one hand, $\Omega$ acts as the scalar $\langle \lambda,\lambda+2\rho\rangle$ by~\Cref{thm:casimir_action}, so that\[\mathrm{tr}_{V_\lambda} (\Omega) = \left\langle \lambda,\lambda+2\rho \right\rangle \mathrm{dim}(V_\lambda).\] On the other hand, for $0 \leq i \leq h$ define $\Omega_i = \sum_{j=0}^n X_{i,j}^2.$ By definition, $X_{0,j}$ acts as $\mu(X_{0,j})$ on the $\mu$-weight space of $V_\lambda$, so \begin{align*} \mathrm{tr}_{V_\lambda}(\Omega_0) = \sum_{\mu \in V_\lambda} \mathrm{dim}(V_\lambda(\mu)) \sum_{j=1}^n \mu(X_{0,j})^2 =\sum_{\mu \in V_\lambda} \mathrm{dim}(V_\lambda(\mu)) \langle \mu,\mu\rangle, \end{align*} since $\{X_{0,1},\ldots,X_{0,n}\}$ is an orthonormal basis of $\mathfrak{h}=\mathfrak{h}_0$. Since every $\mathfrak{h}_i$ is conjugate to $\mathfrak{h}_0$ under an inner automorphism of $\mathfrak{g}$ that leaves the Killing form invariant, we have that $\Omega_i$ is conjugate to $\Omega_0$ for all $0\leq i \leq h$. Therefore, \begin{align*}\mathrm{tr}_{V_\lambda} (\Omega) &= \mathrm{tr}_{V_\lambda} \left(\sum_{i=0}^h \Omega_i\right) = \sum_{i=0}^h \mathrm{tr}_{V_\lambda}\left(\Omega_i\right) \\ &= (h+1) \mathrm{tr}_{V_\lambda}\left(\Omega_0\right) = (h+1)\sum_{\mu \in V_\lambda} \mathrm{dim}(V_\lambda(\mu)) \langle \mu,\mu\rangle.\end{align*} The result now follows from equating the two expressions for $\mathrm{tr}_{V_\lambda} (\Omega)$. \end{proof} By~\Cref{thm:orthogonal}, this proof of~\Cref{thm:main_thm} applies to all types except possibly if $\mathfrak{g}=\mathfrak{sl}_n$ for $n$ composite; or if $\mathfrak{g}=\mathfrak{sp}_{2n}$ for $n \neq 2^m$ \section{Coxeter Cumulants} \label{sec:coxeter_cumulants} Let $[n]_q:=1+q+\cdots+q^{n-1}$ be the uniform distribution on $\{0,1,\ldots,n-1\}$. \begin{lemma} Fix $\{a_i\}_{i=1}^k$ and $\{b_i\}_{i=1}^k$ two sets of positive integers with $\prod_{i=1}^k \frac{[a_i]_q}{[b_i]_q}$ a polynomial in $q$, and let $X$ be a random variable with this distribution. Then the $r$th cumulant of $X$ is \[\kappa_r(X)=\frac{B_r}{r}\sum_{i=1}^n \left(a_i^r-b_i^r\right),\] where $B_r$ is the $r$th Bernoulli number. In particular, $\kappa_r(X)=0$ for odd $r>1$. \label{prop:cumulant} \end{lemma} \begin{proof} We first claim that the $r$th cumulant of $[n]_q$ is \begin{equation}\kappa_r([n]_q)=\frac{B_r}{r}(n^r-1).\label{eq:ncum}\end{equation} Let $X$ be a random variable whose distribution is $[n]_q$, and let \[u(t)=\mathrm{ln}\Expt{}{e^{tX}} = \sum_{r=0}^\infty \kappa_r(X) \frac{t^r}{r!}\] be its cumulant exponential generating function. We claim that \begin{equation}u(t)=\mathrm{ln}(n)+\sum_{r=1}^\infty \frac{B_r}{r}(n^r-1)\frac{t^r}{r}, \label{eq:fcumulant}\end{equation} so that the coefficients of $\frac{t^r}{r!}$ are as desired. To prove~\Cref{eq:fcumulant}, we write \[u(t)=\mathrm{ln}\left(\sum_{i=0}^{n-1} e^{it} \right)=\mathrm{ln}\left( \frac{1-e^{nt}}{1-e^t}\right).\] Taking a derivative of~\Cref{eq:fcumulant} using the fact that $\sum_{r=0}^\infty B_r \frac{t^r}{r!}=\frac{t}{1-e^{-t}}$, we obtain \[\sum_{r=1}^\infty B_r(n^r-1)\frac{t^r-1}{r!} = \frac{1}{t}\left(\frac{nt}{1-e^{-nt}}-\frac{t}{1-e^{-t}}\right),\] which matches the derivative $u'(t)$. The constant term is verified by computing $\lim_{t \to 0} u(t)$. Suppose now $X$ is a random variable with distribution $\prod_{i=1}^k \frac{[a_i]_q}{[b_i]_q}$. Using~\Cref{eq:ncum}, we compute \begin{align*}\kappa_r(X) &= \kappa_r\left(\prod_{i=1}^k \frac{[a_i]_q}{[b_i]_q}\right) =\sum_{i=1}^k \kappa_r([a_i]_q)-\kappa_r([b_i]_q) = \frac{B_r}{r}\sum_{i=1}^k \left(a_i^r-b_i^r\right). \qedhere\end{align*} \end{proof} The remainder of this section is devoted to corollaries of \Cref{prop:cumulant}. \subsection{Coxeter numerology} A finite irreducible Coxeter group $W$ of rank $n$ has roots $\Phi$, positive roots $\Phi^+$, degrees $d_1 < \cdots < d_n$, exponents $e_1<\cdots< e_n$, and a Coxeter number $h$. These satisfy $e_i = d_i-1$, $h=d_n$, and \[\sum_{i=1}^n e_i=\frac{nh}{2}= |\Phi^+|.\] Following~\cite{burns2012power,suter1998coxeter}, for $W$ a finite Weyl group (a crystallographic Coxeter group), define \[\gamma=\frac{\langle \widetilde{\alpha},\widetilde{\alpha} \rangle}{\langle \widetilde{\alpha}_s,\widetilde{\alpha}_s\rangle}g \chk{g},\] where $\widetilde{\alpha}$ is the highest root, $\widetilde{\alpha}_s$ is the highest short root, $g$ is the dual Coxeter number of $\Phi$, and $\chk{g}$ is the dual Coxeter number of the dual root system $\chk{\Phi}$. For noncrystallographic types, $\gamma$ is defined in~\Cref{fig:gamma}. With this definition, R.~Suter found the striking uniform formulas~\cite{suter1998coxeter} \begin{equation}\sum_{i=1}^n e_i^2 = \frac{n(h^2+\gamma-h)}{6} \text{ and } \sum_{i=1}^n e_i^3 = \frac{nh(\gamma-h)}{4}.\label{eq:sut}\end{equation} \begin{figure}[htbp] \[\begin{array}{|c|cccccccc|}\hline \text{Type} & A_n & B_n/C_n & D_n & E_6 & E_7 & E_8 & F_4 & G_2 \\ \gamma & (n+1)^2 & 4n^2+2n-2 & (2n-2)^2 & 144 & 324 & 900 & 162 & 48 \\ \hline \end{array}\] \[\begin{array}{|c|ccc|}\hline \text{Type} &H_3 & H_4 & I_2(m)\\ \gamma & 124 & 1116 & 2m^2-5m+6 \\ \hline \end{array}\] \caption{Values of $\gamma$ for finite Coxeter groups. The definition of $\gamma$ is uniform for Weyl groups. \label{fig:gamma} \end{figure} \subsection{Inversions in Finite Coxeter Groups} Let $(W,S)$ be a finite irreducible Coxeter system. The \defn{length} of the shortest word in simple reflections for an element $w \in W$ is written $\ell(w)$. It is well-known that the generating function for length is \[\mathrm{Inv}(W;q):=\sum_{w\in W} q^{\ell(w)}=\prod_{i=1}^n [d_i]_q.\] Applying \Cref{prop:cumulant} gives the following generalization of~\cite[Theorem 3.1]{kahle2018counting}. \begin{corollary} Let $W$ be a finite irreducible Coxeter group, and let $X$ be a random variable with distribution $\mathrm{Inv}(W;q)$. Then \[\kappa_r(X) = \frac{B_r}{r} \sum_{i=1}^n \left(d_i^r-1\right).\] In particular, the expectation and variance are \[\mathbb{E}(X)=\frac{|\Phi^+|}{2} \text{ and } \mathbb{V}(X)=\frac{n(\gamma+5h+h^2)}{72}.\] \end{corollary} \begin{proof} The formula for expectation is immediate from the symmetry $w \mapsto w w_\circ$, where $w_\circ$ is the longest element of $W$. The formula for variance follows from \Cref{eq:sut}. \end{proof} \subsection{Rational Catalan Numbers} Let $W$ be a finite well-generated irreducible complex reflection group with exponents $e_1,\ldots,e_n$ and $p$ coprime to to the Coxeter number $h=d_n$. The \defn{rational Catalan number} \[\mathrm{Cat}^{[p]}(W;q):=\prod_{i=1}^n \frac{[p+(pe_i \mod h)]_q}{[d_i]_q}\] is the graded character of the representation $eL_{p/h}(\mathsf{triv})$ of the rational Cherednik algebra corresponding to $W$~\cite{gordon2012catalan,stump2015cataland}. When $W$ is a Coxeter group, multiplication by $p$ simply permutes the exponents modulo $h$. Applying \Cref{prop:cumulant} gives the following. \begin{corollary} Let $W$ be a finite Coxeter group, and let $X$ be a random variable with distribution $\mathrm{Cat}^{[p]}(W;q)$. Then \[\mathbb{E}(X)=\frac{n(p-1)}{2}, \hspace{2em} \mathbb{V}(X) = \frac{n(p-1)(p+h+1)}{12},\] and for $W$ a crystallographic Coxeter group $\kappa_4(X)=-\frac{n(p-1)(h+p+1)(p^2+ph+\gamma+1)}{120}$. \label{prop:rat_cat} \end{corollary} \begin{proof} The formulas follow from simple computations and \Cref{eq:sut}. \end{proof} Interestingly, the variance in~\Cref{prop:rat_cat} is exactly the expectation computed in~\cite{thiel2017strange} for simply-laced crystallographic Coxeter groups. \subsection{Minuscule Posets} Recall that a dominant weight $\lambda$ is called \defn{minuscule} if the weights in its highest weight representation $V_\lambda$ coincides with weights in its Weyl group orbit $\{w \lambda : w \in W\}$. A \defn{minuscule poset} may then be defined as the order filter of positive roots generated by simple root whose corresponding fundamental weight is minuscule. A \defn{plane partition of height at most $k$} in a poset $P$ is an order-preserving map $P \to \{0,1,\ldots,k\}$. It is well-known that the generating function for the number of boxes in plane partitions of height at most $k$ in a minuscule poset has the product formula~\cite{proctor1984bruhat} \[\mathrm{PP}^{[p]}(P;q):=\prod_{p\in P} \frac{[k+\mathrm{ht}(p)]_q}{[\mathrm{ht}(p)]_q}.\] \begin{corollary} Let $P$ be a minuscule poset, and let $X$ be a random variable with distribution $\mathrm{PP}^{[p]}(P;q)$. Then \[\mathbb{E}(X)=\frac{k}{2}|P| \text{ and } \mathbb{V}(X) = \frac{k(k+h)}{12}|P|.\ \label{prop:min_pp} \end{corollary} \begin{proof} We have that \[\kappa_r(X)=\frac{B_r}{r}\sum_{p \in P} \Big( (k+\mathrm{ht}(p))^r-\mathrm{ht}(p)^r\Big).\] Expectation is immediate. For variance, we compute \begin{align*} \frac{1}{12}\sum_{p \in P} \Big( (k+\mathrm{ht}(p))^2-\mathrm{ht}(p)^2 \Big) &= \frac{1}{12}\sum_{p \in P} \left( k^2 + 2k \mathrm{ht}(p) \right)\\ &=\frac{1}{12}\left(|P|k^2+2k \sum_{p \in P} \mathrm{ht(p)} \right)\\ &=\frac{1}{12}\left(|P|k^2+|P|hk\right)=\frac{k(k+h)}{12}|P|, \end{align*} where $\sum_{p \in P} \mathrm{ht(p)}=\frac{h}{2}|P|$ using the symmetry of $P$. \end{proof} Specializing to type $A$ gives the following. \begin{corollary} The variance for the number of boxes in plane partitions fitting inside an $a \times b \times c$ box is $\frac{1}{12}abc(a+b+c).$ \end{corollary} \medskip A \defn{linear extension} of a poset $P$ is an order-preserving bijection from $P$ to $\{1,2,\ldots,|P|\}$. Relative to a fixed linear extension $\ell$ of a poset, the \defn{major index} of a second linear extension $\ell'$ is the sum of the positions of the descents of $\ell'$---that is, the sum $\sum i,$ where the sum is over all $i$ for which $\ell'\left(\ell^{-1}(i)\right)>\ell'\left(\ell^{-1}(i+1)\right)$. Recall that the generating function for major index of linear extensions of a minuscule poset is given by \[\mathrm{SYT}(P;q):=\frac{[|P|]!_q}{\prod_{p \in P} [\mathrm{ht}(p)]_q}.\] \begin{corollary} Let $P$ be a minuscule poset, and let $X$ be a random variable with distribution $\mathrm{SYT}(P;q)$. Then $\mathbb{E}(X)=\frac{|P|(|P|+1-h)}{4}$ \label{prop:min_le} \end{corollary} \begin{proof} Compute using \[\kappa_r(X)=\frac{B_r}{r}\left(\sum_{i=1}^{|P|} i^r-\sum_{p \in P}\mathrm{ht}(p)^r\right).\] \end{proof} Specializing to type $A$ gives the following. \begin{corollary} The expected value for major index of standard Young tableaux of rectangular shape is given by $\frac{a(a-1)b(b-1)}{4}$. \end{corollary} \subsection{Descending Plane Partitions} The generating function for descending plane partitions by number of boxes is given by \[\mathrm{DPP}(q)=\prod_{i=0}^{n-1} \frac{[3i+1]!_q}{[n+i]!_q}.\] \begin{corollary} Let $X$ be a random variable with distribution $\mathrm{DPP}(q)$. Then \[\mathbb{E}(X)= \frac{1}{6} n(n^2-1) \text{ and } \mathbb{V}(X)= \frac{1}{12} n^2(n^2-1).\ \label{prop:dpp} \end{corollary} \section{Open problems} \begin{itemize} \item The problem of determining uniform formulas for higher moments for the expected norm of a weight in a highest weight representation is open. \item By the Harish-Chandra isomorphism, $\mathfrak{g}$ has $n$ Casimir elements. Since the Casimir elements live in the center of $U(\mathfrak{g})$, they generalize the degree two Casimir by acting as a scalar on any highest weight representation $V(\lambda)$. It might be interesting to write down explicit formulas.\footnote{After our presentation of this work at FPSAC 2018, we understand that Richard Stanley has made some progress on this problem in type $A$.} \item The expectations arising from evaluating other natural $W$-symmetric functions besides the norm on the weights of $V_\lambda$ could be worth looking at. \item Since Schubert polynomials generalize Schur polynomials---which are equivalent to the Weyl character formula in type $A$---one could ask for expectation of polynomial functions of the exponents in a Schubert polynomial. \item For $P$ a minuscule poset, it would be interesting to find uniform expressions for $\sum_{p \in P} \mathrm{ht}(p)^r$. \end{itemize} \section*{Acknowledgments} The second author warmly thanks Dennis Stanton for precious help with~\Cref{sec:typea} and Paul Garrett for explaining to him where the Casimir element lives. An extended abstract of this work was presented at FPSAC 2018~\cite{thielwinnie}. \bibliographystyle{amsalpha}

https://arxiv.org/abs/1212.1506

Single Layer Potentials on Surfaces with Small Lipschitz constant

This paper considers to the equation [\int_{S} \frac{U(Q)}{|P-Q|^{N-1}} dS(Q) = F(P), P \in S,] where the surface S is the graph of a Lipschitz function \phi on R^N, which has a small Lipschitz constant. The integral in the left-hand side is the single layer potential corresponding to the Laplacian in R^{N+1}. Let \Lambda(r) be a Lipschitz constant of \phi on the ball centered at the origin with radius 2r. Our analysis is carried out in local L^p-spaces and local Sobolev spaces, where 1 < p < \infty, and results are presented in terms of \Lambda(r). Estimates of solutions to the equation are provided, which can be used to obtain knowledge about the behaviour of the solutions near a point on the surface. The estimates are given in terms of seminorms. Solutions are also shown to be unique if they are subject to certain growth conditions. Local estimates are provided and some applications are supplied.

\section{Introduction} Let~$S$ be the graph in~$\ensuremath{\mathbf{R}}^{N+1}$ of a Lipschitz function~$\ensuremath{\varphi} \colon \ensuremath{\mathbf{R}}^N \rightarrow \ensuremath{\mathbf{R}}$, where~$N \geq 2$. We define~$\ensuremath{\Lambda}$ to be a function on~$(0,\infty)$ such that \begin{equation} \label{eq:def_lc} |\ensuremath{\varphi}(x) - \ensuremath{\varphi}(y)| \leq \lc[r] |x-y| \quad \text{for} \quad |x|,\;|y| \leq 2r \end{equation} for every~$r > 0$. The function~$\ensuremath{\Lambda}$ is assumed to be increasing and bounded: \begin{equation} \label{eq:i:lip_cond} \lc[r] \leq \ensuremath{\Lambda}_0 \quad \text{for every } r > 0. \end{equation} We will assume that~$\ensuremath{\Lambda_0}$ is sufficiently small. One can choose~$\lc[r]$ to be the optimal constant in~(\ref{eq:def_lc}) and then~$\ensuremath{\Lambda}_0$ is the (global) Lipschitz constant of~$\ensuremath{\varphi}$. We consider the {\it single layer potential} on the surface~$S$: \begin{equation} \label{eq:simplelayerpot} \int_{S} \frac{U(Q)}{|P-Q|^{N-1}} \, dS(Q) \text{,} \quad P \in S \text{,} \end{equation} where~$dS$ is the Euclidean surface measure. This object is important since, for instance, it appears when one applies the direct approach to solve Laplace's equation corresponding boundary integral equation; see, e.g., Hsiao and Wendland~\cite{wendland}. If the surface is the hyperplane~$x_{N+1} = 0$, then we obtain the classical Riesz potential of order one. The main objective of this article is to find a solution~$u$ to the equation \begin{equation} \label{eq:maineq} \S u (x) = f(x) \text{,} \quad x \in \ensuremath{\mathbf{R}}^N \text{,} \end{equation} where \[ \S u(x) = \int_{\ensuremath{\mathbf{R}}^N} \frac{u(y) \sqrt{1 + |\nabla \ensuremath{\varphi}(y)|^2}}{|\ld[x] - \ld[y]|^{N-1}} \, dy \text{,} \quad x \in \ensuremath{\mathbf{R}}^N \text{.} \] Here,~$\ld[x] = (x,\ensuremath{\varphi}(x))$ for~$x \in \ensuremath{\mathbf{R}}^N$, and~$\S u$ is the parametrization of the single layer potential in~(\ref{eq:simplelayerpot}). More specifically, we will consider equation~(\ref{eq:maineq}) for~$u \in \Lloc[p]$, where~$1 \leq p < \infty$, and~$f \in \ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$, where~$1 < p < \infty$. We will formulate our results in terms of the family of seminorms defined by \[ \Np{u}{r} = \biggl( \frac{1}{r^N} \int_{r \leq |x| < 2r} |u(x)|^p \, dx \biggr)^{1/p} \text{,} \quad r > 0 \text{.} \] To simplify upcoming notation, let $\Qw{m}{n}(t) = t^{m}$ if~$0 < t \leq 1$, and $\Qw{m}{n}(t) = t^{n}$ if~$t > 1$, for non-negative~$m$ and~$n$. For~$1 \leq p < \infty$, the Banach space~$\Xp[1]$ consists of all functions~$u$ that belong to~$\Lloc[p]$ and satisfy \begin{equation} \label{eq:defXp} \int_{0}^{\infty} \Qw{N}{1}(\rho) \, \Np{u}{\rho} \, \frac{d\rho}{\rho} < \infty \text{.} \end{equation} We take this expression as the norm on~$\Xp[1]$. This space is the natural domain, in terms of the seminorms~$\Npw$, for the operator~$\S$ in the case that~$S$ is the hyperplane~$x_{N+1} = 0$; this is discussed further by the authors in~\cite{thim1}. We also remark that, if~$1 \leq p < N$, then~$L^p(\ensuremath{\mathbf{R}}^N) \subset \Xp[1]$. If~$p \geq N$, there exist functions in~$L^p(\ensuremath{\mathbf{R}}^N)$ which do not belong to~$\Xp$. For~$1 < p < \infty$ and~$0 \leq \ensuremath{M} \leq N$, the normed space~$\ensuremath{Y^{1,p}_{M}(\R^N)}$ consists of all functions~$f$ in~$\ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$ such that \begin{equation} \label{eq:Yonew} \int_0^{\infty} \Qw{\ensuremath{M}}{1}(\rho) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} < \infty \end{equation} and~$\lim_{r \rightarrow \infty} \int_{S^{N-1}} f(r\theta) \, dS(\theta) = 0$, where~$S^{N-1}$ is the unit sphere in~$\ensuremath{\mathbf{R}}^N$. The left-hand side of~(\ref{eq:Yonew}) defines the norm on this Banach space. The condition in~(\ref{eq:Yonew}) implies that the limit in the definition exists. This limit ensures that, e.g., constants do not belong to~$\ensuremath{Y^{1,p}_{M}(\R^N)}$. The reason that functions of this type are excluded is that we cannot expect to find a solution to~$\S u = f$ in this case; indeed, if~$f$ is a nonzero constant, then the solution~$u$ in Theorem~\ref{t:i:exist} below would have to be~$u = 0$, which obviously does not solve~$\S u = f$ in any reasonable way. Furthermore, if~$f \in \ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$ such that~$|\nabla f| \in L^p(\ensuremath{\mathbf{R}}^N)$ for some~$N/\ensuremath{M} < p < N$, then~$f \in \ensuremath{Y^{1,p}_{M}(\R^N)}$. In Section~\ref{s:proof_exist}, we prove the following existence result. \begin{theorem} \label{t:i:exist} There exist positive constants\/~$\ensuremath{\Lambda_{*}}$,\/~$\ensuremath{c_1}$,\/~$\ensuremath{c_2}$, and\/~$\ensuremath{c_3}$, depending only on\/~$N$ and\/~$p$, such that if\/~$\ensuremath{\Lambda_0} \leq \ensuremath{\Lambda_{*}}$ and if\/~$f \in \ensuremath{Y^{1,p}_{M}(\R^N)}$ with\/~$M = N - \ensuremath{c_2} \ensuremath{\Lambda_0}$ and~\/$1 < p < \infty$, then the equation~{\rm(}\ref{eq:maineq}{\rm)} has a solution\/~$u \in \Xp$. For\/~$r > 0$, this solution satisfies \begin{equation} \label{eq:i:est_sol} \begin{aligned} \Np{u}{r} \leq {} & \ensuremath{c_3} \int_0^{r} \left( \frac{\rho}{r} \right)^{\ensuremath{M}} \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho}\\ & + \ensuremath{c_3} \int_r^{\infty} \exp \biggl( \ensuremath{c_1} \int_r^{\rho} \ensuremath{\Lambda}(\nu) \, \frac{d\nu}{\nu} \biggr) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho}. \end{aligned} \end{equation} \end{theorem} \noindent Even if the objects are different, Theorem~\ref{t:i:exist} is closely related to results obtained by V.A.~Kozlov and V.G.~Maz'ya for ordinary differential equations and ordinary differential equations with operator coefficients; see Section 6.4 in~\cite{kozlovmazyaSLE} and Section 6.3 in~\cite{kozlovmazyaDEOC}. It is possible to use~(\ref{eq:i:est_sol}) to obtain two-weighted estimates for solutions to~(\ref{eq:maineq}) in weighted~$L^p$-spaces and weighted Sobolev spaces similar to those found in Section~7.5 in~\cite{kozlovmazyaDEOC}; see Section~8 in the authors' article~\cite{thim1} for an example of this procedure when the surface~$S$ is the hyperplane~$x_{N+1} = 0$. Furthermore, one can also compare with the boundedness results for Riesz potentials in local Morrey-type spaces found in Burenkov et al.~\cite{burenkov2,burenkov1} and references cited therein. If~$\ensuremath{\Lambda_0} = 0$, we recover the same estimate in~(\ref{eq:i:est_sol}) for the solution as in the case when~$S$ is the hyperplane~$x_{N+1} = 0$; compare with~(\ref{eq:NpR}) below. Furthermore, if~$\ensuremath{\Lambda_0} \rightarrow 0$, the condition in Theorem~\ref{t:i:exist} that~$f \in \ensuremath{Y^{1,p}_{M}(\R^N)}$ reduces to the corresponding requirement for the hyperplane-case; see Corollary~\ref{c:cont_req_sol}. In Section~\ref{s:proof_uniq}, we prove that solutions to~(\ref{eq:maineq}) are unique if they satisfy certain properties. More specifically, we have the following result. \begin{theorem} \label{t:i:uniq} Suppose that\/~$u \in \Lloc[p]$,\/~$1 < p < \infty$, satisfies~{\rm(}\ref{eq:defXp}{\rm)} and \begin{equation} \label{eq:est_uniq_infty} \Np{u}{r} = O\biggl( \exp \biggl( -\ensuremath{c_1} \int_1^r \ensuremath{\Lambda}(\nu) \, \frac{d\nu}{\nu} \biggr) \biggr) \quad \text{as } r \rightarrow \infty \text{,} \end{equation} and \begin{equation} \label{eq:est_uniq_zero} \Np{u}{r} = O\bigl( r^{-\ensuremath{M}} \bigr) \quad \text{as } r \rightarrow 0 \text{,} \end{equation} where\/~$\ensuremath{c_1}$,\/~$\ensuremath{c_2}$,\/~$\ensuremath{M}$, and\/~$\ensuremath{\Lambda_{*}}$ are in Theorem~\ref{t:i:exist}, and\/~$\ensuremath{\Lambda_0} \leq \ensuremath{\Lambda_{*}}$. If\/~$\S u = 0$, then it follows that\/~$u = 0$. \end{theorem} \noindent It should be noted that the solution in Theorem~\ref{t:i:exist} satisfies the conditions in Theorem~\ref{t:i:uniq}; see Remark~\ref{r:soluniq} in Section~\ref{s:uniq}. In Section~\ref{s:assymp}, we prove a local version of Theorem~\ref{t:i:exist}. \begin{theorem} \label{t:i:local} Let the constants\/~$\ensuremath{c_1}$,\/~$\ensuremath{c_2}$,\/~$\ensuremath{M}$, and\/~$\ensuremath{\Lambda_{*}}$ be as in Theorem~\ref{t:i:exist} and suppose that\/~$\ensuremath{\Lambda_0} \leq \ensuremath{\Lambda_{*}}$. Suppose also that\/~$u \in \Xp$, where\/~$1 < p < \infty$, satisfies~{\rm(}\ref{eq:est_uniq_zero}{\rm)} and\/~$\S u(x) = f(x)$ for\/~$|x| \leq 2r_0$, where\/~$r_0$ is a positive constant and\/~$f \in \ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$ satisfies \begin{equation} \label{eq:req_local} \int_0^{2r_0} \Qw{\ensuremath{M}}{1}(\rho) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} < \infty. \end{equation} Then\/~$\Np{u}{r}$ is bounded by \begin{equation} \label{eq:i:assymp} \begin{aligned} & C \int_0^r \left({\frac{\rho}{r}}\right)^{\ensuremath{M}} \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} + C \int_r^{2r_0} \exp \biggl( \ensuremath{c_1} \int_{r}^{\rho} \ensuremath{\Lambda}(\nu) \, \frac{d\nu}{\nu} \biggr) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho}\\ & \qquad + C \biggl( \| u \|_{\Xp} + \int_{r_0/2}^{2r_0} \Np{f}{\rho} \, \frac{d\rho}{\rho} \biggr) \exp \biggl( \ensuremath{c_1} \int_r^{r_0} \ensuremath{\Lambda}(\nu) \, \frac{d\nu}{\nu} \biggr), \end{aligned} \end{equation} for\/~$0 < r < r_0$, where\/~$C$ only depends on\/~$N$ and\/~$p$. \end{theorem} \noindent An application of this theorem can be found in Section~\ref{s:est_Np_ralpha}, where we show that if~$\Np{\nabla f}{r} \leq C r^{-\alpha}$ for small~$r$ and some~$\alpha \in (0,\ensuremath{M})$ and~$\ensuremath{\Lambda}(r) \rightarrow 0$ as~$r \rightarrow 0$, then~$\Np{u}{r} \leq C r^{-\alpha}$ (for small~$r$). We also show that if~$\ensuremath{\Lambda}$ and~$\Np{\nabla f}{\, \cdot \, }$ satisfy Dini-type conditions at $0$, then~$\Np{u}{r} \leq C$ for small~$r$. The existence and uniqueness of solutions to equation~(\ref{eq:maineq}) follows by applying a fixed point theorem for locally convex spaces, which was developed by the authors in~\cite{thim2}. To apply this theorem, we approximate the operator~$\S$ with a Riesz potential operator and control the behaviour of the remainder. The natural approach of approximating the surface~$S$ with the hyperplane~$x_{N+1} = 0$ does not yield sufficiently strong estimates for our purposes. Instead, we use a weighted Riesz potential to match the behaviour of~$\S u$ at the origin. The choice of weight is not obvious since we need to estimate both the potential and its derivative. The fact that the solution to the fixed point problem solves~(\ref{eq:maineq}) follows from results derived earlier by the authors for the hyperplane case in~\cite{thim1}. A summary of the hyperplane case can be found in Section~\ref{s:rieszpot}. \section{Properties of the Single Layer Potential} \subsection{Riesz Potentials on~\boldmath{$\ensuremath{\mathbf{R}}^N$}} \label{s:prelim} \label{s:rieszpot} We start by recalling some properties of the potential~$\S u$ in the case when~$S$ is the hyperplane~$x_{N+1} = 0$. These results were derived by the authors in~\cite{thim1}. Equation~(\ref{eq:maineq}) reduces in this case to \begin{equation} \label{eq:maineq_riesz} \ensuremath{\mathcal{I}} u(x) = \int_{\ensuremath{\mathbf{R}}^N} \frac{u(y)}{|x-y|^{N-1}} \,dy = f(x)\text{,} \quad x \in \ensuremath{\mathbf{R}}^N \text{,} \end{equation} where the operator $\ensuremath{\mathcal{I}}$ is the Riesz potential operator of order~1 defined for~$u$ in~$\Xpett$. The space~$\Xpett$ is the natural domain if~$\ensuremath{\mathcal{I}} u$ is interpreted as an absolutely convergent integral. For solvability results in~$L^p$-spaces, we refer to Rubin~\cite{rubin} and references cited therein. The following continuity properties for~$\ensuremath{\mathcal{I}}$ hold; see Theorem 1.3 in~\cite{thim1}. \begin{theorem} \label{t:cont_I1} The operator\/~$\ensuremath{\mathcal{I}}$ maps\/~$\Xp[1]$ into\/~$\ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$ for\/~$1 < p < \infty$. Moreover, there exist two constants\/~$C_1$ and\/~$C_2$, depending only on\/~$N$ and\/~$p$, such that \begin{equation} \label{eq:Np_est_I_1} \Np{\ensuremath{\mathcal{I}} u}{r} \leq C_1r \int_{0}^{\infty} \Qw{N}{1}\left(\frac{\rho}{r}\right) \, \Np{u}{\rho} \, \frac{d\rho}{\rho} \text{,} \end{equation} and \begin{equation} \label{eq:Np_est_diff_I_1} \Np{\nabla \ensuremath{\mathcal{I}} u}{r} \leq C_2 \int_{0}^{\infty} \Qw{N}{0}\left(\frac{\rho}{r}\right) \, \Np{u}{\rho} \, \frac{d\rho}{\rho} \text{,} \end{equation} for every function\/~$u \in \Xp$ and every\/~$r > 0$. In the first inequality,\/~$p=1$ is also allowed. \end{theorem} \noindent A solution to~(\ref{eq:maineq_riesz}) is given by \begin{equation*} \ensuremath{\mathcal{R}} f(x) = \frac{c_N}{N-1} \sum_{k=1}^{N} R_k \partial_k f(x) \text{,} \quad x \in \ensuremath{\mathbf{R}}^N \text{,} \end{equation*} where~$c_N = \Gamma\left( (N+1)/2 \right)\pi^{-({N+1})/{2}}$,~$\Gamma$ is the gamma function, and~$R_k$ is the~$k$th Riesz transform (cf.~Stein~\cite[p. 57]{stein}). For~$1 < p < \infty$, the space~$\ensuremath{Y^{1,p}(\R^N)}$ consists of all functions~$f$ in~$\ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$ such that \begin{equation} \label{eq:y1pdef} \| f \|_{\ensuremath{Y^{1,p}(\R^N)}} = \biggl| \int_{1 \leq |x| < 2} f(x) \, dx \biggr| + \int_{0}^{\infty} \Qw{N}{0}(\rho) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} < \infty \text{,} \end{equation} where this expression is the norm on~$\ensuremath{Y^{1,p}(\R^N)}$. The next theorem shows that the operator~$\ensuremath{\mathcal{R}}$ maps this space continuously into~$\Lloc[p]$; see Theorem~1.5 in~\cite{thim1}. \begin{theorem} \label{t:cont_R} The operator\/~$\ensuremath{\mathcal{R}}$ is defined on\/~$\ensuremath{Y^{1,p}(\R^N)}$ for\/~$1 < p < \infty$, and there exists a constant\/~$C$, depending only on\/~$N$ and\/~$p$, such that \begin{equation} \label{eq:NpR} \Np{\ensuremath{\mathcal{R}} f}{r} \leq C \int_{0}^{\infty} \Qw{N}{0}\left( \frac{\rho}{r} \right) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho}, \quad r > 0, \end{equation} for every function\/~$f \in \ensuremath{Y^{1,p}(\R^N)}$. \end{theorem} \noindent We define\/~$\ensuremath{Y^{1,p}_{0}(\R^N)}$ as the proper subspace of\/~$\ensuremath{Y^{1,p}(\R^N)}$ consisting of those functions\/~$f$ that satisfy \begin{equation} \label{eq:y1p0ineq} \int_{0}^{1} \rho^N(1-\log \rho) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} + \int_{1}^{\infty} \Np{\nabla f}{\rho} \, d\rho < \infty \end{equation} and $\lim_{r \rightarrow \infty} \int_{S^{N-1}} f(r\theta) \, d\theta = 0$. The expression in~{\rm(}\ref{eq:y1p0ineq}{\rm)} is taken as the norm on\/~$\ensuremath{Y^{1,p}_{0}(\R^N)}$. We have the following solvability result; see Theorem 1.7 and 1.1 in~\cite{thim1}. \begin{theorem} \label{t:inv} \label{t:injectivity} \mbox{} \begin{enumerate} \item[{\rm(i)}] The operator\/~$\ensuremath{\mathcal{R}}$ is bounded from\/~$\ensuremath{Y^{1,p}_{0}(\R^N)}$ to\/~$\Xp[1]$ for\/~$1 < p < \infty$. \item[{\rm(ii)}] If\/~$f \in \ensuremath{Y^{1,p}_{0}(\R^N)}$, then there exists a solution\/~$u \in \Xp$ to~{\rm(}\ref{eq:maineq_riesz}{\rm)} which satisfies \[ \Np{u}{r} \leq C \int_{0}^{\infty} \Qw{N}{0}\left( \frac{\rho}{r} \right) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho}, \quad r > 0. \] \item[{\rm(iii)}] Suppose that\/~$u$ is a locally integrable function such that \begin{equation} \label{eq:cond_uniq_plane} \int_{|y| < 1} |u(y)| \, dy + \int_{|y| \geq 1} \frac{|u(y)| \, dy}{|y|^{N-1}} < \infty \text{.} \end{equation} If\/~$\ensuremath{\mathcal{I}} u = 0$, then it follows that\/~$u = 0$. \end{enumerate} \end{theorem} \noindent It can be verified that the condition in~(\ref{eq:cond_uniq_plane}) coincides with the definition of~$\Xpett$; see~\cite{thim1}. \subsection{Approximation of~\boldmath{$\S$}} \label{s:appr_S} For~$k = 1,2,\ldots,N+1$, let~$T_k$ be the singular integral operator defined by \[ T_{k} u(x) = \pval \int_{\ensuremath{\mathbf{R}}^N} \frac{(\ld[x] - \ld[y])_k}{|\ld[x] - \ld[y]|^{N+1}} \, u(y) \sqrt{1 + |\nabla \ensuremath{\varphi}(y)|^2} \, dy \text{,} \quad x \in \ensuremath{\mathbf{R}}^N \text{,} \] where~$u \in L^p(\ensuremath{\mathbf{R}}^N)$ with~$1 < p < \infty$. These operators are bounded on~$L^p(\ensuremath{\mathbf{R}}^N)$ for~$1 < p < \infty$; see for instance Dahlberg~\cite{dahlberg}. Moreover, analogously with Lemma~3.1 in the authors' article~\cite{thim1}, one can show that if~$u \in \Lloc[p]$ satisfies \begin{equation} \label{eq:defTk} \int_0^{\infty} \Qw{N}{0}(\rho) \, \Np{u}{\rho} \, \frac{d\rho}{\rho} < \infty, \end{equation} then~$T_k u$ is defined almost everywhere and \begin{equation} \label{eq:est_Tk} \Np{T_k u}{r} \leq C \int_0^{\infty} \Qw{N}{0} \left( \frac{\rho}{r} \right) \, \Np{u}{\rho} \, \frac{d\rho}{\rho}, \quad r > 0, \end{equation} for~$k=1,2,\ldots,N+1$, where~$C$ only depends on~$N$ and~$p$. Using Stokes' theorem, it is straightforward to show that~$\S u$ is weakly differentiable if~$u \in \Ck$, and that \begin{equation} \label{eq:diff_Su} \partial_k \S u(x) = (1-N)\bigl( T_k u(x) + \partial_k \ensuremath{\varphi} (x)T_{N+1} u(x) \bigr) \text{,} \quad x \in \ensuremath{\mathbf{R}}^N \text{, } \end{equation} for~$k =1,2,\ldots,N$. Furthermore, one can show that~$\Ckc$ is dense in the Banach space defined by~{\rm(}\ref{eq:defTk}{\rm)}, so inequality~{\rm(}\ref{eq:est_Tk}{\rm)} and~(\ref{eq:diff_Su}) imply that~$\partial_k \S u$ is defined for~$u \in \Xp$ and given by~{\rm(}\ref{eq:diff_Su}{\rm)}. \noindent To simplify the notation, let~$\ensuremath{\omega}(y) = \ensuremath{\sqrt{1 + |\nabla \varphi(y)|^2}}$ for~$y \in \ensuremath{\mathbf{R}}^N$. We wish to approximate~$\S u$ with a Riesz potential. Put \[ \ensuremath{\psi}(y) = \frac{|y|^{N+1}\ensuremath{\omega}(y)}{(|y|^2 + \ensuremath{\varphi}(y)^2)^{(N+1)/2}}, \quad y \in \ensuremath{\mathbf{R}}^N \setminus \{ 0 \}. \] We define the operators~$\ensuremath{\Ie^{\psif}}$ and~$\ensuremath{R_k^{\psif}}$ by \[ \ensuremath{\Ie^{\psif}} u = \ensuremath{\mathcal{I}}(\ensuremath{\psi} u), \quad u \in \Xp, \] and \[ \ensuremath{R_k^{\psif}} u = R_k(\ensuremath{\psi} u), \quad u\text{ satisfying~(\ref{eq:defTk})}. \] For smooth~$u$, it follows from the fact that~$\pks{x} \ensuremath{\Ie^{\psif}} u = (1-N) c_N^{-1} \ensuremath{R_k^{\psif}} u$ (see Stein~\cite[p.~126]{stein}) and~(\ref{eq:diff_Su}) that \begin{equation} \label{eq:diff_S_I_1u} \begin{aligned} \partial_k (\ensuremath{\Ie^{\psif}} - \S )u &= (N-1) \bigl( T_k u + \partial_k \ensuremath{\varphi} T_{N+1} u - c_N^{-1} \ensuremath{R_k^{\psif}} u \bigr) \\ &= (N-1) \bigl( ( T_k u - c_N^{-1} \ensuremath{R_k^{\psif}} u ) + \partial_k \ensuremath{\varphi} T_{N+1} u \bigr) \text{.} \end{aligned} \end{equation} Equation~(\ref{eq:diff_S_I_1u}) also holds for~$u \in \Xp$. Indeed,~(\ref{eq:est_Tk}) and the corresponding estimate for~$\ensuremath{R_k^{\psif}}$ imply that~(\ref{eq:diff_S_I_1u}) remains valid since~$C^{\infty}_c(\ensuremath{\mathbf{R}}^N)$ is dense in the involved spaces. \begin{lemma} \label{l:S_I_bound_L2} Let~$r > 0$ and define~$B_r = B(0 \, ; \, 2r)$. There exists a constant~$C$, depending only on~$N$, such that for every~$u \in L^2(B_r)$, that is supported in~$B_r$, \begin{enumerate} \item[{\rm (i)}] $\| (T_{k} - c_N^{-1} \ensuremath{R_k^{\psif}}) u \|_{L^2(B_r)} \leq C \lc[r]^2 \, \| u \|_{L^2(B_r)}${\rm;} \item[{\rm (ii)}] $\| T_{N+1} u \|_{L^2(B_r)} \leq C \lc[r] \, \| u \|_{L^2(B_r)}${\rm;} \end{enumerate} \end{lemma} \begin{proof} Put~$\ensuremath{\widetilde{\psi}}(y) = \ensuremath{\psi}(y)\,\ensuremath{\omega}(y)^{-1}$ for~$y \in \ensuremath{\mathbf{R}}^N \setminus \{ 0 \}$. Define the operator~$T$ by \[ Tu = T_k u - c_N^{-1} R_k ( \ensuremath{\widetilde{\psi}} u ) \] for~$u \in L^2(B_r)$. Let~$F_1$ be the function \[ F_1(z) = \frac{(1+z^2)^{(N+1)/2}-1}{(1+z^2)^{(N+1)/2}}, \quad z \in \ensuremath{\mathbf{C}} \setminus \{ \pm i \} \text{.} \] Then~$F_1$ is analytic in the band~$\{ \, z \in \ensuremath{\mathbf{C}} \sa |\mathrm{Im} \, z| < 1 \, \}$ and \[ \begin{aligned} T_k u(x) - c_N^{-1} \ensuremath{R_k^{\psif}} u(x) = {} & T u(x) + \int_{\ensuremath{\mathbf{R}}^N} \frac{x_k - y_k}{|x-y|^{N+1}} \left( 1 - \ensuremath{\widetilde{\psi}}(y) \right) u(y) \, \ensuremath{\omega}(y) \, dy\\ = {} & -\int_{\ensuremath{\mathbf{R}}^N} \frac{x_k - y_k}{|x-y|^{N+1}} F_1\biggl(\frac{\ensuremath{\varphi}(x) - \ensuremath{\varphi}(y)}{|x-y|} \biggr) u(y) \, \ensuremath{\omega}(y) \, dy\\ & + \int_{\ensuremath{\mathbf{R}}^N} \frac{x_k - y_k}{|x-y|^{N+1}} \left( 1 - \ensuremath{\widetilde{\psi}}(y) \right) \ensuremath{\omega}(y) \, dy \end{aligned} \] for~$1 \leq k \leq N$. First, we consider~$T u$. According to McShane's extension theorem (see~\cite{mcshane}), we may assume that~$\ensuremath{\varphi}$ is Lipschitz on~$\ensuremath{\mathbf{R}}^N$ with Lipschitz constant~$L = \lc[r]$ since we only need to consider~$T u (x)$ for~$x \in B_r$ and~$u$ has its support in~$B_r$ as well. Let~$\gamma > 0$ and define \begin{equation} \label{eq:def_gamma} A = \{ \, z \in \ensuremath{\mathbf{C}} \sa |\mathrm{Re} \, z| \leq L(1+\gamma) \text{, } |\mathrm{Im} \, z| \leq L\gamma \, \} \text{.} \end{equation} We let~$\Gamma$ denote the rectangular boundary of the set~$A$ in~(\ref{eq:def_gamma}) and assume that~$L = \lc[r]$ is sufficiently small, e.g.,~$L(1 + 2\gamma) \leq 1/2$, so that~$F_1$ is analytic in a neighbourhood of the set in~(\ref{eq:def_gamma}). If~$\phi \sa \ensuremath{\mathbf{R}} \rightarrow \ensuremath{\mathbf{R}}$ is Lipschitz with Lipschitz-constant~$L$ and~$K$ is the Calder\'on-Zygmund kernel (see Stein~\cite{steinH}, Section~1.5) \[ K(s,t) = \frac{1}{s-t} F_1\biggl( \frac{\phi(s) - \phi(t)}{s-t} \biggr) \text{,} \quad s,t \in \ensuremath{\mathbf{R}} \text{,} \quad s \neq t \text{,} \] and~$V$ is the corresponding principal value operator: \begin{equation} \label{eq:pvo_tf} V g(t) = \pval \int_{-\infty}^{\infty} K(t,s) \, g(s) \, ds \text{,} \quad t \in \ensuremath{\mathbf{R}} \text{,} \end{equation} it follows from well-known results for singular integral operators on Lipschitz curves that the operator~$V$ is bounded on~$L^2(\ensuremath{\mathbf{R}})$: \begin{equation} \label{eq:onedim_Tw} \| V w \|_{L^2(\ensuremath{\mathbf{R}})} \leq C (1 + 1/\gamma)^{3/2} \, \sup_{\omega \in \Gamma} |F(\omega)| \, \| w \|_{L^2(\ensuremath{\mathbf{R}})} \text{,} \quad w \in L^2(\ensuremath{\mathbf{R}}) \text{,} \end{equation} where~$C$ is independent of~$\gamma$ and~$F$. The boundedness is a result by Calder\'on~\cite{calderon} for small~$L$ and Coifman, McIntosh, and Meyer~\cite{coifman} in the general case. The constant in~(\ref{eq:onedim_Tw}) can be derived from the argument presented in Dahlberg~\cite[pp.~47--49]{dahlberg} together with the optimal estimate given in David~\cite{david} for Cauchy integrals on Lipschitz curves. Employing the method of rotations and~(\ref{eq:onedim_Tw}) (see Dahlberg~\cite[pp.~49--50]{dahlberg} for the details), it follows that \[ \| T u \|_{L^2(\ensuremath{\mathbf{R}}^N)} \leq C \, \sup_{z \in \Gamma} |F_1(z)| \| u \|_{L^2(\ensuremath{\mathbf{R}}^N)} \] for~$u \in L^2(\ensuremath{\mathbf{R}}^N)$. It can be verified directly that~$|F_1(z)| \leq C \, \lc[r]^2$ for~$z \in \Gamma$. Moreover, since \[ |1 - \ensuremath{\widetilde{\psi}}(y) | \ensuremath{\omega}(y) \leq C \, |\nabla \ensuremath{\varphi}(y)|^2 \leq C \, \lc[r]^2, \quad y \in B_r, \] we obtain that \[ \biggl( \int_{\ensuremath{\mathbf{R}}^N} \! \left| \int_{\ensuremath{\mathbf{R}}^N} \frac{x_k - y_k}{|x-y|^{N+1}} \left( 1 - \ensuremath{\widetilde{\psi}}(y) \right) u(y) \, \ensuremath{\omega}(y) \, dy \right|^2 dx \biggr)^{1/2} \leq C \, \lc[r]^2 \| u \|_{L^2(\ensuremath{\mathbf{R}}^N)}. \] Thus, we have proved the inequality in~(i). To prove~(ii), we utilise the same method but with the function~$F_2(z) = z(1+z^2)^{-(N+1)/2}$,~$z \in \ensuremath{\mathbf{C}} \setminus \{ \pm i \}$, instead of~$F_1$. \end{proof} \begin{corollary} \label{c:S_I_bound_Lp} Let\/~$1 < p < \infty$,~$r > 0$, and define~$B_r = B(0 \, ; \, 2r)$. Then there exists a constant~$C$, which only depends on~$N$ and~$p$, such that \begin{enumerate} \item[{\rm (i)}] $\| (T_{k} - c_N^{-1} \ensuremath{R_k^{\psif}}) u \|_{L^p(B_r)} \leq C \lc[r]^2 \, \| u \|_{L^p(B_r)}${\rm;} \item[{\rm (ii)}] $\| T_{N+1} u \|_{L^p(B_r)} \leq C \lc[r] \, \| u \|_{L^p(B_r)}${\rm;} \end{enumerate} for every function~$u \in L^p(B_r)$ that is supported in~$B_r$. \end{corollary} \begin{proof} By Lemma~\ref{l:S_I_bound_L2}, we know that these estimates hold for~$p=2$. The operators involved have Calder\'on--Zygmund kernels of the type \[ K(x,y) = \frac{x_k-y_k}{|x-y|^{N+1}} F \left( \frac{\ensuremath{\varphi}(x) - \ensuremath{\varphi}(y)}{|x-y|} \right) \text{,} \quad x \neq y \text{,} \] where~$F$ is analytic in some (complex) neighbourhood of the interval~$I = [-L,L]$ and~$L$ is the Lipschitz constant of~$\ensuremath{\varphi}$. Indeed, these kernels satisfy the properties in Section~1.5 of Stein~\cite{steinH}. In particular, for~$c > 1$ and~$\delta > 0$, there exists a constant~$D$ such that \begin{equation*} \sup_{|z-y| < \delta} \; \int_{ |x-y| \geq c\delta} |K(x,y) - K(x,z)| \, dx \leq D. \quad \end{equation*} Here, $ D = C \bigl( \|F\|_{L^{\infty}(I)} + \lc[r] \| F' \|_{L^{\infty}(I)} \bigr), $ where~$C$ only depends on~$c$ and~$N$, and~$I = [-\lc[r],\lc[r]]$. The function~$F$ is one of the two functions~$F_1$ and~$F_2$ in the proof of Lemma~\ref{l:S_I_bound_L2}. As in the proof of that lemma, we can assume that~$\ensuremath{\varphi}$ is Lipschitz continuous with constant~$\lc[r]$ since we only estimate the operator on~$B_r$ and~$u$ has its support in~$B_r$ as well. Since \[ |F_1'(s)| \leq (N+1)|s| \qquad \text{and} \qquad |F_2'(s)| \leq 1 \] for~$s \in \ensuremath{\mathbf{R}}$, the result now follows from Marcinkiewicz interpolation and the weak~$L^1$-estimate that can be derived from the~$L^2$-estimate. Indeed, the $L^p$-norm for the interpolated operator can be shown to have the form~$C \max \{ \, B, \, D_2 \, \}$, with~$B$ being the~$L^2$-norm of the operator and~$C$ depending only on~$N$ and~$p$. \end{proof} \subsection{Estimate of~\boldmath{$\ensuremath{\Ie^{\psif}} - \S$}} To simplify the notation, we introduce the following quotients: \[ \ensuremath{L_x} = \frac{\ensuremath{\varphi}(x)}{|x|}, \qquad \ensuremath{L_y} = \frac{\ensuremath{\varphi}(y)}{|y|}, \quad \text{and} \quad \ensuremath{L_{xy}} = \frac{\ensuremath{\varphi}(x) - \ensuremath{\varphi}(y)}{|x-y|}, \] defined for~$x \neq 0$,~$y \neq 0$, and~$x \neq y$, respectively. It is clear that \[ |\ensuremath{L_x}| \leq \ensuremath{\Lambda}(2^{-1}|x|), \quad |\ensuremath{L_y}| \leq \ensuremath{\Lambda}(2^{-1}|y|), \quad |\ensuremath{L_{xy}}| \leq \max\{ \ensuremath{\Lambda}(2^{-1}|x|) \, , \; \ensuremath{\Lambda}(2^{-1}|y|) \}. \] Furthermore, let us define~$\kernd{x}{y}$ by \[ \kernd{x}{y} = \frac{\ensuremath{\psi}(y)}{|x-y|^{N-1}} - \frac{\ensuremath{\sqrt{1 + |\nabla \varphi(y)|^2}}}{|\Phi(x) - \Phi(y)|^{N-1}}, \quad x,y \in \ensuremath{\mathbf{R}}^N, \quad x \neq y. \] This is the kernel in the operator~$\ensuremath{\Ie^{\psif}} - \S$. Let us collect some properties of this kernel. \begin{lemma} \label{l:est_LxyLy} If\/ $|x| < 2|y|$, then\/ $|\ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2| \leq C\ensuremath{\Lambda}(2^{-1}|x|)^2$, and if\/ $|y| > 2|x|$, then\/ $|\ensuremath{L_{xy}}^2-\ensuremath{L_y}^2| \leq C \ensuremath{\Lambda}(|y|)^2|x|/|y|$. \end{lemma} \begin{proof} First, observe that \[ \begin{aligned} \ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2 &= \ensuremath{L_y}^2 \frac{2x\cdot y - |x|^2}{|x-y|^2} - 2 \frac{\ensuremath{\varphi}(x)\ensuremath{\varphi}(y)}{|x-y|^2} + \frac{\ensuremath{\varphi}(x)^2}{|x-y|^2}. \end{aligned} \] Hence, if~$|x| > 2|y|$, then $ |\ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2| \leq 8 \ensuremath{L_y}^2 + 8|\ensuremath{L_x}\ensuremath{L_y}||y|/|x| + 4\ensuremath{L_x}^2 \leq 16 \ensuremath{\Lambda}(2^{-1}|x|)^2 $ and if~$|y| < 2|x|$, then $ |\ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2| \leq 6 \ensuremath{\Lambda}(2^{-1}|y|)^2|x|/|y|. $ \end{proof} \begin{lemma} \label{l:est_kernd} Suppose that\/~$|x| > 2|y|$ or\/~$|y| < 2|x|$. Then there exists a constant\/~$C$, depending only on\/ $N$, such that\/ $|\kernd{x}{y}| \leq C \ensuremath{\psi}(y)\ensuremath{L_{xy}}^2|x-y|^{-(N-1)}$ and, if also\/ $\ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2 \geq -1/2$, \[ \left| \frac{\partial}{\partial x_k} \kernd{x}{y} \right| \leq C \frac{\ensuremath{\psi}(y)}{|x-y|^N} \bigl( |\ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2| + \ensuremath{\Lambda}(2^{-1}|x|)|\ensuremath{L_{xy}}| \bigr). \] \end{lemma} \begin{proof} We rewrite~$\kernd{x}{y}$ and use the triangle inequality to obtain that \[ \begin{aligned} |\kernd{x}{y}| &= \frac{\ensuremath{\psi}(y)}{|x-y|^{N-1}} \left| 1 - \frac{1}{\psi(y)} \frac{1}{(1 + \ensuremath{L_{xy}}^2)^{(N-1)/2}} \right|\\ &\leq \frac{\ensuremath{\psi}(y)}{|x-y|^{N-1}} \left( \left| 1 - \frac{1}{(1 + \ensuremath{L_{xy}}^2)^{(N-1)/2}} \right| + \frac{\bigl| (1 + \ensuremath{L_y}^2)^{(N+1)/2} - 1 \bigr|}{(1 + \ensuremath{L_{xy}}^2)^{(N-1)/2}} \right)\\ &\leq C \frac{\ensuremath{\psi}(y) \bigl( \ensuremath{L_{xy}}^2 + \ensuremath{L_y}^2 \bigr)}{|x-y|^{N-1}}. \end{aligned} \] Moreover, since \[ \begin{aligned} \frac{\ensuremath{\sqrt{1 + |\nabla \varphi(y)|^2}}}{|\ld[x] - \ld[y]|^{N+1}} &= \frac{\ensuremath{\psi}(y)}{|x-y|^{N+1}} \biggl(1 + \frac{\ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2}{1+\ensuremath{L_y}^2} \biggr)^{-(N+1)/2}, \end{aligned} \] we obtain that \[ \pks{x} \kernd{x}{y} = \frac{(1-N) \ensuremath{\psi}(y)}{|x-y|^{N+1}} \biggl( \frac{x_k - y_k + \partial_k \ensuremath{\varphi} (x)\bigl(\ensuremath{\varphi}(x) - \ensuremath{\varphi}(y)\bigr)} {\bigl(1 + \ensuremath{a(x,y)} \bigr)^{(N+1)/2}} - \bigl( x_k - y_k \bigr) \biggr), \] where~$\ensuremath{a(x,y)} = (\ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2)/(1+\ensuremath{L_y}^2)$. Thus, \[ \begin{aligned} |\pks{x} \kernd{x}{y} | &\leq \frac{(N-1)\ensuremath{\psi}(y)}{|x-y|^{N}} \biggl( \biggl| \frac{1}{(1+\ensuremath{a(x,y)})^{(N+1)/2}} - 1 \biggr| + |\partial_k \ensuremath{\varphi}(x)\ensuremath{L_{xy}}| \biggr) \\ &\leq C \frac{\ensuremath{\psi}(y)}{|x-y|^{N}} \bigl( |\ensuremath{L_{xy}}^2 - \ensuremath{L_y}^2| + \ensuremath{\Lambda}(2^{-1}|x|)|\ensuremath{L_{xy}}| \bigr), \\ \end{aligned} \] which is the estimate we want. \end{proof} \noindent We now have the tools necessary to derive the following important inequality. \begin{theorem} \label{t:est_diff_I_1_S} Suppose that\/~$u \in \Ckc$. Then \begin{equation} \label{eq:est_diff_I_1_S} \begin{aligned} \Np{\pks{x} (\ensuremath{\Ie^{\psif}} - \S)u}{r} & \leq C \int_0^{\infty} \Ewt[r]{\rho} \, \Np{u}{\rho} \, \frac{d\rho}{\rho} \end{aligned} \end{equation} for~$k=1,2,\ldots,N$ and~$r > 0$, where~$\Ewt[r]{\rho}$ is defined by \[ \Ewt[r]{\rho} = \begin{cases} r^{-N} \rho^N \lc[r]^2, & 0 < \rho \leq r,\\ \lc[r] \bigl( \lc[r] r \rho^{-1} + \lc[\rho] \bigr), & \rho > r.\\ \end{cases} \] The constant~$C$ only depends on~$N$ and~$p$. \end{theorem} \begin{proof} We introduce two cut-off functions~$\ensuremath{\eta_{\pm}} \in C^{\infty}(0,\infty)$ satisfying \[ \ensuremath{\eta_{+}}(t) = \begin{cases} 1, & 0 \leq t \leq \frac{1}{8},\\ 0, & t \geq \frac{1}{4}, \end{cases} \qquad \text{and} \qquad \ensuremath{\eta_{-}}(t) = \begin{cases} 0, & 0 \leq t \leq 4,\\ 1, & t \geq 8. \end{cases} \] Put~$\ensuremath{\eta_0} = 1 - \ensuremath{\eta_{-}} - \ensuremath{\eta_{+}}$, so \[ \ensuremath{\eta_0}(t) = \begin{cases} 1, & \frac{1}{4} \leq t \leq 4,\\ 0, & 0 \leq t \leq \frac{1}{8} \text{ or } t \geq 8. \end{cases} \] Using these cut-off functions, we split the integral in three parts: \[ \begin{aligned} (\ensuremath{\Ie^{\psif}} - \S)u(x) &= \int_{\ensuremath{\mathbf{R}}^N} \kernd{x}{y} \biggl( \ensuremath{\eta_{+}} \biggl(\ensuremath{\frac{|x|}{|y|}} \biggr) + \ensuremath{\eta_{-}}\bgp{\ensuremath{\frac{|x|}{|y|}}} + \ensuremath{\eta_0}\bgp{\ensuremath{\frac{|x|}{|y|}}} \biggr) u(y) \, dy\\ &= \ensuremath{J_{+}}(x) + \ensuremath{J_{-}}(x) + \ensuremath{J_{0}}(x). \end{aligned} \] In~$\ensuremath{J_{\pm}}(x)$, the kernel is smooth, so \[ \begin{aligned} \pks{x} \ensuremath{J_{\pm}}(x) &= \int_{\ensuremath{\mathbf{R}}^N} \biggl( \pk{x} \kernd{x}{y} \biggr) \ensuremath{\eta_{\pm}}\bgp{\ensuremath{\frac{|x|}{|y|}}} u(y) \, dy + \ensuremath{\Xi_{\pm}}(x), \end{aligned} \] where \[ \ensuremath{\Xi_{\pm}}(x) = \int_{\ensuremath{\mathbf{R}}^N} \kernd{x}{y} \ensuremath{\eta_{\pm}}'\bgp{\ensuremath{\frac{|x|}{|y|}}} \frac{x_k/|x|}{|y|} u(y) \, dy. \] It is clear from Lemmas~\ref{l:est_kernd} and~\ref{l:est_LxyLy} that \[ \begin{aligned} \Nppp{\pks{x} \ensuremath{J_{\pm}}}{r} &\leq \Nppp{\int_{\ensuremath{\mathbf{R}}^N} \ensuremath{\eta_{\pm}}\bgp{\ensuremath{\frac{|x|}{|y|}}} \biggl| \pk{x} \kernd{x}{y} \biggr| |u(y)| \, dy}{r} + \Np{\ensuremath{\Xi_{\pm}}}{r}\\ &\leq C \Nppp{\int_{\ensuremath{\mathbf{R}}^N} \frac{g_{\pm}(x,y) |u(y)|}{\max\{{|y|^N, |x|^N\}}} \, dy}{r} + \Np{\ensuremath{\Xi_{\pm}}}{r}, \end{aligned} \] where \[ g_{-}(x,y) = \ensuremath{\Lambda}^2\bgp{\hlf{|y|}} \ensuremath{\frac{|x|}{|y|}} + \ensuremath{\Lambda}\bgp{\hlf{|x|}} \ensuremath{\Lambda} \bgp{\hlf{|y|}}, \quad \ensuremath{\frac{|x|}{|y|}} \leq \frac{1}{4}, \] and \[ g_{+}(x,y) = \ensuremath{\Lambda}^2\bgp{\hlf{|x|}}, \quad \ensuremath{\frac{|x|}{|y|}} \geq 4, \] and~$g_{\pm}(x,y) = 0$ elsewhere. Suppose that~$r \leq |x| < 2r$ and~$0 \leq a \leq b \leq \infty$. Minkowski's inequality implies that \[ \Nppp{\int_{a \leq \ensuremath{\frac{|x|}{|y|}} \leq b} \frac{g_{\pm}(x,y) |u(y)|}{\max\{{|y|^N, |x|^N\}}} \, dy}{r} \leq \int_{ar \leq |y| \leq 2br} \frac{\Np{g_{\pm}(\, \cdot \,,y)}{r} |u(y)|}{\max\{|y|^N,r^N\}} \, dy, \] Since~$\ensuremath{\Lambda}$ is increasing, \[ \Np{g_{-}(\, \cdot \,,y)}{r} \leq C \int_{|y|/2}^{|y|} \biggl( \ensuremath{\Lambda}^2(\rho) \frac{r}{\rho} + \ensuremath{\Lambda}(\rho)\ensuremath{\Lambda}(r) \biggr) \, \frac{d\rho}{\rho} , \quad y \in \ensuremath{\mathbf{R}}^N, \] and \[ \Np{g_{+}(\, \cdot \,,y)}{r} \leq C \ensuremath{\Lambda}(r)^2, \quad y \in \ensuremath{\mathbf{R}}^N. \] Now, if~$G$ is a measurable function, it is true that \[ \begin{aligned} \int_{a \leq |y| \leq b} \biggl( \int_{|y|/2}^{|y|} G(\rho) \, \frac{d\rho}{\rho} \biggr) |u(y)| \, dy &= \int_{a}^{b} h(s) s^{N} \biggl( \int_{s/2}^{s} G(\rho) \, \frac{d\rho}{\rho} \biggr) \, \frac{ds}{s}\\ &\leq \int_{a/2}^{b} G(\rho) \int_{\rho}^{2\rho} s^{N} h(s) \, \frac{ds}{s} \, \frac{d\rho}{\rho}\\ &= \int_{a/2}^{b} G(\rho) \, \Np[1]{u}{\rho} \, \frac{d\rho}{\rho},\\ \end{aligned} \] where~$h(s) = \int_{S^{N-1}} |u(s\theta)| \, d\theta$ for~$s > 0$, and the inequality follows from changing the order of integration and over-estimating the domain of integration. Thus, \begin{equation} \label{eq:est_pkjm} \begin{aligned} \Nppp{\pks{x} \ensuremath{J_{-}}}{r} \leq {} & \int_{2r}^{\infty} \biggl( \ensuremath{\Lambda}(\rho)^2 \frac{r}{\rho} + \ensuremath{\Lambda}(\rho)\ensuremath{\Lambda}(r) \biggr) \, \Np{u}{\rho} \, \frac{d\rho}{\rho} + \Np{\ensuremath{\Xi_{-}}}{r} \end{aligned} \end{equation} and \begin{equation} \label{eq:est_pkjp} \begin{aligned} \Nppp{\pks{x} \ensuremath{J_{+}}}{r} \leq {} & \int_0^{r/2} \biggl( \frac{\rho}{r}\biggr)^N \ensuremath{\Lambda}^2(r) \, \Np{u}{\rho} \, \frac{d\rho}{\rho} + \Np{\ensuremath{\Xi_{+}}}{r}. \end{aligned} \end{equation} According to Lemma~\ref{l:est_kernd}, we can estimate the terms~$\ensuremath{\Xi_{\pm}}(x)$ by \[ |\ensuremath{\Xi_{-}}(x)| \leq C \int_{\frac{1}{8} \leq \ensuremath{\frac{|x|}{|y|}} \leq \frac{1}{4}} \frac{|\kernd{x}{y}|}{|y|} |u(y)| dy \leq C \int_{\frac{1}{8} \leq \ensuremath{\frac{|x|}{|y|}} \leq \frac{1}{4}} \frac{|u(y)|\ensuremath{L_{xy}}^2}{|y|^N} \, dy \] and \[ |\ensuremath{\Xi_{+}}(x)| \leq C \int_{4 \leq \ensuremath{\frac{|x|}{|y|}} \leq 8} \frac{|\kernd{x}{y}|}{|y|} |u(y)| dy \leq C \int_{4 \leq \ensuremath{\frac{|x|}{|y|}} \leq 8} \frac{|u(y)|\ensuremath{L_{xy}}^2}{|x||y|^{N-1}} \, dy, \] and analogously with the derivation of~(\ref{eq:est_pkjm}) and~(\ref{eq:est_pkjp}) above, we can then obtain a bound for~$\Np{\ensuremath{\Xi_{\pm}}}{r}$: \begin{equation} \label{eq:est_pkjrests_p} \Np{\ensuremath{\Xi_{+}}}{r} \leq \lc[r]^2 \int_{r/16}^{r/2} \Np{u}{\rho} \, \frac{d\rho}{\rho} \leq C \int_0^{\infty} \Ewt[r]{\rho} \, \Np{u}{\rho} \, \frac{d\rho}{\rho} \end{equation} and \begin{equation} \label{eq:est_pkjrests_m} \Np{\ensuremath{\Xi_{-}}}{r} \leq \lc[r]^2 \int_{2r}^{16r} \Np{u}{\rho} \, \frac{d\rho}{\rho} \leq C \int_0^{\infty} \Ewt[r]{\rho} \, \Np{u}{\rho} \, \frac{d\rho}{\rho}. \end{equation} Turning our attention to~$\ensuremath{J_{0}}$, we let~$B_r = \{ y \in \ensuremath{\mathbf{R}}^N \sa r/2 \leq |y| \leq 4r \}$ and~$\chi_{B_r}$ be the corresponding characteristic function. For~$r \leq |x| < 2r$, \[ \begin{aligned} \ensuremath{J_{0}}(x) = {} & \int_{\ensuremath{\mathbf{R}}^N} \kernd{x}{y} \chi_{B_r}(y)u(y) \, dy \\ & + \int_{\ensuremath{\mathbf{R}}^N} \ensuremath{\eta_0}\bgp{\ensuremath{\frac{|x|}{|y|}}} \kernd{x}{y} (1 - \chi_{B_r}(y)) u(y) \, dy\\ = {} & \ensuremath{J_{0}^{\mbox{\scriptsize s}}}(x) + \ensuremath{J_{0}^{\mbox{\scriptsize ns}}}(x). \end{aligned} \] The integrand in~$\ensuremath{J_{0}^{\mbox{\scriptsize ns}}}(x)$ is smooth since it is only nonzero when~$2|y| \leq |x| \leq 8|y|$ or~$2|x| \leq |y| \leq 8|x|$. Hence, \[ \begin{aligned} \pks{x} \ensuremath{J_{0}^{\mbox{\scriptsize ns}}} = {} & \int_{\ensuremath{\mathbf{R}}^N} \ensuremath{\eta_0}'\biggl( \ensuremath{\frac{|x|}{|y|}} \biggr) \frac{x_k/|x|}{|y|} \kernd{x}{y} (1 - \chi_{B_r}(y))u(y) \, dy\\ &+ \int_{\ensuremath{\mathbf{R}}^N} \ensuremath{\eta_0}\biggl( \ensuremath{\frac{|x|}{|y|}} \biggr) \frac{x_k/|x|}{|y|} \frac{\partial}{\partial x_k} \kernd{x}{y} (1 - \chi_{B_r}(y))u(y) \, dy. \end{aligned} \] Using the same argument as above, we obtain that \begin{equation} \label{eq:est_pkjcns} \begin{aligned} \Np{\pks{x} \ensuremath{J_{0}^{\mbox{\scriptsize ns}}}}{r} \leq {} & C \Nppp{\mathop{\int_{r/8 \leq |y| \leq r/2}}_{4r \leq |y| \leq 16r} \biggl( \frac{|\kernd{x}{y}|}{|y|} + | \pks{x} \kernd{x}{y}| \biggr) |u(y)| dy}{r}\\ \leq {} & C \int_0^{\infty} \Ewt[r]{\rho} \, \Np{u}{\rho} \, \frac{d\rho}{\rho}. \end{aligned} \end{equation} Moreover,~(\ref{eq:diff_Su}) implies that \[ \pks{x} \ensuremath{J_{0}^{\mbox{\scriptsize s}}} = (1-N)\bigl( T_k( \chi_{B_r} u) - c_N^{-1} \ensuremath{R_k^{\psif}} (\chi_{B_r} u ) + \partial_k \ensuremath{\varphi} T_{N+1} (\chi_{B_r} u) \bigr), \] and an application of Corollary~\ref{c:S_I_bound_Lp} shows that \[ \Np{\pks{x} \ensuremath{J_{0}^{\mbox{\scriptsize s}}}}{r} \leq C r^{-N/p} \lc[r]^2 \| u \chi_{B_r} \|_{L^p(\ensuremath{\mathbf{R}}^N)}. \] Thus, it is clear that \begin{equation} \label{eq:est_pkjcs} \Np{\pks{x} \ensuremath{J_{0}^{\mbox{\scriptsize s}}}}{r} \leq C \int_0^{\infty} \Ewt[r]{\rho} \, \Np{u}{\rho} \, \frac{d\rho}{\rho}. \end{equation} The estimates in~(\ref{eq:est_pkjm}),~(\ref{eq:est_pkjp}),~(\ref{eq:est_pkjrests_p}),~(\ref{eq:est_pkjrests_m}),~(\ref{eq:est_pkjcns}), and~(\ref{eq:est_pkjcs}), imply that the desired result holds. \end{proof} \begin{remark} Analogously with the authors' proof of Lemma 3.1 in~{\rm\cite{thim1}}, it is possible to show that the to\/~$\partial_k (\ensuremath{\Ie^{\psif}} - \S)u(x)$ corresponding ``truncated'' operator\/~$T_{\epsilon} u(x)$ converges both pointwise and in\/~$\Lloc[p]$ to the right-hand side of~{\rm(}\ref{eq:diff_S_I_1u}{\rm)} if\/~$u \in \Xp$ for\/~$1 < p < \infty$. \end{remark} \section{Reduction to a Fixed Point Problem} \label{s:red_fixpt} \subsection{The Fixed Point Equation} Suppose that~$(\ensuremath{\Ie^{\psif}} - \S)u$ and~$f$ both belong to~$\ensuremath{Y^{1,p}_{0}(\R^N)}$. Then, by Theorem~\ref{t:inv}, \[ \ensuremath{\Ie^{\psif}} \bigl( \ensuremath{\psi}^{-1} \ensuremath{\mathcal{R}} ( (\ensuremath{\Ie^{\psif}} - \S)u + f) ) = (\ensuremath{\Ie^{\psif}} - \S)u + f \text{.} \] Formally, let~$\K$ be the operator given by \begin{equation} \label{eq:def_K} \K(u) = \ensuremath{\psi}^{-1} \ensuremath{\mathcal{R}} ((\ensuremath{\Ie^{\psif}} - \S)u + f) \text{.} \end{equation} If~$u$ is a fixed point of~$\K$, then \[ \ensuremath{\Ie^{\psif}} u = \ensuremath{\Ie^{\psif}} \K(u) = (\ensuremath{\Ie^{\psif}} - \S)u + f \text{.} \] Thus,~$u$ is a solution to~(\ref{eq:maineq}). To find a solution to~$\K(u) = u$, we will employ the following fixed point theorem. We refer to a previous article by the authors~\cite{thim1} for details and proofs. \subsection{A Fixed Point Theorem in Locally Convex Spaces} \label{s:fixpt} We let~$\X$ denote a locally convex topological space, where the topology is given by a family~$\{ \pa{\,\cdot\,} \}_{\alpha \in \I}$ of seminorms that separates points. We want to solve the equation \begin{equation} \label{eq:fixpt_eq} \K(u) = u \text{,} \quad u \in \DK \text{,} \end{equation} where~$\K \colon \DK \rightarrow \DK$ is a mapping defined on a subset~$\DK$ of~$\X$. We assume that~$0 \in \DK$ and that there exists an auxiliary linear operator~$\KK \colon \DKK \rightarrow \ensuremath{\R^{\I}}$, where~$\DKK \subset \ensuremath{\R^{\I}}$ is a linear subspace. By~$\ensuremath{\R^{\I}}$ we denote the set of all real-valued functions on~$\I$, endowed with the topology of pointwise convergence. \subsubsection*{Existence of Fixed Points} The operator~$\KK$ is subject to the following assumptions. \begin{enumerate}[label=(\KK\arabic{*}), ref=(\KK\arabic{*})] \item \label{i:kkpos} {\it Positivity.} The operator~$\KK$ is positive, i.e., if~$\eta \in \DKK$ is non-negative, then~$\KK \eta \geq 0$. \item \label{i:kkz} {\it Fixed point inequality.} The function~$\KZ(\,\cdot\,) = \pa[\,\cdot\,]{\K(0)} \in \DKK$, and there exists a non-negative function~$z \in \DKK$ such that \begin{equation} \label{eq:tvsp_zineq} z(\alpha) \geq \KK z(\alpha) + \KZ(\alpha) \text{,} \quad \alpha \in \Omega \text{.} \end{equation} \item \label{i:kkclosed} {\it Monotone closedness.} The operator~$\KK$ is closed for non-negative, increasing sequences: if~$\{ \eta_n \}$ is a non-negative sequence in~$\DKK$ such that~$\eta_n$ increases to~$\eta$, where~$\eta \leq z$, and~$\KK \eta_n \rightarrow \zeta$, then~$\eta \in \DKK$ and~$\KK \eta = \zeta$. \item \label{i:kkdomain} {\it Invariance.} If~$\eta \in \DKK$ is non-negative and~$\eta \leq z$, then~$\KK \eta \in \DKK$. \end{enumerate} The existence of a non-negative solution~$z$ to~(\ref{eq:tvsp_zineq}) enables us to prove the existence of a non-negative solution to the equation \begin{equation} \label{eq:tvsp_sigmaeq} \sigma(\alpha) = \KK \sigma(\alpha) + \KZ(\alpha) \text{,} \quad \alpha \in \Omega \text{,} \end{equation} which is minimal in the sense that if~$\eta \in \DKK$ is another non-negative solution to~{\rm(}\ref{eq:tvsp_sigmaeq}{\rm)}, then~$\sigma \leq \eta$; see Lemma~1 in~\cite{thim2}. Suppose that the operator~$\K$ maps~$\DK$ into~$\DK$. We let~$\sigma$ be the minimal solution to~(\ref{eq:tvsp_sigmaeq}), and put \[ \DKS = \{ u \in \DK \sa \pa{u} \leq \sigma(\alpha) \text{ for every } \alpha \in \I \} \text{.} \] We require the following properties to hold. \begin{enumerate}[label=(\ensuremath{\K}\arabic{*}), ref=(\ensuremath{\K}\arabic{*})] \item \label{i:DKS_est_K_KK} {\it Subordination to~$\KK$.} If~$u,v$ belong to~$\DKS$, then~$\pa[\,\cdot\,]{u-v}$ belongs to~$\DKK$, and we have \begin{equation} \label{eq:est_K_KK} \pa{\K(u) - \K(v)} \leq \KK\apa{u - v}(\alpha) \text{,} \quad \alpha \in \I \text{.} \end{equation} \item \label{i:DKS_closed} {\it Closedness of~$\DKS$.} If~$\{v_k\}_{k=0}^{\infty}$ is a sequence in~$\DKS$ such that~$v_0 = 0$ and \begin{equation} \label{eq:sumcond} \sum_{k=0}^{\infty} \pa{v_{k+1}-v_k} \leq \sigma(\alpha) \text{,} \quad \alpha \in \I \text{,} \end{equation} then the limit of~$v_k$ exists and belongs to~$\DKS$. \end{enumerate} Since~$0 \in \DKS$,~\ref{i:DKS_est_K_KK} implies that~$\pa[\,\cdot\,]{u} \in \DKK$ for~$u \in \DK$. \begin{theorem} \label{t:tvsp_fixpt} Suppose that\/~$\KK$ satisfies~\ref{i:kkpos} to~\ref{i:kkdomain} and that\/~$\K$ satisfies~\ref{i:DKS_est_K_KK} and~\ref{i:DKS_closed}. Then there exists a fixed point of\/~$\K$ in\/~$\DKS$. \end{theorem} \subsubsection*{Uniqueness of Fixed Points} Suppose that the operator~$\K$ maps~$\DK$ into itself. We assume that the following conditions hold. \begin{enumerate}[label=(\Roman{*}), ref=(\Roman{*})] \item \label{i:KKnu} If~$u \in \DK$, then~$\KK^{n} \apa{u}$ is defined and belongs to~$\DKK$ for~$n=1,2,\ldots$, and~$\lim_{n \rightarrow \infty} \KK^{n} \apa{u} = 0$. \item \label{i:KKfp} If~$u \in \DK$ and~$\eta \in \DKK$ satisfy~$0 \leq \eta(\alpha) \leq \pa{u}$ for every~$\alpha \in \I$, then~$\KK \eta$ belongs to~$\DKK$. \item \label{i:DK_est_K_KK} If~$u,v$ belong to~$\DK$, then the function~$\pa[\,\cdot\,]{u-v}$ belongs to~$\DKK$, and~(\ref{eq:est_K_KK}) holds. \end{enumerate} \begin{theorem} \label{t:tvsp_fixptu} Suppose that the operators\/~$\K$ and\/~$\KK$ satisfy~\ref{i:kkpos} and~\ref{i:KKnu} to~\ref{i:DK_est_K_KK}, respectively. Then there exists at most one fixed point of\/~$\K$ in\/~$\DK$. \end{theorem} \subsection{Construction of~\boldmath{$\K$} and~\boldmath{$\KK$}} \label{s:fixpt_defs} To apply the fixed point theorem, we define an operator~$\K$ formally by~(\ref{eq:def_K}) and an auxiliary linear operator~$\KK$, and verify the properties in~\ref{i:kkpos}--\ref{i:kkdomain} along with~\ref{i:DKS_est_K_KK} and~\ref{i:DKS_closed}. The locally convex space~$\X$ will be~$\Lloc[p]$ with seminorms~$\pa{\,\cdot\,}$ given by~$\Np{\,\cdot\,}{r}$,~$r > 0$. Let \[ \Ew[r]{\rho}{\xi} = \Qw{N}{0}\left(\frac{\rho}{r}\right) \Ewt[\rho]{\xi} \text{,} \quad r,\rho,\xi > 0 \text{,} \] where~$\Ewt[\rho]{\xi}$ is defined as in Theorem~\ref{t:est_diff_I_1_S}. We define the linear operator~$\KK$ by \begin{equation} \label{eq:def_kk_r} \begin{aligned} \KK \zeta(r) &= \ensuremath{C_K} \int_0^{\infty} \!\! \int_0^{\infty} \Ew[r]{\rho}{\xi} \, \zeta(\xi) \, \frac{d\xi}{\xi} \, \frac{d\rho}{\rho} \text{,} \quad r > 0 \text{,} \end{aligned} \end{equation} with domain~$\DKK$ given by those measurable~$\zeta$ that satisfy \begin{equation} \label{eq:cond_domK} \Ewn{|\zeta(\xi)|} < \infty \text{.} \end{equation} The constant~$\ensuremath{C_K}$ is the one given in~(\ref{eq:def_ck}) below. Since \[ \Ew[r]{\rho}{\xi} \leq \max \bigl\{ \, 1, \, r^{-N} \, \bigr\} \Ew[1]{\rho}{\xi} \text{,} \quad r,\rho,\xi > 0 \text{,} \] it follows that~$\KK \zeta$ is defined if~$\zeta \in \DKK$. This of course implies that~$\KK \zeta(r)$ is finite for every~$r > 0$ if~$\zeta \in \DKK$. By~$\DK$, we denote the set of those~$u \in \Lloc[p]$ such that \begin{equation} \label{eq:cond_domK_Np} \Ewn{\Np{u}{\xi}} < \infty \text{.} \end{equation} This condition is equivalent to requiring that~$\Np{u}{\, \cdot \,} \in \DKK$. Obviously~$\K(u)$ is defined for~$u \in \Ckc$, and the following two lemmas show that~$\K$ can be extended continuously from~$\Ckc$ to~$\DK$. \begin{lemma} \label{l:Npu_DKK_Yone} Let\/~$u \in \Ckc$. Then\/~$(\ensuremath{\Ie^{\psif}} - \S)u \in \ensuremath{Y^{1,p}(\R^N)}$ and \begin{equation} \label{eq:est_RIu} \Np{\ensuremath{\mathcal{R}} ((\ensuremath{\Ie^{\psif}} - \S)u)}{r} \leq \KK ( \Np{u}{\, \cdot \,} )(r) \text{,} \quad r > 0 \text{.} \end{equation} \end{lemma} \begin{proof} Since~$u$ has compact support,~$u \in \Xp$. Hence,~$(\ensuremath{\Ie^{\psif}} - S)u$ is defined and~$\partial_k (\ensuremath{\Ie^{\psif}} - \S)u$,~$k=1,\ldots,N$, is given by~(\ref{eq:diff_S_I_1u}). It is a straightforward calculation to verify that~$(\ensuremath{\Ie^{\psif}} - \S)u$ belongs to~$\ensuremath{Y^{1,p}(\R^N)}$ by using~(\ref{eq:est_diff_I_1_S}) and changing the order of integration. By Theorem~\ref{t:cont_R},~$\ensuremath{\mathcal{R}}$ is defined on~$\ensuremath{Y^{1,p}(\R^N)}$ and \begin{equation} \label{eq:def_ck} \begin{aligned} \Np{\ensuremath{\mathcal{R}} ((\ensuremath{\Ie^{\psif}} - \S)u)}{r} &\leq C \, \int_0^{\infty} \Qw{N}{0} \left(\frac{\rho}{r}\right) \Npp{\nabla \bigl((\ensuremath{\Ie^{\psif}} - \S)u\bigr)}{\rho} \, \frac{d\rho}{\rho}\\ & \leq \ensuremath{C_K} \int_0^{\infty} \! \! \int_0^{\infty} \Ew[r]{\rho}{\xi} \, \Np{u}{e^{-\xi}} \, \frac{d\xi}{\xi} \, \frac{d\rho}{\rho}\\ &= \KK(\Np{u}{\, \cdot \,})(r) \end{aligned} \end{equation} for every~$r > 0$, where we used Theorem~\ref{t:est_diff_I_1_S} in the last inequality. \end{proof} \begin{corollary} Let\/~$u,v \in \Ckc$ and\/~$f \in \ensuremath{Y^{1,p}(\R^N)}$. Then \begin{equation} \label{eq:NpKuKv} \Np{\K(u) - \K(v)}{r} \leq \KK(\Np{u-v}{\, \cdot \,})(r) \text{,} \quad r > 0 \text{.} \end{equation} \end{corollary} \begin{lemma} \label{l:ext_to_B} Let\/~$f \in \ensuremath{Y^{1,p}(\R^N)}$. The operator\/~$\K$ in~{\rm(}\ref{eq:def_K}{\rm)} can be extended to\/~$\DK$ so that~{\rm(}\ref{eq:NpKuKv}{\rm)} holds for all\/~$u,v \in \DK$. \end{lemma} \begin{proof} Let~$\ensuremath{\mathscr{B}}$ be the space of functions in~$\DK$ with topology defined by the norm \begin{equation} \label{eq:norm_in_B} \| u \|_{\ensuremath{\mathscr{B}}} = \Ewn{\Np{u}{\xi}} \text{.} \end{equation} This is a Banach space and one can check that~$\Ckc$ is a dense subspace of~$\ensuremath{\mathscr{B}}$. The operator~$\ensuremath{\mathcal{R}} ((\ensuremath{\Ie^{\psif}} - \S))$ is defined for~$u \in \Ckc$ and \begin{equation} \label{eq:A_bnd} \Np{\ensuremath{\mathcal{R}} ((\ensuremath{\Ie^{\psif}} - \S)u)}{r} \leq C \max \{1,\, r^{-N} \} \, \| u \|_{\ensuremath{\mathscr{B}}} \text{.} \end{equation} By density, this allows us to extend~$\ensuremath{\mathcal{R}} ((\ensuremath{\Ie^{\psif}} - \S))$ uniquely to all of~$\ensuremath{\mathscr{B}}$ so that~(\ref{eq:A_bnd}) holds for all~$u$ in~$\ensuremath{\mathscr{B}}$. Obviously this gives an extension of the operator~$\K$ to~$\DK$ as well. \end{proof} \subsection{Verification of~($\mathbf{K2}$)} \label{s:aux_eq} Next, we are going to show that~\ref{i:kkz} holds. More specifically, we will prove that there exists a function~$z \in \DKK$ such that \begin{equation} \label{eq:z} \begin{aligned} z(r) \geq {} & \ensuremath{C_K} \biggl( \int_0^{\infty} \!\! \int_0^{\infty} \Ew[r]{\rho}{\xi} \, z(\xi) \, \frac{d\xi}{\xi} \, \frac{d\rho}{\rho} + \int_0^{\infty} \Qw{N}{0}\left(\frac{\rho}{r}\right) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} \biggr) \end{aligned} \end{equation} for every~$r > 0$. This will imply that~\ref{i:kkz} is satisfied since~$\Np{\K(0)}{r}$ is bounded by the second term in the right-hand side of~(\ref{eq:z}) for~$r > 0$. \newcommand{\ensuremath{\zeta}}{\ensuremath{\zeta}} We construct a solution in the following manner. Let~$\ensuremath{c_1},\,\ensuremath{c_3},\,$ and~$\ensuremath{c_2}$ be positive constants depending only on~$N$ and~$p$. We require that~$\ensuremath{c_2} \ensuremath{\Lambda_0} \leq (N-1)/2$ and~$\ensuremath{c_1} \ensuremath{\Lambda_0} \leq 1/2$. Put~$\ensuremath{M} = N - \ensuremath{c_2} \ensuremath{\Lambda_0}$ and~$\ensuremath{\lambda}(\nu) = \ensuremath{\Lambda}(\exp(-\nu))$ for~$\nu \in \ensuremath{\mathbf{R}}$. The function~$\ensuremath{{v}}$ is given by \begin{equation} \label{eq:lu} \begin{aligned} \ensuremath{{v}}(t) = \ensuremath{c_3} \int_{-\infty}^{t} \exp\biggl( \ensuremath{c_1} \int_{s}^{t} \ensuremath{\lambda}(\nu) \, d\nu \biggr) \ensuremath{\zeta}(s) \, ds + \ensuremath{c_3} \int_t^{\infty} \exp \bigl( \ensuremath{M}(t-s) \bigr) \ensuremath{\zeta}(s) \, ds, \end{aligned} \end{equation} where~$\ensuremath{\zeta}(s) = \Np{\nabla f}{\exp(-s)}$ for~$s \in \ensuremath{\mathbf{R}}$. We require that~$\ensuremath{\zeta}$ satisfies \begin{equation} \label{eq:req_lu} \int_{-\infty}^{0} \exp\biggl( \ensuremath{c_1} \int_{s}^{0} \ensuremath{\lambda}(\nu) \, d\nu \biggr) \ensuremath{\zeta}(s) \, ds + \int_0^{\infty} \exp \bigl( -\ensuremath{M} s \bigr) \ensuremath{\zeta}(s) \, ds < \infty. \end{equation} It is clear that if~$\ensuremath{\zeta}$ satisfies~\ref{eq:req_lu}, then~$\ensuremath{{v}}(t)$ is finite for every~$t \in \ensuremath{\mathbf{R}}$. Moreover, if~$f \in \ensuremath{Y^{1,p}_{M}(\R^N)}$, then~(\ref{eq:req_lu}) is also valid. \newcommand{\intlam}[2]{\ensuremath{\exp \biggl( \ensuremath{c_1} \int_{{#1}}^{{#2}} \ensuremath{\lambda}(\nu) \, d\nu \biggr)}} \newcommand{\ensuremath{D}}{\ensuremath{D}} We change variables in the definition of the operator~$\KK$ to~$r = e^{-t}$, $\rho = e^{-\tau}$, and~$s = e^{-\sigma}$, where~$t,\tau,\sigma \in \ensuremath{\mathbf{R}}$. We allow earlier functions of the variables~$r,\rho,s$ to keep the same name when it is clear from the context what we are referring to. The action of~$\KK$ on~$\ensuremath{{v}}$ can be expressed with help from the functions \begin{equation} \ensuremath{E}(\tau, \sigma) = \Biggl\{ \begin{aligned} &\ensuremath{\lambda}^2(\tau) \exp(N(\tau - \sigma)), & \tau \leq \sigma, \\ &\ensuremath{\lambda}(\tau)\bigl( \ensuremath{\lambda}(\tau)\exp(\sigma - \tau) + \ensuremath{\lambda}(\sigma) \bigr), & \tau \geq \sigma,\\ \end{aligned} \end{equation} and \begin{equation} \label{eq:Awtpm} \ensuremath{\Sigma^{\pm}}(s, \sigma) = \left\{ \begin{aligned} &\exp \biggl( \pm \ensuremath{c_1} \int_s^{\sigma} \ensuremath{\lambda}(\nu) \, d\nu \biggr) , & s \leq \sigma, \\ &\exp \bigl( \pm \ensuremath{M}(\sigma - s) \bigr), & s \geq \sigma.\\ \end{aligned} \right. \end{equation} To see why, observe that \[ \begin{aligned} \KK \ensuremath{{v}}(t) &= C_K \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} E(t,\tau,\sigma) \ensuremath{{v}}(\sigma) \, d\sigma \, d\tau\\ &= C_K \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} E(t,\tau,\sigma) \int_{-\infty}^{\infty} \ensuremath{\Sigma^{+}}(s,\sigma) \ensuremath{\zeta}(s) \, ds \, d\sigma \, d\tau\\ &= C_K \int_{-\infty}^{\infty} \ensuremath{\zeta}(s) \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} \Qw{N}{0}(e^{t-\tau}) \ensuremath{E}(\tau,\sigma) \ensuremath{\Sigma^{+}}(s,\sigma) \, d\sigma \, d\tau \, ds.\\ \end{aligned} \] The following lemma provides estimates we need. \begin{lemma} \label{l:est_kern_Ku} Let\/~$s,t \in \ensuremath{\mathbf{R}}$. Then \begin{equation} \label{eq:est_kern_Ku} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} E(t,\tau,\sigma) \ensuremath{\Sigma^{+}}(s,\sigma) \, d\sigma \, d\tau \leq c \ensuremath{\Sigma^{+}}(s,t), \end{equation} and \begin{equation} \label{eq:est_kern_Ku_adj} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} E(t,\tau,\sigma) \ensuremath{\Sigma^{-}}(\sigma,s) \, d\sigma \, d\tau \leq c \ensuremath{\Sigma^{-}}(t,s), \end{equation} where \[ c = 2\Biggl( \frac{3\ensuremath{\Lambda_0}}{\ensuremath{M}} + \frac{1}{\ensuremath{c_1}} + \frac{1}{\ensuremath{c_2}} \Biggr)^2. \] \end{lemma} \begin{proof} First, we prove that, for~$\tau \in \ensuremath{\mathbf{R}}$, \begin{equation} \label{est:ES1} \int_{-\infty}^{\infty} \ensuremath{E}(\tau,\sigma) \ensuremath{\Sigma^{+}}(s,\sigma) \, d\sigma \leq 2 \ensuremath{\lambda}(\tau) \Biggl( \frac{\ensuremath{\Lambda_0}}{\ensuremath{M}} + \frac{1}{\ensuremath{c_1}} + \frac{1}{\ensuremath{c_2}} + \frac{\ensuremath{\lambda}(\tau)}{N} \Biggr) \ensuremath{\Sigma^{+}}(s,\tau). \end{equation} To simplify the notation, we let \begin{equation} \label{eq:def_intl} \ensuremath{D}(a,b) = \intlam{a}{b}, \quad a,b \in \ensuremath{\mathbf{R}}. \end{equation} Let~$s \leq \tau$. Then \begin{equation} \label{eq:aa1} \begin{aligned} \int_{-\infty}^s \ensuremath{E}(\tau,\sigma) \ensuremath{\Sigma^{+}}(s,\sigma) \, d\sigma = {} & \int_{-\infty}^s e^{\ensuremath{M}(\sigma - s)} \ensuremath{\lambda}(\tau) \bigl( \ensuremath{\lambda}(\tau)e^{\sigma - \tau} + \ensuremath{\lambda}(\sigma) \bigr) \, d\sigma\\ \leq {} & \ensuremath{\lambda}(\tau)^2 \frac{e^{s - \tau}}{\ensuremath{M} + 1} + \frac{\ensuremath{\Lambda_0}}{\ensuremath{M}} \ensuremath{\lambda}(\tau) \leq \frac{2\ensuremath{\Lambda_0}}{\ensuremath{M}} \ensuremath{\lambda}(\tau) \end{aligned} \end{equation} and \begin{equation} \label{eq:aa2} \begin{aligned} \int_{\tau}^{\infty} \ensuremath{E}(\tau,\sigma) \ensuremath{\Sigma^{+}}(s,\sigma) \, d\sigma = {} & \int_{\tau}^{\infty} \ensuremath{\lambda}^2(\tau) e^{N(\tau - \sigma)}\ensuremath{D}(s,\sigma) \, d\sigma\\ \leq {} & \ensuremath{\lambda}(\tau)^2 \int_{\tau}^{\infty} \frac{1}{N - \ensuremath{c_1}\ensuremath{\lambda}(\sigma)} \cdot \frac{\partial}{\partial \sigma} \bigl( -e^{N(\tau - \sigma)}\ensuremath{D}(s,\sigma) \bigr) d\sigma \\ \leq {} & \frac{2\ensuremath{\lambda}^2(\tau)}{N} \ensuremath{D}(s,\tau), \end{aligned} \end{equation} since~$(\partial/\partial \sigma) \bigl( e^{N(\tau - \sigma)}\ensuremath{D}(s,\sigma) \bigr) \leq 0$ and~$\ensuremath{c_1} \ensuremath{\Lambda_0} \leq 1/2$. Furthermore, using estimates similar to the one used in~(\ref{eq:aa2}), we obtain that \[ \begin{aligned} \int_{s}^{\tau} \ensuremath{E}(\tau,\sigma) \ensuremath{\Sigma^{+}}(s,\sigma) \, d\sigma = {} & \ensuremath{\lambda}^2(\tau) \int_{s}^{\tau} e^{\sigma - \tau}\ensuremath{D}(s,\sigma) \, d\sigma + \ensuremath{\lambda}(\tau) \int_{s}^{\tau} \ensuremath{\lambda}(\sigma) \ensuremath{D}(s,\sigma) \, d\sigma\\ \leq {} & \ensuremath{\lambda}^2(\tau) \bigl( \ensuremath{D}(s,\tau) - e^{s - \tau} \bigr) + \frac{\ensuremath{\lambda}(\tau)}{\ensuremath{c_1}} \bigl( \ensuremath{D}(s,\tau) - 1 \bigr)\\ \leq {} & \frac{2\ensuremath{\lambda}(\tau)}{\ensuremath{c_1}} \ensuremath{D}(s,\tau). \end{aligned} \] If~$s \geq \tau$, using the same estimates as in~(\ref{eq:aa1}) and~(\ref{eq:aa2}) (where the limits~$s$ and~$\tau$ changes positions), we obtain \[ \int_{\ensuremath{\mathbf{R}} \setminus [\tau,s]} \ensuremath{E}(\tau,\sigma)\ensuremath{\Sigma^{+}}(s,\sigma) \, d\sigma \leq \frac{2\ensuremath{\Lambda_0}\ensuremath{\lambda}(\tau)}{\ensuremath{M}} e^{\ensuremath{M}(\tau - s)} + \frac{2\ensuremath{\lambda}^2(\tau)}{N} e^{N(\tau - s)}. \] Moreover, \[ \begin{aligned} \int_{\tau}^s \ensuremath{E}(\tau,\sigma)\ensuremath{\Sigma^{+}}(s,\sigma) \, d\sigma \leq {} & e^{N(\tau - s)} \ensuremath{\lambda}^2(\tau) \int_{\tau}^s e^{-\ensuremath{c_2} \ensuremath{\Lambda_0}(\sigma - s)} \, d\sigma \leq \frac{\ensuremath{\lambda}(\tau)}{\ensuremath{c_2}} e^{\ensuremath{M}(\tau - s)}. \end{aligned} \] In total, we obtain the estimate in~(\ref{est:ES1}) for all~$s,\tau \in \ensuremath{\mathbf{R}}$. Let us show that, for~$s,t \in \ensuremath{\mathbf{R}}$, \begin{equation} \label{est:ES2} \int_{-\infty}^{\infty} \ensuremath{\lambda}(\tau) \ensuremath{\Sigma^{+}}(s,\tau) \Qw{N}{0}(e^{t-\tau}) \, d\tau \leq \Biggl( \frac{3\ensuremath{\Lambda_0}}{\ensuremath{M}} + \frac{1}{\ensuremath{c_1}} + \frac{1}{\ensuremath{c_2}} \Biggr) \ensuremath{\Sigma^{+}}(s,t). \end{equation} Let~$s \leq t$. Then \[ \int_{-\infty}^{s} \ensuremath{\lambda}(\tau) \ensuremath{\Sigma^{+}}(s,\tau) \Qw{N}{0}(e^{t-\tau}) \, d\tau = \int_{-\infty}^{s} \ensuremath{\lambda}(\tau) e^{\ensuremath{M}(\tau - s)} \, d\tau \leq \frac{\ensuremath{\Lambda_0}}{\ensuremath{M}}, \] and, similarly with~(\ref{eq:aa2}), \[ \begin{aligned} \int_{t}^{\infty} \ensuremath{\lambda}(\tau) \ensuremath{\Sigma^{+}}(s,\tau) \Qw{N}{0}(e^{t-\tau}) \, d\tau &= \int_t^{\infty} \ensuremath{\lambda}(\tau) e^{N(t - \tau)}\ensuremath{D}(s,\tau) \, d\tau\\ \leq {} & \ensuremath{\Lambda_0} \int_t^{\infty} \frac{1}{N - \ensuremath{c_1} \ensuremath{\lambda}(\tau)} \biggl( - \frac{\partial}{\partial \tau} \bigl( e^{N(t-\tau)}\ensuremath{D}(s,\tau) \bigr) \biggr) d\tau\\ \leq {} & \frac{2\ensuremath{\Lambda_0}}{N} \ensuremath{D}(s,t). \end{aligned} \] For the middle part, \[ \int_{s}^{t} \ensuremath{\lambda}(\tau) \ensuremath{\Sigma^{+}}(s,\tau) \Qw{N}{0}(e^{t-\tau}) \, d\tau = \int_{s}^{t} \ensuremath{\lambda}(\tau) \ensuremath{D}(s,\tau) \, d\tau \leq \frac{1}{\ensuremath{c_1}} \ensuremath{D}(s,t). \] If~$s \geq t$, then \[ \int_{\ensuremath{\mathbf{R}} \setminus [t,s]} \ensuremath{\lambda}(\tau) \ensuremath{\Sigma^{+}}(s,\tau) \Qw{N}{0}(e^{t-\tau}) \, d\tau \leq \Biggl( \frac{\ensuremath{\Lambda_0}}{\ensuremath{M}} + \frac{2\ensuremath{\Lambda_0}}{N} \Biggr) e^{\ensuremath{M}(t - s)} \] and \[ \begin{aligned} \int_{t}^s \ensuremath{\lambda}(\tau) \ensuremath{\Sigma^{+}}(s,\tau) \Qw{N}{0}(e^{t-\tau}) \, d\tau = {} & \int_t^s \ensuremath{\lambda}(\tau) e^{N(t-s) - \ensuremath{c_2}\ensuremath{\Lambda_0}(\tau - s)} \, d\tau\\ \leq {} & \frac{\ensuremath{\Lambda_0} e^{N(t-s)}}{\ensuremath{\Lambda_0} \ensuremath{c_2}} \biggl( e^{-\ensuremath{c_2} \ensuremath{\Lambda_0}(t - s)} - 1 \biggr) \leq \frac{1}{\ensuremath{c_2}} e^{\ensuremath{M}(t - s)}. \end{aligned} \] Thus, we obtain~(\ref{est:ES2}) for all~$s$ and~$t$ in~$\ensuremath{\mathbf{R}}$. It is clear that~(\ref{est:ES1}) and~(\ref{est:ES2}) imply~(\ref{eq:est_kern_Ku}) given in the lemma. The inequality in~(\ref{eq:est_kern_Ku_adj}) can be derived analogously, using the same technique as above. \end{proof} \begin{lemma} \label{l:lu_solves} There exists positive constants\/~$\ensuremath{\Lambda_{*}}$,\/~$\ensuremath{c_1}$,\/~$\ensuremath{c_2}$, and\/~$\ensuremath{c_3}$, depending only on\/~$N$ and\/~$p$, such that if\/~$\ensuremath{\Lambda_0} \leq \ensuremath{\Lambda_{*}}$ and\/~$\ensuremath{\zeta}(s) = \Np{\nabla f}{e^{-s}}$,\/~$s \in \ensuremath{\mathbf{R}}$, satisfies~{\rm(}\ref{eq:req_lu}{\rm)}, then \[ \KK \ensuremath{{v}}(t) + \Np{\K(0)}{e^{-t}} \leq \ensuremath{{v}}(t), \quad t \in \ensuremath{\mathbf{R}}, \quad 1 < p < \infty. \] \end{lemma} \begin{proof} Let~$\ensuremath{\Lambda_{*}}$ be sufficiently small for Lemma~\ref{l:S_I_bound_L2} and Corollary~\ref{c:S_I_bound_Lp} to hold when~$\ensuremath{\Lambda_0} \leq \ensuremath{\Lambda_{*}}$. Moreover, let~$\ensuremath{c_1}\ensuremath{\Lambda_{*}} \leq 1/2$ and~$\ensuremath{c_2} \ensuremath{\Lambda_{*}} \leq (N-1)/2$. It follows from Lemma~\ref{l:est_kern_Ku} that \[ \KK \ensuremath{{v}}(t) \leq 2C_K \Biggl( \frac{3\ensuremath{\Lambda_{*}}}{\ensuremath{M}} + \frac{1}{\ensuremath{c_1}} + \frac{1}{\ensuremath{c_2}} \Biggr)^2 u(t). \] Furthermore,~$\K(0) = \ensuremath{\psi}^{-1} \ensuremath{\mathcal{R}} f$, and from Theorem~\ref{t:cont_R}, we know that \[ \begin{aligned} \Np{\K(0)}{e^{-t}} &\leq C_K \int_{-\infty}^{\infty} \Qw{N}{0}(e^{t-s}) \, \ensuremath{\zeta}(s) \, ds \\ &\leq C_K \int_{-\infty}^{\infty} \ensuremath{\Sigma^{+}}(s,t) \, \ensuremath{\zeta}(s) \, ds = \frac{C_K}{\ensuremath{c_3}} u(t). \end{aligned} \] Thus, \[ \KK \ensuremath{{v}}(t) +\Np{\K(0)}{e^{-t}} \leq C_K \Biggl( 2\Biggl( \frac{3\ensuremath{\Lambda_{*}}}{\ensuremath{M}} + \frac{1}{\ensuremath{c_1}} + \frac{1}{\ensuremath{c_2}} \Biggr)^2 + \frac{1}{\ensuremath{c_3}} \Biggr) u(t). \] By choosing~$\ensuremath{c_1}$ and~$\ensuremath{c_3}$ large, and~$\ensuremath{c_2}$ large enough but smaller than~$N/(2\ensuremath{\Lambda_{*}})$, we can see that it is possible to bound the constant by \begin{equation} \label{eq:krav_ci} C_K \Biggl( 2\Biggl( \frac{3\ensuremath{\Lambda_{*}}}{\ensuremath{M}} + \frac{1}{\ensuremath{c_1}} + \frac{1}{\ensuremath{c_2}} \Biggr)^2 + \frac{1}{\ensuremath{c_3}} \Biggr) \leq 1. \end{equation} \end{proof} \begin{lemma} \label{l:z_solves} Suppose that\/~$f \in \ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$ satisfies \begin{equation} \label{eq:req_f_z} \int_{0}^{1} \! \rho^{\ensuremath{M}} \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} + \int_1^{\infty} \exp\biggl( \ensuremath{c_1} \int_{1}^{\rho} \ensuremath{\Lambda}(\nu) \, \frac{d\nu}{\nu} \biggr) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} < \infty, \end{equation} where\/~$1 < p < \infty$ and the constants\/~$\ensuremath{\Lambda_{*}}$,\/~$\ensuremath{c_1}$,\/~$\ensuremath{c_2}$, and\/~$\ensuremath{c_3}$, are given by Lemma~\ref{l:lu_solves}. Then the function\/~$z$ defined by\/~$z(r) = \ensuremath{{v}}(\log r)$ for\/~$r > 0$ belongs to\/~$\DKK$ and \[ \KK z(r) + \Np{\nabla f}{r} \leq z(r), \quad r > 0. \] \end{lemma} \begin{proof} Since~$f$ satisfies~(\ref{eq:req_f_z}) and~(\ref{eq:krav_ci}) holds, the function~$\ensuremath{{v}}(t)$ exists and solves the inequality in Lemma~\ref{l:lu_solves}. This also implies that \[ \KK z(r) \leq \ensuremath{{v}}(-\log r) < \infty, \quad r > 0. \] In particular,~$\KK z(1) < \infty$, which implies that~$z \in \DKK$; see~(\ref{eq:cond_domK}). \end{proof} \subsection{Verification of~($\mathbf{K1}$),~($\mathbf{K3}$),~($\mathbf{K4}$), (\boldmath{$\mathscr{K}1$}), and~(\boldmath{$\mathscr{K}2$})} \label{s:verif_rest} Obviously~$\KK$ is linear and positive, so~\ref{i:kkpos} holds. Furthermore,~\ref{i:kkclosed} follows from the monotone convergence theorem. If~$\eta \in \DKK$ satisfies~$0 \leq \eta \leq z$, then by~\ref{i:kkpos} and~\ref{i:kkz}, we obtain \[ \KK \eta(r) \leq \KK z(r) \leq z(r) - \KZ(r) \leq z(r) \text{,} \quad r > 0 \text{.} \] Since~$z \in \DKK$, this shows that~\ref{i:kkdomain} holds. Now, since~$(K1)$--$(K4)$ holds, there exists a minimal solution~$\sigma$ in~$\DKK$ to~(\ref{eq:tvsp_sigmaeq}). Suppose that~$u,v \in \DKS$. By Lemma~\ref{l:ext_to_B}, \begin{equation} \label{eq:NpKuKv2} \Np{\K(u) - \K(v)}{r} \leq \KK(\Np{u - v}{\, \cdot \,})(r) \text{,} \end{equation} which is the condition in~\ref{i:DKS_est_K_KK}. Since~$\K(u)$ is measurable and \[ \begin{aligned} \Np{\K(u)}{r} &\leq \Np{\K(u) - \K(0)}{r} + \KZ(r)\\ &\leq \KK \Np{u}{\,\cdot\,}(r) + \KZ(r)\\ &\leq \KK \sigma (r) + \KZ(r)\\ &= \sigma(r) \end{aligned} \] for every~$r > 0$, we see that~$\K$ maps~$\DKS$ into~$\DKS$. Suppose that~$\{v_k\}_{k=0}^{\infty}$ is a sequence in~$\DKS$ that satisfies~(\ref{eq:sumcond}). Then this is a Cauchy sequence in~$\Lloc[p]$, so it converges to a measurable function~$v$. It follows that~$\Np{v}{r} \leq \sigma(r)$ for~$r > 0$, so~$v \in \DKS$. Hence,~\ref{i:DKS_closed} holds. \subsection{Existence of a Fixed Point} \label{s:exist_fixpt} We now apply Theorem~\ref{t:tvsp_fixpt}, which shows that there exists a function~$u$ in~$\DKS$ such that~$\K(u) = u$. We have thus derived the following result. \begin{lemma} \label{l:exist_K} Suppose that the conditions of Lemma~\ref{l:z_solves} are satisfied. Then there exists\/~$u \in \Lloc[p]$ such that\/~$\K(u) = u$ and \begin{equation} \label{eq:est_sol_t} \Np{u}{e^{-t}} \leq \ensuremath{{v}}(t) \text{,} \quad t \in \ensuremath{\mathbf{R}} \text{.} \end{equation} \end{lemma} \noindent The estimate in~(\ref{eq:est_sol_t}) implies the following asymptotic behaviour of the fixed point. \begin{lemma} \label{l:lu_assymp} Let\/~$u \in \Xp$ satisfy~{\rm(}\ref{eq:est_sol_t}{\rm)}. Then\/~$\Np{u}{e^{-t}} = o(\ensuremath{\Sigma^{-}}(t,0))$ as\/~$t \rightarrow \pm \infty$. \end{lemma} \begin{proof} Let~$t < -m < 0$ for some~$m > 0$. Then \[ \frac{\ensuremath{\Sigma^{+}}(\tau,t)}{\ensuremath{\Sigma^{-}}(t,0)} = \begin{cases} \ensuremath{D}(\tau,0), & \tau \leq t,\\ e^{\ensuremath{M}(t-\tau)} \ensuremath{D}(t,0), & \tau > t, \end{cases} \] where~$\ensuremath{D}$ is given by~(\ref{eq:def_intl}), and thus, for every~$\epsilon > 0$, we can choose~$m$ large enough so that for~$t < -m$, \[ \begin{aligned} \frac{\ensuremath{{v}}(t)}{\ensuremath{\Sigma^{-}}(t,0)} \leq {} & \int_{-\infty}^{t} \ensuremath{D}(\tau,0) \, \ensuremath{\zeta}(\tau) \, d\tau + \int_{t}^{-m} e^{\ensuremath{M}(t-\tau)} \ensuremath{D}(t,0) \ensuremath{\zeta}(\tau) \, d\tau \\ & + \int_{-m}^{\infty} e^{\ensuremath{M} t} \ensuremath{D}(t,0) e^{-\ensuremath{M} \tau} \ensuremath{\zeta}(\tau) \, d\tau\\ \leq {} & \frac{\epsilon}{3} + \int_{-\infty}^{-m} e^{\ensuremath{M}(t-\tau)} \ensuremath{D}(t,\tau) \ensuremath{D}(\tau,0) \ensuremath{\zeta}(\tau) \, d\tau + e^{-\beta t} \int_{-m}^{\infty} e^{-\ensuremath{M} \tau} \ensuremath{\zeta}(\tau) \, d\tau\\ \leq {} & \frac{2\epsilon}{3} + \int_{-\infty}^{-m} e^{-\beta(t-\tau)} \ensuremath{D}(\tau,0) \ensuremath{\zeta}(\tau) \, d\tau \leq \epsilon, \end{aligned} \] where~$\beta = N - \ensuremath{c_1} \ensuremath{\Lambda_0} - \ensuremath{c_2} \ensuremath{\Lambda_0} > 0$ if~$\ensuremath{c_1}\ensuremath{\Lambda_0} \leq 1/2$ and~$\ensuremath{c_2}\ensuremath{\Lambda_0} \leq (N-1)/2$. Similarly, one can show that if~$t > 0$, we obtain that~$\ensuremath{{v}}(t)/\ensuremath{\Sigma^{-}}(t,0) \leq \epsilon$ if~$t > m$. \end{proof} \subsection{Uniqueness of Fixed Points} \label{s:uniq} We can now prove the following uniqueness result. \begin{lemma} \label{t:uniq_K} Suppose that the conditions in Lemma~\ref{l:z_solves} are satisfied. Then there exists at most one solution\/~$u$ in\/~$\Xp$ to the equation\/~$\K(u) = u$ such that \begin{equation} \label{eq:cond_uniq} \Np{u}{e^{-t}} = O(\ensuremath{\Sigma^{-}}(t,0)), \quad t \rightarrow \pm \infty. \end{equation} \end{lemma} \begin{proof} To simplify notation, let \[ \pa[t]{u} = \Np{u}{e^{-t}}, \quad t \in \ensuremath{\mathbf{R}}. \] Choose~$\DK$ as the linear space of functions~$u \in \Lloc[p]$ such that~(\ref{eq:cond_uniq}) holds. Suppose that~$u \in \DK$. Then there exists constants~$A' > 0$ and~$m > 0$ such that \[ \pa[t]{u} \leq A' \ensuremath{\Sigma^{-}}(t,0), \quad |t| \geq m. \] Let the constant~$A''$ be given by \[ A'' = \sup_{|t| \leq m} \ensuremath{\Sigma^{+}}(t,0) \pa[t]{u}. \] By continuity, it is clear that~$A'' < \infty$. Define~$A = \max\{\,A',A''\,\}$. We find that \begin{equation} \label{eq:uniq_K1} \begin{aligned} \KK \pa[ \, \cdot \, ]{u}(t) &= C_K \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} E(t,\tau,\sigma) \pa[\sigma]{u} \, d\sigma \, d\tau\\ &\leq A C_K \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} E(t,\tau,\sigma) \ensuremath{\Sigma^{-}}(\sigma,0) \, d\sigma \, d\tau\\ &\leq A C_K c \ensuremath{\Sigma^{-}}(t,0) \end{aligned} \end{equation} for~$t \in \ensuremath{\mathbf{R}}$. The last inequality follows from~(\ref{eq:est_kern_Ku_adj}) in Lemma~\ref{l:est_kern_Ku}, where the constant~$c$ is also defined. Inequality~(\ref{eq:uniq_K1}) implies that if~$u$ belongs to~$\DK$, then~$\pa[\, \cdot \,]{u}$ belongs to~$\DKK$, and thus,~$\K(u)$ is defined. This inequality also implies that~$\K$ maps~$\DK$ into~$\DK$. To apply Theorem~\ref{t:tvsp_fixptu}, we need to verify that~\ref{i:KKnu}--\ref{i:DK_est_K_KK} hold. Inequality~(\ref{eq:uniq_K1}) shows that~\ref{i:KKfp} holds: if~$u$ belongs to~$\DK$ and~$\eta$ belongs to~$\DKK$ such that~$0 \leq \eta(t) \leq \pa[t]{u}$ for~$t \in \ensuremath{\mathbf{R}}$, then the monotonicity of~$\KK$ implies that~$\KK \eta(t) = O(\ensuremath{\Sigma^{-}}(t,0))$ as~$t \rightarrow \pm \infty$. Hence,~$\KK\eta$ belongs to~$\DKK$. The fact that~\ref{i:DK_est_K_KK} holds follows from Lemma~\ref{l:ext_to_B}. Furthermore, by applying~(\ref{eq:uniq_K1})~$n$ times, we obtain that \[ \KK^n \pa[\, \cdot \,]{u}(t) \leq (C_K c)^n A \ensuremath{\Sigma^{-}}(t,0). \] Since~$C_K c < 1$, which follows from~(\ref{eq:krav_ci}), and~$\ensuremath{\Sigma^{-}}(t,0) < \infty$, \[ \KK^n \pa[\, \cdot \,]{u}(t) \rightarrow 0, \quad \text{as } n \rightarrow \infty. \] Hence, Theorem~\ref{t:tvsp_fixptu} implies that the fixed point is indeed unique. \end{proof} \begin{remark} \label{r:soluniq} The condition in~{\rm(}\ref{eq:cond_uniq}{\rm)} is a natural condition if we consider the solution\/~$u \in \DK$ in Lemma~\ref{l:exist_K}, which satisfies\/~$\Np{u}{e^{-t}} \leq \ensuremath{{v}}(t)$, so by Lemma~\ref{l:lu_assymp},~{\rm(}\ref{eq:cond_uniq}{\rm)} is valid. \end{remark} \section{Proof of the Main Results} In the previous section, we proved that there exists a fixed point of~$\K$, and that it is unique under certain conditions on~$f$. We will now use these theorems and results from Section~\ref{s:rieszpot} to prove similar results for solutions to~(\ref{eq:maineq}). \subsection{Existence of Solutions to~(\ref{eq:maineq})} We start with deriving two technical lemmas which will show that we can use Theorem~\ref{t:inv} to recover~$(\ensuremath{\Ie^{\psif}} - \S)u + f$ from the equation~$\K(u) = u$. First we find integrated estimates for~$\ensuremath{\Sigma^{+}}$, where~$\ensuremath{\Sigma^{+}}$ is defined by~(\ref{eq:Awtpm}). \begin{lemma} \label{l:sol_in_Xp_estimates} Suppose that\/~$\ensuremath{c_1} \ensuremath{\Lambda_0} \leq 1/2$ and\/~$\ensuremath{c_2} \ensuremath{\Lambda_0} \leq (N-1)/2$. Then, \begin{equation} \label{eq:sol_in_Xp_cond} \int_{-\infty}^{\infty} \Qw{N}{1}(e^{-t}) \ensuremath{\Sigma^{+}}(s,t) dt \leq \ensuremath{C_{\lcwz}(s)} \, \Qw{\ensuremath{M}}{1} (e^{-s}) \end{equation} and \begin{equation} \label{eq:sol_in_Yonez_cond} \int_{-\infty}^{0} e^{-t} \, \ensuremath{\Sigma^{+}}(s,t) \, dt + \int_0^{\infty} (1+t)e^{-Nt} \, \ensuremath{\Sigma^{+}}(s,t) \, dt \leq \ensuremath{C_{\lcwz}(s)} \, \Qw{\ensuremath{M}}{1}(e^{-s}) \end{equation} for\/~$s \in \ensuremath{\mathbf{R}}$. The function\/~$\ensuremath{C_{\lcwz}(s)}$ will depend on\/~$\ensuremath{\Lambda_0}$ and\/~$s$, but for fixed\/~$\ensuremath{\Lambda_0}$ it is uniformly bounded with respect to\/~$s$, and \begin{equation} \label{eq:lcwz_zero} \limsup_{\ensuremath{\Lambda_0} \rightarrow 0} \ensuremath{C_{\lcwz}(s)} \leq \biggl\{ \begin{array}{ll} C, & s \leq 0, \\ C(1+s), & s > 0, \end{array} \end{equation} where\/~$C$ only depends on\/~$N$ and\/~$p$. \end{lemma} \begin{proof} Using notation from Section~\ref{s:aux_eq}, we show~(\ref{eq:sol_in_Xp_cond}) first. Proceeding similarly with the proof of Lemma~\ref{l:est_kern_Ku}, we consider two cases. First, let~$s \leq 0$. The left-hand side in~(\ref{eq:sol_in_Xp_cond}) is given by \begin{equation} \label{eq:sol_in_Xp_cond_sm} \begin{aligned} &\int_{-\infty}^s e^{-t} e^{\ensuremath{M}(t-s)} \, dt + \int_s^0 e^{-t} \ensuremath{D}(s,t) \, dt + \int_0^{\infty} e^{-Nt} D(s,t) \, dt\\ & \qquad \leq \frac{e^{-s}}{\ensuremath{M} - 1} + 2 e^{-s} + \frac{2}{N} D(s,0) \leq C e^{-s}, \end{aligned} \end{equation} where we exploited that \[ \ensuremath{c_1} \int_s^0 \ensuremath{\lambda}(\nu) \, d\nu \leq - \ensuremath{c_1} \ensuremath{\Lambda_0} s \leq -s, \qquad s < 0, \] so~$D(s,0) \leq e^{-s}$. The constant~$C$ in~(\ref{eq:sol_in_Xp_cond_sm}) only depends on~$N$ and~$p$. Let~$s \geq 0$. In the same manner as above, the left-hand side in~(\ref{eq:sol_in_Xp_cond}) is bounded by \[ \begin{aligned} &\int_{-\infty}^0 e^{-t} e^{\ensuremath{M}(t-s)} \, dt + \int_0^s e^{-Nt} e^{\ensuremath{M}(t-s)} \, dt + \int_s^{\infty} e^{-Nt} D(s,t) \, dt\\ & \qquad \leq \frac{e^{-\ensuremath{M} s}}{\ensuremath{M} - 1} + \frac{e^{-\ensuremath{M} s} - e^{-Ns}}{N - \ensuremath{M}} + \frac{2}{N} e^{-Ns} \leq \left( \frac{4}{N-1} + \frac{1 - e^{-\ensuremath{c_2}\ensuremath{\Lambda_0} s}}{\ensuremath{c_2}\ensuremath{\Lambda_0}} \right) e^{-\ensuremath{M} s}, \end{aligned} \] which completes the proof of~(\ref{eq:sol_in_Xp_cond}) since \[ \limsup_{\ensuremath{\Lambda_0} \rightarrow 0} \left( \frac{4}{N-1} + \frac{1 - e^{-\ensuremath{c_2}\ensuremath{\Lambda_0} s}}{\ensuremath{c_2}\ensuremath{\Lambda_0}} \right) \leq C(1+s). \] To prove~(\ref{eq:sol_in_Yonez_cond}), we proceed similarly. Let~$s \leq 0$. Then the left-hand side in~(\ref{eq:sol_in_Yonez_cond}) is given by \[ \int_{-\infty}^{s} e^{-t} e^{\ensuremath{M}(t-s)} \, dt + \int_s^{0} e^{-t} \ensuremath{D}(s,t) \, dt + \int_0^{\infty} (1+t)e^{-Nt}D(s,t) \, dt. \] The first two terms can be estimated in the same manner as~(\ref{eq:sol_in_Xp_cond_sm}): \[ \int_{-\infty}^{s} e^{-t} e^{\ensuremath{M}(t-s)} \, dt + \int_s^{0} e^{-t} \ensuremath{D}(s,t) \, dt \\ \leq \frac{e^{-s}}{\ensuremath{M} - 1} + 2 e^{-s}. \] To investigate the third term, we use integration by parts: \begin{equation} \label{eq:kpp1_pi1} \begin{aligned} \int_0^{\infty} (1+t)e^{-Nt}D(s,t) \, dt &= \int_0^{\infty} (1+t) \frac{1}{N - \ensuremath{c_1} \ensuremath{\lambda}(t)} \biggl( - \frac{\partial}{\partial t} e^{-Nt} \ensuremath{D}(s,t) \biggr) \, dt\\ &\leq \frac{2}{N} \int_0^{\infty} (1+t) \biggl( - \frac{\partial}{\partial t} e^{-Nt} \ensuremath{D}(s,t) \biggr) \, dt\\ &\leq \frac{2}{N} \ensuremath{D}(s,0) + \frac{4}{N^2} \int_0^{\infty} \biggl( - \frac{\partial}{\partial t} e^{-Nt} \ensuremath{D}(s,t) \biggr) \, dt\\ &\leq C \ensuremath{D}(s,0), \end{aligned} \end{equation} where~$C$ only depends on~$N$ and~$p$. Since~$\ensuremath{D}(s,0) \leq e^{-s/2} \leq e^{-s}$, we obtain that~(\ref{eq:sol_in_Yonez_cond}) holds for~$s \leq 0$. Suppose that~$s \geq 0$. We proceed analogously with the case when~$s \leq 0$. The left-hand side of~(\ref{eq:sol_in_Yonez_cond}) is given by \[ \int_{-\infty}^{0} e^{-t} e^{\ensuremath{M}(t-s)} \, dt + \int_{0}^s (1+t) e^{-Nt} e^{\ensuremath{M}(t-s)} \, dt + \int_s^{\infty} (1+t)e^{-Nt}D(s,t) \, dt. \] The first term can be calculated as \[ \int_{-\infty}^{0} e^{-t} e^{\ensuremath{M}(t-s)} \, dt = \frac{e^{-\ensuremath{M} s}}{\ensuremath{M} - 1}. \] The other terms can be bounded in the same manner as~(\ref{eq:kpp1_pi1}) above using integration by parts: \[ \begin{aligned} \int_{0}^s (1+t) e^{-Nt} e^{\ensuremath{M}(t-s)} \, dt &= e^{-\ensuremath{M} s} \int_0^s (1+t)e^{-\ensuremath{c_2} \ensuremath{\Lambda_0} t} \, dt \\ &= \left( \frac{1 - (1+s)e^{-\ensuremath{c_2} \ensuremath{\Lambda_0} s}}{\ensuremath{c_2} \ensuremath{\Lambda_0}} + \frac{1 - e^{-\ensuremath{c_2}\ensuremath{\Lambda_0} s}}{\ensuremath{c_2}^2\ensuremath{\Lambda_0}^2} \right) e^{-\ensuremath{M} s} \end{aligned} \] and \[ \int_s^{\infty} (1+t)e^{-Nt}D(s,t) \, dt \leq \frac{2}{N}(1+s)e^{-Ns} + \frac{4}{N^2} e^{-Ns}. \] Since \[ \limsup_{\ensuremath{\Lambda_0} \rightarrow 0} \left( \frac{1 - (1+s)e^{-\ensuremath{c_2} \ensuremath{\Lambda_0} s}}{\ensuremath{c_2} \ensuremath{\Lambda_0}} + \frac{1 - e^{-\ensuremath{c_2}\ensuremath{\Lambda_0} s}}{\ensuremath{c_2}^2\ensuremath{\Lambda_0}^2} \right) \leq C(1+s), \quad s > 0, \] and \[ (1+s) e^{-Ns} \leq \frac{C}{\ensuremath{c_2} \ensuremath{\Lambda_0}} e^{-\ensuremath{M} s}, \quad s > 0, \] it follows that~(\ref{eq:sol_in_Yonez_cond}) and~(\ref{eq:lcwz_zero}) hold for~$s \geq 0$ as well. \end{proof} \begin{lemma} \label{l:sol_in_Xp} Let\/~$u \in \Lloc[p]$ satisfy\/~$\Np{u}{e^{-t}} \leq C \, \ensuremath{{v}}(t)$ for\/~$t \in \ensuremath{\mathbf{R}}$ and suppose that\/~$\ensuremath{c_1} \ensuremath{\Lambda_0} \leq 1/2$ and\/~$\ensuremath{c_2} \ensuremath{\Lambda_0} \leq (N-1)/2$. If\/~$f \in \ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$ satisfies \begin{equation} \label{eq:cond_u_Xp_simp} \int_0^{\infty} \Qw{\ensuremath{M}}{1} (\rho) \, \Np{\nabla f}{\rho} \frac{d\rho}{\rho} < \infty \text{,} \end{equation} then\/~$u \in \Xp$ and\/ $(\ensuremath{\Ie^{\psif}} -\S)u \in \ensuremath{Y^{1,p}_{0}(\R^N)}$. \end{lemma} \begin{proof} Using notation from Section~\ref{s:aux_eq}, we see that \begin{equation} \begin{aligned} \| u \|_{\Xp} &= \int_0^{\infty} \Qw{N}{1}(\rho) \, \Np{u}{\rho} \, \frac{d\rho}{\rho} = \int_{-\infty}^{\infty} \Qw{N}{1}(e^{-t}) \, \Np{u}{e^{-t}} \, dt\\ &\leq C \int_{-\infty}^{\infty} \Qw{N}{1}(e^{-t}) \int_{-\infty}^{\infty} \ensuremath{\Sigma^{+}}(s,t) \ensuremath{\zeta}(s) ds \, dt\\ &= C \int_{-\infty}^{\infty} \ensuremath{\zeta}(s) \int_{-\infty}^{\infty} \Qw{N}{1}(e^{-t}) \ensuremath{\Sigma^{+}}(s,t) dt \, ds. \end{aligned} \end{equation} Inequality~(\ref{eq:sol_in_Xp_cond}) now implies that~$u \in \Xp$. Turning our attention to the second part of the lemma, i.e., that~$(\ensuremath{\Ie^{\psif}} - \S) u$ belongs to~$\ensuremath{Y^{1,p}_{0}(\R^N)}$, we need to prove two things: \begin{equation} \label{eq:cond_ipsi_yonez} \int_0^{\infty} H(\rho) \, \Np{\nabla (\ensuremath{\Ie^{\psif}} - \S)u}{\rho} \, \frac{d\rho}{\rho} < \infty \end{equation} and \[ h_R(r) \rightarrow 0, \quad \text{as } r \rightarrow \infty, \] where~$H(\rho) = \rho^N(1 - \log \rho)$ if~$0 < \rho \leq 1$ and~$H(\rho) = \rho$ if~$\rho \geq 1$, and the function~$h(x) = (\ensuremath{\Ie^{\psif}} - \S)u(x)$ for~$x \in \ensuremath{\mathbf{R}}^N$. We know that~$\partial_k (\ensuremath{\Ie^{\psif}} - \S)u$,~$k=1,2,\ldots,N$, are defined and the representation in~(\ref{eq:diff_S_I_1u}) holds. Moreover, by Lemma~\ref{l:ext_to_B} it is possible to extend~(\ref{eq:est_diff_I_1_S}) for all~$u$ in~$\Xp$. Hence, since~$\Np{u}{e^{-t}} \leq C \ensuremath{{v}}(t)$ and~$\KK \ensuremath{{v}} \leq \ensuremath{{v}}$, \[ \Np{\nabla(\ensuremath{\Ie^{\psif}} - \S)u}{r} \leq \KK(\Np{u}{\, \cdot \,})(r) \leq C \KK \ensuremath{{v}} (r) \leq C \ensuremath{{v}}(r), \quad r > 0, \] which implies that \[ \begin{aligned} \int_0^{\infty} H(r) \, \Np{\nabla (\ensuremath{\Ie^{\psif}} - \S)u}{r} \, \frac{dr}{r} &\leq C \, \int_{-\infty}^{\infty} H(e^{-t}) \int_{-\infty}^{\infty} \ensuremath{\zeta}(s) \, \ensuremath{\Sigma^{+}}(s,t) \, ds \, dt\\ &\leq C \, \int_{-\infty}^{\infty} \ensuremath{\zeta}(s) \int_{-\infty}^{\infty} H(e^{-t}) \, \ensuremath{\Sigma^{+}}(s,t) \, dt \, ds.\\ \end{aligned} \] Since~(\ref{eq:sol_in_Yonez_cond}) holds, it follows that~(\ref{eq:cond_ipsi_yonez}) is valid. We will now verify that \begin{equation} \label{eq:lim_hr} \lim_{r \rightarrow \infty} h_R(r) = \lim_{r \rightarrow \infty} \int_{S^{N-1}} h(r\theta) \, d\theta = 0 \text{,} \end{equation} where~$h(x) = (\ensuremath{\Ie^{\psif}} - \S)u(x)$,~$x \in \ensuremath{\mathbf{R}}^N$. By Lemma~2.4 in~\cite{thim1}, the fact that \[ \int_{1}^{\infty} \Np{\nabla h}{\rho} \, d\rho < \infty \] implies that the limit in~(\ref{eq:lim_hr}) exists. We now obtain that \[ \begin{aligned} \inf_{r \leq \rho < 2r} |h_R(\rho)| &\leq \frac{1}{r} \int_r^{2r} |h_R(\rho)| \, d\rho\\ &\leq \frac{1}{r^N} \int_r^{2r} \rho^{N-1} \int_{S^{N-1}} |h(\rho\theta)| \, d\theta \, d\rho\\ & \leq C \, r^{-N/p} \biggl( \int_{r \leq |x| < 2r} |h(x)|^p \, dx \biggr)^{1/p}\\ &= C \, \Np{h}{r} \text{.} \end{aligned} \] Obviously~$|h(x)| \leq C \, \ensuremath{\Ie^{\psif}} |u|(x)$, so~(\ref{eq:Np_est_I_1}) in Theorem~\ref{t:cont_I1} implies that \[ \begin{aligned} \Np{h}{r} \leq C \biggl( &r^{1-N} \int_0^1 \rho^{N} \Np{u}{\rho} \, \frac{d\rho}{\rho} + r^{1-N} \int_1^r \rho^{N} \Np{u}{\rho} \, \frac{d\rho}{\rho}\\ &+ \int_r^{\infty} \Np{u}{\rho} \, d\rho \biggr) \end{aligned} \] for~$r > 1$. Since~$u \in \Xp$, the first and last term in the right-hand side tend to zero as~$r \rightarrow \infty$. To show that this is also true for the middle term, let~$m > 1$ be fixed. Then \[ \limsup_{r \rightarrow \infty} r^{1-N} \int_1^r \rho^{N} \, \Np{u}{\rho} \, \frac{d\rho}{\rho} \leq \int_m^{\infty} \Np{u}{\rho} \, d\rho \text{.} \] The number~$m$ is arbitrary, so the limit must be zero since~$u \in \Xp$. Thus,~$h_R(\rho_n) \rightarrow 0$ for some sequence~$\rho_n$ such that~$\rho_n \rightarrow \infty$. Hence~(\ref{eq:lim_hr}) must hold, so~$(\ensuremath{\Ie^{\psif}} - \S)u \in \ensuremath{Y^{1,p}_{0}(\R^N)}$. \end{proof} \subsection{Proof of Theorem~\ref{t:i:exist}} \label{s:proof_exist} If~$f \in \ensuremath{Y^{1,p}_{M}(\R^N)}$, then the solution~$z$ to~(\ref{eq:z}) exists. Applying Lemma~\ref{l:exist_K}, we obtain a fixed point~$u \in \Lloc[p]$ of~$\K$ such that \[ \Np{u}{r} \leq z(r), \quad r > 0. \] Lemma~\ref{l:sol_in_Xp} shows that~$u \in \Xp$ and that the function~$(\ensuremath{\Ie^{\psif}} - \S)u$ belongs to~$\ensuremath{Y^{1,p}_{0}(\R^N)}$. Clearly,~$f \in \ensuremath{Y^{1,p}_{M}(\R^N)}$ implies that~$f \in \ensuremath{Y^{1,p}_{0}(\R^N)}$, so Theorem~\ref{t:inv} proves that \[ \ensuremath{\Ie^{\psif}} u = \ensuremath{\Ie^{\psif}} \K(u) = \ensuremath{\Ie^{\psif}} \ensuremath{\mathcal{R}} ((\ensuremath{\Ie^{\psif}} - \S)u + f) = \ensuremath{\Ie^{\psif}} u - \S u + f \text{,} \] or equivalently, that~$\S u = f$. \begin{corollary} \label{c:cont_req_sol} If\/~$\ensuremath{\Lambda_0} \rightarrow 0$, then the condition that\/~$f \in \ensuremath{Y^{1,p}_{M}(\R^N)}$ reduces to\/~$f \in \ensuremath{Y^{1,p}_{0}(\R^N)}$. \end{corollary} \noindent In other words, we recover the authors' previous result (see Theorem~\ref{t:inv}). \begin{proof} Letting~$\ensuremath{\Lambda_0} \rightarrow 0$ in~(\ref{eq:sol_in_Xp_cond}) and~(\ref{eq:sol_in_Yonez_cond}) and using~(\ref{eq:lcwz_zero}) shows that the right-hand sides of~(\ref{eq:sol_in_Xp_cond}) and~(\ref{eq:sol_in_Yonez_cond}) tend to~$C (1+s) \Qw{N}{1}(e^{-s})$ if~$s > 0$ and~$C \Qw{N}{1}(e^{-s})$ if~$s \leq 0$. The condition in~(\ref{eq:cond_u_Xp_simp}) in Lemma~\ref{l:sol_in_Xp} can now be replaced by \[ \int_0^1 (1-\log \rho)\rho^N \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} + \int_1^{\infty} \Np{\nabla f}{\rho} \, d\rho < \infty. \] \end{proof} \subsection{Uniqueness of Solutions: Proof of Theorem~\ref{t:i:uniq}} \label{s:proof_uniq} We show that~$\S u = 0$ implies that~$\K(u) = u$ and use Theorem~\ref{t:uniq_K} to deduce that~$u = 0$. Let~$\ensuremath{\mathscr{B}}$ be the Banach space introduced in the proof of Lemma~\ref{l:ext_to_B} and pick a sequence~$u_n \in \Ckc$ such that~$u_n \rightarrow u$ in~$\ensuremath{\mathscr{B}}$. It is clear that~$u_n \rightarrow u$ in~$\Lloc[p]$. The definition of~$\K$ implies that \[ \K(u_n) = \ensuremath{\mathcal{R}} (\ensuremath{\mathcal{I}} - \S)u_n \rightarrow \K(u) \] in~$\Lloc[p]$. It is straightforward to verify that~$\ensuremath{\mathcal{R}} \ensuremath{\mathcal{I}} u_n = u_n$, for example by means of the Fourier transform. As in Lemma~\ref{l:ext_to_B}, we extend the operator~$\ensuremath{\mathcal{R}} \S$ to~$\ensuremath{\mathscr{B}}$ so that \begin{equation} \label{eq:est_RS} \Np{\ensuremath{\mathcal{R}} \S u}{r} \leq C \int_0^{\infty} \int_0^{\infty} E(r,\rho,\sigma) \, \Np{u}{\sigma} \, \frac{d\sigma}{\sigma} \, \frac{d\rho}{\rho} \end{equation} for every~$u \in \ensuremath{\mathscr{B}}$. Clearly~$\ensuremath{\mathcal{R}} \S u = 0$ if~$u \in \ensuremath{\mathscr{B}}$ satisfies~$\S u = 0$. From~(\ref{eq:est_RS}), we also obtain that~$\ensuremath{\mathcal{R}} \S u_n \rightarrow \ensuremath{\mathcal{R}} \S u$ in~$\Lloc[p]$. Hence \[ \K(u_n) = \ensuremath{\mathcal{R}} (\ensuremath{\mathcal{I}} - \S) u_n = \ensuremath{\mathcal{R}} \ensuremath{\mathcal{I}} u_n - \ensuremath{\mathcal{R}} \S u_n = u_n - \ensuremath{\mathcal{R}} \S u_n \rightarrow u - \ensuremath{\mathcal{R}} \S u = u \] in~$\Lloc[p]$. By uniqueness of limits, we have~$\K(u) = u$. Now, Theorem~\ref{t:uniq_K} states that the solutions to~$\K(u) = u$ that satisfy~(\ref{eq:cond_uniq}) are unique. Since~$u = 0$ is one such solution, we must conclude that~$u = 0$ is the only possibility. \newcommand{\ensuremath{\left[ \S, \eta \right] \! u}}{\ensuremath{\left[ \S, \eta \right] \! u}} \newcommand{\ensuremath{\partial_k \left[ \S, \eta \right] \! u}}{\ensuremath{\partial_k \left[ \S, \eta \right] \! u}} \newcommand{\ensuremath{\left[ \partial_k \S, \eta \right] \! u}}{\ensuremath{\left[ \partial_k \S, \eta \right] \! u}} \newcommand{\ensuremath{\left[ \nabla \S, \eta \right] \! u}}{\ensuremath{\left[ \nabla \S, \eta \right] \! u}} \newcommand{\ensuremath{\chi}}{\ensuremath{\chi}} \newcommand{\ensuremath{\Psi}}{\ensuremath{\Psi}} \section{Local Estimates} \label{s:assymp} Let~$r_0$ be positive and let~$\eta \in C^{\infty}(\ensuremath{\mathbf{R}})$ be a cut-off function such that~$\eta(r) = 1$ if~$r < r_0$ and~$\eta(r) = 0$ if~$r > 2r_0$. Let~$u \in \Xp$,~$1 < p < \infty$, solve the equation~$\S u(x) = f(x)$ for~$|x| \leq 2r_0$, where~$f \in \ensuremath{W^{1,p}_{\mathrm{loc}}(\R^N \setminus \{ 0 \})}$. Furthermore, let~$\ensuremath{\chi} \in C_c^{\infty}(\ensuremath{\mathbf{R}})$ satisfy~$\ensuremath{\chi}(r) = 0$ if~$r < r_0$ or~$r > 2r_0$. We will require that \[ \int_{0}^{\infty} \ensuremath{\chi}(\rho) \, \frac{d\rho}{\rho} = \frac{N}{|S^{N-1}|}, \] where~$|S^{N-1}|$ is the surface measure of the unit sphere in~$\ensuremath{\mathbf{R}}^N$. We define~$\ensuremath{\Psi}(y)$ for~$y \in \ensuremath{\mathbf{R}}^N \setminus \{0\}$ by \[ \ensuremath{\Psi}(y) = - \ensuremath{\chi}(|y|) \sum_{i=1}^N \gamma_i \frac{y_i}{|y|}, \quad \text{where } \gamma_i = \int_{\ensuremath{\mathbf{R}}^N} \frac{(1-\eta(y))u(y)}{|y|^N} \frac{y_k}{|y|} \, dS(y). \] Here,~$dS(y) = \sqrt{1 + |\nabla \ensuremath{\varphi}(y)|^2} \, dy$. For~$k=1,2,\ldots,N$, the function~$\ensuremath{\Psi}$ satisfies \begin{equation} \label{eq:comcomp_kill} \begin{aligned} \int_{\ensuremath{\mathbf{R}}^N} \frac{\ensuremath{\Psi}(y) y_k \, dS(y)}{|y|^{N+1}} &= - \sum_{i=1}^N \gamma_i \int_0^{\infty} \ensuremath{\chi}(\rho)\int_{S^{N-1}} \theta_i \theta_k \, d\theta \, \frac{d\rho}{\rho} = -\gamma_k. \end{aligned} \end{equation} We multiply the equation~$\S u = f$ by~$\eta(|x|)$, and add~$\S (\eta u)$ and~$\S \ensuremath{\Psi}$ to both sides and rearrange: \begin{equation} \label{eq:Suf_co} \S (\eta u + \ensuremath{\Psi}) = \eta f + \ensuremath{\left[ \S, \eta \right] \! u} + \S \ensuremath{\Psi}, \end{equation} where~$\ensuremath{\left[ \S, \eta \right] \! u}(x) = \S (\eta u) (x) - \eta(x) \S u (x)$,~$x \in \ensuremath{\mathbf{R}}^N$. We wish to estimate the gradient of the right-hand side of~(\ref{eq:Suf_co}). \begin{lemma} \label{l:est_Np_comm} Let\/~$u \in \Xp$,\/~$1 < p < \infty$, and suppose that\/~$r_0 > 0$. Then \[ \Np{\nabla (\ensuremath{\left[ \S, \eta \right] \! u} + \S \ensuremath{\Psi}) }{r} \leq \Biggl\{ \begin{aligned} &C(r + \ensuremath{\Lambda}(r)) \|u\|_{\Xp} , & r \leq r_0,\\ &C r^{-N} \, \|u \|_{\Xp}, & r > r_0, \end{aligned} \] where\/~$C$ depends on\/~$N$,\/~$p$, and\/~$r_0$. \end{lemma} \begin{proof} Let~$0 < r < r_0/2$ and~$r \leq |x| < 2r$. Then~$\ensuremath{\partial_k \left[ \S, \eta \right] \! u} = \partial_k \S ((\eta - 1)u)$, and from the representation in~(\ref{eq:diff_Su}) it follows that \[ \begin{aligned} \frac{\partial_k \bigl( \left[ \S, \eta \right] \! u + \S \ensuremath{\Psi} \bigr)(x)}{1-N} = {} & T_k\bigl((1 - \eta)u + \ensuremath{\Psi} \bigr)(x) + \partial_k \ensuremath{\varphi} (x) T_{N+1}\bigl( (1-\eta)u + \ensuremath{\Psi} \bigr)(x)\\ = {} & \int_{\ensuremath{\mathbf{R}}^N} \frac{(x_k - y_k)\bigl( (1 - \eta(y))u(y) + \ensuremath{\Psi}(y) \bigr) \, dy}{|\ld[x] - \ld[y]|^{N+1}}\\ & + \partial_k \ensuremath{\varphi} (x) \int_{\ensuremath{\mathbf{R}}^N} \frac{(\ensuremath{\varphi}(x) - \ensuremath{\varphi}(y))\bigl( (1 - \eta(y))u(y) + \ensuremath{\Psi}(y) \bigr) \, dy}{|\ld[x] - \ld[y]|^{N+1}}. \end{aligned} \] The second term in the right-hand side can be estimated by \[ C \ensuremath{\Lambda}(r) \int_{|y| > r_0} \frac{ |u(y)| + |\ensuremath{\Psi}(y)|}{|y|^N} \, dy \leq C \ensuremath{\Lambda}(r) \int_{|y| > r_0} \frac{|u(y)|}{|y|^N} \, dy \] since \[ |\ensuremath{\Psi}(y)| \leq |\ensuremath{\chi}(|y|)| \sum_{i=1}^N |\gamma_i| \leq C|\ensuremath{\chi}(|y|)| \int_{|z| > r_0} \frac{|u(z)| \, dz}{|z|^N}, \quad y \in \ensuremath{\mathbf{R}}^N. \] Let~$\tau = |x|/|y|$. We assume that~$\tau \leq 1/2$. The kernel in~$T_k$ can be expressed by \[ \frac{x_k - y_k}{|\ld[x] - \ld[y]|^{N+1}} = \frac{1}{|y|^N} g(\tau), \quad \text{where } g(\tau) = \frac{\frac{x_k}{|x|} \tau - \frac{y_k}{|y|}}{\left|\frac{\ld[x]}{|x|}\tau - \frac{\ld[y]}{|y|} \right|^{N+1}}. \] We see that~$g(0) = -(y_k/|y|)(1+\ensuremath{L_x}^2)^{-(N+1)/2}$, and also that~$g'(\tau)$ is uniformly bounded,~$|g'(\tau)| \leq C$, for~$\tau \leq 1/2$. The constant depends only on~$N$. Thus,~$|g(\tau) - g(0)| \leq C \tau$ for~$\tau \leq 1/2$. Hence, \[ \begin{aligned} |T_k((1-\eta)u + \ensuremath{\Psi})(x)| \leq {} & \frac{1}{(1+\ensuremath{L_x}^2)^{\frac{N+1}{2}}} \biggl| \int_{\ensuremath{\mathbf{R}}^N} \frac{y_k}{|y|} \cdot \frac{(1-\eta(y))u(y) + \ensuremath{\Psi}(y)}{|y|^N} \, dS(y) \biggr| \\ & + C |x| \int_{|y| > r_0} \frac{|u(y)| + |\ensuremath{\Psi}(y)|}{|y|^{N+1}} \, dy. \end{aligned} \] It is clear from~(\ref{eq:comcomp_kill}) that \[ \int_{\ensuremath{\mathbf{R}}^N} \frac{y_k}{|y|} \cdot \frac{(1-\eta(y))u(y) + \ensuremath{\Psi}(y)}{|y|^N} \, dS(y) = 0, \quad k=1,2,\ldots,N. \] Moreover, \[ |x|\int_{|y| > r_0} \frac{|u(y)| + |\ensuremath{\Psi}(y)|}{|y|^{N+1}} \, dy \leq C r \int_{|y| > r_0} \frac{|u(y)|}{|y|^{N}} dy, \] so \[ \Np{T_k((1-\eta)u + \ensuremath{\Psi})}{r} \leq C \bigl( r + \ensuremath{\Lambda}(r) \bigr) \int_{|y| > r_0} \frac{|u(y)| }{|y|^N} dy. \] The right-hand side is finite since~$u \in \Xp$. Thus, we obtain that \begin{equation} \label{eq:est_Np_comm_m} \Np{\nabla \bigl( \left[ \S, \eta \right] \! u + \S \ensuremath{\Psi} \bigr)}{r} \leq C \bigl( r + \ensuremath{\Lambda}(r) \bigr) \|u\|_{\Xp}, \quad 0 < r < \frac{r_0}{2}. \end{equation} Let~$r > 4r_0$. Then~$\ensuremath{\partial_k \left[ \S, \eta \right] \! u} = \partial_k \S(\eta u)$, and \[ |\ensuremath{\partial_k \left[ \S, \eta \right] \! u}(x) | \leq \frac{C}{|x|^N} \int_{|y| < 2r_0} |u(y)| \, dy \leq \frac{C}{|x|^{N}} \|u\|_{\Xp}. \] Furthermore, \[ \begin{aligned} |\partial_k \S \ensuremath{\Psi} (x)| &\leq C |T_k \ensuremath{\Psi} (x)| + C |T_{N+1} \ensuremath{\Psi}(x)| \\ & \leq \frac{C}{|x|^N} \int_{r_0 \leq |y| \leq 2r_0} |\ensuremath{\Psi}(y)| \, dy \leq \frac{C}{|x|^N} \int_{|y| > r_0} \frac{|u(y)|}{|y|^N} dy. \end{aligned} \] This implies that \begin{equation} \label{eq:est_Np_comm_p} \Np{\nabla \bigl( \left[ \S, \eta \right] \! u + \S \ensuremath{\Psi} \bigr)}{r} \leq C r^{-N} \|u\|_{\Xp}, \quad r > 4r_0. \end{equation} If~$r_0/2 \leq r \leq 4r_0$, then \[ \begin{aligned} \ensuremath{\partial_k \left[ \S, \eta \right] \! u}(x) &= \partial_k \S \eta u(x) - \eta'(|x|) \frac{x_k}{|x|} \S u (x) - \eta(|x|) \partial_k \S u(x)\\ &= \ensuremath{\left[ \partial_k \S, \eta \right] \! u}(x) - \eta'(|x|)\frac{x_k}{|x|} \S u(x). \end{aligned} \] From the representation in~(\ref{eq:diff_Su}), we obtain that \[ \begin{aligned} |\ensuremath{\left[ \partial_k \S, \eta \right] \! u}(x)| &\leq C \int_{\ensuremath{\mathbf{R}}^N} \frac{|\eta(x) - \eta(y)|}{|\ld[x] - \ld[y]|^N} \, |u(y)| \, dy \leq C \int_{\ensuremath{\mathbf{R}}^N} \frac{|u(y)|}{|x-y|^{N-1}} \, dy.\\ \end{aligned} \] Thus,~$|\ensuremath{\left[ \partial_k \S, \eta \right] \! u}(x)| \leq C \ensuremath{\mathcal{I}}|u|(x)$, so \[ \begin{aligned} \Np{\ensuremath{\left[ \nabla \S, \eta \right] \! u}}{r} & \leq C r \, \int_0^{\infty} \Qw{N}{1} \left( \frac{\rho}{r} \right) \, \Np{u}{\rho} \, \frac{d\rho}{\rho} \leq C \| u \|_{\Xp}. \end{aligned} \] Similarly, since~$|\eta' \S u| \leq |\eta'| \ensuremath{\mathcal{I}}|u|$, we obtain that $ \Np{\eta' \S u}{r} \leq C \, \|u\|_{\Xp} $ for~$r_0/2 \leq r \leq 2r_0$, and that~$\eta' \S u = 0$ otherwise. Since also \[ \begin{aligned} \Np{\nabla \S \ensuremath{\Psi}}{r} &\leq C \int_{0}^{\infty} \Qw{N}{0} \left( \frac{\rho}{r} \right) \, \Np{\ensuremath{\Psi}}{\rho} \, \frac{d\rho}{\rho} \leq C \int_{|y|>r_0} \frac{|u(y)|}{|y|^N} \, dy , \end{aligned} \] it is clear that \begin{equation} \label{eq:est_Np_comm_c} \Np{\nabla \bigl( \left[ \S, \eta \right] \! u + \S \ensuremath{\Psi} \bigr)}{r} \leq C \|u\|_{\Xp}, \quad \frac{r_0}{2} \leq r \leq 4r_0. \end{equation} The desired result now follows from~(\ref{eq:est_Np_comm_m}),~(\ref{eq:est_Np_comm_p}), and~(\ref{eq:est_Np_comm_c}). \end{proof} \subsection{Proof of Theorem~\ref{t:i:local}} Lemma~\ref{l:est_Np_comm} implies, for~$u$ in~$\Xp$, that~$\ensuremath{\left[ \S, \eta \right] \! u}$ belongs to~$\ensuremath{Y^{1,p}_{M}(\R^N)}$ and that the radial part~$(\ensuremath{\left[ \S, \eta \right] \! u})_R$ tends to zero as~$r \rightarrow \infty$. Similarly, the same is true for~$\S \ensuremath{\Psi}$. Moreover, since~$f$ satisfies~(\ref{eq:req_local}),~$\eta f \in \ensuremath{Y^{1,p}_{M}(\R^N)}$ and~$(\eta f)_R \rightarrow 0$. Thus, by Theorem~\ref{t:i:exist}, \begin{equation} \label{eq:Sw} \S w = \eta f + \left[ \S, \eta \right] \! u + \S \ensuremath{\Psi} \end{equation} has a solution~$w \in \Xp$ which satisfies \begin{equation} \label{eq:est_w} \begin{aligned} \Np{w}{r} \leq {} & \ensuremath{c_3} \int_0^{r} \left( \frac{\rho}{r} \right)^{\ensuremath{M}} \Np{\nabla ( \eta f + \ensuremath{\left[ \S, \eta \right] \! u} + \S \ensuremath{\Psi})}{\rho} \, \frac{d\rho}{\rho}\\ & + \ensuremath{c_3} \int_r^{\infty} \exp \biggl( \ensuremath{c_1} \int_r^{\rho} \ensuremath{\Lambda}(\nu) \, \frac{d\nu}{\nu} \biggr) \, \Np{\nabla (\eta f + \ensuremath{\left[ \S, \eta \right] \! u} + \ensuremath{\Psi})}{\rho} \, \frac{d\rho}{\rho}. \end{aligned} \end{equation} Since also~$\S u = f$ for~$|x| \leq 2r_0$, equation~(\ref{eq:Sw}) can be rewritten as \[ \S w = \eta f + \S \eta u - \eta \S u + \S \ensuremath{\Psi} = \eta f + \S \eta u - \eta f + \S \ensuremath{\Psi} = \S ( \eta u + \ensuremath{\Psi}) . \] In other words,~$\S (w - \eta u - \ensuremath{\Psi}) = 0$, and since~$w - \eta u - \ensuremath{\Psi}$ satisfies the conditions in Theorem~\ref{t:i:uniq}, it follows that~$w = \eta u + \ensuremath{\Psi}$. Thus, \[ \Np{u}{r} = \Np{\eta u + \ensuremath{\Psi}}{r} = \Np{w}{r}, \quad 0 < r < r_0. \] We now turn to prove~(\ref{eq:i:assymp}) in Theorem~\ref{t:i:local}. We split the right-hand side of~(\ref{eq:est_w}) in two parts: one which deals with~$\ensuremath{\left[ \S, \eta \right] \! u} + \S \ensuremath{\Psi}$, and one for~$f$. From Lemma~\ref{l:est_Np_comm}, we obtain that \[ \begin{aligned} \int_0^r \left( \frac{\rho}{r} \right)^{\ensuremath{M}} \Np{\nabla (\ensuremath{\left[ \S, \eta \right] \! u} + \S \ensuremath{\Psi})}{\rho} \, \frac{d\rho}{\rho} &\leq C \| u \|_{\Xp} r^{-\ensuremath{M}} \int_0^{r} \rho^{\ensuremath{M}}(\rho + \ensuremath{\Lambda}(\rho)) \, \frac{d\rho}{\rho}\\ & \leq C(r + \ensuremath{\Lambda}(r)) \| u \|_{\Xp}. \end{aligned} \] Let~$D(a,b)$ be defined by \[ \ensuremath{D}(a,b) = \exp \biggl( \ensuremath{c_1} \int_a^{b} \ensuremath{\Lambda}(\nu) \, \frac{d\nu}{\nu} \biggr), \quad a,b \geq 0. \] It is true that \[ \int_r^{r_0} \ensuremath{D}(r,\rho) \, \Np{\nabla (\ensuremath{\left[ \S, \eta \right] \! u} + \S \ensuremath{\Psi})}{\rho} \, \frac{d\rho}{\rho} \leq C \| u \|_{\Xp} \int_{r}^{r_0} (\rho + \ensuremath{\Lambda}(\rho)) \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho}. \] Since \[ \int_r^{r_0} \rho \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho} = \ensuremath{D}(r,r_0) \int_{r}^{r_0} \ensuremath{D}(r_0,\rho) \, d\rho \leq C \ensuremath{D}(r,r_0) \] and \[ \int_r^{r_0} \ensuremath{\Lambda}(\rho) \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho} \leq \frac{1}{\ensuremath{c_1}} \int_r^{r_0} \frac{\partial}{\partial \rho} \ensuremath{D}(r,\rho) \, d\rho \leq \frac{\ensuremath{D}(r,r_0)}{\ensuremath{c_1}}, \] we obtain that \[ \int_r^{r_0} \ensuremath{D}(r,\rho) \, \Np{\nabla (\ensuremath{\left[ \S, \eta \right] \! u} + \S \ensuremath{\Psi})}{\rho} \, \frac{d\rho}{\rho} \leq C \| u \|_{\Xp} \ensuremath{D}(r,r_0). \] It is clear that~$r + \ensuremath{\Lambda}(r)$ is small whenever~$r$ is small, and also that~$\ensuremath{D}(r,r_0) \geq 1$ for~$r < r_0$. Hence,~$r + \ensuremath{\Lambda}(r) \leq C\ensuremath{D}(r,r_0)$ if~$r < r_0$. Finally, \[ \begin{aligned} \int_{r_0}^{\infty} \ensuremath{D}(r,{\rho}) \, \Np{\nabla (\ensuremath{\left[ \S, \eta \right] \! u} + \S \ensuremath{\Psi})}{\rho} \, \frac{d\rho}{\rho} & \leq C \| u \|_{\Xp} \int_{r_0}^{\infty} e^{-N\log \rho} \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho} \\ & \leq \frac{C}{r_0^N(N - \ensuremath{c_1} \ensuremath{\Lambda_0})} \| u \|_{\Xp} \ensuremath{D}(r,r_0), \end{aligned} \] where we used the argument from~(\ref{eq:aa2}) in Lemma~\ref{l:est_kern_Ku} to estimate the last integral. Now, since~$\nabla(\eta f) = \nabla \eta f + \eta \nabla f$, and~$\nabla \eta$ only lives for~$r_0 < |x| < 2r_0$, it follows that the contribution from~$f$ is bounded by \[ \begin{aligned} & Cr^{-\ensuremath{M}} \int_0^r {\rho}^{\ensuremath{M}} \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} + C \ensuremath{D}(r,r_0) \int_r^{2r_0} \ensuremath{D}(r_0,\rho) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho}\\ & \qquad \qquad + C \ensuremath{D}(r,r_0) \int_{r_0/2}^{2r_0} \Np{f}{\rho} \, \frac{d\rho}{\rho}.\\ \end{aligned} \] This completes the proof. \subsection{The Case of Summable~\boldmath{$\ensuremath{\Lambda}$}} Let the conditions of Theorem~\ref{t:i:local} be satisfied and suppose that~$(\nu \mapsto \ensuremath{\Lambda}(\nu)/\nu)$ belongs to~$L^1(0,2r_0)$. Then~$\ensuremath{D}(r,\rho) \leq C$ for~$0 \leq r \leq \rho \leq 1$, and from Theorem~\ref{t:i:local} it follows that \[ \begin{aligned} \Np{u}{r} &\leq C \int_0^r \left( \frac{\rho}{r} \right)^{\ensuremath{M}} \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} + C \int_r^{2r_0} \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} + C(u,f)\\ &= C \int_0^{2r_0} \Qw{\ensuremath{M}}{0} \left( \frac{\rho}{r} \right) \, \Np{\nabla f}{\rho} \, \frac{d\rho}{\rho} + C(u,f).\\ \end{aligned} \] If also~$(\rho \mapsto \Np{\nabla f}{\rho}/\rho) \in L^1(0,2r_0)$, then~$\Np{u}{r} \leq C$ for~$r < r_0$, where~$C$ depends on~$f$. \subsection{Example:~\boldmath{$\Np{\nabla f}{r} \leq C r^{-\alpha}$}} \label{s:est_Np_ralpha} Let~$0 < \alpha < \ensuremath{M}$. Suppose that~$\ensuremath{\Lambda}(r) \rightarrow 0$ as~$r \rightarrow 0$. Then there exists~$r_1 > 0$ such that~$\ensuremath{\Lambda}(r) \leq \alpha/(2\ensuremath{c_1})$ for~$0 < r < r_1$. Now, suppose that~$\Np{\nabla f}{r} \leq C r^{-\alpha}$ for~$r < 2r_0$. Let~$r < r_0$. Then Theorem~\ref{t:i:local} implies that \begin{equation*} \label{eq:assymp_ex1} \begin{aligned} \Np{u}{r} \leq {} & \frac{C}{\ensuremath{M} - \alpha} r^{-\alpha} + C \int_r^{2r_0} \rho^{-\alpha} \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho} + C(u,f) \ensuremath{D}(r,r_0) , \end{aligned} \end{equation*} where \[ C(u,f) = C \biggl( \|u \|_{\Xp} + \int_{r_0/2}^{2r_0} \Np{f}{\rho} \, \frac{d\rho}{\rho} \biggr) \] and~$C$ only depends on~$N$ and~$p$. Now, \[ \int_r^{2r_0} \rho^{-\alpha} \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho} = r^{-\alpha} \int_r^{2r_0} \left( \frac{r}{\rho} \right)^{\alpha} \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho}. \] Suppose that~$r < r_1$. We split the domain of integration in two parts. For the first part, we obtain that \[ \begin{aligned} \int_r^{r_1} \left( \frac{r}{\rho} \right)^{\alpha} \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho} &= \int_r^{r_1} \exp\left( \ensuremath{c_1} \int_r^{\rho} \left( \ensuremath{\Lambda}(\nu) - \frac{\alpha}{\ensuremath{c_1}} \right) \frac{d\nu}{\nu} \right) \, \frac{d\rho}{\rho}\\ &\leq \int_{r}^{r_1} \left( \frac{r}{\rho} \right)^{\alpha/2} \frac{d\rho}{\rho} \leq \frac{2}{\alpha} \end{aligned} \] since~$\ensuremath{\Lambda}(\nu) - \alpha/\ensuremath{c_1} \leq - \alpha/(2\ensuremath{c_1})$. For the second part, \[ \begin{aligned} \int_{r_1}^{2r_0} \left( \frac{r}{\rho} \right)^{\alpha} \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho} &= \int_{r_1}^{2r_0} \left( \frac{r}{r_1} \right)^{\alpha} \ensuremath{D}(r,r_1) \left( \frac{r_1}{\rho} \right)^{\alpha} \ensuremath{D}(r_1,\rho) \, \frac{d\rho}{\rho}\\ &\leq \int_{r_1}^{2r_0} \left( \frac{r}{r_1} \right)^{\alpha/2} \left( \frac{r_1}{\rho} \right)^{\alpha} \ensuremath{D}(r_1,\rho) \, \frac{d\rho}{\rho}\\ &\leq \frac{r_1^{-\alpha/2} \ensuremath{D}(r_1,2r_0)}{\alpha} r^{\alpha/2}, \end{aligned} \] similarly with the estimate for the first part. Hence, \[ \int_r^{2r_0} \rho^{-\alpha} \ensuremath{D}(r,\rho) \, \frac{d\rho}{\rho} \leq \frac{2}{\alpha} r^{-\alpha} + \frac{r_1^{-\alpha/2} \ensuremath{D}(r_1,2r_0)}{\alpha} r^{-\alpha/2} \leq C_{\alpha} r^{-\alpha}. \] Thus, \[ \begin{aligned} \Np{u}{r} \leq {} & C({\alpha}) r^{-\alpha} \leq C(u,f,{\alpha}) r^{-\alpha} + C(u,f) \ensuremath{D}(r,r_0) , \end{aligned} \] where~$C({\alpha})$ is a constant that depends on~$\alpha$. The last inequality follows from the fact that it is possible to estimate~$\ensuremath{D}(r,r_0) \leq C({\epsilon})r^{-\epsilon}$ for every~$\epsilon > 0$, which follows analogously with the argument used above. \defReferences{References}

https://arxiv.org/abs/1308.2046

A remark on the space of 7-gons with a fixed total length in $\R^3$

Based on the model of the space $Pol_3(n)$ of polygons in $R^3$ with limited number of vertex, which was proposed by Jean-Claude Hausmann and Allen Knutson, and developed by several authors: Jason Cantarella, Alexander Y. Grosberg, Robert Kusner, and Clayton Shonkwiler, we prove that there exists an isometric isotopy of $Pol_3(n)$, $n=7$, into itself, which transforms an arbitrary polygon to its mirror copy, and, additionally, preserves lengths of projections of polygons into the two coordinate planes, and keeps projection of polygons onto the line. The proof is based on elementary arguments with Cayley numbers. A possible generalization of the statement for greater $n$ is related with a theorem by I.James on strong Kervaire invariants in stable homotopy of spheres.

\section*{Introduction} In the paper \cite{C-G-K-Sh} the symmetric measure on closed polygons of fixed total length is investigated. The space $Pol_3(n)$ consists of vectors $(\vec{e}_1, \dots, \vec{e}_n) \in (\R^3)^n$, which satisfy the conditions: --1. $\sum_{i=1}^n \vec{e}_i = 0$. --2. $\sum_{i=1}^n \vert \vec{e}_i \vert = 2$. Such a vector $(\vec{e}_1, \dots, \vec{e}_n)$ represents a closed $n$-gone in $\R^3$ with a fixed (twice unite) total length. The authors apply an approach by Jean-Claude Hausmann and Allen Knutson to describe a parametrization of the space $Pol_3(n)$ by the homogeneous Stiefel manifold $V_2(\C,n)$. This allow to introduce the symmetric measure on $Pol_3(n)$ using the quaternions. In the present paper we use elementary properties of the Cayley numbers to prove the following theorem. \begin{theorem}\label{main} There exists a one-parameter isometric isotopy $\Theta$ of $Pol_3(n)$, $n=7$, to itself, which joins each closed polygon of the length $2$ with its mirror image. Additionally: --I this isotopy preserves lengths of the projections of polygons into the two coordinate planes $(x,y,0) \subset \R^3$, $(x,0,z) \subset \R^3$, --II this isotopy keeps the projection of polygons onto the line $(x,0,0) \subset \R^3$. \end{theorem} A generalization of Theorem \ref{main} for spaces $Pol_3(n)$ with larger $n \ge 8$ could be interesting. I conjecture, such a generalization is possible for $n=15$, using an application by I.James of a Toda's result to the strong Kervaire Invariant One Conjecture, see \cite{J}. The theorem by James is proved using homotopy theory and explicit formulas are unknown. I am not able to formulate a conjecture precisely, because I am not able to formulate analogs of properties I, II in the case $n=15$. A partial solution of the Kervaire Invariant One Problem in stable homotopy of spheres by M.A. Hill, M.J. Hopkins, and D.C. Ravenel (2010), and a particular solution of a weaker result by the author of the Kervaire Invariant One Problem (my approach in a final form is not published), which is based on alternative arguments, allows to formulate the following conjecture. \begin{conjecture}\label{arf} For an arbitrary $n$, $n \ge n_0$, where $n_0$ is sufficiently large, Theorem $\ref{main}$ (even without additional properties --I, --II) is not valid. \end{conjecture} The paper is organized as following. All required constructions from the paper \cite{C-G-K-Sh} are collected if Section 1. In Section 2 new elementary computations are presented and Theorem is proved. This results was initiated by several discussions with the first and the forth authors of \cite{C-G-K-Sh}. The results was prepared at Prof. D.D.Sokoloff's seminaire in Moscow, May (2013). \section{The model spaces} \subsection{The model space of $Pol_2(n)$} Denote by $Pol_2(n;\lambda)$ the space of $n$-gons (possibly, with self-intersections) of the total length $2\lambda$ on the plane. In the case $\lambda=1$ we will write $Pol_2(n)$ for short. This space consists of collections of $n$ vectors $(\vec{e}_1, \dots, \vec{e}_n) \in (\R^2)^n$, which satisfy Conditions --1 and --2 (see Introduction). Denote by $V_2(\R^n)$ the Stiefel manifold of all pairs of orthonormal vectors in $\R^n$. A point $(Z_1,Z_2) \in V_2(\R^n)$ admits coordinates $Z_j=(z_{1,j}, \dots, z_{n,j})$, $z_{i,j} \in \R$, $j=1,2$, which satisfy the properties: --a. $\sum_{i=1}^n (z_{i,1},z_{i,2}) = 0$, $1 \le i \le n$. --b. $\sum_{i=1}^n z_{i,j}^2 = 1$, $j=1,2$. Define the quadratic mapping $$H: V_2(\R^n) \to Pol_2(n) $$ \begin{eqnarray}\label{H} H(Z_1,Z_2) = ((z_{1,1}^2-z_{1,2}^2,2z_{1,1}z_{1,2}), \dots ,(z_{n,1}^2-z_{n,2}^2,2z_{n,1}z_{n,2})). \end{eqnarray} One may extend this mapping to the space $Pol_2(n;\lambda)$, obviously. It is convenient to write the coordinates $(z_{i,1},z_{i,2})$ as a point at the imaginary complex plane in the quaternion space as follows: $\z_i = z_{i,1} + \j z_{i,2} \in \C \subset \H = \C \oplus \j \C$, $\j^2=-1$. Then for the mapping $H$ we get: $$H(\Z)=(\z_1^2,\dots,\z_n^2),$$ $\Z = (\z_1, \dots, \z_n)$. Let us check that the image of $H$ belongs to the subspace $Pol_2(n) \subset (\R^2)^n$. The condition --1 (see Introduction) is deduced from the conditions --a and --b: $$ \sum_{i=1}^n \vec{e}_i = \sum_{i=1}^n \z_i^2 = (\sum_{i=1}^n (z_{i,1}^2-z_{i,2}^2), 2\sum_{i=1}^n z_{i,1}z_{i,2})=(0,0).$$ The condition --2 is proved as following: $$ \sum_{i=1}^n \vert \vec{e}_i \vert = \sum_{i=1}^n (\z_i \bar{\z}_i)^2 = \sqrt{\sum_{i=1}^n (z^2_{i,1} - z^2_{i,2})^2 + 4z_{i,1}z_{i,2}}.$$ using --a and --b, we get: $$\sum_{i=1}^n \vert \vec{e}_i \vert = \sqrt{\sum_{i=1}^n (z_{i,1}^2 + z_{i,2}^2)^2} = 2.$$ It is easy to see that $H$ is not a homeomorphism, this is a ramified covering of the multiplicity $2^n$. The space $V_2(\R^n)$ admits the standard fibration: $$ S^{n-2} \subset V_2(\R^n) \mapsto S^{n-1}, $$ the standard metric in the image and in the fiber of this fibration induces a metric on $V_2(\R^n)$. This metric determines the probability measure on the space $Pol_2(n)$: the measure of a subset $V \subset Pol_2(n)$ is well-defined as the volume of its preimage $H^{-1}(V) \subset V_2(\R^n)$. \subsection{The model space of $Pol_3(n)$} The model space of $Pol_3(n)$ is defined by a "complexification"' of the model space $Pol_2(n)$. Let us consider the following mapping, which is called the Hopf mapping: \begin{eqnarray}\label{Hopf} Hopf: \H \mapsto \R^3, \quad Hopf(\q) = \bar{\q}\i \q, \end{eqnarray} where $\q = q_0 + q_1\i +q_2\j + q_3 \k \in \H$ is a quaternion, $\bar{\q} = q_0 - q_1\i - q_2\j - q_3 \k$, where $\i,\j,k, \in \H$ are the standard generators. The restriction of the mapping $Hopf$ to the plane $(q_0 + q_2 \j)$ is given by the formula: $$Hopf(q_0 + q_2 \j) = (q_0^2 - q_2^2, 0, 2q_0q_2).$$ This mapping coincides with the mapping $$q_0 + q_2 \j \mapsto (q_0 + q_2 \j)^2,$$ where the image is identified with the plane $(z_1,0,z_2 \j) \subset \R^3$. The restriction of the mapping $Hopf$ to the plane $(q_0 + q_3 \k)$ is given by the formula: $$Hopf(q_0 + q_3 \k) = (q_0^2 - q_3^2, -2q_0q_3, 0).$$ This mapping coincides with the mapping $$q_0 + q_3 \k \mapsto (q_0 + q_2 \k)^2,$$ where the image is identified with the plane $(z_1,-z_2 \k,0) \subset \R^3$. Denote by $V_2(\C^n)$ the complex Stiefel manifold of all pairs of Hermitian orthonormal vectors in $\C^n$. A point $(Q_{0,1},Q_{2,3}) \in V_2(\C^n)$ admits coordinates $Q_{0,1}=(q_{1,0} + \i q_{1,1}, \dots, q_{n,0}+ \i q_{n,1})$, $Q_{2,3}=(q_{1,2} + \i q_{1,3}, \dots, q_{n,2}+ \i q_{n,3})$, where $q_{i,1}=q_{i,0}+\i q_{i,1} \in \C$, which satisfy the properties: --a'. $\sum_{i=1}^n (q_{i,0}+\i q_{i,1})(q_{i,2} - \i q_{i,3}) = 0$, $1 \le i \le n$. --b'. $\sum_{i=1}^n q^2_{i,0} + q^2_{i,1} = 1$, $\sum_{i=1}^n q_{i,2}^2 + q_{i,3}^2 = 1$. In \cite{C-G-K-Sh} by a straightforward calculation is proved that the coordinatewise extension of the mapping $(\ref{Hopf})$ defines the required mapping: \begin{eqnarray}\label{Hopff} Hopf: V_2(\C^n) \to Pol_3(n) \end{eqnarray} $$H(Q_0,Q_1,Q_2,Q_3) = (q_{1,0}^2+q_{1,1}^2-q_{1,2}^2-q_{1,3}^2, \quad 2q_{1,1}q_{1,2}- 2q_{1,0}q_{1,3},\quad 2q_{1,0}q_{1,2} + 2q_{1,1}q_{1,3}, \dots ,$$ $$q_{n,0}^2+q_{n,1}^2-q_{n,2}^2-q_{n,3}^2,\quad 2q_{n,1}q_{n,2}- 2q_{n,0}q_{n,3},\quad 2q_{n,0}q_{n,2} + 2q_{n,1}q_{n,3}).$$ \section{Proof of the main result} \subsection{The normalization of preimages of an $n$-gone in the model space} The restriction of the mapping $(\ref{Hopf})$ to the sphere $S^3 \subset \H$ of the radius $\sqrt{2}$ (as well as to a sphere of an arbitrary radius) is invariant with respect to the (-right, or -left) multiplication on a complex number $\cos(\theta) + \i \sin(\theta)$. For the left multiplication we get: $$\overline{(\cos(\theta) + \i \sin(\theta) \q)}\i (\cos(\theta) + \i \sin(\theta))\q = $$ $$\bar{\q} (\cos(\theta) - \i \sin(\theta)) \i (\cos(\theta) + \i \sin(\theta)) \q = \bar {q} \i \q.$$ For the right multiplication the proof is analogous. We shall call a pair $(Q_{0,1},Q_{2,3})$ is normalized, if $q_{i,1}=0$ for an arbitrary $i$, $1 \le i \le n$. For a normalized pair one could write $Q_{0,1}=(Q_0,0)$. Let us prove that an arbitrary pair $(Q_{0,1},Q_{2,3})$, which satisfy properties --a', --b', admits the normalization. Define a complex number $\alpha_i$, $\vert \alpha_i\vert=1$, by the following formula: $$\alpha_i = \frac{\sqrt{q_{i,0}^2 + q_{i,1}^2}}{q_{i,0} + \i q_{i,1}},$$ if $q_{i,1} \ne 0$, and $\alpha_i = \sign(q_{i,2})$, if $q_{i,1}=0$. Define the normalization $(Q_{0,1},Q_{2,3}) \mapsto (Q'_{0,1}=(Q'_0,0),Q'_{2,3})$ by the following formula: $$Q'_{0,1}=(q_{1,0} + \i q_{1,1})\alpha_1, \dots (q_{n,0} +\i q_{n,1})\alpha_n,$$ $$Q'_{2,3}=(q_{1,2} + \i q_{1,3})\alpha_1, \dots (q_{n,2} +\i q_{n,3})\alpha_n.$$ The images by the mapping $(\ref{Hopff})$ of the pair of vectors $(Q_{0,1},Q_{2,3})$ and of its normalization $(Q'_0,Q'_{2,3})$ coincide. For the normalized pair $(Q'_0,Q'_{2,3})$ the formula of the mapping $(\ref{Hopff})$ is much simple. We get: \begin{eqnarray}\label{Hopfnorm} \begin{array}c H(Q'_0,Q'_{2,3}) = (q_{1,0}^2-q_{1,2}^2-q_{1,3}^2, \quad - 2q_{1,0}q_{1,3},\quad 2q_{1,0}q_{1,2}, \dots , \\ q_{n,0}^2-q_{n,2}^2-q_{n,3}^2,\quad - 2q_{n,0}q_{n,3}, \quad 2q_{n,0}q_{n,2}). \end{array} \end{eqnarray} Let us prove that the formula $(\ref{Hopfnorm})$ is defined as a superposition of two formulas $(\ref{H})$. More precisely, a point in $Pol_3(n)$, which parametrizes an $n$-gon $L \subset \R^3$ of the length $2$, is completely reconstructed from the two projections $L_{x,y} \subset \R^2(x,y,0) \subset \R^3$, $L_{x,z} \subset \R^2(x,0,z) \subset \R^3$ onto the standard planes as follows. Consider the projection $L_{x,z}$ of $L$ and denote by $l_{x,z}$ the total length of this projection, $l_{x,y} \le 2$. The projection $L_{x,z}$ is a collection of vectors $\vec{e}_i$, which satisfy Condition --a, and a modification of Condition --b, where in the right side of the first and the second formulas we get the constant $l_{x,y}$ instead of $1$. Take a preimage $(Q'_0,Q'_2)$ of $L_{x,z}$ with respect to the mapping $H$, $H(Q'_0,Q'_2)=L_{x,z}$, $Q'_0=(q'_{1,0}, \dots, q'_{n,0})$, $Q'_2=(q'_{1,2}, \dots q'_{n,2})$. Let us fix a pair $(Q'_0,Q'_2)$ by the condition $Q'_{i,0} \ge 0$, $1 \le i \le n$ (in the case $Q'_{i,0} =0$ let us agree that $Q'_{i,2} \ge 0$ to have a well-defined lift of $H$). Each vector of the pair $(Q'_1,Q'_2)$ belongs to the sphere of the radius $\sqrt{l_{x,y}}$ in $\R^n$. Apply the same construction to the projection $L_{x,z}$ of $L$. Denote by $l_{x,z}$ the total length of this projection, $l_{x,z} \le 2$. The following equation, obviously, is satisfied: $$ l_{x,y} + l_{x,z} - \lambda_0 = 2, $$ where $$\lambda_0 = \sum_{i=1}^n (q')^2_{i,0}. $$ Define the vector $Q'_3$, $Q'_3=(q'_{1,3}, \dots q'_{n,3})$, such that for the pair of vector $(Q'_0,Q'_3)$ the equation $H^{conj}(Q'_0,Q'_3)=L_{x,y}$, where $H^{conj}$ is the composition of $H$ with the reflexion $(x,y) \mapsto (x,-y)$, is satisfied. The triple $(Q'_0,0,Q'_2,Q'_3)$ of vectors determines the normalized preimage of $L$. For an arbitrary preimage $(Q_0,Q_1,Q_2,Q_4)=Hopf^{-1}(L)$ the normalized preimage $(Q'_0,0,Q'_2,Q'_3)$ is well-defined. However, the correspondence $L \mapsto (Q'_0,0,Q'_2,Q'_3)$ is not continuous. \subsection{A construction with Cayley numbers related with $Pol_2(7)$} In this subsection we shall prove the following result, which is, probably, interesting by itself. \begin{lemma}\label{lemma} There exists a one-parameter isometric isotopy of the space $Pol_2(7;\lambda)$ to itself, which joins each closed polygon of the length $\lambda$ in the standard plane with its mirror image with the axis $\{x=0\}$. Moreover, this isotopy keeps the projection of each polygon onto the axis $\{x=0\}$. \end{lemma} \subsubsection*{Proof of Lemma $\ref{lemma}$} Assume, for simplicity, that $\lambda =1$, the general case is analogous. Let us assume that the each vector in the pair $(Z_1,Z_2)$, which determines a point of $V_2(\R^7)$ is represented as an imaginary Cally unit, $\vert\vert Z_1 \vert \vert = \vert\vert Z_2 \vert \vert = 1$: $$Z_j = z_{j,1} \i + z_{j,2} \j + z_{j,3} \k + z_{j,4} \e + z_{j,5} \f + z_{j,6} \g + z_{j,7} \h, \quad j=1,2. $$ Then the vector $Z_1$ determines a one-parameter subgroups in the Cally units $S^7 \subset \Ca$, which contains the elements $\{1,Z_1\}$. Let $Z(\theta)$, $0 \le \theta \le 2\pi$, is the natural parameter of elements of this subgroup, such that $Z(\frac{\pi}{2})=Z_1$. Define the following $\theta$--family $Y(\theta)$ of Cally numbers by the formula: $$ Y(\theta) = Z_{2} \cdot Z(\theta).$$ Obviously, we have $Y(0)=Z_2$, $Y(\pi) = -Z_2$. Moreover, $Y(\theta)$ is an imaginary Cally number for an arbitrary $\theta$. To proof this, recall that the vector $Z_2$ is orthogonal to $Z^{-1}(\theta)$ for an arbitrary $\theta$. The Cally product is represented by an orthogonal transformation of $\R^8$, therefore, $Z_2 \cdot Z(\theta)$ is orthogonal to $Z^{-1}(\theta) \cdot Z(\theta) = 1$. By the same argument $Y(\theta)$ is orthogonal to $Z_1$. The pair $(Z_1,Y(\theta))$ represents a path in $V_2(\R^7)$, which joins the point $(Z_1,Z_2)$ with the point $(Z_1,-Z_2)$. The projection of the polygon $H(Z_1,Y(\theta))$ onto the line $\{x=0\}$ keeps fixed, because this projection is completely determined by the vector $Z_1 \in \R^7$, and this vector is not changed along the path. But $H(Z_1,-Z_2)$ represents a 7-gone, which is obtained from $H(Z_1,Z_2)$ by the reflection with respect to the line $\{ x=0\} \subset \R^2(x,y)$. Let us complete the proof. Take an arbitrary 7-gone $L \subset \R^2$, $L \in Pol_2(7)$, which is represented by a collection of coordinates, satisfied Conditions --a and --b. Take an arbitrary preimage $(Z_1,Z_2) \in H^{-1}(L) \in V_2(\R^7)$. A path $Z_1,Y(\theta))$ is well-defined. To prove that the path $(Z_1,Y(\theta))$ determines an isometric isotopy of $Pol_2(7)$ to itself, it is sufficient to proof that the projection $H(Z_1,Y(\theta))$ of the path depends no of a lift $(Z_1,Z_2)$ of $L$. The subset $H^{-1}(L) \in V_2(\R^7)$ contains $2^7$ pairs of vectors (probably, less, if a $7$-gone $L$ is degenerated). An arbitrary two vectors of this collection is related by an automorphism $\alpha \in Aut(\Ca)$, which sends a subset of the standard generators of $\Ca$ to the antipodal, and all the last standard generators sends to itself. Obviously, because $H$ is quadratic over each coordinate, by the formula $(\ref{H})$ for an arbitrary $\alpha$ we get: $\alpha \circ H = H$. Lemma $\ref{lemma}$ is proved. \subsubsection*{Proof of Theorem $\ref{main}$} Let $L \in Pol_3(7)$ a 7-gons, $(Q_0,Q_1,Q_2,Q_3) \in V_2(\C^7)$ be a preimage of $L$, $Hopf((Q_0,Q_1,Q_2,Q_3))=L$. Consider the normalization $(Q'_0,0,Q'_2,Q'_3)$ of this preimage. Apply Lemma \ref{lemma} twice: for the projection $L_{x,z}$ with the preimage $(Q'_0,Q'_2)$, $H(Q'_0,Q'_2)=L_{x,z}$, $L_{x,z} \in Pol_2(7;l_{x,z})$, and for the projection $L_{x,y}$ with the preimages $(Q'_0,Q'_3)$, $H^{conj}(Q'_0,Q'_3)=L_{x,y}$, $L_{x,y} \in Pol_2(7;l_{x,y})$. The homotopies for the polygons $(Q'_0,Q'_2)$, $(Q'_0,Q'_3)$ denote by $L_{x,z}(\theta)$, $L_{x,y}(\theta)$ correspondingly. By construction $L_{x,z}(0)=L_{x,z}$, $L_{x,z}(\pi)=-L_{x,z}$, $L_{x,y}(0)=L_{x,y}$, $L_{x,z}(\pi)=-L_{x,y}$, where $-L_{x,z}$, $-L_{x,y}$ are the mirror images of $L_{x,z}$, $L_{x,y}$ with respect to the line $\{x=0\}$, correspondingly. Moreover, the projections of $L_{x,z}(\theta)$ and of $L_{x,y}(\theta)$ onto the axis $\{x=0\}$ coincide and depend no of $\theta$. Therefore for an arbitrary $L \in Pol_3(7)$ the required homotopy $L(\theta)$ from $L=L(0)$ to its mirror image $-L=L(\pi)$ is well-defined. To prove that the family of homotopies $Y(\theta)$ determines an isotopy $\Theta$ of the space $Pol_3(7)$ into itself, it is sufficiently to prove that the homotopies continuously depend on $L$. This follows from the fact, that $L(\theta)$ depends no of a lift $Hopf^{-1}(L)$. By Lemma $\ref{lemma}$ the homotopy of the each projections depends no of its lifts $H^{-1}$. The construction of the homotopies $L(\theta)_{x,z}$, $L(\theta)_{x,y}$ depends continuously of its initial points $H^{-1}(L(\theta)_{x,z})$, $(H^{conj})^{-1}(L(\theta)_{x,y})$. Then, even if for two closed polygons $L_1, L_2 \in Pol_3(7)$ the reduced vectors of the preimages jump, the projections of reduced vectors coincide with $L_1$, $L_2$ correspondingly, and $L(\theta)$ continuously depends of $\theta$. The homotopy $\Theta$ is an isometric isotopy by the construction, evidently, this homotopy is a bijection. Theorem $\ref{main}$ is proved. \[ \] \[ \] Moscow, Troitsk, IZMIRAN $\qquad \qquad \qquad \qquad$ [email protected] $\qquad \qquad \qquad \qquad$

https://arxiv.org/abs/1703.08433

Metric random matchings with applications

Let $(\{1,2,\ldots,n\},d)$ be a metric space. We analyze the expected value and the variance of $\sum_{i=1}^{\lfloor n/2\rfloor}\,d({\boldsymbol{\pi}}(2i-1),{\boldsymbol{\pi}}(2i))$ for a uniformly random permutation ${\boldsymbol{\pi}}$ of $\{1,2,\ldots,n\}$, leading to the following results: (I) Consider the problem of finding a point in $\{1,2,\ldots,n\}$ with the minimum sum of distances to all points. We show that this problem has a randomized algorithm that (1) always outputs a $(2+\epsilon)$-approximate solution in expected $O(n/\epsilon^2)$ time and that (2) inherits Indyk's~\cite{Ind99, Ind00} algorithm to output a $(1+\epsilon)$-approximate solution in $O(n/\epsilon^2)$ time with probability $\Omega(1)$, where $\epsilon\in(0,1)$. (II) The average distance in $(\{1,2,\ldots,n\},d)$ can be approximated in $O(n/\epsilon)$ time to within a multiplicative factor in $[\,1/2-\epsilon,1\,]$ with probability $1/2+\Omega(1)$, where $\epsilon>0$. (III) Assume $d$ to be a graph metric. Then the average distance in $(\{1,2,\ldots,n\},d)$ can be approximated in $O(n)$ time to within a multiplicative factor in $[\,1-\epsilon,1+\epsilon\,]$ with probability $1/2+\Omega(1)$, where $\epsilon=\omega(1/n^{1/4})$.

\section{Introduction} A metric space is a nonempty set $M$ endowed with a metric, i.e., a function $d\colon M\times M\to[\,0,\infty\,)$ such that \begin{itemize} \item $d(x,y)=0$ if and only if $x=y$ (identity of indiscernibles), \item $d(x,y)=d(y,x)$ (symmetry), and \item $d(x,y)+d(y,z)\ge d(x,z)$ (triangle inequality) \end{itemize} for all $x$, $y$, $z\in M$~\cite{Rud76}. For all $n\in\mathbb{Z}^+$, define $[n]\equiv\{1,2,\ldots,n\}$. Given $n\in\mathbb{Z}^+$ and oracle access to a metric $d\colon [n]\times[n]\to[\,0,\infty\,)$, {\sc metric $1$-median} asks for $\mathop{\mathrm{argmin}}_{y\in[n]}\,\sum_{x\in[n]}\,d(y,x)$, breaking ties arbitrarily. It generalizes the classical median selection on the real line and has a brute-force $\Theta(n^2)$-time algorithm. More generally, {\sc metric $k$-median} asks for $c_1$, $c_2$, $\ldots$, $c_k\in[n]$ minimizing $\sum_{x\in[n]}\,\min_{i=1}^k\,d(x,c_i)$. Because $d(\cdot,\cdot)$ defines $\binom{n}{2}=\Theta(n^2)$ nonzero distances, only $o(n^2)$-time algorithms are said to run in sublinear time~\cite{Ind99}. For all $\alpha\ge1$, an $\alpha$-approximate $1$-median is a point $p\in[n]$ satisfying $$\sum_{x\in[n]}\,d\left(p,x\right)\le\alpha\cdot \min_{y\in[n]}\,\sum_{x\in[n]}\,d\left(y,x\right).$$ For all $\epsilon>0$, {\sc metric $1$-median} has a Monte Carlo $(1+\epsilon)$-approximation $O(n/\epsilon^2)$-time algorithm~\cite{Ind99, Ind00}. Guha et al.~\cite{GMMMO03} show that {\sc metric $k$-median} has a Monte Carlo, $O(\exp(O(1/\epsilon)))$-approximation, $O(nk\log n)$-time, $O(n^{\epsilon})$-space and one-pass algorithm for all small $k$ as well as a deterministic, $O(\exp(O(1/\epsilon)))$-approximation, $O(n^{1+\epsilon})$-time, $O(n^{\epsilon})$-space and one-pass algorithm. Given $n$ points in $\mathbb{R}^D$ with $D\ge 1$, the Monte Carlo algorithms of Kumar et al.~\cite{KSS10} find a $(1+\epsilon)$-approximate $1$-median in $O(D\cdot\exp(1/\epsilon^{O(1)}))$ time and a $(1+\epsilon)$-approximate solution to {\sc metric $k$-median} in $O(Dn\cdot\exp((k/\epsilon)^{O(1)}))$ time. All randomized $O(1)$-approximation algorithms for {\sc metric $k$-median} take $\Omega(nk)$ time~\cite{MP04, GMMMO03}. Chang~\cite{Cha15} shows that {\sc metric $1$-median} has a deterministic, $(2h)$-approximation, $O(hn^{1+1/h})$-time and nonadaptive algorithm for all constants $h\in\mathbb{Z}^+\setminus\{1\}$, generalizing the results of Chang~\cite{Cha13} and Wu~\cite{Wu14}. On the other hand, he disproves the existence of deterministic $(2h-\epsilon)$-approximation $O(n^{1+1/(h-1)}/h)$-time algorithms for all constants $h\in\mathbb{Z}^+\setminus\{1\}$ and $\epsilon>0$~\cite{Cha16COCOON, Cha17}. In social network analysis, the closeness centrality of a point $v$ is the reciprocal of the average distance from $v$ to all points~\cite{WF94}. So {\sc metric $1$-median} asks for a point with the maximum closeness centrality. Given oracle access to a graph metric, the Monte-Carlo algorithms of Goldreich and Ron~\cite{GR08} and Eppstein and Wang~\cite{EW04} estimate the closeness centrality of a given point and those of all points, respectively. All known sublinear-time algorithms for {\sc metric $1$-median} are either deterministic or Monte Carlo, the latter having a positive probability of failure. For example, Indyk's Monte Carlo $(1+\epsilon)$-approximation algorithm outputs with a positive probability a solution without approximation guarantees. In contrast, we show that {\sc metric $1$-median} has a randomized algorithm that {\em always} outputs a $(2+\epsilon)$-approximate solution in expected $O(n/\epsilon^2)$ time for all $\epsilon\in(0,1)$. So, excluding the known deterministic algorithms (which are Las Vegas only in the degenerate sense), this paper gives the {\em first} Las Vegas approximation algorithm for {\sc metric $1$-median} with an expected sublinear running time. Note that deterministic sublinear-time algorithms for {\sc metric $1$-median} can be $4$-approximate but not $(4-\epsilon)$-approximate for any constant $\epsilon>0$~\cite{Cha13, Cha17}. So our approximation ratio of $2+\epsilon$ beats that of any deterministic sublinear-time algorithm. Inheriting Indyk's algorithm, our algorithm outputs a $(1+\epsilon)$-approximate $1$-median in $O(n/\epsilon^2)$ time with probability $\Omega(1)$ for all $\epsilon\in(0,1)$. \comment{ In case our algorithm fails to output a $(1+\epsilon)$-approximate $1$-median, it nonetheless outputs a $(2+\epsilon)$-approximate $1$-median. \comment{ As a by-product of our derivations, we present a Monte Carlo $O(n/\epsilon^2)$-time algorithm for estimating $\sum_{u,v\in [n]}\,d(u,v)$ to within an additive error of $\epsilon\cdot\sum_{u,v\in [n]}\,d(u,v)$ with an $\Omega(1)$ probability of success, for each constant $\epsilon>0$. Previously, the best algorithm for such estimation needs $O(n/\epsilon^{7/2})$ time~\cite[Sec.~8]{Ind99}. \comment{ So we have the first Las Vegas approximation algorithm for {\sc metric $1$-median} with an expected sublinear running time. Because the best approximation ratio achievable by deterministic sublinear-time algorithms for {\sc metric $1$-median} is $4$~\cite{Cha13, Cha15}, Las Vegas approximation algorithms have a better approximation ratio. Indyk~\cite{Ind99, Ind00} gives a Monte-Carlo $O(n/\epsilon^{3.5})$-time algorithm that approximates the average distance in any metric space $([n],d)$ to within a multiplicative factor in $[\,1-\epsilon,1+\epsilon\,]$, for all $\epsilon>0$. Barhum, Goldreich and Shraibman~\cite{BGS07} improve Indyk's time complexity of $O(n/\epsilon^{3.5})$ to $O(n/\epsilon^2)$. This paper gives a Monte-Carlo $O(n/\epsilon)$-time algorithm that approximates the average distance in $([n],d)$ to within a multiplicative factor in $[\,1/2-\epsilon,1\,]$, for all $\epsilon>0$. For all $\epsilon=\omega(1/n^{1/4})$, we present a Monte-Carlo $O(n)$-time algorithm approximating the average distance of any graph metric to within a multiplicative factor in $[\,1-\epsilon,1+\epsilon\,]$. But for general metrics, we do not know whether the $O(n/\epsilon^2)$ running time of Barhum, Goldreich and Shraibman can be improved to $O(n/\epsilon^{2-\Omega(1)})$. \section{Definitions and preliminaries} For a metric space $([n],d)$, \begin{eqnarray} \bar{r}&\equiv& \frac{1}{n^2}\cdot\sum_{x,y\in[n]}\,d\left(x,y\right), \label{averagedistance}\\ p^* &\equiv& \mathop{\mathrm{argmin}}_{p\in[n]}\,\sum_{x\in [n]}\, d\left(p,x\right),\label{optimalpoint} \end{eqnarray} breaking ties arbitrarily in equation~(\ref{optimalpoint}). So $\bar{r}$ is the average distance in $([n],d)$, and $p^*$ is a $1$-median. \comment{ For brevity, $$ B^2\left(x,R\right)\equiv B\left(x,R\right)\times B\left(x,R\right). $$ The pairs in $B^2(x,R)$ are ordered. An algorithm $A$ with oracle access to $d\colon [n]\times[n]\to[\,0,\infty\,)$ is denoted by $A^d$ and may query $d$ on any $(x,y)\in[n]\times[n]$ for $d(x,y)$. In this paper, all Landau symbols (such as $O(\cdot)$, $o(\cdot)$, $\Omega(\cdot)$ and $\omega(\cdot)$) are w.r.t.\ $n$. The following result is due to Indyk. \begin{fact}[\cite{Ind99, Ind00}]\label{Indykfact} For all $\epsilon>0$, {\sc metric $1$-median} has a Monte Carlo $(1+\epsilon)$-approximation $O(n/\epsilon^2)$-time algorithm with a failure probability of at most $1/e$. \end{fact} Henceforth, denote Indyk's algorithm in Fact~\ref{Indykfact} by {\sf Indyk median}. It is given $n\in\mathbb{Z}^+$, $\epsilon>0$ and oracle access to a metric $d\colon [n]\times[n]\to[\,0,\infty\,)$. The following fact on estimating the average distance is due to Barhum, Goldreich and Shraibman. \begin{fact}[\cite{BGS07}]\label{averagedistancepriorresult} Given $n\in\mathbb{Z}^+$, $\epsilon>0$ and oracle access to a metric $d\colon[n]\times[n] \to[\,0,\infty\,)$, a real number in $\left[\,(1-\epsilon)\bar{r},(1+\epsilon)\bar{r}\,\right]$ can be found in $O(n/\epsilon^2)$ time with probability at least $1/2+\Omega(1)$. \end{fact} \comment{ \begin{fact}[Implicit in~{\cite[Theorem~22]{Cha12}}]\label{uniformlyrandompointisgood} A uniformly random point of $[n]$ is a $4$-approximate $1$-median with probability at least $1/2$. \end{fact} \comment{ \begin{markov} [\cite{MR95}] For a non-negative random variable $X$ and $a>0$, $$ \Pr \left|\, X\ge a \,\right]\le \frac{\mathop{\mathrm{E}}[X]}{a}. $$ \end{markov} \begin{chebyshev} [\cite{MR95}] Let $X$ be a random variable with a finite expected value and a finite nonzero variance. Then for all $k\ge1$, $$ \Pr\left[\, \left|\, X-\mathop{\mathrm E}[X]\,\right|\ge k\sqrt{\mathop{\mathrm{var}}(X)} \,\right]\le \frac{1}{k^2}. $$ \end{chebyshev} \section{Las Vegas approximation for metric $1$-median selection}\label{mainresultsection} \comment{ Clearly, \begin{eqnarray} \sum_{x\in[n]}\,d\left(p^*,x\right) \stackrel{\text{(\ref{optimalpoint})}}{\le} \frac{1}{n}\cdot\sum_{p\in[n]}\,\sum_{x\in[n]}\,d\left(p,x\right) \stackrel{\text{(\ref{averagedistance})}}{=} nr. \end{eqnarray} This section presents a randomized algorithm that {\em always} outputs a $(2+\epsilon)$-approximate $1$-median, where $\epsilon\in(0,1)$. Clearly, \begin{eqnarray} \sum_{x\in[n]}\,d\left(p^*,x\right) \stackrel{\text{(\ref{optimalpoint})}}{=} \min_{p\in[n]}\,\sum_{x\in[n]}\,d\left(p,x\right) \le \frac{1}{n}\cdot \sum_{p\in[n]}\,\sum_{x\in[n]}\,d\left(p,x\right) \stackrel{\text{(\ref{averagedistance})}}{=} n\bar{r}.\label{bestisnoworsethanaverage} \end{eqnarray} For each permutation $\pi\colon[n]\to[n]$, \begin{eqnarray} \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\pi\left(2i-1\right),\pi\left(2i\right)\right) \le \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(p^*,\pi\left(2i-1\right)\right)+d\left(p^*,\pi\left(2i\right)\right) \le \sum_{x\in[n]}\,d\left(p^*,x\right), \label{matchingsizeandoptimalsumofdistances} \end{eqnarray} where the first and the second inequalities follow from the triangle inequality and the injectivity of $\pi$. \begin{lemma}\label{lemmaforapproximationratio} When line~5 of {\sf Las Vegas median} in Fig.~\ref{lasvegasalgorithm} is run, $z$ is a $(2+\epsilon)$-approximate $1$-median. \comment{ For all $y\in[n]$ and permutations $\pi$, if $$ \sum_{x\in[n]}\,d\left(y,x\right) \le \left(2+\epsilon\right) \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\pi\left(2i-1\right),\pi\left(2i\right)\right), $$ then $y$ is a $(2+\epsilon)$-approximate $1$-median. \end{lemma} \begin{proof} The condition in line~4 of {\sf Las Vegas median} implies $$ \sum_{x\in[n]}\, d\left(z,x\right) \stackrel{\text{(\ref{matchingsizeandoptimalsumofdistances})}}{\leq} \left(2+\epsilon\right) \sum_{x\in[n]}\, d\left(p^*,x\right) \stackrel{\text{(\ref{optimalpoint})}}{=} \left(2+\epsilon\right) \min_{p\in[n]}\,\sum_{x\in[n]}\, d\left(p,x\right). $$ So when line~5 is run, it returns a $(2+\epsilon)$-approximate $1$-median. \end{proof} \begin{figure} \begin{algorithmic}[1] \WHILE{\text{\sf true}} \STATE $z\leftarrow\text{\sf Indyk median}^d(n,\epsilon/8)$; \STATE Pick independent and uniformly random permutations $\bs{\pi}_1$, $\bs{\pi}_2$, $\ldots$, $\bs{\pi}_{80\lceil 1/\epsilon\rceil}\colon[n]\to[n]$; \IF{there exists $j\in[\lceil 1/\epsilon\rceil]$ satisfying $\sum_{x\in[n]}\,d(z,x)\le (2+\epsilon)\sum_{i=1}^{\lfloor n/2\rfloor}\,d(\bs{\pi}_j(2i-1),\bs{\pi}_j(2i))$} \RETURN $z$; \ENDIF \ENDWHILE \end{algorithmic} \caption{Algorithm {\sf Las Vegas median} with oracle access to a metric $d\colon [n]\times[n]\to[\,0,\infty\,)$ and with inputs $n\in\mathbb{Z}^+$ and $\epsilon\in(0,1)$} \label{lasvegasalgorithm} \end{figure} Inequalities~(\ref{bestisnoworsethanaverage})--(\ref{matchingsizeandoptimalsumofdistances}) yield the following. \begin{lemma}\label{maximummatchingsize} For each permutation $\pi\colon[n]\to[n]$, $$ \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\pi\left(2i-1\right),\pi\left(2i\right)\right) \le n\bar{r}. $$ \end{lemma} \comment{ \begin{proof} We have \begin{eqnarray*} \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\pi\left(2i-1\right),\pi\left(2i\right)\right) \stackrel{\text{(\ref{optimalpoint})--(\ref{matchingsizeandoptimalsumofdistances})}}{\le} \frac{1}{n}\cdot\sum_{p\in[n]}\,\sum_{x\in[n]}\,d\left(p,x\right) \stackrel{\text{(\ref{averagedistance})}}{=} nr. \end{eqnarray*} \end{proof} \begin{lemma}\label{expectedmatchingsize} For a uniformly random permutation $\bs{\pi}\colon [n]\to[n]$, $$ \mathop{\mathrm{E}}_{\bs{\pi}} \left[\,\sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \right] =\left\lfloor\frac{n}{2}\right\rfloor\cdot\frac{n\bar{r}}{n-1}. $$ \end{lemma} \begin{proof} For each $i\in[\lfloor n/2\rfloor]$, $\{\bs{\pi}(2i-1),\bs{\pi}(2i)\}$ is a uniformly random size-$2$ subset of $[n]$, implying \begin{eqnarray*} \mathop{\mathrm{E}}_{\bs{\pi}} \left[\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \right] &=&\frac{1}{n\cdot(n-1)}\cdot\sum_{\text{distinct $x$, $y\in[n]$}}\, d\left(x,y\right)\\ &=&\frac{1}{n\cdot(n-1)}\cdot\sum_{x,y\in[n]}\,d\left(x,y\right)\\ &\stackrel{\text{(\ref{averagedistance})}}{=}& \frac{n\bar{r}}{n-1}, \end{eqnarray*} where the second equality follows from the identity of indiscernibles. Finally, use the linearity of expectation. \end{proof} \comment{ \begin{lemma}\label{probabilitythatarandommatchingislarge} For a uniformly random permutation $\bs{\pi}\colon [n]\to[n]$ and all sufficiently large $n$, \begin{eqnarray} \Pr_{\bs{\pi}} \left[\,\sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \ge\left(\frac{1}{2}-\frac{\epsilon}{8}\right)nr \right] \ge \frac{\epsilon}{8}. \label{probabilityofasmallmatching} \end{eqnarray} \end{lemma} \begin{proof} Denote the left-hand side of inequality~(\ref{probabilityofasmallmatching}) by $p$. Then by Lemma~\ref{maximummatchingsize}, $$ \mathop{\mathrm{E}}_{\bs{\pi}} \left[\, \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \,\right] \le p nr + \left(1-p\right)\left(\frac{1}{2}-\frac{\epsilon}{8}\right)nr. $$ This and Lemma~\ref{expectedmatchingsize} complete the proof. \end{proof} \begin{lemma}\label{probabilitythatoneoftherandommatchingsislarge} For all $\epsilon\in(0,1)$ and in each iteration of the {\bf while} loop of {\sf Las Vegas median}, \begin{eqnarray} \Pr \left[\, \exists j\in\left[80\cdot\left\lceil\frac{1}{\epsilon}\right\rceil\right],\, \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}_j\left(2i-1\right),\bs{\pi}_j\left(2i\right)\right) \ge\left(\frac{1}{2}-\frac{\epsilon}{8}\right)n\bar{r} \right] \ge0.9, \label{probabilityofalargematching} \end{eqnarray} where the probability is taken over $\bs{\pi}_1$, $\bs{\pi}_2$, $\ldots$, $\bs{\pi}_{80\lceil 1/\epsilon\rceil}$ in line~3 of {\sf Las Vegas median}. \end{lemma} \begin{proof} Let $\bs{\pi}\colon[n]\to[n]$ be a uniformly random permutation and \begin{eqnarray} \alpha=\Pr_{\bs{\pi}} \left[\,\sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \ge\left(\frac{1}{2}-\frac{\epsilon}{8}\right)n\bar{r} \right]. \end{eqnarray} So by Lemma~\ref{maximummatchingsize}, $$ \mathop{\mathrm{E}}_{\bs{\pi}} \left[\, \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \,\right] \le \alpha n\bar{r} + \left(1-\alpha\right)\left(\frac{1}{2}-\frac{\epsilon}{8}\right) n\bar{r}. $$ This and Lemma~\ref{expectedmatchingsize} imply $\alpha\ge \epsilon/8$. So the left-hand side of inequality~(\ref{probabilityofalargematching}) is at least $1-(1-\epsilon/8)^{80\lceil 1/\epsilon\rceil}\ge0.9$. \end{proof} \comment{ \begin{lemma} For each $y\in[n]$ and a uniformly random permutation $\bs{\pi}$, if $$ \sum_{x\in[n]}\,d\left(y,x\right) \le \left(2+\epsilon\right) \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right), $$ then $y$ is a $(2+\epsilon)$-approximate $1$-median. \end{lemma} \comment{ For all $y\in[n]$, $$\sum_{x\in[n]}\,d\left(y,x\right) \ge \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(y,\bs{\pi}\left(2i-1\right)\right) +d\left(y,\bs{\pi}\left(2i\right)\right) \ge \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right), \bs{\pi}\left(2i\right)\right), $$ where the first and the second inequalities follow from the injectivity of $\bs{\pi}$ and the triangle inequality. So inequalities~(\ref{qualityofanearoptimalsolution})--(\ref{lowerboundofarandommatchingsize}) imply the approximation ratio of $z$ to be at most $$ \frac{1+\epsilon/8}{} $$ \comment{ \begin{lemma}\label{probabilitythatIndykalgorithmgivesagoodsolution} In each execution of line~2 of {\sf Las Vegas median}, $$ \Pr\left[\, \sum_{x\in[n]}\,d\left(z,x\right) \le \left(1+\frac{\epsilon}{8}\right) nr \,\right] =\Omega(1), $$ where the probability is taken over the random coin tosses of {\sf Indyk median}. \end{lemma} \begin{proof} By Fact~\ref{Indykfact}, $$ \Pr\left[\, \sum_{x\in[n]}\,d\left(z,x\right) \le \left(1+\frac{\epsilon}{8}\right) \sum_{x\in[n]}\,d\left(p^*,x\right) \,\right] =\Omega(1). $$ This and inequality~(\ref{bestisnoworsethanaverage}) complete the proof. \end{proof} \begin{lemma}\label{probabilitythatIndykalgorithmisgoodandmatchingislarge} For all $\epsilon\in(0,1)$ and in each iteration of the {\bf while} loop of {\sf Las Vegas median}, \begin{eqnarray} &&\Pr\left[\, \left( \sum_{x\in[n]}\,d\left(z,x\right) \le \left(1+\frac{\epsilon}{8}\right) \min_{p\in[n]}\,\sum_{x\in[n]}\,d\left(p,x\right) \right)\right.\nonumber\\ &&\left. \land\left( \exists j\in\left[80\cdot\left\lceil\frac{1}{\epsilon}\right\rceil\right],\, \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}_j\left(2i-1\right),\bs{\pi}_j\left(2i\right)\right) \ge \left(\frac{1}{2}-\frac{\epsilon}{8}\right)n\bar{r} \right) \right.\nonumber\\ &&\left. \land \left(\exists j\in\left[80\cdot\left\lceil\frac{1}{\epsilon}\right\rceil\right],\, \sum_{x\in[n]}\,d\left(z,x\right) \le \left(2+\epsilon\right)\sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}_j\left(2i-1\right),\bs{\pi}_j\left(2i\right)\right) \right) \,\right]\nonumber\\ &=&\frac{1}{2}+\Omega(1),\label{ultimateequation} \end{eqnarray} where the probability is taken over $\bs{\pi}_1$, $\bs{\pi}_2$, $\ldots$, $\bs{\pi}_{80\lceil 1/\epsilon\rceil}$ and the random coin tosses of {\sf Indyk median}. \end{lemma} \begin{proof} By Fact~\ref{Indykfact} and line~2 of {\sf Las Vegas median}, the first condition within $\Pr[\cdot]$ in equation~(\ref{ultimateequation}) holds with probability at least $1-1/e$ over the random coin tosses of {\sf Indyk median}. By Lemma~\ref{probabilitythatoneoftherandommatchingsislarge}, \comment{ \begin{eqnarray*} \sum_{x\in[n]}\,d\left(z,x\right) &\le& \left(1+\frac{\epsilon}{8}\right) nr,\\ \exists j\in\left[\frac{1}{\epsilon}\right],\, \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}_j\left(2i-1\right),\bs{\pi}_j\left(2i\right)\right) &\ge& \left(\frac{1}{2}-\frac{\epsilon}{8}\right)nr \end{eqnarray*} the second condition holds with probability at least $0.9$ over $\bs{\pi}_1$, $\bs{\pi}_2$, $\ldots$, $\bs{\pi}_{80\lceil 1/\epsilon\rceil}$. In summary, the first two conditions hold simultaneously with probability at least $(1-1/e)\cdot 0.9=1/2+\Omega(1)$ (note that the random coin tosses of {\sf Indyk median} are independent of $\bs{\pi}_1$, $\bs{\pi}_2$, $\ldots$, $\bs{\pi}_{80\lceil 1/\epsilon\rceil}$). Finally, the first two conditions together imply the third by inequality~(\ref{bestisnoworsethanaverage}) and the easy fact that $$ \left(1+\frac{\epsilon}{8}\right)\le\left(2+\epsilon\right) \left(\frac{1}{2}-\frac{\epsilon}{8}\right). $$ \end{proof} \comment{ \begin{proof} By Fact~\ref{Indykfact}, $$ \Pr\left[\, \sum_{x\in[n]}\,d\left(z,x\right) \le \left(1+\frac{\epsilon}{8}\right) \sum_{x\in[n]}\,d\left(p^*,x\right) \,\right] =\Omega(1), $$ where the probability is taken over the random coin tosses of {\sf Indyk median}. This and inequality~(\ref{bestisnoworsethanaverage}) give $$ \Pr\left[\, \sum_{x\in[n]}\,d\left(z,x\right) \le \left(1+\frac{\epsilon}{8}\right) nr \,\right] =\Omega(1), $$ which together with Lemma~\ref{probabilitythatarandommatchingislarge} completes the proof (note that $\bs{\pi}$ is independent of the random coin tosses of {\sf Indyk median}). \end{proof} \begin{theorem}\label{maintheorem} For all $\epsilon\in(0,1)$, {\sc metric $1$-median} has a randomized algorithm that (1)~{\em always} outputs a $(2+\epsilon)$-approximate solution in expected $O(n/\epsilon^2)$ time and (2)~outputs a $(1+\epsilon)$-approximate solution in $O(n/\epsilon^2)$ time with probability $\Omega(1)$. \end{theorem} \begin{proof} By Lemma~\ref{probabilitythatIndykalgorithmisgoodandmatchingislarge}, each execution of lines~4--5 of {\sf Las Vegas median} returns with probability $1/2+\Omega(1)$. So the expected number of iterations is $O(1)$. By Fact~\ref{Indykfact}, line~2 takes $O(n/\epsilon^2)$ time. Line~3 takes $80\lceil 1/\epsilon\rceil\cdot O(n)$ time by the Knuth shuffle. Clearly, lines~4--5 take $O(n/\epsilon)$ time. In summary, the expected running time of {\sf Las Vegas median} is $O(1)\cdot O(n/\epsilon^2)=O(n/\epsilon^2)$. To prevent {\sf Las Vegas median} from running forever, find a $1$-median by brute force (which obviously takes $O(n^2)$ time) after $n^2$ steps of computation. By Lemma~\ref{lemmaforapproximationratio}, {\sf Las Vegas median} is $(2+\epsilon)$-approximate. By Lemma~\ref{probabilitythatIndykalgorithmisgoodandmatchingislarge}, $z$ is $(1+\epsilon/8)$-approximate and is also returned in line~5 with probability $\Omega(1)$ in the first (in fact, any) iteration. Finally, the previous paragraph has shown each iteration to take $O(n/\epsilon^2)$ time. \end{proof} \comment{ \begin{eqnarray} \sum_{x\in[n]}\,d\left(z,x\right) &\le& \left(1+\frac{\epsilon}{8}\right) nr,\label{qualityofanearoptimalsolution}\\ \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) &\ge& \left(\frac{1}{2}-\frac{\epsilon}{8}\right)nr, \label{lowerboundofarandommatchingsize} \end{eqnarray} By Fact~\ref{Indykfact}, {\sf Indyk median} satisfies condition~(2) in Theorem~\ref{maintheorem}. But it does not satisfy condition~(1). We now justify the optimality of the ratio of $2+\epsilon$ in Theorem~\ref{maintheorem}. Let $A$ be a randomized algorithm that always outputs a $(2-\epsilon)$-approximate $1$-median. Furthermore, denote by $p\in [n]$ (resp., $Q\subseteq [n]\times [n]$) the output (resp., the set of queries as unordered pairs) of $A^{d_1}(n)$, where $d_1$ is the discrete metric (i.e., $d_1(x,y)=1$ and $d_1(x,x)=0$ for all distinct $x$, $y\in [n]$). Without loss of generality, assume $(p,y)\in Q$ for all $y\in [n]\setminus\{p\}$ by adding dummy queries. So the queries in $Q$ witness that \begin{eqnarray} \sum_{y\in [n]\setminus\{p\}}\,d_1\left(p,y\right)=n-1. \label{outputunderthediscretemetric} \end{eqnarray} Assume without loss of generality that $A$ never queries for the distance from a point to itself. In the sequel, consider the case that $|Q|<\epsilon\cdot(n-1)^2/8$. By the averaging argument, there exists a point $\hat{p}\in [n]\setminus \{p\}$ involved in at most $2\cdot|Q|/(n-1)$ queries in $Q$ (note that each query involves two points). Because every function $f\colon [n]\times[n]\to[\,0,\infty\,)$ with $$\left\{f\left(x,y\right)\mid \left(x,y\in[n]\right)\land\left( x\neq y\right)\right\}\subseteq\left\{\frac{1}{2},1\right\}$$ satisfies the triangle inequality, $A$ cannot exclude the possibility that $d_1(\hat{p},y)=1/2$ for all $y\in[n]\setminus\{\hat{p}\}$ satisfying $(\hat{p},y)\notin Q$. In summary, $A$ cannot rule out the case that \begin{eqnarray} \sum_{y\in[n]}\,d_1\left(\hat{p},y\right)&\le& \frac{2\cdot |Q|}{n-1}\cdot 1 +\left(n-1-\frac{2\cdot |Q|}{n-1}\right)\cdot \frac{1}{2} < \left(\frac{1}{2}+\frac{\epsilon}{8}\right)\cdot(n-1).\,\,\,\,\, \label{acasethatcannotberuledout} \end{eqnarray} Equations~(\ref{outputunderthediscretemetric})--(\ref{acasethatcannotberuledout}) contradict the guarantee that $p$ is $(2-\epsilon)$-approximate. Consequently, the case that $|Q|<\epsilon\cdot(n-1)^2/8$ should {\em never} happen. The next theorem summarizes the above. \begin{theorem} {\sc Metric $1$-median} has no randomized algorithm that always outputs a $(2-\epsilon)$-approximate solution and that makes fewer than $\epsilon\cdot (n-1)^2/8$ queries with a positive probability given oracle access to the discrete metric, for any constant $\epsilon\in(0,1)$. \end{theorem} Lemmas~\ref{maximummatchingsize}~and~\ref{probabilitythatoneoftherandommatchingsislarge} yield the following estimation of the average distance. \begin{theorem}\label{theoremforaveragedistance} Given $n\in\mathbb{Z}^+$, $\epsilon>0$ and oracle access to a metric $d\colon[n]\times[n]\to[\,0,\infty\,)$, a real number in $[\,(1/2-\epsilon)\bar{r},\bar{r}\,]$ can be found in $O(n/\epsilon)$ time with probability $1/2+\Omega(1)$. \end{theorem} \begin{proof} By Lemmas~\ref{maximummatchingsize}~and~\ref{probabilitythatoneoftherandommatchingsislarge}, \begin{eqnarray} \frac{1}{n}\cdot \max_{j\in[80\cdot\lceil 1/\epsilon\rceil]}\, \sum_{i=1}^{\lfloor n/2\rfloor}\,d\left(\bs{\pi}_j\left(2i-1\right), \bs{\pi}_j\left(2i\right)\right) \in\left[\,\left(\frac{1}{2}-\frac{\epsilon}{8}\right)\bar{r}, \bar{r}\,\right] \label{rangeofestimationofaveragedistance} \end{eqnarray} with probability $1/2+\Omega(1)$. The Knuth shuffle picks $\bs{\pi}_1$, $\bs{\pi}_2$, $\ldots$, $\bs{\pi}_{80\lceil 1/\epsilon\rceil}$ in $80\lceil 1/\epsilon\rceil\cdot O(n)$ time. Then the left-hand side of relation~(\ref{rangeofestimationofaveragedistance}) can be calculated in $O(n/\epsilon)$ time. \end{proof} Note that the estimation of the average distance in Theorem~\ref{theoremforaveragedistance} has only one-sided error. The time complexity (resp., approximation ratio) in Theorem~\ref{theoremforaveragedistance} is better (resp., worse) than that in Fact~\ref{averagedistancepriorresult}. \comment{ But we do not know whether the time complexity in Theorem~\ref{theoremforaveragedistance} and the approximation ratio in Fact~\ref{averagedistancepriorresult} can be achieved simultaneously. \section{Estimating the average distance of a graph metric}\label{approximationratiosection} Throughout this section, take any $\epsilon=\omega(1/n^{1/4})$ less than a small constant, e.g., $\epsilon=10^{-100}$. Define \begin{eqnarray} \delta&\equiv& \frac{\epsilon^2}{10^{10}},\label{thedeltavalue}\\ r&\equiv&\frac{1}{n}\cdot\sum_{x\in[n]}\,d\left(p^*,x\right), \label{theaveragedistancefromanoptimalpoint} \end{eqnarray} where $p^*$ is as in equation~(\ref{optimalpoint}). As $\epsilon=\omega(1/n^{1/4})$, $\delta=\omega(1/\sqrt{n})$ by equation~(\ref{thedeltavalue}). \comment{ By line~1 of {\sf average distance} in Fig.~\ref{mainalgorithm}, $\delta>0$ is likewise small, and $\delta=1/n^{o(1)}$. Furthermore, take $\bs{u}$, $r$ and $\bs{\pi}$ as in lines~2--4 of {\sf average distance}. \comment{ The following lemma implies that $z$ in line~3 of {\sf Las Vegas median} is a solution (to {\sc metric $1$-median}) no worse than those in $[n]\setminus B(z,8r)$, where $r$ is as in line~4. \comment{ This section analyzes the probability of running line~7 in any particular iteration of the {\bf while} loop of {\sf Las Vegas median}. The following lemma uses an easy averaging argument. \comment{ Assume $n$ to be sufficiently large so that $|B(z,\delta nr)|\ge 4$ by Lemma~\ref{thesmallradiusballislarge}. \comment{ Define \begin{eqnarray} r' \equiv \frac{1}{|B(z,\delta nr)|^2} \cdot \sum_{u, v\in B(z,\delta nr)}\, d\left(u,v\right) \label{smallerballaveragedefinition} \end{eqnarray} to be the average distance in $B(z,\delta nr)$. \comment{ As $z\in B(z,\delta nr)$, the denominator in the right-hand side of equation~(\ref{smallerballaveragedefinition}) is nonzero. \begin{lemma}\label{inneraverageandoverallaverage} $\bar{r}\leq 2r$. \end{lemma} \begin{proof} By equation~(\ref{averagedistance}) and the triangle inequality, \begin{eqnarray} \bar{r} &\le& \frac{1}{n^2} \cdot \sum_{x, y\in [n]}\, \left(d\left(p^*,x\right)+d\left(p^*,y\right)\right) \nonumber\\ &=& \frac{1}{n^2} \cdot n\cdot\left( \sum_{x\in [n]}\, d\left(p^*,x\right) +\sum_{y\in [n]}\, d\left(p^*,y\right) \right)\nonumber\\ &=& \frac{2}{n} \cdot \sum_{x\in [n]}\, d\left(p^*,x\right). \label{needanumberhere} \end{eqnarray} Equations~(\ref{theaveragedistancefromanoptimalpoint})--(\ref{needanumberhere}) complete the proof. \end{proof} \begin{figure} \begin{algorithmic}[1] \STATE Pick a uniformly random permutation $\bs{\pi}\colon [n]\to [n]$; \RETURN $\sum_{i=1}^{\lfloor n/2\rfloor}\,d(\bs{\pi}(2i-1),\bs{\pi}(2i))\cdot 2/n$; \end{algorithmic} \caption{Algorithm {\sf average distance} with oracle access to a metric $d\colon [n]\times[n]\to[\,0,\infty\,)$ and with inputs $n\in\mathbb{Z}^+$ and $\epsilon=\omega(1/n^{1/4})$.} \label{mainalgorithm} \end{figure} As in line~1 of {\sf average distance} in Fig.~\ref{mainalgorithm}, let $\bs{\pi}\colon[n]\to[n]$ be a uniformly random permutation. Clearly, \begin{eqnarray} &&\mathop{\mathrm E}_{\bs{\pi}}\left[\, \left(\sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right)\right)^2 \,\right] \label{startofthemeanofsquare}\\ &=& \mathop{\mathrm E}_{\bs{\pi}}\left[\, \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \cdot \sum_{j=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2j-1\right),\bs{\pi}\left(2j\right)\right) \,\right]\nonumber\\ &=& \sum_{\text{distinct $i,j=1$}}^{\lfloor n/2\rfloor}\, \mathop{\mathrm E}_{\bs{\pi}}\left[\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \cdot d\left(\bs{\pi}\left(2j-1\right),\bs{\pi}\left(2j\right)\right) \,\right]\nonumber\\ &+& \sum_{i=1}^{\lfloor n/2\rfloor}\, \mathop{\mathrm E}_{\bs{\pi}}\left[\, d^2\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \,\right],\label{endofthemeanofsquare} \end{eqnarray} where the last equality follows from the linearity of expectation and the separation of pairs $(i,j)$ according to whether $i=j$. The next three lemmas analyze the variance of $$ \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right). $$ \begin{lemma}\label{distinctdistancesproduct} {\small \begin{eqnarray*} \sum_{\text{\rm distinct $i,j=1$}}^{\lfloor n/2\rfloor}\, \mathop{\mathrm E}_{\bs{\pi}}\left[\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \cdot d\left(\bs{\pi}\left(2j-1\right),\bs{\pi}\left(2j\right)\right) \,\right] \le\frac{1}{4}\cdot \left(1 + O\left(\frac{1}{n}\right) \right)n^2 \bar{r}^2. \end{eqnarray*} \end{lemma} \begin{proof} Pick any distinct $i$, $j\in[\,\lfloor n/2\rfloor\,]$. Clearly, $$\left\{\bs{\pi}\left(2i-1\right), \bs{\pi}\left(2i\right), \bs{\pi}\left(2j-1\right), \bs{\pi}\left(2j\right)\right\}$$ is a uniformly random size-$4$ subset of $[n]$. So \begin{eqnarray*} &&\mathop{\mathrm E}_{\bs{\pi}}\left[\, d\left(\bs{\pi}(2i-1),\bs{\pi}(2i)\right) \cdot d\left(\bs{\pi}(2j-1),\bs{\pi}(2j)\right) \,\right]\\ &=& \frac{1}{n \cdot(n-1)\cdot(n-2) \cdot(n-3)} \cdot \sum_{\text{distinct $u$, $v$, $x$, $y\in [n]$}}\, d\left(u,v\right)\cdot d\left(x,y\right). \end{eqnarray*} Clearly, \begin{eqnarray*} \sum_{\text{distinct $u$, $v$, $x$, $y\in [n]$}}\, d\left(u,v\right)\cdot d\left(x,y\right) &\le& \sum_{u, v, x, y\in [n]}\, d\left(u,v\right)\cdot d\left(x,y\right)\\ &=& \sum_{u, v\in [n]}\, d\left(u,v\right) \cdot \sum_{x, y\in [n]}\, d\left(x,y\right)\\ &=& \left(\sum_{x, y\in [n]}\, d\left(x,y\right) \right)^2. \end{eqnarray*} In summary, \begin{eqnarray*} && \sum_{\text{\rm distinct $i,j=1$}}^{\lfloor n/2\rfloor}\, \mathop{\mathrm E}_{\bs{\pi}}\left[\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \cdot d\left(\bs{\pi}\left(2j-1\right),\bs{\pi}\left(2j\right)\right) \,\right]\\ &\le& \left\lfloor\frac{n}{2}\right\rfloor \left(\left\lfloor\frac{n}{2}\right\rfloor-1\right) \cdot \frac{1}{n \cdot(n-1)\cdot(n-2) \cdot(n-3)} \cdot \left( \sum_{x, y\in [n]}\, d\left(x,y\right) \right)^2. \end{eqnarray*} This and equation~(\ref{averagedistance}) complete the proof. \end{proof} \comment{ Lemmas~\ref{thesmallradiusballislarge}~and~\ref{distinctdistancesproduct} and equation~(\ref{smallerballaveragedefinition}) yield the following. \begin{lemma} \begin{eqnarray*} \sum_{\text{\rm distinct $i,j=1$}}^{|B(z,\delta nr)|/2}\, \mathop{\mathrm E}\left[\, d\left(\pi(2i-1),\pi(2i)\right) \cdot d\left(\pi(2j-1),\pi(2j)\right) \,\right] =\frac{1}{4}\cdot n^2\left(r'\right)^2\left(1\pm o(1)\right). \end{eqnarray*} \end{lemma} Define $$ \Delta\equiv\max_{x,y\in[n]}\,d(x,y) $$ to be the diameter of $([n],d)$. \begin{lemma}\label{thehardestparttoboundlemma} If \begin{eqnarray} \delta nr \ge \Delta \label{shallhaveanequationnumberfortheballequalsuniversething} \end{eqnarray} then \begin{eqnarray} \sum_{i=1}^{\lfloor n/2\rfloor}\, \mathop{\mathrm E}_{\bs{\pi}}\left[\, d^2\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \,\right] \le \left(\frac{1}{2}+ O\left(\frac{1}{n}\right) \right)\left(\delta n^2r\bar{r}+\delta^2 n r^2\right). \label{boundforthedistancesquaressummed} \end{eqnarray} \end{lemma} \begin{proof} Clearly, $\{\bs{\pi}(2i-1),\bs{\pi}(2i)\}$ is a uniformly random size-$2$ subset of $[n]$ for each $i\in[\,\lfloor n/2\rfloor\,]$. Therefore, \begin{eqnarray} \sum_{i=1}^{\lfloor n/2\rfloor}\, \mathop{\mathrm E}_{\bs{\pi}}\left[\, d^2\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \,\right] &=&\sum_{i=1}^{\lfloor n/2\rfloor}\, \frac{1}{n\cdot(n-1)} \cdot \sum_{\text{distinct $x$, $y\in [n]$}}\,d^2\left(x,y\right) \,\,\,\,\,\,\, \label{startofobjective}\\ &\le&\sum_{i=1}^{\lfloor n/2\rfloor}\, \frac{1}{n\cdot(n-1)} \cdot \sum_{x, y\in [n]}\,d^2\left(x,y\right) \nonumber\\ &=&\left\lfloor\frac{n}{2}\right\rfloor\cdot \frac{1}{n\cdot(n-1)} \cdot \sum_{x, y\in [n]}\,d^2\left(x,y\right). \nonumber \end{eqnarray} By inequality~(\ref{shallhaveanequationnumberfortheballequalsuniversething}), \begin{eqnarray} d\left(x,y\right) \le \delta nr \label{constraint} \end{eqnarray} for all $x$, $y\in [n]$. By equations~(\ref{averagedistance})~and~(\ref{startofobjective})--(\ref{constraint}), the left-hand side of inequality~(\ref{boundforthedistancesquaressummed}) cannot exceed the optimal value of the following problem, called {\sc max square sum}:\\ \begin{quote} Find $d_{x,y}\in \mathbb{R}$ for all $x$, $y\in [n]$ to maximize \begin{eqnarray} \left\lfloor\frac{n}{2}\right\rfloor\cdot \frac{1}{n\cdot(n-1)} \cdot \sum_{x, y\in [n]}\,d_{x,y}^2 \label{objectiveofoptimization} \end{eqnarray} subject to \begin{eqnarray} \frac{1}{n^2}\cdot \sum_{x,y\in [n]}\, d_{x,y} = \bar{r},\label{averagedistanceconstraint}\\ \forall x, y\in [n],\,\, 0\le d_{x,y} \le \delta nr.\label{largestdistanceconstraint} \end{eqnarray} \end{quote} Above, constraint~(\ref{averagedistanceconstraint}) (resp., (\ref{largestdistanceconstraint})) mimics equation~(\ref{averagedistance}) (resp., inequality~(\ref{constraint}) and the non-negativeness of distances). Appendix~\ref{analyzingthemaximizationproblem} bounds the optimal value of {\sc max square sum} from above by \begin{eqnarray} \left\lfloor\frac{n}{2}\right\rfloor \frac{1}{n\cdot(n-1)} \cdot \left(\left\lfloor\frac{n \bar{r}}{\delta r}\right\rfloor+1\right) \cdot \left(\delta nr\right)^2. \nonumber \end{eqnarray} This evaluates to be at most $$\left(\frac{1}{2}+O\left(\frac{1}{n}\right)\right) \left(\delta n^2 r\bar{r}+\delta^2 n r^2\right).$$ \comment{ Finally, $$\left(1\pm o(1)\right)\delta n^2 rr'\le 2\left(1+o(1)\right)\delta n^2 r^2$$ by Lemma~\ref{inneraverageandoverallaverage}. \end{proof} Recall that the variance of any random variable $X$ equals $\mathop{\mathrm E}[X^2]-(\mathop{\mathrm E}[X])^2$. \begin{lemma}\label{boundonthevarianceofthelengthofthematching} If $\delta nr\ge \Delta$, then $$ \mathop{\mathrm{var}}_{\bs{\pi}}\left( \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \right) \le \left(1+ o(1) \right)\delta n^2 r^2. $$ \end{lemma} \begin{proof} By equations~(\ref{startofthemeanofsquare})--(\ref{endofthemeanofsquare}) and Lemmas~\ref{distinctdistancesproduct}--\ref{thehardestparttoboundlemma}, \begin{eqnarray*} &&\mathop{\mathrm E}_{\bs{\pi}}\left[\, \left(\sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right)\right)^2 \,\right]\\ &\le& \frac{1}{4}\cdot \left(1 + O\left(\frac{1}{n}\right) \right) n^2 \bar{r}^2 +\left( \frac{1}{2} + O\left(\frac{1}{n}\right) \right)\left(\delta n^2 r\bar{r}+\delta^2 nr^2\right). \end{eqnarray*} This and Lemma~\ref{expectedmatchingsize} imply \begin{eqnarray*} &&\mathop{\mathrm{var}}_{\bs{\pi}}\left( \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \right)\\ &\le& O\left(\frac{1}{n}\right) \cdot n^2 \bar{r}^2 +\left( \frac{1}{2} + O\left(\frac{1}{n}\right) \right)\left(\delta n^2 r\bar{r}+\delta^2 nr^2\right). \end{eqnarray*} Finally, invoke Lemma~\ref{inneraverageandoverallaverage} and recall that $\delta=\omega(1/\sqrt{n})$. \end{proof} \begin{lemma}\label{randommatchingconcentrationlemma} If $\delta nr \ge \Delta$, then {\small \begin{eqnarray*} \Pr_{\bs{\pi}}\left[ \, \left| \, \left( \sum_{i=1}^{\lfloor n/2\rfloor} d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \right) - \frac{1}{2}\cdot\left(1\pm O\left(\frac{1}{n}\right) \right)n\bar{r} \, \right| \ge k\sqrt{\left(1+ o(1) \right)\delta}\, nr \, \right] \le \frac{1}{k^2} \end{eqnarray*} for all $k>1$. \end{lemma} \begin{proof} Use Chebyshev's inequality and Lemmas~\ref{expectedmatchingsize}~and~\ref{boundonthevarianceofthelengthofthematching}. \end{proof} \comment{ \section{Estimating the average distance on a graph} This section presents an efficient estimation of \begin{eqnarray} \bar{r}\equiv \frac{1}{n^2}\cdot \sum_{x,y\in[n]}\,d\left(x,y\right) \label{averagedistanceingeneral} \end{eqnarray} when $d$ is a graph metric. As in Sec.~\ref{approximationratiosection}, take $\epsilon=1/n^{o(1)}$ less than a small constant and $\delta$ as in line~1 of {\sf Las Vegas median}. Note that the proofs of Lemmas~\ref{pointsoutofballarebad}--\ref{innerballmediantotaldistance} are independent from the choice of $z$ in line~3 of {\sf Las Vegas median}. In particular, they remain to hold when {\sf pick point} in line~3 returns a uniformly random point in $[n]$. \comment{ \begin{lemma}\label{averagedistancecannotbetoolargewhp} $$\Pr_{\bs{u}}\left[\,r\ge 9\bar{r}\,\right]\le\frac{1}{9}.$$ \end{lemma} \begin{proof} By lines~2--3 of {\sf average distance}, $$ \mathop{\mathrm{E}}_{\bs{u}}\left[\,r\,\right] =\frac{1}{n}\cdot \sum_{u\in[n]}\,\frac{1}{n}\cdot \sum_{x\in[n]}\, d\left(u,x\right) \stackrel{\text{(\ref{averagedistance})}}{=} \bar{r}. $$ Now use Markov's inequality. \end{proof} \begin{lemma}\label{thekeyconcentration} If $\delta nr\ge \Delta$, then {\footnotesize $$ \Pr_{\bs{\pi}}\left[\, \left|\, \left( \sum_{i=1}^{\lfloor n/2\rfloor}\, d\left(\bs{\pi}\left(2i-1\right),\bs{\pi}\left(2i\right)\right) \right) - \frac{1}{2}\cdot\left(1\pm O\left(\frac{1}{n}\right) \right)n\bar{r} \,\right| \ge k\sqrt{\left(1+o(1) \right)\delta}\, n\bar{r} \,\right] \le\frac{1}{k^2} $$ for all $k>1$. \end{lemma} \begin{proof} By inequalities~(\ref{bestisnoworsethanaverage})~and~(\ref{theaveragedistancefromanoptimalpoint}), $$r\le\bar{r}.$$ This and Lemma~\ref{randommatchingconcentrationlemma} complete the proof. \end{proof} We now arrive at an efficient estimation of the average distance on a graph. \begin{theorem}\label{theoremonestimatinggraphaveragedistance} Given $n\in\mathbb{Z}^+$, $\epsilon=\omega(1/n^{1/4})$ and oracle access to a graph metric $d\colon[n]\times[n]\to\mathbb{N}$, a real number in $[\,(1-\epsilon)\bar{r},(1+\epsilon)\bar{r}\,]$ can be found in $O(n)$ time with probability $1/2+\Omega(1)$. \end{theorem} \begin{proof} Let $G=([n],E)$ be an undirected unweighted graph inducing the distance function $d$. Then pick $x$, $y\in [n]$ with $d(x,y)=\Delta$, i.e., $(x,y)$ is a furthest pair of vertices of $G$. Find a simple shortest $x$-$y$ path, denoted $(v_0=x,v_1,\ldots,v_\Delta=y)$, in $G$. By equation~(\ref{theaveragedistancefromanoptimalpoint}), \begin{eqnarray} r \ge \frac{1}{n}\cdot\sum_{i=0}^\Delta\,d\left(p^*,v_i\right). \label{totaldistancetoverticesonalongestpath} \end{eqnarray} Now, {\small \begin{eqnarray} \sum_{i=0}^\Delta\,d\left(p^*,v_i\right) = \frac{1}{2}\cdot \sum_{i=0}^\Delta d\left(p^*,v_i\right)+d\left(p^*,v_{\Delta-i}\right) \ge\frac{1}{2}\cdot\sum_{i=0}^\Delta\, d\left(v_i,v_{\Delta-i}\right) =\frac{1}{2}\cdot \sum_{i=0}^\Delta\, \left|\,\Delta-2i\,\right| \ge \frac{\Delta^2}{4}, \label{totaldistancestopointsonapath} \end{eqnarray} where the first inequality (resp., the second equality) follows from the triangle inequality (resp., $(v_0,v_1,\ldots,v_\Delta)$ being a shortest $v_0$-$v_\Delta$ path).\footnote{It is easy to verify that $\sum_{i=0}^\Delta\,|\,\Delta-2i\,| =(\Delta+2)\Delta/2$ if $\Delta\equiv 0\pmod{2}$ and $\sum_{i=0}^\Delta\,|\,\Delta-2i\,| =(\Delta+1)^2/2$ otherwise.} By inequalities~(\ref{totaldistancetoverticesonalongestpath})--(\ref{totaldistancestopointsonapath}), \begin{eqnarray} nr \ge\frac{\Delta^2}{4}. \label{totaldistancetoverticesonalongpath} \end{eqnarray} Because $d$ is a graph metric, $d(x,y)\ge1$ for all distinct $x$, $y\in[n]$. So by equation~(\ref{theaveragedistancefromanoptimalpoint}), \begin{eqnarray} r \ge \frac{1}{n}\cdot \sum_{x\in[n]\setminus\{p^*\}}\,1\ge\frac{1}{2} \label{trivialaveragedistancelowerbound} \end{eqnarray} for all $n\ge2$. By inequalities~(\ref{totaldistancetoverticesonalongpath})--(\ref{trivialaveragedistancelowerbound}), $$ \delta nr \ge \delta \cdot\max\left\{\frac{\Delta^2}{4},\frac{n}{2}\right\}. $$ So \begin{eqnarray} \delta nr \ge \Delta \label{thesmallballwillbetheuniverse} \end{eqnarray} for all sufficiently large $n$.\footnote{If $\Delta\ge 4/\delta$, then $\delta\Delta^2/4\ge \Delta$. Otherwise, $\delta n/2\ge \Delta$ for all $n>8/\delta^2$. Finally, recall that $\delta=\omega(1/\sqrt{n})$.} By equation~(\ref{thedeltavalue}), \begin{eqnarray} 3\sqrt{\left(1+o(1)\right)\delta}\le 0.1\,\epsilon \label{theerrortermcalculated} \end{eqnarray} for all sufficiently large $n$. By inequalities~(\ref{thesmallballwillbetheuniverse})--(\ref{theerrortermcalculated}), Lemma~\ref{thekeyconcentration} with $k=3$ and recalling that $\epsilon=\omega(1/n^{1/4})$, \begin{eqnarray} \Pr_{\bs{\pi}}\left[\, \sum_{i=1}^{\lfloor n/2\rfloor}\,d\left(\bs{\pi}(2i-1),\bs{\pi}(2i)\right) \in \left[\,\left(\frac{1}{2}-\frac{\epsilon}{2}\right)n\bar{r}, \left(\frac{1}{2}+\frac{\epsilon}{2}\right)n\bar{r} \,\right] \,\right]\ge 1-\frac{1}{9} \label{concentrationofmatchingconcluded} \end{eqnarray} for all sufficiently large $n$. Consequently, the output of line~2 of {\sf average distance} in Fig.~\ref{mainalgorithm} is in $[\,(1-\epsilon)\bar{r},(1+\epsilon)\bar{r}\,]$ with probability $1/2+\Omega(1)$. Line~1 takes $O(n)$ time by the Knuth shuffle. Clearly, line~2 also takes $O(n)$ time. \end{proof} The time complexity of $O(n)$ in Theorem~\ref{theoremonestimatinggraphaveragedistance} is independent of $\epsilon$. But for general metrics, we do not know whether the time complexity of $O(n/\epsilon^2)$ in Fact~\ref{averagedistancepriorresult} can be improved to $O(n/\epsilon^{2-\Omega(1)})$. \comment{ Pick $\tilde{z}$ from $[n]$ uniformly at random. Analogous to line~4 of {\sf Las Vegas median}, define \begin{eqnarray} \tilde{r}\equiv \frac{1}{n}\cdot \sum_{x\in[n]}\,d\left(\tilde{z},x\right). \tilde{r}\equiv \end{eqnarray} Observe that the proofs of Lemmas~\ref{pointsoutofballarebad}--\ref{randommatchingconcentrationlemma} are independent from the picking of $z$. So Lemma~\ref{randommatchingconcentrationlemma} remains to hold with $z$ and $r$ replaced by $\tilde{z}$ and $\tilde{r}$, respectively. That is, \comment{ \section{Estimating the sum of distances} Using our results in Sec.~\ref{expectedtimesection}, we present a Monte-Carlo algorithm, called {\sf sum distances} in Fig.~\ref{averagedistancealgorithm}, for estimating $\sum_{u,v\in[n]}\,d(u,v)$. \begin{figure} \begin{algorithmic}[1] \STATE $\tilde{\delta}\leftarrow \epsilon^2/10000$; \STATE $\tilde{z}\leftarrow\text{\sf Indyk median}^d(n,\epsilon)$; \STATE $\tilde{r}\leftarrow \sum_{x\in[n]}\,d(\tilde{z},x)/n$; \STATE Pick a uniformly random bijection $\tilde{\pi}\colon [\,|B(\tilde{z},\tilde{\delta} n\tilde{r})|\,] \to B(\tilde{z},\tilde{\delta} n\tilde{r})$; \STATE $S\leftarrow (2\,|B(\tilde{z},\tilde{\delta} n\tilde{r})|^2/n) \cdot \sum_{i=1}^{\lfloor |B(\tilde{z},\tilde{\delta} n \tilde{r})|/2\rfloor}\, d(\tilde{\pi}(2i-1),\tilde{\pi}(2i))$; \STATE $T\leftarrow \sum_{\text{$u$, $v\in[n]$ s.t.\ $\{u,v\}\not\subseteq B(\tilde{z},\tilde{\delta} n\tilde{r})$}}\, d(u,v)$; \RETURN $S+T$; \end{algorithmic} \caption{Algorithm {\sf sum distances} with oracle access to a metric $d\colon [n]\times[n] \to[\,0,\infty\,)$ and with inputs $n\in\mathbb{Z}^+$, and $\epsilon\in(0,1)$} \label{averagedistancealgorithm} \end{figure} \begin{lemma}\label{generalformofestimationerror} In line~7 of {\sf sum distances}, $$ S+T-\sum_{u,v\in[n]}\,d\left(u,v\right) = \left( \frac{2\,|B(\tilde{z},\tilde{\delta} n\tilde{r})|^2}{n} \cdot \sum_{i=1}^{\lfloor |B(\tilde{z},\tilde{\delta} n \tilde{r})|/2\rfloor}\, d\left(\tilde{\pi}\left(2i-1\right),\tilde{\pi}\left(2i\right)\right) \right) -\sum_{u,v\in B(\tilde{z},\tilde{\delta}n\tilde{r})}\,d\left(u,v\right). $$ \end{lemma} \begin{proof} We have \begin{eqnarray*} \sum_{u,v\in[n]}\,d\left(u,v\right) &=&\sum_{u,v\in B(\tilde{z},\tilde{\delta}n\tilde{r})}\,d\left(u,v\right) +\sum_{\text{$u$, $v\in[n]$ s.t.\ $\{u,v\}\not\subseteq B(\tilde{z},\tilde{\delta} n\tilde{r})$}}\, d\left(u,v\right)\\ &=&\left(\sum_{u,v\in B(\tilde{z},\tilde{\delta}n\tilde{r})}\,d\left(u,v\right)\right) +T, \end{eqnarray*} where the last equality follows from line~6 of {\sf sum distances}. Now use line~5. \end{proof} Similarly to equation~(\ref{smallerballaveragedefinition}), define \begin{eqnarray} \tilde{r}' \equiv \frac{1}{|B(\tilde{z},\tilde{\delta}n\tilde{r})|^2} \cdot \sum_{u,v\in B(\tilde{z},\tilde{\delta}n\tilde{r})}\,d\left(u,v\right). \label{averagedistanceinthenewball} \end{eqnarray} Observe that the proofs of Lemmas~\ref{thesmallradiusballislarge}--\ref{randommatchingconcentrationlemma} use only the following facts: \begin{enumerate}[(i)] \item\label{originalcondition1} $$r = \frac{1}{n}\cdot \sum_{x\in[n]}\,d\left(z,x\right).$$ \item $$r' = \frac{1}{|B(z,\delta nr)|^2}\cdot \sum_{u,v\in B(z,\delta nr)}\,d\left(u,v\right).$$ \item\label{originalcondition3} $\pi\colon [\,|B(z,\delta nr)|\,]\to B(z,\delta nr)$ is a uniformly random bijection. \end{enumerate} In particular, they do not rely on the choices of $z\in[n]$ and $\delta>0$. By lines~3--4 of {\sf sum distances} and equation~(\ref{averagedistanceinthenewball}), conditions~(\ref{originalcondition1})--(\ref{originalcondition3}) hold with $z$, $r$, $r'$, $\delta$ and $\pi$ replaced by $\tilde{z}$, $\tilde{r}$, $\tilde{r}'$, $\tilde{\delta}$ and $\tilde{\pi}$, respectively. Therefore, Lemmas~\ref{thesmallradiusballislarge}--\ref{randommatchingconcentrationlemma} remain true with $z$, $r$, $r'$, $\delta$ and $\pi$ replaced by $\tilde{z}$, $\tilde{r}$, $\tilde{r}'$, $\tilde{\delta}$ and $\tilde{\pi}$, respectively. So we have the following analogies to Lemmas~\ref{thesmallradiusballislarge},~\ref{inneraverageandoverallaverage}~and~\ref{randommatchingconcentrationlemma}. \begin{lemma}\label{thesmallradiusballislargeanalogy} $$ \left|\, [n]\setminus B\left(\tilde{z},\tilde{\delta}n\tilde{r}\right)\,\right|\le\frac{1}{\tilde{\delta}} $$ and, therefore, $$ \left|\,B\left(\tilde{z},\tilde{\delta}n\tilde{r}\right)\,\right|\ge n-\frac{1}{\tilde{\delta}} =\left(1-o(1)\right)n. $$ \end{lemma} \comment{ \begin{proof} Clearly, $$ \sum_{x\in [n]}\,d\left(\tilde{z},x\right) \ge \sum_{x\in [n]\setminus B(\tilde{z},\tilde{\delta} n\tilde{r})}\,d\left(\tilde{z},x\right) \ge \sum_{x\in [n]\setminus B(\tilde{z},\tilde{\delta} n\tilde{r})}\,\tilde{\delta} n\tilde{r} =\left|\,[n]\setminus B\left(\tilde{z},\tilde{\delta} n\tilde{r}\right)\,\right|\cdot \tilde{\delta} n\tilde{r}. $$ Then use line~3 of {\sf sum distances}. \end{proof} \begin{lemma}\label{inneraverageandoverallaveragealternative} $\tilde{r}'\leq 2\tilde{r}$. \end{lemma} \comment{ \begin{proof} By equation~(\ref{averagedistanceinthenewball}) and the triangle inequality, \begin{eqnarray} \tilde{r}' &\le& \frac{1}{|B(\tilde{z},\tilde{\delta} n\tilde{r})|^2} \cdot \sum_{u, v\in B(\tilde{z},\tilde{\delta} n\tilde{r})}\, \left(d\left(\tilde{z},u\right)+d\left(\tilde{z},v\right)\right) \label{frominnerdistancetowholedistancenew}\\ &=& \frac{1}{|B(\tilde{z},\tilde{\delta} n\tilde{r})|^2} \cdot \left|B(\tilde{z},\tilde{\delta} n\tilde{r})\right|\cdot\left( \sum_{u\in B(\tilde{z},\tilde{\delta} n\tilde{r})}\, d\left(\tilde{z},u\right) +\sum_{v\in B(\tilde{z},\tilde{\delta} n\tilde{r})}\, d\left(\tilde{z},v\right) \right)\nonumber\\ &=& \frac{2}{|B(\tilde{z},\tilde{\delta} n\tilde{r})|} \cdot \sum_{u\in B(\tilde{z},\tilde{\delta} n\tilde{r})}\, d\left(\tilde{z},u\right).\nonumber \end{eqnarray} Obviously, the average distance from $\tilde{z}$ to the points in $B(\tilde{z},\tilde{\delta} n\tilde{r})$ is at most that from $\tilde{z}$ to all points, i.e., \begin{eqnarray} \frac{1}{|B(\tilde{z},\tilde{\delta} n\tilde{r})|} \cdot \sum_{u\in B(\tilde{z},\tilde{\delta} n\tilde{r})}\, d\left(\tilde{z},u\right) \le \frac{1}{n}\cdot \sum_{u\in [n]}\, d\left(\tilde{z},u\right). \label{frominnerdistancetowholedistance2new} \end{eqnarray} Inequalities~(\ref{frominnerdistancetowholedistancenew})--(\ref{frominnerdistancetowholedistance2new}) and line~3 of {\sf sum distances} complete the proof. \end{proof} \begin{lemma}\label{randommatchingconcentrationlemmaanalogy} For all $k>1$, $$ \Pr\left[\, \left|\, \left( \sum_{i=1}^{\lfloor|B(\tilde{z},\tilde{\delta} n\tilde{r})|/2\rfloor}\, d\left(\tilde{\pi}\left(2i-1\right),\tilde{\pi}\left(2i\right)\right) \right) - \frac{1}{2}\cdot\left(1\pm o(1) \right)n \tilde{r}' \,\right| \ge k\sqrt{2\left(1+o(1) \right)\tilde{\delta}}\, n\tilde{r} \,\right] \le\frac{1}{k^2}, $$ where the probability is taken over $\tilde{\pi}$. \end{lemma} By Lemma~\ref{randommatchingconcentrationlemmaanalogy}, {\small $$ \Pr\left[\, \left|\, \left( \sum_{i=1}^{\lfloor|B(\tilde{z},\tilde{\delta} n\tilde{r})|/2\rfloor}\, d\left(\tilde{\pi}\left(2i-1\right),\tilde{\pi}\left(2i\right)\right) \right) - \frac{1}{2}\cdot n \tilde{r}' \,\right| \ge k\sqrt{2\left(1+o(1) \right)\tilde{\delta}}\, n\tilde{r} +\frac{1}{2}\cdot o(1) n \tilde{r}' \,\right] \le\frac{1}{k^2} $$ for all $k>1$.\footnote{It is easy to verify that $$\Pr\left[\,\left|X-a\right|\ge c+|b|\,\right] \le\Pr\left[\,\left|X-\left(a+b\right)\right|\ge c\,\right]$$ for all $a$, $b$, $c\in\mathbb{R}$ and for each random variable $X$. Then take $X=\sum_{i=1}^{\lfloor|B(\tilde{z},\tilde{\delta}n\tilde{r})|/2\rfloor}\, d(\tilde{\pi}(2i-1),\tilde{\pi}(2i))$, $a=(1/2)\cdot n\tilde{r}'$, $b=\pm (1/2)o(1)n\tilde{r}'$ (as within $\Pr[\cdot]$ in Lemma~\ref{randommatchingconcentrationlemmaanalogy}) and $c=k\sqrt{2(1+o(1))\tilde{\delta}}\,n\tilde{r}$.} This is equivalent to {\small \begin{eqnarray} && \Pr\left[\, \left|\, \left( \frac{2\,|B(\tilde{z},\tilde{\delta}n\tilde{r})|^2}{n} \cdot\sum_{i=1}^{\lfloor|B(\tilde{z},\tilde{\delta} n\tilde{r})|/2\rfloor}\, d\left(\tilde{\pi}\left(2i-1\right),\tilde{\pi}\left(2i\right)\right) \right) - \sum_{u,v\in B(\tilde{z},\tilde{\delta} n\tilde{r})}\,d\left(u,v\right) \,\right|\right.\nonumber\\ && \left. \phantom{\left|\left(\sum_{i=1}^{\lfloor|B(\tilde{z},\tilde{\delta} n\tilde{r})|/2\rfloor}\,\right)\right| \ge \frac{2\,|B(\tilde{z},\tilde{\delta}n\tilde{r})|^2}{n}\cdot \left( k\sqrt{2\left(1+o(1) \right)\tilde{\delta}}\, n\tilde{r} +\frac{1}{2}\cdot o(1) n \tilde{r}' \right) \,\right]\nonumber\\ &\le&\frac{1}{k^2}\label{theestimationerrorforthetotaldistance} \end{eqnarray} by equation~(\ref{averagedistanceinthenewball}), for all $k>1$. Lemma~\ref{generalformofestimationerror} and inequality~(\ref{theestimationerrorforthetotaldistance}) with $k=5$ imply the following. \begin{lemma}\label{theerrorandprobabilitycomplicatedform} {\small $$\Pr\left[\, \left|\, S+T-\sum_{u,v\in[n]}\,d\left(u,v\right) \,\right| \ge \frac{2\,|B(\tilde{z},\tilde{\delta}n\tilde{r})|^2}{n}\cdot \left( 5\sqrt{2\left(1+o(1) \right)\tilde{\delta}}\, n\tilde{r} +\frac{1}{2}\cdot o(1) n \tilde{r}' \right) \,\right] \le\frac{1}{25}.$$ \end{lemma} \begin{lemma}\label{theerrortermasymptotics} For all sufficiently large $n$, {\small $$ \Pr\left[\, \frac{2\,|B(\tilde{z},\tilde{\delta}n\tilde{r})|^2}{n}\cdot \left( 5\sqrt{2\left(1+o(1) \right)\tilde{\delta}}\, n\tilde{r} +\frac{1}{2}\cdot o(1) n \tilde{r}' \right) \le 100\sqrt{\tilde{\delta}}\cdot \sum_{u,v\in[n]}\,d\left(u,v\right) \,\right]\ge1-\frac{1}{e}. $$ \end{lemma} \begin{proof} By Lemmas~\ref{thesmallradiusballislargeanalogy}--\ref{inneraverageandoverallaveragealternative}, \begin{eqnarray} &&\frac{2\,|B(\tilde{z},\tilde{\delta}n\tilde{r})|^2}{n}\cdot \left( 5\sqrt{2\left(1+o(1) \right)\tilde{\delta}}\, n\tilde{r} +\frac{1}{2}\cdot o(1) n \tilde{r}' \right)\nonumber\\ &\le& 2\left(1-o(1)\right) \left(5\sqrt{2\left(1+o(1)\right)\tilde{\delta}}+o(1) \right)n^2\tilde{r}.\label{aquickestimation} \end{eqnarray} By Fact~\ref{Indykfact} and line~2 of {\sf sum distances}, $$ \Pr\left[ \sum_{x\in[n]}\,d\left(\tilde{z},x\right) \le\left(1+\epsilon\right)\cdot\min_{y\in[n]}\,\sum_{x\in[n]}\,d\left(y,x\right) \right]\ge 1-\frac{1}{e}. $$ Equivalently, \begin{eqnarray} \Pr\left[ n\tilde{r} \le\left(1+\epsilon\right)\cdot\min_{y\in[n]}\,\sum_{x\in[n]}\,d\left(y,x\right) \right]\ge 1-\frac{1}{e}\label{theindykresultisprobablygood} \end{eqnarray} by line~3. By the averaging argument, $$ \min_{y\in[n]}\,\sum_{x\in[n]}\,d\left(y,x\right) \le \frac{1}{n}\cdot\sum_{y\in[n]}\,\sum_{x\in[n]}\,d\left(y,x\right). $$ This and inequality~(\ref{theindykresultisprobablygood}) imply \begin{eqnarray} \Pr\left[ n^2\tilde{r} \le\left(1+\epsilon\right)\cdot\sum_{y\in[n]}\,\sum_{x\in[n]}\,d\left(y,x\right) \right]\ge 1-\frac{1}{e}\label{theindykresultisprobablygoodcomparedtotheaverage} \end{eqnarray} Inequalities~(\ref{aquickestimation})~and~(\ref{theindykresultisprobablygoodcomparedtotheaverage}) complete the proof.\footnote{Note that the right-hand side of inequality~(\ref{aquickestimation}) is at most $(100\sqrt{\tilde{\delta}}/(1+\epsilon))\cdot n^2\tilde{r}$ for a small $\epsilon>0$ and all sufficiently large $n$. Also note that $\sum_{y\in[n]}\,\sum_{x\in[n]}\,d(x,y)=\sum_{u,v\in[n]}\,d(u,v)$.} \end{proof} We now show an efficient estimation of $\sum_{u,v\in[n]}\,d(u,v)$. \begin{theorem}\label{theoremonestimationofsumofdistances} Given $n\in\mathbb{Z}^+$, a constant $\epsilon>0$ and oracle access to a metric $d\colon[n]\times[n]\to[\,0,\infty\,)$, $\sum_{u,v\in[n]}\,d(u,v)$ can be estimated to within an additive error of $\epsilon\cdot\sum_{u,v\in[n]}\,d(u,v)$ in $O(n/\epsilon^2)$ time and with an $\Omega(1)$ probability of success. \end{theorem} \begin{proof} By Lemmas~\ref{theerrorandprobabilitycomplicatedform}--\ref{theerrortermasymptotics}, $$ \Pr\left[\, \left|\, S+T-\sum_{u,v\in[n]}\,d\left(u,v\right) \,\right|\le 100\sqrt{\tilde{\delta}}\cdot \sum_{u,v\in[n]}\,d\left(u,v\right) \,\right] \ge 1-\frac{1}{25}-\frac{1}{e}=\Omega(1). $$ So by line~1 of {\sf sum distances}, line~7 estimates $\sum_{u,v\in[n]}\,d(u,v)$ to within an additive error of $\epsilon\cdot\sum_{u,v\in[n]}\,d(u,v)$ with probability $\Omega(1)$. By Fact~\ref{Indykfact}, line~2 of {\sf sum distances} takes $O(n/\epsilon^2)$ time. Line~4 takes time $O(|B(\tilde{z},\tilde{\delta} n\tilde{r})|)=O(n)$ by the Knuth shuffle. Because lines~6 queries for all the distances incident to any point in $[n]\setminus B(\tilde{z},\tilde{\delta} n\tilde{r})$, it takes time $$O\left(\left|\,[n]\setminus B\left(\tilde{z},\tilde{\delta} n\tilde{r}\right)\,\right| \cdot n\right) \stackrel{\text{Lemma~\ref{thesmallradiusballislargeanalogy}}}{=}O\left(\frac{n}{\tilde{\delta}}\right).$$ By line~1, $\tilde{\delta}=\Theta(\epsilon^2)$. The other lines of {\sf sum distances} clearly take $O(n)$ time. \end{proof} Prior to this paper, the best Monte-Carlo algorithm for estimating $\sum_{u,v\in[n]}\,d(u,v)$ to within an additive error of $\epsilon\cdot \sum_{u,v\in[n]}\,d(u,v)$ takes time $O(n/\epsilon^{7/2})$ when the probability of success is set to $\Omega(1)$~\cite{Ind99}. So Theorem~\ref{theoremonestimationofsumofdistances} implies an algorithm with a better running time in terms of $\epsilon$.

https://arxiv.org/abs/1505.01388

Riemann-Liouville Fractional Cosine Functions

In this paper, a new notion, named Riemann-Liouville fractional cosine function is presented. It is proved that a Riemann-Liouville $\alpha$-order fractional cosine function is equivalent to Riemann-Liouville $\alpha$-order fractional resolvents introduced in [Z.D. Mei, J.G. Peng, Y. Zhang, Math. Nachr. 288, No. 7, 784-797 (2015)].

\section{Introduction} Assume that $X$ is a Banach space, $A: D(A)\subset X\rightarrow X$ and $B: D(B)\subset X\rightarrow X$ are closed linear operators. It is well-known that $C_0$-semigroup is an important tool to study the following abstract Cauchy problem of first order \begin{align}\label{firstorder} \left\{ \begin{array}{ll} \frac{du(t)}{dt}=Au(t), & \hbox{} t>0\\ u(0)=x, & \hbox{ } \end{array} \right. \end{align} and cosine function essentially characterizes the abstract Cauchy problem of second order described by \begin{align}\label{secondorder} \left\{ \begin{array}{ll} \frac{d^2u(t)}{dt^2}=Bu(t), & \hbox{} t>0\\ u(0)=x, u'(0)=0. & \hbox{ } \end{array} \right. \end{align} Here a $C_0$-semigroup is a family $\{T(t)\}_{t\geq 0}$ of strongly continuous and bounded linear operators defined on $X$ satisfying $T(0)=I$ and $T(t+s)=T(t)T(s),\ t,s\geq 0$; a cosine function is a family $\{S(t)\}_{t\geq 0}$ of strongly continuous and bounded linear operators defined on $X$ satisfying $S(0)=I$ and $2S(t)S(s)=S(t)+S(s),\ t\geq s\geq 0.$ Concretely, system (\ref{firstorder}) is well-posed if and only if $A$ generates a $C_0$-semigroup $\{T(t)\}_{t\geq 0}$, namely, $Ax=\lim_{t\rightarrow 0^+}t^{-1}(T(t)x-x)$ with domain $D(A)=\{x\in D(A):\lim_{t\rightarrow 0^+}t^{-1}(T(t)x-x)$ exists $\}$; system (\ref{secondorder}) is well-posed if and only if $B$ generates a cosine function $\{S(t)\}_{t\geq 0}$, namely, $Bx=2\lim_{t\rightarrow 0^+}t^{-2}(S(t)x-x)$ with domain $D(B)=\{x\in D(B):\lim_{t\rightarrow 0^+}t^{-2}(S(t)x-x)$ exists $\}$. Therefore, pure algebraic methods can be used to study abstract Cauchy problems of first and second orders. For details, we refer to \cite{Engel2000,Goldstein1985}. However, equations of integer order such as (\ref{firstorder}) and (\ref{secondorder}) can't exactly describe the behavior of many physical systems; fractional differential equations maybe more suitable for describing anomalous diffusion on fractals (physical objects of fractional dimension, like some amorphous semiconductors or strongly porous materials; see \cite{Anh2001,Metzler2001} and the references therein), fractional random walk \cite{Germano2009,Scalas2006}, etc. Fractional derivatives appear in the theory of fractional differential equations; they describe the property of memory and heredity of materials, and it is the major advantage of fractional derivatives compared with integer order derivatives. Let $\alpha>0$ and $m=[\alpha]$, The smallest integer larger than or equal to $\alpha$. There are mainly two types of $\alpha$-order fractional differential equations, which are most used in the real problems. 1) Caupto fractional abstract Cauchy problem \begin{align}\label{caputo} \left\{ \begin{array}{ll} ^CD_t^\alpha u(t)=Au(t), & \hbox{ } t>0,\\ u(0)=x, u^{(k)}(0)=0,k=1,2, ..., m-1. \end{array} \right. \end{align} where $^CD_t^\alpha$ is the Caupto fractional differential operator defined as follows: \begin{align*} ^CD_t^\alpha u(t)=\frac{1}{\Gamma(m-\alpha)}\int_0^t(t-\sigma)^{-\alpha}u^{(m)}(\sigma)d\sigma; \end{align*} 2) Riemann-Liouville fractional abstract Cauchy problem \begin{align}\label{R-L} \left\{ \begin{array}{ll} D_t^\alpha u(t)=Au(t), & \hbox{ } \\ (g_{2-\alpha}*u)(0)=\lim_{s\rightarrow 0^+}\int_0^s\frac{(s-\sigma)^{m-1-\alpha}}{\Gamma(2-\alpha)}u(\sigma)d\sigma= x, & \hbox{ } \\ (g_{2-\alpha}*u)^{(k)}(0)=\lim_{s\rightarrow 0^+}\int_0^s\frac{d^k}{dt^k}\frac{(s-\sigma)^{m-1-\alpha}}{\Gamma(m-\alpha)}u(\sigma)d\sigma =0, k=1,2, ..., m-1.& \hbox{ } \end{array} \right. \end{align} where $D_t^\alpha$ is the Riemann-Liouville fractional differential operator defined by \begin{align*} D_t^\alpha u(t)=\frac{1}{\Gamma(m-\alpha)}\frac{d}{dt}\int_0^t(t-\sigma)^{m-1-\alpha}u(\sigma)d\sigma. \end{align*} Obviously, (\ref{firstorder}) is just the limit state of equations (\ref{caputo}) and (\ref{R-L}) as $\alpha\rightarrow 1$, and (\ref{secondorder}) is just the limit state of equations (\ref{caputo}) and (\ref{R-L}) as $\alpha\rightarrow 2$. Initial conditions for the Caputo fractional derivatives are expressed in terms of initials of integer order derivatives \cite{Eidelman2004,Mei2014,Peng2012}. For some real materials, initial conditions should be expressed in terms of Riemann-Liouville fractional derivatives, and it is possible to obtain initial values for such initial conditions by appropriate measurements \cite{Heymans2006, Hilfer2000}. In order to study Caputo fractional abstract Cauchy problem (\ref{caputo}), Bajlekova \cite{Bazhlekova2001} introduced the important notion of solution operator for equations (\ref{caputo}) as follows. \begin{definition}\label{solutiona} A family $\{T(t)\}_{t\geq 0}$ of bounded linear operators of $X$ is called a solution operator for (\ref{caputo}) if the following three conditions are satisfied: (a) $T(t)$ is strongly continuous for $t\geq 0$ and $T(0)=I$, (b) $T(t)D(A)\subset D(A)$ and $AT(t)x=T(t)Ax$ for all $x\in D(A)$ and $t\geq 0$, (c) for any $x\in D(A)$, there holds \begin{align*} T(t)x=x+J_t^\alpha T(t)Ax, t\geq 0. \end{align*} Here the notation $J_t^\alpha f(t)$ is defined by \begin{align*} J_t^\alpha f(t)=\frac{1}{\Gamma(\alpha)}\int_0^t(t-\sigma)^{\alpha-1}f(t)dt. \end{align*} \end{definition} Chen and Li developed in \cite{Chen2010} a notion of $\alpha$-resolvent operator function, which was proved to be a new characteristic of solution operator. Hence, Caputo fractional abstract Cauchy problem can be studied by pure algebraic methods. The definition of $\alpha$-resolvent operator function is as follows. \begin{definition}\label{lichen} Let $\{S(t)\}_{t\geq 0}$ be a family of bounded linear operators on $X$. Then $\{S(t)\}_{t\geq 0}$ is called to be an $\alpha$-resolvent operator function, if the following assumptions are satisfied: 1) $S(t)$ is strongly continuous and $S(0)=I.$ 2) $S(s)S(t)=S(t)S(s)$ for all $t,s\geq 0$. 3) $S(s)J_t^\alpha S(t)-J_s^\alpha S(s)S(t)=J_t^\alpha S(t)-J_s^\alpha S(s)$ for all $t,s\geq 0$. \end{definition} In \cite{Li2012}, Li and Peng proposed the following notion of fractional resolvent to study Riemann-Liouville $\alpha$-order fractional abstract Cauchy problem (\ref{R-L}) with $\alpha\in (0,1)$. \begin{definition}\cite{Li2012}\label{de} Let $0<\alpha< 1$. A family $\{T(t)\}_{t>0}$ of bounded linear operators on Banach space $X$ is called an $\alpha$-order fractional resolvent if it satisfies the following assumptions: (P1) for any $x\in X$, $\ T(\cdot)x\in C((0,\infty),X)$, and \begin{equation}\label{clear} \lim_{t\rightarrow 0+}\Gamma(\alpha)t^{1-\alpha}T(t)x=x \ \ \mbox{for all } \ x\in X; \end{equation} (P2) $T(s)T(t)=T(t)T(s) \ \ \mbox{for all } t,s>0;$ (P3) for all $t,s>0$, there holds \begin{equation}\label{fi} \ T(t) J_{s}^{\alpha}T(s)- J_{t}^{\alpha}T(t)T(s)=\frac{t^{\alpha-1}}{\Gamma(\alpha)}J_{s}^{\alpha}T(s) -\frac{s^{\alpha-1}}{\Gamma(\alpha)}J_{t}^{\alpha}T(t). \end{equation} \end{definition} Recently, we studied in \cite{Mei2015} Riemann-Liouville $\alpha$-order fractional Cauchy problem (\ref{R-L}) with order $\alpha\in (1,2)$ through the study of Riemann-Liouville $\alpha$-order fractional resolvent defined as follows: \begin{definition}\label{de} A family $\{T(t)\}_{t> 0}$ of bounded linear operators is called Riemann-Liouviille $\alpha$-order fractional resolvent if it satisfies the following assumptions: (a) For any $x\in X$, $\ T_{\alpha}(\cdot)x\in C((0,\infty),X)$, and \begin{equation}\label{clear} \lim_{t\rightarrow 0^+}\Gamma(\alpha-1)t^{2-\alpha}T(t)x=x \ \ \mbox{for all } \ x\in X; \end{equation} (b) $ \ T(s)T_{\alpha}(t)=T(t)T_{\alpha}(s) \ \ \mbox{for all } t,s>0;$ (c) for all $t,s>0$, there holds \begin{equation}\label{fi} \ T(s) J_{t}^{\alpha}T(t)- J_{s}^{\alpha}T(s)T(t)=\frac{s^{\alpha-2}}{\Gamma(\alpha-1)}J_{t}^{\alpha}T(t) -\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}J_{s}^{\alpha}T(s). \end{equation} \end{definition} The linear operator $A$ defined by \begin{align*} D(A)=\{x\in X: \lim_{t\rightarrow 0^+}\frac{t^{1-\alpha}T(t)x-\frac{1}{\Gamma(\alpha)}x}{t^{2\alpha}}\ \mbox {exists}\} \end{align*} and \begin{align*} Ax=\lim_{t\rightarrow 0^+}\frac{t^{1-\alpha}T(t)x-\frac{1}{\Gamma(\alpha)}x}{t^{2\alpha}} \ \ \mbox {for}\ x \in D(A) \end{align*} is the generator of the Riemann-Liouville $\alpha$-order fractional resolvent $\{T(t)\}_{t>0}$ in Definition \ref{de} with $D(A)$ the domain of $A$. Moreover, we proved that $\{T(t)\}_{t>0}$ is a Riemann-Liouville $\alpha$-order fractional resolvent if and only if it is a solution operator defined as follows: \begin{definition}\label{solutiona} A family $\{T(t)\}_{t>0}$ of bounded linear operators of $X$ is called a solution operator for (\ref{R-L}) if the following three conditions are satisfied: (a) $T(t)$ is strongly continuous for $t> 0$ and $\lim_{t\rightarrow 0^+}\Gamma(\alpha-1)t^{2-\alpha}T(t)x=x,\ x\in X$, (b) $T(t)D(A)\subset D(A)$ and $AT(t)x=T(t)Ax$ for all $x\in D(A)$ and $t> 0$, (c) for any $x\in D(A)$, there holds \begin{align*} T(t)x=\frac{t^{1-\alpha}}{\Gamma(2-\alpha)}x+J_t^\alpha T(t)Ax, t> 0. \end{align*} \end{definition} However, the above functional equations for fractional differential equations are not expressed in terms of the sum of time variables: $s+t$. This is very important in concrete applications of the functional equation, just like $C_0$-semigroups, cosine functions. Motivated by this, Peng and Li \cite{Peng2012} established the characteristic of $\alpha$-order fractional semigroup with $\alpha\in (0,1)$: \begin{align*} \nonumber &\int_0^{t+s}\frac{T(\tau)}{(t+s-\tau)^\alpha}d\tau- \int_0^t\frac{T(\tau)}{(t+s-\tau)^\alpha}d\tau- \int_0^s\frac{T(\tau)}{(t+s-\tau)^\alpha}d\tau\\ =&\alpha\int_0^t\int_0^s\frac{T(r_1)T(r_2)}{(t+s-r_1-r_2)^{1+\alpha}}dr_1dr_2,t,s\geq 0, \end{align*} where the integrals are in the sense of strong operator topology. Concretely, they proved that $\alpha$-order fractional semigroup is closely related to the solution operator of Caputo fractional abstract Cauchy problem (\ref{caputo}). Mei, Peng and Zhang \cite{Mei2013} developed the notion of Riemann-Liouville fractional semigroup as follows: \begin{definition} We call a family $\{T(t)\}_{t>0}$ of bounded linear operators to be a Riemann-Liouville $\alpha$-order fractional semigroup on Banach space $X$, if the following conditions are satisfied: i) for any $x\in X$, $t\mapsto T(t)x$ is continuous over $(0,\infty)$ and \begin{equation}\label{clear} \lim_{t\rightarrow 0+}\Gamma(\alpha)t^{1-\alpha}T(t)x=x; \end{equation} ii) for all $t,s> 0$, there holds \begin{align}\label{cosin} \Gamma(1-\alpha)T(t+s)=\alpha\int_0^t\int_0^s \frac{T(r_1)T(r_2)}{(t+s-r_1-r_2)^{1+\alpha}}dr_1dr_2, \end{align} where the integrals are in the sense of strong operator topology. \end{definition} It is proved in \cite{Mei2013} that $A$ generates a Rimann-Liouville fractional semigroup if and only if it generates a fractional resolvent developed in \cite{Li2010}. In order to study Caputo fractional Cauchy problem of order $\alpha\in (1,2)$, we recently studied in \cite{Mei2014} the notion of fractional cosine function as follows \begin{definition} A family $\{T(t)\}_{t\geq 0}$ of bounded and strongly continuous operators is called an $\alpha$-fractional cosine function if $T(0)=I$ and there hold \begin{align}\label{cosin} \nonumber&\int_0^{t+s}\int_0^\sigma\frac{T(\tau)}{(t+s-\sigma)^{\alpha-1}}d\tau d\sigma-\int_0^t\int_0^\sigma\frac{T(\tau)}{(t+s-\sigma)^{\alpha-1}}d\tau d\sigma\\ \nonumber &-\int_0^s\int_0^\sigma\frac{T(\tau)}{(t+s-\sigma)^{\alpha-1}}d\tau d\sigma\\ \nonumber =&\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma\\ &-\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma, \ t,s\geq 0, \end{align} where the integrals are in the sense of strong operator topology. \end{definition} We proved that $A$ generates a fractional cosine function $\{T(t)\}_{t\geq 0}$ if and only if it generates an $\alpha$-resolvent operator function, that is, the following equalities holds: \begin{equation*} T(s)J_t^\alpha T(t)-J_s^\alpha T(s)T(t)=J_t^\alpha T(t)-J_s^\alpha T(s), \ t,s\geq 0. \end{equation*} As is stated above, functional equations involving $t$, $s$ and $t+s$ have been discussed for Caputo fractional differential equations (\ref{caputo}) with $\alpha\in (0,1)$ and $\alpha\in (1,2)$, Riemann-Liouville fractional equation (\ref{R-L}) with $\alpha\in (0,1)$. To close the gap, we will discuss the residual case, that is, functional equations involving $t$, $s$ and $t+s$ for Riemann-Liouville fractional equation (\ref{R-L}) with $\alpha\in (1,2)$. To this end, we first consider the special case that $T(\cdot)$ is exponentially bounded (hence it is Laplace transformable). Take laplace transform on both sides of (\ref{fi}) with respect to $s$ and $t$ to get \begin{align}\label{lap} (\lambda^{-\alpha}-\mu^{-\alpha})\hat{T}(\mu)\hat{T}(\lambda) =\lambda^{1-\alpha}\mu^{1-\alpha}(\lambda^{-1}\hat{T}(\lambda) -\mu^{-1}\hat{T}(\mu)). \end{align} It follows from \cite[(3.8)]{Mei2014} that the Laplace transform of the right side of (\ref{cosin}) satisfies \begin{align}\label{cosin1} \nonumber&\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s}\bigg(\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma\\ \nonumber &-\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma\bigg) dsdt\\ =&\frac{\Gamma(2-\alpha)(\lambda^\alpha-\mu^\alpha)}{\lambda\mu(\lambda-\mu)} \hat{T}(\mu)\hat{T}(\lambda). \end{align} The combination of (\ref{lap}) and (\ref{cosin1}) implies that \begin{align*} &\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s}\bigg(\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma\\ &-\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma\bigg) dsdt\\ =&\frac{\Gamma(2-\alpha)(\lambda^{-1}\hat{T}(\lambda) -\mu^{-1}\hat{T}(\mu))}{\mu-\lambda}. \end{align*} Let $m(t)=\int_0^t T(\sigma)d\sigma$, by similar proof of \cite[ (4.2)]{Keyantuo2009}, there holds \begin{align*} \int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} m(t+s)dsdt=\frac{\hat{m}(\mu)-\hat{m}(\lambda)}{\lambda-\mu} =\frac{\lambda^{-1}\hat{T}(\lambda) -\mu^{-1}\hat{T}(\mu)}{\mu-\lambda}. \end{align*} By virtue of Laplace transform, it follows that \begin{align}\label{cosin2} \nonumber&\Gamma(2-\alpha)\int_0^{t+s}T(\sigma)d\sigma\\ \nonumber =&\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma\\ &-\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma, \end{align} In the following two sections, we will show that (\ref{cosin2}) also holds without the assumption that $\{T(t)\}_{t>0}$ is exponentially bounded and it essentiality describes a Riemann-Liouville fractional resolvent. \section{Riemann-Liouviller Fractional Cosine Function} Equality (\ref{fi}) is an important functional equation for the solution of equation (\ref{R-L}) with $\alpha\in(1,2)$. However, as is stated in the introduction, (\ref{fi}) does not write the functional equation in terms of the sum of time variables: $s+t$. This is very important in concrete applications of the algebraic functional equation. Therefore, it is very valuable to study functional equation (\ref{cosin2}), which appears in the following definitions. \begin{definition}\label{DEF} We call a family $\{T(t)\}_{t\geq 0}$ of bounded linear operators to be a Riemann-Liouville $\alpha$-order fractional cosine function on Banach space $X$, if the following conditions are satisfied: i) $T(t)$ is strongly continuous, that is, for any $x\in X$, the mapping $t\mapsto T(t)x$ is continuous over $(0,\infty)$; ii) there holds that \begin{equation}\label{clear} \lim_{t\rightarrow 0+}t^{2-\alpha} T(t)x =\frac{x}{\Gamma(\alpha-1)} \ \ \mbox{for all } \ x\in X; \end{equation} iii) for all $t,s> 0$, there holds \begin{align}\label{sin} \nonumber& \Gamma(2-\alpha)\int_0^{t+s}T(\sigma)d\sigma\\ \nonumber =&\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma\\ &-\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma, \end{align} where the integrals are in the sense of strong operator topology. \end{definition} \begin{lemma}\label{commutative} Let $\{T(t)\}_{t> 0}$ be a Riemann-Liouville $\alpha$-order fractional cosine on Banach space $X$. Then $\{T(t)\}_{t> 0}$ is commutative, i.e. $T(t)T(s)=T(s)T(t)$ for all $t,s > 0$. \end{lemma} \begin{proof}\ \ Observe that the left side of (\ref{sin}) is symmetric with respect to $t$ and $s$. Hence we can obtain the following equality. \begin{align*} &\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma -\int_0^t\int_0^s\frac{T(\sigma)T(\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma\\ =& \int_0^s\int_0^t\frac{T(\sigma)T(\tau)}{(s-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^s\int_0^t\frac{T(\sigma)T(\tau)}{(t-\tau)^{\alpha-1}}d\tau d\sigma -\int_0^s\int_0^t\frac{T(\sigma)T(\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma,\ t,s> 0. \end{align*} The commutative is proved by the same procedure of \cite[Proposition 3.4]{Mei2013}. \end{proof} \begin{definition}\label{sss} Let $\{T(t)\}_{t>0}$ be a Riemann-Liouville $\alpha$-order fractional cosine function on Banach space $X$. Denote by $D(A)$ the set of all $x\in X$ such that the limit \begin{align*} \lim_{t\rightarrow 0^+}\Gamma(\alpha+1)t^{-\alpha}J_t^{2-\alpha}\bigg(T(t)x- \frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x\bigg) \end{align*} exists. Then, the operator $A:D(A)\rightarrow X$ defined by \begin{align*} Ax=\lim_{t\rightarrow 0^+}\Gamma(\alpha+1)t^{-\alpha}J_t^{2-\alpha}\bigg(T(t)x- \frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x\bigg) \end{align*} is called the generator of $\{T(t)\}_{t> 0}.$ \end{definition} \begin{proposition}\label{pr2} Assume $\{T(t)\}_{t>0}$ to be a Riemann-Liouville $\alpha$-order fractional cosine function on Banach space $X$. Suppose that $A$ is the generator of $\{T(t)\}_{t>0}$. Then, (a) For any $x\in X$ and $t>0$, there holds $J_t^\alpha T(t)x\in D(A)$ and \begin{align}\label{aabbcc} T(t)x=\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x+AJ_t^\alpha T(t)x; \end{align} (b) $T(t)D(A)\subset D(A)$ and $T(t)Ax=AT(t)x$, for all $x\in D(A)$. (c) For all $x\in D(A)$, we have \begin{align*} T(t)x=\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x+J_t^\alpha T(t)Ax; \end{align*} (d) $A$ is equivalently defined by \begin{align}\label{ssf} Ax=\Gamma(2\alpha-1)\lim_{t\rightarrow 0^+}\frac{T(t)x-\frac{t^{\alpha-2}} {\Gamma(\alpha-1)}x}{t^{2\alpha-2}} \end{align} and $D(A)$ is just consists of those $x\in X$ such that the above limit exists. (e) $A$ is closed and densely defined. (f) $A$ admits at most one Riemann-Liouville $\alpha$-order fractional cosine function. \end{proposition} \begin{proof}\ \ (a) Let $x\in X$ and $b>0$ be fixed. Denote by $g_{b}(\cdot)$ the truncation of $T(\cdot)$ at $b$, that is, \begin{align*} g_{b}(\sigma)=\left\{ \begin{array}{ll} T(\sigma), & \hbox{ } 0< \sigma\leq b\\ 0, & \hbox{} \sigma>b. \end{array} \right. \end{align*} Define the function $H_b(r,s)$ for $r,s> 0$ by \begin{align}\label{Htrs} H_b(r,s)=\bigg(g_b(r)-\frac{r^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg)J_s^\alpha g_b(s)x. \end{align} Obviously, for $0< r\leq t$, \begin{align}\label{HET} H_t(r,t)=\bigg(T(r)-\frac{r^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg)J_t^\alpha T(t)x. \end{align} Take Laplace transform with respect to $r$ and $s$ successively for both sides of (\ref{Htrs}) to derive \begin{align}\label{htl} \hat{H}_b(\mu,\lambda)=\lambda^{-\alpha}\hat{g}_b(\mu)\hat{g}_b(\lambda)x -\lambda^{-\alpha}\mu^{1-\alpha}\hat{g}_b(\lambda)x. \end{align} Denote by $L(t,s)$ and $R(t,s)$ the left and right sides of equality (\ref{sin}), respectively. Moreover, denote by $R_b(t,s)$, and $L_b(t,s)$ the quantities resulted by replacing $T(t)$ with $g_b(t)$ in $R(t,s)$, $L(t,s)$, respectively. It follows from (3.7) of \cite{Mei2014} that the Laplace transform of $R_b(t,s)$ with respect to $t$ and $s$ is given by \begin{align}\label{R} \hat{R}_b(\mu,\lambda) =\frac{\Gamma(2-\alpha)(\lambda^\alpha-\mu^\alpha)}{\lambda\mu(\lambda-\mu)}\hat{g}_b(\mu)\hat{g}_b(\lambda). \end{align} For all $t> 0$, the Laplace transform of $\hat{L}_b(t,s)$ with respect to $s$ and $t$ can be obtained as follows: \begin{align}\label{L} \hat{L}_b(\mu,\lambda) =\Gamma(2-\alpha)\frac{\lambda^{-1}\hat{g_b}(\lambda) -\mu^{-1}\hat{g_b}(\mu)}{\mu-\lambda}. \end{align} Combine (\ref{htl}), (\ref{R}) and (\ref{L}) to derive \begin{align*} \hat{H}_b(\mu,\lambda)=&\mu^{-\alpha}\hat{g}_b(\mu)\hat{g}_b(\lambda)x -\mu^{-\alpha}\lambda^{1-\alpha}\hat{g}_b(\mu)x\\ &+\frac{\lambda^{1-\alpha}\mu^{1-\alpha}(\lambda-\mu)}{\Gamma(2-\alpha)}(\hat{L}_b(\mu,\lambda)-\hat{R}_b(\mu,\lambda))x. \end{align*} Take inverse Laplace transform to obtain \begin{align*} H_b(r,s)=&\bigg(g_b(s)-\frac{s^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg)J_r^\alpha g_b(r)x\\ &+\frac{[(D_s^{2-\alpha})J_r^{\alpha-1}-(D_r^{2-\alpha})J_s^{\alpha-1}]\cdot[L_b(r,s)-R_b(r,s)]x}{\Gamma(2-\alpha)}. \end{align*} Here the Laplace transform formulas \begin{align*} \widehat{D_b^\beta f}(\lambda)=\lambda^\beta \hat{f}(\lambda)-\lim_{t\rightarrow 0^+}J_t^{\alpha-1}f(t), 0<\beta<1, f\in C([0,\infty),X) \end{align*} is used. By the definition of $g_b$, it follows that $L_b(r,s)=R_b(r,s)$ for all $0< s,r\leq b$, we have that \begin{align*} H_b(r,s)=\bigg(T(s)-\frac{s^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg)J_r^\alpha T(r)x, \forall \mbox{ }0< r,s \leq b. \end{align*} This implies that \begin{align}\label{HTT} H_t(r,t)=\bigg(T(t)-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg)J_r^\alpha T(r)x,\forall \mbox{ } 0< r\leq t. \end{align} Combining (\ref{HET}) and (\ref{HTT}), we derive that \begin{align*} &\lim_{r\rightarrow 0^+}\Gamma(\alpha+1)r^{-\alpha}J_r^{2-\alpha}\bigg(T(r)-\frac{r^{\alpha-2}}{\Gamma(\alpha-1)}\bigg)J_t^\alpha T(t)x\\ =&\lim_{r\rightarrow 0^+}\Gamma(\alpha+1)r^{-\alpha}\bigg(T(t)-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg)J_r^2 T(r)x\\ =&\Gamma(\alpha+1)\bigg(T(t)-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg)\lim_{r\rightarrow 0^+}r^{-\alpha}\int_0^r(r-\sigma) T(\sigma)xd\sigma\\ =&\Gamma(\alpha+1)\bigg(T(t)-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg) \\ \cdot&\lim_{r\rightarrow 0^+}\int_0^1(1-\sigma)\sigma^{\alpha-2} (r\sigma)^{2-\alpha}T(r\sigma)xd\sigma. \end{align*} By the dominated convergence theorem and (b) of Definition \ref{DEF}, it follows that \begin{align*} &\lim_{r\rightarrow 0^+}\Gamma(\alpha+1)r^{-\alpha}J_r^{2-\alpha}\bigg(T(r)-\frac{r^{\alpha-2}}{\Gamma(\alpha-1)}\bigg)J_t^\alpha T(t)x\\ =&\Gamma(\alpha+1)\bigg(T(t)-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg) \int_0^1(1-\sigma)\sigma^{\alpha-2} \lim_{r\rightarrow 0^+}(r\sigma)^{2-\alpha}T(r\sigma)xd\sigma\\ =&\frac{\Gamma(\alpha+1)}{\Gamma(\alpha-1)}\bigg(T(t)-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg) \int_0^1(1-\sigma)\sigma^{\alpha-2}d\sigma x\\ =&\frac{\Gamma(\alpha+1)}{\Gamma(\alpha-1)}\bigg(T(t)-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}I\bigg) \frac{\Gamma(\alpha-1)\Gamma(2)}{\Gamma(\alpha+1)} x\\ =&T(t)x-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x. \end{align*} This implies that $J_t^\alpha T(t)x\in D(A)$ and $$AJ_t^\alpha T(t)x=T(t)x-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x.$$ (b) and (c) are directly obtained by Lemma \ref{commutative} and (a). (d) Denote by $D$ the set of those $x\in X$ such that the limit $$\lim_{t\rightarrow 0^+}\frac{T(t)x- \frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x}{t^{2\alpha-2}}$$ exists. Let $x\in D(A).$ Then, by (b), we have that \begin{align*} &\Gamma(2\alpha-1)\lim_{t\rightarrow 0^+}\frac{T(t)x- \frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x}{t^{2\alpha-2}}\\ =&\Gamma(2\alpha-1)\lim_{t\rightarrow 0^+}\frac{J_t^\alpha T(t)Ax}{t^{2\alpha-2}}\\ =&\frac{\Gamma(2\alpha-1)}{\Gamma(\alpha)}\lim_{t\rightarrow 0^+}\frac{1}{t^{2\alpha-2}}\int_0^t(t-\sigma)^{\alpha-1}T(\sigma)Axd\sigma\\ =&\frac{\Gamma(2\alpha-1)}{\Gamma(\alpha)} \lim_{t\rightarrow 0^+}\int_0^1(1-\sigma)^{\alpha-1}\sigma^{\alpha-2} (t\sigma)^{2-\alpha}T(t\sigma)Axd\sigma. \end{align*} The combination of the dominated convergence theorem and (b) of Definition \ref{DEF} indicates that \begin{align*} &\Gamma(2\alpha-1)\lim_{t\rightarrow 0^+}\frac{T(t)x- \frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x}{t^{2\alpha-2}}\\ =&\frac{\Gamma(2\alpha-1)}{\Gamma(\alpha)} \int_0^1(1-\sigma)^{\alpha-1}\sigma^{\alpha-2} \lim_{t\rightarrow 0^+}(t\sigma)^{2-\alpha}T(t\sigma)Axd\sigma\\ =&\frac{\Gamma(2\alpha-1)}{\Gamma(\alpha-1)\Gamma(\alpha)} \int_0^1(1-\sigma)^{\alpha-1}\sigma^{\alpha-2}Axd\sigma\\ =&\frac{\Gamma(2\alpha-1)}{\Gamma(\alpha-1)\Gamma(\alpha)} \frac{\Gamma(\alpha-1)\Gamma(\alpha)}{\Gamma(2\alpha-1)}Ax\\ =&Ax. \end{align*} This implies that $x\in D$ and then $D(A)\subset D$. Now we prove the converse inclusion. Let $x\in D$, that is, the limit $$\lim_{t\rightarrow 0^+}\frac{T(t)x- \frac{t^{\alpha-2}}{\Gamma(\alpha-1))}x}{t^{2\alpha-2}}.$$ exists. By the dominated convergence theorem, it follows that \begin{align*} &\lim_{t\rightarrow 0^+}\Gamma(\alpha+1)t^{-\alpha}J_t^{2-\alpha}\bigg(T(t)x- \frac{t^{\alpha-2}}{\Gamma(\alpha-1)}x\bigg)\\ =&\lim_{t\rightarrow 0^+}\frac{\Gamma(\alpha+1)}{\Gamma(2-\alpha)} \int_0^1(1-\sigma)^{1-\alpha}\sigma^{2\alpha-2} \frac{T(t\sigma)x-\frac{(t\sigma)^{\alpha-2}}{\Gamma(\alpha-1)}x} {(t\sigma)^{2\alpha-2}}d\sigma\\ =&\frac{\Gamma(\alpha+1)}{\Gamma(2-\alpha)} \int_0^1(1-\sigma)^{1-\alpha}\sigma^{2\alpha-2} \lim_{t\rightarrow 0^+}\frac{T(t\sigma)x-\frac{(t\sigma)^{\alpha-2}}{\Gamma(\alpha-1)}x} {(t\sigma)^{2\alpha-2}}d\sigma\\ =&\frac{\Gamma(\alpha+1)}{\Gamma(2-\alpha)} \frac{\Gamma(2-\alpha)\Gamma(2\alpha-1)}{\Gamma(\alpha+1)} \lim_{t\rightarrow 0^+}\frac{T(t)x-\frac{(t)^{\alpha-2}}{\Gamma(\alpha-1)}x} {t^{2\alpha-2}}. \end{align*} Hence, $x\in D(A)$ and \begin{align}\label{absc} Ax=\Gamma(2\alpha-1)\lim_{t\rightarrow 0^+}\frac{T(t)x- \frac{t^{\alpha-2}}{\Gamma(\alpha-1))}x}{t^{2\alpha-2}}. \end{align} (e) The properties that $A$ is closed and densely defined are followed directly from the combination of (d) and \cite{Li2012}. (f) Assume that both $\{T(t)\}_{t>0}$ and $\{S(t)\}_{t> 0}$ are Riemann-Liouville $\alpha$-order fractional resolvent generated by $A$. Then, by (c), for all $x\in D(A)$, we have \begin{align*} \frac{t^{\alpha-2}}{\Gamma(\alpha-1)}*T(t)x =&(S(t)-J_t^\alpha AS(t))*T(t)x\\ =&S(t)*T(t)x-(J_t^\alpha AS(t))*T(t)x\\ =&S(t)*(T(t)x-J_t^\alpha AT(t)x)\\ =& \frac{t^{\alpha-2}}{\Gamma(\alpha-1)}*S(t)x. \end{align*} By Titchmarsh's Theorem, for any $t> 0$, $T(t)=S(t)$ on $D(A)$. The result is obtained by the density of $A$. \end{proof} \begin{corollary}\label{zheng} Assume that $A$ generates a Rimann-Liouville $\alpha$-order fractional cosine function on Banach space $X$. Then $\{T(t)\}_{t> 0}$ is a Riemann-Liouville $\alpha$-order fractional resolvent. \end{corollary} \begin{proof}\ \ In (a) of Theorem \ref{pr2}, replacing $x$ with $J_s^\alpha T(s)x$, and using Lemma \ref{commutative}, we obtain that \begin{align*} T(t)J_s^\alpha T(s)x=&\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}J_s^\alpha T(s)x +AJ_t^\alpha T(t)J_s^\alpha T(s)x\\ =&\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}J_s^\alpha T(s)x +AJ_s^\alpha T(s)J_t^\alpha T(t)x\\ =&\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}J_s^\alpha T(s)x +\bigg(T(s)-\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}\bigg)J_t^\alpha T(t)x, \end{align*} which is just (\ref{fi}). The proof is therefore completed. \end{proof} \section{Equivalent to Riemann-Liouville fractional resolvent} In this section, we will prove that equality (\ref{fi}) essentially describes a Rimann-Liouville $\alpha$-order fractional cosine function. \begin{theorem}\label{relation} Suppose that $\{T(t)\}_{t> 0}$ is a Riemann-Liouville $\alpha$-order fractional resolvent on Banach space $X$. Then, the family is a Riemann-Liouville $\alpha$-order fractional cosine function. \end{theorem} \begin{proof}\ \ Denote by $L(t,s)$ and $R(t,s)$ the left and right sides of equality (\ref{sin}), respectively. Obviously, what we need is to prove that $L(t,s)=R(t,s)$ for all $t,s> 0.$ For brevity, we introduce the following notations. Let \begin{align*} H(t,s)=&T(t)J_s^\alpha T(s)-J_t^\alpha T(t)T(s),\\ K(t,s)=&\frac{t^{\alpha-2}}{\Gamma(\alpha-1)}J_{s}^{\alpha}T(s) -\frac{s^{\alpha-2}}{\Gamma(\alpha-1)}J_{t}^{\alpha}T(t), t,s> 0. \end{align*} Moreover, for sufficiently large $b>0$ denote by $g_b(t)$ the truncation of $T(t)$ at $b$, and by $R_b(t,s)$, $L_b(t,s)$, $H_b(t,s)$ and $K_b(t,s)$ the quantities resulted by replacing $T(t)$ with $g_b(t)$ in $R(t,s)$, $L(t,s)$, $H(t,s)$ and $K(t,s)$, respectively. We set \begin{align*} P_b(t,s)=&\int_0^t\int_0^s\frac{H_b(\sigma,\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{H_b(\sigma,\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma\\ \nonumber &-\int_0^t\int_0^s\frac{H_b(\sigma,\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma \end{align*} and \begin{align}\label{limi} Q_b(t,s)=&\int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma\\ \nonumber &-\int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma. \end{align} Observe that the equality (\ref{fi}) implies $H(t,s)=K(t,s)$ for any $t,s> 0$. Thus, for all $t,s>0$, \begin{align}\label{li} \lim_{b\rightarrow \infty}P_b(t,s)=\lim_{b\rightarrow \infty}Q_b(t,s). \end{align} By \cite[(3.13)]{Mei2014}, it follows that \begin{align}\label{P} P_b(t,s)=(J_s^\alpha-J_t^\alpha)R_b(t,s), \forall\ t,s> 0. \end{align} We now compute Laplace transform of the first term of $Q_b(t,s)$ with respect to $s$ and $t$ as follows, \begin{align*} &\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} \int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma dsdt\\ =&\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} \int_0^t\int_0^s\frac{\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)} J_{\tau}^{\alpha}g_b(\tau) -\frac{\tau^{\alpha-2}}{\Gamma(\alpha-1)}J_{\sigma}^{\alpha}g_b(\sigma)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma dsdt\\ =&\int_0^\infty e^{-\mu t}\int_0^t\int_0^\infty e^{-\lambda s} \int_0^s\frac{\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)}J_{\tau}^{\alpha}g_b(\tau) -\frac{\tau^{\alpha-2}}{\Gamma(\alpha-1)}J_{\sigma}^{\alpha}g_b(\sigma)}{(t-\sigma)^{\alpha-1}}d\tau ds d\sigma dt\\ =&\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} \int_0^t\int_0^s\frac{\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)}J_{\tau}^{\alpha}g_b(\tau) -\frac{\tau^{\alpha-2}}{\Gamma(\alpha-1)}J_{\sigma}^{\alpha}g_b(\sigma)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma dsdt\\ =&\int_0^\infty e^{-\mu t}\int_0^t\frac{\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)} }{(t-\sigma)^{\alpha-1}}\int_0^\infty e^{-\lambda s} \int_0^sJ_{\tau}^{\alpha}g_b(\tau)d\tau ds d\sigma dt\\ &-\int_0^\infty e^{-\mu t}\int_0^t\frac{J_{\sigma}^{\alpha}g_b(\sigma) }{(t-\sigma)^{\alpha-1}}\int_0^\infty e^{-\lambda s} \int_0^s\frac{\tau^{\alpha-2}}{\Gamma(\alpha-1)}d\tau ds d\sigma dt\\ =&\Gamma(2-\alpha)\mu^{-1}\lambda^{-\alpha-1}\hat{g}_b(\lambda) -\Gamma(2-\alpha)\mu^{-2}\lambda^{-\alpha}\hat{g}_b(\mu), \end{align*} The Laplace transform of the second term of $Q_b(t,s)$ with respect to $s$ and $t$ is computed as follows \begin{align*} &\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} \int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma dsdt\\ =&\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} \int_0^t\int_0^s\frac{\frac{\sigma^{\alpha-2}} {\Gamma(\alpha-1)}J_{\tau}^{\alpha}g_b(\tau) -\frac{\tau^{\alpha-2}}{\Gamma(\alpha-1)}J_{\sigma}^{\alpha}g_b(\sigma)}{(s-\tau)^{\alpha-1}}d\tau d\sigma dsdt\\ =&\int_0^\infty e^{-\mu t}\int_0^t\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1))} \int_0^\infty e^{-\lambda s} \int_0^s\frac{J_{\tau}^{\alpha}g_b(\tau) }{(s-\tau)^{\alpha-1}}d\tau dsd\sigma dt\\ &-\int_0^\infty e^{-\mu t}\int_0^tJ_{\sigma}^{\alpha}g_b(\sigma)\int_0^\infty e^{-\lambda s} \int_0^s\frac{\frac{\tau^{\alpha-2}}{\Gamma(\alpha-1)} }{(s-\tau)^{\alpha-1}}d\tau dsd\sigma dt\\ =&\Gamma(2-\alpha)\mu^{-\alpha}\lambda^{-2}\hat{g}_b(\lambda)- \Gamma(2-\alpha)\lambda^{-1}\mu^{-\alpha-1}\hat{g}_b(\mu). \end{align*} We compute the Laplace transform of the third term of $Q_b(t,s)$ with respect to $s$ and $t$ as follows. \begin{align*} &-\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} \int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma dsdt\\ =&-\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} \int_0^t\int_0^s\frac{\frac{\sigma^{\alpha-2}} {\Gamma(\alpha-1)}J_{\tau}^{\alpha}g_b(\tau) -\frac{\tau^{\alpha-2}}{\Gamma(\alpha-1)}J_{\sigma}^{\alpha}g_b(\sigma)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma dsdt\\ =&-\int_0^\infty e^{-\mu t}\int_0^t\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)}\int_0^\infty e^{-\lambda s} \int_0^s\frac{J_{\tau}^{\alpha}g_b(\tau) }{(t+s-\sigma-\tau)^{\alpha-1}}d\tau dsd\sigma dt\\ &+\int_0^\infty e^{-\mu t}\int_0^tJ_{\sigma}^{\alpha}g_b(\sigma)\int_0^\infty e^{-\lambda s} \int_0^s\frac{\frac{\tau^{\alpha-2}}{\Gamma(\alpha-1)}}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau dsd\sigma dt\\ =&-\int_0^\infty e^{-\mu t}\int_0^t\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)}\int_0^\infty e^{-\lambda s}\frac{1}{(t+s-\sigma)^{\alpha-1}} dsd\sigma dt\lambda^{-\alpha}g_b(\lambda)\\ &+\lambda^{1-\alpha}\int_0^\infty e^{-\mu t}\int_0^tJ_{\sigma}^{\alpha}g_b(\sigma)\int_0^\infty e^{-\lambda s} \frac{1}{(t+s-\sigma)^{\alpha-1}} dsd\sigma dt\\ =&-\int_0^\infty e^{-\mu t}\int_0^t\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)} e^{\lambda(t-\sigma)}\bigg(\int_0^\infty e^{-\lambda r}r^{1-\alpha} dr-\int_0^{t-\sigma} e^{-\lambda r}r^{1-\alpha} dr\bigg)d\sigma dt\lambda^{-\alpha}g_b(\lambda)\\ &+\lambda^{1-\alpha}\int_0^\infty e^{-\mu t}\int_0^tJ_{\sigma}^{\alpha}g_b(\sigma)e^{\lambda(t-\sigma)} \bigg(\int_0^\infty e^{-\lambda r} r^{1-\alpha} dr-\int_0^{t-\sigma} e^{-\lambda r} r^{1-\alpha} dr\bigg)d\sigma dt\\ =&-\Gamma(2-\alpha)\lambda^{\alpha-2} \int_0^\infty e^{-\mu t}\int_0^t\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)} e^{\lambda(t-\sigma)}d\sigma dt\lambda^{-\alpha}g_b(\lambda)\\ &+\int_0^\infty e^{-\mu t}\int_0^t\frac{\sigma^{\alpha-2}}{\Gamma(\alpha-1)} \int_0^{t-\sigma} e^{\lambda(t-\sigma-r)}r^{1-\alpha} drd\sigma dt\lambda^{-\alpha}g_b(\lambda)\\ &+\Gamma(2-\alpha)\lambda^{\alpha-2}\lambda^{1-\alpha} \int_0^\infty e^{-\mu t}\int_0^tJ_{\sigma}^{\alpha}g_b(\sigma)e^{\lambda(t-\sigma)}d\sigma dt\\ &-\lambda^{1-\alpha}\int_0^\infty e^{-\mu t}\int_0^tJ_{\sigma}^{\alpha}g_b(\sigma) \int_0^{t-\sigma} e^{\lambda(t-\sigma-r)} r^{1-\alpha} drd\sigma dt\\ =&-\Gamma(2-\alpha)\lambda^{\alpha-2}\lambda^{-\alpha} \frac{\mu^{1-\alpha}}{\mu-\lambda}g_b(\lambda) +\Gamma(2-\alpha)\frac{\mu^{-1}}{\mu-\lambda}\lambda^{-\alpha}g_b(\lambda)\\ &+\Gamma(2-\alpha)\lambda^{\alpha-2}\lambda^{1-\alpha} \frac{\mu^{-\alpha}}{\mu-\lambda}\hat{g}_b(\mu) -\Gamma(2-\alpha)\lambda^{1-\alpha}\mu^{\alpha-2} \frac{\mu^{-\alpha}}{\mu-\lambda}\hat{g}_b(\mu). \end{align*} Using (\ref{L}), we can obtain that \begin{align}\label{dd} \nonumber\hat{Q}_b(\mu,\lambda)=&\int_0^\infty e^{-\mu t}\int_0^\infty e^{-\lambda s} \bigg(\int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(t-\sigma)^{\alpha-1}}d\tau d\sigma+\int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(s-\tau)^{\alpha-1}}d\tau d\sigma\\ \nonumber &-\int_0^t\int_0^s\frac{K_b(\sigma,\tau)}{(t+s-\sigma-\tau)^{\alpha-1}}d\tau d\sigma)\bigg)dsdt\\ =&(\lambda^{-\alpha}-\mu^{-\alpha})\hat{L_b}(\mu,\lambda). \end{align} Taking inverse Laplace transform on both sides of (\ref{dd}), we derive \begin{align}\label{Q} Q_b(t,s)=(J_s^\alpha-J_t^\alpha)L_b(t,s),\ \forall \ t, \ s> 0. \end{align} Form (\ref{P}) and (\ref{Q}), we have that \begin{align*} (J_s^\alpha-J_t^\alpha)L(t,s)=(J_s^\alpha-J_t^\alpha)R(t,s),\ \forall \ t, \ s>0. \end{align*} Therefore, $L(t,s)=R(t,s)$. This completes the proof. \end{proof} Combining Corollary \ref{zheng} and Theorem \ref{relation}, we can obtain the equivalent of Riemann-Liouville $\alpha$-order fractional resolvents and Riemann-Liouville $\alpha$-order fractional cosine functions.

https://arxiv.org/abs/1607.06033

Canonical bases of quantum Schubert cells and their symmetries

The goal of this work is to provide an elementary construction of the canonical basis $\mathbf B(w)$ in each quantum Schubert cell~$U_q(w)$ and to establish its invariance under modified Lusztig's symmetries. To that effect, we obtain a direct characterization of the upper global basis $\mathbf B^{up}$ in terms of a suitable bilinear form and show that $\mathbf B(w)$ is contained in $\mathbf B^{up}$ and its large part is preserved by modified Lusztig's symmetries.

\section{Introduction and main results} Let $\lie g=\lie b^-\oplus \lie n^+$ be a symmetrizable Kac-Moody Lie algebra. For any $w$ in its Weyl group~$W$, define the algebra $U_q(w)$ by \begin{equation}\label{eq:quantum-Schubert-cell} U_q(w):=T_w(U_q(\lie b^-))\cap U_q(\lie n^+) \end{equation} which we refer to as a quantum Schubert cell (see \S\ref{subs:prelim} for notation). This terminology is justified in Remark~\ref{rem:2}. The definition~\eqref{eq:quantum-Schubert-cell} for an infinite (affine) type first appeared in~\cite{BCP}*{Proposition~2.3}. In~\cite{BG-dcb} we conjectured that this definition coincides with Lusztig's one which was proved for all Kac-Moody algebras by Tanisaki in \cite{T}*{Proposition~2.10} and independently by Kimura (\cite{Kim1}*{Theorem~1.3}). In a remarkable paper~\cite{Kim} Kimura proved that each $U_q(w)$ is compatible with the upper global basis $\mathbf B^{up}$ of~$U_q(\lie n^+)$. The aim of the present work is twofold: \begin{itemize} \item to construct the basis $\mathbf B(w)$ of~$U_q(w)$ explicitly using a generalization of Lusztig's Lemma. \item to compute the action of Lusztig symmetries on these bases, thus partially verifying Conjecture~1.16 from~\cite{BG-dcb}. \end{itemize} To achieve the first goal, first we provide an independent definition (see~\S\ref{subs:defn-basis} and~\S\ref{subs:chos basis}) of the global crystal basis $\mathbf B^{up}$ (which coincides with the dual canonical basis). For reader's convenience, we put all necessary definitions and results in Section~\ref{sec:char bas}. Let $\bar\cdot$ be the anti-linear anti-involution of~$U_q(\lie g)$ which maps $q^{\frac12}$ to~$q^{-\frac12}$ and fixes Chevalley generators. It should be noted that we use a slightly different presentation of~$U_q(\lie g)$ (see~\cite{BG-dcb} and \S\ref{subs:prelim}) and accordingly modified~$T_w$ so that they commute with $\bar\cdot$ (\cite{BG-dcb}*{}). Let $\ii=(i_1,\dots,i_m)\in R(w)$. Generalizing~\cites{DKP,LS}, we show (see~\S\ref{subs:some prop q S c}) that the $\AA:=\ZZ[q^{\frac12},q^{-\frac12}]$-subalgebra of~$U_q(\lie n^+)$ generated by the $X_{\mathbf i,k}:=T_{s_{i_1}\cdots s_{i_{k-1}}}(E_{i_k})$, $1\le k\le m$ of~$U_q(\lie n^+)$ is in fact independent of~$\ii$, hence is denoted $U^\AA(w)$, and by Lemma~\ref{lem:lus PBW elt} has an $\AA$-basis $\{X_\ii^{\mathbf a}\,:\, \mathbf a\in \ZZ_{\ge 0}^m\}$ where $X_\ii^{\mathbf a}=q_{\mathbf i,\mathbf a}X_{\ii,1}^{a_1}\cdots X_{\ii,m}^{a_m}$ and $q_{\mathbf i,\mathbf a}\in q^{\frac12\ZZ}$ is defined in~\eqref{eq:q_i a defn}. The importance of this choice of the $q_{\ii,\mathbf a}$ is highlighted by the following version of Lusztig's Lemma. \begin{theorem}\label{thm:bas-from-Lusztig-lemma} Let $w\in W$ and $\ii\in R(w)$. For every $\mathbf a\in \ZZ_{\ge 0}^m$ there exists a unique $b_{\mathbf a}=b_{\mathbf i,\mathbf a}\in U^\AA(w)$ such that $\overline{b_{\mathbf a}}=b_{\mathbf a}$ and $ b_{\mathbf a}-X_\ii^{\mathbf a}\in\sum_{\mathbf a'\not=\mathbf a} q^{-1} \ZZ[q^{-1}] X_\ii^{\mathbf a'}$. \end{theorem} We prove Theorem~\ref{thm:bas-from-Lusztig-lemma} in~\S\ref{subs:pf-bas-from-Lus-lemma}. In particular, elements $b_{\ii,\mathbf a}$, $\mathbf a\in\ZZ_{\ge 0}^m$ form a basis $\mathbf B(\ii)$ of $U^\AA(w)$ which a priori depends on~$\ii$. However, the following result implies that this is not the case. \begin{theorem}\label{thm:inclusion of base} Let $w\in W$ and $\mathbf i\in R(w)$. Then for all $\mathbf a\in \ZZ_{\ge 0}^{\ell(w)}$ we have $b_{\mathbf i,\mathbf a}\in \mathbf B^{up}$. \end{theorem} Theorem~\ref{thm:inclusion of base} is proved in~\S\ref{subs:pf inclusion of base}. It implies that for any $\ii,\ii'\in R(w)$ we have $\mathbf B(\ii)=\mathbf B(\ii')$ and thus can introduce $\mathbf B(w)$. As a consequence, we recover the main result (Theorem~4.22) of~\cite{Kim}. \begin{corollary}\label{cor:Kim} $\mathbf B(w)=\mathbf B^{up}\cap U_q(w)$ for all $w\in W$. \end{corollary} \begin{remark} To obtain his result, Kimura used a rather elaborate theory of global crystal bases. By contrast, our proofs of Theorems~\ref{thm:bas-from-Lusztig-lemma} and~\ref{thm:inclusion of base} are quite elementary and short. \end{remark} Now we turn our attention to the second goal, that is, to the action of Lusztig's symmetries on~$U_q(w)$. \begin{theorem}[\cite{BG-dcb}*{Conjecture~1.17}]\label{thm:T_w(B(w'))} Let $w,w'\in W$ be such that $\ell(ww')=\ell(w)+\ell(w')$. Then $$ \mathbf B(w)\subset \mathbf B(ww'),\qquad T_w(\mathbf B(w'))\subset \mathbf B(w w'). $$ \end{theorem} \begin{remark} \begin{enumerate}[{\rm(a)}] \item In~\cite{BG-dcb} we constructed a basis $\mathbf B_{\lie g}$ of~$U_q(\lie g)$ containing $\mathbf B^{up}$ and conjectured (\cite{BG-dcb}*{Conjecture~1.16}) that $T_w(\mathbf B^{up})\subset \mathbf B_{\lie g}$. Thus, Theorem~\ref{thm:T_w(B(w'))} provides supporting evidence for that conjecture. \item It would be interesting to compare the symmetries discussed above with the quantum twist computed in~\cite{KimOya}. \end{enumerate} \end{remark} We deduce Theorem~\ref{thm:T_w(B(w'))} from Theorem~\ref{thm:bas-from-Lusztig-lemma} in~\S\ref{subs:pf T_w(B(w'))}. All these results are obtained using the following striking property of~$\mathbf B^{up}$ which is parallel to a highly non-trivial result of Lusztig (\cite{Lus-adv}). \begin{theorem}\label{thm:T_i bas} $T_{s_i}(b)\in\mathbf B^{up}$ whenever $b\in \mathbf B^{up}\cap T_{s_i}^{-1}(U_q(\lie n^+))$. \end{theorem} We prove Theorem~\ref{thm:T_i bas} in~\S\ref{subs:proof of thm T_i bas}. Our proof, which is quite elementary and short, relies on the notion of {\em decorated algebras} (Definition~\ref{defn:decorated alg}) to which we generalize $T_{s_i}$ and obtain an explicit formula (Theorem~\ref{thm:tau-homomorphism}) for it. We conclude this section with the following curious application of the above constructions. It is well-known (see e.g. Remark~\ref{rem: star op partial}) that the natural linear anti-involution ${}^*$ on~$U_q(\lie n^+)$ fixing the Chevalley generators (see~\S\ref{subs:prelim}) preserves $\mathbf B^{up}$. Since $T_w\circ {}^*={}^*\circ T_{w^{-1}}^{-1}$ (cf.~\cite{BG-dcb}) it follows that $$ U_q(w)^*=T_{w^{-1}}^{-1}(U_q(\lie b^-))\cap U_q(\lie n^+) $$ and Corollary~\ref{cor:Kim} implies that $U_q(w)^*$ has a basis $\mathbf B(w)^*=U_q(w)^*\cap \mathbf B^{up}$. In particular, one can consider the algebras $$ U_q(w,w'):=U_q(w)\cap U_q(w')^*,\qquad w,w'\in W $$ which is natural to call bi-Schubert algebras. The following is immediate. \begin{corollary} For any $w,w'\in W$, the bi-Schubert algebra $U_q(w,w')$ has a basis $\mathbf B(w,w'):=\mathbf B(w)\cap \mathbf B(w')^*=U_q(w,w')\cap \mathbf B^{up}$. \end{corollary} \begin{remark} \begin{enumerate}[\rm(a)] \item \label{rem:intersect sc.a} Based on numerous examples (see~\S\ref{subs:bi-schub}) one can conjecture that bi-Schubert algebras are Poin\-ca\-r\'e-Birkhoff-Witt (PBW), i.e. have bases consisting of ordered monomials (similarly to Lemma~\ref{lem:lus PBW elt}). \item\label{rem:intersect sc.b} One can also consider intersections $U_q(w)\cap U_q(w')$; however, in this case it appears (and is probably well-known) that the corresponding algebra is always $U_q(w'')$ where $w''$ is less than both $w$ and~$w'$ in the weak right Bruhat order and is maximal with that property. \end{enumerate}\label{rem:intersect sc} \end{remark} \addtocontents{toc}{\SkipTocEntry} \subsection*{Acknowledgements} The main part of this paper was written while both authors were visiting Universit\'e de Gen\`eve (Geneva, Switzerland). We are happy to use this opportunity to thank A.~Alekseev for his hospitality. We also benefited from the hospitality of Max-Planck-Institut f\"ur Mathematik (Bonn, Germany), which we gratefully acknowledge. \section{Definition and characterization of~\texorpdfstring{$\mathbf B^{up}$}{B\^up}}\label{sec:char bas} \subsection{Preliminaries}\label{subs:prelim} Let $\lie g$ be a symmetrizable Kac-Moody algebra with the Cartan matrix $A=(a_{ij})_{i,j\in I}$. Let $\{\alpha_i\}_{i\in I}$ be the standard basis of~$Q=\ZZ^I$. Fix $d_i\in \ZZ_{>0}$, $i\in I$ such that the matrix $(d_i a_{ij})_{i,j\in I}$ is symmetric and define a symmetric bilinear form $(\cdot,\cdot):Q\times Q\to\ZZ$ by $(\alpha_i,\alpha_j)=d_i a_{ij}$; clearly, $(\gamma,\gamma)\in2\ZZ$ for any $\gamma\in Q$. We will write $(\alpha_i^\vee,\gamma)$, $\gamma\in Q$ as an abbreviation for $(\alpha_i,\gamma)d_i^{-1}$. The quantized enveloping algebra $U_q(\lie g)$ is an associative algebra over $\kk=\mathbb Q(q^{\frac12})$ generated by the $E_i$, $F_i$, $K_i^{\pm 1}$, $i\in I$ subject to the relations \begin{gather} \label{eq:commutation g} [E_i,F_j]=\delta_{ij}(q_i^{-1}-q_i)(K_{i}-K_{i}^{-1}) ,\quad K_i E_j = q_i^{a_{ij}}E_j K_i,\,\, K_i F_j =q_i^{-a_{ij}}F_jK_{i},\,\, K_iK_j=K_jK_i \\ \label{eq:qserre} \sum\limits_{r,s\ge 0,\,r+s=1-a_{ij}}\mskip-8mu (-1)^s E_i^{\la r\ra}E_j E_i^{\la s\ra}= \sum\limits_{r,s\ge 0,\,r+s=1-a_{ij}}\mskip-8mu (-1)^s F_i^{\la r\ra}F_j F_i^{\la s\ra}=0,\quad i\not=j \end{gather} for all $i,j\in I$, where $q_i=q^{d_i}$, $X_i^{\la k\ra}:=\big(\prod\limits_{s=1}^k \la s\ra_{q_i}\big)^{-1} X_i^k$ and $\la s\ra_v=v^s-v^{-s}$. We also set $$ (n)_v=\frac{\la n\ra_v}{\la 1\ra_v},\quad \la n\ra_v!=\prod_{t=1}^n \la t\ra_v,\quad (n)_v!=\frac{\la n\ra_v!}{\la 1\ra_v^n},\quad \binom{n}{k}_v=\frac{\prod_{t=0}^{k-1} (n-t)_v}{(k)_v!}=\frac{\prod_{t=0}^{k-1} \la n-t\ra_v}{\la k\ra_v!} $$ and $X_i^{(n)}:=X_i^n/(n)_{q_i}!$. We denote by $U_q(\lie n^+)$ (respectively, $U_q(\lie n^-)$) the subalgebra of $U_q(\lie g)$ generated by the $E_i$ (respectively, the $F_i$), $i\in I$. Let $\mathcal K$ be the subalgebra of $U_q(\lie g)$ generated by the $K_i^{\pm 1}$, $i\in I$ and set $U_q(\lie b^\pm)=\mathcal KU_q(\lie n^\pm)$. It is easy to see from the presentation that $U_q(\lie g)$ admits anti-involutions ${}^t$ and ${}^*$, where ${}^t$ interchanges $E_i$ and $F_i$ for each $i\in I$ and preserves the $K_i^{\pm 1}$ while ${}^*$ preserves the $E_i$ and $F_i$ while $K_i^*=K_i^{-1}$. Furthermore, $U_q(\lie g)$ admits an anti-linear anti-involution $\bar\cdot$ which preserves all generators and maps $q^{\frac12}$ to $q^{-\frac12}$. The algebra $U_q(\lie n^+)$ is naturally graded by $Q^+:=\bigoplus_{i\in I}\ZZ_{\ge 0}\alpha_i$ via $\deg E_i=\alpha_i$. We denote the homogeneous component of $U_q(\lie n^+)$ of degree~$\gamma\in Q^+$ by $U_q(\lie n^+)_\gamma$. This can be extended to a $Q$-grading on~$U_q(\lie g)$ via $\deg F_i=-\alpha_i$, $\deg K_i=0$. \subsection{Modified Lusztig symmetries}\label{subs:Weyl-Lusztig symm} Let~$W$ be the Weyl group of~$\lie g$. It is generated by the simple reflections $s_i$, $i\in I$ which act on~$Q$ via $s_i(\alpha_j)=\alpha_j-a_{ij}\alpha_i$. Given $w\in W$, denote $R(w)$ the set of reduced words for $w$, that is, the set of $\mathbf i=(i_1,\dots,i_m)\in I^m$ of minimal length $m:=\ell(w)$ such that $w=s_{i_1}\cdots s_{i_m}$. It is well-known that the form $(\cdot,\cdot)$ is $W$-invariant. The following essentially coincides with Theorem~1.13 from~\cite{BG-dcb}. \begin{lemma} \begin{enumerate}[{\rm(a)}] \item For each $i\in I$ there exists a unique automorphism~$T_i$ of $U_q(\lie g)$ which satisfies $T_i(K_j)=K_jK_i^{-a_{ij}}$ and \begin{gather*} T_i(E_j)= \begin{cases} q_i^{-1} K_{i}^{-1} F_i,&i=j\\ \sum\limits_{r+s=-a_{ij}}\mskip-8mu (-1)^r q_i^{s+\frac12 a_{ij}} E_i^{\la r\ra}E_j E_i^{\la s\ra}, &i\ne j\\ \end{cases} \\ T_i(F_j)= \begin{cases} q_i^{-1} K_{i} E_i,& i=j\\ \sum\limits_{r+s=-a_{ij}}\mskip-8mu (-1)^r q_i^{s+\frac12 a_{ij}} F_i^{\la r\ra}F_j F_i^{\la s\ra}, &i\ne j\\ \end{cases} \end{gather*} \item For all $x\in U_q(\lie g)$, $\overline{T_i(x)}=T_i(\overline x)$, $(T_i(x))^*=T_i^{-1}(x^*)$ and $(T_i(x))^t=T_i^{-1}(x^t)$. \item The $T_i$, $i\in I$ satisfy the braid relations on $U_q(\lie g)$, that is, they define a representation of the Artin braid group $\operatorname{Br}_{\lie g}$ of~$\lie g$ on~$U_q(\lie g)$. \end{enumerate} \label{lem:T_i-defn-prop} \end{lemma} \subsection{Bilinear forms} Following~\cites{Lus-book,BG-dcb}, we define a symmetric bilinear form $\la\cdot,\cdot\ra$ on $U_q(\lie n^+)$. Let $V=\bigoplus_{i\in I} \kk E_i$ and let $\la\cdot,\cdot\ra$ be the bilinear form on~$V$ defined by $\la E_i,E_j\ra=\delta_{ij}(q_i-q_i^{-1})$. Extend it naturally to $T(V)$ via $$ \la v_1\tensor\cdots\tensor v_k,v'_1\tensor\cdots\tensor v'_l\ra' =\delta_{kl}\prod_{r=1}^k \la v_r,v'_r\ra ,\qquad v_r,\,v'_r\in V,\,1\le r\le k. $$ Define a linear map $\Psi:V\tensor V\to V\tensor V$ by $\Psi(E_i\tensor E_j)=q^{(\alpha_i,\alpha_j)}E_j\tensor E_i$. Finally, define $\la\cdot,\cdot\ra_\Psi$ via $$ \la u,v\ra_\Psi=\delta_{k,l} \la [k]_\Psi!(u),v\ra'=\delta_{k,l}\la u,[k]_\Psi!(v)\ra',\qquad u\in V^{\tensor k},\, v\in V^{\tensor l} $$ and $[k]_\Psi!\in\End_\kk V^{\tensor k}$ is the standard notation for the braided $k$-factorial (see e.g.~\cite{BG-dcb}*{\S A.1}). It is well-known (see e.g.~\cite{Lus-book}) that the kernel~$J$ of the canonical map $T(V)\to U_q(\lie n^+)$, $E_i\mapsto E_i$ is the radical of $\la\cdot,\cdot\ra_\Psi$. Thus, we have a well-defined non-degenerate symmetric bilinear form $\la\cdot,\cdot\ra$ on~$U_q(\lie n^+)$ given by $\la u+J,v+J\ra=\la u,v\ra_\Psi$. This form in fact coincides with the form $\la\cdot,\cdot\ra$ we introduced in~\cite{BG-dcb}*{\S A.3} if we identify $U_q(\lie n^-)$ with $U_q(\lie n^+)$ via ${}^{*t}$. We will often use the following obvious \begin{lemma}\label{lem:form-non-zero-deg} Let $x,x'\in U_q(\lie n^+)$ be homogeneous. Then $\la x,x'\ra\not=0$ implies that $\deg x=\deg x'$. \end{lemma} Define $\fgfrm{\cdot}{\cdot}:U_q(\lie n^+)\tensor U_q(\lie n^+)\to \kk$ by $$ \fgfrm{x}{y}= \mu(\gamma)q^{-\frac12(\gamma,\gamma)} \la x,y\ra,\qquad x,y\in U_q(\lie n^+)_\gamma $$ where \begin{equation}\label{eq:defn-mu} \mu(\gamma)=q^{\frac14(\gamma,\gamma)+\frac12\eta(\gamma)},\qquad \gamma\in Q \end{equation} and $\eta\in\Hom_\ZZ(Q,\ZZ)$ is defined by $\eta(\alpha_i)=d_i$. Note the following properties of~$\mu$ which will be often used in the sequel \begin{equation}\label{eq:mu-prop} \mu(r\alpha_i)=q_i^{\binom{r+1}2},\quad \mu(s_i\gamma)=\mu(\gamma)q^{-\frac12(\alpha_i,\gamma)},\quad \mu(\gamma+\gamma')= \mu(\gamma)\mu(\gamma')q^{\frac12(\gamma,\gamma')} \end{equation} Define an anti-linear automorphism $\tilde\cdot$ of $U_q(\lie g)$ by $$ \tilde x=(\sgn\gamma)\bar x{}^*,\qquad x\in U_q(\lie g)_\gamma $$ where $\sgn:Q\to\{\pm 1\}$ is the homomorphism of abelian groups defined by $\sgn(\alpha_i)=-1$. Then (cf.~\cite{BG-dcb}*{}) \begin{equation}\label{eq:bar tilde} \overline{\fgfrm{x}{y}}=\fgfrm{\bar x}{\tilde y}=\fgfrm{\tilde x}{\bar y}. \end{equation} \subsection{Lattices and signed basis in \texorpdfstring{$U_q(\lie n^+)$}{Uq(n+)}}\label{subs:defn-basis} Let $\AA=\ZZ[q^{\frac12},q^{-\frac12}]$ which is a subring of~$\QQ(q^{\frac12})$. Denote $\AA_0=\ZZ[q,q^{-1}]$ and $\AA_1=q^{\frac12}\AA_0$; clearly, $\AA=\AA_0\oplus \AA_1$ as an $\AA_0$-module. Following~\cite{BG-dcb}*{\S3.1}, for any $J\subset I$, let $U_\ZZ(\lie n^+)_J$ (respectively, $U_\ZZ(\lie n^-)_J$) be the $\AA_0$-subalgebra of~$U_q(\lie n^+)$ (respectively, $U_q(\lie n^-)$) generated by the $E_i^{\la n\ra}$ (respectively, $F_i^{\la n\ra}$), $i\in J$, $n\in\ZZ_{\ge0}$. We abbreviate $U_\ZZ(\lie n^\pm):=U_\ZZ(\lie n^+)_I$. Set $$ U^\ZZ(\lie n^+)=\{ x\in U_q(\lie n^+)\,:\, \fgfrm{x}{U_\ZZ(\lie n^+)} \subset \AA_0\}. $$ Clearly, $U^\ZZ(\lie n^+)$ is an $\AA_0$-submodule of~$U_q(\lie n^+)$. \begin{lemma}\label{lem:prod in U^Z} We have $q^{\frac12(\gamma,\gamma')}xx'\in U^\ZZ(\lie n^+)$ for all $x\in U^\ZZ(\lie n^+)_\gamma$, $x'\in U^\ZZ(\lie n^+)_{\gamma'}$. In particular, all powers of a homogeneous element of~$U^\ZZ(\lie n^+)$ are in $U^\ZZ(\lie n^+)$ and $U^\AA(\lie n^+):=U^\ZZ(\lie n^+)\tensor_{\AA_0}\AA$ is an $\AA$-algebra. \end{lemma} \begin{proof} Following~\cite{Lus-book}*{\S1.2}, let $\ul\Delta:U_q(\lie n^+)\to U_q(\lie n^+)\ul\tensor U_q(\lie n^+)$ be the braided co-multiplication defined by $\ul\Delta(E_i)=E_i\tensor 1+1\tensor E_i$, where $U_q(\lie n^+)\ul\tensor U_q(\lie n^+)=U_q(\lie n^+)\tensor U_q(\lie n^+)$ endowed with an algebra structure via $(x\tensor y)(x'\tensor y')=q^{(\gamma,\gamma')}xx'\tensor yy'$ for all $x,y'\in U_q(\lie n^+)$, $y\in U_q(\lie n^+)_\gamma$, $x'\in U_q(\lie n^+)_{\gamma'}$. Then $\la xx',y\ra=\la x,\ul y_{(1)}\ra\la x',\ul y_{(2)}\ra$ and so $$ \fgfrm{xx'}{y}=q^{-\frac12(\gamma,\gamma')}\fgfrm{x}{\ul y_{(1)}}\fgfrm{x'}{\ul y_{(2)}},\qquad x\in U_q(\lie n^+)_\gamma,\, x'\in U_q(\lie n^+)_{\gamma'}, $$ where $\ul\Delta(y)=\ul y_{(1)}\tensor \ul y_{(2)}$ in Sweedler-like notation. It follows from~\cite{Lus-book}*{Lemma~1.4.2} that $\ul\Delta(U_\ZZ(\lie n^+))\subset U_\ZZ(\lie n^+)\tensor_{\AA_0} U_\ZZ(\lie n^+)$, hence we can assume that $\ul y_{(1)},\ul y_{(2)}\in U_\ZZ(\lie n^+)$ provided that $y\in U_\ZZ(\lie n^+)$. All assertions are now immediate. \end{proof} Define, for any $\gamma\in Q^+$ \begin{equation}\label{eq:def-bas} \mathbf B^{\pm up}{}_\gamma=\{ b\in U^\ZZ(\lie n^+)_\gamma\,:\,\bar b=b,\, \mu(\gamma)^{-1}\fgfrm{b}{b}\in 1+q^{-1}\ZZ[[q^{-1}]])\} \end{equation} and set $\mathbf B^{\pm up}=\bigsqcup_{\gamma\in Q^+} \mathbf B^{\pm up}{}_\gamma$. \subsection{\texorpdfstring{Signed basis is $(K_-,\mu)$}{(K-,mu)}-orthonormal}\label{subs:char sign bas} Let $R$ be a commutative unital subring of a field~$\kk$. Following~\cite{Lus-book}*{\S14.2.1}, a subset $\mathbf B^\pm$ of a free $R$-module~$L$ is a {\em signed basis} of~$L$ if $\mathbf B^\pm=\mathbf B\sqcup (-\mathbf B)$ for some basis $\mathbf B$ of~$L$. Let $K_-$ be a subring of~$\kk$ not containing~$1$. We say that $\mathbf B$ is $(K_-,\mu)$-orthonormal for some $\mu\in R^\times$ with respect to a fixed symmetric bilinear pairing $\fgfrm{\cdot}{\cdot}:L\tensor_R L\to \kk$ if $$ \mu \cdot\fgfrm{b}{b'}\in \delta_{b,b'}+K_-,\qquad b,b'\in\mathbf B. $$ Accordingly, we say that a signed basis $\mathbf B^\pm$ is $(K_-,\mu)$-orthonormal if it contains a $(K_-,\mu)$-orthonormal basis of~$L$. The following result is parallel to~\cite{Lus-book}*{Theorem 14.2.3}. \begin{theorem}\label{thm:sign-bas} $\mathbf B^{\pm up}$ is a signed basis of~$U^\ZZ(\lie n^+)$ with $R=\AA_0$. Moreover, for each $\gamma\in Q^+$, $\mathbf B^{\pm up}{}_\gamma$ is a $(K_-,\mu(\gamma)^{-1})$-orthonormal basis where~$\mu$ is defined by~\eqref{eq:defn-mu} and $K_-=q^{-1}\ZZ[[q^{-1}]]\cap \QQ(q)$. \end{theorem} \begin{proof} We need the following general setup. We say that a domain $R_0$ is {\em strongly integral} if a sum of squares of its non-zero elements is never zero and if $c_1^2+\cdots+c_n^2=1$, $c_r\in R_0$, implies that for all $1\le i\le n$, $c_i=\pm\delta_{ij}$ for some~$1\le j\le n$. Let $R$ be a domain with a subdomain $R_0$. Given a totally ordered additive monoid~$\Gamma$, a map $\nu:R\to \Gamma\sqcup\{-\infty\}$ is called an {\em $R_0$-linear valuation} if the following hold for all $f,g\in R$ \begin{enumerate}[$(V_1)$] \item\label{val-prop.1} $\nu(f)=-\infty$ if and only if $f=0$ \item\label{val-prop.2} $\nu(R_0\setminus\{0\})=0$, \item\label{val-prop.3} $\nu(fg)=\nu(f)+\nu(g)$; \item\label{val-prop.4} $\nu(f+g)\le \max(\nu(f),\nu(g))$. \end{enumerate} It follows that \begin{equation}\label{eq:diff-valuations} \nu(f)\not=\nu(g)\implies\nu(f+g)=\max(\nu(f),\nu(g)). \end{equation} Furthermore, for each $a\in\Gamma$, set $R_{\le a}=\{ r\in R\,:\, \nu(r)\le a\}$ and $R_{<a}=\{ r\in R\,:\, \nu(r)< a\}$. Clearly, $R_{\le a}$ and $R_{<a}$ are $R_0$-submodules of $R$ and $R_{<a}\subset R_{\le a}$. In the spirit of~\cite{KaKh}*{\S2.1}, we call the $R_0$-module $R_{\le a}/R_{<a}$ the {\em leaf of $\nu$ at~$a$}; we say that $\nu$ {\em has one-dimensional leaves} if for each $a\in\nu(R)$, the leaf of~$\nu$ at~$a$ is a non-zero cyclic $R_0$-module. Let $M$ be a free $R$-module with a basis~$\mathbf B$. Then we can define $\nu_{\mathbf B}:M\to \Gamma\cup\{-\infty\}$ by \begin{equation}\label{eq:pre-val} \nu_{\mathbf B}(\sum_{b\in\mathbf B}c_b b)=\max_b \nu(c_b). \end{equation} Clearly~\prpref{val-prop.4}{V} holds and we also have $\nu_{\mathbf B}(f x)=\nu(f)+\nu_\mathbf B(x)$, $f\in R$, $x\in M$. We will need the following Lemma. \begin{lemma}\label{lem:euclidean} Suppose that $\nu:R\to\Gamma\cup\{-\infty\}$ has one-dimensional leaves and let $M$ be a free $R$ module with a basis~$\mathbf B$. Then every $x\in M$ with $\nu_{\mathbf B}(x)>0$ can be written as $x=f x_0+x_1$ where $f\in R$ with $\nu(f)=\nu(x)$, $0\not=x_0\in \sum_{b\in \mathbf B} R_0 b$ and $x_1\in M$ satisfies $\nu_{\mathbf B}(x_1)<\nu_{\mathbf B}(x)$. \end{lemma} \begin{proof} Let $x\in M$ with $a=\nu_{\mathbf B}(x)>0$ and write $$ x=\sum_{b\in\mathbf B} x_b b=\sum_{b\in\mathbf B\,:\, \nu(x_b)=a} x_b b+\sum_{b\in\mathbf B\,:\, \nu(x_b)<a}x_b b. $$ Since $\nu$ has one-dimensional leaves and $R_{\le a}\not=R_{<a}$, $R_{\le a}/R_{<a}$ is a non-zero cyclic $R_0$-module. Let $f\in R_{\le a}$ be any element whose image generates~$R_{\le a}/R_{<a}$ as an $R_0$-module. Then $\nu(f)=a$ and for every $b\in\mathbf B$ with $\nu(x_b)=a$ there exists $r_b\in R_0$ such that $\nu(x_b-r_b f)<a$. Set $$ x_0=\sum_{b\in \mathbf B\,:\,\nu(x_b)=a} r_b b,\qquad x_1=x-f x_0. $$ Clearly, $\nu_{\mathbf B}(x_1)<a$, whence $x_0\not=0$. \end{proof} Henceforth \begin{itemize} \item $R_0$ is a strongly integral domain \item $\kk$ is a field containing $R_0$; \item $R_0\subset R\subset \kk$ as subrings \item $\nu:\kk\to \Gamma\cup\{-\infty\}$ is an $R_0$-linear valuation; \item $K_-$ is an $R_0$-subalgebra of $\kk$ such that $\nu(f)<0$ for all $f\in K_-$ (note that this implies that $K_-\cap R_0=\emptyset$) and $(1+K_-)^{-1}\subset 1+K_-$. \item There is a field involution $\bar\cdot$ of~$\kk$ which restricts to~$R$ and is identity on~$R_0$, while $\overline{K_-}\cap K_-=\emptyset$ \item $\nu(R^{\bar\cdot}\setminus R_0)>0$ where $R^{\bar\cdot}=\{f\in R\,:\, \bar f=f\}$; \item The restriction of $\nu$ to $R^{\bar\cdot}$ is a valuation $\nu:R^{\bar\cdot}\to\Gamma\sqcup\{-\infty\}$ with one-dimensional leaves; \end{itemize} For an $R$-module~$L$, an endomorphism of $\ZZ$-modules $\varphi:L\to L$ is called anti-linear if for all $r\in R$, $x\in L$ we have $\overline{r\cdot x}=\overline r\cdot \overline x$. Anti-linear endomorphisms of a $\kk$-vector space~$V$ are defined similarly. Let $V$ be a $\kk$-vector space with a non-degenerate symmetric bilinear form $\fgfrm\cdot\cdot$. Suppose that $\varphi$, $\varphi'$ are anti-linear involutions on~$V$ satisfying $\overline{\fgfrm xy}=\fgfrm{\varphi(x)}{\varphi'(y)}$, $x,y\in V$. Let $L$ be a free $R$-module such that $V=\kk\tensor_R L$. Denote $L^\vee=\{ x\in V\,:\, \fgfrm{x}{L}\subset R\}$. Clearly, $L^\vee$ is a free $R$-module and $V=\kk\tensor_R L^\vee$. Given $\mu\in R^\times$ define \begin{gather*} \mathbf B^\pm(\mu)=\{ b\in L\,:\,\varphi'(b)=b,\, \mu\cdot \fgfrm{b}{b}\in 1+K_-\}\\ \mathbf B^\vee_\pm(\mu)=\{ b\in L^\vee\,:\,\varphi(b)=b,\, \mu\cdot \fgfrm{b}{b}\in 1+K_-\}. \end{gather*} \begin{proposition}\label{prop:gen-dual-basis} Suppose that $\dim_\kk V<\infty$. The following are equivalent \begin{enumerate}[{\rm(a)}] \item\label{prop:gen-dual-basis.a} $\mathbf B^\pm(\mu)$ is a $(K_-,\mu)$-orthonormal signed $R$-basis of~$L$. \item\label{prop:gen-dual-basis.b} $\mathbf B^\vee_\pm(\mu^{-1})$ is a $(K_-,\mu^{-1})$-orthonormal signed $R$-basis of~$L^\vee$. \end{enumerate} In that case, $\mathbf B^\pm(\mu)$ and $\mathbf B^\vee_\pm(\mu)$ are dual to each other with respect to $\fgfrm{\cdot}{\cdot}$. \end{proposition} \begin{proof} \eqref{prop:gen-dual-basis.a}$\implies$\eqref{prop:gen-dual-basis.b} Let $\ul{\mathbf B}^\pm(\mu)$ be any basis of~$L$ contained in~$\mathbf B^\pm(\mu)$. Since $\fgfrm{\cdot}{\cdot}$ is non-degenerate, for each $b\in \ul{\mathbf B}^\pm(\mu)$ there exists a unique $\delta_b\in L^\vee$ such that $\fgfrm{\delta_b}{b'}=\delta_{b,b'}$. Clearly, the set $\ul{\mathbf B}^\pm(\mu)^\vee:=\{ \delta_b\,:\, b\in \ul{\mathbf B}^\pm(\mu)\}$ is a basis of~$L^\vee$. Note that $\varphi(\delta_b)=\delta_b$. \begin{lemma}\label{lem:scal-quad-delta func} The set $\ul{\mathbf B}^\pm(\mu)^\vee$ is $(K_-,\mu^{-1})$-orthonormal basis of $L^\vee$. In particular, $\nu(\mu^{-1}\fgfrm{\delta_b}{\delta_{b'}})\le 0$ with the equality if and only if~$b=b'$. \end{lemma} \begin{proof} We need the following \begin{lemma} Let $G=(G_{r,s})_{1\le r,s\le n}$ be a matrix over $\kk$ such that $\mu G_{rs}\in\delta_{rs}+K_-$. Then $G$ is invertible and $M=(M_{rs})_{1\le r,s\le n}=G^{-1}$ satisfies $\mu^{-1} M_{rs}\in\delta_{rs}+K_-$. \end{lemma} \begin{proof} Let $\Delta_{r,s}(G)$ be the minor of~$G$ obtained by removing the $r$th row and the $s$th column. Then it is easy to see that $\mu^{n-1}\Delta_{r,s}(G)\in \delta_{rs}+K_-$. Similarly, $\mu^n\det G\in 1+K_-$ hence $G$ is invertible. Moreover, $\mu^{-n}(\det G)^{-1}\in 1+K_-$. Since $M_{rs}=(-1)^{r+s}(\det G)^{-1}\Delta_{s,r}(G)$, the assertion follows. \end{proof} Since $\ul{\mathbf B}^\pm(\mu)$ is $(K_-,\mu)$-orthogonal, the above Lemma applies to the (finite) Gram matrix $G=(\fgfrm{b}{b'})_{b,b'\in \ul{\mathbf B}^\pm(\mu)}$ of $\fgfrm{\cdot}{\cdot}$ with respect to the basis $\ul{\mathbf B}^\pm(\mu)$ and hence $\mu^{-1} M_{b,b'}\in\delta_{b,b'}+K_-$ where $M=G^{-1}$. Since $\mu^{-1}\delta_b=\sum_{b'\in\ul{\mathbf B}^\pm(\mu)} \mu^{-1} M_{b,b'}b'$, we have $\mu^{-1} \delta_b\in b+K_-\cdot \ul{\mathbf B}^\pm(\mu)$. This implies that \begin{equation*} \mu^{-1} \fgfrm{\delta_b}{\delta_{b'}}=\fgfrm{\mu^{-1}\delta_b}{\delta_{b'}}\in \fgfrm{b}{\delta_{b'}}+K_-\fgfrm{\ul{\mathbf B}^\pm(\mu)}{\delta_{b'}} =\delta_{b,b'}+K_-,\qquad b,b'\in \ul{\mathbf B}^\pm(\mu) \end{equation*} This proves Lemma~\ref{lem:scal-quad-delta func}. \end{proof} Note that for any $x=\sum_{b} x_b \delta_b$, $y=\sum_{b'} y_{b'}\delta_{b'}$ in~$L^\vee$ we have $$ \mu^{-1}\fgfrm{x}{y}=\sum_{b,b'} x_b y_{b'}\mu^{-1}\fgfrm{\delta_b}{\delta_{b'}}. $$ Define $\nu=\nu_{\ul{\mathbf B}^\pm(\mu)^\vee}:L^\vee\to \Gamma\sqcup\{-\infty\}$ as in~\eqref{eq:pre-val}. Since $\nu(\mu^{-1}\fgfrm{\delta_b}{\delta_{b'}})\le 0$ for all $b,b'\in \ul{\mathbf B}^\pm(\mu)$ by Lemma~\ref{lem:scal-quad-delta func}, it follows from \prpref{val-prop.3}{V} and~\prpref{val-prop.4}{V} that \begin{equation}\label{eq:prod-ineq} \nu(\mu^{-1}\fgfrm{x}{y})\le \nu(x)+\nu(y),\qquad x,y\in L^\vee. \end{equation} Clearly, for $x\in L^\vee$ we have \begin{equation}\label{eq:invar-form} \varphi(x)=x\iff x\in \sum_{b\in\ul{\mathbf B}^\pm(\mu)} R^{\bar\cdot}\delta_b. \end{equation} Thus, the set $(L^\vee)^\varphi$ of $\varphi$-invariant elements in~$L^\vee$ is a free $R^{\bar\cdot}$-module with a basis $\ul{\mathbf B}^\pm(\mu)^\vee$. The following Lemma is the crucial point of our argument. \begin{lemma}\label{lem:key Lusztig's lemma} Let $x\in L^\vee$ and suppose that $\varphi(x)=x$. Then \begin{enumerate}[{\rm(a)}] \item\label{lem:key Lusztig's lemma.c} If $\nu(x)=0$, that is, $x=\sum_{b} x_b \delta_b$ with $x_b\in R_0$, then $\mu^{-1}\fgfrm{x}{x}-\sum_{b} x_b^2\in K_-$ and $\nu(\mu^{-1}\fgfrm{x}{x})=0$. \item\label{lem:key Lusztig's lemma.b} If $\nu(x)>0$ then $\nu(\mu^{-1}\fgfrm{x}{x})>0$ \end{enumerate} \end{lemma} \begin{proof} Write $x=\sum_{b} x_b \delta_b$, $x_b\in R^{\bar\cdot}$. To prove~\eqref{lem:key Lusztig's lemma.c}, note that $\nu(x)=0$ and $\varphi(x)=x$ imply that $x_b\in R_0$ for all $b\in\ul{\mathbf B}^\pm(\mu)$. We have $$ \mu^{-1}\fgfrm{x}{x}=\sum_{b} x_b^2 \mu^{-1}\fgfrm{\delta_b}{\delta_b}+\sum_{b\not=b'} x_bx_{b'}\mu^{-1}\fgfrm{\delta_b}{\delta_{b'}}. $$ By Lemma~\ref{lem:scal-quad-delta func}, the first sum belongs to $\sum_{b}x_b^2+K_-$ while the second sum belongs to $K_-$. Since $R_0$ is strongly integral, $\sum_{b} x_b^2\not=0$. Thus, $\nu(\mu^{-1}\fgfrm{x}{x})=0$. To prove~\eqref{lem:key Lusztig's lemma.b}, let $a=\nu(x)>0$. Applying Lemma~\ref{lem:euclidean} to $M=(L^\vee)^{\varphi}$ and the ring~$R^{\bar\cdot}$, we can write $x=f x_0+x_1$ where $f\in R^{\bar\cdot}$, $\nu(f)=a$, $\varphi(x_0)=x_0$ (and so $\varphi(x_1)=x_1$), $\nu(x_0)=0$ and $\nu(x_1)<a$. Then $$ \nu(\mu^{-1}\fgfrm{x}{x})=\nu(f^2 \mu^{-1}\fgfrm{x_0}{x_0}+2 f\mu^{-1}\fgfrm{x_0}{x_1}+\mu^{-1}\fgfrm{x_1}{x_1}) =\nu(f^2\mu^{-1}\fgfrm{x_0}{x_0})=2a>0 $$ since $\nu(\mu^{-1}\fgfrm{x_1}{x_1}),\nu(f\mu^{-1}\fgfrm{x_0}{x_1})<2 a$ by~\eqref{eq:prod-ineq} and \prpref{val-prop.3}{V}, \prpref{val-prop.4}{V} while $\nu(\mu^{-1}\fgfrm{x_0}{x_0})=0$ by part~\eqref{lem:key Lusztig's lemma.c}. This proves~\eqref{lem:key Lusztig's lemma.b}. \end{proof} It follows from Lemma~\ref{lem:key Lusztig's lemma}(\ref{lem:key Lusztig's lemma.c},\ref{lem:key Lusztig's lemma.b}) that if $x\in L^\vee$ is fixed by~$\varphi$ and $\mu^{-1}\fgfrm{x}{x}\in 1+K_-$ then $x=\pm\delta_b$ for some $b\in \ul{\mathbf B}^\pm(\mu)$ by the strong integrality of~$R_0$. Thus, $\mathbf B^\vee_\pm(\mu)=\ul{\mathbf B}^\pm(\mu)^\vee\bigsqcup(-\ul{\mathbf B}^\pm(\mu)^\vee)$. This completes the proof of the implication \eqref{prop:gen-dual-basis.a}$\implies$\eqref{prop:gen-dual-basis.b} and the last assertion. The opposite implication follows by the symmetry between $L$ and $L^\vee$ and $\varphi$ and $\varphi'$. \end{proof} We now apply Proposition~\ref{prop:gen-dual-basis} with $L=U_\ZZ(\lie n^+)_\gamma$, $v=q^{\frac12}$, $\varphi=\bar\cdot$, $\varphi'=\tilde\cdot$, $R=\AA_0=\ZZ[v^2,v^{-2}]$, $\kk=\QQ(v)$ and $K_-=v^{-2}\ZZ[[v^{-2}]]\cap \QQ(v)$. We define $\nu:\QQ(v)\to\ZZ\cup\{-\infty\}$ via $$ \nu\Big(c v^n\frac{1+f}{1+g}\Big)=n $$ where $c\in\QQ^\times$, $n\in\ZZ$ and $f,v\in v^{-1}\ZZ[v^{-1}]$. Note that $R^{\bar\cdot}=\ZZ[q+q^{-1}]$ and $\nu$ has one-dimensional leaves on~$R^{\bar\cdot}$ since $\nu((v+v^{-1})^n)=n$. By \cite{Lus-book}*{Theorem~14.2.3}, $\mathbf B^{\pm can}\cap U_\ZZ(\lie n^+)$ is a $(K_-,\mu(\gamma))$-orthonormal signed basis of $L$. Since $\mathbf B^{\pm up}{}_\gamma=(\mathbf B^{\can}\cap U_\ZZ(\lie n^+))^\vee_\pm$ in the notation of Proposition~\ref{prop:gen-dual-basis}, it is a signed $(K_-,\mu(\gamma)^{-1})$-orthonormal basis of $L^\vee=U^\ZZ(\lie n^+)_\gamma$. This completes the proof of Theorem~\ref{thm:sign-bas}. \end{proof} \subsection{Choosing~\texorpdfstring{$\mathbf B^{up}$}{Bup} inside the signed basis}\label{subs:chos basis} It remains to describe a canonical way to choose $\mathbf B^{up}$ inside~$\mathbf B^{\pm up}$. Needless to say, it can be taken as the dual basis of $\mathbf B^{can}$ with respect to $\fgfrm{\cdot}{\cdot}$. However, it more instructive to provide an intrinsic definition. To that effect, following~\cite{BG-dcb}*{\S3.5} and also~\cite{Lus-book}*{Proposition~3.1.6}, define $\kk$-linear endomorphisms $\partial_i,\partial_i^{op}$, $i\in I$ of~$U_q(\lie n^+)$ by \begin{equation}\label{eq:partial-def} [F_i,x ]=(q_i-q_i^{-1})( q^{-\frac12(\alpha_i,\gamma-\alpha_i)} K_i\partial_i(x )-q^{\frac12(\alpha_i,\gamma-\alpha_i)} K_i^{-1}\partial_i^{op}(x )),\qquad x \in U_q(\lie n^+)_\gamma. \end{equation} We need the following properties of these operators (cf.~\cite{BG-dcb}*{Lemmata~3.18 and~3.20}). \begin{lemma}\label{lem:partial-prop} For all $x \in U_q(\lie n^+)_\gamma$ and $i\in I$ we have \begin{enumerate}[{\rm(a)}] \item \label{lem:partial-prop.i} $\overline{\partial_i(x )}=\partial_i(\overline{x })$, $\overline{\partial_i^{op}(x )}=\partial_i^{op}(\overline{ x })$, $\partial_i(x{}^*)^*=\partial_i^{op}(x)$ and $\partial_i\partial_i^{op}(x)=\partial_i^{op}\partial_i(x)$. \item \label{lem:partial-prop.ii} for all $y\in U_q(\lie n^+)$, $n\in\ZZ_{\ge 0}$ $$ \fgfrm{x}{yE_i^{\la n\ra}}=\fgfrm{\partial_i^{(n)}(x)}{y},\qquad \fgfrm{x}{E_i^{\la n\ra}y}=\fgfrm{(\partial_i^{op})^{(n)}(x )}{y}, $$ where $f_i^{(n)}=(q_i-q_i^{-1})^n f_i^{\la n\ra}$. \item \label{lem:partial-prop.iii} $\partial_i$, $\partial_i^{op}$ are quasi-derivations. Namely, for $x\in U_q(\lie n^+)_\gamma$, $y \in U_q(\lie n^+)_{\gamma'}$ we have \begin{equation}\label{eq:partial-inv} \begin{gathered} \partial_i(x y )=q^{\frac12(\alpha_i,\gamma' )}\partial_i(x )y + q^{-\frac12(\alpha_i,\gamma )}x \partial_i(y ), \\ \partial_i^{op}(x y )=q^{-\frac12(\alpha_i,\gamma' )}\partial_i^{op}(x )y +q^{\frac12(\alpha_i,\gamma )}x \partial_i^{op}(y ). \end{gathered} \end{equation} \end{enumerate} \end{lemma} It is easy to see that \begin{equation}\label{eq:partial E_i^r} \partial_i^{(n)}(E_i^r)=\binom{r}{n}_{q_i} E_i^{r-n}=(\partial_i^{op})^{(n)}(E_i^r) \end{equation} whence \begin{equation}\label{eq:q-der preserve Lusztig lattice} ((q_i-q_i^{-1})\partial_i)^n(E_i^{\la r\ra})=E_i^{\la r-n\ra}=((q_i-q_i^{-1})\partial_i^{op})^n(E_i^{\la r\ra}) \end{equation} The following is an immediate consequence of this identity and Lemma~\ref{lem:partial-prop}. \begin{corollary}\label{cor:preserv-lattice} For all $i\in I$ we have \begin{enumerate}[\rm(a)] \item\label{cor:preserv-lattice.a} if $x\in U_\ZZ(\lie n^+)_\gamma$ then $\la 1\ra_{q_i}\partial_i(x),\la 1\ra_{q_i}\partial_i^{op}(x)\in q^{\frac12(\alpha_i,\gamma)}U_\ZZ(\lie n^+)$; \item\label{cor:preserve-lattice.b} the $\partial_i^{(n)}$, $(\partial_i^{op})^{(n)}$, $n\in \ZZ_{\ge0}$ restrict to operators on $U^\ZZ(\lie n^+)$. \end{enumerate} \end{corollary} By degree considerations it is clear that $\partial_i$, $\partial_i^{op}$ are locally nilpotent, that is, for any $x\in U_q(\lie n^+)$ we have $\partial_i^k(x)=(\partial_i^{op})^k(x)=0$ for $k\gg0$. Thus, for each $x\in U_q(\lie n^+)\setminus\{0\}$ we can define $\ell_i(x)$ as the maximal $k> 0$ such that $\partial_i^k(x)\not=0$. Define $\partial_i^{(top)},(\partial_i^{op})^{(top)}:U_q(\lie n^+)\setminus\{0\}\to U_q(\lie n^+)\setminus\{0\}$ by $\partial_i^{(top)}(x)=\partial_i^{(\ell_i(x))}(x)$ and $(\partial_i^{op})^{(top)}(x)=(\partial_i^{(top)}(x^*))^*=(\partial_i^{op})^{(\ell_i(x^*))}(x)$. Similar notation will be used for other locally nilpotent operators in the sequel. For any sequence $\mathbf i=(i_1,\dots,i_m)\in I^m$ set $\partial_{\mathbf i}^{(top)}=\partial_{i_m}^{(top)}\cdots \partial_{i_1}^{(top)}$. \begin{proposition}\label{prop:string char} For every $b\in \mathbf B^{\pm up}$ there exists $\ii=(i_1,\dots,i_m)$ such that $\partial_\ii^{(top)}(b)\in\{\pm 1\}$. Moreover, if $\ii'=(i'_1,\dots,i'_{m'})$ also satisfies $\partial_{\ii'}^{(top)}(b)\in\{\pm 1\}$ then $\partial_\ii^{(top)}(b)=\partial_{\ii'}^{(top)}(b)\in\{\pm 1\}$. \end{proposition} Thus, we can define $\mathbf B^{up}$ to be the set of all $b\in \mathbf B^{\pm up}$ such that $\partial_\ii^{(top)}(b)=1$ for some $\ii=(i_1,\dots,i_m)$. \begin{proof} By Proposition~\ref{prop:gen-dual-basis}, $\mathbf B^{\pm up}$ contains the dual basis $\mathbf B'$ of~$\mathbf B^{\can}$. Our goal is to prove that $\mathbf B^{up}=\mathbf B'$. We need the following result. \begin{lemma}\label{lem:tops} $\partial_i^{(top)}(b)\in\mathbf B'$ for all $b\in \mathbf B'$, $i\in I$. Moreover, if $\partial_i^{(top)}(b)=\partial_i^{(top)}(b')$ and $\ell_i(b)=\ell_i(b')$ for some $b'\in\mathbf B'$ then $b=b'$. \end{lemma} \begin{proof} Following~\cite{Lus-book}*{\S14.3}, denote $\mathbf B^{\can}_{i;\ge r}=\mathbf B^{\can}\cap E_i^r U_q(\lie n^+)$ and $\mathbf B^{\can}_{i;r}=\mathbf B^{\can}_{i;\ge r}\setminus \mathbf B^{\can}_{i;\ge r+1}$. It follows from~\cite{Lus-book}*{\S14.3} that for all~$i\in I$, \begin{equation}\label{eq:Lus-decomp} \mathbf B^{\can}=\bigsqcup_{r\ge 0} \mathbf B^{\can}_{i;r}. \end{equation} Let $b\in\mathbf B^{\can}$ and let $n=\ell_i(\delta_b)$, $u=\partial_i^{(top)}(\delta_b)= \partial_i^{(n)}(\delta_b)$, where $\delta_b$ is the element of~$\mathbf B'$ satisfying $\fgfrm{\delta_b}{b'}=\delta_{b,b'}$. Then $u\in\ker \partial_i$ which, by Lemma~\ref{lem:partial-prop}\eqref{lem:partial-prop.iii}, is orthogonal to $\mathbf B^{\can}_{i;s}$, $s>0$. Thus, we can write $$ u=\sum_{b'\in \mathbf B^{\can}_{i;0}} \fgfrm{u}{b'} \delta_{b'}=\sum_{b'\in \mathbf B^{\can}_{i;0}} \fgfrm{\delta_b}{E_i^{\la n\ra}b'}\delta_{b'}. $$ By~\cite{Lus-book}*{Theorem~14.3.2}, for each $b'\in \mathbf B^{\can}_{i;0}$ there exists a unique $\pi_{i,n}(b')\in\mathbf B^{\can}_{i;n}$ such that $E_i^{\la n\ra} b'-\pi_{i;n}(b')\in\sum_{r>n} \ZZ[q,q^{-1}]\mathbf B^{\can}_{i;r}$. Using Lemma~\ref{lem:partial-prop}\eqref{lem:partial-prop.iii} again, we conclude that for any $b''\in\mathbf B^{\can}_{i;r}$ with~$r>n$, $\fgfrm{ \delta_b}{b''}\in \fgfrm{\delta_b}{E_i^{\la r\ra}U_q(\lie n^+)}=\fgfrm{\partial_i^{(r)}(\delta_b)}{U_q(\lie n^+)}=0$. Thus, $$ u=\sum_{b'\in \mathbf B^{\can}_{i;0}} \fgfrm{\delta_b}{\pi_{i;n}(b')}\delta_{b'}. $$ Note that, since~$u\not=0$, we cannot have $\fgfrm{\delta_b}{\pi_{i;n}(b')}=0$ for all~$b'\in\mathbf B^{\can}_{i;0}$. Since~$\fgfrm{\delta_b}{b''}=\delta_{b,b''}$, we conclude that there exists a unique $b'\in\mathbf B^{\can}_{i;0}$ such that $\pi_{i;n}(b')=b$ and then $u=\partial_i^{(top)}(\delta_b)=\delta_{b'}$. Since $\pi_{i;n}:\mathbf B^{\can}_{i;0}\to\mathbf B^{\can}_{i;n}$ is a bijection by~\cite{Lus-book}*{Theorem~14.3.2}, the first assertion follows. The second assertion is immediate from~\eqref{eq:Lus-decomp}. \end{proof} This implies that for every element $b\in \mathbf B'$, there exists $\ii=(i_1,\dots,i_m)$ such that $\partial_\ii^{(top)}(b)=1$. Since $1$ is the unique element of $\mathbf B'$ of degree~$0$, for any sequence $\ii'$ such that $\partial_{\ii'}^{(top)}(b)=c\in\kk^\times$, one has $c=1$. This completes the proof of Proposition~\ref{prop:string char}. \end{proof} \begin{remark}\label{rem: star op partial} Since $\mathbf B^{can}$ is preserved by~${}^*$ by~\cite{Lus-book}*{Theorem~14.4.3} and ${}^*$ is self-adjoint with respect to $\fgfrm{\cdot}{\cdot}$, it follows that $\mathbf B^{up}$ is preserved by~${}^*$. In particular, we can replace $\partial_i$ by $\partial_i^{op}$ in Lemma~\ref{lem:tops} and Proposition~\ref{prop:string char}. \end{remark} Note that Lemma~\ref{lem:tops} and Remark~\ref{rem: star op partial} immediately yield the following well-known fact. \begin{corollary}\label{cor:top-decomp-B up} Let $x\in U_q(\lie n^+)$ and write $x=\sum_{b\in\mathbf B^{up}} c_b(x) b$. Then $c_b(x)\not=0$ implies that $\ell_i(b)\le \ell_i(x)$ and $\ell_i(b^*)\le \ell_i(x^*)$ and $$ \partial_i^{(top)}(x)=\sum_{b\in\mathbf B^{up}\,:\,\ell_i(b)=\ell_i(x)} c_b(x) \partial_i^{(top)}(b), \quad (\partial_i^{op})^{(top)}(x)=\sum\limits_{b\in\mathbf B^{up}\,:\,\ell_i(b^*)=\ell_i(x^*)} c_b(x) (\partial_i^{op})^{(top)}(b) $$ are the decompositions of~$(\partial_i)^{(top)}(x)$ and $(\partial_i^{op})^{(top)}(x)$, respectively, in the basis~$\mathbf B^{up}$. \end{corollary} \section{Decorated algebras and proof of Theorem~\ref{thm:T_i bas}} \subsection{Decorated algebras}\label{subs:decorated alg} Let $\mathcal A$ be an associative $\ZZ$-graded algebra over~$\kk=\mathbb Q(v^{\frac12})$. Denote the degree of a homogeneous $u\in\mathcal A$ by~$|u|$. \begin{definition}\label{defn:decorated alg} We say that $\mathcal A=\mathcal A(E,\ul F_+,\ul F_-)$ is {\em decorated} if it contains an element $E$ with $|E|=2$ and admits mutually commuting locally nilpotent $\kk$-linear endomorphisms $\ul F_+,\ul F_-:\mathcal A\to \mathcal A$ of degree~$-2$ satisfying $\ul F_\pm(E)=1$ and \begin{equation}\label{eq:skew Leibnitz rule} \ul F_\pm(xy)=v^{\pm\frac12 |y|}\ul F_\pm(x)y+v^{\mp\frac12|x|}x\ul F_\pm(y) \end{equation} for $x,y\in\mathcal A$ homogeneous. \end{definition} Denote $\mathcal A_\pm=\ker D_\mp$ and set $\mathcal A_0=\mathcal A_+\cap\mathcal A_-$. Clearly, $\ul F_\pm$ restricts to an endomorphism of $\mathcal A_\pm$ which will also be denoted by $\ul F_\pm$. Since $F_\pm$ are skew derivations, $\mathcal A_\pm$ are subalgebras of~$\mathcal A$. The following is a basic example of a decorated algebra. Let $\mathcal F_{m,n}=\kk\la E,x,y\ra$ with the $\ZZ$-grading defined by $|x|=-m$, $|y|=-n$, $|E|=2$. The following is immediate \begin{lemma}\label{lem:free subalgebra} There exists unique operators $\ul F_\pm\in\End_\kk \mathcal F_{m,n}$ such that $\ul F_\pm(E)=1$, $\ul F_\pm(x)=\ul F_\pm(y)=0$ and~\eqref{eq:skew Leibnitz rule} holds. In particular, $\mathcal F_{m,n}$ is a decorated algebra and for any decorated algebra $\mathcal A$ and any $x',y'\in\mathcal A_0$ homogeneous the natural homomorphism of graded associative algebras $\phi_{|x'|,|y'|}:\mathcal F_{|x'|,|y'|}\to \mathcal A$, $x\mapsto x'$, $y\mapsto y'$ is a homomorphism of decorated algebras, that is, it commutes with $\ul F_\pm$. \end{lemma} Define $\ul K^{\frac12},\ul E_\pm\in\End_\kk \mathcal A$ by $$ \ul K^{\frac12}(x)=v^{\frac12|x|} u,\quad \ul E_\pm(x)=\pm\la 1\ra_{v}^{-1} (v^{\pm\frac12 |x|}Ex-v^{\mp\frac12|x|}xE), $$ for $x\in\mathcal A$ homogeneous. Clearly $\ul E_\pm$ are of degree~$2$. The following is easily checked. \begin{lemma} \begin{enumerate}[{\rm(a)}] \item\label{lem:algebra A sl2 action.a} For $x,y\in\mathcal A$ homogeneous we have \begin{gather}\label{eq:divided power E_i} \begin{aligned} &\ul E_+^{(r)}(x)=\sum_{r'+r''=r}(-1)^{r'}v^{\frac12(r+|x|-1)(r'-r'')}E^{\la r'\ra}x E^{\la r''\ra}\\ &\ul E_-^{(r)}(x)=\sum_{r'+r''=r}(-1)^{r''}v^{\frac12(r+|x|-1)(r''-r')}E^{\la r'\ra}x E^{\la r''\ra} \end{aligned} \\ \label{eq:powers E_i action} \ul E_\pm^{(r)}(xy)=\sum_{r'+r''=r} v^{\pm\frac12(r'|y|-r''|x|)} \ul E_\pm^{(r')}(x)\ul E_\pm^{(r'')}(y). \end{gather} and $[\ul E_\pm,\ul F_\pm]=\la 1\ra_v{}^{-1}(\ul K-\ul K^{-1})$. In particular, $\ul E_\pm$, $\ul F_\pm$ and $\ul K$ provide actions of Chevalley generators of $U_v(\lie{sl}_2)$ in its {\em standard} presentation on~$\mathcal A$; \item\label{lem:algebra A sl2 action.b} $\ul E_\pm$ restrict to endomorphisms of $\mathcal A_\pm$. In particular, $\mathcal A_\pm$ is a $U_v(\lie{sl}_2)_\pm$-submodule of~$\mathcal A$ and $\mathcal A_0$ is the space of lowest weight vectors for both actions. \item\label{lem:algebra A sl2 action.c} A homomorphism of decorated algebras $\mathcal A\to \mathcal A'$ is a homomorphism of $U_v(\lie{sl}_2)_+$- and $U_v(\lie{sl}_2)_-$-modules. \end{enumerate} \label{lem:algebra A sl2 action} \end{lemma} \begin{remark} Suppose that~$y\in\mathcal A_0$. The following is rather standard an is an obvious consequence of say~\cite{Lus-book}*{Corollary~3.1.9}. \begin{equation}\label{eq:lowest weight action} \ul F_\pm^{(a)}\ul E_\pm^{(b)}(y)=\begin{cases}\displaystyle\binom{a-b-|y|}{a}_v \ul E_\pm^{(b-a)}(y),& 0\le a\le b\\ 0,& a>b \end{cases} \end{equation} \end{remark} Suppose that $\ul E_\pm$ are locally nilpotent on~$\mathcal A_\pm$. Then $\mathcal A_\pm$ are direct sums of finite dimensional $U_v(\lie{sl}_2)_\pm$-modules and if $x\in\mathcal A_0$ is homogeneous then $|x|\le 0$. We need following \begin{lemma} \begin{enumerate}[{\rm(a)}] \item\label{lem:sigma eta defn.a} There exists unique isomorphisms of~$U_v(\lie{sl}_2)$-modules $\sigma_\pm:\mathcal A_\pm\to \mathcal A_\mp$ such that $\sigma_\pm|_{\mathcal A_0}=\id_{\mathcal A_0}$, where $\mathcal A_\pm$ is regarded as a $U_v(\lie{sl}_2)_\pm$-module. In particular, $\sigma_\pm\circ \sigma_\mp=\id_{\mathcal A_\mp}$. \item\label{lem:sigma eta defn.b} There exists unique $\kk$-linear involution $\eta_\pm:\mathcal A_\pm\to\mathcal A_\pm$ such that \begin{equation}\label{eq:intertwiner} \eta_\pm\circ\ul E_\pm=\ul F_\pm\circ\eta_\pm,\quad \eta_\pm\circ \ul F_\pm=\ul E_\pm\circ\eta_\pm,\quad \eta_\pm\circ \ul K=\ul K^{-1}\circ\eta_\pm \end{equation} and $\eta_\pm(x)=\ul E_\pm^{(top)}(x)=\ul E_\pm^{(-|x|)}(x)$ for $x\in\mathcal A_0$ homogeneous. \end{enumerate} \label{lem:sigma eta defn} \end{lemma} \begin{proof} Part~\eqref{lem:sigma eta defn.a} is immediate from the semi-simplicity of~$\mathcal A_\pm$ as $U_v(\lie{sl}_2)_\pm$-modules and the fact that any endomorphism of any lowest weight $U_v(\lie{sl}_2)$-module fixing all lowest weight vectors is identity on that module. To prove~\eqref{lem:sigma eta defn.b} recall that every simple finite dimensional $U_v(\lie{sl}_2)$-module $V_\lambda$ of type~$1$ has a basis $\{z_k\}_{0\le k\le}$ such that $\ul E(x_k)=(k)_v z_{k-1}$, $\ul F(z_k)=(\lambda-k)_v z_{k+1}$, $\ul K(z_k)=v^{\lambda-2k}z_k$. Then it is easy to see that $\eta_\lambda\in\End_\kk V_\lambda$ defined by $\eta(z_k)=z_{\lambda-k}$ is the unique linear map satisfying~\eqref{eq:intertwiner} and such that $\eta_\lambda(z)=\ul E^{(\lambda)}(z)$ for any lowest weight vector~$z$ of~$V_\lambda$. It remains to observe that $\eta_\lambda$ can be extended uniquely to any semi-simple $U_v(\lie{sl}_2)$-module. \end{proof} \subsection{An isomorphism between \texorpdfstring{$\mathcal A_-$}{A-} and~\texorpdfstring{$\mathcal A_+$}{A+}} The following is quite surprising. \begin{theorem}\label{thm:tau-homomorphism} Let $\mathcal A$ be a decorated algebra such that the operators $\ul E_\pm$ are locally nilpotent on~$\mathcal A_\pm$. Then the map $\tau:=\eta_+\circ\sigma_-:\mathcal A_-\to \mathcal A_+$ is an isomorphism of algebras. \end{theorem} \begin{proof} Let $x,y\in\mathcal A_0$ be homogeneous and let $m=-|x|$, $n=-|y|$. For $r\ge 0$, define $x*_r y\in\mathcal A$ by $$ x*_r y= \sum_{\substack{r',r''\ge 0\\ r'+r''\le r}} (-1)^{r'+r''} \prod_{t=1}^{r'} (n-r+t)_{v} \prod_{t=1}^{r''}(m-r+t)_v\prod_{t=r'+r''+1}^{r}\mskip-7mu (m+n-2r+t+1)_v E^{\la r'\ra}x E^{\la r-r'-r''\ra}y E^{\la r''\ra}. $$ Clearly $x*_r y$ is homogeneous of degree $2r-m-n$. \begin{proposition}\label{prop:common lowest vector} Let $\mathcal A$ be a decorated algebra and let $x,y\in\mathcal A_0$ be homogeneous with $|x|=-m$, $|y|=-n$. For all $r\ge 0$ we have $x*_r y\in \mathcal A_0$ and \begin{equation}\label{eq:two expressions for x*_r y} \begin{aligned} x*_r y&=\sum_{t'+t''=r}(-1)^{t''}v^{\frac12(m t'-n t''+(r-1)(t''-t'))} \frac{ (m-t')_v!(n-t'')_v!}{(n-r)_v!(m-r)_v!}\ul E_+^{(t')}(x)\ul E_+^{(t'')}(y) \\&=\sum_{t'+t''=r}(-1)^{t'}v^{\frac12(nt''-m t'+(r-1)(t'-t''))} \frac{ (m-t')_v!(n-t'')_v!}{(n-r)_v!(m-r)_v!}\ul E_-^{(t')}(x)\ul E_-^{(t'')}(y) \end{aligned} \end{equation} \end{proposition} \begin{proof} By Lemma~\ref{lem:free subalgebra} it suffices to prove the proposition for the decorated algebra $\mathcal F_{m,n}$. Let $\mathcal V_{m,n}$ be the subspace of~$\mathcal F_{m,n}$ with the basis $\{ E^{\la a\ra} x E^{\la b\ra} y E^{\la c\ra}\,:\,a,b,c\in\ZZ_{\ge 0}\}$. Clearly, $\ul E_\pm(\mathcal V_{m,n})$, $\ul F_\pm(\mathcal V_{m,n})\subset\mathcal V_{m,n}$ and $x*_r y\in \mathcal V_{m,n}$. We need the following \begin{lemma}\label{lem:free kernels} $(\mathcal F_{m,n})_0\cap \mathcal V_{m,n}$ is spanned by the $x*_r y$, $r\ge 0$ as a $\kk$-vector space. \end{lemma} \begin{proof} It is easy to check that $x*_r y\in (\mathcal F_{m,n})_0$. Conversely, let $u\in(\mathcal F_{m,n})_0\cap \mathcal V_{m,n}$ be homogeneous of degree $2r-m-n$ and write $$ u=\sum_{r',r''\ge 0,r'+r''\le r} c_{r',r''} E^{\la r'\ra}x E^{\la r-r'-r''\ra}y E^{\la r''\ra}. $$ Then \begin{align*} \la 1\ra_v\ul F_\pm(u)&=\sum_{r'\ge 1,\,r'\ge 0,\,r'+r''\le r} c_{r',r''} v^{\pm\frac12(|x|+|y|+2(r-r'))} E^{\la r'-1\ra}x E^{\la r-r'-r''\ra}y E^{\la r''\ra} \\&\quad+\sum_{r',r''\ge0,r'+r''\le r-1} c_{r',r''}v^{\pm\frac12(|y|-|x|+2(r''-r'))}E^{\la r'\ra}x E^{\la r-r'-r''-1\ra}y E^{\la r''\ra} \\&\quad+ \sum_{r'\ge0,\,r''\ge 1,\,r'+r''\le r} c_{r',r''}v^{\mp\frac12(|x|+|y|+2(r-r''))}E^{\la r'\ra}x E^{\la r-r'-r''\ra}y E^{\la r''-1\ra} \\ &=\sum_{r',r''\ge 0,\,r'+r''\le r-1} (c_{r'+1,r''}v^{\pm\frac12(|x|+|y|+2(r-r'-1))}+c_{r',r''}v^{\pm\frac12(|y|-|x|+2(r''-r'))}\\&\quad+c_{r',r''+1}v^{\mp\frac12(|x|+|y|+2(r-r''-1))}) E^{\la r'\ra}x E^{\la r-r'-r''\ra}y E^{\la r''\ra}. \end{align*} It follows that $$ c_{r'+1,r''}\la |x|+|y|+2r-r'-r''-2\ra_v +c_{r',r''}\la |y|-r'+r-1\ra_v=0 $$ and $$ c_{r',r''+1}\la |x|+|y|+2r-r'-r''-2\ra_v +c_{r',r''}\la |x|-r''+r-1\ra_v=0. $$ Thus, \begin{multline*} c_{r',r''} =(-1)^{r'} \frac{\prod_{t=1}^{r'} (n-r+t)_v}{\prod_{t=1}^{r'} (m+n-2r+r''+t)_v} c_{0,r''}\\=(-1)^{r'+r''} \frac{ \prod_{t=1}^{r'}(n-r+t)_v\prod_{t=1}^{r''}(m-r+t)_v}{\prod_{t=1}^{r'+r''} (m+n-2r+t+1)_v} c_{0,0}. \end{multline*} Therefore, $u=\prod_{t=1}^r (m+n-2r-t+1)_v{}^{-1} c_{0,0} x*_r y$. \end{proof} Thus, $x*_ry\in \mathcal A_0$. Furthermore, it is easy to check, using~\eqref{eq:skew Leibnitz rule}, that right hand sides of~\eqref{eq:two expressions for x*_r y} are in $(\mathcal F_{m,n})_0\cap \mathcal V_{m,n}$ and hence proportional to $x*_r y$ by Lemma~\ref{lem:free kernels}. It remains then to compare the coefficient of $E^{\la r\ra} x y$ in both expressions, which is easily calculated using~\eqref{eq:divided power E_i}. \end{proof} \begin{remark}\label{rem:Verma} Clearly, there exist unique injective homomorphisms $j_{\pm,m}$ and $j_{\pm,n}$ from the lowest weight Verma $U_v(\lie{sl}_2)_\pm$-modules $M^\pm_{-m}$, $M^\pm_{-n}$ of lowest weight $-m$ (respectively, $-n$) to $\mathcal V_{m,n}$ sending a fixed lowest weight vector to $x$ (respectively, to $y$). This yields natural injective homomorphisms of $U_v(\lie{sl}_2)_\pm$-modules $j_{\pm,m,n}:M^\pm_{-m}\tensor M^\pm_{-n}\to \mathcal V_{m,n}$ where the comultiplication on~$U_v(\lie{sl}_2)_\pm$ is defined by $\Delta_\pm(\ul E_\pm)=\ul E_\pm\tensor \ul K^{\pm\frac12}+\ul K^{\pm\frac12}\tensor \ul E_\pm$. In particular, Proposition~\ref{prop:common lowest vector} implies that $j_{\pm,m,n}(M^\pm_{-m}\tensor M^{\pm}_{-n})$ share lowest weight vectors of weight $-m-n+2r$. \end{remark} The following Lemma is essentially concerned with quantum Clebsch-Gordan coefficients (also known as $3j$-symbols, see e.g.~\cite{Kassel}*{Chapter VII}). \begin{lemma} \begin{enumerate}[{\rm(a)}] Let $\mathcal A$, $x,y\in\mathcal A_0$ be as in Proposition~\ref{prop:common lowest vector}. \item For $r\le \min(m,n)$ we have \begin{equation}\label{eq:CG-defn} \begin{aligned} &\ul E_+^{(a)}(x*_r y)=\sum_{t'+t''=a+r} C_{r;t',t''}(v)\ul E_+^{(t')}(x)\ul E_+^{(t'')}(y),\\ &\ul E_-^{(a)}(x*_r y)=\sum_{t'+t''=a+r} (-1)^r C_{r;t',t''}(v^{-1})\ul E_-^{(t')}(x)\ul E_-^{(t'')}(y) \end{aligned} \end{equation} where $$ C_{r;t',t''}(v)=v^{\frac12(m t''-n t')} \sum_{k+l=r} (-1)^l v^{l t'-k t''+\frac12(k-l)(1+m+n-r)} \frac{ (n-l)_v! (m-k)_v!}{(n-r)_v!(m-r)_v!}\binom{t'}{k}_v\binom{t''}{l}_v\in\ZZ[v,v^{-1}]. $$ \item\label{lem:Clebsch-Gordan.b} The $C_{r;t',t''}(v)$ satisfy the following recurrence relations \begin{gather} \label{eq:recurr-relation for CG} \begin{split} (m+n-r-t'-t'')_v C_{r;t',t''}(v)&=v^{t''-\frac12 n}(m-t')_v C_{r;t'+1,t''}(v)\\&\qquad\qquad+v^{\frac12 m-t'}(n-t'')_v C_{r;t',t''+1}(v), \end{split}\\ \label{eq:recurr-relation for CG.2} (t'+t''-r)_v C_{r;t',t''}(v)=v^{t''-\frac12n}(t')_v C_{r;t'-1,t''}(v)+v^{\frac12 m-t'}(t'')_v C_{r;t',t''-1}(v). \end{gather} \item\label{lem:Clebsch-Gordan.c} For all $0\le t'\le m$, $0\le t''\le n$, $0\le r\le \min(m,n)$ we have \begin{equation}\label{eq:key-combinatorial-identity} C_{r;m-t',n-t''}(v)=(-1)^r C_{r;t',t''}(v^{-1}). \end{equation} \end{enumerate} \label{lem:Clebsch-Gordan} \end{lemma} \begin{proof} Let $a=0$. Then $$ C_{r;t',t''}(v)=(-1)^{t''}v^{\frac12(m t'-n t''+(r-1)(t''-t'))} \frac{ (m-t')_v!(n-t'')_v!}{(n-r)_v!(m-r)_v!} $$ and the assertion follows from~\eqref{eq:two expressions for x*_r y}. The case of arbitrary $a$ is then easily deduced by applying $\ul E_\pm^{(a)}$ to~\eqref{eq:two expressions for x*_r y} and using Lemma~\ref{lem:algebra A sl2 action}\eqref{lem:algebra A sl2 action.a}. To prove~\eqref{eq:recurr-relation for CG} (respectively, \eqref{eq:recurr-relation for CG.2}) it suffices to apply $\ul F$ (respectively, $\ul E$) to both sides of the first identity in~\eqref{eq:CG-defn}. We leave the details of these computations as an exercise for the reader. We now prove part~\eqref{lem:Clebsch-Gordan.c}. Note first that for all $0\le t'\le m$, $0\le t''\le n$ \begin{equation}\label{eq:boundary-cnd} \begin{aligned} &C_{r;t',0}(v)= v^{\frac12 (r(1+m+n-r)-nt')} \frac{ (n)_v!}{(n-r)_v!}\binom{t'}{r}_v, \\ &C_{r;0,t''}(v)= (-1)^r v^{\frac12(mt''-r(1+m+n-r))} \frac{ (m)_v!}{(m-r)_v!}\binom{t''}{r}_v. \end{aligned} \end{equation} Using~\eqref{eq:recurr-relation for CG} with $t''=n$ we obtain $$ C_{r;t'+1,n}(v)=\frac{(m-r-t')_v}{(m-t')_v}\, v^{-\frac12 n} C_{r;t',n} $$ whence by~\eqref{eq:boundary-cnd} \begin{multline*} C_{r;t',n}(v)=v^{-\frac12 t'n}\frac{ (m-r)_v!(m-t')_v!}{(m)_v!(m-r-t')_v!}\, C_{r;0,n}(v)\\=(-1)^r v^{\frac12(n(m-t')-r(1+m+n-r))}\frac{(n)_v!}{(n-r)_v!}\binom{m-t'}{r}_v=(-1)^r C_{r;m-t',0}(v^{-1}). \end{multline*} Thus, \eqref{eq:key-combinatorial-identity} holds for all $0\le t'\le m$ and for $t''=n$. Suppose that~\eqref{eq:key-combinatorial-identity} was established for all $0\le t'\le m$ and for all $s+1\le t''\le n$. We have by~\eqref{eq:boundary-cnd} \begin{multline*} (-1)^r(n-r-s)_v C_{r;m,s}(v^{-1})=(-1)^r C_{r;m,s+1}(v^{-1})(n-s)_v v^{\frac12 m}=v^{\frac12 m}(n-s)_v C_{r;0,n-s-1}(v) \\ =(-1)^r v^{\frac12(m(n-s)-r(1+m+n-r))} \frac{ (m)_v!(n-s)_v}{(m-r)_v!}\binom{n-s-1}{r}_v =(n-r-s)_v C_{r;0,n-s}(v). \end{multline*} Finally, assume that~\eqref{eq:key-combinatorial-identity} is established for $k+1\le t'\le m$ and for $t''=s$. Then using~\eqref{eq:recurr-relation for CG} and~\eqref{eq:recurr-relation for CG.2} we obtain \begin{align*} (-1)^r(m+&n-r-k-s)_v C_{r;k,s}(v^{-1})\\&=(-1)^r C_{r;k+1,s}(v^{-1})(m-k)_v v^{-s+\frac12 n}+(-1)^r C_{r;k,s+1}(v^{-1})(n-s)_v v^{-\frac12 m+k} \\&=C_{r;m-k-1,n-s}(v)(m-k)_v v^{-s+\frac12 n}+C_{r;m-k,n-s-1}(n-s)_v v^{-\frac12 m+k} \\&=(m+n-r-k-s)C_{r;m-k,n-s}(v). \end{align*} This proves the inductive step and completes the proof of the Lemma. \end{proof} We can now complete the proof of Proposition~\ref{thm:tau-homomorphism}. By construction, $\tau$ is an isomorphism of $U_v(\lie{sl}_2)$-modules. Explicitly, if $z\in \mathcal A_0$ then $\tau(\ul E_-^{(r)}(z))=\ul E_+^{(-|z|-r)}(z)$. It suffices to prove that for any $x,y\in\mathcal A_0$ homogeneous with $|x|=-m$, $|y|=-n$ we have $$ \tau(E_-^{(k)}(x))\tau(E_-^{(l)}(y))=\ul E_+^{(m-k)}(x)\ul E_+^{(n-k)}(y)=\tau(E_-^{(k)}(x)E_-^{(l)}(y)). $$ It is immediate from the Remark~\ref{rem:Verma} that $C_{r;t',t''}(v)$ (respectively, $(-1)^r C_{r;t',t''}(v^{-1})$) provide the transition matrix between the two bases of $U_v(\lie{sl}_2)_+$-) (respectively, $U_v(\lie{sl}_2)_-$) modules $V_m\tensor V_n=\bigoplus_{0\le k\le\min(m,n)} V_{m+n-2k}$. In particular, there exists $\tilde C_{r;k,l}(v)\in\kk$, $0\le k\le m$, $0\le l\le n$, $0\le r\le \min(m,n,k+l)$ such that \begin{equation}\label{eq:inverse} \sum_{r=0}^{\min(m,n,k+l)}(-1)^r \tilde C_{r;k,l}(v) C_{r;t',t''}(v^{-1})=\delta_{k,t'}\delta_{l,t''}. \end{equation} Then $$ \ul E_-^{(k)}(x)\ul E_-^{(l)}(y)=\sum_{r=0}^{\min(m,n,k+l)} \tilde C_{r;k,l}(v) \ul E_-^{(k+l-r)}(x*_r y) $$ and so \begin{align*} \tau(\ul E_-^{(k)}(x)&\ul E_-^{(l)}(y))=\sum_{r=0}^{\min(m,n,k+l)} \tilde C_{r;k,l}(v)\ul E_+^{(m+n-r-k-l)}(x*_r y) \\ &=\sum_{r=0}^{\min(m,n,k+l)} \tilde C_{r;k,l}(v)\sum_{s'+s''=m+n-k-l}C_{r;s',s''}(v)\ul E_+^{(s')}(x)\ul E_+^{(s'')}(y) \\ &=\sum_{t'+t''=k+l}\Big(\sum_{r=0}^{\min(m,n,k+l)} \tilde C_{r;k,l}(v)C_{r;m-t',n-t''}(v)\Big)\ul E_+^{(m-t')}(x)\ul E_+^{(n-t'')}(y) \\ &=\sum_{t'+t''=k+l}\Big(\sum_{r=0}^{\min(m,n,k+l)} (-1)^r \tilde C_{r;k,l}(v)C_{r;t',t''}(v^{-1})\Big)\ul E_+^{(m-t')}(x)\ul E_+^{(n-t'')}(y) \\ &=\ul E_+^{(m-k)}(x)\ul E_+^{(n-l)}(y)=\tau(E_-^{(k)}(x))\tau(E_-^{(l)}(y)), \end{align*} where we used~\eqref{eq:key-combinatorial-identity} and~\eqref{eq:inverse}. \end{proof} Note that for $x\in \mathcal A_-$ homogeneous, $\tau(x)$ can be calculated explicitly in the following way. First, if $y\in \mathcal A_0$ is homogeneous and $x=\ul E_-^{(r)}(y)$ then \begin{equation}\label{eq:tau expl}\tau(x)=\ul E_+^{(-|y|-r)}(y) =\ul E_+^{(r-|x|)}(y)=\binom{2r-|x|}{r}_v^{-1}\ul E_+^{(r-|x|)}\ul F_-^{(r)}(x). \end{equation} By linearity, it remains to observe that any homogeneous element of~$\mathcal A_-$ can be written, uniquely, as $x=\sum_{r\ge \max(0,|x|)} \ul E_-^{(r)}(x_r)$ where $x_r\in \mathcal A_0$ and $|x_r|=|x|-2r$. We will also need the following property of~$\tau$. \begin{lemma} $\ul F_+^{(top)}\circ\tau=\ul F_-^{(top)}$. \label{lem:tau properties} \end{lemma} \begin{proof} Given $x\in\mathcal A_-$, write $x=\sum_{r\ge \max(0,|x|)}\ul E_-^{(r)}(x_r)$ where $x_r\in\mathcal A_0$ and $|x_r|=|x|-2r$. Then $\tau(x)=\sum_{r\ge \max(0,|x|)}\ul E_+^{(r-|x|)}(x_r)$. Let $r_0=\max\{r\ge \max(0,|x|)\,:\, x_r\not=0\}$. Then by~\eqref{eq:lowest weight action} $$ \ul F_-^{(top)}(x)=\ul F_-^{(r_0)}(x)=\ul F_-^{(r_0)}\ul E_-^{(r_0)}(x_{r_0})=\binom{2r_0-|x|}{r_0}_v x_{r_0}. $$ On the other hand, \begin{equation*} \ul F_+^{(top)}\tau(x)=\ul F_+^{(r_0-|x|)}\tau(x)=\ul F_+^{(r_0-|x|)}\ul E_+^{(r_0-|x|)}(x_{r_0})=\binom{2r_0-|x|}{r_0}_vx_{r_0}.\qedhere \end{equation*} \end{proof} \subsection{\texorpdfstring{$U_q(\lie n^+)$}{Uq(n+)} as a decorated algebra} Define a comultiplication $\Delta$ on~$U_q(\lie g)$ by $$ \Delta(E_i)=E_i\tensor 1+K_i^{-1}\tensor E_i,\qquad \Delta(F_i)=1\tensor F_i+F_i\tensor K_i,\qquad i\in I. $$ Then (cf.~\cite{Lus-book}*{\S3.1.5}) \begin{equation}\label{eq:delta-angl powers} \Delta(E_i^{\la r\ra})=\sum_{r'+r''=r} q_i^{-r'r''}E_i^{\la r'\ra}K_i^{-r''}\tensor E_i^{\la r''\ra}, \quad \Delta(F_i^{\la r\ra})=\sum_{r'+r''=r} q_i^{r'r''} F_i^{\la r'\ra}\tensor K_i^{r'}F_i^{\la r''\ra}. \end{equation} For any $J,J'\subset I$ denote $U_\ZZ(\lie g)_{J,J'}$ the $\AA_0$-subalgebra of $U_q(\lie g)$ generated by $U_\ZZ(\lie n^-)_J$, $U_\ZZ(\lie n^+)_{J'}$, the $K_i^{\pm 1}$ and the $\binom{K_i;c}{a}_{q_i}$, $a\in \ZZ_{\ge 0}$, $c\in \ZZ$ and $i\in J\cap J'$, where $$ \binom{K;c}{a}_{v}=\prod_{k=0}^{a-1} \frac{K^{-1}v^{k-c}-K v^{c-k}}{v^{k+1}-v^{-k-1}}. $$ We also abbreviate $U_\ZZ(\lie g)_J=U_\ZZ(\lie g)_{J,I}$ and $U_\ZZ(\lie g):=U_\ZZ(\lie g)_{I,I}$. The corresponding $\kk$-subalgebras of~$U_q(\lie g)$ will be denoted $U_q(\lie g)_{J,J'}$. It follows from~\eqref{eq:delta-angl powers} that $U_\ZZ(\lie g)$ is a Hopf $\AA_0$-algebra. Let $\ad$ be the corresponding adjoint action of~$U_q(\lie g)$ on itself. Consider the extension~$\widetilde U_q(\lie g)$ of $U_q(\lie g)$ obtained by adjoining $K_i^{\pm\frac12}$, $i\in I$. Define operators $\ul E_i$, $\ul F_i$ on~$\widetilde U_q(\lie g)$ via \begin{equation}\label{eq:ul definition} \begin{aligned} &\ul E_i(x)=-(\ad E_i^{\la 1\ra}K_i^{\frac12})=\frac{E_i K_i^{\frac12} xK_i^{-\frac12}-K_i^{-\frac12}xK_i^{\frac12}E_i}{q_i^{-1}-q_i},\\ &\ul F_i(x)=(\ad K_i^{-\frac12}F_i^{\la 1\ra})(x)=\partial_i(x)-K_i^{-1}\partial_i^{op}(x)K_i^{-1}. \end{aligned} \end{equation} Clearly, $\ul E_i$ and $\ul F_i$ restrict to operators on $U_q(\lie g)$ and we have \begin{equation}\label{eq:sl_2-rel} [\ul E_i,\ul F_i]=-\la 1\ra_{q_i}{}^{-2}\ad([E_i,F_i])=\la 1\ra_{q_i}{}^{-1}(\ul K_i-\ul K_i^{-1}), \end{equation} where $\ul K_i(x)=K_i x K_i^{-1}$. We will also need operators $\ul E_i^{op}$, $\ul F_i^{op}$ defined by \begin{equation}\label{eq:ul E_i^op} \ul E_i^{op}(x)=(\ul E_i(x^*))^*,\qquad \ul F_i^{op}(x)=(\ul F_i(x^*))^*. \end{equation} We collect some properties of these operators in the following Lemma. \begin{lemma} \begin{enumerate}[{\rm(a)}] \item\label{lem:prop-ul E_i.a'''} $U_q(\lie n^+)$ is a decorated algebra with $E=E_i$, $\ul F_+=\partial_i$, $\ul F_-=\partial_i^{op}$, $v=q_i$ and $|x|=(\alpha_i^\vee,\gamma)$ for $x\in U_q(\lie n^+)_\gamma$; in particular, $\ul E_+=\ul E_i$ and $\ul E_-=\ul E_i^{op}$. \item\label{lem:prop-ul E_i.a'} $\ul E_i$, $\ul F_i$ commute with $\bar\cdot$. \item\label{lem:prop-ul E_i.a''} If $x\in U_\ZZ(\lie g)_\gamma$ then $\ul E_i^{(r)}(x),\ul F_i^{(r)}(x)\in q^{\frac12 (r\alpha_i,\gamma)}U_\ZZ(\lie g)$ for all $r\in\ZZ_{\ge0}$; \item\label{lem:prop-ul E_i.b} For all $x\in U_q(\lie n^+)_\gamma$, $y\in U_q(\lie n^+)$ we have $$ \fgfrm{\ul E_i^{(r)}(x)}{y}= \sum_{r'+r''=r}(-1)^{r'}q_i^{-\frac12(r+(\alpha_i^\vee,\gamma)-1)(r''-r')}\fgfrm{x}{\partial_i^{(r'')} (\partial_i^{op})^{(r')}(y)}\\ $$ \item\label{lem:prop-ul E_i.c} $T_i\circ\ul E_i^{op}=\ul F_i\circ T_i$, $T_i\circ \ul F_i^{op}=\ul E_i\circ T_i$. \end{enumerate}\label{lem:prop-ul E_i} \end{lemma} \begin{proof} Parts~\eqref{lem:prop-ul E_i.a'''} and~\eqref{lem:prop-ul E_i.a'} are obvious from the definitions. Since $U_\ZZ(\lie g)$ is a Hopf $\AA_0$-algebra, the first assertion in~\eqref{lem:prop-ul E_i.a''} follows from $$ \ul E_i^{(r)} =(-1)^r q_i^{\binom{r}2}\ad(E_i^{\la r\ra}K_i^{\frac r2}),\quad \ul F_i^{(r)}=q_i^{\binom r2}(\ad K_i^{-\frac r2}F_i^{\la r\ra}), $$ while the second is immediate from the above formulae and~\eqref{eq:delta-angl powers}. Part~\eqref{lem:prop-ul E_i.b} is immediate from part~\eqref{lem:prop-ul E_i.a'''}, \eqref{eq:divided power E_i} and Lemma~\ref{lem:partial-prop}\eqref{lem:partial-prop.ii}. Part~\eqref{lem:prop-ul E_i.c} is easy to check using Lemma~\ref{lem:T_i-defn-prop}. \end{proof} \subsection{A new formula for~\texorpdfstring{$T_i$}{Ti}}\label{subs:formula T_i} Using the notation from~\cite{BG-dcb}, denote by $U_i:=T_i^{-1}(U_q(\lie n^+))\cap U_q(\lie n^+)$ and ${}_iU:=T_i(U_q(\lie n^+))\cap U_q(\lie n^+)$. It follows from~\cite{Lus-book}*{Proposition~38.1.6} that $U_i=\ker\partial_i$ and ${}_i U=\ker\partial_i^{op}$. Let $U^\ZZ_i=U_i\cap U^\ZZ(\lie n^+)$ and ${}_iU^\ZZ={}_i U\cap U^\ZZ(\lie n^+)$. \begin{lemma} Let $i\in I$. \label{lem:ul E_i adjoint} \begin{enumerate}[{\rm(a)}] \item\label{lem:ul E_i adjoint.a} For all $x\in U_q(\lie n^+)_\gamma$, $y\in {}_i U$, $z\in U_i$ and $r\ge 0$ we have $$ \fgfrm{\ul E_i^{(r)}(x)}{y}=q_i^{-\frac12(r+(\alpha_i^\vee,\gamma)-1)r}\fgfrm{x}{\ul F_i^{(r)}(y)}, \quad \fgfrm{(\ul E_i^{op})^{(r)}(x)}{z}=q_i^{-\frac12(r+(\alpha_i^\vee,\gamma)-1)r}\fgfrm{x}{(\ul F_i^{op})^{(r)}(z)}. $$ \item\label{lem:ul E_i adjoint.b} $\ul E_i$, $\ul F_i$ (respectively, $\ul E_i^{op}$, $\ul F_i^{op}$) restrict to locally nilpotent operators on~${}_i U$ (respectively, on~$U_i$). \item\label{lem:ul E_i adjoint.c} $\ul E_i^{(n)}({}_i U^\ZZ),\ul F_i^{(n)}({}_i U^\ZZ)\subset {}_i U^\ZZ$ (respectively, $(\ul E_i^{op})^{(n)}(U^\ZZ_i), (\ul F_i^{op})^{(n)}(U^\ZZ_i)\subset U^\ZZ_i$) for all $n\ge 0$. \end{enumerate} \end{lemma} \begin{proof} We only prove the assertion for $\ul E_i$ and~$\ul F_i$. The assertion for $\ul E_i^{op}$ and~$\ul F_i^{op}$ is proved similarly using~\eqref{eq:ul E_i^op} and the fact that ${}^*$ is self-adjoint with respect to $\fgfrm{\cdot}{\cdot}$. Since by~\eqref{eq:ul definition} $\ul F_i|_{{}_i U}=\partial_i|_{{}_i U}$, in particular, $\ul F_i$ is a locally nilpotent operator on~${}_i U$. Part~\eqref{lem:ul E_i adjoint.a} is now immediate from Lemma~\ref{lem:prop-ul E_i}\eqref{lem:prop-ul E_i.b}. Suppose that $x\in U_q(\lie n^+)_\gamma$ and $\ul E_i^{(n)}(x)\not=0$ for all $n\ge 0$. Then $T_i(\ul E_i^{(n)}(x))$ is homogeneous of degree~$s_i(\gamma+n\alpha_i)=\gamma-((\alpha_i^\vee,\gamma)+n)\alpha_i\notin Q^+$ for $n\gg 0$. Since $T_i({}_i U)\subset U_q(\lie n^+)$, this is a contradiction. This proves~\eqref{lem:ul E_i adjoint.b}. To prove~\eqref{lem:ul E_i adjoint.c}, note that the assertion for $\ul F_i$ follows from Corollary~\ref{cor:preserv-lattice}. Since $U_\ZZ(\lie n^+)={}_i U_\ZZ\oplus (E_i U_q(\lie n^+)\cap U_\ZZ(\lie n^+))$, it suffices to prove that for $x\in {}_i U^\ZZ\cap U_q(\lie n^+)_\gamma$, $y\in U_i^\ZZ\cap U_q(\lie n^+)_{\gamma+n\alpha_i}$ we have $\fgfrm{\ul E_i^{(n)}(x)}{y}\in \AA_0$. But for such $y$ we have by part~\eqref{lem:ul E_i adjoint.a} and~\eqref{eq:mu-prop} $$ \fgfrm{\ul E_i^{(n)}(x)}{y}=q_i^{-\frac12 n(n+(\alpha_i^\vee,\gamma)-1)}\fgfrm{x}{\ul F_i^{(n)}(y)}=q_i^{\binom{n}2}\fgfrm{x} {q_i^{-\frac12 n(\alpha_i^\vee,\gamma+n\alpha_i)}\ul F_i^{(n)}(y)}\in \AA_0 $$ and $q_i^{-\frac12 n(\alpha_i^\vee,\gamma+n\alpha_i)}\ul F_i^{(n)}(y)\in U_\ZZ(\lie n^+)$ by Lemma~\ref{lem:prop-ul E_i}\eqref{lem:prop-ul E_i.a''}. \end{proof} Thus, given $x\in U_i\cap U_q(\lie n^+)_\gamma$, $y\in {}_i U\cap U_q(\lie n^+)_\gamma$ we can write uniquely \begin{equation}\label{eq:sl_2 decomp} x=\sum_{r\ge \max(0,(\alpha_i^\vee,\gamma))} (\ul E_i^{op})^{(r)}(x_r),\quad y=\sum_{r\ge \max(0,(\alpha_i^\vee,\gamma))} \ul E_i^{(r)}(y_r),\quad x_r,y_r\in {}_i U\cap U_i\cap U_q(\lie n^+)_{\gamma-r\alpha_i}, \end{equation} and $x_r,y_r=0$ for $r\gg 0$. \begin{corollary}\label{cor:form sl2 decomp} Let $x,x'\in U_i\cap U_q(\lie n^+)_\gamma$, $y,y'\in {}_i U\cap U_q(\lie n^+)_\gamma$ and write $x$, $x'$ and $y$, $y'$ as in~\eqref{eq:sl_2 decomp}. Then \begin{equation}\label{eq:form sl2 decomp} \begin{aligned} \fgfrm{x}{x'}=\sum_{r\ge \max(0,(\alpha_i^\vee,\gamma))} q_i^{\frac12 r(r+1-(\alpha_i^\vee,\gamma))}\binom{2r-(\alpha_i^\vee,\gamma)}{r}_{q_i}\fgfrm{x_r}{x'_r}, \\ \fgfrm{y}{y'}=\sum_{r\ge \max(0,(\alpha_i^\vee,\gamma))} q_i^{\frac12 r(r+1-(\alpha_i^\vee,\gamma))}\binom{2r-(\alpha_i^\vee,\gamma)}{r}_{q_i}\fgfrm{y_r}{y'_r}. \end{aligned} \end{equation} \end{corollary} \begin{proof} Let $z,z'\in {}_i U\cap U_i$, $z'\in U_q(\lie n^+)_{\gamma'}$. Let $a\ge b\ge 0$. Then we have by Lemma~\ref{lem:ul E_i adjoint}\eqref{lem:ul E_i adjoint.a} and~\eqref{eq:lowest weight action} \begin{multline*} \fgfrm{\ul E_i^{(a)}(z)}{\ul E_i^{(b)}(z')}=q_i^{-\frac12 b(b+(\alpha_i^\vee,\gamma')-1)}\fgfrm{\ul F_i^{(b)}\ul E_i^{(a)}(z)}{z'}\\ =q_i^{-\frac12 b(b+(\alpha_i^\vee,\gamma')-1)}\binom{b-a-(\alpha_i^\vee,\gamma')}{b}_{q_i}\fgfrm{\ul E_i^{(a-b)}(z)}{z'} =\delta_{a,b}q_i^{-\frac12a(a+(\alpha_i^\vee,\gamma')-1)}\binom{-(\alpha_i^\vee,\gamma')}{a}_{q_i}\fgfrm{z}{z'}. \end{multline*} This yields the second identity in~\eqref{eq:form sl2 decomp}. The first follows from the second one by applying~${}^*$. \end{proof} We now establish a formula for the action of $T_i$ on $U_i$ in terms of $\ul E_i^{op}$ and $\ul E_i$. \begin{theorem}\label{thm:T_i formula} Write $x\in U_i\cap U_q(\lie n^+)_\gamma$ as in~\eqref{eq:sl_2 decomp}. Then $$ T_i(x)=\sum_{r\ge \max(0,(\alpha_i^\vee,\gamma))} \ul E_i^{(r-(\alpha_i^\vee,\gamma))}(x_r). $$ In particular, $\partial_i^{(top)}T_i(x)=(\partial_i^{op})^{(top)}(x)$. \end{theorem} \begin{proof} We apply Proposition~\ref{thm:tau-homomorphism} to~$\mathcal A=U_q(\lie n^+)$ which is a decorated algebra by Lemma~\ref{lem:prop-ul E_i}\eqref{lem:prop-ul E_i.a'''} with locally nilpotent $\ul E_\pm$ on~$\mathcal A_\pm$ by Lemma~\ref{lem:ul E_i adjoint}\eqref{lem:ul E_i adjoint.b}. We claim that $T_i=\tau$. Since both $\tau$ and~$T_i$ are isomorphisms of algebras $U_i\to {}_i U$, it is enough to check that they coincide on generators of~$U_i$. Let $j\not=i$ and define for all $0\le l\le -a_{ij}$ \begin{equation}\label{eq:E_ji^l} E_{ji^{l}}=\binom{-a_{ij}}{l}_{q_i}^{-1}(\ul E_i^{op})^{(l)}(E_j),\qquad E_{i^lj}=(E_{ji^l})^*=\binom{-a_{ij}}{l}_{q_i}^{-1}\ul E_i^{(l)}(E_j). \end{equation} Clearly, $\overline{E_{ji^l}}=E_{ji^l}$ and it is immediate from~\eqref{eq:divided power E_i} that \begin{equation}\label{eq:E_ji^m explicit} E_{ji^l}=\binom{-a_{ij}}{l}_{q_i}^{-1}\sum_{r+s=l} (-1)^r q_i^{\frac12(r-s)(l+a_{ij}-1)} E_i^{\la s\ra}E_j E_i^{\la r\ra}. \end{equation} \begin{lemma}\label{lem:gen-U_i} Let $i\not=j\in I$, $0\le m\le -a_{ij}$. Then \begin{enumerate}[{\rm(a)}] \item\label{lem:gen-U_i.b} $T_i(E_{ji^m})=E_{i^{-a_{ij}-m}j}=\tau(E_{ji^m})$ \item\label{lem:gen-U_i.c} The elements $E_{ji^l}$ (respectively, $E_{i^lj}$), $j\not=i$, $0\le l\le -a_{ij}$ generate the algebra $U_i$ (respectively ${}_i U$). \end{enumerate} \end{lemma} \begin{proof} To prove~\eqref{lem:gen-U_i.b}, note that by Lemma~\ref{lem:T_i-defn-prop} and~\eqref{eq:E_ji^m explicit} we have $T_i(E_j)=E_{i^{-a_{ij}}j}$. On the other hand, $\tau(E_j)=\ul E_i^{(-a_{ij})}(E_j)=E_{i^{-a_{ij}}j}$. Then by Lemma~\ref{lem:prop-ul E_i}\eqref{lem:prop-ul E_i.c} and~\eqref{eq:lowest weight action} $$ T_i(E_{ji^l})= \binom{-a_{ij}}{l}_{q_i}^{-1} \ul F_i^{(l)}(E_{i^{-a_{ij}}j}) =\binom{-a_{ij}}{l}_{q_i}^{-1} \ul F_i^{(l)}\ul E_i^{(-a_{ij})}(E_j)=E_{i^{-a_{ij}-l}j}. $$ Since by construction $\tau$ also satisfies Lemma~\ref{lem:prop-ul E_i}\eqref{lem:prop-ul E_i.c}, it follows that $\tau(E_{ji^l})=T_i(E_{ji^l})$. Part~\eqref{lem:gen-U_i.c} can be easily deduced from~\cite{Lus-book}*{\S38.1.1}. \end{proof} \noindent This implies that $\tau=T_i$ on~$U_i$. The second assertion of Theorem~\ref{thm:T_i formula} follows from Lemma~\ref{lem:tau properties}. \end{proof} We now prove the following \begin{proposition}\label{prop:T_i lattice} For all $i\in I$, $T_i(U^\ZZ_i)={}_i U^\ZZ$. \end{proposition} \begin{proof} We need the following \begin{lemma}\label{lem:Jantzen} Any element $x\in U_\ZZ(\lie n^+)$ can be written as $x=\sum_{r,s\ge 0} E_i^{\la r\ra} x_{rs} E_i^{\la s\ra}$, where $x_{rs}\in {}_i U\cap U_i\cap U_\ZZ(\lie n^+)$ and only finitely many of them are non-zero. \end{lemma} \begin{proof} We need the following elementary fact. \begin{lemma}\label{lem:elem-lin.alg} Let $V$ be a finite dimensional $\kk$-vector space with a non-degenerate bilinear form $\fgfrm{\cdot}{\cdot}:V\tensor V\to \kk$. Assume that we have two orthogonal direct sum decompositions $V=V_1\oplus W_1=V_2\oplus W_2$ with respect to that form. Then $V=(V_1\cap V_2)\oplus (W_1+W_2)=(W_1\cap W_2)\oplus (V_1+V_2)$ (orthogonal direct sum decompositions). \end{lemma} \begin{proof} Clearly $(V_1\cap V_2)$ is orthogonal to $W_1+W_2$ and $(W_1\cap W_2)$ is orthogonal to $V_1+V_2$. Note that for any $v\in V_i$, $\fgfrm{v}{v}=0$ if and only if $v=0$. This implies that the sums $U_1=(V_1\cap V_2)+(W_1+W_2)$, $U_2=(W_1\cap W_2)+(V_1+V_2)$ are direct. It remains to prove that $\dim U_1=\dim U_2=\dim V$. Since \begin{multline*} \dim U_1+\dim U_2=(\dim W_1+\dim W_2-\dim W_1\cap W_2)+\dim V_1\cap V_2\\+(\dim V_1+\dim V_2-\dim V_1\cap V_2)+\dim W_1\cap W_2=2\dim V, \end{multline*} and $\dim U_1,\dim U_2\le \dim V$ the assertion follows. \end{proof} Given $\gamma\in Q^+$, let $n_i(\gamma)$ be the coefficient of~$\alpha_i$ in~$\gamma$. For any $\gamma\in Q^+$ we have two orthogonal direct sum decompositions $U_q(\lie n^+)_\gamma=(\ker\partial_i|_{U_q(\lie n^+)_\gamma}\oplus U_q(\lie n^+)_{\gamma-\alpha_i}E_i)=(\ker\partial_i^{op}|_{U_q(\lie n^+)_\gamma}\oplus E_i U_q(\lie n^+)_{\gamma-\alpha_i})$. Since $U_q(\lie n^+)_\gamma$ is finite dimensional, it follows from Lemma~\ref{lem:elem-lin.alg} that $U_q(\lie n^+)_\gamma=({}_i U\cap U_i\cap U_q(\lie n^+)_\gamma) \oplus (U_q(\lie n^+)_{\gamma-\alpha_i}E_i+E_i U_q(\lie n^+)_{\gamma-\alpha_i})$. Then an obvious induction on~$n_i(\gamma)$ implies that every $x\in U_q(\lie n^+)$ can be written in $x=\sum_{r,s\ge 0} E_i^{\la r\ra} x_{rs} E_i^{\la s\ra}$, where $x_{rs}\in {}_i U\cap U_i$ and only finitely many of the $x_{rs}$ are non-zero. We now prove by induction on~$n_i(\gamma)$ that if $x\in U_\ZZ(\lie n^+)_\gamma$ then $x_{rs}\in U_\ZZ(\lie n^+)\cap {}_i U\cap U_i$. If $n_i(\gamma)=0$ then $x=x_{00}$ and there is nothing to prove. For the inductive step, we have $$ \sum_{r,s\ge 0} E_i^{\la r\ra}(q_i^{-r-1} x_{r+1,s}+q_i^{-(\alpha_i^\vee,\gamma)+s+1} x_{r,s+1})E_i^{\la s\ra} =q_i^{-\frac12(\alpha_i^\vee,\gamma)}\la 1\ra_{q_i}\partial_i(x)\in U_\ZZ(\lie n^+), $$ where we used Lemma~\ref{lem:partial-prop}\eqref{lem:partial-prop.iii} and Corollary~\ref{cor:preserv-lattice}\eqref{cor:preserv-lattice.a}. Then $x_{r+1,s}+q_i^{-(\alpha_i^\vee,\gamma)+r+s+2} x_{r,s+1}\in U_\ZZ(\lie n^+)$ by the induction hypothesis. Let $s_0$ be such that $x_{rs}=0$ for all $r$ and for all $s>s_0$. It follows then that $x_{r,s_0}\in U_\ZZ(\lie n^+)$ for all $r\ge 0$. Suppose now that $x_{rt}\in U_\ZZ(\lie n^+)$ for all $r$ and for all $s+1\le t\le s_0$. Since $x_{r,s}=-q_i^{-(\alpha_i^\vee,\gamma)+s+r+1} x_{r-1,s+1}$ it follows that $x_{r,s}\in U_\ZZ(\lie n^+)$ for all $r,s\ge 0$ with $r+s>0$. It remains to observe that $x_{00}=x-\sum_{r,s\ge 0,r+s>0} E_i^{\la r\ra} x_{rs} E_i^{\la s\ra}$. \end{proof} \begin{lemma}\label{lem:action-Jantzen} Let $x\in U_i^\ZZ\cap U_q(\lie n^+)_\gamma$ and write $x=\sum_{r\ge \max(0,(\alpha_i^\vee,\gamma))} (\ul E_i^{op})^{(r)}(x_r)$ where $x_r\in {}_i U\cap U_i\cap U_q(\lie n^+)_{\gamma-r\alpha_i}$. Then $\binom{-(\alpha_i^\vee,\gamma-r\alpha_i)}{r}_{q_i}x_r\in U_i^\ZZ$. \end{lemma} \begin{proof} The argument is by induction on~$\ell_i(x^*)$. If $\ell_i(x^*)=0$, that is $\partial_i^{op}(x)=0$, then $x=x_0$ and there is nothing to do. If~$\ell_i(x^*)=n$ then $x_r=0$ for all $r>n$. We have $(\partial_i^{op})^{(top)}(x) =(\partial_i^{op})^{(n)}(x)\in U_i^\ZZ$ by Corollary~\ref{cor:preserv-lattice}\eqref{cor:preserve-lattice.b}. On the other hand, $ (\partial_i^{op})^{(n)}(x)=(\partial_i^{op})^{(n)}(\ul E_i^{op})^{(n)}(x_{n})=\binom{2n-(\alpha_i^\vee,\gamma)}{n}_{q_i}x_{n} $ by~\eqref{eq:lowest weight action}. Thus, $\binom{2n-(\alpha_i^\vee,\gamma)}{n}_{q_i}x_n\in U_i^\ZZ\cap {}_i U^\ZZ$. It remains to observe that the induction hypothesis applies to $x-(\ul E_i^{op})^{(n)}(x_n)$. \end{proof} Thus, is suffices to consider $x=(\ul E_i^{op})^{(r)}(z)\in U_i^\ZZ$ where $z\in {}_i U\cap U_i\cap U_q(\lie n^+)_\gamma$. We claim that $T_i(x)=\ul E_i^{(-(\alpha_i^\vee,\gamma)-r)}(z)\in {}_i U^\ZZ$. Given $y\in U_\ZZ(\lie n^+)_{\gamma+(-(\alpha_i^\vee,\gamma)-r)\alpha_i}$, use Lemma~\ref{lem:Jantzen} to write $y=\sum_{s,s'\ge 0} E_i^{\la s'\ra}y_{s's} E_i^{\la s\ra}$ with $y_{s's}\in {}_i U\cap U_i\cap U_\ZZ(\lie n^+)$. Then by Lemma~\ref{lem:partial-prop}\eqref{lem:partial-prop.ii} and~\eqref{eq:lowest weight action} \begin{multline*} \fgfrm{\ul E_i^{(-(\alpha_i^\vee,\gamma)-r)}(z)}{y}=\sum_{s',s\ge 0} \fgfrm{\ul E_i^{(-(\alpha_i^\vee,\gamma)-r)}(z)}{E_i^{\la s'\ra}y_{s's} E_i^{\la s\ra}} =\sum_{s\ge 0} \fgfrm{\partial_i^{(s)}\ul E_i^{(-(\alpha_i^\vee,\gamma)-r)}(z)}{y_{0s}}\\ =\sum_{s\ge 0} \fgfrm{\ul F_i^{(s)}\ul E_i^{(-(\alpha_i^\vee,\gamma)-r)}(z)}{y_{0s}} =\sum_{s=0}^{-(\alpha_i^\vee,\gamma)-r} \binom{s+r}{r}_{q_i} \fgfrm{\ul E_i^{(-(\alpha_i^\vee,\gamma)-r-s)}(z)}{y_{0s}} \\=\binom{-(\alpha_i^\vee,\gamma)}{r}_{q_i}\fgfrm{z}{y_{0,-(\alpha_i^\vee,\gamma)-r}}, \end{multline*} since $\fgfrm{{}_i U}{E_i U_q(\lie n^+)}=0$ and $\fgfrm{\ul E_i^{(a)}(z)}{y_{0s}}=0$ if $a>0$ by Lemma~\ref{lem:prop-ul E_i}\eqref{lem:prop-ul E_i.b}. Since $\binom{-(\alpha_i^\vee,\gamma)}{r}_{q_i}z\in U^\ZZ(\lie n^+)$ by Lemma~\ref{lem:action-Jantzen}, it follows that $\fgfrm{\ul E_i^{(-(\alpha_i^\vee,\gamma)-r)}(z)}{y}\in\AA_0$. \end{proof} \subsection{Proof of Theorem~\ref{thm:T_i bas}}\label{subs:proof of thm T_i bas} We need the following result which can also be deduced from~\cite{Lus-book}*{Proposition 38.2.1}. However, our argument is much shorter. \begin{lemma}\label{lem:orthogonality of T_i} For all $x,x'\in U_q(\lie n^+)_\gamma\cap U_i$, $a,a'\in \ZZ_{\ge 0}$ we have $$ \fgfrm{E_i^{a}T_i(x)}{E_i^{a'}T_i(x')}=q^{\frac12(a-1)(\alpha_i,\gamma)}\delta_{a,a'}\la a\ra_{q_i}! \fgfrm{x}{x'} =q^{\frac12(a-1)(\alpha_i,\gamma)}\delta_{a,a'}\mu(a\alpha_i)\prod_{t=1}^a(1-q_i^{-2t}) \fgfrm{x}{x'}, $$ where~$\mu$ is defined as in Theorem~\ref{thm:sign-bas}. \end{lemma} \begin{proof} It follows immediately from Lemma~\ref{lem:partial-prop}\eqref{lem:partial-prop.ii} and~\eqref{eq:partial E_i^r} that if $y,y'\in\ker\partial_i^{op}$, $y\in U_q(\lie n^+)_{\gamma'}$ and $a,a'\in\ZZ_{\ge0}$ then $$ \fgfrm{E_i^a y}{E_i^{a'}y'}=\delta_{a,a'}\la a\ra_{q_i}!q^{-\frac12 a(\alpha_i,\gamma')} \fgfrm{y}{y'}=\delta_{a,a'}q_i^{\binom{a+1}2-\frac12 a(\alpha_i^\vee,\gamma')}\prod_{t=1}^a (1-q_i^{-2t})\fgfrm{y}{y'}. $$ Let $\gamma_i=(\alpha_i^\vee,\gamma)$. Since $T_i(x),T_i(x')\in \ker\partial_i^{op}\cap U_q(\lie n^+)_{s_i\gamma}$, it remains to prove the assertion for $a=a'=0$, $x=(\ul E_i^{op})^{(a)}(z)$ and $x'=(\ul E_i^{op})^{(b)}(z')$ where $z,z'\in {}_i U\cap U_i$ and $a\ge b\ge \max(0,\gamma_i)$. Since $z\in U_q(\lie n^+)_{\gamma-a\alpha_i}$, $z'\in U_q(\lie n^+)_{\gamma-b\alpha_i}$ we have, by Corollary~\ref{cor:form sl2 decomp} $$ \fgfrm{x}{x'}=\delta_{a,b}q_i^{\frac12 a(1+a-\gamma_i)}\binom{2a-\gamma_i}{a}_{q_i}\fgfrm{z}{z'}. $$ On the other hand, using Theorem~\ref{thm:T_i formula} and Corollary~\ref{cor:form sl2 decomp} we obtain \begin{equation*} \begin{split} \fgfrm{T_i(x)}{T_i(y)}=\fgfrm{\ul E_i^{(a-\gamma_i)}(z)}{\ul E_i^{(b-\gamma_i)}(z')} =q_i^{\frac12(a-\gamma_i)(a+1)}\delta_{a,b}\binom{2a-\gamma_i}{a-\gamma_i}_{q_i}\fgfrm{z}{z'}=q_i^{-\frac12\gamma_i}\fgfrm{x}{y}.\qedhere \end{split} \end{equation*} \end{proof} Let $b\in \mathbf B^{up}{}_\gamma\cap T_i^{-1}(U_q(\lie n^+))$. Since $T_i$ commutes with $\bar\cdot$ we have $\overline{T_i(b)}=T_i(\overline b)=T_i(b)$. By Proposition~\ref{prop:T_i lattice} we have $T_i(b)\in U^\ZZ(\lie n^+)$. Furthermore, by Lemma~\ref{lem:orthogonality of T_i} and~\eqref{eq:mu-prop} $$ \mu(s_i\gamma)^{-1}\fgfrm{T_i(b)}{T_i(b)}=\mu(\gamma)^{-1}q^{\frac12(\alpha_i,\gamma)}\fgfrm{T_i(b)}{T_i(b)}=\mu(\gamma)^{-1}\fgfrm{b}{b}\in 1+K_-. $$ Thus, $T_i(b)\in\mathbf B^{\pm up}$ by~\eqref{eq:def-bas}. It remains to prove that $T_i(b)\in \mathbf B^{up}$. Since $(\partial_i^{op})^{(top)}(b)\in\mathbf B^{up}$ by Remark~\ref{rem: star op partial}, there exists a sequence $\ii'=(i_1,\dots,i_{m})\in I^{m}$ such that $\partial_{\ii'}^{(top)}((\partial_i^{op})^{(top)}(b))=1$. Let $\ii=(i,i_1,\dots,i_m)$. Then $\partial_\ii^{(top)}(T_i(b))=\partial_{\ii'}^{(top)}\partial_i^{(top)}T_i(b) =\partial_{\ii'}^{(top)}((\partial_i^{op})^{(top)}(b)) =1$ by Theorem~\ref{thm:T_i formula}. Thus, $T_i(b)\in\mathbf B^{up}$.\qed \section{Proofs of mains results} \subsection{Properties of quantum Schubert cells}\label{subs:some prop q S c} Let $w\in W$ and $\mathbf i=(i_1,\dots,i_m)\in R(w)$. Set $X_{\mathbf i,k}=T_{i_1}\cdots T_{i_{k-1}}(E_{i_k})$, $1\le k\le m$, and let $U_q(\ii)$ be the subalgebra of $U_q(\lie n^+)$ generated by the $X_{\mathbf i,k}$, $1\le k\le m$ and set $U^\ZZ(\ii)=U_q(\ii)\cap U^\ZZ(\lie n^+)$, $U_\ZZ(\ii)=U_q(\ii)\cap U_\ZZ(\lie n^+)$. The following is well-known. \begin{lemma}[\cite{Lus-book}*{Propositions~40.2.1 and~41.1.4}]\label{lem:lus PBW elt} The elements $X_\ii^{\la \mathbf a\ra}:=X_{\ii,1}^{\la a_1\ra}\cdots X_{\ii,m}^{\la a_m\ra}$, $\mathbf a\in \ZZ_{\ge 0}^m$ form an $\AA_0$-basis of $U_\ZZ(\ii)$ and a $\kk$-basis of~$U_q(\ii)$. \end{lemma} Set $\alpha^{(k)}=\alpha^{(k)}_\ii:=s_{i_1}\cdots s_{i_{k-1}}(\alpha_{i_k})=\deg X_{\mathbf i,k}$ and given $\mathbf a=(a_1,\dots,a_m)\in\ZZ^m$ denote $|\mathbf a|=|\mathbf a|_\ii=\sum_{k=1}^m a_k\deg X_{\ii,k}=\sum_{k=1}^m a_k\alpha^{(k)}_\ii$. Define $$ X_\ii^{\mathbf a}=q_{\mathbf i,\mathbf a} X_{\ii,1}^{a_1}\cdots X_{\ii,m}^{a_m},\qquad \mathbf a=(a_1,\dots,a_m)\in\ZZ_{\ge 0}^m, $$ where \begin{equation}\label{eq:q_i a defn} q_{\mathbf i,\mathbf a}=q^{\frac12 \sum_{1\le k<l\le m}(\alpha_\ii^{(k)},\alpha_\ii^{(l)})a_ka_l}. \end{equation} This choice is justified by the following \begin{proposition}\label{prop:X i^a scalar square} For all $\ii\in R(w)$, $\mathbf a,\mathbf a'\in \ZZ_{\ge 0}^m$ we have $X_{\ii}^{\mathbf a}\in U^\ZZ(\lie n^+)$ and \begin{equation}\label{eq:X i^a scalar square} \mu(|\mathbf a|)^{-1}\fgfrm{X_\ii^{\mathbf a}}{X_\ii^{\mathbf a'}}=\delta_{\mathbf a,\mathbf a'}\prod_{r=1}^m\prod_{t=1}^{a_m}(1-q_{i_r}^{-2t}). \end{equation} Thus, the set $\{ X_\ii^{\mathbf a}\,:\,|\mathbf a|_\ii=\gamma\}$ is a $(K_-,\mu(\gamma)^{-1})$-orthonormal basis of~$U^\ZZ(\ii)_\gamma$ and \begin{equation}\label{eq:dual PBW} \fgfrm{X_\ii^{\mathbf a}}{X_\ii^{\la \mathbf a'\ra}}=\delta_{\mathbf a,\mathbf a'},\qquad \mathbf a,\mathbf a'\in\ZZ_{\ge 0}^m. \end{equation} \end{proposition} \begin{proof} We need the following \begin{lemma}\label{lem:T_w(E_j)} For all $w\in W$, $j\in I$ such that $\ell(ws_j)=\ell(w)+1$ we have $T_w(E_j)\in U^\ZZ(\lie n^+)$. \end{lemma} \begin{proof} The argument is by induction on~$\ell(w)$. If $\ell(w)=0$ there is nothing to prove. Suppose that $w=s_iw'$ with $\ell(w)=\ell(w')+1$. Clearly, $\ell(w's_j)=\ell(w')+1$. Then $T_{w'}(E_j)\in \ker\partial_i$ by~\cite{Lus-book}*{Lemma~40.1.2} and also $T_{w'}(E_j)\in U^\ZZ(\lie n^+)$ by the induction hypothesis. Then by Proposition~\ref{prop:T_i lattice}, $T_w(E_j)=T_i(T_{w'}(E_j))\in U^\ZZ(\lie n^+)$. \end{proof} This implies that $X_{\mathbf i,k}\in U^\ZZ(\lie n^+)$ and hence $X_{\ii}^{\mathbf a}\in U^\ZZ(\lie n^+)$ by Lemma~\ref{lem:prod in U^Z}. To prove~\eqref{eq:X i^a scalar square} we use induction on~$\ell(w)$. The case~$\ell(w)=0$ is trivial. For the inductive step, assume that $\ell(s_iw)=\ell(w)+1$ and note that we have $$ X_{(i,\ii)}^{(a,\mathbf a)}=q^{\frac12 a(\alpha_i,s_i|\mathbf a|_\ii)}E_{i}^{a} T_{i}(X_{\ii}^{\mathbf a})=q^{-\frac12 a(\alpha_i,|\mathbf a|_\ii)} E_{i}^{a} T_{i}(X_{\ii}^{\mathbf a}). $$ Since $T_i(X_{\ii}^{\mathbf a})\in {}_i U$, we have by Lemmata~\ref{lem:partial-prop}, \ref{lem:orthogonality of T_i} and~\eqref{eq:mu-prop} \begin{align*} \fgfrm{X_{(i,\ii)}^{(a,\mathbf a)}}{&X_{(i,\ii)}^{(a',\mathbf a')}} =q^{-\frac12(a(\alpha_i,|\mathbf a|_\ii)+a'(\alpha_i,|\mathbf a'|_\ii)}\fgfrm{E_{i}^{a}T_i(X_{\ii}^{\mathbf a})}{E_{i}^{a'}T_i(X_{\ii}^{\mathbf a'})}\\ &=\delta_{a,a'}\mu(a\alpha_i) q^{-\frac12(a+1)(\alpha_i,|\mathbf a|_\ii)}\prod_{t=1}^a (1-q_i^{-2t})\fgfrm{X_{\ii}^{{\mathbf a}}}{X_{\ii}^{{\mathbf a}'}}\\ &=\delta_{a,a'}\delta_{\mathbf a,\mathbf a'}\mu(|\mathbf a|)\mu(a\alpha_i)q^{-\frac12(a+1)(\alpha_i^\vee,|\mathbf a|_\ii)}\prod_{t=1}^{a}(1-q_i^{-2t}) \prod_{r=1}^m\prod_{t=1}^{a_r}(1-q_{i_r}^{-2r}) \\&=\delta_{a,a'}\delta_{\mathbf a,\mathbf a'}\mu(|(a,\mathbf a)|_{(i,\ii)})\prod_{t=1}^{a}(1-q_i^{-2t}) \prod_{r=1}^m\prod_{t=1}^{a_r}(1-q_{i_r}^{-2r}), \end{align*} since $|(a,\mathbf a)|_{(i,\ii)}=a\alpha_i+s_i(|\mathbf a|_\ii)$. Finally, \eqref{eq:dual PBW} is immediate from~\eqref{eq:X i^a scalar square} and \eqref{eq:mu-prop}. \end{proof} Set $U^\ZZ(w)=U^\ZZ(\lie n^+)\cap U_q(w)$ where $U_q(w)$ is defined by~\eqref{eq:quantum-Schubert-cell}. \begin{proposition}\label{prop:U(w)-U(ii)} For each $\ii\in R(w)$, $\{X_\ii^{\mathbf a}\}_{\mathbf a\in\ZZ_{\ge 0}^m}$ is an $\AA_0$-basis of~$U^\ZZ(w)$. In particular, $U^\ZZ(w)=U^\ZZ(\ii)$ for all $w\in W$, $\ii\in R(w)$. \end{proposition} \begin{proof} Since $U_q(w)=U_q(\ii)$ by \cite{T}*{Proposition~2.10}, for any $x\in U^\ZZ(w)$ we can write $x=\sum_{\mathbf a'} c_{\mathbf a'} X_\ii^{\mathbf a'}$ where $c_{\mathbf a'}\in\kk$. Since $\fgfrm{x}{X_\ii^{\mathbf a}}\in \AA_0$, it follows from~\eqref{eq:dual PBW} that $c_{\mathbf a}\in \AA_0$ for all $\mathbf a\in\ZZ_{\ge 0}^m$. Thus, the $X_\ii^{\mathbf a}$, $\mathbf a\in\ZZ_{\ge 0}^m$ generate $U^\ZZ(w)$ as an $\AA_0$-module. Since they are already linearly independent over~$\kk$, they form its $\AA_0$-basis. \end{proof} \begin{theorem} \label{thm:two-forms} Let $w\in W$, $\ii=(i_1,\dots,i_m)\in R(w)$. Then the algebra $U^\AA(w)=U^\ZZ(w)\tensor_{\AA_0}\AA$ has the following presentation \begin{equation}\label{eq:straight-rels'} q^{-\frac12(\alpha_\ii^{(k)},\alpha_\ii^{(l)})}X_{\mathbf i,l}X_{\mathbf i,k}- q^{\frac12(\alpha_\ii^{(k)},\alpha_\ii^{(l)})} X_{\mathbf i,k}X_{\mathbf i,l}\in (q_{i_k}-q_{i_k}^{-1})\sum_{\mathbf a =(0,\dots,0,a_{k+1},\dots,a_{l-1},0,\dots,0)\in\ZZ_{\ge 0}^m}\mskip-10mu \AA_0 X_{\ii}^{\mathbf a}, \end{equation} for all $1\le k<l\le m$ \end{theorem} \begin{proof} We need the following \begin{lemma}\label{lem:commutator} Let $w'\in W$ and $i,j\in I$ be such that $\ell(s_iw's_j)=\ell(w')+2$. Then \begin{equation}\label{eq:q-comm-E_i} T_{s_iw'}(E_j)E_i-q^{-(\alpha_i,w'\alpha_j)}E_i T_{s_iw'}(E_j)\in (q_i-q_i^{-1}) T_i(U^\AA(w')). \end{equation} \end{lemma} \begin{proof} First we prove that \begin{equation}\label{eq:comm-F_i} K_i[F_i,T_{w'}(E_j)]\in \la 1\ra_{q_i} U^\AA(w'). \end{equation} Our assumption implies that $\ell(s_iw')=\ell(w')+1$, $\ell(w's_j)=\ell(w')+1$, whence $T_{w'}(E_j),T_{s_iw'}(E_j)\in U_q(\lie n^+)$ by~\cite{Lus-book}*{Lemma~40.1.2} and so $T_{w'}(E_j)\in\ker\partial_i$ by~\cite{Lus-book}*{Proposition~38.1.6}. Moreover, by Proposition~\ref{prop:X i^a scalar square} we have $T_{w'}(E_j),T_{s_iw'}(E_j)\in U^\AA(\lie n^+)$. Then $$ K_i[F_i,T_{w'}(E_j)]=-(1-q_i^{-2})q^{-\frac12(\alpha_i,w'\alpha_j)}\partial_i^{op}(T_{w'}(E_j))\in \la 1\ra_{q_i} U^\AA(\lie n^+), $$ where we used \eqref{eq:partial-def}, Lemma~\ref{lem:T_w(E_j)}, Proposition~\ref{prop:T_i lattice} and Corollary~\ref{cor:preserv-lattice}\eqref{cor:preserve-lattice.b}. On the other hand, $T_{w'}^{-1}(F_i)\in U_q(\lie n^-)$, whence $[T_{w'}^{-1}(F_i),E_j]\in U_q(\lie b^-)$. Therefore, $$ T_{w'}^{-1}(K_i[F_i,T_{w'}(E_j)])=T_{w'}^{-1}(K_i)[T_{w'}^{-1}(F_i),E_j]\in U_q(\lie b^-). $$ Thus, \begin{equation*} K_i[F_i,T_{w'}(E_j)]\in \la 1\ra_{q_i} T_{w'}(U_q(\lie b^-))\cap U^\AA(\lie n^+)=\la 1\ra_{q_i} U^\AA(w'). \end{equation*} This proves~\eqref{eq:comm-F_i}. Since $T_i(K_i F_i)=q_i^{-1}E_i$ sand $K_iT_{s_iw'}(E_j)K_i^{-1}=q^{-(\alpha_i,w'\alpha_j)}T_{s_iw'}(E_j)$, \eqref{eq:q-comm-E_i} follows by applying $T_i$ to both sides of~\eqref{eq:comm-F_i}. \end{proof} Now we use induction on~$\ell(w)$, the induction base being trivial. Applying $T_{i_1}\cdots T_{i_{k-1}}$ to~\eqref{eq:q-comm-E_i} with $w'=s_{i_{k+1}}\cdots s_{i_{l-1}}$, $i=i_k$, $j=i_l$ we obtain $$ X_lX_k-q^{-(\alpha_{i_k},s_{i_{k+1}}\cdots s_{i_{l-1}}\alpha_{i_l})}X_kX_l\in \la 1\ra_{q_{i_k}} T_{i_1}\cdots T_{i_{k}}(U^\AA(w')). $$ By Proposition~\ref{prop:U(w)-U(ii)}, $U^\AA(w')$ has an $\AA$-basis $\{X_{\mathbf i',1}^{a_{k+1}}\dots X_{\mathbf i',l-1}^{a_{l-1}}\,:\,a_{k+1},\dots,a_{l-1}\in \ZZ_{\ge 0}\}$ where $\ii'=(i_{k+1},\dots,i_{l-1})$. Applying $T_{i_1}\cdots T_{i_k}$ we conclude that $\{X_{\mathbf i,k+1}^{a_{k+1}}\dots X_{\mathbf i,l-1}^{a_{l-1}}\,:\,a_{k+1},\dots,a_{l-1}\in \ZZ_{\ge 0}\}$ is an $\AA$-basis of $T_{i_1}\cdots T_{i_k}(U^\AA(w'))$. Note that $(\alpha_{i_k},s_{i_{k+1}}\cdots s_{i_{l-1}}\alpha_{i_l})=-(\alpha_\ii^{(k)},\alpha_\ii^{(l)})$ and so we can write $$ q^{-\frac12(\alpha_\ii^{(k)},\alpha_\ii^{(l)})}X_{\mathbf i,l}X_{\mathbf i,k}- q^{\frac12(\alpha_\ii^{(k)},\alpha_\ii^{(l)})} X_{\mathbf i,k}X_{\mathbf i,l}\in \sum_{\mathbf a =(0,\dots,0,a_{k+1},\dots,a_{l-1},0,\dots,0)\in\ZZ_{\ge 0}^m} \la 1\ra_{q_{i_k}}c_{\mathbf a} X_{\ii}^{\mathbf a}, $$ where $c_{\mathbf a}\in \AA$. Repeating the argument from the proof of Proposition~\ref{prop:U(w)-U(ii)} we conclude that $\la 1\ra_{q_{i_k}}c_{\mathbf a}\in \AA_0$. Thus, $c_{\mathbf a}\in \AA\cap (\la 1\ra_{q_{i_k}})^{-1}\AA_0=\AA_0$. Since relations~\eqref{eq:straight-rels'} imply that $U^\AA(w)$ is generated, as an $\AA$-module, by the $X_\ii^{\mathbf a}$, $\mathbf a\in\ZZ_{\ge 0}^m$, it follows that~\eqref{eq:straight-rels'} is a presentation. \end{proof} \begin{remark}\label{rem:2} Let $A(w)$ be the $\ZZ$-algebra defined by $A(w)=U^\ZZ(w)/(q-1)U^\ZZ(w)$. Clearly, $A(w)$ is commutative and identifies with the coordinate algebra $\ZZ[U(w)]$, where $U(w)=U\cap w(U^-)w^{-1}$ is the Schubert cell in the maximal unipotent subgroup~$U$ of the Kac-Moody group~$G$ corresponding to~$\lie g$. This justifies~\eqref{eq:quantum-Schubert-cell} and the name quantum Schubert cell used for~$U_q(w)$. \end{remark} \subsection{Lusztig's Lemma and proof of Theorem~\ref{thm:bas-from-Lusztig-lemma}}\label{subs:pf-bas-from-Lus-lemma} Let $w\in W$, $\ii=(i_1,\dots,i_m)\in R(w)$ and let $\mathbf e_1,\dots,\mathbf e_m$ be the standard basis of~$\ZZ^m$. For each pair $1\le k<l\le m$ with $k+1\le l-1$ define $\mathcal A_{k,l}=\mathcal A_{k,l}(\ii)$ to be the finite set of all tuples $(a_{k+1},\dots,a_{l-1})$ such that $X_{\mathbf i,k+1}^{a_{k+1}}\cdots X_{\mathbf i,l-1}^{a_{l-1}}$ occurs in the right hand side of~\eqref{eq:straight-rels'} with a non-zero coefficient. Let $C_{\mathbf i}$ be the submonoid of $\ZZ^m$ generated by elements $$ \mathbf e_k+\mathbf e_{l}-\sum_{r=k+1}^{l-1} a_r \mathbf e_r $$ for all $1\le k,l\le m$ with $k+1\le l-1$ such that $\mathcal A_{kl}\not=\emptyset$ and for all $(a_{k+1},\dots,a_{l-1})\in \mathcal A_{kl}$. \begin{proposition}\label{prop:order} $C_{\mathbf i}$ is pointed, that is, if $\mathbf x,-\mathbf x\in C_{\mathbf i}$ then $\mathbf x=0$. In particular, the relation $\prec$ on $\mathbb Z^m_{\ge 0}$ defined by $$ \mathbf a\preceq \mathbf a'\iff \mathbf a'-\mathbf a\in C_{\mathbf i} $$ is a partial order. \end{proposition} \begin{proof} The first assertion is a special case of the following \begin{lemma} For each $k<l$ fix $\mathcal A_{k,l}\subset \Big(\bigoplus\limits_{i=k+1}^{l-1} \ZZ_{\ge 0}\mathbf e_i\Big)\setminus\{0\}$. Let $\Gamma$ be the submonoid of $\ZZ^m$ generated by all elements of the form $\mathbf e_k+\mathbf e_l-\mathbf a$, $\mathbf a\in\mathcal A_{k,l}$ for all $k<l$ such that $\mathcal A_{k,l}\not=\emptyset$. Then $\Gamma$ is pointed. \end{lemma} \begin{proof} Let $\mathbf y=\sum_{k<l}\sum_{\mathbf a\in\mathcal A_{k,l}} n_{k,l,\mathbf a_{k,l}}(\mathbf e_k+\mathbf e_l-\mathbf a_{k,l})$ where $n_{k,l,\mathbf a_{k,l}}\in\ZZ_{\ge 0}$ and are not all zero. Let $k$ be minimal such that $n_{k,l,\mathbf a}\not=0$ for some $l>k$, $\mathbf a\in\mathcal A_{k,l}$. Then the coefficient of~$\mathbf e_k$ in~$\mathbf y$ is positive. This immediately implies that $0$ admits a unique presentation in~$\Gamma$. \end{proof} To prove the second assertion, note that the relation~$\prec$ is clearly transitive. Furthermore, if $\mathbf a'\prec \mathbf a$ and $\mathbf a\prec\mathbf a'$ then $\mathbf a'-\mathbf a,\mathbf a-\mathbf a'\in C_{\mathbf i}$ which implies that $\mathbf a=\mathbf a'$. \end{proof} Since $T_w$ commutes with $\bar\cdot$-anti-involution, $\overline{U_q(w)}=U_q(w)$ and $\overline{X_{\ii,k}}=X_{\ii,k}$. Since also $\overline{U^\ZZ(\lie n^+)}=U^\ZZ(\lie n^+)$, it follows that $\overline{U^\ZZ(w)}=U^\ZZ(w)$. Thus, the restriction of $\bar\cdot$ to $U^\ZZ(\ii)$ is the unique anti-linear anti-involution of that algebra fixing its generators $X_{\ii,k}$. Note that for each $\gamma\in Q^+$, the set $\{ \mathbf a\in\ZZ_{\ge 0}^m\,:\, |\mathbf a|_\ii=\gamma\}$ is finite. The following result is crucial for the proof of Theorem~\ref{thm:bas-from-Lusztig-lemma}. \begin{proposition}\label{prop:initial basis} For all $\mathbf a\in\ZZ_{\ge 0}^m$ we have $ \overline{X_\ii^{\mathbf a}}-X_\ii^{\mathbf a}\in\sum_{\mathbf a'\prec\mathbf a} \AA_0 X_\ii^{\mathbf a'}$. \end{proposition} \begin{proof} We need some notation. Let $\mathcal U=U^\ZZ(\ii)$ and let $\mathcal I=[1,m]$. Let $\mathcal B$ be the set of all finite non-decreasing sequences in~$\mathcal I$. Given a sequence ${\mathbf k}=(k_1,\dots,k_N)\in \mathcal I^N$, let $\mathbf e_{\mathbf k}=\sum_{r=1}^N \mathbf e_{k_r}$ and define $$ X({\mathbf k})=q^{\frac12\sum_{1\le r<s\le N} \limits\sign (k_s-k_r)(\alpha^{(k_r)},\alpha^{(k_s)})} X_{\ii,k_1}\cdots X_{\ii,k_N}. $$ In particular, if ${\mathbf k}=(k_1,\dots,k_N)\in\mathcal B$ and $a_k=\#\{ 1\le r\le N\,:\, k_r=k\}$ then $X({\mathbf k})=X_\ii^{\mathbf a}$. Given $\mathbf a\in\ZZ_{\ge 0}^m$, set $$ \mathcal U_{\prec \mathbf a}=\sum_{\mathbf a'\prec\mathbf a}\AA_0 \cdot X_\ii^{\mathbf a'}=\sum_{{\mathbf k}\in\mathcal B\,:\, \mathbf e_{\mathbf k}\prec \mathbf a} \AA_0 \cdot X({\mathbf k}), \qquad \mathcal U'_{\prec \mathbf a}:=\sum_{N\ge 0,\,{\mathbf k}\in \mathcal I^N:\mathbf e_{\mathbf k}\prec\mathbf a} \AA_0\cdot X({\mathbf k}) $$ with the convention that $\mathcal U_{\prec\mathbf a}={\mathcal U}'_{\prec \mathbf a}=\{0\}$ if $\mathbf a$ is minimal with respect to $\prec$. Clearly, both are increasing filtration on $\mathcal U$. Note following immediate \begin{lemma}\label{lem:finite poset} If $\mathbf a'\prec \mathbf a$, $\mathbf a,\mathbf a'\in\ZZ_{\ge 0}^m$ then $|\mathbf a'|_\ii=|\mathbf a|_\ii$. In particular, $\mathcal U_{\prec \mathbf a}$, $\mathcal U'_{\prec \mathbf a}$ are finite dimensional. \end{lemma} \begin{lemma} \label{lem:recursive base transition} For any sequence ${\mathbf k}=(k_1,\ldots,k_N) \in \mathcal I^N$, $N\ge 0$ and for any $\sigma\in S_N$ we have \begin{equation} \label{eq:monomial permutation} X(\sigma({\mathbf k}))-X({\mathbf k})\in {\mathcal U}'_{\prec \mathbf e_{\mathbf k}}, \end{equation} where $\sigma({\mathbf k})=(k_{\sigma(1)},\dots,k_{\sigma(N)})$. \end{lemma} \begin{proof} Clearly, it suffices to prove the assertion for a transposition~$\sigma=(r,r+1)$. Without loss of generality we may assume that $k_r<k_{r+1}$. Let ${\mathbf k}_r^-=(k_1,\ldots,k_{r-1})$, ${\mathbf k}_r^+=(k_{r+2},\ldots,k_N)$. Then the relation \eqref{eq:straight-rels'} taken with $k=k_r$, $l=k_{r+1}$ implies \begin{equation} \label{eq:transposition monomial} X_{\sigma({\mathbf k})}=X_{{({\mathbf k}_r^-,k_{r+1},k_r,{\mathbf k}_r^+)}}= X_{\mathbf k}+\sum_{{\mathbf k}'\in {\mathcal B}\,:\,\mathbf e_{{\mathbf k}'}\prec \mathbf e_{i_r}+\mathbf e_{i_{r+1}}} c_{{\mathbf k}'} X_{({\mathbf k}_r^-,{\mathbf k}',{\mathbf k}_r^+)},\qquad c_{{\mathbf k}'}\in \AA_0. \end{equation} Clearly, $\mathbf e_{({\mathbf k}_r^-,{\mathbf k}',{\mathbf k}_r^+)}=\mathbf e_{{\mathbf k}_r^-}+\mathbf e_{{\mathbf k}'}+\mathbf e_{{\mathbf k}_r^+}\prec \mathbf e_{\mathbf k}$ for all ${\mathbf k}'\in\mathcal B$ such that $\mathbf e_{{\mathbf k}'}\prec \mathbf e_{k_r}+ \mathbf e_{k_{r+1}}$. This implies that each $X_{({\mathbf k}_r^-,{\mathbf k}',{\mathbf k}_r^+)}$ in the right hand side of \eqref{eq:transposition monomial} belongs to ${\mathcal U}'_{\prec e_{\mathbf k}}$ and we obtain \eqref{eq:monomial permutation} for $\sigma=(r,r+1)$. \end{proof} \begin{lemma} \label{lem:decresing filtration A strict} ${\mathcal U}_{\prec \mathbf a}={\mathcal U}'_{\prec\mathbf a}$ for all $\mathbf a\in\ZZ_{\ge 0}^m$. \end{lemma} \begin{proof} The inclusion ${\mathcal U}_{\prec \mathbf a}\subseteq {\mathcal U}'_{\prec \mathbf a}$ is obvious. To prove the opposite inclusion, we use induction on the partial order $\prec$ which is applicable since $\{\mathbf a'\in\ZZ_{\ge 0}^m\,:\,\mathbf a'\prec\mathbf a\}$ is finite for all~$\mathbf a\in\ZZ_{\ge 0}^m$. If $\mathbf a\in \ZZ_{\ge 0}^m$ is minimal with respect to $\prec$, then ${\mathcal U}_{\prec \mathbf a}=\{0\}$ and we have nothing to prove. Assume now that $\mathbf a$ is not minimal. Then for each ${\mathbf k}\in \mathcal I^N$, $N\ge 0$ such that $\mathbf e_{\mathbf k}\prec \mathbf a$ we have ${\mathcal U}'_{\prec \mathbf e_{\mathbf k}}={\mathcal U}_{\prec \mathbf e_{\mathbf k}}$ by the induction hypothesis. Using this and Lemma \ref{lem:recursive base transition}, we conclude that for any $\sigma\in S_N$ $$X({\mathbf k})-X(\sigma({\mathbf k}))\in {\mathcal U}_{\prec \mathbf e_{{\mathbf k}}}.$$ Taking $\sigma$ such that $\sigma({\mathbf k})\in {\mathcal B}$, that is, is non-decreasing, implies that $X({\mathbf k})\in {\mathcal U}_{\prec \mathbf a}$. \end{proof} Combining Lemmata~\ref{lem:recursive base transition} and \ref{lem:decresing filtration A strict} we obtain the following obvious corollary: \begin{corollary}\label{cor:perm-bas} For any ${\mathbf k}\in {\mathcal B}$ and any $\sigma\in S_N$, we have $ X(\sigma({\mathbf k}))-X({\mathbf k})\in {\mathcal U}_{\prec \mathbf e_{\mathbf k}}$. \end{corollary} Note that $\overline{X(\mathbf k)}=X(\mathbf k^{op})$ for any $\mathbf k\in\mathcal I^N$ where $\mathbf k^{op}$ is~$\mathbf k$ written in the reverse order, and $X(\mathbf k)=X_\ii^{\mathbf e_\mathbf k}$ for $\mathbf k\in\mathcal B$. Since for any $\mathbf a\in\ZZ_{\ge 0}$ there exists a unique $\mathbf k\in\mathcal B$ such that $\mathbf e_\mathbf k=\mathbf a$, these observations together with the above Corollary complete the proof Proposition~\ref{prop:initial basis}. \end{proof} Proposition~\ref{prop:initial basis} implies that for each $\gamma\in Q^+$ the assumptions of~\cite{BZ}*{Theorem~1.1} with ${(L,\prec)}=(\{\mathbf a\in\ZZ_{\ge 0}^{m}\,:\,|\mathbf a|_\ii=\gamma\},\prec)$ and $v=q^{-1}$ are satisfied. The assertion of Theorem~\ref{thm:bas-from-Lusztig-lemma} now follows. \qed Note the following useful fact, which is immediate from the proof of Proposition~\ref{prop:initial basis}. \begin{corollary}\label{cor:skew-symm form} Define $\Lambda=\Lambda_\ii:\ZZ^m\tensor_\ZZ \ZZ^m\to\ZZ$ by $\Lambda(\mathbf e_k,\mathbf e_l)=\sign(l-k)(\alpha_\ii^{(k)},\alpha_\ii^{(l)})$. Then for all $\mathbf a,\mathbf a'\in\ZZ_{\ge 0}^m$ \begin{gather*} X_\ii^{\mathbf a}X_\ii^{\mathbf b}-q^{-\frac12\Lambda(\mathbf a,\mathbf b)}X_\ii^{\mathbf a+\mathbf b}\in q^{-\frac12\Lambda(\mathbf a,\mathbf b)}\sum_{\mathbf a'\prec\mathbf a+\mathbf b} \AA_0 X_\ii^{\mathbf a'} \\ X_\ii^{\mathbf b}X_\ii^{\mathbf a}-q^{\Lambda(\mathbf a,\mathbf b)}X_\ii^{\mathbf a}X_\ii^{\mathbf b}\in q^{\frac12\Lambda(\mathbf a,\mathbf b)}\sum_{\mathbf a'\prec\mathbf a+\mathbf b} \AA_0 X_\ii^{\mathbf a'}. \end{gather*} \end{corollary} We note an obvious property of~$\Lambda$ which will be used in the sequel. \begin{lemma}\label{lem:Lambda pairing} For any $1\le k\le m$, $\mathbf a=(a_1,\dots,a_m)\in\ZZ^m$ we have $ \Lambda_\ii(\mathbf e_k,\mathbf a)=(\alpha_\ii^{(k)},|\mathbf a_{>k}|_\ii-|\mathbf a_{<k}|_\ii) $, where $\mathbf a_{<k}=\sum_{t=1}^{k-1} a_t \mathbf e_t$, $\mathbf a_{>k}=\sum_{t=k+1}^{m} a_t \mathbf e_t$. \end{lemma} \subsection{Containment of \texorpdfstring{$\mathbf B(\ii)$}{B(i)} in~\texorpdfstring{$\mathbf B^{up}$}{Bup} and proof of Theorem~\ref{thm:inclusion of base}} \label{subs:pf inclusion of base} Let $w\in W$, $\mathbf i=(i_1,\dots,i_m)\in R(w)$. Let $\gamma=|\mathbf a|_\ii$. Since $b_{\mathbf i,\mathbf a}\in X_{\mathbf i}^{\mathbf a}+\sum\limits_{\mathbf a'\not=\mathbf a,\,|\mathbf a'|_\ii=\gamma}K_- X_{\ii}^{\mathbf a'}$, it follows from~\eqref{eq:X i^a scalar square} that $$ \mu(\gamma)^{-1}\fgfrm{b_{\ii,\mathbf a}}{b_{\ii,\mathbf a}}\in \mu(\gamma)^{-1}\fgfrm{X_{\ii}^{\mathbf a}}{X_{\ii}^{\mathbf a}} +\sum_{\mathbf a'\in\ZZ_{\ge 0}^m\setminus\{\mathbf a\}\,:\,|\mathbf a'|=\gamma} K_-\mu(\gamma)^{-1}\fgfrm{X_{\ii}^{\mathbf a'}}{X_{\ii}^{\mathbf a'}} \in 1+K_-. $$ Since $b_{\mathbf i,\mathbf a}\in U^\ZZ(\lie n^+)$ and $\overline{b_{\mathbf i,\mathbf a}}=b_{\mathbf i,\mathbf a}$, it follows from~\eqref{eq:def-bas} that $b_{\ii,\mathbf a}\in \mathbf B^{\pm up}$. To prove that~$b_{\ii,\mathbf a}\in\mathbf B^{up}$, we use induction on~$m$. The induction base is trivial. For the inductive step, write $X_\ii^{\mathbf a}=\sum_{b\in \mathbf B^{up}} c_{\mathbf a,b} b$. Since $\pm b_{\ii,\mathbf a}\in \mathbf B^{up}$, it follows that $c_{\mathbf a,b}\in K_-$ for all $b\not=b_0=\pm b_{\ii,\mathbf a}$ and $c_{\mathbf a,b_0}=\pm 1$. Thus, we only need to prove that $c_{\mathbf a,b_0}=1$ for some $b_0\in \mathbf B^{up}$. Let $i=i_1$ and $a=a_1$. Since $ X_\ii^{\mathbf a}= q^{-\frac12a(\alpha_{i},|\mathbf a'|_{\ii'})}E_{i}^{a}T_{i}(X_{\ii'}^{\mathbf a'}) $ where $\ii'=(i_2,\dots,i_m)$, $\mathbf a'=(a_2,\dots,a_m)$, $T_{i}(X_{\ii'}^{\mathbf a'})\in\ker\partial_i^{op}$ and $(\partial_i^{op})^{(top)}(E^a) =(\partial_i^{op})^{(a)}(E^a)=1$, we have $$ (\partial_i^{op})^{(top)}(X_\ii^{\mathbf a})=(\partial_i^{op})^{(a)}(X_\ii^{\mathbf a})=T_i(X_{\ii'}^{\mathbf a'}) =\sum_{b\in \mathbf B^{up}\,:\,\ell_i(b^*)=a} c_{\mathbf a,b} (\partial_i^{op})^{(top)}(b), $$ where we used Corollary~\ref{cor:top-decomp-B up}. Since $T_i^{-1}((\partial_i^{op})^{(top)}(b))\in\mathbf B^{up}$ for any $b\in\mathbf B^{up}$ by Theorem~\ref{thm:T_i bas}, we obtain from the above that $$ X_{\ii'}^{\mathbf a'}=T_i^{-1}((\partial_i^{op})^{(top)}(X_\ii^{\mathbf a})) =\sum_{b\in \mathbf B^{up}\,:\, \ell_i(b^*)=a } c_{\mathbf a,b} T_i^{-1}((\partial_i^{op})^{(top)}(b)) $$ is the decomposition of~$X_{\ii'}^{\mathbf a'}$ with respect to~$\mathbf B^{up}$. By the induction hypothesis, $b_{\ii',\mathbf a''}\in\mathbf B^{up}$ for all $\mathbf a''\in\ZZ_{\ge0}^{m-1}$ and therefore precisely one of the $c_{\mathbf a,b}$, $\ell_{i}(b^*)=a$ is not in~$K_-$ and is equal to~$1$. \qed \begin{remark} Note that for any $w\in W$, $\ii\in R(w)$, $1\le k\le \ell(w)$ and $a\ge 0$ we have $X_{\ii,k}^a\in\mathbf B^{up}$. \end{remark} \subsection{Embeddings of bases and proof of Theorem~\ref{thm:T_w(B(w'))}}\label{subs:pf T_w(B(w'))} Note that $U_q(w)\subset U_q(ww')$. Since $\mathbf B(w)=U_q(w)\cap \mathbf B^{up}$ and $\mathbf B(ww')=U_q(ww')\cap \mathbf B^{up}$, the first assertion follows. To establish the second assertion, it suffices to prove that for $i\in I$ such that $\ell(s_iw)=\ell(w)+1$ we have $T_i(\mathbf B(w))\subset\mathbf B(s_iw)$. The assumption implies that $T_i(\mathbf B(w))\subset U_q(\lie n^+)$ and therefore is contained in~$\mathbf B^{up}$ by Theorem~\ref{thm:T_i bas}. Since $T_i(U_q(w))\subset U_q(s_i w)$, it follows that $T_i(\mathbf B(w))\subset U_q(s_i w)\cap \mathbf B^{up}=\mathbf B(s_i w)$.\qed \section{Examples}\label{sec:examples} In this section we compute bases $\mathbf B(w)$ for various Schubert cells $U_q(w)$. We denote by $E_{i_1^{a_1}\cdots i_r^{a_r}}$ the unique element~$b$ of~$\mathbf B^{up}$ for which $\partial_{\ii}^{(top)}(b)=\partial_{i_r}^{(a_r)}\cdots\partial_{i_1}^{(a_1)}(b)=1$ where $\ii=(i_1,\dots,i_r)$. Note that this element also satisfies $(\partial_{\ii^{op}}^{op})^{(top)}(b)=(\partial_{i_1}^{op})^{(a_1)}\cdots (\partial_{i_r}^{op})^{(a_r)}(b)=1$. We use the notation from~\S\ref{subs:pf-bas-from-Lus-lemma}. \subsection{Bases for repetition free elements}\label{subs:rep-free} We say that $w\in W$ is repetition-free if $w=s_{i_1}\dots s_{i_m}$ where $\ii=(i_1,\dots,i_m)\in R(w)$ is repetition free. Clearly, if $w$ is repetition free then so is each $\ii\in R(w)$. Such an element is called a {\em Coxeter element} if $\ell(w)=|I|$, that is, any $\ii\in R(w)$ is an ordering of~$I$. \begin{lemma}\label{lem:Coxeter} Let $w\in W$ be repetition free and let $\ii \in R(w)$. Then in the notation of~\S\ref{subs:some prop q S c}: \begin{enumerate}[{\rm(a)}] \item\label{lem:Coxeter.a} $U_q(w)$ is a quantum plane of rank~$\ell(w)$ with presentation \begin{equation}\label{eq:Coxeter.pres} q^{-\frac12(\alpha_\ii^{(k)},\alpha_\ii^{(l)})}X_{\ii,l}X_{\ii,k}=q^{\frac12(\alpha_\ii^{(k)},\alpha_\ii^{(l)})}X_{\ii,k}X_{\ii,l},\qquad 1\le k<l\le \ell(w). \end{equation} \item\label{lem:Coxeter.b} $\mathbf B(w)=\{ X_\ii^{\mathbf a}\,:\,\mathbf a\in\ZZ_{\ge 0}^{\ell(w)}\}$. \item\label{lem:Coxeter.c} $X_{\ii,k}=E_{i_1^{m_{k1}}\cdots i_{k-1}^{m_{k,k-1}}i_k}=\ul E_{i_1}^{(m_{k1})}\cdots \ul E_{i_{k-1}}^{(m_{k,k-1})}(E_{i_k})$ where $m_{kr}= -(\alpha_{i_r}^\vee,s_{i_{r+1}}\cdots s_{i_{k-1}}(\alpha_{i_k}))=d_{i_r}^{-1}(\alpha_\ii^{(k)},\alpha_\ii^{(r)})$. \end{enumerate} \end{lemma} \begin{proof} Note that the coefficient of $\alpha_{i_k}$ in every element of the submonoid of $Q^+$ generated by $\alpha_\ii^{(r)}$, $k<r<l$ is zero. Since the algebra $U_q(w)$ is $Q^+$-graded, it follows that the right hand side of~\eqref{eq:straight-rels'} is zero. This proves part~\eqref{lem:Coxeter.a}. In particular, it follows that $\overline{X_\ii^{\mathbf a}}=X_{\ii}^{\mathbf a}$ for all $\mathbf a\in\ZZ_{\ge 0}^{\ell(w)}$, hence $b_{\ii,\mathbf a}=X_{\ii}^{\mathbf a}$. To prove~\eqref{lem:Coxeter.b} it remains to apply Theorems~\ref{thm:bas-from-Lusztig-lemma} and~\ref{thm:inclusion of base}. To prove part~\eqref{lem:Coxeter.c}, let $u_r=T_{i_{r+1}}\cdots T_{i_{k-1}}(E_{i_k})$ and observe that the coefficient of~$\alpha_{i_r}$ in~$\deg u_r=s_{i_{r+1}}\cdots s_{i_{k-1}}(\alpha_{i_k})$ is zero if~$\ii$ is repetition free. Therefore, $u_r\in {}_{i_r} U\cap U_{i_r}$, $T_{i_r}(u_r)=\ul E_i^{(-(\alpha_{i_r}^\vee,\deg u_r))}(u_r)$ by Theorem~\ref{thm:T_i formula} and so $\ell_i(T_{i_r}(u_r))=-(\alpha_{i_r}^\vee,\deg u_r)$. The assertion now follows by induction on $k-r$. \end{proof} \begin{remark} The assertion of Lemma~\ref{lem:Coxeter}\eqref{lem:Coxeter.a} holds for any $w\in W$, $\ii\in R(w)$ and $1\le k<l\le \ell(w)$ such that the subsequence $(i_k,\dots,i_l)$ is repetition free. \end{remark} \subsection{Bases for elements with a single repetition}\label{subs:single rep} We say that $w\in W$ is an element with a single repetition if there exists $\ii=(i_1,\dots,i_m)\in R(w)$ with $i_k\not=i_l$, $k<l$ unless $k=r$ and $l=r'$ for some $1\le r<r'\le m$. Clearly, all $\ii'\in R(w)$ have that property. \begin{proposition}\label{prop:single repetition presentation} Let $w\in W$ be an element with a single repetition and let $\ii=(i_1,\dots,i_m)\in R(w)$, where the $i_k$, $k\not=r,r'$, $1\le k\le m$ are distinct and $i_r=i_{r'}=i$, $1\le r<r'\le m$. Then $U_q(w)$ is generated by the $X_{\ii,k}$, $1\le k\le m$ where \begin{equation}\label{eq:string-nomenclature-single rep} X_{\ii,k}=\begin{cases} E_{i_1^{m_{k,1}}\cdots i_{k-1}^{m_{k,k-1}}i_k},&k\not=r'\\ E_{i_1^{m_{r',1}}\cdots i_{r-1}^{m_{r',r-1}}i_r^{1+m_{r',r}}\cdots i_{r'-1}^{m_{r',r'-1}}},&k=r' \end{cases} \end{equation} with $m_{kl}= -(\alpha_{i_l}^\vee,s_{i_{l+1}}\cdots s_{i_{l-1}}(\alpha_{i_l}))=d_{i_l}^{-1}(\alpha_\ii^{(k)},\alpha_\ii^{(l)})$, subject to the relations \begin{equation}\label{eq:single repetition presentation} \begin{aligned} &q^{-\frac12(\alpha_\ii^{(k)},\alpha_\ii^{(l)})}X_{\ii,l}X_{\ii,k}=q^{\frac12(\alpha_\ii^{(k)},\alpha_\ii^{(l)})}X_{\ii,k}X_{\ii,l},\qquad 1\le k<l\le \ell(w), \, k\not=r,\,l\not=r' \\ &q^{-\frac12(\alpha_\ii^{(r')},\alpha_\ii^{(r)})}X_{\mathbf i,r'}X_{\mathbf i,r}= q^{\frac12(\alpha_\ii^{(r)},\alpha_\ii^{(r')})} X_{\mathbf i,r}X_{\mathbf i,r'}+(q_{i}-q_{i}^{-1})X_\ii^{\mathbf n(r,r')}, \quad \mathbf n(r,r')=-\sum_{k=r+1}^{r'-1} a_{i_ki}\mathbf e_k. \end{aligned} \end{equation} \end{proposition} \begin{proof} Clearly, the sequences $(i_1,\dots,i_{r'-1})$ and $(i_{r+1},\dots,i_m)$ are repetition free. In particular, for $1\le k\le r'-1$ we have $X_{\ii,k}=E_{i_1^{m_{k1}}\cdots i_{k-1}^{m_{k,k-1}}i_k}$ by Lemma~\ref{lem:Coxeter}\eqref{lem:Coxeter.c}. Furthermore, $$ X_{\ii,r'}=T_{i_1}\cdots T_{i_{r-1}}T_iT_{i_{r+1}}\cdots T_{i_{r'-1}}(E_i)=T_{i_1}\cdots T_{i_{r-1}}T_i( E_{i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}i}) $$ where $i=i_r=i_{r'}$. Clearly, $(\partial_i^{op})^2(E_{i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}i})=0$, hence $$ E_{i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}i}=(2+m_{r',r})_{q_i}{}^{-1}\ul E_i^{op}(E_{i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}})+ x_0$$ where $x_0\in {}_i U\cap U_i$. This implies that $$ T_i(E_{i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}i})=(2+m_{r',r})_{q_i}^{-1} \ul E_i^{(1+m_{r',r})}(E_{i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}}) +\ul E_i^{(m_{r',r})}(x_0), $$ and so $T_i(E_{i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}i})=E_{i^{1+m_{r',r}} i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}}$, whence $$ X_{\ii,r'}=E_{i_1^{m_{r',1}}\cdots i_{r-1}^{m_{r',r-1}}i^{1+m_{r',r}} i_{r+1}^{m_{r',r+1}}\cdots i_{r'-1}^{m_{r',r'-1}}}. $$ Since the sequence $(i_{r+1},\dots,i_k)$, $r'+1\le k\le m$ is repetition free, we have $T_{i_{r+1}}\cdots T_{i_{k-1}}(E_{i_k})=E_{i_{r+1}^{m_{k,r+1}}\cdots i_{k-1}^{m_{k,k-1}}i_k}$ by Lemma~\ref{lem:Coxeter}\eqref{lem:Coxeter.c} and hence is in~${}_i U\cap U_i$. Then $X_{\ii,k}=E_{i_1^{m_{k,1}}\cdots i_{k-1}^{m_{k,k-1}}i_k}$ by Theorem~\ref{thm:T_i formula}. This proves~\eqref{eq:string-nomenclature-single rep}. The first identity in~\eqref{eq:single repetition presentation} is proved similarly to~\eqref{eq:Coxeter.pres}. To prove the second, we need the following combinatorial fact similar to~\cite{BR}*{Lemma~4.8}. \begin{lemma}\label{lem:root-comb} Let $w\in W$ and suppose that $\ii=(i_1,\dots,i_m)\in R(w)$ has a single repetition $i_r=i_{r'}=i$. Then \begin{equation}\label{eq:deg-eq} \alpha_\ii^{(r)}+\alpha_\ii^{(r')}=-\sum_{k=r+1}^{r'-1} a_{i_k i}\alpha_\ii^{(k)} \end{equation} and any proper subset of $\{\alpha_\ii^{(k)}\}_{r\le k\le r'}$ is linearly independent. \end{lemma} \begin{proof} Fix $r<k\le r'$. Then \begin{multline}\label{eq:telescope.a} -\sum_{t=r+1}^{k-1} a_{i_t,i}\alpha_\ii^{(t)}=-\sum_{t=r+1}^{k-1} (\alpha_{i_t}^\vee,\alpha_i)s_{i_1}\cdots s_{i_{t-1}}(\alpha_{i_t}) =\sum_{t=r+1}^{k-1} s_{i_1}\cdots s_{i_{t-1}}(s_{i_t}(\alpha_i)-\alpha_i) \\ =s_{i_1}\cdots s_{i_{k-1}}(\alpha_i)-s_{i_1}\cdots s_{i_r}(\alpha_i) =s_{i_1}\cdots s_{i_{k-1}}(\alpha_i)+\alpha_\ii^{(r)}. \end{multline} The first assertion of the Lemma is now immediate. To prove the second, suppose that $\sum_{t=r}^{r'} c_t \alpha_\ii^{(t)}=0$. Using~\eqref{eq:deg-eq} we may assume that~$c_{r'}=0$ and let $r< k<r'$ be maximal such that $c_k\not=0$. Then $\alpha_{i_k}$ occurs with coefficient~$1$ in~$\alpha_\ii^{(k)}$ and does not occur in~$\alpha_\ii^{(t)}$ with~$t<k$, whence $c_k=0$ which contradicts with the choice of~$k$. \end{proof} It follows from~\eqref{eq:straight-rels'} and Lemma~\ref{lem:root-comb} that \begin{equation}\label{eq:tmp-determ-c} q^{-\frac12(\alpha_\ii^{(r')},\alpha_\ii^{(r)})}X_{\mathbf i,r'}X_{\mathbf i,r}- q^{\frac12(\alpha_\ii^{(r)},\alpha_\ii^{(r')})} X_{\mathbf i,r}X_{\mathbf i,r'}=(q_{i}-q_{i}^{-1})c X_\ii^{\mathbf n(r,r')}, \end{equation} for some $c\in \AA_0$. We may assume, without loss of generality, that $r=1$. Then $\ell_i(X_{\ii,k})=m_{k,1}$, $2\le k\le r'-1$, hence by~\eqref{eq:deg-eq} $$ \ell_i(X_{\ii}^{\mathbf n(r,r')})=-\sum_{k=2}^{r'-1} m_{k,1}a_{i_k,i}=-\sum_{k=2}^{r'-1} (\alpha_i^\vee,\alpha_\ii^{(k)})a_{i_k,i} =(\alpha_i^\vee,\alpha_i+\alpha_\ii^{(r')})= m_{r',1}+2. $$ Applying $\partial_i^{(m_{r',1}+2)}$ to both sides of~\eqref{eq:tmp-determ-c} and taking into account that $\ell_i(X_{\ii,r'})=m_{r',1}+1$ we obtain $$ \partial_i^{(top)}X_{\mathbf i,r'}=c\partial_i^{(top)}X_\ii^{\mathbf n(r,r')}. $$ Since $X_{\ii,r'}$ and $X_\ii^{\mathbf n(r,r')}$ are in~$\mathbf B^{up}$ this implies that~$c=1$. \end{proof} \begin{theorem}\label{thm:single rep basis} Let $w\in W$ and suppose that $\ii=(i_1,\dots,i_m)\in R(w)$ has a single repetition $i_r=i_{r'}=i$. Then $$ \mathbf B(w)=\{ q^{\frac12 a\Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'})}X_\ii^{\mathbf a}Y_\ii^a \,:\, \mathbf a=(a_1,\dots,a_m)\in\ZZ_{\ge 0}^{m},\,\min(a_r,a_{r'})=0,\, a\in\ZZ_{\ge 0}\} $$ where $\Lambda=\Lambda_\ii$ is defined as in Corollary~\ref{cor:skew-symm form} and \begin{equation}\label{eq:Y-defn} Y_\ii =q^{\frac12(\alpha_\ii^{(r)},\alpha_\ii^{(r')})} X_{\mathbf i,r}X_{\mathbf i,r'}-q_{i}^{-1} X_\ii^{\mathbf n(r,r')} =E_{i_1^{m_{r',1}}\cdots i_{r-1}^{m_{r',r-1}}i_r^{1+m_{r',r}}\cdots i_{r'-1}^{m_{r',r'-1}}i_1^{m_{r,1}}\cdots i_{r-1}^{m_{r,r-1}}i_r}. \end{equation} \end{theorem} \begin{proof} By~\eqref{eq:deg-eq} $Y_\ii \in U_q(\lie n^+)_{\alpha_\ii^{(r)}+\alpha_\ii^{(r')}}$. It is immediate from~\eqref{eq:single repetition presentation} that $\overline{Y_\ii }=Y_\ii $, whence $Y_\ii \in\mathbf B(w)$ by Theorems~\ref{thm:bas-from-Lusztig-lemma} and~\ref{thm:inclusion of base}. It is easy to see that $(\partial_{i_1}^{op})^{(m_{r,1})}\cdots(\partial_{i_{r-1}}^{op})^{(m_{r,r-1})}\partial_{i_r}^{op}(Y_\ii )=X_{\ii,r'}$ whence $Y_\ii =E_{i_1^{m_{r',1}}\cdots i_{r-1}^{m_{r',r-1}}i_r^{1+m_{r',r}}\cdots i_{r'-1}^{m_{r',r'-1}}i_1^{m_{r,1}}\cdots i_{r-1}^{m_{r,r-1}}i_r}$. Furthermore, we need the following \begin{lemma}\label{lem:comm-rels Y} $X_{\ii}^{\mathbf a}Y_\ii =q^{-\Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'})}Y_\ii X_{\ii}^{\mathbf a}$ for all $\mathbf a\in\ZZ_{\ge 0}^m$. \end{lemma} \begin{proof} It suffices to prove the assertion for $\mathbf a=\mathbf e_k$, $1\le k\le m$. By Corollary~\ref{cor:skew-symm form} and Lemma~\ref{lem:Lambda pairing} we have for $k\not=r,r'$ \begin{equation}\label{eq:comm-rel quantum plane} X_{\ii,k}X_\ii^{\mathbf a}=q^{-\Lambda(\mathbf e_k,\mathbf a)}X_\ii^{\mathbf a}X_{\ii,k} =q^{(\alpha_\ii^{(k)},|\mathbf a_{<k}|_\ii-|\mathbf a_{>k}|_\ii)}X_{\ii}^{\mathbf a}X_{\ii,k}. \end{equation} This immediately yields the assertion for $k<r$ or~$k>r'$. If~$r<k<r'$ then \begin{multline}\label{eq:n(r,r')><} (\alpha_\ii^{(k)},|\mathbf n(r,r')_{<k}|_\ii-|\mathbf n(r,r')_{>k}|_\ii) =(\alpha_\ii^{(k)},\alpha_\ii^{(r)}-\alpha_\ii^{(r')}+s_{i_1}\cdots s_{i_{k-1}}(s_{i_k}(\alpha_i)+\alpha_i)) \\=(\alpha_\ii^{(k)},\alpha_\ii^{(r)}-\alpha_\ii^{(r')})+(\alpha_{i_k},s_{i_k}(\alpha_i)+\alpha_i) =(\alpha_\ii^{(k)},\alpha_\ii^{(r)}-\alpha_\ii^{(r')}), \end{multline} where we used~\eqref{eq:deg-eq} and~\eqref{eq:telescope.a}. Thus, $X_{\ii,k}X_{\ii}^{\mathbf n(r,r')}=q^{(\alpha_\ii^{(k)},\alpha_\ii^{(r)}-\alpha_\ii^{(r')})}X_\ii^{\mathbf n(r,r')}X_{\ii,k}$. Since we also have $$ X_{\ii,k}X_\ii^{\mathbf e_r+\mathbf e_{r'}}=q^{(\alpha_\ii^{(k)},\alpha_\ii^{(r)}-\alpha_\ii^{(r')})}X_\ii^{\mathbf e_r+\mathbf e_{r'}}X_{\ii,k}, $$ we conclude that the assertion holds in this case. Furthermore, \begin{multline*} X_{\ii,r'}Y_\ii =q^{\frac12(\alpha_\ii^{(r)},\alpha_\ii^{(r')})} X_{\ii,r'}X_{\ii,r}X_{\ii,r'}-q_i^{-1} X_{\ii,r'}X_\ii^{\mathbf n(r,r')} \\=q^{\frac12(\alpha_\ii^{(r)},\alpha_\ii^{(r')})}(q^{(\alpha_\ii^{(r)},\alpha_\ii^{(r')})} X_{\mathbf i,r}X_{\mathbf i,r'} +q^{\frac12(\alpha_\ii^{(r')},\alpha_\ii^{(r)})}(q_{i}-q_{i}^{-1})X_\ii^{\mathbf n(r,r')})X_{\ii,r'}- q_i^{-1} q^{(\alpha_\ii^{(r')},\alpha_\ii^{(r)}+\alpha_\ii^{(r')})}X_\ii^{\mathbf n(r,r')}X_{\ii,r'} \\ =q^{(\alpha_\ii^{(r)},\alpha_\ii^{(r')})}Y_\ii X_{\ii,r'} \end{multline*} and similarly $Y_\ii X_{\ii,r}=q^{(\alpha_\ii^{(r)},\alpha_\ii^{(r')})}X_{\ii,r} Y_\ii$. Since $(\alpha_\ii^{(r)},\alpha_\ii^{(r')})= \Lambda(\mathbf e_r,\mathbf e_r+\mathbf e_{r'})=-\Lambda(\mathbf e_{r'},\mathbf e_r+\mathbf e_{r'})$ this completes the proof of Lemma~\ref{lem:comm-rels Y}. \end{proof} \begin{proposition}\label{prop:trans-matrix} For all $\mathbf a\in \ZZ_{\ge 0}^m$ we have $$ X_\ii^{\mathbf a}=\sum_{k+l=\min(a_r,a_{r'})} q_i^{-k(k+|a_r-a_{r'}|)}\qbinom[q_i^{-2}]{\min(a_r,a_{r'})}{k} b(\mathbf a-\min(a_r,a_{r'})(\mathbf e_r+\mathbf e_{r'})+k\mathbf n(r,r'),l) $$ where for $\mathbf n\in\ZZ_{\ge 0}^m$ with $\min(n_r,n'_r)=0$ we set $$ b(\mathbf n,l)=q^{\frac12 l\Lambda(\mathbf n,\mathbf e_r+\mathbf e_{r'})}X_\ii^{\mathbf n}Y_\ii^{l} $$ and $\qbinom[v]{m}{n}\in 1+v \ZZ[v]$ is the Gaussian binomial coefficient defined by $$ \qbinom[v]{m}{n}=\prod_{t=0}^{n-1} \frac{[m-t]_v}{[t+1]_v},\quad [k]_v=\sum_{l=0}^{k-1} v^l. $$ \end{proposition} \begin{proof} We need the following \begin{lemma}\label{lem:special case trans matrix} $ \displaystyle X_{\ii}^{m(\mathbf e_r+\mathbf e_{r'})}=\sum_{k+l=m} q_i^{-k^2}\qbinom[q_i^{-2}]{m}{k}X_\ii^{k\mathbf n(r,r')} Y_\ii ^{l}$ for all $m\ge 0$. \end{lemma} \begin{proof} The argument is by induction on~$m$. The case~$m=0$ is obvious. For the inductive step, note that we have, by the definition of~$Y_\ii $ \begin{multline*} X_{\ii}^{(m+1)(\mathbf e_r+\mathbf e_{r'})}=q^{\frac12(m+1)^2(\alpha_\ii^{(r)},\alpha_\ii^{(r')})} X_{\ii,r}^m X_{\ii,r}X_{\ii,r'}X_{\ii,r'}^{m} =q^{(\frac12m^2+m)(\alpha_\ii^{(r)},\alpha_\ii^{(r')})}X_{\ii,r}^m(Y_\ii +q_i^{-1}X_\ii^{\mathbf n(r,r')})X_{\ii,r'}^m \\=X_\ii^{m(\mathbf e_r+\mathbf e_{r'})}Y_\ii +q_i^{-1-2m} X_\ii^{\mathbf n(r,r')}X_\ii^{m(\mathbf e_r+\mathbf e_{r'})}. \end{multline*} By Corollary~\ref{cor:skew-symm form} we have $X_\ii^{\mathbf a}X_\ii^{k\mathbf a}=X_\ii^{(k+1)\mathbf a}$ if $\mathbf a\in\sum_{t=r+1}^{r'-1}\ZZ_{\ge 0}\mathbf e_t$, whence by the induction hypothesis \begin{align*} X_\ii^{(m+1)(\mathbf e_r+\mathbf e_{r'})}&=\sum_{k+l=m} q_i^{-k^2}\qbinom[q_i^{-2}]{m}{k}X_\ii^{k\mathbf n(r,r')}Y_\ii ^{l+1} +\sum_{k+l=m} q_i^{-k^2-1-2m}\qbinom[q_i^{-2}]{m}{k} X_\ii^{(k+1)\mathbf n(r,r')}Y_\ii ^l \\ &= \sum_{k+l=m+1} q_i^{-k^2}\Big(\qbinom[q_i^{-2}]{m}{k}+q_i^{-2(m+1-k)}\qbinom[q_i^{-2}]{m}{k-1}\Big)X_\ii^{k\mathbf n(r,r')}Y_\ii ^l \\&=\sum_{k+l=m+1} q_i^{-k^2}\qbinom[q_i^{-2}]{m+1}{k}X_\ii^{k\mathbf n(r,r')}Y_\ii ^l.\qedhere \end{align*} \end{proof} Using Corollary~\ref{cor:skew-symm form} we can write $$ X_\ii^{\mathbf a}=q^{\frac12\Lambda(\mathbf a,a_r \mathbf e_r+a_{r'}\mathbf e_{r'})}X_\ii^{\mathbf a-a_r\mathbf e_r-a_{r'}\mathbf e_{r'}} X_\ii^{a_r\mathbf e_r+a_{r'}\mathbf e_{r'}} =q^{-\frac12\Lambda(\mathbf a,a_r \mathbf e_r+a_{r'}\mathbf e_{r'})}X_\ii^{a_r\mathbf e_r+a_{r'}\mathbf e_{r'}} X_\ii^{\mathbf a-a_r\mathbf e_r-a_{r'}\mathbf e_{r'}}. $$ If $a_r\ge a_{r'}$ then \begin{multline*} X_\ii^{\mathbf a}=q^{\frac12(\Lambda(\mathbf a,a_r \mathbf e_r+a_{r'}\mathbf e_{r'})+\Lambda((a_r-a_{r'})\mathbf e_r,a_{r'}\mathbf e_{r'}))} X_\ii^{\mathbf a-a_r\mathbf e_r-a_{r'}\mathbf e_{r'}} X_\ii^{(a_r-a_{r'})\mathbf e_r}X_{\ii}^{a_{r'}(\mathbf e_r+\mathbf e_{r'})} \\ =q^{\frac12 a_{r'}\Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'})}X_\ii^{\mathbf a-a_{r'}(\mathbf e_r+\mathbf e_{r'})} X_{\ii}^{a_{r'}(\mathbf e_r+\mathbf e_{r'})}. \end{multline*} Note that $\Lambda(\mathbf e_r+\mathbf e_{r'},\mathbf n(r,r'))=0$. Then for $0\le k\le a'_r$ \begin{multline*} q^{\frac12 a_{r'}\Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'})}X_\ii^{\mathbf a-a_{r'}(\mathbf e_r+\mathbf e_{r'})} X_\ii^{k\mathbf n(r,r')}Y_\ii ^{a_{r'}-k}\\ =q^{\frac12 \Lambda(\mathbf a,a_{r'}(\mathbf e_r+\mathbf e_{r'})-k\mathbf n(r,r'))}X_\ii^{\mathbf a-a_{r'}(\mathbf e_r+\mathbf e_{r'})+ k\mathbf n(r,r')}Y_\ii ^{a_{r'}-k} \\ =q^{\frac12 k\Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'}-\mathbf n(r,r'))}b(\mathbf a-a_{r'}(\mathbf e_r+\mathbf e_{r'})+k\mathbf n(r,r'),a_{r'}-k). \end{multline*} If $t<r$ or~$t>r'$ then $$ \Lambda(\mathbf e_t,\mathbf e_r+\mathbf e_{r'}-\mathbf n(r,r'))=\pm(\alpha_i^{(t)},\alpha_\ii^{(r)}+\alpha_\ii^{(r')}-|\mathbf n(r,r')|_\ii)=0 $$ by~\eqref{eq:deg-eq}. For $r<t<r'$ it follows from~\eqref{eq:n(r,r')><} that $$ \Lambda(\mathbf e_t,\mathbf e_r+\mathbf e_{r'}-\mathbf n(r,r'))=(\alpha_\ii^{(t)},\alpha_\ii^{(r')}-|\mathbf n(r,r')_{>t}|_\ii -\alpha_\ii^{(r)}+|\mathbf n(r,r')_{<t}|_\ii)=0. $$ Since by Lemma~\ref{lem:Lambda pairing} $$ \Lambda(\mathbf e_r,\mathbf e_r+\mathbf e_{r'}-\mathbf n(r,r'))=(\alpha_\ii^{(r)},\alpha_\ii^{(r')}-\mathbf n(r,r'))= -(\alpha_\ii^{(r)},\alpha_\ii^{(r)})=-(\alpha_i,\alpha_i) $$ while $$ \Lambda(\mathbf e_{r'},\mathbf e_r+\mathbf e_{r'}-\mathbf n(r,r'))=(\alpha_\ii^{(r')},-\alpha_\ii^{(r)}+\mathbf n(r,r'))=(\alpha_\ii^{(r')},\alpha_\ii^{(r')}) =(\alpha_i,\alpha_i) $$ we conclude that $\Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'}-\mathbf n(r,r'))=(\alpha_i,\alpha_i)(a_{r'}-a_r)$. Thus, by Lemma~\ref{lem:special case trans matrix} we have $$ X_\ii^{\mathbf a}=\sum_{k+l=a_{r'}} q_i^{-k(k+a_r-a_{r'})}\qbinom[q_i^{-2}]{a_{r'}}{k} b(\mathbf a-a_{r'}(\mathbf e_r+\mathbf e_{r'})+ k\mathbf n(r,r'),l). $$ For $a_r\le a_{r'}$ we obtain in a similar way $$ X_\ii^{\mathbf a}= q^{-\frac12\Lambda(\mathbf a,a_r \mathbf e_r+a_{r'}\mathbf e_{r'})}X_\ii^{a_r\mathbf e_r+a_{r'}\mathbf e_{r'}} X_\ii^{\mathbf a-a_r\mathbf e_r-a_{r'}\mathbf e_{r'}} =q^{-\frac12a_r \Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'})} X_\ii^{a_r(\mathbf e_r+\mathbf e_{r'})} X_\ii^{\mathbf a-a_r(\mathbf e_r+\mathbf e_{r'})}. $$ Since for $0\le k\le a_r$ \begin{multline*} q^{-\frac12a_r \Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'})} X_\ii^{k\mathbf n(r,r')}Y_\ii^{a_r-k} X_\ii^{\mathbf a-a_r(\mathbf e_r+\mathbf e_{r'})} \\ =q^{-\frac12 k\Lambda(\mathbf a,\mathbf e_r+\mathbf e_{r'}-\mathbf n(r,r'))}b(\mathbf a-a_r(\mathbf e_r+\mathbf e_{r'})+k\mathbf n(r,r'),a_r-k), \end{multline*} it follows that $$ X_\ii^{\mathbf a}=\sum_{k+l=a_{r}} q_i^{-k(k+a_{r'}-a_{r})}\qbinom[q_i^{-2}]{a_{r}}{k} b(\mathbf a-a_{r}(\mathbf e_r+\mathbf e_{r'})+ k\mathbf n(r,r'),l). $$ Proposition~\ref{prop:trans-matrix} is proved. \end{proof} By Lemma~\ref{lem:comm-rels Y}, $\overline{\mathbf b(\mathbf n,l)}=\mathbf b(\mathbf n,l)$ provided that $\min(n_r,n_{r'})=0$. Then Proposition \ref{prop:trans-matrix} and Theorems \ref{thm:bas-from-Lusztig-lemma}, \ref{thm:inclusion of base} imply that $b(\mathbf a-\min(a_r,a'_r)(\mathbf e_r+\mathbf e_{r'}),\min(a_r,a'_r))=b_{\ii,\mathbf a}\in\mathbf B(w)$. Clearly this gives the $b_{\ii,\mathbf a}$ for all $\mathbf a\in\ZZ_{\ge 0}^m$, which completes the proof of Theorem~\ref{thm:single rep basis}. \end{proof} \subsection{Bases for type \texorpdfstring{$A_3$}{A\_3}}\label{subs:ex A_3} Let $w_\circ$ be the longest element in~$W$. We have $E_{ij}=T_i(E_j)$, $\{i,j\}=\{1,2\}$ or $\{2,3\}$, $E_{123}=T_1T_2(E_3)=T_3^{-1}T_2^{-1}(E_1)$, $E_{321}=T_3T_2(E_1)=T_1^{-1}T_2^{-1}(E_3)$, $E_{132}=T_1T_3(E_2)$, $E_{213}=T_2T_1T_3(E_2)=E_{132}^*=T_1^{-1}T_3^{-1}(E_2)$ and $E_{2132}=Y_{(2,1,3,2)}=E_2 E_{213}-q^{-1}E_{21}E_{23}$ as defined in Theorem~\ref{thm:single rep basis}. The following was essentially proved in~\cite{BZ1}, although with a slightly different definition of $\bar\cdot$ and hence with different powers of~$q$ (see also Theorems~1.4.1 and~3.1.3 in a recent work~\cite{Qin}). \begin{theorem} $\mathbf B^{up}=\mathbf B(w_\circ)$ consists of monomials $$ q^{\frac12 f(\mathbf a)} E_1^{m_1}E_2^{m_2}E_3^{m_3}E_{12}^{m_{12}}E_{21}^{m_{21}} E_{23}^{m_{23}}E_{32}^{m_{32}}E_{213}^{m_{213}}E_{132}^{m_{132}}E_{123}^{m_{123}}E_{321}^{m_{321}}E_{2132}^{m_{2132}} $$ where \begin{multline*} f(\mathbf a)=(m_1-m_2)(m_{12}-m_{21})+(m_3-m_2)(m_{32}-m_{23})+(m_1+m_3)(m_{132}-m_{213})\\ +(m_1+m_{12}+m_{21}-m_3-m_{23}-m_{32})(m_{123}-m_{321}) -(m_{12}+m_{32})m_{132}+(m_{21}+m_{23})m_{213}, \end{multline*} and $\min(m_\alpha,m_\beta)=0$ if $E_\alpha$, $E_\beta\notin\{ E_{123},E_{321},E_{2132}\}$ and are not connected by an edge in the following graph (see~\cite{BZ1}*{\S9.4, Fig~2}) $$ \xymatrix@!0@R=6ex@C=4ex{ &&&&&&&E_1\ar@{-}[dd]\ar@{-}[lldd]\ar@{-}[rrdd]&&&\\ \\ &&&&&E_{132}\ar@{-}[rr]\ar@{-}[rd]&&E_3\ar@{-}[ld]\ar@{-}[rd]&&E_{213}\ar@{-}[ld]\ar@{-}[ll]\\ &&&&&&E_{32}\ar@{-}[rd]\ar@{-}[rr]&&E_{23}\ar@{-}[ld]\\ &&&&&&&E_2 \\ E_{12}\ar@{-}[rrrrrrrrrrrrrr] \ar@{-}[rrrrruuu]\ar@{-}[rrrrrrruuuuu]\ar@{-}[rrrrrruu]\ar@{-}[rrrrrrru]&&&&&&&&&&&&&&E_{21}\ar @{-}[llllluuu]\ar @{-}[llllllluuuuu]\ar @{-}[lllllluu]\ar @{-}[lllllllu]} $$ \end{theorem} We have the following table for the action of the $T_i^{-1}$, $1\le i\le 3$ on the~$E_\alpha$ $$ \begin{array}{c|cccccccccccc} &E_1&E_2&E_3&E_{12}&E_{21}&E_{23}&E_{32}&E_{132}&E_{213}&E_{123}&E_{321}&E_{2132}\\\hline T_1^{-1}& &E_{21}&E_3&E_2&&E_{213}&E_{321}&E_{32}&&E_{23}&&E_{2132}\\ T_2^{-1}&E_{12}&&E_{32}&&E_1&E_3&&&E_{132}&E_{123}&E_{321}&\\ T_3^{-1}&E_1&E_{23}&&E_{123}&E_{213}&&E_2&E_{12}&&&E_{21}&E_{2132}\\ \end{array} $$ where the entry is empty if $T_i^{-1}(E_\alpha)\notin U_q(\lie n^+)$. Using Theorem~\ref{thm:T_w(B(w'))} we conclude that $\mathbf B(s_1w_\circ)$ (respectively, $\mathbf B(s_2w_\circ)$) consists of monomials of the form \begin{multline*} q^{\frac12(m_2m_{21}+(m_3-m_2)(m_{32}-m_{23})-(m_{21}-m_3-m_{23}-m_{32})m_{321} -m_3m_{213}+(m_{21}+m_{23})m_{213})}\times\\ E_2^{m_2}E_3^{m_3}E_{21}^{m_{21}} E_{23}^{m_{23}}E_{32}^{m_{32}}E_{213}^{m_{213}}E_{321}^{m_{321}}E_{2132}^{m_{2132}} \end{multline*} and, respectively \begin{multline*} q^{\frac12(m_1m_{12}+m_3m_{32}+(m_1+m_{12}-m_3-m_{32})(m_{123}-m_{321}) +(m_1+m_3-m_{12}+m_{32})m_{132})}\times\\ E_1^{m_1}E_3^{m_3}E_{12}^{m_{12}} E_{32}^{m_{32}}E_{132}^{m_{132}}E_{123}^{m_{123}}E_{321}^{m_{321}} \end{multline*} where $\min(m_\alpha,m_\beta)=0$ if $E_\alpha,E_\beta\notin\{E_{123},E_{321},E_{2132}\}$ are not connected by an edge in the following respective graphs $$ \xymatrix@!0{ &E_3\ar@{-}[ld]\ar@{-}[rd]&&E_{213}\ar@{-}[ld]\ar@{-}[ll]\\ E_{32}\ar@{-}[rd]\ar@{-}[rr]&&E_{23}\ar@{-}[ld]\\ &E_2&&E_{21}\ar@{-}[ll]\ar@{-}[lu]\ar@{-}[uu]} \qquad \xymatrix@!0@C=7ex{&E_1\ar@{-}[rd]\ar@{-}[d]\\E_{12}\ar@{-}[r]\ar@{-}[ru]\ar@{-}[rd]&E_{132}\ar@{-}[r]\ar@{-}[d]&E_3\ar@{-}[dl]\\&E_{32} } $$ The basis $\mathbf B(s_3w_\circ)$ is easy to obtain from $\mathbf B(s_1w_\circ)$ using the diagram automorphism of~$U_q(\lie n^+)$ which interchanges $E_1$ and~$E_3$, $E_{12}$ and~$E_{32}$, $E_{21}$ and $E_{23}$ and $E_{123}$, $E_{321}$ and fixes all other elements $E_\alpha$. Thus, $U_q(s_1w_\circ)$ is generated by $E_2$, $E_3$, $E_{21}$ subject to the relations \begin{gather*} [E_i,[E_i,E_j]_q]_{q^{-1}}=0,\quad [E_2,E_{21}]_{q^{-1}}=0,\quad [E_3,[E_3,E_{21}]_q]_{q^{-1}}=0=[E_{21},[E_{21},E_3]_q]_{q^{-1}}, \end{gather*} where $[x,y]_t=xy-t yx$, $x,y\in U_q(\lie n^+)$, $t\in\kk^\times$ and $\{i,j\}=\{2,3\}$, while $U_q(s_2w_\circ)$ is generated by $E_1$, $E_3$, $E_{12}$, $E_{32}$ subject to the relations $$ [E_1,E_3]=0,\, [E_i,E_{i2}]_{q^{-1}}=0,\, [E_{12},E_{32}]=0,\quad [E_i,[E_i,E_{j2}]_q]_{q^{-1}}=0,\, [E_{i2},[E_{i2},E_j]_q]_{q^{-1}}=0, $$ where $\{i,j\}=\{1,3\}$. Since all elements $w\in W$ with $\ell(w)\le 4$ are either repetition free or with a single repetition, all remaining Schubert cells have already been described in~\S\ref{subs:rep-free} and~\S\ref{subs:single rep}. For example, $$ \mathbf B(s_2s_1s_3s_2)=\{ q^{\frac12(m_2+m_{213})(m_{21}+m_{23})} E_2^{m_2}E_{21}^{m_{21}} E_{23}^{m_{23}}E_{213}^{m_{213}}E_{2132}^{m_{2132}}\,:\, \min(m_2,m_{213})=0\} $$ and $U_q(s_2s_1s_3s_2)$ is generated by $E_2$, $E_{21}$, $E_{23}$, $E_{213}$ subject to the relations $$ [E_2,E_{2i}]_{q^{-1}}=0,\quad [E_{21},E_{23}]=0, \quad [E_{2i},E_{213}]_{q^{-1}}=0,\quad [E_2,E_{213}]=(q^{-1}-q)E_{21}E_{23},\, i\in\{1,3\} $$ and coincides with the algebra of quantum $2\times 2$-matrices. \subsection{Bases for type \texorpdfstring{$C_2$}{C\_2}} We have $E_{12}=T_2^{-1}(E_1)$, $E_{1^22}=T_1(E_2)$, $E_{21}=T_2(E_1)$, $E_{21^2}=T_1^{-1}(E_2)$, $E_{121}=Y_{(1,2,1)}$ and $E_{21^2 2}=Y_{(2,1,2)}$ as defined in Theorem~\ref{thm:single rep basis}. The following is apparently well-known (and can be deduced for instance from~\cite{Qin}*{Theorems~1.4.1 and~3.1.3}). \begin{theorem} $\mathbf B^{up}$ consists of all monomials $$ q^{m_1 (m_{1^22}-m_{21^2})+m_2(m_{21}-m_{12})-m_{12}m_{1^22}+m_{21}m_{21^2}} E_1^{m_1}E_2^{m_2}E_{12}^{m_{12}}E_{21}^{m_{21}}E_{1^22}^{m_{1^22}}E_{21^2}^{m_{21^2}} E_{121}^{m_{121}}E_{21^22}^{m_{21^22}} $$ where $\min(m_\alpha,m_\beta)=0$ if $E_\alpha,E_\beta\notin\{E_{121},E_{21^22}\}$ are not connected by an edge in the following graph $$ \xymatrix{ E_{12}\ar@{-}[d]\ar@{-}[r]&E_2&E_{21}\ar@{-}[l]\ar@{-}[d]\\ E_{112}&E_1\ar@{-}[r]\ar@{-}[l]&E_{211} } $$ \end{theorem} All other Schubert cells have already been described in~\S\ref{subs:rep-free} and~\S\ref{subs:single rep}. \subsection{Bases for bi-Schubert algebras}\label{subs:bi-schub} Let $\lie g=\lie{sl}_4$. Using the computations from~\S\ref{subs:ex A_3} we obtain $$ \mathbf B(s_1w_\circ,s_1w_\circ)= \{ q^{\frac12(m_3-m_2)(m_{32}-m_{23})}E_2^{m_2}E_3^{m_3} E_{23}^{m_{23}}E_{32}^{m_{32}}E_{2132}^{m_{2132}}\,:\,\min(m_2,m_3)=0\} $$ and $U_q(s_1w_\circ,s_1w_\circ)\cong U_q(\lie{sl}_3^+)\tensor \kk[E_{2132}]$, \begin{multline*} \mathbf B(s_1w_\circ,s_2w_\circ)=\{ q^{\frac12(-m_3(m_{23}+m_{213})-(m_{21}-m_3-m_{23})m_{321}+(m_{21}+m_{23})m_{213})}E_3^{m_3}E_{21}^{m_{21}} E_{23}^{m_{23}}E_{213}^{m_{213}}E_{321}^{m_{321}}\,:\\\,\min(m_3,m_{21})=0\} \end{multline*} and $U_q(s_1w_\circ,s_2w_\circ)$ is generated by $E_3$, $E_{21}$ and $E_{23}$ subject to the relations $$ [E_3,E_{23}]_q=0,\quad [E_3,[E_3,E_{21}]_q]_{q^{-1}}=0=[E_{21},[E_{21},E_3]_q]_{q^{-1}}, $$ $$ \mathbf B(s_1w_\circ,s_3w_\circ)=\{ q^{\frac12(m_2-m_{321})(m_{21}-m_{32})}E_2^{m_2}E_{21}^{m_{21}} E_{32}^{m_{32}}E_{321}^{m_{321}}E_{2132}^{m_{2132}}\,:\,\min(m_{21},m_{32})=0\} $$ and $U_q(s_1w_\circ,s_3w_\circ)$ is generated by $E_2$, $E_{21}$, $E_{32}$, $E_{321}$ subject to the relations $$ [E_2,E_{21}]_{q^{-1}}=[E_2,E_{32}]_q=[E_{21},[E_{21},E_{32}]]_{q^2}=[E_{32},[E_{32},E_{21}]]_{q^{-2}}=0 $$ and $[E_2,E_{321}]=[E_{21},E_{321}]_q=[E_{32},E_{321}]_{q^{-1}}=0$, $$ \mathbf B(s_2w_\circ,s_2w_\circ)=\{q^{\frac12(m_1-m_3)(m_{123}-m_{321})}E_1^{m_1}E_3^{m_3}E_{123}^{m_{123}}E_{321}^{m_{321}}\} $$ and $U_q(s_2w_\circ,s_2w_\circ)$ is a quantum plane. In particular, all these algebras are PBW in the sense of Remark~\ref{rem:intersect sc}\eqref{rem:intersect sc.a}. \begin{bibdiv} \begin{biblist} \bib{BCP}{article}{ author = {Beck, J.}, author = {Chari, Vyjayanthi}, author = {Pressley, A.}, title = {An algebraic characterization of the affine canonical basis}, journal = {Duke Math. J.}, volume = {99}, date = {1999}, number = {3}, pages = {455\ndash487}, issn = {0012-7094}, } \bib{BG-dcb}{article}{ author={Berenstein, Arkady}, author={Greenstein, Jacob}, title={Double canonical bases}, eprint={1411.1391} } \bib{BR}{article}{ author={Berenstein, Arkady}, author={Rupel, Dylan}, title={Quantum cluster characters of Hall algebras}, journal={Selecta Math. (N.S.)}, volume={21}, date={2015}, number={4}, pages={1121--1176}, } \bib{BZ1}{article}{ author={Berenstein, Arkady}, author={Zelevinsky, Andrei}, title={String bases for quantum groups of type $A_r$}, conference={ title={I. M. Gel\cprime fand Seminar}, }, book={ series={Adv. Soviet Math.}, volume={16}, publisher={Amer. Math. Soc., Providence, RI}, }, date={1993}, pages={51--89}, review={\MR{1237826}}, } \bib{BZ}{article}{ author={Berenstein, Arkady}, author={Zelevinsky, Andrei}, title={Triangular bases in quantum cluster algebras}, journal={Int. Math. Res. Not.}, volume={2014}, date={2014}, number={6}, pages={1651--1688} } \bib{DKP}{article}{ author={De Concini, C.}, author={Kac, V. G.}, author={Procesi, C.}, title={Some quantum analogues of solvable Lie groups}, conference={ title={Geometry and analysis}, address={Bombay}, date={1992}, }, book={ publisher={Tata Inst. Fund. Res., Bombay}, }, date={1995}, pages={41--65}, } \bib{Kim}{article}{ author={Kimura, Yoshiyuki}, title={Quantum unipotent subgroup and dual canonical basis}, journal={Kyoto J. Math.}, volume={52}, date={2012}, number={2}, pages={277--331}, issn={2156-2261}, } \bib{Kim1}{article}{ author={Kimura, Yoshiyuki}, title={Remarks on quantum unipotent subgroup and dual canonical basis}, eprint={1506.07912}, } \bib{KimOya}{article}{ author={Kimura, Yoshiyuki}, author={Oya, Hironori}, title={Quantum twists and dual canonical bases}, eprint={1604.07748} } \bib{Kas}{article}{ author={Kashiwara, Masaki}, title={Global crystal bases of quantum groups}, journal={Duke Math. J.}, volume={69}, date={1993}, number={2}, pages={455--485}, } \bib{Kassel}{book}{ author={Kassel, Christian}, title={Quantum groups}, series={Graduate Texts in Mathematics}, volume={155}, publisher={Springer-Verlag, New York}, date={1995}, } \bib{KaKh}{article}{ author={Kaveh, Kiumars}, author={Khovanskii, A. G.}, title={Newton-Okounkov bodies, semigroups of integral points, graded algebras and intersection theory}, journal={Ann. of Math. (2)}, volume={176}, date={2012}, number={2}, pages={925--978}, } \bib{LS}{article}{ author={Levendorski{\u\i}, Serge}, author={Soibelman, Yan}, title={Algebras of functions on compact quantum groups, Schubert cells and quantum tori}, journal={Comm. Math. Phys.}, volume={139}, date={1991}, number={1}, pages={141--170}, } \bib{Lus-root1}{article}{ author={Lusztig, George}, title={Quantum groups at roots of $1$}, journal={Geom. Dedicata}, volume={35}, date={1990}, number={1-3}, pages={89--113}, issn={0046-5755}, } \bib{Lus-fdhopf}{article}{ author={Lusztig, George}, title={Finite-dimensional Hopf algebras arising from quantized universal enveloping algebra}, journal={J. Amer. Math. Soc.}, volume={3}, date={1990}, number={1}, pages={257--296}, issn={0894-0347}, } \bib{Lus-book}{book}{ author={Lusztig, George}, title={Introduction to quantum groups}, series={Progress in Mathematics}, volume={110}, publisher={Birkh\"auser Boston, Inc., Boston, MA}, date={1993}, } \bib{Lus-adv}{article}{ author={Lusztig, George}, title={Braid group action and canonical bases}, journal={Adv. Math.}, volume={122}, date={1996}, number={2}, pages={237--261}, issn={0001-8708}, } \bib{Qin}{article}{ author={Qin, Fan}, title={Compare triangular bases of acyclic quantum cluster algebras}, eprint={1606.05604} } \bib{T}{article}{ author={Tanisaki, Toshiyuki}, title={Modules over quantized coordinate algebras and PBW-bases}, eprint={1409.7973}, } \end{biblist} \end{bibdiv} \end{document}

https://arxiv.org/abs/2104.12929

Central limit theorems for high dimensional dependent data

Motivated by statistical inference problems in high-dimensional time series data analysis, we first derive non-asymptotic error bounds for Gaussian approximations of sums of high-dimensional dependent random vectors on hyper-rectangles, simple convex sets and sparsely convex sets. We investigate the quantitative effect of temporal dependence on the rates of convergence to a Gaussian random vector over three different dependency frameworks ($\alpha$-mixing, $m$-dependent, and physical dependence measure). In particular, we establish new error bounds under the $\alpha$-mixing framework and derive faster rate over existing results under the physical dependence measure. To implement the proposed results in practical statistical inference problems, we also derive a data-driven parametric bootstrap procedure based on a kernel estimator for the long-run covariance matrices. We apply the unified Gaussian and bootstrap approximation results to test mean vectors with combined $\ell^2$ and $\ell^\infty$ type statistics, change point detection, and construction of confidence regions for covariance and precision matrices, all for time series data.

\section{Introduction} High-dimensional dependent data are frequently encountered in current practical problems of finance, biomedical sciences, geological studies and many more areas. Due to the complicated dependency among different components and nonlinear dynamical behaviors in the series, there have been tremendous challenges in developing principled statistical inference procedures for such data. Most existing methods require certain parametric assumptions on the underlying data generation mechanism or structural assumptions on the dependency among different components in order to derive asymptotically pivotal distributions of the involved statistics. Assumptions of this kind are not only difficult to be verified but also often violated in real data. How to derive statistically valid inference procedures that do not rely on specific structural assumptions imposed on the dependency among different components of high-dimensional dependent data has been an urgent demand. In this paper, we focus on quantitative high-dimensional Central Limit Theorems (CLTs) and related bootstrap approximations for dependent (and possibly non-stationary) data. Let $\mathcal{X}_n = \{X_1,\ldots,X_n\}$ be a sequence of $p$-dimensional dependent random vectors with mean zero, i.e., $\mathbb{E}(X_t)=0$. Write $S_{n,x} = n^{-1/2} \sum_{t=1}^n X_t$. Denote the instantaneous covariance matrix of $X_t$ at time point $t$ by $\Sigma_t=\mathrm{Cov}(X_t)$ and the long-run covariance matrix of $\{X_t\}_{t=1}^n$ by $\Xi=\mathrm{Cov}(S_{n,x})$. Our main goal is to bound \begin{align}\label{eq:gb} \rho_n(\mathcal{A}) := \sup_{A\in\mathcal{A}}|\mathbb{P}(S_{n,x} \in A)-\mathbb{P}(G\in A)|\,, \end{align} where $G\sim N(0,\Xi)$ and $\mathcal{A}$ is a class of Borel subsets in $\mathbb{R}^p$. Gaussian and bootstrap approximation results over a rich index class $\mathcal{A}$ for large $p$ are fundamental tools in developing downstream statistical inference procedures for a wide spectrum of problems in the high-dimensional setting, including inference of mean vector, change point detection, structure checking of instantaneous covariance matrix, and testing white noise hypothesis. We refer the readers to Section \ref{sec:apps} for more details of these applications. When $X_1,\dots,X_n$ are {\it independent} random vectors in $\R^p$, the problem of bounding $\rho_n(\mathcal{A})$ for a variety of choices $\mathcal{A}}\newcommand{\cB}{\mathcal{B}}\newcommand{\cC}{\mathcal{C}$ is a classical research topic in probability theory \citep{Petrov_1995,BhattacharyaRao_2010}. Bounds on $\rho_n(\mathcal{A})$ under mild assumptions yield useful test statistics for a variety of high-dimensional inference problems. For instance, \cite{ChenQin_2010} and \cite{CaiLiuXia_2014} studied the $\ell^2$-type statistic and $\ell^\infty$-type statistic for testing high-dimensional mean vectors. Asymptotic validity of those test procedures relies on restrictive assumptions such as weak dependence in covariance matrix or sparsity in precision matrix. In contrast, Gaussian approximation results derived in this paper impose no explicit structural assumptions on the component-wise dependence structure, thus allowing to derive associated bootstrap testing procedures for arbitrary dependence among different components of high-dimensional data. Recent years has witnessed a renewed interest in the accuracy of Gaussian approximation with explicit dependence on the dimension $p$ since such results are particularly useful in modern large-scale statistical inference problems such as change point detection \citep{YuChen_2019,YuChen_2020} and multiple testing for high-dimensional data \citep{ChangZhengZhouZhou_2017, ChangZhouZhouWang_2017}. For isotropic distributions with bounded third moments, \cite{Bentkus_2003} derived a Berry-Esseen type bound $O(p^{7/4} n^{-1/2})$ and $O(p^{3/2} n^{-1/2})$ over the class of convex subsets and Euclidean balls in $\R^p$, respectively. For independent (not necessarily identically distributed) sums, \cite{CCK_2013} considered the problem of approximating maxima of $S_{n,x}$ by its Gaussian analogue and established an error bound that allows the dimension $p$ to grow sub-exponentially fast in the sample size $n$ to ensure the validity of Gaussian and bootstrap approximations. Since the seminal work \cite{CCK_2013}, there has been substantial progresses being made in several directions. For instances, generalization of the index set from the max-rectangles to hyper-rectangles with improved rates of convergence to normality can be found in \cite{CCK_2017}, \cite{LopesLinMueller_2020}, \cite{DengZhang_2020}, \cite{CCKK_2019}, \cite{CCK_2020}, \cite{DasLahiri_2020}, \cite{Deng_2020}, \cite{Koike_2019b}, \cite{FangKoike_2020}, \cite{Lopes_2020}, and \cite{KuchibhotlaRinaldo_2020}; extension from linear sums to $U$-statistics with nonlinear kernels can be found in \cite{Chen_2018}, \cite{ChenKato_2019}, \cite{SongChenKato_2019}, and \cite{Koike_2019c}; generalization to dependent random vectors over max-recetangles can be found in \cite{ZhangWu_2017}, \cite{ZhangCheng_2018}, and \cite{CCK_2014}. In the literature, some popular assumptions imposed on the temporal dependence of the sequence $\mathcal{X}_n$ include: (i) strong-mixing (or $\alpha$-mixing) \citep{Rosenblatt_1956}, (ii) $m$-dependent sequence \citep{HoeffdingRobbins_1948}, and (iii) physical (or functional) dependence measure for casual time series \citep{Wu_2005}. Various CLTs for univariate (or fixed dimensional) dependent data have been developed under these dependence frameworks, see \cite{DoukhanMasssartRio_1994}, \cite{Bradley_2007}, \cite{Wu_2007}, and \cite{BerkesLiuWu_2014}. We remark that there are many other mixing coefficients measuring the temporal dependence of the past and future, among which the $\alpha$-mixing coefficient (see Definition \ref{defn:alpha-mixing-coef} in Section \ref{sec:alpha}) is the weakest one in the literature \citep{Bradley_2005}. Note that neither a dependence framework in (i)-(iii) implies the others. Thus there is a pressing call for a unified collection for Gaussian approximation tools under these temporal dependence frameworks. Previous related works on high-dimensional CLTs for dependent data in the literature are complementary results for different dependence frameworks on max-rectangles, a subclass of hyper-rectangles. For examples, \cite{ZhangCheng_2018} studied the Gaussian approximation for $m$-dependent sequences with extension to dependent random vectors satisfying a geometric moment contraction condition \citep{WuShao_2004}; \cite{ZhangWu_2017} derived the Gaussian approximation result for causal {\it stationary} time series under a polynomial decay of the physical dependence measure; \cite{CCK_2014} studied the validity of a block multiplier bootstrap under the $\beta$-mixing condition. All the aforementioned papers are only applicable to approximating the distributions of the $\ell^\infty$-type statistics and not applicable to approximating the distributions of some more general and complicated statistics involved in high-dimensional statistical inference. See Section \ref{sec:apps} for details. We conclude the introduction by summarizing our main contributions. Specifically, we develop a comprehensive and off-the-shelf probability toolbox containing the explicit rates of convergence of the high-dimensional CLTs for a combination of different index sets (including hyper-recetangles, simple convex sets, and sparsely convex sets) and different dependence frameworks (including $\alpha$-mixing, $m$-dependent, and physical dependence measure). Our error bounds are non-asymptotic in all key parameters, including sample size $n$, data dimension $p$ and sparsity $s$. In particular, our results established under the $\alpha$-mixing framework are new in the literature, while results established under the physical dependence measure improve over existing results. In addition, we provide a parametric bootstrap procedure to implement the proposed results with a kernel-type estimator for the long-run covariance matrix. For both Gaussian and bootstrap approximations, the data dimension $p$ is allowed to grow subexponentially fast in sample size $n$. The rest of the paper is organized as follows. Section \ref{sec:main_results} presents the error bounds of $\rho_n(\mathcal{A})$ defined as \eqref{eq:gb} with selecting $\mathcal{A}$ as hyper-recetangles, simple convex sets, and sparsely convex sets, respectively. Section \ref{sec:parametric_boostrap} proposes a data-driven parametric bootstrap to approximate the probability $\mathbb{P}(S_{n,x}\in A)$ uniformly over $A\in\mathcal{A}$. Section \ref{sec:apps} discusses how to implement the proposed results in several statistical inference problems of interest. All technical proofs are given in Appendix. \section{High-dimensional central limit theorems} \label{sec:main_results} We define some notation first. For any positive integer $m$, we write $[m]:=\{1,\ldots,m\}$. Denote by $I(\cdot)$ the indicator function. For two sequences of positive numbers $\{a_n\}$ and $\{b_n\}$, we write $a_n\lesssim b_n$ or $b_n\gtrsim a_n$ if there exists a universal constant $c>0$ such that $a_n/b_n\leq c$. For any two $p$-dimensional vectors $v=(v_1,\ldots,v_p)^{ \mathrm{\scriptscriptstyle T} }$ and $u=(u_1,\ldots,u_p)^{ \mathrm{\scriptscriptstyle T} }$, $v \leq u$ means that $v_j \leq u_{j}$ for all $j\in[p]$. Given $\alpha>0$, we define the function $\psi_{\alpha}(x):=\exp(x^{\alpha})-1$ for any $x>0$. For a real-valued random variable $\xi$, we define $\|\xi\|_{\psi_{\alpha}}:=\inf[\lambda>0:\mathbb{E}\{\psi_{\alpha}(|\xi|/\lambda)\}\leq 1]$ and write $\xi\in\mathcal{L}^q$ for some $q>0$ if $\|\xi\|_q:=\{\mathbb{E}(|\xi|^q)\}^{1/q}<\infty$. For a thricely differentiable function $f:\mathbb{R}^p\rightarrow\mathbb{R}$, we write $\partial_{j}f(x)=\partial f(x)/\partial x_{j}$, $\partial_{jk}f(x)=\partial^2f(x)/\partial x_{j}\partial x_{k}$ and $\partial_{jkl}f(x)=\partial^3 f(x)/\partial x_{j}\partial x_{k}\partial x_{l}$ for any $j,k,l\in [p]$. For a $q_1\times q_2$ matrix $B=(b_{i,j})_{q_1\times q_2}$, let $|B|_\infty=\max_{i\in[q_1],j\in[q_2]}|b_{i,j}|$ be the super-norm, and $\|B\|_2=\lambda_{\max}^{1/2}(BB^{ \mathrm{\scriptscriptstyle T} })$ be the spectral norm. Specifically, if $q_2=1$, we use $|B|_0=\sum_{i=1}^{q_1}I(b_{i,1}\neq0)$, $|B|_1=\sum_{i=1}^{q_1}|b_{i,1}|$ and $|B|_2=(\sum_{i=1}^{q_1}b_{i,1}^2)^{1/2}$ to denote the $\ell^0$-norm, $\ell^1$-norm and $\ell^2$-norm of the $q_1$-dimensional vector $B$, respectively. Recall $S_{n,x}=n^{-1/2}\sum_{t=1}^nX_t$. Let $G\sim N(0,\Xi)$ with $\Xi={\rm Cov}(n^{-1/2}\sum_{t=1}^nX_t)$. Without loss of generality, we assume $G$ is independent of $\mathcal{X}_n=\{X_1,\ldots,X_n\}$. We will first consider in Section \ref{sec:hyper} the upper bounds for \begin{align}\label{eq:varrhon} \varrho_n:=\sup_{u\in\mathbb{R}^p,\nu\in[0,1]}|\mathbb{P}(\sqrt{\nu}S_{n,x}+\sqrt{1-\nu}G\leq u)-\mathbb{P}(G\leq u)| \end{align} when $\{X_t\}$ is (i) an $\alpha$-mixing sequence, (ii) an $m$-dependent sequence, and (iii) a physical dependence sequence, respectively. Based on such derived upper bounds, we can easily translate them to the upper bounds for $\rho_n(\mathcal{A})$ when $\mathcal{A}$ is selected as the class of all hyper-rectangles in $\mathbb{R}^p$. In Sections \ref{sec:simple} and \ref{sec:sparse}, we will consider the upper bounds for $\rho_n(\mathcal{A})$ when $\mathcal{A}$ is selected as the class of simple convex sets and $s$-sparsely convex sets, respectively. Write $X_t=(X_{t,1},\ldots,X_{t,p})^{ \mathrm{\scriptscriptstyle T} }$. Throughout the rest of this paper (unless otherwise explicitly stated), we shall focus on the high-dimensional scenario by assuming that $p\geq n^{\kappa}$ for some universal constant $\kappa>0$. \subsection{High-dimensional CLT for hyper-rectangles}\label{sec:hyper} Let $\mathcal{A}^{{\rm re}}$ be the class of all hyper-rectangles in $\mathbb{R}^p$; that is, $\mathcal{A}^{{\rm re}}$ consists of all sets $A$ of the form $A=\{(w_1,\ldots,w_p)^{ \mathrm{\scriptscriptstyle T} }\in\mathbb{R}^p:a_j\leq w_j\leq b_j\textrm{ for all }j\in[p]\}$ with some $-\infty\leq a_j\leq b_j\leq \infty$. Define $S_{n,\check{x}}=n^{-1/2}\sum_{t=1}^n\check{X}_t$ with $\check{X}_t=(X_t^{ \mathrm{\scriptscriptstyle T} },-X_t^{ \mathrm{\scriptscriptstyle T} })^{ \mathrm{\scriptscriptstyle T} }$ and let $\check{G}\sim N(0,\check{\Xi})$ with $\check{\Xi}={\rm Cov}(n^{-1/2}\sum_{t=1}^n\check{X}_t)$. We then have \[ \rho_n(\mathcal{A}^{\rm re})\leq \sup_{u\in\mathbb{R}^{2p},\nu\in[0,1]}|\mathbb{P}(\sqrt{\nu}S_{n,\check{x}}+\sqrt{1-\nu}\check{G}\leq u)-\mathbb{P}(\check{G}\leq u)|\,, \] where the term on the right-hand side is a $(2p)$-dimensional analogue of $\varrho_n$ defined as \eqref{eq:varrhon} over one-sided hyper-rectangles. To derive the convergence rate of $\rho_n(\mathcal{A}^{{\rm re}})$, it suffices to consider that for $\varrho_n$. \subsubsection{$\alpha$-mixing sequence}\label{sec:alpha} \begin{definition}[$\alpha$-mixing coefficient] \label{defn:alpha-mixing-coef} Let $\{X_t\}$ be a random sequence. Denote by $\mathcal{F}_{-\infty}^u$ and $\mathcal{F}_{u}^\infty$ the $\sigma$-fields generated respectively by $\{X_t\}_{t\leq u}$ and $\{X_t\}_{t\geq u}$. The {\it $\alpha$-mixing coefficient} at lag $k$ of the sequence $\{X_t\}$ is defined as \begin{align} \label{eqn:alpha_mixing_coefficient} \alpha_n(k):=\sup_t\sup_{A\in\mathcal{F}_{-\infty}^t, B\in\mathcal{F}_{t+k}^\infty}|\mathbb{P}(AB)-\mathbb{P}(A)\mathbb{P}(B)|\,. \end{align} We call the sequence $\{X_t\}$ is {\it $\alpha$-mixing} if $\alpha_n(k) \to 0$ as $k\rightarrow\infty$. \end{definition} The {\it long-run variance} of the $j$-th coordinate marginal sequence $\{X_{t,j}\}_{t=1}^n$ is defined as \begin{align}\label{eq:Vtj} V_{n,j}={\rm Var}\bigg(\frac{1}{\sqrt{n}}\sum_{t=1}^{n}X_{t,j}\bigg)\,. \end{align} To investigate the convergence rate of $\varrho_n$ defined as \eqref{eq:varrhon} for the $\alpha$-mixing sequence $\{X_t\}$, we need the following regularity conditions. \begin{ass}[Subexponential moment]\label{as:tail} There exists a sequence of constants $B_{n}\geq 1$ and a universal constant $\gamma_{1}\geq 1$ such that $\|X_{t,j}\|_{\psi_{\gamma_{1}}}\leq B_{n}$ for all $t\in [n]$ and $j\in [p]$. \end{ass} \begin{ass}[Decay of $\alpha$-mixing coefficients]\label{as:alpha-mixing} There exist some universal constants $K_1>1$, $K_2>0$ and $\gamma_2>0$ such that $ \alpha_n(k)\leq K_1\exp(-K_2k^{\gamma_2})$ for any $k\geq 1$. \end{ass} \begin{ass}[Non-degeneracy]\label{as:longrun} There exists a universal constant $K_3>0$ such that $V_{n,j}\geq K_3$ for any $j\in[p]$. \end{ass} Since Condition \ref{as:tail} implies that $\mathbb{E}\{\exp(|X_{t,j}|^{\gamma_{1}}B_{n}^{-\gamma_{1}})\}\leq 2$, it follows from Markov's inequality that $\mathbb{P}(|X_{t,j}|>u) \leq 2 \exp(-u^{\gamma_{1}}B_{n}^{-\gamma_{1}})$ for all $u > 0$. If each $X_{t,j}$ is sub-gaussian, we have $\gamma_1=2$ and $B_n=O(1)$. The $\alpha$-mixing assumption in Condition \ref{as:alpha-mixing} is mild. Causal ARMA processes with continuous innovation distributions are $\alpha$-mixing with exponentially decaying $\alpha_n(k)$. So are stationary Markov chains satisfying certain conditions. See Section 2.6.1 of \cite{FanYao_2003} and references within. In fact stationary GARCH models with finite second moments and continuous innovation distributions are also $\alpha$-mixing with exponentially decaying $\alpha_n(k)$. Condition \ref{as:longrun} assumes the partial sum $n^{-1/2}\sum_{t=1}^{n}X_{t,j}$ is non-degenerated which is required when we apply Nazarov's inequality (Lemma \ref{la:nazarov_ineq} in Appendix) to bound the probability of a Gaussian vector taking values in a small region. When $\{X_{t,j}\}_{t\geq1}$ is stationary, we know $V_{n,j}=\Gamma_j(0)+2\sum_{k=1}^{n-1}(1-kn^{-1})\Gamma_j(k)$, where $\Gamma_j(k)={\rm Cov}(X_{1,j}X_{1+k,j})$ is the autocovariance of $\{X_{t,j}\}_{t\geq1}$ at lag $k$. If each component sequence $\{X_{t,j}\}$ is stationary, Condition \ref{as:longrun} holds if $\Gamma_j(0)+2\sum_{k=1}^\infty \Gamma_j(k)\geq C$ holds for any $j\in[p]$, where $C>0$ is a universal constant. Based on Conditions \ref{as:tail}--\ref{as:longrun}, Theorem \ref{tm:1} gives an upper bound for $\varrho_n$ when the underlying sequence $\{X_t\}$ is $\alpha$-mixing. \begin{theorem}[Rate of convergence under $\alpha$-mixing] \label{tm:1} Assume $\{X_t\}$ is an $\alpha$-mixing sequence with $p\geq n^{\kappa}$ for some universal constant $\kappa>0$. Under Conditions {\rm\ref{as:tail}}--{\rm\ref{as:longrun}}, it holds that \[ \varrho_n\lesssim\frac{B_{n}^{2/3}(\log p)^{(1+2\gamma_2)/(3\gamma_2)}}{n^{1/9}}+\frac{B_{n}(\log p)^{7/6}}{n^{1/9}} \] provided that $(\log p)^{3-\gamma_2}=o(n^{\gamma_2/3})$. \end{theorem} \begin{remark}[Comparison with existing results under mixing dependence measure] Appendix B of \cite{CCK_2014} derived the validity of a block multiplier bootstrap (BMB) under the $\beta$-mixing assumption. There are several differences between our Theorem 1 and Theorem B.1 in \cite{CCK_2014}. First, since the $\beta$-mixing assumption implies the $\alpha$-mixing assumption, our Theorem 1 is applicable for wider class of dependent data. Second, Theorem B.1 in \cite{CCK_2014} is proved and stated with the ``large-and-small-blocks" argument, where conditions of their Theorem B.1 involve the ``tuning parameter" of the block sizes. It is empirically known that BMB has unstable finite sample performance. Thus, it is an undesirable feature of Theorem B.1 in \cite{CCK_2014} to rule out bootstraps without a hard truncation block size to estimate the long-run covariance matrices. In Section \ref{sec:parametric_boostrap}, we consider the kernel-type estimator of \cite{Andrews_1991} to approximate the long-run covariance matrix of $S_{n,x}$, which is more appealing from a practical standview (e.g., with the optimal quadratic spectral kernel). Third, result from \cite{CCK_2014} holds for max-norm statistics, while this paper derives the convergence rates of Gaussian and bootstrap approximations under much broader classes of index sets for high-dimensional dependent data (see Sections \ref{sec:simple} and \ref{sec:sparse} below) that can be applied to approximate the distributions of more general and complicated statistics used in high-dimensional statistical inference. \end{remark} \begin{remark} Theorem \ref{tm:1} extends the Gaussian approximation result for independent data in \cite{CCK_2017} to dependent data. When the eigenvalues of $\Xi$ is bound below from zero (i.e., strongly non-degenerate case), \cite{CCK_2020} derived a nearly optimal rate of convergence for independent data. Our analysis can be adapted with the sharper results from \cite{CCK_2020} to yield an improved error bound in Theorem \ref{tm:1} under stronger conditions. \end{remark} \subsubsection{$m$-dependent sequence} \label{subsec:m-dependent} Based on the temporal dependency among $\{X_t\}_{t=1}^n$, we can define an undirected graph $G_{n} = (V_{n}, E_{n})$, where $V_{n} = [n]$ is a set of nodes with node $t$ denoting $X_t$, and $E_{n}$ is a set of undirected edges connecting the nodes such that $X_{t}$ and $X_{s}$ are independent whenever $(t,s) \notin E_{n}$. Here we adopt the convention $(t,t)\in E_n$ for any $t\in V_n$. We call such defined $G_n$ is the {\it dependency graph} of the sequence $\{X_t\}_{t=1}^n$. The dependence graph is a flexible model to study CLTs with increasing dependence strength \citep{BaldiRinott_1989} that covers the $m$-dependent sequence as a special case. For any $t\in[n]$, let $\mathcal{N}_{t} = \{s\in V_n: (t,s) \in E_{n}\}$ be the neighbor nodes of node $t$ in $G_n$. Define $D_{n} = \max_{t \in[n]} \sum_{s=1}^{n} I\{(t,s) \in E_{n}\}$ as the maximum degree of the first-degree connections in $G_n$, and $D_{n} ^*= \max_{t \in[n]} \sum_{s=1}^{n} I\{s\in\cup_{\ell\in\mathcal{N}_t}\mathcal{N}_\ell\}$ as the maximum degree of the second-degree connections in $G_n$. Theorem \ref{lem:GA_dependency_graph} gives an upper bound for $\varrho_n$ defined as \eqref{eq:varrhon} based on the maximum degrees $D_n$ and $D^*_n$ of the dependency graph determined by the underlying sequence $\{X_t\}_{t=1}^n$. \begin{theorem}[Rate of convergence under dependence graph] \label{lem:GA_dependency_graph} Assume $p\geq n^\kappa$ for some universal constant $\kappa>0$. Under Conditions {\rm \ref{as:tail}} and {\rm\ref{as:longrun}}, it holds that \begin{align* \varrho_n \lesssim\frac{B_{n}(D_nD_n^*)^{1/3}(\log p)^{7/6}}{n^{1/6}}\,, \end{align*} where $D_n$ and $D^*_n$ are the maximum degrees of the first-degree and second-degree of connections in the dependency graph generated by the sequence $\{X_t\}_{t=1}^n$, respectively. \end{theorem} If $\{X_{t}\}_{t=1}^n$ is a centered $m$-dependent sequence, i.e., $X_{t}$ and $X_{s}$ are independent for all $|t-s|>m$, then $\{X_t\}_{t=1}^n$ has a dependency graph with $D_{n} = 2m+1$ and $D_n^*=4m+1$. The next corollary states a result for $m$-dependent sequences. \begin{cy}[Rate of convergence under $m$-dependence] \label{cor:GA_m-dependent_subexp} Assume $\{X_t\}_{t=1}^n$ is an $m$-dependent sequence with $p\geq n^\kappa$ for some universal constant $\kappa>0$. Under Conditions {\rm \ref{as:tail}} and {\rm\ref{as:longrun}}, it holds that \begin{align \label{eqn:GA_m-dependent_subexp} \varrho_n \lesssim\frac{B_{n}m^{2/3}(\log p)^{7/6}}{n^{1/6}}\,. \end{align} \end{cy} Corollary \ref{cor:GA_m-dependent_subexp} is a stepping stone to study the Gaussian approximation under the physical dependence framework with better rate of convergence than the best known results in \cite{ZhangWu_2017} based on the large-and-small-blocks technique in the weaker temporal dependence regime. See Theorem \ref{prop:GA_weakly-dependent} and the discussions in Section \ref{sec:comparison_physical_dependence} for more details. \subsubsection{Sequence with physical dependence}\label{sec:physical} Let $\{\varepsilon_{i}\}_{i \in \mathbb{Z}}$ be a sequence of independent and identically distributed random elements. Consider the (causal) time series model \begin{equation} \label{eqn:weakly_dependent_ts_fdp} X_{t} =f_{t}(\varepsilon_{t},\varepsilon_{t-1},\dots)\,, \quad t\geq1\,, \end{equation} where $f_{t}(\cdot)$ is a jointly measurable function taking values in $\R^{p}$ and $\mathbb{E}(X_{t})= 0$. Here $\{\varepsilon_{i}\}_{i \in \Z}$ are {\it innovations} that can be viewed as the input of the non-linear system (\ref{eqn:weakly_dependent_ts_fdp}). Since the data generation mechanism $f_{t}(\cdot)$ may change over time, $X_{t}$ is allowed to be {\it non-stationary}. Non-linear time series of the form (\ref{eqn:weakly_dependent_ts_fdp}) are first introduced in \cite{Wu_2005} for $p=1$ and $f_{t}(\cdot)\equiv f(\cdot)$ for some measurable function $f(\cdot)$ (i.e., stationary univariate time series), and their temporal dependence can be quantified by the {\it functional dependence measure} based on the idea of coupling. In particular, let $\varepsilon'_{i}$ be an independent copy of $\varepsilon_{i}$ and \begin{align*} X'_{t,\{m\}} =f_{t}(\varepsilon_{t},\dots,\varepsilon_{t-m+1},\varepsilon'_{t-m},\varepsilon_{t-m-1},\dots) \end{align*} be the coupled version of $X_{t}$ at the time lag $m$ with $\varepsilon_{t-m}$ replacing by $\varepsilon'_{t-m}$. By causality, $X'_{t,\{m\}} = X_{t}$ for $m < 0$. Write $X'_{t,\{m\}}=(X'_{t,1,\{m\}},\ldots,X'_{t,p,\{m\}})^{ \mathrm{\scriptscriptstyle T} }$. The (uniform) functional dependence measure for the $j$-th coordinate marginal sequence $\{X_{t,j}\}$ is defined as \begin{align*} \theta_{m,q,j} = \sup_{t \geq 1} \|X_{t,j} - X'_{t,j,\{m\}}\|_{q}\,, \quad q > 0\,. \end{align*} In essence, $\theta_{m,q,j}$ quantifies the uniform impact of coupling on the $j$-th coordinate marginal time series at lag $m$. For any $m\geq0$, write $\Theta_{m,q,j} = \sum_{i=m}^{\infty} \theta_{i,q,j}$. For $\alpha \in (0,\infty)$, define the {\it dependence adjusted norm} introduced in \cite{WuWu_2016} as \begin{align*} \|X_{.,j}\|_{q,\alpha} = \sup_{m \geq 0} (m+1)^{\alpha} \Theta_{m,q,j} \quad \text{and} \quad \|X_{.,j}\|_{\psi_{\nu},\alpha} = \sup_{q \geq 2} q^{-\nu} \|X_{.,j}\|_{q,\alpha}\,, \end{align*} whenever the supremums are finite. Define further the aggregated norms as follows: \begin{align}\label{eq:aggnorm} \Psi_{q,\alpha} = \max_{j \in[p]} \|X_{.,j}\|_{q,\alpha} \quad \text{and} \quad \Phi_{\psi_{\nu},\alpha} = \max_{j \in[p]} \|X_{.,j}\|_{\psi_{\nu},\alpha}\,. \end{align} \begin{theorem}[Rate of convergence under physical dependence] \label{prop:GA_weakly-dependent} Assume $\{X_t\}$ satisfies the model \eqref{eqn:weakly_dependent_ts_fdp} with $p\geq n^\kappa$ for some universal constant $\kappa>0$. Let $\Phi_{\psi_{\nu},\alpha} < \infty$ for some $\alpha, \nu \in (0,\infty)$. \begin{itemize} \item[{\rm(i)}] Under Condition {\rm\ref{as:longrun}}, it holds that \begin{align}\label{eqn:GA_weakly-dependent_blocking} \varrho_n \lesssim \frac{\Phi_{\psi_\nu,0}(\log p)^{7/6}}{n^{\alpha/(3+9\alpha)}}+\frac{\Psi_{2,\alpha}^{1/3} \Psi_{2,0}^{1/3}(\log p)^{2/3}}{n^{\alpha/(3+9\alpha)}}+\frac{\Phi_{\psi_\nu,\alpha}(\log p)^{1+\nu}}{n^{\alpha/(1+3\alpha)}} \end{align} provided that $(\log p)^{\max\{6\nu-1,(5+6\nu)/4\}}= o\{n^{\alpha/(1+3\alpha)}\}$. \item[{\rm(ii)}] Under Conditions {\rm \ref{as:tail}} and {\rm\ref{as:longrun}}, it holds that \begin{align} \label{eqn:GA_weakly-dependent} \varrho_n\lesssim \frac{B_{n}(\log p)^{7/6}}{n^{\alpha/(12+6\alpha)}}+\frac{\Psi_{2,\alpha}^{1/3} \Psi_{2,0}^{1/3}(\log p)^{2/3}}{n^{\alpha/(12+6\alpha)}}+\frac{\Phi_{\psi_{\nu},\alpha}(\log p)^{1+\nu}}{n^{\alpha/(4+2\alpha)}}\,. \end{align} \end{itemize} \end{theorem} We remark that the two rates of convergence given by \eqref{eqn:GA_weakly-dependent_blocking} and \eqref{eqn:GA_weakly-dependent} of Theorem \ref{prop:GA_weakly-dependent} are based on the {\it large-and-small-blocks} and $m$-dependent approximation techniques, respectively. The large-and-small-blocks technique is widely used in time series analysis to approximate the sum of a time series sequence by the sum over its large blocks. It is interesting to note that the large-and-small-blocks technique gives a faster (or slower) rate than the $m$-dependent argument when $0 < \alpha < 3$ (or $\alpha > 3$). In particular, when the temporal dependence is weak (for large values of $\alpha$), the improvement of \eqref{eqn:GA_weakly-dependent} than \eqref{eqn:GA_weakly-dependent_blocking} is more significant. The intuition is that the large-and-small-blocks technique used to establish \eqref{eqn:GA_weakly-dependent_blocking} may lose sample size efficiency when the temporal dependence is weak. In such regime, throw away the data in small blocks may reduce the {\it effective} sample size and \eqref{eqn:GA_weakly-dependent} improves over \eqref{eqn:GA_weakly-dependent_blocking}, while the $m$-dependent approximation directly approximate the sequence $\mathcal{X}_n$ without throw away data. On the other hand, when the temporal dependence is strong (for small values of $\alpha$), we need to use much larger values of $m$ for constructing an $m$-dependent sequence in Section \ref{subsec:m-dependent}, so the $m$-dependent approximation becomes less effective than throwing a reasonable amount of small blocks to reduce the dependence. Now combining the two parts of Theorem \ref{prop:GA_weakly-dependent}, we obtain the overall rate of convergence under the physical dependence measure. \begin{cy}[Overall rate of convergence under physical dependence] \label{cor:eqn:GA_weakly-dependent_combined} Assume $\{X_t\}$ satisfies the model \eqref{eqn:weakly_dependent_ts_fdp} with $p\geq n^\kappa$ for some universal constant $\kappa>0$, and $\Phi_{\psi_{\nu},\alpha} < \infty$ for some $\alpha, \nu \in (0,\infty)$. Let $\alpha' = \alpha / \min\{ 1+3\alpha, 4+2\alpha \}$. Under Conditions {\rm \ref{as:tail}} and {\rm\ref{as:longrun}}, we have \begin{align} \label{eqn:GA_weakly-dependent_combined} \varrho_n \lesssim \frac{\max\{\Phi_{\psi_\nu,0}, B_n \} (\log p)^{7/6}}{n^{\alpha'/3}}+\frac{\Psi_{2,\alpha}^{1/3} \Psi_{2,0}^{1/3}(\log p)^{2/3}}{n^{\alpha'/3}}+\frac{\Phi_{\psi_\nu,\alpha}(\log p)^{1+\nu}}{n^{\alpha'}} \end{align} provided that $(\log p)^{\max\{6\nu-1,(5+6\nu)/4\}}= o\{n^{\alpha/(1+3\alpha)}\}$. \end{cy} \subsubsection{Comparison with existing result under physical dependence measure}\label{sec:comparison_physical_dependence} Under the physical dependence and a subexponential moment condition, \cite{ZhangWu_2017} derived a Gaussian approximation result for the $\ell^\infty$-norm of normalized sums of a class of stationary time series: \begin{equation*} \omega_n := \sup_{u\geq 0}| \mathbb{P}}\newcommand{\bQ}{\mathbb{Q}}\newcommand{{\mathbf R}}{\mathbb{R}(|D^{-1} S_{n,x}|_\infty \geq u) - \mathbb{P}}\newcommand{\bQ}{\mathbb{Q}}\newcommand{{\mathbf R}}{\mathbb{R}(|D^{-1} G|_\infty \geq u) |\,, \end{equation*} where $D= \{{\rm diag}(\Xi)\}^{1/2}$ and $G \sim N(0,\Xi)$. Specifically, Theorem 7.4 in \cite{ZhangWu_2017} gives the following error bound: for any $\lambda \in (0,1)$ and $\eta > 0$, \begin{equation} \label{eqn:zhang_wu_rate_phystical_dependence} \omega_n \lesssim f^{\diamond}(\sqrt{n} \eta) + \eta (\log{p})^{1/2} + h\{\lambda, u_m^{\diamond}(\lambda)\} + \pi\{\chi(m,M)\} \end{equation} with $ f^{\diamond}(y) = p \exp(-C_\beta y^\beta m^{\alpha\beta}n^{-\beta/2}\Phi_{\psi_\nu,\alpha}^{-\beta}) + p \exp\{ -C_\beta y^\beta (mw)^{-\beta/2} \Phi_{\psi_\nu,0}^{-\beta}\}$, $h\{\lambda, u_m^{\diamond}(\lambda)\} = \lambda+w^{-1/8}\max\{\Psi_{3,0}^{3/4},\Psi_{4,0}^{1/2}\}\log^{7/8}(pw\lambda^{-1})+ w^{-1/2}\max\{ \Phi_{\psi_\nu,0}\log^{1/\beta}(pw\lambda^{-1}), \log^{1/2}(pw\lambda^{-1})\}\\ \cdot\log^{3/2}(pw\lambda^{-1})$, $\pi(x) = x^{1/3} \max\{1, \, \log^{2/3}(px^{-1})\}$, $\chi(m,M) = \Psi_{2,\alpha} \Psi_{2,0}\{m^{-\alpha}+v(M)\} + wmn^{-1}$, where $\beta = 2/(1+2\nu)$, $v(M) = M^{-1}I(\alpha > 1)+(M^{-1} \log{M})I(\alpha = 1)+M^{-\alpha}I(0 < \alpha < 1)$, and $(m, M, w)$ are tuning parameters involved in the ``large and small blocks" technique for deriving \eqref{eqn:zhang_wu_rate_phystical_dependence} with $m$ and $M$ being, respectively, the size of small blocks and large blocks satisfying $m=o(M)$, and $w=\lfloor n/(M+m)\rfloor$. They first approximated $S_{n,x}$ by the sum of an $m$-dependent sequence, and then applied the large-and-small-blocks technique to approximate the sum of the $m$-dependent sequence by the sum over large blocks. To simplify the convergence rate of $\omega_n$ specified in \eqref{eqn:zhang_wu_rate_phystical_dependence}, we assume $\Phi_{\psi_\nu,\alpha} = O(1)$ for some $\alpha,\nu \in(0,\infty)$ and $p\geq n^\kappa$ for some $\kappa>0$. Choose $\lambda = n^{-c_1}, \eta = n^{-c_2}$, $w \asymp n^{c_3}$ and $m\asymp n^{c_4}$ for some constants $c_1,c_2,c_3, c_4>0$. By optimizing $(c_1,c_2,c_3,c_4)$ according to the right-hand side of \eqref{eqn:zhang_wu_rate_phystical_dependence}, we have the following proposition whose proof is given in Appendix \ref{sec:xic1c4}. \begin{proposition}[Rate of convergence under physical dependence in \cite{ZhangWu_2017}] \label{pn:zhangwu2017} Assume $\Phi_{\psi_\nu,\alpha} = O(1)$ for some $\alpha,\nu \in(0,\infty)$ and $p\geq n^\kappa$ for some $\kappa>0$. Then the upper bound of $\omega_n$ given in \eqref{eqn:zhang_wu_rate_phystical_dependence} can be simplified as \begin{equation} \label{eqn:zhang_wu_rate_phystical_dependence_simplified} \omega_n \lesssim {\rm polylog}(p) \cdot n^{-\alpha/(3+11\alpha)}\,, \end{equation} where ${\rm polylog}(p)$ is a polynomial factor of $\log p$. \end{proposition} Contrasting Proposition \ref{pn:zhangwu2017} with Corollary \ref{cor:eqn:GA_weakly-dependent_combined} specialized to max-rectangles, we see that, up to a $\mbox{polylog}(p)$ factor, our rate of convergence in \eqref{eqn:GA_weakly-dependent_combined} reads ${\rm polylog}(p) \cdot n^{-\alpha/[3 \min\{ 1+3\alpha, 4+2\alpha \}]}$, which is uniformly faster than \eqref{eqn:zhang_wu_rate_phystical_dependence_simplified} for all $\alpha > 0$. In other words, our rate has a better sample size dependence than \eqref{eqn:zhang_wu_rate_phystical_dependence_simplified}. The reason can be seen that the optimal choice of \cite{ZhangWu_2017} throws away $w \asymp n^{8\alpha/(3+11\alpha)}$ small blocks of size $m\asymp n^{3/(3+11\alpha)}$, which leads a total reduction of $O\{n^{(3+8\alpha)/(3+11\alpha)}\}$ data points in the sample size. In our result, we only throws away $w \asymp n^{2\alpha/(1+3\alpha)}$ small blocks of size $m \asymp n^{1/(1+3\alpha)}$, leading to a total reduction of $O\{n^{(1+2\alpha)/(1+3\alpha)}\}$ data points in the sample size. Moreover, the improvement of \eqref{eqn:GA_weakly-dependent_combined} over \eqref{eqn:zhang_wu_rate_phystical_dependence_simplified} is more significant for larger values of $\alpha > 3$. \subsection{High-dimensional CLT for simple convex sets}\label{sec:simple} In this section, we consider the class of {\it simple convex sets} introduced by \cite{CCK_2017}. Formally, a simple convex set can be well approximated by a convex polytope with a controlled number of facets. Simple convex sets serve an important intermediate step to derive similar error bounds in Gaussian approximation on the class of $s$-sparsely convex sets considered in Section \ref{sec:sparse}. Geometrically, $s$-sparsely convex sets can be represented as an interaction of possibly many convex sets whose indicator functions depend at most on $s$ elements of their coordinates. For a closed convex set $A\subset\mathbb{R}^p$, we define its support function: \begin{align*} \mathcal{S}_{A}:\mathbb{S}^{p-1}\mapsto\mathbb{R}\cup\{\infty\}\,,~~v\mapsto\mathcal{S}_{A}(v):=\sup\{w^{{ \mathrm{\scriptscriptstyle T} }}v:w\in A\}\,, \end{align*} where $\mathbb{S}^{p-1}$ is the unit sphere in $\R^p$. Specially, if $A^{K}$ is \emph{$K$-generated} (that is, $A^{K}$ is generated by the intersection of $K$ half-spaces), we could characterize $A^{K}$ by its support function for the set $\mathcal{V}(A^K)$ consisting $K$ unit normal vectors outward to the facets of $A^K$: \begin{align*} A^{K}=\bigcap_{v\in\mathcal{V}(A^{K})}\{w\in\mathbb{R}^{p}:w^{{ \mathrm{\scriptscriptstyle T} }}v\leq\mathcal{S}_{A^{K}}(v)\}\,. \end{align*} Moreover, for $\epsilon>0$ and a $K$-generated convex set $A^K$, we also define \begin{align*} A^{K,\epsilon}=\bigcap_{v\in\mathcal{V}(A^{K})}\{w\in\mathbb{R}^{p}:w^{{ \mathrm{\scriptscriptstyle T} }}v\leq\mathcal{S}_{A^{K}}(v)+\epsilon\}\,. \end{align*} \begin{definition}[Simple convex set] \label{defn:simple_convex_set} We say $A$ is a \textit{simple convex set}, if there exist two constants $a\geq0$, $d>0$ and an $K$-$generated$ $A^{K}$ satisfying $K\leq (pn)^d$ such that \begin{align}\label{eq:inclusion} A^{K}\subset A\subset A^{K,\epsilon} \end{align} with $\epsilon=a/n$. In this case, $A^K$ provides an approximation to $A$ with precision $\epsilon$. \end{definition} Let $\mathcal{A}^{\rm si}(a,d)$ be the class of all sets $A$ satisfying \eqref{eq:inclusion} with $K\leq(pn)^d$ and $\epsilon=a/n$. For any $v\in \mathbb{R}^{p}$, define \begin{align}\label{eq:Vtv} V_{n}(v)={\rm Var}\bigg(\frac{1}{\sqrt{n}}\sum_{t=1}^{n}v^{{ \mathrm{\scriptscriptstyle T} }}X_{t}\bigg)\,. \end{align} In the sequel, we shall slightly abuse the notation by using $A^K(\cdot)$ to also denote the operator defined on $\mathcal{A}^{\rm si}(a,d)$ such that $A^K(A)=A^K$ with $A^K$ specified in \eqref{eq:inclusion} for any $A\in\mathcal{A}^{\rm si}(a,d)$. To construct the upper bounds for $\rho_n(\mathcal{A})$ for some $\mathcal{A}\subset\mathcal{A}^{\rm si}(a,d)$, we need the following condition that imposes the moment assumption on $v^{ \mathrm{\scriptscriptstyle T} } X_t$ for $v\in\mathcal{V}\{A^K(A)\}$. \begin{ass}\label{as:ExtendedTail} There exist a sequence of constants $B_{n}\geq 1$ and a universal constant $\gamma_{1}\geq 1$ such that $\|v^{{ \mathrm{\scriptscriptstyle T} }}X_{t}\|_{\psi_{\gamma_{1}}}\leq B_{n}$ for all $t\in [n]$ and $v\in\mathcal{V}\{A^{K}(A)\}$. \end{ass} \subsubsection{$\alpha$-mixing sequence} To obtain an upper bound for $\rho_n(\mathcal{A})$ for some $\mathcal{A}\subset\mathcal{A}^{\rm si}(a,d)$, we need the next condition that requires the long-run variance of the sequence $\{v^{ \mathrm{\scriptscriptstyle T} } X_{t}\}_{t=1}^n$ is not degenerated for any $v\in\mathcal{V}\{A^K(A)\}$. \begin{ass}\label{as:longrun_simple_convex} There exists a universal constant $K_{4}>0$ such that $V_{n}(v)\geq K_{4}$ for any $v\in\mathcal{V}\{A^{K}(A)\}$. \end{ass} Condition \ref{as:longrun_simple_convex} holds automatically if the smallest eigenvalue of the long-run covariance $\Xi={\rm Cov}(n^{-1/2}\sum_{t=1}^nX_t)$ is uniformly bounded away from zero. \begin{theorem}\label{pn:GA_simple_convex} Assume $\{X_t\}$ is an $\alpha$-mixing sequence with $p\geq n^{\kappa}$ for some universal constant $\kappa>0$ and Condition {\rm\ref{as:alpha-mixing}} being satisfied. Let $\mathcal{A}$ be a subclass of $\mathcal{A}^{\rm si}(a,d)$ such that Conditions {\rm\ref{as:ExtendedTail}} and {\rm\ref{as:longrun_simple_convex}} are satisfied for any $A\in\mathcal{A}$. Then \begin{align*} \rho_n(\mathcal{A})\lesssim \frac{a(d\log p)^{1/2}}{n} +\frac{B_{n}^{2/3}(d\log p)^{(1+2\gamma_2)/(3\gamma_2)}}{n^{1/9}}+\frac{B_{n}(d\log p)^{7/6}}{n^{1/9}} \end{align*} provided that $(d\log p)^{3-\gamma_2}=o(n^{\gamma_2/3})$. \end{theorem} \subsubsection{$m$-dependent sequence} Theorem \ref{pn:GA_dependency_graph_simple_convex} gives the high-dimensional CLT for simple convex sets based on the maximum degrees $D_n$ and $D_n^*$ of the dependency graph determined by the underlying sequence $\{X_t\}_{t=1}^n$. See the definition of dependency graph and its associated maximum degrees in Section \ref{subsec:m-dependent}. \begin{theorem} \label{pn:GA_dependency_graph_simple_convex} Assume $p\geq n^\kappa$ for some universal constant $\kappa>0$. Let $\mathcal{A}$ be a subclass of $\mathcal{A}^{\rm si}(a,d)$ such that Conditions {\rm\ref{as:ExtendedTail}} and {\rm\ref{as:longrun_simple_convex}} are satisfied for any $A\in\mathcal{A}$. Then \begin{align*} \rho_n(\mathcal{A})\lesssim \frac{a(d\log p)^{1/2}}{n}+\frac{B_{n}(D_{n}D_n^*)^{1/3}(d\log p)^{7/6}}{n^{1/6}}\,, \end{align*} where $D_n$ and $D^*_n$ are the maximum degrees of the first-degree and second-degree of connections in the dependency graph generated by the sequence $\{X_t\}_{t=1}^n$, respectively. \end{theorem} The next corollary states a result for $m$-dependent sequences. \begin{cy} \label{cor:GA_m-dependent_subexp_simple_convex} Assume $\{X_t\}_{t=1}^n$ is an $m$-dependent sequence with $p\geq n^\kappa$ for some universal constant $\kappa>0$. Let $\mathcal{A}$ be a subclass of $\mathcal{A}^{\rm si}(a,d)$ such that Conditions {\rm\ref{as:ExtendedTail}} and {\rm\ref{as:longrun_simple_convex}} are satisfied for any $A\in\mathcal{A}$. Then \begin{align*} \label{eqn:GA_rate_dependency_graph_simple_convex} \rho_n(\mathcal{A})\lesssim \frac{a(d\log p)^{1/2}}{n}+\frac{B_{n}m^{2/3}(d\log p)^{7/6}}{n^{1/6}}\,. \end{align*} \end{cy} \subsubsection{Sequence with physical dependence}\label{sec:phydepsim} For a $K$-generated convex set $A^K$, write $\mathcal{V}(A^K)=\{v_1,\ldots,v_K\}$. Given $\mathcal{V}(A^K)$ and $\{X_t\}$, we can define a new $K$-dimensional sequence $X_t(A^K)=(v_1^{ \mathrm{\scriptscriptstyle T} } X_t,\ldots,v_K^{ \mathrm{\scriptscriptstyle T} } X_t)^{ \mathrm{\scriptscriptstyle T} }$. If $\{X_{t}\}_{t=1}^n$ satisfies model \eqref{eqn:weakly_dependent_ts_fdp} with the jointly measurable function $f_t(\cdot)$, $\{X_t(A^K)\}$ also satisfies \eqref{eqn:weakly_dependent_ts_fdp} with a jointly measurable function $\tilde{f}_t(\cdot)=\{v_1^{ \mathrm{\scriptscriptstyle T} } f_t(\cdot),\ldots,v_K^{ \mathrm{\scriptscriptstyle T} } f_t(\cdot)\}^{ \mathrm{\scriptscriptstyle T} }$. We further define $\Psi_{q,\alpha}(A^K)$ and $\Phi_{\psi_{\nu},\alpha}(A^K)$ for any $K$-generated convex set $A^K$ in the same manner as $\Psi_{q,\alpha}$ and $\Phi_{\psi_{\nu},\alpha}$ in \eqref{eq:aggnorm} by replacing $\{X_{t}\}$ with $\{X_t(A^K)\}$. Given $\mathcal{A}\subset\mathcal{A}^{{\rm si}}(a,d)$, let \begin{equation}\label{eq:cobphy} \Psi_{q,\alpha,\mathcal{A}}=\sup_{A\in\mathcal{A}}\Psi_{q,\alpha}\{A^K(A)\}\quad\text{and}\quad\Phi_{\psi_{\nu},\alpha,\mathcal{A}}=\sup_{A\in\mathcal{A}}\Phi_{\psi_{\nu},\alpha}\{A^K(A)\}\,. \end{equation} \begin{theorem} \label{pn:GA_weakly-dependent_simple_convex} Assume the sequence $\{X_t\}$ satisfies the model \eqref{eqn:weakly_dependent_ts_fdp} with $p\geq n^\kappa$ for some universal constant $\kappa>0$. Let $\mathcal{A}$ be a subclass of $\mathcal{A}^{\rm si}(a,d)$ such that Condition {\rm\ref{as:longrun_simple_convex}} is satisfied for any $A\in\mathcal{A}$, and $\Phi_{\psi_{\nu},\alpha,\mathcal{A}}< \infty$ for some $\alpha, \nu \in (0,\infty)$. \begin{itemize} \item[{\rm (i)}] It holds that \begin{align*} \rho_n(\mathcal{A})\lesssim&\, \frac{a(d\log p)^{1/2}}{n} +\frac{\Phi_{\psi_{\nu},0,\mathcal{A}}(d\log p)^{7/6}}{n^{\alpha/(3+9\alpha)}}\\ &\,~~~+\frac{\Psi_{2,\alpha,\mathcal{A}}^{1/3}\Psi_{2,0,\mathcal{A}}^{1/3}(d\log p)^{2/3}}{n^{\alpha/(3+9\alpha)}}+\frac{\Phi_{\psi_\nu,\alpha,\mathcal{A}}(d\log p)^{1+\nu}}{n^{\alpha/(1+3\alpha)}} \end{align*} provided that $(d\log p)^{\max\{6\nu-1,(5+6\nu)/4\}}=o\{n^{\alpha/(1+3\alpha)}\}$. \item[{\rm (ii)}] If Condition {\rm\ref{as:ExtendedTail}} is satisfied for any $A\in\mathcal{A}$, we have \begin{align*} \rho_n(\mathcal{A})\lesssim&\, \frac{a(d\log p)^{1/2}}{n} +\frac{B_{n}(d\log p)^{7/6}}{n^{\alpha/(12+6\alpha)}}\\ &~~~~+\frac{\Psi_{2,\alpha,\mathcal{A}}^{1/3}\Psi_{2,0,\mathcal{A}}^{1/3}(d\log p)^{2/3}}{n^{\alpha/(12+6\alpha)}}+\frac{\Phi_{\psi_{\nu},\alpha,\mathcal{A}}(d\log p)^{1+\nu}}{n^{\alpha/(4+2\alpha)}}\,. \end{align*} \end{itemize} \end{theorem} \subsection{High-dimensional CLT for sparsely convex sets}\label{sec:sparse} We consider sparsely convex sets here, as a generalization of hyper-rectangles, that can be represented as intersections of convex sets whose indicator functions depend only on a small subset of their coordinates. \begin{definition}[$s$-sparsely convex set] \label{defn:s-sparsely_convex_set} For an integer $s>0$, we say $A\subset\mathbb{R}^{p}$ is an \textit{$s$-sparsely convex set} if (i) $A$ admits a sparse representation $ A=\cap_{q=1}^{K_*}A_{q}$ for some positive integer $K_*$ and convex sets $A_{1},\ldots,A_{K_*}\subset\mathbb{R}^{p}$, and (ii) the indicator function $I(w\in A_{q})$ depends on at most $s$ components of the vector $w\in\mathbb{R}^{p}$ (which we call the main components of $A_{q}$). \end{definition} Denote by $\mathcal{A}^{\rm sp}(s)$ the class of all $s$-sparsely convex sets in $\mathbb{R}^{p}$. In this section, we target on deriving the upper bounds for $\rho_n\{\mathcal{A}^{\rm sp}(s)\}$ when the observed data $\{X_t\}$ are (i) an $\alpha$-mixing sequence, (ii) an $m$-dependent sequence, and (iii) a physical dependence sequence. \subsubsection{$\alpha$-mixing sequence} \begin{ass}\label{as:longrun_s-sparsely} For $V_{n}(v)$ defined in \eqref{eq:Vtv}, there exists a universal constant $K_{5}>0$ such that $V_{n}(v)\geq K_{5}$ for any $v\in\mathbb{S}^{p-1}$ with $|v|_{0}\leq s$. \end{ass} Condition \ref{as:longrun_s-sparsely} holds automatically if the smallest eigenvalue of the long-run covariance $\Xi={\rm Cov}(n^{-1/2}\sum_{t=1}^nX_t)$ is uniformly bounded away from zero. \begin{theorem}\label{pn:GA_s-sparsely} Assume $\{X_t\}$ is an $\alpha$-mixing sequence with $p\geq n^\kappa$ for some universal constant $\kappa>0$. Under Conditions {\rm\ref{as:tail}}, {\rm\ref{as:alpha-mixing}} and {\rm\ref{as:longrun_s-sparsely}}, it holds that \begin{align*} &\rho_n\{\mathcal{A}^{\rm sp}(s)\}\lesssim \frac{B_{n}^{2/3}s^{(2+6\gamma_{2})/(3\gamma_{2})}(\log p)^{(1+2\gamma_2)/(3\gamma_2)}}{n^{1/9}}+\frac{B_{n}s^{10/3}(\log p)^{7/6}}{n^{1/9}} \end{align*} provided that $(s^2\log p)^{3-\gamma_2}=o(n^{\gamma_2/3})$. \end{theorem} \subsubsection{$m$-dependent sequence} \begin{theorem} \label{pn:GA_dependency_graph_s_sparsely} Assume $p\geq n^\kappa$ for some universal constant $\kappa>0$. Under Conditions {\rm\ref{as:tail}} and {\rm\ref{as:longrun_s-sparsely}}, it holds that \begin{align*} \rho_n\{\mathcal{A}^{\rm sp}(s)\} \lesssim\frac{s^{10/3}B_{n}(D_nD_n^*)^{1/3}(\log p)^{7/6}}{n^{1/6}}\,, \end{align*} where $D_n$ and $D^*_n$ are the maximum degrees of the first-degree and second-degree of connections in the dependency graph generated by the sequence $\{X_t\}_{t=1}^n$, respectively. \end{theorem} The next corollary states a result for $m$-dependent sequences. \begin{cy} \label{cor:GA_m-dependent_subexp_s-sparsely} Assume $\{X_t\}_{t=1}^n$ is an $m$-dependent sequence with $p\geq n^\kappa$ for some universal constant $\kappa>0$. Under Conditions {\rm\ref{as:tail}} and {\rm\ref{as:longrun_s-sparsely}}, it holds that \begin{align*} \rho_n\{\mathcal{A}^{\rm sp}(s)\}\lesssim\frac{s^{10/3}m^{2/3}B_{n}(\log p)^{7/6}}{n^{1/6}}\,. \end{align*} \end{cy} \subsubsection{Sequence with physical dependence} For any $c>0$, define $\Omega_{s,c}=\{A\in\mathcal{A}^{\rm si}(1,cs^2): \max_{v\in\mathcal{V}\{A^{K}(A)\}}|v|_{0}\leq s\}$. Analogous to $\Psi_{q,\alpha,\mathcal{A}}$ and $\Phi_{\psi_{\nu},\alpha,\mathcal{A}}$ defined in \eqref{eq:cobphy}, we define \begin{align*} \Psi_{q,\alpha,\Omega_{s,c}}=\sup_{A\in\Omega_{s,c}}\Psi_{q,\alpha}\{A^K(A)\}\quad\text{and}\quad\Phi_{\psi_{\nu},\alpha,\Omega_{s,c}}=\sup_{A\in\Omega_{s,c}}\Phi_{\psi_{\nu},\alpha}\{A^K(A)\} \end{align*} with $\Psi_{q,\alpha}(A^K)$ and $\Phi_{\psi_{\nu},\alpha}(A^K)$ defined in Section \ref{sec:phydepsim}. It then holds that $\Psi_{q,\alpha,\Omega_{s,c}}\geq \Psi_{q,\alpha}$ and $\Phi_{\psi_{\nu},\alpha,\Omega_{s,c}}\geq \Phi_{\psi_{\nu},\alpha}$. \begin{theorem} \label{pn:GA_weakly-dependent_s_sparsely} Assume the sequence $\{X_t\}$ satisfies the model \eqref{eqn:weakly_dependent_ts_fdp} with $p\geq n^\kappa$ for some universal constant $\kappa>0$. For some sufficiently large constant $c>0$, let $\Phi_{\psi_{\nu},\alpha,\Omega_{s,c}}< \infty$ with some $\alpha, \nu \in (0,\infty)$. \begin{itemize} \item[{\rm (i)}] Under Condition {\rm\ref{as:longrun_s-sparsely}}, it holds that \begin{align*} \rho_n\{\mathcal{A}^{\rm sp}(s)\}\lesssim&\, \frac{\Phi_{\psi_{\nu},0,\Omega_{s,c}}(s^2\log p)^{7/6}}{n^{\alpha/(3+9\alpha)}}+\frac{\Psi_{2,\alpha,\Omega_{s,c}}^{1/3}\Psi_{2,0,\Omega_{s,c}}^{1/3}(s^2\log p)^{2/3}}{n^{\alpha/(3+9\alpha)}}\\ &~~~+\frac{\Phi_{\psi_{\nu},\alpha,\Omega_{s,c}}(s^2\log p)^{1+\nu}}{n^{\alpha/(1+3\alpha)}}+\frac{\Phi_{\psi_\nu,0}^3s^{17/4}}{n^{\alpha/(2+4\alpha)}} \end{align*} provided that $(s^2\log p)^{\max\{6\nu-1,(5+6\nu)/4\}}=o\{n^{\alpha/(1+3\alpha)}\}$. \item[{\rm (ii)}] Under Conditions {\rm \ref{as:tail}} and {\rm\ref{as:longrun_s-sparsely}}, it holds that \begin{align*} \rho_n\{\mathcal{A}^{\rm sp}(s)\}\lesssim&\, \frac{s^{10/3}B_{n}(\log p)^{7/6}}{n^{\alpha/(12+6\alpha)}}+\frac{\Psi_{2,\alpha,\Omega_{s,c}}^{1/3}\Psi_{2,0,\Omega_{s,c}}^{1/3}(s^2\log p)^{2/3}}{n^{\alpha/(12+6\alpha)}}\\ &~~~+\frac{\Phi_{\psi_{\nu},\alpha,\Omega_{s,c}}(s^2\log p)^{1+\nu}}{n^{\alpha/(4+2\alpha)}}\,. \end{align*} \end{itemize} \end{theorem} \section{Parametric bootstrap}\label{sec:parametric_boostrap} In Section \ref{sec:main_results}, we have established the error bounds for $\rho_n(\mathcal{A})$ defined as \eqref{eq:gb} when $\{X_t\}$ is (i) an $\alpha$-mixing sequence, (ii) an $m$-dependent sequence, and (iii) a physical dependence sequence. Since $\mathbb{P}(G\in A)$ with $G\sim N(0,\Xi)$ depends on the unknown long-run covariance matrix $\Xi$, to approximate $\mathbb{P}(S_{n,x}\in A)$ in practice, we need to construct a data-dependent Gaussian analogue $\hat{G}$ of $G$. In this section, we propose a parametric bootstrap procedure to construct $\hat{G}\sim N(0,\hat{\Xi}_{n})$ for some covariance matrix $\hat{\Xi}_n$ that is close to $\Xi$ and establish its theoretical validity. Define \begin{align*} \hat{\rho}_n(\mathcal{A}) := \sup_{A\in\mathcal{A}}|\mathbb{P}(S_{n,x} \in A)-\mathbb{P}(\hat{G}\in A\,|\,\mathcal{X}_n)| \end{align*} with $\mathcal{X}_n=\{X_{1},\ldots,X_{n}\}$. Theorem \ref{tm:pb_rectangle} establishes primitive error bounds for $\hat{\rho}_n(\mathcal{A})$ when $\mathcal{A}$ is selected as the class of all hyper-rectangles in $\mathbb{R}^p$, the class of simple convex sets and the class of $s$-sparsely convex sets, respectively. Write $\Delta_{n,r}=|\hat{\Xi}_{n}-\Xi|_{\infty}$ and define \begin{align*} \Delta_{n}(\mathcal{A}):=\sup_{A\in \mathcal{A}}\sup_{v_{1},v_{2}\in \mathcal{V}\{A^{K}(A)\}}|v_{1}^{{ \mathrm{\scriptscriptstyle T} }}(\hat{\Xi}_{n}-\Xi)v_{2}| \end{align*} for any $\mathcal{A}\subset\mathcal{A}^{\rm si}(a,d)$. \begin{theorem}[Rates of convergence for parametric bootstrap] \label{tm:pb_rectangle} Assume $p\geq n^\kappa$ for some universal constant $\kappa>0$. The following assertions are satisfied: \begin{itemize} \item[{\rm (i)}] Under Condition {\rm\ref{as:longrun}}, it holds that $ \hat{\rho}_n(\mathcal{A}^{\rm re})\lesssim \rho_n(\mathcal{A}^{\rm re})+\Delta_{n,r}^{1/3}(\log p)^{2/3}$. \item[{\rm (ii)}] Let $\mathcal{A}$ be a subclass of $\mathcal{A}^{\rm si}(a,d)$ such that Condition {\rm\ref{as:longrun_simple_convex}} is satisfied for every $A\in\mathcal{A}$. Then $ \hat{\rho}_n(\mathcal{A})\lesssim \rho_n(\mathcal{A})+ an^{-1}(d\log p)^{1/2}+\Delta_{n}^{1/3}(\mathcal{A})(d\log p)^{2/3}$. \item[{\rm (iii)}] Under Condition {\rm\ref{as:tail}} and {\rm\ref{as:longrun_s-sparsely}}, it holds that $\hat{\rho}_n\{\mathcal{A}^{\rm sp}(s)\}\lesssim \rho_n\{\mathcal{A}^{\rm sp}(s)\}+s^{2}\Delta_{n,r}^{1/3}(\log p)^{2/3}+\{B_{n}+s(\log p)^{1/2}\}n^{-1}$. \end{itemize} \end{theorem} \begin{remark}\label{remark:paraboot} (i) For the case of hyper-rectangles, $\Delta_{n,r}=o_{\rm p}\{(\log p)^{-2}\}$ can be used to guarantee $\hat{\rho}_n(\mathcal{A}^{\rm re})=o_{\rm p}(1)$. (ii) For $\mathcal{A}\subset \mathcal{A}^{\rm si}(a,d)$, to make $\hat{\rho}_n(\mathcal{A})=o_{\rm p}(1)$, $\hat{\Xi}_n$ should satisfy $\Delta_n(\mathcal{A})=o_{\rm p}\{(d\log p)^{-2}\}$. Notice that $\Delta_n(\mathcal{A})\leq \Delta_{n,r}\sup_{A\in\mathcal{A}}\sup_{v\in\mathcal{V}\{A^K(A)\}}|v|_1^2$. If the $\ell^1$-norm of the unit normal vectors outwards to the facets of $A^K(A)$ is uniformly bounded away from infinity over $A\in\mathcal{A}$, it suffices to require $\Delta_{n,r}=o_{\rm p}\{(d\log p)^{-2}\}$. (iii) For the case of $s$-sparsely convex sets, to make $\hat{\rho}_n\{\mathcal{A}^{\rm sp}(s)\}=o_{\rm p}(1)$, we need to require $\Delta_{n,r}=o_{\rm p}\{(s^3\log p)^{-2}\}$. \end{remark} As we have discussed in Remark \ref{remark:paraboot}, the validity of our proposed parametric bootstrap only requires the estimated long-run covariance $\hat{\Xi}_n$ satisfies $|\hat{\Xi}_n-\Xi|_\infty=o_{\rm p}(\delta_n)$ for some $\delta_n\rightarrow0$ as $n\rightarrow\infty$, where $\delta_n$ will be different for different selections of $\mathcal{A}$. There are various estimation methods for long-run covariance matrices, including the kernel-type estimators \citep{Andrews_1991} and utilizing moving block bootstraps \citep{Lahiri_2003}. See also \cite{DenHanLevin_1997} and \cite{Kieferetaj_2000}. Since the data sequence $\{X_t\}_{t=1}^n$ may be non-stationary, we suggest to adopt the kernel-type estimator for its long-run covariance matrix, that is \begin{equation} \hat{\Xi}_{n}=\sum_{j=-n+1}^{n-1}\mathcal{K}\bigg(\frac{j}{b_{n}}\bigg)\hat{H}_{j}\,,\label{eq:kernelest} \end{equation} where $\hat{H}_{j}=n^{-1}\sum_{t=j+1}^{n}(X_{t}-\bar{X})(X_{t-j}-\bar{X})^{{ \mathrm{\scriptscriptstyle T} }}$ if $j\geq 0$ and $\hat{H}_{j}=n^{-1}\sum_{t=-j+1}^{n}(X_{t+j}-\bar{X})(X_{t}-\bar{X})^{{ \mathrm{\scriptscriptstyle T} }}$ otherwise, with $\bar{X}=n^{-1}\sum_{t=1}^{n}X_{t}$. Here $\mathcal{K}(\cdot)$ is a symmetric kernel function that is continuous at 0 with $\mathcal{K}(0)=1$, and $b_{n}$ is the bandwidth diverging with $n$. Among a variety of kernel functions that guarantee the positive definiteness of the long-run covariance matrix estimators, \cite{Andrews_1991} derived an optimal kernel, i.e., the quadratic spectral kernel \begin{equation}\label{eq:QSkernel} \mathcal{K}_{\rm QS}(x)=\frac{25}{12\pi^{2}x^{2}}\bigg\{\frac{\sin(6\pi x/5)}{6\pi x/5}-\cos(6\pi x/5) \bigg\}\,, \end{equation} by minimizing the asymptotic truncated mean square error of the estimator. To investigate the property of $\hat{\Xi}_n$ given in \eqref{eq:kernelest}, we need the following condition imposed on the kernel function $\mathcal{K}(\cdot)$. \begin{ass}[Kernel regularity]\label{as:kernel} The kernel function $\mathcal{K}(\cdot):\mathbb{R}\rightarrow [-1,1]$ is continuously differentiable with bounded derivatives on $\mathbb{R}$ and satisfies ${\rm(i)}\,\mathcal{K}(0)=1$, ${\rm(ii)}\,\mathcal{K}(x)=\mathcal{K}(-x)$ for any $x\in\mathbb{R}$, ${\rm(iii)}\,\int_{-\infty}^{\infty}|\mathcal{K}(x)|\,{\rm d}x<\infty$, and ${\rm (iv)}\, |\mathcal{K}(x)|\lesssim |x|^{-\vartheta}$ as $|x|\to\infty$ for some constant $\vartheta>1$. \end{ass} \begin{theorem}[Bounds on $\Delta_{n,r}$] \label{tm:cov_compare} Assume $p\geq n^{\kappa}$ for some universal constant $\kappa>0$. Let Condition {\rm\ref{as:kernel}} hold and $b_n\asymp n^{\rho}$. The following assertions are satisfied: \begin{itemize} \item[{\rm(i)}] For $\alpha$-mixing sequence $\{X_t\}$ with Conditions {\rm\ref{as:tail}} and {\rm\ref{as:alpha-mixing}} being satisfied, if $0<\rho<(\vartheta-1)/(3\vartheta-2)$, there exist two constants $c_1>0$ depending only on $(\rho,\vartheta)$ and $c_2>0$ depending only on $(\gamma_1,\gamma_2,\vartheta)$ such that $\Delta_{n,r}=O_{\rm p}\{B_n^2n^{-c_1}(\log p)^{c_2}\}+O(B_n^2n^{-\rho})$. \item[{\rm(ii)}] For $m$-dependent sequence $\{X_t\}$ with Condition {\rm\ref{as:tail}} being satisfied, if $0<\rho<(\vartheta-1)/(3\vartheta-2)$, there exist two constants $c_1>0$ depending only on $(\rho,\vartheta)$ and $c_2>0$ depending only on $(\gamma_1,\vartheta)$ such that $\Delta_{n,r}=O_{\rm p}\{B_n^2n^{-c_1}(\log p)^{c_2}\}+O(B_n^2m^2n^{-\rho})$. \item[{\rm(iii)}] For the sequence $\{X_t\}$ satisfying the model \eqref{eqn:weakly_dependent_ts_fdp}, assume Condition {\rm\ref{as:tail}} is satisfied and $\Phi_{\psi_\nu,0}<\infty$ and $\Psi_{2,\alpha}<\infty$ for some $\alpha,\nu\in(0,\infty)$, if $0<\rho<\min\{(2\alpha+2\vartheta-3)/(2\vartheta-2),1/(2\alpha)\}$ with $\vartheta>(3-2\vartheta)/2$, there exist two constants $c_1>0$ depending only on $(\rho,\vartheta,\alpha)$ and $c_2>0$ depending only on $(\alpha,\vartheta,\gamma_{1},\nu)$ such that $\Delta_{n,r}=O_{\rm p}\{(B_n\Phi_{\psi_{\nu,0}}+B_n^2+\Phi_{\psi_{\nu,0}}^2)n^{-c_1}(\log p)^{c_2}\}+O(n^{-\rho}\Psi_{2,0}\Psi_{2,\alpha}\varpi_{n})$ with $\varpi_n=(\log n)I(\alpha=1)+n^{1-\alpha}I(\alpha\neq 1)$. \end{itemize} \end{theorem} \begin{remark}[Bias and stochastic components of the kernel-type estimator for the long-run covariance matrix] Define $\Xi^* = \sum_{j=-n+1}^{n-1} \mathcal K(j/b_n)H_j$, where $H_{j}=n^{-1}\sum_{t=j+1}^{n}\mathbb{E}(X_{t}X_{t-j}^{{ \mathrm{\scriptscriptstyle T} }})$ if $j\geq 0$ and $H_{j}=n^{-1}\sum_{t=-j+1}^{n}\mathbb{E}(X_{t+j}X_{t}^{{ \mathrm{\scriptscriptstyle T} }})$ otherwise. The terms $O(B_n^2n^{-\rho})$, $O(B_n^2m^2n^{-\rho})$ and $O(n^{-\rho}\Psi_{2,0}\Psi_{2,\alpha},\varpi_n)$ in (i), (ii) and (iii) are, respectively, bounds on the bias $|\Xi^*-\Xi|_\infty$ in the three cases. Thus the bias terms do not depend on $p$ (at least directly). On the other hand, the terms $O_{\rm p}\{B_n^2n^{-c_1}(\log p)^{c_2}\}$, $O_{\rm p}\{B_n^2n^{-c_1}(\log p)^{c_2}\}$ and $O_{\rm p}\{(B_n\Phi_{\psi_{\nu,0}}+B_n^2+\Phi_{\psi_{\nu,0}}^2)n^{-c_1}(\log p)^{c_2}\}$ are bounds on $|\hat{\Xi}_n-\Xi^*|_\infty$ in the three cases, respectively. \end{remark} Now, combining Remark \ref{remark:paraboot} and Theorem \ref{tm:cov_compare}, we see that our proposed parametric bootstrap procedure is asymptotically valid even if the dimension $p$ grows sub-exponentially fast in the sample size $n$. To implement the proposed parametric bootstrap, we need to solve two problems: (i) How to select bandwidth $b_n$ in practice? and (ii) How to generate $\hat{G}\sim N(0,\hat{\Xi}_n)$ efficiently when $p$ is large? For Problem (i), due to the positive definiteness of $\hat{\Xi}_n$ defined as \eqref{eq:kernelest} with the quadratic spectral kernel $\mathcal{K}_{{\rm QS}}(\cdot)$ defined as \eqref{eq:QSkernel}, we can use this kernel in our parametric bootstrap procedure. For $\mathcal{K}_{{\rm QS}}(\cdot)$, \cite{Andrews_1991} suggested the following data-driven procedure to select $b_n$. \begin{algorithm}[H] \caption{Data-driven procedure of bandwidth selection for $\mathcal{K}_{\rm QS}(\cdot)$} \begin{itemize}[leftmargin=1.55cm, rightmargin=0cm]\label{alg:1} \item[{\bf Step 1.}] For each $j\in[p]$, fit an AR(1) model to the $j$-th coordinate marginal sequence $\{X_{t,j}\}_{t=1}^n$. Denote by $\hat{\rho}_j$ and $\hat{\sigma}_j^2$, respectively, the estimated autoregressive coefficient and innovation variance. \\ \item[{\bf Step 2.}] The data-driven bandwidth $b_n=1.3221(\hat{a}n)^{1/5}$ with \[ \hat{a}=\frac{\sum_{j=1}^p4\hat{\rho}_j^2\hat{\sigma}_j^4(1-\hat{\rho}_j)^{-8}}{\sum_{j=1}^p\hat{\sigma}_j^4(1-\hat{\rho}_j)^{-4}}\,. \] \end{itemize} \end{algorithm} For Problem (ii), to generate a random vector $\hat{G}\sim N(0,\hat{\Xi}_n)$, the standard approach consists of three steps: (a) perform the Cholesky decomposition for the $p\times p$ matrix $\hat{\Xi}_n=L^{ \mathrm{\scriptscriptstyle T} } L$; (b) generate independent standard normal random variables $Z_1,\ldots,Z_p$ and let $Z=(Z_1,\ldots,Z_p)^{ \mathrm{\scriptscriptstyle T} }$; (c) perform the transformation $\hat{G}=L^{ \mathrm{\scriptscriptstyle T} } Z$. However, the computation complexity of the standard approach is $O(np^2+p^3)$ and it also requires a large storage space for $\{X_t\}_{t=1}^n$ and the estimated matrix $\hat{\Xi}_n$. To circumvent the high computing cost with large $p$, we propose an algorithm below which involves generating a random vector from an $n$-dimensional normal distribution instead. It is easy to check that such obtained $\hat{G}\sim N(0,\hat{\Xi}_n)$ conditionally on $\mathcal{X}_n$. \begin{algorithm}[H] \caption{Generating $\hat{G}\sim N(0,\hat{\Xi}_n)$ for $\hat{\Xi}_n$ given in \eqref{eq:kernelest} with $\mathcal{K}_{{\rm QS}}(\cdot)$} \begin{itemize}[leftmargin=1.55cm, rightmargin=0cm] \item[{\bf Step 1.}] Obtain the bandwidth $b_n$ by Algorithm \ref{alg:1}. Define an $n\times n$ matrix $\Theta=(\theta_{i,j})_{n\times n}$ with $\theta_{i,j}=\mathcal{K}_{{\rm QS}}\{(i-j)/b_n\}$. \\ \item[{\bf Step 2.}] Generate $Z=(Z_1,\ldots,Z_n)^{ \mathrm{\scriptscriptstyle T} }\sim N(0,\Theta)$ independent of $\{X_t\}_{t=1}^n$. Define \[ \hat{G}=\frac{1}{\sqrt{n}}\sum_{t=1}^nZ_tX_t\,. \] \end{itemize} \end{algorithm} \section{Applications} \label{sec:apps} In this section, we discuss several statistical applications of the Gaussian and bootstrap approximation results for high-dimensional dependent data developed in Sections \ref{sec:main_results} and \ref{sec:parametric_boostrap}. \subsection{Testing high-dimensional mean vector} \label{ex:test_mean} Given data $\{X_t\}_{t=1}^{n_1}$ with $\mathbb{E}(X_t)= \theta_x\in\mathbb{R}^p$ for any $t\in[n_1]$, it is of general interest in testing the hypothesis \begin{equation}\label{eq:one-sample} H_0 : \theta_x = 0 \quad \mbox{versus} \quad H_1 : \theta_x \neq 0\,. \end{equation} If there are another group of data $\{Y_t\}_{t=1}^{n_2}$ with $\mathbb{E}(Y_t)=\theta_y\in\mathbb{R}^p$ for any $t\in[n_2]$, we are also interested in the hypothesis testing problem \begin{equation}\label{eq:two-sample} H_0 : \theta_x = \theta_y \quad \mbox{versus} \quad H_1 : \theta_x \neq \theta_y\,. \end{equation} Hypotheses \eqref{eq:one-sample} and \eqref{eq:two-sample} are called, respectively, one-sample and two-sample mean testing problems in the literature. Lots of statistical inference problems in practice can be formulated as \eqref{eq:one-sample} and \eqref{eq:two-sample}. Generally, the $\ell^2$-type and $\ell^\infty$-type statistics are used to test the hypotheses \eqref{eq:one-sample} and \eqref{eq:two-sample} in the high-dimensional settings. With independent data, we refer to \cite{ChenQin_2010} and \cite{CaiLiuXia_2014} for the use of $\ell^2$-type statistic and $\ell^\infty$-type statistic in these testing problems, respectively. It has been well known that the $\ell^2$-type statistics are powerful for detecting relatively dense signals while the $\ell^\infty$-type statistics are preferable for detecting relatively sparse signals. In practice, we usually have less knowledge on whether the signals are dense or sparse. Let $\Gamma = \{\mathrm{Cov}(X_t)\}^{-1} = (\gamma_{i,j})_{p\times p}$. When $\Gamma$ is known, to combine the advantages of the $\ell^2$-type and $\ell^\infty$-type statistics, \cite{Zhang_2015} considered the following test statistic for \eqref{eq:one-sample} with independent data $\{X_t\}_{t=1}^{n_1}$: \begin{equation} \label{eqn:test_mean_vector_statistic} T_n(s) =\max_{1 \leq j_1< \cdots < j_s \leq p} \sum_{k=1}^s {n_1\bar{Z}_{j_k}^2 \over \gamma_{j_k,j_k}}\,, \end{equation} where $\bar{Z} = \Gamma \bar{X}:=(\bar{Z}_1,\ldots,\bar{Z}_p)^{ \mathrm{\scriptscriptstyle T} }$ with $\bar{X} = n_1^{-1}\sum_{t=1}^{n_1} X_t$. When $\Gamma$ is unknown, \cite{Zhang_2015} proposed a feasible analogue for $T_n(s)$ by replacing $\Gamma$ by its estimator $\hat{\Gamma}$. To simplify our presentation, we assume $\Gamma$ is known in the rest of this subsection. For dependent data, testing for white noise or serial correlation is a fundamental problem in statistical inference, as many testing problems in linear modelling can be transformed into a white noise test. Let $\{\varepsilon_t\}$ be a $d$-dimensional weakly stationary time series with mean zero. Denote by $\Sigma(k)={\rm Cov}(\varepsilon_{t+k},\varepsilon_t)$ the autocovariance of $\varepsilon_t$ at lag $k$. Given a prescribed integer $K$, the white noise hypothesis of $\{\varepsilon_t\}$ can be formulated as \begin{align}\label{eq:whitenoise} H_{0} : \Sigma(1) = \dots = \Sigma(K) = 0 \quad \mbox{versus} \quad H_1: H_0\,\,\mbox{is not true}\,. \end{align} Let $n_1=n-K$ and $X_t=\{{\rm vec}(\varepsilon_{t+1}\varepsilon_t^{ \mathrm{\scriptscriptstyle T} }),\ldots,{\rm vec}(\varepsilon_{t+K}\varepsilon_t^{ \mathrm{\scriptscriptstyle T} })\}^{ \mathrm{\scriptscriptstyle T} }$, where ${\rm vec}(A)$ denotes a row vector that collecting all the elements in $A$. Then the white noise hypothesis \eqref{eq:whitenoise} can be covered by the hypothesis \eqref{eq:one-sample} with $p=d^2K$. \cite{ChangYaoZhou_2017} proposed a bootstrap test based on the $\ell^{\infty}$-type statistic for the white noise hypothesis \eqref{eq:whitenoise} under the $\beta$-mixing assumption of $\{\varepsilon_t\}$. To enhance the power performance of \cite{ChangYaoZhou_2017}, we can use the test statistic $T_n(s)$ given in \eqref{eqn:test_mean_vector_statistic}. Notice that the distribution function of $T_n(s)$ can be written in terms of probability of the random vector $n_1^{1/2}\bar{Z}$ over a class of convex subsets of the form $\{w \in \R^p : \sum_{j \in \Theta_s} \gamma_{j,j}^{-1}w_j^2 \leq t \}$ with $\Theta_s = \{w \in \R^p : |w|_0 = s\}$ which is a subset of the class of all $s$-sparsely convex sets in $\mathbb{R}^{p}$. Select $\hat{G}=(\hat{G}_1,\ldots,\hat{G}_p)^{ \mathrm{\scriptscriptstyle T} }\sim N(0,\Gamma)$. Using the results developed in Sections \ref{sec:main_results} and \ref{sec:parametric_boostrap}, the null-distribution of $T_n(s)$ can be approximated by that of \[ \max_{1 \leq j_1 < \cdots < j_s \leq p} \sum_{k=1}^s \frac{\hat{G}_{j_k}^2}{\gamma_{j_k,j_k}} \] under both the $\alpha$-mixing assumption and physical dependency assumption, where $s$ can diverge with $n$ at some polynomial rate. For any $\delta\in(0,1)$, let $q_{\delta}$ be the upper $\delta$-quantile of the distribution of $\max_{1 \leq j_1 < \cdots < j_s \leq p} \sum_{k=1}^s \gamma_{j_k,j_k}^{-1}\hat{G}_{j_k}^2$. Given significant level $\delta$, we reject the null hypothesis of the white noise hypothesis \eqref{eq:whitenoise} if the test statistic $T_n(s)$ specified in \eqref{eqn:test_mean_vector_statistic} is larger than $q_\delta$. Our procedure allows arbitrary dependency among the components of $X_t$. \subsection{Change point detection} \label{ex:change_point_detection} Consider the problem of change point detection for high-dimensional distributions in a location family \[ X_t= \theta \cdot I(t > m) + \xi_t\,, \] where $\theta \in \mathbb{R}^p$ is the location-shift parameter and $\{\xi_t\}$ is a sequence of stationary time series noise with mean zero. If $\theta=0$ or $m\geq n$, there is no change point in $\{X_t\}_{t=1}^n$. \cite{YuChen_2019} proposed a procedure to test whether there exists change point in the data based on the $U$-statistic \[ U_n = (U_{n,1},\ldots,U_{n,p})^{ \mathrm{\scriptscriptstyle T} } = {n \choose 2}^{-1} \sum_{1 \leq i < j \leq n} h(X_i, X_j)\,, \] where $h : \mathbb{R}^p \times \mathbb{R}^p \to \mathbb{R}^p$ is an anti-symmetric kernel $h(x,y)=-h(y,x)$. The anti-symmetry of the kernel $h$ plays a key role in testing for the change point in terms of noise cancellations so that after proper normalization the distribution of $U_n$ can be approximated by that of a Gaussian analogue. Specifically, under the null hypothesis that there is no change point and assuming independent and identically distributed noise $\{\xi_t\}$ with common distribution $F$, \cite{YuChen_2019} showed that ${E}(U_n)= 0$ and the distribution of $U_n$ can be well approximated on max-rectangles by $N(0,4\Gamma/3)$, where $\Gamma={\rm Cov}\{g(X_1)\}$ and $g(x) = \mathbb{E}\{h(x,X_2)\}$ is the $\mathcal{L}^{2}(F)$ projection of $h$ onto a linear subspace. On the other hand, under the alternative hypothesis when the change point location $m$ is known, $U_n$ is a two-sample Mann-Whitney test statistic (see e.g., Chapter 12 in \cite{vanderVaart_1998}), and the signal distortion under certain nonlinear kernels can be controlled such that the between-sample change point signal is magnitude preserving. To practically calibrate the distribution of $\max_{j \in[p]} |n^{1/2} U_{n,j}|_\infty$, \cite{YuChen_2019} proposed a jackknife multiple bootstrap, which is powerful against alternatives with strong signals. However, validity of the jackknife multiple bootstrap with a general nonlinear kernel heavily relies on the independent and identically distributed assumption of the noise sequence $\{\xi_t\}_{t=1}^n$ \citep{ChenKato_2020}. For time series data, with the linear kernel $h(x,y)=x-y$ we may write \[ W_n=(W_{n,1},\ldots,W_{n,p})^{ \mathrm{\scriptscriptstyle T} } = {2\over\sqrt{n}(n-1)} \sum_{t=1}^n (n-2t+1) X_t\,, \] which can be viewed as one-pass CUSUM test statistic \citep{YuChen_2020}. Thus we can enhance the power performance of the change point test of \cite{YuChen_2019} in the setting of linear kernel by using the test statistic \[ T_n(s) = \max_{1 \leq j_1 < \dots < j_s \leq p} \sum_{k=1}^s W_{n,j_k}^2\,, \] which allows $s$ to diverge with $n$ at some polynomial rate. Let $\hat{G} = (\hat{G}_1,\ldots,\hat{G}_p)^{ \mathrm{\scriptscriptstyle T} }\sim N(0,\hat\Xi_n)$ with $\hat\Xi_n$ being a kernel type estimate of the long-run covariance matrix ${\rm Cov}(W_n)$ of the weighted sequence $\{2(n-1)^{-1}(n-2t+1)X_t\}_{t=1}^n$. Based on the results developed in Sections \ref{sec:main_results} and \ref{sec:parametric_boostrap}, the null-distribution of $T_n(s)$ can be calibrated by that of \[ \max_{1 \leq j_1 < \dots < j_s \leq p} \sum_{k=1}^s \hat{G}_{j_k}^2\, \] under both the $\alpha$-mixing assumption and physical dependence assumption. For any $\delta\in(0,1)$, let $q_{\delta}$ be the upper $\delta$-quantile of the distribution of $\max_{1 \leq j_1 < \cdots < j_s \leq p} \sum_{k=1}^s \hat{G}_{j_k}^2$. Given significant level $\delta$, we reject the null hypothesis that there is no change point if the test statistic $T_n(s) = \max_{1 \leq j_1 < \dots < j_s \leq p} \sum_{k=1}^s W_{n,j_k}^2>q_\delta$. Our procedure does not need to impose any specific structure assumption on the dependency among different components of $X_t$. \subsection{Confidence regions for the instantaneous covariance matrix and its inverse} \label{ex:test_covariance_matrix} Given $d$-dimensional dependent (possibly non-stationary) data $\{Y_t\}_{t=1}^n$ with mean zero and instantaneous covariance $\Sigma$, i.e., $\mathbb{E}(Y_t)=0$ and ${\rm Cov}(Y_t)=\Sigma$ for any $t\in[n]$, the instantaneous covariance matrix $\Sigma$ and the precision matrix $\Omega=\Sigma^{-1}=(\omega_{i,j})_{d\times d}$ quantify the dependence among the $d$ components of $Y_t$. For the estimation of $\Sigma$ and $\Omega$, we refer to \cite{BickelLevina_2008a,BickelLevina_2008b} and \cite{CaiLiuLuo_2011} for independent data, \cite{ChangQiuYaoZou_2018} and \cite{XuChenWu_2020} for dependent data. Confidence regions for $\Sigma$ and $\Omega$ can quantify the uncertainty in their estimates. For a given index set $\mathcal{S}\subset [d]^2$, denote by $\Sigma_{\mathcal{S}}$ and $\Omega_{\mathcal{S}}$ the vectors consisting, respectively, the entries of $\Sigma$ and $\Omega$ with their indices in $\mathcal{S}$. We are interested in constructing a class of confidence regions $\{\mathcal{C}_{\mathcal{S},\delta}\}_{0<\delta<1}$ for $\Sigma_{\mathcal{S}}$ such that \begin{equation}\label{eq:confidenceregion} \sup_{0<\delta<1}|\mathbb{P}(\Sigma_{\mathcal{S}}\in\mathcal{C}_{\mathcal{S},\delta})-\delta|\rightarrow0\quad\mbox{as $n,d\rightarrow\infty$}\,. \end{equation} Analogously, we can also consider the confidence regions $\{\mathcal{C}_{\mathcal{S},\delta}\}_{0<\delta<1}$ for $\Omega_{\mathcal{S}}$ such that \[ \sup_{0<\delta<1}|\mathbb{P}(\Omega_{\mathcal{S}}\in\mathcal{C}_{\mathcal{S},\delta})-\delta|\rightarrow0\quad\mbox{as $n,d\rightarrow\infty$}\,. \] Given observations $\{Y_t\}_{t=1}^n$, we can estimate $\Sigma$ as $\hat{\Sigma}=n^{-1}\sum_{t=1}^nY_tY_t^{ \mathrm{\scriptscriptstyle T} }$ and estimate $\Omega$ by fitting the node-wise regressions $ Y_{j,t}=\sum_{k\neq j}\beta_{j,k}Y_{k,t}+\epsilon_{j,t}$ for each $j\in[d]$. Denote by $\hat{\Omega}=(\hat{\omega}_{i,j})_{d\times d}$ the associated estimate of $\Omega$. See Equation (11) of \cite{ChangQiuYaoZou_2018} for details. Write $\epsilon_t=(\epsilon_{1,t},\ldots,\epsilon_{d,t})^{ \mathrm{\scriptscriptstyle T} }$ and let $V={\rm cov}(\epsilon_t):=(v_{i,j})_{d\times d}$. Proposition 1 of \cite{ChangQiuYaoZou_2018} indicates that $\hat{\omega}_{i,j}-\omega_{i,j}=-n^{-1}\sum_{t=1}^nv_{i,i}^{-1}v_{j,j}^{-1}(\epsilon_{i,t}\epsilon_{j,t}-v_{i,j})+o_{\rm p}\{(n\log d)^{-1/2}\}$, where the remainder term $o_{\rm p}\{(n\log d)^{-1/2}\}$ holds uniformly over $(i,j)\in[d]^2$. Let $p=|\mathcal{S}|$. We can see the leading terms of $\hat{\Sigma}_{\mathcal{S}}-\Sigma_{\mathcal{S}}$ and $\hat{\Omega}_{\mathcal{S}}-\Omega_{\mathcal{S}}$ can both be formulated as a general form $n^{-1}\sum_{t=1}^nX_t$ for some $p$-dimensional dependent sequence $\{X_t\}_{t=1}^n$. Write $\Xi={\rm Cov}(n^{-1/2}\sum_{t=1}^nX_t)$ and denote by $\hat{\Xi}_n$ the estimate of $\Xi$ given in Section \ref{sec:parametric_boostrap}. Let $\hat{G}^{(\mathcal{S})}=\{\hat{G}_1^{(\mathcal{S})},\ldots,\hat{G}_p^{(\mathcal{S})}\}^{ \mathrm{\scriptscriptstyle T} }\sim N(0,\hat{\Xi}_n)$ and define \[ f_s\{\hat{G}^{(\mathcal{S})}\}=\max_{1\leq j_1<\cdots<j_s\leq p}\sum_{k=1}^sa_k\{\hat{G}_{j_k}^{(\mathcal{S})}\}^2\,, \] where $a_1,\ldots,a_s>0$ denote the prescribed weights. For any $\delta\in(0,1)$, let $q_{\mathcal{S},\delta}$ be the upper $\delta$-quantile of the distribution of $f_s\{\hat{G}^{(\mathcal{S})}\}$, which can be determined by Monte Carlo simulation. Write $\hat{\Sigma}_{\mathcal{S}}=\{\hat{\sigma}_1^{(\mathcal{S})},\ldots,\hat{\sigma}_p^{(\mathcal{S})}\}^{ \mathrm{\scriptscriptstyle T} }$. We can select the confidence region $\mathcal{C}_{\mathcal{S},\delta}$ for $\Sigma_{\mathcal{S}}$ as follows: \begin{equation}\label{eq:conf} \mathcal{C}_{\mathcal{S},\delta}=\bigg\{\xi=(\xi_1,\ldots,\xi_p)^{ \mathrm{\scriptscriptstyle T} }:\max_{1\leq j_1<\cdots<j_s\leq p}\sum_{k=1}^sa_k\{\hat{\sigma}_{j_k}^{(\mathcal{S})}-\xi_{j_k}\}^2\leq q_{\mathcal{S},\delta}\bigg\}\,. \end{equation} Based on the results developed in Sections \ref{sec:main_results} and \ref{sec:parametric_boostrap}, we know such defined $\mathcal{C}_{\mathcal{S},\delta}$ satisfies \eqref{eq:confidenceregion}. Analogously, we can also obtain the confidence region $\mathcal{C}_{\mathcal{S},\delta}$ for $\Omega_{\mathcal{S}}$ in the same manner. \cite{ChangQiuYaoZou_2018} used this idea to construct the confidence region for $\Omega_{\mathcal{S}}$ with selecting $s=1$ in \eqref{eq:conf}, and established its validity under the $\beta$-mixing assumption. Using the results in Sections \ref{sec:main_results} and \ref{sec:parametric_boostrap}, we can show the validity of the confidence region defined as \eqref{eq:conf} in more general $\alpha$-mixing setting and physical dependence setting with diverging $s$. Such defined $\mathcal{C}_{\mathcal{S},\delta}$ can be applied for testing the structures and doing support recovering of $\Sigma$ and $\Omega$, respectively. See \cite{ChangQiuYaoZou_2018} for the usefulness of these confidence regions. As we have discussed in Section \ref{ex:test_mean}, involving a larger $s$ in \eqref{eq:conf} can enhance the power performance in finite-samples in comparison to the $\ell^\infty$-type statistic that with $s=1$. \section*{Appendix}

https://arxiv.org/abs/1910.05083

Sparse Reduced-Rank Regression for Simultaneous Rank and Variable Selection via Manifold Optimization

We consider the problem of constructing a reduced-rank regression model whose coefficient parameter is represented as a singular value decomposition with sparse singular vectors. The traditional estimation procedure for the coefficient parameter often fails when the true rank of the parameter is high. To overcome this issue, we develop an estimation algorithm with rank and variable selection via sparse regularization and manifold optimization, which enables us to obtain an accurate estimation of the coefficient parameter even if the true rank of the coefficient parameter is high. Using sparse regularization, we can also select an optimal value of the rank. We conduct Monte Carlo experiments and real data analysis to illustrate the effectiveness of our proposed method.

\section{Introduction} Reduced-rank regression (RRR), a useful tool for statistics, is based on a multivariate linear regression model with a low-rank constraint for the coefficient parameter. RRR reduces the number of parameters included in the model and enables us to easily interpret the relationship between response and predictor variables. Therefore, RRR is used in various fields of research, including genomics, signal processing, and econometrics. To date, various extensions for RRR have been proposed: high-dimensional RRR with a rank selection criterion (Bunea et al., 2011), RRR with a nuclear norm penalization (Yuan et al., 2007; Negahban and Wainwright, 2011), reduced-rank ridge regression and its kernel extensions (Mukherjee and Zhu, 2011), and reduced-rank stochastic regression with sparse singular value decomposition (Chen et al., 2013). In recent years, the number of response and predictor variables has been increasing. This causes difficulty in the estimating of parameters when the sample size is smaller than the number of the parameters included in the model. One approach for overcoming this problem is to apply a regularization method. During previous decades, sparse regularization methods, such as lasso (Tibshirani, 1996), has been the focus of attention, because they can estimate parameters and exclude irrelevant variables simultaneously. Various studies have considered a multivariate linear regression model with some sparse regularization (see, e.g., Rothman et al. (2010); Peng et al. (2010); Li et al. (2015)). Co-sparse factor regression (SFAR; Mishra et al. (2017)) was proposed in one such study. SFAR is based on both RRR and a factor analysis model by assuming that the coefficient parameter can be decomposed by singular value decomposition with both a low-rank constraint and sparsity for the singular vectors. For the estimation of parameters, Mishra et al. (2017) proposed the sequential factor extraction via co-sparse unit-rank estimation (SeCURE) algorithm. The SeCURE algorithm sequentially estimates the parameters with orthogonality and sparsity for each factor. However, the SeCURE algorithm fails to estimate the parameters when the number of latent factors is large, because the algorithm is a greedy estimation method based on the classical Gram-Schmidt orthogonalization algorithm and it is well known that the classical method does not guarantee that the optimal solution will be obtained (Bj{\"o}rck, 1967). To overcome this problem, we propose a factor extraction algorithm with rank and variable selection via sparse regularization and manifold optimization (RVSManOpt). Manifold optimization has demonstrated excellent performance over decades of study (Bak{\i}r al., 2004; Mishra et al., 2013; Tan et al., 2019). The minimization problem of the SFAR model can be reformulated in terms of manifold optimization. Manifold optimization enables us to solve the minimization problem by taking the geometric structure of the SFAR model into consideration. By estimating the parameters on the manifold, we simultaneously obtain all latent factors. In addition, in order to select the optimal value of the rank, we introduce a regularizer which induces a hard-thresholding operator. The remainder of the paper is organized as follows. In Section 2, we introduce RRR and derive the SFAR model from the factor regression model. In Section 3, we reformulate the minimization problem of the SFAR model based on manifold optimization. In Section 4, we provide the estimation algorithm based on manifold optimization and discuss the selection of tuning parameters. In Sections 5, Monte Carlo experiments and real data analysis support the efficacy of RVSManOpt. Concluding remarks which summarize our study are presented in Section 6. Supplementary materials and source codes of our proposed method are available at \url{https://github.com/yoshikawa-kohei/RVSManOpt}. \section{Preliminaries} Suppose that we obtain $n$ independent observations $\left\{ (\vec{y}_i, \vec{x}_i); i=1,\dots,n \right\}$, where $\vec{y}_i = \trans{\left[ y_{i1}, \ldots, y_{iq} \right]} \in \mathbb{R}^{q}$ is a $q$-dimensional vector of response variables and $\vec{x}_i = \trans{\left[ x_{i1}, \ldots, x_{ip} \right]} \in \mathbb{R}^p$ is a $p$-dimensional vector of predictor variables. When we set $\mat{Y} = \trans{\left[ \vec{y}_1, \ldots, \vec{y}_n \right]} \in \mathbb{R}^{n \times q}$ and $\mat{X} = \trans{\left[ \vec{x}_1, \ldots, \vec{x}_n \right]} \in \mathbb{R}^{n \times p}$, RRR (Anderson, 1951; Izenman, 1975; Reinsel and Velu, 1998) is formulated as \begin{align} \label{eq:RRR_model} \mat{Y} = \mat{X}\mat{C} + \mat{E},\quad \st \rank{{\mat{C}}} \leq r, \end{align} where $\mat{C} \in \mathbb{R}^{p \times q}$ is the coefficient matrix, which has rank at most $r = \min \left(\rank{\mat{X}}, q \right)$, and $\mat{E} = \trans{\left[ \vec{e}_1, \ldots, \vec{e}_n \right]} \in \mathbb{R}^{n \times q}$ is the error matrix, which consists of independent random error vectors $\vec{e}_i$ with mean $\mathrm{E}\left[ \vec{e}_i\right] = \vec{0}$ and covariance matrix $\mathrm{Cov}\left[ \vec{e}_i \right] = \mat{\Sigma} \ (i = 1,\dots,n)$. The estimator of the coefficient matrix $\mat{C}$ can be obtained by solving the minimization problem \begin{equation} \label{eq:min_prob_RRR_model} \min_{\mat{C}}\ \norm{\mat{Y} - \mat{X} \mat{C}}_F^2, \st \rank{\mat{C}} \leq r, \end{equation} where $\norm{\cdot}_F$ denotes the Frobenius norm. Mishra et al. (2017) proposed SFAR by extending RRR in terms of factor analysis. Before introducing SFAR, we describe the relationship between RRR and factor analysis. First, we consider the RRR model with a coefficient matrix $\mat{C}$ that is decomposed as \begin{align} \mat{C} = \mat{U} \trans{{\tilde{\mat{V}}}}, \end{align} where $\mat{U} \in \mathbb{R}^{p \times r}$ and $\tilde{\mat{V}} \in \mathbb{R}^{q \times r}$. Then we obtain the RRR model reformulated by \begin{align} \label{eq:RRR_reformulated} \mat{Y} = \mat{X} \mat{U} \trans{{\tilde{\mat{V}}}} + \mat{E}. \end{align} The equation (\ref{eq:RRR_reformulated}) is related to a factor analysis model: $\mat{XU}$ can be regarded as a common factor matrix and $\tilde{\mat{V}}$ can be regarded as a loading matrix. Furthermore, if we assume $\mathrm{E}[\vec{x}_i] = \vec{0}$ and $\mathrm{cov}[\vec{x}_i] = \mat{\Gamma}_i \ (i=1,\dots,n)$, then $\mathrm{cov}[\trans{{\mat{U}}} \vec{x}] = \trans{{\mat{U}}} \mat{\Gamma} \mat{U} = \mat{I}_r$ is derived. This in turn gives the following SFAR model. \begin{align} \mat{Y} = \mat{X} \mat{U} \mat{D} \trans{{\mat{V}}} + \mat{E}, \st \trans{{\mat{U}}} \mat{\Gamma} \mat{U} = \mat{I}_{r}, \trans{{\mat{V}}} \mat{V} = \mat{I}_{r}. \end{align} Here, the coefficient matrix is $\mat{C} = \mat{U} \mat{D} \trans{{\mat{V}}}$. The estimator of SFAR is obtained by solving the minimization problem \begin{align} \label{eq:co-sparse_factor_regression} \min_{\mat{U}, \mat{D},\mat{V}}\ \frac{1}{2} \norm{\mat{Y} - \mat{X} \mat{U} \mat{D} \trans{{\mat{V}}}}_F^2 + \lambda_1 \sum_{i=1}^{p} \sum_{j=1}^{r} w^{(u)}_{ij}|{u}_{ij}| + \lambda_2 \sum_{i=1}^{q} \sum_{j=1}^{r} w^{(v)}_{ij}|{v}_{ij}|, \nonumber \\ \st \trans{{\mat{U}}} \left(\frac{\trans{\mat{X}} \mat{X}}{n} \right) \mat{U}= \mat{I}_{r}, \trans{{\mat{V}}} \mat{V} = \mat{I}_{r}, \end{align} where $u_{ij}, v_{ij}$ are elements of $\mat{U}$ and $\mat{V}$, respectively, $w^{(u)}_{ij}, w^{(v)}_{ij}$ are adaptive weights with positive values proposed by Zou (2006), and $\lambda_1, \lambda_2 > 0$ are regularization parameters. The second and third terms are penalty functions inducing elementwise sparsity (Tibshirani, 1996). By solving this minimization problem, we obtain the estimator of the coefficient matrix $\hat{\mat{C}} = \hat{\mat{U}} \hat{\mat{D}} \trans{\hat{{\mat{V}}}}$. The minimization problem is solved under orthogonality and sparsity of the parameters. However, it is difficult to estimate the parameters directly. For this reason, Mishra et al. (2017) proposed the SeCURE algorithm. The SeCURE algorithm sequentially solves the minimization problem for the $k$-th latent factor given by \begin{align} \label{eq:Q_problem} \min_{d_k, \vec{u}_k, \vec{v}_k} \frac{1}{2} \norm{\mat{Y}_k - d_k \mat{X} \vec{u}_k \trans{\vec{v}_k}}_F^2 + \sum_{i = 1}^p w_{ki}^{(u)} |u_{ki}| + \sum_{i = 1}^q w_{ki}^{(v)} |v_{ki}|, \nonumber\\ \st d_k \geq 0, \trans{\vec{u}_k} \trans{\mat{X}} \mat{X} \vec{u}_k = n, \trans{\vec{v}_k}\vec{v}_k = 1, \end{align} where $k=1,\dots,r$, $\vec{u}_k$ and $\vec{v}_k$ are the $k$-th column vector of $\mat{U}$ and $\mat{V}$, respectively, and $\mat{Y}_k$ is defined by \begin{align} \mat{Y}_k = \mat{Y} - \sum_{j=1}^{k-1} d_j \mat{X} \vec{u}_j \trans{{\vec{v}_j}}, \end{align} in which $d_j$ is the $j$-th diagonal element of $\mat{D}$ and $\mat{Y}_1 = \mat{Y}$. By sequentially solving the minimization problem (\ref{eq:Q_problem}), we obtain the solutions $\hat{d}_k, \hat{\vec{u}}_k$, and $\hat{\vec{v}}_k$ which satisfy orthogonality and sparsity. When $\hat{\vec{u}}_k = \vec{0}$ or $\hat{\vec{v}}_k = \vec{0}$, the SeCURE algorithm updates $d_k = 0$. This means that the updates are terminated. In addition, the index $k$ that terminates the updates is regarded as the optimal value of the rank of the coefficient matrix $\mat{C}$. It should be noted that the estimation method for the minimization problem (\ref{eq:Q_problem}) is the block coordinate descent algorithm proposed by Chen et al. (2012). \section{Minimization problem of co-sparse factor regression via manifold optimization} The SeCURE algorithm fails to estimate the parameters for the $k$-th latent factor when $k$ is large, because the algorithm is based on the classical Gram-Schmidt orthogonalization algorithm. Note that the classical Gram-Schmidt orthogonalization algorithm does not produce an optimal solution, owing to rounding errors (Bj{\"o}rck, 1967). To overcome this problem, we reconsider this minimization problem in terms of manifold optimization. \subsection{Reformulation of the minimization problem as manifold optimization} To consider the minimization problem (\ref{eq:co-sparse_factor_regression}) in terms of manifold optimization, we use the fundamental geometric structure given by \begin{align} \label{eq:Stiefel} \St{r}{q} &:= \left\{ \mat{V} \in \mathbb{R}^{q \times r} \mid \trans{\mat{V}} \mat{V} = \mat{I}_r \right\}, \end{align} where $q \geq r$. Here, $\St{r}{q}$ is called the Stiefel manifold, which is the set of orthogonal matrices of size $q \times r$. Furthermore, we also use the generalized Stiefel manifold given by \begin{align} \label{eq:gen_Stiefel} \StG{r}{p} &:= \left\{ \mat{U} \in \mathbb{R}^{p \times r} \mid \trans{\mat{U}} \mat{G} \mat{U} = \mat{I}_r \right\}, \end{align} where $p \geq r$ and $\mat{G} \in \mathbb{R}^{p \times p}$ is a symmetric positive definite matrix. In this paper, we use $\mat{G} = \trans{\mat{X}} \mat{X} /n$. By utilizing the geometric structures (\ref{eq:Stiefel}) and (\ref{eq:gen_Stiefel}), the minimization problem (\ref{eq:co-sparse_factor_regression}) can be reformulated as \begin{align} \label{eq:riemannian-co-sparse-factor-regression} \min_{\substack{ \mat{U} \in \StG{r}{p},\\ \mat{D}\in \mathbb{R}^{r \times r},\\ \mat{V}\in \St{r}{q} }} \frac{1}{2} \norm{\mat{Y} - \mat{X} \mat{U} \mat{D} \trans{{\mat{V}}}}_F^2 + n\lambda_1 \sum_{i=1}^{p} \sum_{j=1}^{r} w^{(u)}_{ij}|{u}_{ij}| + n\lambda_2 \sum_{i=1}^{q} \sum_{j=1}^{r} w^{(v)}_{ij}|{v}_{ij}|. \end{align} The minimization problem (\ref{eq:riemannian-co-sparse-factor-regression}) is an unconstrained optimization problem, and solving it allows us to estimate all the parameters for all the latent factors at once. \subsection{Rank selection with sparse regularization} The reformulation of the minimization problem (\ref{eq:co-sparse_factor_regression}) gives us the unconstrained optimization problem (\ref{eq:riemannian-co-sparse-factor-regression}). However, we cannot select the optimal value of the rank of the coefficient matrix $\mat{C}$ because of not using a sequential estimating procedure, such as SeCURE. To overcome this drawback, we propose the following minimization problem: \begin{multline} \label{eq:riemannian-co-sparse-factor-regression-rank-select} \min_{\substack{ \mat{U} \in \StG{r}{p},\\ \mat{D}\in \mathbb{R}^{r \times r},\\ \mat{V}\in \St{r}{q} }} \frac{1}{2} \norm{\mat{Y} - \mat{X} \mat{U} \mat{D} \trans{\mat{V}}}_F^2 + n\lambda_1 \sum_{i=1}^{p} \sum_{j=1}^{r} w^{(u)}_{ij}|{u}_{ij}| \\ + n\alpha \lambda_2 \sum_{i=1}^{q} \sum_{j=1}^{r} w^{(v)}_{ij}|{v}_{ij}| + n\sqrt{q}(1-\alpha)\lambda_2 \sum_{i=1}^r w^{(d)}_{i} \mathds{1} (\vec{v}_i \neq \vec{0}), \end{multline} where $\mathds{1} ( \cdot )$ is an indicator function that returns $1$ if the condition is true and returns $0$ if the condition is false, $w^{(d)}_{i}$ is an adaptive weight with a positive value proposed by Zou (2006), and $\alpha$ is a tuning parameter having a value between zero and one. The group selection in the fourth term plays the role of the rank selection of the coefficient matrix $\mat{C}$. The tuning parameter $\alpha$ adjusts the trade-off between the third term and the fourth term. The two terms can be regarded as Sparse Group Lasso (Wu and Lange, 2008; Puig et al., 2009; Simon et al., 2013). The fourth term is a regularizer which induces a hard-thresholding operator. By imposing this regularization, we can estimate some column vectors of $\mat{V}$ as zero vectors. As a consequence, the model is constructed with a small number of latent factors. In that sense, the indicator function plays the role of selecting the rank of the coefficient matrix $\mat{C}$. The reason why we do not apply Group Lasso, which induces a soft-thresholding operator (Yuan and Lin, 2006), is to avoid a double shrinking effect for the parameter $\mat{V}$. If we assume that the fourth term corresponds to the Group Lasso, then such a double shrinking effect appears to occur. The double shrinking effect reduces the variance of the model, but it excessively increases the bias. To prevent the double shrinking effect for the parameter $\mat{V}$, we use a regularizer which induces a hard-thresholding operator, since it does not shrink the value of the parameter. \section{Implementation} \subsection{Computational algorithm} To estimate the parameters, we employ a manifold optimization method (Edelman et al., 1998; Absil et al., 2008). Manifold optimization can be performed for differentiable functions. However, the minimization problem (\ref{eq:riemannian-co-sparse-factor-regression-rank-select}) includes nondifferentiable penalty terms. For this reason, we handle the nondifferentiability by applying the manifold alternating direction method of multipliers (M-ADMM) proposed by Kovnatsky et al. (2016) to the minimization problem (\ref{eq:riemannian-co-sparse-factor-regression-rank-select}). Letting $\mat{U}^* \in \mathbb{R}^{p \times r}$ and $\mat{V}^*$ and $\mat{V}^{**} \in \mathbb{R}^{q \times r}$ denote variables for splitting nondifferentiable penalty terms from the minimization problem (\ref{eq:riemannian-co-sparse-factor-regression-rank-select}), we consider a minimization problem with equality constraints as follows: \begin{multline} \label{eq:riemannian-co-sparse-factor-regression-equaliity} \min_{\substack{ \mat{U} \in \StG{r}{p},\\ \mat{D}\in \mathbb{R}^{r \times r},\\ \mat{V}\in \St{r}{q} }} \frac{1}{2} \norm{\mat{Y} - \mat{X} \mat{U} \mat{D} \trans{\mat{V}}}_F^2 + n\lambda_1 \sum_{i=1}^{p} \sum_{j=1}^{r} w^{(u)}_{ij}|{u}_{ij}^*| \\ + n\alpha \lambda_2 \sum_{i=1}^{q} \sum_{j=1}^{r} w^{(v)}_{ij}|{v}_{ij}^*| + n\sqrt{q}(1-\alpha)\lambda_2 \sum_{i=1}^r w^{(d)}_{i} \mathds{1} (\vec{v}_i^{**} \neq \vec{0}) ,\\ \st \mat{U} = \mat{U}^*,\ \mat{V} = \mat{V}^* = \mat{V}^{**}, \end{multline} where $u_{ij}^*, v_{ij}^*$ are the $(i,j)$-th elements of $\mat{U}^*$ and $\mat{V}^*$, respectively, and $\vec{v}_i^{**}$ is an $i$-th column vector of $\mat{V}^{**}$. When we let $\mat{\Omega} \in \mathbb{R}^{p \times r}$ and $\mat{\Phi}$ and $\mat{\Psi} \in \mathbb{R}^{q \times r}$ denote the dual variables, we obtain a scaled augmented Lagrangian (Boyd et al., 2011) as follows: \begin{multline} \mathcal{L}(\mat{U}, \mat{D}, \mat{V}, \mat{U}^*, \mat{V}^*, \mat{V}^{**}, \mat{\Omega}, \mat{\Phi}, \mat{\Psi}) = \frac{1}{2} \norm{\mat{Y} - \mat{X} \mat{U} \mat{D} \trans{\mat{V}}}_F^2\\ + n\lambda_1 \sum_{i=1}^{p} \sum_{j=1}^{r} w^{(u)}_{ij}|u_{ij}^*| + n\alpha \lambda_2 \sum_{i=1}^{q} \sum_{j=1}^{r} w^{(v)}_{ij}|v_{ij}^*| + n\sqrt{q}(1-\alpha)\lambda_2 \sum_{i=1}^r w^{(d)}_{i} \mathds{1} (\vec{v}_i^{**} \neq \vec{0}) \\ + \frac{\rho_1}{2} \norm{\mat{U} - \mat{U}^* + \mat{\Omega}}_F^2 + \frac{\rho_2}{2} \norm{\mat{V} - \mat{V}^* + \mat{\Phi}}_F^2 + \frac{\rho_3}{2} \norm{\mat{V} - \mat{V}^{**} + \mat{\Psi}}_F^2, \end{multline} where $\rho_1, \rho_2, \rho_3 > 0 $ are penalty parameters. For this study, we fixed $\rho_1 = \rho_2 = \rho_3 = 1$. M-ADMM alternately updates each parameter to minimize the augmented Lagrangian. The estimators of elements in $\mat{U}^*$ and $\mat{V}^*$ indicate whether each element of the parameter is zero. The estimators of column vectors in $\mat{V}^{**}$ indicate whether each vector of the parameter is a zero vector. In the M-ADMM procedure, we initialize the parameters by using $\tilde{\mat{U}} \in \mathbb{R}^{p \times r}, \tilde{\mat{D}} = \mathrm{diag} (\tilde{d}_1, \dots, \tilde{d}_r), \tilde{\mat{V}} \in \mathbb{R}^{q \times r}$. Here, $\tilde{\mat{U}}$ is calculated by $(\trans{\mat{X}} \mat{X})^{-} \trans{\mat{X}} \mat{Y} \tilde{\mat{V}} \tilde{\mat{D}}^{-1}$, where the $k$-th diagonal element of $\tilde{\mat{D}}^2$ is the $k$-th eigenvalue of $(1/n) \trans{\mat{Y}} \mat{X} (\trans{\mat{X}} \mat{X})^{-} \trans{\mat{X}}\mat{Y}$, and the $k$-th column vector of $\tilde{\mat{V}}$ is the $k$-th eigenvalue of $(1/n) \trans{\mat{Y}} \mat{X} (\trans{\mat{X}} \mat{X})^{-} \trans{\mat{X}}\mat{Y}$. We set the adaptive weights $w_{ij}^{(u)}, w_{ij}^{(v)}, w_{i}^{(d)}$ as \begin{align} w^{(u)}_{ij} &= \frac{1}{\left|\tilde{u}_{ij} \right|^{{\gamma^u}}}, \quad i=1,\dots,p, j=1,\dots,r,\\ w^{(v)}_{ij} &= \frac{1}{\left|\tilde{v}_{ij} \right|^{{\gamma^v}}}, \quad i=1,\dots,q, j=1,\dots,r,\\ w^{(d)}_{i} &= \frac{1}{|\tilde{d}_{i} |^{{\gamma^d}}}, \quad i=1,\dots,r, \end{align} where $\gamma^u$, $\gamma^v$, $\gamma^d > 0$ are tuning parameters. The parameters $\mat{U}$ and $\mat{V}$ are estimated by a gradient descent algorithm based on manifold optimization. For example, the procedure for estimating $\mat{U}$ can be represented by the following. \begin{enumerate} \item At a given iteration $s$, calculate the Euclidean gradient $\nabla \mathcal{L}_{\mat{U}^{(s)}}$. \item Project $\nabla \mathcal{L}_{\mat{U}^{(s)}}$ onto the tangent space $\mathcal{T}_{\mat{U}^{(s)}} \StG{p}{r}$ using orthogonal projection $\mathcal{P}_{\mat{U}^{(s)}}(\cdot)$ to obtain the gradient $\mathrm{grad} \mathcal{L}_{\mat{U}^{(s)}}$ on the manifold. \item Update the parameter $\mat{U}^{(s)}$ by retraction $\mathcal{R}_{\mat{U}^{(s)}}(- t\ \mathrm{grad} \mathcal{L}_{\mat{U}^{(s)}})$ to obtain the parameter $\mat{U}^{(s+1)}$, where $t \in \mathbb{R}$ is an Armijo step size described in Absil et al. (2008). \end{enumerate} The necessary notation is shown in Table \ref{tb:notations_manifold}. In the same way, we estimate the parameter $\mat{V}$ on the manifold. The detailed calculation of the updates is described in the Appendix. This algorithm is called the \textit{factor extraction algorithm with rank and variable selection via sparse regularization and manifold optimization} (RVSManOpt). RVSManOpt is summarized as Algorithm \ref{alg:r-sgfar_admm}. \begin{table}[htbp] \centering \caption{Notation for the manifold optimization algorithm} \begin{tabular}{ll} \hline \multicolumn{2}{c}{Generalized Stiefel manifold for parameter $\mat{U}$}\\ Metric & $\langle \mat{U}_1, \mat{U}_2 \rangle = \Tr ( \trans{\mat{U}_1} \mat{G} \mat{U}_2), \mat{G} = \trans{\mat{X}}\mat{X}/n$\\ Tangent space & $\mathcal{T}_\mat{U} \StG{p}{r} = \{ \mat{Z} \in \mathbb{R}^{p \times r} | \trans{\mat{U}} \mat{G} \mat{Z} + \trans{\mat{Z}} \mat{G} \mat{U} = \mat{0} \}$\\ Projection onto tangent space & $\mathcal{P}_\mat{U}(\mat{Z}) = \mat{Z} - \mat{U} \mathrm{sym}(\trans{\mat{U}}\mat{G} \mat{Z}), \mathrm{sym}(\mat{M}) = \frac{1}{2} (\mat{M} + \trans{\mat{M}} )$\\ Gradient & $\mathrm{grad} \mathcal{L}_\mat{U} = \mathcal{P}_\mat{U}(\nabla \mathcal{L}_\mat{U})$\\ Retraction mapping & $\mathcal{R}_\mat{U}(\mat{Z})= \sqrt{\mat{G}}^{-1} \mathrm{qf} \left(\sqrt{\mat{G}}(\mat{U} + \mat{Z}) \right)$,\\ & $\mathrm{qf}(\mat{A})$ denotes the $\mat{Q}$ factor of the QR decomposition of $\mat{A} = \mat{Q}\mat{R}$\\ \\ \multicolumn{2}{c}{Stiefel manifold for parameter $\mat{V}$}\\ Metric & $\langle \mat{V}_1, \mat{V}_2 \rangle = \Tr ( \trans{\mat{V}_1} \mat{V}_2)$\\ Tangent space & $\mathcal{T}_\mat{V} \St{q}{r} = \{ \mat{Z} \in \mathbb{R}^{q \times r} | \trans{\mat{V}} \mat{Z} + \trans{\mat{Z}} \mat{V} = \mat{0} \}$\\ Projection onto tangent space & $\mathcal{P}_\mat{V}(\mat{Z}) = \mat{Z} - \mat{V} \mathrm{sym}(\trans{\mat{V}} \mat{Z})$\\ Gradient & $\mathrm{grad} \mathcal{L}_\mat{V} = \mathcal{P}_\mat{V}(\nabla \mathcal{L}_\mat{V})$\\ Retraction mapping & $\mathcal{R}_\mat{V}(\mat{Z}) = \mathrm{qf}(\mat{V} + \mat{Z})$\\ \hline \end{tabular} \label{tb:notations_manifold} \end{table} \begin{spacing}{0.9} \begin{algorithm}[H] \caption{Factor Extraction Algorithm with Rank and Variable Selection via Sparse Regularization and Manifold Optimization (RVSManOpt)} \label{alg:r-sgfar_admm} \begin{algorithmic}[1] \Require Initial values $\mat{U}^{(0)} = \tilde{\mat{U}}, \mat{D}^{(0)} = \tilde{\mat{D}} , \mat{V}^{(0)} = \tilde{\mat{V}}, {\mat{U}^* }^{(0)}= \mat{U}^{(0)}, {\mat{V}^{*}}^{(0)} ={\mat{V}^{**}}^{(0)} = \mat{V}^{(0)}, \mat{\Omega}^{(0)} = \mat{0}, \mat{\Phi}^{(0)}, \mat{\Psi}^{(0)} = \mat{0}$ \For{$s=0,1,\ldots$} \State \textbf{$\mat{U}$ Step:} Update $\mat{U}^{(s+1)} \gets \mathcal{R}_{\mat{U}^{(s)}}(-t_u^{(s)} \mathrm{grad} \mathcal{L}_{\mat{U}^{(s)}})$, $t_u^{(s)}$ is the Armijo step size. \State \textbf{$\mat{V}$ Step:} Update $\mat{V}^{(s+1)} \gets \mathcal{R}_{\mat{V}^{(s)}}(-t_v^{(s)} \mathrm{grad} \mathcal{L}_{\mat{V}^{(s)}})$, $t_v^{(s)}$ is the Armijo step size. \State \textbf{$\mat{D}$ Step:} Update $\mat{D}^{(s+1)} \gets \mathrm{diag}\left( \frac{1}{n}\trans{{\mat{V}^{(s+1)}}} \trans{\mat{Y}} \mat{X} \mat{U}^{(s+1)} \right)$. \State \textbf{$\mat{U}^*$ Step:} \For{$i=1,\ldots,p$} \For{$j=1,\ldots,r$} \State Update ${u_{ij}^*}^{(s+1)} \gets \mathrm{S} \left(u_{ij}^{(s+1)} + \omega_{ij}^{(s)}, \frac{n\lambda_1 w_{ij}^{(u)}}{\gamma_1} \right)$. \EndFor \EndFor \State \textbf{$\mat{V}^*$ Step:} \For{$i=1,\ldots,q$} \For{$j=1,\ldots,r$} \State Update ${v_{ij}^{*}}^{(s+1)} \gets \mathrm{S} \left(v_{ij}^{(s+1)} + \phi_{ij}^{(s)}, \frac{n\alpha \lambda_2 w_{ij}^{(v)}}{\gamma_2} \right)$. \EndFor \EndFor \State \textbf{$\mat{V}^{**}$ Step:} \For{$i=0,1,\ldots,r$} \State Update ${\vec{v}^{**}_{i}}^{(s+1)} \gets \mathrm{H} \left({\vec{v}}_{i}^{(s+1)} + \vec{\psi}_{i}^{(s)}, \sqrt{\frac{2n\sqrt{q}(1-\alpha)\lambda_2 w_i^{(d)}}{\gamma_3}} \right)$. \EndFor \State \textbf{$\mat{\Omega}$ Step:} Update $\mat{\Omega}^{(s+1)} \gets \mat{\Omega}^{(s)} + \mat{U}^{(s+1)} - {\mat{U}^*}^{(s+1)}$. \State \textbf{$\mat{\Phi}$ Step:} Update $\mat{\Phi}^{(s+1)} \gets \mat{\Phi}^{(s)} + \mat{V}^{(s+1)} - \mat{{V}^*}^{(s+1)} $. \State \textbf{$\mat{\Psi}$ Step:} Update $\mat{\Psi}^{(s+1)} \gets \mat{\Psi}^{(s)} + \mat{V}^{(s+1)} - \mat{{V}^{**}}^{(s+1)} $. \If{$\mathrm{convergence}$} \State \textbf{break}. \EndIf \EndFor \State $\hat{\mat{U}} \gets \mat{U}^*; \hat{\mat{U}} \gets \hat{\mat{U}}(:, \vec{v}_i^{**} \neq \vec{0})$, $\hat{\mat{D}} \gets \mat{D}(:, \vec{v}_i^{**} \neq \vec{0})$, $\hat{\mat{V}} \gets \mat{V}^*; \hat{\mat{V}} \gets \hat{\mat{V}}(:, \vec{v}_i^{**} \neq \vec{0})$ \State \textbf{return} $\hat{\mat{U}}, \hat{\mat{D}}, \hat{\mat{V}}$ \end{algorithmic} \end{algorithm} \end{spacing} \subsection{Selection of tuning parameters} We have six tuning parameters: $\lambda_1, \lambda_2, \alpha, \gamma^u, \gamma^v$, and $\gamma^d$. To avoid a high computational cost, $\alpha, \gamma^u, \gamma^v$, and $\gamma^d$ are fixed in advance. We set the values of these tuning parameters according to the situation. The tuning parameter $\alpha$ is set to a large value when a sparse regularization is more important than a regularization for selecting the rank of the coefficient matrix $\mat{C}$. Larger values of tuning parameters $\gamma^u, \gamma^v$, and $\gamma^d$ correspond to a higher data dependence. To select the remaining two tuning parameters, $\lambda_1$ and $\lambda_2$, we use the Bayesian information criterion (BIC) given by \begin{align} \label{eq:BIC} \mathrm{BIC} = \log \left\{ \mathrm{SSE}_{\lambda_1, \lambda_2} / nq \right\} + \left\{ \log(qn)/(nq) \right\} df_{\lambda_1, \lambda_2}, \end{align} where $\mathrm{SSE}_{\lambda_1, \lambda_2}$ is the sum of squared errors of prediction defined by \begin{equation} \mathrm{SSE}_{\lambda_1, \lambda_2} = \norm{\mat{Y} - \mat{X} \hat{\mat{U}} \hat{\mat{D}} \trans{\hat{\mat{V}}}}_F^2, \end{equation} and $df_{\lambda_1, \lambda_2}$ is the degree of freedom which evaluates the sparsity of the estimates $\hat{\mat{U}}$ and $\hat{\mat{V}}$ defined by \begin{equation} df_{\lambda_1, \lambda_2} = \sum_{i=1}^{p} \sum_{j=1}^r \mathds{1}(\hat{u}_{ij} \neq 0) + \sum_{i=1}^{q} \sum_{j=1}^r \mathds{1}(\hat{v}_{ij} \neq 0) - 1. \end{equation} We select the tuning parameters $\lambda_1$ and $\lambda_2$ which minimize the BIC. The candidates values of $\lambda_1, \lambda_2$ are taken from equally spaced values in the interval $[\lambda_{\max}, \lambda_{\min}]$. We set $\lambda_{\max}=1$ and $\lambda_{\min}=10^{-15}$ in our numerical studies. \section{Numerical study} \subsection{Monte Carlo simulations} We conducted Monte Carlo simulations to illustrate the efficacy of RVSManOpt. In our simulation study, we generated 50 datasets from the model: \begin{align} \mat{Y} = \mat{XC} + \mat{E}, \end{align} where $\mat{Y} \in \mathbb{R}^{n \times q}$ is a response matrix, $\mat{X} \in \mathbb{R}^{n \times p}$ is a predictor matrix, $\mat{C} \in \mathbb{R}^{p \times q}$ is a coefficient matrix, and $\mat{E} = \trans{[\vec{e}_1, \dots \vec{e}_n]} \in \mathbb{R}^{n \times q}$ is an error matrix. Each row of $\mat{X}$ followed a multivariate normal distribution $\mathcal{N}(\vec{0}, \mat{\Gamma})$, where $\mat{\Gamma} = [\gamma_{ij}]$ is a $p \times p$ covariance matrix with $\gamma_{ij} = 0.5^{|i-j|}$ for $i,j = 1,\dots,p$. We generated each row of $\mat{E}$ by $\vec{e}_i \overset{\mathrm{i.i.d.}}{\sim} \mathcal{N}(\vec{0}, \sigma^2 \mat{\Delta})$, where $\mat{\Delta} = [\delta_{ij}]$ is a $q \times q$ matrix with $\delta_{ij} = \rho^{|i-j|}$ and $\sigma$ is determined according to the signal-to-noise ratio defined by $\mathrm{SNR} = \norm{d_r \mat{X} \vec{u}_r \trans{\vec{v}}}_2 / \norm{\mat{E}}_2 = 0.5$. We considered the ranks of the coefficient matrix as follows: $r \in \{3, 5, 7, 10, 12 \}$. We generated the coefficient matrix $\mat{C} = \mat{UD}\trans{\mat{V}}$, where $\mat{U} = [\vec{u}_1, \dots, \vec{u}_r]$, $\mat{D} = \mathrm{diag}(d_1, \dots, d_r)$, $\mat{V} = [\vec{v}_1, \dots, \vec{v}_r]$. Specifically, we set \begin{align*} d_k &= 5 + 0.1(k-1), \quad k = 1, \dots ,r,\\ \vec{u}_k &= \bar{\vec{u}}_k/\norm{\bar{\vec{u}}_k}_2,\\ \bar{\vec{u}}_1 &= \trans{[\check{\vec{u}}, \mathrm{rep}(0, p-8)]}, \bar{\vec{u}}_k = \trans{[\mathrm{rep}(0,5(k-1)),\check{\vec{u}}, \mathrm{rep}(0, p-(5k+3))]},\\ \check{\vec{u}} &= [1, -1, 1, -1, 0.5, -0.5, 0.5, -0.5],\\ \vec{v}_k &= \bar{\vec{v}}_k/\norm{\bar{\vec{v}}_k}_2,\\ \bar{\vec{v}}_1 &= \trans{[\check{\vec{v}}, \mathrm{rep}(0, q-4)]}, \bar{\vec{v}}_k = \trans{[\mathrm{rep}(0,4(k-1)), \check{\vec{v}}, \mathrm{rep}(0, q-4k)]},\\ \check{\vec{v}} &= [1, -1, 0.5, -0.5], \end{align*} where $\mathrm{rep}(a, b)$ represents the vector of length $b$ with all elements having the value $a$. We considered four cases. In Cases 1 and 2, we set $n=400, p=80$, and $q=50$ in common, and we set the correlation as $\rho=0.3$ (Case 1) or $\rho=0.5$ (Case 2). In Cases 3 and 4, we set $n=400, p=120$, and $q=60$ in common, and we set the correlation as $\rho=0.3$ (Case 3) or $\rho=0.5$ (Case 4). To demonstrate the efficacy of RVSManOpt, we compared RVSManOpt with the SeCURE with an adaptive lasso (SeCURE(AL)), and the SeCURE with an adaptive elastic net (SeCURE(AE)). For 50 datasets, we measured the estimation accuracy $\mathrm{Er}(\mat{XC}) $ and the selected rank absolute error $\mathrm{Er}(r)$. These are defined as \begin{align} \mathrm{Er}(\mat{XC}) &= \frac{1}{50} \sum_{k=1}^{50}\frac{\norm{\mat{\Gamma}^{\frac{1}{2}}(\hat{\mat{C}}^{(k)} - \mat{C}^{(k)})}_F^2}{nq},\\ \mathrm{Er}(r) &= \frac{1}{50} \sum_{k=1}^{50}|\hat{r}^{(k)} - r^{(k)}|, \end{align} where $\mat{C}^{(k)}$ is the true coefficient matrix, $r^{(k)}$ is the true rank of the coefficient matrix $\mat{C}^{(k)}$, $\hat{\mat{C}}^{(k)}$ is an estimated coefficient matrix, and $\hat{r}^{(k)}$ is the selected rank of coefficient matrix $\hat{\mat{C}}^{(k)}$ for the $k$-th dataset. In order to evaluate the sparsity, we computed the F-measure defined by \begin{align*} \textrm{F-measure} &= \frac{1}{50} \sum_{k=1}^{50} 2 \cdot \frac{\mathrm{Recall}^{(k)} \cdot \mathrm{Precision}^{(k)}}{\mathrm{Recall}^{(k)} + \mathrm{Precision}^{(k)}}, \end{align*} where $\mathrm{Recall}^{(k)}$ and $\mathrm{Precision}^{(k)}$ are defined by \begin{align*} \textrm{Recall}^{(k)} &= \frac{ \sum_{ij} \left| \left\{ u_{ij} \neq 0 \wedge \hat{u}_{ij}^{(k)} \neq 0 \right\} \right| }{ \sum_{ij} \left| \left\{u_{ij} \neq 0\right\} \right| } + \frac{ \sum_{ij} \left| \left\{ v_{ij} \neq 0 \wedge \hat{v}_{ij}^{(k)} \neq 0 \right\} \right| }{ \sum_{ij} \left| \left\{v_{ij} \neq 0\right\} \right| },\\ \textrm{Precision}^{(k)} &= \frac{ \sum_{ij} \left| \left\{ u_{ij} \neq 0 \wedge \hat{u}_{ij}^{(k)} \neq 0 \right\} \right| }{ \sum_{ij} \left| \left\{\hat{u}_{ij}^{(k)} \neq 0\right\} \right|} + \frac{ \sum_{ij} \left| \left\{ v_{ij} \neq 0 \wedge \hat{v}_{ij}^{(k)} \neq 0 \right\} \right| }{ \sum_{ij} \left| \left\{\hat{v}_{ij}^{(k)} \neq 0\right\} \right|}, \end{align*} for which $\hat{u}_{ij}^{(k)}$ and $\hat{v}_{ij}^{(k)}$ are respectively elements of the estimated $\mat{U}$ and $\mat{V}$ for the $k$-th dataset and $|\{\cdot \}|$ is the count of the elements of set $\{\cdot \}$. All implementations were done in \texttt{R} (ver. 3.6) (R Core Team, 2018). Tables \ref{tb:resuts_table1}, \ref{tb:resuts_table2}, \ref{tb:resuts_table3}, and \ref{tb:resuts_table4} show summaries of the results for, respectively, Cases 1 to 4 of the Monte Carlo simulations. As shown, when the rank of the coefficient matrix $\mat{C}$ is high, RVSManOpt outperforms other algorithms in terms of both $\mathrm{Er}(\mat{XC})$ and $\mathrm{Er}(r)$. In contrast, when the rank of the coefficient matrix $\mat{C}$ is low, the performances of all algorithm are approximately the same. Moreover, the F-measure gives almost the same value for RVSManOpt, SeCURE(AL) and SeCURE(AE). Therefore, our proposed RVSManOpt achieves performance superior to those of other methods in terms of both the estimation accuracy and rank selection. Fig. \ref{fig:ErXC} shows box-plots of $\mathrm{Er}(\mat{XC})$ for Case 1. The box-plots for the other cases are essentially same and are available as the supplementary materials. When the rank of the coefficient matrix $\mat{C}$ is high, we observe many outliers in the box-plots of SeCURE(AL) and SeCURE(AE). These outliers indicate that SeCURE(AL) and SeCURE(AE) fail to estimate parameters many times. On the other hand, the number of the outliers produced by RVSManOpt is small, and hence RVSManOpt performs the other methods in terms of stable estimation. \begin{table}[H] \begin{center} \caption{Results for Monte Carlo simulations in Case 1. For simplicity, $\mathrm{Er}(\mat{XC})$ is multiplied by $10^4$.} \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} c|lcccc} \hline TrueRank & Method & $\mathrm{Er}(\mat{XC})$ & $\mathrm{Er}(\mat{XC})$(sd) & F-measure & $\mathrm{Er}(r)$\\ \hline \hline & \multicolumn{5}{c}{Case 1 : $n = 400, p=80, q = 50, \rho= 0.3$}\\ \multirow{3}{*}{3} & RVSManOpt & 0.42 & 0.20 & 0.59 & 0.00 \\ & SeCURE(AL) & 0.45 & 1.35 & 0.56 & 0.04 \\ & SeCURE(AE) & 0.45 & 1.35 & 0.56 & 0.04 \\ \\ \multirow{3}{*}{5} & RVSManOpt & 1.00 & 0.42 & 0.41 & 0.00 \\ & SeCURE(AL) & 1.18 & 1.84 & 0.42 & 0.10 \\ & SeCURE(AE) & 0.99 & 1.27 & 0.42 & 0.06 \\ \\ \multirow{3}{*}{7} & RVSManOpt & 1.76 & 0.70 & 0.33 & 0.00 \\ & SeCURE(AL) & 3.53 & 4.97 & 0.34 & 0.42 \\ & SeCURE(AE) & 4.18 & 5.81 & 0.33 & 0.54 \\ \\ \multirow{3}{*}{10} & RVSManOpt & 4.06 & 2.30 & 0.27 & 0.00 \\ & SeCURE(AL) & 7.83 & 8.04 & 0.28 & 0.82 \\ & SeCURE(AE) & 8.37 & 8.33 & 0.28 & 0.92 \\ \\ \multirow{3}{*}{12} & RVSManOpt & 7.25 & 4.50 & 0.24 & 0.00 \\ & SeCURE(AL) & 13.55 & 13.65 & 0.24 & 1.60 \\ & SeCURE(AE) & 14.16 & 14.62 & 0.24 & 1.70 \\ \hline \end{tabular*} \label{tb:resuts_table1} \end{center} \end{table} \begin{figure}[H] \centering \includegraphics[width=\linewidth]{Case1ErXC.pdf} \caption{Box-plots of scaled $\mathrm{Er}(\mat{XC})$ for each rank $r$ in Case 1. } \label{fig:ErXC} \end{figure} \begin{table}[H] \begin{center} \caption{Results for Monte Carlo simulations in Case 2. For simplicity, $\mathrm{Er}(\mat{XC})$ is multiplied by $10^4$.} \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} c|lcccc} \hline TrueRank & Method & $\mathrm{Er}(\mat{XC})$ & $\mathrm{Er}(\mat{XC})$(sd) & F-measure & $\mathrm{Er}(r)$\\ \hline \hline & \multicolumn{5}{c}{Case 2 : $n = 400, p=80, q = 50, \rho= 0.5$}\\ \multirow{3}{*}{3} & RVSManOpt & 0.38 & 0.19 & 0.57 & 0.00 \\ &SeCURE(AL) & 0.30 & 0.98 & 0.56 & 0.02 \\ &SeCURE(AE) & 0.30 & 0.98 & 0.56 & 0.02 \\ \\ \multirow{3}{*}{5} & RVSManOpt & 1.10 & 0.66 & 0.42 & 0.00 \\ &SeCURE(AL) & 1.79 & 3.20 & 0.41 & 0.22 \\ &SeCURE(AE) & 1.69 & 3.15 & 0.41 & 0.20 \\ \\ \multirow{3}{*}{7} & RVSManOpt & 1.85 & 0.81 & 0.34 & 0.00 \\ &SeCURE(AL) & 4.45 & 6.40 & 0.33 & 0.58 \\ &SeCURE(AE) & 4.56 & 6.39 & 0.33 & 0.60 \\ \\ \multirow{3}{*}{10} & RVSManOpt & 4.39 & 2.63 & 0.27 & 0.00 \\ &SeCURE(AL) & 8.53 & 8.62 & 0.28 & 1.00 \\ &SeCURE(AE) & 9.33 & 9.12 & 0.28 & 1.14 \\ \\ \multirow{3}{*}{12} & RVSManOpt & 7.47 & 4.76 & 0.24 & 0.00 \\ &SeCURE(AL) & 12.13 & 11.59 & 0.24 & 1.34 \\ &SeCURE(AE) & 11.92 & 11.48 & 0.24 & 1.30 \\ \hline \end{tabular*} \label{tb:resuts_table2} \end{center} \end{table} \begin{table}[H] \begin{center} \caption{Results for Monte Carlo simulations in Case 3. For simplicity, $\mathrm{Er}(\mat{XC})$ is multiplied by $10^4$.} \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} c|lcccc} \hline TrueRank & Method & $\mathrm{Er}(\mat{XC})$ & $\mathrm{Er}(\mat{XC})$(sd) & F-measure & $\mathrm{Er}(r)$\\ \hline \hline & \multicolumn{5}{c}{Case 3 : $n = 400, p=120, q = 60, \rho= 0.3$}\\ \multirow{3}{*}{3} & RVSManOpt & 0.40 & 0.18 & 0.56 & 0.00 \\ &SeCURE(AL) & 0.33 & 0.94 & 0.56 & 0.04 \\ &SeCURE(AE) & 0.33 & 0.94 & 0.56 & 0.04 \\ \\ \multirow{3}{*}{5} & RVSManOpt & 0.73 & 0.21 & 0.42 & 0.00 \\ &SeCURE(AL) & 0.84 & 1.58 & 0.41 & 0.10 \\ &SeCURE(AE) & 1.18 & 1.88 & 0.41 & 0.18 \\ \\ \multirow{3}{*}{7} & RVSManOpt & 1.77 & 1.09 & 0.34 & 0.00 \\ &SeCURE(AL) & 3.92 & 5.63 & 0.34 & 0.64 \\ &SeCURE(AE) & 3.81 & 5.62 & 0.34 & 0.62 \\ \\ \multirow{3}{*}{10} & RVSManOpt & 3.41 & 1.50 & 0.27 & 0.00 \\ &SeCURE(AL) & 8.87 & 9.18 & 0.27 & 1.44 \\ &SeCURE(AE) & 9.01 & 9.27 & 0.27 & 1.46 \\ \\ \multirow{3}{*}{12} & RVSManOpt & 4.59 & 2.22 & 0.23 & 0.00 \\ &SeCURE(AL) & 12.02 & 12.44 & 0.24 & 1.84 \\ &SeCURE(AE) & 14.93 & 14.47 & 0.24 & 2.48 \\ \hline \end{tabular*} \label{tb:resuts_table3} \end{center} \end{table} \begin{table}[H] \begin{center} \caption{Results for Monte Carlo simulations in Case 4. For simplicity, $\mathrm{Er}(\mat{XC})$ is multiplied by $10^4$.} \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} c|lcccc} \hline TrueRank & Method & $\mathrm{Er}(\mat{XC})$ & $\mathrm{Er}(\mat{XC})$(sd) & F-measure & $\mathrm{Er}(r)$\\ \hline \hline & \multicolumn{5}{c}{Case 4 : $n = 400, p=120, q = 60, \rho= 0.5$}\\ \multirow{3}{*}{3} & RVSManOpt & 0.42 & 0.34 & 0.57 & 0.06 \\ &SeCURE(AL) & 0.31 & 0.94 & 0.56 & 0.04 \\ &SeCURE(AE) & 0.31 & 0.94 & 0.56 & 0.04 \\ \\ \multirow{3}{*}{5} & RVSManOpt & 1.24 & 0.75 & 0.43 & 0.08 \\ &SeCURE(AL) & 0.96 & 1.93 & 0.41 & 0.14 \\ &SeCURE(AE) & 1.31 & 2.19 & 0.40 & 0.22 \\ \\ \multirow{3}{*}{7} & RVSManOpt & 2.28 & 1.11 & 0.35 & 0.04 \\ &SeCURE(AL) & 4.20 & 5.27 & 0.33 & 0.72 \\ &SeCURE(AE) & 5.14 & 6.07 & 0.32 & 0.94 \\ \\ \multirow{3}{*}{10} & RVSManOpt & 4.14 & 1.99 & 0.28 & 0.10 \\ &SeCURE(AL) & 5.92 & 5.64 & 0.27 & 0.86 \\ &SeCURE(AE) & 7.10 & 7.08 & 0.27 & 1.10 \\ \\ \multirow{3}{*}{12} & RVSManOpt & 6.46 & 3.61 & 0.24 & 0.14 \\ &SeCURE(AL) & 13.70 & 12.58 & 0.24 & 2.22 \\ &SeCURE(AE) & 13.49 & 12.66 & 0.24 & 2.18 \\ \hline \end{tabular*} \label{tb:resuts_table4} \end{center} \end{table} \subsection{Application to yeast cell cycle dataset} We applied RVSManOpt to yeast cell cycle data (Spellman et al., 1998). The dataset was available in the \texttt{secure} package (Mishra et al., 2017) in the software \texttt{R}. The analysis of the yeast cell cycle enables us to identify transcription factors (TFs) which regulate ribonucleic acid (RNA) levels within the eukaryotic cell cycle. The dataset contains two components: the chromatin immunoprecipitation (ChIP) data and eukaryotic cell cycle data. The binding information of a subset of 1790 genes and 113 TFs was included in the ChIP data (Lee et al., 2002). The cell cycle data were obtained by measuring the RNA levels every 7 minutes for 119 minutes, thus a total of 18 time points, to cover two cycles. Since the dataset contained missing values, we complemented them by using the \texttt{imputeMissings} package in \texttt{R}. By complementing the dataset, we can use all $n = 1790$ genes and analyze the relationship between the RNA levels in the $q = 18$ time points and $p = 113$ TFs. We compared RVSManOpt with SeCURE(AL) and SeCURE(AE) by computing the number of selected experimentally confirmed TFs among the total number of the selected TFs and the proportion of experimentally confirmed TFs. It is known that there are 21 TFs which have been experimentally confirmed to be involved in the cell cycle regulation (Wang et al., 2007). Table \ref{tb:yeast_cell_results} gives the results of a real data analysis. In RVSManOpt, the proportion of experimentally confirmed TFs is larger than both SeCURE(AL) and SeCURE(AE). RVSManOpt estimated $\hat{r} = 5$, while SeCURE(AL) and SeCURE(AE) estimated $\hat{r} = 4$. This result means that RVSManOpt may capture the latent structure of the yeast cell cycle data more precisely by identifying 5 latent factors. Fig. \ref{fig:Yeast} shows estimated transcription levels of three of the experimentally confirmed TFs selected by RVSManOpt. The rest of the 12 experimentally confirmed TFs are available as the supplementary materials. Fig. \ref{fig:Yeast} indicates that the estimated transcription levels followed two cycles. It was experimentally confirmed that the transcription levels in the cell cycle did cover a two cycle time period. Thus, RVSManOpt was demonstrated to accurately estimate the cycles of data. \begin{table}[H] \begin{center} \caption{Results of analysis of yeast cell cycle dataset.} \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} c|ccc} \hline Method & \begin{tabular}{c}Total number of\\ selected TFs \end{tabular}& \begin{tabular}{c}Total number of\\ selected confirmed TFs\end{tabular} & \begin{tabular}{c}Proportion of\\ experimentally confirmed TFs\end{tabular}\\ \hline \hline RVSManOpt & 15 & 67 & 0.224\\ SeCURE(AL) & 17 & 83 & 0.205\\ SeCURE(AE) & 17 & 83 & 0.205\\ \hline \end{tabular*} \label{tb:yeast_cell_results} \end{center} \end{table} \begin{figure}[H] \centering \subcaptionbox{ACE2\label{fig:ACE2}} {\includegraphics[width=.32\linewidth]{ACE2.pdf}} \subcaptionbox{BAS1\label{fig:BAS1}} {\includegraphics[width=.32\linewidth]{BAS1.pdf}} \subcaptionbox{MBP1\label{fig:MBP1}} {\includegraphics[width=.32\linewidth]{MBP1.pdf}} \caption{Plots of estimated transcription levels of 3 experimentally confirmed TFs selected by RVSManOpt.}\label{fig:Yeast} \end{figure} \section{Concluding Remarks} We proposed a minimization problem of SFAR on a Stiefel manifold and developed the factor extraction algorithm with rank and variable selection via sparse regularization and manifold optimization (RVSManOpt). RVSManOpt surpassed the traditional estimation procedure, which fails when the rank of the coefficient matrix is high. Numerical comparisons including Monte Carlo simulations and a real data analysis supported the usefulness of RVSManOpt. In general, it is challenging to estimate parameters while preserving both orthogonality and sparsity. Mishra et al. (2017) indicates that enforcing orthogonality collapses sparsity and does not work from the viewpoint of prediction. Therefore, it may be unnecessary to construct a model with perfect orthogonality if we focus on prediction. Also, the recent paper by Absil and Hosseini (2019) discusses a theory of manifold optimization for non-smooth functions. It would be interesting to develop RVSManOpt based on this theory. We leave these as future topics.

https://arxiv.org/abs/1005.0181

Discontinuity of the Lyapunov Exponent

We study discontinuity of the Lyapunov exponent. We construct a limit-periodic Schrödinger operator, of which the Lyapunov exponent has a positive measure set of discontinuities. We also show that the limit-periodic potentials, whose Lyapunov exponent is discontinuous, are dense in the space of limit-periodic potentials.

\section{Introduction} In this paper, we construct examples of ergodic Schr\"odinger operators $H_{\omega}$, whose Lyapunov exponent $L(E)$ is discontinuous. In addition to being a result of interest of its own, we were motivated by the following observation. Denote by \begin{equation} \mathcal{Z} = \{E: \quad L(E) = 0\} \end{equation} the set where the Lyapunov exponent vanishes. It was shown by Deift and Simon in \cite{deisi} that the measure of this set satisfies \begin{equation}\label{eq:deisi} |\mathcal{Z}| \leq 4. \end{equation} Introduce the essential closure of $\mathcal{Z}$ by \begin{equation} \overline{\mathcal{Z}}^{\mathrm{ess}} = \{E:\quad \forall \varepsilon > 0:\ |\mathcal{Z} \cap (E - \varepsilon, E + \varepsilon)| > 0\}. \end{equation} Denote by $\sigma_{\mathrm{ac}}(H_{\omega})$ the absolutely continuous spectrum of $H_{\omega}: \ell^2({\mathbb Z}) \to \ell^2({\mathbb Z})$. By Kotani theory, we have that for almost every $\omega$ \begin{equation}\label{eq:kotani} \sigma_{\mathrm{ac}}(H_{\omega}) = \overline{\mathcal{Z}}^{\mathrm{ess}}. \end{equation} This result can for example be found in \cite{da1}, \cite{sikot}, or Theorem~5.17 in \cite{tjac}. Being naive, one might assume that from these two statements that \begin{equation}\label{eq:bigquestion} |\sigma_{\mathrm{ac}}(H_{\omega})| \leq 4 \end{equation} holds for almost every $\omega$. However, we do not know if this is true and trying to show this was a big motivation to write this paper. To see that \eqref{eq:bigquestion} cannot be a simple consequence of \eqref{eq:deisi} and \eqref{eq:kotani} let $\mathcal{Z}$ be the complement of a Cantor set of measure $> 4$ in $[-4,4]$. Then \eqref{eq:deisi} holds, but $\overline{\mathcal{Z}}^{\mathrm{ess}} = [-4,4]$ is a set of measure $8$. Let us now discuss our main result and its relevance to this conjecture. We begin by introducing some notation. Let $(\Omega,\mu)$ be a probability space, $T:\Omega\to\Omega$ an invertible ergodic transformation, and $f:\Omega\to {\mathbb R}$ a bounded measurable function. For $\omega\in\Omega$ we introduce the potentials $V_\omega(n)=f(T^n\omega)$. $H_{\omega} = \Delta + V_{\omega}$ denotes the Schr\"odinger operator with potential $V_{\omega}$. Here $\Delta u (n) = u(n+1) + u(n-1)$ denotes the discrete Laplacian. We introduce the transfer matrices \begin{equation} A_{N}(E, V_{\omega}) = \prod_{n=1}^{N} \begin{pmatrix} E - V_{\omega}(N- n) & -1 \\ 1 & 0 \end{pmatrix}. \end{equation} The Lyapunov exponent is then defined by \begin{equation}\label{eq:defLE} L(E) = \lim_{N\to\infty} \frac{1}{N} \int_{\Omega} \log\|A_{N}(E, V_{\omega})\| d\mu(\omega). \end{equation} Our main result is \begin{theorem}\label{thm:intA} There exists $(\Omega,T,f)$ such that $L(E)$ vanishes for generic $E \in [-4,4]$. Furthermore for every $\omega\in\Omega$, we have that $[-4,4] \subseteq \sigma(H_{\omega})$. \end{theorem} Our proof does not give any control on the measure of the set $\mathcal{Z}$, where the Lyapunov exponent vanishes. However, it is sufficient to show that \begin{corollary}\label{cor:intA} We have that $L(E)$ is discontinuous on a set of $E$ of positive Lebesgue measure. \end{corollary} \begin{proof} Denote by $A$ the set of $E_0 \in [-4,4]$ such that $L(E)$ is continuous at $E_0$. By Theorem~\ref{thm:intA}, we have that $A \subseteq \mathcal{Z}$. Hence by \eqref{eq:deisi}, we have that $|A| \leq 4$. In particular, the Lyapunov exponent is discontinuous on $B = [-4,4] \setminus A$ with measure $|B| \geq 4$. \end{proof} Our proof of Theorem~\ref{thm:intA} is done by constructing a limit-periodic potential explicitly with these properties. We will discuss the details in the next section, where we also show that for a dense set of limit-periodic potentials the Lyapunov exponent has at least one discontinuity. This is achieved by adapting the argument of Johnson \cite{john} to the discrete setting. We furthermore wish to point out that Thouvenot has constructed in \cite{thuv} discontinuity points of the Lyapunov exponent for matrix-valued cocycles. \bigskip Let us now discuss the relevance to whether \eqref{eq:bigquestion} holds. One can show that if $A$ is a closed set, then $|\overline{A}^{\mathrm{ess}}| = |A|$. Now, the existence of large sets of discontinuity of the Lyapunov exponent stops us from concluding that $\mathcal{Z}$ is closed and thus \eqref{eq:bigquestion}. \bigskip Artur Avila has informed us of an alternative proof of Theorem~\ref{thm:intA}, which is based on the work of Bochi and Viana \cite{bv}. The main idea is that if $T$ has plenty of periodic points, then one can ensure that $|\sigma(\Delta + f(T^n \omega))| > 4$ for almost every $\omega$. Then one uses to show the results of \cite{bv} to conclude that by perturbing $f$ slightly the Lyapunov exponent $L(E)$ vanishes on a dense set in $\sigma(\Delta + f(T^n \omega))$, and still $|\sigma(\Delta + f(T^n \omega))| > 4$. Then discontinuity of the Lyapunov exponent follows as in the proof of Corollary~\ref{cor:intA}. \section{Limit-periodic potential and further results} We will begin this section by discussing some basics about limit-periodic potentials. For further informations, we refer to the appendix of the paper \cite{as} of Avron and Simon, and to the survey by Gan \cite{g}. Given a bounded sequence $V: {\mathbb Z} \to {\mathbb R}$, we denote by $\Omega_V$ the hull of its translates. That is \begin{equation} \Omega_V = \overline{\{V_m,\quad m \in {\mathbb Z}\}}^{\ell^{\infty}({\mathbb Z})}, \end{equation} where $V_m(n) = V(n - m)$. If $\Omega_V$ is compact in the $\ell^{\infty}({\mathbb Z})$ topology, then $V$ is called {\em almost-periodic}. The shift map on $\ell^\infty({\mathbb Z})$ becomes a translation on the group $\Omega_V$ and it is uniquely ergodic with respect to the Haar measure of $\Omega_V$. $V$ is called {\em limit-periodic}, if there exists a sequence of periodic potentials $V^k$ such that \begin{equation} V = \lim_{k \to \infty} V^k \end{equation} in the $\ell^\infty({\mathbb Z})$ topology. It should be remarked that limit-periodic $V$ are almost-periodic. In fact, then $\Omega_V$ has the extra structure of being a Cantor group, see \cite{g} for details. The proof of Theorem~\ref{thm:intA} will proceed by explicitly constructing such sequences of $V^k$. We will furthermore denote by $L(E, V^k)$ the Lyapunov exponent of such a sequence of periodic potentials. Let us recall some properties of a $p_k$ periodic potentials $V^k$. \begin{enumerate} \item The spectrum $\sigma(\Delta + V^k)$ is a finite union of intervals. That is \begin{equation} \sigma(\Delta + V^k) = \bigcup_{j=1}^{g} [\alpha_j, \beta_j], \end{equation} where $g \geq 2$ and $\alpha_j < \beta_j$. \item The Lyapunov exponent $L(E, V^k)$ is continuous and vanishes on $L(E, V^k)$. \end{enumerate} It should furthermore be pointed out that, one always has $|\sigma(\Delta + V^k)| \leq 4$. Let us now comment further on Theorem~\ref{thm:intA}. \begin{remark} The potential constructed in Theorem~\ref{thm:intA} is limit-periodic. In particular, we provide an example of a limit-periodic potential whose spectrum contains an interval. \end{remark} This is not the first known example with this property, since P\"oschel provided some in \cite{poesch}. However, our construction has the advantage of being relatively elementary in comparison to P\"oschel's KAM-type proof. \bigskip In order to state our other result, we denote by $\mathcal{L}$ the space of all limit-periodic potentials. We have \begin{theorem} \label{thm:main2} There is a dense set of $V$ in $\mathscr{L}$ for which the Lyapunov exponent $L(E)$ is discontinuous. \end{theorem} Here the dense set cannot be improved to a generic set, since Damanik and Gan in \cite{dg} show that there is a generic set of $V$ in $\mathscr{L}$ such that the Lyapunov exponent is continuous. Furthermore, we should point out that the proof of the previous theorem largly parallels the construction of Johnson in \cite{john} of similar examples in the continuum setting. However, the observation of denseness is new. \bigskip Define the individual Lyapunov exponent by \begin{equation}\label{eq:defolLE} \overline{L}(E,\omega) = \limsup_{N\to\infty} \frac{1}{N} \log\|A_{N}(E, V_{\omega})\| . \end{equation} An important aspect of our construction will be to replace the Lyapunov exponent $L(E)$ defined in \eqref{eq:defLE} by $\overline{L}(E, \omega)$. This is possible by the following result. \begin{proposition} Assume that $T: \Omega\to\Omega$ is uniquely ergodic. For each $\omega\in \Omega$, there exists a set $\mathcal{E}_\omega$ of zero Lebesgue measure such that for $E \in {\mathbb R}\setminus\mathcal{E}_{\omega}$ \begin{equation} L(E) = \overline{L}(E,\omega). \end{equation} \end{proposition} \begin{proof} This is a consequence of Theorem~2.1. in \cite{k}. \end{proof} Thus, we obtain that for fixed $\omega$ \begin{equation}\label{eq:lyapequal} L(E) = \lim_{N \to \infty} \frac{1}{N} \log\|A_{N}(E, V_{\omega})\| \end{equation} for almost every $E$. We will furthermore need Thouless' formula \begin{equation}\label{eq:thouless} L(E) = \int \log|t - E| dN(t), \end{equation} where $N(t)$ is the integrated density of states. See for example Theorem~5.15 in Teschl's book \cite{tjac}. \section{Proof of Theorem~\ref{thm:intA}} In order to simplify the notation, we introduce \begin{definition}\label{def:epsdense} A collection of open intervals $\Sigma$ is called $\varepsilon$-dense in $[-4,4]$ if for $E \in [-4,4]$, there exists $I \in \Sigma$ such that $I \subseteq [E - \varepsilon, E + \varepsilon]$. \end{definition} We will construct a sequence of periodic potentials $V^k$ with the following properties: \begin{enumerate} \item $V^k$ is $p_k$ periodic and $\|V^{k} - V^{k-1}\|_{\infty} \leq \frac{1}{2^{k-1}}$. \item $V^{k}(n) = V^{k-1}(n)$ for $0 \leq n \leq p_{k-1}$. \item For $1 \leq l \leq k$ and $E \in \sigma(\Delta + V^l)$ \begin{equation} \frac{1}{p_k} \log\|A_{p_k}(E, V^{k})\| \leq \sum_{s = l+1}^{k+1} \frac{1}{2^s}. \end{equation} \item There is a set $\Sigma_k$ consisting of open intervals $I \subseteq \sigma(\Delta + V^k)$ which is $2^{-k}$-dense in $[-4,4]$. \item For each $I_{k-1} \in \Sigma_{k-1}$ there exists $I_{k} \in \Sigma_k$ such that $I_k \subseteq I_{k-1}$. \end{enumerate} We begin by showing that the existence of such $V^k$ implies Theorem~\ref{thm:intA}. From (i), one obtains that the limit \begin{equation} V(n) = \lim_{k\to\infty} V^{k}(n) \end{equation} exists uniformly and is bounded. \begin{lemma} Let $l \geq 1$, then for $E \in \sigma(\Delta + V^l)$ \begin{equation} \liminf_{N \to \infty} \frac{1}{N} \log\|A_N(E, V)\| \leq \frac{1}{2^l}. \end{equation} \end{lemma} \begin{proof} By (ii), we have $\frac{1}{p_k} \log\|A_{p_k}(E, V)\| = \frac{1}{p_k} \log\|A_{p_k}(E, V^k)\|$. By (iii), we obtain that for $E \in \sigma(\Delta + V^l)$ and $k \ge l$ $$ \frac{1}{p_k} \log\|A_{p_k}(E, V)\| \leq \sum_{s = l+1}^{k+1} \frac{1}{2^s} \leq \frac{1}{2^l}, $$ which implies the claim. \end{proof} Combining this lemma with \eqref{eq:lyapequal}, we obtain that for almost every $E \in \sigma(\Delta + V^l)$ \begin{equation}\label{eq:LEleq2l} L(E)\leq \frac{1}{2^l}. \end{equation} We also have \begin{proposition} For $E_0 \in [-4,4]$ we have \begin{equation}\label{eq:liminfle} \liminf_{E \to E_0} L(E) = 0. \end{equation} \end{proposition} \begin{proof} Pick any $E_0 \in [-4,4]$ and $k \geq 1$. By (iv), we can choose $I_k \in \Sigma_k$ such that $I_k \subseteq [E_0 - \frac{1}{2^k}, E_0 + \frac{1}{2^k}]$. By (v), we can choose a sequence of intervals $$ I_k \supseteq I_{k+1} \supseteq \dots $$ Let $\hat{E} \in \bigcap_{l\geq k} \overline{I_l}$. Since $\hat{E} \in \overline{I_k}$, we have $|E_0 - \hat{E}| < \frac{1}{2^k}$. By \eqref{eq:LEleq2l}, we can choose $E_l \in I_l$ with $|E_l - \hat{E}| \leq \frac{1}{2^l}$ and $L(E_l) \leq \frac{1}{2^l}$. Hence $$ 0 \leq \liminf_{E \to \hat{E}} L(E) \leq \lim_{l \to \infty} L(E_l) = 0, $$ showing \eqref{eq:liminfle}. Since $k \geq 1$ was arbitary, the claim follows. \end{proof} We now come to \begin{proof}[Proof of Theorem~\ref{thm:intA}] Let $$ L_N(E) = \frac{1}{2^N} \int_{\Omega} \log\| A_{2^N}(E,V_\omega) \| d\mu(\omega). $$ $L_N(E)$ is continuous in $E$ and decreasing in $N$. In particular, this implies $$ L(E) = \inf_{N \geq 1} L_N(E). $$ Introduce $$ U_l = \bigcup_{N\geq 1} \left\{E:\quad L_N(E) < \frac{1}{l} \right\}. $$ Since $L_N(E)$ is continuous, $U_l$ is open. Let us now show that $U_l$ is also dense. Let $E_0 \in [-4,4]$ and $\varepsilon > 0$. By the previous proposition, there exists $E \in [E_0-\varepsilon, E_0+\varepsilon]$ such that $L(E) < \frac{1}{2 l}$. Furthermore, there must be an $N_0 \geq 1$ such that $|L(E) - L_{N_0}(E)| < \frac{1}{2 l}$ and thus that $L(E) < \frac{1}{l}$. This implies $E \in U_l$ and thus that $U_l$ is dense. Introduce $$ U = \bigcap_{l \geq 1} U_l. $$ Since each of the $U_l$ is open and dense, we have by the Baire category theorem that also $U$ is dense. The claim follows. \end{proof} In the following subsections, we will explain how to construct $V^k$ given $V^{k-1}$. In the next subsection, we will construct a sequence of potentials $\hat{V}_m$ such that properties (i), (ii), (iv), and (v) hold. Then, we will show in Subsection~\ref{sub:ensuringiii} that (iii) holds if $m$ is chosen large enough. \subsection{A sequence of potentials such that (i), (ii), (iv) and (v) hold} We will need the following lemma, describing the generalized eigenfunctions of a periodic operator. A proof can be given using that $A_p(E, V)$ has an eigenvalue of modulus $1$ for $E \in \sigma(\Delta + V)$ and we omit it for brevity. \begin{lemma} \label{lem:perisolu} Let $V$ be a $p$ periodic potential and $E \in \sigma(\Delta + V)$. There is a bounded solution $\psi$ of $(\Delta + V)\psi = E \psi$ such that for every $m \in {\mathbb Z}$ \begin{equation} \label{eq:perisolu} \sum^{p}_{k=1} |\psi(m + k)|^2 = 1. \end{equation} \end{lemma} Let $I^{1}, \dots, I^{L}$ be an enumeration of the intervals in $\Sigma_{k-1}$. For each $I^{l}$, we choose a subinterval $\tilde{I}^{l}$ of length $< \frac{1}{2^{k+1}}$. Denote by $\widetilde{\Sigma}_{k-1}$ the collection of intervals $\tilde{I}^l$. Introduce $\delta$ by \begin{equation} \delta = \min_{I \in \widetilde{\Sigma}} |I|. \end{equation} Clearly, $0 < \delta < \frac{1}{2^{k+1}}$. Choose $m_0$ so large that $\delta \sqrt{ m_0} > 4$ and let $\tilde{p}_{k-1} = m_0 p_{k-1}$. We will treat $V^{k-1}$ as a $\tilde{p}_{k-1}$ periodic potential. Write \begin{equation} \Lambda_{j,m} = [m \tilde{p}_{k-1} + (j + 4) \tilde{p}_{k-1}, m \tilde{p}_{k-1} +(j + 5) \tilde{p}_{k-1} - 1]. \end{equation} Introduce the potentials $\hat{V}_m$ by \begin{equation} \hat{V}_m(n) = \begin{cases} V^{k-1}(n), & 0 \leq n \leq m \tilde{p}_{k-1} - 1 ; \\ V^{k-1}(n) + \frac{j}{2^{k+1}}, & n \in \Lambda_{j,m}\quad -4\leq j \leq 3. \end{cases} \end{equation} Note that $\hat{V}_m$ will be $\hat{p}_m = m \tilde{p}_{k-1} + 8 \tilde{p}_{k-1}$ periodic. We will later let $V^{k} = \hat{V}_m$ for some large $m$. We see that the claimed properties (i) and (ii) are straightforward. It remains to prove (iv) and (v). \begin{lemma} \label{lem:subinterval} For each $I \in \widetilde{\Sigma}_{k-1}$ and $-4\leq j\leq 3$, there exists an open interval $J$ such that \begin{equation}\label{eq:propJ} J \subseteq I + \frac{j}{2^{k+1}} \text{ and } J \subseteq \sigma(\Delta + \hat{V}_m). \end{equation} \end{lemma} \begin{proof} Let $I = (E_-,E_+)$ and $\hat{E} = \frac{E_- + E_+}{2}$. By Lemma~\ref{lem:perisolu}, there exists a function $\psi$ such that $$ (\Delta + V^{k-1}) \psi = \hat{E} \psi,\quad \text{and}\quad \sum^{p_{k-1}}_{n = 1} |\psi(n)|^2 = 1. $$ Let $\varphi$ be the restriction of $\psi$ to $\Lambda_{j,m}$. A computation shows $\|\varphi\| = \sqrt{m_0}$ and $$ (\Delta + \hat{V}_{m} - \hat{E} - \frac{j}{2^{k+1}}) \varphi(n) = \begin{cases} \psi(m \tilde{p}_{k-1} + (j + 4) \tilde{p}_{k-1}), & n = m \tilde{p}_{k-1} + (j + 4) \tilde{p}_{k-1} - 1;\\ - \psi(m \tilde{p}_{k-1} + (j + 4) \tilde{p}_{k-1} - 1), & n = m \tilde{p}_{k-1} + (j + 4) \tilde{p}_{k-1}; \\ - \psi(m \tilde{p}_{k-1} + (j + 5) \tilde{p}_{k-1} ), & n = m \tilde{p}_{k-1} + (j + 5) \tilde{p}_{k-1} -1;\\ \psi(m \tilde{p}_{k-1} + (j + 5) \tilde{p}_{k-1} - 1), & n = m \tilde{p}_{k-1} + (j + 5) \tilde{p}_{k-1};\\ 0, & \text{otherwise}. \end{cases} $$ So we have \begin{equation} \frac{\|(\Delta + \hat{V}_m - \hat{E} - \frac{j}{2^{k+1}}) \varphi \|}{\| \varphi \|} \leq \frac{2}{\sqrt{m_0}} < \frac{\delta}{2}. \end{equation} The above inequality implies that $$ \dist(\hat{E} + \frac{j}{2^{k+1}},\sigma(\Delta + \hat{V}_m)) < \frac{\delta}{2}. $$ Hence, we see that $\sigma(\Delta + \hat{V}_m) \cap (I + \frac{j}{2^{k+1}})$ is non empty since $|I + \frac{j}{2^{k+1}}| \geq \delta$. Since $\sigma(\Delta + \hat{V}_m)$ consists of bands, we may choose an open interval $J$ such that \eqref{eq:propJ} holds. \end{proof} We denote by $\Sigma_k$ the collection of open intervals $J$ obtained from the previous lemma for all possible choices of $I$ and $j$. It is clear that property (v) holds. \begin{proof}[Proof of Property (iv)] Given any $E \in [-4,4]$, we may find an $I \in \widetilde{\Sigma}_{k-1}$ such that $$ I \subseteq [E - \frac{1}{2^{k-1}}, E + \frac{1}{2^{k-1}}]. $$ Let $I = (E_-,E_+)$ and $\hat{E} =\frac{E_- + E_+}{2}$. Since $\hat{E} \in I$ and $$ [E - \frac{1}{2^{k-1}}, E + \frac{1}{2^{k-1}}] = \bigcup_{j=-4}^{3} \left[E + \frac{j}{2^{k+1}}, E + \frac{j+1}{2^{k+1}}\right], $$ we can find $-4 \leq j \leq 3$ such that $$ \hat{E} \in \left[E - \frac{j}{2^{k+1}}, E - \frac{j+1}{2^{k+1}}\right] $$ By the construction of $\Sigma_{k}$ in Lemma~\ref{lem:subinterval}, there exists $J \in \Sigma_{k}$ such that $J \subseteq I + \frac{j}{2^{k+1}}$. By $|I| \leq \frac{1}{2^{k+1}}$, we obtain $$ J \subseteq I + \frac{j}{2^{k+1}} \subseteq \left[\hat{E} + \frac{j-1}{2^{k+1}}, \hat{E} + \frac{j+1}{2^{k+1}}\right] \subseteq \left[E - \frac{1}{2^{k}}, E + \frac{1}{2^k}\right], $$ which finishes the proof. \end{proof} \subsection{Ensuring (iii)}\label{sub:ensuringiii} \begin{lemma}\label{lem:transmatsmall} Let $V$ be a $p$ periodic potential. For any $\mu > 0$, there exists $M \geq 1$ such that for $E \in \sigma(\Delta + V)$ and $N \geq M$, we have \begin{equation} \frac{1}{N} \log\|A_{N}(E, V)\| \leq \mu. \end{equation} \end{lemma} \begin{proof} Introduce the continuous functions $$ f_i(E) = \frac{1}{2^i} \log \| A_{2^i} (E,V)\|. $$ One can easily check that $f_{i+1}(E) \leq f_i(E)$ and that for $E \in \sigma(\Delta + V)$, we have $\lim_{i \to \infty} f_i(E) = L(E,V) = 0$. Hence, we obtain by Dini's therorem that there exists $M_1 \geq 1$ such that for $i \geq M_1$ and $E \in \sigma(\Delta +V)$ $$ \frac{1}{2^i} \log \| A_{2^i} (E,V)\| \leq \frac{\mu}{2}. $$ Let $N = q 2^{M_1} + r$, where $q \geq 1$ and $0 \leq r \leq 2^{M_1} -1$. For $E \in \sigma(\Delta +V)$, we get \begin{align*} \frac{1}{N} \log \|A_{N} (E,V)\| & \leq \frac{1}{q 2^{M_1} + r} \left ( \log \| A_{q 2^{M_1}} (E,V)\| + \log\| A_{r} (E,V)\|\right) \\ & \leq \frac{\mu}{2} + \frac{\log\| A_{r} (E,V)\|}{q 2^{M_1} + r} . \end{align*} Choose $M_2 \geq 1$ so large that for $E \in \sigma(\Delta +V)$ and $0 \leq r \leq 2^{M_1} - 1$ $$ \frac{\log\parallel A_{r} (E,V)\parallel }{M_2 2^{M_1} + r} \leq \frac{\mu}{2}. $$ The claim follows by taking $q \geq M_2$ or equivalently $N \geq M = M_2 \cdot 2^{M_1}$. \end{proof} By this lemma, we can ensure (iii) for $l = k$ as long as $m$ large enough. Let $1 \leq l \leq k - 1$. By assumption, $$ \frac{1}{p_{k-1}} \log\|A_{p_{k-1}}(E, V^{k-1})\| \leq \sum_{s = l+1}^{k} \frac{1}{2^s}. $$ By submultiplicativity, we get $$ \frac{1}{\tilde{p}_{k-1}} \log\|A_{\tilde{p}_{k-1}}(E, V^{k-1})\| \leq \sum_{s = l+1}^{k} \frac{1}{2^s}. $$ We recall that $\hat{V}_m$ is a $\hat{p}_m = (m + 8)\tilde{p}_{k-1}$ periodic potential, such that $\hat{V}_m(n) = V^{k-1}(n)$ for $0 \leq m \tilde{p}_{k-1}-1$. Thus, we have \begin{equation} A_{\hat{p}_m}(E, \hat{V}_m) = (A_{\tilde{p}_{k-1}}(E, V^{k-1}))^m \cdot \tilde{A}, \end{equation} where $\tilde{A}$ is some fixed matrix, whose norm is independent of $m$. Hence, we obtain $$ \frac{1}{\hat{p}_m} \log\|A_{\hat{p}_m}(E, \hat{V}_m)\| \leq \sum_{s = l+1}^{k} \frac{1}{2^s} + \frac{1}{\hat{p}_m} \log\|\tilde{A}\|. $$ The claim now follows by choosing $m$ large enough. \subsection{Construction of the initial potential.} Last, we construct an initial potential $V^0$. Define for $L \geq 1$, the potential $\tilde{V}_L$ by \begin{equation} \tilde{V}_L (n) = \begin{cases} 2, & 0 \leq n \leq L-1 ; \\ -2, & L \leq n \leq 2L-1 \end{cases} \end{equation} and continue it to be $2L$ periodic. For $k \in [0,\pi]$ the function $\varphi(n) = \mathrm{e}^{i kn}$ solves $\Delta \varphi = 2 \cos(k) \varphi$. Set $E = 2 \cos(k)$ and \begin{equation} \psi(n) = \begin{cases} \varphi(n), & 0 \leq n \leq L-1 ; \\ 0, & \text{otherwise}. \end{cases} \end{equation} Clearly, $\| \psi \| = \sqrt{L}$, and $$ \| (\Delta + V^0 - E-2) \psi \| \leq 2. $$ This implies that $\mathrm{dist}(E+2,\sigma(\Delta+V^0)) \leq \frac{2}{\sqrt{L}}.$ When $L > 4$, we have $$ \mathrm{dist}(E+2,\sigma(\Delta+V^0)) < 1. $$ So for any $E_0 \in [0,4]$ we can find an interval $I \subset \sigma(\Delta + V^0)$ such that $I \subseteq [E_0 - 1, E_0 + 1].$ Similarly, we conclude that for any $E_0 \in [-4,0]$ there is an interval $I \subset \sigma(\Delta + V^0)$ such that $I \subseteq [E_0 - 1, E_0 + 1].$ Thus, for $V^0$ we can find a collection $\Sigma_0$ of intervals $I \subset \sigma(\Delta + V^0)$ which is $1$-dense in $[-4,4]$. (iv) follows. By Lemma~\ref{lem:transmatsmall}, (iii) follows since we can treat $V_0$ as a $p_0$ periodic where $p_0 = 2 L m$, and $m \in {\mathbb Z}^+$ is large. There is nothing to check for (i), (ii), and (v). \section{Proof of Theorem~\ref{thm:main2}} The first step in the proof of Theorem~\ref{thm:main2} is to prove the following proposition, which we postpone to a later subsection. \begin{proposition}\label{prop:perturb} Let $V_0$ be a $p_0$ periodic potential such that \begin{equation} E_0 = \inf(\sigma(\Delta + V_0)) > 0. \end{equation} Introduce $\gamma = \frac{1}{2} L(0, V^0)$. There exists a limit-periodic potential $V$ such that $\|V - V_0\| \leq 5 E_0$, $\inf(\sigma(\Delta + V)) = 0$, and \begin{equation} \limsup_{E \to 0} L(E) \geq \gamma > \liminf_{E \to 0} L(E) = 0. \end{equation} In particular, the Lyapunov exponent $L(E, V)$ is discontinuous at $0$. \end{proposition} We are now ready for \begin{proof}[Proof of Theorem~\ref{thm:main2}] First note that the periodic potentials are dense in the space $\mathscr{L}$ of all limit periodic potentials. Let $V_0$ be any periodic potential and $\varepsilon > 0$. Introduce the potential $$ \tilde{V}_0 = V_0 - \inf(\Delta + V_0) + \frac{\varepsilon}{5}. $$ One can check that $\tilde{V}_0$ satisfies the assumptions of the previous proposition, and we thus obtain a potential $\tilde{V}$ such that the Lyapunov exponent of $\tilde{V}$ is discontinuous and $\|\tilde{V}_0 - \tilde{V}\| \leq \varepsilon$. We now see that $V = \tilde{V} + \inf(\Delta + V_0) - \frac{\varepsilon}{5}$ satisfies that its Lyapunov exponent is discontinuous and $\|V - V_0\| \leq \varepsilon$. Hence, the potentials with discontinuous Lyapunov exponent are dense. \end{proof} \subsection{Proof of Proposition~\ref{prop:perturb}} In the following subsection, we will construct a sequence of potentials $V^k$ with the following properties. \begin{enumerate} \item $V^k$ is $p_k$ periodic and satisfies \begin{equation} \|V^{k} - V^{k-1}\| \leq \frac{2}{5} E_{k-1}, \end{equation} and $V^k(n) = V^{k-1}(n)$ for $0 \leq n \leq p_{k - 1}$. \item The bottom of the spectrum $E_k = \inf \sigma(\Delta + V^k)$ satisfies \begin{equation} \left(\frac{3}{5}\right)^k E_0 \leq E_k \leq \left(\frac{4}{5}\right)^k E_0. \end{equation} \item The Lyapunov exponent at energy $0$ satisfies \begin{equation} L(0, V^k) \geq \left(2 - \sum_{s = 1}^{k} \frac{1}{2^s} \right) \gamma. \end{equation} \item For $1 \leq l \leq k$ and every $E \in \sigma(\Delta + V^l)$ \begin{equation} \frac{1}{p_k} \log\|A_{p_k}(E, V^k)\| \leq \sum_{s = l+1}^{k+1} \frac{1}{2^s}. \end{equation} \end{enumerate} From properties (i) and (ii), we see that \begin{equation} \|V^{k} - V^{k-1}\| \leq \left(\frac{4}{5}\right)^k E_0. \end{equation} Hence the limit \begin{equation} V = \lim_{k \to \infty} V^k, \end{equation} exists in $\ell^\infty({\mathbb Z})$ and $\|V - V^0\| \leq 5 E_0$. \begin{lemma} We have that \begin{equation} L(0) \geq \gamma. \end{equation} \end{lemma} \begin{proof} By (ii), we have that $\inf(\sigma(\Delta + V^k)) > 0$. Hence by Thouless' formula \eqref{eq:thouless} $$ L(E, V^k) = \int \log|t - E| dN^k(t), $$ we have that $E \mapsto L(E, V^k)$ is decreasing in $E \leq 0$. This and property (iii) imply for $E < 0$ and $k \geq 1$ that $$ L(E, V^k) \geq \gamma. $$ Since $N^k \to N$ weakly and $\log|t - E|$ is a bounded and continuous function for $E < 0$, we also obtain $$ L(E) \geq \gamma $$ for $E < 0$. Thouless formula even implies that $$ L(0) = \lim_{E\uparrow 0} L(E). $$ This implies the claim. \end{proof} Similarly as for \eqref{eq:LEleq2l} in the previous section, we have that properties (i) and (iv) imply that \begin{equation} L(E) \leq \frac{1}{2^{k}} \end{equation} for almost every $E \in \sigma(\Delta + V^k)$. Hence, we have that \begin{equation} \limsup_{E \to 0} L(E) > \liminf_{E \to 0} L(E) = 0. \end{equation} This finishes the proof of Proposition~\ref{prop:perturb}. \subsection{Construction of the sequence of potentials} In order to construct the sequence of potentials $V^k$, we will prove the following lemmas. These imply the existence of $V^k$ given $V^{k-1}$ by applying them to $V = V^{k-1}$. \begin{lemma} \label{lem:lowerspec} Let $V$ be a $p$ periodic potential and $E_0 = \inf\sigma(\Delta + V).$ Let $\tilde{p} = m_0 p$ for sufficiently large $m_0$. Define for $1 \leq l \leq h$ the intervals $I_l$ by \begin{equation} I_l = [m \tilde{p} + (l-1) \tilde{p}, m \tilde{p} + l \tilde{p} - 1]. \end{equation} Define the $\hat{p} = (m + h ) \tilde{p}$ periodic potential $\hat{V}_{m,h}$ by \begin{equation}\label{eq:defhatV} \hat{V}_{m,h} (n) = \begin{cases} V(n) & [0, m \tilde{p} -1];\\ V(n) - \frac{2 E_0}{5} & n \in I_l,\quad 1 \leq l \leq h. \end{cases} \end{equation} Then we have \begin{equation} \frac{3E_0}{5} \leq \inf\sigma(\Delta + \hat{V}_{m,h}) \leq \frac{4 E_0}{5}. \end{equation} \end{lemma} \begin{proof} As in the proof of Lemma~\ref{lem:subinterval}, let $E_{-} = E_0$ and $\delta = \frac{E_0}{5}$. Pick $\hat{E}$ in the first band of $\sigma(\Delta+V)$ such that $\hat{E} - E_{0} < E_{0}/10$. Also, $m_0$ will be picked sufficiently large so that $\delta \sqrt{m_0} > 4$. Then, we will get \begin{equation} \dist(\hat{E}- \frac{2E_0}{5}, \sigma(\Delta + \hat{V}_{m,h})) \leq \frac{\delta}{2} = \frac{E_0}{10}. \end{equation} The lemma follows. \end{proof} We can view the $\hat{p}$ period of $\hat{V}_{m,h}$ as a concatenation of two parts: \begin{itemize} \item $m$ pieces of the $\tilde{p}$ period of $V$. \item $h$ pieces of the $\tilde{p}$ period of $\tilde{V}$, where $\tilde{V} = V - \frac{2E_0}{5}.$ \end{itemize} Denote $\tilde{A} = A_{\tilde{p}}(E,\tilde{V})$, then we have \begin{equation} A_{\hat{p}}(E,\hat{V}_{m,h}) = (\tilde{A})^h \cdot (A_{\tilde{p}}(E, V))^{m}. \end{equation} Let $E \notin \inf(\sigma(\Delta + V))$. Then we can chose two normalized vectors $v,u \in C^2$, such that \begin{equation} A_{\tilde{p}}(E, V) v = \mathrm{e}^{\tilde{p} L(E,V)} v,\qquad A_{\tilde{p}}(E, V) u = \mathrm{e}^{-\tilde{p} L(E,V)} u. \end{equation} Denote by $v^{\perp}= a v + b u$ a vector orthonormal to $v$. We will first show \begin{lemma} There exists $h \in \{1,2\}$ such that \begin{equation} \spr{v}{\tilde{A}^h v} + a \spr{v^{\perp}}{\tilde{A}^h v} \neq 0. \end{equation} \end{lemma} \begin{proof} From the Cayley--Hamilton theorem, we have that $\tilde{A}^2 - \mathrm{tr}(\tilde{A}) \tilde{A} + I = 0$. Taking $\spr{v}{. v}$ and $\spr{v^{\perp}}{ . v}$, we obtain \begin{align*} \spr{v}{\tilde{A}^2v} - \mathrm{tr}(\tilde{A}) \spr{v}{ \tilde{A} v} + 1 &= 0, \\ \spr{v^{\perp}}{\tilde{A}^2v} - \mathrm{tr}(\tilde{A}) \spr{v^{\perp}}{ \tilde{A} v} &= 0. \end{align*} Multiplying the second equation by $a$ and adding the two together, we obtain $$ \Big(\spr{v}{\tilde{A}^2v} + a \spr{v^{\perp}}{\tilde{A}^2v}\Big) - \mathrm{tr}(\tilde{A}) \Big(\spr{v}{ \tilde{A} v} + a \spr{v^{\perp}}{ \tilde{A} v}\Big) = -1. $$ Hence, the claim follows. \end{proof} We now come to \begin{lemma} \label{lem:lowle} Let $E \notin \sigma(\Delta + V)$. Then there exists $h \in \{1,2\}$ such that \begin{equation} L(E, \hat{V}_{m,h}) \to L(E, V) \end{equation} as $m \to \infty$. \end{lemma} \begin{proof} Let $h$ be as in the previous lemma. The lower bound can be obtained by a similar argument as in Subsection~\ref{sub:ensuringiii}. By (7.10) in \cite{tjac} and a computation, we have \begin{align*} L(E, \hat{V}_{m,h}) & = \frac{1}{\hat{p}} \log \left (\left |\frac{\mathrm{tr}(A_{\hat{p}}(E,\hat{V}_{m,h}))}{2} \right | + \sqrt {\left (\frac{\mathrm{tr}(A_{\hat{p}}(E,\hat{V}_{m,h}))}{2}\right )^2 - 1} \right )\\ & \geq \frac{1}{\hat{p}} \log{\left | \mathrm{tr}(\tilde{A}^h (A_{\tilde{p}}(E,V))^m ) \right |} - \frac{\log(2)}{\hat{p}}. \end{align*} Let us now evaluate $\mathrm{tr}(\tilde{A}^h (A_{\tilde{p}}(E,V))^m )$. We have \begin{align*} \mathrm{tr}(\tilde{A}^h (A_{\tilde{p}}(E,V))^m ) &=\spr{v}{\tilde{A}^h (A_{\tilde{p}}(E,V))^m v} + \spr{v^{\perp}}{\tilde{A}^h (A_{\tilde{p}}(E,V))^m v^{\perp}} \\ &= \mathrm{e}^{m\tilde{p} L(E, V)}\spr{v}{\tilde{A}^h v} + a \mathrm{e}^{m\tilde{p} L(E, V)} \spr{v^{\perp}}{\tilde{A}^h v}\\ &\qquad +b \mathrm{e}^{- m\tilde{p} L(E, V)}\spr{v^{\perp}}{\tilde{A}^h u} \\ &= \mathrm{e}^{m\tilde{p} L(E, V)} \left(\spr{v}{\tilde{A}^h v} + a \spr{v^{\perp}}{\tilde{A}^h v}\right) + o(1). \end{align*} Combining this with the previous formula, we obtain that $$ L(E, \hat{V}_{m,h}) \geq L(E,V) + o(1). $$ Hence, we see that the claim holds, if we choose $m$ large enough. \end{proof} In order to show the existence of $V^k$ such that (i) to (iv) hold. Use Lemma~\ref{lem:lowerspec} to find a sequence of potentials $\hat{V}_{m,h}$ such that properties (i) and (ii) hold. The previous lemma implies the existence of $h \in \{1,2\}$ such that property (iii) for $m \geq 1$ large enough. Finally, by arguments similar to the ones in Subsection~\ref{sub:ensuringiii}, we can show that property (iv) holds for $m \geq 1$ large enough. This finishes the proof of the existence of the sequence $V^k$ and so also of Proposition~\ref{prop:perturb}. \section*{Acknowledgments} We thank Artur Avila and David Damanik for useful discussions and Svetlana Jitomirskaya for informing us of \cite{thuv}.

https://arxiv.org/abs/1903.12153

On the optimal map in the 2-dimensional random matching problem

We show that, on a $2$-dimensional compact manifold, the optimal transport map in the semi-discrete random matching problem is well-approximated in the $L^2$-norm by identity plus the gradient of the solution to the Poisson problem $-\Delta f^{n,t} = \mu^{n,t}-1$, where $\mu^{n,t}$ is an appropriate regularization of the empirical measure associated to the random points. This shows that the ansatz of Caracciolo et al. (Scaling hypothesis for the Euclidean bipartite matching problem) is strong enough to capture the behavior of the optimal map in addition to the value of the optimal matching cost.As part of our strategy, we prove a new stability result for the optimal transport map on a compact manifold.

\section{Introduction}\label{sec:intro} Let $(X_1, X_2,\dots, X_n)$ be $n$ independent random points uniformly distributed on the square $\cc01^2$. The semi-discrete random matching problem concerns the study of the properties of the optimal coupling (with respect to a certain cost) of these $n$ points with the Lebesgue measure $\restricts{\Leb^2}{\cc01^2}$. More precisely, denoting $\mu^n \defeq \frac1n \sum_{i=1}^n\delta_{X_i}$ the empirical measure and $\m\defeq\restricts{\Leb^2}{\cc01^2}$, we want to investigate the optimal transport from $\m$ to $\mu^n$. The ultimate goal is understanding both the distribution of the random variable associated to the optimal transport cost and the properties of the (random) optimal map. In the present paper we will show that the optimal transport map can be well-approximated by the identity plus the gradient of the solution of a Poisson problem. In the large literature devoted to the matching problem, we believe that (except for the 1-dimensional case) this is one of the few results describing the behavior of the optimal map, and not only of the transport cost, see also \cite{Goldman2018} in connection with the behavior of the optimal transport map in the Lebesgue-to-Poisson problem on large scales. Before going on, let us briefly recall the definitions of optimal transport and Wasserstein distance. We suggest the monographs \cite{Villani08,Santambrogio15} for an introduction to the topic. \begin{definition}[Wasserstein distance] Let $(X, d)$ be a compact metric space and let $\mu,\nu\in\prob(X)$ be probability measures. Given $p\in\co{1}{\infty}$, we define the $p$-Wasserstein distance between $\mu$ and $\nu$ as \begin{equation}\label{eq:defWp} W_p^p(\mu, \nu) \defeq \inf_{\gamma\in\Gamma(\mu, \nu)} \int_{X\times X} d^p(x, y) \de\gamma(x, y) \comma \end{equation} where $\Gamma(\mu, \nu)$ is the set of all $\gamma\in\prob(X\times X)$ such that the projections $\pi_i$, $i=1,2$, on the two factors are $\mu$ and $\nu$, that is $(\pi_1)_\#\gamma = \mu$ and $(\pi_2)_\#\gamma=\nu$. \end{definition} \begin{remark} The infimum in the previous definition is always attained (\cite[Theorem 1.4]{Santambrogio15}). Moreover, if $(X, d)$ is a Riemannian manifold and $\mu\ll\m$, where $\m$ is the volume measure of the manifold, the Wasserstein distance is realized by a map (\cite{McCann01}). Namely, the infimum \cref{eq:defWp} is attained and the unique minimizer is induced by a Borel map $T:M\to M$, so that $T_\#\mu = \nu$ and \begin{equation*} W_p^p(\mu, \nu) = \int_M d^p(x, T(x)) \de\mu(x) \fullstop \end{equation*} \end{remark} Even though the square is a fundamental example, the random matching problem makes perfect sense even in more general spaces (changing the reference measure $\m$ accordingly). Historically, in the combinatorial literature\footnote{In the combinatorial literature the problem considered was the bipartite matching problem, in which two independent random point clouds have to be matched. The semi-discrete matching and the bipartite matching are tightly linked and, given that we will consider only the former, we are going to talk about the combinatorial literature as if it were considering the semi-discrete matching.}, the most common ambient space was $\cc01^d$ for some $d\ge 1$ and the aspect of the problem that attracted more attention was estimating the expected value of the $W_1$ cost. In the papers \cite{Ajtai1984,Talagrand1992,Dobric1995,Ledoux2017} (and possibly in other ones) the problem was solved in all dimensions and for all $1\le p < \infty$, obtaining the growth estimates\footnote{The notation $f(n)\approx g(n)$ means that there exists a positive constant $C>0$ such that $C^{-1}g(n) \le f(n)\le C g(n)$ for every $n$.} \begin{equation*} \E{W_p^p(\m, \mu^n)} \approx \begin{cases} n^{-\frac p2} &\text{ if $d=1$,}\\ \left(\frac{\log(n)}n\right)^{\frac p2} &\text{ if $d=2$,}\\ n^{-\frac pd} &\text{ if $d\ge 3$.} \end{cases} \end{equation*} As might be clear from the presence of a logarithm, the matching problem exhibits some unexpected behavior in dimension $2$. See the introductions of \cite{ambrosio-glaudo2018,Ledoux2017} or \cite[Chapter 4, 14, 15]{Talagrand14} for a more in-depth description of the history of the problem. Nowadays the topic is active again (\cite{holden2018,Talagrand2018,Goldman2018,Ledoux2017,ambrosio-glaudo2018,Ledoux18,Ledoux19}), also as a consequence of \cite{Ambrosio-Stra-Trevisan2018}, in which the authors, following an ansatz suggested in \cite{CaraccioloEtAl2014}, manage to obtain the leading term of the asymptotic expansion of the expected matching cost in dimension $2$ with respect to the quadratic distance\footnote{The notation $f(n)\sim g(n)$ means that $\frac{f(n)}{g(n)}\to 1$ when $n\to\infty$.}: \begin{equation}\label{eq:limit_value} \E{W_2^2(\m, \mu^n)} \sim \frac{\log(n)}{4\pi n} \fullstop \end{equation} The approach is far from being combinatorial, indeed it relies on a first-order approximation of the Wasserstein distance with the $H^{-1}$ negative Sobolev norm. Their proof works on any closed compact $2$-dimensional manifold. Given that we will build upon it, let us give a brief sketch of the approach. What we are going to describe is simpler than the original approach of \cite{Ambrosio-Stra-Trevisan2018} and can be found in full details in \cite{ambrosio-glaudo2018}. For simplicity we will assume to work on the square. Let $T^n$ be the optimal map from $\m$ to $\mu^n$, whose existence is ensured by Brenier's Theorem (see \cite{Brenier91}). Still by Brenier's Theorem, we know that $T^n = \id + \nabla \tilde f^n$, where $\id$ is the identity map and $\tilde f^n:\cc01^2\to M$ is a convex function. With high probability $\mu^n$ is well-spread on the square, thus we expect $\nabla \tilde f^n$ to be \emph{very small}. We know $(T^n)_\#\m=\mu^n$ and we would like to apply the change of variable formula to deduce something on the Hessian of $\tilde f^n$. The issue is that the singularity of $\mu^n$ prevents a direct application of the change of variable formula. Anyhow, proceeding formally we obtain $\det(\id+\nabla^2 \tilde f^n)^{-1}=\mu^n$. Going on with the formal computation, if we consider only the first order term of the left hand side, the previous identity simplifies to \begin{equation*} -\lapl \tilde f^n \approx \mu^n-1 \fullstop \end{equation*} Somewhat unexpectedly, this last equation makes perfect sense. Therefore we might claim that if we define $f^n:\cc01^2\to\R$ as the solution of $-\lapl f^n=\mu^n-1$ (with null Neumann boundary condition), then $T^n$ is well-approximated by $\id+\nabla f^n$ and furthermore the transport cost is well-approximated by $\int \abs{\nabla f^n}^2\de\m$. This conjecture is appealing, but false, if taken literally. Indeed, it is very easy to check that the integral $\int \abs{\nabla f^n}^2\de\m$ diverges. The ingredient that fixes this issue is a regularization argument. More precisely, let $\mu^{n,t}\defeq P_t^*\mu^n$ be the evolution at a certain small time $t>0$ of the empirical measure through the heat semigroup (see \cite[Chapter 6]{Chavel84}). If we repeat the ansatz with $\mu^n$ replaced by $\mu^{n,t}$ we obtain a function $f^{n,t}:\cc01^2\to\R$ that solves \begin{equation*} -\lapl f^{n,t} = \mu^{n,t}-1 \end{equation*} with null Neumann boundary conditions. Let us remark that in fact $f^{n,t}=P_t f^n$. Once again, we can hope that $\id+\nabla f^{n,t}$ approximates very well $T^n$ and furthermore that the transport cost from $\m$ to $\mu^n$ is well-approximated by $\int \abs{\nabla f^{n,t}}^2\de\m$. This time the predictions are sound. Choosing carefully the time $t=t(n)$, we can show that, with high probability, the map $\id+\nabla f^{n,t}$ is optimal from $\m$ to $\left(\id+\nabla f^{n,t}\right)_\#\m$ and the Dirichlet energy of $f^{n,t}$ approximate very well $W_2^2(\m, \mu^n)$. Only one part of the conjecture is left unproven by \cite{ambrosio-glaudo2018}: is it true that $\id+\nabla f^{n,t}$ approximates, in some adequate sense, the optimal map $T^n$? The goal of the present paper is to answer positively this question. We are going to prove the following. \begin{theorem}\label{thm:main_theorem} Let $(M,\metric)$ be a $2$-dimensional closed compact Riemannian manifold (or the square $\cc01^2$) whose volume measure $\m$ is a probability. We will denote with $d:M\times M\to\co0\infty$ the Riemannian distance on $M$. Given $n\in\N$, let $X_1, X_2,\dots, X_n$ be $n$ independent random points $\m$-uniformly distributed on $M$. Let us denote $\mu^n\defeq \frac1n \sum_i \delta_{X_i}$ the empirical measure associated to the random point cloud and let $T^n$ be the optimal transport map from $\m$ to $\mu^n$. For a fixed time $t>0$, let $\mu^{n,t}\defeq P_t^*\mu^n\in\prob(M)$ and let $f^{n,t}:M\to\R$ be the unique null-mean solution\footnote{If $M=\cc01^2$ we ask also that $f$ satisfies the null Neumann boundary conditions.} of the Poisson problem $-\lapl f^{n,t}=\mu^{n,t}-1$. If we set $t=t(n)=\frac{\log(n)^4}n$, on average $T^n$ is very close to $\exp(\nabla f^{n,t})$ in the $L^2$-norm, that is \begin{equation}\label{eq:main-quantitative} \frac{\E{\int_M d^2(T^n, \exp(\nabla f^{n,t}))\de\m}}{\frac{\log(n)}{n}} \ll \sqrt{\frac{\log \left(\log (n)\right)}{ \log(n) }} \fullstop \end{equation} In particular, \begin{equation*} \lim_{n\to\infty}\frac{\E{\int_M d^2(T^n, \exp(\nabla f^{n,t}))\de\m}} {\E{\int_M d^2(T^n, \id)\de\m}} = 0 \fullstop \end{equation*} \end{theorem} \begin{remark} To handle the case of the square $M=\cc01^2$ some care is required. Indeed the presence of boundary makes things more delicate. This is the reason why only the square is considered in the theorem and not any $2$-dimensional compact manifold with boundary. See \cite[Subsection 2.1 and Remark 3.10]{ambrosio-glaudo2018} for some further details on this matter. \end{remark} \begin{remark}\label{rem:distance-tangent} By McCann’s Theorem \cite{McCann01} we can write $T^n = \exp(\nabla f^n)$, hence a natural question is if \cref{eq:main-quantitative} holds with $|\nabla (f^n - f^{n,t})|$ in place of $d(T^n, \exp(\nabla f^{n,t}))$. Using the fact that the exponential map restricted to a sufficiently small neighbourhood of the null vector field is a global diffeomorphism with its image, it would be sufficient to show that, for every $\varepsilon>0$, $\P{\|d(T^{n},\id)\|_\infty > \varepsilon} \ll \log(n)/n$, as $n \to \infty$. We will prove this estimate in \cref{prop:linf_is_small}, that provides the desired approximation at the level of the gradients \begin{equation}\label{eq:main_gradients} \lim_{n\to\infty} \frac{\E{\norm{\nabla f^n-\nabla f^{n,t}}_{L^2(M)}^2}} {\E{\norm{\nabla f^n}_{L^2(M)}^2}} = 0\fullstop \end{equation} \end{remark} The strategy of the proof is to show that the information that we already have on $\exp(\nabla f^{n,t})$ (namely that it is an optimal map between $\m$ and some measure $\hat\mu^{n,t}$ that is very close to $\mu^{n,t}$) is enough to deduce that it must be near to the optimal map $T^n$. As part of the strategy of proof, we obtain, in \cref{sec:stability}, a new stability result for the optimal transport map on a general compact Riemannian manifold (not only of dimension $2$). This is the natural generalization to Riemannian manifolds of \cite{gigli2011}. The said stability result follows rather easily from the study of the short-time behavior of the Hopf-Lax semigroup we perform in \cref{sec:hopflax}. The Hopf-Lax semigroup comes up in our investigation as, when $t=1$, it becomes the operator of $c$-conjugation and thus produce the second Kantorovich potential once the first is known (see \cite[Section 1.2]{Santambrogio15} for the theory of Kantorovich potentials and $c$-conjugation). The main theorem is established in \cref{sec:random_matching}. \vspace{2mm} \noindent {\textit{Acknowledgments. } F. Glaudo has received funding from the European Research Council under the Grant Agreement No 721675. L. Ambrosio acknowledges the support of the MIUR PRIN 2015 project. \subsection{Notation for constants} We will use the letters $c$ and $C$ to denote constants, whose dependencies are denoted by $c=c(A, B,\dots)$. The value of such constants can change from one time to the other. Moreover we will frequently use the notation $A\lesssim B$ to hide a constant that depends only on the ambient manifold $M$. This expression means that there exists a constant $C=C(M)$ such that $A\le C\cdot B$. \section{Short-time behavior of the Hopf-Lax semigroup with datum in \texorpdfstring{$C^{1,1}$}{C(1,1)}} \label{sec:hopflax} Let us begin recalling the definition of the Hopf-Lax semigroup (also called Hamilton-Jacobi semigroup). \begin{definition}[Hopf-Lax semigroup] Let $(X, d)$ be a compact length space\footnote{A metric space is a length space if the distance between any two points is the infimum of the length of the curves between the two points. Let us remark that for the definition we need neither the compactness nor the length property of $X$, but without these assumptions many of the properties of the Hopf-Lax semigroup fail (first of all the fact that it is a semigroup).}. For any function $f\in C(X)$ and any $t\geq 0$, let $Q_t f:X\to\R$ be defined by \begin{equation*} Q_t f(y) = \min_{x\in X} \frac1{2t} d^2(x,y) + f(x) \quad (t>0),\qquad Q_0f=f\fullstop \end{equation*} \end{definition} Without additional assumptions on $X$ or $f$ it is already possible to deduce many properties of the Hopf-Lax semigroup. Let us give a very short summary of the most important ones. \begin{itemize} \item When $t\to 0$ the functions $Q_t f$ converge uniformly to $f$. \item The Hopf-Lax semigroup is indeed a semigroup, that is $Q_{s+t}f = Q_sQ_t f$ for any $s,\, t \geq 0$. \item In a \emph{suitable weak sense},the Hamilton-Jacobi equation \begin{equation*} \frac{\de}{\de t} Q_t f + \frac12\abs{\nabla Q_t f}^2 = 0 \end{equation*} holds. Let us emphasize that the mentioned equation does not make sense if we don't give an appropriate definition of norm of the gradient as we are working in a metric setting. \end{itemize} See \cite{Lott-Villani07}, in particular Theorem~2.5, for a detailed proof of the mentioned properties. There is a vast literature investigating the regularity properties of the Hopf-Lax semigroup and its connection with the Hamilton-Jacobi equation, in particular that it is the unique solution in the viscosity sense (see for instance \cite{lions1982,benton1977,bardi2008}). Nonetheless we could not find a complete reference for the short-time behavior of the Hopf-Lax semigroup on a Riemannian manifold (as the majority of the results are stated on the Euclidean space) with a relatively regular initial datum (namely $C^{1,1}$). This is exactly the topic of this section. What we are going to show, apart from \cref{it:hopflax_convexity}, is not new. For instance, in \cite[Section 5]{fathi2003}, the author proves the validity of the method of characteristics in a way very similar to ours. In that paper more general Lagrangians are considered and as a consequence the proofs are more involved and require much more geometric tools and notation. For us, the ambient space is a compact Riemannian manifold $(M, \metric)$ and the function $f\in C^{1,1}(M)$ is differentiable with Lipschitz continuous gradient. Moreover, either $M$ is closed or it is the square $\cc01^2$. For a general manifold with boundary the results are false, the square is special because its boundary is piecewise geodesic. Handling all manifolds with totally geodesic boundary would be possible, but would require some additional care. In order to simplify the exposition we decided to state the results only for the square. Throughout this section we will often use implicitly that a Lipschitz continuous function is differentiable almost everywhere (see \cite[Theorem 3.2]{Evans-Gariepy}). We will show that, up to a small time that depends on the $C^{1,1}$-norm of $f$, the Hopf-Lax semigroup is \emph{as good as one might hope}. We will describe explicitly the minimizer $x=x_t(y)$ of the variational problem that defines $Q_t f(y)$ deducing some \emph{explicit} formulas for $Q_tf$ and its gradient and we will show that $Q_t f$ solves the Hamilton-Jacobi equation in the classical sense. Finally we will be able to control the $C^{1,1}$-norm of $Q_t f$ and the $C^{0,1}$-norm of $Q_t f-f$. How can we achieve these results for short times when $f\in C^{1,1}$? The main ingredient is the possibility to identify the minimizer $x=x_t(y)$ in the definition of $Q_t f(y)$. Given $x\in M$, let $\gamma:\co{0}{\infty}\to M$ be the unique geodesic with $\gamma(0) = 0$ and $\gamma'(0)=\nabla f(x)$. If $y=\gamma(t)$, then the minimizer in the definition of $Q_t f(y)$ is exactly $x$. This approach is exactly the method of characteristics when applied on a Riemannian manifold (\emph{straight lines on a manifold are geodesics}). Let us begin with a technical lemma. \begin{lemma}\label{lem:exp_is_diffeo} Let $(M, \metric)$ be a closed compact Riemannian manifold (or the square $\cc01^2$). There exists a constant $c=c(M)$ such that the following statement holds. Let $X\in\chi(M)$ be a Lipschitz continuous vector field\footnote{If $M=\cc01^2$ we ask also that $X$ is tangent to the boundary.} with $\norm{X}_\infty\le c$ and $\norm{\nabla X}_\infty\le c$ and, for any $0\le t\le 1$, let $\varphi_t:M\to M$ be the map defined as $\varphi_t(x) \defeq \exp(tX(x))$, where $\exp:TM\to M$ denotes the exponential map. For any $0\le t\le 1$, the map $\varphi_t$ is a homeomorphism such that $\Lip(\varphi_t)$, $\Lip(\varphi_t^{-1}) \le 2$ and the vector field $X_t\in\chi(M)$ defined as \begin{equation*} X_t \defeq \frac{\partial \varphi_s}{\partial s}\Big|_{s=t} \end{equation*} is Lipschitz continuous with $\norm{\nabla X_t}_{\infty}\lesssim\norm{\nabla X}_{\infty}$. \end{lemma} \begin{proof} We will give only a sketch of the proof of the first part of the statement as the argument is well-known. Let us begin by proving the result when $M$ is closed (in particular we exclude only $M=\cc01^2$). We can deduce the first part of the statement from the fact that $\varphi=\varphi_1$ is injective and locally (i.e. on sufficiently small balls) it is a bi-Lipschitz transformation with its image. Working in a suitably chosen finite atlas (whose existence follows from the compactness of $M$), the fact that $\varphi$ is a bi-Lipschitz diffeomorphism is a consequence of the following very well-known lemma about perturbations of the identity (see \cite[Theorem 9.24]{rudin1976} or \cite[Theorem 5.3]{fathi2003}). If $T:\Omega\subseteq \R^d\to\R^d$ is such that $T-\id$ is $L$-Lipschitz with $L<1$, then $T$ is locally invertible and $\Lip(T) \le 1+L,\ \Lip(T^{-1}) \le (1-L)^{-1}$. The global injectivity follows directly from the fact that it is locally bi-Lipschitz. Indeed if $\varphi(x_1)=\varphi(x_2)$ then $d(x_1, x_2) \le 2\norm{X}_\infty$ and therefore we can exploit the local injectivity of $\varphi$. When $M=\cc01^2$ we need only a simple additional remark. Given that $X$ is tangent to the boundary, the map $\varphi$ is a homeomorphism of the boundary. As a consequence of this fact, it is not difficult to prove (by injectivity) that the image of the interior of the square is mapped by $\varphi$ in itself. From here on we can simply mimic the proof described above for closed manifolds and achieve the result also for the case of the square. We move our attention to the second part of the statement. By a simple homogeneity argument, it is sufficient to prove that $\norm{\nabla X_t}_{\infty}\lesssim 1$. Once again we work in chart. Let $\Omega\subseteq\R^d$ be the domain of the chart. As usual, $X_t$ can be understood as a vector field on $\Omega$ and $\varphi_t$ as a map from $\Omega'\Subset\Omega$ into $\Omega$. Choosing the chart appropriately, we can assume that the Euclidean distance is bi-Lipschitz equivalent to the distance induced by the metric $\metric$. The Lipschitz continuity of $X_t$ with respect to the metric $\metric$ is equivalent to proving that, for any $x,y\in\Omega$, it holds \begin{equation*} \abs{X_t(x)-X_t(y)} \lesssim \abs{x-y} \comma \end{equation*} where all the absolute values are with respect to the standard Euclidean norm. Since $\varphi_t$ is surjective, it is sufficient to prove that, for any $x,y\in\Omega'$, it holds \begin{equation}\label{eq:exp_diffeotmp1} \abs{X_t(\varphi_t(x))-X_t(\varphi_t(y))} \lesssim \abs{\varphi_t(x)-\varphi_t(y)} \fullstop \end{equation} Given that $\varphi_t^{-1}$ is Lipschitz, we already know \begin{equation}\label{eq:exp_diffeotmp2} \abs{x-y} \lesssim \abs{\varphi_t(x)-\varphi_t(y)} \quad\text{and}\quad \abs{X(x)-X(y)} \lesssim \abs{\varphi_t(x)-\varphi_t(y)} \fullstop \end{equation} Let $\gamma_x:\cc01\to\Omega$ be the unique geodesic, with respect to $\metric$, such that $\gamma_x(0)=x$ and $\gamma_x'(0)=X(x)$. Let $\gamma_y:\cc01\to\Omega$ be defined analogously. By definition, it holds \begin{equation}\label{eq:exp_diffeotmp3} X_t(\varphi_t(x)) = \gamma_x'(t) \quad\text{and}\quad X_t(\varphi_t(y)) = \gamma_y'(t) \fullstop \end{equation} Taking into account \cref{eq:exp_diffeotmp1}, \cref{eq:exp_diffeotmp2} and \cref{eq:exp_diffeotmp3}, the Lipschitz continuity of $X_t$ would follow from the inequality \begin{equation}\label{eq:exp_diffeotmp4} \abs{\gamma_x'(t)-\gamma_y'(t)} \lesssim \abs{\gamma_x(0)-\gamma_y(0)} + \abs{\gamma_x'(0)-\gamma_y'(0)} \fullstop \end{equation} The curves $\gamma_x,\gamma_y$ are geodesics, hence the vectors $(\gamma_x, \gamma_x')$ and $(\gamma_y,\gamma_y')$ solve the same autonomous ordinary differential equation with different initial data. Hence \cref{eq:exp_diffeotmp4} follows from the well-known Lipschitz dependence of the solution from the initial data (see \cite[Theorem 2.6]{teschl2012}) and therefore the proof is concluded. \end{proof} We can now state and prove the main theorem of this section. The technically demanding part of these notes is entirely enclosed in the following theorem. \begin{theorem}\label{thm:hopflax_properties} Let $(M, \metric)$ be a closed compact Riemannian manifold (or the square $\cc01^2$). Let $f\in C^{1,1}(M)$ be a scalar function\footnote{If $M=\cc01^2$ we ask also that $f$ satisfies the null Neumann boundary conditions.} and, for any positive time $t>0$, let us define the map $\varphi_t:M\to M$ as $\varphi_t(x) \defeq \exp(t\nabla f(x))$. There exists a constant $c=c(M)$ such that the following properties hold for any time $0\le t\le c \left(\norm{\nabla f}_\infty + \norm{\nabla^2 f}_\infty\right)^{-1}$: \begin{enumerate}[ref={(\arabic*)}] \item \label{it:varphi_t_diffeo} The map $\varphi_t$ is a bi-Lipschitz homeomorphism such that $\Lip(\varphi_t),\, \Lip(\varphi_t^{-1}) \leq 2$. \item \label{it:hopflax_explicit} For any $y\in M$, it holds \begin{equation*} Q_t f(y) = \frac1{2t} d^2(\varphi_t^{-1}(y), y) + f(\varphi_t^{-1}(y)) \fullstop \end{equation*} \item \label{it:hopflax_convexity} For any $y,\,y'\in M$, one has the (strict-convexity-like) estimate \begin{equation*} \frac{d^2(y, y')}t \lesssim Q_tf(y)-Q_tf(y') +\frac1{2t}\left[d^2(\varphi_t^{-1}(y), y') - d^2(\varphi_t^{-1}(y), y)\right]\fullstop \end{equation*} \item \label{it:hopflax_regularity} The function $Q_tf$ is Lipschitz continuous in time and $C^{1,1}(M)$ in space. In particular we have $\norm{\partial_t Q_t f}_{\infty}\le \norm{\nabla f}_{\infty}$ and $\norm{\nabla^2Q_tf}_{\infty} \lesssim \norm{\nabla^2 f}_{\infty}$. \item \label{it:hamilton_jacobi} The function $Q_tf$ is a classical solution of the Hamilton-Jacobi equation \begin{equation*} \frac{\de}{\de t} Q_t f + \frac12 \abs{\nabla Q_t f}^2 = 0 \fullstop \end{equation*} \item \label{it:gradient_conservation} For any $x\in M$, if $\gamma:\cc01\to M$ is the geodesic such that $\gamma(0)=x$ and $\gamma'(0) = \nabla f(x)$, then it holds \begin{equation*} Q_tf(\gamma(t)) = f(x) + \frac t2 \abs{\nabla f}^2(x) \quad\text{ and }\quad \nabla Q_t f(\gamma(t)) = \gamma'(t) \fullstop \end{equation*} \item \label{it:hopflax_lip} One has \begin{equation*} \Lip(Q_tf-f) \le t\norm{\nabla f}_\infty\cdot\norm{\nabla^2 f}_\infty \fullstop \end{equation*} \end{enumerate} \end{theorem} \begin{proof} Thanks to the following homogeneity, for any $t>0$ and $\lambda>0$, of the Hopf-Lax semigroup \begin{equation*} Q_t(\lambda f)(y) = \lambda Q_{\lambda t} f(y) \comma \end{equation*} we can assume without loss of generality that $\norm{\nabla f}_\infty + \norm{\nabla^2 f}_\infty \le c$ and prove that the statements hold up to time $1$. Thus, we will implicitly assume that the time variable satisfies $0\le t\le 1$. We will choose the value of the constant $c$ during the proof, it should be clear that all constraints we impose depend only on the manifold $M$ and not on the function $f$. The statement of \cref{it:varphi_t_diffeo} follows from \cref{lem:exp_is_diffeo}. To prove \cref{it:hopflax_explicit} we need some preliminary observations. If $c=c(M)$ is sufficiently small (so that the constraint on $f$ is sufficiently strong), thanks to the compactness of $M$ we can find a radius $r=r(M)>0$ such that: \begin{enumerate}[label=(\alph*)] \item If $p,\,q\in M$ satisfy $d(p,q)\le r$ then \begin{equation*} \nabla^2 d^2(\emptyparam,p)(q) \ge \frac12 \metric \fullstop \end{equation*} \item For any $y\in M$, to compute $Q_tf(y)$ it is sufficient to minimize on $B(y,r)$: \begin{equation*} Q_tf(y) = \inf_{x\in B(y, r)} \frac1{2t}d^2(x, y)+f(x) \fullstop \end{equation*} \item For any $y\in M$ it holds the inequality $d(y, \varphi_t^{-1}(y)) \le r$. In particular we can assume that $\varphi_t^{-1}(y)$ is not in the cut-locus of $y$. \item For any $y\in M$ it holds the identity \begin{equation*} \nabla \left(\frac1{2t}d^2(\emptyparam, y) + f(\emptyparam)\right)(\varphi_t^{-1}(y)) = 0 \fullstop \end{equation*} This identity can be shown computing the gradient of the distance from $y$ squared, since we know that $y=\exp(t\nabla f(x))$ where $x=\varphi_t^{-1}(y)$. Indeed, given that $x$ does not belong to the cut-locus of $y$, we know \begin{equation*} \nabla \left(\frac12 d^2(\emptyparam, y)\right)(x) = -t\nabla f(x) \end{equation*} and the desired identity follows. \end{enumerate} With these observations at our disposal, the proof of \cref{it:hopflax_explicit} is straight-forward. Given a time $0\le t\le 1$ and a point $y\in M$, let us consider the function $w_{t,y}:M\to\R$ defined as \begin{equation*} w_{t,y}(x) = \frac1{2t}d^2(x,y) + f(x) \fullstop \end{equation*} We know that $Q_tf(y) = \min_{x\in B(y, r)} w_{t,y}(x)$. Moreover $\nabla w_{t,y}(\varphi_t^{-1}(y))=0$ and, if the constraint on $\norm{\nabla^2 f}_\infty$ is sufficiently small, we also know $\nabla^2 w_{t,y}\ge \frac1{3t}\metric$ in $B(y, r)$. Hence, by convexity, we deduce that $\varphi_t^{-1}(y)$ is the global minimum point of $w_{t,y}$ and \cref{it:hopflax_explicit} follows. Let us now move to the proof of \cref{it:hopflax_convexity}. Let $x,\, x'\in M$ be such that $\varphi_t(x) = y$ and $\varphi_t(x')=y'$. Applying \cref{it:hopflax_explicit} and recalling that $\varphi_t$ is a bi-Lipschitz diffeomorphism, we can see that the inequality we want to prove is equivalent to \begin{equation*} \frac1t d^2(x, x') \lesssim f(x)-f(x') + \frac 1{2t}\left(d^2(x, y')-d^2(x', y') \right) \end{equation*} and, using the same notation as above, this becomes \begin{equation*} \frac1t d^2(x, x') \lesssim w_{t, y'}(x) - w_{t, y'}(x') \fullstop \end{equation*} The latter inequality follows from the strict convexity of $w_{t,y'}$ that we have already shown while proving \cref{it:hopflax_explicit}. Showing from scratch that $Q_tf$ solves the Hamilton-Jacobi equation would not be hard, but for this we refer to \cite[Theorem 2.5, viii]{Lott-Villani07}, where the authors show that $Q_tf$ is a \emph{suitably weak} solution of the Hamilton-Jacobi equation. From their statement, we can deduce that if $Q_tf$ is differentiable at $x\in M$, then \begin{equation}\label{eq:hamilton_jacobi_ae} \frac{\de}{\de t}Q_tf(x) + \abs{\nabla Q_tf(x)}^2 = 0 \fullstop \end{equation} Since we will show that $Q_tf$ is $C^{1,1}(M)$, the validity of \cref{it:hopflax_regularity,it:hamilton_jacobi} is a consequence of \cref{eq:hamilton_jacobi_ae}. The first part of \cref{it:gradient_conservation}, namely $Q_tf(\gamma(t)) = f(x) + \frac t2\abs{\nabla f}^2(x)$, is implied by \cref{it:hopflax_explicit}. To obtain the identity involving the gradient, let us differentiate the previous equality with respect to the time variable. If $Q_t f$ is differentiable at $\gamma(t)$, it holds \begin{equation}\label{eq:almost_gradient_conservation} \frac{\de}{\de t} (Q_t f)(\gamma(t)) + \scalprod{\nabla Q_t f(\gamma(t))}{\gamma'(t)} = \frac{\de}{\de t} (Q_tf(\gamma(t))) = \frac12\abs{\nabla f}^2(x)\fullstop \end{equation} Applying \cref{it:hamilton_jacobi} and the fact that $\abs{\gamma'(t)} = \abs{\nabla f}(x)$, from \cref{eq:almost_gradient_conservation} we can deduce \begin{equation}\label{eq:hopflax_tmp} -\frac12 \abs{\nabla Q_tf}^2(\gamma(t)) + \scalprod{\nabla Q_t f(\gamma(t))}{\gamma'(t)} = \frac12\abs{\gamma'}^2(x) \iff \abs{\nabla Q_tf(\gamma(t)) - \gamma'(t)}^2 = 0 \fullstop \end{equation} This does not imply directly \cref{it:gradient_conservation} since we have shown the identity only if $Q_t f$ is differentiable at $\gamma(t)$. As a byproduct of \cref{it:hopflax_explicit}, we know that $Q_t f$ is Lipschitz continuous and therefore, from \cref{eq:hopflax_tmp}, we can deduce that, fixed $t$, for almost every $x\in M$ it holds \begin{equation*} \nabla Q_t f(\varphi_t(x)) = \frac{\partial \varphi_s(x)}{\partial s}\Big|_{s=t} \fullstop \end{equation*} Since the right-hand side is Lipschitz continuous (see \cref{lem:exp_is_diffeo}) it follows that $Q_tf\in C^{1,1}(M)$ and, as anticipated, this concludes the proofs of \cref{it:hopflax_regularity},\cref{it:hamilton_jacobi} and \cref{it:gradient_conservation}. Finally let us tackle \cref{it:hopflax_lip}. Given $y\in M$, let $x=\varphi_t^{-1}(y)$. Thanks to \cref{it:gradient_conservation}, if we consider the geodesic $\gamma:\cc01\to M$ such that $\gamma(0)=x$ and $\gamma'(0)=\nabla f(x)$, we know that $\gamma(t)=y$ and $\gamma'(t)=\nabla Q_tf(y)$. Thus we have \begin{equation*} \abs{\nabla f(y) -\nabla Q_tf(y)} \le \int_0^t \abs{\nabla_{\gamma'}\left(\nabla f(\gamma)-\gamma'\right)}\de s \le t\abs{\nabla f(x)}\cdot\norm{\nabla^2 f}_\infty \end{equation*} and this is the desired statement. \end{proof} \begin{remark} Let us emphasize that the only statement contained in \cref{thm:hopflax_properties} that we are going to use is \cref{it:hopflax_convexity}. Indeed it will be crucial when studying the stability of optimal maps. Furthermore, such a statement should be seen more like as a property of the $c$-conjugate (see \cite[Section 1.2]{Santambrogio15}) than as a property of the Hopf-Lax semigroup. We have proven all other statements in order to give a complete reference on the short-time behavior of the Hopf-Lax semigroup when the initial datum is in $C^{1,1}(M)$. \end{remark} \section{Quantitative Stability of the Optimal Map}\label{sec:stability} In this section we will always refer to the optimal transport with respect to the quadratic cost between two probability measures in $\prob(M)$ that are absolutely continuous with respect to the volume measure $\m$ of a compact Riemannian manifold $(M, \metric)$. The duality theory of optimal transport can be seen as a tool to bound from above and from below the optimal transport cost. Indeed, simply producing a transport map we can bound the cost from above, whereas with a pair of potentials we can bound it from below. Estimating the optimal cost is the best one can desire for a generic convex problem, but for the optimal transport problem we know that the optimal map is unique (see \cite{McCann01}) and thence we would like to be able to approximate it. In details, we want to investigate the following problem. \begin{problem} Let $\nu, \,\mu_1,\, \mu_2\in \prob(M)$ be probability measures with $\nu\ll \m$. Let $S,\, T$ be the optimal transport maps from $\nu$ to $\mu_1$ and $\mu_2$ respectively. Estimate the $L^2(\nu)$-distance $\norm{d(S, T)}^2_{L^2(\nu)}$ between the two maps. \end{problem} The approach we are going to adopt builds upon the method, suggested to N.Gigli by the first author, who used it in \cite[Proposition 3.3 and Corollary 3.4]{gigli2011}. In the proof of the mentioned results, the author obtains (even if not stated in this way) exactly the same inequality we are going to obtain. The substantial difference is that those results (and their proofs) work only when the ambient is the Euclidean space. Transporting the proofs from the flat to the curved setting is not straight-forward. The proof of Proposition~3.3 of the mentioned paper does not work on a Riemannian manifold, because curvature comes into play when comparing tangent vectors at different points. To overcome this difficulty we have come up with \cref{it:hopflax_convexity} of \cref{thm:hopflax_properties}. On the contrary, the proof of Corollary 3.4 is easily adapted on a compact Riemannian manifold. Let us also mention the recent result \cite[Theorem 4.1]{berman2018}. In the said theorem the author obtain a quantitative stability of the optimal map when, instead of changing the target measure as we are doing, the source measure is changed. The proof is totally different from ours and is mainly based on complex analytic tools. Also in that paper only the Euclidean setting (and the flat torus) is considered. We will attack the stability problem only in the \emph{perturbative setting}, namely when the optimal map from $\nu$ to $\mu_1$ is the identity up to the first order. Working only in the perturbative setting might look like an extremely strong assumption that would yield no applications at all. This is not the case, indeed what we call \emph{perturbative setting} is more or less equivalent to requiring only that the optimal transport map $T$ is local (meaning that $T-\id$ is uniformly small) and well-behaved. For example, and this is the whole point of \cite{ambrosio-glaudo2018}, the optimal map from the reference measure to a random point cloud is (with high probability) a perturbation of the identity. We don't need any hypothesis on the optimal map between $\nu$ and $\mu_2$. \begin{theorem}\label{thm:optimal_map_stability} Let $(M,\metric)$ be a closed compact Riemannian manifold (or the square $\cc01^2$) and let us denote by $\m$ its volume measure. Let $\nu, \mu_1, \mu_2\in\prob(M)$ be three probability measures with $\nu\ll\m$ and let $S, T:M\to M$ be the optimal transport maps respectively for the pairs of measures $(\nu, \mu_1)$ and $(\nu, \mu_2)$. We assume that $S=\exp(\nabla f)$ where $f:M\to\R$ is a $C^{1,1}$-function\footnote{If $M=\cc01^2$ we ask also that $f$ satisfies the null Neumann boundary conditions.} such that $\norm{\nabla f}_\infty + \norm{\nabla^2 f}_\infty \le c$ where $c=c(M)$ is the constant considered in the statement of \cref{thm:hopflax_properties}. Then it holds \begin{equation*} \int_M d^2(S, T)\de\nu \lesssim W_2^2(\mu_1, \mu_2) + W_2(\mu_1, \mu_2)W_2(\nu, \mu_1) \fullstop \end{equation*} \end{theorem} \begin{proof} Let us consider a generic transport map $S':M\to M$ from $\nu$ to $\mu_1$ and recall that, according to \cite{Glaudo19}, if $c(M)$ is small enough, then the map $S$ is optimal. Given $x\in M$, let us apply \cref{it:hopflax_convexity} of \cref{thm:hopflax_properties} with $y=S(x)$ and $y'=S'(x)$ and $t=1$ \begin{equation*} d^2(S(x),S'(x)) \lesssim Q_1f(S(x))-Q_1f(S'(x)) + \frac12\left(d^2(x, S'(x))-d^2(x, S(x))\right) \fullstop \end{equation*} Integrating this inequality with respect to $\nu$ we obtain \begin{equation}\label{eq:stability_same_measures} \int_M d^2(S, S')\de\nu \lesssim \norm{d(S', \id)}_{L^2(\nu)}^2 - W_2^2(\nu, \mu_1) \end{equation} as the first two terms cancel thanks to the fact that both $S$ and $S'$ sends $\nu$ into $\mu_1$. We can now prove the main statement under the additional assumption that there exists an optimal map $R:M\to M$ from $\mu_2$ to $\mu_1$. Applying \cref{eq:stability_same_measures} with $S' = R\circ T$ we get \begin{equation}\label{eq:temporary_ineq} \int_M d^2(S, R\circ T)\de\nu \lesssim \norm{d(R\circ T, \id)}_{L^2(\nu)}^2 - W_2^2(\nu, \mu_1) \fullstop \end{equation} Thanks to the triangle inequality, it holds \begin{align*} \norm{d(R\circ T, \id)}_{L^2(\nu)} &\le \norm{d(R\circ T, T)}_{L^2(\nu)} + \norm{d(T, \id)}_{L^2(\nu)} = \norm{d(R, \id)}_{L^2(\mu_2)} + W_2(\nu, \mu_2) \\ &\le 2 W_2(\mu_1, \mu_2) + W_2(\nu, \mu_1) \fullstop \end{align*} Applying this last inequality into \cref{eq:temporary_ineq} yields \begin{align*} \int_M d^2(S, R\circ T)\de\nu &\lesssim \left[2 W_2(\mu_1, \mu_2) + W_2(\nu, \mu_1)\right]^2 - W_2^2(\nu, \mu_1) \\ &\lesssim W_2^2(\mu_1, \mu_2) + W_2(\mu_1, \mu_2)W_2(\nu, \mu_1) \end{align*} and the desired statement follows from the triangle inequality \begin{align*} \int_M d^2(S, T) &\lesssim \int_M d^2(S, R\circ T)\de\nu + \int_M d^2(R\circ T, T)\de\nu \\ &\lesssim W_2^2(\mu_1, \mu_2) + W_2(\mu_1, \mu_2)W_2(\nu, \mu_1) + \int_M d^2(R, \id)\de\mu_2 \\ &= 2 W_2^2(\mu_1, \mu_2) + W_2(\mu_1, \mu_2)W_2(\nu, \mu_1) \fullstop \end{align*} It remains to drop the assumption on the existence of the optimal map $R$. Given that our ambient manifold is compact, we can apply the nonquantitative strong stability (see \cite[Corollary 5.23]{Villani08}). Let us take a sequence of absolutely continuous probability measures $\mu_2^n$ that weakly converges to $\mu_2$. Thanks to McCann's Theorem (see \cite{McCann01}) the optimal map $R^n$ from $\mu_2^n$ to $\mu_1$ exists and thanks to the strong stability we know that the optimal maps $T^n$ from $\nu$ to $\mu_1^n$ converge strongly in $L^2(\nu)$ to $T$. Hence it is readily seen that the result for $\mu_2$ can be obtained by passing to the limit the result for $\mu_2^n$. \end{proof} \begin{remark} The first part of the proof of \cref{thm:optimal_map_stability} might seem a bit magical. Let us describe what is happening under the hood. The function $f$ is the Kantorovich potential of the couple $(\nu, \mu_1)$ and hence, by standard theory in optimal transport, it must be $c$-concave. Our hypotheses ensure us that it is not only $c$-concave, but even \emph{strictly} $c$-concave. Furthermore, the theory we have developed on the Hopf-Lax semigroup tells us that even the other potential $f^c=Q_1 f$ is strictly $c$-concave (this is exactly \cref{it:hopflax_convexity}). The result follows integrating the strict $c$-concavity inequality with respect to the measure $\nu$. \end{remark} \begin{remark} The main use of \cref{thm:optimal_map_stability} is the following one. Assume that the optimal map from $\nu$ to $\mu_1$ is local and well-behaved (this ensures the validity of the hypotheses of the theorem) and furthermore that $\mu_2$ is much closer to $\mu_1$ than to $\nu$. In this situation, the theorem tells us \begin{equation*} \int_M d^2(S, T)\de\nu \ll \int_M d^2(S, \id)\de\nu \comma \end{equation*} and this conveys exactly the information that $S$ approximates very well $T$. Notice also that the improvement from $C^{0,1/2}$ dependence of \cite{gigli2011} to the kind of Lipschitz dependence is due to the fact that we are working in a perturbative regime, close to the reference measure. \end{remark} \section{Optimal map in the random matching problem}\label{sec:random_matching} We want to apply our result on the stability of the optimal map in the perturbative setting to the semi-discrete random matching problem. In this section we will work on a compact closed Riemannian manifold $(M, \metric)$ of dimension $2$ (or the square $\cc01^2$). We will denote with $\m$ the volume measure, with the implicit assumption that it is a probability. In this setting, the semi-discrete random matching problem can be formulated as follows. For a fixed $n\in\N$, consider $n$ independent random points $X_1, X_2, \dots, X_n$ $\m$-uniformly distributed on $M$. Study the optimal transport map $T^n$ (with respect to the quadratic cost) from $\m$ to the empirical measure $\mu^n = \frac1n\sum_{i} \delta_{X_i}$. Since we want to attack the problem applying \cref{thm:optimal_map_stability}, first of all we have to choose $\nu,\, \mu_1$ and $\mu_2$. The choices of $\nu$ and $\mu_2$ are very natural, indeed we set $\nu=\m$ and $\mu_2=\mu^n$. This way the map $T$ is $T^n$ . Far less obvious is the choice of $\mu_1$, $S$ and $f$. As one might expect from the statement of \cref{thm:main_theorem} and from the ansatz described in the introduction, our choice is $f=f^{n,t}$. Thus $S=\exp(\nabla f^{n,t})$ (for some appropriate $t=t(n)$). Furthermore, keeping the same notation of \cite{ambrosio-glaudo2018}, the measure $\mu_1=S_\#\m$ will be denoted by $\hat\mu^{n,t}$. First of all it is crucial to understand whether we are in position to apply \cref{thm:optimal_map_stability}. Indeed we need to check if $\nabla^2 f^{n,t}$ and $\nabla f^{n,t}$ are sufficiently small. Moreover we have to obtain a strong estimate on $W_2^2(\mu_1, \mu_2)$. Both this facts are among the main results obtained in \cite{ambrosio-glaudo2018}. Hence let us state them in the following proposition. \begin{proposition}[Summary of results from \texorpdfstring{\cite{ambrosio-glaudo2018}}{[AG18]}]\label{prop:ag18} Let $(M, \metric)$ be a closed compact $2$-dimensional Riemannian manifold (or the square $\cc01^2$) whose volume measure $\m$ is a probability. Given $n\in\N$, let $X_1,\dots, X_n$ be $n$ independent random points $\m$-uniformly distributed on $M$ and denote $\mu^n=\frac1n\sum_i \delta_{X_i}$ the associated empirical measure. For a choice of the time $t>0$, let $\mu^{n,t}=P^*_t(\mu^n)$ be the evolution through the heat flow of the empirical measure and let $f^{n,t}:M\to\R$ be the unique null-mean solution\footnote{If $M=\cc01^2$ we ask also that $f$ satisfies the null Neumann boundary conditions.} to the Poisson equation $-\lapl f^{n,t} = \mu^{n,t}-1$. Finally, let us define the probability measure $\hat\mu^{n,t}$ as the push-forward of $\m$ through the map $\exp(\nabla f^{n,t})$. For any $\xi>0$, let $A^{n,t}_\xi$ be the probabilistic event $\{\norm{\nabla^2 f^{n,t}}_\infty < \xi\}$. If $t=t(n)=\frac{\log^4(n)}{n}$ and $\xi=\xi(n) = \frac1{\log(n)}$, the following statements\footnote{In \cite{ambrosio-glaudo2018} the time $t(n)$ is chosen as $t(n)=\gamma\frac{\log^3(n)}{n}$, where $\gamma$ is a constant. As we clarify in \cref{rem:time_is_flexible}, the choice of the exponent of the logarithm in the definition of $t(n)$ is not rigid. We choose the exponent $4$ instead of $3$ since it lets us get some estimates in a cleaner form and makes it possible to avoid inserting a constant in the definition of $t(n)$.} hold \begin{itemize} \item We know the asymptotic behavior of the expected matching cost \begin{equation}\label{eq:matching_cost_limit} \lim_{n\to\infty} \E{W_2^2(\m, \mu^n)}\left(\frac1{4\pi}\frac{\log(n)}n\right)^{-1} = 1 \fullstop \end{equation} \item The probability of the complement of $A^{n,t}_\xi$ decays faster than any power. In formulas, for any $k>0$ there exists a constant $C=C(M, k)$ such that \begin{equation}\label{eq:exceptional_set_rare} \P{\left(A^{n,t}_\xi\right)^\complement} \le C(M, k) n^{-k} \fullstop \end{equation} \item One has the refined contractivity estimate\footnote{This does not follow from the well-known contractivity property for the heat semigroup. Indeed the standard contractivity would yield an estimate of order $t=\gamma\frac{\log^4(n)}{n}\gg \frac{\log(n)}{n}$ and such magnitude is too large for our purposes.} \begin{equation}\label{eq:diffusion_error} \E{W_2^2(\mu^n, \mu^{n,t})} \lesssim \frac{\log(\log(n))}n \left(\ll \frac{\log(n)}{n}\right)\fullstop \end{equation} \item We are able to control the perturbation error with \begin{equation}\label{eq:perturbation_error} \E{W_2^2(\mu^{n,t}, \hat\mu^{n,t})} \lesssim \frac{1}{n\log(n)} \left(\ll \frac{\log(n)}{n}\right)\fullstop \end{equation} \item \label{it:exp_is_optimal} When $n$ is sufficiently large, in the event $A^{n,t}_\xi$ the map $\exp(\nabla f^{n,t})$ is optimal from $\m$ to $\hat\mu^{n,t}$. \end{itemize} \end{proposition} \begin{proof} All of these results are contained in \cite{ambrosio-glaudo2018} and thus we will only give a precise reference for them. All references are to propositions contained in \cite{ambrosio-glaudo2018}. The validity of \cref{eq:matching_cost_limit} is contained in Theorem 1.2. The fact that the event $A^{n,t}_\xi$ has overwhelming probability follows from Theorem 3.3. The refined contractivity estimate \cref{eq:diffusion_error} is Theorem 5.2. The estimate \cref{eq:perturbation_error} follows from Equation 6.2 and Lemma 3.14. More specifically Equation 6.2 tells us that in the event $A^{n,t}_\xi$ it holds \begin{equation*} \E{W_2^2(\mu^{n,t}, \hat\mu^{n,t})} \lesssim \xi^2 \int_M\abs{\nabla f^{n,t}}^2\de\m \end{equation*} and Lemma 3.14 gives us the expected value of the Dirichlet energy of $f^{n,t}$. The behavior in the complementary of $A^{n,t}_\xi$ can be ignored thanks to \cref{eq:exceptional_set_rare}. It remains to show that in the event $A^{n,t}_\xi$, the map $\exp(\nabla f^{n,t})$ is optimal. This follows directly from \cite[Theorem 1.1]{Glaudo19}. \end{proof} \begin{remark}\label{rem:repeat_old_remark} Let us repeat the elementary observation made in \cite[Remark 5.3]{ambrosio-glaudo2018}, as it will be useful. Let $X, Y$ be two random variables such that, in an event $E$, it holds $X\le Y$. Then \begin{equation*} \E{X} \le \E{Y} + (\norm{X}_\infty + \norm{Y}_\infty)\P{E^\complement} \fullstop \end{equation*} In particular, if the infinity norm of $X, Y$ is suitably controlled and the probability of $E^\complement$ is exceedingly small, we can assume $\E{X}\le \E{Y}$ up to a small error. This observation allows us to restrict our study to the \emph{good} event $A^{n,t}_\xi$. Indeed all quantities involved in our computations have at most polynomial growth, whereas $\P{(A^{n,t}_\xi)^\complement}$ decays faster than any power. \end{remark} Once we have these results in our hands, the proof of the main theorem follows rather easily. Indeed we just have to check that all assumptions of our stability result are satisfied. \begin{proof}[Proof of \cref{thm:main_theorem}] Let us assume to be in the event $A^{n,t}_\xi$ with $\xi = \frac1{\log(n)}$. Hence, thanks to \cref{it:exp_is_optimal}, we can apply \cref{thm:optimal_map_stability} to the triple of measures $\nu=\m$, $\mu_1=\hat\mu^{n,t}$ and $\mu_2=\mu^n$ (with $S=\exp(\nabla f^{n,t})$ and $T=T^n$). We obtain \begin{align*} \int_M d^2(\exp(\nabla f^{n,t}), T^n)\de\m &\lesssim W_2^2(\mu^n, \hat\mu^{n,t}) + W_2(\mu^n, \hat\mu^{n,t})W_2(\m, \hat\mu^{n,t}) \\ &\lesssim W_2^2(\mu^n, \hat\mu^{n,t}) + W_2(\mu^n, \hat\mu^{n,t})W_2(\m, \mu^n) \fullstop \end{align*} Recalling \cref{rem:repeat_old_remark} and \cref{eq:exceptional_set_rare}, if we consider the expected value we can apply the latter inequality as if it were true unconditionally and not only in the event $A^{n,t}_\xi$. Thus, taking the expected value and applying Cauchy-Schwarz's inequality, we get \begin{equation*} \E{\int_M d^2(\exp(\nabla f^{n,t}), T^n)\de\m} \lesssim \E{W_2^2(\mu^n, \hat\mu^{n,t})} + \sqrt{\E{W^2_2(\mu^n, \hat\mu^{n,t})}\cdot \E{W^2_2(\m, \mu^n)}} \fullstop \end{equation*} The desired statement follows directly applying \cref{eq:matching_cost_limit,eq:diffusion_error,eq:perturbation_error}. \end{proof} \begin{remark}\label{rem:time_is_flexible} It might seem that our choice of the time $t=\log^4(n)/n$ is a little arbitrary, and indeed it is. Any time $t=t(n)$ of order $\log^\alpha(n)/n$, for some $\alpha > 3$, would have worked flawlessly. \end{remark} It remains to justify \cref{rem:distance-tangent}. As already said, the desired estimate boils down to the validity of \begin{equation}\label{eq:linf_distance_bounded} \P{\|d(T^{n},\id)\|_\infty > \varepsilon} \ll \frac{\log(n)}n \end{equation} for any fixed $\eps>0$. The strategy of the proof is as follows. With \cref{lem:linf_l2_bound} (see also \cite[Lemma 4.1]{goldman2017}) we reduce the hard task of controlling the $L^\infty$-distance between $T^n$ and $\id$ to the easier task of controlling $W_2^2(\m,\mu^n)$. This latter estimate is then shown to be a consequence of \cref{eq:exceptional_set_rare}. \begin{lemma}\label{lem:linf_l2_bound} Let $(M,\metric)$ be a $d$-dimensional compact Riemannian manifold (possibly with Lipschitz boundary) and let $\m$ be the volume measure on $M$. If $T:M\to M$ is the optimal map with respect to the quadratic cost from $\m$ to $T_\#\m$, then one has \begin{equation*} \norm{d(\id, T)}_{L^\infty(M)} \lesssim \left(\int_M d^2(\id, T)\de\m\right)^{\frac1{d+2}} \fullstop \end{equation*} \end{lemma} \begin{proof} Since the map $T$ is optimal, its graph is essentially contained in $c$-cyclically monotone set (see~\cite[Theorem 1.38]{Santambrogio15}). More precisely, there exists a Borel set $C\subseteq M$ such that $\{(x,T(x)):\ x\in C\}$ is $c$-cyclically monotone and $M\setminus C$ is $\m$-negligible. We will reduce our considerations to points in $C$ in order to exploit the $c$-cyclical monotonicity. Let us fix a point $x_0\in C$ and let us define $\alpha \defeq \frac12 d(x_0, T(x_0))$. Let us define the point $p\in M$ as the middle point between $x_0$ and $T(x_0)$, that is $d(x_0, p) = d(p, T(x_0)) = \alpha$. Let us consider a point $x\in B(p, \eps\alpha)\cap C$ where $\eps>0$ is a small constant that will be chosen a posteriori. Finally let us define $\beta\defeq d(x, T(x))$. We want to show that $\beta$ cannot be much smaller than $\alpha$. \begin{figure}[htb] \centering \input{linf_l2_bound.tikz} \caption{The points considered in the proof of of \cref{lem:linf_l2_bound}.} \end{figure} Thanks to the $c$-cyclical monotonicity of $C$, it holds \begin{equation*} d^2(x_0, T(x_0)) + d^2(x, T(x)) \le d^2(x, T(x_0)) + d^2(x_0, T(x)) \end{equation*} and thus, applying repeatedly the triangle inequality, we deduce \begin{align*} 4\alpha^2 + \beta^2 &\le \left(d(x, p) + d(p, T(x_0))\right)^2 + \left(d(x_0, p) + d(p, x) + d(x, T(x))\right)^2 \\ &\le (\eps\alpha + \alpha)^2 + (\alpha + \eps\alpha + \beta)^2 = 2(1+\eps)^2\alpha^2 + \beta^2 + 2(1+\eps)\alpha\beta \\ &\hphantom{asdasdasdasdasdasdasd}\Updownarrow \\ &\hphantom{asdasdasd}(2-(1+\eps)^2) \alpha \le (1+\eps)\beta \fullstop \end{align*} If $\eps$ is chosen sufficiently small (i.e. $\eps = 1/3$), the desired estimate $\alpha\lesssim\beta$ follows. Since $x$ can be chosen arbitrarily in $B(p, \eps\alpha)\cap C$, the estimate $\alpha\lesssim \beta$ implies \begin{align*} \int_M d^2(x, T(x))\de\m(x) &\ge \int_{B(p, \eps\alpha)} d^2(x, T(x))\de\m(x) \gtrsim \m(B(p, \eps\alpha)) d^2(x_0, T(x_0)) \\ &\gtrsim \eps\alpha^d d^2(x_0, T(x_0) \gtrsim \bigl(d(x_0, T(x_0))\bigr)^{d+2} \comma \end{align*} where we have used that a ball with radius $r$ not larger than the diameter of $M$ has measure comparable to $r^d$ (follows from the Ahlfors-regularity of compact Riemannian manifolds with Lipschitz boundary). This completes the proof since $x_0$ can be chosen arbitrarily in a set with full measure. \end{proof} \begin{remark} The previous lemma holds, with the same proof, on any Ahlfors-regular metric measure space that is also a length space. \end{remark} \begin{remark} If we apply \cref{lem:linf_l2_bound} on a $2$-dimensional manifold with $T^n$ being the optimal map (with respect to the quadratic cost) from $\m$ to the empirical measure $\mu^n$, we obtain \begin{equation*} \norm{d(\id,T^n}_{L^{\infty}(M)} \lesssim W_2(\m, \mu^n)^{\frac{1}{2}} \fullstop \end{equation*} Since we know (as a consequence of \cref{eq:limit_value}) that with high probability $W^2_2(\m, \mu^n) \lesssim n^{-1}\log(n)$, we deduce that with high probability it holds \begin{equation*} \norm{d(\id,T^n}_{L^{\infty}(M)} \lesssim \left(\frac{\log(n)}{n}\right)^{\frac14} \fullstop \end{equation*} This estimate does not match the asymptotic behavior of the $\infty$-Wasserstein distance between $\m$ and $\mu^n$. In fact, as proven in \cite{leighton1989,shor1991,trillos2015}, with high probability it holds \begin{equation*} W_{\infty}(\m, \mu^n) \approx \frac{\log(n)^{\frac34}}{n^{\frac12}} \fullstop \end{equation*} \end{remark} We are now ready to show \cref{eq:linf_distance_bounded} (to be precise we prove a much stronger estimate). \begin{proposition}\label{prop:linf_is_small} Using the same notation and definitions of the statement of \cref{thm:main_theorem}, for any $\eps>0$ and any $k>0$ there exists a constant $C=C(M, \eps, k)$ such that \begin{equation} \P{\|d(T^{n},\id)\|_\infty > \varepsilon} \le C(M, \eps, k)n^{-k} \fullstop \end{equation} \end{proposition} \begin{proof} We show that for any $\eps>0$ and any $k>0$ there exists a constant $C=C(M, \eps, k)$ such that \begin{equation}\label{eq:wasserstein_is_small} \P{W_2(\m, \mu^n) > \varepsilon} \le C(M, \eps, k)n^{-k} \fullstop \end{equation} In fact, if we are able to prove \cref{eq:wasserstein_is_small}, then the statement of the proposition follows applying \cref{lem:linf_l2_bound} with $T=T^n$ (changing adequately $\eps,k$ and the value of the constant $C$). The triangle inequality gives us \begin{equation}\label{eq:linf_tmp1} W_2(\m, \mu^n) \le W_2(\mu^{n,t}, \mu^n) + W_2(\m, \mu^{n,t}) \fullstop \end{equation} The first term can be bounded using the contractivity property of the heat semigroup, obtaining \begin{equation}\label{eq:linf_tmp2} W_2(\mu^{n,t}, \mu^n) \lesssim \sqrt{t} \fullstop \end{equation} For the second term we employ the transport inequality \cite[(4.1)]{ambrosio-glaudo2018} and get \begin{equation}\label{eq:linf_tmp3} W^2_2(\m, \mu^{n,t}) \lesssim \int_M \abs{\nabla f^{n,t}}^2\de\m \fullstop \end{equation} If we assume to be in the event $A^{n,t}_\xi$ (that is defined in the statement of \cref{prop:ag18}) with $\xi=\xi(n)=\frac1{\log(n)}$, we have \begin{equation}\label{eq:linf_tmp4} \int_M \abs{\nabla f^{n,t}}^2\de\m \lesssim \norm{\nabla f}_{L^{\infty}(M)}^2 \lesssim \norm{\nabla^2 f}_{L^{\infty}(M)}^2 \le \xi^2 \fullstop \end{equation} Joining \cref{eq:linf_tmp1,eq:linf_tmp2,eq:linf_tmp3,eq:linf_tmp4} we deduce that in the event $A^{n,t}_\xi$ it holds \begin{equation*} W_2(\m, \mu^n) \lesssim \sqrt{t} + \xi \fullstop \end{equation*} Since $t(n)\to 0$ and $\xi(n)\to 0$ as $n\to\infty$, this implies (for $n$ sufficiently large) that in the event $A^{n,t}_\xi$ it holds $W_2(\m, \mu^n) \le \eps$. Hence \cref{eq:wasserstein_is_small} is a consequence of \cref{eq:exceptional_set_rare} and this concludes of the proof. \end{proof} \printbibliography \end{document}

https://arxiv.org/abs/1205.6032

A Remark on Chern-Weil Theory

We show that Chern-Weil theory for tensor bundles over manifolds is a consequence of the existence of natural closed differential forms on total spaces of torsion free connections on frame bundles.

\section{Introduction} Consider $n$-dimensional real manifold $M$ and the frame bundle $\mathcal{R}_M$ over $M$. We identify bundles and their total spaces. There is the affine bundle \[ \pi : \mathbf{Conn}(\mathcal{R}_M) \to M \] of torsion free connections on $\mathcal{R}_M$ (sections of bundle $\pi$ are torsion free connections on $\mathcal{R}_M$). \begin{theorem}\label{main th} There are natural nonzero closed complex valued differential forms \[ \omega_k (M) \in \Omega^{2k} (\mathbf{Conn}(\mathcal{R}_M), \mathbf{C}), \qquad k = 1, \ldots , \left[ \frac{n}{2}\right]. \] \end{theorem} In the theorem the adjective \emph{natural} means that \it for every $n$-di\-men\-sio\-nal real manifold $N$ and open embedding $f : N \to M$ we have \[ F^* (\omega_k (M)) = \omega_k (N), \] where \[ F : \mathbf{Conn}(\mathcal{R}_N) \to \mathbf{Conn}(\mathcal{R}_M) \] is the embedding defined by $f$ \cite{curv}. \rm Suppose that $D$ is a connection on $\mathcal{R}_M$ and $D_0$ is the corresponding torsion free connection on $\mathcal{R}_M$. So $D_0$ is a differential mapping \[ D_0 : M \to \mathbf{Conn}(\mathcal{R}_M). \] Consider $2k$-forms \[ D_0^*(\omega_k (M)) \in \Omega^{2k}(M, \mathbf{C}), \qquad k = 1, \ldots , \left[ \frac{n}{2}\right]. \] From theorem \ref{main th} and from the fact that every connection on $\mathcal{R}_M$ can be deformed to any other one it follows that \begin{enumerate} \item $d (D_0^*(\omega_k (M))) = 0$; \item cohomology class $[D_0^*(\omega_k (M))] \in H^{2k}(M, \mathbf{C})$ does not depend on the connection $D$. \end{enumerate} In fact, forms $D_0^*(\omega_k (M))$ are exactly $2k$-forms constructed in Chern-Weil theory \cite{chern-weil}. So theorem \ref{main th} means that Chern-Weil theory for tensor bundles over manifolds is a consequence of existence of differential forms $\omega_k (M)$. \section{Proof of theorem \ref{main th}} Let $\{ x^i \}$ be local coordinates on an open subset $X$ of the manifold $M$. The coordinates $\{ x^i \}$ define the local coordinates $\left\{ x^i, \Gamma_{ij}^k \right\}$ on the open subset $\pi^{-1}(X)$ of the manifold $\mathbf{Conn}(M)$. On $\pi^{-1}(X)$ consider 1-forms \[ \widehat\theta_\beta^\alpha := \Gamma_{i \beta}^\alpha d x^i \] and construct the matrix \[ \widehat\theta := \left( \widehat\theta_\beta^\alpha \right). \] Let $\{ {x'}^i \}$ be local coordinates on an open subset $X$ of the manifold $M$. By definition of the manifold $\mathbf{Conn}(M)$, on the intersection $X \cap X'$ we have the transition formula \[ \widehat\theta' = \left( \frac{\partial x'}{\partial x} \right)^{-1} d \left( \frac{\partial x'}{\partial x} \right) + \left( \frac{\partial x'}{\partial x} \right)^{-1} \widehat\theta \left( \frac{\partial x'}{\partial x} \right). \] On $\pi^{-1}(X)$ consider the matrix \[ \widehat\Theta := d \widehat\theta + \widehat\theta \wedge \widehat\theta \] whose entries are 2-forms on $\pi^{-1}(X)$: \begin{equation}\label{theta} \widehat\Theta_\beta^\alpha = d\Gamma_{i \beta}^\alpha \wedge d x^i \ + \ \Gamma_{i k}^\alpha \Gamma_{j \beta}^k dx^i \wedge dx^j. \end{equation} We claim that on the intersection $X \cap X'$ we have the transition formula \begin{equation}\label{tr formula} \widehat\Theta' = \left( \frac{\partial x'}{\partial x} \right)^{-1} \widehat\Theta \left( \frac{\partial x'}{\partial x} \right). \end{equation} We omit an easy check-up of this formula. In fact, it exactly looks like the check-up of the analogous classical transitional formula. Define the forms $\omega_k (M)$ by the formula \[ 1 + \sum_{k \ge 1} \omega_k (M) = \det\left(\mathbf{E} + \frac{\sqrt{-1}}{2 \pi} \widehat\Theta \right), \] where $\mathbf{E}$ is the identity matrix of size $n \times n$. From \eqref{tr formula} it follows that forms $\omega_k (M)$ are natural one on the manifold $M$. Let us prove that forms $\omega_k (M)$ are closed. Let $A \in \mathbf{Conn}(\mathcal{R}_M)$, $\{ x^i \}$ be local coordinates on a neighbourhood $X \subset M$ of the point $\pi (A)$, and $\left\{ x^i, \Gamma_{ij}^k \right\}$ be the corresponding local coordinates on the neighbourhood $\pi^{-1}(X) \subset \mathbf{Conn}(\mathcal{R}_M)$ of the point $A$. Replacing if needed local coordinates $\{ x^i \}$ by new ones we pass to the case when \[ \left. \Gamma_{ij}^k \right|_A = 0. \] In this case from \eqref{theta} it follows that \[ \left. d \widehat\Theta_\beta^\alpha \right|_A= 0 \] whence we get \[ \left. d \det\left(\mathbf{E} + \frac{\sqrt{-1}}{2 \pi} \Theta \right) \right|_A = 0. \] From the above it follows that \[ d \det\left(\mathbf{E} + \frac{\sqrt{-1}}{2 \pi} \Theta \right) = 0 \] and thus forms $\omega(M)$ are closed.

https://arxiv.org/abs/1608.06793

Lie algebras with nilpotent length greater than that of each of their subalgebras

The main purpose of this paper is to study the finite-dimensional solvable Lie algebras described in its title, which we call {\em minimal non-${\mathcal N}$}. To facilitate this we investigate solvable Lie algebras of nilpotent length $k$, and of nilpotent length $\leq k$, and {\em extreme} Lie algebras, which have the property that their nilpotent length is equal to the number of conjugacy classes of maximal subalgebras. We characterise the minimal non-${\mathcal N}$ Lie algebras in which every nilpotent subalgebra is abelian, and those of solvability index $\leq 3$.

\section{Introduction} Let $L$ be a Lie algebra, and let $L^{(0)}=L$, $L^{(i+1)}=[L^{(i)},L^{(i)}]$ be its derived series. Recall that $L$ is {\em solvable} if there exists $r$ such that $L^{(r)}=0$; the smallest such $r$ is called the {\em derived length} of $L$. Similarly, $L^1=L$, $L^{i+1}=[L^i,L]$ is the lower central series of $L$; $L$ is {\em nilpotent} of {\em nilpotency index $r$} if $L^{r+1}=0$ but $L^r \neq 0$. Throughout, $L$ will denote a finite-dimensional solvable Lie algebra over a field $F$. The symbol `$\oplus$' will denote an algebra direct sum, whilst `$\dot{+}$' will denote a direct sum of the underlying vector space structure alone. If $U$ is a subalgebra of $L$ we define $U_L$, the {\em core} (with respect to $L$) of $U$, to be the largest ideal of $L$ contained in $U$. We say that $U$ is {\em core-free} in $L$ if $U_L = 0$. \par We denote the nilradical of $L$ by $N(L)$. We define the {\em upper nilpotent series} of $L$ by \[ N_0(L)=0, \hspace{.5cm} N_i(L)/N_{i-1}(L)=N(L/N_{i-1}(L)) \hbox{ for } i=1,2, \ldots \] The {\em nilpotent length}, $n(L)$, of $L$ is the smallest integer $n$ such that $N_n(L)=L$. \par We define the {\em nilpotent residual}, $\gamma_{\infty}(L)$, of $L$ be the smallest ideal of $L$ such that $L/\gamma_{\infty}(L)$ is nilpotent. Clearly this is the intersection of the terms of the lower central series for $L$. Then the {\em lower nilpotent series} for $L$ is the sequence of ideals $\Gamma_i(L)$ of $L$ defined by $\Gamma_0(L) = L$, $\Gamma_{i+1}(L) = \gamma_{\infty}(\Gamma_i(L))$ for $i \geq 0$. First we note that this series has the same length as that of the upper nilpotent series. \begin{lemma}\label{l:length} Suppose that $r$ is the smallest integer such that $N_r(L)=L$, and that $s$ is the smallest integer such that $\Gamma_s(L)=0$. Then \begin{itemize} \item[(i)] $\Gamma_{s-i}(L)\subseteq N_i(L)$ for $i=0,\ldots,r$; \item[(ii)] $\Gamma_i(L)\subseteq N_{r-i}(L)$ for $i=0,\ldots,s$; and \item[(iii)] $r=s$. \end{itemize} \end{lemma} \begin{proof}\begin{itemize} \item[(i)] Clearly $\Gamma_{s-1}(L)$ is a nilpotent ideal of $L$ and so $\Gamma_{s-1}(L)\subseteq N_1(L)$. Suppose that $\Gamma_{s-k}(L)\subseteq N_k(L)$ for some $k\geq 1$. then \[ \frac{\Gamma_{s-k-1}(L)+N_k(L)}{N_k(L)} \] is a nilpotent ideal of $L/N_k(L)$ and so is contained in $N_{k+1}(L)/N_k(L)$. Hence $\Gamma_{s-k-1}(L)\subseteq N_{k+1}(L)$, and the result follows. \item[(ii)] Clearly $\Gamma_1(L)\subseteq N_{r-1}(L)$. Suppose that $\Gamma_k(L)\subseteq N_{r-k}(L)$ for some $k\geq 1$. Then \[ \frac{\Gamma_k(L)+N_{r-k-1}(L)}{N_{r-k-1}(L)}\cong \frac{\Gamma_k(L)}{\Gamma_k(L)\cap N_{r-k-1}(L)} \] is nilpotent, whence $\Gamma_{k+1}(L)\subseteq N_{r-k-1}(L)$, and the result follows. \item[(iii)] From (i) we have that $\Gamma_{s-r}(L)\subseteq N_r(L)$, so $s-r\geq 0$; from (ii) we see that $\Gamma_s(L)\subseteq N_{r-s}(L)$, so $r-s\geq 0$. Thus $r=s$. \end{itemize} \end{proof} \bigskip Although the upper and lower nilpotent series have equal lengths, $n$ say, we do not necessarily have that $\Gamma_{n-i}(L)=N_i(L)$ for $i=0,\ldots,n$, as the following example shows. \begin{ex} Let $L$ be the metabelian Lie algebra over $\C$ with basis $x_1$, $x_2$, $x_3$, $x_4$ and non-zero products $[x_1,x_2]=x_3$, $[x_1,x_3]=x_3$, $[x_1,x_4]=x_4$, $[x_2,x_3]=x_4$. Then $N_1(L)=\C x_2+\C x_3+\C x_4$, $N_2(L)=L$, $\Gamma_1(L)=\C x_3+\C x_4$, $\Gamma_2(L)=0$. \end{ex} From now on we choose to work with the upper nilpotent series. In section $2$ we investigate properties of this series, particularly its relationship to maximal subalgebras, and of Lie algebras with nilpotent length $k$ or $\leq k$. In considering factor algebras, a complication arises because, unlike the situation in group theory, there are solvable Lie algebras $L$ in which, for an ideal $I$ of $L$, $N(I)$ may not be contained in $N(L)$. To overcome this obstacle we introduce the notions of nilregular and strongly nilregular subalgebras of $L$; in particular, it is shown that if a maximal subalgebra of $L$ has a strongly nilregular core, its nilpotent length is at most one less than that of $L$. The section concludes with a fundamental decomposition theorem for Lie algebras with a given nilpotent length. \par In section $3$ we introduce the class of extreme Lie algebras in which $N_i(L)/\phi_i(L)$ is a chief factor for each $i=1, \ldots, n(L)$. These are characteriseded in relation to the decomposition result from the previous section and described explicitly in two special cases. \par The final section then considers the algebras in the title of the paper. By considering their relationship to extreme Lie algebras the minimal non-${\mathcal N}$ Lie algebras in which every nilpotent subalgebra is abelian, and those of solvability index $\leq 3$, are characterised. The last result is that a homomorphic image of a minimal non-${\mathcal N}$ Lie algebra is minimal non-${\mathcal N}$ if it has a complemented minimal ideal. Much of this is inspired by corresponding work in group theory in \cite{frick} and \cite{c-f-h}, but there are significant differences encountered in the Lie case. \section{Properties of the upper nilpotent series} The {\em Frattini subalgebra} of $L$, $\phi(L)$, is the intersection of the maximal subalgebras of $L$. Since $L$ is solvable, this is an ideal of $L$ (\cite[Lemma 3.4]{b-g-h}). The {\em Frattini series} of $L$ is given by \[ \phi_i(L)/N_{i-1}(L)= \phi(L/N_{i-1}(L)) \hbox{ for } i=1,2, \ldots \] \begin{lemma}\label{l:nat} Let $B$ be an ideal of $L$ with $B \subseteq \phi(L)$. Then $N_i(L/B)=N_i(L)/B$ and $\phi_i(L/B)=\phi_i(L)/B$ for every $i=1,2, \ldots, n(L)$. \end{lemma} \begin{proof} We have $N(L)/B=N(L/B)$, by \cite[Theorem 5]{barnes}, and $\phi(L)/B=\phi(L/B)$, by \cite[Proposition 4.3]{frat}. Suppose that $N_k(L)/B=N_k(L/B)$ and $\phi_k(L)/B= \phi_k(L/B)$. Then $B \subseteq \phi_{k+1}(L)$ and \[ \frac{N_{k+1}(L/B)}{N_k(L/B)}=N\left(\frac{L/B}{N_k(L)/B}\right)=\frac{N_{k+1}(L)/B}{N_k(L)/B}, \] whence $N_{k+1}(L)/B=N_{k+1}(L/B)$. Similarly, \[ \frac{\phi_{k+1}(L/B)}{N_k(L/B)}=\phi\left(\frac{L/B}{N_k(L)/B}\right)=\frac{\phi_{k+1}(L)/B}{N_k(L)/B}, \] which yields that $\phi_{k+1}(L)/B=\phi_{k+1}(L/B)$. \end{proof} \begin{lemma}\label{l:quot} If $A$ is an ideal of $L$ with $N_{r-1}(L)\subseteq A \subseteq N_r(L)$, then $n(L/A)=n(L)-r$ or $n(L)-r+1$. \end{lemma} \begin{proof} Put $K_i/A=N_i(L/A)$. Then it is easy to see that $K_i/K_{i-1}=N(L/K_{i-1})$, and a straightforward induction argument shows that \[ N_{r-1}(L)\subseteq A\subseteq N_r(L)\subseteq K_1\subseteq N_{r+1}(L)\subseteq \ldots \subseteq K_{n(L)-r}\subseteq N_{n(L)}(L). \] If $K_{n(L)-r}=N_{n(L)}(L)$ we have $n(L/A)=n(L)-r$; otherwise, $n(L/A)=n(L)-r+1$. \end{proof} \begin{lemma}\label{l:max} Let $M$ be a maximal subalgebra of the solvable Lie algebra $L$. Then \begin{itemize} \item[(i)] $N_i(L) \cap M \subseteq N_i(M)$; \item[(ii)] $N_i(M)_L \subseteq N_i(L)$; \item[(iii)] if $N_i(L) \subseteq M$ then $N_i(M)_L=N_i(L)$; \item[(iv)] if $k$ is the smallest positive integer such that $N_k(L) \not \subseteq M$ then $N_k(M)_L=N_k(L) \cap M$; and \item[(v)] if $N(L) \subseteq M$ then $N(M)$ acts nilpotently on $L$. \end{itemize} \end{lemma} \begin{proof} \begin{itemize} \item[(i)] This is a straightforward induction proof. \item[(ii)] It is easy to see that this holds for $i=1$. So suppose it holds for $i<k$ ($k \geq 2$). Then there is an $r \in \N$ such that $(N_k(M)_L)^r \subseteq N_k(M)^r \subseteq N_{k-1}(M)$, so $(N_k(M)_L)^r \subseteq N_{k-1}(M)_L \subseteq N_{k-1}(L)$ by the inductive hypothesis. Thus $N_k(M)_L \subseteq N_k(L)$ and the result follows by induction. \item[(iii)] This follows from (i) and (ii). \item[(iv)] Suppose that $k$ is the smallest positive integer such that $N_k(L) \not \subseteq M$. Then $L=N_k(L)+M$ and $N_{k-1}(L) \subseteq M$, so $\phi_k(L) \subseteq M$. Moreover, $N_k^2 \subseteq \phi_k(L) \subseteq M$, by Lemma \ref{l:nat} and \cite{frat}. It follows that $N_k(L) \cap M$ is an ideal of $L$ and hence that $N_k(L) \cap M \subseteq N_k(M)_L$. The reverse inclusion follows as in (ii). \item[(v)] Let $N(L) \subseteq M$. We show that $N(M)$ acts nilpotently on $L$. Suppose not, and let $L=L_0 \dot{+} L_1$ be the Fitting decomposition of $L$ relative to $N(M)$. Then $L_0=M$ and $[N(L),L_1] \subseteq M \cap L_1=0$. Hence $L_1 \subseteq C_L(N(L)) \subseteq N(L)$, giving $L=M$, a contradiction. \end{itemize} \end{proof} \bigskip In general, over a field of characteristic $p>0$, $N(L)$ is not a characteristic ideal of $L$. The best known example is due to Jacobson and first appeared in \cite{sel}; however, it is not solvable. We shall see next that $N(L)$ is characteristic in $L$ whenever $\phi(L)$ is. First we need some lemmas. \begin{lemma}\label{l:ab} Let $L$ be a Lie algebra over a field of characteristic $p>2$, let $I$ be an abelian ideal of $L$ and let $D$ be a derivation of $L$. Then $I+D(I)\subseteq N(L)$. \end{lemma} \begin{proof} This follows easily from \cite[Theorem 1]{mak}. \end{proof} \begin{lemma}\label{l:charfac} Let $I$ be a characteristic ideal of $L$, and let $D$ be a derivation of $L$. Then $\bar{D}:L/I\rightarrow L/I:x+I \mapsto D(x)+I$ is a derivation of $L/I$. \end{lemma} \begin{proof} This is easy to check. \end{proof} \begin{propo}\label{p:phifree} Let $L$ be a $\phi$-free Lie algebra over a field of characteristic $p>2$. Then $N(L)$ is a characteristic ideal of $L$. \end{propo} \begin{proof} Since $L$ is $\phi$-free, $N(L)=A_1\oplus \ldots \oplus A_r$, where $A_1, \ldots, A_r$ are the minimal abelian ideals of $L$, by \cite[Theorem 7.4]{frat}. But $D(A_i)\subseteq N(L)$ for all $D\in$ Der$(L)$ and $i=1, \ldots, r$, whence the result. \end{proof} \begin{coro}\label{c:phifree} Let $L$ be a Lie algebra over a field of characteristic $p>2$, and suppose that $\phi(L)$ is characteristic in $L$. Then $N(L)$ is characteristic in $L$. \end{coro} \begin{proof} This follows easily from Lemma \ref{l:charfac} and Proposition \ref{p:phifree}. \end{proof} \bigskip However, there are solvable Lie algebras $L$ in which $N(L)$ is not a characteristic ideal, as the following example shows. \begin{ex}\label{e:nonchar} Let $L$ be the four-dimensional Lie algebra with basis $x_1$, $x_2$, $x_3$, $x_4$ and non-zero products $[x_4,x_2]=x_1$, $[x_3,x_1]=x_1$ and $[x_3,x_2]=x_2$ over a field $F$ of characteristic $2$. Then $N(L)=Fx_1+Fx_2+Fx_4$ and if $D(x_1)=x_2$, $D(x_2)=0$, $D(x_3)=0$, $D(x_4)=x_3$ then the extension of $D$ to $L$ by linearity is a derivation of $L$. Clearly, $D(N(L))=Fx_2+Fx_3$. Note that $\phi(L)=Fx_1$, so this leaves open the question of whether Proposition \ref{p:phifree} holds over a field of characteristic $2$. \par If we form the split extension $X=Fd\dot{+} L$, where $[d,x]=D(x)$ for all $x\in L$, then $L$ is an ideal of $X$, but $N(L)$ is not. \end{ex} Consequently, we shall need the following result from \cite{nmax}. \begin{propo}\label{p:nil} Let $I$ be a nilpotent subideal of a Lie algebra $L$ over a field $F$. If $F$ has characteristic zero, or has characteristic $p$ and $L$ has no subideal with nilpotency class greater than or equal to $p-1$, then $I \subseteq N$, where $N$ is the nilradical of $L$. \end{propo} Let $L$ be a Lie algebra over a field $F$ and let $U$ be a subalgebra of $L$. We call the largest integer $r$ such that $N_{r}(L)\subseteq U$ the {\em compatibility index} of $U$. As in \cite{gennil}, if $F$ has characteristic $p>0$, we will call $U$ {\em nilregular} if the nilradical of $U$ has nilpotency class less than $p-1$. If $U$ has compatibility index $r$, we say that $U$ is {\em strongly nilregular} if $N_k(U)/N_{k-1}(U)$ has nilpotency class less than $p-1$ for $k=1, \ldots, r$. If $F$ has characteristic zero we regard every subalgebra of $L$ as being nilregular. Then we have the following result. \begin{propo}\label{p:nilregular} If $I$ is a nilregular ideal of $L$ then $N(I)\subseteq N(L)$. \end{propo} \begin{proof} This is \cite[Proposition 2.1]{gennil}. \end{proof} \bigskip We shall call $L$ {\em primitive} if it has a core-free maximal subalgebra. It is said to be {\em primitive of type $1$} if it has a unique minimal ideal that is abelian; since $L$ is solvable this is the only type that can occur here (see \cite{chief}). If $B$ is an ideal of $L$ and $U/B$ is a subalgebra of $L/B$, the {\em centraliser} of $U/B$ in $L$ is $C_L(U/B)=\{x\in L : [x,U]\subseteq B\}$. \begin{propo}\label{p:nmax} Let $L$ be a solvable Lie algebra over a field $F$, and let $M$ be a maximal subalgebra of $L$ of compatibility index $r$. If $M_L$ is strongly nilregular, then $N_i(M)=N_i(M_L)=N_i(L)$ for $i=1,\ldots , r$. \end{propo} \begin{proof} First we show that if $M_L$ is nilregular then $N(M)=N(M_L)=N(L)$. We have that $N(M)$ acts nilpotently on $L$, by Lemma \ref{l:max} (v). Now $L/M_L$ is primitive of type $1$, and so $L/M_L=A/M_L \dot{+} M/M_L$, where $A/M_L$ is the unique minimal ideal of $L/M_L$ and is self-centralising, by \cite[Theorem 1.1]{chief}. Since $L=A+M$, we have that $A+N(M)$ is an ideal of $L$, and so $[A,N(M)]+M_L=[A,A+N(M)]+M_L=M_L$ or $A$, since $A/M_L$ is a chief factor of $L$. \par The former implies that $[A,N(M)] \subseteq M_L$, whence \[ \frac{N(M)+M_L}{M_L} \subseteq C_{L/M_L}\left( \frac{A}{M_L} \right) = \frac{A}{M_L}. \] Thus $N(M) \subseteq A \cap M \subseteq M_L$. \par The latter gives that $[A,N(M)]+M_L=A$. But then an easy induction shows that $A \subseteq A$ (ad\,$N(M))^r +M_L$ for every $r \in \N$, whence $A=M_L$ since $N(M)$ acts nilpotently on $L$. But this is impossible, so $N(M) \subseteq M_L$. \par It follows that $N(M)\subseteq N(M_L)$ and hence $N(M) \subseteq N(L)$, by Proposition \ref{p:nilregular}. Hence $N(M)=N(M_L)=N(L)$. \par So suppose now that $N_k(M)=N_k(M_L)=N_k(L)$ for some $1\leq k<r$. Then $(M/N_k(L))_L=M_L/N_k(M_L)$ is nilregular, so, by the above, \[ N\left(\frac{M}{N_k(M)}\right)=N\left(\frac{M}{N_k(L)}\right) =N\left(\frac{M_L}{N_k(M_L)}\right)=N\left(\frac{L}{N_k(L)}\right), \] whence \[ \frac{N_{k+1}(M)}{N_k(M)}=\frac{N_{k+1}(M_L)}{N_k(M_L)}=\frac{N_{k+1}(L)}{N_k(L)} \] and $N_{k+1}(M)=N_{k+1}(M_L)=N_{k+1}(L)$. The result follows by induction. \end{proof} \bigskip Note that the above result is not true for all maximal subalgebras, as is shown in the next example. \begin{ex}\label{e:nonchar2} Let $X$ be as in Example \ref{e:nonchar}. Then $M=Fd+Fx_1+Fx_2$ and $L$ are both maximal subalgebras of $X$ of compatibility index $1$. However, $N(L) \neq N(X)=Fx_1+Fx_2$, and $M_X=Fx_1+Fx_2$, so $N(M_L)=M_L\neq M=N(M)$. \end{ex} Let ${\mathcal N(k)}$, ${\mathcal N({\leq k)}}$ denote the classes of Lie algebras of nilpotent length $k$ and of nilpotent length $\leq k$ respectively. Of course, over a field of characteristic zero, every Lie algebra $L \in {\mathcal N({\leq 2})}$. However, over a field of characteristic $p>0$ it is easy to construct Lie algebras $L \in {\mathcal N(k)}$ for any $k \in \N$. \par A class ${\mathcal H}$ of finite-dimensional solvable Lie algebras is called a {\em homomorph} if ${\mathcal H}$ contains, along with an algebra $L$, all epimorphic images of $L$. A homomorph ${\mathcal H}$ is called a {\em formation} if $L/A$, $L/B\in {\mathcal H}$ implies that $L/A\cap B \in {\mathcal H}$, where $A$, $B$ are ideals of $L$. A formation ${\mathcal H}$ is said to be {\em saturated} if $L/\phi(L) \in {\mathcal H}$ implies that $L \in {\mathcal H}$. \begin{propo}\label{p:satfor} The class ${\mathcal N({\leq k})}$ is saturated formation for each $k \geq 1$. \end{propo} \begin{proof} It is shown that ${\mathcal N(1)}$ is a saturated formation in \cite[Lemma 3.7]{b-g-h}. Suppose that it holds for $k=r$. Then ${\mathcal N}(\leq r+1)$ is clearly a homomorph. Suppose that $L/A$, $L/B \in {\mathcal N}(\leq r+1)$. Let $S/A=N(L/A)$ and $T/B=N(L/B)$. Then $L/S$, $L/T \in {\mathcal N}(\leq r)$, so $L/S\cap T \in {\mathcal N}(\leq r)$. But there exist $m$, $n\in \N$ such that $S^m\subseteq A$ and $T^n\subseteq B$, so $(S\cap T)^{m+n}\subseteq A\cap B$. Hence $L/A\cap B \in {\mathcal N}(\leq r+1)$, and ${\mathcal N}(\leq r+1)$ is a formation. \par Suppose now that $L/\phi(L) \in {\mathcal N}(\leq r+1)$. Then $N(L/\phi(L))=N(L)/\phi(L)$, by \cite[Theorem 6.1]{frat}, so $L/N(L) \in {\mathcal N}(\leq r)$ and $L \in {\mathcal N}(\leq r+1)$. It follows that ${\mathcal N}(\leq r+1)$ is saturated. \end{proof} \begin{coro}\label{c:satfor} Let $L \in {\mathcal N(k)}$ have more than one minimal ideal. Then there is at least one minimal ideal $A$ of $L$ such that $L/A \in {\mathcal N(k)}$. \end{coro} \begin{proof} If $A_1, \ldots, A_n$ are minimal ideals of $L$, where $n>1$, and $L/A_i \in {\mathcal N({\leq k-1})}$ for all $1 \leq k \leq n$, then $L \in {\mathcal N({\leq k-1})}$, by Proposition \ref{p:satfor}. \end{proof} \begin{propo}\label{p:max} Let $L$ be a solvable Lie algebra over a field $F$. If $M$ is a maximal subalgebra of $L$ for which $M_L$ is strongly nilregular, then $n(M)=n(L)-i$ where $i \in \{0,1\}$. \end{propo} \begin{proof} Let $r$ be the compatibility index of $M$. Then $L=M+N_{r+1}(L)$ and $L/N_{r+1}(L)\cong M/M\cap N_{r+1}(L)=M/N_{r+1}(M)_L$, by Lemma \ref{l:max} (iv). Now $N_r(L)=N_r(M)$ by Proposition \ref{p:max}, so $N_r(M)\subseteq N_{r+1}(M)_L\subseteq N_{r+1}(M)$. It follows from Lemma \ref{l:quot} that $n(M/N_{r+1}(M)_L)=n(M)-r-1$ or $n(M)-r$. Hence $n(L)-r-1=n(M)-r-1$ or $n(M)-r$, which gives the result. \end{proof} \begin{lemma}\label{l:decomp} Let $L \in {\mathcal N}(n)$. Then we can write $L=N_k(L)+U_k$, where $U_k$ is a subalgebra of $L$, $N_k(L)\cap U_k \subseteq \phi(U_k)=\phi_{k+1}(L)\cap U_k$, $U_k\subseteq U_{k-1}$ for each $k=1,\ldots,n$. Moreover, $N(U_k)=N_{k+1}(L)\cap U_k$ for each $k=0,\ldots,n-1$. \end{lemma} \begin{proof} Put $U_0=L$. Then $L=N_1(L)+ U_1$ for some subalgebra $U_1$ of $L$ with $N(L)\cap U_1\subseteq \phi(U_1)$, by \cite[Lemma 4.1]{frat}. Having constructed $U_j$ we construct $U_{j+1}$ such that $U_j=N_{j+1}(L)\cap U_j+U_{j+1}$ and $N_{j+1}(L)\cap U_{j+1}\subseteq \phi(U_{j+1})$, which we can do, by using \cite[Lemma 4.1]{frat} again. Now, it is easy to see inductively that $L=N_k(L)+U_k$, $N_k(L)\cap U_k\subseteq \phi(U_k)$ and $U_k\subseteq U_{k-1}$ for each $k=1,\ldots,n$. Furthermore \begin{align} \frac{\phi_{k+1}(L)\cap U_k}{N_k(L)\cap U_k} & \cong \frac{N_k(L)+\phi_{k+1}(L)\cap U_k}{N_k(L)}= \frac{\phi_{k+1}(L)}{N_k(L)} \nonumber \\ & = \phi \left(\frac{L}{N_k(L)}\right) \cong \phi\left(\frac{U_k}{N_k(L)\cap U_k}\right)= \frac{\phi(U_k)}{N_k(L)\cap U_k}, \nonumber \end{align} so $\phi(U_k)=\phi_{k+1}(L)\cap U_k$. \par Clearly $N_k(L)+N(U_k)\subseteq N_{k+1}(L)$ so $N(U_k)\subseteq N_{k+1}(L)\cap U_k$. But also, there is a natural number $r$ such that \[ (N_{k+1}(L)\cap U_k)^r \subseteq N_k(L)\cap U_k \subseteq \phi(U_k), \] so $N_{k+1}(L)\cap U_k/\phi(U_k)$ is nilpotent. It follows from \cite[Theorem 6.1]{frat} that $N_{k+1}(L)\cap U_k$ is a nilpotent ideal of $U_k$, so $N_{k+1}(L)\cap U_k \subseteq N(U_k)$ and equality results. \end{proof} \bigskip Next we have a fundamental decomposition result. \begin{theor}\label{t:decomp} Let $L \in {\mathcal N}(n)$ if and only if there are nilpotent subalgebras $B_i$ of $L$ for $i=1,\ldots,n$ such that \begin{itemize} \item[(i)] $N_i(L)=B_1+\ldots +B_i$ for $i=1,\ldots, n$, \item[(ii)] $L=B_1+\ldots +B_n$, \item[(iii)] $[B_i,B_j]\subseteq B_i$ for $1\leq i\leq j\leq n$, and \item[(iv)]$N_i(L)\cap U_i\subseteq \phi(U_i)=\phi_{i+1}(L)\cap U_i$, where $U_i= B_{i+1}+\ldots +B_n$, for $i=1,\ldots,n-1$. \end{itemize} \end{theor} \begin{proof} Let $L \in {\mathcal N}(n)$, $U_i$ be as in Lemma \ref{l:decomp} and put $B_i=N(U_{i-1})$, where $U_0=L$. \begin{itemize} \item[(i)] Clearly $B_1=N_1(L)$. Suppose that $B_1+ \ldots +B_k=N_k(L)$. Then \[ B_1+\ldots +B_{k+1}=N_k(L)+N(U_k)=N_k(L)+N_{k+1}(L)\cap U_k=N_{k+1}(L). \] \item[(ii)] $B_1+\ldots +B_n=N_n(L)=L$. \item[(iii)] $[B_i,B_j]=[N(U_i),N(U_j)]\subseteq [N(U_i),U_i] \subseteq N(U_i)=B_i$. \item[(iv)] We have $U_k=U_{k+1}+N_{k+1}(L)\cap U_k=U_{k+1}+N(U_k)=U_{k+1}+B_{k+1}$. Hence $U_i=U_{i+1}+B_{i+1}=U_{i+2}+B_{i+2}+B_{i+1}= \dots = B_{i+1}+\dots +B_n$, and $N_i(L)\cap U_i\subseteq \phi(U_i)$ from Lemma \ref{l:decomp}. \end{itemize} The converse is clear. \end{proof} \section{Extreme Lie Algebras} The Lie algebra $L$ is {\em monolithic} if it has a unique minimal ideal $A$, the {\em monolith} of $L$. If $B$ is an ideal of $L$ and $A/B$ is a minimal ideal of $L/B$ we say that $A/B$ is a {\em chief factor} of $L$. The series \[ \{0\} = A_0 \subset A_1 \subset \ldots \subset A_n=L \] is called a {\em chief series} if $A_i/A_{i-1}$ is a chief factor of $L$ for each $1\leq i\leq n$. \begin{lemma}\label{l:chief} Let $L$ be a Lie algebra such that $N(L)/\phi(L)$ is a chief factor of $L$. Then $L$ has at most one complemented minimal ideal, and if $A$ is one such, then $\phi(L/A)=\phi_2(L)/A$. \end{lemma} \begin{proof} If $\phi(L)=0$, then $N(L)$ is the monolith and the result follows easily from \cite[Theorems 7.3 and 7.4]{frat}. So assume that $\phi(L) \neq 0$ and let $A$ be a complemented minimal ideal of $L$. Then $A \not \subseteq \phi(L)$, so $A+\phi(L)=N(L)$. Let $M$ be a maximal subalgebra of $L$, and suppose that $\phi_2(L) \not \subseteq M$. Then $N(L) \not \subseteq M$, so $L=M+N(L)= M+A+\phi(L)$, giving $M+A=L$. Thus $A \not \subseteq M$. It follows that every maximal subalgebra of $L/A$ contains $\phi_2(L)/A$. Put $T/A= \phi(L/A)$. Then $\phi_2(L) \subseteq T$. \par Let $M/N(L)$ be a maximal subalgebra of $L/N(L)$ and suppose that $(T+N(L))/N(L) \not \subseteq M/N(L)$. Then $M+T+N(L)=M+T=L$. But now $T \not \subseteq M$, so $\phi(L/A) \not \subseteq M$, whence $A \not \subseteq M$. It follows that $L=M+A=M$, a contradiction. Hence $\phi_2(L)/N(L)= \phi(L/N(L)) \supseteq (T+N(L))/N(L)$, which yields that $T= \phi_2(L)$. \par Now suppose that $B$ is another minimal ideal of $L$ with $B \neq A$. Then \[ B \cong (A+B)/A \subseteq N_2(L)/A=N(L/A), \] by the above. It follows that $N_2(L) \subseteq C_L(B)$. Suppose that $B \not \subseteq \phi(L)$. Then, as before, we have that $B+ \phi(L)=N(L)$ and hence that $B \cong N(L)/\phi(L)$. But $C_L(N(L)/\phi(L))=N(L)$, by \cite[Theorem 7.4]{frat}, since $N(L/\phi(L))=N(L)/\phi(L)$. This yields that $N(L)=C_L(B)$, a contradiction. Hence $B \subseteq \phi(L)$ and $L$ has, at most, one complemented minimal ideal. \end{proof} \bigskip We call $L$ {\em extreme} if $N_i(L)/\phi_i(L)$ is a chief factor of $L$ for each $i=1,2, \ldots , n(L).$ \begin{lemma}\label{l:factor} Every factor algebra of an extreme Lie algebra is extreme. \end{lemma} \begin{proof} Let $B$ be an ideal of the extreme Lie algebra $L$. Suppose first that $B \subseteq \phi(L)$. Then $N_i(L/B)/\phi_i(L/B)$ is a chief factor of $L/B$ for each $i$, by Lemma \ref{l:nat}, and $L/B$ is extreme. So suppose that $B \not \subseteq \phi(L)$. Then $\phi(L/B)=\phi_2(L)/B$, by Lemma \ref{l:chief}, and the result again follows. \end{proof} \bigskip We say that the chief factor $A/B$ is {\em complemented} if there is a maximal subalgebra $M$ of $L$ such that $L=A+M$ and $A\cap M=B$. We define $c(L)$ to be the number of complemented chief factors in a chief series for $L$. This is independent of the particular chief series chosen, by \cite[Theorem 2.3]{chief}. \par Let $x \in L$ and let ad\,$x$ be the corresponding inner derivation of $L$. If $F$ has characteristic zero suppose that (ad\,$x)^n = 0$ for some $n$; if $F$ has characteristic $p$ suppose that $x \in I$ where $I$ is a nilpotent ideal of $L$ of class less than $p$. Put \[ \hbox{exp(ad\,}x) = \sum_{r=0}^{\infty} \frac{1}{r!}(\hbox{ad\,}x)^r. \] Then exp(ad\,$x)$ is an automorphism of $L$. We call the group ${\mathcal I}(L)$ generated by all such automorphisms the group of {\em inner automorphisms} of $L$. Two subsets $U, V$ are {\em conjugate in $L$} if $U = \alpha(V)$ for some $\alpha \in {\mathcal I}(L)$. \par Then we have the following characterisation of extreme Lie algebras. \begin{theor}\label{t:extreme} Let $L$ be a solvable Lie algebra. Then the following statements are equivalent: \begin{itemize} \item[(i)] $L$ is extreme; \item[(ii)] $n(L)=m(L)$, the number of conjugacy classes of maximal subalgebras of $L$; \item[(iii)] $n(L)=c(L)$; and \item[(iv)] if $B$ is an ideal of $L$, then $L/B$ has at most one complemented minimal ideal. \end{itemize} \end{theor} \begin{proof} \begin{itemize} \vspace{-.2cm} \item[(i) $\Rightarrow$ (ii)]: Let $L$ be extreme and consider the series \[ 0 \subset \phi_1(L) \subset N_1(L) \subset \ldots \subset \phi_i(L) \subset N_i(L) \subset \ldots. \] There is a unique conjugacy class of maximal subalgebras of $L$ complementing the chief factor $N_i(L)/\phi_i(L)$ for each $i=1,2, \ldots, n(L)$, by \cite{barnes}. But each maximal subalgebra of $L$ must complement one of the complemented chief factors in the above series, and must, therefore, belong to one of these $n(L)$ conjugacy classes. Hence $n(L)=m(L)$. \item[(ii) $\Rightarrow$ (iii)]: We use induction on the dimension of $L$. Suppose that $L$ is a Lie algebra satisfying $n(L)=m(L)$ and assume that the implication holds for Lie algebras of smaller dimension than that of $L$. If $\phi(L) \neq 0$, we have $n(L/\phi(L))=n(L)=m(L)=m(L/\phi(L)$, and so, by induction, $n(L)=n(L/\phi(L))=m(L/\phi(L))=c(L/\phi(L))=c(L)$. \par So suppose that $\phi(L)=0$. Then $N(L)=Asoc(L)$ and each of the $r$ (say) minimal ideals in $Asoc(L)$ is complemented, by \cite[Theorem 7.4 and Lemma 7.2]{frat}. It follows that $n(L)=m(L) \geq m(L/N(L))+r \geq n(L)-1+r$. Hence $r=1$ and, by induction, $c(L)=1+c(L/N(L))=1+n(L/N(L))=n(L)$. \item[(iii) $\Rightarrow$ (i)]: This follows from the fact that there is at least one complemented chief factor $A/B$ satisfying $\phi_i(L) \leq B < A \leq N_i(L)$ for each $i=1,2, \ldots, n(L)$. \item[(i) $\Leftrightarrow$ (iv)]: If $L$ is extreme, then so is $L/B$, by Lemma \ref{l:factor}. Hence $L/B$ has at most one complemented minimal ideal, by Lemma \ref{l:chief}. \par Conversely, suppose that $L$ satisfies (iv) and consider $L/\phi_i(L)$. Since $N_i(L)/\phi_i(L)$ is the direct sum of complemented minimal ideals of $L/\phi_i(L)$, as above, it follows that $N_i(L)/\phi_i(L)$ is a chief factor of $L/\phi_i(L)$. Hence $L$ is extreme. \end{itemize} \end{proof} \bigskip \begin{lemma}\label{l:extreme} Let $L$ be an extreme Lie algebra. \begin{itemize} \item[(i)] If $L$ is nilpotent, then $\dim L=1$. \item[(ii)] If $L \in {\mathcal N}(n)$ then $N_{n-1}(L)=\Gamma_1(L)=L^k$ has codimension one in $L$, for all $k\geq 2$. \end{itemize} \end{lemma} \begin{proof}\begin{itemize} \item[(i)] Since $L$ is nilpotent, $\phi(L)=L^2$, and so $\dim L/L^2=1$. The result follows. \item[(ii)] This follows from Lemma \ref{l:factor} and (i). \end{itemize} \end{proof} \begin{theor}\label{t:extremedecomp} The Lie algebra $L \in {\mathcal N}(n)$ is extreme if and only if $L$ has the decomposition given in Theorem \ref{t:decomp}, $\dim B_n=1$ and $N(U_k)/\phi(U_k)$ is a chief factor of $U_k$ for each $k=0, \ldots,n-1$. \end{theor} \begin{proof} We have that \begin{align} \frac{N_{k+1}(L)}{\phi_{k+1}(L)} & =\frac{N_{k+1}(L)\cap U_k+N_k(L)}{\phi_{k+1}(L)\cap U_k+N_k(L)}\cong \frac{(N_{k+1}(L)\cap U_k+N_k(L))/N_k(L)}{(\phi_{k+1}(L)\cap U_k+N_k(L))/N_k(L)} \nonumber \\ & \cong \frac{N_{k+1}(L)\cap U_k/N_k(L)\cap U_k}{\phi_{k+1}(L)\cap U_k/N_k(L)\cap U_k} \cong \frac{N_{k+1}(L)\cap U_k}{\phi_{k+1}(L)\cap U_k}= \frac{N(U_k)}{\phi(U_k)}. \nonumber \end{align} Also $\dim B_n=1$ by Lemma \ref{l:extreme} (ii). \end{proof} \bigskip If $S$ is a subalgebra of $L$, we will denote by $\overline{S}$ the image of $S$ under the canonical homomorphism from $L$ onto $L/\phi(L)$. We have the following characterisation of those Lie algebras $L \in {\mathcal N({\leq 2})}$ that are extreme, which includes all extreme Lie algebras over a field of characteristic zero. \begin{coro}\label{c:extreme2} Let $L \in {\mathcal N({\leq 2})}$. Then $L$ is extreme if and only if one of the following holds. \begin{itemize} \item[(i)] $\dim L=1$; or \item[(ii)] $\overline{L}=\overline{A} \dot{+} \overline{U}$ where $\overline{A}=\overline{N(L)}$ is the monolith of $\overline{L}$ and $\overline{U}$ is a one-dimensional subalgebra of $\overline{L}$ which acts irreducibly on $\overline{A}$. \end{itemize} \end{coro} \begin{proof} This is just the cases $n=1,2$ in Theorem \ref{t:extremedecomp}. \end{proof} \begin{coro}\label{c:supsolv} Let $L$ be supersolvable. Then $L$ is extreme if and only if one of the following holds. \begin{itemize} \item[(i)] $\dim L=1$; or \item[(ii)] $L/\phi(L)$ is the two-dimensional non-abelian Lie algebra. \end{itemize} \end{coro} \begin{proof} Let $L$ be supersolvable and extreme. Then $\dim \overline{N(L)}=1$ and so $\dim \overline{L}/C_{\overline{L}}(\overline{N(L)})=1$. But $\overline{N(L)}=N(\overline{L})$, so $C_{\overline{L}}(\overline{N(L)}) \subseteq \overline{N(L)}$. It follows that $\dim (L/N(L)) \leq 1$ and $L \in {\mathcal N({\leq 2})}$. But now either $\dim L=1$ or $\overline{L}=\overline{A} \dot{+} \overline{U}$ where $\dim \overline{A}= \dim \overline{U}=1$, by Theorem \ref{t:extremedecomp}. In the latter case $\dim \overline{L}=2$ and $\overline{L}$ cannot be abelian. \par Conversely, if $\dim (L/\phi(L)) \leq 2$ then $L/\phi(L)$ is supersolvable, and so $L$ is supersolvable, by \cite[Theorem 6]{barnes}. Clearly $L$ is also extreme in each of cases (i) and (ii). \end{proof} \begin{ex}\label{e:supsolv} It is easy to check that every three-dimensional Lie algebra as described in Corollary \ref{c:supsolv} has a basis $x, y, z$ with non-zero products $[x,y]=y+z$, $[x,z]=\alpha z$ for some $0 \neq \alpha \in F$. Moreover, no two of these with different values of $\alpha$ are isomorphic. \end{ex} \section{Minimal non-${\mathcal N}$ algebras} If $\mathcal X$ is a class of Lie algebras, we say that $L$ is {\em minimal non-$\mathcal X$} if every proper subalgebra of $L$, but not $L$ itself, belongs to $\mathcal X$. We say that $L$ is {\em minimal non-$\mathcal N$} if it is minimal non-$\mathcal N(\leq k)$ for some $k$; in other words, if its nilpotent length is greater than that of any of its proper subalgebras. Over a field of characteristic zero a Lie algebra can only be minimal non-${\mathcal N(1)}$ and these are described in \cite{nilp}. \begin{lemma}\label{l:cod} Let $L$ be minimal non-${\mathcal N(\leq k-1)}$ and let $M$ be a maximal subalgebra of $L$. Then \begin{itemize} \item[(i)] $L^2=N_{k-1}(L)$ has codimension one in $L$; \item[(ii)] if $N_i(M) \not \subseteq N_{k-1}(L)$ then $N_{k-1}(L) \cap N_i(M)$ has codimension one in $N_i(M)$ for $i=1, \ldots, k-1$; and \item[(iii)] if $N(L) \subseteq M$ then $N(M) \subseteq N_{k-1}(L)$. \end {itemize} \end{lemma} \begin{proof} \begin{itemize} \item[(i)] Let $M$ be a maximal subalgebra of $L$ containing $N_{k-1}(L)$. Since $L/N_{k-1}(L)$ is nilpotent, $M$ is an ideal of $L$ and has codimension one in $L$. But $N_{k-1}(M)_L=N_{k-1}(L)$, by Lemma \ref{l:max} (ii), so, if $N_{k-1}(L) \neq M$, then $M$ has nilpotent length $k$, contradicting the fact that it is minimal non-${\mathcal N(k)}$. Hence $M=N_{k-1}(L)$. \par Now let $M$ be any maximal subalgebra containing $L^2$, so $M$ is an ideal of codimension one in $L$. We have $M \cap N_i(L)=N_i(M)_L$ for each $i=1, \ldots k-1$ by Lemma \ref{l:max} (i) and (ii). It follows that $M \cap N_{k-1}(L) = N_{k-1}(M)_L=M$, so $M \subseteq N_{k-1}(L)$, whence $M=N_{k-1}(L)=L^2$. \item[(ii)]Suppose $N_i(M) \not \subseteq N_{k-1}(L)$. Then $L=N_{k-1}(L)+N_i(M)$, so \[ \frac{L}{N_{k-1}(L)} \cong \frac{N_i(M)}{N_{k-1}(L) \cap N_i(M)}, \] whence the result. \item[(iii)] Let $N(L) \subseteq M$. Then $N(M)$ acts nilpotently on $L$, by Lemma \ref{l:max} (v). If $N(M) \not \subseteq N_{k-1}(L)$ then $L=N_{k-1}(L)+N(M)$ and $L/N_{k-2}(L)$ is nilpotent, a contradiction. \end{itemize} \end{proof} In group theory, every minimal non-${\mathcal N(k)}$ group is extreme, and so a natural question is whether this holds for Lie algebras. We show next that this is `usually' the case for Lie algebras. We call a class ${\mathcal H}$ of Lie algebras a {\em semi-homomorph} if, for all $L \in {\mathcal H}$, \begin{itemize} \item[(i)] $L/N(L) \in {\mathcal H}$; and \item[(ii)] if $A$ is an ideal of $L$ and $A\subseteq \phi(L)$, then $L/A \in {\mathcal H}$. \end{itemize} \begin{lemma}\label{l:crit} Let ${\mathcal H}$ be the class of Lie algebras $L$ in which all maximal subalgebras of $L$ have strongly nilregular cores. Then ${\mathcal H}$ is a semi-homomorph. \end{lemma} \begin{proof} \begin{itemize} \item[(i)] Let $M/N(L)$ be a maximal subalgebra of $L/N(L)$. Clearly we have that $(M/N(L))_{L/N(L)}=M_L/N(L)$ and $M_L/N(L)$ has compatibility index one less than that of $M_L$, $r-1$ say. Then \[ \frac{N_i(M_L/N(L))}{N_{i-1}(M_L/N(L))}\cong \frac{N_{i+1}(M_L)}{N_i(M_L)}, \] which has nilpotency class $<p-1$ for $i=1, \ldots, r-1$, since $M_L$ is strongly nilregular. Hence $M/N(L)$ has a strongly nilregular core. \item[(ii)] Let $M/A$ be a maximal subalgebra of $L/A$. Then $(M/A)_{L/A}=M_L/A$. Also, $N_i(L/A)=N_i(L)/A$, by Lemma \ref{l:nat}, so $M_L/A$ has the same compatibility index as $M_L$, $r$ say. Also $N_i(M)=N_i(M_L)=N_i(L)$ for $i=1, \ldots, r$, by Proposition \ref{p:nmax}. \par We claim that $N_i(M_L/A)=N_i(M_L)/A$ for each $i=1, \ldots, n(M_L)$. Put $K_i/A=N_i(M_L/A)$, so $K_{i+1}/K_i=N(M_L/K_i)$. Then $K_1$ is a subideal of $L$, $A\subseteq K_1\cap \phi(L)$ and $K_1/A$ is nilpotent, so $K_1$ is nilpotent, by \cite[Theorem 6.1]{frat}. It follows that $K_1\subseteq N(M_L)$ and $N(M_L/A)=N(M_L)/A$. Hence \[ K_1 \subseteq N(M_L) \subseteq K_2 \subseteq \ldots \subseteq K_i \subseteq N_i(M_L) \subseteq K_{i+1} \subseteq N_{i+1}(L) \subseteq \ldots \] and $K_{i+1}/K_i\subseteq N_{i+1}(L)/K_i=N_{i+1}(M_L)/K_i$ for $i=1, \ldots, r-1$. It follows that $N_{i+1}(M_L/A)=K_{i+1}/A\subseteq N_{i+1}(L)/A$. The reverse inclusion is clear and the claim is established. \par Now we have that \[ \frac{N_i(M_L/N(L))}{N_{i-1}(M_L/N(L))}\cong \frac{N_i(M_L)}{N_{i-1}(M_L)}, \] which has nilpotency class $<p-1$ for $i=1, \ldots, r$, since $M_L$ is strongly nilregular. Hence $M/N(L)$ has a strongly nilregular core. \end{itemize} \end{proof} A solvable primitive algebra has a unique minimal, self-centralising, ideal $A$ such that $L=A\dot{+} U$ (see \cite{chief}). We shall say that a class of Lie algebras ${\mathcal H}$ has the {\em primitive quotient property} if, for every primitive algebra $L$ in ${\mathcal H}$ with minimal ideal $A$, $L/A$ is minimal non-${\mathcal N}$. \begin{theor}\label{t:crit} Let ${\mathcal H}$ be a semi-homomorph with the primitive quotient property, and let $L \in {\mathcal H}$ be a Lie algebra which is minimal non-${\mathcal N(\leq n)}$. Then \begin{itemize} \item[(i)] $L$ is extreme; and \item[(ii)] $L/N(L)$ is minimal non-${\mathcal N(\leq n-1)}$. \end{itemize} \end{theor} \begin{proof} \begin{itemize} \item[(i)] We use induction on $\dim L$. Suppose first that $\phi(L) \neq 0$. Let $A$ be a minimal ideal of $L$ contained in $\phi(L)$. Then $n(L/A)=n(L)$, so $L/A$ is minimal non-$\mathcal N$. By the inductive hypothesis, $L/A$, and hence $L$ is extreme. So suppose that $\phi(L)=0$. If there are at least two minimal ideals then there is at least one, $A$ say, such that $n(L/A)=n(L)$, by Corollary \ref{c:satfor}. But then $L=A \dot{+} M$ for some maximal subalgebra $M$ of $L$, and $n(M)=n(L/A)=n(L)$, a contradiction. \par Thus there is a unique minimal ideal $A=N(L)$, $L$ is primitive and $n(L/A)=n(L)-1$. Since ${\mathcal H}$ has the primitive quotient property, $L/A$ is minimal non-${\mathcal N}$. Moreover, since ${\mathcal H}$ is a semi-homomorph, $L/A$, and thus $L$, is extreme, by induction. \item[(ii)] Consider the series \[ 0 \subset \phi_1(L) \subset N_1(L) \subset \ldots \subset \phi_i(L) \subset N_i(L) \subset \ldots, \] and let $M$ be a maximal subalgebra of $L$ containing $N_1(L)$. Then $M$ must complement one of the complemented chief factors $N_k(L)/\phi_k(L)$ for some $2 \leq k \leq n$ in the above series. But then $L=N_k(L)+M$, $M \cap N_k(L)=\phi_k(L)$ and $N_i(M)=N_i(L)$ for $i=1, \ldots, k-1$, by Lemma \ref{l:max}. Thus $n(M)-k+1=n(M/N_{k-1}(L))=n(L/N_k(L))=n(L)-k$, whence $n(M/N_1(L))=n(M)-1=n(L)-2=n(L/N_1(L))-1$ and $L/N_1(L)$ is minimal non-${\mathcal N(n-1)}$. \end{itemize} \end{proof} \begin{coro}\label{c:crit} Let $L$ be a Lie algebra in which all maximal subalgebras have strongly nilregular cores and which is minimal non-${\mathcal N(\leq n)}$. Then $L$ is extreme and $L/N(L)$ is minimal non-${\mathcal N(\leq n-1)}$. \end{coro} \begin{proof} Let ${\mathcal H}$ be the class of Lie algebras whose maximal subalgebras have strongly nilregular cores. Then ${\mathcal H}$ is a semi-homomorph, by Lemma \ref{l:crit}.Let $L$ be a primitive algebra in ${\mathcal H}$ with minimal ideal $A=N(L)$, and let $M$ be a maximal subalgebra containing $A$. Then $N(M)=A$ by Proposition \ref{p:nmax}, so $L/A$ is minimal non-${\mathcal N(\leq n-1)}$ and ${\mathcal H}$ satisfies the quotient primitive property. The result now follows from Theorem \ref{t:crit}. \end{proof} \bigskip A Lie algebra $L$ is called an $A$-algebra if all of its nilpotent subalgebras are abelian. These arise in the study of constant Yang–Mills potentials and in relation to the problem of describing residually finite varieties. The structure of solvable Lie $A$-algebras was studied in some detail in \cite{Aalg}. In the case of an $A$-algebra the lower nilpotent series and the derived series coincide (\cite[Lemma 2.3]{Aalg}), and so the terms ``derived length" and ``nilpotent length" are identical. \begin{coro}\label{c:Acrit} If $L$ is an $A$-algebra which is minimal non-${\mathcal N(\leq n)}$, then $L$ is extreme and $L/N(L)$ is minimal non-${\mathcal N(\leq n-1)}$. \end{coro} \begin{proof} Let ${\mathcal H}$ be the class of $A$-algebras. Then ${\mathcal H}$ is a semi-homomorph, by \cite[Lemma 2.1 (iii)]{Aalg}. Let $L \in {\mathcal H}$ be primitive with minimal ideal $A=N(L)$ and let $M$ be a maximal subalgebra containing $A$. Then $N(M)$ is abelian and so $[N(M),A]=0$, giving $N(M) \subseteq C_L(A)=A$. It follows that $N(M)=A$ and so ${\mathcal H}$ has the primitive quotient property. \end{proof} \begin{coro}\label{c:2crit} Let $L$ be minimal non-${\mathcal N(\leq 2)}$ and have solvability index $\leq 3$. Then $L$ is extreme and $L/N(L)$ is minimal non-${\mathcal N(\leq 1)}$. \end{coro} \begin{proof} Let ${\mathcal H}$ be the class of Lie algebras of solvability index $\leq 3$. Then ${\mathcal H}$ is clearly a semi-homomorph. Let $L \in {\mathcal H}$ be primitive with minimal ideal $A=N(L)$ and let $M$ be a maximal subalgebra containing $A$. If $L$ has solvability index $\leq 2$ it is clear that $L/A$ is minimal non-${\mathcal N}$, so assume that $L$ has index $3$. \par We have $L^{(1)}=N_2(L)$, by Lemma \ref{l:cod} (i) and $L^{(2)} \subseteq N(L)=A$, so $N_2(L)/A$ is abelian. Now $N(M) \subseteq N_2(L)$ by Lemma \ref{l:cod} (iii), and so $$[N_2(L),N(M)] \subseteq N_2(L)^2 \subseteq A \subseteq M.$$ Then either $M \neq N_2(L)$ or $L=N_2(L)+M$, in which case $N(M)$ is an ideal of $L$ and $N(M)=A$. In either case $M/A$ is nilpotent and so ${\mathcal H}$ has the primitive quotient property. \end{proof} \begin{ex} Note that there are extreme Lie algebras which are not minimal non-${\mathcal N}$. For example, let $L$ be the Lie algebra over any field $F$ with basis $x, y, z$ with non-zero products $[x,y]=y+z$, $[x,z]=z$. Then $\phi(L)=Fz$ and $N(L)=Fy+Fz$, so $L$ is extreme. However, if $M=Fx+Fz$ then $n(M)=2=n(L)$. \end{ex} \medskip Next we seek to characterise the algebras considered in Corollaries \ref{c:Acrit} and \ref{c:2crit}. We can characterise the $A$-algebras that are also minimal non-${\mathcal N}$ as follows. \begin{theor}\label{Aalg} Let $L$ be a Lie $A$-algebra of derived length $n+1$ over a field $F$. Then $L$ is minimal non-${\mathcal N}$ if and only if the following hold. \begin{itemize} \item[(i)] $L= A_n \dot{+} A_{n-1} \ldots \dot{+} A_1 \dot{+} Fx$ where $A_i$ is an abelian subalgebra of $L$ for each $1 \leq i \leq n$; \item[(ii)] $L^{(i)}= A_n \dot{+} A_{n-1} \ldots \dot{+} A_i$ for each $1 \leq i \leq n$; \item[(iii)] $[A_i,A_j] \subseteq A_j$ for $j>i$; \item[(iv)] $A_i$ is an irreducible $L/L^{(i+1)}$-module for each $0 \leq i \leq n$; and \item[(v)] $N_{n-i+1}(L)=L^{(i)}$ for each $0 \leq i \leq n$. \end{itemize} \end{theor} \begin{proof} Suppose first that $L$ is minimal non-${\mathcal N}$. Since $L$ is an $A$-algebra of derived length $n+1$, $L= A_n \dot{+} A_{n-1} \ldots \dot{+} A_1 \dot{+} A_0$ and $L^{(i)}= A_n \dot{+} A_{n-1} \ldots \dot{+} A_i$, where $A_i$ is an abelian subalgebra of $L$ for each $0 \leq i \leq n$, by \cite[Corollary 3.2]{Aalg}. But $\dim A_0=1$, by Lemma \ref{l:cod} (i), so $A_0 =Fx$ for some $x \in L$. This gives (i) and (ii). The decomposition in (i) follows from the splitting of a Lie $A$-algebra over each term in its derived series (\cite[Theorem 3.1]{Aalg}), so $L=A_n \dot{+} B_n$, where $A_n=L^{(n)}$, $B_n=A_{n-1} \dot{+} B_{n-1}$, where $A_{n-1}=B_n^{(n-1)}$, and so on. But now $[A_i,A_j] \subseteq L^{(j)} \cap B_{j+1} \subseteq A_j$ if $j>i$, giving (iii). \par A straightforward induction argument shows that $A_i \subseteq N_{n-i-1}$ for $0 \leq i \leq n$. We now establish (iv) and (v) by induction on $n$. Then (iv) clearly holds for $i=0$, and (v) holds for $i=0$ by Lemma \ref{l:cod} (i), so suppose that they hold for all $i \geq k$ ($k \geq 0$). Then $N_{n-k+1}(L)=L^{(k)}=A_n \dot{+} \ldots \dot{+} A_k$ and $L^{(k+1)}=A_n \dot{+} \ldots \dot{+} A_{k+1} \subseteq N_{n-k}(L)$. It follows that $N_{n-k}(L)=L^{(k+1)}$ by the irreducibility of $A_k$ and the fact that $N_{n-k}(L) \neq N_{n-k+1}(L)$ (since $L$ has nilpotent length $n+1$). \par Also $\phi_{n-k}(L) \subseteq L^{(k+1)}=N_{n-k}(L)$, and $N_{n-k}(L)/\phi_{n-k}(L)$ is irreducible, since $L$ is extreme, by Theorem \ref{t:crit}. Hence $M=\phi_{n-k}(L) \dot{+} A_{k} \dot{+} \ldots \dot{+} A_1 \dot{+} Fx$ is a maximal subalgebra of $$A_{k+1} \dot{+} A_k \dot{+} \ldots \dot{+} A_1 \dot{+} Fx \cong L/L^{(k+2)} = L/N_{n-k-1}(L).$$ But $L/N_{n-k-1}(L)$ is minimal non-${\mathcal N}(k+1)$, by Theorem \ref{t:crit} (ii), so $M \in {\mathcal N}(k+1)$. It follows that $N(M)= \phi_{n-k}(L) \dot{+} A_k$, so $[\phi_{n-k}(L),A_k]=0$. But $A_k$ is a Cartan subalgebra of $A_{k+1} \dot{+} A_k$, by \cite[Theorem 3.1]{Aalg}, so $\phi_{n-k}(L)=0$ and $A_k$ is an irreducible $L/L^{(k+1)}$-module. This establishes (iv) and (v). \par Conversely, suppose that (i)-(v) hold and let $M$ be a maximal subalgebra of $L$. Clearly $L \in {\mathcal N}(n+1)$. Let $i$ be the smallest integer such that $L^{(i)} \not \subseteq M$. Then $L^{(i)}/L^{(i+1)}$ is a minimal ideal of $L/L^{(i+1)}$, so $M/L^{(i+1)}$ complements $L^{(i)}/L^{(i+1)}$ in $L/L^{(i+1)}$. Hence $M^{(i)} \subseteq M \cap L^{(i)} \subseteq L^{(i+1)}$. But $n(L^{(i+1)})<n-i+1$ by (v). Hence $n(M)<n+1$ and $L$ is minimal non-${\mathcal N}$. \end{proof} \bigskip Recall the following result from \cite{aN-by-A}. \begin{theor}(\cite[Theorem 4]{aN-by-A})\label{t:aN-by-A} Let $L$ be solvable and $\phi$-free. Then $L$ is minimal non-(nilpotent-by-abelian) if and only if $F$ has characteristic $p>0$ and $L=A\dot{+} B$, where $A$ is the unique minimal ideal of $L$, $\dim A \geq 2$, $A^2=0$, and either $B=M\dot{+} Fx$, where $M$ is a minimal ideal of $B$ such that $M^2=0$ (type I), or $B$ is the three-dimnsional Heisenberg algebra (type II). Moreover, if $p \geq 3$ then $\dim A$ is divisible by $p$. \end{theor} \bigskip Then we have the following characterisation of the algebras of solvability index $3$ which are minimal non-${\mathcal N}(2)$. \begin{theor}\label{t:solv3} Let $L$ be a solvable Lie algebra of solvability index $3$. Then $L$ is minimal non-${\mathcal N}(2)$ if and only if it is minimal non-(nilpotent-by-abelian) of type I. \end{theor} \begin{proof} Let $L$ be minimal non-${\mathcal N}(2)$, and denote the image of a subalgebra $S$ of $L$ under the natural homomorphism onto $L/\phi(L)$ by $\overline{L}$. Then $\overline{L}=\overline{N(L)}\dot{+} \overline{U}$ where $U$ is a subalgebra of $L$, by \cite[Theorems 7.3 and 7.4]{frat}. Moreover, $\overline{U}= \overline{A}\dot{+}F\overline{x}$ where $\overline{A}$ is abelian, $N_2(L)=N(L)+A$ and $L/N(L)$ is minimal non-nilpotent, by Corollary \ref{c:2crit}. It follows from \cite[Theorem 2.1]{nilp} that $N_2(L)/N(L)$ is irreducible. Now $U$ is a maximal subalgebra of $L$, so $U \in {\mathcal N}(2)$, which yields that $N(U)=A$ and $U$ is nilpotent-by-abelian. Let $M$ be any maximal subalgebra of $L$. Then either $M\cong U$ or $N(L)\subseteq M$. Suppose that $N(L)\subseteq M$. Then $M=N_2(L)$ or $L=N(L)+Fx$. In either case $M$ is nilpotent-by-abelian. Hence $L$ is minimal non-(nilpotent-by-abelian). It is of type I, since otherwise $L \in {\mathcal N}(2)$. \par Conversely, suppose that $L$ is minimal non-(nilpotent-by-abelian) of type I. Then $L$ has solvability index $3$ and the maximal subalgebras of $L$ are nilpotent-by-abelian, as in the paragraph above. Clearly $L$ itself is not nilpotent-by-abelian. \end{proof} \bigskip Lie algebras as described in Theorem \ref{t:solv3} do exist over every field of characteristic $p>0$, as is shown in \cite{aN-by-A}; over an algebraically closed field they are minimal non-supersolvable (\cite[Theorem 5]{aN-by-A}). Finally we show that a homomorphic image of a minimal non-${\mathcal N}$ Lie algebra is minimal non-${\mathcal N}$ if it has a complemented minimal ideal. \begin{theor}\label{t:chief} If $L$ is a minimal non-${\mathcal N}$ Lie algebra in which all maximal subalgebras have nilregular cores, and $A/B$ is a complemented chief factor of $L$, then $L/B$ is minimal non-${\mathcal N}$. \end{theor} \begin{proof} Suppose there is a chief series of $L$ through $A$ and $B$ in which $A/B$ is the $k$th complemented factor, where $1 \leq k \leq c(L)$. If $k=1$, then $B \subseteq \phi(L)$ and so $n(L/B)=n(L)$, in which case $L/B$ is minimal non-${\mathcal N}$ because $L$ is. \par So let $k >1$ and assume that the theorem holds for the $(k-1)$th complemented factor $C/D$. Without loss of generality we may assume that $D=0$. Then $A/B$ is the second complemented factor in some chief series of $L$. It follows from Theorem \ref{t:crit} that $L$ and $L/N(L)$ are extreme, and so every chief series of $L$ has only one complemented chief factor below $N(L)$, by Theorem \ref{t:extreme}. If $N(L) \cap A \not \subseteq B$ then $N(L) \cap A/N(L) \cap B$ is a complemented chief factor of $L$ and $L$ would have a chief series with two complemented chief factors below $N(L)$, a contradiction. Hence $N(L) \cap A \subseteq B$. \par Let $n=n(L)$. Then $L/A$ and $L/B$ are both extreme, by Lemma \ref{l:factor} and so \[ n(L/B)=c(L/B)=n-1 \hspace{2mm} \hbox{ and } \hspace{2mm} n(L/A)=c(L/A)=n-2, \] by Theorem \ref{t:extreme}. Let $M/B$ be a maximal subalgebra of $L/B$. We have $B \not \subseteq \phi(L)$ and $N(L)/\phi(L)$ is a chief factor of $L$, so $M \supseteq \phi(L)+B \supseteq N(L)$. But $L/N(L)$ is minimal non-${\mathcal N}$, by Theorem \ref{t:crit}, so $n(M/N(L)) \leq n-2$. It follows that $N_{n-2}(M) \subseteq N(L) \cap A \subseteq B$, whence $n(M/B) \leq n-2 < n(L/B)$ and $L/B$ is minimal non-${\mathcal N}$, as claimed. \end{proof} \bigskip

https://arxiv.org/abs/2202.04463

Dual involutions in finite Coxeter groups

There is a well-known classification of conjugacy classes of involutions in finite Coxeter groups, in terms of subsets of nodes of their Coxeter graphs. In many cases, the product of an involution with the longest element is again an involution. We identify the conjugacy class of this product involution in terms of said classification.

\subsubsection*{Overview} Arguably, the most straight-forward path from the above questions to the below results would be via elementary brute force computations. However, this seems both tedious and unenlightening. We leave it as an exercise for the reader to verify the results that we have tabulated for the dihedral groups \(W(G_2(n))\) in this way, and also our results for two specific involutions in \(W(D_{2k})\) (see details at the end of \cref{sec:folding}). All other results can be obtained more elegantly using the following ideas, spelled out in more detail in the body of the paper. \begin{idea}\label{idea:eigenspaces} View \(W(\Sigma)\) a Euclidean reflection group as above, and consider the dimensions of the eigenspaces of the involutions. \end{idea} For a standard involution \(\standardinvolution{I}\), the dimension of the \(-1\)-eigenspace is equal to \(\cardinality{I}\), the number of nodes of \(I\), and the dimension of the \(+1\)-eigenspace is equal to \(\cardinality{\Sigma}-\cardinality{I}\). Multiplication with \({w_o} = -1\) interchanges these two eigenspaces. So any subgraph \(J\) in \cref{question1} must have \(\cardinality{\Sigma}-\cardinality{I}\) nodes. In many cases, this alone determines the conjugacy class of \({w_o} \standardinvolution{I}\). For example, \cref{question1} can be completely answered for \(W(H_3)\) and \(W(H_4)\) in this way. In other cases, this yields at least partial information. For example, in \(W(BC_5)\) we have two non-conjugate standard involutions defined by subgraphs with a single vertex, and two standard involutions defined by subgraphs with four vertices: \begin{align*} c_{\{1\}} & = c_{\dynkin C{*oooo}} & c_{\{5\}} & = c_{\dynkin C{oooo*}} \\ c_{\{1,3,4,5\}} & = c_{\dynkin C{*o***}} & c_{\{2,3,4,5\}} & = c_{\dynkin C{o****}} \end{align*} Multiplication with \({w_o}\) must send each conjugacy class of an involution in the first line to a conjugacy class of an involution in the second line, but further analysis is needed to resolve the ambiguity. \begin{idea}[sub-root systems]\label{example:sub-root-systems} Refine \cref{idea:eigenspaces} by considering not just the dimensions of eigenspaces but the closed sub-root systems they contain. \end{idea} For example, consider \(W(BC_5)\) as the Euclidean reflection group determined by the root system \(C_5\). The fixed-point space of \(\standardinvolution{\{5\}}\) contains the sub-root system generated by the simple roots $\alpha_1,\alpha_2,\alpha_3$, a closed sub-root system of type $A_3$. So the \(-1\)-eigenspace of (any conjugate of) \(-\standardinvolution{\{5\}}\) must also contain a sub-root system of type $A_3$. Using the Borel--de Siebenthal classification of closed sub-root systems, we can therefore rule out one of the two candidate representatives of \(-\standardinvolution{\{5\}}\) from above, and find that \(-\standardinvolution{\{5\}} \sim \standardinvolution{\{2,3,4,5\}}\), resolving the ambiguity. Similar arguments are used in \cref{sec:E7E8} to resolve all ambiguities for \(W(E_7)\) and \(W(E_8)\). \begin{idea}\label{idea:folding} Refine \cref{idea:eigenspaces} using „folding“. \end{idea} Following Steinberg \cite{steinberg:endomorphisms}, we can “fold” the root system of type \(A_{10}\) onto \(C_{5}\) and obtain an embedding \(\iota\colon W(BC_5)\hookrightarrowW(A_{10})\). This embedding sends the longest element of \(W(BC_5)\) to the longest element of \(W(A_{10})\), and induces a bijection between the set of conjucagy classes of involutions in \(W(BC_5)\) and the set of conjugacy classes of involutions in \(\Weyl_o(A_{10})\). We can therefore consider, for each involution \(c\in W(BC_5)\), not just the dimensions of the eigenspaces of \(c\) but also the dimensions of the eigenspaces of \(\iota c\). This yields an alternative proof of the above claim that \(-\standardinvolution{\{5\}} \sim \standardinvolution{\{2,3,4,5\}}\). This approach is expanded in \cref{sec:folding} to complete the results for the classical families \(A_n\), \(BC_n\), \(D_n\), and for \(F_4\) and \(E_6\). \subsubsection*{Acknowledgements} The author has previously discussed some of these ideas on mathoverflow \cite{mathoverflow:collected}. The results presented are motivated by previous results on the real K-theory of certain homogeneous spaces \cite{hz:exterior}, which they make more explicit. \section{Calculations} In the following, we upgrade the Coxeter graph \(\Sigma\) to a Dynkin diagram without changing notation. For most finite Coxeter graphs, there is a unique Dynkin diagram that has the same underlying Coxeter graph, but for the Coxeter graph of type \(BC_n\), we can choose between the Dynkin diagrams \(B_n\) and \(C_n\). We view \(W(\Sigma)\) as the finite Euclidean reflection group determined by a root system of Dynkin type \(\Sigma\). Thus, \(W(\Sigma)\) is a subgroup of the isometry group of a Euclidean vector space \((V, \innerproduct{-}{-})\). Each node of of the Dynkin diagram \(\Sigma\) corresponds to a simple root in \(V\), and we use the same notation \(\Sigma \subset V\) for this set of roots. Each node of \(\Sigma\) also corresponds to a generator of \(W(\Sigma)\), namely to the reflection in the hyperplane orthogonal to the corresponding root. \subsection{Dimensions of eigenspaces} For an arbitrary element \(w\inW(\Sigma)\), we denote by \(\dim^\pm(w)\) the dimensions of the \(\pm1\)-eigenspaces of \(w\). Recall that \({w_o}\) maps the set of simple roots \(\Sigma\subset V\) to the set \(-\Sigma\subset V\), so that \(-{w_o}\) is an automorphism of the set \(\Sigma\), and more precisely an automorphism of the Dynkin diagram. More generally, any subset of roots \(I\subset \Sigma\) is permuted by the involution \(-\longestparabolic{I}\). \begin{prop}\label{lem:minusdim} The dimensions of eigenspaces of various involutions can be obtained by counting orbits, as follows. \begin{enumerate}[(1)] \item For any set of simple roots \(I\subset\Sigma\), \begin{flalign*} &\hspace{4em}\dim^-(\longestparabolic{I}) = \cardinality{\{\text{orbits of } -\longestparabolic{I} \text{ on } I\}}& \end{flalign*} \item For any set of simple roots \(I\subset\Sigma\) such that \(-{w_o} I = I\), we have \(\longestparabolic{I} \in \Weyl_o(\Sigma)\), and \begin{flalign*} &\hspace{4em}\dim^-({w_o} \longestparabolic{I}) = \cardinality{\{\text{non-trivial orbits of } {w_o} \longestparabolic{I} \text{ on } I \}} & \\ &\hspace{10em} + \cardinality{\{\text{orbits of } -{w_o} \text{ on } \Sigma\setminus I\}} \end{flalign*} \item More generally, suppose \(\sigma\) is an involution isometry of \((V,\innerproduct{-}{-})\) such that \(\sigma(\Sigma) = \Sigma\). Assume further that \(\sigma(I) = I\), and that the restrictions of \(\sigma\) and \(-\longestparabolic{I}\) to \(I\) commute. Then \(\sigma\) and \(-\longestparabolic{I}\) commute on all of \(V\), and \begin{flalign*} &\hspace{4em}\dim^-(\sigma(-\longestparabolic{I})) = \cardinality{\{\text{non-trivial orbits of } \sigma(-\longestparabolic{I}) \text{ on } I \}} & \\ &\hspace{11em} + \cardinality{\{\text{orbits of } \sigma \text{ on } \Sigma\setminus I\}} \end{flalign*} \end{enumerate} \end{prop} \begin{proof} (1) Decompose the ambient vector space as \(V = \mathbb{R} I \perp (\mathbb{R} I)^\perp\). (Explicitly, \(\mathbb{R} I\) is the span of the simple roots \(\alpha \in I\), and \((\mathbb{R} I)^\perp\) is the span of the fundamental weights \(\omega_\beta\) with \(\beta \in \Sigma\setminus I\).) Any element of \(\Weylparabolic{I}\) maps \(\mathbb{R} I\) to \(\mathbb{R} I\) and restricts to the identity on \((\mathbb{R} I)^\perp\), so the claim follows. For (3), note that by assumption \(\sigma\) also maps \(\mathbb{R} I\) to \(\mathbb{R} I\), and hence, as an isometry, also maps \((\mathbb{R} I)^\perp\) to \((\mathbb{R} I)^\perp\). The restrictions of \(\sigma\) and \(\longestparabolic{I}\) to \(\mathbb{R} I\) commute by assumption, and their restrictions to \((\mathbb{R} I)^\perp\) commute in any case. For (2), it suffices to check that \(-{w_o}\) satisfies all assumptions in (3). Here, the main point is that the condition \(-{w_o} I = I\) already implies that \(-{w_o}\) and \(-\longestparabolic{I}\) commute on \(I\). This is trivial in the cases when \({w_o} = -\mathrm{id}\), listed in \cref{fact:central}. In the remaining cases, this can easily be verified on a case-by-case basis. Note that the action of \(-\longestparabolic{I}\) on a root \(\alpha\in I\) is determined by the connected component of \(I\) that contains \(\alpha\), and that \(-{w_o}\) maps this component isomorphically to the component containing \(-\longestparabolic{I}\alpha\). \end{proof} Part~(1) gives, in particular, the dimensions of eigenspaces of \({w_o} = \longestparabolic{\Sigma}\). As we have already noted in \cref{fact:central}, \({w_o}\) is central if and only if \(\dim^-({w_o}) = \dim V\), and in this case it is itself a standard involution. In the remaining cases, we find that \({w_o}\) is conjugate to (one or more) standard involutions as follows: \newcommand{\mathllap{\textcolor{white}{\rule{1em}{1ex}}}}{\mathllap{\textcolor{white}{\rule{1em}{1ex}}}} \begin{align*} W(A_n)&\colon {w_o} = \dynkin A{*****.x}\mathllap{\textcolor{white}{\rule{1em}{1ex}}} \sim \dynkin A{*o*o*.x}\mathllap{\textcolor{white}{\rule{1em}{1ex}}} = c_{\{1,3,5,\dots\}}\quad (n\geq 2)\\ W(D_{2m+1})&\colon {w_o} = \dynkin D{*****.***} \sim \dynkin D{o****.***} = c_{\{2,3,4,\dots,2m+1\}}\\ W(E_6)& \colon {w_o} = \dynkin E{******} \sim \dynkin E{o****o} = c_{\{2,3,4,5\}}\\ W(G_2(2m+1)) & \colon {w_o} = \dynkin[gonality=k] G{**} \sim \dynkin[gonality={k}] G{*o} = c_{\{1\}} \sim \dynkin[gonality={k}] G{o*} = c_{\{2\}} \quad (k := 2m+1) \end{align*} Indeed, in in each case, there is a unique conjugacy class of involutions \(\tau\) for which \(\dim^-(\tau)\) is maximal. More generally, we can easily find a standard involution \(\standardinvolution{I}\) conjugate to the involution \(\longestparabolic{H}\) associated with any given subdiagram \(H\subset \Sigma\). Indeed, using the above conjugacy relations, we can find a subdiagram \(I\subset H\) such that \(\standardinvolution{I} \sim \longestparabolic{H}\) in \(\Weylparabolic{H}\), and then a fortiori \(\standardinvolution{I}\sim\longestparabolic{H}\) in \(W(\Sigma)\). \subsection{Sub-root systems} \label{sec:E7E8} \label{sec:subroots} For the Borel--de Siebenthal classification of closed sub-root systems, see for example \cite[\S\,12]{kane}. The only other observation we need is: \begin{observation}\label{obs:subroots} The \(+1\)-eigenspaces of a standard involution \(\standardinvolution{I}\) contains a closed sub-root system whose Dynkin graph is the subgraph of \(\Sigma\) spanned by all nodes that are not connected by an edge to the nodes in \(I\). The \(-1\)-eigenspaces of a standard involution \(\standardinvolution{I}\) contains a closed sub-root system with Dynkin graph~\(I\). \end{observation} \begin{proof} The \(-1\)-eigenspace of \(\standardinvolution{I}\) is spanned by the roots in \(I\). The \(+1\)-eigenspace is spanned by the \emph{fundamental weights} corresponding to roots in \(\Sigma\setminus I\), and thus does not necessarily contain a sub-root system of type \(\Sigma\setminus I\). However, all nodes of \(\Sigma\) not connected to \(I\) correspond to simple roots that are orthogonal to the roots in \(I\), and hence lie in the fixed spaces of \(\standardinvolution{I}\). \end{proof} For example, for $E_7$, the dimensions of eigenspaces only leave one ambiguity, concerning the lines (4) and (5) in the results below. By the above observation, the fixed space of the involution on the left in line (5) contains a sub-root system of type \(A_2\) (generated by the simple roots \(\alpha_1\) and \(\alpha_2)\). Of the involutions on the right of lines (4) and (5), only the involution in line (5) has a \(-1\)-eigenspace that can contain a sub-root system of type \(A_2\). Similary, for $E_8$, the dimensions of eigenspaces only leave one ambiguity, concerning the lines (5) and (6) in the results below. By the observation above, both the \(+1\)-eigenspace and the \(-1\)-eigenspace of the involution in line (6) contains a sub-root system of type \(A_2\), but the \(-1\)-eigenspaces of the involutions in line (5) do not. \subsection{Folding} \label{sec:folding} Consider an isometry \(\sigma\) of \((V,\innerproduct{-}{-})\) that maps \(\Sigma\) to \(\Sigma\). We do \emph{not} assume that simple roots in the same orbit are necessarily orthogonal. Steinberg shows that there is a ``folded'' root system \(\Sigma^\sigma\) in the fixed point space \(V^\sigma\), with one node for every \(\sigma\)-orbit in \(\Sigma\) \cite[1.32\,(b) and surrounding discussion]{steinberg:endomorphisms}. It comes with a canonical surjection of Dynkin diagrams \(\pi\colon \Sigma \twoheadrightarrow \Sigma^\sigma\) and an associated inclusion of Weyl groups \renewcommand{\Weyl^\sigma}{W(\Sigma^\sigma)} \begin{equation}\label{eq:folding-subgroup} \iota\colon \Weyl^\sigma \hookrightarrow W(\Sigma) \end{equation} The image of \(\iota\) is precisely the subgroup of elements of \(W(\Sigma)\) that are invariant under conjugation by \(\sigma\). In particular, for \(\sigma = -{w_o}\), we have \(\iota \Weyl^\sigma = \Weyl_o(\Sigma)\). Given a subdiagram \(I\subset \Sigma\) invariant under \(\sigma\), we write \(\folded{I}\) for the associated ``folded'' subdiagram \(\pi(I)\subset \Sigma^\sigma\). \begin{prop}\label{prop:folding-longest-words} Given a subset \(I\subset \Sigma\) such that \(\sigma(I) = I\), and such that the restrictions of \(\sigma\) and \(-\longestparabolic{I}\) to \(I\) agree, we have \(\iota(\longestparabolic{\folded{I}}) = \longestparabolic{I}\). \end{prop} \begin{proof} Observe the following: \begin{compactitem} \item[(1)] \(\longestparabolic{I}\) is the unique element of \(W(\Sigma)\) satisfying \(\longestparabolic{I}(I) = -I\);\\ \(\longestparabolic{\folded{I}}\) is the unique elmenent of \(\Weyl^\sigma\) satisfying \(\longestparabolic{\folded{I}}(\folded{I}) = -\folded{I}\). \item[(2)] By \cite[1.32(a)]{steinberg:endomorphisms}, it suffices to show that \(\longestparabolic{I}\) agrees with \(\longestparabolic{\folded{I}}\) when restricted to \(V^\sigma\). \item [(3)] By assumption, \(\sigma\) and \(\longestparabolic{I}\) commute (compare proof of \cref{lem:minusdim}\,(3)). \end{compactitem} Combining (1) and (2), we see that it suffices to show that \(\longestparabolic{I}(\folded{I}) = -\folded{I}\). Given a simple root \(\alpha\in I\), let us write the \(\sigma\)-orbit of \(\alpha\) as \(\{\alpha, \sigma\alpha, \dots, \sigma^{l-1}\alpha\}\), for some minimal integer~\(l\). The associated simple root \(a_{[\alpha]} \in \folded{\Sigma}\) is of the form \[ a_{[\alpha]} = r(\alpha + \sigma\alpha + \sigma^2 \alpha + \dots + \sigma^{l-1} \alpha) \] for some real \(r>0\) \cite[above Corollary~1.30]{steinberg:endomorphisms}. As \(-\longestparabolic{I}\) sends \(\sigma\)-orbits to \(\sigma\)-orbits, by (3), we find that \begin{align*} -\longestparabolic{I}(a_{[\alpha]}) &= -\longestparabolic{I}(r\alpha + \sigma\alpha + \sigma^2\alpha + \dots + \sigma^{l-1} \alpha) \\ &= r(-\longestparabolic{I}\alpha + \sigma(-\longestparabolic{I}\alpha) + \sigma^2(-\longestparabolic{I}\alpha) + \dots + \sigma^{l-1}(-\longestparabolic{I}\alpha))\\ &= \tfrac{r}{s} a_{[-\longestparabolic{I}\alpha]} \end{align*} for some real \(s>0\). As a (reduced) root system cannot contain proper multiples of its roots, it follows that \[ -\longestparabolic{I}(a_{[\alpha]}) = a_{[-\longestparabolic{I}\alpha]}. \] As \(-\longestparabolic{I}(\alpha) \in I\), this shows that \(\longestparabolic{I}(a_{[\alpha]}) \in -\folded{I}\), as desired. \end{proof} The following proposition clarifies the relation between parabolic subgroups of \(W(\Sigma)\) and parabolic subgroups of \(\Weyl^\sigma\): \begin{prop} For any subset \(I\subset \Sigma\) such that \(\sigma(I) = I\), the embedding \(\iota\) in \eqref{eq:folding-subgroup} identifies the parabolic subgroup \(\Weylparabolic{\folded{I}}\) of \(\Weyl^\sigma\) with \(\Weylparabolic{I}\cap \iota\Weyl^\sigma\). \end{prop} \begin{proof} The subgroup \(\Weylparabolic{\folded{I}}\) is generated by the reflections in the roots \(a_{[\alpha]}\) with \(\alpha\in I\). By \cite[1.30\,(b)]{steinberg:endomorphisms}, these reflections map to the longest element of the parabolic subgroup of \(W\) corresponding to the \(\sigma\)-orbit of \(\alpha\) in \(I\). (This is a special case of \cref{prop:folding-longest-words}.) So clearly \(\iota\Weylparabolic{\folded{I}}\subset \Weylparabolic{I} \cap \iota(\Weyl^\sigma)\). Conversely, let us consider an arbitrary element \(w\in\Weyl^\sigma\) such that \(\iota w \in \Weylparabolic{I}\). We need to show that \(w\) acts as the identity on any vector \(x\) in the orthogonal complement of \(\mathbb{R}\{a_{[\alpha]} \mid [\alpha] \in \folded{I}\}\) in \(\folded{V}\). By construction, the action of \(w\) on \(\folded{V}\) is the restriction to \(\folded{V}\) of the action of \(\iota w\) on \(V\). Moreover, with notation as in the previous proof, we find that: \begin{align*} \innerproduct{x}{a_{[\alpha]}} &= r\textstyle\sum_{i=1}^l \innerproduct{x}{\sigma^i\alpha}\\ &= rl \innerproduct{x}{\alpha} && \text{ as } x \in V^\sigma \\ \end{align*} for some \(r > 0\). Thus, the assumption that \(x\in\folded{V}\) is orthogonal to \(a_{[\alpha]}\) is equivalent to the assumption that \(x\) is orthogonal to \(\alpha\). Thus, \(\iota w \) acts trivially on \(x\). Altogether, this shows that \(w.x = (\iota w). x = x\), as needed. \end{proof} \begin{proof}[Calculations for \(A_n\), \(BC_n\), \(D_{\mathrm{odd}}\), \(E_6\) and \(F_4\)] For \(\sigma= -{w_o}\), we have an isomorphism \[ \iota\colon \Weyl^\sigma \xrightarrow{\cong} \Weyl_o(\Sigma)\subset W(\Sigma). \] It induces a bijection: \[ \frac{\{\text{involutions in \(\Weyl^\sigma\)}\}}{\text{conjugation}} \xrightarrow{\quad 1:1\quad } \frac{\{\text{involutions in \(\Weyl_o(\Sigma)\)}\}}{\text{conjugation}} \] By \cref{prop:folding-longest-words}, multiplication with \(-1\) (= the longest element in the folded group \(\Weyl^\sigma\)) on the set on the left corresponds to multiplication with \({w_o}\) on the set on the right. Consider the complete sets of representatives of conjugacy clases listed in \cref{sec:opposite-involutions-results}. Using \cref{lem:minusdim}, we can easily compute the dimensions of the \(-1\)-eigenspaces of each representative \(w\), its corresponding representative \(\iota w\), and of \(-w\) and \(\iota(-w) = {w_o} \iota(w)\), see \cref{table:minusdims}. The involutions \(\inv{k,l}{}\) are chosen such that \(\inv{k,l}{}\mapsto\inv{k,l}{}\) under the following isomorphisms induced by folding along \(-{w_o}\): \begin{align*} W(BC_n)&\xrightarrow{\cong} \Weyl_o(A_{2n})\\ W(BC_n)&\xrightarrow{\cong} \Weyl_o(A_{2n-1})\\ W(BC_n)&\xrightarrow{\cong} \Weyl_o(D_{n+1}) \quad \text{ for odd \(n+1\)} \end{align*} By simultaneously considering the entries for \(BC_n\) and \(A_{2n}\) in the table, we easily see that \(-\inv{k,l}{} \sim \inv{k,n-2k-l}{}\) in \(W(BC_n)\) and that \({w_o} \inv{k,l}{} \sim \inv{k,n-2k-l}{}\) in \(\Weyl_o(A_{2n})\). The results for \(\Weyl_o(A_{2n-1})\) and \(W(D_{\mathrm{odd}})\) immediately follow. For \(W(F_4)\) and \(\Weyl_o(E_6)\), the results are obtained by an analogous argument comparing the dimensions of \(-1\)-eigenspaces of involutions corresponding to each other under the folding isomorphism \(W(F_4)\cong\Weyl_o(E_6)\). \end{proof} \begin{table} \centering \newcolumntype{C}{>{$}c<{$}} \newcolumntype{L}{>{$}l<{$}} \begin{tabular}{LLCCCL} \toprule & w & \dim^-(w) & \dim^-({w_o} w) \\% & {w_o} w \\ \midrule A_{2n} & \inv{k,l}{} & 2k+l & n-l \\% & \inv{k,n-2k-l}{} \\ A_{2n-1} & \inv{k,l}{} & 2k+l & n-l \\% & \inv{k,n-2k-l}{} \\ BC_n & \inv{k,l}{} & k+l & n-k-l \\% & \inv{k,n-2k-l}{} \\ D_{n+1}, n+1 \text{ odd} & \inv{k,l}{}, l \text{ even } & k+l & n-k-l \\% & \inv{k,n-2k-l}{} \\ D_{n+1}, n+1 \text{ odd} & \inv{k,l}{}, l \text{ odd } & k+l+1 & n-k-l+1 \\% & \inv{k,n-2k-l}{} \\ \bottomrule \end{tabular} \caption{Dimensions of the \(-1\)-eigenspaces}\label{table:minusdims} \end{table} \begin{proof}[Callculations for \(D_{\mathrm{even}}\)] In this case, \({w_o} = -1\), so \(-{w_o}\) does not define a non-trivial folding. But we still have an automorphism \(\sigma\) of order \(2\) that commutes with \({w_o}\), and that defines a folding embedding \(W(BC_n)\hookrightarrow W(D_{n+1})\). In this case, the subgroup \(\Weyl_o(\Sigma)\) in the discussion above has to replaced by the subgroup \(W(\Sigma)^\sigma\) of elements fixed under conjugation by \(\sigma\). The arguments go through, because \({w_o}\) commutes with \(\sigma\), and hence multiplication by \({w_o}\) takes involutions in \(W(\Sigma)^\sigma\) to involultions in \(W(\Sigma)^\sigma\). All chosen representatives of conjugacy classes of involutions are contained in \(W(\Sigma)^\sigma\), except for the two involutions \(\inv{-}{}\) and \(\inv{+}{}\). These two exceptional involutions are not conjugate to any other standard involution, nor to each other. As multiplication with \({w_o}\) restricts to the set of \emph{other} conjugacy classes, it must also restrict to the two-element set consisting of these two conjugacy classes. The result presented below is obtained via an explicit computation. \end{proof} \newpage \newgeometry{left=3.5cm,right=3.5cm,top=3cm,bottom=2cm} \section{Results} \label{sec:opposite-involutions-results} \begin{@empty} \newcommand{\par\vspace{4pt}\rule{\linewidth}{0.4pt}\vspace{4pt}}{\par\vspace{4pt}\rule{\linewidth}{0.4pt}\vspace{4pt}} \parskip0pt \parindent0pt \newcommand{-1}{-1} \newcommand{\quad\xleftrightarrow{\quad \arrowlabel \quad}\quad}{\quad\xleftrightarrow{\quad -1 \quad}\quad} \newcommand{\quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel}}{\quad{\rotatebox{90}{$\circlearrowleft$}}^{-1}} \newcommand{\class }[1]{\left\{ #1 \right\}} \newcommand{\classes }[2]{\left\{ \substack{ #1 \\ #2} \right\}} \newcommand{\classeses}[2]{\left\{ \substack{ #1 \\ #2\\\\\dots} \right\}} \newcommand{\lineno}[1]{(\text{#1})} \newenvironment{result}[2]{% \par \noindent\parbox{3cm}{\framebox{\parbox{2cm}{\centering #1\\#2}}} \begin{minipage}{\linewidth-3cm}% }{% \end{minipage} \par\vspace{4pt}\rule{\linewidth}{0.4pt}\vspace{4pt} } We indicate a subgraph \(I\subset \Sigma\) by colouring the nodes of \(I\) black. For the families of types \(A\), \(BC\) and \(D\), the longest element \({w_o}\) acts on conjugacy classes of certain involutions \(c_{k,l}\) as follows: \begin{equation}\label{eq:common-involution-involution} \tag{$\dagger$} \renewcommand{-1}{{w_o}} \class{\substack{\text{involutions conjugate to}\\\inv{k,l}{}} } \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class{\substack{\text{involutions conjugate to}\\\inv{k,n-2k-l}{}} } \end{equation} In each case, the involutions \(\inv{k,l}{}\) are indexed by integers \(l \in \{0, \dots, n\}\) and \(k \in \{0, \dots, \left\lfloor \tfrac{n-l}{2} \right\rfloor\}\). \par\vspace{4pt}\rule{\linewidth}{0.4pt}\vspace{4pt} \begin{result}{$\Weyl_o(A_{2n})$}{} Each involution is conjugate to precisely one of the involutions \(\inv{k,l}{}\), defined as follows: \begin{align*} \inv{k,l}{} : ~& \begin{tikzpicture} \dynkin[ply=2,baseline] A{*o*o.o*o*.oo.oo.***.**.***.oo.oo.*o*o.o*o*} \dynkinBrace*[\text{two times } k \text{ marked roots}]{1}{8} \dynkinBrace*[2l \text{ marked roots}]{13}{16} \end{tikzpicture} \end{align*} The longest element \({w_o}\) acts on conjugacy classes as in \eqref{eq:common-involution-involution} above. \end{result} \begin{result}{$\Weyl_o(A_{2n-1})$}{\(n\geq 2\)} Each involution is conjugate to precisely one of the involutions \(\inv{k,l}{}\), defined as follows: \begin{align*} \inv{k,l}{} : ~& \begin{tikzpicture} \dynkin[ply=2,baseline] A{*o*o.o*o*.oo.oo.***.***.***.oo.oo.*o*o.o*o*} \dynkinBrace*[\text{two times } k \text{ marked roots}]{1}{8} \dynkinBrace*[\min\{2l-1,0\} \text{ marked roots}]{13}{16} \end{tikzpicture} \end{align*} The longest element \({w_o}\) acts on conjugacy classes as in \eqref{eq:common-involution-involution} above. \end{result} \begin{result}{$W(BC_n)$}{} Each involution is conjugate to precisely one of the involutions \(\inv{k,l}{}\), defined as follows: \begin{align*} \inv{k,l}{} : ~& \begin{tikzpicture} \dynkin[arrows=false,baseline] B{*o*o.o*o*.oo.oo.**.***} \dynkinBrace[k \text{ marked roots}]{1}{8} \dynkinBrace[l \text{ marked roots}]{13}{17} \end{tikzpicture} \end{align*} The element \({w_o} = -1\) acts on conjugacy classes as in \eqref{eq:common-involution-involution} above. \end{result} \begin{result}{$\Weyl_o(D_{n+1})$}{($n+1$ odd)} Each involution is conjugate to precisely one of the involutions \(\inv{k,l}{}\), defined as follows: \begin{align*} \inv{k,l}{} : ~& \begin{tikzpicture} \dynkin[baseline] D{*o*o.o*o*.oo.oo.**.***} \dynkinBrace[k \text{ marked roots}]{1}{8} \dynkinBrace[\substack{\text{no marked roots if } l = 0\\l + 1\text{ marked roots if } l > 0}]{13}{17} \end{tikzpicture} \end{align*} The longest element \({w_o}\) acts on conjugacy classes as in \eqref{eq:common-involution-involution} above. \end{result} \begin{result}{$W(D_{n+1})$}{($n+1$ even)} Each involution is conjugate to at least one of the involutions \(\inv{k,l}{}\) defined as in the case when \(n+1\) is odd, or to precisely one of the following two additional involutions: \begin{align*} \inv{-}{}&\colon \dynkin D{*o*o*o*.o*o*o*o}\\ \inv{+}{}&\colon \dynkin D{*o*o*o*.o*o*oo*} \end{align*} Here, \(\inv{k,l}{}\) is a standard involution if and only if \(l\) is odd. For even \(l\geq 2\), \(\inv{k,l}{}\sim \inv{k,l-1}{}\). The longest element \({w_o} = -1\) acts on the conjugacy classes of the involutions \(\inv{k,l}{}\) as in \eqref{eq:common-involution-involution} above. The conjugacy classes of \(\inv{-}{}\) and \(\inv{+}{}\) are fixed by multiplication with \({w_o}\) when \(n+1\equiv 0 \mod 4\), and interchanged when \(n+1\equiv 2 \mod 4\). \end{result} \newpage \mbox{}\vfill \par\vspace{4pt}\rule{\linewidth}{0.4pt}\vspace{4pt} \begin{result}{$\Weyl_o(E_6)$}{} \renewcommand{-1}{{w_o}} \begin{flalign*} \class {\dynkin[ply=2,backwards] E{oooooo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin[ply=2,backwards] E{******}} & & \lineno{1} \\ \classeses{\dynkin[ply=2,backwards] E{o*oooo}}{\dynkin[ply=2,backwards] E{ooo*oo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin[ply=2,backwards] E{*o****}} & & \lineno{2} \\ \classeses{\dynkin[ply=2,backwards] E{*oooo*}}{\dynkin[ply=2,backwards] E{oo*o*o}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin[ply=2,backwards] E{o****o}} & & \lineno{3} \\ \classeses{\dynkin[ply=2,backwards] E{*oo*o*}}{\dynkin[ply=2,backwards] E{**ooo*}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel} & & \lineno{4} \\ \class {\dynkin[ply=2,backwards] E{oo***o}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel} & & \lineno{5} \end{flalign*} \end{result} \begin{result}{$W(E_7)$}{} \begin{flalign*} \class {\dynkin E{ooooooo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin E{*******}} & & \lineno{1} \\ \classeses{\dynkin E{*oooooo}}{\dynkin E{oo*oooo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin E{o******}} & & \lineno{2} \\ \classeses{\dynkin E{*oo*ooo}}{\dynkin E{*ooo*oo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin E{o****o*}} & & \lineno{3} \\ \classeses{\dynkin E{*oo*o*o}}{\dynkin E{**oooo*}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \classes{\dynkin E{**oo*o*}}{\dynkin E{o**o*o*}} & & \lineno{4} \\ \class {\dynkin E{o*oo*o*}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin E{o****oo}} & & \lineno{5} \end{flalign*} \end{result} \vfill \newpage \mbox{}\vfill \par\vspace{4pt}\rule{\linewidth}{0.4pt}\vspace{4pt} \begin{result}{$W(E_8)$}{} \begin{flalign*} \class {\dynkin E{oooooooo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin E{********}} && \lineno{1} \\ \classeses{\dynkin E{*ooooooo}}{\dynkin E{oo*ooooo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin E{*******o}} && \lineno{2} \\ \classeses{\dynkin E{*oo*oooo}}{\dynkin E{*ooo*ooo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin E{o******o}} && \lineno{3} \\ \classeses{\dynkin E{*oo*o*oo}}{\dynkin E{**ooooo*}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \classes{\dynkin E{**oo*o*}}{\dynkin E{o**o*o*}} && \lineno{4} \\ \classeses{\dynkin E{*oo*o*o*}}{\dynkin E{**ooo*o*}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel} && \lineno{5} \\ \class {\dynkin E{o****ooo}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel} && \lineno{6} \end{flalign*} \end{result} \vfill \newpage \mbox{}\vfill \par\vspace{4pt}\rule{\linewidth}{0.4pt}\vspace{4pt} \begin{result}{$W(F_4)$}{} \begin{flalign*} \class {\dynkin F{oooo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin F{****}} && \lineno{1} \\ \classes {\dynkin F{*ooo}}{\dynkin F{o*oo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin F{o***}} && \lineno{2} \\ \classes {\dynkin F{oo*o}}{\dynkin F{ooo*}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin F{***o}} && \lineno{3} \\ \classeses{\dynkin F{*o*o}}{\dynkin F{*oo*}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel} && \lineno{4} \\ \class {\dynkin F{o**o}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel} && \lineno{5} \end{flalign*} \end{result} \begin{result}{$W(G_2(n))$}{(\(n\equiv 0 \mod 4\))} \begin{flalign*} \class {\dynkin[gonality=n] G{oo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin[gonality=n] G{**}} && \lineno{1} \\ \class {\dynkin[gonality=n] G{*o}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin[gonality=n] G{o*}} && \lineno{2} \end{flalign*} \end{result} \begin{result}{$W(G_2(n))$}{(\(n\equiv 2 \mod 4\))} \begin{flalign*} \class {\dynkin[gonality=n] G{oo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin[gonality=n] G{**}} & & \lineno{1} \\ \class {\dynkin[gonality=n] G{*o}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel} & & \lineno{2} \\ \class {\dynkin[gonality=n] G{o*}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel} & & \lineno{3} \end{flalign*} An alternative common notation for \(G_2(n)\) is \(I_2(n)\).\\ The Weyl group \(G_2\) is \(G_2 := G_2(6)\). \end{result} \begin{result}{$\Weyl_o(G_2(n))$}{(\(n\) odd)} \renewcommand{-1}{{w_o}} \begin{flalign*} \class {\dynkin[gonality=n] G{oo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin[gonality=n] G{**}} && \lineno{1} \end{flalign*} An alternative common notation for \(G_2(n)\) is \(I_2(n)\). Note that $\Weyl_o(G_2(n)) = \{\mathrm{id},{w_o}\} \cong W(A_1)$. \end{result} \begin{result}{$W(H_3)$}{} \begin{flalign*} \class {\dynkin H{ooo}} &\quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin H{***}} & & \lineno{1}\\ \classeses {\dynkin H{*oo}}{\dynkin H{o*o}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin H{*o*}} & & \lineno{2} \end{flalign*} \end{result} \begin{result}{$W(H_4)$}{} \begin{flalign*} \class {\dynkin H{oooo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin H{****}} && \lineno{1} \\ \classeses {\dynkin H{*ooo}}{\dynkin H{o*oo}} & \quad\xleftrightarrow{\quad \arrowlabel \quad}\quad \class {\dynkin H{***o}} && \lineno{2} \\ \classeses {\dynkin H{*o*o}}{\dynkin H{*oo*}} & \quad{\rotatebox{90}{$\circlearrowleft$}}^{\arrowlabel}&&\lineno{3} \end{flalign*} \end{result} \end{@empty} \vfill \restoregeometry \newpage \bibliographystyle{alpha}

https://arxiv.org/abs/1210.4829

Critical values of Gaussian SU(2) random polynomials

In this note, we will get the estimate of the expected distribution of critical values of Gaussian SU(2) random polynomials as the degree large enough. The result is a direct application of the Kac-Rice formula. The critical values will accumulate at infinity, we further study the rate of this convergence and its rescaling limit as $n\to\infty$.

\section{Introduction} Random polynomials and random holomorphic functions are studied as ways to gain insight for problems arising in string theory and analytic number theory \cite{DSZ, H, S}. In \cite{K}, Kac studied and determined a formula for the expected distribution of zeros of some real Gaussian random polynomials. His work was generalized to complex random polynomials and random analytic functions throughout the years, we refer to \cite{BSZ, DPSZ, EK, HKPV, ST} for more backgrounds and results. \subsection{$SU(2)$ polynomials} When the random polynomial is defined invariant with respect to some group action, the problem can turn out to be particularly interesting, we refer \S 2.3 in \cite{HKPV} for examples. In this article, we will study a special family: the Gaussian $SU(2)$ random polynomial. This is of particular interest in the physics literature as the zeros describe a random spin state for the ~Majorana~ representation (modulo phase) on the unit sphere \cite{H}. Given a probability space $\Omega$ and $\{a_j\}_{j=1}^\infty$ a collection of i.i.d complex Gaussian random variables with mean $0$ and variance $1$ on it, the family of $SU(2)$ random polynomials is defined as, \begin{equation}\label{su2}p_n(z)=\sum_{j=0}^n a_j \sqrt{ {{n}\choose{j}}}z^j.\end{equation} Although this polynomial is defined on $\mathbb{C}$, we may also view it as an analytic function on $\mathbb{CP}^1 = \mathbb{C}\cup{\infty}$ with a pole at $\infty$. Various properties of the zeros of random $SU(2)$ polynomials have been studied such as the distribution of zeros and the two points correlation function \cite{BSZ, H}. First, zeros of this polynomials are uniformly distributed on $S^2\cong \mathbb{CP}^1$ with respect to the Fubini-Study metric, i.e., the average distribution of zeros is invariant under the $SU(2)$ action on $\mathbb{CP}^ 1$ \cite{HKPV}. To be more precise, let's denote $$\mathcal Z_{p_n}=\sum_{z\in\mathbb{CP}^1:\,\,p_n(z)=0}\delta_z$$ as the empirical measure of zeros of Gaussian $SU(2)$ random polynomials and define the pairing, $$\langle \mathcal Z_{p_n},\phi\rangle =\sum_{z\in\mathbb{CP}^1:\,\,p_n(z)=0} \phi(z), \,\,\,\, \mbox{where}\,\,\, \phi\in C^\infty(\mathbb{CP}^1)$$ We define the expectation, $$\langle \mathbb{E} \mathcal Z_{p_n}, \phi\rangle:=\mathbb{E} \langle \mathcal Z_{p_n},\phi\rangle=\frac1{\pi^{n+1}}\int_{\mathbb{C}^{n+1}}\left(\sum_{z\in\mathbb{CP}^1:\,\,p_n(z)=0} \phi(z)\right)e^{-\frac{|a|^2}2}d\ell_{a_0}\cdots d\ell_{a_n}$$ where $d\ell_{a_j}=\frac 1{2i}da_j\wedge d\bar a_j$ is the Legesgue measure on $\mathbb{C}$. Then the expected density of zeros is \cite{BSZ}, $$\mathbb{E} \mathcal Z_{p_n}=n\omega_{FS},$$ in the sense that, $$\mathbb{E}\langle \mathcal Z_{p_n},\phi\rangle =n\int_{\mathbb{CP}^1}\phi\omega_{FS}, \,\,\,\, \mbox{where}\,\,\, \phi\in C^\infty(\mathbb{CP}^1)$$ where $\omega_{FS}$ is the Fubini-Study metric on $\mathbb{CP}^1$ \cite{GH}. We can also study the two points correlation function of zeros of $SU(2)$ polynomials and its scaling property. We define the two points correlation function as \cite{BSZ}, $$ K_n(z,w):=\mathbb{E} \left(Z_{p_n}(z)\otimes Z_{p_n}(w)\right)$$ such that for any smooth test function $\phi_1(z)\otimes\phi_2(w) $, we have the pairing, $$\left\langle K_n(z,w),\phi_1(z)\otimes\phi_2(w)\right\rangle=\mathbb{E} \left(\langle \mathcal Z_{p_n}, \phi_1\rangle\right)\left(\langle \mathcal Z_{p_n}, \phi_2\rangle\right)$$ If we scale the two points correlation function by a factor $\frac 1 {\sqrt n}$, then we have, $$K_n(\frac{z}{\sqrt n},\frac{w}{\sqrt n})=\frac{(\sinh t^2+t^2)\cosh t-2t\sinh t}{\sinh t^3}+O(\frac 1 {\sqrt n})$$ where $t=\frac {|z-w|^2}{2}$ and $|z-w|$ is the geodesic distance of $z$ and $w$ on $\mathbb{CP}^1$. It's easy to see, $$K_n(\frac{z}{\sqrt n},\frac{w}{\sqrt n})=t-\frac 2 9 t^3+O(t^5) \,\,\,\mbox{as}\,\,\, t\rightarrow 0$$ which implies zeros repel each other. We refer to \cite{BSZ, H} for more details. \subsection{Main results} In this article, we will study the expected distribution of nonvanishing critical values of $|p_n|$ as $n$ large enough. Note that the modulus $|p_n|$ is a subharmonic function, thus there is no local maxima; local minima are all zeros and thus nonvanishing critical values are obtained only at saddle points \cite{FS}. Hence the expected density of nonvanishing critical values of $|p_n|$ we study in this article is in fact the expected density of values of saddle points of $|p_n|$. The nonvanishing critical values of $|p_n|$ are obtained at points, \begin{equation}\label{set}\{z\in\mathbb{C}:\,\,\,p_n'=0 \,\,\mbox{and}\,\,\, p_n\neq 0\}. \end{equation} For a random polynomial $p_n$, it has no repeated zeros almost surely, which implies that the set (\ref{set}) is almost surely equivalent to, \begin{equation}\{z\in\mathbb{C}:\,\,\,p_n'=0\}.\end{equation} i.e., (nonvanishing) $|p_n|$ and $p_n$ have the same critical points almost surely. Hence, we will first get the expected density of critical values of $p_n$ in Theorem \ref{main}, as a direct consequence, we can apply the polar coordinate to get the expected density of nonvanishing critical values of $|p_n|$ in Theorem \ref{main2}. We denote the empirical measure of critical values of $p_n$ as, \begin{equation}\label{empirical}\mathcal C_{p_n}=\sum_{z: \,\, p'_n(z)=0}\delta_{p_n(z)}.\end{equation} We now define the pairing, \begin{equation}\label{pair}\langle \mathcal C_{p_n}, \phi\rangle =\sum_{z: \,\, p'_n(z)=0}\phi(p_n(z)), \,\,\,\,\,\,\forall \phi(x)\in C_c^\infty(\mathbb{R}^2)\end{equation} where $C_c^\infty(\mathbb{R}^2)$ is the space of smooth functions on $\mathbb{R}^2$ with compact support. We denote $\mathbb D_{p_n}(x)$ as the expected density of critical values of $p_n$ in the sense that, \begin{equation}\label{expect}\mathbb{E}\langle \mathcal C_{p_n}, \phi\rangle =\int_\mathbb{C} \phi(x)\mathbb D_{p_n}(x)d\ell_x , \,\,\,\,\,\,\forall \phi(x)\in C_c^\infty(\mathbb{R}^2)\end{equation} whereas $d\ell_ x$ is the ~Lebesgue~ measure of $\mathbb{C}$. Those definitions also apply to the empirical measure of the nonvanishing critical values of $|p_n|$, \begin{equation}\label{em3}\mathcal C_{|p_n|}=\sum_{z: \,\, p'_n=0}\delta_{|p_n|}.\end{equation} which is a measure defined on the nonnegative real line $\mathbb{R}_+$. We define its expectation as, \begin{equation}\langle \mathbb{E}\mathcal C_{|p_n|}, \phi\rangle:=\mathbb{E}\langle \mathcal C_{|p_n|}, \phi\rangle= \int_0^\infty\phi(x)\mathbb D_{|p_n|}dx,\,\,\,\forall\,\,\phi(x)\in C^\infty_c(\mathbb{R}_+)\end{equation} where $dx$ is the ~Lebesgue~ measure on $\mathbb{R}$. In this article, we will first get the exact formula for the expected density $\mathbb D_{p_n}$ in the Proposition \ref{Dn} by the Kac-Rice formula (see section \S \ref{KR}), then we study the asymptotic behavior of $\mathbb{D}_{p_n}$ as $n\rightarrow \infty$. Our main results are, \begin{thm} \label{main}The expected density $\mathbb D_{p_n}$ of the empirical measure $\mathcal C_{p_n}$ of the critical values of $p_n$ satisfies the estimate, \begin{equation} \label{mainequation} \mathbb D_{p_n}=\frac {1-e^{-|x|^2}}{\pi|x|^2}+\frac 1\pi\int _0^1 e^{-(s-s\log s)|x|^2}ds+o(1) \,\,\,\mbox{as}\,\,\, n\rightarrow \infty \end{equation} for $x\in \mathbb{C}$. \end{thm} It seems that the growth $\mathbb{D}_{p_n}$ is $\frac1{|x|^2}$ as $|x|$ is away from $0$ as $n\rightarrow \infty$, below is the computer graphic of this function with the variable $t:=|x|$, \begin{figure}[h!] \begin{center} \includegraphics[scale=0.30]{Df.pdf} \end{center} \end{figure} As proved in Proposition \ref{Dn}, the density $\mathbb{D}_{p_n}d\ell_x$ only depends on $|x|$, i.e., the modulus of $|p_n|$, thus we can rewrite it as $\mathbb{D}_{p_n}(|x|)|x|d|x|d\theta$ under the polar coordinate. If we integrate on $\theta$ variable, then $$\mathbb{D}_{|p_n|}=\int_0^{2\pi} \mathbb{D}_{p_n}(|x|)|x|d\theta=2\pi|x| \mathbb{D}_{p_n}(|x|)$$ will be the density of critical values of $|p_n|$. Thus as a direct consequence of Theorem \ref{main}, we get, \begin{thm} \label{main2}The expected density $\mathbb D_{|p_n|}$ of the empirical measure $\mathcal C_{|p_n|}$ of the nonvanishing critical values of $|p_n|$ satisfies the estimate, \begin{equation} \label{mainequation2} \mathbb D_{|p_n|}(x)=\frac {2(1-e^{-x^2})}{x}+2x\int _0^1 e^{-(s-s\log s)x^2}ds+o(1) \,\,\,\mbox{as}\,\,\, n\rightarrow \infty \end{equation} for $x\in \mathbb{R}_+$. \end{thm} The growth of $\mathbb D_{|p_n|}$ is also of order $1/x$ as $x$ is large, but there is a peak when $x$ is small, the following is the computer graphic, \begin{figure}[h] \begin{center} \includegraphics[scale=0.30]{DMf.pdf} \end{center} \end{figure} \subsection{Further remarks} First note that our setting is different from the one in \cite{DSZ}. For example, in \cite{DSZ}, critical points of $SU(2)$ polynomials are defined to be the points $$\{z\in\mathbb{CP}^1:\,\, \nabla'p_n=0\}$$ where $\nabla'=\frac \d{\d z}-\frac{n\bar zdz}{1+|z|^2}$ is the smooth Chern connection on the line bundle $\mathcal O(n)\rightarrow \mathbb{CP}^1$ with respect to the ~Fubini~-Study metric and $p_n$ is a global holomorphic section of the line bundle $\mathcal O(n)\rightarrow \mathbb{CP}^1$ \cite{GH}. By choosing such smooth Chern connection, the expected distribution of critical points are also invariant under the $SU(2)$ action \cite{DSZ}. But in this article, the critical points are defined by the usual derivative $$\{z\in\mathbb{C}:\,\, \frac{\d p_n}{\d z}=0\}.$$ In fact, the derivative $\frac \d{\d z}$ is a meromorphic flat Chern connection on $\mathcal O(n)\rightarrow \mathbb{CP}^1$ with a pole at $\infty$. Under this setting, the expected density of critical points is not $SU(2)$ invariant, we refer to \cite{Ha} for more details. Our second remark is as following. In \cite{FZ}, the authors studied the expected density of nonvanishing critical values of the pointwise norm of Gaussian random holomorphic sections of the positive holomorphic line bundle over compact K\"ahler manifolds. Now let's briefly explain the main result in \cite{FZ} and compare it with Theorem \ref{main2}. Take Gaussian $SU(2)$ random polynomials (sections) $p_n$ for example. We equip the line bundle $\mathcal O(n)\rightarrow \mathbb{CP}^1$ with a Hermitian metric $h^n=e^{-n\phi}$ where $\phi=\log ( 1+|z|^2)$ is the K\"ahler potential of Fubini-Study metric. Then the pointwise $h$-norm of the holomorphic section $|p_n|_{h^n}=|p_n|e^{-\frac {n\phi} 2}$ is global defined on $\mathbb{CP}^1$ \cite{GH} and hence the critical points of $|p_n|_{h^n}$ is defined as, $$\Sigma_n=\{z\in \mathbb{CP}^1: \,\, \frac {\d |p_n|_{h^n}}{\d z}=0\}$$ We define the (normalized) empirical measure of critical value of $|p_n|_{h^n}$ as, $$\mathcal C_{|p_n|_{h^n}}:=\frac 1 n\left(\sum_{z\in \Sigma_n}\delta_{|p_n|_{h^n}}\right)$$ which is also a measure defined on $\mathbb{R}_+$. Then the expectation of $\mathcal C_{|p_n|_{h^n}}$ satisfies the estimate, \begin{equation}\label{timed}\mathbb{E} \mathcal C_{|p_n|_{h^n}}=x\left(2x^2-4+8e^{-\frac{x^2}2}\right) e^{-x^2}+O(\frac 1 n),\,\,\,\,x\in \mathbb{R}_+\end{equation} as $n$ large enough. In fact, this estimate is universal: it holds on any Riemannian surfaces \cite{FZ}. Thus the (normalized) density $\mathbb{E}\mathcal C_{|p_n|_{h^n}}$ is exponent decaying as $x$ large enough which is quite different from the behavior of (non-normalized) density $\mathbb{E}\mathcal C_{|p_n|}$ in Theorem \ref{main2}. This is mainly because of the connection we choose: the usual derivative $\frac \d{dz}$ in this article is a meromorphic flat connection on $\mathbb{CP}^1$ with a pole at $\infty$ while in \cite{FZ}, the proof of (\ref{timed}) relies on a choice of smooth Chern connection $\nabla'=\frac \d{\d z}-\frac{n\bar zd z}{1+|z|^2}$. \bigskip \textbf{Acknowledgements}: The first named author would like to thank S. Zelditch for suggesting this problem and many helpful discussions. \section{Kac-Rice formula}\label{KR} In this section, we first review the Kac-Rice formula for a stochastic process, referring to \cite{AT, K, R} for more details. Then we generalize the formula to the expected distribution of critical values of $p_n$. The Kac-Rice formula is as follows: let $f(z)$ be a real valued stochastic process indexed by a compact interval $I\subset \mathbb{R}$. Then the Kac-Rice formula for the expected number of zeros is, $$\label{kr}\mathbb{E}\#\{z\in I: \,\,f(z)=0\}=\int_I \int_\mathbb{R} |y|p_z(0,y)dydz $$ where $p_z(0,y)$ is the joint density $p_z(x,y)$ of $(f, f')$ evaluated at $(0,y)$. If $f$ is a Gaussian process, then the joint density $p_z(x,y)$ is determined by the covariance matrix of $(f,f')$ \cite{AT}. The proof of this formula is explained in more details in \cite{AT}. The idea of the proof is based on the following observation, $$\#\{z\in I: \,\,f(z)=0\}=\int_I \delta_0(f(z))|f'(z)|dz$$ We take expectation on both sides to get, \begin{align*}\mathbb{E}\#\{z\in I: \,\,f(z)=0\}=&\int_I \int_{\mathbb{R}_y}\int_{\mathbb{R}_x} \delta_0(x)p_z(x,y)|y|dxdydz\\=&\int_I\int_\mathbb{R} |y|p_z(0,y)dydz\end{align*} Thus the expected density of zeros of $f$ is given by, \begin{equation}\label{realcase}\mathbb{E} \left(\sum_{z\in I:\,\, f(z)=0}\delta_z\right)=\left(\int_\mathbb{R} |y|p_z(0,y)dy\right)dz\end{equation} If $f(z)$ is a complex stochastic process indexed by a compact complex domain, then the above formula reads, \begin{equation}\label{complexcase}\mathbb{E} \left(\sum_{z\in I:\,\, f(z)=0}\delta_z\right)=\left(\int_\mathbb{C} |y|^2p_z(0,y)d\ell_y\right)d\ell_z\end{equation} where $d\ell_y$ and $d\ell_ z$ are Lebesgue measures on $\mathbb{C}$. Compared with (\ref{realcase}), we get $|y|^2$ since a $1$-dimensional complex random process is a $2$-dimensional real random process. In fact, this formula is based on the definition of the delta function, $$\#\{z\in I: \,\,f(z)=0\}=\int_I \delta_0(f(z))\frac 1{2i }df\wedge d\bar f=\int_I \delta_0(f(z))|f'|^2d\ell_ z$$ The formula arises when we take expectation on both sides. \subsection{Kac-Rice formula: Revisit} In this subsection, let's get the formula for the expected density of critical values of a (real or complex) stochastic process $f$ by the method of Kac-Rice. For simplicity, let's first consider a smooth real stochastic process $f\in C^\infty(I)$ where $I$ is a compact subset in $\mathbb{R}$. Let $\Theta\subset \mathbb{R}$ be a compact subset. Let's denote the set of critical values in $\Theta$ as, $$\mathcal C_\Theta=\{z\in I: \,\,f(z)\in \Theta, \,\,\,f'(z)=0\}.$$ Let's denote the measure $\mu(x)dx$ on $\Theta$ as, $$\mu(x)dx=\mathbb{E}\left(\sum_{z\in \mathcal C_\Theta}\delta_{f(z)}\right)$$ in the sense that, $$\mathbb{E}\left(\langle\sum_{z\in \mathcal C_\Theta}\delta_{f(z)},\phi\rangle\right)=\int_\Theta \phi \mu(x)dx$$ where $\phi$ is any smooth test function defined on $\Theta$. Then we have the following lemma, \begin{lem}\label{easycase}Let's denote $p_z(x,y,\xi)$ as the joint probability of $(f,f',f'')$. Then, $$\mu(x)dx=\left(\int_{I}\int_\mathbb{R} |\xi|p_z(x,0,\xi)d\xi dz\right)dx$$ where $dx$, $d\xi$ and $dz$ are ~Lebesgue~ measures on $\mathbb{R}$. \end{lem}\begin{proof}First note that, $$\langle \sum_{f\in\Theta,\,\, f'=0}\delta_{f(z)}, \phi(x)\rangle=\sum_{f\in\Theta,\,\, f'=0}\phi(f(z))=\int_I \chi_{\{f\in\Theta\}} \phi(f(z))\delta(f') df'$$ By taking expectation on both sides, \begin{align*}& \mathbb{E}\left\langle \sum_{f\in\Theta,\,\, f'=0}\delta_{f(z)}, \phi(x)\right\rangle \\=&\int_{\mathbb{R}_x}\int_I\int_{\mathbb{R}_\xi}\int_{ \mathbb{R}_y} p_z(x,y,\xi)\chi_{\{x\in\Theta\}} \phi(x)\delta(y) |\xi|dy d\xi dzdx\\ =&\int_{\Theta}\left(\int_I\int_{\mathbb{R}_\xi} p_z(x, 0,\xi) |\xi| d\xi dz \right)\phi(x)dx\\=&\int_\Theta \phi(x)\mu(x)dx\end{align*} \end{proof} In the proof of Lemma \ref{easycase}, we have assumed $I$ and $\Theta$ being compact subsets in $\mathbb{R}$. But the proof of Lemma \ref{easycase} can be generalized to the $SU(2)$ random polynomials $p_n$ which are a collection of complex Gaussian stochastic processes indexed by $\mathbb{C}$. The generalization of $\Theta$ to be $\mathbb{C}$ only requires picking up a sequence of discs centered at the origin with radius $m\in\{1,2,\ldots\}$ and taking limit in weak sense. And the generalization from $I$ to $\mathbb{C}$ is the same. However, we do need to modify the pairing by choosing the test functions $\phi(z)$ in the smooth compact supported space $\mathcal C_c^\infty(\mathbb{R}^2)$ in order to change the order of the integration on $\mathbb{C}$. Following the proof of Lemma \ref{easycase}, we have, \begin{lem}\label{complexcase} The expected density of critical values of $p_n$ is, \begin{equation}\label{density}\mathbb{D}_{p_n}d\ell_x=\left(\int_{\mathbb{C}}\int_\mathbb{C} |\xi|^2p_z(x,0,\xi)d\ell_\xi d\ell_z\right)d\ell_ x\end{equation} where $d\ell_x, d\ell_\xi$ and $d\ell_ z$ are Lebesgue measures on $\mathbb{C}$ and \begin{equation}\label{density}p_z(x,0,\xi)=\frac 1{\pi^3\det\Delta_z}\exp\left\{-\left\langle \begin{pmatrix}x\\ 0 \\ \xi \end{pmatrix}, \Delta_z^{-1}\begin{pmatrix}\bar x\\ 0 \\ \bar \xi \end{pmatrix}\right \rangle\right\}\end{equation} is the joint density of $(p_n,p_n',p_n'')$ where $\Delta_z$ is the covariance matrix of $(p_n,p_n',p_n'')$. \end{lem} The proof of this formula is the same as the one in Lemma \ref{easycase}. Note that we get $|\xi|^2$ in the formula since $p_n$ is a 1-dimensional complex Gaussian process which is a 2-dimensional real process. Moreover, $p_n$ is a Gaussian process, hence the joint density $p_z(x,y,\xi)$ is uniquely determined by the covariance matrix of $(p_n,p_n',p_n'')$, we refer to \cite{AT, BSZ, DSZ} for more details. \section{Proof of main Theorems} \subsection{The Density $\mathbb{D}_{p_n}$} In this subsection, we will derive the exact formula for $\mathbb{D}_{p_n}$ based on Lemma \ref{complexcase}. We prove, \begin{prop}\label{Dn}The expected density of the empirical measure of $\mathcal C_{p_n}$ is given by the formula, \begin{equation}\label{dpn}\mathbb{D}_{p_n}=\frac{n-1}{\pi}\int_1^\infty \frac{n(r-1)+1}{r^{n+2}}e^{-\frac{n(r-1)+1}{r^n}|x|^2}dr\end{equation} Thus $\mathbb{D}_{p_n}$ is a function only depending on $|x|$. \end{prop} \begin{proof} By Lemma \ref{complexcase}, in order to compute the expected density of critical values of $p_n$, we first need to compute the covariance matrix of $(p_n,p_n',p_n'')$. By definition, the covariance matrix of the Gaussian process $(p_n,p_n',p_n'')$ is given by \cite{AT, DSZ}, $$ \Delta=\begin{pmatrix} \mathbb{E}(p_n\overline{p_n}) & \mathbb{E}(p'_n\overline{p_n}) & \mathbb{E}(p_n''\overline{p_n}) \\ \mathbb{E}(p_n\overline{p_n'}) & \mathbb{E}(p'_n\overline{p_n'}) & \mathbb{E}(p''_n\overline{p'_n}) \\ \mathbb{E}(p_n\overline{p''_n}) &\mathbb{E}(p'_n\overline{p''_n}) &\mathbb{E}(p''_n\overline{p''_n}) \end{pmatrix} $$ The covariance kernel for the Gaussian process $p_n$ is, $$\mathbb{E}(p_n(z)\overline{p_n}(w)): =\Pi_n(z,w)=(1+z\bar w)^n$$ Then we can express each entry in the covariance matrix as, $$\mathbb{E}(p_n\overline{p_n})=\Pi_n(z,z), \,\,\, \mathbb{E}(p'_n\overline{p_n})=\frac {\partial\Pi_n(z,w)}{\partial z}|_{z=w}$$ $$ \mathbb{E}(p_n''\overline{p_n})=\frac {\partial^2\Pi_n(z,w)}{\partial^2 z}|_{z=w},\,\,\,\mathbb{E}(p'_n\overline{p'_n})=\frac {\partial^2\Pi_n(z,w)}{\partial z\partial {\bar w}}|_{z=w}$$ $$\mathbb{E}(p_n''\overline{p_n})=\frac {\partial^3\Pi_n(z,w)}{\partial^2 z\partial \bar w}|_{z=w},\,\,\,\mathbb{E}(p''_n\overline{p''_n})=\frac {\partial^4\Pi_n(z,w)}{\partial^2 z\partial^2 \bar w}|_{z=w}$$ Some straight forward computations show that the covariance matrix is, $$ \Delta_z=(1+|z|^2)^{n}\begin{pmatrix} 1 & \frac{n\bar z}{1+|z|^2}& \frac{n(n-1)\bar z^2}{(1+|z|^2)^2}\\ \frac{n z}{1+|z|^2} & \frac{n+n^2| z|^2}{(1+|z|^2)^2}& \frac{2n(n-1)\bar z+(n-1)n^2\bar z |z|^2}{(1+|z|^2)^3}\\ \frac{n(n-1) z^2}{(1+|z|^2)^2} & \frac{2n(n-1)z+(n-1)n^2 z |z|^2}{(1+|z|^2)^3}& \frac{2n(n-1)+4n(n-1)^2|z|^2+n^2(n-1)^2|z|^4}{(1+|z|^2)^4} \end{pmatrix}\;, $$ Hence, \begin{equation}\label{deter}\det \Delta_z= (1+|z|^2)^{3n}\frac{2n^3-2n^2}{(1+|z|^2)^6}.\end{equation} Now we denote, $$Q_z(x,\xi)=:\left\langle \begin{pmatrix}x\\ 0 \\ \xi \end{pmatrix}, \Delta_z^{-1}\begin{pmatrix}\bar x\\ 0 \\ \bar \xi \end{pmatrix}\right \rangle $$ Then by direct computations, we rewrite, $$Q_z(x,\xi)=\frac{(1+|z|^2)^{2n}}{\det \Delta_z}\left\langle \begin{pmatrix}x \\ \xi \end{pmatrix}, \begin{pmatrix} \frac{n^5|z|^4+2n^4(|z|^2-|z|^4)+n^3(|z|^4-2|z|^2+2)-2n^2}{(1+|z|^2)^6}& \frac{(n^3-n^2)\bar z^2}{(1+|z|^2)^4}\\\frac{(n^3-n^2) z^2}{(1+|z|^2)^4}& \frac n {(1+|z|^2)^2}\end{pmatrix} \begin{pmatrix}\bar x\\ \bar \xi \end{pmatrix} \right\rangle.$$ We expand this expression and further rewrite $Q_z(x,\xi)$ as, \begin{equation}\label{vv}\frac {1}{2(1+|z|^2)^n}\left(\left|\sqrt{n^2-n}\bar z^2x+\frac 1{\sqrt{n^2-n}}\xi (1+|z|^2)^2\right|^2+2(n|z|^2+1)|x|^2\right).\end{equation} By Lemma \ref{complexcase}, the expected density of critical values of $p_n$ is given by the formula, \begin{equation}\label{ddddd}\mathbb{D}_{p_n}(x)=\frac 1{\pi^3}\int_{\mathbb{C}}\int_\mathbb{C} \frac {e^{-Q_z(x,\xi)}}{\det \Delta_z}|\xi|^2d\ell_\xi d\ell_z.\end{equation} Let's integrate $\xi$ variable first. Plug (\ref{vv}) into (\ref{ddddd}), we can rewrite (\ref{ddddd}) as, \begin{equation}\label{ttt}\mathbb{D}_{p_n}(x)=\frac 1{\pi^3}\int_{\mathbb{C}}K_z \frac {e^{-\frac{n|z|^2+1}{(1+|z|^2)^n}|x|^2}}{\det \Delta_z} d\ell_z. \end{equation} where $K_z$ is the following integral in $\xi$ variable, $$K_z=\int_\mathbb{C} \exp \left\{-\frac 1{2(1+|z|^2)^n}\left|\sqrt{n^2-n}\bar z^2x+\frac 1{\sqrt{n^2-n}}\xi (1+|z|^2)^2\right|^2\right\}|\xi|^2d\ell_\xi .$$ Change variables $\xi\rightarrow \frac 1 {\sqrt{n^2-n}} \xi (1+|z|^2)^2$ to get, $$K_z=\frac {(n^2-n)^2}{(1+|z|^2)^8}\int_\mathbb{C} \exp \left\{-\frac 1{2(1+|z|^2)^n}\left|\sqrt{n^2-n}x\bar z^2+\xi\right|^2\right \}|\xi|^2d\ell_\xi .$$ Further change variable, $\xi\rightarrow \sqrt{n^2-n}x\bar z^2+\xi$ to get, $$K_z=\frac {(n^2-n)^2}{(1+|z|^2)^8}\int_\mathbb{C} \exp \left\{-\frac {|\xi|^2}{2(1+|z|^2)^n}\right\}\left|\xi-\sqrt{n^2-n}x\bar z^2\right|^2d\ell_\xi ,$$ which equals to$$K_z=\frac {(n^2-n)^2}{(1+|z|^2)^8}\int_\mathbb{C} \exp \left\{-\frac {|\xi|^2}{2(1+|z|^2)^n}\right\}\left(|\xi|^2+(n^2-n)|x z^2|^2\right)d\ell_\xi .$$ Integrate it out, we have, $$K_z=\pi \frac {(n^2-n)^2}{(1+|z|^2)^8}\left[2(1+|z|^2)^n(n^2-n)|x|^2|z|^4+4(1+|z|^2)^{2n}\right].$$ Now we change variable $r=1+|z|^2$, rewrite $$K_z=\pi \frac {(n^2-n)^2}{r^8}\left[2r^n(n^2-n)|x|^2(r-1)^2+4r^{2n}\right]$$ and $$\det\Delta_z=r^{3n-6}(2n^3-2n^2), \,\,\,\, e^{-\frac{n|z|^2+1}{(1+|z|^2)^n}|x|^2}=e^{-\frac{n(r-1)+1}{r^n}|x|^2}$$ Plug these two lines back into the formula of (\ref{ttt}) and use the polar coordinate $d\ell_z=\frac 1 2drd\theta$, integrate on $\theta$ variable, we can rewrite $\mathbb{D}_{p_n}$ as, \begin{equation}\label{ttttt}\mathbb{D}_{p_n}=\frac{n-1}{\pi}\int_1^\infty \frac{(n^2-n)r^n(r-1)^2|x|^2+2r^{2n}}{r^{3n+2}}e^{-\frac{n(r-1)+1}{r^n}|x|^2} dr\end{equation} There are two parts in the numerator, we use integration by part to simplify the first part. Note that $$de^{-\frac{n(r-1)+1}{r^n}|x|^2}=e^{-\frac{n(r-1)+1}{r^n}|x|^2}[r^{-n-1}(n^2-n)(r-1)|x|^2]dr,$$ then the first part is equal to, \[ \begin{aligned}&\frac{n-1}{\pi}\int_1^\infty \frac{(n^2-n)r^n(r-1)^2|x|^2}{r^{3n+2}}e^{-\frac{n(r-1)+1}{r^n}|x|^2} dr\\=&\frac{n-1}{\pi}\int_1^\infty \frac{(r-1)}{r^{n+1}}de^{-\frac{n(r-1)+1}{r^n}|x|^2}\\=&\frac{n-1}{\pi}\int_1^\infty [\frac n{r^{n+1}}-\frac{n+1}{r^{n+2}}]e^{-\frac{n(r-1)+1}{r^n}|x|^2}dr\end{aligned}\] Hence the density (\ref{ttttt}) is further simplified to be, $$\mathbb{D}_{p_n}(|x|^2)=\frac{n-1}{\pi}\int_1^\infty \frac{n(r-1)+1}{r^{n+2}}e^{-\frac{n(r-1)+1}{r^n}|x|^2}dr$$ which completes the proof. \end{proof} \subsection{Proof of Theorem \ref{main}} Now we turn to the proof of our main Theorem \ref{main}. We denote $t=\frac 1{r}$ and $$y_n(t)=\frac{n(r-1)+1}{r^n}=nt^{n-1}-(n-1)t^n,$$ then we have $t\in [0,1]$ and $y_n(t)\in [0,1]$ with $y_n(0)=0$ and $y_n(1)=1$. Substitute $\frac{n(r-1)+1}{r^n}$ by $y_n(t)$, we rewrite $\mathbb{D}_{p_n}$ in Proposition (\ref{Dn}) as, \[ \begin{aligned}\mathbb{D}_{p_n}=&\frac{n-1}{\pi}\int_1^\infty \frac{y_n(t)}{r^2}e^{-y_n(t)|x|^2}dr\\=&\frac{n-1}{\pi}\int_0^1 y_n(t)e^{-y_n(t)|x|^2}dt\end{aligned}\] where in the last step, we change variable $t\rightarrow \frac 1 r$. The trick to estimate $\mathbb{D}_{p_n}$ is to calculate $$g_n(|x|^2):=\int_0^1 e^{-y_n(t)|x|^2}dt,$$ By integration by part, we have, \[ \begin{aligned}g_n(|x|^2)=&\int_0^1 t' e^{-y_n(t)|x|^2}dt\\=&e^{-|x|^2}+\int_0^1 t y'_n(t)|x|^2e^{-y_n(t)|x|^2}dt\\ =&e^{-|x|^2} +n(n-1)|x|^2\int_0^1(t^{n-1}-t^n)e^{-y_n(t)|x|^2}dt\\=&e^{-|x|^2} +n|x|^2\int_0^1(nt^{n-1}-(n-1)t^n)e^{-y_n(t)|x|^2}dt-n|x|^2\int_0^1 t^{n-1}e^{-y_n(t)|x|^2}dt\\=&e^{-|x|^2} +n|x|^2\int_0^1y_n(t)e^{-y_n(t)|x|^2}dt-|x|^2h_n(|x|^2)\\ =& e^{-|x|^2} +\frac {\pi n |x|^2}{n-1}\mathbb{D}_{p_n}-|x|^2h_n(|x|^2),\end{aligned}\] where we denote \begin{equation}\label{hn}h_n(|x|^2):=n\int_0^1 t^{n-1}e^{-y_n(t)|x|^2}dt.\end{equation} Thus, \begin{equation}\label{firstestimate}\mathbb{D}_{p_n}=\frac{n-1}{n\pi}\left(\frac{g_n(|x|^2)-e^{-|x|^2}}{|x|^2} +h_n(|x|^2) \right) \end{equation} We claim $$\lim\limits_{n\rightarrow\infty}g_n(|x|^2)=1.$$ This is quite straight forward, as $\forall\epsilon\in(0,1)$, we rewrite, $$g_n(|x|^2)=\int_0^1 e^{-y_n(t)|x|^2}dt=\int_0^{1-\epsilon}+\int_{1-\epsilon}^1$$ Since $y_n(t)\rightarrow 0$ uniformly on $[0,1-\epsilon]$ as $n\rightarrow\infty$, thus $$\lim_{n\rightarrow \infty} \int_0^{1-\epsilon}e^{-y_n(t)|x|^2}dt=\int_0^{1-\epsilon}\lim_{n\rightarrow \infty} e^{-y_n(t)|x|^2}dt=1-\epsilon$$ For the second integration, since $y_n(t)\geq 0$ on $[0,1]$, we have $\int_{1-\epsilon}^1 e^{-y_n(t)|x|^2}dt\leq \epsilon$ Hence we get $$1-\epsilon\leq\varliminf\limits_{n\rightarrow\infty}\int_0^1 e^{-y_n(t)|x|^2}dt\leq\varlimsup\limits_{n\rightarrow\infty}\int_0^1 e^{-y_n(t)|x|^2}dt\leq 1$$ As $\epsilon$ is chosen arbitrarily, letting $\epsilon\to0^+$ yields the claim. Now we estimate (\ref{firstestimate}) to be, \begin{equation}\label{ddd}\begin{aligned}\mathbb{D}_{p_n}=&\frac{n-1}{n\pi}\left(\frac{1-e^{-|x|^2}}{|x|^2} +h_n(|x|^2) +o(1)\right)\\=&\frac{1-e^{-|x|^2}}{\pi |x|^2} +\frac {1} {\pi} h_n(|x|^2) +o(1).\end{aligned}\end{equation} as $n \rightarrow \infty$. We now turn to estimate $h_n(|x|^2)$. Change variable $s=t^n$, $h_n$ will be rewritten as $$\int_0^1e^{z_n(s)|x|^2}ds$$ where $$z_n(s)=-ns^{\frac{n-1}{n}}+(n-1)s.$$ It's easy to check that $$z_n(s)\leq z_{n+1}(s) $$ for any fixed $s\in [0,1]$. Thus we have $z_n(s)$ monotone increasing to $-(s-s\log s)$ as $n\rightarrow\infty$, hence, $h_n(|x|^2)$ will satisfy $$\lim\limits_{n\rightarrow \infty}h_n(|x|^2)=\int_0^1e^{-(s-s\log s)|x|^2}ds $$ This will give us the estimate $$h_n(|x|^2)=\int_0^1e^{-(s-s\log s)|x|^2}ds+o(1)$$ as $n\rightarrow \infty$. Hence we further estimate (\ref{ddd}) to be, $$\mathbb{D}_{p_n}=\frac{1-e^{-|x|^2}}{\pi|x|^2}+\frac 1 \pi\int_0^1e^{-(s-s\log s)|x|^2}ds +o(1) \,\,\,\mbox{as}\,\,\, n\rightarrow \infty$$ which completes the proof of Theorem \ref{main}.

https://arxiv.org/abs/1602.08997

Scaling limit and ageing for branching random walk in Pareto environment

We consider a branching random walk on the lattice, where the branching rates are given by an i.i.d. Pareto random potential. We show that the system of particles, rescaled in an appropriate way, converges in distribution to a scaling limit that is interesting in its own right. We describe the limit object as a growing collection of "lilypads" built on a Poisson point process in $\mathbb{R}^d$. As an application of our main theorem, we show that the maximizer of the system displays the ageing property.

\section{Introduction and main results} \subsection{Introduction} Consider a branching random walk in random environment defined on $\mathbb{Z}^d$, starting with a single particle at the origin. Given a collection $\xi = \{ \xi(z) \, : \, z \in \mathbb{Z}^d\}$ of non-negative random variables, when at site $z$ each particle branches into two particles at rate $\xi(z)$. Besides this, each particle moves independently as a simple random walk in continuous time on $\mathbb{Z}^d$. This model was introduced in~\cite{GM90}, and most of the analysis thus far has concentrated on the expected number of particles. Fix a realisation of the environment $\xi$ and write \[ u(z,t) = E^\xi [ \# \{ \mbox{particles at site } z \mbox{ at time } t \} ] , \] where the expectation $E^\xi$ is only over the branching and random walk mechansims and $\xi$ is kept fixed. Then $u(z,t)$ solves the stochastic partial differential equation, known as the \emph{parabolic Anderson model} (PAM), \[ \begin{aligned} \partial_t u(z,t) & = \Delta u(z,t) + \xi(z) u(z,t) , & \quad \mbox{for }z \in \mathbb{Z}^d, t \geq 0, \\ u(z,0) & = 1\hspace{-0.098cm}\mathrm{l}_{\{z = 0 \}} &\quad \mbox{for } z \in \mathbb{Z}^d . \end{aligned}\] Here, $\Delta$ is the discrete Laplacian defined for any function $f : \mathbb{Z}^d \rightarrow \mathbb{R}$ as \[ \Delta f(z) = \sum_{y \sim z} (f(y) - f(z)) , \quad z \in \mathbb{Z}^d, \] where we write $y \sim z$ if $y$ is a neighbour of $z$ in $\mathbb{Z}^d$. We are particularly interested in the case when the potential is Pareto distributed, i.e.~$\mathrm{Prob}(\xi(z)>x) = x^{-\alpha}$ for all $x\geq1$ and some $\alpha>0$. In this case, the evolution of the PAM is particularly well understood, including asymptotics for the total mass, one point localisation and a scaling limit: see~\cite{HMS08, KLMS09, MOS11,OR14}. In general much less is known about the branching system itself (without taking expectations). Some of the earlier results include~\cite{ABMY00} and~\cite{GKS13}, who look at the asymptotics of the expectation (with respect to $\xi$) of higher moments of the number of particles. The real starting point for this article is our recent article~\cite{OR14}. We showed that---in the Pareto case---the hitting times of sites, the number of particles, and the support in an appropriately rescaled system are well described by a process defined purely in terms of the environment $\xi$ (that is, given $\xi$, the process is deterministic), which we called the lilypad model. Our central aim in this article is to show that this lilypad process, and therefore the branching system itself, has a scaling limit. This limit object is entirely new, and interesting in its own right: it is neither deterministic, as for example in~\cite{CP07b} for another variant of branching random walk in random environment, nor is it a stochastic (partial) differential equation. Rather the limit is a system of interacting and growing $L^1$ balls in $\mathbb{R}^d$, centred at the points of a Poisson point process. We call this the {\em Poisson lilypad model}, and to avoid confusion we will refer to the lilypad model from~\cite{OR14} as the {\em discrete lilypad model} from now on. As an application of this characterization, we show that the dominant site in the branching process---that is, the site that has more particles than any other site---remains constant for long periods of time, in fact for periods that increase linearly as time increases. This phenomenon is known as {\em ageing}, and was demonstrated for the PAM in \cite{MOS11}. \subsection{Definitions and notation} Before we can state our results precisely, we need to develop some machinery. Throughout this article we write $|\cdot |$ for the $L^1$-norm on $\mathbb{R}^d$. $B(z,R) = \{ x \in \mathbb{R}^d : |x-z| < R \}$ denotes the open ball of radius $R$ about $z$ in $\mathbb{R}^d$, and $\mathcal{B}(z,R) = \{x\in \mathbb{R}^d : |x-z|\leq R\}$ the closed ball. For any measure $\nu$, we write $\supp\nu$ for the (measure theoretic) support of $\nu$. We take a collection of independent and identically distributed random variables $\{ \xi(z),\, z \in\mathbb{Z}^d\}$ satisfying \[ \mathrm{Prob} ( \xi(z) > x ) = x^{- \alpha} \quad \mbox{for all } x \geq 1 , \] for a parameter $\alpha > 0$ and any $z \in \mathbb{Z}^d$. We will also assume that $\alpha > d$, which is known to be necessary for the total mass of the PAM to remain finite \cite{GM90}. For a fixed environment $\xi$, we denote by $P_y^\xi$ the law of the branching simple random walk in continuous time with binary branching and branching rates $\{\xi(z)\, , \, z \in \mathbb{Z}^d\}$ started with a single particle at site $y$. Finally, for any measurable set $F \subset \Omega$, we define \[ \P_y ( F \times \cdot) = \int_F P_y^\xi ( \cdot) \,\mathrm{Prob}(d \xi) . \] If we start with a single particle at the origin, we omit the subscript $y$ and simply write $P^\xi$ and $\P$ instead of $P_0^\xi$ and $\P_0$. We define $Y(z,t)$ to be the set of particles at the point $z$ at time $t$, and let $N(z,t) = \# Y(z,t)$. We introduce a rescaling of time by a parameter $T > 0$, and then also rescale space and the potential. Setting $q=\frac{d}{\alpha-d}$, the right scaling factors turn out to be \[ a(T) = \left(\frac{T}{\log T}\right)^q \quad \mbox{and} \quad r(T) = \left(\frac{T}{\log T}\right)^{q+1} \] for the potential and space respectively. We then define the rescaled lattice as \[ L_T = \{ z \in \mathbb{R}^d \, : \, r(T) z \in \mathbb{Z}^d \}, \] and for $z\in\mathbb{R}^d$, $R\geq 0$ define $L_T(z,R) = L_T\cap B(z,R)$. For $z \in L_T$, the rescaled potential is given by \[ \xi_T(z) = \frac{\xi(r(T) z)}{a(T)}, \] and we set $\xi_T(z) = 0$ for $z \in \mathbb{R}^d \setminus L_T$. \textbf{The branching system} We are interested primarily in three functions: \[H_T(z) = \inf\{t\geq0 : Y(r(T)z,tT)\neq \emptyset\}, \] \[M_T(z,t) = \frac{1}{a(T)T}\log_+ N(r(T)z,tT), \] and \[S_T(t) = \{y\in\mathbb{R}^d : H_T(y)\leq t\}, \] for $z \in L_T$, $t \geq 0$, which we extend to $z \in \mathbb{R}^d$ by linear interpolation. We call these functions the (rescaled) hitting times, numbers of particles, and support, respectively, of the branching system. \textbf{The scaling limit: the Poisson lilypad model} In order to describe the limits of these functions as $T\to\infty$, we suppose that under $\P$ there is an independent Poisson point process $\Pi$ on $\mathbb{R}^d\times[0,\infty)$ with intensity measure $dz \otimes \alpha x^{-(\alpha+1)}dx$. We let $\Pi^\ssup{1}$ be the first marginal of $\Pi$, and write \[\Pi = \sum_{i=1}^\infty \delta_{(z_i, \xi_\Pi(z_i))}\] where $z_i$, $i=1,2,\ldots$ are the points in $\supp\Pi^\ssup{1}$. We define, for $z\in\mathbb{R}^d$ and $t\geq0$, \[ h(z) = \inf_{ y_1,y_2, \ldots \in \supp \Pi^\ssup{1}, y_n \rightarrow 0 } \bigg\{ \sum_{j=1}^\infty q \frac{| y_{j+1} - y_j|}{\xi_\Pi(y_{j+1})} + q \frac{|y_1 - z|}{\xi_\Pi(y_1)} \bigg\}, \] \[ m(z,t) = \sup_{y \in \supp\Pi^\ssup{1}} \big\{ \xi_\Pi(y) (t - h(y)) - q |y-z| \big\} \vee 0, \] and \[ s(t) = \{y\in\mathbb{R}^d : h(z)\leq t\}.\] We recall that here and throughout $|\cdot|$ denotes the $L^1$-norm on $\mathbb{R}^d$. We call these functions the hitting times, numbers of particles, and support, respectively, of the Poisson lilypad process. We think of each site $y \in \supp \Pi^\ssup{1}$ as being home to a lilypad, which grows at speed $\xi_\Pi(y)/q$. However, these lilypads only begin to grow once they are touched by another lilypad. A simulation of the process can be seen at \url{http://tiny.cc/lilypads}. We will see in Lemma~\ref{le:lilypad_nontrival} that these quantities are non-trivial, so that in particular the system does manage to start growing from the origin, and does not explode in finite time. \textbf{Topologies} Write $C(A,B)$ for the set of continuous functions from $A$ to $B$. We use the following topologies: \begin{itemize} \item For the hitting times: $\mathcal{C}^d := C(\mathbb{R}^d , [0,\infty))$, equipped with the topology of uniform convergence on compacts, i.e.\ induced by the metric \[ d_{U} ( f, g) = \sum_{n \geq 1} 2^{-n} \Big( \sup_{x \in [-n,n]^d} \{|f(x) - g(x)| \} \wedge 1\Big), \quad f, g \in \mathcal{C}^d. \] \item For the number of particles: \[\mathcal{C}_0^{d+1} := \{f\in C(\mathbb{R}^d\times[0,\infty), [0,\infty)) : f(x,t)\to 0 \hbox{ as } x\to\infty\,\,\, \forall t\in[0,\infty)\},\] equipped with the topology induced by the metric \[ d_{P} (f,g) = \sum_{n \geq 1} 2^{-n} \Big( \sup_{x\in\mathbb{R}^d, t \in [0,n]} \{|f(x,t) - g(x,t)| \} \wedge 1\Big), \quad f, g \in \mathcal{C}_0^{d+1}. \] \item For the support: $\mathcal{C}_F := C([0,\infty), F(\mathbb{R}^d))$, equipped with the topology induced by the metric \[ d_{F} ( f, g) = \sum_{n \geq 1} 2^{-n} \Big( \sup_{t \in [0,n]} \{d_H(f(t),g(t)) \} \wedge 1\Big), \quad f, g \in \mathcal{C}_F, \] where $F(\mathbb{R}^d)$ is the space of non-empty compact subsets of $\mathbb{R}^d$ and $d_H$ is the Hausdorff distance on $F(\mathbb{R}^d)$. \end{itemize} Finally, we consider $(H_T, M_T, S_T)$ and $(h,m,s)$ as elements in the product space $\mathcal{C}^{\ssup{ \times 3}} := \mathcal{C}^d \times \mathcal{C}^{d+1}_0 \times \mathcal{C}_F$ equipped with the product topology, which is, for example, induced by the metric \[ d^\ssup{\times 3} ((H, M, S), ( H', M', S')) = d_{U} ( H, H') + d_{P} ( M, M') + d_{F}( S, S') , \] for any $(H, M, S), ( H', M', S') \in \mathcal{C}^\ssup{\times 3}$. \subsection{Main results} Our main theorem states that the rescaled branching system (hitting times, number of particles and support) converges weakly to the Poisson lilypad model. For background on weak convergence, we refer to~\cite{Bil99, EthierKurtz86}. \begin{thm}\label{thm:weak_conv} The triple $(H_T, M_T, S_T)$ converges weakly in $\mathcal{C}^\ssup{\times 3}$ as $T \rightarrow \infty$ to $(h,m,s)$. \end{thm} As an application, we show that the maximal site in the branching system---that is, the site with the most particles at a given time---shows {\em ageing} behaviour. Denote by $Z^{\rm max}(t)$ this site: that is, \[N(Z^{\rm max}(t),t) \geq N(z,t) \quad \forall z\in\mathbb{Z}^d;\] in case of a tie choose the point with larger potential. Introduce the rescaled version \[ W_T(t) := Z^{\rm max}(tT) / r(T). \] Also let $w(t)$ be the maximizer in the Poisson lilypad model, \[ m(w(t),t) \geq m(z,t) \quad \forall z\in\mathbb{R}^d;\] again in the case of a tie we choose the site with larger potential (although we will show in Lemma \ref{le:Poisson_max_unique} that for any $t\geq0$ there is almost surely a unique maximizer for the Poisson lilypad model). \begin{thm}\label{thm:ageing} \emph{Ageing.} For any $\theta > 0$, \[ \P ( Z^{\rm max}_{T} = Z^{\rm max}_{(1+\theta) T} ) = \P( W_T(1) = W_T(1+ \theta)) \rightarrow \P ( w(1) = w(1+ \theta) ) . \] \end{thm} In the companion paper~\cite{onepoint}, we show that with high probability, the total mass of the branching process is concentrated in a single point, so the theorem really describes ageing, i.e.\ the temporal slow-down, of this maximizer. The strategy of proof of Theorem~\ref{thm:weak_conv} relies on our previous result from~\cite{OR14}, which shows that the branching system is well described by a functional purely of the environment, which we call the discrete lilypad model and recall in Section~\ref{ssn:discrete_lilypad}. Then, our main task is to show that the discrete lilypad model converges to the Poisson lilypad model that we described above. The underlying reason is that the rescaled environment converges to a Poisson process; see Section~\ref{ssn:point_pr} for some background. The proof of Theorem~\ref{thm:weak_conv} is then an application of the continuous mapping theorem for a suitable continuous approximation of the lilypad models, which we describe in Section~\ref{sec:proof_scaling}. This approach allows us to avoid some of the technicalities involved with a more traditional approach of showing tightness combined with the convergence of finite dimensional distributions. The proof of Theorem~\ref{thm:ageing} in Section~\ref{sec:proof_ageing} is then an application of the scaling limit. Throughout the article, the ideas remain fairly simple, but there are many technicalities due to the highly sensitive nature of the model. For example, if one site of large potential is hit slightly earlier or later than it should be, the whole system could be affected dramatically. We have to keep track of several events that could, feasibly, occur; show that they have small probability; and show that if these events do not occur then the system behaves as we claim. \subsection{The discrete lilypad model}\label{ssn:discrete_lilypad} In~\cite{OR14}, we showed that the branching system is well-approximated by certain functionals of the environment, which we will refer to as the \emph{discrete lilypad model}. For any site $z \in L_T$, we set \[ h_T(z) = \inf_{ \substack{y_0,\ldots,y_n \in L_T:\\ y_0 = z, y_n = 0}} \Bigg( \sum_{j=1}^{n} q\frac{| y_{j-1} - y_{j}|}{\xi_T(y_{j})} \Bigg) .\] We call $h_T(z)$ the first hitting time of $z$ in the discrete lilypad model. We think of each site $y$ as being home to a lilypad, which grows at speed $\xi_T(y)/q$. Note that $h_T(0)=0$. For convenience, we interpolate $h_T$ linearly to define the values for $z \notin L_T$. The rescaled number of particles in the discrete lilypad model is defined as \[m_T(z,t) = \sup_{y \in L_T}\{\xi_T(y)(t-h_T(y)) - q|z-y|\}\vee 0. \] Also, we define the support of particles at time $t$ in the discrete lilypad model as \[ s_T(t) = \{ z \in \mathbb{R}^d \, : \, h_T(z) \leq t \}. \] We recall here the main result from~\cite{OR14}, which can be phrased as: \begin{thm}[\cite{OR14}]\label{thm:discrete_lilypad} For any $t_\infty > 0$, as $T \rightarrow \infty$, \[ \begin{aligned} \sup_{t \leq t_\infty} \sup_{z \in L_T} | M_T(z,t) - m_T(z,t) | \rightarrow 0 \quad \mbox{ in } \P\mbox{-probability} . \end{aligned} \] Moreover, for any $R > 0$, as $T \rightarrow \infty$, \[ \sup_{z \in L_T(0,R)} | H_T(z) - h_T(z) | \rightarrow 0 \quad \mbox{ in } \P\mbox{-probability} , \] and for any $t_\infty >0$, as $T \rightarrow \infty$, \[ \sup_{t \leq t_\infty} d_H( S_T(t), s_T(t)) \rightarrow 0 \quad \mbox{in } \P\mbox{-probability}. \] \end{thm} We reiterate here the general idea behind this article: we know from Theorem \ref{thm:discrete_lilypad} that the branching system is well-approximated (with high probability) by the discrete lilypad model, which is a deterministic functional of the environment $\xi$. We can check that the distribution of $\xi$ (suitably rescaled) converges weakly to that of a Poisson point process; and this allows us to show that the discrete lilypad model converges weakly to the Poisson lilypad model. \subsection{Background on point processes}\label{ssn:point_pr} The proof of our main result, Theorem~\ref{thm:weak_conv}, is a consequence of the convergence of the rescaled environment to a Poisson process. In this section we recall some of the standard definitions concerning point processes. We consider the point process \[ \Pi_T := \sum_{ z\in \mathbb{Z}^d} \delta_{( \frac{z}{r(T)}, \frac{\xi_T(z)}{a(T)}) } . \] on $\mathbb{R}^d \times (0, \infty) $. A classical result in extreme value theory shows that $\Pi_T$ converges in law to the Poisson point process $\Pi$ on $\mathbb{R}^d \times (0,\infty)$ with intensity measure \[ \pi( d (z,x) ) = dz \otimes \, \frac{\alpha}{x^{\alpha+1}}\,d x . \] In order to formalize this convergence we follow the basic setup from~\cite{HMS08}, which is based on~\cite{Res08}. Let $E$ be a locally compact space with a countable basis and let $\mathcal{E}$ denote the Borel-$\sigma$-algebra on $E$. A Radon measure is a Borel measure that is locally finite. If in addition $\mu = \sum_{ i \geq 1} \delta_{x_i}$ for a countable collection of points $\{ x_i, i\ge 1 \} \subset E$, then $\mu$ is called a point measure. We write $M_p(E)$ for the set of all point measures on $E$. We equip the set of Radon measures $M_+(E)$ with the vague topology: i.e.\ $\mu_n \rightarrow \mu$ vaguely, if for any continuous function $f : E \rightarrow \mathbb{R}$ with compact support $\int f \mu_n \rightarrow \int f \mu$. Note that $M_p(E)$ is vaguely closed in $M_+(E)$ (cf.~\cite[Prop.~3.14]{Res08}). In our case we set $E = \mathbb{R}^d \times (0,\infty]$, where the topology on $(0,\infty]$ is understood such that closed neighbourhoods of $\infty$ are compact. Note that $\Pi_T$ and $\Pi$ are elements of $M_p(E)$ for this choice. Then the above convergence means that $\Pi_T \Rightarrow \Pi$ in the topology on $M_p(E)$ induced by vague convergence. This fact is a direct application of~\cite[Prop.~3.21]{Res08} (where $\mathbb{R}^d$ replaces $\mathbb{R}^+$ as the index set). \section{Proof of the scaling limit}\label{sec:proof_scaling} In this section we prove the main scaling limit, Theorem~\ref{thm:weak_conv}. By our previous result on the approximation via the discrete model, Theorem~\ref{thm:discrete_lilypad}, it suffices to show convergence of the discrete lilypad model. Our main strategy is to use the continuous mapping theorem to deduce the convergence of $(h_T, m_T,s_T)$ from the convergence of the point process $\Pi_T$ to $\Pi$. Unfortunately, however, it is not clear that $(h_T,m_T,s_T)$ is a continuous function of the underlying point process. Our way around this problem is to define an $\delta$-approximate lilypad model for both the discrete space version and the Poisson model. By ignoring potential values less than $\delta$---and, later, restricting in space to $B(0,1/\delta)$---we obtain functionals that only depend on a finite set of points and are therefore continuous. We can treat both the discrete space and the Poisson case in the same way. Thus, for $\nu = \Pi_T$, for some $T>0$, or $\nu = \Pi$, we write $\nu\in M_p(E)$ as \[ \nu = \sum_{i \geq 1} \delta_{(z_i , \xi_\nu(z_i))} ,\] and write $\nu^\ssup{1} (\cdot) := \nu ( \,\,\cdot \times [0,\infty))$ for the first marginal of $\nu$. For $r>0$, we write $B_\nu(0,r) = \supp(\nu^\ssup{1})\cap B(0,r)$ and $\mathcal{B}_\nu(0,r)=\supp(\nu^\ssup{1})\cap \mathcal{B}(0,r)$. Where it is clear which point process we are referring to, we write $\xi(z)$ in place of $\xi_\nu(z)$ for conciseness. (Of course, we have already defined $\{\xi(z):z\in\mathbb{Z}^d\}$ to be a collection of i.i.d.~Pareto random variables; but since we already know from Theorem \ref{thm:discrete_lilypad} that the branching process is well approximated by the discrete lilypad model, which can be described via the point process $\Pi_T$, we no longer need this original meaning and $\xi(z)$ will always refer to $\xi_\nu(z)$ for some point process $\nu$.) For a general point process $\nu$, we define the hitting times by setting $h_\nu(0)=0$ and, for $z\in\mathbb{R}^d\setminus\{0\}$, \[h_\nu(z) = \inf \Big\{ q \sum_{i=1}^\infty \frac{|y_i - y_{i-1}|}{\xi(y_i)} \, : \, y_0 = z, y_i \in \supp\nu^\ssup{1} \ \forall i \geq 1\, , \, |y_i | \rightarrow 0 \Big\} . \] The number of particles is defined as \[m_\nu(z,t) = \sup_{y \in \supp \nu^\ssup{1}} \Big\{ \xi(y) ( t - h_\nu(y) ) - q |y-z| \Big\} \vee 0 , \quad z \in \mathbb{R}^d, t \geq 0, \] and the support is defined as \[ s_\nu(t) = \{ z \in \mathbb{R}^d \, : \, h_\nu(z) \leq t \} , \quad t \geq 0 . \] We also define the $\delta$-hitting times by setting \[ h^\delta_\nu(z) = \inf \Big\{ \sum_{j=1}^n q \frac{|y_{j-1} - y_j|}{\xi(y_j)} + q\frac{|y_n|}{\delta} \, : \,n\in\mathbb{N}_0, y_0 = z \mbox{ and }y_1, \ldots, y_{n} \in \supp\nu^\ssup{1} \Big\} \] for any $z\in\mathbb{R}^d$ (note that we allow $n = 0$, in which case we do not insist on $y_n \in \supp(\nu^\ssup{1})$). Effectively, considering $h^\delta_\nu(z)$ rather than $h_\nu(z)$ gives all lilypads a ``minimum speed'' $\delta/q$, which helps in showing the continuity of the process as a function of the point measure $\nu$. In analogy with the definitions above, we also define the $\delta$-number of particles and the $\delta$-support via \[ m_\nu^\delta(z,t) = \sup_{y \in \supp \nu^\ssup{1}} \Big\{ \xi(y) ( t - h_\nu^\delta(y) ) - q |y-z| \Big\} \vee 0 , \quad z \in \mathbb{R}^d, t \geq 0, \] and \[ s_\nu^\delta(t) = \{ z \in \mathbb{R}^d \, : \, h_\nu^\delta(z) \leq t \} , \quad t \geq 0 . \] We write $(h_T^\delta, m_T^\delta, s_T^\delta) := (h_{\Pi_T}^\delta, m_{\Pi_T}^\delta, s_{\Pi_T}^\delta)$ and $(h^\delta, m^\delta, s^\delta) := (h_{\Pi}^\delta, m_{\Pi}^\delta, s_{\Pi}^\delta)$. The main technical result of this section is the following proposition. \begin{prop}\label{prop:appr_works} For any $\varepsilon > 0$, \[\lim_{\delta \downarrow 0} \limsup_{T\to\infty}\P \Big( d^{(\times 3)} \big( ( h_T^\delta, m_T^\delta, s_T^\delta ), (h_T,m_T,s_T) \big) \geq \varepsilon \Big) = 0, \] and analogously for the Poisson point process \[\lim_{\delta \downarrow 0} \P \Big( d^{(\times 3)} \big( ( h^\delta, m^\delta, s^\delta ), (h,m,s) \big) \geq \varepsilon \Big) = 0 . \] \end{prop} The remainder of this section is organised as follows. In Section~\ref{ssn:delta_hitting}, we give general criteria on the point process $\nu$ that ensure that the $\delta$-hitting times approximate well the actual hitting times. Then in Section~\ref{ssn:delta_support} we show that this result can be transferred to the number of particles and the support. In Section~\ref{ssn:delta_works} we show that these general criteria are satisfied by the point processes $\Pi_T$ and $\Pi$, and we prove Proposition~\ref{prop:appr_works}. Finally, in Section~\ref{ssn:proof_scaling}, we show that the $\delta$-processes for $\Pi_T$ converge to the $\delta$-processes for $\Pi$, and we combine these results to show the statement of the main scaling limit Theorem~\ref{thm:weak_conv}. \subsection{The $\delta$-approximation of the hitting times} \label{ssn:delta_hitting} We now state certain assumptions on the point process $\nu$ under which $h_\nu^\delta$ and $h_\nu$ will be close when $\delta$ is small. Let $\gamma = \frac{d+\alpha}{2\alpha}$. \begin{enumerate}[label={(A\arabic*)}] \item \label{asmp:pot_large} For all $R\geq R_0$, $\sup_{y \in B_\nu(0,R)} \xi(y) \leq q R^\gamma$. \item \label{asmp:small_pts} For all $r\leq r_0$, for all $k\in \mathbb{N}_0$, there exists $Z_k \in B_\nu(0,r2^{-k})$ such that $\xi(Z_k) \geq r^\gamma 2^{-k\gamma}$. \end{enumerate} We write \ref{asmp:pot_large}$_{R_0}$ and \ref{asmp:small_pts}$_{r_0}$ to emphasize the dependence of the conditions on the parameters. The main result in this subsection states that the hitting times are approximated well by the $\delta$-hitting times, provided $\nu$ satisfies the above conditions. \begin{prop}\label{prop:del_app_general} Suppose that $\nu$ satisfies~\ref{asmp:pot_large}$_{R_0}$ and~\ref{asmp:small_pts}$_{r_0}$. Then for any $\varepsilon>0$, there exists $\delta>0$ (depending only on $\gamma$, $\varepsilon$, $R_0$ and $r_0$) such that \[h_\nu^\delta(z) \leq h_\nu(z) \leq h_\nu^\delta(z)+\varepsilon \quad \forall z\in \mathbb{R}^d.\] \end{prop} We will also need the following two simple lemmas, which prove upper and lower bounds on the hitting times. \begin{lem}\label{le:small_times} Suppose that $\nu$ satisfies ~\ref{asmp:small_pts}$_{r_0}$. Then for any $r\leq r_0$, \[ \max_{z \in B(0,r)} h_\nu(y) \leq \frac{ 4 q r^{1- \gamma}}{1 - 2^{\gamma-1}} \] and moreover, for any $z\in\mathbb{R}^d$, \[ h_\nu(z) \leq \frac{ 4 q r_0^{1- \gamma}}{1 - 2^{\gamma-1}} + q ( r_0^{-\gamma} |z| + r_0^{1 - \gamma} ). \] \end{lem} \begin{lem}\label{le:large_hitting} Suppose that $\nu$ satisfies \ref{asmp:pot_large}$_{R_0}$. Then for any $R\geq R_0$ and any $\delta>0$, \[ \inf_{y \not\in B(0,R) } h^\delta_\nu(y) \geq \min\{R^{1-\gamma}, qR/\delta\}.\] \end{lem} The lemmas lead easily to two useful corollaries. \begin{cor}\label{cor:restrict} Suppose that $\nu$ satisfies~\ref{asmp:pot_large}$_{R_0}$. Then for any $z\in\mathbb{R}^d$ and any $\delta>0$, there exists $R>0$ (depending only on $\gamma$, $R_0$ and $\delta$) such that the infimum in the definition of $h^\delta_\nu(z)$ can be restricted to points $y_1,\ldots,y_n\in B_\nu(0,R)$. \end{cor} \begin{cor}\label{cor:nontriv} Suppose that $\nu$ satisfies~\ref{asmp:pot_large}$_{R_0}$ and~\ref{asmp:small_pts}$_{r_0}$ for some $R_0$ and $r_0$. Then for all $z\in\mathbb{R}^d\setminus\{0\}$ and all $\delta>0$, we have \[0<h_\nu^\delta(z)\leq h_\nu(z)<\infty.\] \end{cor} We delay the proofs of the lemmas and corollaries for a moment to concentrate on Proposition \ref{prop:del_app_general}. \begin{proof}[Proof of Proposition \ref{prop:del_app_general}] The fact that $h_\nu^\delta(z) \leq h_\nu(z)$ for all $z\in\mathbb{R}^d$ follows immediately from the definitions, so we aim to prove that $h_\nu(z) \leq h_\nu^\delta(z)+\varepsilon$. Since $\gamma<1$ we may choose $\delta>0$ small enough so that \begin{equation}\label{eq:del1} 4q\delta^{1-\gamma}\Big(\frac{1}{1-2^{\gamma-1}}\Big) \leq \varepsilon, \end{equation} \begin{equation}\label{eq:del2} (\delta/4)^{\gamma} \geq 2\delta \quad \mbox{ and } \quad \delta\leq r_0. \end{equation} By Corollary \ref{cor:restrict}, there exists $R>0$ such that the infimum in the definition of $h_\nu^\delta(z)$ is taken over points $y_1,\ldots,y_n\in B_\nu(0,R)$; we also note from the definition that necessarily $\xi(y_i)\geq\delta$ for each $i=1,\ldots,n$. Since the set $B(0,R)\times[\delta,\infty)$ is relatively compact in $E$, and $\nu$ is a Radon measure, there are only finitely many such points. Thus the infimum is actually a minimum, and we can find points $y_0=z,y_1,\ldots,y_n$ such that \begin{equation}\label{eq:2212-2}h_\nu^\delta(z) = \sum_{i=1}^{n} q \frac{|y_{i-1}- y_i|}{\xi(y_i)} + q \frac{|y_n|}{\delta}.\end{equation} Note from the definition of $h_\nu$ that \[h_\nu(z) \leq h_\nu(y_n) + q\sum_{i=1}^{n} \frac{|y_i-y_{i-1}|}{\xi(y_i)} \leq h_\nu(y_n) + h_\nu^\delta(z),\] so it remains to prove that $h_\nu(y_n)\leq \varepsilon$. By Lemma \ref{le:small_times} and the fact that $\delta\leq r_0$, together with \eqref{eq:del1}, we have \[ \max_{y\in B(0,\delta)} h_\nu(y) \leq \varepsilon.\] Thus it suffices to prove that $|y_n|<\delta$. By \ref{asmp:small_pts}$_{r_0}$ with $r=\delta\leq r_0$ and $k=2$, we can choose $Z\in B(0,\delta/4)$ such that $\xi(Z)\geq (\delta/4)^{\gamma} \geq 2\delta$ by~\eqref{eq:del2}. Suppose that $|y_n|\geq\delta$. Then \[ \frac{|Z|}{\delta} + \frac{|Z - y_n|}{\xi(Z)} \leq \frac{|Z|}{\delta} + \frac{|Z - y_n|}{2 \delta} \leq \frac{|Z|}{\delta} + \frac{|Z|}{2\delta} + \frac{|y_n|}{2 \delta} \leq \frac{3}{2} \frac{\delta/4}{\delta} + \frac{|y_n|}{2 \delta} < \frac{|y_n|}{\delta} . \] Thus by including $Z$ in the approximating sequence we get a smaller value of $h_\nu^\delta(z)$ than~\eqref{eq:2212-2}, contradicting the optimality of the sequence $y_0,\ldots,y_n$. We deduce that $|y_n| < \delta$ as required. \end{proof} We now proceed with the proofs of the lemmas. Lemma \ref{le:small_times} follows easily from the assumption~\ref{asmp:small_pts}$_{r_0}$: \begin{proof}[Proof of Lemma \ref{le:small_times}] Fix $r\leq r_0$ and let $Z_k$, $k \geq 0$, be as in~\ref{asmp:small_pts}$_{r_0}$. Then by definition, for any $z \in B(0,r)$, we have \[ h_\nu(z) \leq q \frac{|z - Z_0|}{\xi(Z_0)} + q \sum_{j = 1}^\infty \frac{|Z_{j-1} - Z_{j}|}{\xi(Z_{j})} \leq 2q \frac{r}{r^\gamma} + q \sum_{j=1}^\infty \frac{ 2 r 2^{-(j-1)} }{ r^\gamma 2^{- \gamma j }} \leq 4 q r^{1- \gamma} \frac{1}{1 - 2^{\gamma-1}} .\] For the second claim, taking $r=r_0$ in the above, we have that for any $z$, \[ h_\nu(z) \leq h_\nu(Z_0)+ q\frac{|z- Z_0|}{\xi(Z_0)} \leq \frac{ 4 q r_0^{1- \gamma}}{1 - 2^{\gamma-1}} + \frac{q (|z| + r_0)}{r_0^\gamma}.\qedhere \] \end{proof} Lemma \ref{le:large_hitting} is slightly more fiddly. \begin{proof}[Proof of Lemma \ref{le:large_hitting}] It is easy to see from the definition that $z\mapsto h^\delta_\nu(z)$ is continuous. Therefore there exists a point $\tilde z \in \partial B(0,R)=\{z\in\mathbb{R}^d : |z|=R\}$ that minimizes $h^\delta_\nu$, i.e.\ \[h^\delta_\nu(\tilde z) = \inf_{ y \in \partial B(0,R)} h_\nu^\delta(y).\] We claim that $h^\delta_\nu(\tilde z) = \inf_{|y| \geq R} h_\nu^\delta(y)$. Indeed, suppose there exists $y\not\in B(0,R)$ with $h_\nu^\delta(y) < h_\nu^\delta(\tilde z)$. Then we can choose $y_0=z, y_1,\ldots, y_n$ with \[q\sum_{j=1}^n \frac{|y_j-y_{j-1}|}{\xi(y_j)} + q\frac{|y_n|}{\delta} < h_\nu^\delta(\tilde z).\] We may assume without loss of generality that $y_1\in B(0,R)$ (since clearly $h_\nu^\delta(y_1) < h_\nu^\delta(\tilde z)$, so we can otherwise use $y_1$ in place of $y$). Therefore there exists $a\in(0,1)$ such that $\tilde y:=y_1 + a(y-y_1)\in \partial B(0,R)$. Then \begin{align*} h_\nu^\delta(\tilde y) \leq q\frac{|y_1-\tilde y|}{\xi(y_1)} + q\sum_{j=2}^n \frac{|y_j-y_{j-1}|}{\xi(y_j)} + q\frac{|y_n|}{\delta} &= qa\frac{|y_1- y|}{\xi(y_1)} + q\sum_{j=2}^n \frac{|y_j-y_{j-1}|}{\xi(y_j)} + q\frac{|y_n|}{\delta}\\ & < q\sum_{j=1}^n \frac{|y_j-y_{j-1}|}{\xi(y_j)} + q\frac{|y_n|}{\delta} = h_\nu^\delta(\tilde z), \end{align*} contradicting the choice of $\tilde z$. Therefore the claim holds. Since $h_\nu^\delta(y) > h_\nu^\delta(\tilde z)$ for all $y\not\in B(0,R)$, we see that the infimum in the definition of $h_\nu^\delta(\tilde z)$ can be restricted to points within $B(0,R)$: that is, \[h_\nu^\delta (\tilde z) = \inf\Big\{q \sum_{j=1}^n \frac{|y_{j-1} - y_j|}{\xi(y_j)} + q \frac{|y_n|}{\delta} \, : \,n\in\mathbb{N}_0, y_0 = x \mbox{ and }y_1, \ldots, y_{n} \in B_\nu(0,R) \Big\}.\] In particular, \[h_\nu^\delta(\tilde z) \geq \min\Big\{ \frac{qR}{\max_{y \in B_\nu(0,R)} \xi(y)} , \frac{qR}{\delta} \Big\},\] and therefore by~\ref{asmp:pot_large}$_{R_0}$, if $R\geq R_0$ then \[\inf_{y\not\in B(0,R)} h_\nu^\delta(y) \geq \min\{ R^{1-\gamma}, qR/\delta\}.\qedhere\] \end{proof} \begin{proof}[Proof of Corollary \ref{cor:restrict}] Fix $z\in\mathbb{R}^d$. By Lemma \ref{le:large_hitting}, we can choose $R$ large enough such that \[\inf_{y\not\in B(0,R)} h_\nu^\delta(y) > h_\nu^\delta(z).\] Therefore the infimum in the definition of $h_\nu^\delta(z)$ can be restricted to points within $B(0,R)$. \end{proof} \begin{proof}[Proof of Corollary \ref{cor:nontriv}] Take any $z\in\mathbb{R}^d\setminus\{0\}$ and $\delta>0$. By Corollary \ref{cor:restrict}, there exists $R>0$ such that the infimum in the definition of $h^\delta_\nu(z)$ can be restricted to points within $B(0,R)$, so \[h_\nu^\delta(z) \geq \min\Big\{\frac{q|z|}{\max_{y\in B_\nu(0,R)}\xi(y)},\frac{q|z|}{\delta}\Big\} > 0.\] The fact that $h_\nu^\delta(z)\leq h_\nu(z)$ follows directly from the definitions; and $h_\nu(z)<\infty$ by Lemma \ref{le:small_times}. \end{proof} \subsection{The $\delta$-approximation of the support and number of particles}\label{ssn:delta_support} We recall that \[ m_\nu(z,t) = \sup_{y \in \supp \nu^\ssup{1}} \Big\{ \xi(y) ( t - h_\nu(y) ) - q |y-z| \Big\} \vee 0 , \quad z \in \mathbb{R}^d, t \geq 0 \] and \[ s_\nu(t) = \{ z \in \mathbb{R}^d \, : \, h_\nu(z) \leq t \} , \quad t \geq 0, \] and that $m_\nu^\delta(z,t)$ and $s^\delta_\nu(t)$ are defined similarly by replacing $h_\nu$ by $h_\nu^\delta$. In this subsection, we show that under~\ref{asmp:pot_large}$_{R_0}$ and~\ref{asmp:small_pts}$_{r_0}$, the $\delta$-approximations $m_\nu^\delta$ and $s_\nu^\delta$ are close to $m_\nu$ and $s_\nu$ respectively. We start by showing that the growth of the support $s_\nu$ is well-controlled. This will be key to controlling the Hausdorff distance between $s_\nu$ and $s_\nu^\delta$. \begin{lemma}\label{le:le61} Suppose that $\nu$ satisfies~\ref{asmp:pot_large}$_{R_0}$. For any $\varepsilon>0$ and any $t_0>0$, there exists $\eta \in (0,1)$ (depending only on $\gamma$, $\varepsilon$, $t_0$ and $R_0$) such that \[ s_\nu(t+\eta) \subseteq \bigcup_{y \in s_\nu(t)} B(y,\varepsilon) \quad \forall t\leq t_0. \] \end{lemma} \begin{proof} By Lemma \ref{le:large_hitting}, together with the fact that $h_\nu(z)\geq h_\nu^\delta(z)$ for all $z$, we can choose $R\geq R_0$ such that $h_\nu(y)>t_0+1$ for all $y\not\in B(0,R)$. Then set $\eta = \frac{\varepsilon}{2R^\gamma}\wedge \frac12$. Suppose that $z\in s_\nu(t+\eta)\setminus s_\nu(t)$; then $h_\nu(z)\in(t,t+\eta]$, so we can find $y_0=z,y_1,y_2\ldots\to0$ with $h_\nu(z)\leq q\sum_{i=1}^\infty \frac{|y_i-y_{i-1}|}{\xi(y_i)} \leq t+2\eta$. Since $h_\nu(y)>t_0+1$ for all $y\not\in B(0,R)$, we must have $y_1,y_2,\ldots \in B(0,R)$. Choose $k$ such that $q\sum_{i=k+1}^\infty \frac{|y_i-y_{i-1}|}{\xi(y_i)} \leq t$ and $q\sum_{i=k}^\infty \frac{|y_i-y_{i-1}|}{\xi(y_i)} > t$. Then choose $a\in[0,1)$ such that \[q\sum_{i=k+1}^\infty \frac{|y_i-y_{i-1}|}{\xi(y_i)} + aq\frac{|y_k-y_{k-1}|}{\xi(y_k)} = t.\] Setting $\tilde y = y_k + a(y_k-y_{k-1})$, by the above we have $h_\nu(\tilde y)\leq t$, so $\tilde y\in s_\nu(t)$. On the other hand, \begin{multline*} q\sum_{i=k+1}^\infty \frac{|y_i-y_{i-1}|}{\xi(y_i)} + aq\frac{|y_k-y_{k-1}|}{\xi(y_k)}\\ = q\sum_{i=1}^\infty \frac{|y_i-y_{i-1}|}{\xi(y_i)} - (1-a)q\frac{|y_k-y_{k-1}|}{\xi(y_k)} - q\sum_{i=1}^{k-1} \frac{|y_i-y_{i-1}|}{\xi(y_i)}, \end{multline*} so (since the left-hand side equals $t$ and the first sum on the right-hand side is at most $t+2\eta$) we must have \[(1-a)q\frac{|y_k-y_{k-1}|}{\xi(y_k)} + q\sum_{i=1}^{k-1} \frac{|y_i-y_{i-1}|}{\xi(y_i)} \leq 2\eta.\] By the triangle inequality, we get \[|\tilde y-z|=\Big|(1-a)(y_k-y_{k-1}) + \sum_{i=1}^{k-1} (y_i-y_{i-1})\Big| \leq \frac{2\eta}{q}\sup_{y\in B_\nu(0,R)}\xi(y),\] and by~\ref{asmp:pot_large}$_{R_0}$ and the fact that $\eta \leq \varepsilon/(2R^\gamma)$, we have $|\tilde y-z|\leq \varepsilon$. Since $\tilde y\in s_\nu(t)$ this completes the proof. \end{proof} We can now apply Proposition \ref{prop:del_app_general} together with Lemma \ref{le:le61} to prove our main result for this section. \begin{prop}\label{prop:del_apr_number} Suppose that $\nu$ satisfies~\ref{asmp:pot_large}$_{R_0}$ and~\ref{asmp:small_pts}$_{r_0}$. For any $\varepsilon > 0$ and $t_0 > 0$, there exists $\delta>0$ (depending only on $\gamma$, $\varepsilon$, $t_0$, $R_0$ and $r_0$) such that \[m_\nu(z,t) \leq m_\nu^\delta(z,t)\leq m_\nu(z,t)+\varepsilon \quad \mbox{for all } z \in \mathbb{R}^d,\] and \[d_H(s_\nu(t), s_\nu^\delta(t))\leq \varepsilon\] for all $t\in[0,t_0]$ and $z\in\mathbb{R}^d$. \end{prop} \begin{proof} We start by showing the statement about the supports, $s_\nu$ and $s_\nu^\delta$. By Lemma \ref{le:le61} we can choose $\eta>0$ such that \[s_\nu(t+\eta) \subseteq \bigcup_{y \in s_\nu(t)} B(y,\varepsilon) \quad \forall t\leq t_0.\] Then by Proposition \ref{prop:del_app_general} we can choose $\delta>0$ such that \[h_\nu^\delta(z) \leq h_\nu(z) \leq h_\nu^\delta(z)+\eta \quad \forall z\in\mathbb{R}^d.\] We get \[z\in s_\nu(t) \quad\Rightarrow\quad h_\nu(z)\leq t \quad\Rightarrow\quad h_\nu^\delta(z)\leq t \quad\Rightarrow\quad z\in s_\nu^\delta(t),\] and \[z\in s_\nu^\delta(t) \quad\Rightarrow\quad h_\nu^\delta(z)\leq t \quad\Rightarrow\quad h_\nu(z)\leq t+\eta \quad\Rightarrow\quad z\in s_\nu(t+\eta),\] so \[s_\nu(t) \subset s_\nu^\delta(t) \subset \bigcup_{y\in s_\nu(t)} B(y,\varepsilon).\] This implies that $d_H(s_\nu(t),s_\nu^\delta(t))\leq \varepsilon$ as required. We now turn our attention to the numbers of particles, $m_\nu$ and $m_\nu^\delta$. By Lemma~\ref{le:large_hitting} we can choose $R>R_0$ such that $h_\nu^\delta(z)>t_0$ for all $z\not\in B(0,R)$ and all $\delta\in(0,1]$. Then by Proposition \ref{prop:del_app_general} we can choose $\delta\in(0,1]$ such that $h_\nu^\delta(z)\leq h_\nu(z)\leq h_\nu^\delta(z)+\varepsilon/(qR^\gamma)$ for all $z\in\mathbb{R}^d$. Then, straight from the definitions, we have \[m_\nu(z,t)\leq m_\nu^\delta(z,t) \leq m_\nu(z,t) + \sup_{y\in B_\nu(0,R)}\xi(y)\frac{\varepsilon}{qR^\gamma}\] for all $z\in\mathbb{R}^d$ and $t\leq t_0$. By~\ref{asmp:pot_large}$_{R_0}$ the right-hand side is at most $m_\nu(z,t)+\varepsilon$. Finally, since $h_\nu^\delta(z) \leq h_\nu(z)$ for all $z \in \mathbb{R}^d$ and $h_\nu^\delta$ is increasing as $\delta \downarrow 0$, the event $\{ h_\nu^\delta(z) \leq h(z) \leq h_\nu^\delta(z) + \varepsilon \}$, and therefore the events $\{m_\nu(z,t) \leq m_\nu^\delta(z,t)\leq m_\nu(z,t)+\varepsilon\}$ and $\{s_\nu(t) \subset s_\nu^\delta(t) \subset \bigcup_{y\in s_\nu(t)} B(y,\varepsilon)\}$, are increasing as $\delta\downarrow 0$ for any $\varepsilon > 0$. In particular, we can choose the same $\delta$ for both the support and the number of particles. \end{proof} \subsection{The $\delta$-approximation works}\label{ssn:delta_works} Our aim in this section is to show that the $\delta$-approximations converge (in a suitable sense) as $\delta\downarrow0$ to the quantities they are supposed to approximate. In particular we will prove Proposition \ref{prop:appr_works}. We first show that conditions \ref{asmp:pot_large} and \ref{asmp:small_pts} hold with high probability for both $\Pi_T$ and $\Pi$. \begin{lem}\label{le:conditions_satisfied} As $R_0\to\infty$, \[\P\big( \Pi \hbox{ satisfies \ref{asmp:pot_large}$_{R_0}$}) \to 1 \quad\mbox{ and }\quad \liminf_{T\to\infty}\P\Big( \Pi_T \hbox{ satisfies \ref{asmp:pot_large}$_{R_0}$}\Big) \to 1,\] and as $r_0\to0$, \[\P\big( \Pi \hbox{ satisfies \ref{asmp:small_pts}$_{r_0}$}) \to 1 \quad\mbox{ and }\quad \liminf_{T\to\infty}\P\Big( \Pi_T \hbox{ satisfies \ref{asmp:small_pts}$_{r_0}$}\Big) \to 1.\] \end{lem} \begin{proof} Define the event $A_k(\nu) = \{ \max_{z \in A_\nu(0,2^{k}) } \xi(z) \leq q 2^{(k-1) \gamma}\}$. By~\cite[Lemma 2.7(ii)]{OR14}, there exists a constant $C$ such that for any $T>e$ and any $k\geq 0$, \[ \P (A_k(\Pi_T)^c) \leq C 2^{dk} (q2^{\gamma (k-1) } )^{-\alpha} = C 2^{\alpha\gamma} q^{-\alpha} 2^{(d-\gamma\alpha)k}. \] Similarly, by direct calculation, there exists a constant $C$ such that for any $R\geq 1$, \[ \P (A_k(\Pi)^c) \leq 1-e^{-C 2^{\alpha\gamma} q^{-\alpha} 2^{(d-\gamma\alpha) k }} \leq C 2^{\alpha\gamma} q^{-\alpha} 2^{(d-\gamma\alpha)k}.\] Note that $d-\gamma\alpha<0$, so that in both cases the probabilities are summable over $k$. In particular, we can choose $K$ large enough so that the event $\cap_{k \geq K } A_k(\nu)$ holds with probability arbitrarily close to $1$ (for $\nu = \Pi$ or for $\nu = \Pi_T$ and uniformly in $T > e$). Now on the event $\cap_{k \geq K } A_k(\nu)$, we can take any $R \geq 2^{K}$ and choose $k$ such that $2^k \leq R \leq 2^{k+1}$. Then, we have that \[ \sup_{z \in B(0,R)} \xi(z) \leq \sup_{z \in B(0,2^{k+1})} \xi(z) \leq q 2^{k\gamma} \leq q R^\gamma ,\] so that the first statement follows. To show~\ref{asmp:small_pts}$_{r_0}$, we define $\tilde A_k(\nu) = \{ \exists z \in B_\nu(0,2^{-k}) \, : \, \xi(z) \geq 2^{-\gamma (k-1)} \}$. For $\nu = \Pi_T$, we have from~\cite[Lemma~2.7(i)]{OR14} that there exists $c > 0$ such that for $T > e$, \[\P ( \tilde A_k(\Pi_T)^c ) = \P \Big( \max_{y \in \supp \Pi_T^\ssup{1} \cap B(0,2^{-k})}\xi(y) \leq 2^{-\gamma (k-1)} \Big) \leq e^{-c 2^{-\alpha \gamma} 2^{k(\alpha \gamma - d)} }.\] Similarly, by direct calculation, there exists a constant $c>0$ such that \[\P ( \tilde A_k(\Pi)^c ) = \P \Big( \max_{y \in \supp \Pi^\ssup{1} \cap B(0,2^{-k})}\xi(y) \leq 2^{-\gamma (k-1)} \Big) \leq e^{-c 2^{-\alpha \gamma} 2^{k(\alpha \gamma - d)} }.\] Note that $\alpha \gamma - d>0$, so for any $\varepsilon>0$ we can choose $K$ such that for all large $T$, \[ \P\Big(\bigcup_{k\geq K} \tilde A_k(\Pi_T)^c \Big) \leq \varepsilon \quad\mbox{ and }\quad \P\Big(\bigcup_{k\geq K} \tilde A_k(\Pi)^c \Big) \leq \varepsilon.\] The result follows. \end{proof} We also note the following easy lemma. \begin{lem}\label{le:lilypad_nontrival} Almost surely, $h_\Pi(z) \in (0,\infty)$ and $h_{\Pi_T}(z)\in(0,\infty)$ for any $z \not= 0$. \end{lem} \begin{proof} The statement follows by combining Corollary~\ref{cor:nontriv} with Lemma~\ref{le:conditions_satisfied}. \end{proof} The next corollary is the key tool in proving Proposition \ref{prop:appr_works}. \begin{cor}\label{cor:approx_works} For any $\varepsilon > 0$, $T>e$ and $t_0 > 0$, \begin{eqnarray} &\lim_{\delta \downarrow 0 } \P \Big( \sup_{z \in \mathbb{R}^d} | h_\Pi(z) - h_\Pi^\delta(z) | \geq \varepsilon \Big) = 0, \\ &\lim_{\delta \downarrow 0 } \P \Big( \sup_{t \leq t_0} \sup_{z \in \mathbb{R}^d} | m_\Pi(z,t) - m_\Pi^\delta(z,t) | \geq \varepsilon \Big) = 0,\\ &\lim_{\delta \downarrow 0 } \P \Big( \sup_{t \leq t_0} d_{H} (s_\Pi(t) , s_\Pi^\delta(t) ) \geq \varepsilon \Big) = 0, \end{eqnarray} and similarly \begin{eqnarray} &\lim_{\delta \downarrow 0 } \limsup_{T\to\infty} \P \Big( \sup_{z \in \mathbb{R}^d} | h_{\Pi_T}(z) - h_{\Pi_T}^\delta(z) | \geq \varepsilon \Big) = 0 , \\ &\lim_{\delta \downarrow 0 } \limsup_{T\to\infty} \P \Big( \sup_{t \leq t_0} \sup_{z \in \mathbb{R}^d} | m_{\Pi_T}(z,t) - m_{\Pi_T}^\delta(z,t) | \geq \varepsilon \Big) = 0,\\ &\lim_{\delta \downarrow 0 } \limsup_{T\to\infty} \P \Big( \sup_{t \leq t_0} d_{H} (s_{\Pi_T}(t) , s_{\Pi_T}^\delta(t) ) \geq \varepsilon \Big) = 0. \end{eqnarray} \end{cor} \begin{proof} First, since $h_\Pi^\delta(z)\leq h_\Pi(z)$ for all $z\in\mathbb{R}^d$ and $\delta>0$, and $h_\Pi^\delta(z)$ is increasing as $\delta\downarrow0$, the events $\{h_\Pi^\delta(z)\leq h(z)\leq h_\Pi^\delta(z)+\varepsilon\}$ are increasing as $\delta\downarrow0$. By Lemma \ref{le:conditions_satisfied} and Proposition \ref{prop:del_app_general}, we know that for any $\varepsilon>0$, \[\lim_{\delta\downarrow 0} \P( h_\Pi^\delta(z)\leq h_\Pi(z)\leq h_\Pi^\delta(z)+\varepsilon \quad \forall z\in\mathbb{R}^d) = 1\] and \[\lim_{\delta\downarrow 0} \liminf_{T\to\infty}\P( h_{\Pi_T}^\delta(z)\leq h_{\Pi_T}(z)\leq h_{\Pi_T}^\delta(z)+\varepsilon \quad \forall z\in\mathbb{R}^d) = 1;\] the first and fourth statements follow. The proofs of the statements for $m$ and $s$ are almost identical, using Proposition \ref{prop:del_apr_number} in place of Proposition \ref{prop:del_app_general}. \end{proof} From Corollary~\ref{cor:approx_works}, we can easily deduce our main technical result Proposition~\ref{prop:appr_works}. \begin{proof}[Proof of Proposition~\ref{prop:appr_works}] We consider first the case of the hitting times. Recall that we defined, for any $f , g\in \mathcal{C}^d := C(\mathbb{R}^d,[0,\infty))$, \[ d_{U}(f,g) = \sum_{k \geq 1} 2^{-k} \Big( \sup_{x \in [-k,k]^d} \big\{|f(x) - g(x)|\big\} \wedge 1 \Big). \] For any $\varepsilon > 0$, we choose $N$ such that $2^{-N} \leq \varepsilon/2$. Then we have \[ \begin{aligned} \P \big( d_{U} (h^\delta_T, h_T) \geq \varepsilon \big) & \leq \P \Big( \sum_{k=1}^N \sup_{z \in [-k,k]^d} |h_T^\delta(z) - h_T(z)| \geq \varepsilon/2 \Big) \\ & \leq \sum_{k=1}^N \P\Big( \sup_{z \in [-k,k]^d} |h_T^\delta(z) - h_T(z)| \geq \varepsilon/(2N) \Big) \end{aligned} \] Letting first $T \rightarrow \infty$ and then $\delta \downarrow 0$, we obtain by Corollary~\ref{cor:approx_works} that \[ \lim_{\delta \downarrow 0} \limsup_{T \rightarrow\infty} \P \big( d_{U} (h^\delta_T, h_T) \geq \varepsilon \big) = 0 . \] The argument for the numbers of particles and the support of the discrete lilypad model as well as the analogous statements for the Poisson lilypad model also follow from Corollary~\ref{cor:approx_works} in exactly the same way. If we combine these statements, we obtain Proposition~\ref{prop:appr_works}. \end{proof} \subsection{Proof of Theorem~\ref{thm:weak_conv}}\label{ssn:proof_scaling} We would like to apply the continuous mapping theorem to deduce the weak convergence of the $\delta$-truncated lilypad models. To facilitate this application, we introduce some slightly different $\delta$-approximations: define, for $z\in\mathbb{R}^d$ and $\delta>0$, \[ \tilde h^\delta_\nu(z) = \inf \Big\{ \sum_{j=1}^n q \frac{|y_{j-1} - y_j|}{\xi(y_j)} + q\frac{|y_n|}{\delta} \, : \,n\in\mathbb{N}_0, y_0 = z \mbox{ and }y_1, \ldots, y_{n} \in \mathcal{B}_\nu(0,1/\delta) \Big\}.\] Note that the only difference from our previous definition $h^\delta_\nu$ is that the points $y_1\ldots,y_n$ must now be within the closed ball $\mathcal{B}(0,1/\delta)$. We also define \[ \tilde m_\nu^\delta(z,t) = \sup_{y \in \mathcal{B}_\nu(0,1/\delta)} \Big\{ \xi(y) ( t - \tilde h_\nu^\delta(y) ) - q |y-z| \Big\} \vee 0 , \quad z \in \mathbb{R}^d, t \geq 0, \] and \[ \tilde s_\nu^\delta(t) = \{ z \in \mathbb{R}^d \, : \, \tilde h_\nu^\delta(z) \leq t \} , \quad t \geq 0 . \] We recall that $h_T^\delta$ is shorthand for $h_{\Pi_T}^\delta$, $h^\delta$ for $h_\Pi^\delta$, and so on; and we similarly write $\tilde h_T^\delta$ for $\tilde h_{\Pi_T}^\delta$, $\tilde h^\delta$ for $\tilde h_\Pi^\delta$ and so on. The benefit of introducing these new quantities is that applying the continuous mapping theorem to them is straightforward. \begin{prop}\label{prop:d_apr_conv} For any $\delta > 0$, as $T \rightarrow \infty$ \[ (\tilde h_T^\delta, \tilde m_T^\delta, \tilde s_T^\delta) \Rightarrow (\tilde h^\delta, \tilde m^\delta, \tilde s^\delta) . \] \end{prop} \begin{proof} As discussed in Section~\ref{ssn:point_pr}, we know that \[ \Pi_T \Rightarrow \Pi . \] By the continuous mapping theorem, \cite[Theorem 5.1]{Bil68}, we only have to show that each of the maps \[ \nu \mapsto \tilde h_\nu^\delta, \quad \nu \mapsto \tilde m_\nu^\delta, \quad \nu \mapsto \tilde s^\delta_\nu, \] are continuous as functions from $M_p(E)$ (equipped with the vague topology) into the target spaces equipped with the topologies described before Theorem~\ref{thm:weak_conv}. We note that the definitions of $\tilde h^\delta_\nu$, $\tilde m^\delta_\nu$, and $\tilde s^\delta_\nu$ only depend on the point process through the values in $\mathcal{B}(0,1/\delta) \times [\delta , \infty)$, which is a compact set in $E$. The same is true for $\tilde m^\delta_\nu$ and $\tilde s^\delta_\nu$. Therefore, we can use Proposition 3.31 in~\cite{Res08}: given that $\nu_n$ converges vaguely to $\nu$, we can label atoms of $\nu_n$ and $\nu$ restricted to any compact set such that the finitely many atoms converge pointwise. This implies in particular that $\tilde h_{\nu_n}^\delta \rightarrow \tilde h_\nu^\delta$, $\tilde m_{\nu_n}^\delta \rightarrow \tilde m_\nu^\delta$, and $\tilde s_{\nu_n}^\delta \rightarrow \tilde s_\nu^\delta$. \end{proof} Write \[ A_T = (H_T, M_T, S_T),\quad a_T = (h_T, m_T, s_T), \quad a_T^\delta = (h_T^\delta, m_T^\delta, s_T^\delta), \quad \tilde{a}_T^\delta = (\tilde h_T^\delta, \tilde m_T^\delta, \tilde s_T^\delta), \] \[\tilde{a}^\delta = (\tilde h^\delta,\tilde m^\delta, \tilde s^\delta), \quad a^\delta = (h^\delta ,m^\delta, s^\delta) , \quad a = (h,m,s) .\] We now need to check that $\tilde{a}^\delta_T$ is close to $a^\delta_T$, and $\tilde{a}^\delta$ is close to $a^\delta$. \begin{lem}\label{le:tildegood} For any $\varepsilon>0$, \[\lim_{\delta\downarrow0} \limsup_{T\to\infty}\P( d^\ssup{\times 3}(\tilde{a}^\delta_T, a^\delta_T)>\varepsilon) = 0 \quad\mbox{ and }\quad \lim_{\delta\downarrow0} \P( d^\ssup{\times 3}(\tilde{a}^\delta, a^\delta)>\varepsilon) = 0.\] \end{lem} \begin{proof} Fix $\eta>0$; by Lemma \ref{le:conditions_satisfied} we may choose $R_0, r_0>0$ such that both $\Pi_T$ (for any large $T$) and $\Pi$ satisfy \ref{asmp:pot_large}$_{R_0}$ and \ref{asmp:small_pts}$_{r_0}$ with probability at least $1-\eta$. By Lemmas \ref{le:small_times} and \ref{le:large_hitting}, for any point measure $\nu$ satisfying \ref{asmp:pot_large}$_{R_0}$ and \ref{asmp:small_pts}$_{r_0}$, and any $R>0$ and $t_0>0$, there exists $\delta_0>0$ such that for all $\delta\in(0,\delta_0)$, \[\inf_{y\not\in \mathcal{B}(0,1/\delta)} h^\delta_\nu (y) > \max\Big\{ \sup_{z\in B(0,R)} h^\delta_\nu(z), t_0\Big\}.\] Then for all $\delta\in(0,\delta_0)$, $z\in B(0,R)$ and $t\leq t_0$, we have \[\tilde h^\delta_\nu(z) = h^\delta_\nu(z), \quad \tilde m^\delta_\nu(z,t) = m^\delta_\nu(z,t) \quad \mbox{and}\quad \tilde s^\delta_\nu(t) = s^\delta_\nu(t).\] From the definition of $d^\ssup{\times 3}$ (choosing $R$ and $t_0$ large enough that the distance is guaranteed to be small) we get that for all large $T$, \[\P\Big( d^\ssup{\times 3}(\tilde{a}^\delta_T, a^\delta_T)>\varepsilon\Big) \leq\eta \quad\mbox{ and }\quad \P\Big( d^\ssup{\times 3}(\tilde{a}^\delta, a^\delta)>\varepsilon\Big)\leq \eta\] for all $\delta\in(0,\delta_0)$. Since $\eta>0$ was arbitrary, this completes the proof. \end{proof} We can now combine the various parts of this section to deduce the main scaling limit, Theorem~\ref{thm:weak_conv}. \begin{proof}[Proof of Theorem~\ref{thm:weak_conv}] By the portmanteau theorem it suffices to show that for any bounded and Lipschitz-continuous function $f : \mathcal{C}^\ssup{\times 3} \rightarrow \mathbb{R}$, we have that \begin{equation}\label{eq:wk_conv_lip} \mathbb{E} [ f(H_T, M_T, S_T) ] \rightarrow \mathbb{E} [ f ( h, m , s) ] \quad \mbox{as } T \rightarrow \infty. \end{equation} Suppose that $f : \mathcal{C}^3 \rightarrow \mathbb{R}$ is bounded by $\|f\|$ and Lipschitz continuous with Lipschitz constant $L$, and let $\varepsilon > 0$. We have that \[\begin{aligned} \big| \mathbb{E} [ f(A_T) ] - \mathbb{E} [f (a) ] \big| & \leq \mathbb{E} \big[ | f(A_T) - f(a_T) | \big] + \mathbb{E} \big[ | f(a_T) - f (a_T^\delta) |\big] + \mathbb{E}\big[|f(a_T^\delta)-f(\tilde{a}_T^\delta)|\big] \\ &\quad + \big| \mathbb{E} [ f(\tilde{a}_T^\delta )] - \mathbb{E} [ f(\tilde{a}^\delta) ] \big| + \mathbb{E}\big[ | f(\tilde{a}^\delta ) - f(a^\delta) | \big] + \mathbb{E}\big[|f(a^\delta)-f(a)|\big] \\ & \leq 5 L \varepsilon + 2\| f \| \, \P ( d^\ssup{\times 3}( A_T, a_T) > \varepsilon ) + 2\| f \| \, \P ( d^\ssup{\times 3}( a_T, a_T^\delta) > \varepsilon ) \\ & \qquad + 2\|f\|\,\P(d^\ssup{\times 3}(a^\delta_T,\tilde{a}^\delta_T)>\varepsilon) + \big| \mathbb{E} [ f(\tilde{a}_T^\delta )] - \mathbb{E} [ f(\tilde{a}^\delta) ] \big| \\ &\qquad + 2\|f\|\,\P(d^\ssup{\times 3}(\tilde{a}^\delta,a^\delta)>\varepsilon) + 2\| f \|\, \P ( d^\ssup{\times 3}(a^\delta, a) > \varepsilon ). \\ \end{aligned} \] We now take a $\limsup$ as $T\to\infty$: by Theorem \ref{thm:discrete_lilypad}, \[\P ( d^\ssup{\times 3}( A_T, a_T) > \varepsilon ) \to 0;\] and by Proposition \ref{prop:d_apr_conv}, \[\big| \mathbb{E} [ f(\tilde{a}_T^\delta )] - \mathbb{E} [ f(\tilde{a}^\delta) ] \big|\to 0.\] Thus \begin{multline*} \limsup_{T\to\infty} \big| \mathbb{E} [ f(A_T) ] - \mathbb{E} [f (a) ] \big|\\ \leq 5 L \varepsilon + 2\| f \| \, \limsup_{T\to\infty}\P ( d^\ssup{\times 3}( a_T, a_T^\delta) > \varepsilon ) + 2\|f\|\,\limsup_{T\to\infty}\P(d^\ssup{\times 3}(a^\delta_T,\tilde{a}^\delta_T)>\varepsilon) \\ \qquad + 2\|f\|\,\P(d^\ssup{\times 3}(\tilde{a}^\delta,a^\delta)>\varepsilon) + 2\| f \|\, \P ( d^\ssup{\times 3}(a^\delta, a) > \varepsilon ). \end{multline*} Finally, by Proposition \ref{prop:appr_works} and Lemma \ref{le:tildegood}, taking a limit as $\delta\downarrow0$ on the right-hand side, we get \[\limsup_{T\to\infty} \big| \mathbb{E} [ f(A_T) ] - \mathbb{E} [f (a) ] \big| \leq 5 L \varepsilon,\] and since $\varepsilon>0$ was arbitrary the proof is complete. \end{proof} \section{Proof of the ageing result}\label{sec:proof_ageing} In this section we prove Theorem~\ref{thm:ageing}. Before we start with the main proof, we need to collect several auxiliary lemmas, where we show that the lilypad models are rather `discrete': once two maximizing points are close, they are in fact the same. \begin{lemma}\label{le:compact_max} For any $t > 0$ \[ \lim_{n \rightarrow \infty} \limsup_{T \rightarrow \infty } \P ( S_T(\cdot,t) \not\subseteq B(0,n) ) = 0 . \] \end{lemma} \begin{proof} Follows for $s_T$ instead of $S_T$ by combining Lemma~\ref{le:large_hitting} with Lemma~\ref{le:conditions_satisfied} and thus for $S_T$ by Theorem~\ref{thm:discrete_lilypad}. \end{proof} \begin{lemma}\label{le:Poisson_simple} We have: \begin{itemize} \item[(i)] For any $t > 0$, $\lim_{n \rightarrow \infty} \P (\supp m(\cdot, t) \not\subseteq B(0,n)) = 0 )$. \item[(ii)] For any $t > 0$, $ \lim_{\varepsilon \downarrow 0 } \P (\xi(w(t)) \leq \varepsilon ) = 0$. \item[(iii)] For any $n \in \mathbb{N}, \varepsilon > 0$, \[ \lim_{\delta \downarrow 0} \P ( \exists z_1 \neq z_2 \in B_\Pi(0,n) \, : \, |z_1 - z_2|< \delta \mbox{ and } \xi(z_1) \geq \varepsilon , \xi(z_2) \geq \varepsilon ) = 0 . \] \end{itemize} \end{lemma} \begin{proof} (i) Follows by combining Lemma~\ref{le:large_hitting} with Lemma~\ref{le:conditions_satisfied}. (ii) By monotone convergence \[ \lim_{\varepsilon \downarrow 0 } \P (\xi(w(t)) \leq \varepsilon ) = \P (\xi(w(t)) = 0 ) = \P ( m(x,t) = 0 \mbox{ for all } x) . \] But by Lemma~\ref{le:lilypad_nontrival} we know that the lilypad model is almost surely non-trivial, so the latter probability is $0$. (iii) By the standard Palm calculus for Poisson processes we know that, conditionally on $\Pi(\{ (x,y) \}) = 1$, the process $\Pi - \delta_{(x,y)}$ is again a Poisson process with intensity $\pi$, see e.g.~\cite[Theorem 3.1]{Baddeley}. Therefore, we can write \[ \begin{aligned} \P ( \exists z_1 & \neq z_2 \in B(0,n) \, : \, z_2 \in B(z_1, \delta) \mbox{ and } \xi(z_1) \geq \varepsilon , \xi(z_2) \geq \varepsilon ) \\ & = \int_{B(0,n) \times [\varepsilon, \infty)} \P( \exists z_2 \in B(z_1, \delta)\setminus \{z_1\} \, : \, \xi(z) \geq \varepsilon ) \pi( d (x_1,y_2) ) \\ & = \int_{B(0,n) \times [\varepsilon, \infty)} \P ( \Pi( B(z_1, \delta) ) \times [\varepsilon, \infty) \neq 0 ) ) \pi( d (x_1,y_2) ) \\ \end{aligned} \] However, we know that \[ \P ( \Pi( B(z_1, \delta) ) \times [\varepsilon, \infty) \neq 0 ) = 1 - e^{ - \pi( B(z_1, \delta) ) \times [\varepsilon, \infty) } \rightarrow 0 , \] as $\delta \downarrow 0$. The claim follows by dominated convergence, since $\pi( B(0,n) \times [\varepsilon, \infty)) < \infty$. \end{proof} \begin{lem}\label{le:close_not_equal} For any $0 \leq s < t$, \[\lim_{\delta \downarrow 0} \limsup_{T \rightarrow \infty} \P ( | W_T(t) - W_T(s)| < \delta ; W_T(t) \neq W_T(s) ) = 0 \] and \[\lim_{\delta \downarrow 0} \P ( | w(t) - w(s)| < \delta ; w(t) \neq w(s) ) = 0. \] \end{lem} \begin{proof} We begin with the first statement. From Theorem~1.1 in~\cite{onepoint}, we know that for any $t$, with probability tending to $1$, the branching random walk is localised in the maximizer $w_T(t)$ of $m_T(\cdot, t)$. Therefore it suffices to show the corresponding statement for $w_T(t)$. Note that for any $t > 0$ \[ \lim_{\varepsilon \downarrow 0} \limsup_{T \rightarrow \infty} \P ( \xi(w_T(t)) \geq \varepsilon ) \leq \lim_{\varepsilon\downarrow 0}\P ( m(w(t), t) \leq \varepsilon ) = 0, \] since the limiting model $m(\cdot, t)$ is almost surely non-trivial by Lemma~\ref{le:lilypad_nontrival}. Also, by Lemma~\ref{le:compact_max}, we have for any $t$ that \[ \lim_{n \rightarrow \infty} \limsup_{T \rightarrow \infty} \P ( |w_T(t)| \geq n) = 0 . \] Now, for fixed $s$ and $t$, under the assumptions that $\xi(w_T(t)) \wedge \xi(w_T(s)) > \varepsilon$ and $|w_T(t)| \vee |w_T(s)| < n$, the event $\{ | w_T(t) - w_T(s)| < \delta ;\, w_T(t) \neq w_T(s)\}$ implies that there exist $w \neq w' \in L_T(0,n)$ with $|w - w'| \leq \delta$ such that $\xi_T(w), \xi_T(w') \geq \varepsilon$. Thus, by the above, we are done if we can show that for any $n \in \mathbb{N}, \varepsilon > 0$, \[ \lim_{\delta \downarrow 0} \limsup_{T\to\infty}\P ( \exists w \neq w' \in L_T(0,n) \, : \, |w - w'| \leq \delta \, : \, \xi_T(w), \xi_T(w') \geq \varepsilon ) = 0 . \] However, this follows from an explicit calculation: for some constant $C$, \begin{multline*}\P ( \exists w \neq w' \in L_T(0,n) \, : \, |w - w'| \leq \delta \, : \, \xi_T(w), \xi_T(w') \geq \varepsilon ) \\ \leq C r(T)^{2d} a(T)^{- 2\alpha} n^d \delta^d \varepsilon^{-2\alpha} = C n^d \delta^d \varepsilon^{-2\alpha} , \end{multline*} and letting $T\to\infty$ and then $\delta \downarrow 0$ completes the proof of the first statement. The second is almost identical, using Lemma \ref{le:Poisson_simple}. \end{proof} We now check that the maximizer for the Poisson lilypad model behaves sensibly. For $x\in\mathbb{R}^d$ and $\delta>0$, let $\partial B(x,\delta) = \{z : |z-x|=\delta\}$, the boundary of the ball of radius $\delta$ about $x$. \begin{lemma}\label{le:Poisson_max_unique} The following are true: \begin{itemize} \item[(i)] For any $t \geq 0$, almost surely, there is a single maximizer in the Poisson model $m(\cdot, t)$. \item[(ii)] For any fixed $x\in\mathbb{R}^d$, $\delta > 0$ and $t> 0$, $P^{\xi}\big(w(t) \in \partial B(x,\delta)\big) =0$. \end{itemize} \end{lemma} \begin{proof} (i) The basic idea is the following: if both $w$ and $w'$ are maximizers, we have $m(w,t)=m(w',t)$, which means $\xi(w) = \xi(w')(t-h(w'))/(t-h(w))$. Suppose without loss of generality that $h(w)\geq h(w')$. Then from the definition of $h$, if $w\neq w'$, the values of $\xi(w')$, $h(w')$ and $h(w)$ are independent of $\xi(w)$. So the probability that $\xi(w)$ takes on the exact value $\xi(w')(t-h(w'))/(t-h(w))$ is zero. However, since our point process $\Pi$ has infinitely many atoms, we need to be careful. Fix for a moment $z\in\mathbb{R}^d$, $\delta>0$ and $\varepsilon>0$, and let $\hat \Pi$ be the point process obtained by taking $\Pi$ and removing all of the points in $B(z,\delta)\times(\varepsilon,\infty)$ and $\tilde \Pi$ be the point process consisting of only those points of $\Pi$ in $B(z,\delta)\times(\varepsilon,\infty)$. Clearly $\hat \Pi$ and $\tilde \Pi$ are independent. Note from the definition of $h$ that for any $w\in B(z,\delta)$, if $B_{\tilde \Pi}(z,\delta)=\{w\}$, then $h_\Pi(w) = h_{\hat\Pi}(w)$. Furthermore, for any other point $w'\in\mathbb{R}^d$, if additionally $h_\Pi(w)\geq h_\Pi(w')$ then $h_\Pi(w') = h_{\hat\Pi}(w')$. Therefore \begin{align*} &\P\big( \exists w\in\supp\tilde\Pi^\ssup{1},\, w'\in\supp\hat\Pi^\ssup{1} \,:\, B_{\tilde \Pi}(z,\delta)=\{w\},\, h_\Pi(w)\geq h_\Pi(w'), \\ &\hspace{70mm} \xi_\Pi(w)(t-h_\Pi(w)) = \xi_\Pi(w')(t-h_\Pi(w'))\big)\\ &\leq \P\big(\exists w\in\supp\tilde\Pi^\ssup{1},\, w'\in\supp\hat\Pi^\ssup{1} \,:\, B_{\tilde \Pi}(z,\delta)=\{w\}, \\ &\hspace{70mm} \xi_{\tilde\Pi}(w)(t-h_{\hat\Pi}(w)) = \xi_{\hat\Pi}(w')(t-h_{\hat\Pi}(w'))\big)\\ &= 0, \end{align*} since $\hat\Pi$ and $\tilde\Pi$ are independent. Returning to our usual notation, this tells us that \begin{multline*} \P(\exists w\in B_\Pi(z,\delta),\, w'\in\supp \Pi^\ssup{1} \,:\, h(w)>h(w'),\,\xi(w)>\varepsilon,\\ \xi(y)\leq \varepsilon\,\,\forall y\in B_\Pi(z,\delta)\setminus\{w\},\, m(w,t)=m(w',t)) = 0 \end{multline*} (where no subscript means we are using the point process $\Pi$). Now, taking a sum over all $z$ such that $z/\delta\in\mathbb{Z}^d\cap B(0,n)$, we deduce that \begin{multline*} \P(\exists w, w'\in B_{\Pi}(0,n) \,:\, h(w)>h(w'),\,\xi(w)>\varepsilon,\\ \xi(y)\leq \varepsilon\,\,\forall y\in B_\Pi(w,2\delta)\setminus\{w\},\, m(w,t)=m(w',t)) = 0. \end{multline*} Taking a limit as $\delta\downarrow0$, we get by Lemma \ref{le:Poisson_simple} (iii) that \[\P(\exists w, w'\in B_{\Pi}(0,n) \,:\, h(w)>h(w'),\,\xi(w)>\varepsilon,\, m(w,t)=m(w',t)) = 0.\] Now taking $n\to\infty$, by Lemma \ref{le:Poisson_simple} (i), we have \[\P(\exists w, w'\in \supp \Pi^\ssup{1} \,:\, h(w)>h(w'),\,\xi(w)>\varepsilon,\, m(w,t)=m(w',t)>0) = 0.\] Finally, taking $\varepsilon\downarrow0$, by Lemma \ref{le:Poisson_simple} (ii), we get \[\P(\exists w, w'\in \supp \Pi^\ssup{1} \,:\, h(w)>h(w'),\, m(w,t)=m(w',t)=\sup_{x\in\mathbb{Z}^d}m(x,t)) = 0.\] This completes the proof of (i). (ii) To show the second statement, we note that by construction the maximizer $w(t)$ is in $\supp \Pi^\ssup{1}$. Also, note that the event $\{ \xi(w(t)) \geq \varepsilon \}$ is an increasing event as $\varepsilon \downarrow 0$. Thus, we have by Lemma~\ref{le:Poisson_simple} and monotone convergence \[ \begin{aligned} \P ( w(t) \in \partial B(x,\delta) ) & = \lim_{\varepsilon \downarrow 0} \P ( w(t) \in \partial B(x,\delta) , \xi(w(t) )\geq \epsilon )\\ & \leq \limsup_{\varepsilon \downarrow 0 } \P ( w(t) \in \partial B(x,\delta) , \xi(w(t)) \geq \varepsilon ) \\ & \leq \limsup_{\varepsilon \downarrow 0} \P ( \Pi( \partial B(x,\delta) \times [\varepsilon,\infty)) \geq 1 ) = 0, \end{aligned} \] since $\pi( \partial B(x,\delta) \times [\varepsilon,\infty) ) = 0$. \end{proof} \begin{lemma}\label{le:Poisson_not_equal} For any $\theta > 0$, \[ \lim_{\delta \downarrow 0}\P \big( |w(1) - w(1+\theta)| \leq 2\delta,\,\, w(1) \neq w(1+\theta)\big) = 0.\] \end{lemma} \begin{proof} Let $n \in \mathbb{N}$ and $\varepsilon > 0$. Then \[\begin{aligned} \P ( & |w(1) - w(1+\theta)| \leq 2\delta, w(1) \neq w(1+\theta)) \\ & \leq \P ( |w(1) - w(1+\theta)| \leq 2\delta, w(1) \neq w(1+\theta), \xi(w(1)) \geq \varepsilon, \xi(w(1+\theta)) \geq \varepsilon) \\ & \hspace{2cm} + \P (\min \{ \xi (w(1)), \xi(w(1+\theta))\} \leq \varepsilon) \\ & \leq \P ( \exists z_1 \neq z_2 \in B_\Pi(0,n) \, : \, | z_2 - z_1| \leq 2 \delta , \xi(z_1) \geq \varepsilon , \xi(z_2) \geq \varepsilon ) \\ & \hspace{2cm} + \P (\min \{ \xi (w(1)), \xi(w(1+\theta))\} \leq \varepsilon) + \P ( \max \{ |w(1)|, |w(1+\theta)|\} \geq n ) . \end{aligned} \] Now, letting $\delta \downarrow 0$, we obtain from Lemma~\ref{le:Poisson_simple}(iii) that \[ \begin{aligned} \lim_{\delta \downarrow 0} \P ( & |w(1) - w(1+\theta)| \leq 2\delta, w(1) \neq w(1+\theta)) \\ & \leq \P (\min \{ \xi (w(1)), \xi(w(1+\theta))\} \leq \varepsilon) + \P ( \max \{ |w(1), |w(1+\theta)| \geq n ) . \end{aligned} \] Finally, letting $\varepsilon \downarrow 0$ and $n \rightarrow \infty$, we obtain the statement from Lemma~\ref{le:Poisson_simple} (i), (ii). \end{proof} We are now finally ready to prove the ageing result, Theorem~\ref{thm:ageing}. \begin{proof}[Proof of Theorem~\ref{thm:ageing}] We start with a lower bound. For any $\theta > 0, \delta > 0$, define the open set \[\begin{aligned} \mathcal{O}_\theta^\delta := \Big\{ f \in \mathcal{C}^{d+1}_0 \, : \, \exists y \in \mathbb{R}^d \, \hbox{ with } & \max_{z \in \mathbb{R}^d \setminus B(y,\delta) } f(z,1) < f(y,1), \\ & \max_{z \in \mathbb{R}^d \setminus B(y,\delta) } f(z,1+\theta) < f( y,1+\theta) \Big\} . \end{aligned} \] From the weak convergence $M_T \Rightarrow m$, we know that \begin{equation}\label{eq:Oli} \liminf_{T \rightarrow \infty} \P ( M_T \in \mathcal{O}_\theta^\delta ) \geq \P (m \in \mathcal{O}_\theta^\delta) . \end{equation} Note that if $w(1)=w(1+\theta)$, then $m\in\mathcal{O}_\theta^\delta$ for any $\delta>0$; so \begin{align*} \P(w(1) = w(1+\theta)) &= \P(w(1) = w(1+\theta), m\in\mathcal{O}_\theta^\delta)\\ &= \P(m\in\mathcal{O}_\theta^\delta)-\P(m\in\mathcal{O}_\theta^\delta, w(1)\neq w(1+\theta)). \end{align*} Note that on the event $\{m\in\mathcal{O}_\theta^\delta\}$, if there are two different maximizers at times $1$ and $1+\theta$ then they must be within distance $\delta$. Thus by Lemma \ref{le:close_not_equal}, $\lim_{\delta\downarrow0} \P(m\in\mathcal{O}_\theta^\delta, w(1)\neq w(1+\theta)) = 0$, and therefore \begin{equation}\label{eq:wo1} \P(w(1) = w(1+\theta)) = \lim_{\delta\downarrow0} \P(m\in\mathcal{O}_\theta^\delta). \end{equation} Similarly, for any $\delta>0$, \begin{align*} \P(W_T(1) = W_T(1+\theta)) &= \P(W_T(1) = W_T(1+\theta), M_T\in\mathcal{O}_\theta^\delta)\\ &= \P(M_T\in\mathcal{O}_\theta^\delta)-\P(M_T\in\mathcal{O}_\theta^\delta, W_T(1)\neq W_T(1+\theta)), \end{align*} and by Lemma \ref{le:close_not_equal}, \[\lim_{\delta\downarrow 0} \limsup_{T\to\infty}\P(M_T\in\mathcal{O}_\theta^\delta, W_T(1)\neq W_T(1+\theta)) = 0\] since on the event $\{M_T\in\mathcal{O}_\theta^\delta\}$, if there are two different maximizers at times $1$ and $1+\theta$ then they must be within distance $\delta$. Therefore \[\liminf_{T\to\infty} \P(W_T(1) = W_T(1+\theta)) = \lim_{\delta\downarrow0}\liminf_{T\to\infty} \P(M_T\in \mathcal{O}_\theta^\delta).\] Combining this with \eqref{eq:Oli} and \eqref{eq:wo1}, we get \[\lim_{\delta\downarrow 0} \liminf_{T\to\infty} \P(W_T(1) = W_T(1+\theta)) \geq \P(w(1) = w(1+\theta)),\] which is the required lower bound. We now continue with an upper bound. Recall that $\mathcal{B}(z,r)$ is the closed ball of radius $r$ about $z$. For $z \in \mathbb{R}^d$, $\delta > 0$ and $\theta > 0$, we consider the set \[\begin{aligned} C_\theta(z,\delta) := \Big\{ f \in \mathcal{C}_0^{d+1} & \, : \, \max_{x\in \mathcal{B}(z,\delta)} f(x,1) = \max_{x \in \mathbb{R}} f(x,1) , \\ & \max_{x\in \mathcal{B}(z,\delta)} f(x,1+\theta) = \max_{x \in \mathbb{R}} f(x,1+\theta) \Big\} . \end{aligned} \] This set is closed, so since $M_T \Rightarrow m$ we know that \begin{equation}\label{eq:0206-1} \limsup_{T \rightarrow \infty} \P ( M_T \in C_\theta(z,\delta) ) \leq \P (m \in C_\theta(z,\delta) ) . \end{equation} Now let $n \in \mathbb{N}, \delta > 0$ and take $\Gamma_n^\delta$ to be a collection of points such that $ \mathcal{B}(0,n\delta) = \bigcup_{z \in \Gamma_n^\delta} \mathcal{B}(z,\delta)$, but the collection $\{B(z,\delta) : z \in \Gamma_n^\delta\}$ is disjoint (recall that we are working with $L^1$-balls so that this is possible). Then \[\P ( W_T(1) = W_T(1+\theta) ) \leq \sum_{z \in \Gamma_n^\delta} \P ( M_T \in C_\theta(z,\delta)) + \P (W_T(1) \notin \mathcal{B}(0,n\delta) ),\] and combining with \eqref{eq:0206-1} and Lemma \ref{le:compact_max} we get that for any $\delta>0$, \begin{equation}\label{eq:wc} \limsup_{T\to\infty} \P ( W_T(1) = W_T(1+\theta) ) \leq \limsup_{n\to\infty}\sum_{z \in \Gamma_n^\delta} \P( m\in C_\theta(z,\delta)). \end{equation} On the other hand, since by Lemma \ref{le:Poisson_max_unique} the maximizers for the Poisson lilypad model at times $1$ and $1+\theta$ are almost surely unique and not located on the boundary of any of the balls $B(z,\delta)$ for $z\in \Gamma_n^\delta$, we have \begin{align*} \sum_{z\in\Gamma_n^\delta} \P(m\in C_\theta(z,\delta)) &\leq \sum_{z\in\Gamma_n^\delta} \P(|w(1)-z|\leq \delta, |w(1+\theta)-z|\leq \delta)\\ &\leq \P(\exists z\in B(0,n\delta) : |w(1)-z|\leq \delta, |w(1+\theta)-z|\leq \delta). \end{align*} But, for any $n$, \begin{multline*} \P(\exists z\in B(0,n\delta) : |w(1)-z|\leq \delta, |w(1+\theta)-z|\leq \delta)\\ \leq \P(w(1) = w(1+\theta)) + \P(w(1)\neq w(1+\theta), |w(1)-w(1+\theta)|\leq 2\delta), \end{multline*} and by Lemma \ref{le:Poisson_not_equal}, the limit of the latter probability as $\delta\downarrow0$ is zero. Thus \[\lim_{\delta\downarrow0}\limsup_{n\to\infty}\sum_{z\in\Gamma_n^\delta} \P(m\in C_\theta(z,\delta)) \leq \P(w(1) = w(1+\theta)).\] Combining this with \eqref{eq:0206-1} and \eqref{eq:wc}, we obtain \[\limsup_{T\to\infty} \P ( W_T(1) = W_T(1+\theta) ) \leq \P(w(1) =w(1+\theta )), \] which is the required upper bound and completes the proof. \end{proof} \bibliographystyle{alpha}

https://arxiv.org/abs/1305.1268

A Contraction Analysis of the Convergence of Risk-Sensitive Filters

A contraction analysis of risk-sensitive Riccati equations is proposed. When the state-space model is reachable and observable, a block-update implementation of the risk-sensitive filter is used to show that the N-fold composition of the Riccati map is strictly contractive with respect to the Riemannian metric of positive definite matrices, when N is larger than the number of states. The range of values of the risk-sensitivity parameter for which the map remains contractive can be estimated a priori. It is also found that a second condition must be imposed on the risk-sensitivity parameter and on the initial error variance to ensure that the solution of the risk-sensitive Riccati equation remains positive definite at all times. The two conditions obtained can be viewed as extending to the multivariable case an earlier analysis of Whittle for the scalar case.

\section{Introduction} \label{sec:intro} Starting with Kalman and Bucy's paper \cite{KB}, the convergence of the Kalman filter has been examined in detail, and it soon became clear that if the state-space model is stabilizable and detectable, the filter is asymptotically stable and the error covariance converges to the unique non-negative definite solution of a matching algebraic Riccati equation. However the classical Kalman filter convergence analysis \cite{AM,KSH} is rather intricate and involves several steps, including first showing that the error covariance is upper bounded, next proving that with a zero initial value, it is monotone increasing, so it has a limit, and then establishing that the corresponding filter is stable and that the limit is the same for all initial covariances. In 1993, Bougerol \cite{Bou} proposed a more direct convergence proof based on establishing that the discrete-time Riccati iteration is a contraction for the Riemmanian metric associated to the cone of positive definite matrices. Although this result attracted initially little notice in the systems and control community, this approach was adopted by several researchers \cite{LaL1,LaL2,LeL,LYD} to study the convergence of a number of nonlinear matrix iterations. We will use this viewpoint here to analyze the convergence of risk-sensitive estimation filters. Unlike the Kalman filter which minimizes the mean square estimation error, risk-sensitive filters \cite{SDJ,Whi} minimize the expected value of an exponential of quadratic error index, which ensures a higher degree of robustness \cite{HSK1,HS} against modelling errors. Unfortunately, in spite of extensive studies on risk-sensitive and related $H^{\infty}$ filters, results concerning their convergence remain fragmentary \cite[Chap. 9]{Whi}, \cite[Sec. 14.6]{HSK1}, \cite{BS}. In particular, one question that remains unresolved is whether there exists an a-priori upper bound on the risk-sensitivity parameter ensuring the convergence of the solution of the risk-sensitivite Riccati equation to a positive definite solution associated to a stable filter. The contraction anaysis presented in this paper relies on a block implementation of Kalman (risk-neutral) and risk-sensitive filters. When the system is reachable and observable and the block length $N$ exceeds the number of states, it is shown that in the risk-neutral case, the Riccati equation corresponding to the block filter is strictly contractive, which allows us to conclude that the Riccati equation of Kalman filtering has a unique positive definite fixed point. This analysis is equivalent to the derivation of \cite{Bou} which relied on showing that the $N$-fold composition of the Hamiltonian map associated to the risk-neutral Riccati operator is strictly contractive. However, it has the advantage that it can be extended easily to the risk-sensitive case by using the Krein-space formulation of risk-sensitive and $H^{\infty}$ filtering developed in \cite{HSK2,HSK3}. With this approach, it is is shown that the $N$-block risk-sensitive Riccati equation remains strictly contractive as long as a corresponding observability Gramian is positive definite. This Gramian is shown to be a monotone decreasing function of the risk-sensitivity parameter $\theta$ with respect to the partial order of non-negative definite matrices. Accordingly, it is possible to identify a priori a range $[0,\tau_N)$ of values of the risk-sensitivity parameter $\theta$ for which the block Riccati equation is strictly contractive. This result is used to show that the risk-sensitive Riccati equation has a unique positive definite fixed-point, but because the image of the cone ${\cal P}$ of positive definite matrices under the risk-sensitive Riccati map is not entirely contained in ${\cal P}$ a second condition must be placed on $\theta$ and the initial variance $P_0$ of the filter to ensure that the evolution of the risk-sensitive Riccati equation stays in ${\cal P}$. The two conditions obtained can be viewed as extensions to the multivariable case of those presented in \cite[Chap. 9]{Whi} for scalar risk-sensitive Riccati equations. The paper is organized as follows. The properties of the Riemann distance for positive definite matrices and of contraction mappings are reviewed in Section \ref{sec:riemann}. The block-update filtering interpretation of the $N$-fold Riccati equation of Kalman filtering is described in Section \ref{sec:block} and is extended to the risk-sensitive case in Section \ref{sec:contrisk}. This formulation is used to estimate the range of values of the risk-sensitivity parameter for which the risk-sensitive Riccati equation is contractive. A second condition on the risk-sensitivity parameter and initial condition ensuring that the solution of the Riccati equation remains positive is obtained in Section \ref{sec:posit}. An illustrative example is studied in Section \ref{sec:examp} and conclusions as well as a possible extension are presented in Section \ref{sec:conc}. \section{Riemann distance and contraction mappings} \label{sec:riemann} Let ${\cal P}$ denote the cone of positive definite symmetric matrices of dimension $n$, and let $\bar{\cal P}$ denote the cone of non-negative definite matrices forming the closure of ${\cal P}$. If $P$ is an element of ${\cal P}$ with eigendecomposition \begin{equation} P = U \Lambda U^T \label{2.1} \end{equation} where $U$ is an orthogonal matrix formed by normalized eigenvectors of $P$ and $\Lambda = \mbox{diag} \, \{ \lambda_1 , \ldots , \lambda_n \}$ is the diagonal eigenvalue matrix of $P$, the symmetric positive square-root of $P$ is defined as \[ P^{1/2} = U \Lambda^{1/2} U^T \] where $\Lambda^{1/2}$ is diagonal, with entries $\lambda_i^{1/2}$ for $1 \leq i \leq n$. Similarly, the logarithm of $P$ is the symmetric, not necessarily positive definite, matrix specified by \[ \log(P) = U \log(\Lambda ) U^T \: , \] where $\log(\Lambda)$ is diagonal with entries $\log(\lambda_i)$ for $1 \leq i \leq n$. Let $P$ and $Q$ be two positive definite matrices of ${\cal P}$. Then $P^{-1}Q$ is similar to $P^{-1/2}QP^{-1/2}$, so they have the same eigenvalues, and $P^{-1/2}QP^{-1/2}$ is positive definite. Let $s_1 \geq s_2 \geq \ldots \geq s_n >0$ denote the eigenvalues of $P^{-1}Q$ sorted in decreasing order. The Riemann distance between $P$ and $Q$ is defined as \begin{equation} d(P,Q) = ||\log(P^{-1/2}QP^{-1/2})|| = \Big( \sum_{i=1}^n ( \log (s_i))^2 \Big)^{1/2} \: , \label{2.2} \end{equation} where $||.||$ denotes the matrix Frobenius norm. In addition to having all the traditional properties of a distance, $d$ has the feature that it is invariant under matrix inversion and congruence transformations. Specifically, if $M$ denotes an arbitrary real invertible matrix of dimension $n$, \begin{equation} d(P,Q) = d(P^{-1},Q^{-1}) = d(MPM^T,MQM^T) \: . \label{2.3} \end{equation} Furthermore, it was also shown by Bougerol \cite{Bou} that the translation of ${\cal P}$ by a non-negative definite symmetric matrix $S$ is a non-expansive map. Specifically, \begin{equation} d(P+S,Q+S) \leq \frac{\alpha}{\alpha+\beta} d(P,Q) \label{2.4} \end{equation} where $\alpha = \max (\lambda_1(P),\lambda_1(Q))$ and $\beta = \lambda_n (S)$. In these definitions, it is assumed that the eigenvalues of $P$, $Q$ and $S$ are sorted in decreasing order, so that $\lambda_1 (P)$ is the largest eigenvalue of $P$, i.e., its spectral norm, and $\lambda_n (S)$ is the smallest eigenvalue of $S$. Recall that if $f(\cdot)$ is an arbitrary mapping of ${\cal P}$, $f$ is non-expansive if \[ d(f(P),f(Q)) \leq d(P,Q) \: , \] and strictly contractive if \[ d(f(P),f(Q)) \leq c d(P,Q) \] with $0 \leq c < 1$. The least contraction coefficient or Lipschitz constant of a non-expansive mapping $f$ is defined as \begin{equation} c(f) = \sup_{P,Q \in {\cal P}, P \neq Q} \frac{d(f(P),f(Q)}{d(P,Q)} \: . \label{2.5} \end{equation} Clearly, if $f$ and $g$ denote two non-expansive mappings, the contraction coefficient $c(f \circ g)$ of the composition of $f$ and $g$ satisfies $c (f \circ g) \leq c(f) c(g)$, so if at least one of the two maps is strictly contractive, the composition is also strictly contractive. From inequality (\ref{2.4}), we deduce that if $\tau_S (P) = P+S$ denotes the translation by a positive definite matrix $S$, $\tau_S$ is non-expansive, but the bound (\ref{2.4}) does not allow us to conclude that $c(\tau_S) < 1$, since when the largest eigenvalue of either $P$ or $Q$ goes to infinity, the constant $\alpha/(\alpha+\beta)$ tends to one. The key result that will be used in this paper is that if $f$ is a strict contraction of ${\cal P}$ for the distance $d$, by the Banach fixed point theorem \cite[p. 244]{AE}, there exists a unique fixed point $P$ of $f$ in $\bar{\cal P}$ satisfying $P=f(P)$. Furthermore this fixed point can be evaluated by performing the iteration $P_{n+1} = f(P_n)$ starting from any initial point $P_0$ of ${\cal P}$. Also if the $N$-fold composition $f^N$ of a non-expansive map $f$ is strictly contractive, then $f$ has a unique fixed point. We will consider in particular the Riccati-type map over ${\cal P}$ defined by \begin{equation} f(P) = M[P^{-1} + \Omega ]^{-1}M^T + W \label{2.6} \end{equation} where $P$, $\Omega$ and $W$ are symmetric real positive definite matrices and $M$ is a square real, but not necessarily invertible, matrix. For this mapping the following result was established in \cite[Th. 4.4]{LeL}. \vskip 2ex \begin{lemma} \label{lem1} $f$ is a strict contraction with \begin{equation} c(f) \leq \frac{\lambda_1 (M\Omega^{-1}M^T)}{ \lambda_n (W) + \lambda_1(M\Omega^{-1}M^T)} \: , \label{2.7} \end{equation} where we use again the convention that the eigenvalues of positive definite matrices are sorted in decreasing order. \end{lemma} \vskip 2ex Note that although the results presented in this paper use the Riemann distance over ${\cal P}$, other metrics such as the Thompson metric \begin{equation} d_T (P,Q) = \max ( \lambda_1 (\log(P^{-1/2}QP^{-1/2})), \lambda_1 (\log(Q^{-1/2} P Q^{-1/2}))) \: . \label{2.8} \end{equation} have similar properties \cite{Bha,ISY}, and the Thompson metric is in fact often used to analyze the convergence of nonlinear matrix iterations \cite{LaL2,LYD}. \section{Block update filter} \label{sec:block} Consider a Gauss-Markov state space model \begin{eqnarray} x_{t+1} =A x_t + B u_t \label{3.1} \\ y_t = C x_t + v_t \: , \label{3.2} \end{eqnarray} where the state $x_t \in \mathbb{R}^n$, the process noise $u_t \in \mathbb{R}^m$ and the observation noise $v_t \in \mathbb{R}^p$. The noises $u_t$ and $v_t$ are assumed to be independent zero-mean WGN processes with normalized covariance matrices, so \[ E \Big[ \left[ \begin{array} {c} u_t \\ v_t \end{array} \right] \left[ \begin{array} {cc} u_s^T & v_s^T \end{array} \right] \Big] = \left[ \begin{array} {cc} I_m & 0\\ 0 & I_p \end{array} \right] \delta_{t-s} \: , \] where \[ \delta_t = \left \{ \begin{array} {cc} 1 & t=0 \\ 0 & t \neq 0 \end{array} \right. \] denotes the Kronecker delta function. The initial state vector $x_0$ is assumed independent of noises $u_t$ and $v_t$ and ${\cal N} (\hat{x}_0,P_0)$ distributed. Since we are interested in the asymptotic behavior of Kalman and risk-sensitive filters, the matrices $A$, $B$, and $C$ specifying the state-space model are assumed to be constant. Then if ${\cal Y}_{t-1}$ denotes the sigma field generated by observations $y(s)$ for $0 \leq s \leq t-1$, the least-squares estimate $\hat{x}_t = E[x_t|{\cal Y}_{t-1}]$ depends linearly on the observations and can be evaluated recursively by the predicted form of the Kalman filter specified by \begin{equation} \hat{x}_{t+1} = A \hat{x}_t + K_t \nu_t \: , \label{3.3} \end{equation} where the innovations process \begin{equation} \nu_t \stackrel{\triangle}{=} y_t - C\hat{x}_t \: . \label{3.4} \end{equation} In (\ref{3.3}), the Kalman gain matrix \begin{equation} K_t = AP_tC^T(R_t^{\nu})^{-1} \: , \label{3.5} \end{equation} where \begin{equation} R_t^{\nu} = E[\nu_t \nu_t^T] = CP_tC^T +I_p \label{3.6} \end{equation} represents the variance of the innovations process, and if $\tilde{x}_t = x_t - \hat{x}_t$ denotes the state prediction error, its variance matrix $P_t = E[\tilde{x}_t \tilde{x}_t^T]$ obeys the Riccati equation \begin{equation} P_{t+1}= r(P_t) \stackrel{\triangle}{=} A[ P_t^{-1} + C^TC ]^{-1} A^T + BB^T \label{3.7} \end{equation} with initial condition $P_0$. This equation can also be rewritten in the equivalent form \cite[p. 325]{KSH} \begin{equation} P_{t+1} = (A-K_tC)P_t(A-K_t C)^T + BB^T + K_t K_t^T \: , \label{3.8} \end{equation} which will be used later in our analysis. The Riccati mapping $r(P)$ specified by (\ref{3.7}) has the form (\ref{2.6}). Unfortunately the matrices $C^TC$ and $BB^T$ are not necessarily invertible, so Lemma \ref{lem1} is not directly applicable. Under the assumption that the pairs $(A,B)$ and $(C,A)$ are reachable and observable, respectively, Bougerol \cite{Bou} was able to show that the $n$-fold map $r^n$ is a strict contraction. This was achieved by considering the $n$-fold composition of the symplectic Hamiltonian mapping associated to $r$ (see \cite{LaL2} for a study of the contraction properties of symplectic Hamiltonian mappings). We present below an equivalent derivation of of Bougerol's result which relies on a block update implementation of the Kalman filter. The starting point is the observation that since $x_t$ is Gauss-Markov, the downsampled process $x_k^d = x_{kN}$ with $N$ integer is also Gauss-Markov with state-space model \begin{eqnarray} x_{k+1}^d &=& A^N x_k^d + {\cal R}_N {\bf u}_k^N \label{3.9} \\ {\bf y}_k^N &=& {\cal O}_N x_k^d + {\bf v}_k^N + {\cal H}_N {\bf u}_k^N \label{3.10} \end{eqnarray} where \begin{eqnarray} {\bf u}_k^N &=& \left[ \begin{array} {cccc} u_{kN+N-1}^T & u_{kN+N-2}^T & \ldots & u_{kN}^T \end{array} \right]^T \nonumber \\ {\bf y}_k^N &=& \left[ \begin{array} {cccc} y_{kN+N-1}^T & y_{kN+N-2}^T & \ldots & y_{kN}^T \end{array} \right]^T \nonumber \\ {\bf v}_k^N &=& \left[ \begin{array} {cccc} v_{kN+N-1}^T & v_{kN+N-2}^T & \ldots & v_{kN}^T \end{array} \right]^T \:. \nonumber \end{eqnarray} In the model (\ref{3.9})--(\ref{3.10}) \begin{eqnarray} {\cal R}_N &=& \left[ \begin{array} {cccc} B & AB & \ldots & A^{N-1}B \end{array} \right] \nonumber \\ {\cal O}_N &=& \left[ \begin{array} {cccc} (CA^{N-1})^T & \ldots & (CA)^T & C^T \end{array} \right]^T \nonumber \end{eqnarray} denote respectively the $N$-block reachability and observability matrices of system (\ref{3.1})--(\ref{3.2}, where the blocks forming ${\cal O}_N$ are written from bottom to top instead of the usual top to bottom convention. If the pairs $(A,B)$ and $(C,A)$ are reachable and observable, ${\cal R}_N$ and ${\cal O}_N$ have full rank for $N \geq n$. In (\ref{3.10}), if \[ H_t = \left \{ \begin{array} {cc} CA^{t-1}B & t \geq 1 \nonumber \\ 0 & \mbox{otherwise } \end{array} \right. \] denotes the impulse response representing the response of output $y_t$ in (\ref{3.2}) to the process noise $u_t$ input in (\ref{3.1}), ${\cal H}_N$ is the $Np \times Nm$ block Toeplitz matrix defined by \[ {\cal H}_N = \left[ \begin{array} {cccccc} 0 & H_1 & H_2 & \cdots & H_{N-2} &H_{N-1}\\ 0 & 0 & H_1 & H_2 & \cdots & H_{N-2}\\ 0 & 0 & 0 & H_1 & \cdots & H_{N-3}\\ \vdots & \vdots & \vdots & & \vdots & \vdots\\ 0 & 0 & 0 & \cdots & \cdots & H_1 \\ 0 & 0 & 0 & \cdots & \cdots & 0 \end{array} \right] \: . \] Note however that the noise vectors ${\bf u}_k^N$ and ${\bf w}_k^N \stackrel{\triangle}{=} {\bf v}_k^N +{\cal H}_N {\bf u}_k^N$ are correlated since \[ E \Big[ \left[ \begin{array} {c} {\bf u}_k^N \\ {\bf w}_k^N \end{array} \right] \left[ \begin{array} {cc} {\bf u}_{\ell}^{N T} & {\bf w}_{\ell}^{N T} \end{array} \right] = \left[ \begin{array} {cc} I_{Nm} & {\cal H}_N^T \\ {\cal H}_N & I_{Np} + {\cal H}_N {\cal H}_N^T \end{array} \right] \delta_{k-\ell} \: . \] This correlation can be removed by noting that the estimate of ${\bf u}_k^N$ given ${\bf w}_k^N$ takes the form \[ \hat{u}_k^N = {\cal G}_N {\bf w}_k^N \] where \[ {\cal G}_N ={\cal H}_N^T (I+ {\cal H}_N{\cal H}_N^T)^{-1} \:. \] Then by premultiplying the observation equation (\ref{3.10}) by ${\cal R}_N {\cal G}_N$ and subtracting it from (\ref{3.9}) we obtain the new downsampled state dynamics \begin{equation} x_{k+1}^d = \alpha_N x_k^d + {\cal R}_N \tilde{\bf u}_k^N +{\cal R}_N {\cal G}_N {\bf y}_k^N \label{3.11} \end{equation} with \[ \alpha_N \stackrel{\triangle}{=} A^N - {\cal R}_N {\cal G}_N {\cal O}_N \: , \] where the zero mean white Gaussian noise $\tilde{\bf u}_k^N = {\bf u}_k^N - \hat{\bf u}_k^N$ is now uncorrelated with observation noise ${\bf w}_k^N$, and has the invertible variance matrix \begin{eqnarray} {\cal Q}_N &=& I_{Nm}-{\cal H}_N^T [ I_{Np} + {\cal H}_N {\cal H}_N^T ]^{-1} {\cal H}_N \nonumber \\ &=& [ I_{Nm} + {\cal H}_N^T {\cal H}_N ]^{-1} \: . \nonumber \end{eqnarray} The Kalman filter corresponding to the downsampled state-space model (\ref{3.10})--(\ref{3.11}) can be interpreted as a block update filter, where the state estimate is updated only after a block of $N$ observations has been collected. The Riccati equation corresponding to this Kalman filter is then given by \begin{equation} P_{k+1}^d = r_d (P_k^d) = \alpha_N [ (P_k^d)^{-1} + \Omega_N ]^{-1} \alpha_N^T + W_N \: , \label{3.12} \end{equation} where the $n \times n$ symmetric real matrices \begin{eqnarray} \Omega_N &\stackrel{\triangle}{=}& {\cal O}_N^T [I + {\cal H}_N{\cal H}_N^T]^{-1} {\cal O}_N \label{3.13} \\ \nonumber \\ W_N &\stackrel{\triangle}{=}& {\cal R}_N [I+ {\cal H}_N^T {\cal H}_N]^{-1} {\cal R}_N^T \label{3.14} \end{eqnarray} are positive definite for $N \geq n$ whenever the pairs $(C,A)$ and $(A,B)$ are observable and reachable, respectively. In fact, $\Omega_N$ and $W_N$ can be viewed as observability and reachability Wronskians for the state-space model (\ref{3.1})--(\ref{3.2}). From Lemma \ref{lem1}, we can therefore conclude that $r_d(\cdot)$ is a strict contraction. However, since $P_k^d$ is the variance matrix of the one-step ahead prediction error for state $x_{kN}$, $r_d$ coincides with the $N$-fold composition $r^N$ of Riccati map $r$, which must have therefore a unique fixed point $P$ in ${\cal P}$. This establishes the following classical Kalman filter convergence result \cite{AM,KSH}. \vskip 2ex \begin{theorem} \label{theo1} If in system (\ref{3.1})-(\ref{3.2}) the pairs $(A,B)$ and $(C,A)$ are reachable and observable, respectively, the algebraic Riccati equation $P=r(P)$ admits a unique positive definite solution, and as $t$ tends to infinity, for any positive definite initial condition $P_0$, $P_t$ tends to $P$ as $t$ tends to infinity, and the Kalman gain matrix $K_t$ tends to \[ K = APC^T [CPC^T +I]^{-1} \] which has the property that the matrix $A-KC$ is stable. \end{theorem} \vskip 2ex Given the fixed point $P >0$, the stability of $A-KC$ is obtained by applying the Lyapunov stability theorem to equation \begin{equation} P = (A-KC)P(A-KC)^T +BB^T +KK^T \label{3.15} \end{equation} (see \cite[p. 80]{AM}). One unsatisfactory aspect of the contraction approach to the derivation of Theorem \ref{theo1} is its requirement that the system should be reachable and observable, instead of the weaker stabilizability and detectability conditions required by conventional Kalman filter convergence proofs \cite{AM,KSH}. The stronger conditions conditions are needed to ensure that the Riccati evolution takes place entirely in the cone of positive definite matrices. On the other hand, if the system is reachable and observable, the limit $P$ is guaranteed to be positive definite, instead of just nonnegative definite under the usual assumptions. Finally, note that the block update implementation of the Kalman filter which was used here to show that $r_d=r^N$ is a strict contraction is equivalent to Bougerol's derivation in \cite{Bou}, but as shown below it can be extended more easily to the risk-sensitive case. \section{Contraction property of the risk-sensitive Riccati equation} \label{sec:contrisk} For the state-space model (\ref{3.1})--(\ref{3.2}), the risk-sensitive estimate $\hat{x}_t$ solves the exponential quadratic minimization problem \cite{Whi,SFB} \begin{equation} \hat{x}_t = \mbox{arg} \, \min_{\xi \in \mathbb{R}^n} \frac{1}{\theta} \log \Big( E[ \exp \big( \frac{\theta}{2} ||D(x_t-\xi)||^2 \big) | {\cal Y}_{t-1}] \Big) \: , \label{4.1} \end{equation} where $D \in {\mathbb R}^{q \times n}$ with $q \leq n$ is assumed to have full row rank, and $||z|| = (z^T z)^{1/2}$ denotes the Euclidean vector norm. The parameter $\theta$ appearing in (\ref{4.1}) is called the risk-sensitivity parameter. The resulting estimate $\hat{x}_t$ obeys the recursion (\ref{3.3})-(\ref{3.4}), where \begin{equation} K_t = A(P_t^{-1} -\theta D^TD)^{-1} C^T (R_t^{\nu})^{-1} \label{4.2} \end{equation} with \begin{equation} R_t^{\nu} = C(P_t^{-1}-\theta D^TD)^{-1}C^T + I \:, \label{4.3} \end{equation} and where $P_t$ obeys the risk-sensitive Riccati equation \begin{equation} P_{t+1} = r^{\theta} (P_t) = A[P_t^{-1} +C^TC-\theta D^TD]^{-1}A^T + BB^T \: . \label{4.4} \end{equation} Our analysis will use the fact that the risk-sensitive Riccati equation can be rewritten as \begin{equation} P_{t+1} = (A-K_tC)[P_t^{-1}-\theta D^TD]^{-1}(A-K_t C)^T + BB^T +K_t K_t^T \: . \label{4.5} \end{equation} The values $\theta =0$, $\theta <0$ and $\theta >0$ of the risk-sensitivity parameter correspond respectively to the risk-neutral, risk-seeking, and risk-averse cases. When $\theta =0$, the risk-sensitive filter reduces to the Kalman fillter studied in the previous section, and when $\theta <0$ the matrix $C^TC -\theta D^T D$ is non-negative definite and can be rewritten as $\tilde{C}^T \tilde{C}$ where the pair formed by \[ \tilde{C} \stackrel{\triangle}{=} \left[ \begin{array} {c} C \\ (-\theta)^{1/2} D \end{array} \right] \] and $A$ is necessarily observable if $(C,A)$ is observable. Accordingly, the convergence result of Theorem \ref{theo1} is applicable to this problem, and in the remainder of this paper our attention will focus on the risk-averse case with $\theta >0$. An interesting feature of the risk-sensitive filter is that it can be interpreted as solving a standard least-squares filtering problem in Krein space \cite{HSK2,HSK3}. We will use this viewpoint here to extend the block filtering idea of the previous section to the risk-sensitive case. The Krein-space state-space model consists of dynamics (\ref{3.1}) and observations (\ref{3.2}), to which we must adjoin the risk-sensitive observations \begin{equation} 0 = D x_t + v_t^R \: . \label{4.6} \end{equation} The components of noise vectors $u_t$, $v_t$ and $v_t^R$ now belong to a Krein space and have the inner product \begin{equation} \left \langle \left[ \begin{array} {c} u_t \\ v_t \\ v_t^R \end{array} \right] \, , \, \left[ \begin{array} {c} u_s \\ v_s \\ v_s^R \end{array} \right] \right \rangle = \left[ \begin{array} {ccc} I_m & 0 & 0\\ 0 & I_p & 0 \\ 0 & 0 & -\theta^{-1} I_q \\ \end{array} \right] \delta_{t-s} \label{4.7} \end{equation} The $N$-step observability matrix of the pair $(D,A)$ is denoted as \[ {\cal O}_N^R = \left[ \begin{array} {cccc} (DA^{N-1})^T & \ldots & (DA)^T & D^T \end{array} \right]^T \] and if \[ L_t = \left \{ \begin{array} {cc} DA^{t-1}B & t \geq 1 \\ 0 & \mbox{otherwise} \end{array} \right. \] denotes the impulse response from input $u_t$ to the risk-sensitive observation output, the corresponding $N$-block Toeplitz matrix takes the form \[ {\cal L}_N = \left[ \begin{array} {cccccc} 0 & L_1 & L_2 & \cdots & L_{N-2} &L_{N-1}\\ 0 & 0 & L_1 & L_2 & \cdots & L_{N-2}\\ 0 & 0 & 0 & L_1 & \cdots & L_{N-3}\\ \vdots & \vdots & \vdots & & \vdots & \vdots\\ 0 & 0 & 0 & \cdots & \cdots & L_1 \\ 0 & 0 & 0 & \cdots & \cdots & 0 \end{array} \right] \: . \] Then if \[ {\bf v}_k^{R N} = \left[ \begin{array} {cccc} (v_{kN+N-1}^R)^T & (v_{kN+N-2}^R)^T & \ldots & (v_{kN}^R)^T \end{array} \right]^T \: , \] the $N$-block risk-sensitive observation for the downsampled process $x_k^d$ can be expressed as \begin{equation} {\bf 0} = {\cal O}_N^R x_k^d + {\bf v}_k^{R N} + {\cal L}_N {\bf u}_k^N \: . \label{4.8} \end{equation} The Krein space inner product of observation noise vector \[ \left[ \begin{array}{c} {\bf w}_k^N\\ {\bf w}_k^{R N} \end{array} \right] = \left[ \begin{array} {c} {\bf v}_k^N\\ {\bf v}_k^{R N} \end{array} \right] + \left[ \begin{array} {c} {\cal H}_N \\ {\cal L}_N \end{array} \right] {\bf u}_k^N \] with itself admits the block LDU decomposition \begin{eqnarray} \lefteqn{\left \langle \left[ \begin{array} {c} {\bf w}_k^N \\ {\bf w}_k^{R N} \end{array} \right] \, , \, \left[ \begin{array} {c} {\bf w}_k^N \\ {\bf w}_k^{R N} \end{array} \right] \right \rangle \stackrel{\triangle}{=} {\cal K}_N^{\theta}}\nonumber \\ &=& \left[ \begin{array} {cc} I_{Np} & 0 \\ 0 & -\theta^{-1} I_{Nq} \end{array} \right] + \left[ \begin{array} {c} {\cal H}_N \\ {\cal L}_N \end{array} \right] \left[ \begin{array} {cc} {\cal H}_N^T & {\cal L}_N^T \end{array} \right] \nonumber \\ &=& \left[ \begin{array} {cc} I_{Np} & 0 \\ {\cal L}_N{\cal H}_N^T (I_{Np} + {\cal H}_N{\cal H}_N^T)^{-1} & I_{Nq} \end{array} \right] \left[ \begin{array} {cc} I_{Np} + {\cal H}_N{\cal H}_N^T & 0 \\ 0 & S_N^{\theta} \end{array} \right] \nonumber \\ && \hspace*{1in} \times \left[ \begin{array} {cc} I_{Np} & (I + {\cal H}_N{\cal H}_N^T)^{-1} {\cal H}_N{\cal L}_N^T \\ 0 & I_{Nq} \end{array} \right] \: , \label{4.9} \end{eqnarray} where \begin{equation} S_N^{\theta} \stackrel{\triangle}{=} -\theta^{-1} I_{Nq} + {\cal L}_N(I_{Nm}+{\cal H}_N^T{\cal H}_N)^{-1}{\cal L}_N^T \label{4.10} \end{equation} denotes the Schur complement of the $(1,1)$ block inside ${\cal K}_N^{\theta}$. The projection of noise vector ${\bf u}_k^N$ on the Krein subspace spanned by the observation noise vector $\left[ \begin{array} {cc} ({\bf w}_k^N)^T & ({\bf w}_k^{R N})^T \end{array} \right]^T$ is then given by \[ \hat{\bf u}_k^N = \left[ \begin{array} {cc} {\cal G}_N^{\theta} & {\cal G}_N^{R \theta} \end{array} \right] \left[ \begin{array} {c} {\bf w}_k^N \\ {\bf w}_k^{N R} \end{array} \right] \: , \] where \[ \left[ \begin{array} {cc} {\cal G}_N^{\theta} & {\cal G}_N^{R \theta} \end{array} \right] = \left[ \begin{array} {cc} {\cal H}_N^T & {\cal L}_N^T \end{array} \right] ({\cal K}_N^{\theta})^{-1} \: , \] and the residual $\tilde{\bf u}_k^N = {\bf u}_k^N -\hat{\bf u}_k^N$ has for inner product \begin{eqnarray} \langle \tilde{\bf u}_k^N \, , \, \tilde{\bf u}_k^N \rangle \stackrel{\triangle}{=} {\cal Q}_N^{\theta} &=& I_{Nm} - \left[ \begin{array} {cc} {\cal G}_N^{\theta} & {\cal G}_N^{R \theta} \end{array} \right] {\cal K}_N^{\theta} \left[ \begin{array} {c} ({\cal G}_N^{\theta})^T \\ ({\cal G}_N^{R \theta})^T \end{array} \right] \nonumber \\[2ex] &=& [ I_{Nm} + {\cal H}_N^T {\cal H}_N - \theta {\cal L}_N^T {\cal L}_N ]^{-1} \: . \label{4.11} \end{eqnarray} The matrix ${\cal Q}_N^{\theta} $ will be positive definite if and only if \begin{equation} \theta < \theta_N \stackrel{\triangle}{=} 1/\lambda_1 ( {\cal L}_N (I_{Nm} + {\cal H}_N^T{\cal H}_N)^{-1} {\cal L}_N^T ) \: . \label{4.12} \end{equation} Note that this condition is also necessary and sufficient to ensure that the Schur complement $S_N^{\theta}$ in (\ref{4.10}) is negative definite. Then by multiplying the observation equation obtained by combining equations (\ref{3.10}) and (\ref{4.8}) by ${\cal R}_N \left[ \begin{array} {cc} {\cal G}_N^{\theta} & {\cal G}_N^{R \theta} \end{array} \right]$ and subtracting it from (\ref{3.9}), we obtain the state-space equation \begin{equation} x_{k+1}^d = \alpha_N^{\theta} x_k^d + {\cal R}_N \tilde{\bf u}_k^N +{\cal R}_N {\cal G}_N^{\theta} {\bf y}_k^N \label{4.13} \end{equation} with \[ \alpha_N^{\theta} \stackrel{\triangle}{=} A^N - {\cal R}_N [ {\cal G}_N^{\theta} {\cal O}_N + {\cal G}_N^{R \theta} {\cal O}_N^R ] \: , \] where the driving noise is now orthogonal to the noises ${\bf w}_k^N$ and ${\bf w}_k^{R N}$ appearing in observation equations (\ref{3.10}) and (\ref{4.8}). Accordingly, the Riccati equation associated to the downsampled model takes the form \begin{equation} P_{k+1}^d = r_d^{\theta} (P_k^d) \stackrel{\triangle}{=} \alpha_N^{\theta} [ (P_k^d)^{-1} + \Omega_N^{\theta} ]^{-1} ( \alpha_N^{\theta})^T + W_N^{\theta} \: , \label{4.14} \end{equation} where \begin{eqnarray} \Omega_N^{\theta} &=& \left[ \begin{array} {cc} {\cal O}_N^T & ({\cal O}_N^R)^T \end{array} \right] ({\cal K}_N^{\theta})^{-1} \left[ \begin{array} {c} {\cal O}_N \\ {\cal O}_N^R \end{array} \right] \nonumber \\ &=& \Omega_N + {\cal J}_N^T (S_N^{\theta})^{-1} {\cal J}_N \label{4.15} \end{eqnarray} with \[ {\cal J}_N \stackrel{\triangle}{=} {\cal O}_N^R -{\cal L}_N{\cal H}_N^T [I+{\cal H}_N{\cal H}_N^T]^{-1}{\cal O}_N \] and \begin{equation} W_N^{\theta} = {\cal R}_N{\cal Q}_N^{\theta}{\cal R}_N^T \:. \label{4.16} \end{equation} For $\theta=0$, the matrices $\Omega_N^{\theta}$ and $W_N^{\theta}$ coincide with the risk-neutral Gramians $\Omega_N$ and $W_N$ defined in (\ref{3.13}) and (\ref{3.14}). These matrices are positive definite for $N \geq n$ if and only if the pairs $(C,A)$ and $(A,B)$ are observable and reachable, respectively. Since ${\cal Q}_N^{\theta}$ is positive definite for $0 \leq \theta < \theta_N$, we deduce that $W_N^{\theta} >0$ over this range as long as $(A,B)$ is reachable and $N \geq n$. On the other hand, the Schur complement matrix $S_N^{\theta}$ is negative definite for $0 \leq \theta < \theta_N$, so \[ \Omega_N^{\theta} < \Omega_N \label{4.17} \] over this range. To establish that there exists a range $0 \leq \theta < \tau_N$ over which $\Omega_N^{\theta}$ remains positive definite when $(C,A)$ is observable, we use the following observation. \vskip 2ex \begin{lemma} \label{lem2} Over $0 \leq \theta < \theta_N$, the Gramians $\Omega_N^{\theta}$ and $W_N^{\theta}$ are monotone decreasing, and monotone nondecreasing, respectively, with respect to the partial order defined on nonnegative definite matrices. \end{lemma} \vskip2ex \begin{proof} We have \[ \frac{d~}{d\theta} (S_N^{\theta})^{-1}= -(S_N^{\theta})^{-1} \Big(\frac{d~}{d\theta}S_N^{\theta} \Big) (S_N^{\theta})^{-1} = -(\theta S_N^{\theta})^{-2} < 0 \] and \[ \frac{d~}{d\theta} Q_N^{\theta}= -Q_N^{\theta} \frac{d~}{d\theta} (Q_N^{\theta})^{-1} Q_N^{\theta} = Q_N^{\theta}{\cal L}_N^T {\cal L}_N Q_N^{\theta} \geq 0 \: . \] \qquad \end{proof} \vskip 2ex To understand why $\Omega_N^{\theta}$ and $W_N^{\theta}$ vary in opposite direction as $\theta$ increases, note that $W_N^{\theta}$ can be viewed as a measure of the uncertainty introduced by the process noise in the state-space model, whereas $\Omega_N^{\theta}$ is a measure of the information about the state contained in a block observation. As the risk-sensitivity parameter $\theta$ increases, it is natural that the uncertainty matrix $W_N^{\theta}$ should increase and the information matrix $\Omega_N^{\theta}$ should decrease. Let $\tau_N < \theta_N$ be the first value of $\theta$ for which $\Omega_N^{\theta}$ becomes singular. Then since $\Omega_N^{\theta}$ and $W_N^{\theta}$ are positive definite for $\theta \in [0,\tau_N)$, we conclude that over this range the Riccati map $r_d^{\theta}$ is strictly contractive and has a unique fixed point $P$ in ${\cal P}$. Like the risk-neutral case, we have $r_d^{\theta} = (r^{\theta})^N$. However, because the image $r^{\theta}({\cal P})$ is not completely contained in ${\cal P}$, to ensure that $P$ is also the unique fixed point of $r^{\theta}$, we must also require that $r^{\theta} (P) \in {\cal P}$. Note indeed that if \begin{equation} P = (r^{\theta})^N (P) \: , \label{4.18} \end{equation} by applying $r^{\theta}$ to both sides of (\ref{4.18}), we obtain \[ r^{\theta}(P) = (r^{\theta})^N (r^{\theta} (P)) \] so $r^{\theta}(P)$ is a fixed point of $(r^{\theta})^N$. If $r^{\theta}(P) \in {\cal P}$, we must have \[ P= r^{\theta} (P) \] since $(r^{\theta})^N$ has a unique fixed point in ${\cal P}$. At this point it is worth pointing out that until now we have ignored an important constraint \cite{BS,HSK1} for the risk-sensitive filter, namely that the matrix \begin{equation} V_t = (P_t^{-1} - \theta D^TD)^{-1} \label{4.19} \end{equation} should be positive definite for all $t$. If this condition is satisfied, then the fixed point $P$ of $r_d^{\theta}$ will be in ${\cal P}$, ensuring that it is the unique fixed point of $r^{\theta}$. \section{Positiveness conditions for $V_t$} \label{sec:posit} In this section we identify conditions on the initial covariance $P_0$ and risk-sensitivity parameter $\theta$ which ensure that the trajectory of iteration $P_{t+1} = r^{\theta} (P_t)$ satisfies $V_t >0$ for all $t$. Our analysis will exploit the monotonicity of Riccati operator $r^{\theta} (P)$ with respect to the partial order of positive definite matrices. \vskip 2ex \begin{lemma} \label{lem3} Let $P_1$ and $P_2$ be two matrices in ${\cal P}$ such that $P_1 \geq P_2$ and $P_1^{-1} - \theta D^TD > 0$. Then \begin{equation} r^{\theta} (P_1) \geq r^{\theta} (P_2) \: . \label{5.1} \end{equation} \end{lemma} \begin{proof} The monotonicity of $r^{\theta}$ is due to the fact that the inversion of positive definite matrices reverses their partial order. In addition, congruence transformations and translation by symmetric matrices preserve the partial order. Since the operator $r^{\theta}(P)$ in (\ref{4.4}) can be expressed in terms of two nested inversions of positive definite matrices, two matrix translations and a congruence transformation, it is monotone in $P$.\qquad \end{proof} \vskip 2ex Next, observe that for any $n \times p$ observer gain matrix $G$, the risk-sensitive Riccati equation (\ref{4.5}) can be rewritten as \begin{eqnarray} P_{t+1} &=& (A -GC)(P_{t}^{-1}- \theta D^TD)^{-1} (A-GC)^T + GG^T +BB^T \nonumber \\ && - [(A-GC)(P_t^{-1}-\theta D^TD)^{-1}C^T -G](R_t^{\nu})^{-1} \nonumber \\ && \hspace*{0.5in} \times [(A-GC)(P_t^{-1} - \theta D^T D)^{-1}C^T -G]^T \: . \label{5.2} \end{eqnarray} This expression can be obtained by writing $A= (A-GC)+GC$ in (\ref{4.5}) and performing simple algebraic manipulations. While it may appear surprising that a free matrix gain $G$ can be introduced in the equation, the above modification has actually a simple explanation. Consider the state-space model (\ref{3.1})--(\ref{3.2}). We can always design a preliminary suboptimal observer \begin{equation} \hat{x}_{t+1}^S = A \hat{x}_t^S + G(y_t -C\hat{x}_t^S) \: . \label{5.3} \end{equation} Then the residual $\tilde{x}_t^S = x_t - \hat{x}_t^S$ admits the state-space model \begin{eqnarray} \tilde{x}_{t+1}^S &=& (A-GC) \tilde{x}_t^S + Bu_t -Gv_t \nonumber \\ y_t - C \hat{x}_t^S &=& C \tilde{x}_t^S + v_t \: , \label{5.4} \end{eqnarray} for which the only difference with respect to the original model (\ref{3.1})--(\ref{3.2}) is that the process noise $Bu_t -Gv_t$ and measurement noise $u_t$ are now correlated. The risk-neutral and risk-sensitive problems associated to the original model (\ref{3.1})--(\ref{3.2}) and modified model (\ref{5.4}) are exactly the same since observations $y_t$ and \begin{equation} y_t^S \stackrel{\triangle}{=} y_t - C\hat{x}_t^S \label{5.5} \end{equation} can be obtained causally from each other. In particular, the variance matrices $P_t$ of the error are the same for both models. Thus it should not be a surprise that the solution $P_t$ of Riccati equation (\ref{4.5}) should also solve the risk-sensitive Riccati equation (\ref{5.2}) corresponding to modified model (\ref{5.4}). One important advantage of introducing the free matrix gain $G$ is that when the pair $(C,A)$ is observable, the characteristic polynomial of the closed-loop observer matrix $A-GC$ can be assigned arbitrarily \cite{Kai}. In particular, it is possible to ensure that the matrix $A-GC$ is stable, i.e. all its eigenvalues are strictly inside the unit circle. In this case, let \[ r \stackrel{\triangle}{=} \max_{1 \leq i \leq n} |\lambda_i (A-GC)| \] denote its spectral radius. For $\rho <1/r$, the matrix $\rho (A-GC)$ will also be stable, and when $(A,B)$ is reachable, the algebraic Lyapunov equation (ALE) \begin{equation} \Sigma_{\rho} = \rho^2 (A-GC) \Sigma_{\rho} (A-GC)^T + BB^T + GG^T \label{5.6} \end{equation} admits a unique positive definite solution \begin{equation} \Sigma_{\rho} = \sum_{k=0}^{\infty} \rho^{2k} (A-GC)^k (BB^T + GG^T) ((A-GC)^k)^T \: . \label{5.7} \end{equation} Note that $\Sigma_{\rho}$ is positive definite if and only if the pair $A-GC$, $\left[ \begin{array} {cc} B & G \end{array} \right]$ is reachable. But if this pair is not reachable, by the Popov-Belevich-Hautus (PBH) test \cite[p. 366]{Kai}, there must be a left eigenvector $z^T$ of $A-GC$ which is orthogonal to the column space of $\left[ \begin{array} {cc} B & G \end{array} \right]$, so \[ z^T (A-GC) = \lambda z^T \hspace*{0.15in}, \hspace*{0.15in} z^T B = z^T G = 0 \: . \] This implies $z^T A = \lambda z^T$, so $z^T$ is a left eigenvector of $A$ perpendicular to the column space of $B$, which implies that $(A,B)$ is not reachable, a contradiction. If we select $1 < \rho < 1/r$, the matrix $\Sigma_{\rho}$ is positive definite and the matrix \begin{equation} M \stackrel{\triangle}{=} (1-\rho^{-2}) \Sigma_{\rho}^{-1} - \theta D^T D \label{5.8} \end{equation} will be non-negative definite if and only if the matrix \[ \tilde{M} \stackrel{\triangle}{=} I_n -\theta \frac{\rho^2}{\rho^2 -1} \Sigma_{\rho}^{1/2} D^TD \Sigma_{\rho}^{1/2} \] is non-negative definite. But because the matrices $\Sigma_{\rho}^{1/2} D^T D \Sigma_{\rho}^{1/2}$ and $D\Sigma_{\rho}D^T$ have the same nonzero eigenvalues, $\tilde{M}$ is non-negative definite if and only if \begin{equation} \theta \frac{\rho^2}{\rho^2 -1} D \Sigma_{\rho} D^T \leq I_q \label{5.9} \end{equation} or equivalently \begin{equation} 0 \leq \theta \leq \beta_{\rho} \stackrel{\triangle}{=} \frac{\rho^2-1}{ \rho^2 \lambda_1 (D \Sigma_{\rho} D^T)} \label{5.10} \end{equation} where $\lambda_1 (D \Sigma_{\rho} D^T)$ is the largest eigenvalue of $D\Sigma_{\rho} D^T$. It is strictly positive since $\Sigma_{\rho}$ is positive definite and $D$ has full row rank. \vskip 2ex \begin{lemma} \label{lem4} If the initial variance $P_0$ for the risk-sensitive Riccati equation (\ref{4.4}) (or equivalently (\ref{5.2})) satisfies $0 < P_0 \leq \Sigma_{\rho}$ and $0 \leq \theta \leq \beta_{\rho}$, the entire trajectory of the recursion $P_{t+1} = r^{\theta} (P_t)$ satisfies $0 < P_t \leq \Sigma_{\rho}$, so $V_t >0$. Furthermore, for $P_0 = \Sigma_{\rho}$, the sequence $P_t$ is monotone decreasing. \end{lemma} \vskip 2ex \begin{proof} Suppose first that $P_0 = \Sigma_{\rho}$. Then the non-negative definiteness of the matrix $M$ in (\ref{5.8}) implies $V_0 >0$. By subtracting (\ref{5.2}) for $t=0$ from ALE (\ref{5.6}), we obtain \begin{eqnarray} \lefteqn{\Sigma_{\rho} - P_1 = (A-GC)(\rho^2 \Sigma_{\rho} - (\Sigma_{\rho}^{-1} - \theta D^TD)^{-1})(A-GC)^T} \nonumber \\ &&+ [(A-GC)(\Sigma_{\rho}^{-1}-\theta D^TD)^{-1}C^T -G]\nonumber \\ && \times (R_0^{\nu})^{-1} [(A-GC)(\Sigma_{\rho}^{-1} - \theta D^T D)^{-1}C^T -G]^T \: . \label{5.11} \end{eqnarray} But when $M$ is non-negative definite, the matrix \[ \rho^2 \Sigma_{\rho} - (\Sigma_{\rho}^{-1} -\theta D^T D)^{-1} \] appearing in the first term of the right hand side of (\ref{5.11}) is non-negative definite, which implies \begin{equation} P_1 = r^{\theta} (\Sigma_{\rho}) \leq P_0= \Sigma_{\rho} \:. \label{5.12} \end{equation} By induction, suppose that $P_t \leq P_{t-1}$. The motonicity of $r^{\theta}$ implies \[ P_{t+1} = r^{\theta} (P_t) \leq r^{\theta} (P_{t-1}) =P_t \] so $P_t$ is monotone decreasing. Next, consider the case of an initial condition $P_0 \leq \Sigma_{\rho}$. The monotonicity of $r^{\theta}$ implies \[ P_1 = r^{\theta} (P_0) \leq r^{\theta} (\Sigma_{\rho}) \leq \Sigma_{\rho} \] where the last inequality uses (\ref{5.12}). Proceeding by induction, we deduce that $P_t \leq \Sigma_{\rho}$ for all $t$. This implies \[ P_t^{-1} \geq \Sigma_{\rho}^{-1} > \theta D^TD \] so $V_t > 0$ for all $t$.\qquad \end{proof} \vskip 2ex \noindent {\it Remarks:} \begin{remunerate} \item[1)] For the risk-neutral case ($\theta=0$), the solution $\Sigma_{\rho}$ of the ALE (\ref{5.6}) is similar to an upper bound proposed for the positive definite solution of the algebraic Riccati equation (ARE) in \cite{DSW} (see also \cite{KP}), which was also shown to yield a monotone decreasing sequence of iterates. However the construction of the upper bound given in \cite{DSW} is purely algebraic, whereas for $\rho=1$ the covariance matrix $\Sigma_{\rho}$ can be interpreted as the steady-state error variance of the suboptimal filter (\ref{5.3}). \item[2)] Since the bound $\beta_{\rho}$ for the risk-sensitivity parameter depends on both $G$ and $\rho$, it is of interest to determine if a choice of $G$ and $\rho$ makes the bound as large as possible. Note in this respect that there exists a trade-off between making $\beta_{\rho}$ as large as possible and enlarging the set $0 \leq P_0 \leq \Sigma_{\rho}$ of allowable initial conditions, since from (\ref{5.9}) in order to increase the range of $\theta$ values, $\Sigma_{\rho}$ must be as small as possible, which shrinks the domain of allowable $P_0$s. A clue on how to select $G$ is provided by the scalar case analysis presented in \cite[Chap. 9]{Whi}. With $n=m=p=q=1$, if we select the gain $G=A/C$, $A-GC=0$ so $\rho$ can be selected arbitrarily large, and \[ \Sigma_{\rho} = \frac{A^2}{C^2} + B^2 \] for all $\rho$. Letting $\rho \rightarrow \infty$ in (\ref{5.10}), the bound $\beta_{\rho}$ then coincides with the scalar case bound derived on p. 116 of \cite{Whi}. This suggests that selecting a gain $G$ that moves all the eigenvalues of the closed-loop observer $A-GC$ to zero is likely to yield a satisfactory upper bound $\beta_{\rho}$. Note that in the multivariable case, $A-GC$ cannot in general be set to zero by selecting the gain matrix $G$, but the characteristic polynomial and some additional parameters (when $p>1$) can be assigned arbitrarily \cite[Chap. 7]{Kai}. Unfortunately, as will be demonstrated on an example in the next section, the gain $G$ which assigns all the eigenvalues of $A-GC$ to zero does not necessarily yield the largest possible value of $\beta_{\rho}$ and a comprehensive search over $G$ and $\rho$ is usually required to make $\beta_{\rho}$ as large as possible. \end{remunerate} By assembling the preliminary results of the current and previous sections, we obtain the following convergence theorem for risk-sensitive filters. \vskip 2ex \begin{theorem} \label{theo2} Assume that in system (\ref{3.1})--(\ref{3.2}), the pairs $(A,B)$ and $(C,A)$ are reachable and observable. Then if $0 \leq \theta < \tau_N$ and $\theta \leq \beta_{\rho}$ with $N \geq n$, the risk-sensitive Riccati map $r^{\theta}$ has a unique positive definite fixed point $P$ such that $P^{-1} - \theta D^TD >0$. Furthermore, if the initial condition $P_0$ of the Riccati equation satisfies $0 < P_0 \leq \Sigma_{\rho}$, the entire trajectory of iteration $P_{t+1} = r^{\theta} (P_t)$ stays in ${\cal P}$, satisfies $V_t >0$ and tends to $P$. In this case the limit $K$ of filtering gain $K_t$ as $t \rightarrow \infty$ has the property that $A-KC$ is stable. \end{theorem} \begin{proof} Since the trajectory $P_t$ stays in ${\cal P}$ and satisfies $V_t > 0$, and the $N$-fold operator $r_d^{\theta} = (r^{\theta})^N$ has a unique fixed point $P$ in ${\cal P}$, the sequence $P_t$ must tend to $P$, and $P$ must be such that $P^{-1} - \theta D^TD >0$. Then the stability of $A-KC$ can be established by applying Lyapunov stability theory to the risk-sensitive ARE. \[ P = (A-KC)(P^{-1} -\theta D^TD)^{-1}(A-KC)^T + BB^T + KK^T \: . \] \qquad \end{proof} \vskip 2ex This theorem answers in the affirmative the question posed in \cite{BS} whether it is possible to specify a-priori a range of risk-sensitivity parameters $\theta$ and initial conditions such that the risk-sensitive Riccati equation admits a solution. On the other hand, it leaves open the computation of the maximum value of $\theta$ (its breakdown value in the terminology of \cite{Whi}) for which a solution exists, which corresponds to the optimal $H^{\infty}$ filter. \section{Example} \label{sec:examp} To illustrate our results, we consider a system with \[ A= \left[ \begin{array} {cc} 0.1 & 1\\ 0 & 1.2 \end{array} \right] \hspace*{0.3in} C = \left[ \begin{array} {cc} 1 & -1 \end{array} \right] \] and $B=D=I_2$. Note that $A$ is unstable, $(A,B)$ is reachable, but the pair $(C,A)$ is barely observable, since the eigenvector \[ p= \left[ \begin{array} {c} 1 \\ 1.1 \end{array} \right] \] corresponding to the eigenvalue $\lambda=1.2$ is at a $92.72$ degree angle with respect to $C$. In this case, for $N=2$, the largest eigenvalue of matrix ${\cal L}_2 (I_4 + {\cal H}_2^T{\cal H}_2)^{-1}{\cal L}_2$ equals $1$, so $\theta_2 =2$. To evaluate the value $\tau_2$ for which $\Omega_2^{\theta}$ becomes singular, the smallest eigenvalue of Gramian $\Omega_2^{\theta}$ is plotted in \fig{laOm2} as a function of $\theta$ for $0 \leq \theta \leq 2 \times 10^{-3}$. For this example, it decreases linearly, and becomes negative at $\tau_2 = 0.715 \times 10^{-3}$. For completeness, the smallest eigenvalue of reachability Gramian $W_2^{\theta}$ is plotted in \fig{laW2} over the same range of $\theta$. It is monotone increasing, as expected, but the rate of increase is very small, since $\lambda_2(W_2^{\theta})$ varies from $1.002828$ to $1.02831$. Note that although we have selected $N=2$ here, larger values of $N$ can be considered, and in fact as $N$ increases, $\tau_N$ increases and $\theta_N$ decreases, and for this example both values tend to $1.33 \times 10^{-3}$ for large $N$. \vskip 2ex \begin{figure}[htb] \centering \includegraphics[width =3.5in, height=3in]{laOm2-th.eps} \caption{Smallest eigenvalue of observability Gramian $\Omega_2^{\theta}$ for $0 \leq \theta \leq 2 \times 10^{-3}$.} \label{laOm2} \end{figure} \begin{figure}[htb] \centering \includegraphics[width =3.5in, height=3.5in]{laW2-th.eps} \caption{Smallest eigenvalue of reachability Gramian $W_2^{\theta}$ for $0 \leq \theta \leq 2 \times 10^{-3}$.} \label{laW2} \end{figure} Next, to evaluate $\beta_{\rho}$, we observe that with the gain matrix \begin{equation} G= \left[ \begin{array} {c} -13.1\\ -14.4 \end{array} \right] \label{6.1} \end{equation} the closed-lood matrix \[ A-GC = \left[ \begin{array} {cc} 13.2 & -12.1\\ 14.4 & -13.2 \end{array} \right] \] is nilpotent, i.e., its eigenvalues are zero. Note however that $G$ is rather large, which reflects the weak observability of the system. In this case, if we select $\rho=2$, the solution $\Sigma_2$ of the Lyapunov equation (\ref{5.6}) is \[ \Sigma_2 = 10^{3} \left[ \begin{array} {cc} 1.4622 & 1.5954 \\ 1.5954 & 1.7431 \end{array} \right] \: . \] Its largest eigenvalue is $\lambda_1 (\Sigma_2) = 3.2042 \times 10^3$ and from (\ref{5.10}), we obtain $\beta_2 = 2.3407 \times 10^{-4}$. This bound is significantly smaller than $\tau_2$. To illustrate Lemma \ref{lem4}, the risk-sensitive Riccati iteration $P_{t+1} = r^{\theta} (P_t)$ is simulated with $\theta = \beta_2$ and initial condition $P_0 = \Sigma_2$. The two eigenvalues of $P_t$ and $V_t$ are plotted as a function of $t$ for $0 \leq t \leq 10$ in \fig{eigP} and \fig{eigV}, respectively. As expected, the eigenvalues remain positive and are monotone decreasing. The monotone decreasing property of the eigenvalues is due to the fact that if two $n \times n$ positive definite matrices $P$ and $Q$ are such that $P \geq Q$ and if the eigenvalues of $P$ and $Q$ are sorted in decreasing order, then $\lambda_i (P) \geq \lambda_i (Q)$ for $1 \leq i \leq n$. In other words, the eigenvalues follow the partial order of positive definite matrices. Since according to Lemma \ref{lem4}, the sequence $P_t$ is monotone decreasing, so are its eigenvalues. The figures indicate that the risk-sensitive Riccati equation converges very quickly, after 4 or 5 iterations. Note that if $P$ denotes the limit of $P_t$, its smallest eigenvalue is $1.003$, but the other eigenvalue is much larger and equals $332.4$. This reflects our earlier observation that one of the modes of the system is barely observable. The eigenvalues of the matrix $A-KC$ for the estimation error dynamics are $0.034$ and $0.776$, so the filter is stable, as expected. \vskip 2ex \begin{figure}[htb] \centering \includegraphics[width =4in, height=6in]{laPt.eps} \caption{Eigenvalues of Riccati solution $P_t$ for $0 \leq t \leq 11$ with $\theta = \beta_2$ and initial condition $P_0 = \Sigma_2$.} \label{eigP} \end{figure} \begin{figure}[htb] \centering \includegraphics[width =4in, height=6in]{laVt.eps} \caption{Eigenvalues of $V_t$ for $0 \leq t \leq 11$ with $\theta = \beta_2$ and initial condition $P_0 = \Sigma_2$.} \label{eigV} \end{figure} \noindent Finally, to illustrate the onset of breakdown as $\theta$ increases, the two eigenvalues of the fixed point solution $P^{\theta}$ of $r^{\theta}$ and of the corresponding matrix $V^{\theta} = ((P^{\theta})^{-1} - \theta I_2)^{-1}$ are plotted as a function of $\theta$ in \fig{thdepP} and \fig{thdepV}, respectively, for $0 \leq \theta \leq 0.95 \times 10^{-3}$. It is known \cite[p. 379]{HSK1} that $P^{\theta}$ is a monotone increasing function of $\theta$, and as expected the eigenvalues of $P^{\theta}$ are monotone increasing. However, while the change in the smaller eigenvalue is barely noticeable, the eigenvalue representing the weakly observable mode increases rapidly with $\theta$. As $\theta$ increases, the eigenvalues of $V^{\theta}$ start diverging, and the breakdown value of $\theta$ for this example is just above $0.95 \times 10^{-3}$. This value is significantly higher than the bound $\beta_2$ obtained by applying Lemma \ref{lem4} with the gain (\ref{6.1}), suggesting that the bound can be improved. In fact, an exhaustive search over $G$ and $\rho$ showed that $\beta_{\rho}$ is maximized by selecting \[ G = \left[ \begin{array} {c} -7.2196\\ -7.9753 \end{array} \right] \] and $\rho=1.2849$, in which case $\beta_{\rho} = 0.4824 \times 10^{-3}$. \vskip 2ex \begin{figure}[htb] \centering \includegraphics[width =4in, height=6in]{laP-th.eps} \caption{Eigenvalues of Riccati fixed point $P^{\theta}$ in function of $\theta$ for $0 \leq \theta \leq 0.95 \times 10^{-3}$.} \label{thdepP} \end{figure} \begin{figure}[htb] \centering \includegraphics[width =4in, height=6in]{laV-th.eps} \caption{Eigenvalues of $V^{\theta}$ in function of $\theta$ for $0 \leq \theta \leq 0.95 \times 10^{-3}$.} \label{thdepV} \end{figure} \section{Conclusions} \label{sec:conc} A convergence analysis of risk-sensitive filters has been presented. It relies on extending Bougerol's contraction analysis of risk-neutral Riccati equations to the risk-sensitive case. This was accomplished by considering a block-filtering implementation of the $N$-fold Riccati map and showning that this map is strictly contractive as long as an observability Wronskian depending on the risk-sensitivity parameter remains positive definite. A second condition was derived for the risk-sensitivity parameter and initial error variance to ensure that the trajectory of the risk-sensitive Riccati iteration stays positive definite at all times. The two conditions obtained can be viewed as multivariable versions of conditions obtained earlier by Whittle \cite[Chap. 9]{Whi} for the scalar case. Although the results we have presented concern filters with a constant risk-sensitivity parameter $\theta$, a closely related class of robust filters was derived recently \cite{LN} by assigning a fixed relative entropy tolerance to increments of the state-space model. In this case, the risk-sensitivity parameter is time-varying, but the tolerance is fixed, and based on computer simulations, it appears that the risk-sensitivity parameter and associated filter always converge as long as the relative entropy tolerance remains small. Since Bougerol's analysis \cite{Bou} is applicable to systems with random fluctuations, it is reasonable to wonder if the analysis presented here can be extended to establish the convergence of the filters discussed in \cite{LN}.

https://arxiv.org/abs/1804.11340

Local laws for polynomials of Wigner matrices

We consider general self-adjoint polynomials in several independent random matrices whose entries are centered and have the same variance. We show that under certain conditions the local law holds up to the optimal scale, i.e., the eigenvalue density on scales just above the eigenvalue spacing follows the global density of states which is determined by free probability theory. We prove that these conditions hold for general homogeneous polynomials of degree two and for symmetrized products of independent matrices with i.i.d. entries, thus establishing the optimal bulk local law for these classes of ensembles. In particular, we generalize a similar result of Anderson for anticommutator. For more general polynomials our conditions are effectively checkable numerically.

\section{Introduction} \label{sec:introduction} Polynomials of random matrices have been subject of intensive research in the last thirty years. In the 1980's Voiculescu realized that random matrices and their polynomials can be used to solve some basic problems in operator algebras of free groups, which gave birth to \emph{free probability theory}. Roughly speaking, large independent random matrices serve as concrete approximants to free elements in abstract noncommutative probability spaces, i.e. unital $C^*$-algebras with a tracial state. In other words, \emph{freeness} is the appropriate operator algebraic analogue of independence in classical probability. A classical example for such result is Theorem 2.2 from \cite{Voic} showing that the trace of a self-adjoint polynomial $p(X_1,\ldots,X_k)$ in $k$ independent $N\times N$ standard complex Gaussian (GUE) matrices converges in expectation and almost surely, as the size of the matrices goes to infinity, to the trace of the polynomial $p(\semic_1,\ldots,\semic_k)$ in free semicircular variables. Voiculescu's pioneering result has since been extended in many directions. Convergence in operator norm was proved in \cite{HaagThor}, while convergence of the spectrum, in particular absence of outliers, was established in \cite{HaagSchuThor}. Another direction of generalizations was to replace Gaussian matrices with Wigner matrices, i.e. retain independence of the matrix elements while dropping the special distribution; for the first such result see \cite{Dyke}, followed by many others, e.g. \cite{Ande_2013,BeliBercCapi,BeliCapi,CapiDonaMart,Male} and references therein. Yet another line of research concerns certain qualitative properties of the limiting spectral measure. For example, the limiting spectral measure for self-adjoint polynomials does not contain atoms \cite{MaiSpeiWebe,ShlySkou} and for monomials it is even absolutely continuous \cite{CharShly}. A common feature of all these results, as well as the scope of the underlying methods, is that they describe the spectrum of $p(X_1, \ldots, X_k)$ on the global scale, which is typically by a factor $N$ larger than the scale of the eigenvalue spacing. What happens on scales in between? Recent developments revealed that the eigenvalue density of Wigner and related matrices on \emph{mesoscopic} scales, i.e., scales involving $\sim N^{\gamma}$ eigenvalues for $0<\gamma<1$, also becomes deterministic in the large $N$ limit. Such results are commonly called \emph{local laws} and they have been established in increasing generality for hermitian matrices; with independent entries, see e.g. \cite{AjanErdoKrug_PTRF1,ErdoKnowYauYin,GotzNaumTikhTimu,HeKnowRose,TaoVu}, with general short range correlation structure for their matrix elements \cite{AjanErdoKrug_PTRF2,EKS18}, as well as for adjacency matrices for random regular graphs \cite{BaueHuanYau,BaueHuanKnowYau}. One of the main motivations for local laws is their key role in the proof of the Wigner-Dyson-Mehta conjecture on the local spectral universality, see \cite{ErdoYau_12}. Recent developments on the local ergodicity of the Dyson Brownian motion (DBM) have demonstrated that local laws are the only model-dependent inputs for the universality proofs using the DBM, see \cite{ErdoYauBook} for an overview and newer results in \cite{ErdoSchn,LandoYau,LandYau2}. In this paper we prove optimal local laws for self-adjoint polynomial models, thus connecting two large areas of recent research in random matrices. We will combine methods from free probability theory, most importantly the concept of linearization, with techniques developed for local laws, such as large deviations and fluctuation averaging phenomenon. We point out that mesoscopic spectral properties for general polynomials have not been studied before. Local laws have only been established for a very few specific polynomials such as (i) the anticommutator, $X_1X_2+X_2X_1$, of two independent Wigner matrices in \cite{Ande_2015} and (ii) the (non-Hermitian) product $Y_1Y_2\ldots Y_k$ of several independent \emph{i.i.d.} matrices in \cite{GotzNaumTikh2017,NemiLLP}. We now explain the method and some difficulties. The first major obstacle is that the entries of a general polynomial $P:=p(X_1,\ldots,X_k)$ of, say, independent $N\times N$ Wigner matrices, have a very complex non-local correlation structure. This makes it impossible to apply the tools developed in \cite{AjanErdoKrug_PTRF2} directly in the polynomial setting. However, the well-known \emph{linearization trick} \cite{HaagSchuThor, HaagThor} transforms the polynomial model into a much larger random matrix $\Hb$ with a transparent correlation structure. In fact, the linearized matrix is a tensor linear combination of the independent Wigner matrices with matrix coefficients whose dimension $m\times m$ depends only on the polynomial $p$ and is independent of $N$. This structure exactly corresponds to certain block matrices and more generally \emph{Kronecker random matrices} introduced in \cite{AltErdoKrugNemi_Kronecker}. We remark that the linearization technique has been widely used in the free probability community to study polynomials of random matrices on the global scale, see e.g. \cite{Ande_2013,BeliMaiSpei,HaagThor,HeltMaiSpei,HeltMaccVinn} and \cite[Chapter~5]{AndeGuioZeit} for a pedagogical introduction. Local laws for Kronecker matrix $\Hb$ have been studied in detail in \cite{AltErdoKrugNemi_Kronecker} by proving concentration of its resolvent $(\Hb-z I_m\otimes I_N)^{-1}$ around the solution of corresponding matrix Dyson equation for spectral parameter $z$ in complex upper half-plane. In contrast to the Kronecker case, to study the resolvent $(P-z)^{-1}$ of our polynomial, we have to consider the \emph{generalized resolvent} of the linearized matrix $\Hb$, i.e., $(\Hb - z J\otimes I_N)^{-1}$, where $J$ is a rank-one $m\times m$ matrix. Thus the results on the Kronecker matrices cannot be directly applied, in fact \emph{a priori} it is unclear whether the generalized resolvent is stable. This is the second major obstacle in the study of polynomial models, and we overcome it by simultaneously considering the generalized resolvent of $\Hb$ and its usual regularized version. It turns out that a certain nilpotency structure inherent for linearizations of polynomials yields the boundedness of the generalized resolvents even after the regularization is removed. After these two key obstacles cleared, we can essentially use the local law established for general Kronecker matrices in \cite{AltErdoKrugNemi_Kronecker} under two basic conditions: (i) the solution of the underlying Dyson equation is bounded and (ii) the stability operator is invertible. These conditions are verified for homogeneous polynomials of degree two in Wigner matrices, substantially generalizing the case of anticommutator studied by Anderson in \cite{Ande_2015}. We also verify them for the symmetrized product $Y_1 \cdots Y_k (Y_1 \cdots Y_k)^*$ of independent matrices with \emph{i.i.d.} entries. For more general polynomials, the validity of these conditions depends on the structure of the linearization but they are independent of $N$, so they are numerically checkable. Notice that the linearization of a polynomial is not unique. In fact, any of the standard linearizations, obtained via a simple recursive procedure, typically has unnecessarily large dimension. It is much more effective to use the so-called \emph{minimal linearization}, which is canonical \cite{BersReut,HeltMaccVinn}, and we present numerical examples to demonstrate its advantages. Since both linearizations are nilpotent, our theory equally applies to them. We expect that for any self-adjoint polynomial there exists a linearization for which the conditions (i) and (ii) above hold everywhere in the bulk, i.e., where the density of states is bounded and bounded away from zero, and in fact the minimal linearization is a natural candidate. The task of dealing with the generalized resolvent of the linearization is inherent in other works on polynomials of random matrices that use the resolvent method, see \cite{Ande_2013,CapiDonaMart,HaagThor,Male}. In the most general setup, Anderson in \cite{Ande_2013} used one of the explicit standard linearizations to prove the global law and the convergence of the norm for polynomials in Wigner matrices. The structure of the standard linearization allowed him to control the generalized resolvent directly from the resolvent of $P$ via Schur complement formula. A simpler version of this idea was presented in \cite[Chapter~5.5]{AndeGuioZeit}. For the canonical minimal linearization such simple a priori bound is not available. From algebraic point of view, the main novelty of our work is to identify a nilpotency structure in the minimal linearization and show that this structure is sufficient to control the generalized resolvent. From the analytic point of view, we advocate the method of the stability analysis of the Dyson equation combined with large deviation and fluctuation averaging estimates as presented in \cite{AltErdoKrugNemi_Kronecker}, which itself is a natural extension of many previous works on local laws for Wigner and Wigner-type matrices. This approach substitutes the Poincar\'{e} inequality used in \cite{CapiDonaMart} and the $L^p$ bounds used in \cite{Ande_2013} whose analogue for Wigner and Wishart matrices go back to Bai and Silverstein \cite{Bai93,BaiSilv98}. We close this introduction with a remark on local spectral universality. Our local law is optimal and it provides the necessary input for the customary proofs via the Dyson Brownian motion (DBM) as mentioned above. Thus we could easily prove bulk universality for polynomials that already have a small additive GUE component. We cannot, however, apply the usual DBM argument to the linearized matrix since it would need to assume that a small global Gaussian component is present in $\Hb$, but $\Hb$ has many zero blocks by construction. This fundamental difficulty has been overcome for certain band matrices \cite{BourErdoYauYin,BourYauYinBand18} which also has many zero entries. However, the specific band structure was essential in those proofs. For the local spectral universality for a polynomial $P$ the structure of the linearized matrix needs to be exploited in a similar fashion. We note that apart from the trivial case of hermitian polynomials of a single random matrix, currently the only nontrivial universality results for polynomials are obtained for very special cases and only for Gaussian matrices by exploiting their determinantal structure, see, e.g., the survey \cite{AkerIpse} on products of large Gaussian random matrices. In Section~\ref{sec:polyAndLin} we introduce the concept of nilpotent linearization, the corresponding Dyson equation and we present our main result together with the conditions expressed in terms of the solution to the Dyson equation. Section~\ref{sec:aboutlinearizations} is devoted to control the generalized resolvent by exploiting the nilpotent structure. In Section~\ref{sec:MDE} we present the existence and uniqueness of the solution to the Dyson equation by using semicircular variables. In Section~\ref{sec:locallaw} we give a proof of the local law. Finally, as an application, in Section~\ref{sec:examples} we show the optimal bulk local law for general homogeneous polynomials of degree two in Wigner matrices and for symmetrized products of matrices with \emph{i.i.d.} entries. Additional information on two different linearizations, as well as their numerical comparison are deferred to Appendix~\ref{sec:aLinCM}, while in Appendix~\ref{sec:semicircular} we collected some basic information on semicircular variables for the reader's convenience. \medskip \emph{Acknowledgement.} The authors are grateful to Oskari Ajanki for his invaluable help at the initial stage of this project, to Serban Belinschi for useful discussions and to Alexander Tikhomirov for calling our attention to the model example in Section~\ref{sec:exampl-sing-valu} \section{Main results} \label{sec:polyAndLin} \subsection{Linearization and Dyson equation in $C^*$-algebras}\label{sec:linDyson} Fix $\alpha_*,\beta_*\in \NN$. Let $\mathscr{A}$ be a unital $C^*$-algebra with norm $\| \cdot \|_{\Acc}$ and identity element $\cstarunit$, and let $x_1,\ldots,x_{\alpha_*},y_1,\ldots,y_{\beta_*} \in \mathscr{A}$ with $x_i^*=x_i$ for $1\leq i\leq \alpha_*$. For any $n\in \NN$ and $\rb = (r_1,\ldots,r_n) \in \Acc^{n}$ we define $\|\rb \|: = \max_{1\leq i \leq n} \|r_i\|_{\Acc}$. Denote by \begin{equation*} \CC\la \xb,\yb,\yb^* \ra : = \CC\la x_1,\ldots,x_{\alpha_*},y_1,\ldots,y_{\beta_*},y_1^*,\ldots,y_{\beta_*}^* \ra \end{equation*} the set of polynomials with complex coefficients in noncommutative elements $\{x_{\alpha},y_{\beta},y_{\beta}^*,1\leq\alpha\leq \alpha_*, 1\leq \beta\leq \beta_*\}$. Let $p:= p(\xb, \yb, \yb^*) \in \CC \la \xb,\yb,\yb^* \ra$ and assume that $p$ is self-adjoint, i.e., \begin{equation*} (p(\xb, \yb, \yb^*))^*=p(\xb, \yb, \yb^*) . \end{equation*} It is a common and convenient practice to study the polynomials via their linearizations. Linearization allows to transform polynomial model into a linear one, which is typically easier to analyze. The price for doing this is the increased dimension of the model, which can quickly become prohibitive for more complicated polynomials. \begin{defn}[Self-adjoint linearization]\label{def:saLin} Let $m\in\NN$ and let $\Lb\in (\CC\la \xb,\yb,\yb^* \ra)^{m\times m}$ be a matrix, whose matrix elements are polynomials of degree at most $1$. Suppose that \begin{equation} \label{eq:Lexpr1} \Lb = \begin{pmatrix} \lambda & \ell^* \\ \ell & \widehat{\Lb} \end{pmatrix}, \end{equation} where $\widehat{\Lb}$ is the $(m-1)\times (m-1)$ submatrix of $\Lb$. We call $\Lb$ a \emph{self-adjoint linearization} (or simply \emph{linearization}) of $p\in \CC \la \xb, \yb, \yb^* \ra$ if $\Lb^*=\Lb$ and there exists $\epsD>0$ such that for all $\|\xb\|<\epsD, \|\yb\|< \epsD$ matrix $\widehat{\Lb}$ is invertible and satisfies \begin{equation}\label{eq:Lin11} p = \lambda - \ell^*\widehat{\Lb}^{-1}\ell . \end{equation} We will refer to $m$ as the \emph{dimension} of the linearization $\Lb$. \end{defn} Note that due to the property $\Lb^*=\Lb$ a self-adjoint linearization $\Lb$ can be written as \begin{equation}\label{eq:linearization} \Lb = K_0\otimes\cstarunit - \sum_{\alpha=1}^{\alpha_*} K_\alpha\otimes x_\alpha - \sum_{\beta=1}^{\beta_*}( L_{\beta}\otimes y_\beta+L^*_{\beta} \otimes y^*_\beta) , \end{equation} where $K_\alpha, L_\beta\in \CC^{m\times m}$ and $K_0^*=K_0$, $K_\alpha^*=K_\alpha$. In this paper all linearizations are self-adjoint, so we will not stress self-adjointness all the times. For each polynomial one can write many different linearizations. In the related literature \cite{AndeGuioZeit,BersReut,HaagThor,HeltMaccVinn} one can distinguish two groups of methods used for constructing the linearizations of polynomial (and more generally rational) functions. One group uses very explicit algorithms to build linearizations first for monomials, and then extending them to linear combinations of monomials. These algorithms are well-know, but for the sake of completeness we will give in Appendix~\ref{sec:abiglinearization} a version of such an explicit linearization. This is a standard construction that typically yields a linearization in very high dimension. For many practical reasons it is better to work with smaller linearizations, which naturally leads to the notion of \emph{minimal} linearization. \begin{defn}[Minimal linearization] A linearization of a polynomial is called \emph{minimal} if it has the smallest dimension among all linearizations. \end{defn} Minimal linearization can be obtained by reducing the dimension of some previously constructed linearization (see e.g. \cite[Chapter~2.3]{BersReut}) and then using the symmetrization trick if needed to restore self-adjointness \cite[Lemma~4.1 (3)]{HeltMaccVinn}. For completeness, as well as for the reader's convenience, in Appendix~\ref{sec:minimallinearization} we present a somewhat different algorithmic procedure that directly yields a minimal (self-adjoint) linearization from any (self-adjoint) linearization. Typically the dimension of a minimal linearization is significantly smaller compared to the standard linearization constructed in Appendix~\ref{sec:abiglinearization} (see Appendix~\ref{sec:aLinComp} for comparison), which makes it much more convenient to work with if we want to study the model numerically. In order to use linearizations for studying the resolvents of polynomials of random matrices it will be convenient to work with a special class of \emph{nilpotent} linearizations that we introduce now. \begin{defn}[Nilpotent family] A family of matrices $\{ R_i \in \CC^{m\times m} \; : \; i\in I\}$ is called \emph{nilpotent} if there exists an integer $n$ such that $R_{i_1} R_{i_2} \ldots R_{i_n}=0$ for any $n$-tuple of indices $(i_1, i_2,\ldots , i_n)\in I^n$. \end{defn} Define the matrix $J := e_1 e_1^{t} \in \CC^{m\times m}$, where $e_1= (1,0,\ldots 0)^{t}\in \CC^m$, i.e. $J$ is an $m\times m$ matrix having the $(1,1)$-entry equal to $1$ and all the other entries equal to zero. Let $\la \cdot\, , \cdot \ra\, : \, \CC^{m}\times \CC^{m} \rightarrow \CC$ be the usual scalar product in $\CC^{m}$ linear in the second variable. For brevity we will denote $\llbr n \rrbr := \{1,\ldots , n\}$ for any $n \in \NN$. \begin{defn}[Nilpotent linearization]\label{def:nilpotentlin} A linearization of a polynomial $p$ of the form \eqref{eq:linearization} is called \emph{nilpotent} if \begin{itemize} \setlength\itemsep{0em} \item[(i)] $K_0$ is invertible; \item[(ii)] $\la e_1 , K_0^{-1} e_1\ra=1$ \item[(iii)] The family of matrices \begin{equation*} \Big\{ \proj' K_\alpha K_0^{-1} \proj',\, \proj' L_\beta K_0^{-1} \proj', \,\proj' L_\beta^* K_0^{-1} \proj' \; : \; \alpha \in \llbr \alpha_* \rrbr, \; \beta \in \llbr \beta_* \rrbr\Big\} \end{equation*} is nilpotent, where we set $\proj : = JK_0^{-1}$ and $\proj':=I-\proj$ \end{itemize} \end{defn} One can easily see that (ii) is equivalent to $p(0,0,0)=\cstarunit$. Indeed, let $\la \cdot, \cdot \ra_{\Acc} \, : \, \CC^{m}\otimes \Acc \times \CC^{m}\otimes \Acc\rightarrow \Acc$ be an operator given by \begin{equation*} \la \lb , \rb \ra_{\Acc} := \sum_{k=1}^{m} \lb_{k}^{*} \rb_{k} \end{equation*} where $\rb, \lb \in \CC^{m}\otimes \Acc$, $\rb = \sum_{k=1}^{m} e_k \otimes \rb_{k}$, $\lb = \sum_{k=1}^{m} e_k \otimes \lb_{k}$ and $e_k = (\delta_{ik})_{k=1}^{m}$. Then by the Schur complement formula and \eqref{eq:Lin11} we have that \begin{equation*} \la e_1 \otimes \cstarunit, \Lb^{-1} e_1\otimes \cstarunit \ra_{\Acc} = p^{-1} . \end{equation*} If we now take $\xb=\yb=0$, then $\Lb^{-1} = K_0^{-1}\otimes \cstarunit$ and thus $\la e_1, K_0^{-1}\, e_1 \ra \cstarunit = (p(0,0,0))^{-1}$. Shifting the polynomial by a constant, without loss of generality, we may and will assume in the rest of the paper that the constant term of the polynomial is 1, i.e. we write $p(\xb, \yb, \yb^*)=\cstarunit- q(\xb,\yb,\yb^*)$ for some polynomial $q(\xb,\yb,\yb^*)$ with $q(0,0,0)=0$. Furthermore, note that $\pi$ is a projection by (ii) and $J=e_1 e_1^{t}$, but in general it is not an orthogonal projection. We will show in Section~\ref{sec:Nilp} that for any polynomial of the form $p = \cstarunit - q$ the both linearizations constructed in Appendix~\ref{sec:aLinCM} belong to the class of nilpotent linearizations. This property will be used to obtain an a priori bound for the \emph{generalized resolvent} of the linearization, that we define below. Denote by $\Amxm$ the set of $m\times m$ matrices with elements from $\Acc$. We can look at $\Lb$ as an operator on $\Amxm$ equipped with the Banach space structure from $\mathscr{A}$. For any $z\in\CC_+$ we will consider the \emph{generalized resolvent} of $\Lb$ defined as $(\Lb-zJ \otimes \cstarunit)^{-1}$. From the Schur complement formula \begin{equation}\label{eq:schur} \big\la e_1 \otimes \cstarunit, (\Lb-zJ \otimes \cstarunit )^{-1} e_1\otimes \cstarunit\big\ra_{\Acc} = \big(\lambda - z{\cstarunit} - \ell^*\widehat{\Lb}^{-1}\ell\big)^{-1} = (p-z{\cstarunit})^{-1} \end{equation} i.e. the $(1,1)$-component of the generalized resolvent is the resolvent of $p$, viewed as an element of $\mathscr{A}$. In particular, if we take $\Acc$ to be $\CC^{N\times N}$, then the resolvent of a polynomial $p$ of matrices of size $N\times N$ is given by the {upper left $N\times N$ block} of the generalized resolvent of the corresponding linearization. In Section~\ref{sec:atrivialbound} we show that generalized resolvent of a nilpotent linearization is well defined for all $z\in \CC_{+}$. More precisely, define a norm $\|\, \cdot \, \|$ on $\Amxm$ by \begin{equation*} \| \Rb \| := \max_{1\leq k, l \leq m} \| \Rb_{kl} \|_{\Acc} , \end{equation*} where $\Rb = \sum_{k,l=1}^{m} E_{kl}\otimes \Rb_{kl}$ and $E_{kl} = (\delta_{ik}\delta_{lj})_{1\leq i,j\leq m}$ is the standard basis in $\CC^{m\times m}$. Then the following lemma holds. \begin{lem} \label{lem:trivBound} Let $q\in \CC\la \xb,\yb,\yb^* \ra$ be a self-adjoint polynomial with $q(0,0,0)=0$. Let $\Lb\in \Amxm$ be a nilpotent linearization of $\cstarunit - q(\xb,\yb, \yb^*)$ Then there exist $C_1>0$ and $n_1\in \NN$, depending on $q$, such that for all $z\in \CC_+$ \begin{equation} \label{eq:trivResBound} \| ( \Lb -zJ\otimes \cstarunit)^{-1}\| \leq C_1 \Big(1+\frac{1}{\Im z}\Big) (1+\max_{1\leq \alpha\leq \alpha_*}\|x_{\alpha}\|_{\Acc}^{n_1}+\max_{1\leq \beta \leq \beta_*}\|y_{\beta}\|_{\Acc}^{n_1}) . \end{equation} \end{lem} Suppose now that we have a nilpotent linearization $\Lb$ in the form \eqref{eq:linearization}. Define the linear map $\SuOp:\CC^{m\times m}\rightarrow \CC^{m\times m}$ by \begin{equation}\label{eq:SuOp} \SuOp[R] = \sum_{\alpha=1}^{\alpha_*} K_\alpha R K_\alpha +\sum_{\beta=1}^{\beta_*} (L_{\beta} R L^*_{\beta} +L^*_{\beta} R L_{\beta}) . \end{equation} For any $z\in \CC_+$ (spectral parameter) we consider the equation \begin{equation} \label{eq:MDE} -\frac{1}{M} = z J - K_0 + \SuOp[M] \end{equation} for the unknown matrix $M\in \CC^{m\times m}$. We will always consider solutions with the side condition that $\Im M\ge 0$ where $\Im M = \frac{1}{2i}(M-M^*) $. We call equation \eqref{eq:MDE} the \emph{Dyson equation for linearization (DEL)}. Note that \eqref{eq:MDE} is very similar to the matrix Dyson equations (MDE) extensively studied in the literature in connection with large random matrices (see e.g. \cite{HeltRaFaSpei} and \cite{AjanErdoKrug_PTRF2}). Their solutions typically give the deterministic part of the resolvent of a random matrix. The main difference between \eqref{eq:MDE} and the MDE in \cite{AjanErdoKrug_PTRF2} is that instead of the identity matrix, the spectral parameter $z$ appears with a coefficient matrix $J$ of smaller rank. This makes \eqref{eq:MDE} much harder to analyse, in particular basic boundedness and stability properties do not follow directly from the structure of \eqref{eq:MDE} alone. Nevertheless, the fact that \eqref{eq:MDE} comes from the linearization of a polynomial, especially that it is nilpotent still ensures its good properties. The next lemma states the existence and uniqueness of the solution to \eqref{eq:MDE}, in particular we may denote the solution $M=M(z)$, indicating its dependence on the spectral parameter. \begin{lem}[Existence and uniqueness of solution of DEL]\label{lem:MDEexistence} Let $\Lb$ be a nilpotent linearization of the self-adjoint polynomial $\cstarunit - q(\xb, \yb,\yb^*)$ with $q(0,0,0)=0$ and let $\SuOp:\CC^{m\times m}\rightarrow \CC^{m\times m}$ be defined as in \eqref{eq:SuOp}. There exists a matrix-valued function $\GSol \,: \, \CC_{+}\rightarrow \CC^{m\times m}$ such that for all $z\in \CC_+$ \begin{itemize} \setlength\itemsep{0em} \item[(i)] $\| \GSol (z)\| \leq C \, ( 1+1/\Im z)$ for some $C > 0$ independent of $z$; \item[(ii)] $\GSol (z)$ depends analytically on $z$; \item[(iii)] $\Im \GSol (z) \geq 0$; \item[(iv)] $\GSol (z)$ satisfies the DEL \eqref{eq:MDE}. \end{itemize} This function is the unique solution of \eqref{eq:MDE} in the class of matrix-valued functions with $\Im M(z)\ge 0$ that are analytic in the upper half-plane, i.e. if $M'\,: \, \CC_+ \rightarrow \CC^{m\times m}$ and $M'$ satisfies $(i)-(iv)$, then $M'=\GSol$. \end{lem} For any any matrix $R\in \CC^{m\times m}$ we denoted by $\| R\|$ the operator norm induced by the Euclidean norm in $\CC^{m}$. Lemma~\ref{lem:MDEexistence} will be proven in Section~\ref{sec:MDE}. In the rest of the paper, $\GSol=\GSol(z)$ will always denote the unique solution to \eqref{eq:MDE} obtained in Lemma~\ref{lem:MDEexistence}. \begin{lem}[Stieltjes transform representation] \label{lem:StiltRep} Let $M$ be the unique solution to DEL \eqref{eq:MDE} constructed in Lemma~\ref{lem:MDEexistence}. We then have the following \begin{itemize} \setlength\itemsep{0em} \item[(i)] For any $z\in \CC_+$ \begin{equation}\label{eq:matSt} \GSol(z) = \GSol^{\infty}+\int_{\RR} \frac{V(dx)}{x-z} , \end{equation} where $\GSol^{\infty} \in \CC^{m\times m}$ and $V(dx)$ is a (positive semidefinite) matrix-valued measure on $\RR$; \item[(ii)] For almost every $x \in \RR$ there exists the limit $\lim_{y \rightarrow 0} \pi^{-1} \Im \GSol(x+\imunit y) = V(x) \in \CC^{m\times m}$; if the limit is finite on some interval $I\subset \RR$ everywhere, then $V(dx)$ is absolutely continuous on $I$ and $V(dx)=V(x)dx$; \item[(iii)] There exists $C >0$ such that for any $z\in \CC_+$ \begin{equation*} \Tr \Im \GSol(z) \leq C \la e_1, \Im \GSol(z) \, e_1 \ra . \end{equation*} In particular, we have that $\mathrm{supp} (V_{11}) = \mathrm{supp} (\Tr V)$. \end{itemize} \end{lem} This lemma will be proven in Section~\ref{sec:MDE}. \subsection{Polynomials and linearization of random matrices} \label{sec:RM} In this section we specialize the setup from Section~\ref{sec:linDyson} to the matrix setup, i.e. to the case when $\Acc=\CC^{N\times N}$ for some $N\in \NN$ equipped with the usual Euclidean matrix norm and hermitian conjugation to define the $C^*$-algebra structure. To indicate this special case in the notation, instead of $x_1, x_2, \ldots y_1, y_2, \ldots$ we will use capital letters, $X_1, X_2, \ldots$ and $Y_1, Y_2,\ldots$ for the $N\times N$ matrices. Moreover, we assume that these matrices are random and independent. The self-adjoint matrices $X_\alpha$ will be Wigner-type matrices, i.e. they have independent elements up to hermitian symmetry, while the matrices $Y_\beta$ will have independent entries without any restriction. We assume the matrix elements are centered and their variances are $1/N$. We collect these assumptions in the following list: \begin{ass} Let $\Xb^{(N)}:=\{X_{\alpha}^{(N)}, \alpha \in \llbr \alpha_* \rrbr\}$ and $\Yb^{(N)}:=\{Y_{\beta}^{(N)}, \beta \in \llbr \beta_* \rrbr\}$ be two families of $N\times N$ random matrices such that \begin{itemize} \setlength\itemsep{0em} \item[(\textbf{H1})] the joint family {$\Xb^{(N)}\cup \Yb^{(N)}$ is independent;} \item[(\textbf{H2})] $X_{\alpha}^{(N)}$ are Hermitian random matrices having {independent} centered entries with variance $N^{-1}$; \item[(\textbf{H3})] $Y_{\beta}^{(N)}$ are (non-Hermitian) random matrices having {independent} centered entries with variance $N^{-1}$; \item[(\textbf{H4})] entries of $X_{\alpha}^{(N)}$ and $Y_{\beta}^{(N)}$ satisfy the moment bounds \begin{equation*} \max_{i,j\in \llbr N \rrbr} \Big(\max_{\alpha\in \llbr \alpha_* \rrbr} \ME{ |\sqrt{N} X_{\alpha}^{(N)}(i,j)|^{p}}+ \max_{\beta\in \llbr \beta_* \rrbr} \ME{ |\sqrt{N} Y_{\beta}^{(N)}(i,j)|^{p} }\Big) \leq C_{p} . \end{equation*} \end{itemize} \end{ass} Another set of assumptions concerns the properties of the solution of the Dyson equation for linearization \eqref{eq:MDE}. To this end we introduce the notions of the \emph{$\kappa$-bulk} and the \emph{stability operator}, which plays a crucial role in the analysis of the stability of the solution of \eqref{eq:MDE}. \begin{defn}[Density of states]\label{def:ds} Let $\GSol$ denote the unique solution of the DEL \eqref{eq:MDE} given in Lemma~\ref{lem:MDEexistence}. Define function $\rho : \RR \rightarrow [0, +\infty]$ \begin{equation*} \rho (E) := \lim_{\eta\rightarrow 0}\frac{1}{\pi} \la e_1, \Im \GSol(E+\imunit \eta)\, e_1 \ra , \end{equation*} where the limit exists due to Lemma~\ref{lem:StiltRep}. We will refer to $\rho$ as the (absolutely continuous part of the) \emph{density of states} of $p$. \end{defn} It will follow from the proof of Lemma~\ref{lem:MDEexistence} (see \eqref{eq:ImM11}) that $\rho (E)$ does not depend on the choice of linearization. \begin{defn}[Bulk, $\epsB$-bulk]\label{def:Bkappa} We say that $E\in\RR$ belongs to the \emph{bulk} if $0<\rho(E)<\infty$. For any $\epsB>0$ we define the set $B_{\epsB}:=\{E\in\RR \, : \, \epsB<\rho(E)<\epsB^{-1} \}$, which we will call the $\epsB$-\emph{bulk}. \end{defn} We remark that Definition~\ref{def:ds} slightly differs from the standard definition used for the matrix Dyson equation in \cite{AjanErdoKrug_PTRF2}, where the density of states was defined via the trace of $\Im \GSol$ as $\widetilde \rho (E) := \lim_{\eta\rightarrow 0}\frac{1}{\pi m} \Tr \Im \GSol(E+\imunit \eta)$ and not only its (1,1)-component. The current definition is justified since our main object is the polynomial $p$ and not its linearization $\Lb$. Note, that it follows from $(iii)$ in Lemma~\ref{lem:StiltRep} that $\rho(E)$ and $\widetilde\rho(E)$ are comparable, i.e., the bulk could have been defined using $\rho$ instead of $\widetilde{\rho}$. From now on we fix $\epsB>0$. \begin{defn}[Stability operator] Let $\SuOp$ be defined as in \eqref{eq:SuOp} and $M$ obtained in Lemma~\ref{lem:MDEexistence}. Then the operator \begin{equation*} \GStOp: \CC^{m\times m} \rightarrow \CC^{m\times m} , \quad \GStOp\,[R] = R- \GSol \, \SuOp[R]\,\GSol \end{equation*} is called the \emph{stability operator} corresponding to the DEL \eqref{eq:MDE}. \end{defn} \begin{ass} There exists a constant $C_3$, depending only on $\epsB$ and the polynomial $p$, such that for any $z\in \CC_+$ with $\Re z \in B_{\epsB}$ and $0 < \Im z < \infty$ we have \begin{itemize} \setlength\itemsep{0em} \item[(\textbf{M1})] $\|\GSol(z)\|\leq C_3$; \item[(\textbf{M2})] $\|\GStOp^{-1}(z)\|\leq C_3$. \end{itemize} \end{ass} The local law is formulated using the following the notion of \emph{stochastic domination}. \begin{defn}[Stochastic domination] Let $\mathscr{D}\subset \CC$ and let $(\Phi_w^{(N)})_{N\in \NN}$ and $(\Psi_w^{(N)})_{N\in \NN}$, $w\in \mathscr{D}$, be two sequences of nonnegative random variables. Then we say that $\Phi$ is stochastically dominated by $\Psi$ uniformly on $\mathscr{D}$ if for all $\epsE, D > 0$ there exists $C(\epsE, D)>0$ such that for all $N\in \NN$ \begin{equation*} \Pr{ \Phi_w^{(N)} \geq N^{\epsE} \Psi_w^{(N)}} \leq \frac{C(\epsE, D)}{N^{D}} \end{equation*} with $C(\epsE, D)$ independent of $N$ and $w$. \end{defn} We are now ready to state our main result. \begin{thm}[Local law for polynomials] \label{thm:mainPoly} Let $p\in \CC \la \xb, \yb, \yb^* \ra$ be a self-adjoint polynomial with $p(0,0,0)=\cstarunit$ and let $\Lb$ be a nilpotent linearization of $p$ be defined as in \eqref{eq:linearization}. Let $\GSol(z)$ be a solution of the corresponding \emph{DEL} \eqref{eq:MDE} constructed as in Lemma~\ref{lem:MDEexistence}. Suppose that the families of random matrices $\Xb^{(N)}, \Yb^{(N)}$ satisfy conditions \textup{(\textbf{H1})-(\textbf{H4})} and that $\GSol(z)$ satisfies \textup{(\textbf{M1})-(\textbf{M2})} for some fixed $\epsB>0$. Then the local law holds for $p(\Xb^{(N)},\Yb^{(N)}, [\Yb^{(N)}]^*)$ in the $\epsB$-bulk up to the optimal scale, i.e., for any $\gamma>0$ \begin{equation} \label{eq:locallawpoly} \max_{i,j\in\llbr N \rrbr} \|\PRes_{ij}(z)-\la e_1, \GSol(z)\, e_1 \ra \delta_{ij}\| \prec \sqrt{\frac{1}{N\Im z}} ,\qquad \Big\| \frac{1}{N} \sum_{i=1}^{N}\PRes_{ii}(z)-\la e_1, \GSol(z)\, e_1 \ra \Big\| \prec \frac{1}{N\Im z} \end{equation} uniformly for $z\in D_{\kappa, \gamma}$ with $D_{\kappa, \gamma}:= \{z\in \CC\; : \;\Re z\in B_{\epsB}, \; N^{-1+\gamma} \leq \Im z \leq 1\}$, where $\PRes(z)$ is the resolvent matrix of the polynomial \begin{equation*} \PRes(z) := \big(p(\Xb^{(N)},\Yb^{(N)}, [\Yb^{(N)}]^*)-z\otimes I_N\big)^{-1} . \end{equation*} \end{thm} Note that the typical distance between two adjacent eigenvalues in the bulk is of order $N^{-1}$. Thus the exponent in the bound $\Im z \geq N^{-1+\gamma}$ is the lowest possible that allows for a deterministic limit of the resolvent. In \eqref{eq:locallawpoly} we formulated the local law in the entrywise and in the tracial sense, but it is easy to extend the first result to a more general \emph{anisotropic sense} that approximates $\langle \ub, g(z) \vb\rangle$ for any deterministic vectors $\ub, \vb\in \CC^N$ by adapting the method from \cite[Section 7]{BloeErdoKnowYauYin} or \cite[Section~6.1]{AjanErdoKrug_PTRF1} to Kronecker random matrices in the spirit of \cite{AltErdoKrugNemi_Kronecker}. We now comment on the assumptions \textup{(\textbf{M1})-(\textbf{M2})}. We expect that these hold for an appropriate linearization for any self-adjoint polynomial, but this remains an open question in full generality. However, in Section~\ref{sec:example} we prove \textup{(\textbf{M1})-(\textbf{M2})} for a general homogeneous polynomial of degree two in Wigner matrices. For other polynomials we remark that these two assumptions can be checked numerically since they require the solution $M(z)$ of the Dyson equation for a linearization \eqref{eq:MDE} that can be computed by an effective fixed point iteration. The numerics can be speeded up by reducing the dimension of the \emph{DEL}, e.g. by considering the minimal linearization instead of the standard one, see Section~\ref{sec:aLinComp} for some examples. Local laws provide information that can be used to estimate with relatively high precision the locations of individual eigenvalues of the corresponding random matrix, as well as to show the delocalization of its eigenvalues. These results have been obtained many times in the literature, therefore we state them without proofs and refer the interested reader to, e.g., \cite[Section~5]{AjanErdoKrug_PTRF1}. \begin{cor}[Bulk rigidity] Let $\lambda_i$, $1\leq i\leq N$, be the eigenvalues of $p(\Xb^{(N)},\Yb^{(N)},[\Yb^{(N)}]^*)$ in the increasing order. For each $E\in B_{\epsB}$ denote by $\iota(E)$ the index of the eigenvalue that is typically close to $E$, i.e., \begin{equation}\label{eq:iota} \iota(E) := \Big\lceil N \int_{-\infty}^{E} \rho(dx) \Big\rceil . \end{equation} Then \begin{equation*} \sup \{ |\lambda_{\iota(E)}-E| \, : \, E\in B_{\epsB}\} \prec \frac{1}{N} . \end{equation*} \end{cor} \begin{cor}[Delocalization of bulk eigenvectors] For $1\leq i\leq N$ denote by $\mathbf{u}_i\in \ CC^{N}$ the normalized eigenvector of $p(\Xb^{(N)},\Yb^{(N)},[\Yb^{(N)}]^*)$ that corresponds to the eigenvalue $\lambda_i$. Then for any deterministic unit vector $\mathbf{b}\in \CC^{N}$ and $E\in B_{\epsB}$ we have \begin{equation*} |\mathbf{b} \cdot \mathbf{u}_{\iota(E)}| \prec \frac{1}{\sqrt{N}} , \end{equation*} where $\iota(E)$ is defined as in \eqref{eq:iota}. \end{cor} Although the main focus of this paper is the local law and its consequences, we remark that our method also gives an optimal $1/N$ speed of convergence of the empirical spectral distribution of any self-adjoint polynomial $p(\Xb^{(N)},\Yb^{(N)}, [\Yb^{(N)}]^*)$ to its limiting density on the global scale. More precisely, we have the following: \begin{pr}[Speed of convergence]\label{pr:speed} Let $p\in \CC \la \xb, \yb, \yb^* \ra$ be a self-adjoint polynomial with $p(0,0,0)=\cstarunit$ and let $\rho$ be the density of states. Suppose that the families of random matrices $\Xb^{(N)}, \Yb^{(N)}$ satisfy conditions \textup{(\textbf{H1})-(\textbf{H4})}. Let $\lambda_1, \ldots, \lambda_N$ be the eigenvalues of $p(\Xb^{(N)},\Yb^{(N)},[\Yb^{(N)}]^*)$ and let $f$ be a smooth function on $\RR$. Then \begin{equation}\label{fp} \Big| \frac{1}{N} \sum_{i=1}^N f(\lambda_i) - \int_\RR f(x)\rho(dx)\Big| \prec \frac{1}{N}. \end{equation} In particular, we have \begin{equation}\label{pp} \Big| \frac{1}{N} \Tr p(\Xb^{(N)},\Yb^{(N)},[\Yb^{(N)}]^*) - \int_\RR x\rho(dx)\Big| \prec \frac{1}{N}. \end{equation} \end{pr} Note that this result does not assume the conditions \textup{(\textbf{M1})-(\textbf{M2})}. In fact, \eqref{pp} shows that the speed of convergence in the customary definition of asymptotic freeness of the random variables $(\Xb^{(N)},\Yb^{(N)},[\Yb^{(N)}]^*)$ is of order $1/N$. In the rest of the paper, whenever this does not cause any confusion, we will suppress the $N$-dependence in $\Xb$, $\Yb$ and other $N$-dependent objects. \section{Linearizations: nilpotency and a priori bound} \label{sec:aboutlinearizations} In this section we prove that the linearizations constructed in Appendix~\ref{sec:aLinCM} possess some nice properties. More precisely, we show in Lemmas \ref{lem:bignilp} and \ref{lem:anilpotency} that both the standard and the minimal linearizations are nilpotent, and then, in Section~\ref{sec:atrivialbound}, we prove that the bound \eqref{eq:trivResBound} holds for the generalized resolvents of any nilpotent linearization. Note, that in Appendix~\ref{sec:aLinCM} we consider linearizations of noncommutative polynomials in \emph{self-adjoint} variables only. We start this section with a short remark explaining why this is indeed enough. Define the real and imaginary parts of an element $a\in\Acc$ as \begin{equation*} \Re a := \frac{a+a^*}{2} ,\quad \Im a := \frac{a-a^*}{2\,\imunit} \end{equation*} so that $\Re a$ and $\Im a$ are self-adjoint and $a=\Re a +\imunit \Im a$. Then \eqref{eq:linearization} can be rewritten as \begin{equation}\label{eq:aLsum3} \Lb = K_0\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\otimes x_{\gamma} , \end{equation} where $\gamma_*:= \alpha_*+2\beta_*$ and we defined for $\beta\in \llbr \beta_*\rrbr$ \begin{equation}\label{eq:aNewK} x_{\alpha_*+\beta} := \sqrt{2}\Re x_{\beta} ,\quad x_{\alpha_*+\beta_*+\beta} := \sqrt{2}\Im x_{\beta} ,\quad K_{\alpha_*+\beta} := \sqrt{2}\Re L_{\beta} ,\quad K_{\alpha_*+\beta_*+\beta} := -\sqrt{2} \Im L_{\beta} . \end{equation} We can now use formulas \eqref{eq:linearization}, \eqref{eq:aLsum3} and \eqref{eq:aNewK} to switch between linearizations of $q\in \CC \la \xb, \yb,\yb^*\ra $ and $\tilde{q} \in \CC \la \xb, \Re \yb, \Im \yb \ra$ with $\tilde{q}(\xb,\Re \xb, \Im \xb)=q(\xb,\yb,\yb^*)$. Clearly $q(0,0,0)=0$ is equivalent to $\tilde q(0,0,0)=0$ which we will assume in the sequel. Therefore, in the current section and Section~\ref{sec:MDE}, with a slight abuse of notation, by defining $\xb = (x_{\gamma},\, \gamma \in \llbr \gamma_*\rrbr)$, it will be enough to consider only self-adjoint polynomials only of the form $\tilde q(\xb)$ with $\tilde q(0)=0$ and linearizations $\Lb$ of $\tilde{q}(\xb)$ of the form \eqref{eq:aLsum3}. In Section~\ref{sec:locallaw}, we will go back to the linearization \eqref{eq:linearization}. \subsection{Joint nilpotency} \label{sec:Nilp} In the next lemma we show that the standard linearization constructed in Appendix~\ref{sec:abiglinearization} is nilpotent. \begin{lem}[Nilpotency of the standard linearization]\label{lem:bignilp} Let $\tilde{q} \in \CC \la \xb \ra$ be a self-adjoint polynomial satisfying $\tilde q(0)=0$. Let \begin{equation*} \Lb = K_0\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\otimes x_{\gamma} \end{equation*} be a linearization of $\cstarunit - \tilde{q}$ constructed in Appendix~\ref{sec:abiglinearization}. Then $\Lb$ is nilpotent. \end{lem} \begin{proof} First of all, note that (i) and (ii) in the definition of the nilpotent linearization follow directly from \eqref{eq:apermut}. Thus, in order to finish the proof we need to show that the family \begin{equation*} \Big\{ \proj' K_\gamma K_0^{-1} \proj', \; : \; \gamma \in \llbr \gamma_* \rrbr \Big\} \end{equation*} is nilpotent, where, we recall, \begin{equation*} \proj := J K_0^{-1} ,\quad \proj' := I-\proj . \end{equation*} From the representation \eqref{eq:apermut} we have that $\proj = J K_0^{-1} = J= K_0^{-1}J$, hence $\proj$ commutes with $K_0^{-1}$. This implies that for any $\gamma\in \llbr \gamma_* \rrbr$ \begin{equation*} \proj' K_{\gamma} K_0^{-1} \proj' = \proj' K_{\gamma}(\proj + \proj') K_0^{-1} \proj' = \proj' K_{\gamma} \proj' K_0^{-1} \proj' = \Theta \widehat{K}_{\gamma} \widehat{K}_0^{-1} \Theta^{-1} , \end{equation*} where we recall the structure of $K_0$ and $ K_\gamma$, explicitly indicating their minors after separating the first row and column: \begin{equation*} K_0 = \left( \begin{array}{c|ccc} 1 & 0 & \cdots & 0 \\ \hline 0 & & & \\ \vdots & & \Theta \widehat{K}_0 & \\ 0 & & & \end{array} \right) ,\quad K_{\gamma} = \left( \begin{array}{c|ccc} * &* & \cdots & * \\ \hline * & & & \\ \vdots & & \Theta \widehat{K}_{\gamma} & \\ *& & & \end{array} \right) , \end{equation*} with $\widehat{K}_0,\widehat{K}_{\gamma}\in \{0,1\}^{(m-1)\times (m-1)}$ and $\widehat{K}_0$ being a permutation matrix. Stars indicate arbitrary unspecified matrix elements. The key observation is the following relation between the location of nonzero matrix elements of $\widehat{K}_0$ and $\widehat{K}_{\gamma}$ \begin{equation}\label{eq:alphaPerm} \mbox{if }\quad \widehat{K}_0 = (e_{\tau_1}, e_{\tau_2}, \ldots, e_{\tau_{m-1}})^{t}\quad \mbox{ then }\quad \widehat{K}_{\gamma} = (c^{\gamma}_1 e_{\tau_2}, c^{\gamma}_2 e_{\tau_3}, \ldots, c^{\gamma}_{m-2} e_{\tau_{m-1}},0)^{t} \end{equation} where $\tau$ is the permutation on $\llbr m-1 \rrbr$ determined by the permutation matrix $\widehat{K}_0$, $e_\tau$ is the $\tau$-th coordinate vector in $\CC^{m-1}$ and $c_i^{\gamma} \in \{0,1\}$ are some constants. In other words \eqref{eq:alphaPerm} says that an entry of $\widehat{\Lb}$ may contain $x_{\gamma}$ only if the entries just below it and on the right side contain $\cstarunit$. For the proof, notice that this rule is immediate for the basic block of $\Lb$ for monomials \eqref{eq:abasicblock} and it remains valid after taking the conjugate transpose or applying any of the rules (R1)-(R3). Next, the fact that $\widehat{K}_0$ is a symmetric permutation matrix implies that $\widehat{K}_0^{-1} = \widehat{K}_0$, which means that $\widehat{K}_0^{-1} = (e_{\tau_1}, e_{\tau_2}, \ldots, e_{\tau_{m-1}})$. Therefore, \begin{equation*} \widehat{K}_{\gamma}\widehat{K}_0^{-1} = (c^{\gamma}_1 e_{\tau_2}, c^{\gamma}_2 e_{\tau_3}, \ldots, c^{\gamma}_{m-2} e_{\tau_{m-1}},0)^{t} (e_{\tau_1}, e_{\tau_2}, \ldots, e_{\tau_{m-1}}) = \sum_{i=1}^{m-2} c_i^{\gamma}E_{i,i+1} \end{equation*} is strictly upper-triangular. A family of strictly upper-triangular matrices is obviously nilpotent. This finishes the proof of the lemma. \end{proof} \begin{lem}[Nilpotency of the minimal linearization]\label{lem:anilpotency} Let $\tilde{q} \in \CC \la \xb \ra$ be a self-adjoint polynomial satisfying $\tilde q(0)=0$. Let \begin{equation*} \Lb = K_0\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\otimes x_{\gamma} \end{equation*} be a minimal linearization of $\cstarunit - \tilde{q}$. Then $\Lb$ is nilpotent. \end{lem} \begin{proof} By \eqref{eq:aKmin}, \eqref{eq:aKB} and \eqref{eq:111} properties (i) and (ii) from the definition of the nilpotent linearization are satisfied. Thus it is left to show that the family of matrices \begin{equation*} \Big\{ \proj' K_\gamma K_0^{-1} \proj', \; : \; \gamma \in \llbr \gamma_* \rrbr \Big\} \end{equation*} is nilpotent. Define for brevity $A_{\gamma}:=K_{\gamma}K_0^{-1}$. Assume that $\|\xb \| \leq \epsD $ for $\epsD>0$ small enough, so that \begin{equation}\label{eq:aTBexp0} \Big\la K_0^{-1}e_1 \otimes \cstarunit , \Big(I\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma} \otimes x_{\gamma}\Big)^{-1} e_1 \otimes \cstarunit \Big\ra_{\Acc} = \frac{1}{\cstarunit - p(\xb)} \end{equation} and the objects on both sides can be expanded into a convergent geometric series. Using the notation \begin{equation*} \pproj := \proj\otimes \cstarunit ,\quad \pproj' := I\otimes \cstarunit - \pproj , \end{equation*} and defining the trace operator $ \la \,\cdot \,\ra_{\Acc} : \Amxm \rightarrow \Acc$ by \begin{equation*} \la \Rb \ra_{\Acc} := \sum_{k=1}^{m} \Rb_{kk} \quad \mbox{ for } \Rb \in \Amxm , \end{equation*} equality \eqref{eq:aTBexp0} can be rewritten as \begin{equation*} \Big\la \pproj \Big(I\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma} \otimes x_{\gamma}\Big)^{-1} \Big\ra_{\Acc} = \frac{1}{\cstarunit - \tilde{q}(\xb)} . \end{equation*} Now using the geometric series expansion for $(I\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma} \otimes x_{\gamma})^{-1} $ we have \begin{align} \frac{1}{\cstarunit - \tilde{q}(\xb)} &= \Big\la \pproj \, \Big(I\otimes \cstarunit + \sum_{k=1}^{\infty} \sum_{(\alpha_1, \ldots , \alpha_k)\in \llbr \gamma_*\rrbr^k} A_{\alpha_1} \cdots A_{\alpha_k}\otimes x_{\alpha_1}\cdots x_{\alpha_k} \Big) \Big\ra_{\Acc} \nonumber \\\label{eq:TBexp2} &= \la \proj \ra \otimes \cstarunit + \sum_{k=1}^{\infty} \sum_{(\alpha_1, \ldots , \alpha_k)\in \llbr \gamma_*\rrbr^k} \la \proj A_{\alpha_1} \cdots A_{\alpha_k} \ra \otimes x_{\alpha_1}\cdots x_{\alpha_k} , \end{align} where $\la \, \cdot \, \ra$ denotes the usual trace operator, i.e., $\la R \ra = \Tr R$ for $R\in \CC^{m\times m}$. Since the polynomial $\tilde{q}$ has no constant term, we can write it as $\tilde{q}(\xb) = \sum_{\beta=1}^{\infty} \tilde{q}_{\beta}(\xb)$, where $\tilde{q}_\beta$ is a homogeneous polynomial of degree $\beta$. Clearly this summation is finite since $\tilde{q}_\beta\equiv 0$ whenever $\beta$ is larger than the degree of $\tilde{q}$. In other words, $\tilde{q}_1$ denotes the linear part of $\tilde{q}$, $\tilde{q}_2$ the quadratic part, etc. Then $(\cstarunit - \tilde{q}(\xb))^{-1}$ can be expanded as \begin{equation}\label{eq:TBexp3} \frac{1}{\cstarunit - \tilde{q}(\xb)} = \cstarunit +\sum_{\ell=1}^{\infty} \tilde{q}^{\ell}(\xb) = \cstarunit +\sum_{\ell=1}^{\infty} \sum_{\beta_1,\ldots,\beta_{\ell} =1}^{\mathrm{deg}(\tilde{q})} \tilde{q}_{\beta_1} \cdots \tilde{q}_{\beta_{\ell}} . \end{equation} By construction we know that $\la \proj \ra =1$. If we now compare \eqref{eq:TBexp2} and \eqref{eq:TBexp3} recursively degree by degree, then from degree one terms we get that \begin{equation*} \tilde{q}_1 = \sum_{\gamma=1}^{\gamma_*} \la \proj A_{\gamma} \ra x_{\gamma} . \end{equation*} Similarly, from comparing the degree two terms we have \begin{equation*} \tilde{q}_2+\tilde{q}_1 \tilde{q}_1 = \sum_{\alpha_1, \alpha_2 =1}^{\gamma_*} \la \proj A_{\alpha_1} A_{\alpha_2} \ra x_{\alpha_1} x_{\alpha_2} , \end{equation*} so that \begin{equation*} \tilde{q}_2 = \sum_{\alpha_1, \alpha_2 =1}^{\gamma_*} \la \proj A_{\alpha_1} A_{\alpha_2} \ra x_{\alpha_1} x_{\alpha_2} - \sum_{\alpha_1=1}^{\gamma_*} \la \proj A_{\alpha_1} \ra x_{\alpha_1} \sum_{\alpha_2=1}^{\gamma_*} \la \proj A_{\alpha_2} \ra x_{\alpha_2} = \sum_{\alpha_1,\alpha_2=1}^{\gamma_*} \la \proj A_{\alpha_1} \proj' A_{\alpha_2} \ra x_{\alpha_1} x_{\alpha_2} , \end{equation*} where we used the following factorization property, based upon $J=e_1^* e_1$: for any $B_1, B_2 \in \CC^{m\times m}$ \begin{equation}\label{eq:trCommu} \la \proj B_{1} \ra \la \proj B_{2} \ra = \la e_1, K_0^{-1} \, B_1\, e_1 \ra \la e_1, K_0^{-1}\, B_2 \, e_1 \ra = \la \proj B_1 \proj B_2 \ra . \end{equation} Next, from comparing the degree three terms in \eqref{eq:TBexp2} and \eqref{eq:TBexp3} we get \begin{equation*} \tilde{q}_3 + \tilde{q}_2\tilde{q}_1 + \tilde{q}_1 \tilde{q}_2 + \tilde{q}_1\tilde{q}_1 \tilde{q}_1 = \sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1}A_{\alpha_2} A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} \end{equation*} and thus \begin{align*} \tilde{q}_3 & = \sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1}(\proj+\proj')A_{\alpha_2}(\proj+\proj') A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} -\sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1}\proj'A_{\alpha_2} \ra \la \proj A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} \\ & \quad - \sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1} \ra \la \proj A_{\alpha_2}\proj'A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} -\sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1} \ra \la \proj A_{\alpha_2} \ra \la \proj A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} \\ &= \sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1}(\proj+\proj')A_{\alpha_2}(\proj+\proj') A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} -\sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1}\proj'A_{\alpha_2} \proj A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} \\ &\quad - \sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1} \proj A_{\alpha_2}\proj'A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} -\sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1}\proj A_{\alpha_2}\proj A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} \\ &= \sum_{\alpha_1,\alpha_2,\alpha_3=1}^{\gamma_*} \la \proj A_{\alpha_1}\proj'A_{\alpha_2}\proj' A_{\alpha_3} \ra x_{\alpha_1} x_{\alpha_2} x_{\alpha_3} , \end{align*} where again, similarly as for quadratic terms, we used \eqref{eq:trCommu} to change the order of multiplication and taking trace. Now we prove the general formula for $\tilde{q}_\ell$ by induction on the degree $\ell$. Suppose that for any $k < \ell$ \begin{equation}\label{eq:aIndHyp} \tilde{q}_k = \sum_{\alpha_1,\ldots ,\alpha_k=1}^{\gamma_*} \la \proj A_{\alpha_1}\proj'A_{\alpha_2}\proj'\cdots A_{\alpha_{k-1}}\proj' A_{\alpha_k} \ra x_{\alpha_1} \cdots x_{\alpha_k} . \end{equation} Then from comparing the degree $\ell$ terms in \eqref{eq:TBexp2} and \eqref{eq:TBexp3} we get \begin{align*} &\sum_{\alpha_1,\ldots ,\alpha_{\ell}=1}^{\gamma_*} \la \proj A_{\alpha_1}A_{\alpha_2}\cdots A_{\alpha_{\ell-1}} A_{\alpha_{\ell}} \ra x_{\alpha_1} \cdots x_{\alpha_{\ell}} \\ & \quad = \sum_{j_1,\ldots,j_{\ell-1}\in \{0,1\}} \sum_{\alpha_1,\ldots ,\alpha_{\ell}=1}^{\gamma_*} \la \pi A_{\alpha_1}\kappa_{j_1} A_{\alpha_2}\kappa_{j_2}\cdots A_{\alpha_{\ell-1}}\kappa_{j_{\ell-1}} A_{\alpha_{\ell}} \ra x_{\alpha_1} \cdots x_{\alpha_{\ell}} \\ & \quad = \tilde{q}_{\ell}+\tilde{q}_{\ell-1}\tilde{q}_1+\tilde{q}_{\ell-2}\tilde{q}_2+\tilde{q}_{\ell-2} \tilde{q}_1 \tilde{q}_1 +\cdots + \tilde{q}_1 \tilde{q}_1\cdots \tilde{q}_1 \end{align*} where $\kappa_0= \proj'$ and $\kappa_1=\proj$. Using the factorization property \eqref{eq:trCommu} and the induction hypothesis \eqref{eq:aIndHyp} one can see that for the terms in the last sum can be written as \begin{equation*} \tilde{q}_{i_1}\cdots \tilde{q}_{i_t} = \sum_{\alpha_1,\ldots ,\alpha_{\ell}=1}^{\gamma_*} \la \pi A_{\alpha_1}\kappa_{j_1} A_{\alpha_2}\kappa_{j_2}\cdots A_{\alpha_{\ell-1}}\kappa_{j_{\ell-1}} A_{\alpha_{\ell}} \ra x_{\alpha_1} \cdots x_{\alpha_{\ell}} \end{equation*} with \begin{equation*} j_{s} =\left\{ \begin{array}{ll} 1,& s\in \{i_1,i_1+i_2,\ldots, i_1+i_2+\cdots + i_{t-1}\}, \\ 0, &\mbox{otherwise} \end{array} \right. \end{equation*} and $i_s<m$. Therefore, we deduce by induction that for all $\ell \in \NN$ \begin{equation*} \tilde{q}_{\ell} = \sum_{\alpha_1,\ldots ,\alpha_{\ell}=1}^{\gamma_*} \la \proj A_{\alpha_1}\proj'A_{\alpha_2}\proj'\cdots A_{\alpha_{\ell-1}}\proj' A_{\alpha_{\ell}} \ra x_{\alpha_1} \cdots x_{\alpha_{\ell}} . \end{equation*} In particular, if $\tilde{q}$ is a polynomial of degree $\ell^*$, then for any $\ell>\ell^*$ \begin{equation}\label{eq:anilpotencytrace} \la \proj A_{\alpha_1}\proj'A_{\alpha_2}\proj'\cdots A_{\alpha_{\ell-1}}\proj' A_{\alpha_{\ell}} \ra = 0 . \end{equation} Since $\Lb$ is a minimal linearization and $K_0, K_{\gamma} \in \CC^{m\times m}$, by Proposition~\ref{prop:min} we have that \begin{equation*} \mathrm{span} \Big(\bigcup_{\overline{\alpha} \in \Icc} A_{\overline{\alpha}} e_1 \Big) = \CC^m ,\quad \mathrm{span} \Big(\bigcup_{\overline{\alpha}\in \Icc} A^*_{\overline{\alpha}} K_0^{-1} e_1 \Big) = \CC^m , \end{equation*} where $\Icc$, $A_{\overline{\alpha}}$ and $A^*_{\overline{\alpha}}$ are defined in \eqref{eq:aIcc} and \eqref{eq:Ralpha}. For $\overline{\alpha}= (\alpha_1, \ldots, \alpha_k) \in \Icc$ denote \begin{equation*} \tilde{r}_{\overline{\alpha}} := A_{\alpha_1} \proj' \cdots \proj' A_{\alpha_k} e_1 ,\quad \tilde{l}_{\overline{\alpha}} := A_{\alpha_1}^* {\proj'}^* \cdots {\proj'}^* A_{\alpha_k}^* K_0^{-1}e_1 . \end{equation*} Then we can show that in fact \begin{equation}\label{eq:aeq121} \mathrm{span} \Big(\bigcup_{\overline{\alpha}\in \Icc} \tilde{r}_{\overline{\alpha}} \Big) = \CC^m ,\quad \mathrm{span} \Big(\bigcup_{\overline{\alpha}\in \Icc} \tilde{l}_{\overline{\alpha}} \Big) = \CC^m . \end{equation} Indeed, using the fact that for any $(\alpha_1, \ldots, \alpha_{\ell})\in \llbr \gamma_*\rrbr^{\ell}$ and any $B\in \CC^{m\times m}$ \begin{equation*} A_{\alpha_1}\proj' \cdots \proj' A_{\alpha_{k-1}} \proj B e_1 = A_{\alpha_1} \proj' \cdots \proj' A_{\alpha_{k-1}} e_1 \la \proj B \ra \end{equation*} it can be easily seen that \begin{equation*} A_{\overline{\alpha}} e_1 = A_{\alpha_1} (\proj+\proj') A_{\alpha_{2}}\cdots (\proj+\proj') A_{\alpha_{\ell}} e_1 = \tilde{r}_{\overline{\alpha}} + u , \end{equation*} where \begin{equation*} u \in \mathrm{span} \Big(\{\tilde{r}_{\emptyset}\}\cup \bigcup_{k=1}^{\ell-1} \bigcup_{\overline{\beta}\in \llbr \gamma_*\rrbr^{k}} \tilde{r}_{\overline{\beta}} \Big) , \end{equation*} This means that for any $\ell\in \NN$ \begin{equation*} \mathrm{span} \Big(\{e_1 \}\cup \bigcup_{k=1}^{\ell} \bigcup_{\overline{\alpha}\in \llbr \gamma_*\rrbr^{k}} A_{\overline{\alpha}} e_1 \Big) \subset \mathrm{span} \Big(\{ \tilde{r}_{\emptyset}\}\cup \bigcup_{k=1}^{\ell} \bigcup_{\overline{\alpha}\in \llbr \gamma_*\rrbr^{k}} \tilde{r}_{\overline{\alpha}} \Big) \end{equation*} which implies the first equality in \eqref{eq:aeq121}. The second equality can be shown similarly. After all these preparations, we are ready to prove the nilpotency. Fix $\ell>\ell^*$ and $(\gamma_1, \ldots, \gamma_{\ell})\in \llbr \gamma_*\rrbr^{\ell}$, where $\ell^*$ denotes the degree of $\tilde{q}$. Then for any $\overline{\alpha}, \overline{\beta} \in \Icc$ of lengths $k_{\alpha}$ and $k_{\beta}$ correspondingly, by \eqref{eq:anilpotencytrace} we have \begin{align*} &\la \tilde{l}_{\overline{\alpha}}, \proj'A_{\gamma_1}\proj'\cdots A_{\gamma_{\ell}} \proj' \tilde{r}_{\overline{\beta}}\ra \\ &\quad = \big\la \proj A_{\alpha_{k_\alpha}} \proj' A_{\alpha_{k_\alpha-1}} \proj' \cdots A_{\alpha_1} \proj' A_{\gamma_1} \proj' \cdots A_{\gamma_{\ell}} \proj' A_{\beta_1} \proj' A_{\beta_2} \proj' \cdots A_{\beta_{k_\beta}} \big\ra = 0 , \end{align*} which together with \eqref{eq:aeq121} implies that $\proj'A_{\gamma_1}\proj'\cdots A_{\gamma_{\ell}} \proj'=0$. This completes the proof of the lemma. \end{proof} \subsection{A priori bound}\label{sec:atrivialbound} The a priori bound on the generalized resolvent of any nilpotent linearization was formulated in Lemma~\ref{lem:trivBound}. Now we give its proof using the Schur complement formula. \begin{proof}[Proof of Lemma~\ref{lem:trivBound}] First of all, with the definition \begin{equation}\label{eq:aTscDef} \Tb := \Big(I\otimes \cstarunit -z\proj\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma}\otimes x_{\gamma}\Big)^{-1}, \end{equation} we can rewrite the generalized resolvent as \begin{equation}\label{eq:rewrite} \Big( (K_0-zJ)\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} K_\gamma\otimes x_{\gamma} \Big)^{-1} = (K_0^{-1}\otimes \cstarunit) \, \Tb . \end{equation} Taking the quadratic form of this identity at $e_1$, we have \begin{equation*} \la K_0^{-1} e_1 \otimes \cstarunit, \,\Tb\,e_1 \otimes \cstarunit \ra_{\Acc} = \Big\la e_1 \otimes \cstarunit,\Big((K_0-zJ)\otimes \cstarunit - \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\otimes x_{\gamma}\Big)^{-1} e_1 \otimes \cstarunit \Big\ra_{\Acc}. \end{equation*} From the definition of the linearization and \eqref{eq:schur}, the right hand side is just the resolvent $((1-z) \cstarunit - \tilde{q}(\mathbf{ x}))^{-1}$, hence \begin{equation}\label{eq:aSchurC110} \la K_0^{-1} e_1 \otimes \cstarunit, \,\Tb\,e_1 \otimes \cstarunit \ra_{\Acc} = \frac{1}{\cstarunit - \tilde{q}(\mathbf{ x})-z\cstarunit} . \end{equation} After multiplying this identity by $e_1 \otimes \cstarunit$ on the left and $(K_0^{-1}e_1)^* \otimes \cstarunit$ on the right we obtain \begin{equation}\label{eq:aSchurC11} \pproj \Tb \pproj = \proj \otimes \frac{1}{\cstarunit - \tilde{q}(\mathbf{ x})-z\cstarunit} , \end{equation} recalling $\pi= JK_0^{-1}$ and $J= e_1^* e_1$. With $\pproj=\proj\otimes \cstarunit$, and $\pproj'= I\otimes \cstarunit - \pproj$, we now define \begin{equation*} \Sb := \pproj' + \sum_{k=1}^{\infty} \Big(\pproj' \Big(\sum_{\gamma=1}^{\gamma_* } A_{\gamma}\otimes x_{\gamma} \Big) \pproj'\Big)^{k} = \pproj' + \sum_{k=1}^{\infty} \Big(\sum_{\gamma=1}^{\gamma_* } \proj'A_{\gamma}\proj'\otimes x_{\gamma} \Big)^{k} \end{equation*} where the series are convergent, in fact finite, by the joint nilpotency of the family of matrices $\{ \proj' A_{\gamma} \proj', 1\leq \gamma \leq \gamma_* \}$ (see Lemma~\ref{lem:anilpotency}). In particular, there exists a $k^*\in \NN$ such that \begin{equation}\label{eq:Sbound} \|\Sb\|\le C \big(1+ \max_\gamma \| x_\gamma\|^{k^*}_{\Acc}\big) <\infty. \end{equation} Notice that $\Sb$ is the inverse of $\pproj' \Tb \pproj'$ on the range of $\pproj$, i.e. \begin{equation*} \pproj' \Big( I\otimes \cstarunit -z\pproj - \sum_{\gamma=1}^{\gamma_* } A_{\gamma}\otimes x_{\gamma} \Big)\pproj' \Sb = \Sb \pproj' \Big(I\otimes \cstarunit -z\pproj - \sum_{\gamma=1}^{\gamma_* } A_{\gamma}\otimes x_{\gamma} \Big)\pproj' = \pproj' . \end{equation*} By the generalized Schur complement formula for $\Tb=(\pproj+\pproj')\Tb(\pproj+\pproj')$ we have \begin{align} \pproj \Tb \pproj' &= -\Big(\proj\otimes \frac{1}{\cstarunit - \tilde{q}(\mathbf{ x})-z\cstarunit}\Big) \Big(I\otimes \cstarunit -z\pproj - \sum_{\gamma=1}^{\gamma_*} A_{\gamma}\otimes x_{\gamma}\Big)\pproj' \Sb \nonumber \\ \label{eq:aSchurC12} & = -\Big(\proj\otimes\frac{1}{\cstarunit - \tilde{q}(\mathbf{ x})-z\cstarunit}\Big) \Big(\sum_{\gamma=1}^{\gamma_*} \proj A_{\gamma}\proj'\otimes x_{\gamma}\Big) \Sb ,\\ \pproj' \Tb \pproj & = -\Sb \Big(\sum_{\gamma=1}^{\gamma_*} \proj' A_{\gamma}\proj\otimes x_{\gamma} \Big)\Big(\proj \otimes \frac{1}{\cstarunit - \tilde{q}(\mathbf{ x})-z\cstarunit} \Big) \nonumber ,\\\label{eq:aSchurC22} \pproj' \Tb \pproj' &= \Sb+\Sb \Big(\sum_{\gamma=1}^{\gamma_*} \proj' A_{\gamma}\proj\otimes x_{\gamma}\Big)\Big(\proj \otimes \frac{1}{\cstarunit - \tilde{q}(\mathbf{ x})-z\cstarunit}\Big) \Big(\sum_{\gamma=1}^{\gamma_*} \proj A_{\gamma}\proj'\otimes x_{\gamma}\Big) \Sb . \end{align} Since $\tilde{q}(\mathbf{ x})$ is self-adjoint, we have a bound on the inverse of $\cstarunit - \tilde{q}(\mathbf{ x})-z\cstarunit$ \begin{equation}\label{eq:aTBsc} \Big\|\frac{1}{\cstarunit - \tilde{q}(\mathbf{ x})-z\cstarunit} \Big\|_{\Acc} \leq \frac{1}{\eta} . \end{equation} Using now \eqref{eq:aTBsc}, the boundedness of $S$ and formulas \eqref{eq:aSchurC11}-\eqref{eq:aSchurC22} it can be seen that there exists $C>0$ such that $\|T\|\leq C(1+\eta^{-1})$. The bound \eqref{eq:trivResBound} now follows from \eqref{eq:rewrite} and \eqref{eq:Sbound}. \end{proof} \section{Solution to the polynomial Dyson equation} \label{sec:MDE} Before starting the proof of Lemma~\ref{lem:MDEexistence} we observe that the linear map $\SuOp$ can be written using only self-adjoint matrices. Indeed, if we define (compare with \eqref{eq:aNewK}) \begin{equation} \label{eq:symmetrized} K_{\alpha_*+\beta} := \sqrt{2}\Re L_{\beta} ,\quad K_{\alpha_*+\beta_*+\beta} := -\sqrt{2}\Im L_{\beta} ,\quad 1\leq \beta\leq \beta_* , \end{equation} then for any $R\in \CC^{m\times m}$ \begin{equation*} \SuOp[R]=\sum_{\alpha=1}^{\alpha_*+2\beta_*} K_\alpha R K_{\alpha} . \end{equation*} Therefore, the Dyson equation for linearization \eqref{eq:MDE} can also be written as \begin{equation}\label{eq:MDEothersym} -\frac{1}{M} = zJ-K_0 + \sum_{\alpha=1}^{\gamma_*} K_{\alpha} M K_{\alpha} . \end{equation} where we introduced $\gamma_*:= \alpha_*+2\beta_*$ for brevity. In the sequel we will use the following notations for comparison relations. Let $\mathscr{D}\subset \CC$ and let $(\phi_w^{(N)})_{N\in \NN}$ and $(\psi_w^{(N)})_{N\in \NN}$, $w\in \mathscr{D}$, be two sequences of complex-valued functions on $\mathscr{D}$. We will write $\phi_w^{(N)} \lesssim \psi_w^{(N)}$ (or simply $\phi \lesssim \psi$) if there exists $C>0$ depending only on the polynomial $p$ such that $\phi_w^{(N)} \leq C \psi_w^{(N)}$ uniformly for $w\in \mathscr{D}$ and $N\in \NN$. If $\phi \lesssim$ and $\psi \lesssim \phi$ then we will write $\phi \sim \psi$. Also, from now on we will always denote the real and imaginary parts of the spectral parameter $z$ by $E$ and $\eta$ correspondingly, i.e., $z:=E+\imunit \eta$. \begin{proof}[Proof of Lemma~\ref{lem:MDEexistence}] \emph{Existence.} Let $\{\semic_1, \ldots, \semic_{\gamma_*}\}$ be a family of free semicircular variables in a $C^*$--probability space $(\Scc,\tau)$ (see Appendix~\ref{sec:freeIntro}). Define \begin{equation*} \Lb_{\mathrm{sc}} := K_0 \otimes \freeunit - \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\otimes \semic_{\gamma} , \end{equation*} and for $z\in \CC_{+}$ define a function $ \GSol_{\mathrm{sc}}(z) : \CC_+ \rightarrow \CC^{m\times m}$ by \begin{equation}\label{eq:MscDef} \GSol_{\mathrm{sc}}(z) := (\id\otimes \tau)\big(\Lb_{\mathrm{sc}}-zJ\otimes \freeunit \big)^{-1} . \end{equation} The subscript in $\Lb_{\mathrm{sc}}$ and $\GSol_{\mathrm{sc}}$ refers to the semicircular elements. We will show that the function $\GSol_{\mathrm{sc}}$ is well-defined on $\CC_{+}$ and satisfies \emph{(i)-(iv)} of Lemma~\ref{lem:MDEexistence}. We now introduce some notation that will be used throughout the proof. Let $\proj$ and $ \proj'$ denote as before projections on $\CC^{m\times m}$ given by $\proj = JK_0^{-1}$, $\proj'= I-\proj$, and let $\pproj$ and $\pproj'$ be projections on $\Cmxm$ defined by \begin{equation*} \pproj := \proj\otimes \freeunit ,\quad \pproj' := I\otimes \freeunit - \pproj = \proj'\otimes \freeunit . \end{equation*} Define also the matrices $A_{\gamma}:= K_{\gamma} K_0^{-1}$, $\gamma\in\llbr \gamma_* \rrbr$. Notice that the nilpotency of $\Lb$ implies that $\{ \proj' A_\gamma\proj'\}_{\gamma=1}^{\gamma^*}$ is a nilpotent family. \setcounter{cl}{0} \emph{Step 1:} We first show that $\GSol_{\mathrm{sc}}$ is well-defined and properties $(i)$-$(iii)$ hold. To see this, we apply Lemma~\ref{lem:trivBound} with $\Acc = \Scc$ to $(\Lb_{\mathrm{sc}}-zJ\otimes \freeunit )^{-1}$. Then from \eqref{eq:trivResBound} (assuming only self-adjoint variables) we obtain that for any $z\in \CC_+$ \begin{equation}\label{scbound} \|(\Lb_{\mathrm{sc}}-zJ\otimes \freeunit)^{-1}\| \lesssim 1+\frac{1}{\eta} , \end{equation} Moreover, simple computation shows that \begin{equation*} \Im \GSol_{\mathrm{sc}}(z)= \eta \, (\id\otimes \tau) \left((\Lb_{\mathrm{sc}}-\overline{z}J\otimes \freeunit)^{-1} (J\otimes \freeunit) ( \Lb_{\mathrm{sc}} - z J \otimes \freeunit)^{-1}\right) , \end{equation*} which yields that $\Im \GSol_{\mathrm{sc}} (z)$ is positive semi-definite. \emph{Step 2:} Now we show that $\GSol_{\mathrm{sc}}$ satisfies the DEL \eqref{eq:MDE}. We will need the following technical lemma whose proof is postponed: \begin{lem}\label{lem:IPF0} Let $z\in \CC_+$. Then $\GSol_{\mathrm{sc}}(z)$ satisfies the DEL \eqref{eq:MDE} if and only if \begin{equation}\label{eq:IPF3} (\id \otimes \tau ) \bigg( \Big(I\otimes \freeunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma}^z\otimes s_{\gamma}\Big)^{-1} \sum_{\gamma=1}^{\gamma_*} A_{\gamma}^z\otimes s_{\gamma}\bigg) = \GSol_{\mathrm{sc}}^{\mathrm{ns},z} \sum_{\gamma=1}^{\gamma_*} A_{\gamma}^z \GSol_{\mathrm{sc}}^{\mathrm{ns},z} A_{\gamma}^z , \end{equation} where \begin{equation}\label{eq:nonSymM} \GSol_{\mathrm{sc}}^{\mathrm{ns},z} := D_z \, K_0 \GSol_{\mathrm{sc}}(z) \, D_z ,\quad A_{\gamma}^{z} := D_z^{-1} \,A_{\gamma} D_z^{-1} , \end{equation} $D_z$ and its inverse $D_z^{-1}$ are given by \begin{equation*} D_z := (\sqrt{1-z} \proj + \proj') , \quad D_z^{-1} = \bigg(\frac{1}{\sqrt{1-z}} \proj + \proj'\bigg) , \end{equation*} and we choose the principal branch of the square root. \end{lem} The advantage of working with \eqref{eq:IPF3} is that taking $|z|$ big enough ensures that \begin{equation} \label{eq:expansionT} \Tb_{\mathrm{sc}}^{z} := \Big(I\otimes \freeunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma}^z\otimes s_{\gamma}\Big)^{-1} \end{equation} can be expanded into a convergent geometric series, and thus Lemma~\ref{lem:aIPF} can be applied to the LHS of \eqref{eq:IPF3} with $B_\gamma= A_\gamma^z$. We now check the condition \eqref{eq:aIPFcond} of this lemma. By the nilpotency of the family $\{ \proj' A_\gamma\proj'\}_{\gamma=1}^{\gamma^*}$ it follows that if we take the smallest $k^* \in \NN$ such that $\proj' A_{\gamma_1} \proj' \ldots \proj' A_{\gamma_{k^*}} \proj'=0$ for all $(\gamma_1, \ldots, \gamma_{k^*})\in\llbr \gamma_*\rrbr^{k^*}$, then \begin{equation*} \|(A_{\gamma}^{z})^{k^*}\| = \Big\| \Big(\frac{1}{1-z} \proj A_{\gamma} \proj + \frac{1}{\sqrt{1-z}}(\proj A_{\gamma} \proj' + \proj' A_{\gamma} \proj) + \proj' A_{\gamma} \proj'\Big)^{k^*}\Big\| \leq \frac{2k^* \|A_{\gamma}\|^{k^*}}{\sqrt{|1-z|}} + \OO{\frac{1}{|1-z|}} . \end{equation*} Notice that the right hand side goes to zero as $|z|\to\infty$. Therefore, by choosing $C\ge 4$ big enough, depending on $k^*$ and on the matrices $\{ A_\gamma\}$, we have that \begin{equation*} \max_{1\leq \gamma \leq \gamma_*} \| (A_{\gamma}^z)^{k^*} \| \|\semic_{\gamma}^{ k^*} \| \leq \Big( 1+ \max_{1\leq \gamma \leq \gamma_*} 2\| A_{\gamma}\| \, \|\semic_{\gamma}\| \Big)^{- k^* } \gamma_*^{- 4 k^*} . \end{equation*} uniformly for any $|z|\ge C$. Moreover, $\|A_\gamma^z\|\le 2\|A_\gamma\|$. This implies that for any $\gamma\in \llbr \gamma_* \rrbr$, $\ell\ge 1$ and $0\leq k\leq k^*-1$ \begin{equation*} \| (A_{\gamma}^z)^{\ell k^*+k} \| \|\semic_{\gamma}^{ \ell k^*+k} \| \,\leq\, \| (A_{\gamma}^z)^{k^*} \|^{\ell} \|\semic_{\gamma}^{k^*} \|^{\ell} \| (A_{\gamma}^z) \|^k \|\semic_{\gamma}\|^k \, \leq\, \gamma_*^{-2(\ell k^* +k)} , \end{equation*} so that the condition \eqref{eq:aIPFcond} holds. From this, using Lemma~\ref{lem:aIPF} we conclude that \begin{equation}\label{eq:IPFmain} (\id \otimes \tau ) \Big( \Tb_{\mathrm{sc}}^{z} \, \sum_{\gamma=1}^{\gamma_*} A_{\gamma}^z\otimes s_{\gamma}\Big) = (\id \otimes \tau ) \big(\Tb_{\mathrm{sc}}^{z} \big) \sum_{\gamma=1}^{\gamma_*} A_{\gamma}^z \, (\id \otimes \tau ) \big( \Tb_{\mathrm{sc}}^{z}\big) A_{\gamma}^z , \end{equation} The last ingredient we need in order to show that $\GSol_{\mathrm{sc}}$ satisfies \eqref{eq:IPF3}, hence the DEL \eqref{eq:MDE}, is the following identity \begin{equation}\label{eq:rescaledsolution} (\id \otimes \tau ) \big(\Tb_{\mathrm{sc}}^z\big) = \GSol_{\mathrm{sc}}^{\mathrm{ns},z} . \end{equation} Then from the definition of $A_{\gamma}^{z}$, and the fact that $D_{z}^{-1}=\frac{1}{\sqrt{1-z}}\proj + \proj'$ we have \begin{align*} (\id \otimes \tau ) \big(\Tb_{\mathrm{sc}}^{z}\big) &= D_z \, (\id \otimes \tau ) \Big((I-z\proj)\otimes \freeunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma}\otimes s_{\gamma}\Big)^{-1} D_z \\ & = D_z \, K_0 \, (\id \otimes \tau ) \left( \Lb_{ \mathrm{sc} }- zJ\otimes \freeunit\right)^{-1} D_z \end{align*} which is exactly the definition of $\GSol_{\mathrm{sc}}^{\mathrm{ns},z}$ thus \eqref{eq:rescaledsolution} is proven. Now we can finish the proof of the existence of the solution to the DEL. From the \eqref{eq:IPFmain} and \eqref{eq:rescaledsolution} we see that \eqref{eq:IPF3} holds for all $z\in \CC_{+}$ such that $|z|>C$. By Lemma~\ref{lem:IPF0} this shows that $\GSol_{\mathrm{sc}}(z)$ solves the DEL \eqref{eq:MDE} for any fixed $z\in \CC_+$ with $|z|>C$. Moreover, the matrix-valued function $\GSol_{\mathrm{sc}}(z)$ is analytic on $\CC_{+}$, therefore the function $I - \GSol_{\mathrm{sc}}(z)\, ( K_0-zJ) + \GSol_{\mathrm{sc}}(z)\, \sum_{\gamma=1}^{\gamma_*} K_{\gamma} \GSol_{\mathrm{sc}}(z)\, K_{\gamma}$ is also analytic on $\CC_{+}$. Since \begin{equation*} I - \GSol_{\mathrm{sc}}(z)\, ( K_0-zJ) + \GSol_{\mathrm{sc}}(z)\, \sum_{\gamma=1}^{\gamma_*} K_{\gamma} \,\GSol_{\mathrm{sc}}(z)\, K_{\gamma} = 0 \end{equation*} for all $z\in \CC_{+}$ such that $|z|>C$, we conclude that \eqref{eq:MDEothersym} holds for all $z\in \CC_{+}$. \emph{Uniqueness.} Suppose that $M_1, M_2 \, : \CC_+ \rightarrow \CC_+$ are two analytic solutions of \eqref{eq:MDEothersym} satisfying $\Im M_1(z) \geq 0$ and $\Im M_2(z) \geq 0$. It is easy to see that both $M_{1,2}(z)$ are solutions of \eqref{eq:MDEothersym} on $\CC_{+}$ if and only if for all $z\in \CC_{+}$ functions $M^{\mathrm{ns}}_{1,2}(z):= K_0 M_{1,2} (z)$ satisfy \begin{equation}\label{eq:MDEotherR1} M^{\mathrm{ns}}_{1,2}(z) = \frac{1}{1-z}\proj + \proj' + M^{\mathrm{ns}}_{1,2}(z) \sum_{\gamma=1}^{\gamma_*} A_{\gamma} M^{\mathrm{ns}}_{1,2}(z) A_{\gamma} \Big(\frac{1}{1-z}\proj + \proj'\Big) . \end{equation} If we recursively replace $M^{\mathrm{ns}}_{1,2}(z)$ in the RHS by the expression given in the RHS of \eqref{eq:MDEotherR1}, we obtain a series which is convergent for large $z$ due to nilpotency of the linearization. Indeed, if we assume for simplicity that $\gamma_*=1$, then $M^{\mathrm{ns}}_{1,2}(z)$ can be rewritten as \begin{equation}\label{eq:Mexpan} M^{\mathrm{ns}}_{1,2}(z) = \sum_{\ell=0}^{\infty} C_{\ell} \Big(\Big(\frac{1}{1-z}\proj + \proj'\Big)A_{1}\Big)^{2\ell} \Big(\frac{1}{1-z}\proj + \proj'\Big) \end{equation} where $C_{\ell}$ denotes the $\ell$th Catalan number. Since $C_{\ell} = \frac{1}{\ell +1}\binom{2\ell}{\ell} \le 4^{\ell}$, we conclude that the RHS of \eqref{eq:Mexpan} contains $\OO{16^{\ell}}$ products of type \begin{equation}\label{eq:MexpanProd} \sigma_{1} A_1 \cdots \sigma_{2\ell} A_1 \sigma_{2\ell +1} \end{equation} with $\sigma_{i}\in \{\frac{1}{1-z} \proj, \proj'\}$. Now we collect all terms of type \eqref{eq:MexpanProd} that behave asymptotically like $(1-z)^{-i}$ for some $i\in \NN$ as $|z|\rightarrow \infty$. By the nilpotency of $\{ \proj' A_{\gamma} \proj', 1\leq \gamma\leq \gamma_*\}$ there exists $k\in \NN$ such that $\proj' A_{\gamma_1} \proj'\cdots \proj' A_{\gamma_k} \proj'=0$. Therefore, the maximum $\ell$ for which there exists a product \eqref{eq:MexpanProd} that behaves asymptotically as $(1-z)^{-i}$ is less than $k(i+1)/2$. This implies that if $|1-z|>2 (4\| A_1 \|)^{k}$, then the RHS in \eqref{eq:Mexpan} converges and thus $M^{\mathrm{ns}}_{1}(z)= M^{\mathrm{ns}}_{2}(z)$ on $\CC_+$ by analiticity. If $\gamma_* > 1$, then it can be shown similarly that on the set $|z|> 2(4 \gamma_* \max_{1\leq \gamma \leq \gamma_*}\|A_{\gamma}\|)^{k}$ functions $M^{\mathrm{ns}}_{1}(z)$ and $M^{\mathrm{ns}}_{2}(z)$ coincide, which again implies $M^{\mathrm{ns}}_{1}(z)= M^{\mathrm{ns}}_{2}(z)$ on $\CC_+$. This concludes the proof that $\GSol_{\mathrm{sc}}(z)$ defined in \eqref{eq:MscDef} is the unique solution to the DEL \eqref{eq:MDE}, i.e. $M(z)=\GSol_{\mathrm{sc}}(z)$. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:IPF0}] We start with a trivial identity \begin{equation*} 0 = I\otimes \freeunit -\left(\Lb_{\mathrm{sc}}-zJ\otimes \freeunit \right)^{-1} \Big((K_0-zJ)\otimes \freeunit - \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\otimes s_{\gamma}\Big) . \end{equation*} Applying $(\id \otimes \tau)$ to the above equality leads to \begin{equation*} 0 = I -\GSol_{\mathrm{sc}}(z) (K_0-zJ) + (\id \otimes \tau) \Big( \left( \Lb_{\mathrm{sc}}-zJ\otimes \freeunit \right)^{-1} \, \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\otimes s_{\gamma} \Big) . \end{equation*} Therefore, after rewriting equation \eqref{eq:MDEothersym} as \begin{equation*} 0 = I + M (zJ - K_0) + M \sum_{\gamma=1}^{\gamma_*} K_{\gamma} M K_{\gamma} \end{equation*} we see that $\GSol_{\mathrm{sc}}(z)$ solves the DEL \eqref{eq:MDEothersym} for some $z\in \CC_{+}$ if and only if \begin{equation}\label{eq:IPF1} (\id \otimes \tau)\Big(\left(\Lb_{\mathrm{sc}}-zJ\otimes \freeunit \right)^{-1}\, \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\otimes s_{\gamma}\Big) = \GSol_{\mathrm{sc}}(z) \sum_{\gamma=1}^{\gamma_*} K_{\gamma} \GSol_{\mathrm{sc}}(z) K_{\gamma} . \end{equation} Now, by multiplying equation \eqref{eq:IPF1} with $K_0$ on the left and $K_0^{-1}$ on the right, we see that \eqref{eq:IPF1} can be rewritten as \begin{equation}\label{eq:IPF2} (\id \otimes \tau ) \Big( \Big((I-z\proj)\otimes \freeunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma}\otimes s_{\gamma}\Big)^{-1} \, \sum_{\gamma=1}^{\gamma_*} A_{\gamma}\otimes s_{\gamma}\Big) = K_0\GSol_{\mathrm{sc}}(z) \sum_{\gamma=1}^{\gamma_*} A_{\gamma} K_0 \GSol_{\mathrm{sc}} (z) A_{\gamma} . \end{equation} Finally, if we multiply equation \eqref{eq:IPF2} with $D_z$ on the left and $D_z^{-1}$ on the right we see that formula \eqref{eq:IPF2} is equivalent to \begin{equation*} (\id \otimes \tau ) \Big( \Tb_{\mathrm{sc}}^{z} \, \sum_{\gamma=1}^{\gamma_*} A_{\gamma}^z\otimes s_{\gamma}\Big) = \GSol_{\mathrm{sc}}^{\mathrm{ns},z} \sum_{\gamma=1}^{\gamma_*} A_{\gamma}^z \GSol_{\mathrm{sc}}^{\mathrm{ns},z} A_{\gamma}^z . \end{equation*} \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:StiltRep}] $\GSol(z)$ is a matrix-valued Herglotz function, therefore, from \cite[(1.1)-(1.3)]{GeszTsek} it has the following representation \begin{equation}\label{eq:StieltRep} \GSol(z) = B_1 z + B_0 + \int_{\RR} \Big( \frac{1}{x-z} - \frac{x}{1+x^2}\Big) dV(x) , \end{equation} where $B_1, B_0 \in \CC^{m\times m}$, $B_1=\lim_{\eta\uparrow \infty}(\frac{1}{\imunit \eta} \GSol(\imunit \eta))$ and $dV(x)$ is a matrix-valued measure satisfying \begin{equation*} \int_{\RR} \frac{ \la \cb , dV(x) \,\cb \ra}{1+x^2} < \infty \quad \mbox{and}\quad \int_{I} dV(x) \geq 0 \end{equation*} for any $\cb \in \CC^{m}$ and Borel $I \subset \RR$. From Lemma~\ref{lem:MDEexistence} we know that $\lim_{z\rightarrow \infty} \| \GSol(z)\| <\infty$, which implies that $B_1=0$. From the properties of scalar-valued Herglotz functions (formula S1.1.9 in \cite{KacKrei}) and polarization (as in the proof of Lemma~5.3 in \cite{GeszTsek}) we know that \begin{equation}\label{eq:VofR} \int_{\RR} dV(x) = \lim_{\eta\uparrow \infty} \eta \Im \GSol(\imunit \eta) . \end{equation} By definition \eqref{eq:MscDef} and the conclusion of the proof of Lemma~\ref{lem:MDEexistence} \begin{equation*} \GSol(z)=\GSol_{\mathrm{sc}}(z) = (\id\otimes \tau)\big( \Lb_{\mathrm{sc}}-zJ\otimes \freeunit\big)^{-1} , \end{equation*} therefore \begin{align*} \Im \GSol(z) &= (\id\otimes \tau) \big( \Im \big( \Lb_{\mathrm{sc}}-zJ\otimes \freeunit\big)^{-1} \big) \\ &= (\id\otimes \tau) \Big( \eta \big( \Lb_{\mathrm{sc}}-zJ\otimes \freeunit\big)^{-1} \, (J\otimes \freeunit) \, \big( \Lb_{\mathrm{sc}}-\overline{z}J\otimes \freeunit\big)^{-1} \Big). \end{align*} If we define (similarly to \eqref{eq:aTscDef}) \begin{equation*} \Tb_{\mathrm{sc}}^{z} := \Big(I\otimes \freeunit -z\proj\otimes \freeunit - \sum_{\gamma=1}^{\gamma_*} A_{\gamma}\otimes \semic_{\gamma}\Big)^{-1} , \end{equation*} then \begin{equation}\label{eq:aImM} \Im \GSol(z) = \eta \, (\id \otimes \tau) \big( (K_0^{-1} \otimes \freeunit)\, \Tb_{\mathrm{sc}}^{z} \, \pproj \, \Tb_{\mathrm{sc}}^{\overline{z}} \big) , \end{equation} where $\pproj= JK_0^{-1} \otimes \freeunit$. We will now use several results from Section~\ref{sec:atrivialbound} by specializing the proof of Lemma~\ref{lem:trivBound} for $\Acc = \Scc$ and $x_{\alpha}=s_{\alpha}$. By \eqref{eq:aSchurC110} we have \begin{equation*} \la K_0^{-1} e_{1} \otimes \freeunit, \Tb_{\mathrm{sc}}^{z} \, e_1\otimes \freeunit \ra_{\Scc} = \frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-z\freeunit} , \end{equation*} with $\tilde{q}\in \CC \la x_1, \ldots, x_{\gamma_*} \ra$ being a self-adjoint polynomial. This implies that \begin{align} \la e_1, \Im \GSol(z) \,e_1 \ra &= \eta \, \tau \big( \la K_0^{-1} e_1 \otimes \freeunit , \Tb_{\mathrm{sc}}^{z} \, e_1\otimes \freeunit \ra_{\Scc} \la K_0^{-1} e_1 \otimes \freeunit , \Tb_{\mathrm{sc}}^{\overline{z}} \, e_1\otimes \freeunit \ra_{\Scc}\big) \nonumber \\ \label{eq:ImM11} &= \eta \, \tau \Big( \frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-z\freeunit} \frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-\overline{z}\freeunit} \Big) , \end{align} where we used that $J=e_1 e_1^{t}$. In particular, this shows that the imaginary part of the upper left entry of $\GSol(z)$ is independent of the linearization. Now write $ \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}} $ from \eqref{eq:aImM} as \begin{equation}\label{eq:TpiT} \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}} = \pproj \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}}\pproj +\pproj \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}}\pproj' + \pproj' \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}}\pproj + \pproj' \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}} \pproj' . \end{equation} From \eqref{eq:aSchurC12}-\eqref{eq:aSchurC22} we have that each entry of $ \pproj \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}}\pproj$, $\pproj \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}}\pproj'$, $\pproj' \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}}\pproj$ and $\pproj' \Tb_{\mathrm{sc}}^{z} \pproj \Tb_{\mathrm{sc}}^{\overline{z}} \pproj'$ is of the form \begin{equation}\label{eq:ImMij} a(\mathbf{\semic}) \frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-z\freeunit} \frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-\overline{z}\freeunit} b(\mathbf{\semic}) , \end{equation} where $a(\mathbf{\semic})$ and $b(\mathbf{\semic})$ are some polynomials in $\mathbf{\semic}$. Now, using, e.g., (G.5), (G.7) and (5.4.28) from \cite{AndeGuioZeit}, $\|\semic_{\alpha}\|=2$ and submultiplicativity of the norm (see Appendix~\ref{sec:freeIntro}) \begin{equation}\label{eq:ImM11bound} \left|\tau \Big(a(\mathbf{\semic}) \frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-z\freeunit} \frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-\overline{z}\freeunit} b(\mathbf{\semic})\Big)\right| \leq \| a(\mathbf{\semic}) \| \| b(\mathbf{\semic}) \| \tau\Big(\frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-z\freeunit} \frac{1}{\freeunit - \tilde{q}(\mathbf{\semic})-\overline{z}\freeunit} \Big) . \end{equation} Therefore, \eqref{eq:aImM}, \eqref{eq:TpiT}, \eqref{eq:aSchurC12}-\eqref{eq:aSchurC22} and \eqref{eq:ImM11bound} imply that $\lim_{\eta\uparrow \infty} \eta \, \|\Im \GSol (\imunit \eta)\| < \infty$, and thus by \eqref{eq:VofR} $\| \int_{\RR} dV(x) \| < \infty$ and both integrands in \eqref{eq:StieltRep} have finite integrals. If we denote now \begin{equation*} \GSol^{\infty} := B_0 - \int_{\RR} \frac{x\, d V(x)}{1+x^2} \in \CC^{m\times m} , \end{equation*} then we obtain \eqref{eq:matSt}. As we now know that $\GSol(z)$ has a Stieltjes transform representation \eqref{eq:matSt}, part $(ii)$ of Lemma~\ref{lem:StiltRep} follows from, e.g., \cite[Lemma~5.4]{GeszTsek}. Since $\Im \GSol\geq 0$, the bound \eqref{eq:ImM11bound} together with \eqref{eq:ImM11} and the representation of the entries of $\Im \GSol$ given in \eqref{eq:ImMij} implies that there exists $C\in \RR$ such that \begin{equation*} \Tr \Im \GSol \leq C \la e_1, \Im \GSol\,e_1 \ra . \end{equation*} Therefore, if $\lim_{\eta \rightarrow 0}\la e_1, \Im \GSol(E+\imunit \eta) e_1 \ra = 0$ then $\lim_{\eta\rightarrow 0} \Tr\Im \GSol(E+\imunit \eta) = 0$, and similarly if $\lim_{\eta\rightarrow 0} \Tr\Im \GSol(E+\imunit \eta) = \infty$ then $\lim_{\eta \rightarrow 0}\la e_1, \Im \GSol(E+\imunit \eta) e_1 \ra = \infty$. We conclude that $\mathrm{supp} ( V_{11}) = \mathrm{supp} (\Tr V)$. \end{proof} \begin{lem}[Stability of the solution of the DEL] There exists $\epsilon>0$ such that \begin{itemize} \setlength\itemsep{0em} \item[(i)] for any $R \in \CC^{m\times m}$ , $\|R\| < \epsilon$, the matrix equation \begin{equation*} -\frac{1}{\GSol} = zJ-K_0 + \SuOp [ \GSol] + R \end{equation*} has a solution, which we denote by $\GSol(R)$; \item[(ii)] for any $R_1, R_2 \in \CC^{m\times m}$, $\|R_1\| < \epsilon$, $\|R_2\| < \epsilon$, we have \begin{equation}\label{eq:SolStab} \| M (R_1)-M (R_2)\| \leq C \|R_1 - R_2\| . \end{equation} \end{itemize} \end{lem} \begin{proof} This follows easily from (\textbf{M1}) and (\textbf{M2}) (see e.g. proof of the Corollary 3.8 in \cite{AltErdoKrugNemi_Kronecker}) \end{proof} \section{Proof of the local law} \label{sec:locallaw} In order to establish the local law for the polynomials we will rely heavily on the linearization technique described in the previous sections. More precisely, given a self-adjoint polynomial $p=p(\Xb, \Yb, \Yb^{*})$ in the variables $\Xb$, $\Yb$ and $\Yb^{*}$, we consider one of its nilpotent linearizations $\Lb$ as defined in Section~\ref{sec:linDyson}. Its generalized resolvent will give the necessary information on the resolvent of $p$ via \eqref{eq:schur}. So from now on our main object of interest will be the linearized random matrix $\Hb$ defined by \begin{equation}\label{eq:Hdef} \Hb = K_0\otimes I_N - \sum_{\alpha=1}^{\alpha_*} K_\alpha\otimes X_\alpha - \sum_{\beta=1}^{\beta_*}\big( L_{\beta}\otimes Y_\beta+L^*_{\beta}\otimes Y_\beta^{*} \big) . \end{equation} This matrix plays the role of $\Lb$ in Section~\ref{sec:linDyson}, but we use a different letter to stress that we are in the random matrix setup. We denote by $I_N$ the unit element of $\mathscr{A}=\CC^{N\times N}$. We remark that random matrices of the form \eqref{eq:Hdef}, in particular their \emph{resolvents}, have been extensively studied in \cite{AltErdoKrugNemi_Kronecker} where they were called {\it Kronecker matrices}. We will denote the \emph{generalized resolvent} of $\Hb$ by $\GRes(z):=(\Hb-zJ\otimes I_N)^{-1}$. By $\GRes_{kl} \in \CC^{N\times N}$ and $\Gm_{ij} \in \CC^{m\times m}$ we will denote the coefficient of $\GRes$ in the standard bases of $\CC^{m \times m}$ and $\CC^{N\times N}$ correspondingly, i.e., \begin{equation*} \GRes = \sum_{k,l=1}^{m} E_{kl}\otimes \GRes_{kl} = \sum_{i,j=1}^{N} \Gm_{ij} \otimes E_{ij} \end{equation*} with $E_{ij}=E_{ij}^{(n)}:= (\delta_{ki}\delta_{jl})_{k,l=1}^{n}\in \CC^{n}$ for corresponding $n\in \NN$. More generally, for any $\Rb \in \CC^{m\times m}\otimes \CC^{N\times N}$ we will denote its coefficients in the standard basis of $\CC^{N\times N}$ by $P_{ij} (\Rb), 1\leq i, j \leq N$, so that \begin{equation*} \Rb = \sum_{i,j=1}^{N} P_{ij}(\Rb)\otimes E_{ij} . \end{equation*} In particular, we have $P_{ij}(\GRes) = \Gm_{ij}$. Here is our main technical result. \begin{thm}[Local law for the linearization] \label{thm:main} Let $p\in \CC \la \xb, \yb, \yb^* \ra$ be a self-adjoint polynomial with $p(0,0,0)=1$ and let $\Lb$ be a nilpotent linearization of $p$ be defined as in \eqref{eq:linearization}. Let $\GSol(z)$ be a solution of the corresponding \emph{DEL} \eqref{eq:MDE} constructed as in Lemma~\ref{lem:MDEexistence}. Let $\Hb$ be defined as in \eqref{eq:Hdef}. Suppose that the families of random matrices $\Xb, \Yb$ satisfy conditions \textup{(\textbf{H1})-(\textbf{H4})} and that $\GSol(z)$ satisfies \textup{(\textbf{M1})-(\textbf{M2})} for some fixed $\epsB>0$. Then the local law holds for $H$ in the $\epsB$-bulk up to the optimal scale, i.e., for any $\gamma>0$ we have \begin{equation} \label{eq:locallaw} \max_{i,j\in\llbr N \rrbr} \|\Gm_{ij}(z)-\GSol(z)\delta_{ij}\| \prec \sqrt{\frac{1}{N \Im z}} ,\qquad \Big\|\frac{1}{N} \sum_{i=1}^{N}\Gm_{ii}(z)-\GSol(z)\Big\| \prec \frac{1}{N \Im z} \end{equation} uniformly for $z\in D_{\kappa, \gamma}$ with $D_{\kappa, \gamma}:= \{z\in \CC\; : \;\Re z\in B_{\epsB}, \; N^{-1+\gamma} \leq \Im z \leq 1\}$. \end{thm} \begin{proof}[Proof of Theorem~\ref{thm:mainPoly}] It follows immediately from Theorem~\ref{thm:main} and the Schur complement formula \eqref{eq:schur}. \end{proof} The rest of this section is devoted to the proof of Theorem~\ref{thm:main}. Throughout this section we will use regularizations of $\GRes$ and $\GSol$. For $z,\omega \in \CC_{+}$ define \begin{equation*} \Ress{\omega}(z) = \big( \Hb - zJ\otimes I_N - \omega I\otimes I_N \big)^{-1} \end{equation*} and let $\Soll{\omega}(z)$ be the solution of the regularized DEL \begin{equation}\label{eq:MregA} -\frac{1}{\Soll{\omega}(z)} = zJ + \omega I - K_0 + \SuOp[\Soll{\omega}(z)] \end{equation} that is analytic in $z$ and $\omega$ and has positive definite imaginary part. Note that the existence of such solution analytic in $\omega$ has been proven in \cite[Lemma~2.2]{AltErdoKrugNemi_Kronecker} with $\omega$ playing the role of the spectral parameter, while $zJ-K_0$ is the expectation matrix with nonnegative semi-definite imaginary part. To show that $\Soll{\omega}(z)$ depends analytically on $z$, differentiate \eqref{eq:MregA} with respect to $z$ so that \begin{equation}\label{Mderiv} -\frac{1}{\Soll{\omega}(z)} \partial_z \Soll{\omega}(z) \frac{1}{\Soll{\omega}(z)} = J + \SuOp[\partial_z \Soll{\omega}(z)] . \end{equation} After rearranging the terms, the above equation can be rewritten as \begin{equation*} \GStOp_{\omega}[\partial_z \Soll{\omega}(z)] = -\Soll{\omega}(z)J\Soll{\omega}(z) , \end{equation*} where \begin{equation*} \GStOp_{\omega}\,:\, \CC^{m\times m} \rightarrow \CC^{m\times m} ,\quad \GStOp_{\omega}(R) := R- \Soll{\omega}(z)\, \Gamma[R] \, \Soll{\omega}(z) . \end{equation*} From \cite[Lemma~3.7]{AltErdoKrugNemi_Kronecker} and the trivial bound $\|\Soll{\omega}(z)\|\leq (\Im \omega)^{-1}$ we have that $\| \GStOp_{\omega}^{-1}\| \leq (\Im \omega)^{-10}$. We conclude that $\|\partial_z \Soll{\omega}(z) \| \leq (\Im \omega)^{-12}$, which yields that $\Soll{\omega}(z)$ is analytic in $z$. The next lemma collects some properties of the regularizations $\Ress{\omega}(z)$ and $\Soll{\omega}(z)$. \begin{lem}\label{lem:RegBnd} There exists $C>0$ such that \begin{itemize} \setlength{\itemsep}{0em} \item[(i)] uniformly on $E\in \RR$, $\eta > 0 $ and $\epsF \geq 0 $ \begin{equation} \label{eq:GregB} \|\Ress{\imunit \epsF}(z)\| \leq C \Big( 1+\frac{1}{\eta } \Big) ; \end{equation} \item[(ii)] if \textup{(\textbf{M1})} holds, then uniformly on $E\in B_{\epsB}$, $\eta \geq 0$ and $\epsF \geq 0 $ \begin{equation} \label{eq:MregB} \| \Soll{\imunit \epsF}(z)\|\leq C ,\quad \|(\Soll{\imunit \epsF}(z))^{-1}\| \leq C (1+|z| + \epsF) ; \end{equation} \item[(iii)] if additionally \textup{(\textbf{M2})} holds, then uniformly on $E\in B_{\epsB}$, $0\leq \eta \leq 1$ and $\epsF \geq 0 $ \begin{equation} \label{eq:LregB} \| (\GStOp_{\imunit \epsF}(z))^{-1}\| \leq C . \end{equation} \end{itemize} \end{lem} \begin{proof} Firstly, by specializing Lemma~\ref{lem:trivBound} for $\Acc = \CC^{m\times m}\otimes \CC^{N\times N}$, $\xb=\Xb $, $\yb=\Yb$ and $\Lb= \Hb$, we obtain that there exists $C_1>0$ such that \begin{equation*} \| \GRes (z) \| \leq C_1 \Big( 1+\frac{1}{\eta} \Big) \end{equation*} By the resolvent identity, for any $E\in B_{\epsB}$, $\eta\geq 0$ and $\epsF\geq 0$ we have that \begin{equation*} \Ress{\imunit \epsF}(z) = \GRes(z) + \imunit \epsF\,\Ress{\imunit \epsF}(z)\,\GRes(z) , \end{equation*} therefore from the trivial bound $\|\Ress{\imunit \epsF}(z) \| \leq \epsF^{-1}$ we obtain \begin{equation*} \|\Ress{\imunit \epsF}(z)\| \leq 2\| \GRes(z)\| . \end{equation*} By the stability of the solution of the DEL \eqref{eq:SolStab} and (\textbf{M1}), there exists $C_2>0$ such that for any $E\in B_{\epsB}$, $\eta\geq 0$ and $0\leq \epsF \leq 1$ \begin{equation*} \| \Soll{\imunit \epsF}(z)\| \leq C_2 . \end{equation*} On the other hand, if we apply the trivial bound $\Soll{\imunit \epsF}(z) \leq \epsF^{-1}$ for $\epsF \geq 1$, we obtain that \begin{equation}\label{eq:MregB2} \|\Soll{\imunit \epsF}(z)\| \leq \max\{1, C_2\} =: C_3 \end{equation} for $E\in B_{\epsB}$, $\eta\geq 0$ and $\epsF\geq 0$. Now, using \eqref{eq:MregA} and \eqref{eq:MregB2}, there exists $C_4>0$ such that for all $E\in B_{\epsB}$, $\eta\geq 0$ and $\epsF\geq 0$ \begin{equation}\label{eq:MinvRegB1} \|(\Soll{\imunit \epsF}(z))^{-1}\| \leq C_4(1+|z| + \epsF) . \end{equation} To obtain \eqref{eq:LregB} note that \begin{equation} \|\GStOp_{\imunit \epsF} - \GStOp \| \leq \|\Soll{\imunit \epsF}(z)-\GSol(z)\|(\|\Soll{\imunit \epsF}(z)\| + \|\GSol(z)\| ) \|\SuOp \| \leq C_5 \epsF \end{equation} for some $C_5>0$. Therefore by (\textbf{M2}) there exists $\epsilon_1>0$ and $C_6>0$ such that for $0\leq \epsF \leq \epsilon_1$ \begin{equation*} \|(\GStOp_{\imunit \epsF})^{-1}\| = \|\GStOp^{-1}(I-(\GStOp-\GStOp_{\imunit \epsF})\GStOp^{-1})^{-1}\| \leq 2\|\GStOp^{-1}\| \leq C_6 . \end{equation*} By the definition of $\GStOp_{\imunit \epsF}(z)$ and the trivial bound $\|\Soll{\imunit \epsF}(z)\|\leq \epsF^{-1}$, there exists $\epsilon_2>0$ such that for $\epsF \geq \epsilon_2$ \begin{equation*} \| (\GStOp_{\imunit \epsF})^{-1}\| \leq \frac{1}{1-\|\Soll{\imunit \epsF}(z)\|^2 \|\SuOp\|} \leq 2 . \end{equation*} Finally, by \cite[Lemma~3.7]{AltErdoKrugNemi_Kronecker}, compactness of $B_{\epsB}$, \eqref{eq:MregB2} and \eqref{eq:MinvRegB1} there exists $C_7>0$ such that for all $E\in B_{\epsB}$, $0\leq \eta \leq 1$ and $\epsilon_1 \leq \epsF \leq \epsilon_2$ \begin{equation*} \| (\GStOp_{\imunit \epsF})^{-1}\| \leq \frac{\| \Soll{\imunit \epsF}(z)\|^2 \|(\Soll{\imunit \epsF}(z))^{-1}\|^{9}}{(\mathrm{dist}(\mathrm{supp} (\rho_{z}),\imunit \epsF))^{8}} \leq C_7 , \end{equation*} where $\rho_z(x):= \lim_{u\downarrow 0} (\pi m)^{-1}\Tr \Im \Soll{ x + \imunit u}(z)$. To finish the proof, take $C>\max\{2,2C_1,C_3,C_4,C_6,C_7\}$. \end{proof} Now we state the local law for the regularized resolvent. \begin{lem} \label{lem:LLgoodB} Uniformly for $E\in B_{\epsB}$, $0\leq \eta \leq 1$ and $\epsF\geq N^{-1+\gamma}$ \begin{equation}\label{eq:LLgoodB} \max_{i,j\in\llbr N \rrbr} \|P_{ij}(\Ress{\imunit \epsF}(E+\imunit \eta))-\Soll{\imunit \epsF}(E+\imunit \eta)\delta_{ij}\| \prec \sqrt{\frac{1}{N\epsF}} . \end{equation} \end{lem} \begin{proof} Follows from Lemma~B.1 in \cite{AltErdoKrugNemi_Kronecker}. Indeed, by \eqref{eq:MregB} and \eqref{eq:LregB} for all $E\in B_{\epsB}$, $0\leq \eta \leq 1$ and $\epsF \geq N^{-1+\gamma}$ we have \begin{equation*} \max_{i,j\in\llbr N \rrbr} \|P_{ij}(\Ress{\imunit \epsF}(E+\imunit \eta))-\Soll{\imunit \epsF}(E+\imunit \eta)\delta_{ij}\| \prec \frac{1}{1+\epsF}\sqrt{\frac{\|\Soll{\imunit \epsF}(E+\imunit \eta)\|}{N\epsF}}+\frac{1}{(1+\epsF^2)N}+\frac{1}{(1+\epsF^2)N\epsF} . \end{equation*} The fact that $\Soll{\imunit \epsF}(E+\imunit \eta)$ is bounded by \eqref{eq:MregB} yields \eqref{eq:LLgoodB}. \end{proof} We are ready to prove the main theorem. \begin{proof}[Proof of Theorem~\ref{thm:main}] By \cite[Lemma~4.4]{AltErdoKrugNemi_Kronecker} and Lemma~\ref{lem:RegBnd}, for $E\in B_{\epsB}$, $0\leq \eta \leq 1$ and $\epsG\geq 0$ \begin{equation*} \max_{1\leq i \leq N}\|P_{ii}(\Ress{\imunit \epsG}(z)) - \Soll{\imunit \epsG}(z) \| \chi(\Lambda^{\epsG} \leq \vartheta^{\epsG}) \prec \frac{1}{\sqrt{N}}+\Lambda^{\epsG}_{\mathrm{hs}}+\| ( \Soll{\imunit \epsG}(z))^{-1}\| (\Lambda^{\epsG}_{\mathrm{w}})^{2} , \end{equation*} where \begin{align*} &\Lambda_{\mathrm{hs}}^{\epsG}(z) := \frac{1}{N}(\Tr \Ress{\imunit \epsG}(z)^* \Ress{\imunit \epsG}(z))^{1/2} , \\ &\Lambda_{\mathrm{w}}^{\epsG}(z) := \frac{1}{\sqrt{2N}}\max_{i}\big(\Tr P_{ii}[\Ress{\imunit \epsG}(z)^*\Ress{\imunit \epsG}(z) +\Ress{\imunit \epsG}(z) \Ress{\imunit \epsG}(z)^*]\big)^{1/2} \end{align*} and \begin{equation*} \vartheta^{\epsG} := \frac{1}{4(\|(\GStOp_{\epsG})^{-1}\| \| \Soll{\imunit \epsG}(z)\| \|\SuOp\| + \|( \Soll{\imunit \epsG}(z))^{-1}\|)} . \end{equation*} To estimate $\Lambda_{\mathrm{hs}}^{\epsG}$ note that $\Lambda_{\mathrm{hs}}^{\epsG}=N^{-1}\normHS{\Ress{\imunit \epsG}(z)}$, where for any $n\in \NN$ we denote by $\normHS{\,\cdot \,}: \CC^{n \times n} \rightarrow [0, +\infty)$ the usual Hilbert-Schmidt norm, i.e., for any $R\in \CC^{n\times n}$ \begin{equation*} \normHS{R}^2 = \Tr R^* R . \end{equation*} By the resolvent identity \begin{align*} \normHS{\Ress{\imunit \epsG}(z)} &\leq \normHS{(\Hb-EJ\otimes I_N - \imunit (\eta + \epsG) I_m \otimes I_N )^{-1}} \\ &\qquad\qquad +\eta\normOp{\Ress{\imunit \epsG}(z)} \normHS{(\Hb-EJ\otimes I_N - \imunit (\eta+\epsG) I_m \otimes I_N )^{-1}} \\ &\leq \normHS{ \Ress{\imunit (\eta+\epsG)}(E) } + C \normHS{\Ress{\imunit (\eta+\epsG)}(E) } \\ &\lesssim \normHS{\Ress{\imunit(\eta+\epsG)}(E)} , \end{align*} where we used \eqref{eq:GregB} to obtain the bound $\eta\normOp{\Ress{\imunit \epsG}(z)}\leq C$ for some $C>0$ uniformly on $E\in B_{\epsB}$, $0 < \eta \leq 1$ and $\epsG \geq 0$. Since $\Ress{\imunit(\eta+\epsG)}(E)$ is a resolvent with spectral parameter $\imunit (\eta + \epsG)$, we can apply to it the Ward identity, which together with Lemma~\ref{lem:LLgoodB} gives \begin{equation}\label{eq:HSbound} \frac{1}{N}\normHS{\Ress{\imunit(\eta + \epsG)}(E)} = \left( \frac{ \Tr \Im \Ress{\imunit (\eta + \epsG)}(E) }{N^2(\eta + \epsG)}\right)^{1/2} \prec \sqrt{\frac{1}{N(\eta + \epsG)}} \end{equation} uniformly for $E\in B_{\epsB}$, $N^{-1+\gamma} \leq \eta \leq 1$ and $\epsG\geq 0$. In order to estimate $\Lambda_{\mathrm{w}}^{\epsG}$, we introduce the norm $\| \, \cdot \, \|_{\mathrm{w}} : \CC^{m\times m} \otimes \CC^{N\times N} \rightarrow [0, +\infty)$ given by \begin{equation*} \| \Rb \|_{\mathrm{w}}^2 = \max_{1\leq i \leq N} \Tr P_{ii}(\Rb \Rb^*) . \end{equation*} One can easily see that $\Lambda_{\mathrm{w}}^{\epsG}\sim N^{-1/2}\| \Ress{\imunit \epsG}(z) \|_{\mathrm{w}}$. Then similarly as for $\normHS{\, \cdot \,}$, \begin{align*} \normW{\Ress{\imunit \epsG}(z)} &\leq \normW{(\Hb-EJ\otimes I_N - \imunit (\eta + \epsG) I_m \otimes I_N )^{-1}} \\ &\qquad\qquad +\eta\normOp{\Ress{\imunit \epsG}(z)} \normW{(\Hb-EJ\otimes I_N - \imunit (\eta+\epsG) I_m \otimes I_N )^{-1}} \\ &\leq \normW{ \Ress{\imunit (\eta+\epsG)}(E) } + C \normW{\Ress{\imunit (\eta+\epsG)}(E) } \\ &\lesssim \normW{\Ress{\imunit(\eta+\epsG)}(E)} . \end{align*} By applying again the Ward identity and Lemma~\ref{lem:LLgoodB} we obtain that uniformly for $E\in B_{\epsB}$, $N^{-1+\gamma}\leq \eta \leq 1$ and $\epsG \geq 0$ \begin{equation*} \frac{1}{N}\normW{\Ress{\imunit (\eta + \epsG)}(E)}^2 \lesssim \max_{1\leq i\leq N}\frac{\Im \Tr P_{ii}(\Ress{\imunit(\eta+\epsG)}(E))}{N(\eta+\epsG)} \prec \frac{1}{N(\eta+\epsG)} . \end{equation*} Together with \eqref{eq:HSbound} and \eqref{eq:MregB} this implies that \begin{equation}\label{eq:errTbndEpsG2} \max_{1\leq i \leq N}\|P_{ii}(\Ress{\imunit \epsG}(z)) - \Soll{\imunit \epsG}(z) \| \chi(\Lambda^{\epsG} \leq \vartheta^{\epsG}) \prec \frac{1}{\sqrt{N}}+\sqrt{\frac{1}{N(\eta+\epsG)}} \end{equation} for $E\in B_{\epsB}$, $N^{-1+\gamma}\leq \eta \leq 1$ and $\epsG \geq 0$. For any $1\leq i, j \leq N, i\neq j,$ we have that \begin{equation*} \| P_{ij}(\Ress{\imunit \epsG}(z))\| \lesssim \|P_{ii}(\Ress{\imunit \epsG}(z))\|\Lambda_{\mathrm{w}}^{\epsG} , \end{equation*} which can be bounded by $ \|\Soll{\imunit (\eta + \epsG)}(E)\| \normW{\Ress{\imunit(\eta+\epsG)}(E)} + \Lambda^{ \epsG} \normW{\Ress{\imunit(\eta+\epsG)}(E)}$. The second term can be absorbed into $\Lambda^{ \epsG}$, so by using \eqref{eq:errTbndEpsG2} and \eqref{eq:MregB} we end up with the bound \begin{equation*} \max_{1\leq i,j \leq N}\|P_{ij}(\Ress{\imunit \epsG}(z)) - \Soll{\imunit \epsG}(z)\delta_{ij} \| \chi(\Lambda^{\epsG} \leq \vartheta^{\epsG}) \prec \frac{1}{\sqrt{N}}+\sqrt{\frac{1}{N(\eta+\epsG)}} \end{equation*} uniformly on $E\in B_{\epsB}$, $N^{-1+\gamma}\leq \eta \leq 1$ and $\epsG \geq 0$ Since $\vartheta^{\epsG}(z) \gtrsim \epsG^{-1}$ by \eqref{eq:MregB} and \eqref{eq:LregB}, and $\Lambda^{ \epsG}(z) \prec \epsG^{-2}$ by \cite[Lemma~4.4, (i)]{AltErdoKrugNemi_Kronecker}, we can choose $\epsG_1>0$ such that for all $E\in B_{\epsB}$ \begin{equation}\label{eq:LetaEpsG} \Lambda^{ \epsG_1}(E+\imunit) \leq \vartheta^{\epsG_1}(E+\imunit) . \end{equation} Then by \cite[Lemma~A.2]{AjanErdoKrug_PTRF1} \eqref{eq:LetaEpsG} holds not only for $\epsG=\epsG_1$, but \emph{a.w.o.p.} for all $0\leq \epsG \leq \epsG_1$. In particular, we will have that \emph{a.w.o.p.} \begin{equation*} \Lambda(E+\imunit) \leq \vartheta(E+\imunit) . \end{equation*} On the other hand, if we take $\epsG=0$ in \eqref{eq:errTbndEpsG2} we will get that for $E\in B_{\epsB}$ and $0\leq \eta \leq 1$ \begin{equation*} \max_{1\leq i \leq N}\|\Gm_{ii}(z) - \GSol(z) \| \chi(\Lambda \leq \vartheta) \prec \frac{1}{\sqrt{N}}+\sqrt{\frac{1}{N\eta}} . \end{equation*} Applying \cite[Lemma~A.2]{AjanErdoKrug_PTRF1} to $\Lambda(E+\imunit \eta)$ and $\vartheta(E+\imunit \eta)$ we get that for $E\in B_{\epsB}$ and $N^{-1+\gamma} < \eta \leq 1$ \begin{equation*} \Lambda(E+\imunit \eta) \leq \vartheta(E+\imunit \eta) , \end{equation*} which yields the first inequality in \eqref{eq:locallaw}. To prove the averaged local law we will use the fluctuations averaging mechanism, proof of which in a suitable form can be found in \cite[Proposition~4.6]{AltErdoKrugNemi_Kronecker}, see also Section~10 of \cite{ErdoYauBook} related previous proofs. To this end, we introduce conditional expectation with respect to the $i$th rows and columns of the matrices $X_{\alpha}$ and $Y_{\beta}$ \begin{equation*} \EE_{i}[\, \cdot \,] := \EE \Big[\, \cdot \, \Big| \Big\{X_{\alpha}(k,l), Y_{\beta}(k,l)\, : \, \alpha\in \llbr \alpha_*\rrbr, \beta\in \llbr \beta_*\rrbr, k,l\in \llbr N \rrbr \setminus \{i\}\Big\}\Big] , \end{equation*} and a family of operators \begin{equation*} \Qcc_{i}[\, \cdot \,] := \mathrm{Id}[\, \cdot \,] - \EE_{i}[\, \cdot \, ] . \end{equation*} By the Schur complement formula (see e.g. \cite[Section~4.1]{AltErdoKrugNemi_Kronecker}), for all $i\in \llbr N \rrbr$ we have that \begin{equation*} -\frac{1}{\Gm_{ii}} = zJ - P_{ii}(\Hb) + \sum_{k,l \neq i} P_{ik}(\Hb) P_{kl}(\GRes^{\{i\}}) P_{li}(\Hb) , \end{equation*} where \begin{equation*} \GRes^{\{i\}}(z): = \Big( K_0\otimes I_N - \sum_{\alpha=1}^{\alpha_*} K_\alpha\otimes X_\alpha^{\{i\}} - \sum_{\beta=1}^{\beta_*}\big( L_{\beta}\otimes Y_\beta^{\{i\}}+L^*_{\beta}\otimes Y_\beta^{\{i\}*} \big) - zJ \otimes I_N \Big)^{-1} \end{equation*} and matrices $X_\alpha^{\{i\}}$ and $Y_\alpha^{\{i\}}$ are obtained from $X_\alpha$ and $Y_\alpha$, respectively, by replacing their $i$th rows and columns by zero. Then the difference between $\Gm_{ii}(z)$ and $\GSol(z)$ can be written as \begin{align}\label{identity} \Gm_{ii}(z)-\GSol(z) &= \GSol(z)\SuOp\Big[\frac{1}{N}\sum_{j=1}^{N}\big(\Gm_{jj}(z)-\GSol(z)\big)\Big]\Gm_{ii}(z) \\ &\quad - \GSol(z) \Qcc_{i}\Big[\frac{1}{\Gm_{ii}(z)}\Big] \Gm_{ii}(z) \nonumber \\ &\quad + \GSol(z)\SuOp\Big[\frac{1}{N}\sum_{j\neq i}\big (P_{jj}(\GRes^{\{i\}}(z)) -\Gm_{jj}(z)\big)\Big]\Gm_{ii}(z) \nonumber \\ &\quad - \GSol(z)\SuOp\Big[\frac{\Gm_{ii}(z)}{N}\Big]\Gm_{ii}(z) \nonumber . \end{align} Note, that by using the large deviation bounds (see e.g. \cite[Lemma~4.3]{AltErdoKrugNemi_Kronecker}) we have \begin{equation*} \Big \| \Qcc_{i}\Big[\frac{1}{\Gm_{ii}(z)}\Big] \Big\| \prec \sqrt{\frac{1}{N\eta}} . \end{equation*} After taking the average over $i\in \llbr N \rrbr$, rearranging the terms in \eqref{identity} and using the entry-wise local law from \eqref{eq:locallaw}, (\textbf{M1}), boundedness of $\SuOp$ and the formula \begin{equation*} \Gm_{jj}(z)- P_{jj}(\GRes^{\{i\}}(z)) = \Gm_{ji}(z)\frac{1}{\Gm_{ii}(z)}\Gm_{ij}(z) ,\quad j\neq i , \end{equation*} we obtain \begin{equation}\label{eq:AvLL1} \GStOp \Big[\frac{1}{N}\sum_{i=1}^{N}\Gm_{ii}(z)-\GSol(z) \Big] = - \GSol(z) \frac{1}{N}\sum_{i=1}^{N}\Qcc_{i}\Big[\frac{1}{\Gm_{ii}(z)}\Big] \GSol(z) +O_{ \prec }\Big(\frac{1}{N}+\frac{1}{N\eta}\Big) , \end{equation} where $O_{\prec}(N^{-1}+(N\eta)^{-1})$ collects terms stochastically dominated by $N^{-1} + (N\eta)^{-1}$ . Applying again the entry-wise local law \eqref{eq:locallaw} and \eqref{eq:MregB}, we have that uniformly on $E\in B_{\epsB}$ and $N^{-1+\gamma} < \eta \leq 1$ \begin{equation}\label{eq:FlAvCond} \max_{i,j\in \llbr N \rrbr}\Big\| \frac{1}{\GSol(z)}\Gm_{ij}(z)-\delta_{ij} I_m \Big\| \prec \sqrt{\frac{1}{N\eta}} \leq N^{-\gamma} . \end{equation} Inequality \eqref{eq:FlAvCond} allows us to improve a bound on the first term on the RHS of \eqref{eq:AvLL1} by using the fluctuation averaging (see \cite[Proposition~4.6]{AltErdoKrugNemi_Kronecker}), which gives that for $E\in B_{\epsB}$ and $N^{-1+\gamma} < \eta \leq 1$ \begin{equation*} \frac{1}{N}\sum_{i=1}^{N}\Qcc_{i}\Big[\frac{1}{\Gm_{ii}(z)}\Big] \GSol(z) \prec \frac{1}{N\eta} . \end{equation*} Now the boundedness of $\GStOp^{-1}$ from (\textbf{M2}) yields the second inequality in \eqref{eq:locallaw}. This completes the proof of Theorem~\ref{thm:main}. \end{proof} Finally we prove the speed of convergence in the global law: \begin{proof}[Proof of Proposition~\ref{pr:speed}] First we prove \eqref{fp} for bounded test functions. In this case, by the Helffer-Sj\"ostrand formula (see, e.g. Section 11.2 of \cite{ErdoYauBook}) it is sufficient to prove \eqref{fp} for all functions of the form $f(x)= (x-z)^{-1}$ with any fixed $z\in \CC_+$. Consider any nilpotent linearization of $p$ and let $M(z)$ be the solution of DEL \eqref{eq:MDE}. Notice that \begin{equation}\label{MM} \| \GSol(z)\| \le C\Big( 1+ \frac{1}{\Im z}\Big), \qquad \| \GStOp^{-1}\| \le C(1+|z|)^2\Big( 1+ \frac{1}{\Im z}\Big)^2, \end{equation} where the first bound was obtained in Lemma~\ref{lem:MDEexistence} (i). The second bound is a consequence of the identity \begin{equation}\label{eq:identity} \GStOp^{-1}[R] =(\id\otimes \tau) \Big( \frac{1}{\Lb-zJ\otimes \freeunit}\big(M^{-1}RM^{-1} \otimes \freeunit \big)\frac{1}{\Lb-zJ\otimes \freeunit}\Big), \end{equation} the a priori bound $\| \GSol^{-1}\|\le C(1+|z|)$, see \eqref{eq:MinvRegB1}, and the bound \eqref{eq:trivResBound} applied to semicircular elements as in \eqref{scbound}. The identity \eqref{eq:identity} follows immediately by expressing the derivative of the function $$ \Phi(A):= (\id\otimes \tau) (\Lb - (zJ-A)\otimes \freeunit )^{-1}, $$ at $A=0$ in two different ways. The derivative on the free probability level gives the right hand side of \eqref{eq:identity}, while the derivative on the level of the Dyson equation \eqref{eq:MDE} perturbed with $A$ gives the left hand side, see \eqref{Mderiv} for a similar calculation. Thus \eqref{MM} shows that the analogues of the conditions \textup{(\textbf{M1})-(\textbf{M2})} hold away from the real axis, i.e. on any compact set $\{ z\in \CC\; : \; |z|\le C^*, \Im z\ge c^*\}$ with fixed positive thresholds $c^*, C^*$. Inspecting the proof of the local law in Section~\ref{sec:locallaw}, we see that the entire argument goes through under these modified assumptions. We leave the details to the reader. Finally, using the boundedness of all $\Xb$ and $\Yb$ matrices with very high probability, a standard cutoff argument yields \eqref{fp} for arbitrary smooth function. \end{proof} \section{Examples} \label{sec:examples} In this section we prove optimal bulk local law (in the sense of Theorem~\ref{thm:mainPoly}) for two concrete families of polynomials of random matrices, namely, for the eigenvalues of quadratic forms in Wigner matrices and for eigenvalues of symmetrized products (i.e. singular values of products) of matrices with \emph{i.i.d.} entries. \subsection{Local law for homogeneous polynomials of degree two in Wigner matrices} \label{sec:example} Consider a family of noncommutative self-adjoint polynomials of degree 2 in $\gamma_*\geq 2$ variables given by \begin{equation*} \tilde{q}(x_1,\ldots,x_{\gamma_*}) = \xb^{t} \,\Xi \, \xb , \end{equation*} where $\xb = (x_1,\ldots, x_{\gamma_*})^{t}$ and $\Xi$ is a Hermitian $\gamma_*\times \gamma_*$ matrix. We will assume that $\Xi$ is invertible. Note that if we take $\Xi = \left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right)$, then $\tilde{q}(x_1, x_2) = x_1 x_2 + x_2 x_1$, the anticommutator, which was studied by Anderson in \cite{Ande_2015}. Suppose that $\Lb$ is the minimal linearization of $\cstarunit-\tilde{q}(\xb)$ and suppose that for any $\epsB>0$ the assumptions (\textbf{M1}) and (\textbf{M2}) hold for $\Lb$ and the corresponding solution of the \emph{DEL} everywhere in the $\epsB$-bulk. This, together with Theorem~\ref{thm:mainPoly}, would imply that the optimal local law holds for the polynomial $\cstarunit-\tilde{q}$ everywhere in the $\epsB$-bulk. Therefore, in order to prove the local law it is enough to find a minimal linearization of $\cstarunit-\tilde{q}$ that satisfies (\textbf{M1}) and (\textbf{M2}). Before proceeding to the proof, we fix a specific linearization of the polynomial $\cstarunit-\tilde{q}$, which is particularly suitable for our computations. More precisely, let $\Lb(\xb) = K_0 - \sum_{\gamma = 1}^{\gamma_*} K_{\gamma} x_{\gamma}$ with \begin{equation}\label{eq:ExLin} K_0 = \left( \begin{array}{c|ccc} 1 & 0 & \cdots & 0 \\ \hline 0 & & & \\ \vdots & &\Xi^{-1} & \\ 0 & & & \end{array} \right) ,\quad K_{\gamma} = \left( \begin{array}{c|ccc} 0 & & e_{\gamma}^{t} & \\ \hline & 0 & \cdots & 0 \\ e_{\gamma} & \vdots& & \vdots \\ & 0 & \cdots & 0 \end{array} \right) . \end{equation} One can easily see that $\Lb(\xb)$ gives a linearization of $\cstarunit-\tilde{q}$. Moreover, this linearization is in fact minimal. To show this latter property of \eqref{eq:ExLin}, note that the matrix representation of the series $(1-\tilde{q})^{-1}$ corresponding to the linearization \eqref{eq:ExLin} (see Remark~\ref{rem:aAdvantage}) is given by $(K_0^{-1} e_1, e_1, K_1 K_0^{-1}, \ldots, K_{\gamma_*} K_0^{-1})$. From the special structure of $K_0^{-1}$ and $K_{\gamma}$ we see that for \begin{equation*} K_{\gamma} K_0^{-1} e_1 = \begin{pmatrix} 0 \\ e_{\gamma} \end{pmatrix} ,\quad 1\leq \gamma \leq \gamma_* . \end{equation*} Therefore, \begin{equation*} \mathrm{span}\{e_1, K_{1} K_0^{-1} e_1, \ldots, K_{\gamma_*} K_0^{-1} e_1\} = \mathrm{span}\{e_1, e_2, \ldots, e_{\gamma_*+1}\} = \CC^{\gamma_*+1} , \end{equation*} which corresponds to condition \eqref{eq:aMinCrit1} in Proposition~\ref{prop:min}. Similarly, one can show that the condition \eqref{eq:aMinCrit2} is satisfied as well, i.e., \begin{equation*} \mathrm{span}\{K_0^{-1}e_1, K_0^{-1}K_{1} K_0^{-1} e_1, \ldots, K_0^{-1}K_{\gamma_*} K_0^{-1} e_1\} = \CC^{\gamma_*+1} . \end{equation*} We then conclude using Proposition~\ref{prop:min} that the linearization \eqref{eq:ExLin} is indeed minimal. Below we show that for this choice of linearization conditions (\textbf{M1}) and (\textbf{M2}) hold everywhere in the $\epsB$-bulk. \subsubsection{Boundedness of $\GSol$ (assumption (\textbf{M1}))} First, we show that the solution $\GSol(z)$ of the corresponding \emph{DEL} has the following structure \begin{equation} \label{eq:Ex1} \GSol(z) = \left( \begin{array}{c|ccc} \GSol_{11}(z) & 0 & \cdots & 0 \\ \hline 0 & & & \\ \vdots & & \widehat{\GSol}(z) & \\ 0 & & & \end{array} \right) , \end{equation} where by $\widehat{\GSol}(z)$ we denote the $\gamma_*\times \gamma_*$ submatrix of $\GSol(z)$. Indeed, recall that by \eqref{eq:nonSymM}-\eqref{eq:expansionT} and \eqref{eq:rescaledsolution} for $|z|>C$, $C>0$ big enough, \begin{align} \nonumber \GSol(z) &= K_0^{-1} D_{z}^{-1} (\id \otimes \tau) \Big(I + \sum_{k=1}^{\infty} \Big( \sum_{\gamma=1}^{\gamma_*}A_{\gamma}^{z} \otimes \semic_{\gamma}\Big)^k\Big) D_{z}^{-1} \\ \label{eq:largeZexp} &= K_0^{-1} D_{z}^{-1} (\id \otimes \tau)\Big(I + \sum_{\ell=1}^{\infty}\Big( \sum_{\gamma=1}^{\gamma_*}A_{\gamma}^{z} \otimes \semic_{\gamma} \Big)^{2\ell} \Big) D_{z}^{-1} , \end{align} where the second equality is due to \eqref{eq:aSemicMoments} and $\semic_1,\ldots,\semic_{\gamma_*}$ being free independent. On the other hand, $K_0^{-1}$, $D_{z}^{-1}$ and the products $A_{i(1)}^{z}A_{i(2)}^{z}\cdots A_{i(2\ell)}^{z}$ for any $\ell \in \NN$ and $i(j)\in \{1,\ldots,\gamma_*\}$ are all of the form \begin{equation*} \left( \begin{array}{c|ccc} * & 0 & \cdots & 0 \\ \hline 0 & * & \cdots & * \\ \vdots & \vdots & & \vdots \\ 0 & * & \cdots & * \end{array} \right) , \end{equation*} where we use stars to designate arbitrary unspecified entries. By analyticity we obtain that \eqref{eq:Ex1} holds for all $z\in \CC_+$. Now, for $\GSol = \GSol(z)$ of the form \eqref{eq:Ex1} we write the \emph{DEL} \begin{equation*} I + (zJ-K_0) \GSol + \sum_{\gamma=1}^{\gamma_*} K_{\gamma}\, \GSol \, K_{\gamma} \, \GSol = 0 , \end{equation*} which can be split into two parts \begin{align} \label{eq:ExMDE1} 1 + (z-1) \GSol_{11} + \GSol_{11} \Tr\widehat{\GSol} &= 0 , \\ I - \Xi^{-1} \widehat{\GSol} + \GSol_{11} \widehat{\GSol} &= 0 . \nonumber \end{align} From \eqref{eq:ExMDE1} we obtain that \begin{equation} \label{eq:ExMDE3} \widehat{\GSol} = (\Xi^{-1} - \GSol_{11}I)^{-1} . \end{equation} Recall, that by the definition of the $\epsB$-bulk there exists $\eta_0>0$ small enough such that for all $\eta \in [0,\eta_0]$ and $E \in B_{\epsB}$ we have $\Im \GSol_{11}(E+\imunit \eta) \geq \epsB/2$. Therefore, since $\Xi$ is self-adjoint, \eqref{eq:ExMDE3} implies that $\| \widehat{\GSol}\| \leq 2/\epsB$ for all $E\in B_{\epsB}$ and $\eta \in [0, \eta_0]$. Moreover, by plugging \eqref{eq:ExMDE3} into \eqref{eq:ExMDE1} we derive an equation for $\GSol_{11}$ \begin{equation*} 1 + (z-1) \GSol_{11} + \sum_{\gamma=1}^{\gamma_*} \frac{\GSol_{11}}{\xi_{\gamma}^{-1}-\GSol_{11}} = 0 , \end{equation*} where by $\xi_{\gamma} \in \RR$, $1\leq \gamma \leq \gamma_*$, we denoted the eigenvalues of $\Xi$. Note that if there exists an unbounded solution $m$ of the equation \begin{equation*} 1 + (z-1) m + \sum_{\gamma=1}^{\gamma_*} \frac{m}{\xi_{\gamma}^{-1}-m} = 0 , \end{equation*} then $m(z)$ can be unbounded only near the point $z=1$. In this case \begin{equation*} \lim_{z\rightarrow 1} (z-1)m(z) = \gamma_*-1 , \end{equation*} which implies that $\Im m(z) < 0$ in the neighborhood of $z=1$. Therefore, function $\GSol_{11}(z)$, whose imaginary part by Lemma~\ref{lem:MDEexistence} (iii) must be nonnegative on $\CC_+$, has absolute value bounded by some $C>0$ for all $z\in \CC_+$. We conclude that the assumption (\textbf{M1}) holds for the linearization \eqref{eq:ExLin} of the polynomial $1-\tilde{q}(\xb)$ with a constant $C_3=\max\{C,2/\epsB \}$ depending only on the model parameters $\epsB$ and $\Xi$. \subsubsection{Boundedness of $\GStOp^{-1}$ (assumption (\textbf{M2}))} In order to prove the stability assumption (\textbf{M2}), we will have to extract additional information from the \emph{DEL} \eqref{eq:MDE} by taking its imaginary part at $\eta = 0$ \begin{equation}\label{eq:ExImMDE} \Im \GSol = \GSol^* \sum_{\gamma = 1}^{\gamma_*} K_{\gamma} \Im \GSol K_{\gamma } \GSol . \end{equation} By using \eqref{eq:ExLin}, \eqref{eq:Ex1} and \eqref{eq:ExMDE3} and comparing the $(1,1)$-components of both sides of \eqref{eq:ExImMDE}, we obtain that for all real $z=E$, $E\in B_{\epsB}$, \begin{equation} \label{eq:ExImMDE3} |\GSol_{11}|^2 \Tr \widehat{M}^* \widehat{M} = 1 . \end{equation} Now, consider the space of $(\gamma_*+1)\times (\gamma_*+1)$ matrices with basis vectors $\{E_{11},E_{12},\ldots\}$, on which the linear operator $\GStOp$ is acting, as $\CC^{(\gamma_*+1)^2}$ with the standard basis $\{e_1,e_2,\ldots\}$. On this latter space $\GStOp$ can be represented by the matrix \begin{equation*} \Ab_{\GStOp} := I_{(\gamma_*+1)^2} - \sum_{\gamma=1}^{\gamma_*} \GSol K_{\gamma} \otimes \GSol^{t} K_{\gamma} , \end{equation*} or more explicitly \begin{equation*} \Ab_{\GStOp} = \begin{pmatrix} I_{\gamma_*+1} & -\Ab_{12} \\ -\Ab_{21} & I_{(\gamma_*+1)\gamma_*} \end{pmatrix} , \end{equation*} where \begin{equation*} \Ab_{12} = \GSol_{11} \sum_{\gamma=1}^{\gamma^*} e_{\gamma}^{t} \otimes \GSol^{t} K_{\gamma} , \quad \Ab_{21} = \sum_{\gamma=1}^{\gamma_*} \widehat{\GSol} e_{\gamma} \otimes \GSol^{t} K_{\gamma} . \end{equation*} Note, that \begin{equation*} \det(\Ab_{\GStOp}) = \det(I_{\gamma_*+1} -\Ab_{12} \Ab_{21} ) , \end{equation*} therefore, in order to prove invertibility of $\GStOp$ it will be enough to show invertibility of \begin{equation}\label{eq:exampleSmallMatrix} I_{\gamma_*+1} -\Ab_{12} \Ab_{21} = \left( \begin{array}{c|ccc} 1-\GSol_{11}^{2} \Tr(\widehat{\GSol})^2 & 0& \cdots& 0 \\ \hline 0 & & & \\ \vdots & & I_{\gamma_*} - \GSol_{11}^{2} (\widehat{\GSol}^{t}\widehat{\GSol}) & \\ 0 & & & \end{array} \right) . \end{equation} Assume that the upper-left entry is not invertible, i.e., \begin{equation*} 1-\GSol_{11}^{2} \Tr (\widehat{\GSol})^2 = 1-\GSol_{11}^{2} \sum_{\gamma =1}^{\gamma_*} \frac{1}{(\xi_{\gamma}^{-1}-M_{11})^{2}} = 0 , \end{equation*} where we used \eqref{eq:ExMDE3}. In this case, from \eqref{eq:ExImMDE3}, we obtain that for all $1\leq \gamma \leq \gamma_*$ \begin{equation*} (\GSol_{11}(\xi_{\gamma}^{-1} - \GSol_{11})^{-1})^2 = |\GSol_{11}(\xi_{\gamma}^{-1} - \GSol_{11})^{-1}|^2 , \end{equation*} so that $\Im (\GSol_{11}(\xi_{\gamma}^{-1} - \GSol_{11})^{-1}) = \Im (\GSol_{11})\xi_{\gamma}^{-1}|\xi_{\gamma}^{-1} - \GSol_{11}|^{-2} = 0$ for all $1\leq \gamma \leq \gamma_*$, which leads to a contradiction with $\Im \GSol_{11} > \epsB$ for $z=E$, $E\in B_{\epsB}$. Consider now the case when $ \det (I_{\gamma_*} - \GSol_{11}^{2} (\widehat{\GSol}^{t}\widehat{\GSol}))=0$, so that the lower-right submatrix of \eqref{eq:exampleSmallMatrix} is singular. This implies that there exists $\omega \in \CC^{\gamma_*}$, $\|\omega \| = 1$, such that \begin{equation} \label{eq:ExStab1} \GSol_{11}^{2} \omega^* \widehat{\GSol}^{t}\widehat{\GSol} \omega = 1 . \end{equation} We can rewrite the LHS of \eqref{eq:ExStab1} as \begin{equation*} \sum_{k=1}^{\gamma_*} \la \omega, \vb_k \ra \la \overline{\omega}, \vb_k \ra = \sum_{k=1}^{\gamma_*} (\la \Re \omega, \vb_k \ra )^2 + (\la \Im \omega, \vb_k \ra)^2 , \end{equation*} where \begin{equation*} \vb_k := (\GSol_{11} \widehat{\GSol}_{k1},\GSol_{11} \widehat{\GSol}_{k2}, \ldots, \GSol_{11} \widehat{\GSol}_{k\gamma_*})^{t} . \end{equation*} Due to \eqref{eq:ExImMDE3}, using triangular and Cauchy-Schwarz inequalities, we have \begin{equation}\label{eq:ExStab3} \Big|\sum_{k=1}^{\gamma_*} (\la \Re \omega, \vb_k \ra )^2 + (\la \Im \omega, \vb_k \ra)^2 \Big| \leq \|\omega \|^2 \sum_{k=1}^{\gamma_*}\|\vb_k\|^2 = |\GSol_{11}|^2 \Tr \widehat{\GSol}^* \widehat{\GSol} = 1 . \end{equation} Assumption \eqref{eq:ExStab1} implies that the first inequality in \eqref{eq:ExStab3} is in fact an equality and that \begin{equation*} |\la \Re \omega, \vb_k \ra|^2 = \|\Re \omega \|^2 \| \vb_k \|^2 ,\quad |\la \Im \omega, \vb_k \ra|^2 = \|\Im \omega \|^2 \| \vb_k \|^2 ,\quad 1\leq k \leq \gamma_* . \end{equation*} Thus there exist $c_1^{(1)},\ldots,c_{\gamma_*}^{(1)},c_1^{(2)},\ldots,c_{\gamma_*}^{(2)} \in \CC$ such that \begin{equation*} \vb_k = c_k^{(1)} \Re \omega = c_k^{(2)} \Im \omega , \end{equation*} and we see that the rows of the matrix $\GSol_{11} \widehat{\GSol}$ are linearly dependent. At the same time we know that since $\Im \GSol_{11}>\epsB$ in the $\epsB$-bulk, by \eqref{eq:ExMDE3} matrix $\widehat{\GSol}$ must be invertible. From the obtained contradiction we conclude that $I_{\gamma_*+1} - \Ab_{12} \Ab_{21}$, $\Ab_{\GStOp}$ and $\GStOp$ are all invertible for $z=E$ with $E\in B_{\epsB}$, so that there exists $C>0$ depending only on $\epsB$ and $\Xi$, such that $\| \GStOp^{-1}(E) \| \leq C$ for all $E\in B_{\epsB}$. Now a simple continuity argument, together with the a priori bound from Lemma~\ref{lem:MDEexistence}, shows that the condition (\textbf{M2}) holds for the model given by \eqref{eq:ExLin} everywhere in the $\epsB$-bulk. \subsection{Local law for singular values of a product of independent non-Hermitian matrices} \label{sec:exampl-sing-valu} Consider $q(y_1,\ldots, y_{\beta_*}, y_1^*, \ldots, y_{\beta_*}^*) = y_1 \cdots y_{\beta_*} y_{\beta_*}^* \cdots y_1^*$. Then a minimal linearization of the polynomial $\cstarunit - q(\yb,\yb^*)$ is given by \begin{equation} \label{eq:7} \Lb = \left( \begin{array}{c|ccccccc} 1&&&&&&&y_1 \\ \hline &&&&&& y_2 &1 \\ &&&&&\iddots&\iddots& \\ &&&&y_{\beta_*}&1&& \\ &&&y_{\beta_*}^*&1&&& \\ &&\iddots&1&&&& \\ & y_2^* &\iddots&&&&& \\ y_1^*&1&&&&&& \end{array} \right) , \end{equation} or, using the representation \eqref{eq:linearization} and the basis vectors $E_{ij}=e_i e_j^t$, by a set of matrices \begin{equation} \label{eq:6} K_0 = E_{1,1} + \sum_{j=2}^{2\beta_*} E_{j,2\beta_*+2-j} , \qquad L_{\beta} = E_{\beta, 2\beta_*+1-\beta} ,\quad 1\leq \beta \leq \beta_* . \end{equation} Before proving that assumptions (\textbf{M1}) and (\textbf{M2}) hold for this model, we show first that the solution matrix $\GSol(z)$ has the following structure \begin{equation} \label{eq:9} \GSol (z) = \sum_{j=1}^{2\beta_*} m_j(z) E_{jj} + \sum_{j=2}^{2\beta_*} m_{\beta_*+1}(z) E_{j,2\beta_*+2-j} -m_{\beta_*+1}(z) E_{\beta_*+1,\beta_*+1} , \end{equation} for some $m_j : \CC_+\rightarrow \CC$, $1\leq j \leq 2\beta_*$. In order to do so, similarly as in Section~\ref{sec:example}, we use the large $z$ expansion \eqref{eq:largeZexp} for $\GSol (z)$. With $K_{\gamma}$ as in \eqref{eq:symmetrized}, $\alpha_*=0$, and \eqref{eq:6} we have that \begin{equation} \label{eq:10} \sum_{\gamma=1}^{2\beta_*} K_{\gamma} \otimes \semic_{\gamma} = \sum_{\beta=1}^{\beta_*} (E_{\beta,2\beta_*+1-\beta} \otimes c_{\beta} + E_{2\beta_*+1-\beta, \beta} \otimes c_{\beta}^*) , \end{equation} where we denoted $c_{\beta}:= \frac{1}{\sqrt{2}}(s_{\beta} + \imunit s_{\beta_*+\beta})$ and $c_{\beta}^* := \frac{1}{\sqrt{2}}(s_{\beta} - \imunit s_{\beta_*+\beta})$ for $1\leq \beta \leq \beta_*$. Therefore, recalling the definitions of $A_{\gamma}^z$ \eqref{eq:nonSymM}, we obtain \begin{equation*} \sum_{\gamma=1}^{2\beta_*} A_{\gamma}^z \otimes s_{\gamma} = \frac{1}{\sqrt{\zeta}} (E_{1,2} \otimes c_{1}+ E_{2\beta_*,1} \otimes c_{1}^*) + \sum_{\beta=2}^{\beta_*} (E_{\beta,\beta+1} \otimes c_{\beta} + E_{2\beta_*+1-\beta, 2\beta_*+2-\beta} \otimes c_{\beta}^*) \end{equation*} with $\zeta := 1-z$. Due to the cyclic structure of the above matrix, its $2\beta_*$-th power is diagonal with non-zero entries equal to $\zeta^{-1} c_1\cdots c_{\beta_*} c_{\beta_*}^* \cdots c_1^*$. If we now write each power appearing in the expansion \eqref{eq:largeZexp} using $\ell = t\beta_* + k$ with $k=0,1,\ldots,\beta_*-1$, we see that after applying $(\id \otimes \tau)$ the only non-vanishing off-diagonal terms are those containing the multiples of $E_{\beta_*+1-k,\beta_*+1-k}$ or $E_{2\beta_*+1-k,1+k}$. Thus, we conclude that \begin{equation*} D_z K_0 \GSol(z) D_z = \left( \begin{array}{c|ccccccc} \nu &&&&&&& \\ \hline &\nu&&&&&&* \\ &&\ddots&&&&\iddots& \\ &&&\nu&&*&& \\ &&&&\nu&&& \\ &&&*&&\nu&& \\ &&\iddots&&&&\ddots & \\ &*&&&&&&\nu \end{array} \right) , \end{equation*} where $\nu(z) = 1 + \sum_{t=1}^{\infty}\zeta^{-t} \tau((c_1 \cdots c_{\beta_*} c_{\beta_*}^* \cdots c_1^*)^{t})$ and $*$ represent some unspecified functions. From $D_z = \sqrt{\zeta} E_{11} + \sum_{j=2}^{2\beta_*} E_{jj}$, $K_0=K_0^{-1}$ and \eqref{eq:6} we obtain \eqref{eq:9}. \subsubsection{Boundedness of $\GSol$ (assumption (\textbf{M1}))} We now prove that $\GSol$ is bounded everywhere in the $\epsB$-bulk. Using the structure of $\GSol$ \eqref{eq:9}, the \emph{DEL} \eqref{eq:MDE} can be reduced to the following system of equations for $m_{\beta}, 1\leq \beta\leq 2\beta_*$, \begin{equation*} \left\{ \begin{array}{ll} 1-\zeta m_1 + m_{2\beta_*} m_1 = 0, & \\ 1 - m_{\beta_*+1} + m_{2\beta_*+1-\beta} m_{\beta} = 0, & 2\leq \beta \leq 2\beta_*, \\ - m_{2\beta_*+2-\beta} + m_{2\beta_*+1-\beta} m_{\beta+1} = 0, & 2\leq \beta \leq 2\beta_*, i \neq \beta_*+1. \end{array} \right. \end{equation*} From these equations we obtain that all $m_{\beta}$, $\beta\geq 2$, can be expressed in terms of $m_1$: \begin{equation} \label{eq:3} \begin{array}{ll} m_{\beta_*+1} = \zeta m_1 ,& \\ m_{\beta_*+1+\beta} = m_{\beta_*+1}^{\beta+1} = (\zeta m_1)^{\beta+1} ,& 0\leq \beta \leq \beta_*-1 , \\ m_{\beta} = m_1 m_{\beta_*+1}^{\beta-1} = \zeta^{\beta-1}m_1^{\beta} ,& 2\leq \beta \leq \beta_* , \end{array} \end{equation} and $m_1(z)=\la e_1, M(z) e_1 \ra$ satisfies the following polynomial equation \begin{equation} \label{eq:11} 1 - \zeta m_1 + \zeta^{\beta_*} m_1^{\beta_*+1} = 0 ,\qquad \zeta = 1-z. \end{equation} From \eqref{eq:11} it is easy to see that $|m_1(z)|$ can be unbounded only in the neighborhood of $z=1$. Moreover, we will show that there exists $c(\epsB)= c(\epsB,\beta_*)>0$ small enough such that \begin{equation} \label{eq:2} B_{\epsB} \subset [1-C(\beta_*), 1-c(\epsB,\beta_*)] , \end{equation} where $C(\beta_*)\geq 2$ comes from the boundedness of the support of the density of states. In order to prove the upper bound in \eqref{eq:2} we may, without loss of generality, consider only $z\in \CC_+$ with $|\zeta|=|1-z|\leq 4^{-(\beta_*+1)}$ and $\eta=\Im z$ small. We will show that for such $z$ the condition $E=\Re z \in B_{\epsB}$ implies $|\zeta|\geq c(\epsB)$, where $c(\epsB)$ will be specified below. Rewrite \eqref{eq:11} as \begin{equation} \label{eq:14} (\zeta m_1)^{\beta_*+1} = -\zeta (1-\zeta m_1) , \end{equation} from which it follows that $ |\zeta m_1|^{\beta_*+1} \leq |\zeta|(1+|\zeta m_1|)$, so that $|\zeta m_1| \leq 2|\zeta|^{\frac{1}{\beta_*+1}}$. This last bound implies that $|\zeta m_1|< 1/2$, which, together with \eqref{eq:14}, yields \begin{equation} \label{eq:15} |m_1| \sim |\zeta|^{-\frac{\beta_*}{\beta_*+1}} . \end{equation} Suppose that $|\zeta|^{-\frac{\beta_*}{\beta_*+1}} \geq C'\epsB^{-1}$ with some large constant $C'$. For $E$ in the $\epsB$-bulk and $\eta$ small we have $ \Im m_1 \leq 2 \epsB^{-1}$, which together with \eqref{eq:15} gives $ |\Re m_1| \sim |\zeta|^{-\frac{\beta_*}{\beta_*+1}}$. In the regime when $\eta \ll |1-E|$ and $|\zeta| \sim |1-E|$, by taking the imaginary part of the equation \eqref{eq:11} and dividing it through by $\Im m_1 (\Re m_1)^{-1}$ we obtain \begin{equation} \label{eq:17} 0 = -(1-E)\Re m_1 + (\beta_*+1)(1-E)^{\beta_*}(\Re m_1)^{\beta_*+1} + O\bigg(\bigg| \frac{\Im m_1}{\Re m_1}\bigg| + \frac{\eta}{|\zeta|^2}\bigg| \frac{\Re m_1}{\Im m_1}\bigg| \bigg) . \end{equation} Choosing $C'$ sufficiently large and $\eta$ sufficiently small (depending on $\epsB$) the error term becomes negligible, using $\Im m_1\geq \epsB$ since we are in $B_{\epsB}$. Using the scaling of $1-E$ and $\Re m_1$ in $\zeta$, we obtain $|(1-E) \Re m_1| \sim |\zeta|^{\frac{1}{\beta_*+1}}$ and $|(\beta_*+1)(1-E)^{\beta_*}(\Re m_1)^{\beta_*+1}|\sim 1$. Since $\zeta$ is small, this leads to a contradiction in \eqref{eq:17}, hence $|\zeta|^{-\frac{\beta_*}{\beta_*+1}} \geq C'\epsB^{-1}$ cannot hold. This finishes the proof of \eqref{eq:2} with $c(\epsB)=(\epsB/C')^{\frac{\beta_*+1}{\beta_*}} $. Now, for any $E\in B_{\epsB}$ and $\eta>0$ small enough, \eqref{eq:11} and \eqref{eq:2} imply that $ |m_1| \leq (c(\epsB))^{-1}C$, which gives an effective bound on $m_1$. Boundedness of $|m_1|$ together with \eqref{eq:3} implies that assumption (\textbf{M1}) holds everywhere in $B_{\epsB}$ with $C_3 = c(\epsB)^{3\beta_*}$. \subsubsection{Boundedness of $\GStOp^{-1}$ (assumption (\textbf{M2}))} In this section we show that assumption (\textbf{M2}) holds everywhere in the $\epsB$-bulk, which together with Theorem~\ref{thm:mainPoly} implies optimal bulk local law for the singular values of a product of matrices $Y_1\cdots Y_{\beta_*}$ satisfying \textup{(\textbf{H1})-(\textbf{H4})}. By \eqref{eq:6}, which, in particular, gives that $\GSol^{t}(z)=\GSol(z)$, and \eqref{eq:9}, matrix $\Ab_{\GStOp} =I- \sum_{\beta=1}^{\beta_*} (\GSol L_{\beta} \otimes \GSol L_{\beta} + \GSol L_{\beta}^{t} \otimes \GSol L_{\beta}^{t})$ representing the stability operator $\GStOp$ in the standard basis of $\CC^{(2\beta_*)^2}$ can be written as \begin{align*} \Ab_{\mathscr{L}} = I_{(2 \beta_*)^2 } & - (m_1 E_{1,2\beta_*})^{\otimes 2} - \sum_{\beta=2}^{\beta_*} (m_\beta E_{\beta,2\beta_*+1-\beta} + m_{\beta_*+1} E_{2\beta_*+2-\beta,2\beta_*+1-\beta})^{\otimes 2} \\ & - \sum_{\beta=1}^{\beta_*-1} (m_{2\beta_*+1-\beta} E_{2\beta_*+1-\beta,\beta} + m_{\beta_*+1} E_{\beta+1,\beta})^{\otimes 2} - (m_{\beta_*+1} E_{\beta_*+1,\beta_*})^{\otimes 2} , \end{align*} where for any $k\in \NN$ and $R \in \CC^{k\times k}$ we denote $R^{\otimes 2}:= R\otimes R$. After removing from $\Ab_{\GStOp}$ rows and columns for all indices such that either row or column of the corresponding index has only one non-zero entry equal to $1$, we obtain that $\det(\Ab_{\mathscr{L}}) = \det(\widehat{\Ab}_{\mathscr{L}})$, where \begin{equation*} \widehat{\Ab}_{\mathscr{L}} = -I_{2\beta_*} + \sum_{\beta=1}^{2\beta_*} m_{\beta}^{2} E_{\beta,2\beta_*+1-\beta} + m_{\beta_*+1}^2 \sum_{\beta=1}^{\beta_*-1}(E_{\beta+1, \beta } + E_{\beta_*+\beta+1, \beta_*+\beta}) . \end{equation*} Divide $\widehat{\Ab}_{\mathscr{L}}$ into four blocks of equal size $\Ab_{ij}, 1\leq i,j\leq 2$, so that, e.g., $\Ab_{11}$ denotes the upper-left $\beta_*\times \beta_*$ submatrix of $\Ab$. Then by the lower-triangular structure of $\Ab_{11}$ and $\Ab_{22}$ having $-I_{\beta_*}$ on their diagonals, skew-diagonal shape of $\Ab_{12}$ and $\Ab_{21}$, and \eqref{eq:3} we have that $\det (\widehat{\Ab}_{\mathscr{L}})$ is equal to \begin{equation*} \det (\Ab_{11}-\Ab_{12} \Ab_{22}^{-1} \Ab_{21}) = \det \left( \begin{array}{cccccc} \upsilon-1&\upsilon& \upsilon & \upsilon&\cdots &\upsilon \\ \omega&\upsilon-1&\upsilon&\upsilon &\cdots & \upsilon \\ & &\ddots& & &\vdots \\ &&\omega&\upsilon-1&\upsilon&\upsilon \\ & & &\omega&\upsilon-1&\upsilon \\ & & &&\omega &\upsilon-1 \end{array} \right) \end{equation*} with $\upsilon = \zeta^{2\beta_*} m_1^{2(\beta_*+1)}$ and $\omega=(\zeta m_1)^2$. One can easily see that the above determinant is equal to the determinant of the following tridiagonal matrix \begin{equation*} \left( \begin{array}{cccccc} \upsilon-1&1 &&&& \\ \omega &\upsilon-1-\omega& 1 &&& \\ & \omega &\upsilon-1-\omega& 1&& \\ &&\ddots&\ddots&\ddots& \\ && & \omega &\upsilon-1- \omega&1 \\ && & & \omega &\upsilon-1 -\omega \end{array} \right) , \end{equation*} that is equal to \begin{equation} \label{eq:4} (\upsilon-1) \det(\TT_{\beta_*-1}(\upsilon-1-\omega,1,\omega )) - \omega \det(\TT_{\beta_*-2}(\upsilon-1-\omega, 1, \omega )), \end{equation} where $\TT_{k}(a,b,c)$ denotes a $k\times k$ Toeplitz tridiagonal matrix with $a$ on the main diagonal, and $b$ and $c$ above and below the main diagonal respectively. From \eqref{eq:11} we have \begin{equation} \label{eq:12} \upsilon-1-\omega = -2\zeta m_1 , \end{equation} and thus $(\upsilon-1-\omega)^2=4\omega $. Note, that under the condition $a^2=4bc$ the determinant of the Toeplitz tridiagonal matrix takes a particularly simple form \begin{equation*} \det(\TT_{k}(a,b,c )) = (k+1) \left(\frac{a}{2}\right)^k . \end{equation*} A simple calculation from \eqref{eq:4} and \eqref{eq:12} gives that \begin{equation*} \det( \Ab_{\mathscr{L}}) = (\beta_*+1) (-\zeta m_1)^{\beta_*} + (\zeta m_1)^2 \beta_* (-\zeta m_1)^{\beta_*-1} . \end{equation*} Hence, $\det( \Ab_{\mathscr{L}})=0$ implies that (since $\zeta m_1 \neq 0$ in the $\epsB$-bulk) \begin{equation}\label{eq:5} \zeta m_1 = \frac{\beta_*+1}{\beta_*} . \end{equation} Now if we plug \eqref{eq:5} into \eqref{eq:3} we obtain \begin{equation*} m_1 = \Big(\frac{\beta_*}{\beta_*+1}\Big)^{\beta_*}\frac{1}{\beta_*} ,\qquad \zeta = \frac{(\beta_*+1)^{\beta_*+1}}{{\beta_*}^{\beta_*}} ,\qquad z = 1-\frac{(\beta_*+1)^{\beta_*+1}}{{\beta_*}^{\beta_*}} . \end{equation*} Since at $z=1-(\beta_*+1)^{\beta_*+1}/{\beta_*}^{\beta_*}$ the imaginary part of $m_1(z)$ vanishes, this point does not belong to the $\epsB$-bulk. This argument, which can be made effective, yields that \textup{(\textbf{M2})} holds.

https://arxiv.org/abs/1903.02103

Extremes of Chi triangular array from the Gaussian $β$-Ensemble at high temperature

We study the extreme point process associated to the off-diagonal components in the matrix representation of the Gaussian $\beta$-Ensemble and prove its convergence to Poisson point process as $n\to +\infty$ when the inverse temperature $\beta$ scales with $n$ and tends to $0$. We consider two main high temperature regimes: $\displaystyle{\beta\ll \frac{1}{n}}$ and $\displaystyle{n\beta= 2\gamma \geq 0}$. The normalizing sequences are explicitly given in each cases. As a consequence, we estimate the first order asymptotic of the largest eigenvalue of the Gaussian $\beta$-Ensemble.

\section{Introduction} Gaussian ensembles, namely Hermitian matrices with independent Gaussian entries whose joint distribution invariant is under conjugation by appropriate unitary matrices, play a central role in the realm of Random Matrix Theory. Regarding their spectrum, the joint distribution of the eigenvalues is available through the formula: \begin{align}\label{def_beta_ens} P_{n,\beta}(d\lambda_1,...,d\lambda_n):= \frac{1}{Z_{n,\beta}} \exp\left( -\frac{\beta}{4} \sum_{i=1}^{n}\lambda^2_i\right) \prod_{i<j}^{n}\left| \lambda_j-\lambda_i\right| ^\beta \prod_{i=1}^{n}\mathrm{d}\lambda_i. \end{align} According to the Dyson index $\beta$, the density (\ref{def_beta_ens}) describes the eigenvalues of the Gaussian Orthogonal Ensemble $(\beta=1)$, the Gaussian Unitary Ensemble $(\beta=2)$, or the Gaussian Symplectic Ensemble $(\beta=4)$. Although it seemed that no natural matrix model could be associated to the Gaussian $\beta$-Ensemble, a collection of $n$ points with distribution (\ref{def_beta_ens}) and general $\beta>0$, the gap is bridged many years later by Edelman and Dumitriu in \cite{MatrixModelBetaEnsemble}. Through the Househölder method, they constructed the following tridiagonal symmetric random matrix: \begin{align}\label{H} H_{n,\beta}:=\frac{1}{\sqrt{\beta}} \begin{pmatrix} \mathcal{N}(0,2) &\chi\left( \left( n-1\right) \beta\right) & && \\ \chi\left( \left( n-1\right) \beta\right) &\mathcal{N}(0,2) &\chi\left( \left( n-2\right) \beta\right) &&\\ &\chi\left( \left( n-2\right) \beta\right) &\mathcal{N}(0,2) & \chi\left( \left( n-3\right) \beta\right) & \\ & & \ddots &\ddots & \ddots & \\ & &\qquad \ddots & \qquad \ddots &\chi\left( \beta\right) & \\ & & &\chi\left( \beta\right) & \mathcal{N}(0,2) \end{pmatrix} \end{align} with independence between all variables, such that for any $n\geq 1$ and $\beta> 0$, the eigenvalues $(\lambda_1,...,\lambda_n)$ of $H_{n,\beta}$ have exactly joint distribution $P_{n,\beta}$ from (\ref{def_beta_ens}). Let us consider the off-diagonal sequence $\left( X_{i,n}\right)_{i\leq n} $ of $H_{n,\beta}$, namely $\left( X_{i,n}\right) $ are independent random variables with distribution $\displaystyle{X_{i,n}\sim \chi\left( i\beta\right) }$ for $\displaystyle{i\leq n}$. The scope of this paper is to study the asymptotic behavior of the largest $X_{i,n}$, namely of the extreme point process $\displaystyle{\sum_{i=1}^{n}\delta_{a_n\left( X_{i,n}-b_n\right)}}$ where $(a_n)$, $(b_n)$ are suitably chosen normalizing sequences and when the inverse temperature $\beta:=\beta_n$ scales with $n$ in such a way that $\beta \xrightarrow[n\infty]{} 0$. We consider two regimes in this context. In first place, we study the case $\displaystyle{n^{-2}\ll \beta \ll {n}^{-1}}$ which itself divides into two subregimes around $\displaystyle{\left( n\log\log n\right)^{-1} }$. Then, the intermediate high temperature regime $\displaystyle{n\beta =2\gamma}$ is analyzed for $\gamma \in (0,1)$. Both cases lead to an inhomogeneous limiting Poisson point process on the half line $\mathbb{R}^+$ with intensity measure of exponential kind. The point process approach is borrowed from the Extreme Value Theory (see \cite{Leadbetter,Resnick}) as it shows to be a convenient way to encompass many informations. In this vein, our results are stated in terms of Poisson convergence. One consequence of this fact is that one can derive the limiting distribution of the rescaled largest component to be Gumbel. We exploit here the quantitative growth rate of the largest entries in (\ref{H}) to provide an estimate on the first order asymptotic of the largest eigenvalue in the high temperature regime. The question of the normalizing (or scaling) sequences is fundamental. They indicate the precise growth rate of the maximum of the set of random variables considered, which, roughly speaking, provides the right way to look at the extreme values. Although they are not always explicit or with closed form, we furnish here exact formulae. The case $\displaystyle{n\beta =2\gamma}$ has a nice interpretation: the scaling sequences for $\displaystyle{\max_{1\leq i \leq n}X_{i,n}}$ thoroughly match with those of the maximum of $\displaystyle{\frac{n}{\gamma \log\log(n)}}$ i.i.d. copies of $X_{n,n}\sim \chi(2\gamma)$. We now state our main result in the case $\displaystyle{n^{-2}\ll \beta \ll {n}^{-1}}$: \begin{Th} \label{Main_th} Let $\beta := \beta_n \ll 1$ such that $\displaystyle{\frac{1}{n^2}\ll \beta \ll \frac{1}{n } }$ and $\displaystyle{X_{i,n}\sim \chi \left( i\beta\right) }$, with $1\leq i \leq n$, a triangular array of independent random variables. Assume $\displaystyle{\frac{1}{n^2}\ll \beta \ll \frac{1}{n\log \log \left( n\right) } }$. Let the scaling sequences: \begin{align}\label{defTH_scaling_1} a_n={\sqrt{2\log \left( n^2 \beta \right) }},\qquad b_n=\sqrt{2\log \left( n^2 \beta \right) } - \frac{\log \log\left( n^2 \beta\right) }{\sqrt{2\log \left( n^2 \beta \right) }}. \end{align} Then the point process $\displaystyle{\sum_{i=1}^{n}\delta_{a_n\left( X_{i,n}-b_n\right)}}$ converges weakly to an inhomogeneous Poisson point process on $\mathbb{R}^+$ with intensity measure $\displaystyle{\frac{e^{-x}}{4}\mathrm{d}x}$. Assume $\displaystyle{ \frac{1}{n\log \log \left( n\right) } \ll \beta \ll \frac{1}{n} }$. Let the scaling sequences: \begin{align}\label{defTH_scaling_2}a_n=\sqrt{2\log \left( n \right) },\qquad b_n=a_n - \frac{n\beta \log \log\left( n\right) }{\sqrt{2\log \left( n \right) }}-\frac{\log \log \left( n\right) }{\sqrt{2\log \left( n \right) }}-\frac{\log \log \log \left( n\right) }{\sqrt{2\log \left( n \right) }}.\end{align}Then the point process $\displaystyle{\sum_{i=1}^{n}\delta_{a_n\left( X_{i,n}-b_n\right)}}$ converges weakly to an inhomogeneous Poisson point process on $\mathbb{R}^+$ with intensity measure $\displaystyle{e^{-x}\mathrm{d}x}$. \end{Th} In the intermediate high regime $\displaystyle{n\beta=2\gamma}$, we have the following result: \begin{Th} \label{Main_th2} Let $\displaystyle{\gamma \in (0,1)}$ and $\displaystyle{ \beta_n = {2\gamma}n^{-1}}$. Let $\displaystyle{X_{i,n}\sim \chi \left( i\beta_n\right) }$, with $1\leq i \leq n$, a triangular array of independent random variables. Let the scaling sequences: \begin{align}\label{defTH_scaling_3} a_n:={\sqrt{2\log ( \frac{n}{\gamma\log \log(n)})}},\quad b_n:=a_n + \frac{ \left( \frac{\gamma}{2}-1\right) \log\log(\frac{n}{\gamma\log \log(n)})-\log\left( 2^{-\frac{\gamma}{2}}\Gamma(\gamma)\right) }{\sqrt{2\log ( \frac{n}{\gamma\log \log(n)})}}. \end{align} Then the point process $\displaystyle{\sum_{i=1}^{n}\delta_{a_n\left( X_{i,n}-b_n\right)}}$ converges weakly to an inhomogeneous Poisson point process on $\mathbb{R}^+$ with intensity measure $\displaystyle{{e^{-x}}{}\mathrm{d}x}$. \end{Th} \begin{rmk} On one hand, we consider a size sample of $(X_{i,n})$ growing in $n$ and on the other hand, the greatest chi parameter of $(X_{i,n})$ is $n\beta$. Assuming that $n\beta\ll 1$, one could expect that the maximum of the $(X_{i,n})$ should converge to $0$. Our result shows the other way. In the regime $\displaystyle{n^{-2}\ll \beta\ll n^{-1}}$, the centering $(b_n)$ is still diverging to infinity. Though, one can retrain our work to show it is converging to $0$ when $\displaystyle{\beta\ll n^{-2}}$. We did not consider this special case here since the competition against the maxium of Gaussian variables becomes irrelevant. \end{rmk} Concerning the edge statistics of $H_{n,\beta}$, we derive the following corollary: \begin{cor} Let $\lambda_{\max}\left( H\right) $ the largest eigenvalue of $\displaystyle{H:=\sqrt{\beta}H_{n,\beta}}$ defined in (\ref{H}). Assume the hypothesis of Theorem \ref{Main_th} or \ref{Main_th2}. There exists constants $0<c<c'<+\infty$ such that with probability tending to $1$, as $n\to +\infty$, $$\frac{\lambda_{\max}\left(H \right)}{\sqrt{\log(n)}} \in [c,c']. $$ \end{cor} \begin{proof} Consider the standard basis $e_i:=(\delta_{i,j})$ of $\mathbb{R}^n$. The characterization of the largest eigenvalue with the Rayleigh quotient (see Theorem 4.2.2 in \cite{Horn}) allows us to write: \begin{align}\label{corollary_lowerbound} \lambda_{\max}\left( H \right) = \max_{x\neq 0} \frac{x^t Hx}{x^t x}\geq \frac{e_i^t He_i}{e_i^t e_i} =H\left( {i,i}\right), \quad 1\leq i\leq n . \end{align} Thus we get the lower bound $\displaystyle{\lambda_{\max}\left( H\right)\geq \max_{1\leq i \leq n}H\left(i,i \right) }$. On the other hand, we apply the Gerschgorin circles (see Theorem 6.1.1 in \cite{Horn}) to get an upperbound: \begin{align}\label{corollary_upperbound} \lambda_{\max}\left( H\right)\leq \max_{1\leq i \leq n}H\left(i,i \right) + 2\max_{1\leq i \leq n-1}H\left(i,i+1 \right). \end{align} The claim yields by combining both bounds (\ref{corollary_lowerbound}), (\ref{corollary_upperbound}) with the conclusions of Theorem \ref{Main_th} or \ref{Main_th2}, and the counterpart result for maximum of i.i.d. Gaussian variables, that is the point process $\displaystyle{\sum_{i=1}^{n-1}\delta_{a_n\left(H\left(i,i+1 \right)-b_n\right)}}$ converges weakly in probability to a Poisson point process, where the normalizing sequences are: $$b_n = 2\sqrt{\log(n)}-\frac{\log\log(n)+\log(4\pi)}{2\sqrt{\log(n)}},\quad a_n=2\sqrt{\log(n)}.$$ \end{proof} In Section \ref{section_Multivariate}, we establish a general tool for asymptotic Poissonian statistics of extreme point processes in the independent but non identically distributed context. Theorem \ref{Th_tool} will be the framework of our method and is the analog of the classic Poisson limit theorem for i.i.d. extremes (see for instance Proposition 3.21 in \cite{Resnick}). Sections \ref{section_proof1} and \ref{section_proof2} are devoted to the proof of each case following the steps accordingly to Theorem \ref{Th_tool}, namely we prove appropriate convergence and uniform negligibility. \section{Multivariate Poisson limit}\label{section_Multivariate} For completness, we recall the definition of a Poisson point process. \begin{Def}\label{def_ppp} Let $X$ a topological space endowed with a $\sigma$-finite measure $\mu$. A Poisson point process on $X$ with intensity $\mu$ is a random Radon measure $N$ on $X$ such that for any $k\geq 1$ and $B_1,\ldots,B_k$ pairwise disjoint Borel sets of $X$ with finite $\mu$ measure, the random vector $\left( N\left( B_1\right) ,\ldots, N\left( B_k\right) \right) $ has law $\displaystyle{\op{Po}\left( \mu(B_1\right)\otimes \cdots \otimes \op{Po}\left( \mu(B_k\right) }$. \end{Def} In order to reach the point process level, we need the following finite-dimensional Poisson approximation: \begin{Prop}\label{multivariate_poisson_limit_prop} Let $(X_{i,n})$ a triangular array of independent random variables with same interval support $\mathcal{S}$. Let $I\subset \mathbb{R}$ an interval and $\displaystyle{\varphi_n:I\to \mathcal{S}}$ an increasing homeomorphism. Let $k\geq 1$ and $ x_0<\ldots <x_k\in I$ and the random vector $(N_0,\ldots, N_k)$ defined for $ \ell \leq k$ by: $$N_\ell=\#\{ i=1, \ldots, n\, ;\, X_{i,n}\in (\varphi_n(x_{\ell-1}),\varphi_n(x_{\ell})]\}= \sum_{i=1}^n \delta_{\{ X_{i,n}\in (\varphi_n(x_{\ell-1}),\varphi_n(x_{\ell})] \}}.$$ For $1\leq \ell \leq k<+\infty$, we suppose that: \begin{align}\label{condition_for_poisson1} \sum_{i=1}^n\mathbb{P}\left( X_{i,n}\in (\varphi_n(x_{\ell-1}),\varphi_n(x_{\ell})]\right) \xrightarrow[n\infty]{} \lambda_\ell, \end{align} \begin{align}\label{condition_for_poisson2} \max_{1\leq \ell \leq k} \max_{i\leq n}\left( \mathbb{P}\left(X_{i,n}\in (\varphi_n(x_{\ell-1}),\varphi_n(x_{\ell})] \right) \right) \xrightarrow[n\infty]{}0. \end{align} Then we have the convergence in distribution: $$(N_1, \ldots N_k)\underset{n\to\infty}{\longrightarrow} \op{Po}(\lambda_1)\otimes\cdots\otimes \op{Po}(\lambda_k).$$ \end{Prop} \begin{proof} For convenience, we set $\displaystyle{p^{(n)}_{i,\ell}:=\mathbb{P}\left( X_{i,n}\in (\varphi_n(x_{\ell-1}),\varphi_n(x_{\ell})] \right) }$. Let $t=(t_1,..,t_k) \in \mathbb{R}^k$, $N = (N_1,..,N_k)$. We compute the Laplace transform of $N$: \begin{align*} \op{\mathbb{E}} \big( e^{-<t,N>} \big) &=\op{\mathbb{E}} \big( e^{-\sum_{\ell=1}^k \sum_{i=1}^n t_\ell 1_{\{ X_{i,n}\in (\varphi_n(x_{\ell-1}),\varphi_n(x_{\ell})]\} }} \big) \\ &= \prod_{i=1}^n \op{\mathbb{E}} \big( e^{-\sum_{\ell=1}^k t_\ell 1_{\{ X_{i,n}\in (\varphi_n(x_{\ell-1}),\varphi_n(x_{\ell})]\} }} \big) \qquad \text{ by independence} \\ &= \prod_{i=1}^n \Big( \sum_{\ell=0}^k p^{(n)}_{\ell,i} e^{-t_{\ell}}\Big) \qquad \text{ by transfer and with $t_0 = 0$} \\ &=\prod_{i=1}^n \Big(1+ \sum_{\ell=1}^k p^{(n)}_{\ell,i}( e^{-t_{\ell}}-1)\Big) \\ &= \prod_{i=1}^n \Big(1+ a_{i}^{(n)} \Big) \qquad \text{ with $a_{i}^{(n)} := \sum_{\ell=1}^k p^{(n)}_{\ell,i}( e^{-t_{\ell}}-1)$}. \end{align*} Taking the logarithm and Taylor-expanding, \begin{align*} \log \op{\mathbb{E}} \big( e^{-<t,N>} \big) &= \sum_{i=1}^n \log \Big(1+ \sum_{\ell=1}^k p_{\ell,i}^{(n)}( e^{-t_{\ell}}-1)\Big) \\ &= \sum_{i=1}^n \log \Big(1+ a_{i}^{(n)}\Big) \\ &=\sum_{i=1}^n \Big( a_{i}^{(n)} - \frac{(a_{i}^{(n)})^2}{2(1+c)^2} \Big) \qquad \text{ with $0<c=c_{i,n}<a_{i}^{(n)}$} \\ &= \sum_{i=1}^n a_{i}^{(n)} - \sum_{i=1}^n \frac{(a_{i}^{(n)})^2}{2(1+c)^2} \end{align*} From hypothesis (\ref{condition_for_poisson1}) and (\ref{condition_for_poisson2}), we deduce: $$\displaystyle{\sum_{i=1}^n a_{i}^{(n)} \xrightarrow[n\infty]{} \sum_{\ell=1}^k \lambda_\ell (e^{-t_\ell}-1)},$$ $$\displaystyle{0 \leq \sum_{i=1}^n\frac{(a_{i}^{(n)})^2}{2(1+c)^2} \leq \sum_{i=1}^n (a_{i}^{(n)})^2 \leq \big( \max_{i \leq n} a_{i}^{(n)} \big)\Big( \sum_{i=1}^n a_{i}^{(n)} \Big) \xrightarrow[n\infty]{} 0. }$$ It follows that: \begin{align*} \log \op{\mathbb{E}} \big( e^{-<t,N>} \big) &= \sum_{i=1}^n a_{i}^{(n)} + o(1) \xrightarrow[n\infty]{} \sum_{\ell=1}^k \lambda_\ell (e^{-t_\ell}-1). \end{align*} \end{proof} We are now ready to show the goal of this section: \begin{Th}\label{Th_tool} Let $(X_{i,n})$ a triangular array of independent random variables with same interval support $\mathcal{S}$. Assume there exists $I\subset \mathbb{R}$ interval, $\displaystyle{G:I\to \mathbb{R}^+}$ c\`{a}dl\`{a}g with finite limit at the right boundary of $I$ and $\displaystyle{\varphi_n:I\to \mathcal{S}}$ increasing homeomorphism such that: \begin{align}\label{condition_for_poisson3} \forall x\in I,\qquad \sum_{i=1}^n \mathbb{P}\left( X_{i,n}\geq \varphi_n\left( x\right) \right) \xrightarrow[n\infty]{} G(x),\end{align} and for $x,y\in I$ such that $y<x<+\infty$, \begin{align}\label{condition_for_poisson4} \max_{i \leq n} \mathbb{P}\left( \varphi_n\left(y\right) < X_{i,n} \leq \varphi_n\left( x\right) \right) \xrightarrow[n\infty]{} 0. \end{align} Then the point process $$\sum_{i=1}^n \delta_{\varphi^{-1}_n(X_{i,n})}$$ converges to a Poisson point process on the interval $I$ with intensity $\mu$ defined by $$\displaystyle{\mu\left((x,y] \right):=G(x)-G(y), \qquad x<y .}$$ \end{Th} \begin{proof} Let $\displaystyle{N_n:=\sum_{i=1}^n \delta_{\varphi^{-1}_n(X_{i,n})}}$. By Definition \ref{def_ppp}, we need to show the following convergence in distribution, for any $k\geq 1$ and $x_0<\ldots<x_k\in I$, $$\left( N_n (x_0,x_1] , \ldots N_n(x_{k-1},x_k]\right) \xrightarrow[n\infty]{} \op{Po}(\mu(x_0,x_1])\otimes\cdots\otimes \op{Po}(\mu(x_{k-1},x_k])$$ where for $\lambda\ge 0$, $$\op{Po}(\lambda):=e^{-\lambda}\sum_{i\ge 0}\frac{\lambda^i}{i!}\delta_i.$$ First thing to notice is the following identity for $1\leq \ell\leq k$, $$\displaystyle{N_n (x_{\ell-1},x_\ell]=\#\{ i=1, \ldots, n\, ;\, X_{i,n}\in (\varphi_n(x_{\ell-1}),\varphi_n(x_{\ell})]\}.}$$ Thus, we aim to apply Proposition \ref{multivariate_poisson_limit_prop} in order to complete the proof. Clearly, the condition (\ref{condition_for_poisson2}) is fulfilled thanks to the uniform negligibility hypothesis (\ref{condition_for_poisson4}). Besides, for any $x\in I$, one has $\displaystyle{\mu\left( \left( x,+\infty\right) \cap I \right) <+\infty}$ because of the hypothesis on $G$. Hence, using (\ref{condition_for_poisson3}), we get: $$\forall x<y\in I,\qquad \sum_{i=1}^n \mathbb{P}\left(\varphi_n(x)< X_{i,n}\leq \varphi_n\left( y\right) \right) \xrightarrow[n\infty]{} G(x) -G(y).$$ The assumption (\ref{condition_for_poisson1}) readily follows with $\displaystyle{\lambda_\ell = \mu\left((x_{\ell-1},x_\ell] \right)=G(x_{\ell-1})-G(x_\ell)}$. \end{proof} \section{Proof of Theorem \ref{Main_th}}\label{section_proof1} Before going in the details of the demonstration, let us fix our notation and recall some well-known facts about Gamma function and Gamma distribution. \begin{rmk}\label{gamma_chi_relation_rmk} A real random variable $X$ follows a $\Gamma(\alpha,\beta)$ distribution, ie: $X\sim \Gamma(\alpha,\beta)$, if its density is $\displaystyle{f_X(x)= \frac{\beta^\alpha x^{\alpha-1}e^{-\beta x}}{\Gamma(\alpha)}}$, $\displaystyle{x\in (0,+\infty)}$. Also, $$ \Gamma(a,z):= \mathbb{P}\left(\Gamma\left( a,1\right) \geq z\right) =\int_z^{+\infty}\frac{ x^{\alpha-1}e^{- x}}{\Gamma(\alpha)}\mathrm{d}x .$$ For $\alpha,\beta,\lambda > 0$ and $k>0$, the following equalities holds in distribution: \begin{align}\label{gamma_chi_relation} \chi(k)= \sqrt{\Gamma(\frac{k}{2},\frac{1}{2})},\qquad \lambda \Gamma(\alpha,\beta)= \Gamma(\alpha,\frac{\beta}{\lambda} ). \end{align} \end{rmk} The main step in our proof is the fulfillment of condition (\ref{condition_for_poisson3}) so we need a way to control summation of incomplete Gamma functions. Indeed, previous Remark \ref{gamma_chi_relation_rmk} indicates how chi's survival functions are related to incomplete Gamma functions. As we have to take limit on a sum, we require to work on a finite setting ($n<+\infty$) and estimate each summand by the crucial bound: \begin{lem}\label{bounds_incomplete_lem} For $a<1$ and $z>0$, \begin{align}\label{bounds_incomplete_eq} \frac{z}{z+1-a}\frac{e^{-z}z^{a-1}}{\Gamma(a)} < \frac{\Gamma(a,z)}{\Gamma(a)}\leq \frac{e^{-z}z^{a-1}}{\Gamma(a)}. \end{align} \end{lem} \begin{proof} The upper bound is classic and can be found in \cite{Borwein_Chan}. One can use Pad\'e approximants to prove the lower bound of (\ref{bounds_incomplete_eq}). As a reference for this, we give \cite{nist}. \end{proof} We state two well-known facts which are going to be helpful to accuratly evaluate the whole sum. \begin{rmk} It is well-known that the Gamma function admits the Taylor expansion near $u=0$: $$\Gamma(u)=\frac{1}{u}+\gamma +\frac{6\gamma^2+\pi^2}{12} u+O(u^2).$$ This combines well with the hypothesis $n\beta \ll 1$ since it implies $\displaystyle{i\beta\xrightarrow[n\infty]{}0}$ uniformly in $i\leq n$, ie: $\displaystyle{\max_{i\leq n}\left(i\beta \right)\xrightarrow[n\infty]{}0 }$. Thus, \begin{align}\label{Gamma_near_0} \Gamma\left( \frac{i\beta}{2} \right) &= \frac{2}{i\beta}+\gamma_{\text{Euler}} + O\left(n\beta\right). \end{align} For $u \in \mathbb{R}$, let us give another useful identity: \begin{align}\label{sum} \sum_{k=1}^{n}ku^k &= \frac{(nu-n-1)u^{n+1}+u}{(1-u)^2}. \end{align} \end{rmk} Lastly, we show a technical result which estimates an auxiliary term in the upcoming computations. Its outcome varies depending to $\beta$ regime and accurately shows the differences between the two regimes. \begin{lem}\label{lambda_estimate} Let $z:=z_n\gg 1$ such that $\displaystyle{z\ll \log^\delta(n)} $ for some $\delta>0$ and $\displaystyle{\beta \ll \frac{1}{n}}$. We define: \begin{align*} \Lambda_n(z) &:= n\left(\left( \frac{z^2}{2}\right)^\frac{\beta}{2} -1\right) \left( \frac{z^2}{2}\right)^{\frac{n\beta}{2}} -\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} +1 . \end{align*} Assume $\displaystyle{\frac{1}{n^2}\ll \beta \ll \frac{1}{n\log \log \left( n\right) } }$. Then, \begin{align}\label{lambda_first_case}\Lambda_n(z)&= \frac{1}{2}\left( \frac{n\beta}{2}\log \left( \frac{z^2}{2} \right) \right)^2 \left(1+o(1) \right). \end{align} Assume $\displaystyle{\frac{1}{n\log\log \left( n\right) }\ll \beta \ll \frac{1}{n} }$. Then, \begin{align}\label{lambda_second_case}\Lambda_n(z)&= \left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} \left( \frac{n\beta}{2} \log \left( \frac{z^2}{2}\right)\right) \left( 1+o(1)\right) . \end{align} \end{lem} \begin{proof}We begin with the first case $\displaystyle{n^{-2}\ll \beta \ll \left( n\log \log n\right)^{-1} }$. We treat each three main terms (except $n$) by second order Taylor expansion, so that $\Lambda_n(z)$ equals to: \begin{align*} &\left(\frac{n\beta}{2} \log \left( \frac{z^2}{2}\right)+\frac{n\left(\frac{\beta}{2} \log \left( \frac{z^2}{2}\right)\right)^2 }{2} +nO\left( \beta^3 \log^3(z) \right) \right) \times \\&\qquad \times \left(1+ \frac{n\beta}{2} \log \left( \frac{z^2}{2}\right) + \frac{\left( \frac{n\beta}{2} \log \left( \frac{z^2}{2}\right)\right)^2 }{2} + O\left( ( n\beta)^3 \log^3 \left( z\right)\right) \right) \\&\quad -\frac{n\beta}{2}\log \left( \frac{z^2}{2} \right) - \frac{\left( \frac{n\beta}{2}\log \left( \frac{z^2}{2} \right) \right)^2 }{2}+O\left( n^3\beta^3\log^3 \left( z \right) \right) . \end{align*} Expanding the product, we get: \begin{align*} \Lambda_n(z)&=\frac{n\beta}{2} \log \left( \frac{z^2}{2}\right)+ \frac{n\left(\frac{\beta}{2} \log \left( \frac{z^2}{2}\right)\right)^2 }{2} +nO\left( \beta^3 \log^3(z) \right) \\&+\left( \frac{n\beta}{2} \log \left( \frac{z^2}{2}\right)\right)^2 + \frac{n^2\left(\frac{\beta}{2} \right) ^3 \log^3 \left( \frac{z^2}{2}\right)}{2}+n^2 \beta \log(z)O\left( \beta^3 \log^3(z) \right) \\&-\frac{n\beta}{2}\log \left( \frac{z^2}{2} \right) - \frac{\left( \frac{n\beta}{2}\log \left( \frac{z^2}{2} \right) \right)^2 }{2}+O\left( n^3\beta^3\log^3 \left( z \right)\right) . \end{align*} The first order terms vanishes. Cleaning the highest order quantities, it leads to: $$\frac{n\left(\frac{\beta}{2} \log \left( \frac{z^2}{2}\right)\right)^2 }{2} + \frac{\left( \frac{n\beta}{2}\log \left( \frac{z^2}{2} \right) \right)^2 }{2}+O\left( n^3\beta^3\log^3 \left(z \right) \right) . $$ Thus, the result readily follows from the assumptions made on $z$ and $\beta$. Now, we assume $\displaystyle{\frac{1}{n\log\log \left( n\right) }\ll \beta \ll \frac{1}{n} }$. Likewise, \begin{align*} \Lambda_n(z) &:= n\left(\left( \frac{z^2}{2}\right)^\frac{\beta}{2} -1\right) \left( \frac{z^2}{2}\right)^{\frac{n\beta}{2}} -\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} +1 \\&= \left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} \left( n\left(\left(\frac{z^2}{2} \right)^{\frac{\beta}{2}} -1\right) -1+\frac{1}{\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}}}\right) \\&= \left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} \left( n\left(\frac{\beta}{2} \log \left( \frac{z^2}{2}\right) +O\left(\beta^2 \log^2(z) \right) \right) -1+\frac{1}{\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}}}\right) \\&=\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} \left( \frac{n\beta}{2} \log \left( \frac{z^2}{2}\right) +nO\left(\beta^2 \log^2(z) \right) -1+\frac{1}{\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}}}\right) \\&=\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} \left( \frac{n\beta}{2} \log \left( \frac{z^2}{2}\right)\right) \left( 1+o(1)\right) . \end{align*} \end{proof} We can now turn to the actual proof of Theorem \ref{Main_th} which is in two stages accordingly to conditions (\ref{condition_for_poisson3}) and (\ref{condition_for_poisson4}). We prove the first condition. \begin{Prop}\label{Main_prop} Let $\beta := \beta_n \ll 1$ such that $\displaystyle{n^{-2}\ll \beta \ll n^{-1} }$ and $\left( X_{i,n}\right) $ a triangular array of independent random variables with $\displaystyle{X_{i,n}\sim \chi \left( i\beta\right) }$ for any $1\leq i \leq n$. Assume $\displaystyle{\frac{1}{n^2}\ll \beta \ll \frac{1}{n\log \log \left( n\right) } }$. For $x\in\mathbb{R}^+$, we set: \begin{align}\label{def_scaling_1} \varphi_n(x)&:=\frac{x}{a_n}+b_n= \frac{x}{\sqrt{2\log \left( n^2 \beta \right) }}+\sqrt{2\log \left( n^2 \beta \right) } - \frac{\log \log\left( n^2 \beta\right) }{\sqrt{2\log \left( n^2 \beta \right) }}. \end{align} Then for $0\leq x<+\infty$, $$S_n \left( \varphi_n(x)\right) :=\sum_{i=1}^{n}\mathbb{P}\left( X_{i,n} \geq \varphi_n(x)\right) \xrightarrow[n\infty]{} \frac{e^{-x}}{4}.$$ Assume $\displaystyle{ \frac{1}{n\log \log \left( n\right) } \ll \beta \ll \frac{1}{n} }$. For $x\in\mathbb{R}^+$, we set: \begin{align}\label{def_scaling_2}\varphi_n(x)&:= \frac{x}{\sqrt{2\log \left( n \right) }}+\sqrt{2\log \left( n \right) } - \frac{n\beta \log \log\left( n\right) }{\sqrt{2\log \left( n \right) }}-\frac{\log \log \left( n\right) }{\sqrt{2\log \left( n \right) }}-\frac{\log \log \log \left( n\right) }{\sqrt{2\log \left( n \right) }}.\end{align} Then for $0\leq x<+\infty$, $$S_n \left( \varphi_n(x)\right) :=\sum_{i=1}^{n}\mathbb{P}\left( X_{i,n} \geq \varphi_n(x)\right) \xrightarrow[n\infty]{}e^{-x}.$$ \end{Prop} \begin{proof} For convenience of notation, let $\displaystyle{z:=\varphi_n(x)\gg 1}$. By Remark \ref{gamma_chi_relation_rmk}, we write the quantity of interest in terms of incomplete Gamma functions: \begin{align}\label{aux_equality} S_n(z)&=\sum_{i=1}^{n}\mathbb{P}\left(\chi^2\left( i\beta\right) \geq z^2 \right) =\sum_{i=1}^{n}\mathbb{P}\left(\Gamma\left( \frac{i\beta}{2},1\right) \geq \frac{z^2}{2} \right) = \sum_{i=1}^{n} \frac{\Gamma\left( \frac{i\beta}{2},\frac{z^2}{2}\right) }{\Gamma\left( \frac{i\beta}{2}\right) }. \end{align} By hypothesis $n\beta\ll 1$, we can apply Lemma \ref{bounds_incomplete_lem} to estimate (\ref{aux_equality}): \begin{align*} S_n(z)&\leq \sum_{i=1}^{n} \frac{e^{-\frac{z^2}{2}}z^{i\beta -2}}{\Gamma\left(\frac{i\beta}{2}\right) 2^{\frac{i\beta}{2}-1} } = \frac{2e^{-\frac{z^2}{2}}}{z^2}\sum_{i=1}^{n}\frac{\left( \frac{z^2}{2}\right)^{\frac{i\beta}{2}} }{\Gamma\left( \frac{i\beta}{2}\right) }. \end{align*} Applying the Gamma expansion (\ref{Gamma_near_0}) gives: \begin{align}\label{aux_upperbound2} S_n(z)&\leq \frac{e^{-\frac{z^2}{2}}}{z^2}\sum_{i=1}^{n}\frac{i\beta \left( \frac{z^2}{2}\right)^{\frac{i\beta}{2}} }{1+\frac{i\beta}{2}\gamma_{\text{Euler}} + \frac{i\beta}{2}O\left( n\beta\right) } = \frac{e^{-\frac{z^2}{2}}}{z^2}\frac{\beta}{1+O\left( n\beta\right) }\sum_{i=1}^{n}i\left( \left( \frac{z^2}{2}\right)^{\frac{\beta}{2}} \right)^i . \end{align} We analyze the last term with the identity (\ref{sum}): \begin{align*} \sum_{i=1}^{n}i\left( \left( \frac{z^2}{2}\right)^{\frac{\beta}{2}} \right)^i&=\frac{ \left( n\left( \frac{z^2}{2}\right) ^\frac{\beta}{2} -n-1\right) \left( \frac{z^2}{2}\right)^{\frac{n\beta}{2} + \frac{\beta}{2} } + \left( \frac{z^2}{2}\right)^\frac{\beta}{2} }{\left( 1-\left( \frac{z^2}{2}\right)^\frac{\beta}{2} \right)^2 } \\&= \frac{1}{\left( 1-\left( \frac{z^2}{2}\right)^\frac{\beta}{2} \right)^2 }\left( \frac{z^2}{2}\right)^\frac{\beta}{2} \left( \left( n\left( \frac{z^2}{2}\right) ^\frac{\beta}{2} -n-1\right) \left( \frac{z^2}{2}\right)^{\frac{n\beta}{2} } + 1 \right) \\&= \frac{1}{\left( 1-\left( \frac{z^2}{2}\right)^\frac{\beta}{2} \right)^2 }\left( \frac{z^2}{2}\right)^\frac{\beta}{2} \left( n\left(\left( \frac{z^2}{2}\right)^\frac{\beta}{2} -1\right) \left( \frac{z^2}{2}\right)^{\frac{n\beta}{2}} + 1-\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} \right) \\&= \frac{ n\left(\left( \frac{z^2}{2}\right)^\frac{\beta}{2} -1\right) \left( \frac{z^2}{2}\right)^{\frac{n\beta}{2}} -\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} +1 }{\left( 1-\left( \frac{z^2}{2}\right)^\frac{\beta}{2} \right)^2 }\left( \frac{z^2}{2}\right)^\frac{\beta}{2} . \end{align*} The decay hypothesis on $\beta$ and definitions (\ref{def_scaling_1}), (\ref{def_scaling_2}) imply that $\displaystyle{\beta\log(z)\ll 1}$ and hence, $$\frac{1}{\left( 1-\left( \frac{z^2}{2}\right)^\frac{\beta}{2} \right)^2 }=\frac{1}{\left( \frac{\beta}{2} \log \left(\frac{z^2}{2} \right) \right)^2}\left( 1+o(1)\right), \qquad\left( \frac{z^2}{2}\right)^\frac{\beta}{2}=1+o(1) .$$ Combining everything in (\ref{aux_upperbound2}), we get: $$S_n(z)\leq \frac{4e^{-\frac{z^2}{2}}}{z^2} \frac{\Lambda_n(z)}{\beta \log^2(\frac{z^2}{2})}\left( 1+o(1)\right) ,$$ with $$ \Lambda_n(z) :=n\left(\left( \frac{z^2}{2}\right)^\frac{\beta}{2} -1\right) \left( \frac{z^2}{2}\right)^{\frac{n\beta}{2}} -\left(\frac{z^2}{2} \right)^{\frac{n\beta}{2}} +1 . $$ We turn to the lower bound. For any $\displaystyle{a\in [0,1]}$ and $z\gg 1$, we have: \begin{align}\label{rmk_on_lowerbound} \frac{z}{z+1-a}&\geq \frac{z}{z+1} = 1+o(1). \end{align} Applying the lower bound (\ref{bounds_incomplete_eq}) on (\ref{aux_equality}) and according to (\ref{rmk_on_lowerbound}), we get: \begin{align*} S_n(z)&\geq \sum_{i=1}^{n} \frac{z}{z+1-\frac{i\beta}{2}}\frac{e^{-\frac{z^2}{2}}z^{i\beta -2}}{\Gamma\left(\frac{i\beta}{2}\right) 2^{\frac{i\beta}{2}-1} } \\&=\left(1+o(1) \right) \frac{2e^{-\frac{z^2}{2}}}{z^2}\sum_{i=1}^{n}\frac{\left( \frac{z^2}{2}\right)^{\frac{i\beta}{2}} }{\Gamma\left( \frac{i\beta}{2}\right) }. \end{align*} Repeating the same steps as before, it yields: \begin{align}\label{intermediate_equivalent} S_n(z)&= \frac{4e^{-\frac{z^2}{2}}}{z^2} \frac{\Lambda_n(z)}{\beta \log^2(\frac{z^2}{2})}\left( 1+o(1)\right) . \end{align} Note that we have not made any additional assumptions on $\displaystyle{\beta\ll n^{-1}}$ yet. The next step is to estimate the $\Lambda_n(z)$ term. According to the $\beta$ regime considered, two different outcomes turn out, which is described in Lemma \ref{lambda_estimate}. We begin with the first case and restrict to $\displaystyle{n\beta \log z \ll 1}$, ie: $\displaystyle{ n\beta \log\log n \ll 1 }$. Combining the asymptotic (\ref{lambda_first_case}) with (\ref{intermediate_equivalent}), it follows that: \begin{align}\label{last_step1} S_n(z)& = \exp\left( -\frac{z^2}{2} +\log(n^2\beta )-2\log (z) -\log 2\right)\left(1+o(1) \right) . \end{align} From the definition (\ref{def_scaling_1}) of the scaling function, we compute the following asymptotics: $$-\frac{\varphi_n(x)^2}{2}= -x-\log \left( n^2 \beta \right) + \log \log (n^2\beta)+o(1)$$ $$ \log \varphi_n(x) =\log \left( \sqrt{2\log \left( n^2 \beta \right) } \left( 1+o(1)\right) \right) = \frac{1}{2} \log \log \left( n^2 \beta \right) +\frac{1}{2}\log 2 +o(1).$$ And finally, substituting in (\ref{last_step1}), the result yields: $$S_n(z) = \frac{e^{-x}}{4}(1+o(1)). $$ We consider the second regime case for $\beta$ and assume in the sequel that $\displaystyle{ \frac{1}{n\log \log \left( n\right) } \ll \beta \ll \frac{1}{n} }$. The main difference lies in Lemma \ref{lambda_estimate}. We substitute the asymptotic (\ref{lambda_second_case}) in (\ref{intermediate_equivalent}): \begin{align}\label{last_step2} S_n(z)&= 4\exp\left( -\frac{z^2}{2} +n\beta \log(z)-2\log (z) - \log\log(z) +\log (n)\right)\left(1+o(1) \right) . \end{align} Again, we lastly compute from (\ref{def_scaling_2}) the required asymptotics: \begin{align}\label{asymptotic1} -\frac{z^2}{2}& = -x-\log(n) -n\beta \log\log(n) +\log\log(n)+\log\log\log(n)+o(1) \end{align} \begin{align}\label{asymptotic2} \log(z)&=\log \left(\sqrt{2\log(n)}(1+o(1)) \right)=\frac{1}{2}\log\log(n)+\frac{1}{2}\log(2)+o(1) \end{align} $$-\log\log(z) = -\log\log\log(n)-\log(2)+o(1).$$ Combining these with (\ref{last_step2}), it leads to the desired conclusion: $$S_n(z)=e^{-x}\left( 1+o(1)\right) .$$ \end{proof} Finally, we show condition (\ref{condition_for_poisson4}): \begin{lem}\label{uniforme_negligeabilite_nbeta_petit} Let $0\leq y<x<+\infty$. Under the assumptions of Proposition \ref{Main_prop}, one has: \begin{align}\label{uniform_negl_limit} \max_{i\leq n}\left|\mathbb{P}\left(X_{i,n}\geq \varphi_n(x) \right) - \mathbb{P}\left(X_{i,n}\geq \varphi_n(y) \right)\right| &\xrightarrow[n\infty]{}0.\end{align} \end{lem} \begin{proof} For $\varphi_n$ such as in (\ref{def_scaling_1}) or (\ref{def_scaling_2}), and $i\leq n$, let us write: $$\xi_{i,n}:= \left| \mathbb{P}\left(X_{i,n}\geq \varphi_n(x) \right) - \mathbb{P}\left(X_{i,n}\geq \varphi_n(y) \right)\right|.$$ There exists a constant $c>0$ such that:\begin{align}\label{upperbound_for_negl} \xi_{i,n}&= \int_{\varphi_n(y)}^{\varphi_n(x)} \frac{t^{i\beta-2}e^{-\frac{t^2}{2}}}{2^{\frac{i\beta}{2}-1}\Gamma(\frac{i\beta}{2})} \mathrm{d}t \leq \frac{\varphi_n(x)^{i\beta-2}e^{-\frac{\varphi_n(y)^2}{2}}}{2^{\frac{i\beta}{2}-1}\Gamma(\frac{i\beta}{2})} \leq c \frac{\varphi_n(x)^{n\beta}e^{-\frac{\varphi_n(y)^2}{2}}}{\varphi_n(x)^2\Gamma(\frac{i\beta}{2})}. \end{align} Assume that $\displaystyle{ \left( n\log \log \left( n\right) \right)^{-1} \ll \beta \ll n^{-1}}$ and $\varphi_n$ as in (\ref{def_scaling_2}). In this case, considering the asymptotics (\ref{asymptotic1}), (\ref{asymptotic2}) and the fact that \begin{align}\label{asymptotic_rmk} \varphi_n \gg 1, \qquad \min_{i\leq n} \Gamma\left(\frac{i\beta}{2} \right) \gg 1, \end{align} a little computation shows that the latter quantity in (\ref{upperbound_for_negl}) converges to $0$ uniformly in $i\leq n$ as $n\to +\infty$. When $\displaystyle{n^{-2}\ll \beta \ll \left( n\log\log n\right) ^{-1} }$ and $\varphi_n$ as in (\ref{def_scaling_1}), the conclusion follows from (\ref{asymptotic_rmk}) and $\displaystyle{\varphi_n(x)^{n\beta}=1+o(1)}$. \end{proof} \section{Proof of Theorem \ref{Main_th2}}\label{section_proof2} From the normalization sequences $(a_n),(b_n)$ in Theorem \ref{Main_th2}, we define the scaling function $\varphi_n$ in the same way that (\ref{def_scaling_1}) by: \begin{align}\label{def_vfi_gamma} \varphi_n(x):=\frac{x}{a_n}+b_n, \qquad x\in \mathbb{R}^+. \end{align} This is an increasing homeomorphism on $\mathbb{R}^+$ with inverse $\displaystyle{\varphi^{-1}_n(x)=a_n\left(x-b_n\right) }$. We note that it corresponds to the scaling function of $\displaystyle{\frac{n}{\gamma \log\log(n)}}$ i.i.d. copies of $\chi(2\gamma)$. In other words, for $x<+\infty$, it satisfies: $$\frac{n}{\gamma\log \log(n)}\mathbb{P}\big(X_{n,n}\geq \varphi_n(x)\big) \xrightarrow[n\infty]{} e^{-x}.$$ Likewise Theorem \ref{Main_th}, we follow the path of Theorem \ref{Th_tool} and begin with the assumption (\ref{condition_for_poisson3}). We control the sum of chi's survival functions with the same approach. The regime $\displaystyle{n\beta =2\gamma}$ does not allow an uniform negligibility on the chi's random variables. In Theorem \ref{Main_th}, one takes into consideration all summands, each underlying random variables contributing to the extreme value. This time, the $X_{i,n}$ terms with index $i$ negligible to $n$, ie: $i\ll n$, are inconsequential. In order to get meaninful at the limit, we need to explore the region of $n$. We cut off at the appropriate section of the sum, hence identifying the significant stack of chi's random variables significant to the maximum. \begin{lem}\label{lem_exp_limit_gamma} Let $\displaystyle{\gamma \in (0,1)}$. Let $\beta := \beta_n \ll 1$ such that $\displaystyle{n\beta ={2\gamma} }$ and $\displaystyle{X_{i,n}\sim \chi \left( i\beta\right) }$, with $1\leq i \leq n$, a triangular array of independent random variables and let $\varphi_n$ defined in (\ref{def_vfi_gamma}). Then for $0\leq x<+\infty$, $$\sum_{i=1}^{n} \mathbb{P} \big(X_{i,n}\geq \varphi_n(x)\big)\xrightarrow[n\infty]{}e^{-x}. $$ \end{lem} \begin{proof} For clarity, we will work with $z$ instead of $\varphi_n(x)$, $x\in\mathbb{R}^+$. By Remark \ref{gamma_chi_relation_rmk}, \begin{align}\label{aux_equality2} \sum_{i=1}^{n}\mathbb{P}\left(\chi\left( i\beta\right) \geq z \right) =\sum_{i=1}^{n}\mathbb{P}\left(\Gamma\left( \frac{i\beta}{2},1\right) \geq \frac{z^2}{2} \right) = \sum_{i=1}^{n} \frac{\Gamma\left( \frac{i\beta}{2},\frac{z^2}{2}\right) }{\Gamma\left( \frac{i\beta}{2}\right) }=\sum_{i=1}^{n} \frac{\Gamma\left(\gamma \frac{i}{n},\frac{z^2}{2}\right) }{\Gamma\left(\gamma \frac{i}{n}\right) }. \end{align} Since $\displaystyle{i\frac{\gamma}{n}<1}$ for any $i\leq n$, one can estimate (\ref{aux_equality2}) with Lemma \ref{bounds_incomplete_lem}: $$\frac{z^2}{z^2+2}e^{-\frac{z^2}{2}}\sum_{i=1}^{n}\frac{z^{\gamma\frac{i}{n}-2}}{2^{\frac{i\gamma}{2n}-1}\Gamma(\gamma\frac{i}{n})} <\sum_{i=1}^{n}\mathbb{P}\big(X_{i,n}\geq z\big)\leq e^{-\frac{z^2}{2}}\sum_{i=1}^{n}\frac{z^{\gamma\frac{i}{n}-2}}{2^{\frac{i\gamma}{2n}-1}\Gamma(\gamma\frac{i}{n})}.$$ The extra-term $\displaystyle{\frac{z^2}{z^2+2}}$ in the lower bound converges to $1$ because $z\gg 1$. Hence, it is sufficient to prove that $\displaystyle{S_n(z):=e^{-\frac{z^2}{2}}\sum_{i=1}^{n}\frac{z^{\gamma\frac{i}{n}-2}}{2^{\frac{i\gamma}{2n}-1}\Gamma(\gamma\frac{i}{n})}}$ converges to $\displaystyle{e^{-x}}$. For this purpose, let $\mu :=\mu_n$ such that $\displaystyle{\frac{1}{\log\log(n)}\ll \mu_n \ll 1}$, e.g., $\displaystyle{\mu_n =\big( \log \log (n) \big)^{-\frac{1}{2}} }$. We split $\displaystyle{S_n(z) =S^1_n(z)+S^2_n(z)}$ with $$S^1_n(z):= e^{-\frac{z^2}{2}}\sum_{i=1}^{n(1-\mu)}\frac{z^{\gamma\frac{i}{n}-2}}{2^{\frac{i\gamma}{2n}-1}\Gamma(\gamma\frac{i}{n})},\qquad S^2_n(z):= e^{-\frac{z^2}{2}}\sum_{i=n(1-\mu)}^{n}\frac{z^{\gamma\frac{i}{n}-2}}{2^{\frac{i\gamma}{2n}-1}\Gamma(\gamma\frac{i}{n})} ,$$ and show that: $$S^1_n(z) \xrightarrow[n\infty]{} 0,\qquad S^2_n(z) \xrightarrow[n\infty]{}e^{-x}.$$ We begin with the main core term $S^2_n(z)$. Let us define: \begin{align}\label{m_n_M_n2} {m_n}^{-1} := \min_{n(1-\mu)\leq i \leq n}2^{\frac{i\gamma}{2n}} \Gamma(\gamma\frac{i}{n})\sim 2^{\frac{\gamma}{2}}\Gamma(\gamma),\quad {M_n}^{-1} := \max_{n(1-\mu)\leq i \leq n}2^{\frac{i\gamma}{2n}} \Gamma(\gamma\frac{i}{n})\sim 2^{\frac{\gamma}{2}}\Gamma(\gamma) . \end{align} Consequently, \begin{align}\label{consequently} M_n\frac{2e^{-\frac{z^2}{2} }}{z^2}\sum_{i=n(1-\mu)}^{n}\left(\frac{z^2}{2} \right) ^{\gamma\frac{i}{2n}}\leq S^2_n(z) \leq m_n\frac{2e^{-\frac{z^2}{2} }}{z^2}\sum_{i=n(1-\mu)}^{n}\left(\frac{z^2}{2} \right) ^{\gamma\frac{i}{2n}}. \end{align} We compute the geometric sum in the previous line: \begin{align}\label{compute_geometric_sum} \frac{2e^{-\frac{z^2}{2} }}{z^2}\sum_{i=n(1-\mu)}^{n}\left(\frac{z^2}{2} \right) ^{\gamma\frac{i}{2n}}= e^{-\frac{z^2}{2} }\frac{\left( \frac{z^2}{2}\right) ^{\frac{\gamma}{2}-1} \left( 1-\left(\frac{z^2}{2} \right)^{-\frac{\gamma \mu}{2}}\right) }{\left( \frac{z^2}{2}\right)^{\frac{\gamma}{2n}}-1} . \end{align} In order to estimate (\ref{compute_geometric_sum}), we use Taylor expansion, the hypothesis on $\mu$, the definition (\ref{def_vfi_gamma}) and (\ref{m_n_M_n2}), which provide the asymptotics: \begin{align}\label{intermediate_asymptotics_gamma} \left(\frac{z^2}{2} \right)^{-\frac{\gamma \mu}{2}}\ll 1,\qquad \left( \frac{z^2}{2}\right)^{\frac{\gamma}{2n}}-1 = \frac{\gamma}{2n}\log\log(n)\left( 1+o(1) \right) \end{align} Combining (\ref{m_n_M_n2}), (\ref{compute_geometric_sum}) and (\ref{intermediate_asymptotics_gamma}) with the bound (\ref{consequently}), it provides the asymptotic of $\displaystyle{S^2_n(z)}$: \begin{align}\label{S^2_n(z)2} \exp\left(-\frac{z^2}{2}+\left( \frac{\gamma}{2}-1\right) \log\left( \frac{z^2}{2}\right) +\log(n)-\log^{(3)}(n)-\log\left(\frac{2^{\frac{\gamma}{2}-1}\gamma}{\Gamma(\gamma)} \right)+o(1)\right) . \end{align} Finally, together with (\ref{def_vfi_gamma}) and the fact that \begin{align}\label{log(z^2/2)} \left( \frac{\gamma}{2}-1\right) \log\left( \frac{z^2}{2}\right)=\left( \frac{\gamma}{2}-1\right) \log\log(n) +\left( \frac{\gamma}{2}-1\right) \log(2)+o(1) \end{align} \begin{align}\label{-z^2/2} -\frac{z^2}{2} =-x-\log (n)+\log^{(3)}(n)-\left( \frac{\gamma}{2}-1\right) \log\log({n})+\log\left(\gamma 2^{-\frac{\gamma}{2}}\Gamma(\gamma)\right)+o(1), \end{align} the expression (\ref{S^2_n(z)2}) yields the claim. Regarding the first sum $S^1_n(z)$, there exists a numerical constant $c>0$ such that: $$ S^1_n(z)=e^{-\frac{z^2}{2}}\sum_{i=1}^{n(1-\mu)}\frac{z^{\gamma\frac{i}{n}-2}}{2^{\frac{i\gamma}{2n}-1}\Gamma(\gamma\frac{i}{n})} \leq c\frac{2e^{-\frac{z^2}{2} }}{z^2}\sum_{i=1}^{n(1-\mu)}\left(\frac{z^2}{2} \right) ^{\frac{i\gamma}{2n}} = c\frac{2e^{-\frac{z^2}{2}}}{z^2} \frac{\left( \frac{z^2}{2} \right) ^{ \frac{\gamma}{2}(1-\mu)}-\left( \frac{z^2}{2} \right) ^{\frac{\gamma}{2n}}}{\left(\frac{z^2}{2} \right) ^{\frac{\gamma}{2n}}-1}.$$ By Taylor expansion, we find another constant $c'>0$ such that the previous upper bound asymptoticaly equals to: $$c'e^{-\frac{z^2}{2}} \frac{2n}{\gamma\log\log\left( \frac{z^2}{2}\right) }e^{\left( \frac{ \gamma}{2}-1-\frac{\gamma}{2}\mu\right) \log\left( \frac{z^2}{2}\right) }\left( 1+o(1)\right).$$ Using (\ref{log(z^2/2)}) and (\ref{-z^2/2}), for some constant $c''>0$, we have: $$0\leq S^1_n(z)\leq c''\exp\left( -\frac{\gamma}{2}\mu \log\log(n) +o(1)\right) .$$ The hypothesis on $\mu$ is precisely that $\mu \log\log(n)\gg 1$, thus the proof is complete. \end{proof} To conclude the proof, it only remains to show: \begin{lem} Given the assumptions of Theorem \ref{Main_th2} and $\varphi_n$ defined in (\ref{def_vfi_gamma}), for any $0\leq x<y<+\infty$, one has: $$ \max_{i\leq n}\mathbb{P}\big(\varphi_n(x)\leq X_{i,n}\leq \varphi_n(y)\big)\xrightarrow[n\infty]{}0.$$ \end{lem} \begin{proof} Let $x<y$ and $i\leq n$, then \begin{align*} \mathbb{P}\big(\varphi_n(x)\leq X_{i,n}\leq \varphi_n(y)\big)&= \int_{\varphi_n(x)}^{\varphi_n(y)} \frac{e^{-u}u^{t\frac{i}{n}-1}}{\Gamma(t\frac{i}{n})}du\\&\leq \frac{1}{\min_{\mathbb{R}^+} \Gamma}\big(\varphi_n(y)-\varphi_n(x)\big) e^{-\varphi_n(x)} \varphi_n(y)^{t-1} \\ &=\frac{1}{\min_{\mathbb{R}^+} \Gamma}(y-x) e^{-\varphi_n(x)} \varphi_n(y)^{t-1} \end{align*} The latter term is independent of $i\leq n$ and converges to $0$. \end{proof} \bibliographystyle{plain}

https://arxiv.org/abs/1605.06852

The Greedy Spanner is Existentially Optimal

The greedy spanner is arguably the simplest and most well-studied spanner construction. Experimental results demonstrate that it is at least as good as any other spanner construction, in terms of both the size and weight parameters. However, a rigorous proof for this statement has remained elusive.In this work we fill in the theoretical gap via a surprisingly simple observation: The greedy spanner is \emph{existentially optimal} (or existentially near-optimal) for several important graph families, in terms of both the size and weight. Roughly speaking, the greedy spanner is said to be existentially optimal (or near-optimal) for a graph family $\mathcal G$ if the worst performance of the greedy spanner over all graphs in $\mathcal G$ is just as good (or nearly as good) as the worst performance of an optimal spanner over all graphs in $\mathcal G$.Focusing on the weight parameter, the state-of-the-art spanner constructions for both general graphs (due to Chechik and Wulff-Nilsen [SODA'16]) and doubling metrics (due to Gottlieb [FOCS'15]) are complex. Plugging our observation on these results, we conclude that the greedy spanner achieves near-optimal weight guarantees for both general graphs and doubling metrics, thus resolving two longstanding conjectures in the area.Further, we observe that approximate-greedy spanners are existentially near-optimal as well. Consequently, we provide an $O(n \log n)$-time construction of $(1+\epsilon)$-spanners for doubling metrics with constant lightness and degree. Our construction improves Gottlieb's construction, whose runtime is $O(n \log^2 n)$ and whose number of edges and degree are unbounded, and remarkably, it matches the state-of-the-art Euclidean result (due to Gudmundsson et al.\ [SICOMP'02]) in all the involved parameters (up to dependencies on $\epsilon$ and the dimension).

\section{Introduction} \vspace{3pt} {\bf 1.1~ Graph Spanners.~} Given a (connected and undirected) $n$-vertex $m$-edge graph $G = (V,E,w)$ with positive edge weights and a parameter $t \ge 1$, a subgraph $H = (V,E',w)$ of $G$ ($E' \subseteq E$) is called a \emph{$t$-spanner} for $G$ if for all $u,v \in V$, $\delta_H(u,v) \le t \cdot \delta_G(u,v)$. (Here $\delta_G(u,v)$ and $\delta_H(u,v)$ denote the distances between $u$ and $v$ in the graphs $G$ and $H$, respectively.) The parameter $t$ is called the \emph{stretch} of $H$. Spanners constitute a fundamental graph structure, and have been extensively and intensively studied since they were introduced \cite{PS89,PU89}. In many practical applications one is required to construct a spanner that satisfies a number of useful properties, while preserving a small stretch. First, the spanner $H$ should have a small number of edges. Second, its \emph{weight} $w(H) = \sum_{e \in E} w(E)$ should be close to the weight of a minimum spanning tree (MST) of the graph $G$. We henceforth refer to the normalized notion of weight $\Psi(H) = \frac{w(H)}{w(MST(G))}$, which is called \emph{lightness}; a \emph{light} spanner is one with small lightness. Third, its \emph{degree} $\Delta(H)$, defined as the maximum number of edges incident on a vertex, should be small. Light and sparse spanners are particularly useful for efficient broadcast protocols in the message-passing model of distributed computing \cite{ABP90,ABP91}, where efficiency is measured with respect to both the total communication cost (corresponding to the spanner's size and weight) and the speed of message delivery at all destinations (corresponding to the spanner's stretch). Additional applications of such spanners in distributed systems include network synchronization and computing global functions \cite{Awerbuch85,PU89,ABP90,ABP91,Peleg00}. Light and sparse spanners were also found useful for various data gathering and dissemination tasks in overlay networks \cite{BKRCV02,VWFME03,KV01}, in wireless and sensor networks \cite{RW04,BDS04,SS10}, for VLSI circuit design \cite{CKRSW91,CKRSW292,CKRSW92,SCRS01}, for routing \cite{WCT02,PU89,PU89b,TZ01}, to compute distance oracles and labels \cite{Peleg99,TZ01b,RTZ05}, and to compute almost shortest paths \cite{Coh93,RZ04,Elkin05,EZ06,FKMSZ05}. Low degree spanners are also very useful in many of these applications. For example, the degree of the spanner is what determines local memory constraints when using spanners to construct network synchronizers and efficient broadcast protocols. In compact routing schemes, the use of low degree spanners enables the routing tables to be of small size. More generally, viewing vertices as processors, in many applications the degree of a processor represents its \emph{load}, hence a low degree spanner guarantees that the load on all the processors in the network will be low. The \emph{greedy spanner} by Alth$\ddot{\mbox{o}}$fer et al.\ \cite{ADDJS93} is arguably the simplest and most well-studied spanner construction. Alth$\ddot{\mbox{o}}$fer et al.\ showed that for every weighted $n$-vertex graph $G = (V,E,w)$ and an integer parameter $k \ge 1$, the \emph{greedy algorithm} (see \algref{fig:greedy_alg}) constructs a $(2k-1)$-spanner with $O(n^{1+1/k})$ edges; assuming Erd\H{o}s' girth conjecture \cite{erd64}, this size bound is asymptotically tight. Alth$\ddot{\mbox{o}}$fer et al.\ also showed that the lightness of the greedy spanner is $O(n/k)$. Chandra et al.\ \cite{CDNS92} improved the lightness bound, and showed that the greedy spanner for stretch parameter $t=(2k-1)\cdot(1+\epsilon)$ (here $k>1$, $\epsilon>0$) has lightness $O(k \cdot n^{1/{k}} \cdot (1/\epsilon)^{1+1/k})$. Two decades later, Elkin, Neiman and the second author \cite{ENS14} improved the analysis of \cite{CDNS92} and showed that the greedy $(2k-1)\cdot(1+\epsilon)$-spanner has lightness $O(n^{1/k}\cdot (1+ k/(\epsilon^{1+1/k}\log k)))$. Very recently Chechik and Wulff-Nilsen \cite{CW16} improved the lightness bound all the way to $O(n^{1/k} (1/\eps)^{3+2/k})$. Assuming Erd\H{o}s' girth conjecture \cite{erd64} and ignoring dependencies on $\epsilon$, the bound of \cite{CW16} on the lightness is asymptotically tight, thus resolving a major open question in this area. However, the result of Chechik and Wulff-Nilsen \cite{CW16} is not due to a refined analysis of the greedy spanner. Instead, they devised a different construction, which is far more complex, and bounded the lightness of their own construction. The following question was left open. \begin{question} \label{question1} Is the lightness analysis of \cite{ENS14} for the greedy spanner optimal, or can one refine it to derive a stronger bound? In particular, is the spanner of \cite{CW16} lighter than the greedy spanner? \end{question} \vspace{2pt} \noindent{\bf 1.2~ Spanners for Euclidean and Doubling Metrics.~} Consider a set $P$ of $n$ points in $\mathbb R^d$, $d \ge 2$, and a stretch parameter $t \ge 1$. A graph $G = (P,E,w)$ in which the weight $w(p,q)$ of each edge $e = (p,q) \in E$ is equal to the Euclidean distance $\|p-q\|$ between $p$ and $q$ is called a \emph{Euclidean graph}. We say that the Euclidean graph $G$ is a \emph{$t$-spanner} for $P$ (or equivalently, for the corresponding Euclidean metric $(P,\|\cdot\|)$) if for every pair $p,q \in P$ of distinct points, there exists a path $\Pi(p,q)$ in $G$ between $p$ and $q$ whose weight (i.e., the sum of all edge weights in it) is at most $t \cdot \|p-q\|$. The path $\Pi(p,q)$ is said to be a \emph{$t$-spanner path} between $p$ and $q$. For Euclidean metrics, one usually focuses on the regime $t = 1+\eps$, for $\eps > 0$ being an arbitrarily small parameter. Euclidean spanners were introduced by Chew \cite{Chew86}, and were subject to intensive ongoing research efforts since then. We refer to the book ``Geometric Spanner Networks'' \cite{NS07}, which is devoted almost exclusively to Euclidean spanners and their numerous applications. As with general graphs, it is important to devise Euclidean spanners that achieve small size, lightness and degree. The \emph{doubling dimension} of a metric space $(M,\delta)$ is the smallest value $\ddim$ such that every ball $B$ in the metric space can be covered by at most $2^{\ddim}$ balls of half the radius of $B$. This notion generalizes the Euclidean dimension, since the doubling dimension of the Euclidean space $\mathbb R^d$ is $\Theta(d)$. A metric space is called \emph{doubling} if its doubling dimension is constant. Spanners for doubling metrics were also subject of intensive research \cite{GGN04,CGMZ05,CG06,HPM06,Rod07,GR081,GR082,Smid09,ES13,CLNS13,Sol14}. The basic line of work in this context is to generalize the known Euclidean spanner results for arbitrary doubling metrics. Das et al.\ \cite{DHN93} showed that, in low-dimensional Euclidean metrics, the greedy $(1+\eps)$-spanner has constant degree (and so $O(n)$ edges) and constant lightness. In $n$-point doubling metrics, the greedy $(1+\eps)$-spanner has $O(n)$ edges and lightness $O(\log n)$ \cite{Smid09}. As for the degree, there exist $n$-point metric spaces with doubling dimension 1 for which the greedy spanner has a degree of $n-1$ \cite{HM06,Smid09}. It has been a major open question to determine whether any doubling metric admits a $(1+\eps)$-spanner with sub-logarithmic lightness. Recently Gottlieb \cite{Got15} answered this question in the affirmative by devising such a spanner construction with constant lightness. Again, this result is not due to a refined analysis of the greedy spanner. Instead, Gottlieb devised a different construction, which is far more complex, and bounded the lightness of his own construction. The following question was left open. \begin{question} \label{question2} Is the lightness analysis of \cite{Smid09} for the greedy spanner optimal, or can one refine it to derive a stronger bound? In particular, is the spanner of \cite{Got15} lighter than the greedy spanner? \end{question} The high runtime of the greedy spanner is a major drawback. The state-of-the-art implementation of the greedy spanner in both Euclidean and doubling metrics requires time $O(n^2 \log n)$ \cite{BCFMS10}. Building on \cite{DHN93}, Das and Narasimhan \cite{DN97} devised a much faster algorithm that follows the greedy approach. The runtime of their ``approximate-greedy'' algorithm is $O(n \log^2 n)$, yet its degree and lightness are both bounded by constants (as with the greedy spanner). Gudmundsson et al.\ \cite{GLN02} improved the result of \cite{DN97}, implementing the approximate-greedy algorithm within time $O(n \log n)$. For doubling metrics, however, the only spanner construction with sub-logarithmic lightness is that of \cite{Got15}; the runtime of Gottlieb's construction is $O(n \log^2 n)$ rather than $O(n \log n)$, and the size and degree of his construction are unbounded. Hence, there is a big gap in this context between Euclidean and doubling metrics, leading to the following question. \begin{question} \label{question3} Can one compute $(1+\eps)$-spanners with constant lightness in doubling metrics within time $O(n \log n)$? Furthermore, can one extend the state-of-the-art Euclidean result of \cite{GLN02} to arbitrary doubling metrics? \end{question} There have been numerous experimental studies on Euclidean spanners. (See \cite{FG05,Far08}, and the references therein.) The conclusion emerging from these experiments is that the greedy Euclidean spanner outperforms the other popular Euclidean spanner constructions, with respect to the size and lightness bounds. (Specifically, the greedy spanner was found to be $10$ times sparser and $30$ times lighter than any other examined spanner.) It is reasonable to assume that a similar situation occurs in arbitrary doubling metrics. \vspace{6pt} \\ \noindent{\bf 1.3~ Our Contribution.~} In this work we fill in the theoretical gap by making three important observations. \begin{enumerate} \item Our first observation is surprisingly simple: The greedy spanner is \emph{existentially optimal} with respect to both the size and the lightness, for any graph family that is closed under edge removal. Applying this observation to the family of general weighted graphs, we conclude that the greedy spanner is just as light as the spanner of \cite{CW16}, thus answering \questionref{question1}. \item The first observation does not hold for doubling metrics. Our second observation is that the greedy spanner is existentially \emph{near}-optimal with respect to both the size and the lightness, for the family of doubling metrics. In particular, it is just as light as the spanner of \cite{Got15}, thus answering \questionref{question2}. \item Our third observation concerns the optimality of the approximate-greedy algorithm of \cite{DN97,GLN02} in doubling metrics, and is more intricate than the first two observations. Informally, it states that the approximate-greedy spanner with stretch parameter $t$ is existentially near-optimal with respect to the lightness, for the family of doubling metrics, but when compared to spanners with a slightly smaller stretch parameter $t' < t$. This enables us to conclude that the lightness of the approximate-greedy spanner is close to that of \cite{Got15}. In this way we manage to extend the state-of-the-art Euclidean result of \cite{GLN02} to arbitrary doubling metrics, thus answering \questionref{question3}.\footnote{The $O(n \log n)$ runtime bound of \cite{GLN02} holds in the traditional algebraic computation-tree model with the added power of indirect addressing. Our result applies with respect to the same computation model.} \end{enumerate} To clarify the meaning of \emph{existential optimality}, suppose that the greedy spanner is existentially optimal with respect to the lightness, for some graph family $\mathcal{F}$. This does not imply that for any graph $G \in \mathcal{F}$, the lightness of the greedy spanner for $G$ is bounded by the optimal lightness of any spanner for $G$. It simply means that for any graph $G \in \mathcal{F}$, there \emph{exists} a graph $G' \in \mathcal{F}$, such that the lightness of the greedy spanner for $G$ is bounded by the optimal lightness of any spanner for $G'$. Put in other words, for a graph $G \in \mathcal{F}$, let $l(G) = l_t(G)$ denote the optimal lightness of any $t$-spanner for $G$, for an arbitrary stretch parameter $t \ge 1$, and let $l(\mathcal{F}) = \max\{l(G) ~\vert~ G \in \mathcal{F}\}$ denote the maximum value $l(G)$ over all graphs $G$ in $\mathcal{F}$. Then, for any graph $G \in \mathcal{F}$, although the lightness of the greedy $t$-spanner for $G$ may well exceed $l(G)$, it must be upper bounded by $l(\mathcal{F})$. {For example, let $\mathcal{F}$ be the family of general weighted graphs on $n$ vertices, and let $H$ be an $n$-vertex dense graph of high girth, namely, with girth $t+2$ and $n^{1+\Theta(1/t)}$ edges, where all edge weights are 1. Also, let $S$ be a star on the same vertex set as $H$ rooted at an arbitrary vertex, so that all edges of $S$ that belong to $H$ have weight 1 and all edges of $S$ that do not belong to $H$ have weight $1+\eps$. Finally, let $G$ be the graph containing all edges of $H$ and all edges of $S$ with weight $1+\eps$. Note that the greedy $t$-spanner for $G$ includes all $n^{1+\Theta(1/t)}$ edges of the high girth graph $H$, whereas the optimal $t$-spanner (assuming $t \ge 2+2\eps$) consists of the edges of the star $S$, hence is much sparser and lighter. (See \figureref{fig:girth} for an illustration.) This example, however, does not contradict the existential optimality of the greedy spanner: Although the lightness of the greedy $t$-spanner for $G$ exceeds $l(G)$, it can be shown that it is equal to $l(H)$, which, in turn, is bounded by $l(\mathcal{F})$.} \begin{figure} \begin{center} \includegraphics[width=0.41\textwidth]{figpeterson3} \caption{\small The graph $H$ in the figure is the Petersen graph on 10 vertices, with girth 5 and 15 edges. All edges of $H$ have weight 1, and are colored black. The red dashed edges are the edges of the star $S$ of weight $1+\eps$. The greedy 3-spanner for the graph $G$ obtained as the union of the black and red edges in the figure includes all 15 edges of $H$, whereas the optimal 3-spanner for $G$ consists of the 9 edges of $S$.} \label{fig:girth} \end{center} \end{figure} The meaning of \emph{existential near-optimality} is similar, except that we are allowed to have some \emph{slack}, which may depend on the stretch parameter $t$ as well as on parameters of the graph family of interest $\mathcal{F}$. As mentioned, in our third observation we compare the lightness of the greedy spanner with a certain stretch parameter $t$ to the optimal lightness of any spanner, but with a slightly smaller stretch parameter $t'$. This is just one example of how the slack parameter can be used. Another example is to compare the greedy spanner in some graph family $\mathcal{F}$ to the optimal spanner, but with respect to a different (closely related) graph family $\mathcal{F'}$. In particular, in our second and third observations we compare the lightness of the greedy spanner in metric spaces of bounded doubling dimension to the optimal lightness of any spanner, but with respect to metric spaces of slightly larger doubling dimension. (See \sectionref{sec:pre} for the definition of doubling dimension.) It would be interesting to study additional ways of using the slack parameter, as they may lead to new results in this area. We remark that light spanners were extensively studied in various graph families such as planar graphs \cite{ADDJS93,Klein05}, apex graphs \cite{GS02}, bounded pathwidth graphs \cite{GH12,H12}, bounded catwidth graphs \cite{H12}, bounded genus graphs \cite{Grigni00,GS02,DHM10}, bounded treewidth graphs \cite{DHM10,H12}, and graphs excluding fixed minors \cite{Grigni00,DHM10}. Since all these graph families are closed under edge removal, our first observation implies that the greedy spanner for them is just as good as any other spanner. \vspace{6pt} \\ \noindent{\bf 1.4~ Organization.~} In \sectionref{sec:pre} we present the notation that is used throughout the paper, and summarize some statements from previous work that are most relevant to us. In \sectionref{sec:Greedy_optimal} we show that the greedy spanner is existentially optimal for graph families that are closed under edge removal. The basic optimality argument of \sectionref{sec:Greedy_optimal} is extended to doubling metrics in \sectionref{sec:doubling}. Finally, in \sectionref{sec:Fast_alg} we show that the approximate-greedy spanner in doubling metrics is light. \section{Preliminaries}\label{sec:pre} Let $G=(V,E,w)$ be a (connected and undirected) graph with positive edge weights. The weight $w(P)$ of a path $P$ is the sum of all edge weights in it, i.e., $w(P) = \sum_{e \in P} w(e)$. For a pair of vertices $u,v \in V$, let $\delta_G(u,v)$ denote the distance between $u$ and $v$ in $G$, i.e., the weight of a shortest path between them. We denote by $M_G = (V,\delta_G)$ the (shortest path) metric space \emph{induced} by $G$; we will view $M_G$ as a complete weighted graph $(V, {V \choose 2},w)$ over the vertex set $V$, where the weight $w(u,v)$ of an edge $(u,v)$ is given by the graph distance $\delta_G(u,v)$ between its endpoints. A subgraph $H=(V,E',w)$ of $G$ (where $E'\subseteq E$) is called a \emph{t-spanner} for $G$ if for all $u,v \in V$, $\delta_H(u,v)\le t\cdot \delta_G(u,v)$. The parameter $t$ is called the \emph{stretch} of the spanner $H$. If $\delta_H(u,v)\le t\cdot \delta_G(u,v)$ for all edges $(u,v) \in E$, then it also holds that $\delta_H(u,v)\le t\cdot \delta_G(u,v)$ for all pairs of vertices $u,v \in V$. Therefore, to bound the stretch of the spanner, one may restrict the attention to the edges of the graph. Let $|H| = |E'|$ denote the \emph{size} of $H$, and let $w(H) = w(E') = \sum_{e\in E'}w(e)$ denote its \emph{weight}. The \emph{lightness} $\Psi(H)$ of $H$ is the ratio between the weight of $H$ and the weight of an MST for $G$, i.e., $\Psi(H)=\frac{w(H)}{w(MST(G))}$. (Throughout the paper all logarithms are in base 2.) The result of Chechik and Wulff{-}Nilsen \cite{CW16} is summarized in the following theorem. \begin{theorem}[\cite{CW16}]\label{thm:CW16} For every weighted $n$-vertex graph $G=(V,E,w)$ and parameters $k\ge 1$ and $0<\epsilon<1$, there exists a $(2k-1)\cdot(1+\epsilon)$-spanner with $O(n^{1+1/k})$ edges and lightness $O(n^{1/k} (1/\eps)^{3+2/k})$. Such a spanner can be constructed in polynomial time. \end{theorem} \subsection{Doubling metrics} The following lemma gives the standard packing property of doubling metrics (see, e.g., \cite{GKL03}). \begin{lemma} \label{lem:doubling_packing} Let $(M,\delta)$ be a metric space with doubling dimension $\ddim$. If $S \subseteq M$ is a subset of points with minimum interpoint distance $r$ that is contained in a ball of radius $R$, then $|S| \le \left(\frac{2R}{r}\right)^{O(\ddim)}$. \end{lemma} The following theorem states that any doubling metric admits a constant degree $(1+\eps)$-spanner. \begin{theorem}[\cite{CGMZ05,GR08}]\label{thm:doubling_degree} For any $n$-point metric space $(M,\delta)$ with doubling dimension $\ddim$ and parameter $0<\epsilon<1/2$, there exists a $(1 + \epsilon)$-spanner with degree $\epsilon^{-O(\ddim)}$. The runtime of this construction is $\eps^{-O(\ddim)} (n \log n)$. \end{theorem} The result of Gottlieb \cite{Got15} is summarized in the following theorem. \begin{theorem}[\cite{Got15}]\label{thm:Got15} For any $n$-point metric space $(M,\delta)$ with doubling dimension $\ddim$ and parameter $0<\epsilon<1/2$, there exists a $(1+\epsilon)$-spanner with lightness $(\ddim/\epsilon)^{O(\ddim)}$. The runtime of this construction is $(\ddim/\epsilon)^{O(\ddim)} (n\log^{2}n)$. \end{theorem} \subsection{The Greedy Spanner and its Basic Properties} \begin{algorithm} \caption{$\texttt{Greedy}(G=(V,E,w),t)$}\label{fig:greedy_alg} \begin{algorithmic}[1] \STATE $H=(V,\emptyset,w)$. \FOR {each edge $(u,v)\in E$, in non-decreasing order of weight,} \IF {$\delta_H(u,v)>t\cdot w(u,v)$} \STATE Add the edge $(u,v)$ to $E(H)$. \ENDIF \ENDFOR \end{algorithmic} \end{algorithm} The greedy spanner algorithm is presented in \algref{fig:greedy_alg}. Let $H=\left(V,E_{H},w\right)$ be the output of an arbitrary execution of the greedy algorithm with stretch parameter $t$. It is immediate that $H$ has stretch at most $t$. If the edge weights in the graph are distinct, then $H$ is uniquely defined, but this does not hold in general; nevertheless, by letting $H$ designate an arbitrary such spanner, we may henceforth refer to it as the \emph{greedy $t$-spanner}. The following observation is immediate (see, .e.g., \cite{ENS14,CW16}). \begin{observation} \label{fct:greedy contains MST} $H$ contains all edges of some MST of $G$, denoted $Z$. (Hence $Z$ is also an MST of $H$.) \end{observation} \section{The Basic Optimality Proof}\label{sec:Greedy_optimal} In this section we show that the greedy spanner is existentially optimal, with respect to both the size and the lightness, for any graph family that is closed under edge removal. We start by making the basic observation that the only $t$-spanner of the greedy $t$-spanner is itself. \begin{lemma} \label{obs:spanner_of_greedy} Let $G=(V,E,w)$ be any weighted graph, let $t \ge 1$ be any stretch parameter, and let $H$ be the greedy $t$-spanner of $G$. If $H'$ is a $t$-spanner for $H$, then $H' = H$. \end{lemma} \begin{proof} Assume for contradiction that $H'$ is a $t$-spanner for $H$ yet there is an edge $e \in H \setminus H'$. Let $P$ be a shortest path in $H'$ between the endpoints of $e$. As $H'$ is a $t$-spanner of $H$, it holds that $w(P)\le t\cdot w(e)$. Consider the last edge examined by the greedy algorithm among the edges of $P$ and $e$, denoted $e'$. By the description of the greedy algorithm, we have $w(e) \le w(e')$. Consequently, by the time the greedy algorithm examines edge $e'$, all the edges of the path $(P \cup e) \setminus e'$ must have already been added to the greedy spanner. (See \figureref{fig:lem3} for an illustration.) This path connects the endpoints of $e'$, and its weight is given by $$w(P) - w(e') + w(e) ~\le~ w(P) ~\le~ t \cdot w(e) ~\le~ t \cdot w(e').$$ Hence the greedy algorithm will not add edge $e'$ to $H$, a contradiction. \begin{figure} \begin{center} \includegraphics[width=0.41\textwidth]{figLem3colored} \caption{\small The path $P$ in $H'$ between the endpoints of edge $e$ is depicted by a dashed line. The path $P\cup e \setminus e'$ between the endpoints of edge $e'$, all edges of which have been added to $H$ by the time the greedy algorithm examines edge $e'$, is colored red.} \label{fig:lem3} \end{center} \end{figure} \end{proof} Equipped with \lemmaref{obs:spanner_of_greedy}, we now turn to the basic optimality proof. \begin{theorem}[Greedy is optimal]\label{thm:main} Let $\mathcal{G}$ be any family of $n$-vertex graphs that is closed under edge removal, and let $t = t(n)$ be any stretch parameter. Assume that for every graph $G\in\mathcal{G}$, there exists a $t$-spanner for $G$ with at most $m(n,t)$ edges and lightness at most $l(n,t)$. Then for every graph $G\in\mathcal{G}$, the greedy $t$-spanner $H$ of $G$ has at most $m(n,t)$ edges and lightness at most $l(n,t)$.\end{theorem} \begin{proof} Consider an arbitrary graph $G$ in $\mathcal{G}$, and let $H$ be the greedy $t$-spanner of $G$. Since $\mathcal{G}$ is closed under edge removal and $H$ is a subgraph of $G$, $H$ belongs to $\mathcal{G}$. Hence, there exists a $t$-spanner $\mathcal{H}$ of $H$ with at most $m(n,t)$ edges and lightness at most $l(n,t)$. \lemmaref{obs:spanner_of_greedy} implies that $\mathcal{H}=H$, from which the size bound on $H$ immediately follows. The lightness bound is slightly trickier, as the spanner $\mathcal{H}$ is computed on top of the greedy spanner $H$ rather than the original graph $G$. Nevertheless, \observationref{fct:greedy contains MST} implies that $G$ and $H$ have the same MST $Z$. Since the lightness of $\mathcal{H}$ is at most $l(n,t)$ and $Z$ is an MST for $H$, it follows that $$l(n,t) ~\ge~ \Psi(\mathcal{H}) ~=~ \frac{w(\mathcal{H})}{w(MST(H))} ~=~ \frac{w(\mathcal{H})}{w(Z)},$$ hence $w(\mathcal{H}) \le l(n,t) \cdot w(Z)$. Using the fact that $\mathcal{H}= H$, we conclude that the lightness of $H$ satisfies $$\Psi(H) ~=~ \frac{w(H)}{w(MST(G))} ~=~ \frac{w(H)}{w(Z)} ~=~ \frac{w(\mathcal{H})}{w(Z)} ~\le~ \frac{l(n,t)\cdot w(Z)}{w(Z)} ~=~ l(n,t).$$ \end{proof} As the family of weighted graphs is closed under edge removal, we can apply \theoremref{thm:main} on it. Hence the greedy spanner for general graphs has size and lightness at least as good as in \theoremref{thm:CW16}. \begin{corollary}\label{cor:Greedy_CW} For every weighted graph $G=(V,E,w)$ on $n$ vertices and $m$ edges and parameters $k\ge 1$ and $0<\epsilon<1$, the greedy $(2k-1)\cdot(1+\epsilon)$-spanner has $O(n^{1+1/k})$ edges and lightness $O(n^{1/k} (1/\eps)^{3+2/k})$. (A naive implementation of the greedy algorithm requires $O(m n^{1+1/k})$ time.) \end{corollary} In \cite{BFN16} it was proved that for any parameter $0<\delta<1$ and any stretch parameter $t = t(n)$, if every $n$-vertex weighted graph admits a $t$-spanner with at most $m(n,t)$ edges and lightness at most $l(n,t)$, then for every such graph there also exists a $t/\delta$-spanner with at most $m(n,t)$ edges and lightness at most $1+\delta\cdot l(n,t)$.\footnote{This reduction appears in the full version of \cite{BFN16}, which is currently available via \url{http://www.cs.bgu.ac.il/~arnoldf/BFN16.pdf}.} Applying \theoremref{thm:main} again, we derive the following result. \begin{corollary}\label{cor:light_Greedy} For every weighted $n$-vertex graph $G = (V,E,w)$ and parameter $0<\delta<1$, the greedy $O(\log n/\delta)$-spanner has $O(n)$ edges and lightness at most $1+\delta$. \end{corollary} As mentioned in the introduction, a plethora of graph families that are closed under edge removal were studied extensively in the spanner literature. This includes the families of planar graphs, bounded genus graphs, bounded treewidth graphs, graphs excluding fixed minors, and more. For all these graph families, \theoremref{thm:main} shows that the greedy spanner is existentially optimal. \section{The Optimality Argument in Doubling Metrics}\label{sec:doubling} The basic optimality argument of \sectionref{sec:Greedy_optimal} applies to graph families that are closed under edge removal. Note that metric spaces do not fall into this category. Nevertheless, for metric spaces, the basic optimality argument suffices: On the one hand, the upper bound for general weighted graphs applies to any metric space, and on the other hand, the lower bound due to high girth graphs naturally applies to the induced metric spaces (see, e.g., \cite{ADDJS93,RR98}). In this section we study the optimality of the greedy spanner for \emph{doubling metrics}. For such metric spaces, one would like to obtain spanners with stretch $1+\eps$, where $\eps$ is arbitrarily close to 0. We will show that the greedy $(1+\eps)$-spanner is existentially near-optimal in doubling metrics, with respect to both the size and the lightness. The next observation and subsequent lemma will be used for proving the lightness optimality. \begin{observation} \label{mst-metric} Consider the metric space $M_G$ induced by an arbitrary weighted graph $G = (V,E,w)$. Then any MST of $M_G$ is a spanning tree of $G$. (Hence there is a common MST for $G$ and $M_G$, denoted $Z$.) \end{observation} \begin{proof} Consider an MST $Z$ for $M_G$, and suppose for contradiction that $Z$ contains an edge $e$ outside $G$. Since $e$ belongs to $M_G \setminus G$, any path in $G$ between the endpoints of $e$ consists of at least two edges. Consider the (multi) graph obtained from $Z$ by replacing edge $e$ with a shortest path in $G$ between the endpoints of $e$. It is a spanning subgraph of $M_G$ of weight $w(Z)$, which contains at least $n+1$ edges (some of which may be multiple edges), and thus at least one cycle. By breaking cycles in this subgraph, we obtain a spanning tree of $M_G$ of weight strictly smaller than $w(Z)$, yielding a contradiction to the weight minimality of $Z$. \end{proof} \begin{lemma} \label{lightopt} Let $(M,\delta)$ be any metric space, $t \ge 1$ be some stretch parameter, and $H$ be the greedy $t$-spanner of $M$. For every $t$-spanner $H'$ of the metric space $M_H$ induced by $H$, we have $w(H) \le w(H')$. \end{lemma} \begin{proof} Let $H'$ be a $t$-spanner of $M_H$, and define $H''$ as the subgraph of $H$ obtained from $H'$ by replacing each edge $e$ of $H'$ with a shortest path in $H$ between the endpoints of $e$. Clearly, the distances in $H''$ are no greater than the respective distances in $H'$. Since $H''$ is a subgraph of $H$, it follows that $H''$ is a $t$-spanner for $H$. \lemmaref{obs:spanner_of_greedy} implies that $H'' = H$. Finally, noting that $w(H'') \le w(H')$, we have $w(H) = w(H'') \le w(H')$. \end{proof} The following lemma will be used for proving the size optimality. \begin{lemma}\label{lem:metric_spars} Let $(M,\delta)$ be any metric space, $t < 2$ be some stretch parameter, and $H$ be the greedy $t$-spanner of $M$. For every $t$-spanner $H'$ of the metric space $M_H$ induced by $H$, we have $|H| \le |H'|$. \end{lemma} \begin{proof} For every edge $e'\in H'$, let $P_{e'}$ be a shortest path between the endpoints of $e'$ in $H$. We say that edge $e'\in H'$ \emph{covers} all edges of $P_{e'}$, and all these edges are \emph{covered} by $e'$. (An edge $e' \in H \cap H'$ covers itself.) For each edge $e$ in $H \setminus H'$, let $\mathcal{Q}_{e}$ be a shortest path between the endpoints of $e$ in $H'$. Since $H'$ is a $t$-spanner for $M_H$, we have $w(\mathcal{Q}_{e}) \le t\cdot w(e)$. Observe that the edges in $\cup_{e'\in\mathcal{Q}_{e}} P_{e'}$ form a path $\Pi_e$ in $H$ between the endpoints of $e$. (It will be shown next that the path $\Pi_e$ is not simple.) We have $$w(\Pi_e) ~\le~ \sum_{e'\in\mathcal{Q}_{e}}w(P_{e'}) ~= \sum_{e'\in\mathcal{Q}_{e}}w(e') ~=~ w(\mathcal{Q}_{e}) ~\le~ t\cdot w(e)~.$$ Next, we argue that the edge $e$ must belong to $\Pi_e$. Indeed, otherwise the edges of $\Pi_e$ contain a simple path in $H$ between the endpoints of $e$ of weight bounded by $t \cdot w(e)$, implying that the heaviest edge among the edges of this path and $e$ would not be added to the greedy $t$-spanner $H$. Consequently, at least one edge $e'$ in $\mathcal{Q}_{e}$ must cover $e$. We define an injection $f:H\rightarrow H'$ as follows. For each edge $e\in H \cap H'$, $f(e)$ is defined as $e$; in this case edge $e = f(e)$ covers itself. For each $e \in H \setminus H'$, $f(e)$ is defined to be an arbitrary edge of $\mathcal{Q}_{e}$ that covers $e$. To see that $f$ is injective, suppose for contradiction the existence of two distinct edges $e_{1}$ and $e_{2}$ in $H$ and an edge $e' \in H'$ such that $f(e_{1})=f(e_{2})= e' \in H'$. It must hold that $e_1$ and $e_2$ are in $H \setminus H'$. Assume without loss of generality that $w(e_{1})\le w(e_{2})$. Since both $e_{1}$ and $e_{2}$ are covered by $e'$, it follows that $w(e')\ge w(e_{1})+w(e_{2})\ge2\cdot w(e_{1})$. On the other hand, by the definition of $f$, the shortest path $\mathcal{Q}_{e_1}$ in $H'$ between the endpoints of $e_{1}$ contains the edge $e' = f(e_1)$. Hence the weight of a shortest path in $H'$ between the endpoints of $e_1$ is given by $w(\mathcal{Q}_{e_1}) \ge w(e') \ge2\cdot w(e_{1})>t\cdot w(e_{1})$, which contradicts the fact that $H'$ is a $t$-spanner for $H$. It follows that $f$ is injective, from which we conclude that $|H|\le|H'|$. \end{proof} The following observation shows that a small ``stretching'' of any metric space does not change the doubling dimension of the metric space by much. \begin{observation}\label{lem:doubling_embedding} Let $(M,\delta)$ be a metric space with doubling dimension $\ddim$. Let $H$ be a $t$-spanner of $M$, for $t\le 2$. Then the metric space $M_H$ induced by $H$ has doubling dimension at most $2\cdot\ddim$. \end{observation} \begin{proof} Clearly, any ball of radius $r$ in the ``stretched'' metric space $M_{H}$ is contained in the respective ball of the original metric space $M$. By definition, this ball can be covered by $2^{2\cdot\ddim}$ balls of radius $\frac{r}{4}$ in $M$, and so by $2^{2\cdot\ddim}$ balls of radius $t\cdot\frac{r}{4}\le\frac{r}{2}$ in the stretched metric space $M_H$. \end{proof} The existential near-optimality result for doubling metrics is summarized in the following theorem. \begin{theorem}[Greedy is near-optimal in doubling metrics]\label{theorem:doubling_light} Assume that for every $n$-point metric space $(M,\delta)$ with doubling dimension $\ddim$ and any stretch parameter $t < 2$, there is a $t$-spanner with at most $m(n,\ddim,t)$ edges and lightness at most $l(n,\ddim,t)$. Then for every $n$-point metric space $(M,\delta)$ with doubling dimension $\ddim$, the greedy $t$-spanner has at most $m(n,2\cdot\ddim,t)$ edges and lightness at most $l(n,2\cdot\ddim,t)$. \end{theorem} \begin{proof} Let $M$ be an arbitrary $n$-point metric space with doubling dimension $\ddim$, let $H=\left(V,E\right)$ be the greedy $t$-spanner for $M$, and let $M_H$ be the metric space induced by $H$. By \observationref{lem:doubling_embedding}, the doubling dimension of $M_H$ is bounded by $2\cdot \ddim$. Our assumption implies that there exists a $t$-spanner $\mathcal{H}$ for $M_H$ with at most $m(n,2\cdot\ddim,t)$ edges and lightness at most $l(n,2\cdot\ddim,t)$. \lemmaref{lem:metric_spars} implies that $|H|\le |\mathcal H|$, from which the size bound on $H$ immediately follows. As for the lightness bound on $H$, note that $\mathcal{H}$ is computed on top of $M_H$ rather than the original metric space $M$. Nevertheless, \observationref{fct:greedy contains MST} and \observationref{mst-metric} imply that $M$ and $M_H$ have the same MST $Z$. Since the lightness of $\mathcal{H}$ is at most $l(n,2\cdot\ddim,t)$, it follows that $$l(n,2\cdot\ddim,t) ~\ge~ \Psi(\mathcal{H}) ~=~ \frac{w(\mathcal{H})}{w(MST(M_H))} ~=~ \frac{w(\mathcal{H})}{w(Z)},$$ hence $w(\mathcal{H}) \le l(n,2\cdot\ddim,t) \cdot w(Z)$. By \lemmaref{lightopt}, we have $w(H) \le w(\mathcal H)$, hence the lightness of $H$ satisfies $$\Psi(H) ~=~ \frac{w(H)}{w(MST(M))} ~=~ \frac{w(H)}{w(Z)} ~\le~ \frac{w(\mathcal{H})}{w(Z)} ~\le~ \frac{l(n,2\cdot\ddim,t) \cdot w(Z)}{w(Z)} ~=~ l(n,2\cdot\ddim,t).$$ \end{proof} It is known that the greedy $(1+\eps)$-spanner for $n$-point doubling metrics has $O(n)$ edges and lightness $O(\log n)$ \cite{Smid09}, where the $O$-notation hides a multiplicative term of $(1/\epsilon)^{O(\ddim)}$. Applying \theoremref{theorem:doubling_light} in conjunction with the result of \theoremref{thm:Got15}, we reduce the lightness bound of the greedy spanner to constant. \begin{corollary}\label{cor:Greedy_G} For every metric space $(M,\delta)$ with doubling dimension $\ddim$ and any parameter $0<\epsilon<\frac{1}{2}$, the greedy $(1+\epsilon)$-spanner has $n(1/\epsilon)^{O(\ddim)}$ edges and lightness $(\ddim/\epsilon)^{O(\ddim)}$. \end{corollary} \noindent {\bf Remark.} \corollaryref{cor:Greedy_G} shows that the greedy $(1+\eps)$-spanner in doubling metrics achieves optimal bounds on the size and the lightness, \emph{disregarding dependencies on $\eps$ and the doubling dimension}. However, improving these dependencies is a fundamental challenge of practical importance. By \theoremref{theorem:doubling_light}, any improvement whatsoever in the dependencies on $\eps$ and the doubling dimension on either the size or the lightness of \emph{any spanner construction} for doubling metrics -- would trigger a similar improvement to the greedy spanner. \section{The Approximate-Greedy Spanner in Doubling Metrics is Light}\label{sec:Fast_alg} \corollaryref{cor:Greedy_G} shows that the greedy $(1+\eps)$-spanner in doubling metrics achieves near-optimal bounds on the size and the lightness. Nevertheless, this spanner has two major disadvantages. First, as mentioned in the introduction, there exist metric spaces with doubling dimension 1 for which its degree may be unbounded. (This is in contrast to $d$-dimensional Euclidean metrics, where the greedy $(1+\eps)$-spanner has degree $\eps^{-O(d)}$.) Second, it cannot be constructed within sub-quadratic time in doubling metrics due to a lower bound of \cite{HM06}. In fact, even in $d$-dimensional Euclidean metrics, the state-of-the-art implementation of the greedy $(1+\eps)$-spanner requires time $\eps^{-O(d)} (n^2 \log n)$ \cite{BCFMS10}. Building on \cite{DHN93,DN97}, Gudmundsson et al.\ \cite{GLN02} devised a much faster algorithm that follows the greedy approach, hereafter Algorithm \texttt{Approximate-Greedy}. The runtime of this algorithm is $\eps^{-O(d)}(n \log n)$, yet the degree and lightness of the approximate-greedy spanner produced by the algorithm are both bounded by $\eps^{-O(d)}$, just as with the greedy spanner for Euclidean metrics. The runtime analysis of Algorithm \texttt{Approximate-Greedy} \cite{GLN02} does not exploit any properties of Euclidean geometry. Specifically, it relies on the triangle inequality, which applies to arbitrary metric spaces, and on standard packing arguments (cf.\ \lemmaref{lem:doubling_packing}), which apply to arbitrary doubling metrics. Therefore, the runtime of Algorithm \texttt{Approximate-Greedy} remains $\eps^{-O(\ddim)}(n \log n)$ in arbitrary doubling metrics. Moreover, the degree bound of $\eps^{-O(d)}$ applies to arbitrary doubling metrics as well. (We refer to Chapter 15 in \cite{NS07} for an excellent description of this algorithm and its analysis.) In this section we show that the approximate-greedy spanner of \cite{GLN02} has constant lightness in arbitrary doubling metrics. Consequently, Algorithm \texttt{Approximate-Greedy} provides an $O(n\log n)$-time construction of $(1+\eps)$-spanners in doubling metrics with lightness and degree both bounded by constants. \subsection{A Rough Sketch of Algorithm \texttt{Approximate-Greedy}} In this section we provide a very rough sketch of Algorithm \texttt{Approximate-Greedy}, aiming to highlight the high-level ideas behind it. This outline is not required for the analysis that is given in \sectionref{bounding}; it is provided here for clarity and completeness. In metric spaces, the greedy algorithm sorts the ${n \choose 2}$ interpoint distances and examines the edges by non-decreasing order of weight. For each edge that is examined for inclusion in the spanner, the distance between its endpoints in the current spanner is computed. This is expensive for two reasons: (1) The number of examined interpoint distances is quadratic in $n$. (2) Computing the \emph{exact} spanner distance between two points is costly. Suppose we aim for a stretch of $t = 1+\eps$, and let $t'$ be an appropriate parameter satisfying $t' = 1 + O(\eps) < t$. (Refer to \cite{GLN02,NS07} for the exact constant hiding in the $O$-notation of $O(\eps)$.) Instead of examining all ${n \choose 2}$ interpoint distances, Algorithm \texttt{Approximate-Greedy} computes a bounded degree $\sqrt{t/t'}$-spanner $G' = (M,E',\delta)$ for the input metric space $(M,\delta)$, and simulates the greedy algorithm with stretch parameter $\sqrt{t \cdot t'}$ only on the edges of $G'$. The output of the algorithm is a $\sqrt{t \cdot t'}$-spanner $G = (M,E,\delta)$ for $G'$, which is a $t$-spanner for the original metric space $(M,\delta)$ by the ``transitivity'' of spanners. A spanner $G'$ of degree $\eps^{-O(\ddim)}$ can be constructed in $\eps^{-O(\ddim)} (n \log n)$ time via \theoremref{thm:doubling_degree}. Since the output $t$-spanner $G$ for $(M,\delta)$ is a subgraph of $G'$, its degree will be at most $\eps^{-O(\ddim)}$. The greedy simulation is applied only on the edges of $G'$ that are sufficiently ``heavy''. Formally, let $D$ denote the maximum weight of any edge of the bounded degree spanner $G'$, and let $E_0$ be the set of \emph{light edges} in $E'$, namely, of weight at most $D/n$. As $|E_0| \le |E'| = O(n)$, we have $w(E_0) = O(D) = O(MST(M))$. All light edges are taken to the output spanner $G$, and the greedy simulation is applied only on the edges of $E' \setminus E_0$. (So the output spanner $G$ will contain \emph{all} edges of $E_0$ and \emph{some} edges of $E' \setminus E_0$.) As mentioned, computing the \emph{exact} distance between two points is costly; using Dijkstra's algorithm, it requires $O(n \log n)$ time (see, e.g., Section 2.5 and Corollary 2.5.10 in \cite{NS07}). Since $G'$ has $O(n)$ edges, the overall runtime will be $O(n^2 \log n)$. To speed up the computation time, Algorithm \texttt{Approximate-Greedy} does not compute the exact distance between two points, but rather an approximation of that distance. This is achieved by maintaining a much simpler and coarser \emph{cluster graph} that approximates the original distances, on which the distance queries are performed. More specifically, the algorithm partitions the edge set $E' \setminus E_0$ into $\log_\mu n$ buckets, for an appropriate parameter $1 < \mu = O(\log n)$, such that edge weights within each bucket differ by at most a factor of $\mu$. Then it examines the edges of $E' \setminus E_0$ by going from one bucket to the next, examining edges by non-decreasing order of weight. Whenever all edges of some bucket have been examined, the cluster graph is updated according to the new edges that were added to the spanner. The idea is to periodically make the cluster graph simpler and coarser, so that the shortest path computations made on it will be fast. The bottom-line is that one does not simulate the greedy algorithm (with stretch parameter $\sqrt{t \cdot t'}$) on the edge set $E' \setminus E_0$, but rather an \emph{approximate version} of it. \subsection{Bounding the Lightness of the \texttt{Approximate-Greedy} Spanner} \label{bounding} As mentioned, the runtime of Algorithm \texttt{Approximate-Greedy} is $\eps^{-O(\ddim)}(n \log n)$ in arbitrary doubling metrics. In what follows let $G = (M,E,\delta)$ be the $t$-spanner for $(M,\delta)$ returned by Algorithm \texttt{Approximate-Greedy}. Since $G$ is a subgraph of the bounded degree spanner $G' = (M,E',\delta)$, its degree is $\eps^{-O(\ddim)}$. It remains to bound the lightness of $G$. The lightness argument of \cite{GLN02}, which relies on previous works \cite{DHN93,DN97}, is based on rather deep properties from Euclidean geometry, most notably the \emph{leapfrog property}. In particular, this argument does not apply to arbitrary doubling metrics. Instead, we employ the following lemma, which lies at the heart of the lightness analysis of \cite{GLN02}. While this lemma applies to arbitrary doubling metrics, the way it was used in \cite{GLN02} does not extend to arbitrary doubling metrics. Specifically, it was used in \cite{GLN02} to show that the edge set $E \setminus E_0$ satisfies the leapfrog property. (Recall that $E_0$ is the set of light edges in $G' = (M,E',\delta)$, all of which are taken to the approximate-greedy spanner $G = (M,E,\delta)$.) In Euclidean metrics, it has been proved \cite{DHN93,NS07} that any edge set satisfying the leapfrog property has constant lightness, but this proof does not carry over to arbitrary doubling metrics. \begin{lemma}[Lemma 17 in \cite{GLN02}]\label{lem:GLN_second_path} Let $e=(u,v) \in E \setminus E_0$. The weight of the second shortest path between $u$ and $v$ in the approximate-greedy spanner $G$ is greater than $t' \cdot w(e)$. (If there are multiple shortest paths between $u$ and $v$, then the weight of the second shortest path equals the weight of the shortest path.) \end{lemma} \noindent {\bf Remark.} The parameter $t'$ in the statement of this lemma depends on the stretch parameters of the spanners $G'$ and $G$ that are constructed by Algorithm \texttt{Approximate-Greedy}. Specifically, recall that the output spanner $G$ is a $\sqrt{t \cdot t'}$-spanner for $G'$, which is, in turn, a $\sqrt{t/t'}$-spanner for the input metric space $M$. We make the following simple observation. \begin{observation} \label{mstsimple2} Let $H$ be an arbitrary weighted graph, and let $t$ be any stretch parameter. For any $t$-spanner $H'$ of $H$, $w(MST(H')) \le t \cdot w(MST(H))$. \end{observation} \begin{proof} Consider an MST $Z$ for $H$. Replace each edge of $Z$ by a $t$-spanner path in $H'$ between the endpoints of that edge, and then break cycles. The resulting structure $Z'$ is a spanning tree of $H'$, hence $w(MST(H')) \le w(Z')$, and we have $w(MST(H')) \le w(Z') \le t \cdot w(Z) = t \cdot w(MST(H))$. \end{proof} The following lemma bounds the lightness of $G$. Its proof is based on the somewhat surprising observation that the lightness of the $t$-spanner $G$ produced by Algorithm \texttt{Approximate-Greedy} is existentially near-optimal with respect to stretch parameter $t' < t$ (rather than $t$). We remark that $G$ is not a greedy spanner, but rather an approximate-greedy spanner, and it is inherently different than the greedy $t$-spanner and the greedy $t'$-spanner. In particular, its weight may be larger than the weights of both these greedy spanners. Nevertheless, our existential near-optimality argument suffices to derive the required lightness bound. \begin{lemma} The lightness of $G$ is $(\frac{\ddim}{t'-1})^{O(\ddim)}$. \end{lemma} \begin{proof} Recall that $G = (M,E,\delta)$ is a $t$-spanner for $M$, where $t = 1+\eps, \eps < 1$, and let $M_G$ be the metric space induced by $G$. By \observationref{lem:doubling_embedding}, the doubling dimension of $M_G$ is bounded by $2 \cdot \ddim$. Let $H'$ be the $t'$-spanner of $M_G$ with lightness $(\ddim/(t'-1))^{O(\ddim)}$ that is guaranteed by \theoremref{thm:Got15}, where $t' = 1+O(\eps) < t$ is the parameter appearing in the statement of \lemmaref{lem:GLN_second_path}, which is optimized as part of Algorithm \texttt{Approximate-Greedy}. As in the proof of \lemmaref{lightopt}, we transform $H'$ into a $t'$-spanner $H''$ of $G$ of weight at most $w(H')$. By \observationref{mst-metric} and \observationref{mstsimple2}, the MST weights for all graphs $M,G,M_G,H'$ and $H''$ are the same, up to a factor of $t \cdot t' = O(1)$. We argue that every edge $e \in E \setminus E_0$ belongs to $H''$. Suppose for contradiction that there is an edge $e \in E \setminus E_0$ that does not belong to $H''$. Let $P$ be a shortest path between the endpoints of $e$ in $H''$. Since $H''$ is a $t'$-spanner of $G$, we have $w(P) \le t' \cdot w(e)$. Note that this path is contained in $G$. Since $e \in G$ and $M$ is a metric space, the weight of the second shortest path between $u$ and $v$ is at most $w(P) \le t' \cdot w(e)$. On the other hand, By \lemmaref{lem:GLN_second_path}, the weight of this path is greater than $t'\cdot w(e)$, a contradiction. It follows that \begin{eqnarray*} w(G) & = & w(E\setminus E_{0}) +w(E_{0}) ~\le~ w(H'')+w(E_{0})\\ & \le & w(H')+w(E_{0}) ~=~ \left(\frac{\ddim}{t'-1}\right)^{O(\ddim)} \cdot w(MST(M))\\ \end{eqnarray*} \end{proof} Setting $t = 1+\eps$ and $t' = 1+c \cdot \eps$ (for an appropriate constant $c$; see \cite{GLN02,NS07}), we conclude: \begin{theorem} \label{finish} For any metric space $(M,\delta)$ with doubling dimension $\ddim$ and parameter $0<\epsilon<\frac{1}{2}$, Algorithm \texttt{Approximate-Greedy} returns a $(1+\eps)$-spanner with lightness $\left(\frac{\ddim}{\eps}\right)^{O(\ddim)}$ and degree $\eps^{-O(\ddim)}$. The runtime of Algorithm \texttt{Approximate-Greedy} is $\eps^{-O(\ddim)}(n \log n)$. \end{theorem} \noindent{\bf Remark.} \theoremref{finish} should be compared to \theoremref{thm:Got15} due to \cite{Got15}. Both constructions achieve the same lightness bound, but the degree and number of edges in the spanner construction of \cite{Got15} are unbounded. Moreover, the runtime of the construction of \cite{Got15} is $(\ddim/\epsilon)^{O(\ddim)} (n\log^{2}n)$, whereas that of \theoremref{finish} is $\eps^{-O(\ddim)}(n \log n)$. By combining the light spanner $H_1$ of \cite{Got15} with a bounded degree spanner $H_2$, one can obtain a spanner with constant degree and lightness. Specifically, such a spanner $\mathcal{H}$ is obtained by replacing each edge of $H_1$ with a shortest (or approximately shortest) path in $H_2$ between the endpoints of that edge. The lightness of the resulting spanner $\mathcal{H}$ will not exceed that of $H_1$ by much, whereas the degree bound will follow from that of $H_2$. There is a major problem with this approach: The runtime needed for computing spanner $\mathcal{H}$ may be very high. Indeed, although there are efficient ways to estimate the weight of an approximately shortest path in $H_2$ between two points, we must \emph{compute} the corresponding path in $H_2$. In particular, to achieve the degree bound of $H_2$, one may not use edges outside $H_2$. Moreover, even regardless of this computation time, such a path may contain many edges that already belong to the gradually growing spanner $\mathcal{H}$. Deciding which edges of this path should be added to $\mathcal{H}$ may be very costly by itself. \section{Acknowledgements} We are grateful to Michael Elkin and Ofer Neiman for fruitful discussions. \pagebreak {\small \bibliographystyle{alpha}

https://arxiv.org/abs/1801.10389

Generalized reverse Young and Heinz inequalities

In this paper, we study the further improvements of the reverse Young and Heinz inequalities for the wider range of $v$, namely $v\in \mathbb{R}$. These modified inequalities are used to establish corresponding operator inequalities on a Hilbert space.

\section{Introduction} The weighted arithmetic-geometric mean inequality, which is also called Young's inequality, states \begin{equation*} (1-v) a +v b \geq a^{1-v} b^v \end{equation*} for $a, b \geq 0$ and $v \in [0,1]$. If $v=\frac{1}{2}$, we obtain the arithmetic-geometric mean inequality \begin{equation*} \frac{a+b}{2}\geq \sqrt{ab}. \end{equation*} The Heinz means, introduced in \cite{bb}, are defined by \begin{equation*} H_{v}(a,b)=\frac{a^{1-v}b^{v}+a^{v}b^{1-v}}{2} \end{equation*} for $a,b\geq 0$ and $v\in [0,1]$. It is easy to see that \begin{equation*} \sqrt{ab}\leq H_{v}(a,b)\leq \frac{a+b}{2}\qquad v\in [0,1], \end{equation*} which is called Heinz inequality. Let $B(H)$ denote the $C^{*}$-algebra of all bounded linear operators on a complex Hilbert space $H$. A self-adjoint operator $A\in B(H)$ is called positive, and we write $A\geq 0$ if $\langle Ax,x\rangle \geq 0$ for all $x\in H$. The set of all positive operators is denoted by $B^{+}(H)$. The set of all invertible operators in $B^{+}(H)$ is denoted by $B^{++}(H)$. We say $A\geq B$ if $A-B\geq 0$. Let $A, B\in B^{++}(H)$ and $v\in [0,1]$. The $v$-weighted operator geometric mean of $A$ and $B$, denoted by $A\sharp_{v} B$, is defined as \begin{equation*} A\sharp_{v} B=A^{\frac{1}{2}}\left(A^{-\frac{1}{2}}B A^{-\frac{1}{2}}\right)^{v} A^{\frac{1}{2}} \end{equation*} and the $v$-weighted operator arithmetic mean of $A$ and $B$, denoted by $A\nabla_{v} B$, is \begin{equation*} A\nabla_{v} B= (1-v)A+vB. \end{equation*} When $v=\frac{1}{2}$, $A\sharp_{\frac{1}{2}}B$ and $A\nabla_{\frac{1}{2}} B$ are called operator geometric mean and operator arithmetic mean, and denoted by $A\sharp B$ and $A\nabla B$, respectively \cite{nn}. For positive operators $A, B$ and $v\in[0,1]$, we have \cite[Definition 5.2]{Mond} \begin{equation*} A\sharp_{v} B=B\sharp_{1-v} A. \end{equation*} It is well known that if $A, B\in B^{++}(H)$ and $v\in [0,1]$, then \cite{ee,ff} \begin{align*} A\nabla_{v} B \geq A\sharp_{v} B, \end{align*} which is the operator version of the scalar Young's inequality. An operator version of Heinz means was introduced in \cite{bb} by \begin{equation*} H_{v}(A,B)=\frac{A\sharp_{v} B+A\sharp_{1-v}B}{2}, \end{equation*} where $v\in[0,1]$. In particular $H_{1}(A,B)=H_{0}(A,B)=A\nabla B$. It is easy to see that the Heinz operator means interpolate between the arithmetic and geometric operator means \begin{equation*} A\sharp B\leq H_{v}(A,B)\leq A\nabla B, \end{equation*} which is called the Heinz operator inequality \cite{ii,jj}. We note that we use in Sect. 3 the following notations $$A\natural_v B \equiv A^{\frac{1}{2}}\left(A^{-\frac{1}{2}}B A^{-\frac{1}{2}}\right)^{v} A^{\frac{1}{2}},\quad \hat{H}_v(A,B)\equiv \frac{A\natural_v B+A\natural_{1-v}B}{2},$$ for all $v \in \mathbb{R}$ including the range $v \notin [0,1]$. Improvements of Young and Heinz inequalities and their reverses have been done for the weight $v\in [0,1]$ by many researchers. We refer the reader to \cite{AFK2015,aa,D2015_03,cc,FM2011,dd,xx,gg,hh,ii,jj,kk,ll,mm,qq,SC,SM,SS,oo} as a sample of the extensive use of Young and Heinz inequalities. One of the first refinements is as follows in Lemma \ref{l01} which one positive term was added to the right-hand side of the Young's inequality. \begin{lemma}\textsc{(\cite{kk})}\label{l01} Let $a,b\geq 0$ and $v\in[0,1]$. Then \begin{align*} (1-v)a+vb &\geq a^{1-v}b^{v}+r_{0}\left(\sqrt{a}-\sqrt{b}\right)^{2} \end{align*} where $r_{0}=\min\lbrace v,1-v \rbrace$. \end{lemma} However, in the recent paper \cite{oo}, Zhao and Wu provided two refining terms of Young's inequality in the following way. \begin{lemma}\textsc{(\cite{oo})}\label{p1} Let $a,b\geq 0$ and $v\in[0,1]$. \begin{enumerate}[(i)] \item If $v \in [0,\frac{1}{2}]$, then \begin{align*} (1-v) a+v b \geq a^{1-v} b^v +v (\sqrt{a}-\sqrt{b})^2 +r_{0}(\sqrt{a}-\sqrt[4]{ab})^2, \end{align*} \item If $v \in [\frac{1}{2},1]$, then \begin{align*} (1-v) a+v b \geq a^{1-v} b^v + (1-v) (\sqrt{a}-\sqrt{b})^2 +r_{0}(\sqrt{b}-\sqrt[4]{ab})^2, \end{align*} \end{enumerate} where $r=\min\lbrace v,1-v \rbrace$ and $r_{0}=\min\lbrace 2r,1-2r \rbrace$. \end{lemma} In the same paper, the following refined reverse versions have been proved too. \begin{lemma}\textsc{(\cite[Lemma 2]{oo})}\label{l1} Let $a,b\geq 0$ and $v\in[0,1]$. \begin{enumerate}[(i)] \item If $v \in [0,\frac{1}{2}]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+(1-v)(\sqrt{a}-\sqrt{b})^{2}-r_{0}(\sqrt{b}-\sqrt[4]{ab})^{2}, \end{align*} \item If $v \in [\frac{1}{2},1]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+v(\sqrt{a}-\sqrt{b})^{2}-r_{0}(\sqrt{a}-\sqrt[4]{ab})^{2}, \end{align*} \end{enumerate} where $r=\min\lbrace v,1-v \rbrace$ and $r_{0}=\min\lbrace 2r,1-2r \rbrace$. \end{lemma} Sababheh and Choi \cite{SC} obtained a complete refinement of the Young's inequality by adding as many refining terms as we like. For $a,b> 0$, $n \in \mathbb{N}$ and $v\in[0,1]$ \begin{align*} (1-v)a+vb &\geq a^{1-v}b^{v}\\ &\quad + \sum_{k=1}^n s_{k}(v) \left( \sqrt[2^k]{a^{2^{k-1}-j_{k}(v)} b^{j_{k}(v)}} -\sqrt[2^k]{a^{2^{k-1}-j_{k}(v)-1} b^{j_{k}(v)+1} }\right)^2, \end{align*} where $[x]$ is the greatest integer less than or equal to $x$ and \begin{align*} j_{k}(v)&=[2^{k-1}v],\\ r_{k}(v)&=[2^{k} v],\\ s_{k}(v)&=(-1)^{r_{k}(v)}2^{k-1}v+(-1)^{r_{k}(v)+1}\left[\frac{r_{k}(v)+1}{2}\right]. \end{align*} Quite recently, Sababheh and Moslehian \cite{SM} gave a full description of all other refinements of the reverse Young's inequality in the literature as follows. \begin{lemma}\textsc{( \cite[Theorem 2.1]{SM})}\label{SM} Let $a,b> 0$ and $v\in[0,1]$. \begin{enumerate}[(i)] \item If $v \in [0,\frac{1}{2}]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+(1-v)(\sqrt{a}-\sqrt{b})^{2}-S_{n}(2v, \sqrt{ab}, b). \end{align*} \item If $v \in [\frac{1}{2},1]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+v(\sqrt{a}-\sqrt{b})^{2}-S_{n}(2(1-v), \sqrt{ab}, a). \end{align*} \end{enumerate} Where \begin{align*} S_{n}(v, a, b)&=\sum_{k=1}^n s_{k}(v) \left( \sqrt[2^k]{b^{2^{k-1}-j_{k}(v)} a^{j_{k}(v)}} -\sqrt[2^k]{a^{j_{k}(v)+1} b^{2^{k-1}-j_{k}(v)-1}}\right)^2,\\ j_{k}(v)&=[2^{k-1}v],\\ r_{k}(v)&=[2^{k} v],\\ s_{k}(v)&=(-1)^{r_{k}(v)}2^{k-1}v+(-1)^{r_{k}(v)+1}\left[\frac{r_{k}(v)+1}{2}\right]. \end{align*} \end{lemma} In the study of Young's inequalities, supplemental Young's inequality $$ a^v v^{1-v} \geq v a + (1-v) b $$ for $a, b>0$ and $v \notin[0,1]$ is often discussed. Our main idea in this paper is to extend the range of $v$ and to give the tighter bounds of the reverse Young's inequalities proved in \cite{SM} and \cite{oo}. In Theorem \ref{t2}, we will obtain a new generalization of the reverse Young's inequality which is stronger than the reverse Young's inequalities shown in \cite[Theorem 2.1]{SM} and \cite[Lemma 2]{oo}. Theorem \ref{t11} is another refinement of \cite[Theorem 2.9]{SC} which extend the range of $v$. In Sect. 3, these modified inequalities are used to establish corresponding operator inequalities on a Hilbert space. We emphasize that the significance of the inequalities in this paper is to have the wider range, namely $v\in \mathbb{R}$, and tighter bounds. \section{Generalizations of the Reverse Scalar Young and Heinz Inequalities} In this section, we present the numerical inequalities needed to prove the operator versions. We start from the following lemma to prove our main result. \begin{lemma}\textsc{(\cite[Lemma 2.1]{aa})}\label{l0} Let $a,b> 0$ and $v\notin [0,1]$. Then \begin{align*} (1-v)a+vb \leq a^{1-v}b^{v}. \end{align*} \end{lemma} \begin{corollary}\label{c2} Let $a,b> 0$ and $\frac{1}{2}\neq v\in \mathbb{R}$. \begin{enumerate}[(i)] \item If $v \notin [0,\frac{1}{2}]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+v(\sqrt{a}-\sqrt{b})^{2}. \end{align*} \item If $v \notin [\frac{1}{2},1]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+(1-v)(\sqrt{a}-\sqrt{b})^{2}. \end{align*} \end{enumerate} \end{corollary} \begin{proof} \textit {(i)} If $v \notin [0,\frac{1}{2}]$, then \begin{align*} &(1-v)a+vb-v(\sqrt{a}-\sqrt{b})^{2}\\ &=(1-2v)a+(2v)\sqrt{ab}\\ &\leq a^{(1-2v)} (\sqrt{ab})^{2v}\qquad \text{( by Lemma \ref{l0})}\\ &=a^{1-v}b^{v}. \end{align*} \textit{(ii)} If $v \notin [\frac{1}{2},1]$, then $(1-v) \notin [0,\frac{1}{2}]$. So by changing two elements $a,b$ and two weights $v,1-v$ in \textit {(i)}, the desired inequality is obtained.\qed \end{proof} Next, we represent our main result which is the reverse Young's inequality for $v\in \mathbb{R}$. \begin{theorem}\label{t2} Let $a,b > 0$, $n \in \mathbb{N}$ such that $n\geq 2$ and $\frac{1}{2}\neq v\in \mathbb{R}$. Then, \begin{enumerate}[(i)] \item If $v \notin [\frac{1}{2},\frac{2^{n-1}+1}{2^n}]$, then \begin{align}\label{eq5} (1-v) a+v b &\leq a^{1-v} b^v+(1-v)(\sqrt{a}-\sqrt{b})^{2}\notag\\ &\qquad +(2v-1)\sqrt{ab}\sum_{k=2}^n2^{k-2}\left(\sqrt[2^{k}]{\frac{b}{a}}-1\right)^{2}. \end{align} \item If $v \notin [\frac{2^{n-1}-1}{2^n},\frac{1}{2}]$, then \begin{align}\label{eq6} (1-v) a+v b &\leq a^{1-v} b^v +v(\sqrt{a}-\sqrt{b})^{2}\notag\\ &\qquad +(1-2v)\sqrt{ab}\sum_{k=2}^n2^{k-2}\left(\sqrt[2^{k}]{\frac{a}{b}}-1\right)^{2}. \end{align} \end{enumerate} \end{theorem} \begin{proof} \textit{(i)} If $v \notin [\frac{1}{2},\frac{2^{n-1}+1}{2^n}]$, we have $(2^{n}v-2^{n-1})\notin [0,1]$ and $(2^{n-1}-2^{n}v+1)\notin [0,1]$. Now compute \begin{align*} &(1-v) a+v b-(1-v)(\sqrt{a}-\sqrt{b})^{2}-(2v-1)\sqrt{ab}\left\{\sum_{k=2}^n2^{k-2}\left(\sqrt[2^{k}]{\frac{b}{a}}-1\right)^{2}\right\}\\ &=(1-v) a+v b-(1-v)(\sqrt{a}-\sqrt{b})^{2}\\ &\qquad -(2v-1)\sqrt{ab}\left\{\left(\sqrt[4]{\frac{b}{a}}-1\right)^{2}+2\left(\sqrt[8]{\frac{b}{a}}-1\right)^{2}+4\left(\sqrt[16]{\frac{b}{a}}-1\right)^{2}\right\}\\ &\qquad -\cdots -2^{n-4}(2v-1)\sqrt{ab}\left(\sqrt[2^{n-2}]{\frac{b}{a}}-1\right)^{2}\\ &\qquad -2^{n-3}(2v-1)\sqrt{ab}\left(\sqrt[2^{n-1}]{\frac{b}{a}}-1\right)^{2} -2^{n-2}(2v-1)\sqrt{ab}\left(\sqrt[2^{n}]{\frac{b}{a}}-1\right)^{2}\\ &=(1-v) a+v b-(1-v)(a-2\sqrt{ab}+b)-(2v-1)\sqrt{ab}\left(\sqrt{\frac{b}{a}}-2\sqrt[4]{\frac{b}{a}}+1\right)\\ &\qquad -2(2v-1)\sqrt{ab}\left(\sqrt[4]{\frac{b}{a}}-2\sqrt[8]{\frac{b}{a}}+1\right)-4(2v-1)\sqrt{ab}\left(\sqrt[8]{\frac{b}{a}}-2\sqrt[16]{\frac{b}{a}}+1\right)\\ &\qquad -\cdots -2^{n-4}(2v-1)\sqrt{ab}\left(\sqrt[2^{n-3}]{\frac{b}{a}}-2\sqrt[2^{n-2}]{\frac{b}{a}}+1\right)\\ &\qquad -2^{n-3}(2v-1)\sqrt{ab}\left(\sqrt[2^{n-2}]{\frac{b}{a}}-2\sqrt[2^{n-1}]{\frac{b}{a}}+1\right)\\ &\qquad -2^{n-2}(2v-1)\sqrt{ab}\left(\sqrt[2^{n-1}]{\frac{b}{a}}-2\sqrt[2^{n}]{\frac{b}{a}}+1\right)\\ &=2(1-v)\sqrt{ab}+(1-2v)\sqrt{ab}\sum_{l=0}^{n-2}2^l+2^{n-1}(2v-1)\sqrt{ab}\sqrt[2^{n}]{\frac{b}{a}}\\ &=\left\{2(1-v)+(1-2v)\sum_{l=0}^{n-2}2^l \right\}\sqrt{ab}+2^{n-1}(2v-1)\sqrt{ab}\sqrt[2^{n}]{\frac{b}{a}}\\ &=\left\{2(1-v)+(1-2v)(2^{n-1}-1) \right\}\sqrt{ab}+2^{n-1}(2v-1)\sqrt{ab}\sqrt[2^{n}]{\frac{b}{a}}\\ &=(2^{n-1}-2^{n}v+1)\sqrt{ab}+(2^{n}v-2^{n-1})\sqrt{ab}\sqrt[2^{n}]{\frac{b}{a}}\\ &\leq \left(\sqrt{ab}\right)^{(2^{n-1}-2^{n}v+1)}\left( \sqrt{ab}\sqrt[2^{n}]{\frac{b}{a}} \right)^{(2^{n}v-2^{n-1})}\: \text{( by Lemma \ref{l0})}\\ &=a^{1-v}v^{v}. \end{align*} So we get the following inequality \begin{align*} &(1-v) a+v b-(1-v)(\sqrt{a}-\sqrt{b})^{2}-(2v-1)\sqrt{ab}\left\{\sum_{k=2}^n2^{k-2}\left(\sqrt[2^{k}]{\frac{b}{a}}-1\right)^{2}\right\}\\ & \qquad \leq a^{1-v}v^{v} \end{align*} which is equivalent to \eqref{eq5}. \textit{(ii)} If $v \notin [\frac{2^{n-1}-1}{2^n},\frac{1}{2}]$, then $(1-v)\notin [\frac{1}{2},\frac{2^{n-1}+1}{2^n}]$. Now by changing two elements $a,b$ and replacing the weight $v$ with $(1-v)$ in \textit{(i)}, the desired inequality \eqref{eq6} is deduced.\qed \end{proof} \begin{remark} We would remark that if we rewrite Theorem \ref{t2} for $n=1$, then we get Corollary \ref{c2}. \end{remark} \begin{remark}\label{RM} From the equality of the proof in Theorem \ref{t2}, the inequality \eqref{eq5} is equivalent to \begin{align}\label{ineq0002} a^{1-v} b^v &\geq \sqrt{ab}+2^{n}\left(v-\frac{1}{2}\right)\sqrt{ab} \left(\sqrt[2^n]{\frac{b}{a}}-1\right), \end{align} which gives the following inequality \begin{align*} \left( \frac{b}{a} \right)^{v-\frac{1}{2}}\geq 1+2^n\left(v - \frac{1}{2} \right) \left(\sqrt[2^n]{\frac{b}{a}}-1 \right). \end{align*} Since $\lim_{r\to 0} \frac{t^r -1}{r} =\log t$, by putting $r=\frac{1}{2^{n}}$, we have \begin{align*} \lim_{n\to \infty}2^{n}\left(\sqrt[2^n]{\frac{b}{a}}-1\right) = \log \frac{b}{a}. \end{align*} Thus, we have the following inequality in the limit of $n \to \infty$ for the inequality \eqref{eq5} in Theorem \ref{t2}: \begin{equation} \label{ineq0001} \log \left( \frac{b}{a} \right)^{v-\frac{1}{2}} \leq \left( \frac{b}{a} \right)^{v-\frac{1}{2}}-1, \end{equation} for $a,b > 0$ and $\frac{1}{2} \neq v \in \mathbb{R}$, which comes from the condition $v \notin \left[\frac{1}{2},\frac{2^{n-1}+1}{2^n} \right]$ in the limit of $n \to \infty$. The above inequality recover the equality in the case $v =\frac{1}{2}$. Therefore, we have the inequality \eqref{ineq0001} for all $v \in \mathbb{R}$. We notice that the inequality \eqref{ineq0001} can be proven directly by putting $x=\left(\frac{b}{a}\right)^{v-1/2}$ in the inequality \begin{equation} \label{fundamental_log_ineq} \log x \leq x-1,\quad (x>0). \end{equation} Similarly in the limit of $n \to \infty$ for the inequality \eqref{eq6} in Theorem \ref{t2}, we get the inequality $$ \log \left( \frac{a}{b} \right)^{\frac{1}{2}-v} \leq \left( \frac{a}{b} \right)^{\frac{1}{2}-v}-1, $$ by changing two elements $a,b$ and replacing the weight $v$ with $(1-v)$ in the inequality \eqref{ineq0001}. \end{remark} Next, in Remarks \ref{r1} and \ref{r2}, we show that Theorem \ref{t2} recover the inequalities in Lemma \ref{l1}. To achieve this, we compare Lemma \ref{l1} with Theorem \ref{t2} in the cases such as $n=2$ and $n=3$ where $v\in[0,1]$. First, we notice that Lemma \ref{l1} is equivalent to the following proposition. \begin{prop}\label{p3} Let $a,b\geq 0$ and $v\in [0,1]$. \begin{enumerate}[(i)] \item If $v\in[0,\frac{1}{4}]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+(1-v)(\sqrt{a}-\sqrt{b})^{2}-2v(\sqrt{b}-\sqrt[4]{ab})^{2}. \end{align*} \item If $v\in[\frac{1}{4},\frac{1}{2}]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+(1-v)(\sqrt{a}-\sqrt{b})^{2}+(2v-1)(\sqrt{b}-\sqrt[4]{ab})^{2}. \end{align*} \item If $v\in[\frac{1}{2},\frac{3}{4}]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+v(\sqrt{a}-\sqrt{b})^{2}-(2v-1)(\sqrt{a}-\sqrt[4]{ab})^{2}. \end{align*} \item If $v\in[\frac{3}{4},1]$, then \begin{align*} (1-v)a+vb &\leq a^{1-v}b^{v}+v(\sqrt{a}-\sqrt{b})^{2}+(2v-2)(\sqrt{a}-\sqrt[4]{ab})^{2}. \end{align*} \end{enumerate} \end{prop} \begin{remark}\label{r1} Consider Theorem \ref{t2} in the case $n=2$ with $v\in [0,1]$. For $a,b> 0$, we have the following inequalities \begin{enumerate}[\textit{(i)}] \item If $v\notin[\frac{1}{2},\frac{3}{4}]$, then \begin{align} (1-v)a+vb &\leq a^{1-v}b^{v}+(1-v)(\sqrt{a}-\sqrt{b})^{2}\notag \\ & \qquad +(2v-1)(\sqrt{b}-\sqrt[4]{ab})^{2}.\label{eq7} \end{align} \item If $v\notin[\frac{1}{4},\frac{1}{2}]$, then \begin{align} (1-v)a+vb &\leq a^{1-v}b^{v}+v(\sqrt{a}-\sqrt{b})^{2}\notag \\ & \qquad -(2v-1)(\sqrt{a}-\sqrt[4]{ab})^{2}.\label{eq8} \end{align} \end{enumerate} In our recent paper \cite{xx}, we showed that the right-hand sides of both inequalities (\ref{eq7}) and (\ref{eq8}) give tighter upper bounds of the $v$-weighted arithmetic mean than those in Proposition \ref{p3}. \end{remark} \begin{remark}\label{r2} As a direct consequence of Theorem \ref{t2} in the case $n=3$ with restricted range $v\in[0,1]$, we have \begin{enumerate}[\textit{(i)}] \item If $v\notin[\frac{1}{2},\frac{5}{8}]$, then \begin{align}\label{eq9} (1-v)a+vb &\leq a^{1-v}b^{v}+(1-v)(\sqrt{a}-\sqrt{b})^{2}+(2v-1)(\sqrt{b}-\sqrt[4]{ab})^{2}\notag \\ &\qquad +(4v-2)(\sqrt[8]{ab^{3}}-\sqrt[4]{ab})^{2}. \end{align} \item If $v\notin[\frac{3}{8},\frac{1}{2}]$, then \begin{align}\label{eq10} (1-v)a+vb &\leq a^{1-v}b^{v}+v(\sqrt{a}-\sqrt{b})^{2}-(2v-1)(\sqrt{a}-\sqrt[4]{ab})^{2}\notag \\ &\qquad -(4v-2)(\sqrt[8]{a^{3}b}-\sqrt[4]{ab})^{2}. \end{align} \end{enumerate} We here give advantages of inequalities \eqref{eq9} and \eqref{eq10} in comparison with Proposition \ref{p3}. \begin{itemize} \item[(a)] Firstly, we compare Proposition \ref{p3} with the inequality \eqref{eq9} which holds in the cases $v \in [0,\frac{1}{4}]$, $v \in [\frac{1}{4},\frac{1}{2}]$, $v \in [\frac{5}{8},\frac{3}{4}]$ and $v \in [\frac{3}{4},1]$ . \begin{itemize} \item[(a1)] In the case $v \in [0,\frac{1}{4}]$, we have $(2v-1)<(-2v)$ and $(4v-2)<0$. Indeed, the right-hand side of inequality \eqref{eq9} is less than the right-hand side of \textit{(i)} in Proposition \ref{p3}. For the case of $v \in [\frac{1}{4},\frac{1}{2}]$, we have $(4v-2)<0$. So we easily find that the right-hand side of inequality \eqref{eq9} is less than the right-hand side of \textit{(ii)} in Proposition \ref{p3}. \item[(a2)]For the case of $v \in [\frac{5}{8},\frac{3}{4}]$, we claim that the right-hand side of inequality \eqref{eq9} is less than or equal to the right-hand side of \textit{(iii)} in Proposition \ref{p3}. To prove our claim, we show that the following inequality holds: \begin{align}\label{eq100} &(1-v)(\sqrt{a}-\sqrt{b})^{2}+(2v-1)(\sqrt{b}-\sqrt[4]{ab})^{2}+(4v-2)(\sqrt[8]{ab^{3}}-\sqrt[4]{ab})^{2} \notag \\ &\qquad \leq v(\sqrt{a}-\sqrt{b})^{2}-(2v-1)(\sqrt{a}-\sqrt[4]{ab})^{2}, \end{align} which is equivalent to the inequality \begin{equation*} 2(1-2v) t^{1/4}\left\{3t^{1/4}-1-2t^{3/8}\right\}\geq 0, \end{equation*} for $t>0$ and $v \in [\frac{5}{8},\frac{3}{4}]$. To obtain this, it is enough to prove $(3t^{1/4}-1-2t^{3/8})\leq 0$. If $t^{1/8}=x$, then we easily find that $f(x)=3x^2-2x^3-1$ is increasing for $0<x<1$ and decreasing where $x>1$. Indeed, $f(x)=3x^2-2x^3-1\leq 0$ where $x>0$ and so $3t^{1/4}-1-2t^{3/8} \leq 0$. \item[(a3)]For the case of $v \in [\frac{3}{4},1]$, we claim that the right-hand side of inequality \eqref{eq9} is less than or equal to the right-hand side of \textit{(iv)} in Proposition \ref{p3}. To prove our claim, we show that the following inequality holds: \begin{align}\label{eq101} &(1-v)(\sqrt{a}-\sqrt{b})^{2}+(2v-1)(\sqrt{b}-\sqrt[4]{ab})^{2}+(4v-2)(\sqrt[8]{ab^{3}}-\sqrt[4]{ab})^{2}\notag \\ &\qquad \leq v(\sqrt{a}-\sqrt{b})^{2}-2(1-v)(\sqrt{a}-\sqrt[4]{ab})^{2}, \end{align} which is equivalent to the inequality \begin{equation}\label{eq11} (4v-3)+(3-8v)t^{1/2}+(4-4v)t^{1/4}+(8v-4)t^{5/8}\geq 0, \end{equation} for $t>0$ and $v \in [\frac{3}{4},1]$. To obtain the inequality \eqref{eq11}, it is sufficient to prove $f(x,v) \geq 0$ where $ x=t^{1/8} >0$ and $$ f(x,v) \equiv (8v-4)x^5 +(3-8v) x^4 +(4-4v) x^2 +(4v-3). $$ Since $\frac{df(x,v)}{dv} = 4(x-1)^2 (2x^3+2x^2+2x+1) \geq 0$, we have $f(x,v) \geq f(x,\frac{3}{4}) =x^2 (x-1)^2(2x+1) \geq 0$ for $x>0$. \end{itemize} \item[(b)] Secondly, we compare Proposition \ref{p3} with the inequality \eqref{eq10} which holds in the cases $v \in [0,\frac{1}{4}]$, $v \in [\frac{1}{4},\frac{3}{8}]$, $v \in [\frac{1}{2},\frac{3}{4}]$ and $v \in [\frac{3}{4},1]$. \begin{itemize} \item[(b1)]For the case of $v \in [0,\frac{1}{4}]$, we claim that the right-hand side of inequality \eqref{eq10} is less than or equal to the right-hand side of \textit{(i)} in Proposition \ref{p3}. To prove our claim, we give the following inequality \begin{align*} &v(\sqrt{a}-\sqrt{b})^{2}-(2v-1)(\sqrt{a}-\sqrt[4]{ab})^{2}-(4v-2)(\sqrt[8]{a^{3}b}-\sqrt[4]{ab})^{2}\\ &\qquad \leq (1-v)(\sqrt{a}-\sqrt{b})^{2}-2v(\sqrt{b}-\sqrt[4]{ab})^{2}, \end{align*} which we get it by replacing $v$ with $(1-v)$ and changing the elements $a,b$ in the inequality \eqref{eq101} . \item[(b2)] For the case of $v \in [\frac{1}{4},\frac{3}{8}]$, we claim that the right-hand side of inequality \eqref{eq10} is less than or equal to the right-hand side of \textit{(ii)} in Proposition \ref{p3}. To prove our claim, we give the following inequality \begin{align*} &v(\sqrt{a}-\sqrt{b})^{2}-(2v-1)(\sqrt{a}-\sqrt[4]{ab})^{2}-(4v-2)(\sqrt[8]{a^{3}b}-\sqrt[4]{ab})^{2}\\ &\qquad \leq (1-v)(\sqrt{a}-\sqrt{b})^{2}-(1-2v)(\sqrt{b}-\sqrt[4]{ab})^{2}, \end{align*} which is deduced by replacing $v$ with $(1-v)$ and changing the elements $a,b$ in the inequality \eqref{eq100} . \item[(b3)] For the case of $v \in [\frac{1}{2},\frac{3}{4}]$, we have $-(4v-2)<0$. So we easily find that the right-hand side of inequality \eqref{eq10} is less than the right-hand side of \textit{(iii)} in Proposition \ref{p3}. In the case $v \in [\frac{3}{4},1]$ , we have $-(2v-1)<(2v-2)$ and $-(4v-2)<0$. That is the right-hand side of inequality \eqref{eq10} is less than the right-hand side of \textit{(iv)} in Proposition \ref{p3}. \end{itemize} \end{itemize} \end{remark} Thus, according to Remark \ref{r1} and Remark \ref{r2}, Theorem \ref{t2} recover Lemma \ref{l1}. We notice that the range of the reverse Young's inequalities in Theorem \ref{t2} is wider than Lemma \ref{l1}, namely Theorem \ref{t2} holds in the case $v \in \mathbb{R}$ and Lemma \ref{l1} holds for $v \in [0,1]$. \\ Next, we compare Theorem \ref{t2} with Lemma \ref{SM}. In Theorem \ref{t2}, we have (I) $v\leq 0$, (II) $0\leq v < \frac{2^{n-1}-1}{2^n}$, (III) $\frac{2^{n-1}-1}{2^n} \leq v \leq \frac{1}{2}$, (IV) $\frac{1}{2} \leq v \leq \frac{2^{n-1}+1}{2^n}$, (V) $\frac{2^{n-1}+1}{2^n} < v \leq 1$ and (VI) $v\geq 1$.\\ For the cases of (II) and (V), we claim that Theorem \ref{t2} has tighter upper bounds than those in Lemma \ref{SM}, while Lemma \ref{SM} recover Theorem \ref{t2} in the cases (III) and (IV). Therefore we conclude that Theorem \ref{t2} and Lemma \ref{SM} are different refinements of the reverse Young's inequality which both of them recover Lemma \ref{l1}. However, we emphasize that Theorem \ref{t2} gives the reverse Young's inequality in the wider range than Lemma \ref{SM}, namely in the case $v \in \mathbb{R}$. This justify why our refinement in Theorem \ref{t2} is better than Lemma \ref{SM}.\\ To prove our claims, we compare Theorem \ref{t2} with Lemma \ref{SM} in the same steps such as $n=2$. For this purpose, we list up the following corollaries which are deduced directly from Theorem \ref{t2} and Lemma \ref{SM}, respectively. \begin{corollary}\label{c1} Let $a,b> 0$ and $v\in \mathbb{R}$. \begin{enumerate}[(i)] \item If $v\notin[\frac{1}{2},\frac{3}{4}]$, then \begin{align} (1-v)a+vb - a^{1-v}b^{v} &\leq (1-v)(\sqrt{a}-\sqrt{b})^{2}+(2v-1)(\sqrt{b}-\sqrt[4]{ab})^{2}.\label{eq19} \end{align} \item If $v\notin[\frac{1}{4},\frac{1}{2}]$, then \begin{align} (1-v)a+vb - a^{1-v}b^{v} &\leq v(\sqrt{a}-\sqrt{b})^{2}-(2v-1)(\sqrt{a}-\sqrt[4]{ab})^{2}.\label{eq20} \end{align} \end{enumerate} \end{corollary} \begin{corollary}\label{c11} Let $a,b> 0$ and $v\in[0,1]$. \begin{enumerate}[(i)] \item If $v\in [0,\frac{1}{4}]$, then \begin{align} (1-v)a+vb - a^{1-v}b^{v} &\leq (1-v)(\sqrt{a}-\sqrt{b})^{2}-2v(\sqrt{b}-\sqrt[4]{ab})^{2}\notag \\ &\qquad -4v(\sqrt{b}-\sqrt[8]{ab^3})^{2}.\label{eq15} \end{align} \item If $v\in[\frac{1}{4},\frac{1}{2}]$, then \begin{align} (1-v)a+vb - a^{1-v}b^{v} &\leq (1-v)(\sqrt{a}-\sqrt{b})^{2}+(2v-1)(\sqrt{b}-\sqrt[4]{ab})^{2}\notag \\ &\qquad-(4v-1)(\sqrt[8]{ab^3} -\sqrt[4]{ab})^{2}.\label{eq16} \end{align} \item If $v\in[\frac{1}{2},\frac{3}{4}]$, then \begin{align} (1-v)a+vb - a^{1-v}b^{v} &\leq v(\sqrt{a}-\sqrt{b})^{2}-(2v-1)(\sqrt{a}-\sqrt[4]{ab})^{2}\notag \\ &\qquad+(4v-3)(\sqrt[8]{ab^3} -\sqrt[4]{ab})^{2}.\label{eq17} \end{align} \item If $v\in[\frac{3}{4},1]$, then \begin{align} (1-v)a+vb - a^{1-v}b^{v} &\leq v(\sqrt{a}-\sqrt{b})^{2}+(2v-2)(\sqrt{a}-\sqrt[4]{ab})^{2}\notag \\ &\qquad-4(1-v)(\sqrt[8]{ab^3} -\sqrt{a})^{2}.\label{eq18} \end{align} \end{enumerate} \end{corollary} \begin{remark} Here, we compare the upper bounds in Corollary \ref{c1} with those in Corollary \ref{c11}. Firstly, we compare Corollary \ref{c11} with the inequality (\ref{eq19}) which holds in the cases $v\in[0,\frac{1}{4}]$, $v\in[\frac{1}{4},\frac{1}{2}]$ and $v\in[\frac{3}{4},1]$.\\ For the case of $v\in[0,\frac{1}{4}]$, we can find examples such that the right-hand side of (\ref{eq19}) is tighter than that of (\ref{eq15}) in Corollary \ref{c11}. Actually, take $a=1$, $b=16$ and $v=1/8$, then the right-hand side of (\ref{eq19}) is equal to $4.875$, while the right-hand side of (\ref{eq15}) is nearly equal to $6.2892$. Indeed, the inequality (\ref{eq19}) can recover Corollary \ref{c11} where $v\in[0,\frac{1}{4}]$.\\ In the case $v\in[\frac{3}{4},1]$, we claim that the right-hand side of the inequality (\ref{eq19}) is less than or equal to the right-hand side of (\ref{eq18}). According to (\ref{eq101}), we have \begin{align*} &(1-v)(\sqrt{a}-\sqrt{b})^{2}+(2v-1)(\sqrt{b}-\sqrt[4]{ab})^{2}\notag \\ &\qquad \leq v(\sqrt{a}-\sqrt{b})^{2}+(2v-2)(\sqrt{a}-\sqrt[4]{ab})^{2}-(4v-2)(\sqrt[8]{ab^{3}}-\sqrt[4]{ab})^{2}. \end{align*} So to prove our claim, we show that the following inequality holds \begin{align*} (2-4v)(\sqrt[8]{ab^{3}}-\sqrt[4]{ab})^{2}\leq (4v-4)(\sqrt[8]{ab^{3}}-\sqrt{a})^{2}, \end{align*} which is equivalent to \begin{align*} g(x,v)=(4v-2)x^6+(2-4v)x^5+(2v-1)x^4+(2-4v)x^3+(2v-1)\geq 0, \end{align*} where $x=t^{\frac{1}{8}}>0$ and $v\in[\frac{3}{4},1]$. Since $\frac{dg(x,v)}{dv}=2(x-1)^2(2x^4+2x^3+3x^2+2x+1)\geq 0$, we have $g(x,v)\geq g(x,\frac{3}{4})=(x-1)^2(x^4+x^3+\frac{3}{2}x^2+x+\frac{1}{2})\geq 0$. This justify that the inequality (\ref{eq19}) also recover Corollary \ref{c11} where $v\in[\frac{3}{4},1]$. However, in the case $v\in[\frac{1}{4},\frac{1}{2}]$, we easily find that the right-hand side of (\ref{eq16}) is less than or equal to the right-hand side of (\ref{eq19}). That is Corollary \ref{c11} is a refinement of (\ref{eq19}) where $v\in[\frac{1}{4},\frac{1}{2}]$.\\ Secondly, we compare Corollary \ref{c11} with the inequality (\ref{eq20}) which holds in the cases $v\in[0,\frac{1}{4}]$, $v\in[\frac{1}{2},\frac{3}{4}]$ and $v\in[\frac{3}{4},1]$. The comparison is done by the same way as in the first step, and we omit it. \end{remark} Sababheh and Choi gave the following refinement of Lemma \ref{l0} which it's complete proof can be found in \cite[Theorem 2.2]{SM}. \begin{lemma}\textsc{( \cite[Theorem 2.9]{SC})}\label{t00} Let $a,b>0$. Then, we have \begin{enumerate}[(i)] \item If $v\leq 0$, then \begin{align}\label{SC1} (1-v) a+v b \leq a^{1-v} b^v + v \sum_{k=1}^n 2^{k-1} \left( \sqrt{a} -\sqrt[2^k]{a^{2^{k-1}-1}b}\right)^2. \end{align} \item If $v\geq 1$, then \begin{align}\label{SC2} (1-v) a+v b \leq a^{1-v} b^v + (1-v) \sum_{k=1}^n 2^{k-1} \left( \sqrt{b} -\sqrt[2^k]{ab^{2^{k-1}-1}}\right)^2. \end{align} \end{enumerate} \end{lemma} We can extend the ranges of $v$ in Lemma \ref{t00} to those in following theorem, by the similar way to the line of proof of Theorem \ref{t2}. \begin{theorem}\label{t11} Let $a,b >0$, $n \in \mathbb{N}$ and $v \in \mathbb{R}$. \begin{enumerate}[(i)] \item If $v \notin [0,\frac{1}{2^n}]$, then the inequality \eqref{SC1} holds. \item If $v \notin [\frac{2^{n}-1}{2^n},1]$, then the inequality \eqref{SC2} holds. \end{enumerate} \end{theorem} As we discussed the case of $n \to \infty$ in Remark \ref{RM}, we also give the following remark. \begin{remark}\label{RM2} The inequality (\ref{SC1}) is equivalent to the inequality $$ a+v a 2^n\left( \left( \frac{b}{a}\right)^{1/2^n} -1\right) \leq a^{1-v} b^v $$ by the elementary computations. Using the formula $\lim_{r\to 0} \frac{t^r-1}{r} = \log t$, we have the inequality $$ \log\left( \frac{b}{a} \right)^v \leq \left( \frac{b}{a} \right)^v -1 $$ for $v \neq 0$ in the limit of $n \to \infty$. The above inequality trivially holds for all $v \in \mathbb{R}$. It can be also obtained by the inequality (\ref{fundamental_log_ineq}). Similarly, the inequality (\ref{SC2}) is equivalent to the inequality $$ b+(1-v)b2^n \left( \left( \frac{a}{b}\right)^{1/2^n} -1\right) \leq a^{1-v} b^v $$ so that we have the inequality $$ \log\left( \frac{a}{b} \right)^{1-v }\leq \left( \frac{a}{b} \right)^{1-v} -1 $$ for $v \neq 1$ in the limit of $n \to \infty$. The above inequality trivially holds for all $v \in \mathbb{R}$. It can be also obtained by the inequality (\ref{fundamental_log_ineq}). \end{remark} As a direct consequence of Theorems \ref{t2} and \ref{t11}, we have the following reverse inequalities with respect to the Heinz means. \begin{corollary}\label{c00} Let $a,b > 0$, $n \in \mathbb{N}$ such that $n\geq 2$ and $\frac{1}{2}\neq v\in \mathbb{R}$. \begin{enumerate}[(i)] \item If $v \notin [\frac{1}{2},\frac{2^{n-1}+1}{2^n}]$, then \begin{align*} \frac{a+b}{2} &\leq H_{v}(a,b)+(1-v)(\sqrt{a}-\sqrt{b})^{2}\notag\\ &\qquad +\left(v-\frac{1}{2}\right)\sqrt{ab}\sum_{k=2}^n2^{k-2}\left\{\left(\sqrt[2^{k}]{\frac{a}{b}}-1\right)^{2}+\left(\sqrt[2^{k}]{\frac{b}{a}}-1\right)^{2} \right\} \end{align*} \item If $v \notin [\frac{2^{n-1}-1}{2^n},\frac{1}{2}]$, then \begin{align*} \frac{a+b}{2} &\leq H_{v}(a,b)+v(\sqrt{a}-\sqrt{b})^{2}\notag\\ &\qquad +\left(\frac{1}{2}-v\right)\sqrt{ab}\sum_{k=2}^n2^{k-2}\left\{\left(\sqrt[2^{k}]{\frac{a}{b}}-1\right)^{2}+\left(\sqrt[2^{k}]{\frac{b}{a}}-1\right)^{2} \right\}. \end{align*} \end{enumerate} \end{corollary} \begin{corollary}\label{c0011} Let $a,b > 0$, $n \in \mathbb{N}$ and $v \in \mathbb{R}$. \begin{enumerate}[(i)] \item If $v \notin [0,\frac{1}{2^n}]$, then \begin{align*} \frac{a+b}{2} &\leq H_{v}(a,b)\\ &\qquad + v \sum_{k=1}^n 2^{k-2} \left\lbrace \left( \sqrt{a} -\sqrt[2^k]{a^{2^{k-1}-1}b}\right)^2+\left( \sqrt{b} -\sqrt[2^k]{ab^{2^{k-1}-1}}\right)^2\right\rbrace . \end{align*} \item If $v \notin [\frac{2^{n}-1}{2^n},1]$, then \begin{align*} \frac{a+b}{2} &\leq H_{v}(a,b)\\ &\,\,+ (1-v) \sum_{k=1}^n 2^{k-2} \left\lbrace \left( \sqrt{a} -\sqrt[2^k]{a^{2^{k-1}-1}b}\right)^2+\left( \sqrt{b} -\sqrt[2^k]{ab^{2^{k-1}-1}}\right)^2\right\rbrace . \end{align*} \end{enumerate} \end{corollary} \section{Generalized Reverse Young and Heinz Inequalities for Operators} In this section by applying Kubo--Ando theory \cite{nn} and thanks to Theorems \ref{t2} and \ref{t11}, we have the following operator inequalities. \begin{theorem}\label{t6} Let $A, B\in B^{++}(H)$, $n \in \mathbb{N}$ such that $n\geq 2$ and $\frac{1}{2}\neq v\in \mathbb{R}$. Then, we have the following inequalities. \begin{enumerate}[(i)] \item If $v \notin [\frac{1}{2},\frac{2^{n-1}+1}{2^n}]$, then \begin{align*} A\nabla_{v} B &\leq A\natural_{v} B+2(1-v)(A\nabla_{v} B-A\sharp B)\\ &\qquad +(2v-1)\sum_{k=2}^{n}2^{k-2}\left( A\sharp B-2A\sharp_{\frac{2^{k-1}+1}{2^k}}B+A\sharp_{\frac{2^{k-2}+1}{2^{k-1}}}B\right). \end{align*} \item If $v \notin [\frac{2^{n-1}-1}{2^n},\frac{1}{2}]$, then \begin{align*} A\nabla_{v} B &\leq A\natural_{v} B+2v(A\nabla_{v} B-A\sharp B)\\ &\qquad +(1-2v)\sum_{k=2}^{n}2^{k-2}\left( A\sharp B-2A\sharp_{\frac{2^{k-1}-1}{2^k}}B+A\sharp_{\frac{2^{k-2}-1}{2^{k-1}}}B\right). \end{align*} \end{enumerate} \end{theorem} \begin{proof} \textit{(i)} According to the inequality \eqref{eq5}, the following inequality holds for $t\geq 0$ \begin{align*} (1-v)+v t &\leq t^v+(1-v)(1-\sqrt{t})^{2}\notag\\ &\qquad +(2v-1)\sqrt{t}\sum_{k=2}^n2^{k-2}\left(\sqrt[2^{k}]{t}-1\right)^{2}. \end{align*} By functional calculus, if we replace $t$ with $A^{-\frac{1}{2}}BA^{-\frac{1}{2}}$ and then multiplying both sides of the inequality by $A^{\frac{1}{2}}$, the desired inequality is obtained. \textit{(ii)} The line of proof is similar to \textit{(i)} by applying the inequality \eqref{eq6}.\qed \end{proof} \begin{remark} We notice that Theorem \ref{t6} with $v\in[0,1]$, recover the inequalities obtained in \cite[Theorem 3]{xx}, if we put $n=2$. \end{remark} \begin{theorem}\label{t6-6} Let $A, B\in B^{++}(H)$, $n \in \mathbb{N}$ and $v \in \mathbb{R}$. Then, we have the following inequalities. \begin{enumerate}[(i)] \item If $v \notin [0,\frac{1}{2^n}]$, then \begin{align*} A\nabla_{v} B &\leq A\natural_{v} B +v \sum_{k=1}^n 2^{k-1} \left( A-2A\sharp_{\frac{1}{2^k}}B+A\sharp_{\frac{1}{2^{k-1}}}B\right). \end{align*} \item If $v \notin [\frac{2^{n}-1}{2^n},1]$, then \begin{align*} A\nabla_{v} B &\leq A\natural_{v} B +(1-v) \sum_{k=1}^n 2^{k-1} \left( B-2A\sharp_{\frac{2^k-1}{2^k}}B+A\sharp_{\frac{2^{k-1}-1}{2^{k-1}}}B\right). \end{align*} \end{enumerate} \end{theorem} \begin{proof} \textit{(i)} According to Theorem \ref{t11} in the case $v \notin [0,\frac{1}{2^n}]$, the following inequality holds for $t\geq 0$ \begin{align*} (1-v)+v t &\leq t^v + v \sum_{k=1}^n 2^{k-1} \left( 1 -\sqrt[2^k]{t}\right)^2\\ &= t^v + v \sum_{k=1}^n 2^{k-1} \left( 1-2t^{\frac{1}{2^k}} +t^{\frac{1}{2^{k-1}}}\right). \end{align*} By functional calculus, if we replace $t$ with $A^{-\frac{1}{2}}BA^{-\frac{1}{2}}$ and then multiplying both sides of the inequality by $A^{\frac{1}{2}}$, the desired inequality is deduced. \textit{(ii)} By applying Theorem \ref{t11} for the case of $v \notin [\frac{2^{n}-1}{2^n},1]$, we get the desired inequality in the same way as in \textit{(i)}. \qed \end{proof} As a direct consequence of Theorems \ref{t6} and \ref{t6-6}, we get the generalized reverse Heinz operator inequalities as follows. That is Corollaries \ref{c3} and \ref{c3-3} are operator versions of Corollaries \ref{c00} and \ref{c0011} respectively. \begin{corollary}\label{c3} Let $A, B\in B^{++}(H)$, $n \in \mathbb{N}$ such that $n\geq 2$ and $\frac{1}{2}\neq v\in \mathbb{R}$. Then we have the following inequalities. \begin{enumerate}[(i)] \item If $v \notin [\frac{1}{2},\frac{2^{n-1}+1}{2^n}]$, then \begin{align*} A\nabla B &\leq \hat{H}_{v}(A,B) +2(1-v)(A\nabla B-A\sharp B)\\ &\quad +(2v-1)\sum_{k=2}^{n}2^{k-2}\left( A\sharp B-2H_{\frac{2^{k-1}+1}{2^k}}(A,B)+H_{\frac{2^{k-2}+1}{2^{k-1}}}(A,B)\right). \end{align*} \item If $v \notin [\frac{2^{n-1}-1}{2^n},\frac{1}{2}]$, then \begin{align*} A\nabla B &\leq \hat{H}_{v}(A,B) +2v(A\nabla B-A\sharp B)\\ &\quad +(1-2v)\sum_{k=2}^{n}2^{k-2}\left( A\sharp B-2H_{\frac{2^{k-1}-1}{2^k}}(A,B)+H_{\frac{2^{k-2}-1}{2^{k-1}}}(A,B)\right). \end{align*} \end{enumerate} \end{corollary} \begin{remark} Putting $n=2$ in Corollary \ref{c3} with $v\in[0,1]$ gives the inequalities obtained in \cite[Corollary 6]{xx}. \end{remark} \begin{corollary}\label{c3-3} Let $A, B\in B^{++}(H)$, $n \in \mathbb{N}$ and $v \in \mathbb{R}$. Then, we have the following inequalities. \begin{enumerate}[(i)] \item If $v \notin [0,\frac{1}{2^n}]$, then \begin{align*} A\nabla B &\leq \hat{H}_{v}(A,B)\\ &\qquad +v \sum_{k=1}^n 2^{k-1} \left( A\nabla B-2H_{\frac{1}{2^k}}(A,B)+H_{\frac{1}{2^{k-1}}}(A,B)\right). \end{align*} \item If $v \notin [\frac{2^{n}-1}{2^n},1]$, then \begin{align*} A\nabla B &\leq \hat{H}_{v}(A,B)\\ & \quad +(1-v) \sum_{k=1}^n 2^{k-1} \left( A\nabla B-2H_{\frac{2^k-1}{2^k}}(A,B)+H_{\frac{2^{k-1}-1}{2^{k-1}}}(A,B)\right). \end{align*} \end{enumerate} \end{corollary} \begin{remark} If we put $n=1$, Corollary \ref{c3-3} gives the inequality shown in \cite[Theorem 3.1]{aa} where $v \notin [\frac{1}{2},1]$. \end{remark} \begin{acknowledgements} The authors express their gratitude to the editor-in-chief Prof. Rosihan M. Ali and the anonymous referees for their careful reading and detailed comments which have considerably improved the paper. \end{acknowledgements}

https://arxiv.org/abs/1604.04343

A Generalized Fundamental Matrix for Computing Fundamental Quantities of Markov Systems

As is well known, the fundamental matrix $(I - P + e \pi)^{-1}$ plays an important role in the performance analysis of Markov systems, where $P$ is the transition probability matrix, $e$ is the column vector of ones, and $\pi$ is the row vector of the steady state distribution. It is used to compute the performance potential (relative value function) of Markov decision processes under the average criterion, such as $g=(I - P + e \pi)^{-1} f$ where $g$ is the column vector of performance potentials and $f$ is the column vector of reward functions. However, we need to pre-compute $\pi$ before we can compute $(I - P + e \pi)^{-1}$. In this paper, we derive a generalization version of the fundamental matrix as $(I - P + e r)^{-1}$, where $r$ can be any given row vector satisfying $r e \neq 0$. With this generalized fundamental matrix, we can compute $g=(I - P + e r)^{-1} f$. The steady state distribution is computed as $\pi = r(I - P + e r)^{-1}$. The Q-factors at every state-action pair can also be computed in a similar way. These formulas may give some insights on further understanding how to efficiently compute or estimate the values of $g$, $\pi$, and Q-factors in Markov systems, which are fundamental quantities for the performance optimization of Markov systems.

\section{Introduction}\label{section_intro} Markov decision processes (MDPs) are widely adopted to model the dynamic decision problem in stochastic systems \cite{Bertsekas12,Puterman94}. The fundamental matrix $(\bm I - \bm P + \bm e \bm \pi)^{-1}$ plays a key role in the performance optimization of MDPs. With the fundamental matrix, we can further study the properties of Markov systems. For example, we can use it to compute the performance potential (or called relative value function) of MDPs under the average criterion, i.e., $\bm g=(\bm I - \bm P + \bm e \bm \pi)^{-1} \bm f$. The concept of the fundamental matrix was first proposed by J. G. Kemeny in his book ``\emph{Finite Markov Chains}" coauthored with L. J. Snell in 1960 \cite{Kemeny60}. In this book, the fundamental matrix is defined as $\bm Z^*:=(\bm I - \bm P + \bm e \bm \pi)^{-1}$. Many analysis, such as the mean passage time and the variance of passage time, can be conducted by using this fundamental matrix. The fundamental matrix also plays a key role in the performance sensitivity analysis of Markov systems. The original work about the sensitivity analysis of the steady state distribution and the fundamental matrix with respect to the stochastic matrix of Markov systems can be referred backward to P. Schweitzer's work in 1968 \cite{Schweitzer68}. P. Schweitzer presented a perturbation formalism that shows how the stationary distribution and the fundamental matrix of a Markov chain containing a single irreducible set of states change as the transition probabilities vary. The sensitivity information can be represented in a series of Rayleigh perturbation expansions of the fundamental matrix and other parameters. This is the main target of the perturbation analysis of Markov chains at the early stage. E. Seneta and C. D. Meyer did a lot of work \cite{Meyer94,Seneta93} to study the relation between the eigenvalues of stochastic matrix $\bm P$ and the condition number ($\max\{|a^\#_{i,j}|\}$) of group generalized inverse ($\bm A^\#=(\bm I - \bm P + \bm e \bm \pi)^{-1} - \bm e \bm \pi$) of matrix $\bm A = \bm I - \bm P$, where $a^\#_{i,j}$ is the element of matrix $\bm A^\#$. Some inequalities are derived to quantify the sensitivity of the steady state distribution when $\bm P$ is perturbed to $\bm P'$. Therefore, the sensitivity of the steady state distribution can be analyzed through studying the eigenvalues of stochastic matrix $\bm P$. This is the main idea of the perturbation analysis of Markov chains at that period. More than the sensitivity analysis of the steady state distribution, X. R. Cao proposed the sensitivity-based optimization theory that focuses on the sensitivity analysis of the system performance with respect to the perturbed transition probabilities or policies \cite{Cao97,Cao07}. This approach works well for different system settings, including the average or discounted criterion, the unichain or multichain, Markov or semi-Markov systems. Performance potential $\bm g$ is a fundamental quantity in MDPs. We have to compute or estimate its value before we conduct the policy iteration or sensitivity-based optimization. For an MDP under the discounted criterion, we can directly compute it as $\bm g= (\bm I - \gamma \bm P)^{-1}\bm f$ since the matrix $(\bm I - \gamma \bm P)$ is invertible \cite{Puterman94}, where $\gamma$ is the discount factor and $0<\gamma<1$. For an MDP under the average criterion, the fundamental matrix is used to compute the value of performance potentials and it has the form $\bm g=(\bm I - \bm P + \bm e \bm \pi)^{-1} \bm f$. Since the fundamental matrix can be decomposed as $(\bm I - \bm P + \bm e \bm \pi)^{-1} = \sum_{n=0}^{\infty}(\bm P^n - \bm e \bm \pi)$, we can rewrite the definition of the performance potential as $\bm g= \sum_{n=0}^{\infty}(\bm P^n \bm f - \eta \bm e )$, where $\eta = \bm \pi \bm f$ is the long-run average performance. That is, we can derive the following sample path version of the performance potential as $g(i) = E\{ \sum_{n=0}^{N}(f(X_n) - \hat{\eta})|X_0=i \}$, where $X_n$ is the system state at time $n$, $\hat{\eta}$ is the estimated value of the long-run average performance, $i \in \mathcal S$ is a state of the state space $\mathcal S$, and $N$ is a proper integer that is to control the estimation variance \cite{Cao07}. However, we see that when we compute $\bm g=(\bm I - \bm P + \bm e \bm \pi)^{-1} \bm f$, we have to pre-compute the value of $\bm \pi$ first. This issue was also pointed out by J. G. Kemeny when he studied the computation of the fundamental matrix $\bm Z^*=(\bm I - \bm P + \bm e \bm \pi)^{-1}$. He wrote ``It also suffers from the difficulty that one must compute $\alpha$ (solving $n$ equations) before one can compute $\bm Z^*$" \cite{Kemeny81}, where $\alpha$ is the steady state distribution $\bm \pi$ in our paper. In this paper, we study another form of the fundamental matrix as $\bm Z_r := (\bm I - \bm P + \bm e \bm r)^{-1}$, where $\bm r$ is any row vector satisfying $\bm r \bm e \neq 0$. Using this generalized fundamental matrix, we can compute the value of performance potentials as $\bm g = (\bm I - \bm P + \bm e \bm r)^{-1} \bm f$, where all the parameters are known and no pre-computation is required. The traditional approach $\bm g = (\bm I - \bm P + \bm e \bm \pi)^{-1} \bm f$ is a special case where we choose $\bm r = \bm \pi$. We can also choose $\bm r$ as other vectors. This formula can also be used to compute the value of $\bm \pi$ as $\bm \pi = \bm r (\bm I - \bm P + \bm e \bm r)^{-1}$, where the kernel computation remains the same as the generalized fundamental matrix $(\bm I - \bm P + \bm e \bm r)^{-1}$. There exist two works about the generalization of the fundamental matrix in the literature. After proposing the concept of the fundamental matrix, J. G. Kemeny further studied a generalized form \cite{Kemeny81}. A special case for ergodic Markov chains is that he defined $\bm Z_{\beta} := (\bm I - \bm P + \bm e \bm \beta)^{-1}$, $\bm \beta \bm e = 1$, where $\bm \beta$ is a row vector. In Kemeny's work, it is shown that the above matrix can be used to compute the stationary distribution as $\bm \pi = \bm \beta \bm Z_{\beta}$. Similar result was later reported in J. J. Hunter's book \cite{Hunter83}, which has the form $\bm \pi = \bm u (\bm I - \bm P + \bm e \bm u)^{-1}$, $\bm u \bm e \neq 0$, where $\bm u$ is a row vector. After that, J. J. Hunter further gave a thorough study on varied forms of general inverse of Markovian kernel $(I-P)$ in his recent works \cite{Hunter07,Hunter14}. One form of the general inverse is written as $(\bm I - \bm P + \bm t \bm u)^{-1}$ with condition $\bm \pi \bm t \neq 0$ and $\bm u \bm e \neq 0$, where $\bm t$ is a column vector. Obviously, $(\bm I - \bm P + \bm t \bm u)^{-1}$ is more general than the above results. Although these works have a similar result to ours, they focus on the computation of the steady state distribution or the mean first passage time. There is no study on the computation of performance potentials or relative value functions. Compared with the widely-adopted form $\bm g = (\bm I - \bm P + \bm e \bm \pi)^{-1} \bm f$, our new formula $\bm g = (\bm I - \bm P + \bm e \bm r)^{-1} \bm f$ does not require the pre-computation of $\bm \pi$. It may shed some light on how to efficiently compute or estimate the value of $\bm g$, which is an essential procedure for the policy iteration in MDPs. Moreover, we also study the generalization of the fundamental matrix not only for a discrete time Markov chain, but also for a continuous time Markov process. We further extend the similar idea to the representation and computation of Q-factors that are fundamental quantities for reinforcement learning and artificial intelligence \cite{Silver16, Sutton98}. \section{Main Results}\label{section_result} \subsection{Generalized Fundamental Matrix}\label{subsection_GFM} We focus on the discussion of a Markov chain with finite states. The state space is denoted as $\mathcal S:=\{1,2,\cdots,S\}$. The transition probability matrix is denoted as $\bm P$ and its element is $P(i,j)$ indicating the probability of which the system transits from the current state $i$ to the next state $j$, where $i,j\in \mathcal S$. The steady state distribution of this Markov chain is denoted as an $S$-dimensional row vector $\bm \pi$ and its element $\pi(i)$ is the probability of the system staying at state $i$, $i \in \mathcal S$. Obviously, we have \begin{equation} \bm P \bm e = \bm e, \qquad \bm \pi \bm P = \bm \pi, \qquad \bm \pi \bm e = 1, \end{equation} where $\bm e$ is an $S$-dimensional column vector with all elements 1. First, we give the following lemma about shifting the eigenvalues of a general matrix. \begin{lemma}\label{lemma1} Suppose that $\bm A$ is a square matrix for which $\lambda$ is an eigenvalue having multiplicity 1, and the associated column eigenvector is $\bm v$. Let $\lambda_i$ be the other eigenvalues of $\bm A$ and $\bm \phi_i$ be the associated row eigenvectors. Then, for any row vector $\bm r$, $\lambda + \bm r\bm v$ is an eigenvalue of $\bm A + \bm v\bm r$ and the associated column eigenvector is $\bm v$; other eigenvalues of $\bm A + \bm v\bm r$ are given by $\lambda_i$ with the associated row eigenvectors $\bm \phi_i$. \end{lemma} \begin{proof} Since $\bm \phi_i$ is a \emph{row eigenvector} of $\bm A$ associated with eigenvalue $\lambda_i$, we have \begin{equation} \bm \phi_i \bm A \bm v = \lambda_i \bm \phi_i \bm v. \end{equation} On the other hand, since $\bm v$ is a \emph{column eigenvector} of $\bm A$, we have \begin{equation} \bm \phi_i \bm A \bm v = \lambda \bm \phi_i \bm v. \end{equation} Since $\lambda$ has multiplicity 1, $\lambda_i$ is not equal to $\lambda$. Comparing the above two equations, we directly have \begin{equation}\label{eq_phiv} \bm \phi_i \bm v = 0. \end{equation} Below, we further study the eigenvalue of matrix $\bm A + \bm v\bm r$. Using (\ref{eq_phiv}), we have \begin{equation} \bm \phi_i (\bm A + \bm v \bm r) = \bm \phi_i \bm A = \lambda_i \bm \phi_i. \end{equation} Therefore, $\lambda_ i$ continues to be an eigenvalue of $\bm A + \bm v \bm r$ and $\bm \phi_i$ continues to be the associated row eigenvector of $\bm A + \bm v \bm r$. Moreover, we have \begin{equation} (\bm A + \bm v\bm r) \bm v = \bm A \bm v + \bm v (\bm r\bm v) = (\lambda + \bm r\bm v) \bm v, \end{equation} Therefore, $\lambda + \bm{rv}$ is a new eigenvalue of $\bm A + \bm{vr}$ and $\bm v$ continues to be the associated column eigenvector of $\bm A + \bm{vr}$. The lemma is proved. \end{proof} For the eigenvalues of the transition probability matrix of a Markov chain, we have the following lemma. \begin{lemma}\label{lemma2} If $\bm P$ is the stochastic matrix of an irreducible and aperiodic Markov chain, its spectral radius $\rho(\bm P) = \lambda_{1} = 1$ and $|\lambda_{i}| < 1$ for $i \neq 1$, where $\lambda_{i}$ is the eigenvalues of $\bm P$ descendingly sorted by their modulus. Moreover, $\lambda_{1} = 1$ is a simple eigenvalue and the associated column eigenvector is $\bm x_1 = \bm e$. \end{lemma} This lemma can be obtained directly from the \emph{Perron-Frobeniu theorem} that was separately proposed by Oskar Perron in 1907 \cite{Perron07} and Georg Frobenius in 1908 \cite{Frobenius12}. The original Perron-Frobeniu theorem aims to study the eigenvalues and eigenvectors of nonnegative matrix. Since the stochastic matrix is a special case of nonnegative matrix, we can obtain more specific properties of stochastic matrix, such as the statement in the above lemma. The proof of this lemma is ignored and interested readers can find it from reference books \cite{Berman94,Cao07}. With the above lemmas, we derive the following theorem about the eigenvalues of matrix $\bm I - \bm P + \bm e \bm r$. \begin{theorem}\label{theorem1} Assume that $\bm P$ is the stochastic matrix of an irreducible and aperiodic Markov chain. Denote $\lambda_i$ as the eigenvalues of $\bm P$ in descending order of their modulus, $\bm \phi_i$ and $\bm x_i$ are the associated row and column eigenvectors, respectively. Then, for any row vector $\bm r$, the matrix $\bm I - \bm P + \bm e \bm r$ has the following property: one eigenvalue is $\bm r \bm e$ and the associated column eigenvector is $\bm e$; other eigenvalues are $1-\lambda_i$ for $i \neq 1$, the associated row eigenvectors are $\bm \phi_i$, and the associated column eigenvectors are $\bm x_i + \frac{\bm r \bm x_i}{\lambda_i - \bm r \bm e}\bm e$ if $\bm r \bm e \neq \lambda_i$. \end{theorem} \begin{proof} With Lemma~\ref{lemma2}, we see that $\lambda_1=1$ and $|\lambda_{i}| < 1$ for $i \neq 1$. We denote $\bm A := \bm I - \bm P$. It is easy to verify that the eigenvalues of $\bm A$ are $1-\lambda_i$ and the associated row eigenvectors are the same as $\bm \phi_i$. That is, $0$ is an eigenvalue of $\bm A$ with simplicity 1 and the associated column eigenvector is $\bm e$. Therefore, by applying Lemma~\ref{lemma1}, we see that $\bm r \bm e$ is an eigenvalue of matrix $\bm I - \bm P + \bm e \bm r$ and the associated column vector is $\bm e$; other eigenvalues are $1-\lambda_i$ with the associated row eigenvector $\bm \phi_i$ for $i \neq 1$. The column eigenvectors of $\bm I - \bm P + \bm e \bm r$ can be verified as follows. \begin{eqnarray} &&(\bm I-\bm P+\bm e \bm r) \left(\bm x_i + \frac{\bm r \bm x_i}{\lambda_i-\bm r\bm e} \bm e \right) \nonumber\\ &=& (\bm I - \bm P)\bm x_i + (\bm I - \bm P)\bm e \frac{\bm r \bm x_i}{\lambda_i-\bm r\bm e} + \bm e \bm r \bm x_i + \bm e \bm r \bm e \frac{\bm r \bm x_i}{\lambda_i-\bm r \bm e} \nonumber\\ &=& \lambda_i \bm x_i + \bm 0 + \bm r \bm x_i \bm e + \frac{\bm r\bm e}{\lambda_i-\bm r\bm e} \bm r \bm x_i \bm e \nonumber\\ &=& \lambda_i \left( \bm x_i + \frac{\bm r \bm x_i}{\lambda_i-\bm r \bm e} \bm e \right). \end{eqnarray} Therefore, the theorem is proved. \end{proof} The eigenvalue of a matrix may be a complex number. With Lemma~\ref{lemma2}, we can see that the eigenvalues of $\bm P$ have $|\lambda_i|<1$ for $i \neq 1$ and they are located in the unit circle in the complex plane, as illustrated by the left sub-figure of Fig~\ref{fig_circleP}. With Theorem~\ref{theorem1}, we can see that the eigenvalues of $\bm I - \bm P + \bm e \bm r$ are $1-\lambda_i$ for $i \neq 1$ and they are also located in the unit circle illustrated by the right sub-figure of Fig.~\ref{fig_circleP}. \begin{figure}[htbp] \centering \includegraphics[width=0.9\columnwidth]{circleP.eps} \caption{The distribution area of eigenvalues of matrix $\bm P$ and $\bm I - \bm P + \bm e \bm r$, i.e., $\lambda_i$ and $1-\lambda_i$ for $i \neq 1$.}\label{fig_circleP} \end{figure} Therefore, we can directly derive the following theorem. \begin{theorem}\label{theorem2} The matrix $\bm I-\bm P+\bm e \bm r$ is not singular if and only if $\bm r \bm e \neq 0$, where $\bm P$ is the stochastic matrix of an irreducible and aperiodic Markov chain. \end{theorem} \begin{proof} The proof of this theorem is very straightforward. Based on Theorem~\ref{theorem1} and Lemma~\ref{lemma2}, we see that the eigenvalues of $\bm I - \bm P + \bm e \bm r$ are either $\bm r \bm e$ or $1-\lambda_i$ for $i \neq 1$. Since $\bm r \bm e \neq 0$ and $|\lambda_i|<1$ for $i \neq 1$, we can easily verify that $0$ is not the eigenvalue of $\bm I - \bm P + \bm e \bm r$. Therefore, the matrix is invertible and the theorem is proved. \end{proof} With Theorem~\ref{theorem2}, we see that $\bm I - \bm P + \bm e \bm r$ is invertible if $\bm r \bm e \neq 0$. Therefore, we define a \emph{generalized fundamental matrix} as below. \begin{equation}\label{eq_Zr} \bm{Z_r} := (\bm I - \bm P + \bm e \bm r)^{-1}, \qquad \bm r \bm e \neq 0. \end{equation} Compared with the fundamental matrix $\bm Z^*:=(\bm I - \bm P + \bm e \bm \pi)^{-1}$ defined in the literature \cite{Kemeny60}, $\bm Z^*$ can be viewed as a special case of $\bm{Z_r}$ with $\bm r = \bm \pi$. The generalized fundamental matrix $\bm{Z_r}$ is an important quantity of Markov chains and it can be utilized to compute the performance potential and the steady state distribution, as we will discuss in the following subsections. \subsection{Computation of Performance Potential}\label{subsection_PP} As we know, the value function (or performance potential) is an very important quantity of Markov decision processes. In a standard policy iteration procedure, we have to compute the value function for the current policy, which is called the \emph{policy evaluation} step \cite{Puterman94}. In the approximate dynamic programming, we study various approximation approaches to simplify the computation of the value function to alleviate the curse of dimensionality \cite{Bertsekas12}. Therefore, the efficient computation of the value function is an very important topic in the field of MDPs. Note that the computation of value functions under the discount criterion is easy, because the associated Poisson equation has a unique solution. We focus on the value function under the long-run average criterion of MDPs. The \emph{performance potential} is an alias of the value function and it has a special physical meaning from the perspective of the perturbation analysis and the sensitivity-based optimization \cite{Cao07}. In the following content, we will use the term of performance potential to study how to compute or estimate it. We denote the performance potential as an $S$-dimensional column vector $\bm g$ and its element $g(i)$, $i \in \mathcal S$, is defined as below. \begin{equation}\label{eq_g} g(i) := \lim\limits_{T \rightarrow \infty} E\left\{ \sum_{t=0}^{T-1}[f(X_t) - \eta] | X_0 = i \right\}, \end{equation} where $X_t$ is the system state at time $t$, $f(X_t)$ is the system reward at state $X_t$, and $\eta$ is the long-run average performance defined as below. \begin{equation} \eta := \lim\limits_{T \rightarrow \infty} E\left\{ \frac{1}{T} \sum_{t=0}^{T-1} f(X_t) | X_0 = i \right\} = \lim\limits_{T \rightarrow \infty} \frac{1}{T} \sum_{t=0}^{T-1} f(X_t), \end{equation} where the second equality holds when the Markov chain is a unichain. Extending the right-hand side of (\ref{eq_g}) at time $t=0$ and recursively substituting (\ref{eq_g}), we can obtain \begin{equation}\label{eq_g2} g(i) = f(i) - \eta + \sum_{j \in \mathcal S} p(i,j)g(j). \end{equation} Rewriting the above equation in a matrix form, we obtain the \emph{Poisson equation} as below. \begin{equation}\label{eq_poisson} \bm g = \bm f - \eta \bm e + \bm P \bm g. \end{equation} The above equation can be rewritten as below. \begin{equation} (\bm I - \bm P)\bm g = \bm f - \eta \bm e. \end{equation} However, as we know from Lemma~\ref{lemma2}, the matrix $(\bm I - \bm P)$ has an eigenvalue with value 0 and it is not invertible. Noticing the fact that $\bm g + c \bm e$ is still a solution to (\ref{eq_poisson}) for any constant $c$, we can properly choose $c$ to let $\bm \pi \bm g = \eta$. Therefore, we have \begin{equation} (\bm I - \bm P + \bm e \bm \pi)\bm g = \bm f. \end{equation} With Theorem~\ref{theorem2}, we see that matrix $(\bm I - \bm P + \bm e \bm \pi)$ is invertible. The inverse matrix is called the fundamental matrix defined in the literature \cite{Kemeny60} and it has \begin{equation} (\bm I - \bm P + \bm e \bm \pi)^{-1} = \sum_{n=0}^{\infty} (\bm P - \bm e \bm \pi)^n = \bm I + \sum_{n=1}^{\infty} (\bm P^n - \bm e \bm \pi). \end{equation} Therefore, the solution to the Poisson equation (\ref{eq_poisson}) can be written as below. \begin{equation}\label{eq_g3} \bm g = (\bm I - \bm P + \bm e \bm \pi)^{-1} \bm f. \end{equation} The above formula widely exists in the literature \cite{Cao97,Cao07} and we can use it to numerically compute the value of $\bm g$ for a specific MDP. Note that the value of $\bm g$ computed with (\ref{eq_g3}) satisfies the condition $\bm \pi \bm g = \eta$. However, $\bm \pi$ is not a given parameter in the above equation. We have to compute the value of $\bm \pi$ before we can use (\ref{eq_g3}) to compute $\bm g$. This increases the computation burden. Moreover, if we conduct online estimation, the estimation error of $\bm \pi$ may increase the estimation variance of $\bm g$. Fortunately, we have another way to numerically compute $\bm g$ without the extra computation for $\bm \pi$. Since $\bm g + c \bm e$ is still a solution to (\ref{eq_poisson}) for any constant $c$, for any $S$-dimensional row vector $\bm r$ satisfying $\bm r \bm e \neq 0$, we can choose a proper $c$ to let $\bm r \bm g = \eta$. Therefore, we can rewrite (\ref{eq_poisson}) as below. \begin{equation} (\bm I - \bm P + \bm e \bm r)\bm g = \bm f. \end{equation} Since matrix $(\bm I - \bm P + \bm e \bm r)$ is always invertible as proved in Theorem~\ref{theorem2}, the above equation can be further rewritten as \begin{center} \begin{boxedminipage}{1\columnwidth} \begin{equation}\label{eq_g4} \bm g = (\bm I - \bm P + \bm e \bm r)^{-1} \bm f, \end{equation} \vspace{-10pt} \end{boxedminipage} \end{center} where $\bm r$ is any $S$-dimensional row vector $\bm r$ satisfying $\bm r \bm e \neq 0$. We can see that all the parameters in (\ref{eq_g4}) are given and we can directly compute $\bm g$ with (\ref{eq_g4}) without any pre-computation. \noindent \textbf{Remark 1.} Both (\ref{eq_g3}) and (\ref{eq_g4}) are solutions to the Poisson equation (\ref{eq_poisson}) and they have difference only with a constant column vector $c \bm e$. The value of $\bm g$ in (\ref{eq_g3}) satisfies $\bm \pi \bm g = \eta$, while the value of $\bm g$ in (\ref{eq_g4}) satisfies $\bm r \bm g = \eta$. \noindent \textbf{Remark 2.} (\ref{eq_g3}) can be viewed as a special case of (\ref{eq_g4}) if we choose $\bm r = \bm \pi$. With (\ref{eq_g4}), we have more flexibility to choose different $\bm r$'s, which may give some insights on the computation or estimation of $\bm g$. \subsection{Computation of Steady State Distribution}\label{subsection_SSD} The fundamental matrix can also be used to compute the steady state distribution of Markov chains. We also assume that the Markov chain is irreducible and aperiodic. We know that $\bm \pi$ can be determined by the following set of linear equations \begin{equation} \begin{array}{l} \bm \pi \bm P = \bm \pi. \\ \bm \pi \bm e = 1. \end{array} \end{equation} We can rewrite the above equations according to the standard form of linear equations as below. \begin{equation} \begin{array}{l} (\bm I - \bm P^{T}) \bm \pi^{T} = \bm 0. \\ \bm e^T \bm \pi^T = 1. \end{array} \end{equation} That is, \begin{equation}\label{eq4} \left[ \left( \begin{array}{ccccc} 1 & 0 & 0 & \cdots & 0 \\ 0 & 1 & 0 & \cdots & 0 \\ 0 & 0 & 1 & \cdots & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & 0 & \cdots & 1 \\ \end{array} \right) - \left( \begin{array}{ccccc} P_{1,1} & P_{2,1} & P_{3,1} & \cdots & P_{S,1} \\ P_{1,2} & P_{2,2} & P_{3,2} & \cdots & P_{S,2} \\ P_{1,3} & P_{2,3} & P_{3,3} & \cdots & P_{S,3} \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ P_{1,S} & P_{2,S} & P_{3,S} & \cdots & P_{S,S} \\ \end{array} \right) \right] \left( \begin{array}{c} \pi_1 \\ \pi_2 \\ \pi_3 \\ \vdots \\ \pi_S \\ \end{array} \right) = \left( \begin{array}{c} 0 \\ 0 \\ 0 \\ \vdots \\ 0 \\ \end{array} \right). \end{equation} \begin{equation}\label{eq5} \pi_1 + \pi_2 + \pi_3 + \dots + \pi_S = 1. \end{equation} For any $S$-dimensional row vector $\bm r$ satisfying $\bm r \bm e \neq 0$, we multiply $r(i)$ on both sides of (\ref{eq5}) and summate this equation to the $i$th equation of (\ref{eq4}), $i=1,2,\cdots,S$. We can obtain \begin{equation} (\bm I - \bm P^{T} + \bm r^T \bm e^T) \bm \pi^{T} = \bm r^T. \end{equation} With Theorem~\ref{theorem2}, we know that $\bm I - \bm P^{T} + \bm r^T \bm e^T$ is invertible and $(\bm I - \bm P^{T} + \bm r^T \bm e^T)^{-1} = \bm Z^{T}_{\bm r}$. Therefore, we have \begin{equation} \bm \pi^{T} = (\bm I - \bm P^{T} + \bm r^T \bm e^T)^{-1} \bm r^T, \end{equation} or \begin{center} \begin{boxedminipage}{1\columnwidth} \begin{equation}\label{eq_pi} \bm \pi = \bm r (\bm I - \bm P + \bm e \bm r)^{-1}, \qquad \bm r \bm e \neq 0. \end{equation} \vspace{-10pt} \end{boxedminipage} \end{center} From the above equation, we can see that the computation of $\bm \pi$ has the same key part as the computation of $\bm g$ with (\ref{eq_g4}), i.e., the computation of the generalized fundamental matrix $\bm{Z_r}=(\bm I - \bm P + \bm e \bm r)^{-1}$. Therefore, $\bm{Z_r}$ plays a key role in the analysis of Markov chains. \subsection{Property Analysis and Estimation Algorithm}\label{subsection_Alg} In this subsection, we discuss the properties of the generalized fundamental matrix $\bm{Z_r}$ and the effect on the computation of performance potentials. If the spectral radius of $\bm P - \bm e \bm r$ is smaller than 1, we can rewrite the generalized fundamental matrix as follows. \begin{equation}\label{eq_Zr2} \bm{Z_r} = \sum_{n=0}^{\infty}(\bm P - \bm e \bm r)^n, \qquad \rho(\bm P - \bm e \bm r)<1. \end{equation} According to Lemma~\ref{lemma1}, we see that the eigenvalues of $\bm P - \bm e \bm r$ are $\lambda_1 - \bm r \bm e$ and $\lambda_i$ for $i \neq 1$, where $\lambda_i$ are the eigenvalues of $\bm P$ sorted in the descending order of their modulus. With Lemma~\ref{lemma2}, we see that $\lambda_1=1$ and $|\lambda_i|<1$ for $i \neq 1$. Therefore, we have \begin{equation} \rho(\bm P - \bm e \bm r)<1 \quad \Longleftrightarrow \quad 0<\bm r \bm e <2. \end{equation} Furthermore, we can verify that (\ref{eq_Zr2}) can be rewritten as below. \begin{equation} \bm{Z_r} = \sum_{n=0}^{\infty} \left[ \bm P^n - \bm e \bm r \sum_{j=0}^{n-1} (1 - \bm r \bm e)^j \bm P^{n-1-j} \right], \qquad 0<\bm r \bm e <2. \end{equation} If we choose $\bm r$ such that $\bm r \bm e = 1$, then we can further simplify the above equation as below. \begin{equation}\label{eq_Zr3} \bm{Z_r} = \sum_{n=0}^{\infty} \left[ \bm P^n - \bm e \bm r \bm P^{n-1} \right], \qquad \bm r \bm e = 1, \end{equation} where we define $\bm e \bm r \bm P^{n-1} = 0$ when $n=0$. If we choose $\bm r$ stochastic, then the $(i,j)$'th entry of $(\bm I - \bm P + \bm e \bm r)^{-1}$ has the interpretation that it is a sum over $n$ in which the $n$'th term (for $n \geq 1$) is $P(X_n = j | X_0 = i) - P( X_{n-1} = j | X_0$ having initial distribution $\bm r$). If we manipulate the terms in the summation of (\ref{eq_Zr3}), we can obtain \begin{equation}\label{eq_Zr4} \bm{Z_r} = \sum_{n=0}^{\infty} \left[ \bm P^n - \bm e \bm r \bm P^{n} \right] + \lim\limits_{n \rightarrow \infty}\bm e \bm r \bm P^n, \qquad \bm r \bm e = 1. \end{equation} Since the Markov chain is irreducible, aperiodic, and finite, the limiting probability exists and it equals the steady state distribution. That is, \begin{equation} \lim\limits_{n \rightarrow \infty} \bm P^n = \bm e \bm \pi. \end{equation} Therefore, the above equation (\ref{eq_Zr4}) can be rewritten as below. \begin{equation}\label{eq_Zr5} \bm{Z_r} = \sum_{n=0}^{\infty} \left[ \bm P^n - \bm e \bm r \bm P^{n} \right] + \bm e \bm \pi, \qquad \bm r \bm e = 1. \end{equation} Therefore, neglecting the steady distribution $\bm e \bm \pi$ for simplicity, we can see that the $n$'th term of the above summation is $P(X_n = j | X_0 = i) - P( X_n = j | X_0$ having initial distribution $\bm r$). This interpretation can help develop online estimation algorithms for quantities related to $\bm{Z_r}$ from a viewpoint of sample paths. Substituting (\ref{eq_Zr5}) into (\ref{eq_g4}), we have \begin{eqnarray} \bm g &=& (\bm I - \bm P + \bm e \bm r)^{-1} \bm f \nonumber\\ &=& \sum_{n=0}^{\infty} \left[ \bm P^n - \bm e \bm r \bm P^{n} \right] \bm f + \bm e \bm \pi \bm f \nonumber\\ &=& \sum_{n=0}^{\infty} \bm P^n \bm f - \bm e \bm r \sum_{n=0}^{\infty} \bm P^{n} \bm f + \eta \bm e, \end{eqnarray} where $\bm r \bm e = 1$ and $\bm r \bm g = \eta$. Since $\bm g + c \bm e$ is still a performance potential for any constant $c$, we can neglect the term $\eta \bm e$ in the above equation and rewrite it as below. \begin{equation}\label{eq_42} \bm g = \sum_{n=0}^{\infty} \bm P^n \bm f - \bm e \bm r \sum_{n=0}^{\infty} \bm P^{n} \bm f, \end{equation} where $\bm r \bm e = 1$ and $\bm r \bm g = 0$. From the above equation, we can see that $\sum_{n=0}^{\infty} \bm P^n \bm f$ equals the expectation of the accumulated rewards along the sample path, i.e., $\mathbb E\{\sum_{t=0}^{\infty} f(X_t)\}$. We denote \begin{equation} \tilde{\bm g} = \sum_{n=0}^{\infty} \bm P^n \bm f = \mathbb E\left\{ \sum_{t=0}^{\infty} f(X_t) \right\}. \end{equation} or \begin{equation} \tilde{\bm g}_T = \mathbb E\left\{ \sum_{t=0}^{T} f(X_t) \right\}, \end{equation} for a large constant $T$. We can rewrite (\ref{eq_42}) as below. \begin{equation} \bm g = \tilde{\bm g} - \bm e \bm r \tilde{\bm g}. \end{equation} When $\bm r$ is a probability distribution, the physical meaning of the above equation is that: we sum all the rewards along the sample path, then we use a weighting vector $\bm r$ to obtain a \emph{reference level} $\bm r \tilde{\bm g}$, the gap between $\tilde{\bm g}$ and the reference level $\bm r \tilde{\bm g}$ is exactly the performance potential (this is also the reason that we call it \emph{relative} value function). The insight for the estimation algorithm is that we can just sum all the rewards along the sample path, then we choose an arbitrary reference level determined by a combination of elements of $\tilde{\bm g}$, the gap between them is the estimate of performance potentials. This can help to simplify the online estimation procedure. $\bm r = \bm \pi$ is one of the special cases. For example, we can also set $\bm r = (1, 0, 0, \cdots, 0)$, which lets $\tilde{g}(1)$ as the reference level. Instead of numerically computing $\bm g$ with (\ref{eq_g4}), we can further develop an iterative algorithm to estimate the value of $\bm g$ based on (\ref{eq_g4}). With (\ref{eq_g4}), we have \begin{equation} \bm g = \bm f - \bm e \bm r \bm g + \bm P \bm g. \end{equation} Suppose $\hat{\bm g}$ is an unbiased estimate of $\bm g$, i.e., $E(\hat{\bm g}) = \bm g$. We have \begin{equation} E(\hat{\bm g}) = \bm f - \bm e \bm r E(\hat{\bm g}) + \bm P E(\hat{\bm g}). \end{equation} Rewriting the above equation as a \emph{sample path version}, we derive the following equation that holds from the sense of statistics \begin{eqnarray} E[\hat{g}(X_t)] &=& E[f(X_t) - \bm r \hat{\bm g} + \hat{g}(X_{t+1})] \nonumber\\ &=& E[f(X_t)] - \sum_{i \in \mathcal S}r(i)E[\hat{g}(i)] + E[\hat{g}(X_{t+1})]. \end{eqnarray} Based on the above equation, we develop a \emph{least-squares} algorithm that can online estimate $\bm g$ based on sample paths. Define a quantity $z_{t}$ as below, which can be viewed as the new information learned from the current feedback of the system. \begin{equation} z_{t} := f(X_t) - \sum_{i \in \mathcal S}r(i) \hat{g}(i) + \hat{g}(X_{t+1}) - \hat{g}(X_{t}). \end{equation} With a stochastic approximation framework, we have the following formula to update the value of $\hat{g}$ \begin{equation}\label{eq_lsg} \hat{g}(X_t) \leftarrow \hat{g}(X_t) + \alpha_t z_{t}, \end{equation} where $\alpha_t$ is a positive step-size that satisfies the convergence condition of Robbins-Monro stochastic approximation, i.e., \begin{equation} \sum_{t=0}^{\infty}\alpha_t = \infty, \quad \sum_{t=0}^{\infty} \alpha^2_t < \infty. \end{equation} or $\alpha_t$ satisfies an even looser condition as below \cite{Chen14}. \begin{equation} \sum_{t=0}^{\infty}\alpha_t = \infty, \quad \lim\limits_{t \rightarrow \infty} \alpha_t \rightarrow 0. \end{equation} The name of least-squares of update formula (\ref{eq_lsg}) comes from the fact that we aim to obtain the following estimate of $\bm g$ that can obtain the least squares of $z_t$ in statistics, i.e., \begin{equation} \bm g = \argmin\limits_{\hat{\bm g}} E\{z^2_t\} \end{equation} The idea of least-squares update formula (\ref{eq_lsg}) is similar to that of temporal-difference (TD) algorithm in reinforcement learning \cite{Sutton98}. Below, we give a procedure framework of the online estimation algorithm as illustrated in Fig.~\ref{fig_algo}. In Fig.~\ref{fig_algo}, the algorithm stopping criterion can be set as the norm of two successive estimates is smaller than a given small threshold $\epsilon$, i.e., $\parallel \bm g - \bm g' \parallel < \epsilon$, or just simply stop the algorithm after reaching a large step number $T$, i.e., $t > T$. \begin{figure} \begin{center} \begin{boxedminipage}{1\columnwidth} \vspace{5pt} \begin{itemize} \item[] Initialize $\hat{g}(s)$ arbitrarily, e.g., set $\hat{g}(s)=0$ for all $s \in \mathcal S$. \item[] Repeat (for each time step $t$): \subitem observe the current state $s$, the current reward $f_t$, and the next state $s'$ \subitem calculate $z_t = f_t - \sum_{s \in \mathcal S}r(s)\hat{g}(s) + \hat{g}(s') - \hat{g}(s)$ \subitem update $\hat{g}(s) \leftarrow \hat{g}(s)+\alpha_t z_t$ \subitem $s \leftarrow s'$; $t \leftarrow t+1$ \item[] Until a stopping criterion is satisfied. \end{itemize} \vspace{5pt} \end{boxedminipage} \end{center} \caption{An online least-squares algorithm to estimate $\bm g$ based on (\ref{eq_g4}).}\label{fig_algo} \end{figure} \section{Extension to Other Cases}\label{section_Ext} In the previous section, we discuss the generalized fundamental matrix for the discrete time Markov chain. In this section, we extend the result to other cases, including the continuous time Markov process and the Q-factors in reinforcement learning. \subsection{Continuous Time Markov Process}\label{subsection_CTMP} Consider a continuous time Markov process with transition rate matrix $\bm B$. Assume the Markov process is ergodic, we can similarly prove \begin{theorem}\label{theorem5} The eigenvalue of matrix $\bm B$ has the following property: $\lambda^B_1 = 0$ is a simple eigenvalue and the associated column eigenvector is $\bm x^B_1 = \bm e$; $ |\lambda^B_i + \gamma| < \gamma$ for $i \neq 1$, where $\gamma$ is any constant satisfying $\gamma \geq \max_{s}|B(s,s)|$. \end{theorem} \begin{proof} By using the uniformization technique in MDPs, we can define a matrix $\bm P$ as below. \begin{equation} \bm P := \bm I + \frac{\bm B}{\gamma}. \end{equation} Obviously, $\bm P$ is a stochastic matrix since it satisfies $\bm P \bm e = \bm e$ and all of its elements are nonnegative. The statistical behavior of the Markov chain with transition probability matrix $\bm P$ is equivalent to that of the Markov process with transition rate matrix $\bm B$ \cite{Puterman94}. We have \begin{equation}\label{eq28} \bm B = \gamma(\bm P - \bm I). \end{equation} Since the Markov process with $\bm B$ is ergodic, the equivalent Markov chain with $\bm P$ is also ergodic. Therefore, $\bm P$ is an ergodic stochastic matrix. With Lemma~\ref{lemma2}, we know that the eigenvalue of $\bm P$ has \begin{equation} \left\{ \begin{array}{ll} \lambda^P_i = 1, &i=1;\\ |\lambda^P_i| < 1, &i \neq 1;\\ \end{array} \right. \end{equation} Therefore, with (\ref{eq28}), we see that the eigenvalue of $\bm B$ is $\lambda^B_i = \gamma(\lambda^P_i - 1)$ and the column eigenvector is $\bm x^B_i = \bm x^P_i$, $\forall i$. We have \begin{equation} \left\{ \begin{array}{lll} \lambda^B_i = 0, &\bm x^B_i=\bm e, &i=1;\\ |\lambda^B_i+\gamma| < \gamma, &\bm x^B_i=\bm x^P_i, &i \neq 1;\\ \end{array} \right. \end{equation} The theorem is proved. \end{proof} With the above theorem, we know that the eigenvalue of $\bm B$ is either $0$ or distributed inside of the dotted circle in Fig.~\ref{fig_circleB}. When $\gamma$ is smaller, we can obtain a tighter area describing the distribution area of $\lambda^B_i$. Obviously, the smallest value of $\gamma$ is \[ \gamma = \max_{s \in \mathcal S}|B(s,s)|. \] \begin{figure}[htbp] \centering \includegraphics[width=0.5\columnwidth]{circleQ.eps} \caption{The distribution area of eigenvalues of matrix $B$, i.e., $\lambda^B_i$ for $i \neq 1$.}\label{fig_circleB} \end{figure} Furthermore, we study the eigenvalue of matrix $\bm D := \bm B + \bm e \bm r$ and we can similarly obtain the following theorem. \begin{theorem}\label{theorem6} The eigenvalue of matrix $\bm D := \bm B + \bm e \bm r$ has the following property: $\lambda^D_1 = \bm r \bm e$ is a simple eigenvalue and the associated column eigenvector is $\bm x^D_1 = \bm e$; $\lambda^D_i = \lambda^B_i$ and the associated column eigenvector is $\bm x^D_i = \bm x^B_i + \frac{\bm r \bm x^B_i}{\lambda^B_i-\bm r \bm e}\bm e$ for $i \neq 1$ and $\lambda^B_i \neq \bm r \bm e$. \end{theorem} \begin{proof} The proof of this theorem is similar to that of Theorem~\ref{theorem1}. We only need to verify whether the eigenvalue and eigenvector satisfy the equation $\bm D \bm x^D_i = \lambda^D_i \bm x^D_i$. Such verification is easy and we ignore the details for simplicity. \end{proof} \begin{theorem}\label{theorem7} For any row vector $\bm r$ satisfying $\bm r \bm e \neq 0$, the matrix $\bm D := \bm B + \bm e \bm r$ is invertible, where $\bm B$ is the transition rate matrix of an ergodic Markov process. \end{theorem} \begin{proof} Based on Theorem~\ref{theorem5} and Theorem~\ref{theorem6}, we can easily verify that the eigenvalues of $\bm D$ satisfy: \begin{equation} \left\{ \begin{array}{ll} \lambda^D_i = \bm r \bm e, &i=1;\\ |\lambda^D_i+\gamma| < \gamma, &i \neq 1;\\ \end{array} \right. \end{equation} Therefore, $0$ is not the eigenvalue of the matrix $\bm D$ and $\bm D$ is invertible. \end{proof} With Theorem~\ref{theorem7}, we can also simplify the computation of the steady state distribution and the performance potential (value function) of Markov processes. First, we discuss how to compute the steady state distribution $\bm \pi$ in Markov processes. It is known that $\bm \pi$ can be determined by the following equations \begin{equation} \begin{array}{ll} \bm \pi \bm B = \bm 0 \\ \bm \pi \bm e = 1\\ \end{array} \end{equation} Multiplying $r(i)$ on both sides of the second equation and adding it to the $i$th row of the first equation, we can obtain the following equation after proper manipulations \begin{equation} \bm \pi(\bm B + \bm e \bm r) = \bm r. \end{equation} With Theorem~\ref{theorem7}, for any $S$-dimensional row vector $\bm r$ satisfying $\bm r \bm e \neq 0$, $(\bm B + \bm e \bm r)$ is invertible and we have \begin{center} \begin{boxedminipage}{1\columnwidth} \begin{equation}\label{eq_piQ} \bm \pi = \bm r (\bm B + \bm e \bm r)^{-1}, \qquad \bm r \bm e \neq 0. \end{equation} \vspace{-10pt} \end{boxedminipage} \end{center} Second, we discuss how to compute the performance potential of Markov processes. In a continuous time Markov process, the performance potential $\bm g$ is defined as below. \begin{equation} g(i) = \lim\limits_{T \rightarrow \infty}E\left\{ \int^{T}_{0} [f(X_t) - \eta]dt \Big| X_0=i \right\}. \end{equation} We also have the Poisson equation for the above definition \begin{equation} -\bm B \bm g = \bm f - \eta \bm e. \end{equation} In the literature, it is widely adopted that $\bm g$ can be numerically computed by the following equation \cite{Cao97,Cao07} \begin{equation}\label{eq37} \bm g = -(\bm B - \bm e \bm \pi)^{-1}\bm f. \end{equation} The above equation includes the condition $\bm \pi \bm g = \eta$. Similar to the case of discrete time Markov chain, we can also derive \begin{equation} (\bm B + \bm e \bm r)\bm g = -\bm f, \end{equation} where the condition $\bm r \bm g = \eta$ is required. With Theorem~\ref{theorem7}, for any $S$-dimensional row vector $\bm r$ satisfying $\bm r \bm e \neq 0$, $(\bm B + \bm e \bm r)$ is invertible and we have \begin{center} \begin{boxedminipage}{1\columnwidth} \begin{equation}\label{eq_gQ} \bm g = -(\bm B + \bm e \bm r)^{-1}\bm f, \qquad \bm r \bm e \neq 0. \end{equation} \vspace{-10pt} \end{boxedminipage} \end{center} Therefore, we can directly compute the value of $\bm g$ using the above equation, without the pre-computation of $\bm \pi$ required by (\ref{eq37}). On the other hand, we can also develop an online least-squares algorithm to estimate $\bm g$ based on sample paths, which is similar to the algorithm described in Fig.~\ref{fig_algo} for the case of discrete time Markov chains. For simplicity, we ignore the details. \subsection{Poisson Equation for Q-factors}\label{subsection_Q} It is well known that we have Poisson equation for the performance potential or the value function in MDPs. Similar to the performance potential quantifying the effect of the initial state on the average performance, the Q-factor is an important quantity in reinforcement learning and it quantifies the effect of the state-action pair on the system average performance. Below, we give a Poisson equation for Q-factors in an MDP under the time average criterion. Suppose the current policy is $\mathcal L$. In this subsection, all the quantities of MDPs are assumed for the policy $\mathcal L$ by default, unless we have specific other notations. The Q-factor of the Markov system under this policy $\mathcal L$ is defined as below. \begin{equation}\label{eq_Q1} Q^{\mathcal L}(s,a) := f(s,a) - \eta^{\mathcal L} + \sum_{s' \in \mathcal S} p(s,a,s')g^{\mathcal L}(s'), \end{equation} where $p(s,a,s')$ is the probability of which the system transits from the current state $s$ to the next state $s'$ if the action $a$ is adopted. For a randomized policy, $\mathcal L$ is a mapping $\mathcal L: \mathcal S \rightarrow \Omega(\mathcal A)$, where $\Omega(\mathcal A)$ is the set of probability measurements over the action space $\mathcal A$. That is, $\mathcal L(s,a)$ is the probability of which action $a$ is adopted at state $s$, $s \in \mathcal S$ and $a \in \mathcal A$. For the deterministic case, $\mathcal L$ is a mapping $\mathcal L: \mathcal S \rightarrow \mathcal A$, i.e., $\mathcal L(s)$ indicates the action adopted at state $s$. We can rewrite (\ref{eq_Q1}) in a recursive form as below. \begin{equation}\label{eq_QPoisson} Q^{\mathcal L}(s,a) = f(s,a) - \eta^{\mathcal L} + \sum_{s' \in \mathcal S} p(s,a,s') \sum_{a' \in \mathcal A}\mathcal L(s',a')Q^{\mathcal L}(s',a'). \end{equation} The above equation is a fixed-point equation for Q-factors, which is similar to the Poisson equation for the performance potential or the value function in the classical MDP theory. By sorting the element $Q^{\mathcal L}(s,a)$ in a vector form, we define an $SA$-dimensional column vector $\bm Q^{\mathcal L}$ as below. \begin{equation} \bm Q^{\mathcal L}(s) := \left[ \begin{array}{c} Q^{\mathcal L}(s,a_1)\\ Q^{\mathcal L}(s,a_2)\\ \vdots\\ Q^{\mathcal L}(s,a_A) \end{array} \right], \ \forall s \in \mathcal S. \qquad \bm Q^{\mathcal L} := \left[ \begin{array}{c} \bm Q^{\mathcal L}(1)\\ \bm Q^{\mathcal L}(2)\\ \vdots\\ \bm Q^{\mathcal L}(S) \end{array} \right]. \end{equation} With the same order to sort the element $s,a,s',a'$, we can rewrite (\ref{eq_QPoisson}) in a matrix form as below. \begin{equation}\label{eq_QPoisson2} \bm Q^{\mathcal L} = \bm f - \eta^{\mathcal L} \bm e + \bm P \bm L \bm Q^{\mathcal L}, \end{equation} where $\bm Q^{\mathcal L}$ and $\bm r$ are column vectors with size $SA$, $\bm e$ is a column vector of ones with a proper size (here the size is $SA$), $\bm P$ is a stochastic matrix with size $SA \times S$ \begin{equation} \bm P := \Big [\bm P((s,a),s') \Big]_{SA \times S}, \quad \bm P((s,a),s')=p(s,a,s'), \ \forall s,s' \in \mathcal S, \ a \in \mathcal A, \end{equation} $\bm L$ is a stochastic matrix with size $S \times SA$ \begin{equation} \bm L := \Big [\bm L(s',(s',a')) \Big]_{S \times SA}, \quad \bm L(s',(s',a'))=\mathcal L(s',a'), \ \forall s' \in \mathcal S, \ a' \in \mathcal A, \end{equation} where $\bm L(s,(s',a'))=0$ if $s \neq s'$. Therefore, most of the elements of $\bm L$ is 0 and $\bm L$ is a sparse matrix like below \begin{equation} \bm L = \left[ \begin{array}{ccccc} \mathcal L(1,:) & \bm 0 & \bm 0 & \cdots & \bm 0\\ \bm 0 & \mathcal L(2,:) & \bm 0 & \cdots & \bm 0\\ \vdots & \vdots & \vdots & \ddots & \vdots\\ \bm 0 & \bm 0 & \bm 0 & \cdots & \mathcal L(S,:) \end{array} \right]_{S \times SA}. \end{equation} If we write $\bm {\mathcal L}$ as an $S \times A$ matrix as below. \begin{equation} \bm {\mathcal L} := \left[ \begin{array}{cccc} \mathcal L(1,a_1),&\mathcal L(1,a_2),&\cdots,&\mathcal L(1,a_A)\\ \mathcal L(2,a_1),&\mathcal L(2,a_2),&\cdots,&\mathcal L(2,a_A)\\ \vdots & \vdots & \ddots & \vdots\\ \mathcal L(S,a_1),&\mathcal L(S,a_2),&\cdots,&\mathcal L(S,a_A)\\ \end{array} \right]_{S \times A}. \end{equation} Then matrix $\bm L$ equals the block diagonal of \emph{Kronecker product} (matrix form of tensor-product) of vector $\bm e^T$ and matrix $\bm {\mathcal L}$, where $\bm e^T$ is an $S$-dimensional row vector of ones. That is, we have \begin{equation} \bm L = \diag (\bm e^T \otimes \bm {\mathcal L}) = \left[ \begin{array}{cccc} \bm {\mathcal L}(1,:),&\bm 0,&\cdots,&\bm 0\\ \bm 0,&\bm {\mathcal L}(2,:),&\cdots,&\bm 0\\ \vdots & \vdots & \ddots & \vdots\\ \bm 0,&\bm 0,&\cdots,&\bm {\mathcal L}(S,:)\\ \end{array} \right]_{S \times SA}. \end{equation} It is easy to verify that \begin{equation} \begin{array}{l} \bm L \bm e = \bm e. \\ \bm P \bm e = \bm e. \end{array} \end{equation} Therefore, \begin{equation} \bm P \bm L \bm e = \bm e, \end{equation} and $\bm P \bm L$ is still a stochastic matrix that can be denoted as matrix $\bm {\tilde{P}}$ with size $SA \times SA$. That is, \begin{equation} \bm {\tilde{P}} := \bm P \bm L. \end{equation} Therefore, we obtain a linear form of (\ref{eq_QPoisson2}) as below. \begin{center} \begin{boxedminipage}{1\columnwidth} \begin{equation}\label{eq_QPoisson3} (\bm I - \bm {\tilde{P}} )\bm Q^{\mathcal L} = \bm f - \eta^{\mathcal L} \bm e. \end{equation} \vspace{-10pt} \end{boxedminipage} \end{center} We can see that for any solution of $\bm Q^{\mathcal L}$ satisfying (\ref{eq_QPoisson3}), $\bm Q^{\mathcal L} + c \bm e$ is still a solution to (\ref{eq_QPoisson3}), where $c$ is any constant. Therefore, we can let $\bm Q^{\mathcal L}$ satisfy \begin{equation} \bm r \bm Q^{\mathcal L} = \eta, \end{equation} where $\bm r$ is any $SA$-dimensional row vector satisfying $\bm r \bm e \neq 0$. Substituting the above equation into (\ref{eq_QPoisson3}), we obtain \begin{equation}\label{eq_QPoisson4} (\bm I - \bm {\tilde{P}} + \bm e \bm r)\bm Q^{\mathcal L} = \bm f. \end{equation} Since $\bm {\tilde{P}} $ is a stochastic matrix, with Theorem~\ref{theorem2}, we know that $(\bm I - \bm {\tilde{P}} + \bm e \bm r)$ is invertible. Therefore, we have the following solution of Q-factors \begin{center} \begin{boxedminipage}{1\columnwidth} \begin{equation}\label{eq_QPoisson5} \bm Q^{\mathcal L} = (\bm I - \bm {\tilde{P}} + \bm e \bm r)^{-1} \bm f, \qquad \bm r \bm e \neq 0. \end{equation} \vspace{-10pt} \end{boxedminipage} \end{center} Therefore, we obtain the Poisson equation (\ref{eq_QPoisson3}) for Q-factors, which is a fixed point equation to solve Q-factors. The closed-form solution of Q-factors is also obtained in (\ref{eq_QPoisson5}). Based on these equations, we may also develop numerical computation algorithms or online estimation algorithms for Q-factors, similar to the discussion and algorithms in Subsection~\ref{subsection_Alg}. Recently, the deep reinforcement learning, such as the AlphaGo of Google, is becoming a promising direction of artificial intelligence \cite{Silver16} and the Q-factors are fundamental quantities in reinforcement learning \cite{Sutton98}, how to efficiently compute, estimate or even represent the Q-factors is an interesting topic that deserves further research efforts. \section{Conclusion} \label{section_Conclusion} In this paper, we study a generalized fundamental matrix in Markov systems. Different from the fundamental matrix in the classical MDP theory, the generalized fundamental matrix does not require the pre-computation of $\bm \pi$ and it can provide a more concise form for the computation of some fundamental quantities in Markov systems. Based on the generalized fundamental matrix, we give a closed-form solution to the Poisson equation and represent the values of performance potentials, steady state distribution, and Q-factors of Markov systems. The new representation of these solutions may shed some light on efficiently computing or estimating these fundamental quantities from a new perspective, which is very important for the performance optimization of Markov decision processes. \section*{Acknowledgment} The first author was supported in part by the National Natural Science Foundation of China (61573206, 61203039, U1301254) and would like to thank X. R. Cao, J. J. Hunter, C. D. Meyer, and M. L. Puterman for their helpful discussions and comments.

https://arxiv.org/abs/0910.3456

Coherent tangent bundles and Gauss-Bonnet formulas for wave fronts

We give a definition of `coherent tangent bundles', which is an intrinsic formulation of wave fronts. In our application of coherent tangent bundles for wave fronts, the first fundamental forms and the third fundamental forms are considered as induced metrics of certain homomorphisms between vector bundles. They satisfy the completely same conditions, and so can reverse roles with each other. For a given wave front of a 2-manifold, there are two Gauss-Bonnet formulas. By exchanging the roles of the fundamental forms, we get two new additional Gauss-Bonnet formulas for the third fundamental form. Surprisingly, these are different from those for the first fundamental form, and using these four formulas, we get several new results on the topology and geometry of wave fronts.

\section{Introduction} In this paper, we give a definition of {\it coherent tangent bundles}, which is an intrinsic formulation of wave fronts. The advantage of this formulation is that the first fundamental forms and the third fundamental forms satisfy exactly the same conditions for wave fronts in space forms, and they can reverse roles with each other. Then the two Gauss-Bonnet formulas induce two more Gauss-Bonnet formulas by exchanging the first and third fundamental forms. These turn out to be different from those for the first fundamental form, and using these four formulas, we get new results on fronts. In Section~\ref{sec:coherent}, we generalize the definitions of {\em coherent tangent bundles\/} and {\it singular curvature}, which were given for $2$-dimensional coherent tangent bundles (or fronts) in \cite{SUY1} and \cite{SUY2}, to fronts of general dimension. A coherent tangent bundle on an $m$-manifold $M^m$ induces a positive semi-definite metric on $M^m$ (called the first fundamental form), and can be regarded as a generalization of Riemannian manifolds. Using the concept of coherent tangent bundles, we can give a unified treatment of hypersurfaces and $C^\infty$-maps between the same dimensional manifolds. The points where the metric degenerates are called {\it singular points}. We define a notion of $A_k$ (singular) points, which is a generalization of that of fronts and of $C^\infty$-Morin maps at the same time. We define {\it singular principal curvatures\/} for each $A_2$-point of a given coherent tangent bundle, which are $(m-1)$-tuples of real numbers. One of them diverges to $-\infty$ at $A_3$-points as shown in Theorem \ref{thm:infty}. This is a generalization of the result in \cite{SUY1} for $2$-dimensional fronts. Moreover, as shown in Section~\ref{sec:gauss-bonnet}, our intrinsic setting enables us to introduce the singular curvature on $A_2$-Morin singular points. When $m=2$, as an application of the two intrinsic Gauss-Bonnet formulas, we give a new proof of Quine's formula (cf.\ Proposition~\ref{prop:quine}), and also get a new formula (cf.\ Proposition~\ref{prop:id}) for the total singular curvature of generic smooth maps between $2$-manifolds. Furthermore, in Section~\ref{sec:gauss-bonnet}, we also give several results on compact $2$-dimensional fronts, as an application of our new Gauss-Bonnet formulas for the third fundamental form. In particular, we show that the total negative Gaussian curvature $\int_{M^2}\min(0,K)\,dA$ of a closed immersed surface $M^2$ in $\boldsymbol{R}^3$ is equal to the signed sum of total geodesic curvature on each connected component of the singular set of its Gauss map, see Theorem~\ref{thm:c}. The deepest applications are given for surfaces of bounded Gaussian curvature. For example, the Euler characteristic of a closed wave front with Gaussian curvature $-1/\delta<K<-\delta$ ($\delta>0$) in a $3$-dimensional flat torus vanishes (cf.\ Theorem~\ref{thm:g}). \section{Coherent tangent bundles} \label{sec:coherent} Let $M^m$ be an oriented $m$-manifold ($m\ge 1$) and $\mathcal{E}$ a vector bundle of rank $m$ over $M^m$. Before defining coherent tangent bundles, we define $A_k$-points for each vector bundle homomorphism from the tangent bundle of $M^m $to $\mathcal{E}$. \subsection{$A_k$-points} We assume that $\mathcal{E}$ is orientable, namely there exists a non-vanishing section $\mu:M^m\to \op{det}(\mathcal{E}^*)$, where $\det(\mathcal{E}^*)$ is the determinant line bundle of the dual bundle $\mathcal{E}^*$. This assumption does not create any restriction for the local theory. In fact, for each point $p\in M^m$, there exists a section $\mu$ of $\det(\mathcal{E}^*)\setminus\{0\}$ defined on a neighborhood $U$ of $p$. Now we fix a bundle homomorphism \[ \phi:TM^m\longrightarrow \mathcal{E}, \] where $TM^m$ is the tangent bundle of $M^m$. A point $p\in M^m$ is called a {\it $\phi$-singular point\/} if $\phi_p\colon{}T_pM^m\to\mathcal{E}_p$ is not a bijection, where $\mathcal{E}_p$ is the fiber of $\mathcal{E}$ at $p$. We denote by $\Sigma_{\phi}$ the set of $\phi$-singular points on $M^m$. On the other hand, a point $p\in M^m\setminus\Sigma_{\phi}$ is called a {\it $\phi$-regular point}. We now fix a $\phi$-singular point $p$, and take a positively oriented local coordinate neighborhood $(U;u_1,\dots,u_m)$ of $p$. THen it holds that \[ \varphi^*\mu=\lambda_\varphi\,du_1\wedge\dots\wedge du_m, \] where \begin{equation}\label{eq:jacobian} \lambda_\varphi =\mu(\varphi_1,\ldots,\varphi_n):U\to\boldsymbol{R}% ,\qquad \left( \varphi_j= \phi\left(\frac{\partial}{\partial u_j}\right)% ,\ j=1,\dots,m \right) \end{equation} is a $C^\infty$-function. We call the function $\lambda_\varphi$ the {\em $\varphi$-Jacobian function}. The set of $\phi$-singular points on $U$ is expressed as \begin{equation}\label{eq:singular0} \Sigma_{\phi}\cap U:=\{p\in U\,;\,\lambda_{\phi}(p)=0 \}. \end{equation} We now set \begin{equation}\label{eq:M_pm} M^+_\phi:=\{q\in M^m\,;\, \lambda_\phi(q)>0\},\qquad M^-_\phi:=\{q\in M^m\,;\, \lambda_\phi(q)<0\}. \end{equation} Since the sign of $\lambda_\phi$ does not depend on the choice of the positively oriented local coordinate system, the above definitions of $M^\pm_\phi$ make sense. A $\phi$-singular point $p$ $(\in\Sigma_{\phi})$ is called {\em non-degenerate\/} if $d\lambda_\phi$ does not vanish at $p$. On a neighborhood of a non-degenerate $\phi$-singular point, the $\phi$-singular set consists of an $(m-1)$-submanifold in $M^m$, called the {\em $\phi$-singular submanifold}. If $p$ is a non-degenerate $\phi$-singular point, the rank of $\phi_p$ is $m-1$. The direction of the kernel of $\phi_p$ is called the {\em null direction}. Let $\eta$ be {\it a null vector field}, namely, smooth (non-vanishing) vector field along the $\phi$-singular submanifold $\Sigma_\phi$, which gives the null direction at each point in $\Sigma_\phi$. \begin{definition}[$A_2$-points]% \label{def:a2-point} A non-degenerate $\phi$-singular point $p\in M^m$ is called an {\em $A_2$-point\/}, or an {\em $A_2$-point of $\phi$\/}, if the null direction $\eta(p)$ is transversal to the $\varphi$-singular submanifold. \end{definition} We set \begin{equation}\label{eq:lambda-prime} \lambda_\phi':=d\lambda(\tilde\eta), \end{equation} where $\tilde\eta$ is a vector field on a neighborhood $U$ of $p$ which gives a null vector field on $\Sigma_\phi\cap U$. Then $p$ is an $A_2$-point if and only if the function $\lambda'_\phi$ does not vanish at $p$ (see \cite[Theorem 2.4]{SUY4}). \begin{definition}[$A_3$-points] \label{def:a3-point} A non-degenerate $\varphi$-singular point $p\in M^m$ is called an {\em $A_3$-point} or an {\em $A_3$-point of $\varphi$} if it is not an $A_2$-point of $\varphi$, but $d\lambda_\varphi'(\eta)$ does not vanish. \end{definition} Similarly, we can define $A_{k+1}$-points inductively. We set \begin{equation}\label{eq:lambda-k} \lambda^{(k)}_\varphi := d\lambda^{(k-1)}_\varphi(\tilde{\eta}) : \Sigma_{\varphi}\to{\boldsymbol{R}}\qquad (\lambda_\varphi^{(1)}:=\lambda_\varphi',\ \lambda_\varphi^{(0)}:=\lambda_\varphi). \end{equation} If $(d\lambda^{(k-1)}_\varphi)_p\ne0$ on $T_p\Sigma^{k-1}$, then we can define a subset of the submanifold $\Sigma_\varphi^k$ by \[ \Sigma^{k+1}_\varphi := \{q\in\Sigma_\varphi^k;\lambda^{(k)}_\varphi(q)=0\} = \{q\in\Sigma_\varphi^k;\eta(q)\in T_q\Sigma_\varphi^k\}, \] where we set $\Sigma_\varphi^1:=\Sigma_\varphi$ and $\Sigma_\varphi^0:=M^m$. If a non-degenerate $\varphi$-singular point $p$ is not an $A_k$-point of $\varphi$, but $d\lambda^{(k-1)}_\varphi(\tilde\eta)$ is well-defined inductively and is nonvanishing at $p$, then it is called an {\em $A_{k+1}$-point of} $\varphi$. Here an $A_1$-point means a $\varphi$-regular point. (See [17, Section 2] for details.) \subsection{Coherent tangent bundles} As a generalization of the $2$-dimensional case in \cite{SUY1} and \cite{SUY2}, we give a general setting for intrinsic fronts: Let $M^m$ be an oriented $m$-manifold ($m\ge 1$). A {\em coherent tangent bundle\/} over $M^m$ is a $5$-tuple $(M^m,\mathcal{E},\inner{~}{~},D,\varphi)$, where \begin{enum} \item $\mathcal{E}$ is a vector bundle of rank $m$ over $M^m$ with an inner product $\inner{~}{~}$, \item $D$ is a metric connection on $(\mathcal{E},\inner{~}{~})$, \item $\varphi\colon{}TM^m\to \mathcal{E}$ is a bundle homomorphism which satisfies \begin{equation}\label{eq:c} D^{}_{X}\phi(Y)-D^{}_{Y}\phi(X)-\phi([X,Y])=0 \end{equation} for vector fields $X$ and $Y$ on $M^m$. \end{enum} In this setting, the pull-back of the metric \begin{equation}\label{eq:phi-metric} ds^2_\phi:=\phi^*\inner{~}{~} \end{equation} is called the {\em $\phi$-metric}, which is a positive semidefinite symmetric tensor on $M^m$. It should be remarked that as in \eqref{eq:c}, \begin{equation}\label{eq:torsion} T_\phi(X,Y):=D_X\phi(Y)-D_Y\phi(X)-\phi([X,Y]) \end{equation} gives a skew-symmetric tensor on $M^m$, called the {\it torsion tensor\/} of $D$ with respect to $\phi$. If $(M^m,\mathcal{E},\inner{~}{~},D,\phi)$ is a coherent tangent bundle, then the pull back of the connection $D$ by $\phi$ coincides with the Levi-Civita connection. In this sense, a coherent tangent bundle can be considered as a generalization of Riemannian manifold. In fact, as an application, a duality of conformally flat manifolds with admissible singularities is given in terms of coherent tangent bundles in \cite{LUY}. A coherent tangent bundle $(M^m,\mathcal{E},\inner{~}{~},D,\phi)$ is called {\it co-orientable\/} if the vector bundle $\mathcal{E}$ is orientable, namely, there exists a globally defined smooth section $\mu$ of $\det(\mathcal{E}^*)$ such that \begin{equation}\label{eq:mu} \mu(\vect{e}_1,\dots,\vect{e}_m)=\pm 1 \end{equation} for any orthonormal frame $\{\vect{e}_1,\dots,\vect{e}_m\}$ on $\mathcal{E}$. The form $\mu$ is determined uniquely up to a $\pm$-ambiguity. A {\em co-orientation\/} of the coherent tangent bundle $\mathcal{E}$ is a choice of $\mu$. An orthonormal frame $\{\vect{e}_1,\dots,\vect{e}_m\}$ is called {\em positive\/} with respect to the co-orientation $\mu$ if $\mu(\vect{e}_1,\dots,\vect{e}_m)=+1$. We give here typical examples of coherent tangent bundles: \begin{example}\label{ex:map} Let $M^m$ be an oriented $m$-manifold and $(N^m,g)$ an oriented Riemannian $m$-manifold. A $C^\infty$-map $f:M^m\to N^m$ induces a coherent tangent bundle over $M^m$ as follows: Let $\mathcal{E}_f:=f^*TN^m$ be the pull-back of the tangent bundle $TN^m$ by $f$. Then $g$ induces a positive definite metric $\inner{~}{~}$ on $\mathcal{E}_f$, and the restriction $D$ of the Levi-Civita connection of $g$ gives a connection on $\mathcal{E}$ which is compatible with respect to the metric $\inner{~}{~}$. We set $\phi^{}_f:=df:TM^m\to\mathcal{E}_f$, which gives the structure of the coherent tangent bundle on $M^m$. A necessary and sufficient condition for a given coherent tangent bundle over an $m$-manifold to be realized as a smooth map into the $m$-dimensional space form is given in \cite{SUY6}. \end{example} \begin{example}\label{ex:front} Let $(N^{m+1},g)$ be an $(m+1)$-dimensional Riemannian manifold. A $C^\infty$-map $f:M^m\to N^{m+1}$ is called a {\it frontal\/} if for each $p\in M^m$, there exists a neighborhood $U$ of $p$ and a unit vector field $\nu$ along $f$ defined on $U$ such that $g\bigl(df(X),\nu\bigr)=0$ holds for any vector field $X$ on $U$ (that is, $\nu$ is a unit normal vector field), and the map $\nu\colon{}U\to T_1N^{m+1}$ is a $C^{\infty}$-map, where $T_1N^{m+1}$ is the unit tangent bundle of $N^{m+1}$. Moreover, if $\nu$ can be taken to be an immersion as a smooth section of $T_1N^{m+1}$ for each $p\in M^m$, $f$ is called a {\em front} or a {\em wave front}. We remark that $f$ is a front if and only if $f$ has a lift $L_f:M^m\to P\bigl(T^*N^{m+1}\bigr)$ as a Legendrian immersion, where $P(T^*N^{m+1})$ is a projectified cotangent bundle on $N^{m+1}$ with the canonical contact structure. The subbundle $\mathcal{E}_f$ which consists of the vectors in the pull-back bundle $f^*TN^{m+1}$ perpendicular to $\nu$ gives a coherent tangent bundle. In fact, $\phi_f\colon{}TM^m\ni X \mapsto df(X)\in \mathcal{E}_f$ gives a bundle homomorphism. Let $\nabla$ be the Levi-Civita connection on $N^{m+1}$. Then by taking the tangential part of $\nabla$, it induces a connection $D$ on $\mathcal{E}_f$ satisfying \eqref{eq:c}. Let $\inner{~}{~}$ be a metric on $\mathcal{E}_f$ induced from the Riemannian metric on $N^{m+1}$, then $D$ is a metric connection on $\mathcal{E}_f$. Thus we get a coherent tangent bundle $(M^{m},\mathcal{E}_f,\inner{~}{~},D,\phi_f)$. Since the unit tangent bundle can be canonically identified with unit cotangent bundle, the map $\nu\colon{}U\to T_1N^{m+1}$ can be identified with $L_f|_U$. A frontal $f$ is called {\it co-orientable\/} if there is a unit normal vector field $\nu$ globally defined on $M^m$. When $N^{m+1}$ is orientable, the coherent tangent bundle is co-orientable if and only if so is $f$. \end{example} From now on, we assume that $(M^m,\mathcal{E},\inner{~}{~},D,\phi)$ is a co-orientable coherent tangent bundle, and fix a co-orientation $\mu$ on the coherent tangent bundle. (If $\mathcal{E}$ is not co-orientable, one can take a double cover $\pi\colon{}\widehat M^m\to M^m$ such that the pull-back of $\mathcal{E}$ by $\pi$ is a co-orientable coherent tangent bundle over $\widehat M^m$.) \begin{definition}\label{def:volume} The {\em signed $\phi$-volume form\/} $d\hat A_{\phi}$ and the ({\em unsigned}) {\em $\phi$-volume form \/} $dA_{\phi}$ are defined as \begin{equation}\label{eq:volume-form} \begin{aligned} d\hat A_\phi := \phi^*\mu = \lambda_{\phi}\,du_1\wedge \dots \wedge du_m,\quad dA_\phi := |\lambda_{\phi}|\,du_1 \wedge \dots \wedge du_m, \end{aligned} \end{equation} where $(U;u_1,\dots,u_m)$ is a local coordinate system of $M^m$ compatible with the orientation of $M^m$, and $\lambda_{\phi}$ is the $\phi$-Jacobian function on $U$ given in \eqref{eq:jacobian}. Both $d\hat A_{\phi}$ and $dA_{\phi}$ are independent of the choice of a positively oriented local coordinate system $(U;u_1,\dots,u_m)$, and give two globally defined $m$-forms on $M^m$. ($d\hat A_{\phi}$ is $C^\infty$-differentiable, but $dA_{\phi}$ is only continuous.) When $M^m$ has no $\phi$-singular points, the two forms coincide up to sign. Then the two sets $M^\pm_{\phi}$ as in \eqref{eq:M_pm} are written as \begin{align*} M^+_{\phi}&:=\bigl\{p\in M^m\setminus \Sigma_{\phi} \,;\, d\hat A_{\phi}(p)=dA_{\phi}(p)\bigr\}, \\ M^-_{\phi}&:=\bigl\{p\in M^m\setminus \Sigma_{\phi}\,;\, d\hat A_\phi(p)=-dA_{\phi}(p) \bigr\}. \end{align*} The $\phi$-singular set $\Sigma_{\phi}$ coincides with the boundary $\partial M^+_\phi=\partial M^-_\phi$. \end{definition} \subsection{Singularities of $2$-dimensional coherent tangent bundle}~ When $m=2$ and $(M^2,\mathcal{E},\inner{~}{~},D,\phi)$ comes from a front in $3$-manifold as in Example~\ref{ex:front} (resp.\ a map into $2$-manifold as in Example~\ref{ex:map}), an $A_2$-point corresponds to a cuspidal edge (resp.\ a fold) (cf.\ \cite{SUY3}, see Fig.~\ref{fig:folds}). Though cuspidal cross caps in surfaces in $\boldsymbol{R}^3$ are not singular points of a front, they are also $A_2$-points (see \cite[Remark 1.6]{SUY2}). In this way, our definition of $A_2$-points are wider than the original definition of $A_2$-points on fronts or Morin maps. However, if $(M^2,\mathcal{E},\inner{~}{~},D,\phi)$ is associated to a wave front, $A_2$-points really corresponds to cuspidal edges on its realization as fronts. \begin{figure}[htb] \begin{center} \includegraphics[height=2.5cm]{foldshade.eps}\hspace{1cm} \includegraphics[height=2.5cm]{cuspidaledge.eps} \end{center} \caption{% A fold in a plane (left) and a cuspidal edge in a space (right).} \label{fig:folds} \end{figure} On the other hand, an $A_3$-point corresponds to a swallowtail (resp.\ a cusp) when $m=2$ and $(M^2,\mathcal{E},\inner{~}{~},D,\phi)$ comes from a front in a $3$-manifold (resp.\ a map into a $2$-manifold). This fact was shown in \cite{KRSUY} (see Fig.~\ref{fig:cusps}). \begin{figure}[htb] \begin{center} \includegraphics[height=2.5cm]{cuspshade.eps}\hspace{1cm} \includegraphics[height=2.5cm]{swallowtail.eps} \end{center} \caption{% A cusp in a plane (left) and a swallowtail in a space (right).} \label{fig:cusps} \end{figure} When $\phi$ is induced from a smooth map between $2$-manifolds (resp.\ a front in a $3$-manifold), it is well-known that $A_2$-points and $A_3$-points are generic singular points of $\phi$ (cf.\ \cite{AGV}). However, if one consider deformations of $A_2$ or $A_3$-points, an additional three types of singular points of type $A_3^{+}$ (lips), $A_3^{-}$ (beaks) and of $A_4$ (butterfly) also generically appear. So we mention here an intrinsic characterization of these three singular points. Let $(M^2,\mathcal{E},\inner{~}{~},D,\phi)$ be a coherent tangent bundle over a $2$-manifold $M^2$. \begin{definition}[Lips]\label{def:lips} A $\phi$-singular point $p\in M^2$ is called a {\em lips of $\phi$\/} if it satisfies the following conditions: \begin{enum} \item The rank of $\phi$ at $p$ is $1$, \item the exterior derivative $d\lambda_\phi$ vanishes at $p$, \item the Hessian matrix of the function $\lambda_\phi$ at $p$ is positive definite. \end{enum} \end{definition} \begin{definition}[Beaks]\label{def:beaks} A $\phi$-singular point $p\in M^2$ is called a {\em beaks of $\phi$\/} if it satisfies the following conditions: \begin{enum} \item The rank of $\phi$ at $p$ is $1$, \item the exterior derivative $d\lambda_\phi$ vanishes at $p$, \item the Hessian matrix of the function $\lambda_\phi$ at $p$ is negative definite, \item the second derivative $\lambda''_\phi(=d\lambda'_\phi(\eta))$ does not vanish at $p$, where the prime means the derivative with respect to the null direction (cf. \eqref{eq:lambda-prime} and \eqref{eq:lambda-k}). \end{enum} \end{definition} \begin{definition}[Butterfly]\label{def:b-fly} A non-degenerate $\phi$-singular point $p\in M^2$ is called a {\em butterfly of $\phi$\/} if it satisfies \[ \lambda'_\phi(p)=\lambda''_\phi(p)=0,\quad \lambda'''_\phi(p)\ne 0, \] where the prime means the derivative with respect to the null direction. \end{definition} \begin{remark}\label{fact:saji} When a coherent tangent bundle is induced from a front, lips, beaks and butterfly correspond to cuspidal lips, cuspidal beaks and cuspidal butterfly, respectively (Fig.~\ref{fig:front-peaks}). See \cite{IST} and \cite{IS}. On the other hand, if a coherent tangent bundle is induced from a smooth map between $2$-manifolds, these singular points are corresponding to those on the map (Fig.~\ref{fig:map-peaks}). See \cite{S}. \begin{figure}[htb] \begin{center} \includegraphics[height=3.5cm]{cuspidallipsf.eps}\hspace{5mm} \includegraphics[height=3.5cm]{cuspidalbeaksf.eps}\hspace{5mm} \includegraphics[height=3.5cm]{cuspidalbflyf.eps} \caption{% A cuspidal lips (left), a cuspidal beaks (center) and a cuspidal butterfly (right) in $\boldsymbol{R}^3$.} \label{fig:front-peaks} \end{center} \begin{center} \includegraphics[height=3.2cm]{lipsshadef.eps}\hspace{1cm} \includegraphics[height=3.2cm]{beaksshadef.eps}\hspace{1cm} \includegraphics[height=3.2cm]{bflyshadef.eps} \caption{% A lips (left), a beaks (center) and a butterfly (right) in $\boldsymbol{R}^2$.} \label{fig:map-peaks} \end{center} \end{figure} In this way, our intrinsic formulation can give a unified treatment of singular points on maps and on fronts at the same time. \end{remark} \subsection{Singular curvatures} Let $(M^m,\mathcal{E},\inner{~}{~},D,\phi)$ be a coherent tangent bundle and fix a $\phi$-singular point $p\in\Sigma_{\phi}$ which is an $A_2$-point. Then there exists a neighborhood $U$ of $p$ such that $\Sigma_{\phi}\cap U$ consists of $A_2$-points. Now we define the {\it singular shape operator\/} as follows: Since the kernel of $\phi_p$ is transversal to $\Sigma_\phi$ at $p$, $\phi|_{T(\Sigma_{\phi}\cap U)}$ is injective, where $U$ is a sufficiently small neighborhood of $p$. Then the metric $ds^2_\phi$ is positive definite on $\Sigma_\phi\cap U$. We take an orthonormal frame field $e_1$, $e_2$,\dots, $e_{m-1}$ on $\Sigma_{\phi}\cap U$ with respect to $ds^2_\phi$. Without loss of generality, we may assume that $(e_1,e_2,\dots,e_{m-1})$ is smoothly extended on $U$ as an orthonormal $(m-1)$-frame field. Then we can take a unique smooth section $\vect{n}:U\to \mathcal{E}$ (called the {\it conormal vector field}) so that $(\phi(e_1),\dots,\phi(e_{m-1}), \vect{n})$ gives a positively oriented orthonormal frame field on $\mathcal{E}$. Now, we set \begin{equation}\label{eq:sing-shape} S_\phi(X):= -\operatorname{sgn}\left(d\lambda_\phi\bigl(\eta(q)\bigr)\right) \phi^{-1}(D_X\vect n) \quad (X\in T_q\Sigma_\phi,\,\,q\in \Sigma_\phi\cap U), \end{equation} where the non-vanishing null vector field $\eta$ is chosen so that $(e_1,\dots,e_{m-1},\eta)$ is compatible with respect to the orientation of $M^m$. It holds that \begin{equation}\label{eq:sign2} \operatorname{sgn}\bigl(d\lambda_\phi(\eta(q))\bigr) = \begin{cases} \hphantom{-} 1 & \mbox{if $\eta(q)$ points toward $M^+_{\phi}$},\\ -1 & \mbox{if $\eta(q)$ points toward $M^-_{\phi}$}. \end{cases} \end{equation} Since $\phi$ is injective on each tangent space of $\Sigma_\phi$ and $D_X\vect{n}\in \phi(T\Sigma_\phi)$, the inverse element $\phi^{-1}(D_X\vect{n})$ is uniquely determined. Thus we get a bundle endomorphism $S_\phi:T\Sigma_\phi \to T\Sigma_\phi$ which is called the {\it singular shape operator\/} on $\Sigma_\phi$. \begin{proposition}[A generalization of Theorem 1.6 in \cite{SUY1}] \label{prop:sym} The definition of the singular shape operator $S_\phi$ is independent of the choice of an orthonormal frame field $e_1,\dots,e_{m-1}$, the choice of an orientation of $M^m$, and the choice of a co-orientation of $\mathcal{E}$. Moreover, it holds that \[ ds^2_\phi\bigl(S_\phi(X),Y\bigr)=ds^2_\phi\bigl(X,S_\phi(Y)\bigr) \qquad (X,Y\in T_q\Sigma_\phi,~q\in \Sigma_\phi). \] \end{proposition} \begin{proof} In fact, if one reverses the orientation of the basis $(e_1,\dots,e_{m-1})$, then $\vect{n}$ and $\eta$ change sign at the same time, and thus $S_{\phi}$ is unchanged. Similarly, if we reverse the orientation of $M^m$ (resp.\ $\mathcal{E}$), then $\lambda_\phi$ and $\eta$ (resp.\ $\lambda_\phi$ and $\vect{n}$) change sign at the same time, and $S_\phi$ unchanged. We can get the final assertion from the following identity \begin{align*} ds^2_\phi&(\phi^{-1}(D_X\vect{n}),Y)= \inner{D_X\vect{n}}{\phi(Y)} =-\inner{\vect{n}}{D_X\phi(Y)}\\ &=-\inner{\vect{n}}{D_Y\phi(X)+\phi\bigl([X,Y]\bigr)}\\ &=-\inner{\vect{n}}{D_Y\phi(X)} = \inner{D_Y\vect{n}}{\phi(X)} =ds^2_\phi\bigl(\phi^{-1}(D_Y\vect{n}),X\bigr), \end{align*} where $X$ and $Y$ are both vector fields on $\Sigma_\phi$. \end{proof} Hence $S_\phi$ is symmetric with respect to $ds^2_{\varphi}$. \begin{definition}\label{def:singular-curvature} Let $p\in\Sigma_\phi$ be an $A_2$-point of $\phi$. Then \begin{equation}\label{eq:kappa} \kappa_\phi(X):=ds^2_\phi(S_\phi(X),X)/ds^2_\phi(X,X), \qquad (X\in T_p\Sigma_{\phi}\setminus\{0\}) \end{equation} is called the {\it $\phi$-singular normal curvature\/} at $p$ with respect to the direction $X$. The eigenvalues of $S_\phi$ are called the {\it $\phi$-singular principal curvatures\/}, which give the critical values of the singular normal curvature on $T_p\Sigma_\phi$. \end{definition} When $m=2$, the $\phi$-singular principal curvature is called (simply) the {\it $\phi$-singular curvature}, which is also denoted by $\kappa_\phi$. This definition of the singular curvature is the same as in \cite[(1.7)]{SUY1} and \cite[(1.6)]{SUY2}. More precisely, $\kappa_\phi$ is computed as follows: Let $p\in\Sigma_\phi$ be an $A_2$-point of $\phi$. Then the $\phi$-singular set $\Sigma_\phi$ is parametrized by a regular curve $\gamma(t)$ ($t\in I\subset\boldsymbol{R}$) on $M^2$ on a neighborhood of $p$, and $\gamma(t)$ is an $A_2$-point of $\phi$ for each $t\in I$. Since $\dot\gamma(t)$ ($\dot{~}=d/dt$) is not a null-direction, $\phi\bigl(\dot\gamma(t)\bigr)\neq 0$. Take a section $\vect{n}(t)$ of $\mathcal{E}$ along $\gamma$ such that $\{\phi(\dot\gamma)/|\phi(\dot\gamma)|,\vect{n}\}$ gives a positive orthonormal frame field on $\mathcal{E}$ along $\gamma$, where $|\phi(\dot\gamma)|=\inner{\phi(\dot\gamma)}{\phi(\dot\gamma)}^{1/2}$. Then we have \begin{equation}\label{eq:singular-curvature-2} \kappa_\phi(t) :=\kappa_{\phi}\bigl(\dot\gamma(t)\bigr) = -\operatorname{sgn} \left(d\lambda_\phi\bigl(\eta(t)\bigr)\right) \frac{\inner{D_{d/dt}\vect{n}(t)}{\phi\bigl(\dot\gamma(t)\bigr)}}{% |\phi\bigl(\dot\gamma(t)\bigr)|^2}, \end{equation} where $\eta(t)$ is a null-vector field along $\gamma(t)$ such that $\{\dot\gamma(t),\eta(t)\}$ is compatible with the orientation of $M^2$. By \eqref{eq:sign2}, it holds that \begin{equation}\label{eq:sign3} \operatorname{sgn}\bigl(d\lambda_\phi(\eta(t))\bigr) =\begin{cases} \hphantom{-}1 & \mbox{if $M^+_{\phi}$ lies on the left-hand side of $\gamma$},\\ -1 & \mbox{if $M^-_{\phi}$ lies on the left-hand side of $\gamma$}. \end{cases} \end{equation} \subsection{Behavior of the singular curvatures} We now prove the following: \begin{theorem}[A generalization of {\cite[Corollary 1.14]{SUY1}}] \label{thm:infty} Let $p\in\Sigma_{\phi}$ be a $\phi$-singular point of a coherent tangent bundle $(M^m,\mathcal{E},\inner{~}{~},D,\varphi)$, and assume $p$ is non-degenerate but not an $A_2$-point of $\phi$. Take a regular curve $\gamma\colon [0,1]\ni t\mapsto \gamma(t)\in\Sigma_\phi$ such that such that $\gamma\bigl((0,1]\bigr)$ consists only of $A_2$-points of $\phi$ and $\gamma(0)=p$. Then one of the $\phi$-singular principal curvatures along $\gamma(t)$ diverges to $-\infty$. \end{theorem} \begin{proof} We can take a local coordinate system $(U;u_1,\dots,u_{m})$ on $M^m$ at $p$ satisfying the following properties: \begin{enum} \item\label{item:coord-1} $(u_1,\dots,u_{m})$ is compatible with respect to the orientation of $M^m$, \item\label{item:coord-2} the $\phi$-singular submanifold is characterized by the zeros of the last coordinate function, that is, $\Sigma_\phi\cap U=\{u_m=0\}$. \item\label{item:coord-3} the vector field $\eta:=\frac{\partial}{\partial u_1} + \delta \frac{\partial}{\partial u_m}$ gives the null direction, where $\delta:=\delta(u_1,\dots,u_{m-1})$ is a $C^\infty$-function satisfying $\delta(p)=0$. (In fact, $\eta(p)$ must tangent to the singular manifold, since $p$ is not an $A_2$-point.) \end{enum} For the sake of simplicity, we set \begin{equation}\label{eq:notation} \partial_j:=\frac{\partial}{\partial u_j},\quad \phi_j:=\phi(\partial_j),\quad D_j:=D_{\partial_j}\qquad (j=1,\dots,m). \end{equation} Since $\gamma(t)$ ($t\in (0,1]$) is an $A_2$-point of $\phi$, \ref{item:coord-2} and \ref{item:coord-3} imply that $\delta(t):=\delta\bigl(\gamma(t)\bigr)$ does not vanish for $t\in (0,1]$. Then $\{\partial_1,\dots,\partial_{m-1},\operatorname{sgn}(\delta)\eta\}$ forms a positive frame field on $TM^m$ along $\gamma(t)$ ($t\neq 0$). By definition, the $\phi$-singular normal curvature along $\gamma$ with respect to $\partial_1$ is given by \begin{align} \label{eq:kappa-phi-t} \kappa_\phi(t)\biggl( :&=\kappa_\phi(\partial_1)\biggr) =-\operatorname{sgn} \left(d\lambda_\phi\bigl(\operatorname{sgn}(\delta)\eta\bigr)\right) \frac{\inner{D_1\vect{n}}{\phi_1}}{|\phi_1|^2}\\ &=-\operatorname{sgn}(\delta) \operatorname{sgn} \bigl(d\lambda_\phi(\eta)\bigr) \frac{\inner{\vect{n}}{D_1\phi_1}}{|\phi_1|^2}. \nonumber \end{align} Since $\eta$ is a null vector, it holds that \begin{equation}\label{eq:null} \phi(\eta)=\phi_1+\delta \phi_m=0 \end{equation} on $\Sigma_\phi$. In particular, since $\eta=\partial_1$ at $p$, we have that \begin{equation}\label{eq:phi-m} \phi_m(p)\ne 0. \end{equation} Differentiating \eqref{eq:null} by $u_1$, we have \begin{equation}\label{eq:null2} D_1\phi_1+\delta_1 \phi_m+\delta D_1 \phi_m=0 \qquad \left(\delta_1:=\frac{\partial \delta}{\partial u_1}\right). \end{equation} We can identify $\bigwedge_{j=1}^{m-1}\mathcal{E}_q\cong \mathcal{E}_q$ for each $q\in M^m$ by using the inner product on $\mathcal{E}$, and then we set \begin{equation}\label{eq:use_next_paper} \vect{n}=\frac{\phi_1\wedge\dots \wedge \phi_{m-1}} {|\phi_1\wedge\dots \wedge \phi_{m-1}|}, \end{equation} which gives a conormal vector field such that $\{\phi_1,\dots,\phi_{m-1}, \vect{n}\}$ is a positive frame field on $\mathcal{E}$ along $\gamma(t)$ ($t\neq 0$). By \eqref{eq:null}, we have $\phi_1=-\delta \phi_m$. Then \begin{equation}\label{eq:n-red} \vect{n}= -\frac{\delta}{|\delta|}\cdot \frac{\phi_m\wedge\phi_2\wedge \dots \wedge \phi_{m-1}} {|\phi_m\wedge\phi_2\wedge \dots \wedge \phi_{m-1}|} =(-1)^{m-1} \operatorname{sgn}(\delta) \frac{\phi_2\wedge \dots \wedge \phi_{m}} {|\phi_2\wedge \dots \wedge \phi_{m}|}. \end{equation} Substituting \eqref{eq:null} and \eqref{eq:null2} into \eqref{eq:kappa-phi-t}, we have that \begin{align*} \kappa_\phi(t)&= -\operatorname{sgn}(\delta)\operatorname{sgn} \bigl(d\lambda_\phi(\eta)\bigr) \frac{\inner{\vect{n}}{\delta_1 \phi_m+\delta D_1 \phi_m}} {|\delta \phi_m|^2} \\ &=- \operatorname{sgn} \bigl(d\lambda_\phi(\eta)\bigr) \frac{\inner{\vect{n}}{D_1 \phi_m}}{|\delta|\, |\phi_m|^2}. \end{align*} Set \[ \Delta := \frac{(-1)^{m-1}\operatorname{sgn} \bigl(d\lambda_\phi(\eta)\bigr)} {|\delta|\, |\phi_m|^2|\phi_2\wedge \dots \wedge \phi_{m}|} \qquad\text{and}\qquad \lambda_m:=\frac{\partial\lambda_\phi}{\partial u_m}. \] Then substituting \eqref{eq:n-red} into the above equation and noticing that $\phi_1$ is proportional to $\phi_m$, we have \begin{align*} \frac{\kappa_\phi(t)}{\Delta}&= \inner{\phi_2\wedge \dots \wedge \phi_{m}}{D_m\phi_1}\\ &= \partial_m \inner{\phi_2\wedge\dots\wedge\phi_m}{\phi_1} -\inner{D_m(\phi_2\wedge\dots\wedge\phi_m)}{\phi_1}\\ &= (-1)^{m-1}\partial_m\mu(\phi_1,\dots,\phi_m) -\inner{\phi_2\wedge\dots\wedge D_m\phi_m}{\phi_1}\\ &= (-1)^{m-1}\lambda_m + \inner{\phi_2\wedge\dots\wedge D_m(\delta\phi_1)}{\phi_1}\\ &= (-1)^{m-1}\lambda_m + \delta\inner{\phi_2\wedge\dots\wedge D_m \phi_1}{\phi_1}. \end{align*} Hence \[ \kappa_\phi(t)= \frac{\operatorname{sgn} \bigl(d\lambda_\phi(\eta)\bigr)\operatorname{sgn}(\delta)} {|\delta|\, |\phi_m|^2|\phi_2\wedge \dots \wedge \phi_{m}|} \lambda_m + \text{(a bounded term)}, \] because of \eqref{eq:phi-m}. Since $d\lambda_\phi(\eta)=d\lambda_\phi(\partial_1+\delta \partial_m) =\delta \lambda_m$, we have that \[ \kappa_{\phi}(t)= -\frac{|\lambda_m|}{% |\delta|\,|\phi_m|^2|\phi_2\wedge \dots \wedge \phi_{m}|} +\mbox{(a bounded term)}. \] Since $p$ is non-degenerate and $\Sigma_\phi=\{u_m=0\}$, $\lambda_m\ne 0$ holds. Thus, $\kappa_{\phi}(t)$ tends to $-\infty$ as $t\to +0$. Hence, at least one of the singular principal curvatures tends to $-\infty$. \end{proof} \subsection{Frontal bundles} At the end of this section, we give a definition of frontal bundles as an intrinsic characterization of wave fronts in space forms. Let $M^m$ be an oriented $m$-manifold and $(M^m,\mathcal{E},\inner{~}{~},D,\phi)$ a co-orientable coherent tangent bundle over $M^m$. If there exists another bundle homomorphism $\psi:TM^m\to \mathcal{E}$ such that $(M^m,\mathcal{E},\inner{~}{~},D,\psi)$ is also a coherent tangent bundle and the pair $(\phi,\psi)$ of bundle homomorphisms satisfies a compatibility condition \begin{equation}\label{eq:compati} \inner{\phi(X)}{\psi(Y)}=\inner{\phi(Y)}{\psi(X)}, \end{equation} then $(M^m,\mathcal{E},\inner{~}{~},D,\phi,\psi)$ is called a {\it frontal bundle}. The bundle homomorphisms $\phi$ and $\psi$ are called the {\em first homomorphism\/} and the {\em second homomorphism}, respectively. We set \begin{align*} \operatorname{\mathit{I}}(X,Y)&:=ds^2_{\phi}(X,Y)=\inner{\phi(X)}{\phi(Y)}, \\ \operatorname{\mathit{I\!I}}(X,Y)&:=-\inner{\phi(X)}{\psi(Y)},\\ \operatorname{\mathit{I\!I\!I}}(X,Y)&:=ds^2_{\psi}(X,Y)= \inner{\psi(X)}{\psi(Y)} \end{align*} for $X,Y\in T_pM^m$ ($p\in M^m$), and we call them {\it the first, the second and the third fundamental forms}, respectively. They are all symmetric covariant tensors on $M^m$. \begin{definition}\label{def:front} A frontal bundle $(M^m,\mathcal{E},\inner{~}{~},D, \phi,\psi)$ is called a {\it front bundle\/} if \begin{equation}\label{eq:front} \operatorname{Ker}(\phi_p)\cap \operatorname{Ker}(\psi_p)=\{0\} \end{equation} holds for each $p\in M^m$. \end{definition} \begin{example}\label{ex:front-bundle} Let $\bigl(N^{m+1}(c),g\bigr)$ be an $(m+1)$-dimensional space form, that is, a complete Riemannian $(m+1)$-manifold of constant curvature $c$, and denote by $\nabla$ the Levi-Civita connection on $N^{m+1}(c)$. Let $f:M^m\to N^{m+1}(c)$ be a co-orientable frontal. Then there exists a globally defined unit normal vector field $\nu$. Since the coherent tangent bundle $\mathcal{E}_f$ given in Example~\ref{ex:front} is orthogonal to $\nu$, we can define a bundle homomorphism \[ \psi_f:T_pM^m\ni X\longmapsto \nabla_X\nu \in \mathcal{E}_p \qquad (p\in M^m). \] Then $(M^m,\mathcal{E}_f,\inner{~}{~},D,\phi_f,\psi_f)$ is a frontal bundle. Moreover, this is a front bundle in the sense of Definition~\ref{def:front} if and only if $f$ is a front, which is equivalent to $\operatorname{\mathit{I}}+\operatorname{\mathit{I\!I\!I}}$ being positive definite. As a fundamental theorem for hypersurfaces, the integrability condition for a given frontal bundle to be realized as a frontal in a space form is given in \cite{SUY6}. \end{example} We fix a front bundle $(M^m,\mathcal{E},\inner{~}{~},D,\phi,\psi)$ over an $m$-dimensional manifold $M^m$. \begin{definition}\label{def:ext} When $p\in M^m$ is not a singular point of $\phi$, we define \begin{equation}\label{eq:ext} K^{\operatorname{ext}}(X\wedge Y):= \frac{ \operatorname{\mathit{I\!I}}(X,X) \operatorname{\mathit{I\!I}}(Y,Y) -\operatorname{\mathit{I\!I}}(X,Y)^2 }{ I(X,X) I(Y,Y) -I(X,Y)^2 } \qquad (X,Y\in T_pM^m), \end{equation} which is called the {\em extrinsic curvature\/} at $p$ with respect to the $X\wedge Y$-plane in $T_pM^m$. \end{definition} If a front bundle $(M^m,\mathcal{E},\inner{~}{~},D,\phi,\psi)$ is induced from a front in $N^{m+1}(c)$, then it holds that \begin{equation}\label{eq:ext2} K^{\operatorname{ext}}(X\wedge Y)=K(X\wedge Y)+c \qquad (X,Y\in T_pM^m), \end{equation} where $K(X\wedge Y)$ is the sectional curvature at each $\phi$-regular point $p$ of $M^m$. As a generalization of \cite[Theorem 3.1]{SUY1}, relationships between singular principal curvatures and $K^{\operatorname{ext}}$ of wave fronts in space forms are investigated in \cite{SUY6}. \section{Four Gauss-Bonnet formulas on surfaces} \label{sec:gauss-bonnet} In this section, we give four Gauss-Bonnet formulas on a given front in $3$-dimensional space forms, and will point out several remarkable consequences of them. \subsection{The Gauss-Bonnet formulas for smooth maps} Let $M^m$ be an oriented $m$-manifold and $(N^m,g)$ an oriented Riemannian $m$-manifold. As in Example \ref{ex:map}, a $C^\infty$-map $f\colon{}M^m\to N^m$ induces a coherent tangent bundle $(M^m,\mathcal{E}_f,\inner{~}{~},D,\phi_f=df)$ over $M^m$. In this setting, an $A_k$-point (cf.~Definitions~\ref{def:a2-point} and \ref{def:a3-point}) coincides with an $A_{k}$-Morin singular point of $f$ (see \cite{SUY3}). Now we restrict our attention to the case $m=2$: An $A_2$-point (resp.\ $A_3$-point) of $\phi_f$ on $M^2$ is called a {\it fold\/} (resp. a {\it cusp}); namely, a fold (resp.\ a cusp) is right-left equivalent to the map \[ \boldsymbol{R}^2\ni (u,v)\longmapsto (u^2,v)\in \boldsymbol{R}^2 \quad \bigl(\mbox{resp.\ } \boldsymbol{R}^2\ni (u,v) \longmapsto (uv+v^3,u)\in \boldsymbol{R}^2 \bigr) \] at the origin. Here, two map germs $f_{i}:(\boldsymbol{R}^m,p)\to(\boldsymbol{R}^n,q)$ $(i=1,2)$ are right-left equivalent if there exist diffeomorphism germs $\xi_1:(\boldsymbol{R}^m,p)\to(\boldsymbol{R}^m,p)$ and $\xi_2:(\boldsymbol{R}^n,q)\to(\boldsymbol{R}^n,q)$ such that $\xi_2\circ f_1=f_2\circ\xi_1$ holds. The $\phi_f$-singular curvature $\kappa_f(p)$ of an $A_2$-point $p$ of $\phi_f$ (i.e.\ a fold) is called the {\it singular curvature\/} at a fold. The following assertion follows immediately. \begin{proposition}\label{prop:geod} Let $f:M^2\to N^2$ be a $C^\infty$-map and $p$ a fold singular point. Suppose that $\gamma(t)$ is a regular curve which parametrizes the singular set so that $f(M^2)$ lies on the left-hand side of $f\circ\gamma$. Then the singular curvature $\kappa_f(p)$ at $p$ is equal to the geodesic curvature of $f\circ \gamma$. \end{proposition} \begin{example}\label{ex:parabola} We set $f_\epsilon(u,v):=(u^2+\epsilon v^2,v)$ $(\epsilon:=\pm 1)$. If $\epsilon=1$ (resp. $\epsilon=-1$), then all singular points consist of folds with positive (resp.\ negative) singular curvature, see Fig.\ \ref{fig:pn-folds}. \end{example} \begin{figure}[h] \begin{center} \includegraphics[height=3cm]{pfolds.eps}\hspace{1cm} \includegraphics[height=3cm]{nfolds.eps} \end{center} \caption{% The images of $f_1$ and $f_{-1}$ (the folds are the left edges in both figures).} \label{fig:pn-folds} \end{figure} Now we recall the following two Gauss-Bonnet formulas on a given coherent tangent bundle: \begin{fact}[\cite{SUY1,SUY2}] \label{fact:suy} Let $(M^2,\mathcal{E},\inner{~}{~},D,\phi)$ be a coherent tangent bundle over a compact oriented $2$-manifold $M^2$, and suppose that the $\phi$-singular set $\Sigma_\phi$ consists of $A_2$-points and $A_3$-points. We denote by $K$ the Gaussian curvature of the induced metric $ds^2=\phi^*\langle\,,\,\rangle$. Then it holds that \begin{align} (\chi_{\mathcal{E}}^{}=)&\frac1{2\pi}\int_{M^2}K\, d\hat A_\phi= \chi(M^+_\phi)-\chi(M^-_\phi)+S^+_\phi-S^-_\phi, \label{eq:B}\\ 2\pi\chi(M^2)&= \int_{M^2}K\, dA_\phi+2\int_{\Sigma_\phi} \kappa_\phi\, d\tau_\phi, \label{eq:A} \end{align} where $d\tau_\phi$ is the length element on the $\phi$-singular set with respect to $ds^2_\phi$, and $S^+_\phi$ and $S^-_\phi$ are the numbers of positive and negative $A_3$-points of $\phi$, respectively {\rm(}see {\rm\cite[Figure 4]{SUY1}} and also {\rm\cite[Definition 2.9]{SUY2}}{\rm)}. On the other hand, $\chi_{\mathcal{E}}^{}$ is the Euler characteristic of the $\operatorname{SO}(2)$-vector bundle $\mathcal{E}$. \end{fact} \begin{remark}\label{rem:GB} Let $(\mathcal{E},\inner{~}{~})$ be an oriented vector bundle of rank $2$ with a metric over a compact oriented $2$-manifold $M^2$, and $D$ a metric connection. Suppose that there exists a bundle homomorphism $\phi:TM^2\to \mathcal{E}$. In the first section, $A_2$ and $A_3$-points are defined for such an arbitrary bundle homomorphism $\phi$ and {\it \eqref{eq:B} holds without assuming the compatibility condition \eqref{eq:c}}, as pointed out in \cite{SUY5}: For an oriented orthonormal frame field $\{\vect{e}_1,\vect{e}_2\}$ of $\mathcal{E}$ defined on $U\subset M^2$, there is a unique $1$-form $\omega$ on $U$ such that \begin{equation}\label{eq:conn-form} D_X\vect{e}_1=-\omega(X)\vect{e}_2,\qquad D_X\vect{e}_2=\omega(X)\vect{e}_1. \end{equation} Then $d\omega$ does not depend on the choice of $\{\vect{e}_1, \vect{e}_2\}$, and there is a $C^\infty$-function $K_{\phi,D}$ on $M^2\setminus \Sigma_\phi$ such that \begin{equation}\label{eq:K} d\omega=K_{\phi,D}\, d\hat A_{\phi}. \end{equation} We call $K_{\phi,D}$ the {\it Gaussian curvature\/} of $D$ with respect to $\phi$. If $\mathcal{E}$ is a coherent tangent bundle, the identity \eqref{eq:c} implies that $K_{\phi,D}$ coincides with the Gaussian curvature of the metric $ds^2_\phi=\phi^*\inner{~}{~}$. Thus formulas \eqref{eq:B} and \eqref{eq:A} still hold for $K=K_{\phi,D}$, without assuming \eqref{eq:c}. Although \eqref{eq:A} depends on $D$, \eqref{eq:B} is independent of the choice of metric connections. \end{remark} \begin{remark}\label{rem:peaks} Under the assumption that $\mathcal{E}$ is a coherent tangent bundle, the identities \eqref{eq:B} and \eqref{eq:A} hold even if $\phi$ admits wider class of singular points called `peaks'. (cf.\ \cite[Definition 1.10]{SUY1} and also \cite[Definition 2.1]{SUY2}). $A_3$-points, lips (see Definition~\ref{def:lips}), beaks (see Definition~\ref{def:beaks}), butterflies (see Definition~\ref{def:b-fly}) are all examples of peaks. For a coherent tangent bundle over a compact oriented $2$-manifold whose singular points are at most peaks, the formulas \eqref{eq:B} and \eqref{eq:A} hold (see \cite[Theorem 2.3]{SUY1} and \cite[Theorem~B]{SUY2}), here $S^+_{\phi}$ and $S^-_{\phi}$ in \eqref{eq:B} should be replaced by the numbers of positive (resp.\ negative) peaks (cf.\ \cite[Definition 2.1]{SUY2}). \end{remark} The formula \eqref{eq:B} in our situation induces Quine's formula. \begin{proposition}[Quine \cite{Q}] \label{prop:quine} Let $M^2$ and $N^2$ both be compact oriented connected $2$-manifolds, and $f:M^2\to N^2$ be a $C^\infty$-map whose singular set consists of folds and cusps. Then the topological degree of $f$ satisfies \[ \deg(f)\chi(N^2)=\chi(M^+_f)-\chi(M^-_f) +S^+_f-S^-_f, \] where $M^+_f$ {\rm(}resp.\ $M^-_f${\rm)} is the set of regular points at which $f$ preserves {\rm(}resp. reverses{\rm)} the orientation, and $S^+_f$ {\rm(}resp. $S^-_f${\rm)} is the number of positive cusps {\rm(}resp.\ the number of negative cusps{\rm)}. \end{proposition} On the other hand, the formula \eqref{eq:A} induces the following new formula. \begin{proposition}\label{prop:id} Let $(N^2,g)$ be an oriented Riemannian $2$-manifold, and $M^2$ a compact oriented $2$-manifold. Let $f:M^2\to N^2$ be a $C^\infty$-map whose singular set consists of folds and cusps. Then the total singular curvature $\int_{\Sigma} \kappa\, d\tau$ with respect to the length element $d\tau$ {\rm(}with respect to $g${\rm)} on the singular set $\Sigma$ is bounded, and satisfies the following identity \[ 2\pi\chi(M^2)= \int_{M^2}(\widetilde K\circ f) \,|f^*dA_g|+ 2\int_{\Sigma} \kappa\, d\tau, \] where $\widetilde{K}$ is the Gaussian curvature function on $(N^2,g)$, and $|f^*dA_g|$ is the pull-back of the Riemannian measure of $(N^2,g)$. \end{proposition} In particular, if $(N^2,g)$ is the Euclidean plane $(\boldsymbol{R}^2,g_0)$, we have the following classical result: \begin{corollary}[Levine \cite{L}]\label{cor:id} Let $M^2$ be a compact oriented $2$-manifold, and $f:M^2\to \boldsymbol{R}^2$ a $C^\infty$-map whose singular set consists of folds and cusps. Let $C_1,\dots,C_r$ be the disjoint union of the closed regular curves on $M^2$ such that the singular set of $f$ is $C_1\cup\dots \cup C_r$. Suppose that we give an orientation to each $f(C_j)$ {\rm(}$j=1,\dots,r${\rm)} so that the image $f(M^2)$ lies on the left-hand side of $f(C_j)$. Then the rotation indices $I(C_j)$ {\rm(}which takes values in the set of half-integers{\rm)} of $f(C_j)$ {\rm(}$j=1,\dots,r${\rm)} as fronts satisfy \[ \frac{\chi(M^2)}2=I(C_1)+\ldots +I(C_r). \] \end{corollary} \subsection{The Gauss-Bonnet formulas for wave fronts} We use the same notation as in the previous sections. Let $N^3(c)$ be a $3$-dimensional complete Riemannian manifold with constant curvature $c$. Here, we do not assume that $N^3(c)$ is simply connected, and we use the notation $\widetilde N^3(c)$ when it is in fact assumed to be simply connected, like in the previous section. Let $M^2$ be an oriented closed $2$-manifold and $f:M^2\to N^3(c)$ a front. Then the front bundle $(M^2,\mathcal{E}_f,\inner{~}{~},D,\phi=\phi_f,\psi=\psi_f)$ is defined as in Example~\ref{ex:front-bundle}. We fix a positively oriented local coordinate system $(U;u,v)$ on $M^2$. Denote by $\lambda$ (resp.\ $\lambda_\#$) the $\phi$-Jacobian function (resp.\ $\psi$-Jacobian function) as in \eqref{eq:jacobian}: \begin{equation}\label{eq:2-lambda} \lambda:=\lambda_{\phi}=\mu(\phi_u,\phi_v) \qquad\text{and}\qquad \lambda_\#:=\lambda_{\psi}=\mu(\psi_u,\psi_v), \end{equation} where \[ \phi_u=\phi(\partial /\partial u),\quad \phi_v=\phi(\partial /\partial v),\quad \psi_u=\psi(\partial /\partial u),\quad \psi_v=\psi(\partial /\partial v), \] and $\mu$ is the volume form of $\mathcal{E}_f$ giving co-orientation of $\mathcal{E}_f$. Then the $\phi$-singular set (resp.\ $\psi$-singular set) \[ \Sigma:=\Sigma_{\phi}\qquad (\text{resp.\ }\Sigma_\#:=\Sigma_{\psi}) \] is expressed as \[ \Sigma\cap U = \{p\in U\,;\,\lambda(p)=0\}\qquad (\text{resp.\ } \Sigma_\#\cap U = \{p\in U\,;\,\lambda_\#(p)=0\}). \] \begin{remark}\label{rmk:2} If $N^3(c)=\boldsymbol{R}^3$, $\Sigma_\#$ is the singular set of the Gauss map $\nu:M^2\to S^2$ of $f$. When $N^3(c)=S^3$, $\Sigma_\#$ is the singular set of the map induced by the unit normal vector field $\nu:M^2\to S^3$, where $S^3$ is the unit $3$-sphere. This unit normal vector corresponds to the dual map of $f$ induced by a duality between $S^3$ and itself. When $N^3(c)$ is the hyperbolic $3$-space $H^3$, $\Sigma_\#$ is the singular set of the map induced by the normal vector field $\nu\colon{}M^2\to S^3_1$ into the de Sitter $3$-space $S^3_1$. \end{remark} We write the volume forms as in \eqref{eq:volume-form} as \begin{equation}\label{eq:2-volume-forms} \begin{array}{llll} d\hat{A} &:=d\hat{A}_\phi=\lambda\, du\wedge dv,\qquad &d\hat{A}_\#&:=d\hat{A}_\psi=\lambda_\#\, du\wedge dv, \\ dA &:=dA_\phi=|\lambda|\, du\wedge dv,\qquad &dA_\# &:=dA_\psi=|\lambda_\# |\, du\wedge dv, \end{array} \end{equation} and \begin{align*} M^+ &:= M_{\phi}^+= \{p\in M^2\setminus \Sigma\,;\, d\hat{A}=dA \},\\ M^- &:= M_{\phi}^-= \{p\in M^2\setminus \Sigma\, ;\, d\hat{A}=-dA \}, \\ M^+_\#& := M_{\psi}^+= \{p\in M^2\setminus \Sigma_\#\, ;\, d\hat{A}_\#=dA_\#\},\\ M^-_\#& := M_{\psi}^-= \{p\in M^2\setminus \Sigma_\#\, ;\, d\hat{A}_\#=-dA_\# \}. \end{align*} Denote by $K$ (resp.\ $K_\#$) the Gaussian curvature of the metric $\operatorname{\mathit{I}}=ds^2_{\phi}$ (resp.\ $\operatorname{\mathit{I\!I\!I}}=ds^2_{\psi}$), which is defined on $M^2\setminus\Sigma$ (resp. $M^2\setminus\Sigma_\#$). The extrinsic curvature $K^{\operatorname{ext}}$ in Definition~\ref{def:ext} can be considered as a $C^{\infty}$-function on $M^2\setminus\Sigma$, because we are working with the $2$-dimensional case. Similarly, we denote by $K^{\operatorname{ext}}_{\#}$ the extrinsic curvature with respect to the second homomorphism $\psi$, namely, it is obtained by replacing $\operatorname{\mathit{I}}$ with $\operatorname{\mathit{I\!I\!I}}$ in \eqref{eq:ext}. If we denote by $\widehat\operatorname{\mathit{I}}$, $\widehat\operatorname{\mathit{I\!I}}$ and $\widehat\operatorname{\mathit{I\!I\!I}}$ the representation matrices of $\operatorname{\mathit{I}}$, $\operatorname{\mathit{I\!I}}$ and $\operatorname{\mathit{I\!I\!I}}$, respectively, with respect to $\{\partial/\partial u,\partial/\partial v\}$: \begin{equation}\label{eq:fundamental-matrices} \begin{aligned} \widehat\operatorname{\mathit{I}}&:= \begin{pmatrix} \inner{\phi_u}{\phi_u} & \inner{\phi_u}{\phi_v}\\ \inner{\phi_v}{\phi_u} & \inner{\phi_v}{\phi_v} \end{pmatrix},\\ \widehat\operatorname{\mathit{I\!I}} &:= -\begin{pmatrix} \inner{\phi_u}{\psi_u} & \inner{\phi_u}{\psi_v}\\ \inner{\phi_v}{\psi_u} & \inner{\phi_v}{\psi_v} \end{pmatrix},\\ \widehat\operatorname{\mathit{I\!I\!I}} &:= \begin{pmatrix} \inner{\psi_u}{\psi_u} & \inner{\psi_u}{\psi_v}\\ \inner{\psi_v}{\psi_u} & \inner{\psi_v}{\psi_v} \end{pmatrix}, \end{aligned} \end{equation} then the extrinsic curvatures are expressed as \begin{equation}\label{eq:k-ext} K^{\operatorname{ext}}=\frac{\det\widehat\operatorname{\mathit{I\!I}}}{\det\widehat\operatorname{\mathit{I}}},\qquad K_{\#}^{\operatorname{ext}}=\frac{\det\widehat\operatorname{\mathit{I\!I}}}{\det\widehat\operatorname{\mathit{I\!I\!I}}}. \end{equation} Then \eqref{eq:ext2} is equivalent to \begin{equation}\label{eq:2-gauss} K=c+K^{\operatorname{ext}}. \end{equation} \begin{lemma}\label{lem:unbdd} Let $f:M^2\to N^3(c)$ be a front. Then \begingroup \renewcommand{\theenumi}{{\rm(\alph{enumi})}} \renewcommand{\labelenumi}{{\rm(\alph{enumi})}} \begin{enum} \item\label{item:k:1} $K\,d\hat{A}=K_\#\,d\hat{A}_\#$, \item\label{item:k:1b} $K_\#=1$ if $c=0$, \item\label{item:k:2} $K^{\operatorname{ext}} \,d\hat{A}=d\hat{A}_\#$ and $K^{\operatorname{ext}}_\# \,d\hat{A}_\#=d\hat{A}$, \item\label{item:k:2b} $|K^{\operatorname{ext}}| \,dA=dA_\#$ and $|K^{\operatorname{ext}}_\#| \,d{A}_\#=d{A}$, \item\label{item:k:3} $K^{\operatorname{ext}}\,K^{\operatorname{ext}}_\#=1$. \end{enum} \endgroup \end{lemma} \begin{proof} Take a positive (local) orthonormal frame field $\{\vect{e}_1,\vect{e}_2\}$ of $\mathcal{E}_f$, and take a one-form $\omega$ as in \eqref{eq:conn-form}, that is, $\omega$ is the connection form of $D$ with respect to the frame $\{\vect{e}_1,\vect{e}_2\}$. By \eqref{eq:K}, $K\,d\hat{A}$ is determined by just the connection $D$. Hence we have \ref{item:k:1}. Take $2\times 2$-matrix-valued functions $G$ and $G_\#$ as \begin{equation}\label{eq:matrix-G} (\phi_u,\phi_v)=G(\vect{e}_1,\vect{e}_2),\qquad (\psi_u,\psi_v)=G_\#(\vect{e}_1,\vect{e}_2). \end{equation} By definition, $d\hat A$ and $d\hat A_\#$ in \eqref{eq:2-volume-forms} are expressed as \begin{equation}\label{eq:dA-G} \begin{aligned} d\hat A &= \lambda\,du\wedge dv=(\det G) \,du\wedge dv,\\ d\hat A_\# &=\lambda_\#\,du\wedge dv= (\det G_\#) \,du\wedge dv. \end{aligned} \end{equation} In particular, we have that \begin{equation}\label{eq:dA-G0} dA =|\det G| \,du\wedge dv,\qquad dA_\# =|\det G_\#| \,du\wedge dv. \end{equation} On the other hand, the matrices $\widehat\operatorname{\mathit{I}}$, $\widehat\operatorname{\mathit{I\!I}}$ and $\widehat\operatorname{\mathit{I\!I\!I}}$ as in \eqref{eq:fundamental-matrices} are written as \[ \widehat\operatorname{\mathit{I}} = \trans{G}G,\qquad \widehat\operatorname{\mathit{I\!I}} = -\trans{G}G_\#=-\trans{G_\#}G,\qquad \widehat\operatorname{\mathit{I\!I\!I}} = \trans{G_\#}G_\#, \] where $\trans(~)$ denotes transposition. Thus, by \eqref{eq:k-ext}, we have \begin{equation}\label{eq:K-G} K^{\operatorname{ext}} = \frac{\det G_\#}{\det G\hphantom{_\#}},\quad K_\#^{\operatorname{ext}} = \frac{\det G\hphantom{_\#}}{\det G_\#}. \end{equation} Hence \ref{item:k:2}, \ref{item:k:2b} and \ref{item:k:3} hold. Finally, if $c=0$, $\operatorname{\mathit{I\!I\!I}}$ is just the pull-back metric of canonical metric of the unit sphere by the Gauss map, so \ref{item:k:1b} follows. \end{proof} Now, we assume both of the singular sets $\Sigma$ and $\Sigma_\#$ consist of $A_2$- points and $A_3$-points. Then applying two abstract formulas \eqref{eq:B} and \eqref{eq:A} for $\phi$ and $\psi$ respectively, we have the following four Gauss-Bonnet formulas: \begin{align} \int_{M^2} K\, d\hat{A} &= \int_{M^+}K\, dA-\int_{M^-}K\, dA \label{eq:1p}\\ &= 2\pi\big(\chi(M^+)-\chi(M^-)\big) +2\pi \big(S^+-S^-\big), \nonumber \\ \int_{M^2} K\, dA &=2\pi\chi(M)-2\int_{\Sigma}\kappa\,d\tau, \label{eq:1m}\\ \int_{M^2} K_\#\, d\hat{A}_\# &= \int_{M_\#^+}K_\#\, dA_\#-\int_{M_\#^-}K_\#\, dA_\# \label{eq:2p}\\ &= 2\pi\big(\chi(M_\#^+)-\chi(M_\#^-)\big) +2\pi \big(S_\#^+-S_\#^-\big), \nonumber \\ \int_{M^2} K_\#\, dA_\#&=2\pi\chi(M^2)- 2\int_{\Sigma_\#}\kappa_\#\, d\tau_\#, \label{eq:2m} \end{align} where $\kappa$ (resp.\ $\kappa_\#$) is the singular curvature function along $A_2$-points in $\Sigma$ (resp.\ $\Sigma_\#$) as in \eqref{eq:singular-curvature-2}, $d\tau$ (resp.\ $d\tau_\#$) is the length element on the singular curve with respect to $\operatorname{\mathit{I}}$ (resp.\ $\operatorname{\mathit{I\!I\!I}}$), $S^+$ (resp.\ $S^-$) is the number of positive (resp.\ negative) $A_3$-points of $\phi$, and $S_\#^+$ (resp.\ $S_\#^-$) is the number of positive (resp.\ negative) $A_3$-points of $\psi$. \begin{remark}\label{rem:peak2} As seen in Remark~\ref{rem:peaks}, formulas \eqref{eq:1p}--\eqref{eq:2m} hold for fronts such that the singular sets of $\phi$ and $\psi$ consist of at most peak. Here, $S^+$ (resp.\ $S^-$) in \eqref{eq:1p} should be considered as a number of positive (resp.\ negative) peaks of $\phi$, and $S^+_\#$ (resp.\ $S^-_\#$) in \eqref{eq:2p} should be considered as a number of positive (resp.\ negative) peaks of $\psi$, in this case. \end{remark} \subsection{Applications of the four Gauss-Bonnet formulas} Let $f:M^2\to N^3(c)$ be an immersion of a compact orientable $2$-manifold $M^2$. Then the second homomorphism $\psi_f$ is the shape operator of $f$, and the set of singular points $\Sigma_\#$ of it coincides with the {\it inflection points\/} of $f$ (see \cite{SUY4}). Then $A_2$-points and $A_3$-points in $\Sigma_\#$ are called {\it $A_2$-inflection points\/} and {\it $A_3$-inflection points}, respectively. \begin{theorem}[A generalization of the Bleecker-Wilson formula] \label{thm:a} Let $M^2$ be a compact oriented $2$-manifold and $f:M^2\to N^3(c)$ an immersion. Suppose that the set of inflection points of $f$ consists of $A_2$-points and $A_3$-points. Then we have \begin{equation}\label{eq:BWnew} 2\chi(M^-_\#)=S_\#^+-S_\#^-. \end{equation} Moreover, $M^-_\#$ coincides with the set $\{p\in M^2\,;\,K^{\operatorname{ext}}(p)<0\}$. \end{theorem} \begin{proof} Since $A_2$-inflection points and $A_3$-inflection points are non-degenerate singular points of $\psi$, the set of inflection points $\Sigma_\#$ is a regular submanifold of a compact manifold $M^2$. Hence $\Sigma_\#$ is the disjoint union of a finite number of closed curves in $M^2$, and thus the Euler number $\chi(\Sigma_\#)$ vanishes. Then we have \begin{equation}\label{eq:euler-plus-minus} \chi(M^2) = \chi(M_\#^+)+\chi(M_\#^-)+\chi(\Sigma_\#) = \chi(M_\#^+)+\chi(M_\#^-). \end{equation} Since $f$ is an immersion, $dA=d\hat A$ holds, and then \begin{alignat*}{2} \chi(M^2) &= \frac{1}{2\pi}\int_{M^2} K\,d\hat A \qquad &&\text{(by the Gauss-Bonnet formula)}\\ &= \frac{1}{2\pi}\int_{M^2} K_{\#}\,d\hat A_{\#}\qquad &&\text{(by \ref{item:k:1} in Lemma~\ref{lem:unbdd})}\\ &= \chi(M_\#^+)-\chi(M_\#^-)+S_\#^+-S_\#^-\quad &&\text{(by \eqref{eq:2p})}\\ &= \chi(M^2)-2\chi(M_\#^-)+S_\#^+-S_\#^-\quad &&\text{(by \eqref{eq:euler-plus-minus})}. \end{alignat*} Thus, we have the equality. Here, since $dA=d\hat A$, \ref{item:k:2} and \ref{item:k:2b} of Lemma~\ref{lem:unbdd} implies that $\Sigma_\#=\{p\in M^2\,;\,K^{\operatorname{ext}}=0\}$. Then \begin{alignat*}{2} M_\#^- &= \{p\in M^2\setminus\Sigma_\#\,;\,d\hat A_\#=-dA_\#\} \qquad && \\ &= \{p\in M^2\setminus\Sigma_\#\,;\,K^{\operatorname{ext}}\,d\hat A=-dA_\#\}\qquad && \text{(by \ref{item:k:2} in Lemma~\ref{lem:unbdd})}\\ &= \{p\in M^2\setminus\Sigma_\#\,;\,K^{\operatorname{ext}}\,dA=-dA_\#\} \qquad && \text{(since $d\hat A=dA$)}\\ &= \{p\in M^2\setminus\Sigma_\#\,;\,K^{\operatorname{ext}}<0\} && \text{(by \ref{item:k:2b} in Lemma~\ref{lem:unbdd})}\\ &= \{p\in M^2\,;\,K^{\operatorname{ext}}<0\}. && \end{alignat*} Hence we have the conclusion. \end{proof} \begin{remark}\label{rmk:BW} When $N^3(c)=\boldsymbol{R}^3$, \eqref{eq:BWnew} is the classical formula given in \cite{BW}. In this case, $S_\#^++S_\#^-$ is equal to the total number of cusps appears in the Gauss map $\nu$ of $f$. Romero-Fuster \cite{R} discusses on this number for embedded surfaces. When $N^3(c)=S^3$ or $H^3$, the formula has been also proved in the authors' previous work \cite{SUY4}. However, this formula is a generalization of those results in \cite{BW} and \cite{SUY4}, since $N^3(c)$ might not be simply connected in our setting. It should be also remarked that the formula \eqref{eq:BWnew} can be generalized to any compact immersed surfaces in an arbitrary orientable Riemannian 3-manifold: In fact, \eqref{eq:c} for $\psi_f$ is equivalent to the Codazzi equation for the space form and does not hold for a general Riemannian $3$-manifold. However, \eqref{eq:2p} still holds without assuming \eqref{eq:c}, as mentioned in Remark \ref{rem:GB}, and the above proof works in this general setting (see \cite{SUY5} for details). \end{remark} Here, we give two examples satisfying $\chi(M^-)<0$. \begin{example} It is well-known that there are embedded triply periodic minimal surfaces. Although the Gauss maps of minimal surfaces only have isolated singular points, if we perturb the surface, we get an immersion $f:M^2\to T^3$, where $M^2$ is a compact $2$-manifold with positive genus. Since minimal surfaces have negative Gaussian curvature, the perturbation $f$ satisfies $\chi(M^-)<0$, and the Gauss map of $f$ must have cusps. \end{example} \begin{example}\label{ex:trinoid} Similarly, considering the Jorge-Meeks symmetric trinoid with the three ends rounding off to become closed discs, one can get a closed surface in $\boldsymbol{R}^3$ as in Fig.~\ref{fig:trioid}. Then the resulting surface satisfies $\chi(M^-)<0$, and its Gauss map must have cusps. \end{example} \begin{figure}[htb] \begin{center} \vspace*{2mm} \includegraphics[width=3.3cm]{trinoid-2.eps} \vspace*{2mm} \caption{Trinoid with $\chi(M^-)<0$ in Example~\ref{ex:trinoid}.} \label{fig:trioid} \end{center} \end{figure} Interchanging the roles of $\phi$ and $\psi$, we get the following dual assertion in $\boldsymbol{R}^3$. Let $f\colon{}M^2\to \boldsymbol{R}^3$ be a front with unit normal vector $\nu$. Then for each real number $t$, $f_t:=f+t\nu$ is a front with unit normal $\nu$, which is called the {\em parallel front\/} of signed distance $t$ of the front $f$. \begin{theorem}[The dual version of Theorem~\ref{thm:a}]\label{thm:b} Let $f:S^2\to \boldsymbol{R}^3$ be a strictly convex surface and $f_t\,\,(t\in\boldsymbol{R})$ a parallel front of $f$, Assume that the set of singular points of $f_t$ consists of cuspidal edges and swallowtails. Then $2\chi(M^-_{f_t})=S^+_{f_t}-S^-_{f_t}$ holds, where $S^+_{f_t}$ {\rm(}resp. $S^-_{f_t}${\rm)} is the number of positive {\rm(}resp. negative{\rm)} swallowtails of $f_t$, and $M^{-}_{f_t}$ is the subset of $M^2$ where the Gaussian curvature $K_t$ of $f_t$ is negative. \end{theorem} \begin{proof} Since $f$ is strictly convex, $\nu\colon{}S^2\to S^2$ is a diffeomorphism, and then $\Sigma_\#$ vanishes. Then, exchanging the roles of $\phi$ and $\psi$, the proof of Theorem~\ref{thm:a} implies the conclusion. \end{proof} \begin{example}\label{ex:ellipsoid} Let $f\colon S^2 \to\boldsymbol{R}^3$ be a parametrization of the ellipsoid defined by $(x^2/5)^2+(y/4)^2+z^2=1$ and let $\nu$ be its unit normal vector field. Then we can observe that $f_c:=f+c\nu$ for $c=11/2$ has four negative swallowtails and the Euler number of the set $\{K_c<0\}$ is $-2$, where $K_c$ is the Gaussian curvature of $f_c$, see Fig.~\ref{fig:ellipsoid}. \end{example} \begin{figure}[htb] \begin{center} \includegraphics[height=4cm]{ellisppara.eps}\hspace{1cm} \caption{A parallel surface of an ellipsoid with $\chi(M^-)<0$ in Example~\ref{ex:ellipsoid}.} \label{fig:ellipsoid} \end{center} \end{figure} On the other hand, as a conclusion of \eqref{eq:2m}, we have: \begin{theorem}\label{thm:c} Let $f:M^2\to \boldsymbol{R}^3$ be an immersion and $\nu:M^2\to S^2$ the Gauss map of $f$. Assume that the singular set $\Sigma_\#$ of $\nu$ consists of folds {\rm(}i.e. $A_2$-points{\rm)} and cusps {\rm(}i.e. $A_3$-points{\rm)}. Then \[ \int_{\Sigma_{\#}}\kappa_\#\,d\tau_{\#} = \int_{M_\#^-} K\,d\hat{A} = \int_{M^2}K^-\,dA \] holds, where $K^-=\min(0,K)$. In particular, the total dual singular curvature {\rm(}with respect to $\nu${\rm)} $\int_{\Sigma_{\#}}\kappa_\#\,d\tau_\#$ is non-positive. \end{theorem} \begin{proof} Since $f$ is an immersion into $\boldsymbol{R}^3$ (i.e. $c=0$), $M^-_{\#}=\{p\in M^2\,;\,K^{\operatorname{ext}}<0\}=\{p\in M^2\,;\,K<0\}$ holds. Then we have \begin{equation}\label{eq:negint} \int_{M_\#^-} K\,dA = \int_{M^2} K^-\,dA. \end{equation} Here, \begin{alignat*}{2} \int_{M^2}K_{\#}\,dA_\#&= \int_{M_\#^+}K_\#\,d\hat A_\# - \int_{M_\#^-}K_\#\,d\hat A_\# &&\\ &=\int_{M_\#^+}K\,d\hat A - \int_{M_\#^-}K\,d\hat A \qquad &&\text{(by \ref{item:k:1} of Lemma~\ref{lem:unbdd})}\\ &=\int_{M_\#^+}K\,dA - \int_{M_\#^-}K\,dA \qquad &&\text{(since $f$ is an immersion)}\\ &=\int_{M^2}K\,dA - 2\int_{M_\#^-}K\,dA && \\ &= 2\pi\chi(M^2) -2 \int_{M^2} K^-\,dA. && \text{(by \eqref{eq:negint})}. \end{alignat*} Thus, by \eqref{eq:2m}, we have the conclusion. \end{proof} \begin{example}\label{ex:sin} Consider a rotation of the sine curve \[ f(u,v):=(u \cos v, u \sin v, \cos u) \qquad (0< u<\pi,~0\le v\le 2\pi) \] whose Gauss map has $A_2$-singular points of positive singular curvature. This shows that the singular curvature of Gauss maps can take positive values. \end{example} \begin{corollary}\label{cor:c1} Under the same assumptions as in Theorem~\ref{thm:c}, if there exists a point $p$ satisfying $K(p)<0$ then there exists a fold point such that $\kappa_\#$ is negative. \end{corollary} \begin{example} An embedded Delaunay surface (an unduloid) as a rotationally symmetric periodic surface in $\boldsymbol{R}^3$ can be considered as an immersion into a flat space $f:S^1\times S^1 \to \boldsymbol{R}^2\times S^1$ which has an annular domain having negative Gaussian curvature. As a consequence, the image of its Gauss map of $f$ has folds of negative singular curvature. \end{example} \begin{remark}\label{cor:c2} Let $f:M^2\to \boldsymbol{R}^3$ be an immersion of a compact 2-manifold $M^2$. Then the inequality \[ \frac{1}{2\pi}\int_{M^2}|K|\, dA\ge \beta_0+\beta_1+\beta_2 \] is called the Chern-Lashof inequality, where $\beta_j$ $(j=0,1,2)$ is the $j$-th Betti number of $M^2$. The equality holds if and only if $f$ is tightly immersed in $\boldsymbol{R}^3$. On the other hand, by the Gauss-Bonnet formula, it holds that \[ \frac{1}{2\pi}\int_{M^2}K\, dA = \beta_0-\beta_1+\beta_2. \] Hence, we have that \[ -\int_{\Sigma_\#}\kappa_\#\,d\tau_\#= -\int_{M^2}K^-\,dA\ge 2\pi\beta_1, \] where the equality holds if and only if $f$ is a tight immersion. \end{remark} Similar to Theorem~\ref{thm:b}, the dual version of Theorem~\ref{thm:c} holds: \begin{theorem}\label{thm:d} Let $f:S^2\to \boldsymbol{R}^3$ be a strictly convex immersion and $f_t$ a parallel front of $f$ for $t\in\boldsymbol{R}$. Assume that the set of singular points $\Sigma_{f_t}$ of $f_t$ consists of cuspidal edges and swallowtails. Then \[ \int_{\Sigma_{f_t}}\kappa_{f_t}\,d\tau_{f_t} = \int_{M^{-}_{f_t}}K_\#\,d\hat{A}_\# = \int_{M^{-}_{f_t}}K_{t}\,d\hat{A}_{f_t} \] holds, where $K_t$ is the Gaussian curvature of $f_t$, $M^-_{f_t}\subset M^2$ is the set where $K_t<0$, and $\kappa_{f_t}$ is the singular curvature of cuspidal edges of $f_t$. \end{theorem} \begin{proof} The equality $\int_{\Sigma_{f_t}}\kappa_{f_t}\,d\tau_{f_t} = \int_{M^{-}_{f_t}}K_\#\,d\hat{A}_\#$ follows from the proof of Theorem \ref{thm:c} by exchanging the roles of $\phi$ and $\psi$. On the other hand, in the proof of Theorem~\ref{thm:a}, we proved that $M^{-}_{\#}$ coincides with the set $\{p\in M^2\,;\,K^{\operatorname{ext}}(p)<0\}$ because of $\phi_f$ has no singular points. In our situation, since $\nu\colon{}S^2\to S^2$ is a diffeomorphism, $\psi_f$ has no singular points. Thus, by exchanging the roles of $\phi$ and $\psi$, we have that \[ M^{-}_{f_t}=\{p\in M^2\,;\,(K^{\operatorname{ext}}_t)_\#(p)<0\} =\{p\in M^2\,;\,K^{\operatorname{ext}}_t(p)<0\}, \] because of \ref{item:k:3} in Lemma \ref{lem:unbdd}. Moreover, by \eqref{eq:2-gauss}, we have $K^{\operatorname{ext}}_t=K_t$ and get the assertion. \end{proof} \begin{corollary}\label{cor:d1} Under the same assumptions as in Theorem \ref{thm:d}, if there exists a point $p$ satisfying $K_{f_t}(p)<0$ then there exists a point such that $\kappa_{f_t}$ is negative. \end{corollary} \subsection{The case of bounded Gaussian curvature} From now on, we assume that the Gaussian curvature is bounded. As shown below, there are many such surfaces as wave fronts. The following lemma holds. \begin{lemma}\label{lem:bdd} Let $M^2$ be an oriented $2$-manifold, $f:M^2\to N^3(c)$ a front, and $p$ an $A_2$-singular point of $f$. Let $(M^2,\mathcal{E}_f,\inner{~}{~},D,\phi,\psi)$ be a front bundle associated to $f$ {\rm(}cf.\ Example~\ref{ex:front-bundle}\/{\rm)}. Suppose that there exists a neighborhood $U$ of a $(\phi$-$)$singular point $p$ such that $\log |K^{\operatorname{ext}}|$ is bounded on $U\setminus\Sigma$, where $K^{\operatorname{ext}}$ is the extrinsic Gaussian curvature and $\Sigma$ is the set of singular points of $f$. Then the following holds on $U${\rm:} \begingroup \renewcommand{\theenumi}{{\rm(\alph{enumi})}} \renewcommand{\labelenumi}{{\rm(\alph{enumi})}} \begin{enum} \setcounter{enumi}{4} \item\label{item:k:4} $\Sigma=\Sigma_\#$. \item\label{item:k:5} By setting $\epsilon:=\operatorname{sgn}(K^{\operatorname{ext}}|_U)$, it holds that $M^+=M_\#^\epsilon$ and $\kappa\,d\tau=\epsilon\kappa_\#\,d\tau_\#$, where $K^{\operatorname{ext}}|_U$ is the restriction of the function $K^{\operatorname{ext}}$ on $U$. \end{enum} \endgroup \end{lemma} \begin{proof} We fix a positive orthonormal frame field $\{\vect{e}_1,\vect{e}_2\}$ of $\mathcal{E}_f$ and take matrices $G$ and $G_\#$ as in \eqref{eq:matrix-G}. Since $K^{\operatorname{ext}}$ is bounded on $U$, $\det G_\#=0$ holds on $\Sigma\cap U=\{\det G=0\}$ because of \eqref{eq:K-G}. This implies $\Sigma\subset \Sigma_\#$. On the other hand, by \ref{item:k:3} of Lemma~\ref{lem:unbdd} and the assumptions, $K_\#^{\operatorname{ext}}=1/K^{\operatorname{ext}}$ is also bounded. Then exchanging the roles of $\phi$ and $\psi$, we have $\Sigma\supset \Sigma_\#$. Hence \ref{item:k:4} holds. We shall prove \ref{item:k:5}. Parametrize the singular set $\Sigma=\Sigma_\#$ by a regular curve $\gamma(t)$ on $U$. Since $p$ is an $A_2$-point, $\phi\bigl(\dot\gamma(t)\bigr)$ does not vanish, and one can take a $C^{\infty}$-function $\theta=\theta(t)$ such that \begin{equation}\label{eq:phi-dot} \phi\bigl(\dot\gamma(t)\bigr) = \left|\phi\bigl(\dot\gamma(t)\bigr)\right| \bigl( \cos\theta(t)\vect{e}_1 +\sin \theta(t)\vect{e}_2 \bigr) \end{equation} where $\dot{~}=d/dt$ and $\vect{e}_j=\vect{e}_j\bigl(\gamma(t)\bigr)$ ($j=1,2$). Then the conormal vector $\vect{n}(t)$ along $\gamma(t)$ is expressed as \[ \vect{n}(t)= -\sin\theta(t)\vect{e}_1 + \cos\theta(t) \vect{e}_2. \] Let $\omega$ be the connection form of $D$ with respect to the frame $\{\vect{e}_1,\vect{e}_2\}$ as in \eqref{eq:conn-form}. Then we have \[ D_{d/dt}\vect{n}(t) = \left(\omega\bigl(\dot\gamma(t)\bigr)-\dot\theta(t)\right) \frac{\phi\bigl(\dot\gamma(t)\bigr)}{% \left|\phi\bigl(\dot\gamma(t)\bigr)\right|}. \] Substituting this into \eqref{eq:singular-curvature-2}, we have \[ \kappa(t)= -\frac{\operatorname{sgn}\left(d\lambda\bigl(\eta(t)\bigr)\right)}{% \left|\phi\bigl(\dot\gamma(t)\bigr)\right|} \left(\omega\bigl(\dot\gamma(t)\bigr)-\dot\theta(t)\right), \] where $\eta(t)$ is a null vector field such that $\{\dot\gamma(t),\eta(t)\}$ is positively oriented. Then we have \begin{equation}\label{eq:bounded-curv} \kappa(t)\,d\tau = \kappa(t)\,\left|\phi\bigl(\dot\gamma(t)\bigr)\right|\,dt = -\operatorname{sgn}\left(d\lambda\bigl(\eta(t)\bigr)\right) \left(\omega\bigl(\dot\gamma(t)\bigr)-\dot\theta(t)\right)\,dt. \end{equation} Similarly, if we set \begin{equation}\label{eq:psi-dot} \psi\bigl(\dot\gamma(t)\bigr) = \left|\psi\bigl(\dot\gamma(t)\bigr)\right| \bigl( \cos\theta_\#(t)\vect{e}_1 +\sin \theta_\#(t)\vect{e}_2 \bigr), \end{equation} we have \begin{equation}\label{eq:bounded-curv-dual} \kappa_{\#}(t)\,d\tau_\#= -\operatorname{sgn}\left(d\lambda_\#\bigl(\eta_\#(t)\bigr)\right) \left(\omega\bigl(\dot\gamma(t)\bigr)-\dot\theta_\#(t)\right)\,dt, \end{equation} where $\eta_\#(t)$ is a null vector field with respect to $\psi$ such that $\{\dot\gamma(t),\eta_\#(t)\}$ is positively oriented. Both of the frame fields $\{\dot\gamma(t),\eta(t)\}$ and $\{\dot\gamma(t),\eta_\#(t)\}$ are compatible to the orientation of $M^2$. Since $\lambda=0$ on $\gamma(t)$, the relations \eqref{eq:dA-G} and \eqref{eq:K-G} yield \[ \operatorname{sgn} d\lambda\bigl(\eta(t)\bigr) = \operatorname{sgn} d\lambda\bigl(\eta_\#(t)\bigr) = \left(\operatorname{sgn}(K^{\operatorname{ext}}|_U)\right) \left(\operatorname{sgn} d\lambda_\#\bigl(\eta_\#(t)\bigr)\right). \] Since $\epsilon=\operatorname{sgn}(K^{\operatorname{ext}}|_U)$, we have \begin{equation}\label{eq:sign} \operatorname{sgn} d\lambda\bigl(\eta(t)\bigr) = \epsilon\, d\lambda_\#\bigl(\eta_\#(t)\bigr). \end{equation} Finally, by \cite[Theorem 3.1]{SUY1}, the second fundamental form $\operatorname{\mathit{I\!I}}$ vanishes on $\gamma(t)$. Thus by \eqref{eq:phi-dot} and \eqref{eq:psi-dot}, we have \begin{align*} 0 &= -\operatorname{\mathit{I\!I}}\bigl(\dot\gamma(t),\dot\gamma(t)) = -\inner{\phi\bigl(\dot\gamma(t)\bigr)}{% \psi\bigl(\dot\gamma(t)\bigr)}\\ &= -\left|\phi(\dot\gamma)\right| \left|\psi(\dot\gamma)\right| \cos\bigl(\theta(t)-\theta_\#(t)\bigr), \end{align*} and thus $\theta(t)-\theta_{\#}(t)$ is constant. Hence \begin{equation}\label{eq:dot-theta} \dot\theta(t) = \dot\theta_{\#}(t). \end{equation} Summing up, by \eqref{eq:bounded-curv}, \eqref{eq:bounded-curv-dual}, \eqref{eq:sign} and \eqref{eq:dot-theta}, we have \ref{item:k:5}. \end{proof} \begin{corollary}\label{thm:add} Let $M^2$ be a compact oriented $2$-manifold, and $f:M^2\to \boldsymbol{R}^3$ a front such that $K>0$ and $\log|K|$ is bounded on $M^2\setminus \Sigma$. Then the singular curvature at the $A_2$-points of the Gauss map of $f$ is always negative. \end{corollary} \begin{example}\label{ex:p} We set \[ f(u,v):=\bigl( (2-\cos u)\cos v,(2-\cos u)\sin v ,u-\sin u\bigr). \] which is a rotation of a cycloid. By a straightforward calculation, $K$ is bounded. In particular, the Gauss map of $f$ has a fold of negative singular curvature. \end{example} \begin{theorem}\label{thm:e} Let $M^2$ be a compact oriented $2$-manifold, and $f:M^2\to N^3(c)$ a front such that $\log|K^{\operatorname{ext}}|$ is bounded on $M^2\setminus\Sigma$. Suppose that the singular sets $\Sigma$ of $\phi$ and $\Sigma_\#$ ($=\Sigma$) of $\psi$ consist of $A_2$-points and $A_3$-points. Then \[ S^+ - S^- = \operatorname{sgn}(K^{\operatorname{ext}})(S_\#^+-S_\#^-) \] holds, where $S^+$ {\rm(}resp.\ $S^-${\rm)} is the number of positive {\rm(}resp.\ negative{\rm)} swallowtails of $f$, and $S_\#^+$ {\rm(}resp.\ $S_\#^+${\rm)} is the number of positive {\rm(}resp.\ negative{\rm)} $A_3$-point of the map $\nu$ induced from the unit normal vector field of $f$ as in Remark \ref{rmk:2}. \end{theorem} \begin{proof} Combining \ref{item:k:1} of Lemma~\ref{lem:unbdd}, \eqref{eq:1p}, \eqref{eq:2p} and \ref{item:k:5} of Lemma~\ref{lem:bdd}, we have the conclusion. \end{proof} \begin{example}\label{ex:CMC} Let $f\colon{}M^2\to \boldsymbol{R}^3$ be an immersion of constant mean curvature $1/2$ and $\nu:M^2\to S^2(\subset \boldsymbol{R}^3)$ its Gauss map. Then the parallel surface $f_1=f +\nu$ is a wave front with constant Gaussian curvature $1$. Since the unit normal vector field of $f_1$ is also $\nu$, the singular set $\Sigma_\#$ of $\nu$ coincides with the set $\{K_f=0\}$, where $K_f$ is the Gaussian curvature of $f$. Then by Lemma~\ref{lem:bdd}, $\Sigma=\Sigma_\#=\{K_f=0\}$. It is well-known that there are many constant mean curvature immersed tori in $\boldsymbol{R}^3$. On the other hand, Gro\ss{}e-Brauckmann \cite{GB} constructed triply periodic surfaces in $\boldsymbol{R}^3$ with constant mean curvature $1/2$. Such surfaces are compact surfaces in a $3$-dimensional flat torus, and then Theorem~\ref{thm:e} can be applied for $f +\nu$ of them. \end{example} Moreover, for fronts of negative extrinsic curvature, we have the following: \begin{theorem}\label{thm:g} Let $M^2$ be a compact oriented $2$-manifold, and $f:M^2\to N^3(c)$ a front such that $K^{\operatorname{ext}}<0$ and $\log|K^{\operatorname{ext}}|$ is bounded on $M^2$. Then $\chi(M^2)=0$ holds. \end{theorem} \begin{proof} By Lemma~\ref{lem:bdd}, we have $\Sigma=\Sigma_\#$ and $\kappa\, d\tau=-\kappa_{\#}\,d\tau_\#$. Then by \eqref{eq:1m} and \eqref{eq:2m}, we have \begin{align*} \int_{M^2}\,K\,dA &= 2\pi \chi(M^2) - 2 \int_{\Sigma}\kappa\,d\tau = 2\pi\chi(M^2) + 2\int_{\Sigma}\kappa_{\#}\,d\tau_\# \\ &= 2\pi\chi(M^2) + \left(2\pi\chi(M^2)- \int_{M^2}K_{\#}\,dA_{\#}\right). \end{align*} On the other hand, by \ref{item:k:5} in Lemma~\ref{lem:bdd} and \ref{item:k:1} in Lemma~\ref{lem:unbdd}, we have \begin{align*} \int_{M^2}\!\!K_{\#}\,dA_\# &= \int_{M_\#^+}\!\!K_{\#}\,d\hat A_{\#} - \int_{M_\#^-}\!\!K_{\#}\,d\hat A_{\#} =\int_{M^-}\!\!K_{\#}\,d\hat A_{\#} - \int_{M^+}\!\!K_{\#}\,d\hat A_{\#} \\ &= \int_{M^-}\!\!K\,d\hat A - \int_{M^+}\!\!K\,d\hat A = -\int_{M^2}\!\!K\,dA. \end{align*} Combining these, we have the conclusion. \end{proof} \begin{example}\label{ex:cyc} It is well-known that there is a front of constant Gaussian curvature $-1$ which is diffeomorphic to a torus, that is, Euler number vanishes. Moreover, we can construct an example of a compact front of non-constant negative Gaussian curvature as follows: We set \[ f(u,v):=\bigl((2+\cos u)\cos v ,(2+\cos u)\sin v ,u-\sin u\bigr), \] which is another rotation of a cycloid (cf.\ Example \ref{ex:p}). One can easily check that $K$ is bounded. Since $f$ is periodic, it can be considered as a torus in a flat space form $\boldsymbol{R}^2\times S^1$. \end{example} \begin{example} Let $f\colon{}M^2\to S^3$ be a flat front in the unit $3$-sphere, that is, $f$ has vanishing Gaussian curvature on its regular points. Then the unit normal vector field $\nu:M^2\to S^3$ also gives a flat fronts. If $M^2$ is compact, $f$ is weakly complete in the sense of \cite{KU} and $M^2$ is orientable (see \cite[Section 1]{KU}). On the other hand by Theorem~\ref{thm:g}, the Euler characteristic of $M^2$ vanishes because $K^{\operatorname{ext}}=-1$ for flat surfaces in $S^3$. Thus $M^2$ is diffeomorphic to a torus. Moreover the singular set of $f$ coincides with that of $\nu$, and Theorem~\ref{thm:e} yields that $S^+-S^- = -(S^+_\#-S^{-}_\#)$ holds. This identity is not trivial since the set of swallow tails of $f$ is not equal to that of $\nu$. (In fact, $p$ is an $A_3$-point, the null direction should be tangential direction of the singular set $\Sigma$. On the other hand, null directions of $\phi$ and $\psi$ are linearly independent because $f$ is a front.) Several other properties of flat tori in $S^3$ as fronts are discussed in \cite{KU}. \end{example} \section*{Acknowledgement} The authors thank Wayne Rossman for careful reading of the first draft for giving valuable comments. They also thank the referee for valuable comments.

https://arxiv.org/abs/2001.04781

On generalized Holmgren's principle to the Lamé operator with applications to inverse elastic problems

Consider the Lamé operator $\mathcal{L}(\mathbf{ u} ) :=\mu \Delta \mathbf{u}+(\lambda+\mu) \nabla(\nabla \cdot \mathbf{ u} )$ that arises in the theory of linear elasticity. This paper studies the geometric properties of the (generalized) Lamé eigenfunction $\mathbf{u}$, namely $-\mathcal{L}(\mathbf{ u} )=\kappa \mathbf{ u}$ with $\kappa\in\mathbb{R}_+$ and $\mathbf{ u}\in L^2(\Omega)^2$, $\Omega\subset\mathbb{R}^2$. We introduce the so-called homogeneous line segments of $\mathbf{u}$ in $\Omega$, on which $\mathbf{u}$, its traction or their combination via an impedance parameter is vanishing. We give a comprehensive study on characterizing the presence of one or two such line segments and its implication to the uniqueness of $\mathbf{u}$. The results can be regarded as generalizing the classical Holmgren's uniqueness principle for the Lamé operator in two aspects. We establish the results by analyzing the development of analytic microlocal singularities of $\mathbf{u}$ with the presence of the aforesaid line segments. Finally, we apply the results to the inverse elastic problems in establishing two novel unique identifiability results. It is shown that a generalized impedance obstacle as well as its boundary impedance can be determined by using at most four far-field patterns. Unique determination by a minimal number of far-field patterns is a longstanding problem in inverse elastic scattering theory.

\section{Introduction} Consider the following partial differential operator (PDO) acting on a $\mathbb{C}^2$-valued function $\mathbf{u}=({u}_\ell({\mathbf x}))_{\ell=1}^2$, $\mathbf{x}=(x_\ell)_{\ell=1}^2\in\mathbb{R}^2$: \begin{equation}\label{eq:pdo1} \mathcal{L}(\bmf u) :=\mu \Delta \bmf u+(\lambda+\mu) \nabla(\nabla \cdot \bmf u ). \end{equation} $\mathcal{L}$ is known as the Lam\'e operator that arises in the theory of linear elasticity. $\lambda,\mu$ are the Lam\'e constants satisfying the following strong convexity condition \begin{equation}\label{eq:convex} \mu>0 \text { and } \lambda+\mu>0. \end{equation} Let $\Omega\subset\mathbb{R}^2$ be an open set. $\bmf u=(u_\ell)_{\ell=1}^2\in L^2(\Omega)^2$ is said to be a (generalised) Lam\'e eigenfunction if \begin{equation}\label{eq:lame} -\mathcal{L} (\bmf{u})=\kappa \bmf{u}\ \text { in }\ \Omega,\ \ \kappa \in\mathbb{R}_+. \end{equation} It is noted that there is no boundary condition prescribed on $\partial\Omega$ for $\bmf{u}$. \eqref{eq:lame} arises in the study of the time-harmonic wave scattering. Let $\Gamma_h\Subset\Omega$ be an open connected line segment, where $h\in\mathbb{R}_+$ signifies the length of the line segment. Let \begin{equation}\label{eq:nutau} \bmf{\nu }=(\nu_1,\nu_2)^\top \mbox{ and } \boldsymbol{\tau}=(-\nu_2,\nu_1)^\top \end{equation} respectively, signify the unit normal and tangential vectors to $\Gamma_h$. The traction $T_\nu\bmf{u}$ on $\Gamma_h$ is defined by \begin{equation}\label{eq:Tu} T_{\bmf{\nu }}\bmf{u}=2\mu\partial_{\bmf{\nu }}\bmf{u}+\lambda\bmf{\nu} \left(\nabla \cdot \bmf{u}\right)+\mu\boldsymbol{\tau}(\partial_{2}u_{1}-\partial_{1}u_{2}), \end{equation} where $$ \nabla\bmf u:=\begin{bmatrix} \partial_1 u_1 & \partial_2 u_1 \cr \partial_1 u_2 & \partial_2 u_2 \end{bmatrix}, $$ $\partial_\nu\bmf u:=\nabla\bmf u \cdot \nu$, $\partial_j u_i:=\partial u_i/\partial x_j$. It is noted that $\bmf{u}|_{\Gamma_h}$ and $T_\nu\bmf{u}|_{\Gamma_h}$ consist of the Cauchy data on $\Gamma_h$ to $\bmf{u}$ in \eqref{eq:lame} . We recall the classical Holmgren's theorem for an elliptic PDO $\mathcal{P}$ with real-analytic coefficients (cf. \cite{TF}). If $\mathcal{P}\bmf{u}$ is real analytic in a connected open neighbourhood of $\Omega$, then $\bmf{u}$ is also real-analytic. The Holmgren's theorem applied to $\bmf{u}$ in \eqref{eq:lame}, we immediately see that $\bmf{u}$ is real-analytic in $\Omega$. Suppose that \begin{equation}\label{eq:cond1l} \bmf{u}=\mathbf{0}\quad\mbox{and}\quad T_\nu\bmf{u}=\mathbf{0}\quad\mbox{on}\ \ \Gamma_h, \end{equation} then by the Cauchy-Kowalevski theorem, one readily has that $\bmf{u}\equiv 0$ in $\Omega$. This is known as the Holmgren's uniqueness principle. It also holds when $\Gamma_h$ is replaced to be an analytic curve. In this paper, we shall generalize the Holmgren's principle with the Cauchy data on a line segment to the Lam\'e operator $\mathcal{L}+\kappa$ in two aspects. First, we note that in \eqref{eq:cond1}, both Cauchy data are required to vanish on the line segment $\Gamma_h$. We ask whether this is the minimal/optimal requirement to ensure the uniqueness of $\bmf{u}$. Can the Holmgren's principle still hold, say if one only requires that \[ \bmf{u}(\bmf{x}_0)=\mathbf{0}\quad\mbox{and}\quad T_\nu \bmf{u}|_{\Gamma_h}=\mathbf{0}, \] where $\bmf{x}_0\in\Gamma_h$ is a single point? Clearly, in general, this cannot be true for a generic PDO. However, it is one of the interesting discoveries of the present paper that one of the two homogeneous conditions in \eqref{eq:cond1l} can indeed be replaced by a certain point-value condition. Second, we view \eqref{eq:cond1l} as the existence of two line segments $\Gamma_h^\pm$ such that: (i) $\bmf{u}|_{\Gamma_h^-}=\bmf{0}$ and $T_\nu\bmf{u}|_{\Gamma_h^+}=\mathbf{0}$; (ii) $\angle (\Gamma_h^-, \Gamma_h^+)=\pi$. Hence, a natural generalization is to consider the case that the two line segments are not of a straight intersection, namely, $\angle(\Gamma_h^-, \Gamma_h^+)\neq \pi$. In such a case, we can also establish a certain uniqueness principle for the solution to \eqref{eq:lame}. It is interesting to point out that for the latter generalization, the Cauchy data of $\bmf{u}$ are no longer prescribed on an analytic curve. Furthermore, we would like to emphasize that for both cases mentioned above, we also include the more general Robin-type condition into our study, namely $\big(\bmf{u}+\eta T_\nu\bmf{u}\big)|_{\Gamma_h}=\mathbf{0}$, which is known as an impedance condition with $\eta$ called an impedance parameter. We refer to the above discoveries as the generalized Holmgren's principle to the Lam\'e operator. The implication of the generalized principle to the uniqueness of a solution to the elastic problem \eqref{eq:lame} is obvious. Moreover, our study is clearly related to the geometric structures of the (generalized) Lam\'e eigenfunctions in \eqref{eq:lame}, which is also a central topic in the spectral theory of PDOs; see \cite{CDLZ,CDLZ2} and the references therein for more related discussions. According to our discussion above, the results obtained are clearly of independent interest for their own sake in the PDE theory of elasticity and the spectral theory of the Lam\'e operator. Moreover, as an interesting practical application of our theoretical findings, we apply the results to the inverse scattering problem of determining an elastic obstacle as well as its possible surface impedance parameter by a minimal/optimal number of far-field measurements. This is a challenging problem with a strong applied background. In its abstract formulation, the problem can be roughly described as a nonlinear operator equation, \begin{equation}\label{eq:ipa1} \mathcal{F}(\Omega, \eta)=\mathcal{M}(\hat x; \bmf{u}^i_j),\quad \hat x\in\mathbb{S}^{1}, j=1,2,\ldots, N, \end{equation} where the scattering map $\mathcal{F}$ is defined by a certain PDE system in the exterior of a domain $\Omega$. $\eta$ is a boundary impedance parameter on $\partial\Omega$. Through solving the aforementioned PDE system, the scattering map $\mathcal{F}$ sends $\Omega$ and $\eta$ to a real-analytic function $\mathcal{M}$ on the unit sphere, which signifies the observation data. This correspondence also depends on $\bmf{u}_j^i$, $j=1,2,\ldots, N$, known as the incident fields, that account for the number of measurements in the practical scenario. We shall give more relevant details about \eqref{eq:ipa1} in Section~\ref{sect:6}. We are mainly concerned with the unique identifiability issue for \eqref{eq:ipa1}. That is, we aim to establish the unique one-to-one correspondence between the target object $(\Omega, \eta)$ and the measurement data $\mathcal{M}$, particularly with the minimal/optimal number of measurements. Geometrically speaking, a single measurement, namely $\mathcal{M}(\hat x)$, $\hat x\in\mathbb{S}^1$, corresponding to a single incident field $\bmf{u}^i$ (or at most a few), may serve as a global parametrization for $\partial\Omega$. However, there is very limited progress in the literature on this challenging geometrical problem. The more recent progress is concerned with the case that $\Omega$ is of general polygonal shape \cite{ElschnerYama2010,LiuXiao}. The mathematical machinery therein is mainly based on certain reflection and path arguments, which cannot deal with the more challenging case that $\eta$ is not identically $0$ or $\infty$. Using the generalized Holmgren's principle established in this paper, we can provide a different and unified approach in tackling with the geometrical inverse problem\eqref{eq:ipa1} in the case that $\Omega$ is of general polygonal shape with at most a few measurements. More importantly, our method can deal with the more challenging case that $\eta$ is finite and not identically zero. We derive a comprehensive study for this geometrical inverse problem. It is mentioned in passing that unique determination by a minimal number of far-field patterns is a longstanding problem in the inverse scattering theory. We refer to \cite{AR,CDLZ,CDLZ2,CY,CK,CK18,LPRX,LRX,Liu-Zou,Liu-Zou3} and the references therein for related studies for the inverse acoustic and electromagnetic wave scattering problems. In addition to the application to the inverse problem, we believe that the generalized Holmgren's principle may find more interesting applications in different contexts. Finally, we would like to briefly discuss about the technicality of our study. We shall be mainly based on analyzing the microlocal singularities of the solution $\bmf{u}$ to \eqref{eq:lame} due to the presence of the homogeneous line segments discussed earlier. Clearly, the singularities are developed across the aforementioned line segments and are of analytic type. In the case that there are two intersecting line segments with a non-straight intersecting angle, it is not surprising that the singularities are developed at the intersecting point. However, we shall show that the singularities can even be developed across a single line segment, which are really subtle and tricky to capture. In this paper, we mainly focus on the two-dimensional case. As can be seen that even in the two dimensions, the analyses are highly technical and lengthy with tedious calculations. We shall present the extensions to the three dimensions as well as to the case with Cauchy data on an analytic curve instead of a straight line segment in forthcoming articles. The result of the paper is organized as follows. Sections 2--4 are devoted to establishing the generalized Holmgren's principle in different scenarios. Section 5 presents the unique identifiability results for the inverse elastic obstacle problem \eqref{eq:ipa1}. \section{Auxiliary results}\label{sect:2} We first introduce two important definitions. \begin{defn}\label{def:1} Let $\bmf u=(u_\ell)_{\ell=1}^2$ be a generalized Lam\'e eigenfunction to \eqref{eq:lame} associated with an eigenvalue $\kappa\in\mathbb{R}_+$. An open and connected line segment $\Gamma_{h}\Subset\Omega$ is called {\it a rigid line} of $\bmf{u}$ if $\bmf{u}|_{\Gamma_h}=\bmf{0}$; {\it a traction-free line} if $T_{\bmf{\nu}}\bmf{u}|_{\Gamma_h}=\bmf{0}$; and an {\it impedance line} if \begin{equation}\label{eq:im line} (T_{\bmf{\nu}}\bmf{u}+\eta \bmf{u})\big|_{\Gamma_h}=\bmf{0}, \end{equation} where $\eta\in\mathbb{C}$ is constant and referred to as an impedance parameter. Set ${\mathcal R}_\Omega^{\kappa} $, ${\mathcal T}_\Omega^{\kappa} $ and ${\mathcal I}_\Omega^{\kappa} $ to respectively denote the sets of rigid, traction-free and impedance lines in $\Omega$ of $\bmf{u}$. \end{defn} \begin{defn}\label{def:generalized line} Recall that the unit normal vector $\nu $ and the tangential vector $\boldsymbol{\tau}$ to $\Gamma_h$ are defined in \eqref{eq:nutau}, respectively. Define \begin{subequations} \begin{align} {\mathcal S}\left( {\mathcal R}_\Omega^{\kappa} \right):=& \{\Gamma_h \in {\mathcal R}_\Omega^{\kappa} ~|~ \exists \bmf{x}_0 \in \Gamma_h \mbox{ such that } \boldsymbol{\tau} \cdot \partial_{\nu} \bmf{u} |_{\bmf{x}={\bmf{x}}_0 }=0 \},\label{eq:def1} \\ {\mathcal S}\left( {\mathcal T}_\Omega^{\kappa} \right):=& \{\Gamma_h \in {\mathcal T}_\Omega^{\kappa} ~|~ \exists \bmf{x}_0 \in \Gamma_h \mbox{ such that } \bmf{u}(\bmf{x}_0)=\bmf{0} \mbox{ and }\boldsymbol{\tau} \cdot \partial_{\nu} \bmf{u} |_{\bmf{x}={\bmf{x}}_0 }=0 \},\label{eq:def2}\\ {\mathcal S}\left( {\mathcal I}_\Omega^{\kappa} \right):=& \{\Gamma_h \in {\mathcal I}_\Omega^{\kappa} ~|~ \exists \bmf{x}_0 \in \Gamma_h \mbox{ such that } \bmf{u}(\bmf{x}_0)=\bmf{0} \mbox{ and }\boldsymbol{\tau} \cdot \partial_{\nu} \bmf{u} |_{\bmf{x}={\bmf{x}}_0 }=0 \},\label{eq:def3} \end{align} \end{subequations} where ${\mathcal S}\left( {\mathcal R}_\Omega^{\kappa} \right)$, ${\mathcal S}\left( {\mathcal T}_\Omega^{\kappa} \right)$ and ${\mathcal S}\left( {\mathcal I}_\Omega^{\kappa} \right)$ are named as the sets of the \emph {singular rigid, singular traction-free } and \emph{singular impedance} lines of $\bmf{u}$ respectively. Let ${\mathcal S}( \Omega )= {\mathcal S}\left( {\mathcal R}_\Omega^{\kappa} \right)\cup {\mathcal S}\left( {\mathcal T}_\Omega^{\kappa} \right) \cup {\mathcal S}\left( {\mathcal I}_\Omega^{\kappa} \right) $ be a set of singular lines of $\bmf{u}$ in $\Omega$. \end{defn} It is noted that compared to the homogeneous lines introduced in Definition~\ref{def:1}, the singular lines in Definition~\ref{def:generalized line} are further required to satisfy a number of conditions on a specific point. In what follows, we shall prove that if a (generalised) Lam\'e eigenfunction $\bmf{u}$ to \eqref{eq:lame} possesses a singular line in $\Omega$, then $\bmf{u}$ is identically zero. We prove this by quantitatively characerizing $\bmf{u}$ in the phase space across the lines. This is the reason that we call them (microlocally) singular lines. Furthermore, we show that the generic intersections of the homogeneous lines of Definition~\ref{def:1} shall also generate microlocal singularities, which prevent the occurrence of such intersections unless $\bmf{u}$ is trivially zero. In this article, we provide a comprehensive characterization of all those cases. To our best knowledge, those results are new to the literature. Next we introduce the geometric setup of our study. Consider two line segments respectively defined by (see Fig.~\ref{fig1} for a schematic illustration) \begin{equation}\label{eq:gamma_pm} \begin{split} \Gamma_h^+&=\{\bmf{x} \in \mathbb R^2~|~\bmf{x}=r\cdot (\cos \varphi_0, \sin \varphi_0 )^\top ,\quad 0\leq r\leq h,\quad 0<\varphi_0\leq 2\pi \}, \\ \Gamma_h^-&=\{\bmf{x} \in \mathbb R^2~|~\bmf{x}=r\cdot (1, 0 )^\top,\quad 0\leq r\leq h \},\ \ h\in\mathbb{R}_+. \end{split} \end{equation} Clearly, the intersecting angle between $\Gamma_h^+$ and $\Gamma_h^-$ is \begin{equation}\label{eq:angle1} \angle(\Gamma_h^+,\Gamma_h^{-})=\varphi_0, \quad 0< \varphi_0 \leq 2\pi. \end{equation} It is noted that if $\varphi_0=\pi$ or $2\pi$, $\Gamma_h^+$ and $\Gamma_h^-$ are actually lying on a same line. In such a case, the intersection between $\Gamma_h^+$ and $\Gamma_h^-$ is said to be degenerate. In our subsequent study, $\Gamma_h^\pm$ shall be the homogeneous lines in Definition~\ref{def:1} or the singular lines in Definition~\ref{def:generalized line}. In fact, for any two of such lines that are intersecting in $\Omega$ (or one line in the degenerate case), since the PDO $\mathcal{L}$ defined in \eqref{eq:lame} is invariant under rigid motions, one can always have two lines as introduced in \eqref{eq:gamma_pm} after a straightforward coordinate transformation such that the homogeneous conditions in Definitions~\ref{def:1} and \ref{def:generalized line} are still satisfied on $\Gamma_h^\pm$. We assume that $h\in\mathbb{R}_+$ is sufficiently small such that $\Gamma_h^\pm$ are contained entirely in $\Omega$. Moreover, if $\Gamma_h^\pm$ are impedance lines, we assume that the impedance parameters on $\Gamma_h^\pm$ are respectively two constants $\eta_1$ and $\eta_2$. As also noted before that $\bmf{u}$ is analytic in $\Omega$, it is sufficient for us to consider the case that $0<\varphi_0\leq \pi$. In fact, if $\pi < \varphi_0 \leq 2\pi$, we see that $\Gamma_h^+$ belongs to the half-plane of $x_2<0$ (see Fig.~\ref{fig1}). Let $\widetilde{\Gamma}_h^+$ be the extended line segment of length $h$ in the half-plane of $x_2>0$. By the analytic continuation, we know that $\widetilde{\Gamma}_h^+$ is of the same type of $\Gamma_h^+$, namely $\bmf{u}$ satisfies the same homogeneous condition on $\widetilde{\Gamma}_h^+$ as that on $\Gamma_h^+$. Hence, instead of studying the intersection of $\Gamma_h^+$ and $\Gamma_h^-$, one can study the intersection of $\widetilde{\Gamma}_h^+$ and $\Gamma_h^-$. Clearly, $\angle(\widetilde{\Gamma}_h^+, \Gamma_h^-)\in (0,\pi]$. \begin{figure \centering \includegraphics[width=0.3\textwidth]{fig1.eps} \caption{Schematic of the geometry of two intersecting lines with an angle $\varphi_0$ with $0< \varphi_0\leq \pi $.} \label{fig1} \end{figure} Let $B_h$ be the central disk of radius $h \in \mathbb{R}_+$. Let $\Gamma^\pm$ signify the infinite extension of $\Gamma_{h}^\pm$ in the half-space $x_2\geq 0$. Consider the open sector \begin{equation}\label{eq:K} \mathcal{K} =\left\{\bmf{x}=(x_1,x_2) \in \mathbb{R}^{2}~ |~ \bmf{x}\neq \bmf{0},\quad 0<\arg \left(x_{1}+\mathrm{i} x_{2}\right)<\varphi_{0}\right\},\quad \mathrm{i}:=\sqrt{-1}, \end{equation} which is formed by the two half-lines $\Gamma^-$ and $\Gamma^+$. In the sequel, we set \begin{equation}\label{eq:sh} S_h=\mathcal{K}\cap B_h, \end{equation} where $\partial S_h=\Gamma_h^+\cup\Gamma_h^-\cup\Lambda_h$ and \begin{equation}\label{eq:Lambda_h} \begin{aligned} & \Lambda_h=\mathcal{K}\cap \partial B_h.\\ \end{aligned} \end{equation} Clearly in $S_h$, the unit outward normal vectors to $\Gamma_h^+$ and $\Gamma_h^-$ are respectively \begin{equation}\label{eq:nu} \bmf{\nu }\big |_{\Gamma_h^+}=(-\sin\varphi_0,\cos\varphi_0),\quad \bmf{\nu }\big |_{\Gamma_h^-}=(0,-1). \end{equation} Throughout the rest of the paper, we set \begin{equation}\label{eq:kpks} k_{{p}}=\sqrt{ \frac{\kappa }{\lambda+2 \mu} } \text { and } k_{{s}}=\sqrt{ \frac{\kappa }{\mu}}, \end{equation} which are known as the compressional and shear wave numbers, respectively. We next present a few lemmas that will be needed in our subsequent analysis. The following lemma from \cite{DR95,SP} states the Fourier expansion in terms of the radial wave functions of the solution $\bmf{u}$ to \eqref{eq:lame} around the origin. \begin{lem}\cite{DR95,SP}\label{lem:u exp} Recall that $J_m(t)$ is the first-kind Bessel function of order $m\in \mathbb{N} \cup \{0\} $ and $\bmf{x}=r(\cos \varphi, \sin \varphi )^\top$. $\bmf{u}(\bmf{x})$ to \eqref{eq:lame} has the following radial wave expansion at the origin, \begin{equation}\label{eq:radial} \begin{aligned} \mathbf{u}(\mathbf{x}) =& \sum_{m=0}^{\infty}\left\{a_{m}\left\{k_{p} J_{m}^{\prime}\left(k_{p} r\right) \mathrm{e}^{{\mathrm{i}} m \varphi} \mathbf{\hat{r}}+\frac{{\mathrm{i}} m}{r} J_{m}\left(k_{p} r\right) \mathrm{e}^{{\mathrm{i}} m \varphi} \bm{\hat{\varphi}}\right\}\right.\\ &+b_{m}\left\{\frac{{\mathrm{i}} m}{r} J_{m}\left(k_{s} r\right) \mathrm{e}^{{\mathrm{i}} m \varphi} \mathbf{\hat{r}}-k_{s} J_{m}^{\prime}\left(k_{s} r\right) \mathrm{e}^{{\mathrm{i}} m \varphi} \bm{\hat{\varphi}}\right\}\bigg\}, \end{aligned} \end{equation} where $a_m$ and $b_m$ are constants, $\bm{\hat{\varphi}}=\left(\begin{array}{c}{-\sin{\varphi}}\\ {\cos{\varphi}} \end{array}\right)$, $\bmf{\hat{r}}=\left(\begin{array}{c}{\cos{\varphi}}\\ {\sin{\varphi}} \end{array}\right)$ and the prime denotes the differentiation with respect to $k_a r$, $a=p$ or $a=s$. Note that \eqref{eq:radial} converges uniformly on compact subsets of $\mathbb R^2$. \end{lem} By the analyticity of $\bmf{u}$ in the interior domain of $\Omega$ and the analytic continuation principle, we have the following proposition. \begin{prop}\label{prop:21} Suppose $\mathbf{0}\in \Omega$ and $\bmf{u}$ has the expansion \eqref{eq:radial} around the origin such that $a_m=b_m=0$ for $\forall m \in \mathbb \cup \{0\}$. Then $$ \bmf{u} \equiv \bmf{0} \mbox{ in } \Omega . $$ \end{prop} The recursive relationship of the first-kind Bessel function and its derivative can be found in \cite{Abr}. In fact we have \begin{lem}\cite{Abr}\label{lem:J exp} Recall that $J_m(t)$ is the first-kind Bessel function of the order $m \in \mathbb{N}\cup \{0\}$. Then \begin{equation}\label{eq:lem recursive} J_m'\left(t\right)=\frac{J_{m-1}\left(t\right)-J_{m+1}\left(t\right)}{2},\quad J_m\left(t\right)=\frac{t\left(J_{m-1}\left(t\right)+J_{m+1}\left(t\right)\right). }{2m} \end{equation} Moreover, we have \begin{equation}\label{eq:J-1} J_{-m}(t)=(-1)^m J_m(t) \end{equation} \end{lem} \begin{rem} Using Lemma \ref{lem:J exp}, one can derive that \begin{equation}\label{eq:J2} \begin{aligned} & \frac{J_{m-1}(k_p r)}{r}=\frac{k_p}{2(m-1)}\left(J_{m-2}(k_p r)+J_m(k_p r)\right), \frac{J_{m+1}(k_p r)}{r}=\frac{k_p}{2(m+1)}\left(J_m(k_p r)+J_{m+2}(k_p r)\right),\\ & \frac{J_{m-1}(k_s r)}{r}=\frac{k_s}{2(m-1)}\left(J_{m-2}(k_s r)+J_m(k_s r)\right), \frac{J_{m+1}(k_s r)}{r}=\frac{k_s}{2(m+1)}\left(J_m(k_s r)+J_{m+2}(k_s r)\right). \end{aligned} \end{equation} \end{rem} \begin{lem}\label{lem:uu exp} The radial wave expression of $\mathbf{u}(\bmf{x})$ to \eqref{eq:lame} at the origin can be written as \begin{equation}\label{eq:u} \begin{aligned} \mathbf{u}(\mathbf{x})= \sum_{m=0}^{\infty} & \left\{ \frac{k_p}{2} a_{m} \mathrm{e}^{{\mathrm{i}} m \varphi} \left\{J_{m-1}\left(k_{p} r\right)\mathrm{e}^{-{\mathrm{i}} \varphi}\mathbf{e}_1 -J_{m+1}\left(k_{p}r\right)\mathrm{e}^{{\mathrm{i}} \varphi}\mathbf{e}_2 \right\}\right.\\ & + \frac{{\mathrm{i}} k_s}{2} b_{m} \mathrm{e}^{{\mathrm{i}} m \varphi} \left\{J_{m-1}\left(k_{s} r\right)\mathrm{e}^{-{\mathrm{i}} \varphi}\mathbf{e}_1 +J_{m+1}\left(k_{s} r\right)\mathrm{e}^{{\mathrm{i}} \varphi} \mathbf{e}_2 \right\} \bigg\} . \end{aligned} \end{equation} where and also throughout the rest of the paper, $\mathbf{e}_1:=(1,{\mathrm{i}})^\top \mbox{ and }\mathbf{e}_2:=(1,-{\mathrm{i}})^\top$. \end{lem} \begin{proof} Using \eqref{eq:lem recursive} and Euler's formula, it can be deduced that \begin{equation}\label{eq:lem 110} \begin{aligned} & k_{p} J_{m}^{\prime}\left(k_{p} r\right) \mathrm{e}^{{\mathrm{i}} m \varphi} \mathbf{\hat{r}}+\frac{{\mathrm{i}} m}{r} J_{m}\left(k_{p} r\right) \mathrm{e}^{{\mathrm{i}} m \varphi} \bm{\hat{\varphi}}\\ = & k_p \frac{J_{m-1}\left(k_p r\right)-J_{m+1}\left(k_p r\right)}{2} \mathrm{e}^{{\mathrm{i}} m \varphi} \mathbf{\hat{r}}+ \frac{{\mathrm{i}} k_p }{2}\left(J_{m-1}\left(k_p r\right)+J_{m+1}\left(k_p r\right)\right)\mathrm{e}^{{\mathrm{i}} m \varphi} \bm{\hat{\varphi}}\\ = & \frac{k_p}{2} \mathrm{e}^{{\mathrm{i}} m \varphi} \left\{J_{m-1}\left(k_p r\right)\mathrm{e}^{-{\mathrm{i}}\varphi}\mathbf{e}_1- J_{m+1}\left(k_p r\right)\mathrm{e}^{{\mathrm{i}}\varphi}\mathbf{e}_2\right\}. \end{aligned} \end{equation} Similarly, we have \begin{equation}\label{eq:lem 111} \begin{aligned} & \frac{{\mathrm{i}} m}{r} J_{m}\left(k_{s} r\right) \mathrm{e}^{{\mathrm{i}} m \varphi} \mathbf{\hat{r}}-k_{s} J_{m}^{\prime}\left(k_{s} r\right) \mathrm{e}^{{\mathrm{i}} m \varphi}\bm{\hat{\varphi}} \\ = & \frac{{\mathrm{i}} k_s}{2}\left(J_{m-1}\left(k_s r\right)+J_{m+1}\left(k_s r\right)\right)\mathrm{e}^{{\mathrm{i}} m \varphi} \mathbf{\hat{r}}-k_s \frac{J_{m-1}\left(k_s r\right)-J_{m+1}\left(k_s r\right)}{2} \mathrm{e}^{{\mathrm{i}} m \varphi} \bm{\hat{\varphi}}\\ = & \frac{{\mathrm{i}} k_s}{2} \mathrm{e}^{{\mathrm{i}} m \varphi} \left\{J_{m-1}\left(k_s r\right)\mathrm{e}^{-{\mathrm{i}}\varphi}\mathbf{e}_1+ J_{m+1}\left(k_s r\right)\mathrm{e}^{{\mathrm{i}}\varphi}\mathbf{e}_2\right\}. \end{aligned} \end{equation} Substituting \eqref{eq:lem 110} and \eqref{eq:lem 111} into \eqref{eq:radial}, after some algebraic calculations, we can prove \eqref{eq:u}. \end{proof} \begin{rem} In view of \eqref{eq:u}, we have \begin{equation}\label{eq:u comp} \begin{aligned} u_1 \left(\bmf{x}\right)= \sum_{m=0} ^\infty & \Big [ \frac{k_p}{2} a_m \left(\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-1} \left(k_p r\right) - \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+1} \left(k_p r\right) \right) \\ & + \frac{{\mathrm{i}} k_s}{2} b_m \left(\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-1} \left(k_s r\right) + \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+1} \left(k_s r\right) \right) \Big ],\\ u_2 \left(\bmf{x}\right)= \sum_{m=0} ^\infty & \Big [ \frac{{\mathrm{i}} k_p}{2} a_m \left( \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-1} \left(k_p r\right) + \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+1} \left(k_p r\right) \right) \\ & + \frac{ k_s}{2} b_m \left(-\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-1} \left(k_s r\right) + \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+1} \left(k_s r\right) \right) \Big ]. \end{aligned} \end{equation} \end{rem} Using \eqref{eq:radial}, we can obtain the corresponding Fourier representation of the boundary traction operator $T_{\bmf{\nu }}\mathbf u \Big |_{\Gamma^\pm_h }$ defined in \eqref{eq:Tu} as follows. \begin{lem}\label{lem:Tuu exp} Let $\bmf{u}(\bmf{x})$ be a Lam\'e eigenfunction to \eqref{eq:lame} with the Fourier expansion \eqref{eq:u} and $\Gamma^\pm_h$ is defined in \eqref{eq:gamma_pm}. Then $ T_\nu \mathbf{u}\Big |_{\Gamma^+_h}$ possesses the following radial wave expansion at the origin \begin{equation}\label{eq:Tu1} \begin{aligned} T_\nu \mathbf{u}\Big |_{\Gamma^+_h} & = \sum_{m=0}^{\infty} \left\{\frac{{\mathrm{i}} k_p^2}{2} a_{m} \Big[ \mathrm{e}^{{\mathrm{i}} \left(m-2\right) \varphi} \mathrm{e}^{{\mathrm{i}} \varphi_0} \mu J_{m-2}(k_p r) \mathbf{e}_1 + \mathrm{e}^{{\mathrm{i}} m \varphi} \mathrm{e}^{ - {\mathrm{i}} \varphi_0} (\lambda+\mu) J_m(k_p r) \mathbf{e}_1\right.\\ & - \mathrm{e}^{{\mathrm{i}} (m+2) \varphi} \mathrm{e}^{-{\mathrm{i}} \varphi_0} \mu J_{m+2}\left(k_{p} r\right) \mathbf{e}_2 - \mathrm{e}^{{\mathrm{i}} m \varphi} \mathrm{e}^{{\mathrm{i}} \varphi_0} \left(\lambda+\mu\right) J_m\left(k_{p} r\right) \mathbf{e}_2\Big] \\ &- \frac{k_s^2}{2} b_{m}\Big[ \mathrm{e}^{{\mathrm{i}} \left(m-2\right) \varphi} \mathrm{e}^{{\mathrm{i}} \varphi_0} \mu J_{m-2}\left(k_{s} r\right) \mathbf{e}_1 - \mathrm{e}^{{\mathrm{i}} \left(m+2\right) \varphi} \mathrm{e}^{-{\mathrm{i}} \varphi_0} \mu J_{m+2}\left(k_{s} r\right) \mathbf{e}_2\Big] \bigg\}. \end{aligned} \end{equation} Similarly, the radial wave expansion of $ T_\nu \mathbf{u}\Big |_{\Gamma^-_h}$ at the origin is given by \begin{equation}\label{eq:Tu2} \begin{split} & T_\nu \mathbf{u}\Big |_{\Gamma^-_h}= \sum_{m=0}^{\infty} \left\{-\frac{{\mathrm{i}} k_p^2}{2} a_{m} \mu J_{m-2}(k_p r) \mathbf{e}_1 - \frac{{\mathrm{i}} k_p^2}{2} a_{m} (\lambda+\mu) J_m(k_p r) \mathbf{e}_1\right.\\ & + \frac{k_s^2}{2} b_{m} \mu J_{m-2}\left(k_{s} r\right) \mathbf{e}_1 + \frac{{\mathrm{i}} k_p^2}{2} a_{m} \left(\lambda+\mu\right) J_m\left(k_{p} r\right) \mathbf{e}_2 + \frac{{\mathrm{i}} k_p^2}{2} a_{m} \mu J_{m+2}\left(k_{p} r\right) \mathbf{e}_2 + \frac{k_s^2}{2} b_{m} \mu J_{m+2}\left(k_{s} r\right) \mathbf{e}_2 \bigg\}. \end{split} \end{equation} \end{lem} The proof of Lemma~\ref{lem:Tuu exp} involves rather tedious calculations and it is postponed to be given in Appendix. Combing Lemmas \ref{lem:uu exp} and \ref{lem:Tuu exp}, for impedance boundary conditions defined on $\Gamma_h^\pm$ with the boundary parameters being constant on $\Gamma_h^\pm$, we have \begin{lem}\label{Tu+ exp} Let $\Gamma_h^-$ and $\Gamma_h^+$ be two impedance lines of $\bmf{u}$ with constant boundary parameters $\eta_1$ and $\eta_2$ respectively. We have \begin{equation}\label{eq:Tu3} \begin{aligned} & \left(T_\nu \mathbf{u}+\eta_2 \boldsymbol{u}\right )\Big |_{\Gamma^+_h}= \sum_{m=0}^{\infty} \bigg \{\frac{{\mathrm{i}} k_p^2} {2} a_{m} \Big[ \mathrm{e}^{{\mathrm{i}} \left(m-2\right) \varphi} \mathrm{e}^{{\mathrm{i}} \varphi_0} \mu J_{m-2}(k_p r) \mathbf{e}_1 + \mathrm{e}^{{\mathrm{i}} m \varphi} \mathrm{e}^{ - {\mathrm{i}} \varphi_0} (\lambda+\mu) J_m(k_p r) \mathbf{e}_1 \\ & -\mathrm{e}^{{\mathrm{i}} m \varphi} \mathrm{e}^{{\mathrm{i}} \varphi_0} \left(\lambda+\mu\right) J_m\left(k_{p} r\right) \mathbf{e}_2 -\mathrm{e}^{{\mathrm{i}} (m+2) \varphi} \mathrm{e}^{-{\mathrm{i}} \varphi_0} \mu J_{m+2}\left(k_{p} r\right) \mathbf{e}_2 \Big] \\ & - \frac{k_s^2}{2} b_{m}\Big[ \mathrm{e}^{{\mathrm{i}} \left(m-2\right) \varphi} \mathrm{e}^{{\mathrm{i}} \varphi_0} \mu J_{m-2}\left(k_{s} r\right) \mathbf{e}_1+ \mathrm{e}^{{\mathrm{i}} \left(m+2\right) \varphi} \mathrm{e}^{-{\mathrm{i}} \varphi_0} \mu J_{m+2}\left(k_{s} r\right) \mathbf{e}_2\Big] +\frac{\eta_2 k_p}{2} a_{m} \mathrm{e}^{{\mathrm{i}} m \varphi}\times \\ & \bigg[J_{m-1}\left(k_{p} r\right)\mathrm{e}^{-{\mathrm{i}} \varphi}\mathbf{e}_1 -J_{m+1}\left(k_{p}r\right)\mathrm{e}^{{\mathrm{i}} \varphi}\mathbf{e}_2 \bigg] + \frac{{\mathrm{i}} \eta_2 k_s}{2} b_{m} \mathrm{e}^{{\mathrm{i}} m \varphi} \left[J_{m-1}\left(k_{s} r\right)\mathrm{e}^{-{\mathrm{i}} \varphi}\mathbf{e}_1 +J_{m+1}\left(k_{s} r\right)\mathrm{e}^{{\mathrm{i}} \varphi} \mathbf{e}_2 \right] \bigg\}, \end{aligned} \end{equation} and \begin{equation}\label{eq:Tu4} \begin{aligned} & \left( T_\nu \mathbf{u}+\eta_1 \boldsymbol{u}\right) \Big|_{\Gamma_h^-}=\sum_{m=0}^{\infty} \biggl \{-\frac{{\mathrm{i}} {k}_p^2}{2} a_{m} \Big[ \mu J_{m-2}(k_p r) \mathbf{e}_1 + (\lambda+\mu) J_m(k_p r) \mathbf{e}_1\\ & - \left(\lambda+\mu\right) J_m\left(k_{p} r\right) \mathbf{e}_2 - \mu J_{m+2}\left(k_{p} r\right) \mathbf{e}_2\Big] + \frac{k_s^2}{2} b_{m}\Big[ \mu J_{m-2}\left(k_{s} r\right) \mathbf{e}_1 + \mathrm{e}^{-{\mathrm{i}} \varphi_0} \mu J_{m+2}\left(k_{s} r\right) \mathbf{e}_2 \Big] \\ & +\frac{\eta_1 k_p}{2} a_{m} \left[J_{m-1}\left(k_{p} r\right)\mathbf{e}_1 -J_{m+1}\left(k_{p}r\right) \mathbf{e}_2 \right] + \frac{{\mathrm{i}} \eta_1 k_s}{2} b_{m} \left[J_{m-1}\left(k_{s} r\right)\mathbf{e}_1 +J_{m+1}\left(k_{s} r\right)\mathbf{e}_2 \right] \bigg\}. \end{aligned} \end{equation} \end{lem} \begin{lem}{\cite{CDLZ}}\label{lem:co exp} Suppose that for $0<h \ll 1$ and $t \in (0,h)$, $\sum_{n=0}^{\infty}\alpha_{n}J_{n}(t)=0,$ where $J_{n}\left(t\right)$ is the $n$-th Bessel function of the first kind. Then $\alpha_{n}=0,n=0,1,2,...$ \end{lem} Next we set \begin{equation}\label{eq:v} \bmf{v}(\bmf{x} )=\left(\begin{array}{c}{\exp (-s \sqrt{r} \exp({\mathrm{i}} \varphi/2))} \\ {\mathrm{i}} \cdot {\operatorname{exp}(-s \sqrt{r} \exp({\mathrm{i}} \varphi/2) )}\end{array}\right):= \left(\begin{array}{c}{v_1(\bmf{x})} \\ {v_2(\bmf{x})}\end{array}\right)=v_1(\bmf{x}) \bmf{e}_1, \end{equation} where $\bmf{x}=r\cdot (\cos \varphi, \sin \varphi ) $, $s \in \mathbb{R}_{+}$, $-\pi<\varphi\leqslant \pi$ and $\bmf{e}_1$ is defined in Lemma~\ref{lem:uu exp}. $\bmf{v}$ is known as the Complex Geometrical Optics (CGO) solution for the Lam\'e operator and it was first introduced in \cite{EBL}. We have \begin{lem}\cite[Lemma 2.1]{EBL} Let $\Omega \subset { \mathbb R}^2 $ be such that $\Omega \cap\left(\mathbb{R}_{-} \cup\{\bmf{0}\}\right)=\emptyset$ and $\bmf{v}$ be defined in \eqref{eq:v}. Then there holds $\mathcal{L} \bmf{v} =0 \mbox{ in } \Omega$. \end{lem} By direct calculations, one can derive the following lemma. \begin{lem}\label{eq:lem22 t} Let $\bmf{v}$ be defined in \eqref{eq:v}. Denote \begin{equation}\label{eq:zeta} \zeta{(\varphi)}=-\mathrm{e}^{{\mathrm{i}} \varphi/2 }. \end{equation} For any given curve $\Gamma \Subset \mathbb{R}^2$ with a unit normal vector $\nu =(\nu_1,\nu_2 ) $, if $\bmf{v}$ is complex analytic in a neighbourhood of $\Gamma$, then $$ T_\nu \bmf{v}\big |_{\Gamma} (\bmf{x})= \mu (\nu_1+{\mathrm{i}} \nu_2 ) \frac{s \exp ( sr^{1/2} \zeta{(\varphi)} )}{r^{1/2} \zeta{(\varphi)} } \bmf{e}_1, \quad \bmf{x} =r(\cos \varphi, \sin \varphi ) \in \Gamma, $$ where the boundary traction operator $T_\nu \bmf{v}\big |_{\Gamma} $ and $\bmf{e}_1$ are defined in \eqref{eq:Tu} and Lemma~\ref{lem:uu exp}, respectively. \end{lem} Using Lemma \ref{eq:lem22 t} and the fact that the CGO solution $\bmf{v}$ is complex analytic on a neighborhood of $\Gamma_h^\pm \backslash\{ \bmf{0}\} $, and noting that the unit normal vectors on $\Gamma_h^\pm$ are given by \eqref{eq:nu}, we have \begin{lem}\label{lem:lemTv} Let $\bmf{v}$ be defined in \eqref{eq:v}. Recall that $\Gamma_h^\pm$ are given by \eqref{eq:gamma_pm}, where their corresponding unit normal vectors $\nu\big|_{\Gamma_h^+ }$ and $\nu\big|_{\Gamma_h^- }$ are defined in \eqref{eq:nu}. Then we have \begin{equation*} \begin{split} T_{\nu} \bmf{v}\big|_{ \Gamma_h^+ \backslash\{\bmf{0}\} }={\mathrm{i}} s \mu \zeta{(\varphi_0)}\frac{\exp(sr^{1/2} \zeta{(\varphi_0))}}{r^{1/2}}\mathbf{e}_1,\ \ T_{\nu} \bmf{v}\big|_{ \Gamma_h^- \backslash\{\bmf{0}\} }={\mathrm{i}} s \mu \frac{\exp(-sr^{1/2} )}{r^{1/2}} \mathbf{e}_1,\\ \end{split} \end{equation*} where $\zeta(\varphi_0)$ is defined in \eqref{eq:zeta}. \end{lem} Next, we derive the expansions of $ T_\nu \mathbf{u}\cdot \bmf{v} $, $ T_\nu \mathbf{v}\cdot \bmf{u} $ and $ \mathbf{u}\cdot \bmf{v} $ on $\Gamma_h^\pm$ around the origin, where $\bmf{u}$ is given by \eqref{eq:u} and $\bmf{v}$ is the CGO solution defined in \eqref{eq:v}. These expansions will be used to analyze the vanishing property of $\bmf{u}$ at the intersecting point $\bmf{0}$ of $\Gamma_h^\pm$. \begin{lem}\cite[Proposition 2.1.7]{krantz}\label{lem:kra} If the power series $\sum_{\mu } a_{\mu } \bmf{x}^\mu $ converges at a point $\bmf{x}_0$, then it converges uniformly and absolutely on compact subsets of $U(\bmf{x}_0)$, where $$ U(\bmf{x}_0)=\{(r_1 x_{0,1},\ldots, r_n x_{0,n}):-1<r_j<1,j=1,\ldots,n\}, \, \bmf{x}_0=(x_{0,1},\ldots, x_{0,n}) \in {\mathbb R}^n. $$ \end{lem} \begin{lem}\label{lem:211 expan} Let $\bmf{u}$ be a Lam\'e eigenfunction to \eqref{eq:lame} and the CGO solution $\bmf{v}$ be defined in \eqref{eq:v}. Recall that the Lam\'e eigenfunction $\mathbf{u}$ to \eqref{eq:lame} has the radial wave expansion \eqref{eq:u} at the origin. Then the following expansions \begin{equation}\label{eq:Tudotv} \begin{split} T_\nu \mathbf{u}\Big |_{\Gamma^+_h} \cdot \bmf{v} \Big |_{\Gamma^+_h} & = - {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) }\bigg\{{\mathrm{i}} k_p^2 (\lambda+\mu) \mathrm e^{{\mathrm{i}} \varphi_0} a_0 + \frac{{\mathrm{i}}}{2} k_p^3 (\lambda+\mu) \mathrm e^{2 {\mathrm{i}} \varphi_0} a_1 r\\ &\quad +\frac{1}{8}({\mathrm{i}} k_p^4 (\lambda+\mu) \mathrm e^{3 {\mathrm{i}} \varphi_0} a_2 - {\mathrm{i}} k_p^4 (2 \lambda+\mu) \mathrm e^{ {\mathrm{i}} \varphi_0} a_0+ k_s^4 \mu \mathrm e^{ {\mathrm{i}} \varphi_0} b_0)r^2 + R_{1,\Gamma_h^+} \bigg\},\\ T_\nu \mathbf{u}\Big |_{\Gamma^-_h} \cdot \bmf{v} \Big |_{\Gamma^-_h} & = {\mathrm e}^{-s \sqrt{ r}}\bigg\{{\mathrm{i}} k_p^2 (\lambda+\mu) a_0 + \frac{{\mathrm{i}}}{2} k_p^3 (\lambda+\mu) a_1 r\\ &\quad +\frac{1}{8}({\mathrm{i}} k_p^4 (\lambda+\mu) a_2 - {\mathrm{i}} k_p^4 (2 \lambda+\mu) a_0+ k_s^4 \mu b_0)r^2 + R_{1,\Gamma_h^-} \bigg\}, \end{split} \end{equation} converge uniformly and absolutely in $r\in (0,h]$, where \begin{equation}\label{eq:RTuv+} \begin{aligned} & R_{1,\Gamma_h^+} = r^3 \Bigg\{ {\mathrm{i}} k_p^2 (\lambda+\mu)\left[a_0 \mathrm e^{ {\mathrm{i}} \varphi_0} \sum_{k=2}^{\infty} \frac{(-1)^k k_p^{2k}}{2^{2k}k! k!}r^{2k-3} \right.\\ &+\sum_{m=1}^2\sum_{k=1}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1) \varphi_0} \frac{(-1)^k k_p^{2k+m}}{2^{2k+m}k! (k+m)!}r^{2k+m-3}+\sum_{m=3}^{\infty}\sum_{k=0}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1) \varphi_0} \frac{(-1)^k k_p^{2k+m}}{2^{2k+m}k! (k+m)!}r^{2k+m-3} \bigg]\\ & + {\mathrm{i}} k_p^2 \mu \left[a_0 \mathrm e^{ {\mathrm{i}} \varphi_0} \sum_{k=1}^{\infty}\frac{(-1)^k k_p^{2k+2}}{2^{2k+2}k! (k+2)!}r^{2k-1}\right. + \sum_{m=1}^{\infty}\sum_{k=0}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1) \varphi_0} \frac{(-1)^k k_p^{2k+m+2}}{2^{2k+m+2}k! (k+m+2)!}r^{2k+m-1}\bigg]\\ & + k_s^2 \mu \left[b_0 \mathrm e^{ {\mathrm{i}} \varphi_0} \sum_{k=1}^{\infty}\frac{(-1)^k k_S^{2k+2}}{2^{2k+2}k! (k+2)!}r^{2k-1}\right.+ \sum_{m=1}^{\infty}\sum_{k=0}^{\infty} b_m \mathrm e^{ {\mathrm{i}} (m+1) \varphi_0} \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+2}k! (k+m+2)!}r^{2k+m-1}\bigg]\Bigg\}, \end{aligned} \end{equation} and \begin{equation}\label{eq:RTuv-} \begin{aligned} & R_{1,\Gamma_h^-} = r^3 \Bigg\{ {\mathrm{i}} k_p^2 (\lambda+\mu)\left[a_0 \sum_{k=2}^{\infty} \frac{(-1)^k k_p^{2k}}{2^{2k}k! k!}r^{2k-3} \right.+\sum_{m=1}^2\sum_{k=1}^{\infty} a_m \frac{(-1)^k k_p^{2k+m}}{2^{2k+m}k! (k+m)!}r^{2k+m-3}\\ &+\sum_{m=3}^{\infty}\sum_{k=0}^{\infty} a_m \frac{(-1)^k k_p^{2k+m}}{2^{2k+m}k! (k+m)!}r^{2k+m-3} \bigg]\\ & + {\mathrm{i}} k_p^2 \mu \left[a_0 \sum_{k=1}^{\infty}\frac{(-1)^k k_p^{2k+2}}{2^{2k+2}k! (k+2)!}r^{2k-1}\right. + \sum_{m=1}^{\infty}\sum_{k=0}^{\infty} a_m \frac{(-1)^k k_p^{2k+m+2}}{2^{2k+m+2}k! (k+m+2)!}r^{2k+m-1}\bigg]\\ & + k_s^2 \mu \left [b_0 \sum_{k=1}^{\infty}\frac{(-1)^k k_s^{2k+2}}{2^{2k+2}k! (k+2)!}r^{2k-1}\right. + \sum_{m=1}^{\infty}\sum_{k=0}^{\infty} b_m \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+2}k! (k+m+2)!}r^{2k+m-1}\bigg]\Bigg\}. \end{aligned} \end{equation} Furthermore, the following expansions \begin{equation}\label{eq:Tvu} \begin{split} & T_{\nu} \bmf{v}\big|_{ \Gamma_h^+ \backslash\{\bmf{0}\} } \cdot \bmf{u}\big|_{ \Gamma_h^+ \backslash\{\bmf{0}\} }= {\mathrm{i}} s \mu \zeta(\varphi_0 ) {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) }\bigg\{\frac{1}{2}(-k_p^2 a_0 + {\mathrm{i}} k_s^2 b_0) {\mathrm e}^{{\mathrm{i}} \varphi_0} r^{\frac{1}{2}}+ \frac{1}{8}(-k_p^3 a_1 + {\mathrm{i}} k_s^3 b_1)\\ &\times {\mathrm e}^{2 {\mathrm{i}} \varphi_0} r^{\frac{3}{2}} +\frac{1}{48}(-k_p^4 a_2 + {\mathrm{i}} k_s^4 b_2) {\mathrm e}^{3 {\mathrm{i}} \varphi_0} r^{\frac{5}{2}} + \frac{1}{16}(k_p^4 a_0 - {\mathrm{i}} k_s^4 b_0) {\mathrm e}^{{\mathrm{i}} \varphi_0} r^{\frac{5}{2}} + R_{2,\Gamma_h^+}\bigg\},\\ & T_{\nu} \bmf{v}\big|_{ \Gamma_h^- \backslash\{\bmf{0}\} } \cdot \bmf{u}\big|_{ \Gamma_h^- \backslash\{\bmf{0}\} }= {\mathrm{i}} s \mu {\mathrm e}^{- s \sqrt{ r}}\bigg\{\frac{1}{2}(-k_p^2 a_0 + {\mathrm{i}} k_s^2 b_0) r^{\frac{1}{2}}\\ & + \frac{1}{8}(-k_p^3 a_1 + {\mathrm{i}} k_s^3 b_1) r^{\frac{3}{2}} +\frac{1}{48}(-k_p^4 a_2 + {\mathrm{i}} k_s^4 b_2) r^{\frac{5}{2}} + \frac{1}{16}(k_p^4 a_0 - {\mathrm{i}} k_s^4 b_0) r^{\frac{5}{2}} + R_{2,\Gamma_h^-}\bigg\},\\ & \bmf{u}\big|_{\Gamma_h^+} \cdot \bmf{v}\big|_{\Gamma_h^+} ={\mathrm e}^{s\sqrt{r} \zeta{(\varphi_0} )} \bigg\{\frac{1}{2} (-k_p^2 a_0+ {\mathrm{i}} k_s^2 b_0) \mathrm{e}^{{\mathrm{i}} \varphi_0} r + R_{0.\Gamma_h^+}\bigg\}, \\ & \bmf{u}\big|_{\Gamma_h^-} \cdot \bmf{v}\big|_{\Gamma_h^-} = {\mathrm e}^{-s\sqrt{r}} \bigg\{\frac{1}{2} (-k_p^2 a_0+ {\mathrm{i}} k_s^2 b_0) r + R_{0.\Gamma_h^+}\bigg\}, \end{split} \end{equation} converge uniformly and absolutely in $r\in (0,h]$, where \begin{equation}\label{eq:RTvu+} \begin{aligned} &R_{2,\Gamma_h^+} = r^{\frac{7}{2}}\Bigg\{ a_0 \mathrm e^{ {\mathrm{i}} \varphi_0} \sum_{k=2}^{\infty} \frac{(-1)^{k+1} k_p^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-3} \\ &+ \sum_{m=1}^{2} \sum_{k=1}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} \\ & + \sum_{m=3}^{\infty} \sum_{k=0}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} +{\mathrm{i}} b_0 \mathrm e^{ {\mathrm{i}} \varphi_0} \sum_{k=2}^{\infty} \frac{(-1)^k k_s^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-3} \\ & + {\mathrm{i}} \sum_{m=1}^{2} \sum_{k=1}^{\infty} b_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3}\\ & + {\mathrm{i}} \sum_{m=3}^{\infty} \sum_{k=0}^{\infty} b_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3}\Bigg\}, \end{aligned} \end{equation} and \begin{equation}\label{eq:RTvu-} \begin{aligned} & R_{2,\Gamma_h^-} = r^{\frac{7}{2}}\Bigg\{ a_0 \sum_{k=2}^{\infty} \frac{(-1)^{k+1} k_p^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-3} + \sum_{m=1}^{2} \sum_{k=1}^{\infty} a_m \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} \\ & + \sum_{m=3}^{\infty} \sum_{k=0}^{\infty} a_m \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} +{\mathrm{i}} b_0 \sum_{k=2}^{\infty} \frac{(-1)^k k_s^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-3} \\ & + {\mathrm{i}} \sum_{m=1}^{2} \sum_{k=1}^{\infty} b_m \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} + {\mathrm{i}} \sum_{m=3}^{\infty} \sum_{k=0}^{\infty} b_m \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3}\Bigg\}, \end{aligned} \end{equation} and \begin{equation}\label{eq:Ruv+} \begin{aligned} & R_{0,\Gamma_h^+} = r^2 \bigg\{ a_0 \mathrm e^{ {\mathrm{i}} \varphi_0} \sum_{k=1}^{\infty} \frac{(-1)^{k+1} k_p^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-1} + \sum_{m=1}^{\infty} \sum_{k=0}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-1} \\ & + {\mathrm{i}} b_0 \mathrm e^{ {\mathrm{i}} \varphi_0} \sum_{k=1}^{\infty} \frac{(-1)^{k} k_s^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-1} + {\mathrm{i}} \sum_{m=1}^{\infty} \sum_{k=0}^{\infty} b_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^{k} k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-1} \bigg\}, \end{aligned} \end{equation} and \begin{equation}\label{eq:Ruv-} \begin{aligned} & R_{0,\Gamma_h^-}= r^2 \bigg\{ a_0 \sum_{k=1}^{\infty} \frac{(-1)^{k+1} k_p^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-1} + \sum_{m=1}^{\infty} \sum_{k=0}^{\infty} a_m \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-1} \\ & + {\mathrm{i}} b_0 \sum_{k=1}^{\infty} \frac{(-1)^{k} k_s^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-1} + {\mathrm{i}} \sum_{m=1}^{\infty} \sum_{k=0}^{\infty} b_m \frac{(-1)^{k} k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-1} \bigg\}. \end{aligned} \end{equation} \end{lem} \begin{proof} Recall that \begin{equation}\label{eq:Jm ex} J_{m+1}(t)=\sum_{\ell=0}^{\infty} \frac{(-1)^\ell }{2^{2\ell+m+1} \ell! (m+\ell+1 )!} t^{2\ell+m+1} \end{equation} Since \begin{equation}\label{eq:orthogonal} \bmf{e}_1\cdot \bmf{e}_1=\bmf{e}_2\cdot \bmf{e}_2=0,\quad \bmf{e_1} \cdot \bmf{e}_2=2, \end{equation} using \eqref{eq:Jm ex}, \eqref{eq:Tu1} and \eqref{eq:v}, we obtain \eqref{eq:Tudotv}. Using Lemma \ref{lem:lemTv} and \eqref{eq:orthogonal} and in view of \eqref{eq:u} and \eqref{eq:v}, we can derive \eqref{eq:Tvu}. Recall that $\mathcal K$ is defined in \eqref{eq:K}. Since $\bmf{u}$, $ T_\nu \mathbf{u}$, $ \bmf{v} $ and $ T_\nu \mathbf{v}$ are analytic in $S_{2h}$, where $S_{2h}=\mathcal K \cap B_{2h}$ and $B_{2h}$ is a disk centered at the origin with the radius $2h$, from Lemma \ref{lem:kra}, we know that \eqref{eq:Tudotv} and \eqref{eq:Tvu} are convergent uniformly and absolutely in $r\in (0,h]$. \end{proof} \begin{lem}\label{lem:uv} Let $\bmf{u}$ and $\bmf{v}$ be respectively given by \eqref{eq:u} and \eqref{eq:v}. Then the following expansion \begin{equation}\label{eq:uv old} \begin{aligned} \mathbf{u} \cdot \mathbf{v} ={\mathrm e}^{s \sqrt{ r} \zeta(\varphi ) } \sum_{m=0}^{\infty} & \mathrm{e}^{{\mathrm{i}} (m+1) \varphi} \left[- k_p a_{m} J_{m+1}\left(k_{p}r\right) +{\mathrm{i}} k_s b_{m} J_{m+1}\left(k_{s} r\right) \right ] \end{aligned} \end{equation} convergences uniformly in $S_{2h}:=\mathcal{K}\cap B_{2h} $, where $\mathcal K$ is defined in \eqref{eq:K}. For $0\leqslant r\leqslant h $, it holds that \begin{equation}\label{eq:uv} \begin{aligned} &\left| \sum_{m=0}^{\infty} \mathrm{e}^{{\mathrm{i}} (m+1) \varphi} \left[- k_p a_{m} J_{m+1}\left(k_{p}r\right) +{\mathrm{i}} k_s b_{m} J_{m+1}\left(k_{s} r\right) \right ] \right| \leqslant \frac{r\left|k_p ^2 a_0 - {\mathrm{i}} k_s ^2 b_0 \right|}{2} + r^2\cdot S_1, \end{aligned} \end{equation} where \begin{equation*} \begin{aligned} S_1&= \ \left| \sum_{k=1}^{\infty} \frac{(-1)^k }{2^{2k+1} k! (k+1)!} \left(- k_p ^{2k+2} a_0 + {\mathrm{i}} k_s ^{2k+2} b_0 \right) h^{2k-1} \right| \\ &\quad +\sum_{m=1}^{\infty} \left| \sum_{k=0}^{\infty} \frac{(-1)^k }{2^{2k+m+1} k! (m+k+1 )!} \left( - k_p ^{2k+m+2} a_m + {\mathrm{i}} k_s ^{2k+m+2} b_m \right) h^{2k+m-1} \right| . \end{aligned} \end{equation*} Furthermore, if $a_0=b_0=\ldots=a_{\ell-1}=b_{\ell-1}=0$, we can conclude that \begin{equation}\label{eq:uv1} \begin{aligned} \left| \sum_{m=\ell}^{\infty} \mathrm{e}^{{\mathrm{i}} (m+1) \varphi} \left[- k_p a_{m} J_{m+1}\left(k_{p}r\right) +{\mathrm{i}} k_s b_{m} J_{m+1}\left(k_{s} r\right) \right ] \right| & \leqslant \frac{r^{\ell+1}\left|k_p ^{\ell+2} a_\ell - {\mathrm{i}} k_s ^{\ell+2} b_\ell \right|}{2^{\ell+1}(\ell+1)!} + r^{\ell+2}\cdot S_1(\ell), \end{aligned} \end{equation} where \begin{equation*} \begin{aligned} S_1(\ell)&= \ \left| \sum_{k=1}^{\infty} \frac{(-1)^k }{2^{2k+\ell+1} k! (k+\ell+1)!} \left( - k_p ^{2k+\ell+2} a_\ell + {\mathrm{i}} k_s ^{2k+\ell+2} b_\ell \right) h^{2k-1} \right| \\ &\quad +\sum_{m=\ell+1}^{\infty} \left| \sum_{k=0}^{\infty} \frac{(-1)^k }{2^{2k+m+1} k! (m+k+1 )!} \left( - k_p ^{2k+m+2} a_m + {\mathrm{i}} k_s ^{2k+m+2} b_m \right) h^{2k+m-\ell-1} \right| . \end{aligned} \end{equation*} \end{lem} \begin{proof} Since $\bmf{v}$ defined in \eqref{eq:v} is analytic in $S_{2h} $ and $\bmf{u}$ has the expansion \eqref{eq:u}, by noting \eqref{eq:orthogonal}, we have \eqref{eq:uv old}. Since $\Re(\zeta(\varphi ) )<0$ if $\varphi\in [0,\varphi_0 ]$, when $s$ is sufficient large we have ${\mathrm e}^{s \sqrt{ r} \zeta(\varphi ) } \leqslant 1$. Since \eqref{eq:uv old} is convergent at $\bmf{x}_0 \in \partial B_{2h} \cap {\mathcal K}$ from Lemma \ref{lem:kra} and noting \eqref{eq:Jm ex}, we can obtain \eqref{eq:uv}. \end{proof} \begin{lem}\label{lem:Tvu} Let $\bmf{u}$ be given in \eqref{eq:u} and $T_\nu \bmf{v}$ be defined in Lemma \ref{lem:lemTv}. If $a_0=b_0=\cdots=a_{\ell-1}=b_{\ell-1}=0$, then the following expansion{ \begin{equation}\label{eq:Tvu guina1new} \begin{split} & T_{\nu} \bmf{v}\big|_{ \Gamma_h^+ \backslash\{\bmf{0}\} } \cdot \bmf{u}\big|_{ \Gamma_h^+ \backslash\{\bmf{0}\} }= {\mathrm{i}} s \mu \zeta(\varphi_0 ) {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) }\bigg\{\frac{{\mathrm e}^{{\mathrm{i}} (\ell+1) \varphi_0}}{2^{\ell+1}(\ell+1)!}(-k_p^{\ell+2} a_\ell + {\mathrm{i}} k_s^{\ell+2} b_\ell) r^{\ell+\frac{1}{2}}\\ &+ \frac{{\mathrm e}^{{\mathrm{i}} (\ell+2) \varphi_0}}{2^{\ell+2}(\ell+2)!}(-k_p^{\ell+3} a_{\ell+1} + {\mathrm{i}} k_s^{\ell+3} b_{\ell+1}) r^{\ell+\frac{3}{2}} + \frac{{\mathrm e}^{{\mathrm{i}} (\ell+3) \varphi_0}}{2^{\ell+3}(\ell+3)!}(-k_p^{\ell+4} a_{\ell+2} + {\mathrm{i}} k_s^{\ell+4} b_{\ell+2}) r^{\ell+\frac{5}{2}}\\ &+ \frac{{\mathrm e}^{{\mathrm{i}} (\ell+) \varphi_0}}{2^{\ell+3}(\ell+2)!}(k_p^{\ell+4} a_{\ell} - {\mathrm{i}} k_s^{\ell+4} b_{\ell}) r^{\ell+\frac{5}{2}} + \hat{R}_{2,\Gamma_h^+}\bigg\},\\ & T_{\nu} \bmf{v}\big|_{ \Gamma_h^- \backslash\{\bmf{0}\} } \cdot \bmf{u}\big|_{ \Gamma_h^- \backslash\{\bmf{0}\} }= {\mathrm{i}} s \mu {\mathrm e}^{- s \sqrt{ r} }\bigg\{\frac{1}{2^{\ell+1}(\ell+1)!}(-k_p^{\ell+2} a_\ell + {\mathrm{i}} k_s^{\ell+2} b_\ell) r^{\ell+\frac{1}{2}}\\ & + \frac{1}{2^{\ell+2}(\ell+2)!}(-k_p^{\ell+3} a_{\ell+1} + {\mathrm{i}} k_s^{\ell+3} b_{\ell+1}) r^{\ell+\frac{3}{2}} + \frac{1}{2^{\ell+3}(\ell+3)!}(-k_p^{\ell+4} a_{\ell+2} + {\mathrm{i}} k_s^{\ell+4} b_{\ell+2}) r^{\ell+\frac{5}{2}}\\ & + \frac{1}{2^{\ell+3}(\ell+2)!}(k_p^{\ell+4} a_{\ell} - {\mathrm{i}} k_s^{\ell+4} b_{\ell}) r^{\ell+\frac{5}{2}} + \hat{R}_{2,\Gamma_h^-}\bigg\},\\ \end{split} \end{equation} } converge uniformly and absolutely with respect to $r\in (0,h]$, where { \begin{equation}\label{eq:RTvu++} \begin{aligned} & \hat{R}_{2,\Gamma_h^+} = r^{\ell+\frac{7}{2}}\Bigg\{ a_\ell \mathrm e^{ {\mathrm{i}} (\ell+1) \varphi_0} \sum_{k=2}^{\infty} \frac{(-1)^{k+1} k_p^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}r^{2k-3} + \sum_{m=\ell+1}^{\ell+2} \sum_{k=1}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0}\times \\ & \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-\ell-3} + \sum_{m=\ell+3}^{\infty} \sum_{k=0}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-\ell-3} \\ & +{\mathrm{i}} b_0 \mathrm e^{ {\mathrm{i}}(\ell+1) \varphi_0} \sum_{k=2}^{\infty} \frac{(-1)^k k_s^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}r^{2k-3} + {\mathrm{i}} \sum_{m=\ell+1}^{\ell+2} \sum_{k=1}^{\infty} b_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0}\times\\ & \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} + {\mathrm{i}} \sum_{m=\ell+3}^{\infty} \sum_{k=0}^{\infty} b_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-\ell-3}\Bigg\}, \end{aligned} \end{equation} } and { \begin{equation}\label{eq:RTvu--} \begin{aligned} & \hat{R}_{2,\Gamma_h^-} = r^{\ell+\frac{7}{2}}\Bigg\{ a_0 \sum_{k=2}^{\infty} \frac{(-1)^{k+1} k_p^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-3} + \sum_{m=\ell+1}^{\ell+2} \sum_{k=1}^{\infty} a_m \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} \\ & + \sum_{m=\ell+3}^{\infty} \sum_{k=0}^{\infty} a_m \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} +{\mathrm{i}} b_0 \sum_{k=2}^{\infty} \frac{(-1)^k k_s^{2k+2}}{2^{2k+1}k! (k+1)!}r^{2k-3} \\ & + {\mathrm{i}} \sum_{m=\ell+1}^{\ell+2} \sum_{k=1}^{\infty} b_m \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3} + {\mathrm{i}} \sum_{m=\ell+3}^{\infty} \sum_{k=0}^{\infty} b_m \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-3}\Bigg\}. \end{aligned} \end{equation} } Furthermore, we have the following expansions \begin{equation}\label{eq:Ruv guina} \begin{aligned} \bmf{u}\big|_{\Gamma_h^+} \cdot \bmf{v}\big|_{\Gamma_h^+} &={\mathrm e}^{s\sqrt{r} \zeta{(\varphi_0} )} \bigg\{\frac{{\mathrm e}^{{\mathrm{i}} (\ell+1) \varphi_0 }}{2^(\ell+1) (\ell+1)!} (-k_p^{\ell+2} a_\ell+ {\mathrm{i}} k_s^{\ell+2} b_\ell) r^{\ell+1} + \hat{R}_{0.\Gamma_h^+}\bigg\}, \\ \bmf{u}\big|_{\Gamma_h^-} \cdot \bmf{v}\big|_{\Gamma_h^-} &= {\mathrm e}^{-s\sqrt{r}} \bigg\{\frac{1}{2^(\ell+1) (\ell+1)!} (-k_p^{\ell+2} a_\ell+ {\mathrm{i}} k_s^{\ell+2} b_\ell) r^{\ell+1} + \hat{R}_{0.\Gamma_h^-}\bigg\}, \\ \end{aligned} \end{equation} which converge uniformly and absolutely in $r\in (0,h]$, where \begin{equation}\label{eq:Ruv++} \begin{aligned} & \hat{R}_{0,\Gamma_h^+} = r^{\ell+2} \bigg\{ a_\ell \mathrm e^{ {\mathrm{i}} (\ell+1)\varphi_0} \sum_{k=1}^{\infty} \frac{(-1)^{k+1} k_p^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}r^{2k-1} + \sum_{m=\ell+1}^{\infty} \sum_{k=0}^{\infty} a_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0}\times\\ & \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-\ell-1} + {\mathrm{i}} b_0 \mathrm e^{ {\mathrm{i}} (\ell+1)\varphi_0} \sum_{k=1}^{\infty} \frac{(-1)^{k} k_s^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}r^{2k-1} \\ & + {\mathrm{i}} \sum_{m=\ell+1}^{\infty} \sum_{k=0}^{\infty} b_m \mathrm e^{ {\mathrm{i}} (m+1)\varphi_0} \frac{(-1)^{k} k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-\ell-1} \bigg\}, \end{aligned} \end{equation} and \begin{equation}\label{eq:Ruv--} \begin{aligned} & \hat{R}_{0,\Gamma_h^-} = r^{\ell+2} \bigg\{ a_\ell \sum_{k=1}^{\infty} \frac{(-1)^{k+1} k_p^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}r^{2k-1} + \sum_{m=\ell+1}^{\infty} \sum_{k=0}^{\infty} a_m \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-\ell-1}\\ & + {\mathrm{i}} b_0 \sum_{k=1}^{\infty} \frac{(-1)^{k} k_s^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}r^{2k-1} + {\mathrm{i}} \sum_{m=\ell+1}^{\infty} \sum_{k=0}^{\infty} b_m \frac{(-1)^{k} k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}r^{2k+m-\ell-1} \bigg\}. \end{aligned} \end{equation} \end{lem} \begin{lem} Recall that $R_{1,\Gamma_h^+}$, $R_{1,\Gamma_h^-} $, $R_{2,\Gamma_h^+} $ , $R_{2,\Gamma_h^-} $ ,$R_{0,\Gamma_h^+}$ , $ R_{0,\Gamma_h^-}$, $\hat{R}_{2,\Gamma_h^+}$ , $ \hat{R}_{2,\Gamma_h^-}$ $ \hat{R}_{0,\Gamma_h^+}$ and $ \hat{R}_{0,\Gamma_h^-}$ are defined in \eqref{eq:RTuv+}, \eqref{eq:RTuv-}, \eqref{eq:RTvu+}, \eqref{eq:RTvu-}, \eqref{eq:Ruv+}, \eqref{eq:Ruv-}, \eqref{eq:RTvu++}, \eqref{eq:RTvu--}, \eqref{eq:Ruv++} and \eqref{eq:Ruv--} respectively. Then we have \begin{subequations} \begin{align} &\left |R_{1,\Gamma_h^+}\right| \leq r^3 S_{2},\quad \left |R_{1,\Gamma_h^-}\right| \leq r^3 S_{2},\label{eq:R1 s} \\ &\left |R_{2,\Gamma_h^+}\right| \leq r^{7/2} S_{3},\quad \left |R_{2,\Gamma_h^-}\right| \leq r^{7/2} S_{3},\label{eq:R2 s}\\ &\left |R_{0,\Gamma_h^+}\right| \leq r^2 S_{0},\quad \left |R_{0,\Gamma_h^-}\right| \leq r^2 S_{0},\label{eq:R3 s}\\ &\left |\hat{R}_{2,\Gamma_h^+}\right| \leq r^{\ell+7/2} \hat{S}_{3},\quad \left |\hat{R}_{2,\Gamma_h^-}\right| \leq r^{\ell+7/2} \hat{S}_{3},\label{eq:R4 s}\\ &\left |\hat{R}_{0,\Gamma_h^+}\right| \leq r^{\ell+2} \hat{S}_{0},\quad \left |\hat{R}_{0,\Gamma_h^-}\right| \leq r^{\ell+2} \hat{S}_{0},\label{eq:R5 s} \end{align} \end{subequations} where \begin{equation}\label{eq:RTuv+S} \begin{aligned} S_2 =& k_p^2 (\lambda+\mu)\bigg [|a_0| \sum_{k=2}^{\infty} \frac{ k_p^{2k}}{2^{2k}k! k!}h ^{2k-3} +\sum_{m=1}^2\sum_{k=1}^{\infty} |a_m| \frac{ k_p^{2k+m}}{2^{2k+m}k! (k+m)!} h^{2k+m-3}\bigg] \\ & + k_p^2 \mu |a_0| \sum_{k=1}^{\infty}\frac{ k_p^{2k+2}}{2^{2k+2}k! (k+2)!}h^{2k-1} +k_s^2 \mu |b_0| \sum_{k=1}^{\infty}\frac{ k_s^{2k+2}}{2^{2k+2}k! (k+2)!} h ^{2k-1} \\ & + k_p^2 \mu \sum_{m=1}^{2}\sum_{k=0}^{\infty} |a_m| \frac{ k_p^{2k+m+2}}{2^{2k+m+2}k! (k+m+2)!}h^{2k+m-1} \\ & + k_s^2 \mu \sum_{m=1}^{2}\sum_{k=0}^{\infty} |b_m| \frac{ k_s^{2k+m+2}}{2^{2k+m+2}k! (k+m+2)!}h^{2k+m-1}\bigg]\\ &+\sum_{m=3}^{\infty}\bigg|{\mathrm{i}} k_p^2 (\lambda+\mu) a_m \sum_{k=0}^{\infty} \frac{(-1)^k k_p^{2k+m}}{2^{2k+m}k! (k+m)!}h ^{2k+m-3}\\ &+ {\mathrm{i}} k_p^2 a_m\mu \sum_{k=0}^\infty \frac{(-1)^k k_p^{2k+m+2}}{2^{2k+m+2}k! (k+m+2)!}h^{2k+m-1} \\ &+ k_s^2 b_m \mu \sum_{k=0}^\infty \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+2}k! (k+m+2)!}h^{2k+m-1} \bigg|, \end{aligned} \end{equation} and \begin{equation}\label{eq:RTvu+S} \begin{aligned} S_3 = & |a_0| \sum_{k=2}^{\infty} \frac{ k_p^{2k+2}}{2^{2k+1}k! (k+1)!}h^{2k-3} +| b_0 | \sum_{k=2}^{\infty} \frac{ k_s^{2k+2}}{2^{2k+1}k! (k+1)!}h^{2k-3} \\ & + \sum_{m=1}^{2}\sum_{k=1}^{\infty} |a_m| \frac{ k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-3} \\ & + \sum_{m=1}^{2}\sum_{k=1}^{\infty} |b_m| \frac{ k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-3} \\ & \quad + \sum_{m=3}^{\infty} \bigg| a_m \sum_{k=0}^{\infty} \frac{(-1)^{k} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-3} \\ &\quad + {\mathrm{i}} b_m \sum_{k=0}^{\infty} \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-3}\bigg |, \end{aligned} \end{equation} and \begin{equation}\label{eq:uvS} \begin{aligned} S_0 = & |a_0| \sum_{k=1}^{\infty} \frac{ k_p^{2k+2}}{2^{2k+1}k! (k+1)!}h^{2k-1} +| b_0 | \sum_{k=1}^{\infty} \frac{ k_s^{2k+2}}{2^{2k+1}k! (k+1)!}h^{2k-1} \\ & + \sum_{m=1}^{\infty} \bigg| a_m \sum_{k=0}^{\infty} \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-1}\\ & + {\mathrm{i}} b_m \sum_{k=0}^{\infty} \frac{(-1)^{k} k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-1} \bigg |, \end{aligned} \end{equation} and \begin{equation}\label{eq:RTvu+Shat} \begin{aligned} \hat{S}_3 = & |a_\ell| \sum_{k=2}^{\infty} \frac{ k_p^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}h^{2k-3} +| b_\ell | \sum_{k=2}^{\infty} \frac{ k_s^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}h^{2k-3} \\ & + \sum_{m=\ell+1}^{\ell+2}\sum_{k=1}^{\infty} |a_m| \frac{ k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-\ell-3} \\ & + \sum_{m=\ell+1}^{\ell+2}\sum_{k=1}^{\infty} |b_m| \frac{ k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-\ell-3} \\ & \quad + \sum_{m=\ell+3}^{\infty} \bigg| a_m \sum_{k=0}^{\infty} \frac{(-1)^{k} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-\ell-3} \\ &\quad + {\mathrm{i}} b_m \sum_{k=0}^{\infty} \frac{(-1)^k k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-\ell-3}\bigg |, \end{aligned} \end{equation} \begin{equation}\label{eq:uvShat} \begin{aligned} \hat{S}_0 = & |a_\ell| \sum_{k=1}^{\infty} \frac{ k_p^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}h^{2k-1} +| b_\ell | \sum_{k=1}^{\infty} \frac{ k_s^{2k+\ell+2}}{2^{2k+\ell+1}k! (k+\ell+1)!}h^{2k-1} \\ & + \sum_{m=\ell+1}^{\infty} \bigg| a_m \sum_{k=0}^{\infty} \frac{(-1)^{k+1} k_p^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-\ell-1}\\ & + {\mathrm{i}} b_m \sum_{k=0}^{\infty} \frac{(-1)^{k} k_s^{2k+m+2}}{2^{2k+m+1}k! (k+m+1)!}h^{2k+m-\ell-1} \bigg |, \end{aligned} \end{equation} \end{lem} \begin{proof} Since $ T_\nu \mathbf{u}\Big |_{\Gamma^\pm_h} \cdot \bmf{v} \Big |_{\Gamma^\pm_h} $ are analytic on ${\Gamma^\pm_h}$, from \eqref{eq:Tudotv}, by root test and Lemma \ref{lem:kra}, for $r\in (0,h)$, together with straightforward though tedious calculations, one can prove \eqref{eq:R1 s}. \eqref{eq:R2 s}, \eqref{eq:R3 s}, \eqref{eq:R4 s} and \eqref{eq:R5 s} can be proved in a similar way, and we skip the details. \end{proof} Recall that the open sector $\mathcal K$ and $\Gamma_h^\pm$ are defined in \eqref{eq:K} and \eqref{eq:gamma_pm}, respectively. For $\varepsilon \in {\mathbb R}_{+}$ satisfying $\varepsilon<h$, let \begin{equation}\label{eq:Seps} S_\varepsilon={\mathcal K} \cap B_\varepsilon, \quad \Gamma_{(0,\varepsilon )}^\pm =\Gamma_h^\pm \cap B_\varepsilon,\quad \Lambda_\varepsilon= S_\varepsilon \cap \partial B_\varepsilon. \end{equation} \begin{lem}\label{lem:25 int 0} Let $\bmf{u}$ be a Lam\'e eigenfunction to \eqref{eq:lame} and $\bmf{v}$ be defined in \eqref{eq:v}. Recall that $\Gamma_{(0,\varepsilon )}^\pm$ and $\Lambda_\varepsilon$ are defined in \eqref{eq:Seps}. Then \begin{subequations} \begin{align} &\lim_{\varepsilon\rightarrow 0^+} \int_{\Gamma_{(0,\varepsilon )}^+} T_\nu \bmf{v} \cdot \bmf{u} \mathrm{d} \sigma= \lim_{\varepsilon\rightarrow 0^+} \int_{\Gamma_{(0,\varepsilon )}^-} T_\nu \bmf{v} \cdot \bmf{u} \mathrm{d} \sigma=0,\label{eq:27a} \\ &\lim_{\varepsilon\rightarrow 0^+} \int_{\Gamma_{(0,\varepsilon )}^+} T_\nu \bmf{u} \cdot \bmf{v} \mathrm{d} \sigma= \lim_{\varepsilon\rightarrow 0^+} \int_{\Gamma_{(0,\varepsilon )}^-} T_\nu \bmf{u} \cdot \bmf{v} \mathrm{d} \sigma=0,\label{eq:27b} \\ &\lim_{\varepsilon\rightarrow 0^+} \int_{\Gamma_{(0,\varepsilon )}^+} \bmf{u} \cdot \bmf{v} \mathrm{d} \sigma= \lim_{\varepsilon\rightarrow 0^+} \int_{\Gamma_{(0,\varepsilon )}^-} \bmf{u} \cdot \bmf{v} \mathrm{d} \sigma=0,\label{eq:27c} \end{align} \end{subequations} \end{lem} \begin{proof} From \eqref{eq:Tvu}, it is easy to see that $$ \lim_{\bmf{x} \rightarrow \bmf{0}\atop \bmf{x} \in \Gamma_h^\pm } T_{\nu} \bmf{v}\big|_{ \Gamma_h^\pm \backslash\{\bmf{0}\} } \cdot \bmf{u}\big|_{ \Gamma_h^\pm \backslash\{\bmf{0}\} }=0. $$ Therefore the function $ T_{\nu} \bmf{v}\big|_{ \Gamma_h^\pm \backslash\{\bmf{0}\} } \cdot \bmf{u}\big|_{ \Gamma_h^\pm \backslash\{\bmf{0}\} }$ is continuous at the origin. Hence by the dominant convergent theorem, we can prove \eqref{eq:27a}. Similarly, from \eqref{eq:Tudotv} and \eqref{eq:Tvu}, we know that \eqref{eq:27b} and \eqref{eq:27c} hold via the dominant convergent theorem. \end{proof} The following lemma gives the estimates of the integrals with respect to the CGO solution $\bmf{v}$ \eqref{eq:v} on an open sector and an arbitrary arc, which will be used in the subsequent study. \begin{lem}\cite[Proposition 3.1]{EBL}\label{lem:CGO exp} Let $\bmf{v} : \mathbb{R}^{2} \rightarrow \mathbb{C}^{2}$ be defined by \eqref{eq:v} and \begin{equation}\label{eq:K phi} \mathcal{K}_{\varphi_m,\varphi_M} =\left\{\bmf{x} =(x_1,x_2) \in \mathbb{R}^{2} ~|~\bmf{ x} \neq \bmf{0} , \quad \varphi_{m}<\arg \left(x_{1}+\mathrm{i} x_{2}\right)<\varphi_{M}\right\} \end{equation} for given angles $-\pi<\varphi_{m}<\varphi_{M}<\pi$. Then it holds that \begin{equation}\label{eq:int 29 n} \int_{{\mathcal K}_{\varphi_m,\varphi_M }} v_1 \left(\bmf{x}\right) \mathrm{d} \bmf{x} = 6 {\mathrm{i}} \left(\mathrm{e}^{-2\varphi_M {\mathrm{i}}}-\mathrm{e}^{-2\varphi_m {\mathrm{i}}}\right)s^{-4}. \end{equation} In addition for $\alpha$,$h>0$ and $j\in\left\{1,2\right\}$, we have the upper bounds \begin{subequations} \begin{align} \int_{{\mathcal K}_{\varphi_m,\varphi_M }} \left|v_j \left(\bmf{x}\right)\right| \left|\bmf{x}\right|^\alpha \mathrm{d} \bmf{x}& \leqslant \frac{2\left(\varphi_M-\varphi_m\right)\Gamma\left(2\alpha+4\right)}{\delta_{{\mathcal K}_{\varphi_m,\varphi_M }}^{2\alpha+4}}s^{-2\alpha-4},\label{eq:volume n} \\ \int_{{\mathcal K}_{\varphi_m,\varphi_M } \backslash B_h}\left|v_{j}(\bmf{x})\right| \mathrm{d} \bmf{x} &\leqslant \frac{6\left(\varphi_{M}-\varphi _{m}\right)}{\delta_{\mathcal{K}_{\varphi_m, \varphi_M }}^{4}} s^{-4} \mathrm{e}^{-\delta_{\mathcal{K}_{\varphi_m, \varphi_M } } s \sqrt{h} / 2},\label{eq:volume2 n} \end{align} \end{subequations} where $\delta_{{\mathcal K}_{\varphi_m,\varphi_M }}=\min_{\varphi_m<\varphi<\varphi_M} \cos(\varphi/2)$ is a positive constant depending ${\mathcal K}_{\varphi_m,\varphi_M }$. Furthermore, assume that $\bmf{u} \in H^{2}\left(\mathcal{K}_{\varphi_m, \varphi_M } \cap B_l \right)$ where $B_l=B(\bmf{0}, l)$ for some $l>0$. The it holds that \begin{equation}\label{eq:arc n} \begin{aligned} & \left|\int_{\mathcal{K}_{\varphi_m,\varphi_M} \cap \partial B_l }\left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma\right| \\ & \leq \mathcal{C}_{{\mathcal K}_{{\varphi_m,\varphi_M }},B_l,\mu,\lambda}\|\bmf{u}\|_{H^2\left( {{\mathcal K}_{\varphi_m,\varphi_M }} \cap B_l \right)}\left(1+s\right)\exp\left(-\delta_{{\mathcal K}_{\varphi_m,\varphi_M }} s \sqrt{h}\right), \end{aligned} \end{equation} where $\mathcal{C}_{{\mathcal K}_{{\varphi_m,\varphi_M }},B_l,\mu,\lambda}$ is a positive constant depending on $B_l, \lambda, \mu $ and ${\mathcal K}_{\varphi_m,\varphi_M }$. \end{lem} \begin{rem} Recall that $\mathcal K$ is defined in \eqref{eq:K}. Then ${\mathcal K}_{\varphi_m, \varphi_M }$ defined in \eqref{eq:K phi} degenerates to $\mathcal K$ whenever $\varphi_M:=\varphi_0$, $\varphi_m:=0$. In this situation, the constant $\delta_{{\mathcal K}_{\varphi_m,\varphi_M }}$ given in \eqref{eq:volume n} and \eqref{eq:volume2 n} is denoted by $\delta_{{\mathcal K}}$ in the remainder of this paper. Indeed, setting $\varphi_M:=\varphi_0$ and $\varphi_m:=0$, from \eqref{eq:int 29 n}, \eqref{eq:volume n}, \eqref{eq:volume2 n} and \eqref{eq:arc n}, we have \begin{subequations} \begin{align} &\int_{{\mathcal K} } {\mathrm e}^{s \sqrt{ r} \zeta(\varphi ) } r \mathrm{d} {r} \mathrm{d} \varphi = 6 {\mathrm{i}} \left(\mathrm{e}^{-2\varphi_0 {\mathrm{i}}}-1\right)s^{-4},\\ &\int_{{\mathcal K} } {\mathrm e}^{s \sqrt{ r} \Re(\zeta(\varphi )) } r^{\alpha+1} \mathrm{d} r \mathrm{d} \varphi \leqslant \frac{2\varphi_0\Gamma\left(2\alpha+4\right)}{\delta_{{\mathcal K}}^{2\alpha+4}}s^{-2\alpha-4},\label{eq:volume} \\ & \left|\int_{\Lambda_h }\left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma\right| \leqslant \mathcal{C}_{{\mathcal K },B_h,\mu,\lambda}\|\bmf{u}\|_{H^2\left( {{\mathcal K}} \cap B_h \right)}\left(1+s\right){\mathrm e}^{-\delta_{{\mathcal K} } s \sqrt{h}} , \label{eq:arc} \end{align} \end{subequations} where $\delta_{{\mathcal K}}=\min_{0<\varphi<\varphi_0} \cos(\varphi/2)$ and $\Lambda_h$ is defined in \eqref{eq:Lambda_h}. \end{rem} We next derive several crucial integral identities. \begin{lem}\label{lem:part exp} Suppose $D$ is a bounded Lipschitz domain in $\mathbb R^2$ and $\bmf{u}$, $\bmf{v}$ are $H_{loc}^2(\mathbb{R}^2)$ functions. Let $\bmf{v}$ be the CGO solution defined in \eqref{eq:v}, which satisfies $\mathcal{L}\bmf v=\bmf{0}$. Then \begin{equation}\label{eq:CGO1} \int_D \left(\mathcal{L} \bmf{u}\right) \cdot \bmf{v} \mathrm{d} x=\int_{\partial D} \left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma. \end{equation} If $\bmf{u}$ is a Lam\'e eigenfunction to \eqref{eq:lame}, then the following integral identity holds \begin{equation}\label{eq:CGO2} I_3=I_1^+ + I_1^-+I_2, \end{equation} where \begin{subequations} \begin{align} I_1^\pm&=\int_{\Gamma_h^\pm } \left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma, \label{eq:CGO6}\\ I_2&=\int_{\Lambda_h} \left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma, \label{eq:CGO6a}\\ I_3 =-\kappa \int_{S_h} \bmf{u} \cdot \bmf{v} \mathrm{d} \bmf{x}. \label{eq:I3 int} \end{align} \end{subequations} Furthermore, we have \begin{equation}\label{eq:I2I4} \begin{split} \left|I_2\right|& \leqslant \mathcal{C}_{{\mathcal K },B_h,\mu,\lambda}\|\bmf{u}\|_{H^2\left( {{\mathcal K}} \cap B_h \right)}\left(1+s\right){\mathrm e}^{-\delta_{{\mathcal K} } s \sqrt{h}} \end{split} \end{equation} which is exponentially decays as $s\rightarrow +\infty$. Here $\delta_{\mathcal{K}} $ is a positive constant defined in \eqref{eq:volume}. \end{lem} \begin{proof} We first prove \eqref{eq:CGO2}. Recall that $S_\varepsilon$ is defined in \eqref{eq:Seps}. Since $\bmf{u}$ and $\bmf{v}$ are $H^2( S_h\backslash S_\varepsilon )$, from \eqref{eq:CGO1}, we have \begin{equation}\label{eq:CGO3} \int_{S_h \backslash{S_\varepsilon}} \left(\mathcal{L} \bmf{u}\right) \cdot \bmf{v} \mathrm{d} x=\int_{\partial ({S_h}\backslash{S_\varepsilon})} \left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma. \end{equation} Since $\mathbf{u}$ is a Lam\'e eigenfunction to \eqref{eq:lame}, then $\mathcal{L} (\bmf{u})=-\kappa \bmf{u} \text { in } \Omega \subset \mathbb{R}^{2}$. Moreover, $\bmf{u}$ and $\bmf{v}$ are $H_{loc}^2(\mathbb{R}^2)$ functions, it yields that \begin{equation}\label{eq:CGO4} \lim_{\varepsilon\rightarrow 0^+ } \int_{S_\varepsilon} \left(\mathcal{L} \bmf{u}\right) \cdot \bmf{v} \mathrm{d} x=-\kappa \lim_{\varepsilon\rightarrow 0^+ } \int_{S_\varepsilon} \bmf{u} \cdot \bmf{v} \mathrm{d} x =0. \end{equation} Recall that $\Lambda_\varepsilon$ is defined in \eqref{eq:Seps}. It is easy to see that \begin{equation}\label{eq:CGO5} \int_{\partial ({S_h}\backslash{S_\varepsilon})} \left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma=I_1^+ + I_1^- + I_2 -I_\varepsilon^+- I_\varepsilon^- + I_{\Lambda_\varepsilon}. \end{equation} where \begin{subequations} \begin{align} I_\varepsilon^\pm&=\int_{\Gamma_{(0,\varepsilon)}^\pm } \left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma, \label{eq:CGO7} \\ I_{\Lambda_\varepsilon}&=\int_{\Lambda_\varepsilon} \left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma. \label{eq:CGO9} \end{align} \end{subequations} Here the line segements $\Gamma_{(0,\varepsilon)}^\pm $ are defined in \eqref{eq:Seps}. From \eqref{eq:27a} and \eqref{eq:27b}, we know that \begin{equation}\label{eq:220 eps} \lim_{\varepsilon \rightarrow 0^+ } I_\varepsilon^\pm =0. \end{equation} By setting $h=\varepsilon$ in \eqref{eq:arc}, it is readily seen that \begin{equation}\label{eq:221 lam} \lim_{\varepsilon \rightarrow 0^+ } I_{\Lambda_\varepsilon} =0. \end{equation} In \eqref{eq:CGO3}, letting $\varepsilon \rightarrow 0^+$, using \eqref{eq:CGO4}, \eqref{eq:220 eps} and \eqref{eq:221 lam}, we obtain that \begin{equation*} I_3=-\kappa \int_{S_h} \bmf{u} \cdot \bmf{v} \mathrm{d} x =\int_{\partial S_h} \left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma, \end{equation*} from which we prove \eqref{eq:CGO2}. The proof is complete. \end{proof} \begin{lem}\label{lem:the order of r exp2} For a given $\zeta(\varphi)=-\mathrm{e}^{{\mathrm{i}} \frac{\varphi}{2}} \in \mathbb{C}$ and $\ell=\frac{m}{2},m=0,1,2,...$. we have \begin{equation*} \begin{aligned} \int_{0}^{h}r^\ell \mathrm{e}^{s \sqrt{r} \zeta(\varphi)}\mathrm{d} r= & \frac{2}{s^{2 \ell+2}}\left(\frac{\left(-1\right)^{2\ell}\left(2 \ell+1\right)!}{\zeta(\varphi)^{2l+2}}\right.\\ & + \mathrm{e}^{s \sqrt{h} \zeta(\varphi)}\sum_{j=0}^{2 \ell+1}\frac{\left(-1\right)^j\left(2 \ell+1\right)!}{\left(2 \ell+1-j\right)!\zeta(\varphi)^{j+1}} \left(s^2 h\right)^\frac{\left(2\ell+1-j\right)}{2}\bigg),\\ \int_{0}^{h}r^\ell \mathrm{e}^{s \sqrt{r} \Re(\zeta(\varphi)) }\mathrm{d} r= & \frac{2}{s^{2 \ell+2}}\left(\frac{\left(-1\right)^{2\ell}\left(2 \ell+1\right)!}{\Re( \zeta(\varphi)) ^{2l+2}}\right.\\ & + \mathrm{e}^{s \sqrt{h} \Re( \zeta(\varphi)) }\sum_{j=0}^{2 \ell+1}\frac{\left(-1\right)^j\left(2 \ell+1\right)!}{\left(2 \ell+1-j\right)! \Re( \zeta(\varphi))^{j+1}} \left(s^2 h\right)^\frac{\left(2\ell+1-j\right)}{2}\bigg),\\ \end{aligned} \end{equation*} Furthermore, if $\mathfrak{R}\left(\zeta\left(\varphi\right)\right)<0$, we have the following asymptotic expansion: \begin{subequations} \begin{align} \int_{0}^{h}r^\ell \mathrm{e}^{s \sqrt{r} \zeta(\varphi)}\mathrm{d} r&=\frac{2}{s^{2 \ell+2}}\frac{\left(-1\right)^{2\ell}\left(2 \ell+1\right)!}{\zeta(\varphi)^{2 \ell+2}}+{\mathcal O} \left(s^{-\frac{1}{2}} \mathrm{e}^{\sqrt{sh} \zeta(\varphi)}\right),\label{eq:268a int} \\ \int_{0}^{h}r^\ell \mathrm{e}^{s \sqrt{r} \Re(\zeta(\varphi))}\mathrm{d} r&=\frac{2}{s^{2 \ell+2}}\frac{\left(-1\right)^{2\ell}\left(2 \ell+1\right)!}{\Re( \zeta(\varphi))^{2 \ell+2}}+{\mathcal O} \left(s^{-\frac{1}{2}} \mathrm{e}^{\sqrt{sh} \Re( \zeta(\varphi))}\right). \label{eq:268b int} \end{align} \end{subequations} as s $\rightarrow +\infty$, where $\ell=\frac{m}{2},m=0,1,2, \ldots$. \end{lem} \begin{proof} By induction and direct verifications, one can show the lemma. We skip the details. \end{proof} In the next lemma, we derive an upper bound for the integral $I_3$ defined in \eqref{eq:I3 int}. \begin{lem}\label{lem:I3 exp} Recall that the Lam\'e eigenfunction $\mathbf{u}$ to \eqref{eq:lame} has the radial wave expression \eqref{eq:u} at the origin and $I_3$ is defined by \eqref{eq:I3 int}, then one has \begin{equation}\label{eq:I3} |I_3| \leqslant \left|k_p ^2 a_0 - {\mathrm{i}} k_s ^2 b_0 \right| \frac{\kappa \varphi_0\Gamma\left(6\right) }{\delta_{{\mathcal K}}^{6} } s^{-6} + \frac{2\kappa \varphi_0\Gamma\left(8\right) S_1 }{\delta_{{\mathcal K}}^{8} } s^{-8}, \end{equation} as $s\rightarrow +\infty$, where $\delta_{\mathcal{K}} $ is a positive constant defined in \eqref{eq:volume} and $S_1$ is defined in \eqref{eq:uv}. Recall that $S_1(\ell) $ is defined in \eqref{eq:uv1}. Furthermore, if $a_0=b_0=\cdots=a_{\ell-1}=b_{\ell-1}=0$, \begin{equation}\label{eq:I3+} |I_3| \leqslant \left|k_p ^{\ell+2} a_{}\ell - {\mathrm{i}} k_s ^{\ell+2} b_{\ell} \right| \frac{ \kappa \varphi_0\Gamma(2 \ell+6) }{2^{\ell}(\ell+1)!\delta_{{\mathcal K}}^{2 \ell+6} } s^{-2 \ell-6} + \frac{2 \kappa \varphi_0\Gamma(2 \ell+8) S_1(\ell) }{\delta_{{\mathcal K}}^{8} } s^{-2\ell-8}, \end{equation} as $s\rightarrow +\infty$. \end{lem} \begin{proof} Submitting \eqref{eq:uv old} into \eqref{eq:I3 int}, using \eqref{eq:volume} and \eqref{eq:uv}, we derive that \begin{equation*} \begin{aligned} |I_3| &\leqslant \kappa \int_{\mathcal K} |\bmf{u} \cdot \bmf{v}| \mathrm{d} \bmf{x} \\ &\leqslant \kappa \frac{\left|k_p ^2 a_0 - {\mathrm{i}} k_s ^2 b_0 \right| }{4} \int_{\mathcal K} r^2 {\mathrm e}^{s \sqrt{ r} \Re(\zeta(\varphi )) } \mathrm{d} r \mathrm{d} \varphi + \kappa S_1 \int_{\mathcal K} r^3 {\mathrm e}^{s \sqrt{ r} \Re(\zeta(\varphi )) } \mathrm{d} r % \end{aligned} \end{equation*} from which we complete the proof of \eqref{eq:I3}. By virtue of \eqref{eq:uv1}, \eqref{eq:I3+} can be proved in a similar way. \end{proof} \section{Generalized Holmgren's principle with the presence of singular lines}\label{sect:3} In this section, we prove that if a generic Lam\'e eigenfunction $\bmf{u}$ to \eqref{eq:lame} possesses a singular line $\Gamma_h$ in $\Omega$ as defined in Definition~\ref{def:generalized line}, then $\bmf{u}$ is identically zero. According to our discussion made at the beginning of Section~\ref{sect:2}, without loss of generality, we can assume that $\Gamma_h$ is $\Gamma_h^-$ as defined in \eqref{eq:gamma_pm}. Moreover, we can assume that the point $\bmf{x}_0$ involved in Definition~\ref{def:generalized line} is the origin, namely $\bmf{x}_0=\bmf{0}$. It is clear that the unit normal vectors $\nu $ to $\Gamma_h^-$ is $(0,\pm 1)^\top$. In this paper we choose $(0,-1)^\top$ as the unit normal vector $\nu $ to $\Gamma_h^-$. In such a case, the following conditions involved in Definition~\ref{def:generalized line} \begin{equation}\label{eq:31} \bmf{u}(\bmf{x}_0 )=\bmf{0} \mbox{ and/or } \boldsymbol{\tau} \cdot \partial_{\nu} \bmf{u} |_{\bmf{x}={\bmf{x}}_0 }=0, \end{equation} turn into $$ \bmf{u}(\bmf{0} )=\bmf{0} \mbox{ and/or } \partial_2 u_1(\bmf{0} )=0. $$ \subsection{The case with a singular rigid line} \begin{lem}\label{lem:rigid} Let $\mathbf{u}$ be a generalized Lam\'e eigenfunction to \eqref{eq:lame} with the radial wave expansion given in \eqref{eq:u} around the origin. Suppose there exists $\Gamma_h^-\in {\mathcal R}_\Omega^{\kappa} $. Then one has \begin{equation}\label{eq:1} \bigg\{ \begin{array}{l} k_p a_1 + {\mathrm{i}} k_s b_1=0,\\ k_p^3a_1-{\mathrm{i}} k_s^3 b_1=0, \end{array} \end{equation} and \begin{equation}\label{eq:2} \bigg\{ \begin{array}{l} k_p ^2 a_0+ {\mathrm{i}} k_s^2 b_0- k_p^2 a_2- {\mathrm{i}} k_s ^2 b_2=0,\\ k_p ^2 a_0- {\mathrm{i}} k_s^2 b_0=0. \end{array} \end{equation} Moreover, it holds that \begin{equation}\label{eq:a1b1 lem} a_1=b_1=0. \end{equation} Furthermore, suppose that \begin{equation}\label{eq:lem cond} a_\ell=b_\ell=0 \end{equation} where $\ell=0,\ldots,m-1$ and $m\in {\mathbb N}$ with $m\geq 2 $, then $$ a_\ell=b_\ell=0, \quad \forall \ell \in \mathbb N \cup \{0\}. $$ \end{lem} \begin{proof} Since $\Gamma_h^-$ is a rigid line of $\bmf{u}$, then $\bmf{u}\big|_{\Gamma_h^- }=\bmf{0}$. Therefore from \eqref{eq:u} and by noting $\varphi=0$ on $\Gamma_h^-$, we have for $0 \leqslant r\leqslant h$ that \begin{equation}\label{eq:u=0} \begin{aligned} \mathbf{0}= \sum_{m=0}^{\infty} & \left\{ \frac{k_p}{2} a_{m} \left\{J_{m-1}\left(k_{p} r\right) \mathbf{e}_1 -J_{m+1}\left(k_{p}r\right)\mathrm{e}^{{\mathrm{i}} \varphi}\mathbf{e}_2 \right\}\right. + \frac{{\mathrm{i}} k_s}{2} b_{m} \left\{J_{m-1}\left(k_{s} r\right) \mathbf{e}_1 +J_{m+1}\left(k_{s} r\right) \mathbf{e}_2 \right\} \bigg\}. \end{aligned} \end{equation} From \eqref{eq:J-1}, we know that $J_{-1}(k_pr)=-J_{1}(k_pr) $ and $J_{-1}(k_sr)=-J_{1}(k_sr) $. Using Lemma \ref{lem:co exp}, comparing the coefficients of the term $r^0$ in both sides of \eqref{eq:u=0}, we obtain \eqref{eq:1}. Similarly, from Lemma \ref{lem:co exp}, we compare the coefficients of the term $r^1$ in both sides of \eqref{eq:u=0}, and obtain that \begin{equation}\nonumber \left(-k_p^2 a_0-{\mathrm{i}} k_s^2 b_0+k_p^2 a_2+{\mathrm{i}} k_s^2 b_2\right)\mathbf{e}_1 -\left(k_p^2 a_0-{\mathrm{i}} k_s^2 b_0\right)\mathbf{e}_2=\bmf{0}. \end{equation} Since $\mathbf{e}_1$ and $\mathbf{e}_2$ are linearly independent, we can obtain \eqref{eq:2}. Similarly, comparing the coefficients of the term $r^2$ in both sides of \eqref{eq:u=0}, we have \begin{equation}\nonumber \left(-2k_p^3 a_1-2 {\mathrm{i}} k_s^3 b_1+k_p^3 a_3+{\mathrm{i}} k_s^3 b_3\right)\mathbf{e}_1 -\left(k_p^3 a_1-{\mathrm{i}} k_s^3 b_1\right)\mathbf{e}_2=\bmf{0}, \end{equation} which implies that the second equation of \eqref{eq:1} holds. The determinant of the coefficient matrix of \eqref{eq:1} is $$ \left| \begin{matrix} k_p & {\mathrm{i}} k_s\\ k_p^3 & -{\mathrm{i}} k_s^3 \end{matrix} \right|=-{\mathrm{i}} k_pk_s (k_p^2+k_s^2) \neq 0, $$ then we know that \eqref{eq:a1b1 lem} holds. Substituting \eqref{eq:lem cond} into \eqref{eq:u=0}, it yields that \begin{equation}\label{eq:u=0 new} \begin{aligned} \mathbf{0}= \sum_{\ell=m}^{\infty} & \left\{ \frac{k_p}{2} a_{\ell } \left\{J_{\ell -1}\left(k_{p} r\right) \mathbf{e}_1 -J_{\ell +1}\left(k_{p}r\right)\mathrm{e}^{{\mathrm{i}} \varphi}\mathbf{e}_2 \right\}\right.\\ & + \frac{{\mathrm{i}} k_s}{2} b_{\ell } \left\{J_{\ell -1}\left(k_{s} r\right) \mathbf{e}_1 +J_{\ell +1}\left(k_{s} r\right) \mathbf{e}_2 \right\} \bigg\},\quad 0 \leqslant r\leqslant h. \end{aligned} \end{equation} Therefore the lowest order of $r$ in the right hand side of \eqref{eq:u=0 new} is $m-1$. Hence comparing the coefficients of the term $r^{m-1}$ in both sides of \eqref{eq:u=0 new}, we obtain that \begin{equation}\label{eq:6} k_p ^{m } a_m+ {\mathrm{i}} k_s^{m} b_m=0. \end{equation} Comparing the coefficients of the term $r^{m+1}$ in both sides of \eqref{eq:u=0 new}, which are related to $J_{m-1}(k_p r)$, $J_{m-1}(k_s r)$, $J_{m+1}(k_p r)$ and $J_{m+1}(k_s r)$, one has \begin{equation}\nonumber \begin{aligned} & \left(-\left(m+1\right) k_p^{m+2} a_m-{\mathrm{i}} (m+1) k_s^{m+2} b_m+k_p^{m+2} a_{m+2}+{\mathrm{i}} k_s^{m+2} b_{m+2} \right) \mathbf{e}_1\\ & - \left(k_p^{m+2} \mathrm{e}^{2 {\mathrm{i}}\varphi_0} a_m-{\mathrm{i}} k_s^{m+2} b_m\right)\mathbf{e}_2=\bmf{0}. \end{aligned} \end{equation} Since $\mathbf{e}_1$ and $\mathbf{e}_2$ are linear independent, it yields that \begin{equation}\label{eq:7} \left\{ \begin{array}{l} -\left(m+1\right) k_p^{m+2} a_m-{\mathrm{i}} (m+1) k_s^{m+2} b_m+k_p^{m+2} a_{m+2}+{\mathrm{i}} k_s^{m+2} b_{m+2} =0,\\ k_p^{m+2} a_m-{\mathrm{i}} k_s^{m+2} b_m=0,\\ \end{array} \right. \end{equation} Combining \eqref{eq:6} and \eqref{eq:7}, we obtain that \begin{equation}\notag \left\{ \begin{array}{l} k_p ^m a_m+ {\mathrm{i}} k_s^m b_m=0,\\ k_p^{m+2} a_m-{\mathrm{i}} k_s^{m+2} b_m=0. \end{array} \right. \end{equation} Since \begin{equation}\notag \left| \begin{array}{cc} k_p^m & {\mathrm{i}} k_s^m\\ k_p^{m+2} & -{\mathrm{i}} k_s^{m+2}\\ \end{array} \right|=-{\mathrm{i}} k_p^m k_s^m \left(k_p^2 +k_s^2\right)\neq 0, \end{equation} one readily has that $a_m=b_m=0$. In an inductive manner, we can prove that $a_\ell=b_\ell=0$ for $\ell=m+1,\ldots$. The proof is complete. \end{proof} \begin{lem}\label{lem:34 three conds} Let $\mathbf{u}=(u_\ell)_{\ell=1}^2$ be a Lam\'e eigenfunction to \eqref{eq:lame} with the radial wave expansion \eqref{eq:u} around the origin. If $\bmf{u}(\bmf{0})=\bmf{0}$, then \begin{equation}\label{eq:u0} k_p a_1 + {\mathrm{i}} k_s b_1=0. \end{equation} Recall that $u_1$ has the expansion \eqref{eq:u comp}. If $\partial_2 u_1 ({\mathbf 0})=0$, then \begin{equation}\label{eq:u1} -2 k_s^2 b_0+{\mathrm{i}} k_p^2 a_2 -k_s^2 b_2=0. \end{equation} \end{lem} \begin{proof} Since $\bmf{u}(\bmf{0})=\bmf{0}$, substituting $r=0$ in \eqref{eq:u} we can prove \eqref{eq:u0}. From \eqref{eq:u3 par} in the Appendix, it is easy to know that \begin{equation}\notag \partial_1 u_1=\frac{\partial u_1}{\partial r} \end{equation} at $\varphi=0$. Substituting \eqref{eq:J2} into \eqref{eq:u1 par}, we can obtain that \begin{equation}\label{eq:u3 new} \begin{aligned} \frac{1}{r}\frac{\partial u_1}{\partial \varphi}\Big|_{\varphi=0}=&\sum_{m=0}^\infty \bigg\{ {\mathrm{i}} k_p^2 a_m J_{m-2}(k_p r)-{\mathrm{i}} k_p^2 a_m J_{m+2}(k_p r)- k_s^2 b_m J_{m-2}(k_s r)\\ &\quad -2 k_s^2 b_m J_m(k_s r)-k_s^2 b_m J_{m+2}(k_s r)\bigg\}. \end{aligned} \end{equation} Substituting \eqref{eq:u3 new} into \eqref{eq:u3 par} and evaluating the resulting equality of \eqref{eq:u3 par} at $r=0$ and $\varphi=0$ since $\partial_2 u_1 (\bmf{0})=0$, one can prove \eqref{eq:u1}. \end{proof} \begin{thm}\label{Thm:31 singular rigid} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. If there exits a singular rigid line $\Gamma_h\subset\Omega$ of $\bmf{u}$, then $\bmf{u}\equiv \bmf{ 0}$. \end{thm} \begin{proof} Suppose that there exits a singular rigid line $\Gamma_h$ of $\bmf{u}$ as described at the beginning of this section. Then we have \begin{equation}\label{eq:311 cond} \partial_2 \bmf{u}_1(\bmf{x}_0)=0,\quad \bmf{x}_0\in \Gamma_h. \end{equation} By virtue of \eqref{eq:1} we know that \begin{equation}\label{eq:a1b1=0} a_1=b_1=0 \end{equation} since $$ \left|\begin{matrix} k_p & {\mathrm{i}} k_s\\ k_p^3 &-{\mathrm{i}} k_s^3 \end{matrix} \right|=-{\mathrm{i}} k_pk_s(k_p^2+k_s^2)\neq 0. $$ Substituting \eqref{eq:a1b1=0} into \eqref{eq:u=0}, comparing the coefficients of the term $r^3$ in both sides of \eqref{eq:u=0}, we have \begin{equation}\label{eq:313 eq} 3k_p^4 a_0-3{\mathrm{i}} k_s^4 b_0-k_p^4 a_2+{\mathrm{i}} k_s^4 b_2=0. \end{equation} \eqref{eq:311 cond} implies that \eqref{eq:u1} is satisfied. Combing \eqref{eq:2} \eqref{eq:u1}, \eqref{eq:313 eq}, we have $$ \begin{bmatrix} k_p^2 &{\mathrm{i}} k_s^2 & -k_p^2 & -{\mathrm{i}} k_s^2\\ k_p^2&-{\mathrm{i}} k_s^2 &0 &0\\ 0&-2k_s^2&{\mathrm{i}} k_p^2& -k_s^2\\ 3k_p^4&-3{\mathrm{i}} k_s^4 &-k_p^4 & {\mathrm{i}} k_s^4 \end{bmatrix} \begin{bmatrix} a_0\\b_0\\a_2\\ b_2 \end{bmatrix}=\bmf{0}, $$ whose determinant is $-4{\mathrm{i}} k_p^4 k_s^4 (k_p^2+k_s^2) \neq 0$, then we know that $a_0=b_0=a_2=b_2=0$. Using Lemma \ref{lem:rigid}, we can prove $a_\ell=b_\ell=0$ for $\ell\in \mathbb N$ and $\ell\geq 3$, which induces that $\bmf{u}\equiv \bmf{0} $ in $\Omega$ by Proposition \ref{prop:21}. The proof is complete. \end{proof} \subsection{The case with a singular traction-free line} \begin{lem}\label{lem:tranction free} Let $\mathbf{u}$ be a Lam\'e eigenfunction to \eqref{eq:lame} with the radial wave expansion \eqref{eq:u} around the origin. Suppose that $\Gamma_h^- \in {\mathcal T}_\Omega^{\kappa} $. Then we have \begin{equation}\label{eq:8} \left\{ \begin{array}{l} {\mathrm{i}} k_p^2 a_2- k_s^2 b_2=0,\\ a_0 =0, \end{array} \right. \end{equation} and \begin{equation}\label{eq:10} \left\{ \begin{array}{l} k_s^3 b_1+{\mathrm{i}} k_p^3 a_3-k_s^3 b_3=0,\\ a_1=0. \end{array} \right. \end{equation} Furthermore, suppose that $a_\ell=b_\ell=0$ for $\ell=0,1$, then \begin{equation}\label{eq:lem33} a_\ell=b_\ell=0,\quad \forall \ell \in \mathbb N\cup \{0\}. \end{equation} \end{lem} \begin{proof} Since $\Gamma_h^-$ is a traction-free line of $\bmf{u}$, then $T_\nu \bmf{u}\big|_{\Gamma_h^- }=\bmf{0}$, we have from \eqref{eq:Tu2} that \begin{equation}\label{eq:Tu2=0} \begin{split} \bmf{0}= \sum_{m=0}^{\infty} & \left\{-\frac{{\mathrm{i}} k_p^2}{2} a_{m} \mu J_{m-2}(k_p r) \mathbf{e}_1 - \frac{{\mathrm{i}} k_p^2}{2} a_{m} (\lambda+\mu) J_m(k_p r) \mathbf{e}_1\right.\\ & + \frac{k_s^2}{2} b_{m} \mu J_{m-2}\left(k_{s} r\right) \mathbf{e}_1 + \frac{{\mathrm{i}} k_p^2}{2} a_{m} \left(\lambda+\mu\right) J_m\left(k_{p} r\right) \mathbf{e}_2\\ & + \frac{{\mathrm{i}} k_p^2}{2} a_{m} \mu J_{m+2}\left(k_{p} r\right) \mathbf{e}_2 + \frac{k_s^2}{2} b_{m} \mu J_{m+2}\left(k_{s} r\right) \mathbf{e}_2 \bigg\}, \end{split} \end{equation} where $r\in [0,h] $. Using \eqref{eq:J-1} and Lemma \ref{lem:co exp}, comparing the coefficients of the term $r^0$ in both sides of \eqref{eq:Tu2=0}, we obtain \begin{equation}\nonumber \left({\mathrm{i}} k_p^2 \left(\lambda+\mu\right) a_0+{\mathrm{i}} k_p^2 \mu a_2-k_s^2 \mu b_2\right)\mathbf{e}_1-{\mathrm{i}} k_p^2 \left(\lambda+\mu\right) a_0 \mathbf{e}_2=\bmf{0}. \end{equation} From the fact that $\mathbf{e}_1$ and $\mathbf{e}_2$ are linearly independent, we prove \eqref{eq:8} by using \eqref{eq:convex}. Again comparing the coefficients of the term $r^1$ in both sides of \eqref{eq:Tu2=0}, we obtain \begin{equation}\nonumber \left({\mathrm{i}} k_p^3 \lambda a_1+k_s^3 \mu b_1+{\mathrm{i}} k_p^3 \mu a_3-k_s^3 \mu b_3\right)\mathbf{e}_1-{\mathrm{i}} k_p^3 \left(\lambda+\mu\right)a_1 \mathbf{e}_2=\bmf{0}, \end{equation} from which one can easily see that \eqref{eq:8} holds by using \eqref{eq:convex}. Suppose that $a_\ell=b_\ell=0$ for $\ell=0,1$, then we want to prove \eqref{eq:lem33}. Substituting $a_\ell=b_\ell=0$ ($\ell=0,1$) into \eqref{eq:Tu2=0} and comparing the coefficients of $r^2$ in the resulting equation \eqref{eq:Tu2=0}, which are related to $J_0(k_p r)$, $J_0(k_s r)$ and $J_2(k_p r)$, $J_2(k_s r)$ for the indexes $m=2$ and $m=4$ in \eqref{eq:Tu2=0}, we obtain that \begin{equation}\nonumber \left({\mathrm{i}} k_p^4 \left(\lambda-\mu\right) a_2+2k_s^4 \mu b_2+{\mathrm{i}} k_p^4\mu a_4-k_s^4 \mu b_4\right)\mathbf{e}_1-{\mathrm{i}} k_p^4 \left(\lambda+\mu\right) a_2\mathbf{e}_2=\bmf{0}. \end{equation} Since $\mathbf{e}_1$ and $\mathbf{e}_2$ are linear independent, we can deduce that \begin{equation}\label{eq:12} \left\{ \begin{array}{l} {\mathrm{i}} k_p^4 \left(\lambda-\mu\right) a_2+2k_s^4 \mu b_2+{\mathrm{i}} k_p^4\mu a_4-k_s^4 \mu b_4=0,\\ a_2=0,\\ \end{array} \right. \end{equation} Since $a_2=0$ from \eqref{eq:12}, we have $b_2=0$ from \eqref{eq:8}. Thus \eqref{eq:12} can be written as $ {\mathrm{i}} k_p^4 a_4-k_s^4 b_4=0$. Repeating the above argument in an inductive manner, we can prove that \begin{equation}\label{eq:albl=0T} a_\ell=b_\ell=0 \end{equation} for $\ell=0,1,2,\ldots, m-1$ where $m\in \mathbb N$ is fixed and $m \geq 3$. Next, we prove $a_m=b_m=0$. Substituting \eqref{eq:albl=0T} into \eqref{eq:Tu2=0}, it yields that \begin{equation}\label{eq:Tu2=0 new} \begin{split} \bmf{0}= \sum_{\ell=m}^{\infty} & \left\{-\frac{{\mathrm{i}} k_p^2}{2} a_{\ell } \mu J_{\ell -2}(k_p r) \mathbf{e}_1 - \frac{{\mathrm{i}} k_p^2}{2} a_{\ell } (\lambda+\mu) J_\ell (k_p r) \mathbf{e}_1\right.\\ & + \frac{k_s^2}{2} b_{\ell } \mu J_{\ell -2}\left(k_{s} r\right) \mathbf{e}_1 + \frac{{\mathrm{i}} k_p^2}{2} a_{\ell } \left(\lambda+\mu\right) J_\ell \left(k_{p} r\right) \mathbf{e}_2\\ & + \frac{{\mathrm{i}} k_p^2}{2} a_{\ell } \mu J_{\ell +2}\left(k_{p} r\right) \mathbf{e}_2 + \frac{k_s^2}{2} b_{ \ell } \mu J_{\ell +2}\left(k_{s} r\right) \mathbf{e}_2 \bigg\}. \end{split} \end{equation} Therefore the lowest order of $r$ in the left sides of \eqref{eq:Tu2=0 new} is $m-2$. Comparing the coefficients of $r^{m-2}$ in both sides of \eqref{eq:Tu2=0 new} and using \eqref{eq:Jm ex}, we have \begin{equation}\label{eq:16} {\mathrm{i}} k_p ^m a_m- k_s^m b_m=0, \end{equation} since $\mu>0$. Again comparing the coefficients of $r^{m}$ in both sides of \eqref{eq:Tu2=0 new}, which are related to $J_{m-2}(k_p r)$, $J_{m-2}(k_s r)$ and $J_m(k_p r)$, $J_m(k_s r)$, we can derive that \begin{equation}\nonumber \begin{aligned} & \left[{\mathrm{i}} k_p^{m+2} \left(\lambda-\left(m-1\right)\mu\right) a_m+m k_s^{m+2} \mu b_m+{\mathrm{i}} k_p^{m+2} \mu a_{m+2}-k_s^{m+2} \mu b_{m+2}\right]\mathbf{e}_1\\ & - {\mathrm{i}} k_p^{m+2} \left(\lambda+\mu\right) a_m \mathbf{e}_2=\bmf{0}. \end{aligned} \end{equation} Since $\mathbf{e}_1$ and $\mathbf{e}_2$ are linearly independent, we can deduce that \begin{equation}\label{eq:17} \left\{ \begin{array}{l} {\mathrm{i}} k_p^{m+2} \left(\lambda-\left(m-1\right)\mu\right) a_m+m k_s^{m+2} \mu b_m+{\mathrm{i}} k_p^{m+2} \mu a_{m+2}-k_s^{m+2} \mu b_{m+2}=0,\\ a_m=0. \end{array} \right. \end{equation} Combining \eqref{eq:16}and \eqref{eq:17}, we derive that $a_m=b_m=0$. Using the above recursive procedure, we can prove that $a_m=b_m=0$ for $m \in \mathbb{N}\cup \{0\}$. The proof is complete. \end{proof} \begin{thm}\label{thm:traction free line} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. If there exits a singular traction-free line $\Gamma_h\subset\Omega$ of $\bmf{u}$, then $\bmf{u}\equiv \bmf{ 0}$. \end{thm} \begin{proof} Suppose that there exits a singular traction-free line $\Gamma_h$ of $\bmf{u}$ as described at the beginning of this section. Recall that $\bmf{u}$ has the expansion \eqref{eq:u} around the origin. From Lemma \ref{lem:tranction free}, we know that \eqref{eq:8} holds. By virtue of the fact that $\bmf{u}(\bmf{0})=\bmf{0} \mbox{ and }\boldsymbol{\tau} \cdot \partial_{\nu} \bmf{u} (\bmf{0})=0$, from Lemma \ref{lem:34 three conds}, we know that \eqref{eq:u1} is satisfied. Substituting \eqref{eq:8} into \eqref{eq:u1}, we can obtain that $ b_0=0. $ Furthermore, substituting the second equation in \eqref{eq:10} into \eqref{eq:u0}, one can derive that $ b_1=0, $ which implies that \eqref{eq:10} can be rewritten as $ {\mathrm{i}} k_p^3 a_3-k_s^3 b_3=0$. By now we have proven that $a_\ell=b_\ell=0$ for $\ell=0,1$. Then from Lemma \ref{lem:tranction free}, we have that \eqref{eq:lem33} holds. Therefore, from Proposition \ref{prop:21}, we know that $\bmf{u}\equiv \bmf{0}$ in $\Omega$. The proof is complete. \end{proof} \subsection{The case with a singular impedance line} \begin{lem}\label{lem:impedance eqn} Let $\mathbf{u}$ be a solution to \eqref{eq:lame} with the radial wave expansion \eqref{eq:u} around the origin. Suppose that there is an impedance line $\Gamma_h^-$ of $\bmf{u}$ with a constant impedance parameter $0\neq \eta \in \mathbb C$. Then we have \begin{equation}\label{eq:Tu+u1} \left\{ \begin{array}{l} \eta k_p a_1+{\mathrm{i}} \eta k_s b_1 -{\mathrm{i}} k_p^2 \mu a_2+k_s^2 \mu b_2=0,\\ a_0=0, \end{array} \right. \end{equation} and \begin{equation}\label{eq:Tu+u2} \left\{ \begin{array}{l} - k_s^3 \mu b_1 + \eta k_p^2 a_2+{\mathrm{i}} \eta k_s^2 b_2 +{\mathrm{i}} k_p^3 \mu a_3-k_s^3 \mu b_3=0,\\ a_1=0. \end{array} \right. \end{equation} Furthermore, if $a_\ell=b_\ell=0$ for $\ell=0,1$, then \begin{equation}\label{eq:a1bl=0} a_\ell=b_\ell=0, \quad \forall \ell \in \mathbb N \cup \{0\}. \end{equation} \end{lem} \begin{proof} Since $(T_\nu \bmf{u}+\eta \bmf{u}) \big|_{\Gamma_h^-}=\bmf{0}$, using \eqref{eq:Tu4} and noting $\varphi=0$ on $\Gamma_h^-$, we have for $0 \leqslant r\leqslant h$ that \begin{equation}\label{eq:Tu+u=0} \begin{aligned} &\bmf{0}=\sum_{m=0}^{\infty} \biggl \{-\frac{{\mathrm{i}} {k}_p^2}{2} a_{m} \Big[ \mu J_{m-2}(k_p r) \mathbf{e}_1 + (\lambda+\mu) J_m(k_p r) \mathbf{e}_1\\ &- \left(\lambda+\mu\right) J_m\left(k_{p} r\right) \mathbf{e}_2 - \mu J_{m+2}\left(k_{p} r\right) \mathbf{e}_2\Big] + \frac{k_s^2}{2} b_{m}\Big[ \mu J_{m-2}\left(k_{s} r\right) \mathbf{e}_1 + \mu J_{m+2}\left(k_{s} r\right) \mathbf{e}_2 \Big] \\ &+\frac{\eta k_p}{2} a_{m} \left[J_{m-1}\left(k_{p} r\right)\mathbf{e}_1 -J_{m+1}\left(k_{p}r\right) \mathbf{e}_2 \right]+ \frac{{\mathrm{i}} \eta k_s}{2} b_{m} \left[J_{m-1}\left(k_{s} r\right)\mathbf{e}_1 +J_{m+1}\left(k_{s} r\right)\mathbf{e}_2 \right] \bigg\}. \end{aligned} \end{equation} Using Lemma \ref{lem:co exp}, comparing the coefficients of the term $r^0$ in both sides of \eqref{eq:Tu+u=0}, which are related to $J_0(k_p r)$ and $J_0(k_s r)$ for the indexes $m=0$, $m=1$ and $m=2$ in \eqref{eq:Tu+u=0}, we have \begin{equation}\nonumber \left[-{\mathrm{i}} k_p^2 (\lambda+\mu) a_0+\eta k_p a_1+{\mathrm{i}} \eta k_s b_1 -{\mathrm{i}} k_p^2 \mu a_2+k_s^2 \mu b_2\right]\mathbf{e}_1+{\mathrm{i}} k_p^2(\lambda+\mu)a_0 \mathbf{e}_2=\bmf{0}. \end{equation} Using the fact that $\mathbf{e}_1$ and $\mathbf{e}_2$ are linearly independent, we can obtain \eqref{eq:Tu+u1} since $k_p \neq 0$ and $\lambda+\mu >0$ from \eqref{eq:convex}. Similarly, comparing the coefficients of the term $r^1$ in both sides of \eqref{eq:Tu+u=0}, we can derive \eqref{eq:Tu+u2}. Now we are in the position to prove \eqref{eq:a1bl=0} under the condition $a_\ell=b_\ell=0$ for $\ell=0,1$. Since $a_1=b_1=0$, \eqref{eq:Tu+u1} can be rewritten as \begin{equation}\label{eq:Tu+u1a} {\mathrm{i}} k_p^2 a_2-k_s^2 b_2=0. \end{equation} Substituting $a_\ell=b_\ell=0$ into \eqref{eq:Tu+u=0}, where $\ell=0,1$, comparing the coefficients of $r^2$ in the resulting equation \eqref{eq:Tu+u=0}, which are related to $J_0(k_p r)$, $J_0(k_s r)$ and $J_2(k_p r)$, $J_2(k_s r)$ for the indexes $m=2$ $m=3$ and $m=4$ in \eqref{eq:Tu+u=0}, we can derive that \begin{equation}\nonumber \left[{\mathrm{i}} k_p^4 (\mu-\lambda) a_2 - 2 k_s^4 \mu b_2+\eta k_p^3 a_3+{\mathrm{i}} \eta k_s^3 b_3 -{\mathrm{i}} k_p^4 \mu a_4 + k_s^4 \mu b_4\right]\mathbf{e}_1+{\mathrm{i}} k_p^4(\lambda+\mu)a_2 \mathbf{e}_2=\bmf{0}, \end{equation} Since $\mathbf{e}_1$ and $\mathbf{e}_2$ are linearly independent, we can obtain that \begin{equation}\label{eq:Tu+u3} \left\{ \begin{array}{l} - 2 k_s^4 \mu b_2+\eta k_p^3 a_3+{\mathrm{i}} \eta k_s^3 b_3 -{\mathrm{i}} k_p^4 \mu a_4+k_s^4 \mu b_4=0,\\ a_2=0,\\ \end{array} \right. \end{equation} Combing \eqref{eq:Tu+u1a} with \eqref{eq:Tu+u3}, we can derive that $b_2=0$. Therefore, it is easy to see that \eqref{eq:Tu+u2} can be rewritten as \begin{equation}\label{eq:Tu+u2a} {\mathrm{i}} k_p^3 a_3-k_s^3 b_3=0. \end{equation} By now we have proven that $a_\ell=b_\ell=0$ ($\ell=0,1,2$) {if $\Gamma_h^- \Subset \Omega $ is an impedance line of $\bmf{u}$.} Repeating the above argument in an inductive manner, suppose that we have proven \begin{equation}\label{eq:albl=0Tu+u} a_\ell=b_\ell=0 \end{equation} for $\ell=0,1,2,\ldots, m-1$ where $m\in \mathbb N$ is fixed and $m \geq 3$. We next prove $a_m=b_m=0$. Substituting \eqref{eq:albl=0Tu+u} into \eqref{eq:Tu+u=0}, it yields that \begin{equation}\label{eq:Tu+u=0new} \begin{aligned} &\bmf{0}=\sum_{\ell=m}^{\infty} \biggl \{-\frac{{\mathrm{i}} {k}_p^2}{2} a_{\ell} \Big[ \mu J_{\ell-2}(k_p r) \mathbf{e}_1 + (\lambda+\mu) J_\ell(k_p r) \mathbf{e}_1\\ &- \left(\lambda+\mu\right) J_\ell\left(k_{p} r\right) \mathbf{e}_2 - \mu J_{\ell+2}\left(k_{p} r\right) \mathbf{e}_2\Big] + \frac{k_s^2}{2} b_{\ell}\Big[ \mu J_{\ell-2}\left(k_{s} r\right) \mathbf{e}_1 + \mu J_{\ell+2}\left(k_{s} r\right) \mathbf{e}_2 \Big] \\ &+\frac{\eta k_p}{2} a_{\ell} \left[J_{\ell-1}\left(k_{p} r\right)\mathbf{e}_1 -J_{\ell+1}\left(k_{p}r\right) \mathbf{e}_2 \right] + \frac{{\mathrm{i}} \eta k_s}{2} b_{\ell} \left[J_{\ell-1}\left(k_{s} r\right)\mathbf{e}_1 +J_{\ell+1}\left(k_{s} r\right)\mathbf{e}_2 \right] \bigg\}. \end{aligned} \end{equation} Therefore the lowest order of $r$ in the left sides of \eqref{eq:Tu+u=0new} is $m-2$. Comparing the coefficients of $r^{m-2}$ in both sides of \eqref{eq:Tu+u=0new}, we have \begin{equation}\label{eq:Tu+u4} {\mathrm{i}} k_p ^m a_m- k_s^m b_m=0, \end{equation} since $\mu>0$. Again comparing the coefficients of $r^{m}$ in both sides of \eqref{eq:Tu+u=0new}, which are related to $J_{m-2}(k_p r)$, $J_{m-2}(k_s r)$ and $J_m(k_p r)$, $J_m(k_s r)$, we can derive that \begin{equation}\label{eq:339 im} \begin{aligned} & \big[-{\mathrm{i}} k_p^{m+2} \left(\lambda+(1-m)\mu\right) a_m - m k_s^{m+2} \mu b_m + \eta k_p^{m+1} a_{m+1} + {\mathrm{i}} \eta k_s^{m+1} \mu b_{m+1} \\ &- {\mathrm{i}} k_p^{m+2} \mu a_{m+2} + k_s^{m+2} \mu a_{m+2} \big]\mathbf{e}_1 + {\mathrm{i}} k_p^{m+2} \left(\lambda+\mu\right) a_m \mathbf{e}_2=\bmf{0}. \end{aligned} \end{equation} Since $\mathbf{e}_1$ and $\mathbf{e}_2$ are linearly independent, from \eqref{eq:339 im}, we can deduce that \begin{equation}\label{eq:Tu+u5} \left\{ \begin{array}{l} - m k_s^{m+2} \mu b_m + \eta k_p^{m+1} a_{m+1} + {\mathrm{i}} \eta k_s^{m+1} \mu b_{m+1} - {\mathrm{i}} k_p^{m+2} \mu a_{m+2} + k_s^{m+2} \mu a_{m+2}=0,\\ a_m=0,\\ \end{array} \right. \end{equation} Combining \eqref{eq:Tu+u4}and \eqref{eq:Tu+u5}, we derive that $a_m=b_m=0$. Using the above recursive procedure, we can prove that $a_m=b_m=0$ for $m \in \mathbb{N}\cup \{0\}$. The proof is complete. \end{proof} \begin{thm}\label{thm:impedance line} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. If there exits a singular impedance line $\Gamma_h\subset\Omega$ of $\bmf{u}$ with a constant impedance parameter $0\neq \eta \in \mathbb C$ as defined in \eqref{eq:def3}, then $\bmf{u}\equiv \bmf{ 0}$. \end{thm} \begin{proof} Suppose that there exits a singular impedance line $\Gamma_h$ of $\bmf{u}$ as described at the beginning of this section with a nonzero constant impedance $\eta$. From \eqref{eq:def3}, we see that \begin{equation}\label{eq:thm 33 cond1} \mathbf{u}(\bmf{0})=\bmf{0} \mbox{ and } \partial_2 u_1(\bmf{0})=0. \end{equation} Moreover, $(T_\nu \bmf{u}+\eta \bmf{u}) \big|_{\Gamma_h^-}=\bmf{0}$ . Hence from Lemma \ref{lem:impedance eqn}, we know that \eqref{eq:Tu+u1} and \eqref{eq:Tu+u2} hold. By virtue of \eqref{eq:thm 33 cond1}, we know that \eqref{eq:u0} and \eqref{eq:u1} hold. Substituting $a_1=0$ in \eqref{eq:Tu+u2} into \eqref{eq:u0} we have $b_1=0$. Therefore from \eqref{eq:Tu+u1} we derive that \begin{equation}\label{eq:339} -\mathrm{i} k^2_p a_2 + k^2_2 b_2=0 \end{equation} by using the fact that $a_1=b_1=0$. Substituting \eqref{eq:339} into \eqref{eq:u1} we have $b_0=0$. Therefore, from Lemma \ref{lem:impedance eqn}, we have $a_\ell=b_\ell=0$ for $\forall \ell \in \mathbb N \cup \{0\}$. Then $\bmf{u}\equiv \bmf{0}$ in $\Omega$ via Proposition \ref{prop:21}. The proof is complete. \end{proof} \section{Generalized Holmgre's principle with the non-degenerate intersection of two homogeneous lines} In this section, we consider the homogeneous lines introduced in Definition~\ref{def:1}. We shall show that the generic non-degenerate intersections of two of such lines also generate microlocal singularities, which prevent the occurrence of such intersections unless the Lam\'e eigenfunction $\bmf{u}$ is identically vanishing. As discussed in the beginning of Section~\ref{sect:2}, we assume throughout this section that the aforementioned two homogeneous lines are given by $\Gamma_h^\pm$ in \eqref{eq:gamma_pm} with the intersecting angle $\varphi_0\in (0, \pi)$. \begin{lem}\label{lem:FB exp} Let $\mathbf{u}$ be a solution to \eqref{eq:lame} with the radial wave expansion \eqref{eq:u} around the origin. Suppose that there are two rigid lines $\Gamma_h^+$ and $\Gamma_h^-$ in $\Omega$ of $\bmf{u}$ that are intersecting with each other in a non-degenerate way. Then $$ - {\mathrm{i}} k_p ^4 (2 \lambda + \mu) a_0 + \mu k_s ^4 b_0=0. $$ \end{lem} \begin{proof} Recall that $I_1^\pm$ are defined in \eqref{eq:CGO6}. Substituting \eqref{eq:Tudotv} into \eqref{eq:CGO6}, since $\bmf{u} \big|_{\Gamma_h^\pm}=\bmf{0}$, we can obtain that \begin{equation}\label{eq:Tudotv1} \begin{split} I_1^+ & = \int_{\Gamma_h^+} \left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}\mathrm{d} \sigma = - \int_0^h {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) }\bigg\{{\mathrm{i}} k_p^2 (\lambda+\mu) \mathrm e^{{\mathrm{i}} \varphi_0} a_0 + \frac{{\mathrm{i}}}{2} k_p^3 (\lambda+\mu) \mathrm e^{2 {\mathrm{i}} \varphi_0} a_1 r\\ &\quad +\frac{1}{8}\left[{\mathrm{i}} k_p^4 (\lambda+\mu) \mathrm e^{3 {\mathrm{i}} \varphi_0} a_2 - {\mathrm{i}} k_p^4 (2 \lambda+\mu) \mathrm e^{ {\mathrm{i}} \varphi_0} a_0+ k_s^4 \mu \mathrm e^{ {\mathrm{i}} \varphi_0} b_0\right]r^2 \bigg\}\mathrm{d} r - r_{I_1^+,1} \\ &= -{\mathrm{i}} \left(\lambda+\mu\right) k_p^2 \left( 2 a_0 s^{-2} - 60 k_p^2 \mathrm{e}^{-2 {\mathrm{i}} \varphi_0} a_0 s^{-6} + 6 k_p a_1 s^{-4} + 30 k_p^2 a_2 s^{-6} \right)\\ &-30 {\mathrm{i}} \mu k_p ^4 \mathrm{e}^{-2 {\mathrm{i}} \varphi_0} a_0 s^{-6} - 30\mu k_s^4 \mathrm{e}^{-2 {\mathrm{i}} \varphi_0} b_0 s^{-6}-r_{I_1^+,1}, \end{split} \end{equation} where $r_{I_1^+,1}= - \int_0^h {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) } R_{1,\Gamma_h^+} \mathrm{d} r $. Similarly, we have \begin{equation}\label{eq:Tudotv2} \begin{split} I_1^- & = \int_{\Gamma_h^-} \left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}\mathrm{d} \sigma = \int_0^h {\mathrm e}^{-s \sqrt{ r}}\bigg\{{\mathrm{i}} k_p^2 (\lambda+\mu) a_0 + \frac{{\mathrm{i}}}{2} k_p^3 (\lambda+\mu) a_1 r\\ &\quad +\frac{1}{8}({\mathrm{i}} k_p^4 (\lambda+\mu) a_2 - {\mathrm{i}} k_p^4 (2 \lambda+\mu) a_0+ k_s^4 \mu b_0)r^2 \bigg\}\mathrm{d} r + r_{I_1^-,1} \\ & = {\mathrm{i}} \left(\lambda+\mu\right) k_p^2 \left( 2 a_0 s^{-2} - 60 k_p^2 a_0 s^{-6} + 6 k_p a_1 s^{-4} + 30 k_p^2 a_2 s^{-6} \right)\\ &\quad +30{\mathrm{i}} \mu k_p ^4 a_0 s^{-6} + 30 \mu k_s^4 b_0 s^{-6}+ r_{I_1^-,1}, \end{split} \end{equation} where $r_{I_1^-,1}= \int_0^h {\mathrm e}^{- s \sqrt{ r} } R_{1,\Gamma_h^-} \mathrm{d} r $. Therefore by straightforward calculations, we can derive that \begin{equation}\label{eq:EFI} I_1^++I_1^-= 30 (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})(-{\mathrm{i}} k_p ^4 (2 \lambda+\mu)a_0+ \mu k_s^4 b_0)s^{-6}+ r_{I_1^-,1}-r_{I_1^+,1}. \end{equation} By virtue of \eqref{eq:R1 s} and \eqref{eq:268b int}, it is easy to see that \begin{equation}\label{eq:54 r} \left|r_{I_1^+,1} \right| \leq S_2 \cdot {\mathcal O}( s^{-8} ),\quad \left|r_{I_1^-,1} \right| \leq S_2 \cdot {\mathcal O}( s^{-8} ). \end{equation} Substituting \eqref{eq:EFI} into \eqref{eq:CGO2}, we can obtain that \begin{equation} \notag \begin{split} 30 (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})(-{\mathrm{i}} k_p ^4 (2 \lambda+\mu)a_0+ \mu k_s^4 b_0)s^{-6}&=I_3-I_2+r_{I_1^-,1}-r_{I_1^+,1}, \end{split} \end{equation} which further implies that \begin{equation}\label{eq:55 int} \begin{split} 30 (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})(-{\mathrm{i}} k_p ^4 (2 \lambda+\mu)a_0+ \mu k_s^4 b_0)&=s^{6}\left(I_3-I_2+r_{I_1^-,1}-r_{I_1^+,1}\right). \end{split} \end{equation} From \eqref{eq:55 int}, it yields that \begin{equation}\label{eq:56 ineq} \left| 30 (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})(-{\mathrm{i}} k_p ^4 (2 \lambda+\mu)a_0+ \mu k_s^4 b_0) \right|\leq s^6\left(|I_3|+\left|I_2\right|+ \left| r_{I_1^-,1} \right|+\left|r_{I_1^+,1}\right| \right). \end{equation} Since $\Gamma_h^-$ is a rigid line of $\bmf{u}$, then from Lemma \ref{lem:rigid}, \eqref{eq:2} holds. Substituting the second equation of \eqref{eq:2} into \eqref{eq:I3}, we can obtain that \begin{equation}\label{eq:57 I_3} \left| I_3\right | \leq \frac{2\kappa \varphi_0 \Gamma(6) S_1 }{\delta_{\mathcal K }^8}s^{-8}. \end{equation} In \eqref{eq:56 ineq}, using \eqref{eq:I2I4}, \eqref{eq:54 r} and \eqref{eq:57 I_3}, letting $s\rightarrow +\infty$, under the condition $\varphi_0\neq \pi$, we can finish the proof of this lemma. \end{proof} \begin{thm}\label{thm:rigid line exp} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. There cannot exit two intersecting rigid lines $\Gamma_h^\pm$ of $\bmf{u}$ with the intersecting angle $\angle(\Gamma_h^+,\Gamma_h^-)=\varphi_0\neq \pi $ unless $\bmf{u}\equiv \bmf{0}$. \end{thm} \begin{proof} Suppose that there are two intersecting rigid lines $\Gamma_h^\pm$ of $\bmf{u}$ with the intersecting angle $\angle(\Gamma_h^+,\Gamma_h^- )=\varphi_0\neq \pi$. Recall that $\bmf{u}$ has the raidal wave expansion \eqref{eq:u}. From \eqref{eq:2} and Lemma \ref{lem:FB exp}, it yields that \begin{equation} \bigg\{ \begin{array}{l} -{\mathrm{i}} k_p^4(2\lambda+\mu)a_0+\mu k_s^4 b_0=0,\\ k_p ^2 a_0- {\mathrm{i}} k_s^2 b_0=0,\\ \end{array} \end{equation} where $k_p$ and $k_s$ are defined in \eqref{eq:kpks}. Moreover, the eigenvalue $\kappa $ of \eqref{eq:lame} is positive. Hence by using \eqref{eq:convex} it is easy to see that $$ (2\lambda+\mu)k_p^2 +\mu k_s^2= \frac{3(\lambda+\mu )\kappa }{\lambda+2\mu }>0. $$ Therefore \begin{equation}\notag \left| \begin{array}{cc} -{\mathrm{i}} k_p^4(2\lambda+\mu) & \mu k_s^4\\ k_p^2 & -{\mathrm{i}} k_s^2\\ \end{array} \right|=-k_p^2 k_s^2 \left[(2\lambda+\mu)k_p^2 +\mu k_s^2\right]< 0, \end{equation} which implies that \begin{equation}\label{eq:a0b0=0} a_0=b_0=0. \end{equation} Substituting \eqref{eq:a0b0=0} into \eqref{eq:u=0}, we compare the coefficients of $r^2$ in both sides of \eqref{eq:u=0} to obtain that \begin{equation}\nonumber \left(-2 k_p^3 a_1-2 {\mathrm{i}} k_s^3 b_1+k_p^3 a_3+{\mathrm{i}} k_s^3 b_3\right)\mathbf{e}_1-\left(k_p^3 a_1-{\mathrm{i}} k_s^3 b_1\right)\mathbf{e}_2=\bmf{0}, \end{equation} which can be used to further derive that \begin{equation}\label{eq:4} \left\{ \begin{array}{l} -2 k_p^3 a_1-2 {\mathrm{i}} k_s^3 b_1+k_p^3 a_3+{\mathrm{i}} k_s^3 b_3=0,\\ k_p^3 a_1-{\mathrm{i}} k_s^3 b_1=0. \end{array} \right. \end{equation} Combining \eqref{eq:1} and \eqref{eq:4}, it yields that \begin{equation}\label{eq:ll1} k_p a_1 + {\mathrm{i}} k_s b_1=0,\ \ k_p^3 a_1-{\mathrm{i}} k_s^3 b_1=0. \end{equation} Since \begin{equation}\notag \left| \begin{array}{cc} k_p & {\mathrm{i}} k_s\\ k_p^3 & -{\mathrm{i}} k_s^3\\ \end{array} \right|=-{\mathrm{i}} k_p k_s \left(k_p^2 +k_s^2\right)\neq 0, \end{equation} which together with \eqref{eq:ll1} readily implies that \begin{equation}\label{eq:a1b1=0} a_1=b_1=0. \end{equation} In view of \eqref{eq:a0b0=0} and \eqref{eq:a1b1=0}, from Lemma \ref{lem:rigid} we obtain that $a_\ell=b_\ell=0$ for $\ell\in \mathbb N$, which implies that $\bmf{u} \equiv \bmf{0}$ in $\Omega$ by Proposition \ref{prop:21}. The proof is complete. \end{proof} \begin{lem}\label{lem:singular line} Let $\mathbf{u}$ be a solution to \eqref{eq:lame} with the radial wave expansion \eqref{eq:u} around the origin. Suppose that there are two traction-free lines $\Gamma_h^+$ and $\Gamma_h^-$. If $\angle(\Gamma_h^+,\Gamma_h^-)= \varphi_0 \neq\pi$ and \begin{equation}\label{eq:512 cond} \frac{4 \varphi_0 }{3 \delta_{\mathcal K}^6 }<1, \end{equation} where $\delta_{\mathcal K}=\min_{ 0<\varphi<\varphi_0 }\cos(\varphi/2 )>0$ is defined in \eqref{eq:volume n}, then $ b_0= 0$. \end{lem} \begin{proof} Recall that $I_1^\pm$ are defined in \eqref{eq:CGO6}. Substituting \eqref{eq:Tvu} into \eqref{eq:CGO6}, since $T_\nu\bmf{u} \big|_{\Gamma_h^\pm}=\bmf{0}$, we can obtain that \begin{equation}\label{eq:Tvdotu1} \begin{split} I_1^+ & = - \int_{\Gamma_h^+} \left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\mathrm{d} \sigma = s \mu\zeta(\varphi_0) \int_0^h {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) }\bigg\{\frac{1}{2}({\mathrm{i}} k_p^2 a_0 + k_s^2 b_0) \mathrm e^{{\mathrm{i}} \varphi_0} r^{1/2}\\ & + \frac{1}{8}({\mathrm{i}} k_p^3 a_1 + k_s^3 b_1) \mathrm e^{2 {\mathrm{i}} \varphi_0} r^{3/2} + \frac{1}{48}({\mathrm{i}} k_p^4 a_2 + k_s^4 b_2) \mathrm e^{3 {\mathrm{i}} \varphi_0} r^{5/2}\\ & - \frac{1}{16}({\mathrm{i}} k_p^4 a_0 + k_s^4 b_0) \mathrm e^{{\mathrm{i}} \varphi_0} r^{5/2}\bigg\}\mathrm{d} r - r_{I_1^+,2} \\ & = -2 \mu({\mathrm{i}} k_p^2 a_0+ k_s^2 b_0)s^{-2}-6 \mu({\mathrm{i}} k_p^3 a_1+ k_s^3 b_1)s^{-4}-120 \mu({\mathrm{i}} k_p^4 a_2+ k_s^4 b_2)s^{-6}\\ & +90 \mathrm{e}^{-2 {\mathrm{i}} \varphi_0}\mu({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6}-r_{I_1^+,2}, \end{split} \end{equation} where \begin{equation}\label{eq:r1+} r_{I_1^+,2}= {\mathrm{i}} s \mu\zeta(\varphi_0) \int_0^h {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) } R_{2,\Gamma_h^+} \mathrm{d} r. \end{equation} Similarly, we have \begin{equation}\label{eq:Tvdotu2} \begin{split} I_1^- & = - \int_{\Gamma_h^-} \left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\mathrm{d} \sigma = s \mu \int_0^h {\mathrm e}^{- s \sqrt{ r} }\bigg\{\frac{1}{2}({\mathrm{i}} k_p^2 a_0 + k_s^2 b_0) r^{1/2}\\ & + \frac{1}{8}({\mathrm{i}} k_p^3 a_1 + k_s^3 b_1) r^{3/2} + \frac{1}{48}({\mathrm{i}} k_p^4 a_2 + k_s^4 b_2) r^{5/2}\\ & - \frac{1}{16}({\mathrm{i}} k_p^4 a_0 + k_s^4 b_0) r^{5/2}\bigg\}\mathrm{d} r - r_{I_1^-,2} \\ & = 2 \mu({\mathrm{i}} k_p^2 a_0+ k_s^2 b_0)s^{-2}+6 \mu({\mathrm{i}} k_p^3 a_1+ k_s^3 b_1)s^{-4}+120 \mu({\mathrm{i}} k_p^4 a_2+ k_s^4 b_2)s^{-6}\\ & -90 \mu({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6} - r_{I_1^-,2}, \end{split} \end{equation} where \begin{equation}\label{eq:r1-} r_{I_1^-,2} = {\mathrm{i}} s \mu\int_0^h {\mathrm e}^{- s \sqrt{ r} } R_{2,\Gamma_h^-} \mathrm{d} r. \end{equation} Therefore, from \eqref{eq:Tvdotu1} and \eqref{eq:Tvdotu2}, after straightforward algebraic calculations, we derive that \begin{equation} \nota I_1^++I_1^-= -90 \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^4 a_0+ k_s^4 b_0)s^{-6} - r_{I_1^-,2}-r_{I_1^+,2}, \end{equation} which can be further reduced to \begin{equation}\label{eq:EFI1} I_1^++I_1^-= -90 \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0}) k_s^4 b_0s^{-6} - r_{I_1^-,2}-r_{I_1^+,2} \end{equation} via the second equality of \eqref{eq:8} since $\Gamma_h^-$ is a traction-free line. By virtue of \eqref{eq:R2 s} and \eqref{eq:268b int}, it is easy to see that \begin{equation}\label{eq:54 r1} |r_{I_1^+,2}| \leq S_3 \cdot{\mathcal O}( s^{-8} ),|r_{I_1^-,2}| \leq S_3 \cdot{\mathcal O}( s^{-8} ), \end{equation} as $s\rightarrow +\infty$. Using the fact that $a_0=0$ and \eqref{eq:I3}, we have as $s\rightarrow +\infty$ that \begin{equation}\label{eq:55 I_3} \left| I_3 \right| \leq k_s^2 |b_0| \frac{\kappa \varphi_0 \Gamma(6) }{\delta_{\mathcal K}^6 }s^{-6}+ \frac{2 \kappa \varphi_0 \Gamma(8) }{\delta_{\mathcal K}^8 } s^{-8}. \end{equation} Substituting \eqref{eq:EFI1} into \eqref{eq:CGO2}, using \eqref{eq:I2I4}, \eqref{eq:54 r1} and \eqref{eq:55 I_3}, we obtain that \begin{equation}\label{eq:518 in} \begin{split} 90 \mu |1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0}| k_s^4 |b_0| s^{-6} &\leq k_s^2 |b_0| \frac{\kappa \varphi_0 \Gamma(6) }{\delta_{\mathcal K}^6 }s^{-6}+ \frac{2 \kappa \varphi_0 \Gamma(8) }{\delta_{\mathcal K}^8 } s^{-8} + S_3\cdot {\mathcal O}(s^{-8})\\ &+ \mathcal{C}_{{\mathcal K },B_h,\mu,\lambda}\|\bmf{u}\|_{H^2\left( {{\mathcal K}} \cap B_h \right)}\left(1+s\right){\mathrm e}^{-\delta_{{\mathcal K} } s \sqrt{h}}. \end{split} \end{equation} From \eqref{eq:512 cond}, recalling that \eqref{eq:kpks}, we can derive that \begin{equation}\label{eq:512 cond con} \mu > \frac{4\kappa \varphi_0 }{3 k_s^2 \delta_{\mathcal K}^6 }, \end{equation} Multiplying $s^6$ in both sides of \eqref{eq:518 in} and letting $s\rightarrow + \infty$, under the condition $\varphi_0\neq \pi$, we can derive that $$ \mu |b_0|\leq \frac{4\kappa \varphi_0 }{3 k_s^2 \delta_{\mathcal K}^6 } |b_0| $$ which implies that $b_0=0$ by virtue of \eqref{eq:512 cond con}. The proof is complete. \end{proof} \begin{rem}\label{rem:varphi} Clearly when $\varphi_0 \in (0,\pi)$, it is easy to see that $\delta_{\mathcal K}=\min_{ 0<\varphi<\varphi_0 }\cos(\varphi/2 )= \cos(\varphi_0/2 ) >0$ and the function $f(\varphi_0)=\varphi_0/ \cos^6(\varphi_0/2 ) $ is a strictly monotone increasing function. Denote $$ g(\varphi_0 ):=\frac{4}{3}\cdot \frac{\varphi_0}{\cos^6(\varphi_0/2 ) }-1. $$ Therefore $g(\varphi_0 )$ is a strictly monotone increasing function. Let \begin{equation}\label{eq:varphi0} \varphi_{\sf root} \end{equation} be the root of $g(\varphi_0 )$. In fact, using a standard root-finding algorithm, the numerical value of $\varphi_{\sf root} $ is $\varphi_{\sf root}^{\sf N}:=0.58043041944310849341051295527519$ and $$ g(\varphi_{\sf root}^{\sf N})=-5.5101297694794726936034525182293\times 10^{-40}. $$ Hence if $\varphi_0 \in (0, \varphi_{\sf root})$ we can claim that \eqref{eq:512 cond} is always fulfilled. \end{rem} \begin{thm}\label{thm:two traction} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. Suppose there exit two intersecting lines $\Gamma_h^\pm$ of $\bmf{u}$ such that $\Gamma_h^\pm $ are two traction-free lines with the intersecting angle $\angle(\Gamma_h^+,\Gamma_h^-)=\varphi_0\neq \pi $, if $\bmf{u}(\bmf{0})=\bmf{0}$ and $\varphi_0 \in (0, \varphi_{\sf root})$ where $\varphi_{\sf root}$ is defined in \eqref{eq:varphi0}, then $\bmf{u}\equiv \bmf{0}$. \end{thm} \begin{proof} Suppose that there exit two intersecting lines $\Gamma_h^\pm$ of $\bmf{u}$ such that $\Gamma_h^\pm $ are two traction-free lines satisfying that $\bmf{u}(\bmf{0})=\bmf{0}$ and $0<\varphi_0<\varphi_{\sf root}$. Hence from Remark \ref{rem:varphi}, we know that \eqref{eq:512 cond} is fulfilled. Since $\Gamma_h^-$ is a traction free line, then \eqref{eq:8} and \eqref{eq:10} hold. Since \eqref{eq:512 cond} is fulfilled, from Lemma \ref{lem:singular line}, we know that $b_0=0$. Furthermore, under the condition $\bmf{u}(\bmf{0})=\bmf{0}$, we know \eqref{eq:u0} holds. Substituting $a_1=0$ of \eqref{eq:10} into \eqref{eq:u0}, one has $b_1=0$. In view of the second equality of \eqref{eq:8}, we have shown that $a_\ell=b_\ell=0$ for $\ell=0,1$. Therefore from Lemma \ref{lem:tranction free} and Proposition \ref{prop:21}, we readily have $\bmf{u}\equiv \bmf{0}$ in $\Omega$, which finishes the proof. \end{proof} \begin{thm}\label{thm:u&Tu exp} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. If there exist two intersecting lines $\Gamma_h^\pm$ of $\bmf{u}$ such that $\Gamma_h^-$ is a rigid line and $\Gamma_h^+$ is a traction-free line with the intersecting angle $\angle(\Gamma_h^+,\Gamma_h^-)=\varphi_0\neq \pi $, then $\bmf{u}\equiv \bmf{0}$. \end{thm} \begin{proof} Since $\Gamma_h^-$ is a rigid line and $\Gamma_h^+$ is a traction free line. Recall that $I_1^\pm$ are defined in \eqref{eq:CGO6}, using the definition of rigid and traction-free lines, we can obtain that \begin{equation}\label{eq:17a} \begin{aligned} & I_1^- =\int_{\Gamma_h^-} \left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}\mathrm{d} \sigma,\quad I_1^+ = -\int_{\Gamma_h^+} \left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\mathrm{d} \sigma. \end{aligned} \end{equation} Using \eqref{eq:Tudotv2} and \eqref{eq:Tvdotu1}, we can deduce that \begin{equation}\label{eq:17b} I_1^+ + I_1^- = 2 ( {\mathrm{i}} \lambda k_p^2 a_0- \mu k_s^2 b_0) s^{-2}+ R_{r,1} -r_{I_1^+,2}+ r_{I_1^-,1}, \end{equation} where \begin{equation}\label{eq:17e} \begin{aligned} R_{r,1} & = 6 ( {\mathrm{i}} k_p^3 a_1- \mu k_s^3 b_1)s^{-4}+ 30 \bigg\{-{\mathrm{i}} k_p^4(\mu+2 \lambda-3 u \mathrm{e}^{-2{\mathrm{i}} \varphi_0})a_0\\ & + k_s^4 \mu (1+3\mathrm{e}^{-2{\mathrm{i}} \varphi_0})b_0 + {\mathrm{i}} k_p^4(\lambda-3\mu)a_2-4\mu k_s^4 b_2 \bigg\}s^{-6}. \end{aligned} \end{equation} Therefore, from \eqref{eq:17e}, we have \begin{equation}\label{eq:17f} \left| R_{r,1} \right| \leq {\mathcal O} (s^{-4}), \end{equation} as $s\rightarrow +\infty$. Substituting \eqref{eq:17b} into \eqref{eq:CGO2}, we can obtain that \begin{equation}\label{eq:17g} \begin{split} 2 ( {\mathrm{i}} \lambda k_p^2 a_0- \mu k_s^2 b_0) s^{-2}&=I_3-I_2-R_{r,1} -r_{I_1^-,1}+r_{I_1^+,2}, \end{split} \end{equation} which can further yield that \begin{equation}\label{eq:17h} \begin{split} 2 ( {\mathrm{i}} \lambda k_p^2 a_0- \mu k_s^2 b_0) &=s^{2} (I_3-I_2-R_{r,1} -r_{I_1^-,1}+r_{I_1^+,2}) \end{split} \end{equation} From \eqref{eq:17h}, it gives that \begin{equation}\label{eq:17i} \left| 2 ( {\mathrm{i}} \lambda k_p^2 a_0- \mu k_s^2 b_0) \right|\leq s^2\left(|I_3|+\left|I_2\right|+ \left|R_{r,1}\right|+\left| r_{I_1^-,1} \right|+\left|r_{I_1^+,2}\right| \right). \end{equation} Since $\Gamma_h^-$ is a rigid line of $\bmf{u}$, then from Lemma \ref{lem:rigid}, we know \eqref{eq:2} holds. Substituting the second equation of \eqref{eq:2} into \eqref{eq:I3}, we can obtain that \begin{equation}\label{eq:57 I_3b} \left| I_3\right | \leq \frac{2\kappa \varphi_0 \Gamma(6) S_1 }{\delta_{\mathcal K }^8}s^{-8}. \end{equation} In \eqref{eq:17i}, using \eqref{eq:I2I4}, \eqref{eq:54 r}, \eqref{eq:54 r1}, \eqref{eq:17f} and \eqref{eq:57 I_3b}, letting $s\rightarrow +\infty$, under the condition $\varphi_0\neq \pi$, we can obtain that \begin{equation}\label{eq:mix up 1} {\mathrm{i}} \lambda k_p^2 a_0- \mu k_s^2 b_0=0. \end{equation} Combing \eqref{eq:2} with \eqref{eq:mix up 1}, it yields that \begin{equation}\nonumber \left\{ \begin{array}{l} k_p ^2 a_0 - {\mathrm{i}} k_s^2 b_0=0,\\ {\mathrm{i}} \lambda k_p^2 a_0- \mu k_s^2 b_0=0. \end{array} \right. \end{equation} In view of \eqref{eq:convex}, we have \begin{equation}\notag \left| \begin{array}{cc} k_p^2 & -{\mathrm{i}} k_s^2\\ {\mathrm{i}} \lambda k_p^2 & -\mu k_s^2 \end{array} \right|=-(\lambda+\mu) k_p^2 k_s^2\neq 0, \end{equation} and therefore $ a_0=b_0=0. $ Again using the fact that $\Gamma_h^-\in {\mathcal R}_\Omega^\kappa$, we have \eqref{eq:a1b1 lem}. From Lemma \ref{lem:rigid} and Proposition \ref{prop:21}, we have $\bmf{u}\equiv \bmf{0}$ in $\Omega$, which completes the proof. \end{proof} \begin{thm}\label{thm:54} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. If there exist two intersecting lines $\Gamma_h^\pm$ of $\bmf{u}$ such that $\Gamma_h^-$ is a rigid line and $\Gamma_h^+$ is an impedance line with the intersecting angle $\angle(\Gamma_h^+,\Gamma_h^-)=\varphi_0\neq \pi $, where the associated impedance parameter is a nonzero constant $\eta_2\in \mathbb C$, then $\bmf{u}\equiv \bmf{0}$. \end{thm} \begin{proof} Since $\Gamma_h^-$ is a rigid line and $\Gamma_h^+$ is an impedance line, using the boundary conditions of $\bmf{u}$ on $\Gamma_h^\pm$ respectively, from the definition \eqref{eq:CGO6} of $I_1^\pm $, we have \begin{equation}\notag \begin{aligned} & I_1^- =\int_{\Gamma_h^-} \left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}\mathrm{d} \sigma,\\ & I_1^+ = \int_{\Gamma_h^+}\left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma =-\int_{\Gamma_h^+}\left[\left(\eta_2 \bmf{u}\right) \cdot \bmf{v}+\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma. \end{aligned} \end{equation} By virtue of \eqref{eq:Tudotv2} and \eqref{eq:Im+}, we know that \begin{equation}\label{eq:IR} \begin{aligned} I_1^- & = {\mathrm{i}} \left(\lambda+\mu\right) k_p^2 \left( 2 a_0 s^{-2} - 60 k_p^2 a_0 s^{-6} + 6 k_p a_1 s^{-4} + 30 k_p^2 a_2 s^{-6} \right)\\ & +30{\mathrm{i}} \mu k_p ^4 a_0 s^{-6} + 30 \mu k_s^4 b_0 s^{-6}+ r_{I_1^-,1},\\ I_1^+& = -2 \mu({\mathrm{i}} k_p^2 a_0+ k_s^2 b_0)s^{-2}-6 \mu({\mathrm{i}} k_p^3 a_1+ k_s^3 b_1)s^{-4}-120 \mu({\mathrm{i}} k_p^4 a_2+ k_s^4 b_2)s^{-6}\\ & +90 \mathrm{e}^{-2 {\mathrm{i}} \varphi_0}\mu({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6}-r_{I_1^+,2} + 6 \eta_2(k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) \mathrm e^{- {\mathrm{i}} \varphi_0} s^{-4} - r_{I_1^+,0}, \end{aligned} \end{equation} where $ r_{I_1^-,1}$, $r_{I_1^+,2} $, $r_{I_1^+,0}$ are defined in \eqref{eq:Tudotv2}, \eqref{eq:r1+} and \eqref{eq:Im+} respectively. Therefore, from \eqref{eq:IR}, after direct algebraic calculations, we can derive that \begin{equation}\label{eq:IR1} \begin{aligned} I_1^- + I_1^+ & = (2 {\mathrm{i}} \lambda k_p^2 a_0-2 \mu k_s^2 b_0) s^{-2} + R_{r,2}- r_{I_1^+,0} + r_{I_1^-,1} -r_{I_1^+,2} \end{aligned} \end{equation} where \begin{equation}\label{eq:r2} \begin{aligned} R_{r,2} & = 6 \big[{\mathrm{i}} \lambda k_p^3 a_1 - \mu k_s^3 b_1 + \eta_2 (k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) \mathrm{e}^{- {\mathrm{i}} \varphi_0}\big] s^{-4}\\ & + 30 \big[- {\mathrm{i}} (2 \lambda + \mu) k_p^4 a_0 + \mu k_s^4 b_0 + {\mathrm{i}} ( \lambda - 3 \mu) k_p^4 a_2 - 4 \mu k_s^4 b_2\\ & + 3 \mu ({\mathrm{i}} k_p^4 a_0 + k_s^4 b_0) \mathrm{e}^{- 2 {\mathrm{i}} \varphi_0} \big] s^{-6} \end{aligned} \end{equation} Substituting \eqref{eq:IR1} into \eqref{eq:CGO2}, one can deduce that \begin{equation}\notag \label{eq:1a} ( 2 {\mathrm{i}} \lambda k_p^2 a_0-2 \mu k_s^2 b_0) s^{-2}= I_3 - I_2 - R_{r,2} + r_{I_1^+,0} - r_{I_1^-,1} + r_{I_1^+,2}, \end{equation} which can be used to further derive that \begin{equation}\label{eq:1b} 2 {\mathrm{i}} \lambda k_p^2 a_0-2 \mu k_s^2 b_0 = s^2 (I_3 - I_2 - R_{r,2} + r_{I_1^+,0} - r_{I_1^-,1} + r_{I_1^+,2}). \end{equation} From \eqref{eq:1b}, one can show that \begin{equation}\label{eq:1c} \left|2 {\mathrm{i}} \lambda k_p^2 a_0-2 \mu k_s^2 b_0 \right|= s^2 (\left|I_3\right| +\left| I_2\right| +\left| R_{r,2}\right| + |r_{I_1^+,0}| + | r_{I_1^-,1}| + |r_{I_1^+,2}|). \end{equation} Since $\Gamma_h^-$ is a rigid line of $\bmf{u}$, then from Lemma \ref{lem:rigid}, \eqref{eq:2} holds. Substituting the second equation of \eqref{eq:2} into \eqref{eq:I3}, we can obtain \eqref{eq:57 I_3b}. In \eqref{eq:1c}, using \eqref{eq:I2I4}, \eqref{eq:54 r}, \eqref{eq:54 r1}, \eqref{eq:57 I_3b}, \eqref{eq:54 r3} and \eqref{eq:r2}, letting $s\rightarrow +\infty$, we can derive that \begin{equation}\label{eq:mix up 2} {\mathrm{i}} \lambda k_p^2 a_0- \mu k_s^2 b_0=0. \end{equation} Recall that $\Gamma_h^-$ is a rigid line of $\bmf{u}$. Combining \eqref{eq:2} with \eqref{eq:mix up 2}, we have \begin{equation}\nonumber \left\{ \begin{array}{l} k_p ^2 a_0 - {\mathrm{i}} k_s^2 b_0=0,\\ {\mathrm{i}} \lambda k_p^2 a_0- \mu k_s^2 b_0=0. \end{array} \right. \end{equation} By virtue of \eqref{eq:convex}, since \begin{equation}\notag \left| \begin{array}{cc} k_p^2 & -{\mathrm{i}} k_s^2\\ {\mathrm{i}} \lambda k_p^2 & - \mu k_s^2 \end{array} \right|=-(\lambda+\mu) k_p^2 k_s^2\neq 0, \end{equation} then $ a_0=b_0=0. $ Since $\Gamma_h^-$ is a rigid line of $\bmf{u}$, from \eqref{eq:a1b1 lem}, we know that $a_1=b_1=0$. Therefore, from Lemma \ref{lem:rigid} and tje strong unique continuation principle, we have $\bmf{u}\equiv \bmf{0} $ in $\Omega$, which readily completes the proof. \end{proof} \begin{thm}\label{thm:55} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. Suppose there exit two intersecting lines $\Gamma_h^\pm$ of $\bmf{u}$ such that $\Gamma_h^-$ is a traction-free line and $\Gamma_h^+$ is an impedance line associated with a nonzero impedance constant $\eta_2\in \mathbb C$, with the property that $\angle(\Gamma_h^+,\Gamma_h^-)=\varphi_0\neq \pi $ and $\bmf{u}$ vanishes at the intersecting point, namely $\bmf{u}(\bmf{0})=\bmf{0}$, then $\bmf{u}\equiv \bmf{0}$. \end{thm} \begin{proof} Since $\Gamma_h^+ \in {\mathcal I}_\Omega^\kappa$, then $T_\nu \bmf{u}=-\eta_2 \bmf{u}$ on $\Gamma_h^+$. Recall that $I_1^+$ is defined in \eqref{eq:CGO6}. Therefore we have \eqref{eq:Im+}. Recall that $\Gamma_h^- \in {\mathcal T}_\Omega^\kappa$ and $I_1^-$ is defined in \eqref{eq:CGO6}. We have \eqref{eq:Tvdotu2}. Using \eqref{eq:Im+} and \eqref{eq:Tvdotu2}, we can derive that \begin{equation}\label{eq:2a} \begin{aligned} I_1^- + I_1^+ & =6 \eta_2(k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) \mathrm e^{- {\mathrm{i}} \varphi_0} s^{-4} -90 \mu (1-\mathrm{e}^{-2 {\mathrm{i}} \varphi_0})({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6} \\ & - r_{I_1^+,2}- r_{I_1^-,2} -r_{I_1^+,0}, \end{aligned} \end{equation} where $r_{I_1^+,2} $, $ r_{I_1^-,2} $ and $ r_{I_1^+,0}$ are defined in \eqref{eq:r1+}, \eqref{eq:r1-} and \eqref{eq:Im+} respectively. Substituting \eqref{eq:2a} into \eqref{eq:CGO2}, one can deduce that \begin{equation}\notag \label{eq:2b} \begin{aligned} 6 \eta_2(k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) \mathrm e^{- {\mathrm{i}} \varphi_0} s^{-4} & = I_3 - I_2 + r_{I_1^+,0} + r_{I_1^+,2} + r_{I_1^-,2} + 90 \mu (1-\mathrm{e}^{-2 {\mathrm{i}} \varphi_0})({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6}, \end{aligned} \end{equation} which can further give that \begin{equation}\label{eq:2c} \begin{aligned} 6 \eta_2(k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) \mathrm e^{- {\mathrm{i}} \varphi_0} & = s^4\big(I_3 - I_2 + r_{I_1^+,0} + r_{I_1^+,2} + r_{I_1^-,2} + 90 \mu (1-\mathrm{e}^{-2 {\mathrm{i}} \varphi_0})({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6}\big). \end{aligned} \end{equation} From \eqref{eq:2c}, it yields that \begin{equation}\label{eq:2d}\ \begin{aligned} \left|6 \eta_2(k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) \right| & \leq s^4 \big(\left|I_3\right| +\left| I_2\right| + |r_{I_1^+,0}| + | r_{I_1^+,2}| + |r_{I_1^-,2}|\\ & +| 90 \mu (1-\mathrm{e}^{-2 {\mathrm{i}} \varphi_0})({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6} | \big), \end{aligned} \end{equation} In \eqref{eq:2d}, using \eqref{eq:I2I4}, \eqref{eq:I3}, \eqref{eq:54 r1} and \eqref{eq:54 r3} and letting $s\rightarrow +\infty$, we can obtain that \begin{equation}\label{eq:18e} k_p^2 a_0- {\mathrm{i}} k_s^2 b_0=0, \end{equation} Since $\Gamma_h^-$ is a traction-free line of $\bmf{u}$, from Lemma \ref{lem:tranction free}, we know that \eqref{eq:8} and \eqref{eq:10} hold. Combining \eqref{eq:8} with \eqref{eq:18e}, we can obtain that $a_0=b_0=0$. Since $\bmf{u}(\bmf{0)}=\bmf{0}$, we know that \eqref{eq:u0} holds. Combining \eqref{eq:10} with \eqref{eq:u0} , it is easy to see that $a_1=b_1=0$. Therefore, from Lemma \ref{lem:tranction free} and Proposition \ref{prop:21}, we have $ \bmf{u} \equiv \bmf{0} \mbox { in } \Omega, $ which completes the proof. \end{proof} \begin{lem}\label{lem:Im} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. Suppose that there are two impedance lines $\Gamma_h^\pm$ intersecting with each other and satisfying \begin{equation*} \begin{split} T_{\bmf{\nu}}\bmf{u}+\eta_1 \bmf{u}&=\bmf{0} \mbox{ on } \Gamma_h^-,\\ T_{\bmf{\nu}}\bmf{u}+\eta_2 \bmf{u}&=\bmf{0} \mbox{ on } \Gamma_h^+, \end{split} \end{equation*} with $\eta_\ell \in \mathbb {C} \backslash \{0\}$. If $\angle(\Gamma_h^+,\Gamma_h^-)=\varphi_0$ and \begin{equation}\label{eq:lem53 cond} \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+ \eta_1 \neq 0, \end{equation} then \begin{equation}\label{eq:lem53 eq1} k_p^2 a_0-{\mathrm{i}} k_s^2 b_0=0. \end{equation} Furthermore, if $a_0=b_0=\ldots=a_{\ell-1}=b_{\ell-1}=0$, under \eqref{eq:lem53 cond}, we have \begin{equation}\label{eq:lem53 eq2} k_p^{\ell+2} a_{\ell}-{\mathrm{i}} k_s^{\ell+2} b_{\ell}=0. \end{equation} \end{lem} \begin{proof} Since $\Gamma_h^+$ is an impedance line of $\bmf{u}$ with the impedance parameter $\eta_2$, then \begin{equation}\label{eq:528Tvu} T_\nu\bmf{u}=-\eta_2 \bmf{u} \end{equation} Recall that $I_1^\pm$ are defined in \eqref{eq:CGO6}. Using \eqref{eq:528Tvu}, substituting \eqref{eq:Tvu} into \eqref{eq:CGO6}, we can obtain that \begin{equation}\label{eq:Im+} \begin{aligned} I_1^+ & = \int_{\Gamma_h^+}\left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma =-\int_{\Gamma_h^+}\left[\left(\eta_2 \bmf{u}\right) \cdot \bmf{v}+\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma\\ & =\frac{\eta_2}{2} (k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) {\mathrm e}^{{\mathrm{i}} \varphi_0} \int_0^h {\mathrm e}^{s\sqrt{r} \zeta{(\varphi_0} )} r \mathrm{d} r- r_{I_1^+,0}\\ &\quad + s \mu\zeta(\varphi_0) \int_0^h {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) }\bigg\{\frac{1}{2}({\mathrm{i}} k_p^2 a_0 + k_s^2 b_0) \mathrm e^{{\mathrm{i}} \varphi_0} r^{1/2}\\ & \quad + \frac{1}{8}({\mathrm{i}} k_p^3 a_1 + k_s^3 b_1) \mathrm e^{2 {\mathrm{i}} \varphi_0} r^{3/2} + \frac{1}{48}({\mathrm{i}} k_p^4 a_2 + k_s^4 b_2) \mathrm e^{3 {\mathrm{i}} \varphi_0} r^{5/2}\\ &\quad - \frac{1}{16}({\mathrm{i}} k_p^4 a_0 + k_s^4 b_0) \mathrm e^{{\mathrm{i}} \varphi_0} r^{5/2}\bigg\}\mathrm{d} r - r_{I_1^+,2} \\ & = -2 \mu({\mathrm{i}} k_p^2 a_0+ k_s^2 b_0)s^{-2}-6 \mu({\mathrm{i}} k_p^3 a_1+ k_s^3 b_1)s^{-4}-120 \mu({\mathrm{i}} k_p^4 a_2+ k_s^4 b_2)s^{-6}\\ & \quad +90 \mathrm{e}^{-2 {\mathrm{i}} \varphi_0}\mu({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6}-r_{I_1^+,2} + 6 \eta_2(k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) \mathrm e^{- {\mathrm{i}} \varphi_0} s^{-4} - r_{I_1^+,0}, \end{aligned} \end{equation} where $r_{I_1^+,0}=\int_0^h {\mathrm e}^{s \sqrt{ r} \zeta(\varphi_0 ) } R_{0,\Gamma_h^+} \mathrm{d} r$ and $r_{I_1^+,2}$ is defined in \eqref{eq:r1+}. Similarly, we have \begin{equation}\label{eq:Im-} \begin{aligned} I_1^- & = \int_{\Gamma_h^-}\left[\left({T}_{\nu} \bmf{u}\right) \cdot \bmf{v}-\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma =-\int_{\Gamma_h^-}\left[\left(\eta_1 \bmf{u}\right) \cdot \bmf{v}+\left({T}_{\nu} \bmf{v}\right) \cdot \bmf{u}\right] \mathrm{d} \sigma\\ & =\frac{\eta_1}{2} (k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) \int_0^h {\mathrm e}^{ - s\sqrt{r}} r \mathrm{d} r + s \mu \int_0^h {\mathrm e}^{- s \sqrt{ r} }\bigg\{\frac{1}{2}({\mathrm{i}} k_p^2 a_0 + k_s^2 b_0) r^{1/2}\\ &\quad + \frac{1}{8}({\mathrm{i}} k_p^3 a_1 + k_s^3 b_1) r^{3/2} + \frac{1}{48}({\mathrm{i}} k_p^4 a_2 + k_s^4 b_2) r^{5/2}- \frac{1}{16}({\mathrm{i}} k_p^4 a_0 + k_s^4 b_0) r^{5/2}\bigg\}\mathrm{d} r \\ &\quad - r_{I_1^-,0}- r_{I_1^-,2} \\ & = 2 \mu({\mathrm{i}} k_p^2 a_0+ k_s^2 b_0)s^{-2}+6 \mu({\mathrm{i}} k_p^3 a_1+ k_s^3 b_1)s^{-4}+120 \mu({\mathrm{i}} k_p^4 a_2+ k_s^4 b_2)s^{-6}\\ &\quad -90 \mu({\mathrm{i}} k_p^4 a_0+ k_s^4 b_0)s^{-6} - r_{I_1^-,2} + 6 \eta_1(k_p^2 a_0 - {\mathrm{i}} k_s^2 b_0) s^{-4} - r_{I_1^-,0}, \end{aligned} \end{equation} where $r_{I_1^-,0}=\int_0^h {\mathrm e}^{- s \sqrt{ r} } R_{0,\Gamma_h^-} \mathrm{d} r$ and $r_{I_1^-,2}$ is defined in \eqref{eq:r1-}. Therefore, combining \eqref{eq:Im+} and \eqref{eq:Im-}, after direct algebraic calculations, we can derive that \begin{equation}\label{eq:EFI2} \begin{aligned} I_1^++I_1^- & =(6 \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+6 \eta_1) (k_p^2 a_0-{\mathrm{i}} k_s^2 b_0) s^{-4}- r_{I_1^-,0}-r_{I_1^+,0}\\ & -90 \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^4 a_0+ k_s^4 b_0)s^{-6} - r_{I_1^-,2}-r_{I_1^+,2}. \end{aligned} \end{equation} By virtue of \eqref{eq:Ruv+} and \eqref{eq:R3 s}, it can be directly verified that \begin{equation}\label{eq:54 r3} |r_{I_1^+,0}| \leq S_0 \cdot{\mathcal O}( s^{-6} ),\ \ |r_{I_1^-,0}| \leq S_0 \cdot{\mathcal O}( s^{-6} ). \end{equation} Substituting \eqref{eq:EFI2} into \eqref{eq:CGO2}, we can obtain that \begin{equation} \notag \begin{split} (6 \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+6 \eta_1) (k_p^2 a_0-{\mathrm{i}} k_s^2 b_0) s^{-4}&=I_3-I_2+r_{I_1^-,0}+r_{I_1^+,0}+r_{I_1^-,2}+r_{I_1^+,2}\\ & + 90 \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^4 a_0+ k_s^4 b_0)s^{-6}, \end{split} \end{equation} which can further give that \begin{equation}\label{eq:56 int} \begin{split} (6 \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+6 \eta_1) (k_p^2 a_0-{\mathrm{i}} k_s^2 b_0)&= s^4\bigg(I_3-I_2+r_{I_1^-,0}+r_{I_1^+,0}+r_{I_1^-,2}+r_{I_1^+,2}\\ & + 90 \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^4 a_0+ k_s^4 b_0)s^{-6}\bigg) \end{split} \end{equation} From \eqref{eq:56 int}, it yields that \begin{equation}\label{eq:56 ineq} \begin{aligned} \left| 6 \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+6 \eta_1) (k_p^2 a_0-{\mathrm{i}} k_s^2 b_0) \right| & \leq s^4 \bigg(|I_3|+\left|I_2\right|+ \left| r_{I_1^-,0} \right|+\left|r_{I_1^+,0}\right| + \left| r_{I_1^-,2} \right|+\left|r_{I_1^+,2}\right|\\ & + \left| 90 \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^4 a_0+ k_s^4 b_0)s^{-6}\right| \bigg). \end{aligned} \end{equation} In \eqref{eq:56 ineq}, using \eqref{eq:I3}, \eqref{eq:arc}, \eqref{eq:54 r1} and \eqref{eq:54 r3}, and letting $s\rightarrow +\infty$, under \eqref{eq:lem53 cond}, we can obtain that \eqref{eq:lem53 eq1}. Suppose that $a_0=b_0=\ldots=a_{\ell-1}=b_{\ell-1}=0$, by direct calculations and using \eqref{eq:Tvu guina1new} and \eqref{eq:Ruv guina}, we can derive that \begin{equation}\label{eq:EFI3} \begin{aligned} I_1^++I_1^- & =\frac{(2 \ell + 3)!}{2^\ell (\ell+1)!}( \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+ \eta_1) (k_p^{\ell+2} a_\ell-{\mathrm{i}} k_s^{\ell+2} b_\ell) s^{-2 \ell-4}- \hat{r}_{I_1^-,0}-\hat{r}_{I_1^+,0}\\ & -\frac{(2 \ell+6)!}{2^{\ell+2}(\ell+2)!} \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^{\ell+4} a_\ell + k_s^{\ell+4} b_\ell)s^{-2 \ell-6} - \hat{r}_{I_1^-,2}-\hat{r}_{I_1^+,2}, \end{aligned} \end{equation} where \begin{equation*} \begin{aligned} & \hat{r}_{I_1^+,0}=\int_0^h \mathrm{e}^{s \sqrt{r} \zeta(\varphi_0)} \hat{R}_{0,\Gamma_h^+} \mathrm{d} r,\quad \hat{r}_{I_1^-,0}=\int_0^h \mathrm{e}^{ - s \sqrt{r}} \hat{R}_{0,\Gamma_h^-} \mathrm{d} r,\\ & \hat{r}_{I_1^+,2}={\mathrm{i}} s \mu \zeta(\varphi_0) \int_0^h \mathrm{e}^{s \sqrt{r} \zeta(\varphi_0)} \hat{R}_{2,\Gamma_h^+} \mathrm{d} r ,\quad \hat{r}_{I_1^-,2}={\mathrm{i}} s \mu \int_0^h \mathrm{e}^{ - s \sqrt{r} } \hat{R}_{2,\Gamma_h^-} \mathrm{d} r. \end{aligned} \end{equation*} Here $\hat{R}_{2,\Gamma_h^+}$, $\hat{R}_{2,\Gamma_h^-}$, $\hat{R}_{0,\Gamma_h^+}$ and $\hat{R}_{0,\Gamma_h^-}$ are defined in \eqref{eq:RTvu++}, \eqref{eq:RTvu--}, \eqref{eq:Ruv++} and \eqref{eq:Ruv--} respectively. From \eqref{eq:R4 s} and \eqref{eq:R5 s}, using \eqref{eq:268b int}, it is readily seen that \begin{equation}\label{eq:54 r4} \begin{aligned} & |\hat{r}_{I_1^+,0}| \leq \hat{S}_0 \cdot{\mathcal O}( s^{-2\ell-6} ),\quad |\hat{r}_{I_1^-,0}| \leq \hat{S}_0 \cdot{\mathcal O}( s^{-2\ell-6} ),\\ & |\hat{r}_{I_1^+,2}| \leq \hat{S}_3 \cdot{\mathcal O}( s^{-2\ell-6} ),\quad |\hat{r}_{I_1^-,2}| \leq \hat{S}_3 \cdot{\mathcal O}( s^{-2\ell-6} ), \end{aligned} \end{equation} where $\hat{S}_0$ and $\hat{S}_3 $ are defined in \eqref{eq:RTvu+Shat} and \eqref{eq:uvShat} respectively. Substituting \eqref{eq:EFI3} into \eqref{eq:CGO2}, we can obtain that \begin{equation} \notag \begin{split} & \frac{(2 \ell + 3)!}{2^\ell (\ell+1)!}( \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+ \eta_1) (k_p^{\ell+2} a_\ell-{\mathrm{i}} k_s^{\ell+2} b_\ell) s^{-2 \ell-4} \\ & =I_3-I_2+\hat{r}_{I_1^-,0}+\hat{r}_{I_1^+,0}+\hat{r}_{I_1^-,2}+\hat{r}_{I_1^+,2}\\ & \quad + \frac{(2 \ell+6)!}{2^{\ell+2}(\ell+2)!} \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^{\ell+4} a_\ell + k_s^{\ell+4} b_\ell)s^{-2 \ell-6} \end{split} \end{equation} which can be further reduced to \begin{equation}\label{eq:57 int} \begin{split} & \frac{(2 \ell + 3)!}{2^\ell (\ell+1)!}( \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+ \eta_1) (k_p^{\ell+2} a_\ell-{\mathrm{i}} k_s^{\ell+2} b_\ell) \\ & =s^{2\ell+4}\bigg(I_3-I_2+\hat{r}_{I_1^-,0}+\hat{r}_{I_1^+,0}+\hat{r}_{I_1^-,2}+\hat{r}_{I_1^+,2}\\ & \quad + \frac{(2 \ell+6)!}{2^{\ell+2}(\ell+2)!} \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^{\ell+4} a_\ell + k_s^{\ell+4} b_\ell)s^{-2 \ell-6}\bigg), \end{split} \end{equation} From \eqref{eq:57 int}, we can readily have \begin{equation}\label{eq:57 ineq} \begin{aligned} & \left| \frac{(2 \ell + 3)!}{2^\ell (\ell+1)!}( \eta_2 \mathrm{e}^{-{\mathrm{i}} \varphi_0}+ \eta_1) (k_p^{\ell+2} a_\ell-{\mathrm{i}} k_s^{\ell+2} b_\ell) \right|\\ & \leq s^{2\ell+4} \bigg(|I_3|+\left|I_2\right|+ \left| \hat{r}_{I_1^-,0} \right|+\left|\hat{r}_{I_1^+,0}\right| + \left| \hat{r}_{I_1^-,2} \right|+\left|\hat{r}_{I_1^+,2}\right|\\ & + \left| \frac{(2 \ell+6)!}{2^{\ell+2}(\ell+2)!} \mu (1-\mathrm{e}^{-2{\mathrm{i}} \varphi_0})({\mathrm{i}} k_p ^{\ell+4} a_\ell + k_s^{\ell+4} b_\ell)s^{-2 \ell-6}\right| \bigg). \end{aligned} \end{equation} In \eqref{eq:57 ineq}, using \eqref{eq:I3+}, \eqref{eq:arc} and \eqref{eq:54 r4}, and letting $s\rightarrow +\infty$, one finally sees that \eqref{eq:lem53 eq2} hold under \eqref{eq:lem53 cond}, which completes the proof. \end{proof} \begin{rem}\label{rem:imp} If the impedance parameters $\eta_1=\eta_2$, one can directly verify that \eqref{eq:lem53 cond} is equivalent to $\varphi_0\neq \pi$. \end{rem} \begin{thm}\label{eq:impedance line exp} Let $\mathbf{u}\in L^2(\Omega)^2$ be a solution to \eqref{eq:lame}. Suppose there exist two impedance lines $\Gamma_h^\pm$ of $\bmf{u}$ such that $\angle(\Gamma_h^+,\Gamma_h^-)=\varphi_0\neq \pi$ satisfying that \eqref{eq:lem53 cond} is fulfilled and $\bmf{u}$ vanishes at the intersecting point, i.e. $\bmf{u}(\bmf{0})=\bmf{0}$, then $\bmf{u}\equiv \bmf{0}$. Here, $T_\nu \bmf{u}+\eta_2\bmf{u}=\mathbf{0}$ on $\Gamma_h^+$, $T_\nu \bmf{u}+\eta_1\bmf{u}=\mathbf{0}$ on $\Gamma_h^-$ and $\eta_\ell\in \mathbb{C}\backslash\{0\}, \ell=1,2.$. \end{thm} \begin{proof} Since that $\Gamma_h^-$ is an impedance line of $\bmf{u}$. From Lemma \ref{lem:impedance eqn}, we know that \eqref{eq:Tu+u1} and \eqref{eq:Tu+u2} hold. Under the condition \eqref{eq:lem53 cond}, from Lemma \ref{lem:Im}, we have \eqref{eq:lem53 eq1}. Combining \eqref{eq:Tu+u1} with \eqref{eq:lem53 eq1}, we can derive that $a_0=b_0=0$. Therefore, using Lemma \ref{lem:Im} again, under \eqref{eq:lem53 cond}, from \eqref{eq:lem53 eq2}, we have \begin{equation}\label{eq:26C} k_p^3 a_1 -{\mathrm{i}} k_s^3 b_1=0. \end{equation} Since $\bmf{u}(\bf{0})=\bmf{0}$, from Lemma \ref{lem:34 three conds}, \eqref{eq:u0} holds. Combining \eqref{eq:u0} and \eqref{eq:26C}, one readily has that \begin{equation}\notag \left\{ \begin{array}{l} k_p^3 a_1 -{\mathrm{i}} k_s^3 b_1=0,\\ k_p a_1 + {\mathrm{i}} k_s b_1=0. \end{array} \right. \end{equation} Since \begin{equation}\notag \left| \begin{array}{cc} k_p^3 & -{\mathrm{i}} k_s^3\\ k_p & {\mathrm{i}} k_s\\ \end{array} \right|={\mathrm{i}} k_p k_s( k_p^2 + k_s^2)\neq 0, \end{equation} we can derive that $a_1=b_1=0$. Using Lemma \ref{lem:impedance eqn} and Proposition \ref{prop:21}, we readily have $\bmf{u}\equiv \bmf{0}$. The proof is complete. \end{proof} \section{Unique identifiability for inverse elastic obstacle problems}\label{sect:6} In this section, as an important and practical application, we apply the theoretical findings in the previous sections to the study of the unique identifiability for the inverse elastic obstacle problem. As mentioned earlier, we shall refer to the theoretical findings in the previous sections as the generalized Holmgren's principle. The inverse problem is concerned with recovering the geometrical shape of a certain unknown object by using the elastic wave probing data. The inverse elastic obstacle problem arises from industrial applications of practical importance, e.g. in the geophysical exploration. We next introduce the mathematical setup of the inverse obstacle problem that expands the abstract formation \eqref{eq:ipa1}. Let $\Omega\subset\mathbb{R}^2$ be a bounded Lipschitz domain such that $\mathbb{R}^2\backslash\bar{\Omega}$ is connected. Let $\bmf{u}^i$ be an incident elastic wave field, which is a time-harmonic elastic plane wave of the form \begin{equation}\label{eq:ui} \bmf{u}^i:=\bmf{u}^i(\mathbf{ x};k_p,k_s, \mathbf{d})=\alpha_{p} \bmf{d} \mathrm{e}^{\mathrm{i} k_{p} \bmf{x} \cdot \bmf{d} }+\alpha_{s} \bmf{d}^{\perp} \mathrm{e}^{\mathrm{i} k_s \bmf{x} \cdot \bmf{d} }, \quad \alpha_{p}, \alpha_{s} \in \mathbb{C}, \quad\left|\alpha_{p}\right|+\left|\alpha_{s}\right| \neq 0 \end{equation} where $\mathbf{d}\in\mathbb{S}^1$ denotes the incident direction, $\bmf{d}^{\perp}$ is orthogonal to $\bmf{d}$, $k_p$ and $k_s$ are compressional and shear wave numbers defined in \eqref{eq:kpks}. Physically speaking, $\bmf{u}^i$ is the detecting wave field and $\Omega$ denotes an impenetrable obstacle which interrupts the propagation of the incident wave and generates the corresponding scattered wave field $\bmf{u}^{\mathrm{sc}}$. The scattered field $\bmf{u}^{\mathrm{sc} }$ in ${\mathbb R}^2 \backslash \Omega$ can be decomposed into the sum of the compressional part $\bmf{u}^{\mathrm{sc}}_p$ and the shear part $\bmf{u}^{\mathrm{sc}}_s$ as follows \begin{equation} \bmf{u}^{\mathrm{s c} }=\bmf{u}_{p}^{\mathrm {s c} }+\bmf{u}_{s}^{\mathrm{s c} }, \quad \bmf{u}_{p}^{\mathrm {s c} }=-\frac{1}{k_{p}^{2}} \nabla\left( \nabla \cdot \bmf{ u}^{\mathrm {s c} }\right ), \quad\bmf{ u}_{s}^{\mathrm{s c}}=\frac{1}{k_{s}^{2}} \bf{curl} \operatorname{curl} u^{\mathrm {s c} }, \end{equation} where \[ {\rm curl}\bmf{ u}=\partial_1 u_2-\partial_2 u_1, \quad {\bf curl}{u}=(\partial_2 u, -\partial_1 u)^\top. \] Let $\omega=\sqrt{ \kappa}$ be the angular frequency., where $\kappa$ is the Lam\'e eigenvalue of \eqref{eq:lame}. Define $\bmf{u}:=\bmf{u}^i+\bmf{u}^{\mathrm{ sc} }$ to be the total wave field, then the forward scattering problem of this process can be described by the following system, \begin{equation}\label{forward} \begin{cases} & {\mathcal L} \bmf{ u} + \omega^2 \bmf{u} = 0\qquad\quad \mbox{in }\ \ \mathbb{R}^2\backslash\overline{\Omega},\medskip\\ & \bmf{u} =\bmf{u}^i+\bmf{u}^{\mathrm{sc} }\hspace*{1.56cm}\mbox{in }\ \ \mathbb{R}^2,\medskip\\ & \mathscr{B}(\bmf{u})=\bmf{0}\hspace*{1.95cm}\mbox{on}\ \ \partial\Omega,\medskip\\ &\displaystyle{ \lim_{r\rightarrow\infty}r^{\frac{1}{2}}\left(\frac{\partial \bmf{u}_\beta^{\mathrm{sc} }}{\partial r}-\mathrm{i}k_\beta \bmf{u}_\beta^{\mathrm{sc} }\right) =\,0,} \quad \beta=p,s, \end{cases} \end{equation} where the last equation is the Kupradze radiation condition that holds uniformly in $\hat{\mathbf{ x}}:=\mathbf{ x}/|\mathbf{ x}|\in\mathbb{S}^1$. The boundary condition $\mathscr{B}(u)$ on $\partial \Omega$ could be either of the following three conditions: \begin{enumerate} \item the Dirichlet condition ($\Omega$ is a rigid obstacle): $\mathscr{B}(\bmf{u})=\bmf{u}$; \item the Neumann condition ($\Omega$ is a traction-free obstacle): $\mathscr{B}(\bmf{u})=T_\nu \bmf{u}$; \item the impedance condition ($\Omega$ is an impedance obstacle): $\mathscr{B}(\bmf{u})=T_\nu \bmf{u}+\eta \bmf{u},\ \Re(\eta)\geq 0 \mbox{ and } \Im(\eta)>0$, \end{enumerate} where $\nu$ denotes the exterior unit normal vector to $\partial\Omega$, $\boldsymbol{\tau}= \nu^\perp$ and the boundary traction operator $T_\nu$ is defined in \eqref{eq:Tu}. Moreover, in the impedance condition given above, $\eta\in L^\infty(\partial\Omega)$, and this is different from our study in the previous sections, where the impedance $\eta$ is always required to be a constant. We would also like to point out that the conditions $\Re(\eta)\geq 0 \mbox{ and } \Im(\eta)>0$ are the physical requirement. The elastic system \eqref{forward} associated with either of the three kinds of boundary conditions is well understood with a unique solution $ \bmf{u} \in H^1_{\mathrm{loc} } ({\mathbb R}^2 \backslash \Omega )$. We refer to \cite{ElschnerYama2010,Kupradze} for the related results. It is known that the compressional and shear parts $\bmf{u}_\beta^{\mathrm {sc} }$ ($\beta=p,s$) of a radiating solution $\bmf{u}^{\mathrm{sc} }$ to the elastic system \eqref{forward} possess the following asymptotic expansions \begin{equation} \begin{aligned} \bmf{u}_{p}^{\mathrm{sc}}(\bmf{x}; k_p,k_s, \mathbf{d}) &=\frac{\mathrm{e}^{\mathrm{i} k_{p} r}}{\sqrt{r}}\left\{u_{p}^{\infty}(\hat{\bmf{x}}; \bmf{d} ) \hat{\bmf{x}}+{\mathcal O}\left(\frac{1}{r}\right)\right\} \\ \bmf{u}_{s}^{\mathrm{sc}}(\bmf{x};k_p,k_s, \mathbf{d}) &=\frac{\mathrm{e}^{\mathrm{i} k_{s} r}}{\sqrt{r}}\left\{u_{s}^{\infty}(\hat{\bmf{x}}; \bmf{d}) \hat{\bmf{x}}^{\perp}+{\mathcal O}\left(\frac{1}{r}\right)\right\} \end{aligned} \end{equation} as $r =|\bmf{x} | \rightarrow \infty$, where $u_{p}^{\infty}$ and $u_{s}^{\infty}$ are both scalar functions defined on $\mathbb S^1$. Hence, a Kupradze radiating solution has the asymptotic behavior $$ \bmf{u}^{\mathrm{sc}}(\bmf{x}; k_p,k_s, \mathbf{d})=\frac{\mathrm{e}^{\mathrm{i} k_{p} r}}{\sqrt{r}} u_{p}^{\infty}(\hat{\bmf{x} }; \bmf{d} ) \hat{\bmf{x}}+\frac{\mathrm{e}^{\mathrm{i} k_{s} r}}{\sqrt{r}} u_{s}^{\infty}(\hat{\bmf{x} }; \bmf{d} ) \hat{\bmf{x} }^{\perp}+{\mathcal O}\left(\frac{1}{r^{3 / 2}}\right) \quad \text { as } \quad r \rightarrow \infty $$ The far-field pattern $\bmf{u}^\infty$ of $\bmf{u}^{\mathrm{sc}} $ is defined as $$ \bmf{u}_t^{\infty}(\hat{\bmf{x}}; \bmf{d} ) :=u_{p}^{\infty}(\hat{\bmf{x}}; \bmf{d} ) \hat{\bmf{x}}+u_{s}^{\infty}(\hat{\bmf{x}}; \bmf{d} ) \hat{\bmf{x}}^{\perp}. $$ Obviously, the compressional and shear parts of the far-field are uniquely determined by $\bmf{u}^{\infty}$ as follows: $$ \bmf{u}_{p}^{\infty}(\hat{\mathbf{x}}; \bmf{d} )=\bmf{u}^{\infty}(\hat{\bmf{x}};\bmf{d} ) \cdot \hat{\bmf{x}}; \quad \bmf{u}_{s}^{\infty}(\hat{\bmf{x}},\bmf{d} )=\bmf{u}^{\infty}(\hat{\bmf{x}}; \bmf{d} ) \cdot \hat{\bmf{x}}^{\perp}. $$ The inverse elastic scattering problem corresponding to \eqref{forward} concerns the determination of the scatterer $\Omega$ (and $\eta$ as well in the impedance case) by knowledge of the far-field pattern $\bmf{u}_\beta^\infty(\hat{\mathbf{ x}},\mathbf{d},k)$, where $\beta=t,p$ or $s$. As in \eqref{eq:ipa1}, we introduce the operator $\mathcal{F}$ which sends the obstacle to the corresponding far-field pattern and is defined by the forward scattering system \eqref{forward}, the aforementioned inverse problem can be formulated as \begin{equation}\label{inverse} \mathcal{F}(\Omega, \eta)=\bmf{u}_\beta^\infty(\hat{\mathbf{ x}}; \mathbf{d}),\quad \beta=t,p, \mbox{ or } s. \end{equation} Next, we show that by using the generalized Holmgren's uniqueness principle, we can establish two novel unique identifiability results for \eqref{inverse} in determining an obstacle without knowing its a priori physical property as well as its possible surface impedance by at most four far-field patterns, namely $\bmf{u}_\beta^\infty(\hat{\mathbf{x}})$ corresponding to four distinct $\mathbf{d}$'s. \begin{defn}\label{def:61} Let $Q\subset\mathbb{R}^2$ be a polygon in $\mathbb{R}^2$ such that \begin{equation}\label{eq:edge} \partial Q=\cup_{j=1}^\ell \Gamma_j, \end{equation} where each $\Gamma_j$ is an edge of $\partial Q$. $Q$ is said to be a generalized impedance obstacle associated with \eqref{forward} if there exists a Lipschitz dissection of $\Gamma_j$, $1\leq j\leq \ell$, \[ \Gamma_j=\Gamma_D^j\cup\Gamma_N^j\cup\Gamma_I^j \] such that \begin{equation}\label{eq:66} \mathbf{ u} =\mathbf{0} \quad \text { on } \Gamma_{D}^j, \quad T_\nu \mathbf{u}=\mathbf{0} \quad \text { on } \Gamma_{N}^j, \quad T_\nu \mathbf{u}+\eta_j \mathbf{u}=\mathbf{0} \quad \text { on } \Gamma_{I}^j, \end{equation} where $\eta_j\in \mathbb{C}$ with $\Im\eta_j\geq 0$. \end{defn} It is emphasized that in \eqref{eq:66}, either $\Gamma_D^j, \Gamma_N^j$ or $\Gamma_I^j$ could be an empty set, and hence a generalized impedance obstacle could be purely a rigid obstacle, a traction-free obstacle, an impedance obstacle or a mixed type. Moreover, one each edge of the polygonal obstacle, the impedance parameter can take different (complex) values. In order to simply notations, we formally write $T_\nu\bmf{u}+\eta\bmf{u}$ with $\eta\equiv \infty$ to signify $T_\nu\bmf{u}=\bmf{0}$. In doing so, \eqref{eq:66} can be unified as $T_\nu\bmf{u}+\eta\bmf{u}=\bmf{0}$ on $\partial\Omega$ with \begin{equation}\label{eq:eta} \eta=0\cdot\chi_{\cup_{j=1}^\ell \Gamma_D^j}+\infty\cdot\chi_{\cup_{j=1}^\ell \Gamma_N^j}+\sum_{j=1}^\ell \eta_j\cdot\chi_{\Gamma_I^j}. \end{equation} We write $(Q,\eta)$ to denote a generalized polygonal impedance obstacle as describe above with $\eta\in L^\infty(\partial Q)\cup\{\infty\}$. In what follows, $(\Omega , \eta)$ is said to be an admissible complex obstacle if \begin{equation}\label{eq:p1} (\Omega, \eta)=\cup_{j=1}^p (\Omega_j, \eta_j), \end{equation} where each $(\Omega_j, \eta_j)$ is a generalized polygonal impedance obstacle such that $\Omega_j, j=1, 2,\ldots, p$ are pairwise disjoint and \begin{equation}\label{eq:r2b} \eta=\sum_{j=1}^p \eta_j\chi_{\partial\Omega_j},\quad \eta_j\in L^\infty(\partial\Omega_j)\cup\{\infty\}. \end{equation} \begin{thm}\label{thm:uniqueness1} Let $(\Omega, \eta)$ and $(\widetilde\Omega, \widetilde\eta)$ be two admissible complex obstacles. Let $\omega\in\mathbb{R}_+$ be fixed and $\mathbf{d}_\ell$, $\ell=1, 2,3, 4$ be four distinct incident directions from $\mathbb{S}^1$. Let $\bmf{u}_\beta^\infty$ and $\widetilde{\bmf{u}}^\infty_\beta$ be, respectively, the far-field patterns associated with $(\Omega, \eta)$ and $(\widetilde\Omega, \widetilde\eta)$, where $\beta=t,p, \mbox{ or } s$. If \begin{equation}\label{eq:cond1} \bmf{u}_\beta^\infty(\hat{\mathbf{ x}}; \mathbf{d}_\ell )=\widetilde{\bmf{u}}_\beta^\infty(\hat{\mathbf{ x}}; \mathbf{d}_\ell), \ \ \hat{\mathbf x}\in\mathbb{S}^1, \ell=1, \ldots, 4, \end{equation} then one has that \begin{equation}\label{eq:u1n} \Omega =\widetilde{\Omega}\mbox{ and } \eta=\widetilde \eta. \end{equation} \end{thm} Before giving the proof of Theorem \ref{thm:uniqueness1}, we first derive an auxiliary lemma as follows. \begin{lem}\label{lem:51} Let $\mathbf {\mathbf d}_{\ell}\in\mathbb{S}^1$, $\ell=1,\ldots, n$, be $n$ vectors which are distinct from each other. Suppose that $\Omega$ is a bounded Lipschitz domain and $\mathbb R^2\backslash \overline{\Omega } $ is connected. Let the incident elastic wave filed $\bmf{u}^i(\mathbf{ x};k_p,k_s, \mathbf{d}_\ell)$ be defined in \eqref{eq:ui}. Furthermore, suppose that the total elastic wave filed $\bmf{u}(\bmf{x}; k_p,k_s, \mathbf{d}_\ell)$ associated with $\bmf{u}^i(\mathbf{ x};k_p,k_s, \mathbf{d}_\ell)$ satisfies \eqref{forward}. Then {the following set of functions is linearly independent:} $$ \{\bmf{u}(\bmf{x}; k_p,k_s, \mathbf{d}_\ell);~\mathbf x \in D , \ \ \ell=1,2,\ldots, n \}, $$ where $D \subset \mathbb R^2 \backslash \overline \Omega $ is an open set. \end{lem} \begin{proof} The lemma can be proved by following a similar argument to the proof of Theorem 5.1 in \cite{CK}, and we skip the detailed calculations. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:uniqueness1}] By an absurdity argument, we first prove that if \eqref{eq:cond1} holds, one must have that $\Omega=\widetilde \Omega$. Suppose that $\Omega $ and $\widetilde{\Omega}$ are two different admissible complex obstacles such that $\Omega\neq \widetilde\Omega$ and \eqref{eq:cond1} holds. Let $\mathbf{G}$ denote the unbounded connected component of $\mathbb{R}^2\backslash\overline{(\Omega\cup\widetilde\Omega)}$. Then by a similar topological argument to that in \cite{Liu-Zou}, one can show that there exists a line segment $\Gamma_h\subset\partial\mathbf{G}\backslash\partial\Omega$ or $\Gamma_h\subset\partial\mathbf{G}\backslash\partial\widetilde\Omega$. Without loss of generality, we assume the former case. Let $\bmf{u}$ and $\widetilde{\bmf{u}}$ respectively denote the total wave fields to \eqref{forward} associated with $(\Omega, \eta)$ and $(\widetilde\Omega, \widetilde\eta)$. By \eqref{eq:cond1} and the Rellich theorem (cf. \cite{CK}), we know that \begin{equation}\label{eq:aa3} \bmf{u} (\mathbf x; k_p,k_s, \mathbf{d}_\ell)=\widetilde{\bmf{u} }(\mathbf x; k_p,k_s,\mathbf{d}_\ell),\quad \mathbf{x}\in\mathbf{G},\ \ell=1, \ldots, 4. \end{equation} By using \eqref{eq:aa3} as well as the generalized impedance boundary condition on $\partial\widetilde\Omega$, we readily have \begin{equation}\label{eq:aa4} T_\nu \bmf{u}+\widetilde\eta \bmf{u}=T_\nu \widetilde {\bmf{u}} +\widetilde\eta\widetilde{ \bmf{u}} =\bmf{0}\quad\mbox{on}\ \ \Gamma_{h}. \end{equation} Consider a fixed point $\mathbf x_0 \in \Gamma_h $. There exits a sufficient small positive number $\varepsilon \in {\mathbb R}_+$ such that $B_{2\varepsilon} (\mathbf x_0) \Subset \mathbf{G} $, where $B_{2\varepsilon }(\mathbf x_0)$ is a disk centered at $\mathbf x_0$ with the radius $2\varepsilon$. Let $\Gamma_\varepsilon = B_\varepsilon (\mathbf x_0) \cap \Gamma_h$, where $B_{\varepsilon }(\mathbf x_0)$ is a disk centered at $\mathbf{x_0}$ with the radius $\varepsilon$. It is also noted that $$ -{\mathcal L} \bmf{u}= \omega^2 \bmf{u} \mbox{ in } B_{2\varepsilon } (\mathbf x_0). $$ Recall that the unit normal vector $\nu $ and the tangential vector $\boldsymbol{\tau}$ to $\Gamma_h$ are defined in \eqref{eq:nutau}, respectively. Due to the linear dependence of four $\mathbb{C}^3$-vectors, it is easy to see that there exist four complex constants $a_\ell$ such that \begin{equation}\notag \sum_{\ell=1}^4 a_{\ell } \begin{bmatrix} \bmf{u} (\mathbf x_0; k_p,k_s, \mathbf{d}_\ell)\cr \boldsymbol{\tau} \cdot \partial_{\nu} \bmf{u} |_{\bmf{x}={\bmf{x}}_0 } \end{bmatrix}=\bmf{0}. \end{equation} Moreover, there exits at least one $a_\ell $ is not zero. Let \begin{equation}\label{eq:u513} \bmf{u}(\bmf{x};k_p,k_s)=\sum_{\ell=1}^4 a_{\ell } \bmf{u} (\mathbf x; k_p,k_s, \mathbf{d}_\ell). \end{equation} Then we have \begin{equation}\label{eq:611} \bmf{u}(\bmf{x}_0;k_p,k_s)=\bmf{0} \mbox{ and } \boldsymbol{\tau} \cdot \partial_{\nu} \bmf{u} |_{\bmf{x}={\bmf{x}}_0 }=0. \end{equation} Next we distinguish two separate cases. The first case is that $ \bmf{u}(\bmf{x};k_p,k_s)\equiv \bmf{0},\, \forall \bmf{x} \in \mathbf{G}$. In view of \eqref{eq:u513}, since $a_\ell $ are not all zero and $\bmf{d}_\ell$ are distinct, we readily have a contradiction by Lemma \ref{lem:51}. For the second case, we suppose that $ \bmf{u}(\bmf{x};k_p,k_s)\equiv\hspace*{-3mm}\backslash\ \bmf{0}$. In view of \eqref{eq:aa4} and \eqref{eq:611}, recalling Definition \ref{def:generalized line}, we know that $\Gamma_\varepsilon $ is a singular line of $\bmf{u}$, which implies that $\Gamma_\varepsilon $ could be a singular rigid, or singular traction-free or singular impedance line of $\bmf{u}$ in Definition \ref{def:generalized line}. Therefore, by the generalized Holmgren's principle (cf. Theorems \ref{Thm:31 singular rigid}, \ref{thm:traction free line} and \ref{thm:impedance line}), we obtain that \begin{equation}\label{eq:613 con} \bmf{u}\equiv \bmf{0} \mbox{ in } B_{2\varepsilon }(\bmf{x}_c), \end{equation} which is obviously a contradiction. Next, we prove that by knowing $\Omega =\widetilde{\Omega}$, one must have that $\eta \equiv \widetilde \eta$. Assume contrarily that $\eta \neq \widetilde \eta$. One can easily show that there exists an open subset $\Sigma$ of $\partial \Omega=\partial \widetilde \Omega$ such that $$ \mathbf{u}=T_\nu \mathbf{u}=\mathbf{0} \mbox { on } \Sigma. $$ Therefore by the classical Holmgren's principle, we know that $\mathbf{u}\equiv \mathbf{0}$ in ${\mathbb R}^2 \backslash \Omega $, which readily yields a contradiction. The proof is complete. \end{proof} \begin{rem} \eqref{eq:u1n} means that one can not only determine the shape of an admissible complex obstacle, but also its physical properties (in the case that $\eta=0$ or $\eta=\infty$). Furthermore, if it is of impedance type, one can determine the surface impedance parameter as well. \end{rem} Finally, we show that if fewer far-field patterns are used, one can establish a local uniqueness result in determining a generic class of admissible complex obstacles. To that end, we first introduce a geometric notion of the degree of an admissible complex obstacle. Let $\Omega$ be defined in \eqref{eq:p1} that consists of finitely many pairwise disjoint polygons. Let $\Gamma, \Gamma'\subset\partial\Omega$ be two adjacent edges of $\partial\Omega$. Extending $\Gamma$ and $\Gamma'$ into straight lines in the plane $\mathbb{R}^2$, we denote them by $\widehat{\Gamma}$ and $\widehat{\Gamma'}$. Clearly, the intersection of $\widehat{\Gamma}$ and $\widehat{\Gamma'}$ forms two angles, with one belonging to $(0,\pi/2]$ and the other one belonging to $[\pi/2, \pi)$. We write $\angle_{\rm acute}(\Gamma,\Gamma')$ to signify the former one. Define \begin{equation}\label{de:dg1} \mathrm{deg}(\Omega):=\max_{\Gamma, \Gamma'\in \partial\Omega}\{\angle_{\rm acute}(\Gamma, \Gamma')|\ \Gamma, \Gamma'\ \mbox{are two adjacent edges of}\ \partial\Omega\}. \end{equation} Moreover, we let $\zeta$ and $\zeta'$ respectively signify the values of $\eta$ on $\Gamma$ and $\Gamma'$ around the vertex formed by those two edges. It is noted that $\zeta$ and $\zeta'$ may be $0, \infty$ or finite and nonzero. An admissible complex obstacle $(\Omega, \eta)$ is said to belong to the class $\mathcal{C}$ if \begin{equation}\label{eq:cond1n} \mathrm{deg}(\Omega)<\varphi_{\sf root}, \end{equation} where $\varphi_{\sf root}$ is defined in \eqref{eq:varphi0}, and \begin{equation}\label{eq:cond2n} \zeta=\zeta'\quad \mbox{if both $\zeta$ and $\zeta'$ are finite and nonzero}, \end{equation} for all two adjacent edges $\Gamma, \Gamma'$ of $\partial\Omega$. \begin{thm}\label{thm:uniqueness2} Let $(\Omega, \eta)$ and $(\widetilde\Omega, \widetilde\eta)$ be two admissible complex obstacles from the class $\mathcal{C}$ as described above. Let $\omega \in\mathbb{R}_+$ be fixed and $\mathbf{d}_\ell$, $\ell=1, 2,3$ be three distinct incident directions from $\mathbb{S}^1$. Let $\mathbf{G}$ denote the unbounded connected component of $\mathbb{R}^2\backslash\overline{(\Omega\cup\widetilde\Omega)}$. Let $\bmf{u}_\beta^\infty$ and $\widetilde{\bmf{u}}^\infty_\beta$ be, respectively, the far-field patterns associated with $(\Omega, \eta)$ and $(\widetilde\Omega, \widetilde\eta)$, where $\beta=t,p, \mbox{ or } s$. If \begin{equation}\label{eq:cond1 corner} \bmf{u}_\beta^\infty(\hat{\mathbf{ x}},\mathbf{d}_\ell )=\widetilde{\bmf{u}}_\beta^\infty(\hat{\mathbf{ x}},\mathbf{d}_\ell), \ \ \hat{\mathbf x}\in\mathbb{S}^1, \ell=1, 2, 3, \end{equation} then one has that $$ \left(\partial \Omega \backslash \partial \overline{ \widetilde{\Omega }} \right )\cup \left(\partial \widetilde{\Omega } \backslash \partial \overline{ \Omega } \right) $$ cannot have a corner on $\partial \mathbf{G}$. \end{thm} \begin{proof} We prove the theorem by contradiction. Assume \eqref{eq:cond1 corner} holds but $ \left(\partial \Omega \backslash \partial \overline{ \widetilde{\Omega }} \right )\cup \left(\partial \widetilde{\Omega } \backslash \partial \overline{ \Omega } \right) $ has a corner $\mathbf x_c$ on $\partial \mathbf{G}$. Clearly, $\mathbf x_c$ is either a vertex of $\Omega$ or a vertex of $\widetilde\Omega$. Without loss of generality, we assume the latter case. Let $h\in\mathbb{R}_+$ be sufficiently small such that $B_h(\bmf{x}_c)\Subset\mathbb{R}^2\backslash\overline \Omega $. Moreover, since $\bmf{x}_c$ is a vertex of $\widetilde\Omega$, we can assume that \begin{equation}\label{eq:aa2} B_h(\mathbf x_c)\cap \partial\widetilde\Omega=\Gamma_h^\pm, \end{equation} where $\Gamma_h^\pm$ are the two line segments lying on the two edges of $\widetilde\Omega$ that intersect at $\mathbf x_c$. Furthermore, on $\Gamma_h^\pm$ the boundary conditions are given by \eqref{eq:66}. By \eqref{eq:cond1 corner} and the Rellich theorem (cf. \cite{CK}), we know that \begin{equation}\label{eq:aa5} \bmf{u} (\mathbf x; k_p,k_s, \mathbf{d}_\ell)=\widetilde{\bmf{u} }(\mathbf x; k_p,k_s, \mathbf{d}_\ell),\quad x\in\mathbf{G},\ \ell=1, 2, 3. \end{equation} It is clear that $\Gamma_h^\pm\subset\partial\mathbf{G}$. Hence, by using \eqref{eq:aa3} as well as the generalized boundary condition \eqref{eq:66} on $\partial\widetilde\Omega$, we readily have \begin{equation}\label{eq:aa4new} \partial_\nu \bmf{u}+\widetilde\eta \bmf{u}=\partial_\nu \widetilde {\bmf{u}} +\widetilde\eta\widetilde{ \bmf{u}} =\bmf{0}\quad\mbox{on}\ \ \Gamma_h^\pm. \end{equation} It is also noted that in $B_h(\mathbf x_c)$, $-{\mathcal L} \bmf{u}=\omega^2 \bmf{u}$. Due to the linear dependence of three $\mathbb{C}^2$-vectors, we see that there exits three complex constants $a_\ell $ such that $$ \sum_{\ell=1}^3 a_\ell \bmf{u}(\bmf{x}_c;k_p,k_s,\bmf{d}_\ell)=\bmf{0}, $$ where there exits at least one $a_\ell $ is not zero. Set $\bmf{u}(\bmf{x};k_p,k_s)=\sum_{\ell=1}^3 a_\ell \bmf{u}(\bmf{x};k_p,k_s,\bmf{d}_\ell)$. Then we know that \begin{equation}\label{eq:611new} \bmf{u}(\bmf{x}_c;k_p,k_s)=\bmf{0}. \end{equation} Similar to the proof of Theorem \ref{thm:uniqueness2}, we consider the following two cases. The first one is $ \bmf{u}(\bmf{x};k_p,k_s)\equiv \bmf{0},\, \forall \bmf{x}\in \bmf{G}$. Since there exits at one $a_\ell$ such that $a_\ell \neq 0$ and $\mathbf{d}_\ell$ are distinct, from Lemma \ref{lem:51}, we can arrive at a contradiction. The other case is that $ \bmf{u}(\bmf{x};k_p,k_s)\equiv\hspace*{-3mm}\backslash\ \bmf{0}$. By \eqref{eq:cond1n} and \eqref{eq:cond2n}, as well as the generalized Holmgren's principle in Theorems \ref{thm:rigid line exp}--\ref{eq:impedance line exp}, one can show that $$ \mathbf{u} \equiv \mathbf{0} \mbox{ in } \mathbf{G} $$ which yields a contradiction again. The proof is complete. \end{proof} \begin{rem} Following a similar argument, one can derive more unique identifiability results similar to Theorem~\ref{thm:uniqueness2}. For example, if one excludes the presence of $T_\nu \mathbf{u}=\mathbf{0}$ on any boundary portion in \eqref{eq:66} of Definition \ref{def:61}, then the assumption \eqref{eq:cond1n} in Theorem \ref{thm:uniqueness2} can be removed. We choose not to discuss the details about those extensions in this article. \end{rem} \section*{Acknowledgement} The research of H Liu was supported by Hong Kong RGC GRF grants, No. 12302017, 12301218 and 12302919. \section*{Appendix} \begin{proof}[Proof of Lemma~\ref{lem:Tuu exp}] We first prove \eqref{eq:Tu1}. Recall that $\bmf{\nu }\big|_{\Gamma_h^+}$ is defined in \eqref{eq:nu}: \begin{equation}\label{eq:normal der} \boldsymbol{\tau}=(-\cos\varphi_0,-\sin\varphi_0). \end{equation} Substituting \eqref{eq:normal der} into \eqref{eq:Tu} yields \begin{equation}\label{eq:Tu new exp} \begin{aligned} T_{\nu} \mathbf{u}\Big |_{\Gamma^+_h} & =2 \mu \left[\begin{array}{cc} \partial_1 u_1 & \partial_2 u_1\\ \partial_1 u_2 & \partial_2 u_2\\ \end{array}\right] \left[\begin{array}{c} -\sin \varphi_0\\ \cos \varphi_0\\ \end{array}\right] +\lambda \left[\begin{array}{c} -\sin \varphi_0\\ \cos \varphi_0\\ \end{array}\right] \left(\partial_1 u_1+\partial_2 u_2\right)\\ & \quad + \mu \left[\begin{array}{c} -\cos \varphi_0\\ -\sin \varphi_0\\ \end{array}\right]\left(\partial_2 u_1-\partial_1 u_2\right):=\left[\begin{array}{c}{T_1 (\bmf{u}) }\big|_{\Gamma_h^+ }\\{T_2 (\bmf{u}) }\big|_{\Gamma_h^+ }\end{array}\right], \end{aligned} \end{equation} where \begin{equation*} \begin{split} T_1( \bmf{u})\big|_{\Gamma_h^+ }&=2\mu(-\sin \varphi_0 \partial_1 u_1+\cos \varphi_0 \partial_2 u_1 ) -\lambda \sin \varphi_0 (\partial_1 u_1+\partial_2 u_2 )-\mu \cos \varphi_0 (\partial_2 u_1-\partial_1 u_2 ) , \\ T_2 (\bmf{u})\big|_{\Gamma_h^+ }&=2\mu(-\sin \varphi_0 \partial_1 u_2+\cos \varphi_0 \partial_2 u_2 ) -\lambda \cos \varphi_0 (\partial_1 u_1+\partial_2 u_2 )-\mu \sin \varphi_0 (\partial_2 u_1-\partial_1 u_2 ). \end{split} \end{equation*} Using \eqref{eq:u comp}, it is readily shown that \begin{equation}\label{eq:u1 par} \begin{aligned} \frac{\partial u_1}{\partial r} &= \sum_{m=0} ^\infty \left\{ \frac{k_p^2}{4}a_m \left\{\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}\left(k_p r\right) - \left[\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi}+\mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi}\right] J_m \left(k_p r\right) \right\}\right.\\ &\quad + \frac{{\mathrm{i}} k_s^2}{4}b_m \left\{\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}\left(k_s r\right) - \left[\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi}-\mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi}\right] J_m \left(k_s r\right) \right\}\\ &\quad + \frac{k_p^2}{4}a_m \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+2} \left(k_p r\right)-\frac{{\mathrm{i}} k_s^2}{4}b_m \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+2} \left(k_s r\right) \bigg\},\\ \frac{\partial u_1}{\partial \varphi} &= \sum_{m=0} ^\infty \left\{\frac{{\mathrm{i}}\left(m-1\right)}{2} k_p \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-1} \left(k_p r\right) a_m - \frac{{\mathrm{i}}\left(m+1\right)}{2} k_p \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+1} \left(k_p r\right) a_m \right.\\ & \quad - \frac{\left(m-1\right)}{2} k_s \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-1} \left(k_s r\right) b_m - \frac{\left(m+1\right)}{2} k_s \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+1} \left(k_s r\right) b_m\bigg\}, \end{aligned} \end{equation} and \begin{equation}\label{eq:u2 par} \begin{aligned} \frac{\partial u_2}{\partial r} & = \sum_{m=0} ^\infty \left\{\frac{{\mathrm{i}} k_p^2}{4}a_m \left\{\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}\left(k_p r\right) - \left[\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi}-\mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi}\right] J_m \left(k_p r\right)\right\}\right.\\ &\quad + \frac{k_s^2}{4}b_m \left\{-\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}\left(k_s r\right) + \left[\mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi}+\mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi}\right] J_m \left(k_s r\right) \right\}\\ &\quad - \frac{{\mathrm{i}} k_p^2}{4}a_m \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+2} \left(k_p r\right) -\frac{k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+2} \left(k_s r\right) \bigg\}, \end{aligned} \end{equation} \begin{equation} \begin{aligned} \frac{\partial u_2}{\partial \varphi} &= \sum_{m=0} ^\infty \left\{-\frac{\left(m-1\right)}{2} k_p \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-1} \left(k_p r\right) a_m - \frac{\left(m+1\right)}{2} k_p \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+1} \left(k_p r\right) a_m \right.\\ & \quad - \frac{{\mathrm{i}} \left(m-1\right)}{2} k_s \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-1} \left(k_s r\right) b_m + \frac{{\mathrm{i}} \left(m+1\right)}{2} k_s \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+1} \left(k_s r\right) b_m\bigg\}. \end{aligned} \end{equation} Using the fact that \begin{equation}\label{eq:u3 par} \begin{split} \frac{\partial u_i}{\partial x_1}&=\cos\varphi \cdot \frac{\partial u_i}{\partial r}- \frac{\sin \varphi}{r} \cdot \frac{\partial u_i}{\partial \varphi},\quad \frac{\partial u_i}{\partial x_2}=\sin \varphi \cdot \frac{\partial u_i}{\partial r}+ \frac{\cos \varphi}{r} \cdot \frac{\partial u_i}{\partial \varphi}, \end{split} i=1,2, \end{equation} as well as \eqref{eq:u1 par} and \eqref{eq:u2 par}, by tedious but straightforward calculations, one can obtain that \begin{equation}\label{eq:u1 partial 127} \begin{aligned} & \partial_1 u_1 \cdot \left(-\sin \varphi_0\right)+\partial_2 u_1 \cdot \left(\cos \varphi_0\right)\\ & = \sum_{m=0} ^\infty \Bigg \{ \sin(\varphi -\varphi_0) \Big[\frac{k_p^2}{4} a_m \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}(k_p r) - \frac{k_p^2}{4} a_m J_m(k_p r)\left(\mathrm{e}^{{\mathrm{i}} (m-1)\varphi}+\mathrm{e}^{{\mathrm{i}} (m+1)\varphi}\right)\\ & + \frac{k_p^2}{4} a_m \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+2}(k_p r) + \frac{{\mathrm{i}} k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}(k_s r) \\ & - \frac{{\mathrm{i}} k_s^2}{4} b_m J_m(k_s r)\left(\mathrm{e}^{{\mathrm{i}} (m-1)\varphi}-\mathrm{e}^{{\mathrm{i}} (m+1)\varphi}\right) - \frac{{\mathrm{i}} k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+2}(k_s r) \Big ] \\ & + \frac{ \cos(\varphi -\varphi) }{r} \Big [\frac{{\mathrm{i}}(m-1)k_p}{2} a_m \mathrm{e}^{{\mathrm{i}} (m-1)\varphi} J_{m-1}(k_p r) \\ & - \frac{{\mathrm{i}}(m+1)k_p}{2} a_m \mathrm{e}^{{\mathrm{i}} (m+1)\varphi} J_{m+1}(k_p r) - \frac{(m-1)k_s}{2} b_m \mathrm{e}^{{\mathrm{i}} (m-1)\varphi} J_{m-1}(k_s r)\\ & - \frac{(m+1)k_s}{2} b_m \mathrm{e}^{{\mathrm{i}} (m+1)\varphi} J_{m+1}(k_s r) \Big] \Bigg \},\\ & \partial_1 u_2 \cdot \left(-\sin \varphi_0\right)+\partial_2 u_2 \cdot \left(\cos \varphi_0\right)\\ & = \sum_{m=0} ^\infty \Bigg \{ \sin(\varphi -\varphi_0 ) \Big [ \frac{{\mathrm{i}} k_p^2}{4} a_m \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}(k_p r) - \frac{{\mathrm{i}} k_p^2}{4} a_m J_m(k_p r)\left(\mathrm{e}^{{\mathrm{i}} (m-1)\varphi}-\mathrm{e}^{{\mathrm{i}} (m+1)\varphi}\right) \\ & - \frac{{\mathrm{i}} k_p^2}{4} a_m \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+2}(k_p r) - \frac{ k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}(k_s r) + \frac{ k_s^2}{4} b_m J_m(k_s r)\left(\mathrm{e}^{{\mathrm{i}} (m-1)\varphi}+\mathrm{e}^{{\mathrm{i}} (m+1)\varphi}\right)\\ & -\frac{ k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} \left(m+1\right) \varphi} J_{m+2}(k_s r) \Big] + \frac{\cos(\varphi -\varphi_0)}{r} \Big[ \frac{-(m-1)k_p}{2} a_m \mathrm{e}^{{\mathrm{i}} (m-1)\varphi} J_{m-1}(k_p r) \\ & - \frac{(m+1)k_p}{2} a_m \mathrm{e}^{{\mathrm{i}} (m+1)\varphi} J_{m+1}(k_p r) - \frac{{\mathrm{i}}(m-1)k_s}{2} b_m \mathrm{e}^{{\mathrm{i}} (m-1)\varphi} J_{m-1}(k_s r)\\ & + \frac{{\mathrm{i}}(m+1)k_s}{2} b_m \mathrm{e}^{{\mathrm{i}} (m+1)\varphi} J_{m+1}(k_s r) \Big] \Bigg \}. \end{aligned} \end{equation} Similarly, from \eqref{eq:u1 par} and \eqref{eq:u2 par}, we have \begin{equation} \begin{aligned} & \partial_1 u_1 +\partial_2 u_2 = \sum_{m=0} ^\infty \Bigg \{ \frac{k_p^2}{4} \mathrm{e}^{{\mathrm{i}} m \varphi} a_m \left( J_{m-2} \left(k_p r\right)-2 J_m \left(k_p r\right) + J_{m+2} \left(k_p r\right)\right ) \\ & +\frac{{\mathrm{i}} k_s^2}{4} \mathrm{e}^{{\mathrm{i}} m \varphi} b_m \left ( J_{m-2} \left(k_s r\right)-J_{m+2} \left(k_s r\right)\right) + \frac{1}{r} \bigg [ -\frac{k_p}{2}\mathrm{e}^{{\mathrm{i}} m \varphi} a_m \big( \left(m-1\right)J_{m-1} \left(k_p r\right)\\ &+\left(m+1\right)J_{m+1} \left(k_p r\right)\big) - \frac{{\mathrm{i}} k_s}{2}\mathrm{e}^{{\mathrm{i}} m \varphi} b_m \left( \left(m-1\right)J_{m-1} \left(k_s r\right)-\left(m+1\right)J_{m+1} \left(k_s r\right)\right) \bigg ]\Bigg\}, \end{aligned} \end{equation} and \begin{equation}\label{eq:u1u2 129} \begin{aligned} & \partial_2 u_1-\partial_1 u_2 = \sum_{m=0} ^\infty \Bigg \{\frac{{\mathrm{i}} k_p^2}{4} \mathrm{e}^{{\mathrm{i}} m \varphi} a_m \left( -J_{m-2} \left(k_p r\right)+J_{m+2} \left(k_p r\right)\right) \\ & +\frac{k_s^2}{4} \mathrm{e}^{{\mathrm{i}} m \varphi} b_m \left( J_{m-2} \left(k_s r\right)-2 J_m \left(k_s r\right)+J_{m+2} \left(k_s r\right)\right) + \frac{1}{r} \bigg [ \frac{{\mathrm{i}} k_p}{2}\mathrm{e}^{{\mathrm{i}} m \varphi} a_m \bigg( \left(m-1\right)J_{m-1} \left(k_p r\right)\\ & -\left(m+1\right)J_{m+1} \left(k_p r\right)\bigg) - \frac{k_s}{2}\mathrm{e}^{{\mathrm{i}} m \varphi} b_m \left( \left(m-1\right)J_{m-1} \left(k_s r\right)+\left(m+1\right)J_{m+1} \left(k_s r\right)\right) \bigg] \Bigg\}. \end{aligned} \end{equation} Plugging \eqref{eq:u1 partial 127}--\eqref{eq:u1u2 129} into \eqref{eq:Tu new exp}, after tedious but straightforward calculations, we have \begin{equation}\label{eq:J1} \begin{aligned} &T_1(\bmf{u})\big |_{\Gamma_h^+ } = \sum_{m=0} ^\infty \left\{\frac{k_p^2}{4} a_m \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}(k_p r) \left[2 \mu \sin(\varphi-\varphi_0)-\lambda \mathrm{e}^{{\mathrm{i}} \varphi} \sin \varphi_0 +{\mathrm{i}} \mu \mathrm{e}^{{\mathrm{i}} \varphi} \cos \varphi_0\right] \right.\\ & + \frac{k_p^2}{2} a_m \mathrm{e}^{{\mathrm{i}} m \varphi} J_m(k_p r)\left[\lambda \sin\varphi_0+2 \mu \cos\varphi \sin(\varphi_0-\varphi)\right]\\ & + \frac{k_p^2}{4} a_m \mathrm{e}^{{\mathrm{i}} m \varphi} J_{m+2}(k_p r) \left[2 \mu \mathrm{e}^{{\mathrm{i}} \varphi} \sin(\varphi-\varphi_0) - \lambda \sin\varphi_0 - {\mathrm{i}} \mu \cos \varphi_0 \right]\\ & + \frac{{\mathrm{i}} k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}(k_s r) \left[2 \mu \sin(\varphi-\varphi_0)-\lambda \mathrm{e}^{{\mathrm{i}} \varphi} \sin \varphi_0 +{\mathrm{i}} \mu \mathrm{e}^{{\mathrm{i}} \varphi} \cos \varphi_0\right]\\ & + \frac{ k_s^2}{2} b_m \mathrm{e}^{{\mathrm{i}} m\varphi} J_m(k_s r)\left[\mu \cos\varphi_0+2 \mu \sin\varphi \sin(\varphi_0-\varphi)\right] \\ & + \frac{{\mathrm{i}} k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} m \varphi} J_{m+2}(k_s r) \left[2 \mu \mathrm{e}^{{\mathrm{i}} \varphi} \sin(\varphi_0-\varphi) + \lambda \sin\varphi_0 + {\mathrm{i}} \mu \cos \varphi_0\right]\\ & + \frac{1}{r} \left\{\frac{(m-1)k_p}{2} a_m \mathrm{e}^{{\mathrm{i}} (m-1)\varphi} J_{m-1}(k_p r)\left[2 {\mathrm{i}} \mu \cos(\varphi_0-\varphi) + \lambda \mathrm{e}^{{\mathrm{i}} \varphi} \sin\varphi_0 - {\mathrm{i}} \mu \mathrm{e}^{{\mathrm{i}} \varphi} \cos\varphi_0\right] \right.\\ & + \frac{(m+1)k_p}{2} a_m \mathrm{e}^{{\mathrm{i}} m \varphi} J_{m+1}(k_p r)\left[-2 {\mathrm{i}} \mu \mathrm{e}^{{\mathrm{i}} \varphi} \cos(\varphi_0-\varphi) + \lambda \sin\varphi_0 + {\mathrm{i}} \mu \cos\varphi_0\right]\\ & + \frac{(m-1)k_s}{2} b_m \mathrm{e}^{{\mathrm{i}} (m-1)\varphi} J_{m-1}(k_s r)\left[-2 \mu \cos(\varphi_0-\varphi) + {\mathrm{i}} \lambda \mathrm{e}^{{\mathrm{i}} \varphi} \sin\varphi_0 + \mu \mathrm{e}^{{\mathrm{i}} \varphi} \cos\varphi_0\right]\\ & + \frac{(m+1)k_s}{2} b_m \mathrm{e}^{{\mathrm{i}} m\varphi} J_{m+1}(k_s r) \left[-2 \mu \mathrm{e}^{{\mathrm{i}} \varphi} \cos(\varphi_0-\varphi) - {\mathrm{i}} \lambda \sin\varphi_0 + \mu \cos\varphi_0\right] \bigg\}\bigg\}. \end{aligned} \end{equation} Substituting \eqref{eq:J2} into \eqref{eq:J1}, we can obtain that \begin{equation}\label{eq:T1 132} \begin{aligned} & T_1( \bmf{u})\big |_{\Gamma_h^+ } = \sum_{m=0} ^\infty \bigg \{ \frac{{\mathrm{i}} k_p^2}{2} a_m \mathrm{e}^{{\mathrm{i}} (m-1) \varphi} \mathrm{e}^{-{\mathrm{i}} (\varphi-\varphi_0)} \mu J_{m-2} \left(k_p r\right)+ k_p^2 a_m \mathrm{e}^{{\mathrm{i}} m \varphi} \left(\lambda+\mu\right) \sin\varphi_0 J_m \left(k_p r\right) \\ & -\frac{{\mathrm{i}} k_p^2}{2} a_m \mathrm{e}^{{\mathrm{i}} (m+1) \varphi} \mathrm{e}^{{\mathrm{i}} (\varphi-\varphi_0)} \mu J_{m+2} -\frac{k_s^2}{2} b_m \mathrm{e}^{{\mathrm{i}} (m-1) \varphi} \mathrm{e}^{-{\mathrm{i}} (\varphi-\varphi_0)} \mu J_{m-2} \left(k_s r\right)\\ &-\frac{k_s^2}{2} b_m \mathrm{e}^{{\mathrm{i}} (m+1) \varphi} \mathrm{e}^{{\mathrm{i}} (\varphi-\varphi_0)} \mu J_{m+2} \left(k_s r\right)\bigg\}. \end{aligned} \end{equation} Similarly, substituting \eqref{eq:u1 partial 127} to \eqref{eq:u1u2 129} into \eqref{eq:Tu new exp}, after tedious but straightforward calculations, we have \begin{equation}\label{eq:J3} \begin{aligned} & T_2(\bmf{u} )\big|_{\Gamma_h^+ }= \sum_{m=0} ^\infty \left\{\frac{k_p^2}{4} a_m \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}(k_p r) \left[2 {\mathrm{i}} \mu \sin(\varphi-\varphi_0)+\lambda \mathrm{e}^{{\mathrm{i}} \varphi} \cos \varphi_0 +{\mathrm{i}} \mu \mathrm{e}^{{\mathrm{i}} \varphi} \sin \varphi_0\right] \right.\\ & + \frac{k_p^2}{2} a_m \mathrm{e}^{{\mathrm{i}} m \varphi} J_m(k_p r)\left[-\lambda \cos\varphi_0+2 \mu \sin\varphi \sin(\varphi_0-\varphi)\right]\\ & + \frac{k_p^2}{4} a_m \mathrm{e}^{{\mathrm{i}} m \varphi} J_{m+2}(k_p r) \left[2 {\mathrm{i}} \mu \mathrm{e}^{{\mathrm{i}} \varphi} \sin(\varphi_0-\varphi) + \lambda \cos\varphi_0 - {\mathrm{i}} \mu \sin \varphi_0 \right]\\ & + \frac{ k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} \left(m-1\right) \varphi} J_{m-2}(k_s r) \left[2 \mu \sin(\varphi_0-\varphi)+ {\mathrm{i}} \lambda \mathrm{e}^{{\mathrm{i}} \varphi} \cos \varphi_0 - \mu \mathrm{e}^{{\mathrm{i}} \varphi} \sin \varphi_0\right]\\ & + \frac{ k_s^2}{2} b_m \mathrm{e}^{{\mathrm{i}} m\varphi} J_m(k_s r)\left[\mu \sin\varphi_0+2 \mu \cos\varphi \sin(\varphi-\varphi_0)\right] \\ & + \frac{ k_s^2}{4} b_m \mathrm{e}^{{\mathrm{i}} m \varphi} J_{m+2}(k_s r) \left[2 \mu \mathrm{e}^{{\mathrm{i}} \varphi} \sin(\varphi_0-\varphi) -{\mathrm{i}} \lambda \cos\varphi_0 - \mu \sin \varphi_0\right]\\ & + \frac{1}{r} \left\{-\frac{(m-1)k_p}{2} a_m \mathrm{e}^{{\mathrm{i}} (m-1)\varphi} J_{m-1}(k_p r)\left[2 \mu \cos(\varphi_0-\varphi) + \lambda \mathrm{e}^{{\mathrm{i}} \varphi} \cos\varphi_0 + {\mathrm{i}} \mu \mathrm{e}^{{\mathrm{i}} \varphi} \sin\varphi_0\right] \right.\\ \end{aligned} \end{equation} \[ \begin{aligned} &\hspace*{-5mm} + \frac{(m+1)k_p}{2} a_m \mathrm{e}^{{\mathrm{i}} m \varphi} J_{m+1}(k_p r)\left[-2 \mu \mathrm{e}^{{\mathrm{i}} \varphi} \cos(\varphi_0-\varphi) - \lambda \cos\varphi_0 + {\mathrm{i}} \mu \sin\varphi_0\right]\\ &\hspace*{-5mm} + \frac{(m-1)k_s}{2} b_m \mathrm{e}^{{\mathrm{i}} (m-1)\varphi} J_{m-1}(k_s r)\left[-2 {\mathrm{i}} \mu \cos(\varphi_0-\varphi) - {\mathrm{i}} \lambda \mathrm{e}^{{\mathrm{i}} \varphi} \cos\varphi_0 + \mu \mathrm{e}^{{\mathrm{i}} \varphi} \sin\varphi_0\right]\\ &\hspace*{-5mm}+ \frac{(m+1)k_s}{2} b_m \mathrm{e}^{{\mathrm{i}} m\varphi} J_{m+1}(k_s r) \left[2 {\mathrm{i}} \mu \mathrm{e}^{{\mathrm{i}} \varphi} \cos(\varphi_0-\varphi) + {\mathrm{i}} \lambda \cos\varphi_0 + \mu \sin\varphi_0\right] \bigg\}\bigg\}, \end{aligned} \] Plugging \eqref{eq:J2} into \eqref{eq:J3}, we can obtain that \begin{equation}\label{eq:T2 134} \begin{aligned} & T_2( \bmf{u})\big |_{\Gamma_h^+ } = \sum_{m=0} ^\infty \left\{- \frac{ k_p^2}{2} a_m \mathrm{e}^{{\mathrm{i}} (m-1) \varphi} \mathrm{e}^{-{\mathrm{i}} (\varphi-\varphi_0)} \mu J_{m-2} \left(k_p r\right)- k_p^2 a_m \mathrm{e}^{{\mathrm{i}} m \varphi} \left(\lambda+\mu\right) \cos\varphi_0 J_m \left(k_p r\right)\right.\\ & -\frac{ k_p^2}{2} a_m \mathrm{e}^{{\mathrm{i}} (m+1) \varphi} \mathrm{e}^{{\mathrm{i}} (\varphi-\varphi_0)} \mu J_{m+2}(k_p r) -\frac{{\mathrm{i}} k_s^2}{2} b_m \mathrm{e}^{{\mathrm{i}} (m-1) \varphi} \mathrm{e}^{-{\mathrm{i}} (\varphi-\varphi_0)} \mu J_{m-2} \left(k_s r\right)\\ &+\frac{{\mathrm{i}} k_s^2}{2} b_m \mathrm{e}^{{\mathrm{i}} (m+1) \varphi} \mathrm{e}^{{\mathrm{i}} (\varphi-\varphi_0)} \mu J_{m+2} \left(k_s r\right)\bigg\}. \end{aligned} \end{equation} Using the fact \[\left[\begin{array}{c}{\sin \varphi}\\{-\cos \varphi}\end{array}\right]=\frac{{\mathrm{i}}}{2}\left(\mathrm{e}^{-{\mathrm{i}} \varphi}\mathbf{e}_1-\mathrm{e}^{{\mathrm{i}} \varphi}\mathbf{e}_2\right), \] substituting \eqref{eq:T1 132} and \eqref{eq:T2 134} into \eqref{eq:Tu new exp}, we can prove \eqref{eq:Tu1}. The proof of \eqref{eq:Tu2} is similar to \eqref{eq:Tu1}, which is omitted here. \end{proof}

https://arxiv.org/abs/2005.06253

Universal central extensions of linear groups over rings of non-commutative Laurent polynomials, associated $K_1$-groups and $K_2$-groups

We prove that linear groups over rings of non-commutative Laurent polynomials $D_{\tau}$ have Tits systems with the corresponding affine Weyl groups and have universal central extensions if $|Z(D)|\geq 5$ and $|Z(D)|\neq 9$. We also determine structures of $K_1$-groups and identify generators of $K_2$-groups.

\section{Introduction} Many researchers have studied the structure of general linear groups and their elementary subgroups over fields $F$ or commutative rings $R$. They also have analyzed associated lower $K$-groups, for example \cite{m} and \cite{s}. Needless to say, general linear groups are important objects and have many applications in various areas of mathematics, but they particularly have much to do with Lie theory; Lie groups, Lie algebras and their representations. Lower $K$-groups also play an important role as a certain invariant. In this paper, we treat some rings $D_{\tau}$ of non-commutative Laurent polynomials over division rings $D$ (cf. Section 2). Here $\tau$ is an automorphism of $D$. We note that the ring we use generalizes the one which is studied in \cite{m} and \cite{s}. When $D=F$ and $\tau=id$, our discussion is just a subject of loop groups which are applied in the theory of affine Kac-Moody Lie algebras, and this is surveyed in \cite{jm} for example. On the other hand, the corresponding linear group was studied in the case when $D$ is the field of formal power series and $\tau$ is not trivial (cf. \cite{ms},\cite{hs}), which is deeply related to the theory of extended affine Lie algebras (cf. \cite{aabgp},\cite{mt},\cite{yy}). Our main object in this study is the following exact sequence. \[1\to K_2(n,D_{\tau})\to St(n,D_{\tau})\xrightarrow{\phi} GL(n,D_{\tau})\to K_1(n,D_{\tau})\to 1\] We reveal the structure of groups in the above sequence. We first describe an existence of a Tits system in the elementary subgroup $E(n,D_{\tau})$ of the general linear group $GL(n,D_{\tau})$ and the associated Steinberg group $St(n,D_{\tau})$ in Section 2 and Section 3 respectively. Using this fact, we show the above homomorphism $\phi$ is a central extension of $E(n,D_{\tau})$, that is, we confirm $\mathrm{Ker}\phi:=K_2(n,D_{\tau})$ is a central subgroup of the Steinberg group $St(n,\Dt)$ in Section 3. It is proved in Section 5 that $\phi$ is actually universal when the center $Z(D)$ of $D$ has at least five elements. Meanwhile, we discuss the structure of the associated $K_1$-group and $K_2$-group in Section 4 and 6. In particular we check that $K_2$-groups are generated by certain products of Steinberg symbols. While we deal with linear groups of rank two for simplicity of discussion, we need to raise a rank of these groups for the sake of the general theory which is possible by \cite{mj}. \begin{center} \textbf{Acknowledgements} \end{center} The author would like to thank his advisor, Professor Jun Morita, for his continuous guidance and encouragement.\vspace{3mm} \section{Linear groups over rings of non-commutative Laurent polynomials} Let $D$ be a division ring, and we fix an automorphism $\tau\in\mathrm{Aut}(D)$. In the following, we denote by $D_{\tau}=D[t,t^{-1}]$ the ring of Laurent polynomials generated by $D$ and an indeterminate $t$, whose ring structure is given by $tat^{-1}=\tau(a)$ for all $a\in D$. Let $M(n,\Dt)$ be the ring of $n\times n$ matrices whose entries are in $\Dt$, and we define the general linear group $GL(n,\Dt)$ as the matrix group of all invertible matrices, that is, $GL(n,\Dt)=M(n,\Dt)^{\times}$.\vspace{3mm} Let $\Delta=\{\epsilon_i-\epsilon_j\ |\ 1\leq i\neq j\leq n\}$ be a root system of type $\mathrm{A}_{n-1}$, where $\{ \epsilon_i\}_{1\leq i\leq n}$ is an orthonormal basis with respect to an inner product $(\cdot,\cdot)$ defined by $(\epsilon_i,\epsilon_j)=\delta_{ij}$. We see that any root in $\Delta$ is expressed as \[ \epsilon_i-\epsilon_j=(\epsilon_i-\epsilon_{i+1})+(\epsilon_{i+1}-\epsilon_{i+2})+\cdots+(\epsilon_{j-1}-\epsilon_j)\] if $i<j$ and as its minus version if $i>j$, so we put $\Pi=\{ \alpha_i:=\epsilon_i-\epsilon_{i+1}\ |\ 1\leq i\leq n-1\}$ and call it a simple system of $\Delta$. We call $\Delta^+:=\text{Span}_{\mathbb{Z}_{\geq 0}}\Pi\cap\Delta$ a set of positive roots and $\Delta^-:=-\Delta^+$ a set of negative roots. Also, let $\Delta_a:=\Delta\times\mathbb{Z}$ be an (abstract) affine root system of type $\mathrm{A}_{n-1}^{(1)}$ and $\Pi_a=\{ \dot{\alpha_i}:=(\alpha_i,0)|1\leq i\leq n-1\}\cup\{\dot{\alpha_0}:=(-\theta,1)\}$ be a simple system of $\Delta_a$, where $\theta=\alpha_1+\dots+\alpha_{n-1}$ is a highest root in $\Delta$. \vspace{3mm} For $\beta=\epsilon_i-\epsilon_j\in\Delta$, $f\in \Dt$ we define \[ e_{\beta}(f)=I+fE_{ij}\] where $I$ is the identity matrix and $E_{ij}$ is the matrix unit. For $\dot{\beta}:=(\beta, m)\in\Delta_a$, $f\in D$ and $s,u\in D^{\times}=D\setminus\{ 0\}$ we put \begin{align*} & x_{\dot{\beta}}(f)=\begin{cases} e_{\beta}(ft^m) & (\beta\in\Delta^+), \\ e_{\beta}(t^mf) & (\beta\in\Delta^-), \end{cases}\\ &w_{\dot{\beta}}(u)=x_{\dot{\beta}}(u)x_{-\dot{\beta}}(-u^{-1})x_{\dot{\beta}}(u),\\ &h_{\dot{\beta}}(s)=w_{\dot{\beta}}(s)w_{\beta}(-1), \end{align*} then we can easily see $x_{\dot{\beta}}(f)^{-1}=x_{\dot{\beta}}(-f)$, $w_{\dot{\beta}}(u)^{-1}=w_{\dot{\beta}}(-u)$. The elementary subgroup $E(n,\Dt)$ is defined to be the subgroup of $GL(n,\Dt)$ generated by $x_{\dot{\beta}}(f)$ for all $\dot{\beta}\in\Delta_a$ and $f\in D$.\vspace{3mm} In a standard way, the Weyl group $W$ of $\Delta$ is generated by all reflections $\sigma_{\beta}$ for $\beta\in\Delta$, and the Weyl group $W_a$ of $\Delta_a$ (the affine Weyl group of $\Delta$) is generated by all reflections $\sigma_{\dot{\beta}}$ for $\dot{\beta}\in\Delta_a$, where the action of $\sigma_{\dot{\beta}}$ is defined as \[\sigma_{\dot{\beta}}(\dot{\gamma})=(\sigma_{\beta}(\gamma), n-\langle \gamma, \beta \rangle m)\] for $\dot{\beta}=(\beta, m),\ \dot{\gamma}=(\gamma, n)\in\Delta_a$ and $\langle \gamma,\beta\rangle=2(\gamma,\beta)/(\beta,\beta)$. Define $\xi_{\dot{\beta}}:=\sigma_{\dot{\beta}}\sigma_{\beta}$ for each $\dot{\beta}\in\Delta_a$, then let $T_a$ be the subgroup of $W_a$ generated by $\xi_{\dot{\beta}}$ for all $\dot{\beta}\in\Delta_a$. In order to use later, we will characterize the group $W_a$ as follow (cf. \cite{jm} Lemma 1.1, Proposition 1.2).\vspace{3mm} {\lem \begin{align*} (1)\ &\text{Let $\dot{\beta}=(\beta,m)$ and $\dot{\gamma}=(\gamma,n)$. Then } \xi_{\dot{\beta}}(\dot{\gamma})=(\gamma, n+\langle\gamma,\beta\rangle m). \\ (2)\ &\text{Let $\alpha\in \Pi$. Then $T_a$ is a free abelian group generated by $\xi_{\dot{\gamma}}$ for all $\dot{\gamma}=(\alpha,1)$.}\\ (3)\ &\text{$\sigma_{\dot{\beta}}$ normalizes $T_a$ for $\dot{\beta}\in\Delta_a$. $\square$} \end{align*} } {\prop $W_a\cong T_aW\cong T_a\rtimes W$. $\square$} \vspace{5mm} As a subgroup of $E(n,\Dt)$ we put \begin{align*} &U_{\dot{\beta}}=\{ x_{\dot{\beta}}(f)\ |\ f\in D\},\\ &U^{\pm}=\langle U_{\dot{\beta}}\ |\ \dot{\beta}\in\Delta_a^{\pm}\rangle,\\ &N=\langle w_{\dot{\beta}}(u)\ |\ \dot{\beta}\in\Delta_a,\ u\in D^{\times}\rangle,\\ &T=\langle h_{\dot{\beta}}(u)\ |\ \dot{\beta}\in\Delta_a,\ u\in D^{\times}\rangle, \end{align*} where $\Delta_a^+=(\Delta^+\times\mathbb{Z}_{\geq 0})\cup(\Delta^-\times\mathbb{Z}_{>0})$ and $\Delta_a^-=(\Delta^+\times\mathbb{Z}_{< 0})\cup(\Delta^-\times\mathbb{Z}_{\leq 0})$. If $h\in T$ is expressed as $h=\mathrm{diag}(u_1,u_2,\dots,u_n)$, the diagonal matrix with $u_i\in\Dt^{\times}=D^{\times}\cdot \{ t^l|l\in\mathbb{Z}\}$, then we define \[ \mathrm{deg}_i(h)=\mathrm{deg}(u_i)=m_i \] for $i=1,2,\dots,n$, where $u_i=s_it^{m_i}$ with $s_i\in D^{\times}$ and $m_i\in\mathbb{Z}$. Then we set \begin{align*} &T_0=\langle h\ |\ h\in T,\ \mathrm{deg}_i(h)=0\ \mathrm{for\ all}\ i=1,2,\dots,n\rangle,\\ &B^{\pm}=\langle U^{\pm},\ T_0\rangle,\\ &S=\{ w_{\dot{\beta}}(1)\ \mathrm{mod}\ T_0\ |\ \dot{\beta}\in\Delta_a\}. \end{align*} In particular $S$ is identified with the set $\{ w_{\dot{\alpha}}\ |\ \dot{\alpha}\in\Pi_a\}$. The main result in this section is the following theorem.\vspace{3mm} {\thm{Notation is as above. $(E(n,\Dt),B^{\pm},N,S)$ is a Tits system with the corresponding affine Weyl group $W_a$.}}\vspace{5mm} Before proving this theorem we give several relations between $e$, $w$ and $h$. \begin{align*} &(R1)\ e_{\beta}(f)e_{\beta}(g)=e_{\beta}(f+g),\\ &(R2)\ [e_{\beta}(f),e_{\gamma}(g)]= \begin{cases} e_{\beta+\gamma}(fg) \hspace{5mm}&\text{if}\ \beta+\gamma\in\Delta,\ j=k,\\ e_{\beta+\gamma}(-gf) \hspace{5mm}&\text{if }\ \beta+\gamma\in\Delta,\ i=l,\\ 1 \hspace{10mm} &\text{otherwise}, \end{cases} \end{align*} where $\beta=\epsilon_i-\epsilon_j$, $\gamma=\epsilon_k-\epsilon_l$, $i\neq l,j\neq k$ and $f,g\in D_{\tau}$. These are fundamental relations in $E(n,\Dt)$ for example. Then we obtain; \begin{align*} &(R3)\ w_{\beta}(u)e_{\gamma}(f)w_{\beta}(u)^{-1}= \begin{cases} e_{\gamma}(f)\hspace{5mm}&\text{if} \hspace{2mm}(\beta,\gamma)=0,\\ e_{\mp\beta}(u^{\mp 1}fu^{\mp 1})\hspace{5mm}&\text{if} \hspace{2mm}\gamma=\pm\beta,\\ e_{\sigma_{\beta}(\gamma)}(-u^{-1}f)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=k,\\ e_{\sigma_{\beta}(\gamma)}(-fu)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=l,\\ e_{\sigma_{\beta}(\gamma)}(uf)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=k,\\ e_{\sigma_{\beta}(\gamma)}(fu^{-1})\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=l,\\ \end{cases}\\ &(R4)\ h_{\beta}(u)e_{\gamma}(f)h_{\beta}(u)^{-1}= \begin{cases} e_{\gamma}(f)\hspace{5mm}&\text{if} \hspace{2mm}(\beta,\gamma)=0,\\ e_{\pm\beta}(-u^{\pm 1}fu^{\pm 1})\hspace{5mm}&\text{if} \hspace{2mm}\gamma=\pm\beta,\\ e_{\gamma}(uf)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=k,\\ e_{\gamma}(fu^{-1})\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=l,\\ e_{\gamma}(u^{-1}f)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=k,\\ e_{\gamma}(fu)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=l,\\ \end{cases}\\ &(R5)\ w_{\beta}(u)w_{\gamma}(s)w_{\beta}(u)^{-1}= \begin{cases} w_{\gamma}(s)\hspace{5mm}&\text{if} \hspace{2mm}(\beta,\gamma)=0,\\ w_{\mp\beta}(-u^{\mp 1}su^{\mp 1})\hspace{5mm}&\text{if} \hspace{2mm}\gamma=\pm\beta,\\ w_{\sigma_{\beta}(\gamma)}(-u^{-1}s)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=k,\\ w_{\sigma_{\beta}(\gamma)}(-su)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=l,\\ w_{\sigma_{\beta}(\gamma)}(us)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=k,\\ w_{\sigma_{\beta}(\gamma)}(su^{-1})\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=l,\\ \end{cases}\\ &(R6)\ w_{\beta}(u)h_{\gamma}(s)w_{\beta}(u)^{-1}= \begin{cases} h_{\gamma}(s)\hspace{5mm}&\text{if} \hspace{2mm}(\beta,\gamma)=0,\\ h_{\mp\beta}(u^{\mp 1}su^{\mp 1})h_{\mp\beta}(u^{\pm 2})\hspace{5mm}&\text{if} \hspace{2mm}\gamma=\pm\beta,\\ h_{\sigma_{\beta}(\gamma)}(u^{-1}s)h_{\sigma_{\beta}(\gamma)}(u)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=k,\\ h_{\sigma_{\beta}(\gamma)}(su)h_{\sigma_{\beta}(\gamma)}(u^{-1})\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=l,\\ h_{\sigma_{\beta}(\gamma)}(us)h_{\sigma_{\beta}(\gamma)}(u^{-1})\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=k,\\ h_{\sigma_{\beta}(\gamma)}(su^{-1})h_{\sigma_{\beta}(\gamma)}(u)\hspace{5mm}&\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=l,\\ \end{cases} \end{align*} for all $\beta,\gamma\in\Delta$, $f,g\in D_{\tau}$ and $s,u\in D_{\tau}^{\times}$, where $\beta=\epsilon_i-\epsilon_j$ and $\gamma=\epsilon_k-\epsilon_l$. {\lem\begin{align*} &(1)\ B^{\pm}=U^{\pm}\rtimes T_0,\\ &(2)\ T_0\lhd N\ and\ T\lhd N,\\ &(3)\ B^{\pm}\cap N=T_0,\\ &(4)\ N/T_0\cong W_a. \end{align*}} \textbf{Proof:}\ From the definition, we know $B^{\pm}=\langle U^{\pm},T_0\rangle$, and we also see $U^{\pm}\cap T_0=\{ I\}$ easily if we consider the degree of an element in $U^{\pm}$. In addition $T_0$ normalizes $U^{\pm}$ from (R4), hence (1) holds. (2) has already been proven by (R6). We prove (4) along with \cite{jm}. If we put $N_0=\langle w_{\alpha}(u)| \alpha\in\Pi,u\in D^{\times}\rangle$ then we have $N=TN_0$ by the definition of $T$, and we know $T_0\lhd T$ and $\ T_0\lhd N$. Therefore \[N/T_0\cong (T/T_0)(N_0/T_0)\cong (T/T_0)W.\] Here we set $\dot{\gamma}=(\alpha,1)$ for $\alpha\in\Pi$. Then the map $T_a\to T/T_0$ defined by $\xi_{\dot{\gamma}}\mapsto h_{\alpha}(t)$ is an isomorphism, hence $N/T_0\cong W_a$ from Proposition 1. By $U^{\pm}\cap T_0=\{ I\}$ and (4), we can see that (3) holds. $\square$\vspace{3mm} To discuss these subgroups more explicitly we put \[ U'_{\pm\dot{\alpha}}=\langle x_{\pm\dot{\alpha}}(g)U_{\dot{\beta}}x_{\pm\dot{\alpha}}(g)^{-1}\ |\ g\in D,\dot{\beta}\in\Delta^{\pm}_a\setminus\{ \pm\dot{\alpha}\}\rangle\] for each $\dot{\alpha}\in\Pi_a$.\vspace{3mm} {\prop Let $\dot{\alpha}\in\Pi_a$. Then; \begin{align*} (1)\ &w_{\pm\dot{\alpha}}(u)U'_{\pm\dot{\alpha}}w_{\pm\dot{\alpha}}(u)^{-1}=U'_{\pm\dot{\alpha}}\ \text{for all $u\in D^{\times}$},\\ (2)\ &U^{\pm}=U'_{\pm\dot{\alpha}}\rtimes U_{\pm\dot{\alpha}}. \end{align*}} \textbf{Proof:} (1) It suffices to show that \[w_{\dot{\alpha}}(u)x_{\dot{\alpha}}(g)x_{\dot{\beta}}(f')x_{\dot{\alpha}}(g)^{-1}w_{\dot{\alpha}}(u)^{-1}\in U_{\dot{\alpha}}'\] for all $\dot{\beta}\in\Delta_a^+\setminus\{\dot{\alpha}\}$. For $\dot{\beta}=(\beta,m)$ we denote its negative root in a first entry by $\dot{-\beta}$, that is, $\dot{-\beta}=(-\beta,m)$. There are two cases:\vspace{3mm} Case 1. If $\dot{\beta}=(\beta, m)\in\Delta_a^+\setminus\{ \dot{\alpha}\}$ and $\beta\neq -\alpha$, then we see \begin{align*} w_{\dot{\alpha}}(u)x_{\dot{\alpha}}(g)x_{\dot{\beta}}(f')x_{\dot{\alpha}}(g)^{-1}w_{\dot{\alpha}}(u)^{-1}&= \begin{cases} w_{\dot{\alpha}}(u)x_{\dot{\alpha}+\dot{\beta}}(uf')x_{\dot{\beta}}(f')w_{\dot{\alpha}}(u)^{-1}\\ w_{\dot{\alpha}}(u)x_{\dot{\alpha}+\dot{\beta}}(-f'u)x_{\dot{\beta}}(f')w_{\dot{\alpha}}(u)^{-1}\\ w_{\dot{\alpha}}(u)x_{\dot{\beta}}(f')w_{\dot{\alpha}}(u)^{-1} \end{cases}\\&= \begin{cases} x_{\sigma_{\dot{\alpha}}(\dot{\alpha}+\dot{\beta})}(*)x_{\sigma_{\dot{\alpha}}(\dot{\beta})}(*)\\ x_{\sigma_{\dot{\alpha}}(\dot{\alpha}+\dot{\beta})}(*)x_{\sigma_{\dot{\alpha}}(\dot{\beta})}(*)\\ x_{\sigma_{\dot{\alpha}}(\dot{\beta})}(*) \end{cases} \end{align*} by the previous relations $(R2)$ and $(R3)$, but these all belong to $U'_{\dot{\alpha}}$ because of the fact $m\in\mathbb{Z}_{>0}$ in any cases.\vspace{3mm} Case 2. If $\dot{\beta}=(\beta, m)\in\Delta_a^+\setminus\{ \dot{\alpha}\}$ and $\beta= -\alpha$, then \begin{align*} &w_{\dot{\alpha}}(u)x_{\dot{\alpha}}(g)x_{\dot{\beta}}(f')x_{\dot{\alpha}}(g)^{-1}w_{\dot{\alpha}}(u)^{-1}\\ &=w_{\dot{\alpha}}(u)x_{\dot{\alpha}}(g)w_{\dot{\alpha}}(u)^{-1}w_{\dot{\alpha}}(u)x_{\dot{\beta}}(f')w_{\dot{\alpha}}(u)^{-1}w_{\dot{\alpha}}(u)x_{\dot{\alpha}}(g)^{-1}w_{\dot{\alpha}}(u)^{-1}\\ &=x_{\dot{-\alpha}}(g')x_{\dot{-\beta}}(f'')x_{\dot{-\alpha}}(g')^{-1}\\ &=x_{\dot{\alpha}}(g'^{-1})w_{\dot{\alpha}}(-g'^{-1})x_{\dot{\alpha}}(g'^{-1})x_{\dot{-\beta}}(f'')x_{\dot{\alpha}}(g'^{-1})^{-1}w_{\dot{\alpha}}(-g'^{-1})^{-1}x_{\dot{\alpha}}(g'^{-1})^{-1}\\ &=x_{\dot{\alpha}}(g'^{-1})x_{\dot{\beta}}(f''')x_{\dot{\alpha}}(g'^{-1})^{-1}\in U_{\dot{\alpha}}'. \end{align*} (2) From the definition, we see $U'_{\pm\dot{\alpha}}<U^{\pm}$ and $U_{\dot{\pm\alpha}}<U^{\pm}$, and that $U_{\pm\dot{\alpha}}$ normalizes $U'_{\dot{\pm\alpha}}$. Suppose $U'_{\pm\dot{\alpha}}\cap U_{\pm\dot{\alpha}}\neq \{ I\}$. Then, for each $x_{\pm\dot{\alpha}}(f)\in U_{\pm\dot{\alpha}}$, there exist $x'_i\in U'_{\pm\dot{\alpha}}$ such that $x_{\pm\dot{\alpha}}(f)=x'_1x'_2\dots x'_r$. By (1), we have \[U^{\mp}\ni w_{\pm\dot{\alpha}}(u)x_{\pm\dot{\alpha}}(f)w_{\pm\dot{\alpha}}(u)^{-1}=w_{\pm\dot{\alpha}}(u)x'_1x'_2\dots x'_rw_{\pm\dot{\alpha}}(u)^{-1}\in U^{\pm}.\] Thus $U'_{\pm\dot{\alpha}}\cap U_{\pm\dot{\alpha}}= \{ I\}$. $\square$\vspace{3mm} We need to check the following two conditions for the sake of the proof; \begin{align*} (T1)\ &sB^{\pm}w\subset B^{\pm}wB^{\pm}\cup B^{\pm}swB^{\pm}\ \text{for all $s\in S$ and $w\in W_a$},\\ (T2)\ &sB^{\pm}s\not\subset B^{\pm}\ \text{for each $s\in S$}. \end{align*} \textbf{Proof of Theorem 1}: We check the ``axiom of Tits systems''. We have already see that $E(n,\Dt)$ is a group, $B^{\pm}$ and $N$ are subgroups of $E(n,\Dt)$ and $S$ is a subset of $N/(B^{\pm}\cap N)$. In addition $E(n,\Dt)=\langle B^{\pm},N\rangle$ and $B^{\pm}\cap N\lhd N$. Moreover $N/(B^{\pm}\cap N)=W_a=\langle S\rangle$, and all elements in $S$ are of order two (modulo $T_0$). Finally we prove the above two conditions (T1) and (T2), but it is enough to check only the case of $B^+$. In the following, we write $B=B^+$.\vspace{3mm} Define the length, called $l(w)$, of an element $w$ in $W_a$ as usual. For $w\in W_a$ and $s=w_{\dot{\alpha}}(1)\in S$, if $l(w)<l(sw)$ then we have $w(\dot{\alpha})\in\Delta_a^+$, hence \begin{align*} wBs&=wU_{\dot{\alpha}}U'_{\dot{\alpha}}T_0s\\ &=wU_{\dot{\alpha}}w^{-1}wsU'_{\dot{\alpha}}sT_0s\\ &=U_{w(\dot{\alpha})}wsU'_{\dot{\alpha}}T_0\\ &\subset BwsB. \end{align*} Also, if $l(ws)<l(w)$ then $w(\dot{\alpha})\in\Delta_a^-$, and we have $l(w')<l(w's)$ for $w'=ws$. Thus \begin{align*} wBs&=w'sBs\\ &\subset w'(B\cup BsB)\\ &=w'B\cup w'BsB\\ &\subset Bw'B\cup Bw'sBB\\ &=BwsB\cup BwB, \end{align*} so (T1) holds. On the other hand, we can easily see that (T2) holds by direct calculation. $\square$\vspace{3mm} {\cor{Notation is as above. Then\vspace{3mm}\\ $(1)$ $E(n,\Dt)$ has a Bruhat decomposition: \begin{center}$E(n,\Dt)=\bigcup_{w\in W_a}B^{\pm}wB^{\pm}=U^{\pm}NU^{\pm}$,\end{center} $(2)$ $E(n,\Dt)$ has a Birkhoff decomposition: \begin{center}$E(n,\Dt)=\bigcup_{w\in W_a}B^{\mp}wB^{\pm}=U^{\mp}NU^{\pm}$,\end{center} (3) $E(n,\Dt)$ has a Gauss decomposition: \begin{center}$E(n,\Dt)=\bigcup_{X\in U^{\pm}}XB^{\mp}B^{\pm}X^{-1}=U^{\pm}B^{\mp}U^{\pm}$. $\square$\end{center}}} \section{Steinberg groups} Let $St(n,\Dt)$ be the Steinberg group, which is defined by the generators $\hat{x}_{ij}(f)$ for all $f\in\Dt$, $1\leq i\neq j\leq n$ and the defining relations \begin{align*} &(ST1)\ \hat{x}_{ij}(f)\hat{x}_{ij}(g)=\hat{x}_{ij}(f+g)\\ &(ST2)\ [\hat{x}_{ij}(f),\hat{x}_{kl}(g)]= \begin{cases} \hat{x}_{il}(fg) &\text{if } j=k,\\ \hat{x}_{kj}(-gf) &\text{if } i=l, \\ 1 &\text{otherwise}, \end{cases} \end{align*} where $f,g\in\Dt$, $1\leq i\neq j\leq 2$ and $1\leq k\neq l\leq n$ with$(i,j)\neq (l,k)$. Then, we see $\hat{x}_{ij}(f)^{-1}=\hat{x}_{ij}(-f)$ from $(ST1)$. When $n=2$, we use \[(ST2)'\ \hat{w}_{ij}(u)\hat{x}_{ij}(v)\hat{w}_{ij}(-u)=\hat{x}_{ji}(-u^{-1}vu^{-1})\] instead of $(ST2)$, where $\hat{w}_{ij}(u)=\hat{x}_{ij}(u)\hat{x}_{ji}(-u^{-1})\hat{x}_{ij}(u)$.\vspace{3mm} We now consider a natural homomorphism $\phi$ from $St(n,\Dt)$ onto $E(n,\Dt)$ defined by $\phi(\hat{x}_{ij}(f))=I+E_{ij}$ for all $f\in\Dt$. Then, we claim that $\phi$ is a universal central extension of $E(n,\Dt)$ under some conditions. The proof is formed in three steps: we first show that $St(n,\Dt)$ has a Tits system, and we check that $\phi$ is a central extension. Finally, we prove the universality of $\phi$. In this section we show $St(n,\Dt)$ has a Tits system.\vspace{3mm} For $\dot{\beta}=(\beta,m)\in\Delta_a$ and $f\in D$, we put \[\hat{x}_{\dot{\beta}}(f)= \begin{cases} \hat{x}_{\beta}(ft^m) & (\beta\in\Delta^+), \\ \hat{x}_{\beta}(t^mf) & (\beta\in\Delta^-), \end{cases}\] and we also put $\hat{w}_{\dot{\beta}}(u)=\hat{x}_{\dot{\beta}}(u)\hat{x}_{-\dot{\beta}}(-u^{-1})\hat{x}_{\dot{\beta}}(u)$ and $\hat{h}_{\dot{\beta}}(u)=\hat{w}_{\dot{\beta}}(u)\hat{w}_{\beta}(-1)$ for $u\in D^{\times}$ as in Section 2. Here, we put, as a subgroup of $St(n,\Dt)$, \begin{align*} &\hat{U}_{\dot{\beta}}=\{ \hat{x}_{\dot{\beta}}(f)\ |\ f\in D\},\\ &\hat{U}^{\pm}=\langle \hat{U}_{\dot{\beta}}\ |\ \dot{\beta}\in\Delta_a^{\pm}\rangle,\\ &\hat{N}=\langle \hat{w}_{\dot{\beta}}(u)\ |\ \dot{\beta}\in\Delta_a,\ u\in D^{\times}\rangle,\\ &\hat{T}=\langle \hat{h}_{\dot{\beta}}(u)\ |\ \dot{\beta}\in\Delta_a,\ u\in D^{\times}\rangle. \end{align*} Then we set \begin{align*} &\hat{T_0}=\langle h\ |\ h\in \hat{T},\ \mathrm{deg}_i(\phi(h))=0\ \mathrm{for\ all}\ i=1,\dots, n\rangle,\\ &\hat{B}^{\pm}=\langle \hat{U}^{\pm},\ \hat{T_0}\rangle,\\ &\hat{S}=\{ \hat{w}_{\dot{\beta}}(1)\ \mathrm{mod}\ \hat{T_0}\ |\ \dot{\beta}\in\Delta_a\}. \end{align*} In particular $\hat{S}$ is identified with the set $\{ \hat{w}_{\dot{\alpha}}\ |\ \dot{\alpha}\in\Pi_a\}$.\vspace{3mm} We also need several relations between $\hat{x}_{\dot{\beta}}(f)$, $\hat{w}_{\dot{\beta}}(u)$ and $\hat{h}_{\dot{\beta}}(s)$ like $(R1)\sim (R6)$, but those actually are the same as in Section 2 except $(R6)$. We give $(R6)$ in Steinberg groups as follows (cf. \cite{s}): \begin{align*} (\hat{R}6)\ &\hat{w}_{\beta}(u)\hat{h}_{\gamma}(s)\hat{w}_{\beta}(u)^{-1}\\&= \begin{cases} \hat{h}_{\gamma}(s)\hspace{2mm}&\text{if} \hspace{2mm}(\beta,\gamma)=0,\\ \hat{h}_{\mp\beta}(-u^{\mp 1}su^{\mp 1})\hat{h}_{\mp\beta}(-u^{\pm 2})^{-1} &\text{if} \hspace{2mm}\gamma=\pm\beta,\\ \hat{h}_{\sigma_{\beta}(\gamma)}(-u^{-1}s)\hat{h}_{\sigma_{\beta}(\gamma)}(-u^{-1})^{-1} &\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=k,\\ \hat{h}_{\sigma_{\beta}(\gamma)}(-su)\hat{h}_{\sigma_{\beta}(\gamma)}(-u)^{-1} &\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ i=l,\\ \hat{h}_{\sigma_{\beta}(\gamma)}(us)\hat{h}_{\sigma_{\beta}(\gamma)}(u)^{-1} &\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=k,\\ \hat{h}_{\sigma_{\beta}(\gamma)}(su^{-1})\hat{h}_{\sigma_{\beta}(\gamma)}(u^{-1})^{-1} &\text{if} \hspace{2mm}\beta\pm\gamma\neq 0\ \text{and}\ j=l,\\ \end{cases} \end{align*} where $\beta=\epsilon_i-\epsilon_j$, $\gamma=\epsilon_k-\epsilon_l\in\Delta$ and $u,s\in\Dt^{\times}$. Then we have the following.\vspace{3mm} {\lem\begin{align*} &(1)\ \hat{B}^{\pm}=\hat{U}^{\pm}\rtimes \hat{T_0},\\ &(2)\ \hat{T_0}\lhd \hat{N}\ and\ \hat{T}\lhd \hat{N},\\ &(3)\ \hat{B}^{\pm}\cap \hat{N}=\hat{T_0},\\ &(4)\ \hat{N}/\hat{T_0}\cong W_a.\hspace{5mm}\square \end{align*}} To discuss these subgroups more explicitly we put \[ \hat{U}'_{\pm\dot{\alpha}}=\langle \hat{x}_{\pm\dot{\alpha}}(g)\hat{U}_{\dot{\beta}}\hat{x}_{\pm\dot{\alpha}}(g)^{-1}\ |\ g\in D,\dot{\beta}\in\Delta^{\pm}_a\setminus\{ \pm\dot{\alpha}\}\rangle\] for each $\dot{\alpha}\in\Pi_a$. Then the following proposition holds.\vspace{3mm} {\prop\begin{align*} (1)\ &\hat{w}_{\pm\dot{\alpha}}(u)\hat{U}'_{\pm\dot{\alpha}}\hat{w}_{\pm\dot{\alpha}}(u)^{-1}=\hat{U}'_{\pm\dot{\alpha}}\ \text{for all $u\in D^{\times}$},\\ (2)\ &\hat{U}^{\pm}=\hat{U}'_{\pm\dot{\alpha}}\hat{U}_{\pm\dot{\alpha}}=\hat{U}_{\pm\dot{\alpha}}\hat{U}'_{\pm\dot{\alpha}},\\ (3)\ &\hat{s}\hat{B}^{\pm}w\subset \hat{B}^{\pm}w\hat{B}^{\pm}\cup \hat{B}^{\pm}\hat{s}w\hat{B}^{\pm}\ \text{for all $\hat{s}\in \hat{S}$ and $w\in W_a$},\\ (4)\ &\hat{s}\hat{B}^{\pm}\hat{s}\not\subset \hat{B}^{\pm}\ \text{for each $\hat{s}\in \hat{S}$}.\hspace{5mm}\square \end{align*}} {\thm{Notation is as above. $(St(n,\Dt),\hat{B}^{\pm},\hat{N},\hat{S})$ is a Tits system with the corresponding affine Weyl group $W_a$.\hspace{5mm}$\square$}}\vspace{5mm} {\cor{Notation is as above. Then\vspace{3mm}\\ $(1)$ $St(n,\Dt)$ has a Bruhat decomposition: \begin{center}$St(n,\Dt)=\bigcup_{w\in W_a}\hat{B}^{\pm}w\hat{B}^{\pm}=\hat{U}^{\pm}\hat{N}\hat{U}^{\pm}$,\end{center} $(2)$ $St(n,\Dt)$ has a Birkhoff decomposition: \begin{center}$St(n,\Dt)=\bigcup_{w\in W_a}\hat{B}^{\mp}w\hat{B}^{\pm}=\hat{U}^{\mp}\hat{N}\hat{U}^{\pm}$,\end{center} (3) $St(n,\Dt)$ has a Gauss decomposition: \begin{center}$St(n,\Dt)=\bigcup_{\hat{X}\in\hat{U}^{\pm}}\hat{X}\hat{B}^{\mp}\hat{B}^{\pm}\hat{X}^{-1 =\hat{U}^{\pm}\hat{B}^{\mp}\hat{U}^{\pm}$.\hspace{3mm}$\square$\end{center}}} \section{Central extensions} We will show that a given natural homomorphism $\phi$ in Section 4 is a central extension, i.e., $\mathrm{Ker}\phi$ is a central subgroup of $St(n,\Dt)$. For this homomorphism $\phi$ we define our $K_2$-groups as $K_2(n,\Dt)=\mathrm{Ker}\phi$ (cf.\cite{m}). We first discuss a central extension of $E(2,\Dt)$.\vspace{3mm} Let $\tilde{E}(2,\Dt)$ be the group presented by the generators $\tilde{x}_{ij}(f)$ for all $f\in\Dt$, $1\leq i\neq j\leq 2$ and defining relations $(ST1)$, $(ST2)'$ and \begin{align*} (ST3)\ \tilde{c}(u_1,v_1)\tilde{c}(u_2,v_2)\dotsm\tilde{c}(u_r,v_r)=1 \end{align*} for all $u_i,v_i\in\Dt^{\times}$ such that $[u_1,v_1][u_2,v_2]\dotsm[u_r,v_r]=1$, where for $u,v\in\Dt^{\times}$ we put \begin{align*} &\tilde{w}_{ij}(u)=\tilde{x}_{ij}(u)\tilde{x}_{ji}(-u^{-1})\tilde{x}_{ij}(u),\\ &\tilde{h}_{ij}(u)=\tilde{w}_{ij}(u)\tilde{w}_{ij}(-1),\\ &\tilde{c}(u,v)=\tilde{h}_{12}(u)\tilde{h}_{12}(v)\tilde{h}_{12}(vu)^{-1}, \end{align*} and where we change $\hat{x}$ and $\hat{w}$ in the defining relations to $\tilde{x}$ and $\tilde{w}$ respectively. That is, $\tilde{E}(2,\Dt)$ is the quotient group of $St(2,\Dt)$ by the corresponding elements $\hat{c}(u_1,v_1)\hat{c}(u_2,v_2)\dotsm\hat{c}(u_r,v_r)$. Here, we note \[\phi(\hat{c}(u,v))=\left(\begin{array}{cc} [u,v] & 0 \\ 0 & 1 \end{array}\right),\] and, in particular, we have $\hat{c}(u,v)\in\hat{T}_0$ and \begin{align*} \phi(\hat{c}(u_1,v_1)&\hat{c}(u_2,v_2)\dotsm\hat{c}(u_r,v_r))\\ &= \left(\begin{array}{cc} [u_1,v_1] & 0 \\ 0 & 1 \end{array}\right) \left(\begin{array}{cc} [u_2,v_2] & 0 \\ 0 & 1 \end{array}\right)\dots \left(\begin{array}{cc} [u_r,v_r] & 0 \\ 0 & 1 \end{array}\right)\\ &=\left(\begin{array}{cc} [u_1,v_1][u_2,v_2]\dots [u_r,v_r] & 0 \\ 0 & 1 \end{array}\right)\\ &=I \end{align*} if $[u_1,v_1][u_2,v_2]\dots [u_r,v_r]=1$. Then we obtain the following theorem.\vspace{3mm} {\thm{Notation is as above. We have $\tilde{E}(2,\Dt)\cong E(2,\Dt)$.}}\vspace{3mm}\\ \textbf{Proof of Theorem 3:} We will show this along with \cite{ms}. The homomorphism $\phi$ induces two canonical homomorphism called $\hat{\phi}$ and $\tilde{\phi}$: \begin{align*} &\hat{\phi}: St(2,\Dt)\to \tilde{E}(2,\Dt),\\ &\tilde{\phi}: \tilde{E}(2,\Dt)\to E(2,\Dt), \end{align*} which are defined by \[ \hat{\phi}(\hat{x}_{ij}(f))=\tilde{x}_{ij}(f)\ \mathrm{and}\ \tilde{\phi}(\tilde{x}_{ij}(f))=I+fE_{ij}.\] We use the same notation of subgroups of $\tilde{E}(2,\Dt)$ as in $St(2,\Dt)$ changing $\hat{}$ to $\tilde{}$. Then, for any $\tilde{x}\in\mathrm{Ker}\tilde{\phi}$, we can choose $\tilde{y}\in\tilde{U}$ and $\tilde{z}\in\tilde{T_0}$ satisfying $\tilde{x}=\tilde{y}\tilde{z}$ because $St(2,\Dt)$ has a Bruhat decomposition and $\mathrm{Ker}\tilde{\phi}\subset \tilde{B}=\tilde{U}\tilde{T_0}$. We know $\tilde{\phi}(\tilde{x})=\tilde{\phi}(\tilde{y})\tilde{\phi}(\tilde{z})=I$ and $U\cap T_0=\{ I\}$, so we have $\tilde{\phi}(\tilde{y})=I$ and $\tilde{\phi}(\tilde{z})=I$, in particular, $\tilde{y}$ and $\tilde{z}$ belong to the kernel of $\tilde{\phi}$. Therefore we need to prove $\tilde{y}=1=\tilde{z}$.\vspace{3mm} \textbf{STEP 1.} We claim $\tilde{z}=1$. We know $\tilde{w}_{21}(u)=\tilde{w}_{12}(-u^{-1})$ from $(\hat{R}3)$, and this gives us the relation $\tilde{h}_{21}(u)=\tilde{h}_{12}(u)^{-1}$. In addition we have $1=\tilde{c}(u,u^{-1})=\tilde{h}_{12}(u)\tilde{h}_{12}(u^{-1})$, that is, $\tilde{h}_{12}(u)^{-1}=\tilde{h}_{12}(u^{-1})$ by $(ST3)$. Therefore we see $\tilde{z}$ is of the form \[ \tilde{z}=\tilde{h}_{12}(u_1)\dotsm\tilde{h}_{12}(u_r)\] for $u_i\in\Dt^{\times}$. We note that $\tilde{\phi}(\tilde{z})=I$ implies $u_1u_2\cdots u_r=1$ and $u_ru_{r-1}\cdots u_1=1$. In addition, we easily see $[u_1,u_2][u_2u_1,u_3]\dotsm [u_{r-1}\dotsm u_1,u_r]=1$ using a similar way in Section 3. Then, by $(ST3)$, \begin{align*} \tilde{z}&=\tilde{h}_{12}(u_1)\tilde{h}_{12}(u_2)\dotsm\tilde{h}_{12}(u_r)\\ &=\tilde{c}(u_1,u_2)\tilde{h}_{12}(u_2u_1)\tilde{h}_{12}(u_3)\dotsm\tilde{h}_{12}(u_r)\\ &=\tilde{c}(u_1,u_2)\tilde{c}(u_2u_1,u_3)\tilde{h}_{12}(u_3u_2u_1)\tilde{h}_{12}(u_4)\dotsm\tilde{h}_{12}(u_r)\\ &\hspace{20mm}\vdots\\ &=\tilde{c}(u_1,u_2)\tilde{c}(u_2u_1,u_3)\dotsm\tilde{c}(u_{r-2}\dotsm u_1,u_{r-1})\tilde{h}_{12}(u_{r-1}\dotsm u_1) \tilde{h}_{12}(u_r)\\ &=\tilde{c}(u_1,u_2)\tilde{c}(u_2u_1,u_3)\dotsm\tilde{c}(u_{r-1}\dotsm u_1,u_r)\tilde{h}_{12}(u_r\dotsm u_1)\\ &=\tilde{c}(u_1,u_2)\tilde{c}(u_2u_1,u_3)\dotsm\tilde{c}(u_{r-1}\dotsm u_1,u_r)\\ &=1.\hspace{5mm}\square \end{align*}\vspace{3mm} \textbf{STEP 2.} We claim $\tilde{y}=1$. We set subgroups of $U$ as follows: \[ U_1=\{ e_{\alpha}(f)|f\in D[t]\}\ \mathrm{and}\ U_2=\{ e_{-\alpha}(g)|g\in D[t]t\}.\] Then $U=\langle U_1,U_2\rangle$, but we will show that this is actually a free product. We now introduce a ``degree map'' $\mathrm{deg}: D[t]\to \mathbb{Z}_{\geq 0}\cup\{ \infty\}$, which is defined by $\mathrm{deg}(f)=m$ for $f=\Sigma_{i=0}^mf_it^i$ and $\mathrm{deg}(0)=\infty$, where $f_i\in D$.\vspace{3mm} Let $x\in U$, and we assume \[x=e_{\beta_1}(q_1)e_{\beta_2}(q_2)\dotsm e_{\beta_r}(q_r),\] where $\beta_i\in\Delta$, $\beta_i\neq\beta_{i+1}$, $q_i\neq 0$ and \[q_i\in \begin{cases} D[t]\ \mathrm{if}\ \beta_i\in\Delta^+,\\ D[t]t\ \mathrm{if}\ \beta_i\in\Delta^-. \end{cases}\] For each $1\leq i\leq r$, we put \[\left(\begin{array}{cc} a_i & b_i \\ c_i & d_i \end{array}\right) =e_{\beta_1}(q_1)e_{\beta_2}(q_2)\dotsm e_{\beta_r}(q_r).\] Suppose $\beta_1=\epsilon_1-\epsilon_2$. Then we have \[\left(\begin{array}{cc} a_1 & b_1 \\ c_1 & d_1 \end{array}\right)= \left(\begin{array}{cc} 1 & q_1 \\ 0 & 1 \end{array}\right)\] and $0=\mathrm{deg}(a_1)\leq\mathrm{deg}(b_1)$. Since \[\left(\begin{array}{cc} a_2 & b_2 \\ c_2 & d_2 \end{array}\right)= \left(\begin{array}{cc} a_1 & b_1 \\ c_1 & d_1 \end{array}\right) \left(\begin{array}{cc} 1 & 0 \\ q_2 & 1 \end{array}\right)= \left(\begin{array}{cc} a_1+b_1q_2 & b_1 \\ c_1+d_1q_2 & d_1 \end{array}\right),\] we obtain $\mathrm{deg}(a_2)>\mathrm{deg}(b_2)=\mathrm{deg}(b_1)$. Moreover, since \[\left(\begin{array}{cc} a_3 & b_3 \\ c_3 & d_3 \end{array}\right)= \left(\begin{array}{cc} a_2 & b_2 \\ c_2 & d_2 \end{array}\right) \left(\begin{array}{cc} 1 & q_3 \\ 0 & 1 \end{array}\right) \left(\begin{array}{cc} a_2 & a_2q_3+b_2 \\ c_2 & c_2q_3+d_2 \end{array}\right),\] we have $\mathrm{deg}(a_2)=\mathrm{deg}(a_3)\leq\mathrm{deg}(b_3)$. Continuing this process, we can reach \begin{align*} &\mathrm{deg}(a_1)\leq\mathrm{deg}(b_1)=\mathrm{deg}(b_2)<\mathrm{deg}(a_2)=\\ &\mathrm{deg}(a_3)\leq\mathrm{deg}(b_3)=\mathrm{deg}(b_4)<\mathrm{deg}(a_4)=\\ &\mathrm{deg}(a_5)\leq\mathrm{deg}(b_5)=\mathrm{deg}(b_6)<\mathrm{deg}(a_6)=\dotsm . \end{align*} Next, suppose $\beta_1=\epsilon_2-\epsilon_1$. Then, in the same way as above, we can get \begin{align*} &\mathrm{deg}(b_1)\leq\mathrm{deg}(a_1)=\mathrm{deg}(a_2)<\mathrm{deg}(b_2)=\\ &\mathrm{deg}(b_3)\leq\mathrm{deg}(a_3)=\mathrm{deg}(a_4)<\mathrm{deg}(b_4)=\\ &\mathrm{deg}(b_5)\leq\mathrm{deg}(a_5)=\mathrm{deg}(a_6)<\mathrm{deg}(b_6)=\dotsm . \end{align*} In any case we can show $x\neq 1$, which means that $U=U_1\star U_2$ is a free product, and we have $\tilde{\phi}(\tilde{y})=I$ implies $\tilde{y}=1$. $\square$\vspace{3mm} {\prop{Notation is as above. We obtain that \[ K_2(2,\Dt)=\langle \hat{c}(u_1,v_1)\hat{c}(u_2,v_2)\dotsm\hat{c}(u_r,v_r) | u_i,v_i\in\Dt^{\times}; [u_1,v_1][u_2,v_2]\dotsm [u_r,v_r]=1\rangle,\] where $\hat{c}(u,v)=\hat{h}_{12}(u)\hat{h}_{12}(v)\hat{h}_{12}(vu)^{-1}$, and $K_2(2,\Dt)$ is a central subgroup of $St(2,\Dt)$.}}\vspace{5mm} \textbf{Proof:} It suffices to check that $\hat{c}(u_1,v_1)\hat{c}(u_2,v_2)\dotsm\hat{c}(u_r,v_r)$ is a central element when $[u_1,v_1][u_2,v_2]\dotsm [u_r,v_r]=1$. We easily see for $f\in\Dt$ and $1\leq i\neq j\leq 2$ \begin{align*} \{ \hat{c}(u_1,v_1)\hat{c}(u_2,v_2)&\dotsm\hat{c}(u_r,v_r)\} \hat{x}_{ij}(f) \{ \hat{c}(u_1,v_1)\hat{c}(u_2,v_2)\dotsm\hat{c}(u_r,v_r)\} ^{-1}\\ &=\hat{x}_{ij}(\{ [u_1,v_1][u_2,v_2]\dotsm [u_r,v_r]\} ^{\pm 1}f\{ [u_1,v_1][u_2,v_2]\dotsm [u_r,v_r]\} ^{\mp 1})\\ &=\hat{x}_{ij}(f).\ \square \end{align*}\vspace{3mm} Using the Proposition 4 and \cite{mj}, we construct a central extension of higher rank. Suppose $n\geq 3$. We now consider the following commutative diagram; \begin{align*} 1 \rightarrow K_2&(2,D_\tau) \rightarrow St(2,D_{\tau}) \rightarrow E(2,D_{\tau}) \rightarrow 1 \hspace{5mm}(\text{exact})\\ &\downarrow\hspace{20mm}\downarrow\hspace{20mm}\downarrow\\ 1 \rightarrow K_2&(n,D_\tau) \rightarrow St(n,D_{\tau}) \rightarrow E(n,D_{\tau}) \rightarrow 1 \hspace{5mm}(\text{exact}) \end{align*} Then we see that a canonical homomorphism of $K_2(2,\Dt)$ into $K_2(n,\Dt)$ is surjective by \cite{mj} since $\Dt$ is a euclidean ring, so we have;\vspace{3mm} {\thm{Let $\tilde{E}(n,\Dt)$ be the group presented by the generators $\tilde{x}_{ij}(f)$ for all $1\leq i\neq j\leq n$ and $f\in\Dt$ with the defining relations $(ST1)$, $(ST2)$ and $(ST3)$. Then $\tilde{E}(n,\Dt)\cong E(n,\Dt)$. $\square$}}\vspace{3mm}\\ {\prop{We have \[ K_2(n,\Dt)=\langle \hat{c}(u_1,v_1)\hat{c}(u_2,v_2)\dotsm\hat{c}(u_r,v_r) | u_i,v_i\in\Dt^{\times}; [u_1,v_1][u_2,v_2]\dotsm [u_r,v_r]=1\rangle,\] where $\hat{c}(u,v)=\hat{h}_{12}(u)\hat{h}_{12}(v)\hat{h}_{12}(vu)^{-1}$, and $K_2(n,\Dt)$ is a central subgroup of $St(n,\Dt)$. $\square$}}\vspace{3mm} \section{Universal central extensions} We will show that the natural homomorphism $\phi$ is universal when $|Z(D)|\geq 5$ and $|Z(D)|\neq 9$, where $Z(D)$ is the center of $D$. We first fix four central elements $a,b,c,d\in D$, which satisfy \[ a^2-1\neq 0,\ b=(a^2-1)^{-1},\ c\neq 0,\ c-1\neq 0,\ c^2-c+1\neq 0,\ d^3-1\neq 0.\] We note that $E(n,\Dt)$ is perfect (also $St(n,\Dt)$) since \[ x_{\beta}(f)=[h_{\beta}(a),x_{\beta}(bf)],\] where $\beta\in\Delta$, $f\in\Dt$. Let $\phi^*:E^*\to E(n,\Dt)$ be any central extension, and set \[M(z)=\{ z^*\in E^* | \phi^*(z^*)=z\} \] for $z\in E(n,\Dt)$. Then we define \[x_{\beta}^*(f)=[h^*,x^*]\] for $h^*\in M(h_{\beta}(a))$ and $x^*\in M(x_{\beta}(bf))$. We see this is well-defined from the following lemma, which is called ``central trick''.\vspace{3mm} {\lem{Let $\psi: K\to G$ be a central extension. Then for $X,X',Y,Y'\in K $, we obtain that $\psi(X)=\psi(X')$ and $\psi(Y)=\psi(Y')$ imply $[X,Y]=[X',Y']$.}}\vspace{3mm}\\ \textbf{Proof:} This holds by direct calculation. $\square$\vspace{3mm} For $u\in \Dt$ we put \begin{align*} w_{\beta}^*(u)&=x_{\beta}^*(u)x_{\beta}^*(-u^{-1})x_{\beta}^*(u),\\ h_{\beta}^*(u)&=w_{\beta}^*(u)w_{\beta}^*(-1). \end{align*} {\lem{Let $\beta\in\Delta$, $f\in\Dt$ and $u\in\Dt^{\times}$. Then the following equations hold. \begin{align*} &(1)\ w_{\beta}^*(u)x_{\beta}^*(f)w_{\beta}^*(u)^{-1}=x_{-\beta}^*(-u^{-1}fu^{-1})\\ &(2)\ h_{\beta}^*(u)x_{\beta}^*(f)h_{\beta}^*(u)^{-1}=x_{\beta}^*(ufu)\\ \end{align*}}}\vspace{3mm} \textbf{Proof:} \begin{align*} (1)\ w_{\beta}^*(u)x_{\beta}^*(f)w_{\beta}^*(u)^{-1} &=w_{\beta}^*(u)[h_{\beta}^*(a),x_{\beta}^*(bf)]w_{\beta}^*(u)^{-1}\\ &=[w_{\beta}^*(u)h_{\beta}^*(a)w_{\beta}^*(u)^{-1},w_{\beta}^*(u)x_{\beta}^*(bf)w_{\beta}^*(u)^{-1}]\\ &=[h_{-\beta}^*(a),x_{-\beta}^*(-bu^{-1}fu^{-1})]\ \text{(by central trick)}\\ &=x_{-\beta}^*(-u^{-1}fu^{-1}). \end{align*} \begin{align*} (2)\ h_{\beta}^*(u)x_{\beta}^*(f)h_{\beta}^*(u)^{-1} &=h_{\beta}^*(u)[h_{\beta}^*(a),x_{\beta}^*(bf)]h_{\beta}^*(u)^{-1}\\ &=[h_{\beta}^*(u)h_{\beta}^*(a)h_{\beta}^*(u)^{-1},h_{\beta}^*(u)x_{\beta}^*(bf)h_{\beta}^*(u)^{-1}]\\ &=[h_{\beta}^*(a),x_{\beta}^*(bufu)]\ \text{(by central trick)}\\ &=x_{\beta}^*(ufu).\hspace{10mm}\square \end{align*}\vspace{3mm} We now define $\pi_{\beta,\gamma}(f,g)$ for $\beta,\gamma\in\Delta$ and $f,g\in \Dt$ by \begin{align*} \pi_{\beta,\gamma}(f,g)= \begin{cases} [x_{\beta}^*(f),x_{\gamma}^*(g)]x_{\beta+\gamma}^*(fg)^{-1}\hspace{5mm}&\text{if $j=k$},\\ [x_{\beta}^*(f),x_{\gamma}^*(g)]x_{\beta+\gamma}^*(-gf)^{-1}&\text{if $i=l$},\\ [x_{\beta}^*(f),x_{\gamma}^*(g)]&\text{otherwise}, \end{cases} \end{align*} where $\beta=\epsilon_i-\epsilon_j$, $\gamma=\epsilon_k-\epsilon_l$, $i\neq j$ and $k\neq l$. We note $\pi_{\beta,\gamma}(f,g)$ is central in $E^*$. Then we can show the following (cf.\cite{s}).\vspace{3mm} {\lem{Let $\beta=\epsilon_i-\epsilon_j,\gamma=\epsilon_k-\epsilon_l\in\Delta$ and $f,f',g,g'\in\Dt$. Then \begin{align*} &(1)\ \pi_{\beta,\gamma}(f+f',g)=\pi_{\beta,\gamma}(f,g)\pi_{\beta,\gamma}(f',g),\\ &(2)\ \pi_{\beta,\gamma}(f,g+g')=\pi_{\beta,\gamma}(f,g)\pi_{\beta,\gamma}(f,g'),\\ &(3)\ \pi_{\beta,\gamma}(f,g)=1,\\ &(4)\ x_{\beta}^*(f)x_{\beta}^*(g)=x_{\beta}^*(f+g). \end{align*}}} \textbf{Proof:} We will prove this lemma dividing into the three steps: we first check (1), (2) and (3) only for $j\neq k$ and $i\neq l$, and then show that (4) holds for all $\beta\in\Delta$, $f,g\in\Dt$. Finally, we check (1), (2) and (3) for the others.\vspace{3mm} \textbf{Step 1.} Let $j\neq k,i\neq l$. Then, by definition of $\pi_{\beta,\gamma}(f,g)$, \begin{align*} \pi_{\beta,\gamma}(f+f',g)&=[x_{\beta}^*(f+f'),x_{\gamma}^*(g)]\\ &=[x_{\beta}^*(f)x_{\beta}^*(f'),x_{\gamma}^*(g)]\\ &=x_{\beta}^*(f)x_{\beta}^*(f')x_{\gamma}^*(g)x_{\beta}^*(f')^{-1}x_{\beta}^*(f)^{-1}x_{\gamma}^*(g)^{-1}\\ &=x_{\beta}^*(f)\pi_{\beta,\gamma}(f',g)x_{\gamma}^*(g)x_{\beta}^*(f)^{-1}x_{\gamma}^*(g)^{-1}\\ &=\pi_{\beta,\gamma}(f,g)\pi_{\beta,\gamma}(f',g). \end{align*} \begin{align*} \pi_{\beta,\gamma}(f,g+g')&=[x_{\beta}^*(f),x_{\gamma}^*(g+g')]\\ &=[x_{\beta}^*(f),x_{\gamma}^*(g)x_{\gamma}^*(g')]\\ &=x_{\beta}^*(f)x_{\gamma}^*(g)x_{\gamma}^*(g')x_{\beta}^*(f)^{-1}x_{\gamma}^*(g')^{-1}x_{\gamma}^*(g)^{-1}\\ &=x_{\beta}^*(f)x_{\gamma}^*(g)x_{\beta}^*(f)^{-1}\pi_{\beta,\gamma}(f,g')x_{\gamma}^*(g)^{-1}\\ &=\pi_{\beta,\gamma}(f,g)\pi_{\beta,\gamma}(f,g'). \end{align*}\vspace{3mm} If $i\neq k$ and $j\neq l$, then \begin{align*} \pi_{\beta,\gamma}(bf,g)&=h^*_{\beta}(a)\pi_{\beta,\gamma}(bf,g)h^*_{\beta}(a)^{-1}\\ &=[h^*_{\beta}(a)x^*_{\beta}(bf)h^*_{\beta}(a)^{-1}, h^*_{\beta}(a)x^*_{\gamma}(g)h^*_{\beta}(a)^{-1}]\\ &=[x^*_{\beta}(a^2bf),x^*_{\gamma}(g)]\hspace{5mm}\text{(by central trick and (R4))}\\ &=\pi_{\beta,\gamma}(a^2bf,g) \end{align*} and \begin{align*} \pi_{\beta,\gamma}(f,g)&=\pi_{\beta,\gamma}(a^2bf-bf,g)\\ &=[x^*_{\beta}(a^2bf)x^*_{\beta}(bf)^{-1},x^*_{\gamma}(g)]\\ &=x^*_{\beta}(a^2bf)x^*_{\beta}(bf)^{-1}\pi_{\beta,\gamma}(bf,g)^{-1}x^*_{\beta}(bf)x^*_{\gamma}(g)x^*_{\beta}(a^2bf)^{-1}x^*_{\gamma}(g)^{-1}\\ &=\pi_{\beta,\gamma}(a^2bf,g)\pi_{\beta,\gamma}(bf,g)^{-1}, \end{align*} so we obtain $\pi_{\beta,\gamma}(f,g)=1$.\vspace{3mm} If $i=k$ and $j\neq l$, then \begin{align*} \pi_{\beta,\gamma}(f,g)&=h^*_{\beta}(d)\pi_{\beta,\gamma}(f,g)h^*_{\beta}(d)^{-1}\\ &=\pi_{\beta,\gamma}(d^2f,dg)\hspace{5mm}\text{(by central trick and (R4))} \end{align*} and \begin{align*} \pi_{\beta,\gamma}(f,g)&=h^*_{\beta-\gamma}(d)\pi_{\beta,\gamma}(f,g)h^*_{\beta-\gamma}(d)^{-1}\\ &=\pi_{\beta,\gamma}(df,d^{-1}g)\hspace{5mm}\text{(by central trick and (R4))}, \end{align*} so we obtain $\pi_{\beta,\gamma}(f,g)=\pi_{\beta,\gamma}(d^2f,dg)=\pi_{\beta,\gamma}(d^3f,g)$, hence $\pi_{\beta,\gamma}((d^3-1)f,g)=1$. Therefore $\pi_{\beta,\gamma}(f,g)=1$.\vspace{3mm} If $i\neq k$ and $j=l$, then using the same way as the previous case, we see \[\pi_{\beta,\gamma}((d^3-1)f,g)=1.\] Hence $\pi_{\beta,\gamma}(f,g)=1$.\vspace{3mm} If $i=k$ and $j=l$, then \[\pi_{\beta,\beta}(f,g)=h_{\beta}^*(c)\pi_{\beta,\beta}(f,g)h_{\beta}^*(c)^{-1}=[x_{\beta}^*(c^2f),x_{\beta}^*(c^2g)]=\pi_{\beta,\beta}(c^2f,c^2g),\] so we see \begin{align*} \pi_{\beta,\gamma}(f,g)&=\pi_{\beta,\gamma}(cf+(1-c)f,g)\\ &=\pi_{\beta,\gamma}(cf,g)\pi_{\beta,\gamma}((1-c)f,g)\\ &=\pi_{\beta,\gamma}(f,g/c)\pi_{\beta,\gamma}(f,g/(1-c))\\ &=\pi_{\beta,\gamma}(f,g/c+g/(1-c))\\ &=\pi_{\beta,\gamma}(f,g/c(1-c))\\ &=\pi_{\beta,\gamma}(c(1-c)f,g), \end{align*} hence $\pi_{\beta,\gamma}((c^2-c+1)f,g)=1$. Therefore we obtain $\pi_{\beta,\gamma}(f,g)=1$, in particular, \[ x^*_{\beta}(f)x^*_{\beta}(g)=x^*_{\beta}(g)x^*_{\beta}(f)\] for all $f,g\in\Dt$.\vspace{3mm} \textbf{Step 2.} We now put $(f,g)=x_{\beta}^*(f)x_{\beta}^*(g)x_{\beta}^*(f+g)^{-1}$. Then, by Step 1, \begin{align*} (bf,bg)&=h_{\beta}^*(a)(bf,bg)h_{\beta}^*(a)^{-1}\\ &=h_{\beta}^*(a)x_{\beta}^*(bf)x_{\beta}^*(bg)x_{\beta}^*(b(f+g))^{-1}h_{\beta}^*(a)^{-1}\\ &=[h_{\beta}^*(a),x_{\beta}^*(bf)]x_{\beta}^*(bf)h_{\beta}^*(a)x_{\beta}^*(bg)x_{\beta}^*(b(f+g))^{-1}h_{\beta}^*(a)^{-1}\\ &=[h_{\beta}^*(a),x_{\beta}^*(bf)]x_{\beta}^*(bf)[h_{\beta}^*(a),x_{\beta}^*(bg)]x_{\beta}^*(bg)\\ &\hspace{10mm}\cdot [h_{\beta}^*(a),x_{\beta}^*(b(f+g))^{-1}]x_{\beta}^*(b(f+g))^{-1}\\ &=x_{\beta}^*(f)x_{\beta}^*(bf)x_{\beta}^*(g)x_{\beta}^*(bg)x_{\beta}^*(f+g)^{-1}x_{\beta}^*(b(f+g))^{-1}\\ &=(f,g)(bf,bg). \end{align*} Thus $(f,g)=1$ for all $f,g\in\Dt$, hence $x_{\beta}^*(f)x_{\beta}^*(g)=x_{\beta}^*(f+g)$.\vspace{3mm} \textbf{Step 3.} Let $j=k$. Then, by Step 1 and Step 2, \begin{align*} \pi_{\beta,\gamma}(f+f',g)&=[x^*_{\beta}(f+f'),x^*_{\gamma}(g)]x^*_{\beta+\gamma}(-(f+f')g)\\ &=x^*_{\beta}(f)x^*_{\beta}(f')x^*_{\gamma}(g)x^*_{\beta}(-f')x^*_{\beta}(-f)x^*_{\gamma}(-g)x^*_{\beta+\gamma}(-fg)x^*_{\beta+\gamma}(-f'g)\\ &=x^*_{\beta}(f)[x^*_{\beta}(f'),x^*_{\gamma}(g)]x^*_{\gamma}(g)x^*_{\beta}(-f)x^*_{\gamma}(-g)x^*_{\beta+\gamma}(-fg)x^*_{\beta+\gamma}(-f'g)\\ &=x^*_{\beta}(f)\pi_{\beta,\gamma}(f',g)x^*_{\beta+\gamma}(f'g)x^*_{\beta}(-f)\pi_{\beta,\gamma}(f,g)x^*_{\beta+\gamma}(-f'g)\\ &=\pi_{\beta,\gamma}(f,g)\pi_{\beta,\gamma}(f',g) \end{align*} and we similarly obtain \begin{align*} \pi_{\beta,\gamma}(f,g+g')&=[x^*_{\beta}(f),x^*_{\gamma}(g+g')]x^*_{\beta+\gamma}(-f(g+g'))\\ &=x^*_{\beta}(f)x^*_{\gamma}(g)x^*_{\gamma}(g')x^*_{\beta}(-f)x^*_{\gamma}(-g')x^*_{\gamma}(-g)x^*_{\beta+\gamma}(-fg)x^*_{\beta+\gamma}(-fg')\\ &=x^*_{\beta}(f)x^*_{\gamma}(g')x^*_{\beta}(-f)[x^*_{\beta}(f),x^*_{\gamma}(g)]x^*_{\gamma}(-g')x^*_{\beta+\gamma}(-fg)x^*_{\beta+\gamma}(-fg')\\ &=\pi_{\beta,\gamma}(f,g')x^*_{\beta+\gamma}(fg')\pi_{\beta,\gamma}(f,g)x^*_{\gamma}(-g)x^*_{\beta+\gamma}(-fg)x^*_{\beta+\gamma}(-fg')\\ &=\pi_{\beta,\gamma}(f,g)\pi_{\beta,\gamma}(f,g'). \end{align*} Also, we see \begin{align*} \pi_{\beta,\gamma}(f,g)&=h^*_{\beta}(d)\pi_{\beta,\gamma}(f,g)h^*_{\beta}(d)^{-1}\\ &=h^*_{\beta}(d)[x^*_{\beta}(f),x^*_{\gamma}(g)]x^*_{\beta+\gamma}(-fg)h^*_{\beta}(d)^{-1}\\ &=[x^*_{\beta}(d^2f),x^*_{\gamma}(d^{-1}g)]x^*_{\beta+\gamma}(-dfg)\\ &=\pi_{\beta,\gamma}(d^2f,d^{-1}g) \end{align*} and \begin{align*} \pi_{\beta,\gamma}(f,g)&=h^*_{\beta+\gamma}(d)\pi_{\beta,\gamma}(f,g)h^*_{\beta+\gamma}(d)^{-1}\\ &=[x^*_{\beta}(df),x^*_{\gamma}(dg)]x^*_{\beta+\gamma}(-d^2fg)\\ &=\pi_{\beta,\gamma}(df,dg). \end{align*} Therefore we obtain $\pi_{\beta,\gamma}((d^3-1)f,g)=1$, in particular, $\pi_{\beta,\gamma}(f,g)=1$.\vspace{3mm} Let $i=l$. Then, using the similar way to the case $j=k$, we obtain \begin{align*} \pi_{\beta,\gamma}(f+f',g)&=\pi_{\beta,\gamma}(f,g)\pi_{\beta,\gamma}(f',g),\\ \pi_{\beta,\gamma}(f,g+g')&=\pi_{\beta,\gamma}(f,g)\pi_{\beta,\gamma}(f,g'),\\ \pi_{\beta,\gamma}(f,g)&=1.\ \square \end{align*}\vspace{3mm} From the above, we have the following (cf. [7, p.51]). {\thm{If $|Z(D)|\geq 5$ and $|Z(D)|\neq 9$, then $\phi$ is a universal central extension of $E(n,\Dt)$.}}\vspace{3mm} \textbf{Proof:} For any central extension $\phi^*$, we can construct $x_{\beta}^*(f)$ for all $\beta\in\Phi$ and $f\in\Dt$ as above. Then the relations give us a homomorphism $\hat{\phi}^*:St(n,\Dt)\to E^*$ defined by $\hat{\phi}^*(e_{\beta}(f))=x_{\beta}^*(f)$. Thus we obtain $\phi=\phi^*\circ \hat{\phi}^*$. $\square$\vspace{5mm} \begin{rem} If the cardinality of $Z(D)$ is less than 5 or is 9, then there exist counter examples which are written in \cite{s}. \end{rem} \begin{rem} Let $F$ be any field and $x_i$ $(i\in\mathbb{Z})$ be countable indeterminates. We put \[K=F(x_i)_{i\in\mathbb{Z}}=\left\{ \frac{f(x_{i_1},\cdots,x_{i_r})}{g(x_{j_1},\dots,x_{j_s})}\ \Bigg|\ f,g\in F[x_i]_{i\in\mathbb{Z}} \right\}.\] We define an automorphism $\tau$ of $K$ to be $\tau(x_i)=x_{i+1}$. Furthermore let $t$ be a new indeterminate, and put \[ D=K((t))=\left\{ \sum_{-\infty<<k<\infty}a_kt^k\ \Bigg|\ a_k\in K\right\}. \] Then, $D$ can be viewed as a division ring, whose product is given by \[\left(\sum_ka_kt^k\right)\left(\sum_lb_lt^l\right)=\sum_m\left(\sum_{k+l=m}a_k\tau^k(b_l)\right)t^m.\] In particular, for any field $F$ there exists a division ring $D$ such that $Z(D)=F$. \end{rem} \section{$K_1$-groups} As a subgroup of $GL(n,\Dt)$, if we put $H=\{ \mathrm{diag}(u_1,\dots, u_n)| u_i\in \Dt^{\times}\}$, then $GL(n,\Dt)=\langle E(n,\Dt),H\rangle$ since $\Dt$ is a euclidean ring (cf. \cite{ms},\cite{hs} Proposition 1.1.2). Put $H_1=\{ \mathrm{diag}(u,1,\dots,1)|u\in\Dt^{\times} \}$, and we have \begin{align*} GL(n,\Dt)&=\langle E(n,\Dt), H\rangle\\ &=\langle E(n,\Dt), H_1\rangle\\ &\rhd E(n,\Dt). \end{align*} Define our $K_1$-group by $K_1(n,\Dt)=GL(n,\Dt)/E(n,\Dt)$ (cf. \cite{m}). Then, we have \[ K_1(n,\Dt)=GL(n,\Dt)/E(n,\Dt)\cong H_1/(E(n,\Dt)\cap H_1).\] In the following we will discuss the subgroup $E(n,\Dt)\cap H_1$. Since $E(n,\Dt)$ has a Bruhat decomposition, we consider $BwB\cap H_1$. From Proposition 1, there exist $\dot{w}\in W$ and $h=\mathrm{diag}(t^{m_1},\dots,t^{m_n})$ such that $w=\dot{w}h$, where $m_i\in\mathbb{Z}$ with $m_1+\cdots+m_n=0$, and then \[BwB\cap H_1=UT_0wT_0U\cap H_1=U\dot{w}hT_0U\cap H_1.\] If we set \[\Dt^{\geq 0}=D[t]\ \text{and}\ \Dt^{>0}=D[t]t,\] then we see \[U\subset \left( \begin{array}{cccc} 1+\Dt^{>0} & \Dt^{\geq 0} & \cdots & \Dt^{\geq 0}\\ \Dt^{>0} & 1+\Dt^{>0} & & \vdots\\ \vdots & & \ddots & \Dt^{\geq 0}\\ \Dt^{>0} & \cdots & \Dt^{>0} &1+\Dt^{>0} \end{array} \right).\] Suppose $BwB\cap H_1\neq\emptyset$. Then \begin{align*} BwB\cap H_1\neq\emptyset &\Rightarrow U\dot{w}hT_0U\cap H_1\neq\emptyset\\ &\Rightarrow U\cap H_1UT_0h^{-1}{\dot{w}}^{-1}\neq\emptyset. \end{align*} Therefore we can take $d\in H_1$ of degree $m$, $x_+\in U$ and $h_0=\mathrm{diag}(u_1,\dots,u_n)\in T_0$ satisfying $x=dx_+h_0h^{-1}{\dot{w}}^{-1}\in U$. Then we see \[x\in \left( \begin{array}{cccc} t^{m-m_1}(1+\Dt^{>0}) & t^{m-m_2}\Dt^{\geq 0} & \cdots & t^{m-m_n}\Dt^{\geq 0} \\ t^{-m_1}\Dt^{>0} & t^{-m_2}(1+\Dt^{>0}) & \cdots & t^{-m_n}\Dt^{\geq 0}\\ \vdots & \vdots & \ddots & \vdots\\ t^{-m_1}\Dt^{>0} & t^{-m_2}\Dt^{>0} & \cdots & t^{-m_n}(1+\Dt^{>0}) \end{array} \right) \dot{w}^{-1}.\] If $m_i>0$ for some $2\leq i\leq n$, then $x$ has an entry with a negative power of $t$ in a diagonal part. This is a contradiction, hence $m_i\leq 0$ for all $2\leq i\leq n$. In particular we obtain $m_1\geq 0$ from $m_1+\cdots+m_n=0$. Suppose that $\dot{w}^{-1}$ is a permutation sending $j$th column to the 1st one for some $2\leq j\leq n$. Then $2\leq \dot{w}^{-1}(k)\leq n$ for all $2\leq k\leq n$ without $k= j$. If $\dot{w}^{-1}(i_1)>\dot{w}^{-1}(i_2)$ for $2\leq i_1<i_2\leq n$ except $j$, then we see $t^{-m_{i_1}}\Dt^{>0}\cap \Dt^{\geq 0}=\emptyset$. This is a contradiction. Thus, we obtain $\dot{w}^{-1}(k)=k$ for such $k$. Hence, it suffices to check the two cases: When $\dot{w}^{-1}$ is a transposition of the 1st column and the $j$th column, and when $\dot{w}^{-1}$ is an identity. Suppose that $\dot{w}^{-1}$ is a transposition $(1,j)$. Then we have $t^{m-m_1}(1+\Dt^{>0})\cap\Dt^{\geq 0}\neq\emptyset$, and which implies $n-m_1\geq 0$. On the other hand, we see $t^{m-m_j}\Dt^{\geq 0}\cap(1+\Dt^{>0})\neq\emptyset$, and which implies $m-m_j\leq 0$. If we consider $m_1+m_j=0$, we obtain \[ 0\leq -m_j\leq m\leq m_j\leq 0,\] hence, $m=m_1=m_j=0$. However, the 1st column cannot be going to the $j$th column from the fact $(1+\Dt^{>0})\cap\Dt^{>0}=\emptyset$. Therefore $\dot{w}=1$. Thus, we obtain \begin{align*} E(n,\Dt)\cap H_1=B\cap H_1&=(U\rtimes T_0) \cap H_1\\ &=T_0\cap H_1\\ &\subset T\cap H_1. \end{align*} In the following, for $\beta=\epsilon_i-\epsilon_j$ we denote the element $h_{\beta}(s)$ in $T$ by $h_{ij}(s)$. From the definition of $T$ and the relation $h_{ij}(s)=h_{1j}(s)h_{1i}(s^{-1})$, we find that $T$ is generated by $h_{1j}(s)$ for all $2\leq j\leq n$ and $s\in\Dt^{\times}$. Thus every element $h$ in $T\cap H_1$ can be of the form \[ h=h_{1,l_1}(s_1)h_{1,l_2}(s_2)\cdots h_{1,l_k}(s_k),\] where $2\leq l_1,\dots, l_k\leq n$ and $s_i\in\Dt^{\times}$. Here, we put $I_j=\{ i| l_i=j\}$ for $2\leq j\leq n$. Then we see $\{ l_1,\dots, l_k\}=I_2\cup I_3\cup\cdots\cup I_n$. Using this notation we find \[h=\left( \begin{array}{cccc} \prod_{i=1}^ks_i & 0 & \cdots & 0 \\ 0 & \prod_{i\in I_2}s_i^{-1} & & \huge{0}\\ \vdots & & \ddots & \\ 0 & \huge{0} & & \prod_{i\in I_n}s_i^{-1} \end{array}\right),\] where $\prod_{i\in I_j}s_i^{-1}=1$ for all $2\leq j\leq n$. Therefore, if we put $s=\prod_{i=1}^ks_i$, we can rewrite $s$ as \[ s=s_1\cdots s_k=s_1\cdots s_ks_{i_1}^{-1}\dots s_{i_k}^{-1},\] where $\{ i_1\dots,i_k\}=\{ 2,\dots,n\}$. If $i_r=1$, then putting $v_1=s_2\cdots s_ks_{i_1}^{-1}\cdots s_{i_{r-1}}^{-1}$ we have \begin{align*} s&=[s_1,v_1]v_1s_1s_{i_r}^{-1}\cdots s_{i_k}^{-1}\\ &=[s_1,v_1]v_1s_{i_{r+1}}^{-1}\cdots s_{i_k}^{-1}\\ &=[s_1,v_1]s_2\cdots s_ks_{i_1}^{-1}\cdots s_{i_{r-1}}^{-1}s_{i_{r+1}}^{-1}\cdots s_{i_k}^{-1}. \end{align*} If $i_{r'}=2$, then putting $v_2=s_3\cdots s_ks_{i_1}^{-1}\cdots s_{i_{r'-1}}^{-1}$ we also have \[ s=[s_1,v_1][s_2,v_2]s_3\cdots s_ks_{i_1}^{-1}\cdots s_{i_{r'-1}}^{-1}s_{i_{r'+1}}^{-1}\cdots s_{i_{r-1}}^{-1}s_{i_{r+1}}^{-1}\cdots s_{i_k}^{-1}.\] Repeating this operation, we finally reach \[ s=[s_1,v_1][s_2,v_2]\cdots [s_k,v_k].\] Thus, we obtain \[T\cap H_1\subset \left( \begin{array}{cccc} [\Dt^{\times},\Dt^{\times}] & 0 & \cdots & 0 \\ 0 & 1 & & 0\\ \vdots & & \ddots & \\ 0 & 0 & & 1 \end{array}\right).\] On the other hand, we see \[\left( \begin{array}{cccc} [u,v] & 0 & \cdots & 0 \\ 0 & 1 & & 0\\ \vdots & & \ddots & \\ 0 & 0 & & 1 \end{array}\right)=h_{12}(u)h_{12}(v)h_{12}(u^{-1}v^{-1}),\] where $u,v\in\Dt^{\times}$. We see $\mathrm{deg}([u,v])=0$, hence, \[ \left( \begin{array}{cccc} [\Dt^{\times},\Dt^{\times}] & 0 & \cdots & 0 \\ 0 & 1 & & 0\\ \vdots & & \ddots & \\ 0 & 0 & & 1 \end{array}\right) \subset T_0\cap H_1.\] These deduce \[E(n,\Dt)\cap H_1= \left( \begin{array}{cccc} [\Dt^{\times},\Dt^{\times}] & 0 & \cdots & 0 \\ 0 & 1 & & 0\\ \vdots & & \ddots & \\ 0 & 0 & & 1 \end{array}\right).\] Therefore we obtain the following theorem.\vspace{3mm} {\thm{Notation is as above. Then, we have \[ K_1(n,\Dt)\cong \Dt^{\times}/[\Dt^{\times},\Dt^{\times}].\hspace{3mm}\square\]}}

https://arxiv.org/abs/1512.05076

Generalized Quantum Statistics and Lie (Super)Algebras

Generalized quantum statistics, such as paraboson and parafermion statistics, are characterized by triple relations which are related to Lie (super)algebras of type B. The correspondence of the Fock spaces of parabosons, parafermions as well as the Fock space of a system of parafermions and parabosons to irreducible representations of (super)algebras of type B will be pointed out. Example of generalized quantum statistics connected to the basic classical Lie superalgebra B(1|1)=osp(3|2) with interesting physical properties, such as noncommutative coordinates, will be given. Therefore the article focuses on the question, addressed already in 1950 by Wigner: do the equation of motion determine the quantum mechanical commutation relation?

\section{INTRODUCTION} Two classes of particles, bosons and fermions, are usually considered in quantum mechanics. Our interest is in their generalization and more precisely in certain generalized quantum systems. The latter have interesting mathematical structures and representation theory of Lie (super)algebras plays an important role. The aim of the present paper is further exploration the connection between representation theory and quantum systems. In 1950, Wigner~\cite{Wigner} asked himself the following question 'Do the equations of motion determine the quantum mechanical commutation relations?' He has shown on the example of a one-dimensional harmonic oscillator with a Hamiltonian $H=\frac{1}{2}(p^2+r^2) \;\;(m=\omega =\hbar=1)$ that one can derive the canonical commutation relations (CCRs) $[p,r]=-i$, assuming that Hamilton's equations $\dot{r}=p,\; \dot{p}=-r$ are identical with the Heisenberg equations $\dot{r}=i[H,r]$ and $\dot{p}=i[H,p]$. It is known~\cite{Ehrenfest} that the inverse is true. From the Heisenberg equations and the CCRs one obtains Hamilton's equations and from Hamilton's equations and the CCRs one derives the Heisenberg equations. As a result the question of Wigner was whether one can generalize the concept of a quantum system. He observed that the Hamilton's equations can be identical to the Heisenberg equations for position and momentum operators, which do not necessary satisfy the CCRs. Nowadays such systems are known as Wigner quantum systems (WQS)~\cite{Palev1}. In most of the cases the geometrical properties of such systems are investigated using representation theory of Lie superalgebras. The work of Wigner~\cite{Wigner} led to the theory of parabosons and parafermions introduced by H.S. Green~\cite{Green} in 1953. Paraboson and parafermion statistics are generalizations of ordinary Bose-Einstein and Fermi-Dirac statistics. Algebraically they can be formulated in terms of generators and relations. The parafermion operators $f_j^\pm$, $j=1,2,\ldots, m$, satisfy the following triple relations $[[f_{ j}^{\xi}, f_{ k}^{\eta}], f_{l}^{\epsilon}]=\frac 1 2 (\epsilon -\eta)^2 \delta_{kl} f_{j}^{\xi} -\frac 1 2 (\epsilon -\xi)^2 \delta_{jl}f_{k}^{\eta}, \label{f-rels}$ where $j,k,l\in \{1,2,\ldots,m\}$ and $\eta, \epsilon, \xi \in\{+,-\}$ (to be interpreted as $+1$ and $-1$ in the algebraic expressions $\epsilon -\xi$ and $\epsilon -\eta$), and can be considered as generating elements of the orthogonal Lie algebra $B_n\equiv so(2m+1)$~\cite{Kamefuchi,Ryan}. The paraboson operators $b_j^\pm$, $j=1,2,\ldots, n$, satisfying $[\{ b_{ j}^{\xi}, b_{ k}^{\eta}\} , b_{l}^{\epsilon}]= (\epsilon -\xi) \delta_{jl} b_{k}^{\eta} + (\epsilon -\eta) \delta_{kl}b_{j}^{\xi}, \label{b-rels} $ are generating elements of the orthosymplectic Lie superalgebra $B(0|n)\equiv osp(1|2n)$~\cite{Ganchev}. The case of $m$ parafermions $f_j^\pm\equiv c_j^\pm$~(\ref{f-rels}) and $n$ parabosons $b_j^\pm \equiv c_{m+j}^\pm$~(\ref{b-rels}) with the so called relative parafermion relations~\cite{Greenberg} \begin{eqnarray} && [\![\lb c_{ j}^{\xi}, c_{ k}^{\eta}]\!] , c_{l}^{\epsilon}]\!] =-2 \delta_{jl}\delta_{\epsilon, -\xi}\epsilon^{\langle l \rangle} (-1)^{\langle k \rangle \langle l \rangle } c_{k}^{\eta} +2 \epsilon^{\langle l \rangle } \delta_{kl}\delta_{\epsilon, -\eta} c_{j}^{\xi}, \label{para} \end{eqnarray} where \begin{equation} [\![ a, b ]\!] = ab-(-1)^{{\mathop{\rm deg}\nolimits(a) \mathop{\rm deg}\nolimits(b)}}ba \;\; \rm{and } \;\; \mathop{\rm deg}\nolimits(c_i^\pm)\equiv \langle i\rangle= \left\{ \begin{array}{lll} {0} & \hbox{if} & j=1,\ldots ,m \\ {1} & \hbox{if} & j=m+1,\ldots ,n+m \end{array}\right. \end{equation} lead to the result that they generate the orthosymplectic Lie superalgebra $B(m|n)\equiv osp(2m+1|2n)$~\cite{Palev2}. As a consequence the parastatistics Fock spaces denoted by $V(p)$ are unitary infinite-dimensional $osp(2m+1|2n)$ representations with lowest weight $[-\frac{p}{2},\ldots, -\frac{p}{2}|\frac{p}{2},\ldots, \frac{p}{2}]$, $p$ being a positive integer. An explicit and elegant construction of the latter were given in~\cite{SJ}. The case with $n=0$ corresponds to the parafermion Fock spaces~\cite{parafermion} and with $m=0$ to the paraboson Fock spaces~\cite{paraboson}. The conclusion for us is that the physical properties of such generalized quantum systems can be investigated applying the representation theory of Lie (super)algebras. In the present article we will apply the results for one pair of parafermion operators $f^\pm\equiv c_1^\pm$ and one pair of paraboson operators $b^\pm \equiv c_{2}^\pm$ generating the Lie superalgebra $osp(3|2)$. Using their parastatistics Fock spaces we will consider and investigate a 3D harmonic oscillator as a WQS. \section{The Lie superalgebras $osp(3|2)$ and corresponding to it parastatistics Fock spaces $V(p)$} The Lie superalgebra $osp(3|2)$~\cite{Kac} consists of matrices of the form \begin{equation} \left(\begin{array}{ccccc} a&0&b&x&u \\ 0&-a&c&y&v\\ -c&-b&0&z&w\\ v&u&w&d&e\\ -y&-x&-z&f&-d \end{array}\right), \label{osp} \end{equation} where the nonzero entries are arbitrary complex numbers. The even subalgebra $so(3) \oplus sp(2)$ consists of all matrices~(\ref{osp}), for which $x=y=z=u=v=w=0$, whereas the odd subspace is obtained taking $a=b=c=d=e=f=0$. Let $e_{ij}$ be a 5-by-5 matrix with 1 on the cross of the $i^{th}$ row and the $j^{th}$ column and zero elsewhere. The Cartan subalgebra $H$ of $osp(3|2)$ is the subspace of diagonal matrices with basis $h_1=e_{11}-e_{22}, \;$ $h_{2}=e_{44}-e_{55}.$ It is easy to verify that the matrices \begin{eqnarray} c_{1}^+=f^+= \sqrt{2}(e_{13}-e_{32}), \quad c_{1}^-=f^-= \sqrt{2}(e_{31}-e_{23}), \\ c_{2}^+=b^+= \sqrt{2}(e_{35}+e_{43}), \quad c_{2}^-=b^-= \sqrt{2}(e_{34}-e_{53}), \label{pb-as-e} \end{eqnarray} satisfy the triple relations~(\ref{para}). Moreover, the following holds \begin{theorem}~\cite{Palev2} As a Lie superalgebra defined by generators and relations, $osp(3|2)$ is generated by elements $c_j^\pm, \; j=1,2$ subject to the parastatistics relations~(\ref{para}). \end{theorem} It is straightforward to see that $ [c_1^-,c_1^+]=-2 h_1, \; {\rm and} \;\; \{c_{2}^-,c_{2}^+]=2 h_{2}. $ The parastatistics Fock space $V(p)$ is defined by: \begin{eqnarray} \langle 0|0\rangle=1, \qquad c_j^- |0\rangle = 0, \qquad (c_j^\pm)^\dagger = c_j^\mp,\qquad [\![ c_j^-,c_k^+ ]\!] |0\rangle = p\delta_{jk}\, |0\rangle , \;\; p=1,2,\ldots \nonumber \end{eqnarray} and therefore it is the unitary irreducible representation of $osp(3|2)$ with lowest weight $(-\frac{p}{2},|\frac{p}{2})$. \begin{theorem}~\cite{SJ} An orthonormal basis for the parastatistics Fock space $V(p)$ of one pair of parafermions and one pair of parabosons is given by the vectors \begin{equation} |\mu) = \left| \begin{array}{l} \mu_{12}, \mu_{22} \\ \mu_{11} \end{array} \right),\; {\rm for} \;\mu_{22}=0 \; \mu_{12}=0,1\ldots,p; {\rm for}\; \mu_{22}=1,2,\ldots \; \mu_{12}=1,2,\ldots,p; \mu_{12}-\mu_{11}=\theta=0,1 ({\rm if}\; \mu_{12}=0, \theta=0). \label{cond} \end{equation} The action of the Cartan algebra elements of $osp(3|2)$ is: \begin{equation} h_{1}|\mu)=\left(-\frac{p}{2}+\mu_{11}\right)|\mu), \quad h_{2}|\mu)=\left(\frac{p}{2}+\mu_{12}+\mu_{22}-\mu_{11}\right)|\mu). \label{h_k} \\ \end{equation} For the action of the parastatistics operators $c_j^\pm, \; j=1,2$ we have: \begin{eqnarray} c^+_1 \left| \begin{array}{l} \mu_{12}, \mu_{22} \\ \mu_{11} \end{array} \right) &=& \left(\frac{\mu_{12}+\mu_{22}}{\mu_{12}+\mu_{22}+1}\right)^{\frac{\theta}{2}}{G}_1(\mu_{12},\mu_{22}) \left| \begin{array}{l} \mu_{12}+1, \mu_{22} \\ \mu_{11}+1 \end{array} \right) \nonumber\\ &-&\theta\left(\frac{1}{\mu_{12}+\mu_{22}+1}\right) {G}_2(\mu_{12},\mu_{22}) \left| \begin{array}{l} \mu_{12},\mu_{22}+1 \\ \mu_{11}+1 \end{array} \right),\label{c1+}\\ c^+_2 \left| \begin{array}{l} \mu_{12}, \mu_{22} \\ \mu_{11} \end{array} \right) &=& (1-\theta)\left(\frac{1}{\mu_{12}+\mu_{22}+1}\right)^{\frac{1}{2}}{G}_1(\mu_{12},\mu_{22}) \left| \begin{array}{l} \mu_{12}+1, \mu_{22} \\ \mu_{11} \end{array} \right) \nonumber\\ &+&(-1)^{\theta}\left(\frac{\mu_{12}+\mu_{22}}{\mu_{11}+\mu_{22}+1}\right)^{\frac{1}{2}} {G}_2(\mu_{12},\mu_{22}) \left| \begin{array}{l} \mu_{12},\mu_{22}+1 \\ \mu_{11} \end{array} \right),\label{c2+} \end{eqnarray} \begin{eqnarray} c^-_1 \left| \begin{array}{l} \mu_{12}, \mu_{22} \\ \mu_{11} \end{array} \right) &=& \left(\frac{\mu_{12}+\mu_{22}-1}{\mu_{12}+\mu_{22}}\right)^{\frac{\theta}{2}}{G}_1(\mu_{12}-1,\mu_{22}) \left| \begin{array}{l} \mu_{12}-1, \mu_{22} \\ \mu_{11}-1 \end{array} \right) \nonumber\\ &-&(1-\theta)\left(\frac{1}{\mu_{12}+\mu_{22}}\right) {G}_2(\mu_{12},\mu_{22}-1) \left| \begin{array}{l} \mu_{12},\mu_{22}-1 \\ \mu_{11}-1 \end{array} \right),\label{c1-}\\ c^-_2 \left| \begin{array}{l} \mu_{12}, \mu_{22} \\ \mu_{11} \end{array} \right) &=& \theta\left(\frac{1}{\mu_{12}+\mu_{22}}\right)^{\frac{1}{2}}{G}_1(\mu_{12}-1,\mu_{22}) \left| \begin{array}{l} \mu_{12}-1, \mu_{22} \\ \mu_{11} \end{array} \right) \nonumber\\ &+&(-1)^{\theta}\left(\frac{\mu_{12}+\mu_{22}-1}{\mu_{11}+\mu_{22}}\right)^{\frac{1}{2}} {G}_2(\mu_{12},\mu_{22}-1) \left| \begin{array}{l} \mu_{12},\mu_{22}-1 \\ \mu_{11} \end{array} \right),\label{c2-} \end{eqnarray} \begin{equation} G_1(\mu_{12},\mu_{22})=\sqrt{ \frac{\mu_{12}(\mu_{12}+\mu_{22}+1)(p-\mu_{12})}{\mu_{12}+\mu_{22}+1-{\cal O}_{\mu_{22}+1}}}, G_2(\mu_{12},\mu_{22})=\sqrt{ \frac{({\cal O}_{\mu_{22}}\mu_{22}+1)({\cal E}_{\mu_{22}+1}(p+\mu_{22})+1)({\cal O}_{\mu_{22}+1}(\mu_{12}+\mu_{22})+1)} {{\cal E}_{\mu_{22}+1}(\mu_{12}+\mu_{22}-1)+1}}, \end{equation} \begin{equation} {\cal E}_{j}=1 \hbox{ if } j \hbox{ is even and 0 otherwise},\quad {\cal O}_{j}=1 \hbox{ if } j \hbox{ is odd and 0 otherwise}. \label{EO} \end{equation} \end{theorem} \section{Wigner quantum oscillator and $osp(3|2)$} Consider a three-dimensional harmonic oscillator, namely a quantum system with a Hamiltonian $ H={{\bf p}^2\over 2m}+ {m \omega^2 \over 2}{\bf r}^2 $ as a Wigner quantum system. This means that we must find the unknown operators {\bf r}=$(r_1, r_2, r_3)$ and {\bf p}=$(p_1, p_2, p_3)$ so that the following conditions (i) The state space of the oscillator $W$ is a Hilbert space. The physical observables are Hermitian operators in $W$. (ii) Hamilton's equations $\dot{\bf p}=-m \omega^2 {\bf r}, \dot{\bf r}={{\bf p}\over m} $ and the Heisenberg equations $\dot {\bf p}=-{i\over \hbar}[{\bf p},H], \dot {\bf r}=-{i\over \hbar}[{\bf r},H], $ are identical (as operator equations) in $W$. (iii) The projections of the angular momentum {\bf M}=$(M_1, M_2, M_3)$ are in the enveloping algebra of the position {\bf r}=$(r_1, r_2, r_3)$ and momentum {\bf p}=$(p_1, p_2, p_3)$ operators. Each $M_k$ is linear in $(r_1, r_2, r_3)$ and $(p_1, p_2, p_3)$, so that {\bf M}, {\bf r} and {\bf p} transform as vectors: $[M_j,c_k]=i \sum_{l=1}^3 \varepsilon_{jkl}c_l, c_k=M_k,r_k,p_k,\; j,k=1,2,3 $ hold. Introduce new unknown operators $a_k^\pm=\sqrt{m \omega \over 2 \hbar} r_k \mp {i \over \sqrt {2m \omega \hbar}}p_k, k=1,2,3. $ In terms of $a_k^\pm$ the Hamiltonian reads $H={1 \over 2}\omega \hbar \sum_{k=1}^3 \{a_k^+, a_k^- \}$ and the condition (ii) yields ($k=1,2,3$): \begin{equation} \sum_{i=1}^3 [ \{a_i^+,a_i^- \},a_k^\pm]=\pm 2a_k^\pm. \label{CCs} \end{equation} Let \begin{equation} a_1^\pm={1\over {2 \sqrt{3}}}[c_1^- -c_1^+, c_2^\pm] \quad a_2^\pm={i\over {2 \sqrt{3}}}[c_1^- +c_1^+, c_2^\pm], \quad a_3^\pm={1\over {\sqrt{3}}}c_2^\pm. \label{CAO} \end{equation} It is straightforward to check that the operators~(\ref{CAO}) satisfy the relations~(\ref{CCs}). Therefore condition (ii) holds. The projections $(M_1, M_2, M_3)$ of the angular momentum can be defined as $M_i=-{3\over 4\hbar}\sum_{j,k=1}^3 \varepsilon_{ijk} \{r_j,p_k\}, i=1,2,3 $ or equivalently $M_1={1\over 2}(c_1^+ +c_1^-), M_2=-{i\over 2}(c_1^+ -c_1^-), M_3={1\over 2}[c_1^-,c_1^+] $ and it is easy to check that condition (iii) is fulfilled. Let $W$ be the $osp(3|2)$ module $V(p)$. In each such module $(c_i^-)^\dagger = c_i^+, i=1,2$. Therefore the position and the momentum operators are Hermitian operators (hence also the Hamiltonian $H$, the square of the angular momentum ${\bf M}^2$ and its projections $M_1$, $M_2$, $M_3$ are Hermitian operators) and also condition (i) holds. From~(\ref{CAO}) and Theorem 2 one immediately derives that \begin{equation} H|\mu) = \frac{\hbar\omega}{2}\{c_2^+,c_2^-\}|\mu)=\frac{\hbar\omega}{2}(p+2\mu_{22}+2\theta)|\mu) \end{equation} Thus, the energy spectrum $E_n, n=0,1,2,\ldots$ of the Wigner quantum oscillator is given by \begin{equation} E_n={\hbar}\omega(n+\frac{p}{2}), \; n=0,1,2,\ldots . \end{equation} The case $p=1$ corresponds to anticommuting pairs of Bose and Fermi operators and the energy spectrum is the same as for the one-dimensional canonical harmonic oscillator. The different coordinates (resp. the different momenta) of the oscillator anticommute \begin{equation} \{ r_i, r_j \}=\{ p_i, p_j \}=0, \;\; i\neq j=1,2,3. \end{equation} Consequently the Wigner quantum oscillator has neither coordinate nor momentum representation. The coordinates, the momenta and the angular momenta operators are on the same footing: the different components do not commute with each other: \begin{equation} [ r_i, r_j ]\neq 0, \; [ p_i, p_j] \neq 0, \;[ M_i, M_j] \neq 0, \; i\neq j=1,2,3. \end{equation} The geometry of the oscillator is noncommutative. Each state $|\mu)$ is an eigenvector of $M_3$: \begin{equation} M_3|\mu)=(\frac{p}{2}-2\mu_{12})|\mu), \quad \mu_{12}\leq p. \end{equation} Consequently the Wigner oscillator has an angular momentum $M=\frac{p}{2}, p=1,2,\ldots.$ \section{ACKNOWLEDGMENTS} The author was supported by Bulgarian NSF grant DFNI T02/6. \nocite{*}

https://arxiv.org/abs/2008.04602

Central limit theorem of Brownian motions in pinched negative curvature

We prove the central limit theorem of random variables induced by distances to Brownian paths and Green functions on the universal cover of Riemannian manifolds of finite volume with pinched negative curvature. We further provide some ergodic properties of Brownian motions and an application of the central limit theorem to the dynamics of geodesic flows in pinched negative curvature.

\section{Introduction} Let $\widetilde{\mathcal{M}}$ be a simply connected complete Riemannian manifold of dimension $d\ge2$ with pinched negative curvature; its sectional curvature is uniformly bounded between two negatives. We further assume that $\widetilde{\mathcal{M}}$ admits a finite-volume quotient $\M$ and the first derivative of the sectional curvature is uniformly bounded. The Brownian motion $(\widetilde\omega_{t})_{t\in\mathbb{R}_{+}}$ on $\widetilde{\mathcal{M}}$ starting from $x$ is transient as $\widetilde{\mathcal{M}}$ is negatively curved. Therefore, the distance $\mathrm{d}(x, \widetilde\omega_{t})$ goes to infinity as $t \to \infty$ with probability 1 and its asymptotic growth is linear (\cite{Guivarch_1981}): there is $\ld>0$ such that \begin{equation*} \ld = \lim_{t\to\infty} \frac{1}{t}\mathrm{d}_{}(x, \widetilde\omega_{t}). \end{equation*} Due to the pinched negative curvature, the Green function $\G(x,y)$ on $\widetilde{\mathcal{M}}$ tends to zero as $\mathrm{d}(x,y) \to \infty$. Hence $\G(x, \widetilde\omega_{t}) \to 0$ as $t\to\infty$ and it decays exponentially fast with probability 1 (\cite{Kai_1986}): there exists $\hs>0$ such that \begin{equation*} \hs = \lim_{t\to\infty} -\frac{1}{t}\log\G(x, \widetilde\omega_{t}). \end{equation*} Even though Brownian motions on manifolds with pinched negative curvature has been studied for a long time, the majority of the results holds for either every Cartan-Hadamard manifolds or co-compact ones and few are known for the cases in between, especially for the co-finite manifolds $\widetilde{\mathcal{M}}$. Our main result, the central limit theorem of random processes $Y^{\ld}_{t}(\widetilde\omega)= \mathrm{d}_{}(x,\widetilde\omega_{t})-t\ld$ and $Y^{\hs}_{t}(\widetilde\omega)= \log \G(x, \widetilde\omega_{t})+t\hs$, is a generalization of the central limit theorem in co-compact manifolds proved by F. Ledrappier in \cite{Led_1995}. \begin{thm} \label{CLT} The distributions of $\frac{1}{\sigma_{\bb}\sqrt{t}}Y^{\ld}_{t}$ and $\frac{1}{\sigma_{\kk}\sqrt{t}}Y^{\hs}_{t}$ are asymptotically normal for some positive constants $\sigma_{\bb}, \sigma_{\kk}$. More precisely, for every $x \in \widetilde{\mathcal{M}}$, \begin{align*} \mathbb{P}_{x} \left[ \frac{Y^{\ld}_{t}}{\sigma_{\bb}\sqrt{t}} \le r \right] , \mathbb{P}_{x} \left[ \frac{Y^{\hs}_{t}}{\sigma_{\kk}\sqrt{t}} \le r \right] \to \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{r} \exp\left( - \frac{s^{2}}{2} \right) ds, \textrm{ as } t\to\infty, \end{align*} where $\mathbb{P}_{x}$ is the probability measures on the space $\C(\mathbb{R}_{+}, \widetilde{\mathcal{M}})$ of continuous sample paths which defines the Brownian motion on $\widetilde{\mathcal{M}}$ starting from $x$. \end{thm} F. Ledrappier introduced a double process to provide a lower bound for the expectation of the Gromov product at Brownian points in \cite{Led_1995}. The lower bound implies the contraction property of the foliated Brownian motion, which plays an important role in the proof of the central limit theorem. However, since the double process argument is not valid in the absence of compactness, we instead provide an argument using the $\C^{2}$-convergence of the normalized distance functions to the Busemann function in pinched negatively curved manifolds. Although the resulting lower bound is less sharp than the lower bound by the double process argument, it is sufficient for the proof of the contraction property. As in \cite{Led_1995}, we use the contraction property of the foliated Brownian motion (Theorem \ref{Ctr}) on H\"older spaces to solve the leafwise heat equation on the unit tangent bundle for the foliated Laplacian. We construct Martingales from the solutions of the heat equation with the initial conditions of the Busemann function and the logarithm of the Martin kernel of the Brownian motion. We prove that they are asymptotically normal and have the same distributions with the random variables of our interest. As a consequence of the central limit theorem, we provide a characterization for the asymptotic harmonicity of $\widetilde{\mathcal{M}}$ with an assumption for thermodynamic formalism. We say that $\widetilde{\mathcal{M}}$ is \textit{asymptotically harmonic} if the mean curvature of the horospheres of $\widetilde{\mathcal{M}}$ is constant. If $\widetilde{\mathcal{M}}$ is asymptotically harmonic then the Liouville measure on the unit tangent bundle of $\M$ has maximal entropy for the geodesic flow. The characterization reveals an interplay between the stochastic properties, the geometry and the dynamics of the geodesic flow of $\widetilde{\mathcal{M}}$. Indeed, an asymptotically harmonic manifold $\widetilde{\mathcal{M}}$ is a symmetric space if it is the universal cover of a compact negatively curved manifold (\cite{FouLab_1992}, \cite{BFL_1992}, \cite{Led_1990}). The Martin kernel of the Brownian motion gives rise to a H\"older continuous function $\K$ on $\T^{1}\M$, which helps us understand the asymptotic behavior of Brownian paths and correlation with geodesics. An equilibrium state of $\K$ is a geodesic flow-invariant Borel probability measure on $\T^{1}\M$ which maximizes the pressure of $\K$. For compact manifolds, every H\"older continuous function admits a unique equilibrium states (\cite{Franco_1977}) while the existence is not always guaranteed for finite-volume manifolds. \begin{thm}\label{AsympH} If $\K$ admits an equilibrium state, then \begin{equation*} \sigma_{\kk}^{2}\ge 2\hs. \end{equation*} The equality holds if and only if $\widetilde{\mathcal{M}}$ is asymptotically harmonic. \end{thm} In Section 2, we introduce the heat kernel and the Brownian motion on $\widetilde{\mathcal{M}}$. We also recall preliminaries of the geometry of manifolds with pinched negative curvature, the ergodic theory and thermodynamic formalisms for their geodesic flow. We prove Theorem \ref{CLT} in Section 3 while Section 4 is devoted to the proof of the contraction property (Theorem \ref{Ctr}). Section 4 also contains a diagonal estimate of the heat kernel and the proof of exponential ergodicity of the Brownian motion on $\M$. In Section 5, we prove ergodic properties of the Brownian motions which generalize the results in \cite{Led_1988}. We conclude the section with the proof of Theorem \ref{AsympH}. \\ \textit{Acknowledgement}. It is a pleasure to thank Fran\c{c}ois Ledrappier and Seonhee Lim for sharing their insights and helpful comments. The work is supported by Samsung Science and Technology Foundation under Project Number SSTF-BA1601-03. \section{Preliminaries} \label{Pre} Let $(\M, g)$ be a complete finite-volume Riemannian manifold of dimension $d\ge2$. We say that $\M$ has \textit{pinched negative curvature} if \begin{equation*} -b^{2}\le \mathrm{sec}_{\M} \le -a^2 \end{equation*} for some positive numbers $b>a>0$. We assume that $\M$ has pinched negative curvature and $|\nabla \sec_{\M}| \le c$ for some $c>0$. Let $\widetilde{\mathcal{M}}\to\M$ be the universal cover with the group of deck transformation $\Gamma$ acting isometrically on $\widetilde{\mathcal{M}}$. We also denote the lift of the metric on $\M$ to $\widetilde{\mathcal{M}}$ by $g$. Let $\mathrm{d}_{}$ be the Riemannian distance of $\widetilde{\mathcal{M}}$ and $\mathrm{vol}:=\mathrm{vol}_{\widetilde{\mathcal{M}}}$ the Riemannian volume on $\widetilde{\mathcal{M}}$. A number of examples can be constructed from noncompact finite-volume hyperbolic manifolds by perturbing the metric near cusps. See \cite{DPPS_2009}, \cite{DPPS_2017} for the detail. \subsection{Geometry of pinched negative curvature} Since $\widetilde{\mathcal{M}}$ has pinched negative curvature, the metric space $(\widetilde{\mathcal{M}}, \mathrm{d}_{})$ is a CAT(0)-space. Hence we consider its boundary at infinity $\partial\widetilde{\mathcal{M}}$, also called the \textit{visual boundary}. Fix $x\in\widetilde{\mathcal{M}}$. A sequence $(z_{n})$ in $\widetilde{\mathcal{M}}$ converges to a point $\xi$ in $\partial\widetilde{\mathcal{M}}$ if and only if $z_{n}\to\infty$ and the sequence of normalized distance functions \begin{equation*} f_{n}(y)=\bb(y, x, z_{n}):=\mathrm{d}_{}(y, z_{n})-\mathrm{d}_{}(x, z_{n}) \end{equation*} converges uniformly on compact sets in $\C(\widetilde{\mathcal{M}})$. We denote the limit function by $\bb(y, x, \xi)$, which we call the \textit{Busemann function} based at $\xi$. The convergence of $z_{n}$ to $\xi$ is independent of the choice of $x$. An important remark is that $f_{n}$ converges to the Busemann function $\C^{2}$-uniformly on compact sets: \begin{prop} \label{CCB} (\cite{Bal_1995}) Let $f_{n}(y)= \mathrm{d}_{}(y, z_{n})-\mathrm{d}_{}(x, z_{n})$ and $z_{n}\to \xi \in \partial\widetilde{\mathcal{M}}$. Then \begin{align*} \nabla f_{n} &\to \nabla \bb(\cdot, x, \xi), \\ \nabla_{\mathrm{v}}\nabla f_{n} &\to \nabla_{\mathrm{v}} \nabla \bb(\cdot, x, \xi), \end{align*} uniformly on compact sets. $\nabla \bb(\cdot, x, \xi)$ means the covariance derivative of $y\mapsto\bb(y,x,\xi)$. \end{prop} Let $\Delta= \mathrm{div} \nabla$ be the Laplace-Beltrami operator on $(\widetilde{\mathcal{M}},g)$. If $\{e_{1}, \dots, e_{d}\}$ is an orthonormal frame on an open set $U$, for each $\C^{2}$-function $f$ on $U$, \begin{equation}\label{Localex} \Delta f = \sum_{j=1}^{d}\langle e_{j}, \nabla_{e_{j}}\nabla f\rangle_{g} \end{equation} on $U$. Applying Proposition \ref{CCB} to each summand of (\ref{Localex}), we obtain the following result. \begin{prop} \label{C2conv} Let $f_{n}(y)= \mathrm{d}_{}(y, z_{n})-\mathrm{d}_{}(x, z_{n})$ and $z_{n}\to \xi \in \partial\widetilde{\mathcal{M}}$. Then $\Delta f_{n}$ converges to $\Delta \bb(\cdot, x, \xi)$ uniformly on compact sets. \end{prop} The visual boundary $\partial\widetilde{\mathcal{M}}$ is equipped with a distance. For $\xi, \eta \in\partial \widetilde{\mathcal{M}}$ with $z_{n}, w_{n}\in\widetilde{\mathcal{M}}$ which converge to $\xi, \eta$ respectively, we define the Gromov product of $\xi$ and $\eta$ at $x\in \widetilde{\mathcal{M}}$ by \begin{equation*} (\xi|\eta)_{x}:=\lim_{n\to\infty} \mathrm{d}(x, z_{n}) + \mathrm{d}(x, w_{n}) -\mathrm{d}(z_{n}, w_{n}). \end{equation*} Then for $\tau>0$ small enough, $d_{\infty}^{x,\tau}(\xi, \eta):=\exp[-\tau(\xi|\eta)_{x}]$ is a distance function on the visual boundary $\partial\widetilde{\mathcal{M}}$ (see \cite{BriHae}). Let $\pi:\T\widetilde{\mathcal{M}}\to\widetilde{\mathcal{M}}$ be the tangent bundle of $\widetilde{\mathcal{M}}$. We endow $\T\widetilde{\mathcal{M}}$ with a Riemannian metric $g_{\T}$ called the \textit{Sasaki metric}, induced by the Riemannian structure $g$ of $\widetilde{\mathcal{M}}$ and its Levi-Civita connection $\nabla$. We consider the unit tangent bundle $\T^{1}\widetilde{\mathcal{M}}=\{ \mathrm{v}\in \T\widetilde{\mathcal{M}}: \|\mathrm{v}\|^{2}=\langle \mathrm{v}, \mathrm{v}\rangle_{g}=1\}$ of $\widetilde{\mathcal{M}}$, which is a submanifold of $\T\widetilde{\mathcal{M}}$ and also a sphere bundle of $\widetilde{\mathcal{M}}$. We denote the geodesic flow on $\T^{1}\widetilde{\mathcal{M}}$ by $\g^{t}:\T^{1}\widetilde{\mathcal{M}} \to \T^{1}\widetilde{\mathcal{M}}$. We also denote by $\g^{t}$ the geodesic flow on the unit tangent bundle $\T^{1}\M$ of $\M$. We introduce the stable foliation $\widetilde\W^{s}$ and the strong unstable foliation $\widetilde\W^{su}$ of $\T^{1}\widetilde{\mathcal{M}}$ which will play an important role in the following sections. Their leaves are defined by \begin{align*} \widetilde\W^{s}(\mathrm{v}) &= \left\{ \mathrm{w}\in\T^{1}\widetilde{\mathcal{M}}: \lim_{t \to \infty} \mathrm{d}_{}(\gamma_{\mathrm{v}}(t+s), \gamma_{\mathrm{w}}(t)) =0, \, \exists s \right\}, \\ \widetilde\W^{su}(\mathrm{v}) &= \left\{ \mathrm{w}\in\T^{1}\widetilde{\mathcal{M}}: \lim_{t\to\infty} \mathrm{d}_{}(\gamma_{\mathrm{v}}(t), \gamma_{\mathrm{w}}(t)) =0 \right\}, \end{align*} Where $\gamma_{\mathrm{v}}$ is the geodesic generated by $\mathrm{v}$. Note that $\widetilde\W^{su}$ consists of unit normal bundles of level sets of Busemann functions and leaves are transversal to the stable foliation with angle uniformly bounded away from zero (Lemma 7.4. in \cite{PPS}). The \textit{stable distribution} $\widetilde{E}^{s}$ of $\T^{1}\widetilde{\mathcal{M}}$ is a rank $d$-subbudle of the tangent bundle $\T\T^{1}\widetilde{\mathcal{M}} \to \T^{1}\widetilde{\mathcal{M}}$ of $\T^{1}\widetilde{\mathcal{M}}$ whose fibers are tangent spaces of stable leaves: $\widetilde{E}^{s}_{\mathrm{v}}:= \T_{\mathrm{v}}\widetilde\W^{s}(\mathrm{v})$. Since $\widetilde\W^{s}(\mathrm{v})$ is diffeomorphic to $\widetilde{\mathcal{M}}$ via $\pi:\T^{1}\widetilde{\mathcal{M}}\to\widetilde{\mathcal{M}}$ for $\mathrm{v} \in \T^{1}\widetilde{\mathcal{M}}$, we endow stable leaves of $\widetilde\W^{s}$ with a metric $g_{s}$ induced from the metric $g$ on $\widetilde{\mathcal{M}}$: for $\mathrm{v}\in\T^{1}_{x}\widetilde{\mathcal{M}}$, define $g_{s}$ on $\widetilde{E}^{s}_{\mathrm{v}}=\T_{\mathrm{v}}\widetilde\W^{s}(\mathrm{v})$ from $g$ on $\T_{x}\widetilde{\mathcal{M}}$. For each point $x\in\widetilde{\mathcal{M}}$ and point at infinity $\xi\in\partial\widetilde{\mathcal{M}}$, there is a unique unit vector $\mathrm{v}$ in $\T^{1}\widetilde{\mathcal{M}}$ such that $\gamma_{\mathrm{v}}(t)$ converges to $\xi$ at $t\to \infty$. Conversely, for every geodesic $\gamma$, $\gamma(t)$ converges to a point $\xi$ in $\partial\widetilde{\mathcal{M}}$. We denote the limit point $\xi$ of $\gamma_{\mathrm{v}}(t)$ by $\mathrm{v}_{+}$. This gives a useful identification of $\T^{1}\widetilde{\mathcal{M}}$ with $\widetilde{\mathcal{M}}\times\partial\widetilde{\mathcal{M}}$. With such identification, we have that for $\mathrm{v}=(x,\xi)$, $\widetilde\W^{s}(\mathrm{v})=\widetilde{\mathcal{M}}\times\{\xi\}$. Moreover, $\nabla_{y} \bb(y, x, \xi)= (y, \xi)$. Let $X: \T^{1}\widetilde{\mathcal{M}} \to \widetilde{E}^{s}$ be a section of the stable distribution which is leafwise $\C^{1}$, i.e., the restriction $X|_{\widetilde\W^{s}(x, \xi)}$ is $\C^{1}$ on $\widetilde\W^{s}(x, \xi)$ for each $(x, \xi)\in \T^{1}\widetilde{\mathcal{M}}$. We identify $X|_{\widetilde\W^{s}(x, \xi)}$ with a $\C^{1}$-vector field $X^{\xi}$ on $\widetilde{\mathcal{M}}$ for each $\xi$. We define the $g_{s}$-divergence $\mathrm{div}_{s}$ by \begin{equation*} \mathrm{div}_{s}X(x, \xi)= \mathrm{div} X^{\xi}(x). \end{equation*} Let $u \in \C(\T^{1}\widetilde{\mathcal{M}})$ be a leafwise $\C^{2}$-function; $u|_{\widetilde\W^{s}(\mathrm{v})}$ is $\C^{2}$ on $\widetilde\W^{s}(\mathrm{v})$. Thus for each $\xi\in\partial\widetilde{\mathcal{M}}$, $u^{\xi}(x):= u(x, \xi)$ is $\C^{2}$ on $\widetilde{\mathcal{M}}$. We define the \textit{foliated Laplacian} $\Delta_{s}$ by \begin{equation*} \Delta_{s} u= \mathrm{div}_{s} \nabla u, \end{equation*} where $\nabla u (x, \xi) := \nabla u^{\xi} (x)$. \subsection{Brownian motions} The heat kernel $\wp:(0,\infty)\times\widetilde{\mathcal{M}}\times\widetilde{\mathcal{M}}\to(0,\infty)$ is the fundamental solution of the heat equation: \begin{align*} \partial_{t}\wp(t,x,y) &= \Delta_{y}\wp(t, x,y),\\ \lim_{t \downarrow 0} \wp(t, x, y) &= \delta_{x}(y). \end{align*} The limit in the last equation means that for each $f\in\C_{b}(\widetilde{\mathcal{M}})$, \begin{equation*} \lim_{t\downarrow0} \int_{\widetilde{\mathcal{M}}} \wp(t, x, y)f(y) d\mathrm{vol}_{\widetilde{\mathcal{M}}}(y) =f(x). \end{equation*} Since the curvature of $\widetilde{\mathcal{M}}$ is negatively pinched, $\Delta$ is (weakly) coercive, i.e., the Green function of $\Delta$ $$\G(x,y):=\int_{0}^{\infty}\wp(t, x, y)dt$$ is finite for $x \ne y\in \widetilde{\mathcal{M}}$. For $\kappa<0$, if $\wp_{\mathbb{H}^{d}(\kappa)}(t, x,y)$ is the heat kernel on the $d$-dimensional hyperbolic space $\mathbb{H}^{d}(\kappa)$ of constant curvature $\kappa$, $\wp_{\mathbb{H}^d(\kappa)} (t, x, y)$ depends only on $t$ and $\mathrm{d}_{\mathbb{H}^{d}(\kappa)}(x,y)$. The following comparison theorem of the heat kernel is also due to the pinched negative curvature. \begin{prop} \label{HCom} (Heat kernel comparison theorem, \cite{Hsu}) \begin{align*} \wp_{\mathbb{H}^{d}(-b^{2})}(t, \mathrm{d}(x,y)) \le \wp(t,x,y) \le \wp_{\mathbb{H}^{d}(-a^{2})}(t, \mathrm{d}(x,y)). \end{align*} \end{prop} Note that $\wp(t,x,y)$ determines a unique family of probability measures on the space $\widetilde\Omega= \C(\mathbb{R}_{+}, \widetilde{\mathcal{M}})$ of sample paths. For each $x\in\widetilde{\mathcal{M}}$, we define the probability measure $\mathbb{P}_{x}$ on the cylinder sets in $\widetilde\Omega$ by \begin{align*} &\mathbb{P}_{x} \left[ \widetilde\omega_{t_{i}}\in A_{i}, t_{1}<\cdots<t_{k} \right] = \\& \int_{A_{k}} \cdots \int_{A_{1}} \wp(t_{1}, x, y_{1}) \wp(t_{2}-t_{1}, y_{1}, y_{2}) \times \cdots \times \wp({t_{k}-t_{k-1}}, y_{k-1}, y_{k}) d \mathrm{vol}(y_{1}) \cdots d \mathrm{vol}(y_{k}). \end{align*} By Kolmogorov extension theorem, $\mathbb{P}_{x}$ extends to a unique probability measure on $\widetilde\Omega$. For $s\ge0$, we denote the projection map $\widetilde\omega \mapsto \widetilde\omega_{s}$ by $\pi_{s}:\widetilde\Omega\to\widetilde{\mathcal{M}}$. Let $\Fi_{t}=\Fi_{t}(\widetilde{\mathcal{M}}):=\sigma\{\pi_{s}\}_{0\le s \le t}$ be the smallest $\sigma$-algebra for which the projections $\pi_{s}$ are measurable. The canonical process $\widetilde{Z}_{t}(\widetilde\omega):= \widetilde\omega_{t}$ of the filtered space $( \widetilde\Omega, \{ \Fi_{t} \}_{0\le t\le\infty} )$ forms a Markov process with respect to $\mathbb{P}_{x}$, which is called the \textit{Brownian motion} on $\widetilde{\mathcal{M}}$ with initial distribution $\delta_{x}$, for each $x\in\widetilde{\mathcal{M}}$. Let $\Omega=\mathcal{C}(\mathbb{R}_{+},\M)$. For each $x\in\M$ and its lift $\widetilde{x}\in\widetilde{\mathcal{M}}$, we also denote the push-forward measure of $\mathbb{P}_{\widetilde{x}}$ by $\mathbb{P}_{x}$. Then the canonical process $Z_{t}$ of $(\Omega, (\Fi_{t}(\M))_{0\le t\le\infty}, (\mathbb{P}_{x})_{x\in\M})$ is a Markov process, which we call the Brownian motion on $\M$. This process is the projected process of the Brownian motion on $\widetilde{\mathcal{M}}$. The stationary measure of the Brownian motion is the probability measure which defines the Brownian motion with initial distribution $m$: $\mathbb{P}_{m}=\int_{\M} \mathbb{P}_{x}\,d\m(x)$ where $\m$ is the normalized Riemannian volume on $\M$. The shift dynamical system on the path space $(\Omega, \mathscr{S}^{t}, \mathbb{P}_{m})$ is ergodic since $\M$ is connected, where $\mathscr{S}^{t}\omega_s =\omega_{t+s}$ for $\omega\in\Omega$. Let $r(\omega, t)= \mathrm{d}_{}(\widetilde\omega_{0}, \widetilde\omega_{t})$ where $\widetilde\omega$ is a lift of $\omega$. Then since $r$ is a sub-additive cocycle, that is, $r(\omega,{t+s})\le r(\omega,t)+r(\mathscr{S}^{t}\omega, s)$ for every $s,t>0$, there exists a positive constant $\ld$, which is called \textit{the linear drift} of the Brownian motion, such that for every $x\in\widetilde{\mathcal{M}}$ and for a.s. $\omega\in\Omega$ \begin{equation*} \ld =\lim_{t\to \infty}\frac{1}{t}r(\omega, t) =\lim_{t\to\infty}\frac{1}{t}\mathrm{d}_{}(x, \widetilde\omega_{t}) \end{equation*} due to the subadditive ergodic theorem (\cite{Kingman_1968}). If $\M$ has constant negative curvature $-a^2$, then $\ld = (d-1)a$. For a fixed $x\in\widetilde{\mathcal{M}}$, the exponential map at $x$ induces a polar coordinate on $\widetilde{\mathcal{M}}\setminus\{x\}$: \begin{align*} (0,\infty)\times \T^{1}_{x}\widetilde{\mathcal{M}} &\to \widetilde{\mathcal{M}}\setminus\{x\}\\ (r, \mathrm{v}) &\mapsto \exp_{x} r\mathrm{v}. \end{align*} Note that $\T^{1}_{x}\widetilde{\mathcal{M}}$ inherits the Riemannian metric $g_{\mathbb{S}}$ of the unit sphere $\mathbb{S}^{d-1}$ from $(\widetilde{\mathcal{M}}, g)$ and write $g$ as \begin{equation*} g= dr^{2}+ \lambda_{x}(r, \mathrm{v}) g_{\mathbb{S}}, \end{equation*} for some smooth function $\lambda_{x}$ on $\widetilde{\mathcal{M}}\setminus\{x\}=(0,\infty)\times\T^{1}_{x}\widetilde{\mathcal{M}}$. For $\widetilde\omega\in \widetilde\Omega$, we write $r(\widetilde\omega, t)= \mathrm{d}_{}(\widetilde\omega_{0}, \widetilde\omega_{t})$ and let $\theta(\widetilde\omega,{t})$ be the unit vector in $\T^{1}_{\widetilde\omega_{0}}\widetilde{\mathcal{M}}$ with $\exp_{\widetilde\omega_{0}} \left[ r(\widetilde\omega,t)\theta(\widetilde\omega,t) \right] = \widetilde\omega_{t}$. \begin{prop} (\cite{Prat_1975}, \cite{Pin_1978}) For every $x\in\widetilde{\mathcal{M}}$ and $\mathbb{P}_{x}$-a.e. $\widetilde\omega$, the limit $\lim_{t\to\infty} \theta(\widetilde\omega, t)$ exists. \end{prop} Since $r(\widetilde\omega, t)\to \infty$ as $t\to\infty$ for $\mathbb{P}_{x}$-a.e. $\widetilde\omega$, the limit $\widetilde\omega_{\infty} := \lim_{t\to\infty} \widetilde\omega_{t}$ exists for $\mathbb{P}_{x}$-a.e. $\widetilde\omega$. In addition, the Brownian path roughly follows the geodesic $\gamma_{\theta(\widetilde\omega, \infty)}$ (\cite{Led_1988}): \begin{equation}\label{roughpath} \lim_{t\to\infty} \frac{1}{t} \mathrm{d} \left( \widetilde\omega_{t}, \exp_{x} \left[ r(\widetilde\omega, t)\theta(\widetilde\omega, \infty) \right] \right) =0. \end{equation} We can replace $r(\widetilde\omega, t)$ by $\ld t$. We denote the asymptotic distribution of Brownian paths starting from $x$ by $\nu_{x}$, i.e., \begin{equation*} \nu_{x}(U):= \mathbb{P}_{x}\left[ \widetilde\omega: \widetilde\omega_{\infty}\in U \right], \textrm{ for }U\subset \partial\widetilde{\mathcal{M}}. \end{equation*} Since the family $(\mathbb{P}_{x})$ is $\Gamma$-equivariant, $(\nu_{x})_{x\in \widetilde{\mathcal{M}}}$ is also $\Gamma$-equivariant: $\gamma_{*}\nu_{x} =\nu_{\gamma x}$ for each $\gamma\in\Gamma$. Moreover, $(\nu_{x})_{x\in\widetilde{\mathcal{M}}}$ is absolutely continuous and we denote the Radon-Nikodym derivative, called the \textit{Martin kernel}, by \begin{equation*} \kk(x, y, \xi):=\frac{d\nu_{y}}{d\nu_{x}}(\xi). \end{equation*} The Martin kernel is also characterized by the limiting behavior of the Green function. \begin{prop} (\cite{AndSch_1985}) For each sequence $(z_{n})$ in $\widetilde{\mathcal{M}}$ with $z_{n} \to \xi \in \partial\widetilde{\mathcal{M}}$, \begin{equation*} \kk(x, y, \xi) = \lim_{n\to\infty} \frac{\G(y, z_{n})}{\G(x, z_{n})}. \end{equation*} \end{prop} We introduce another invariant of the Brownian motion called the \textit{stochastic entropy} of the Brownian motion denoted by $\hs$. The stochastic entropy was first introduced by V. Kaimanovich in \cite{Kai_1986} for co-compact manifolds with negative curvature. The stochastic entropy determines whether the Poisson boundary is trivial or not. The argument in \cite{Led} easily extends to manifolds with finite volume. \begin{prop} For each $x\in\widetilde{\mathcal{M}}$, $\mathbb{P}_{x}$-a.e. $\widetilde\omega$, the following limits exist and coincide: \begin{align*} \hs &= \lim_{t\to\infty} -\frac{1}{t}\log \wp(t,x,\widetilde\omega_{t}) \\ &= \lim_{t\to\infty} -\frac{1}{t} \log \G(x,\widetilde\omega_{t}) . \end{align*} \end{prop} Note that $\hs = (d-1)^{2}a^{2}$ when $\textrm{sec}_{\M}=-a^{2}$. There is another characterization of the stochastic entropy analogous to the definition of the topological entropy as the exponential growth of dynamically separated sets (see \cite{Kai_1986}, \cite{Led}). \begin{prop} For $x\in\widetilde{\mathcal{M}}$, $T>0$ and $0<\delta<1$, \begin{equation*} \hs=\lim_{T\to\infty}\frac{1}{T}\log N(x, T, \delta), \end{equation*} where $N(x, T, \delta):=\inf \left\{ \mathrm{Card}(E): \mathbb{P}_{x}[ d(\widetilde\omega _T, E)\le 1] \ge \delta\right\}$. \end{prop} \begin{proof} Fix $\varepsilon>0$. Let \begin{align*} \mathscr{C}_{T,x} &:= \{ \widetilde\omega_{0}=x, \wp(T, \widetilde\omega_{0}, \widetilde\omega_{T}) \le e^{-T(\hs-\varepsilon)} \},\\ \mathscr{D}_{T, x} &:= \{ \widetilde\omega: \mathrm{d}(\widetilde\omega_{t}, \gamma_{\theta(\widetilde\omega, \infty)}(\ld T))\le \varepsilon T, \wp(T, x, \gamma_{\theta(\widetilde\omega, \infty)}(\ld T)) \ge e^{-T(h+\varepsilon)} \}. \end{align*} Choose a sufficiently large $T$ such that $1-\frac{\delta}{2} \le \mathbb{P}_{x}(\mathscr{C}_{T,x})=\mathbb{P}_{x}[\widetilde\omega_{T}\in\pi_{T}\mathscr{C}_{T,x}]$. We denote by $\mathbb{E}_{x}$ the expectation with respect to $\mathbb{P}_{x}$. For each finite set $E$ such that $\mathbb{P}_{ x}[ d(\widetilde\omega _T, E)\le 1]\ge\delta$, \begin{align*} \delta \le\mathbb{E}_{x}[d(\widetilde\omega_{T}, E)\le1] &=\mathbb{P}_{x}[\{d(\widetilde\omega_{T}, E)\le1\}\cap\mathscr{C}_{T,x}] +\mathbb{P}_{x}[\{d(\widetilde\omega_{T}, E)\le1\}\setminus\mathscr{C}_{T,x}]\\ &\le e^{-T(\hs-\varepsilon)}\sum_{y\in E} \mathrm{vol} B(y,1) + 1-(1-\frac{\delta}{2})\\ &\le C e^{-T(\hs-\varepsilon)} \mathrm{Card}(E) + \frac{\delta}{2}, \end{align*} where $C= \sup_{z} \mathrm{vol} B(z,1)$. Thus, $\frac{\delta}{2C} e^{T(\hs-\varepsilon)}\le \mathrm{Card}(E)$ and we have \begin{equation*} \hs \le \lim_{T\to\infty} \frac{1}{T}\log N(x,T,\delta). \end{equation*} For the converse inequality, Let $E$ be a minimal set satisfying $d(\widetilde\omega_{T}, E)\le 1$ for every $\widetilde\omega\in\mathscr{D}_{T, x}$ and $F\subset \{\gamma_{\theta(\widetilde\omega, \infty)}(\ld T): \widetilde\omega\in\mathscr{D}_{T, x}\}$ a maximal $\frac{1}{2}$-separated set. Note that $\mathrm{Card} (E) \ge N(x, T, \mathbb{P}_{x}(\mathscr{D}_{T, x}))$ and $\mathrm{Card} (F) \le C' e^{T(h+\varepsilon)}$. For each $f\in F$, \begin{equation*} N(f) := \{e \in E: \exists\,\widetilde\omega \in \mathscr{D}_{T, x} \textrm{ s.t. } \mathrm{d}(f, \gamma_{\theta(\widetilde\omega, \infty)}(\ld T)) \le \frac{1}{2}, \, \mathrm{d}(\widetilde\omega_{T}, e)\le1 \}. \end{equation*} Then $\mathrm{Card} N(f) \le e^{C'' \varepsilon T}$. Therefore, we have \begin{equation*} N(x, T, \mathbb{P}_{x}(\mathscr{D}_{T, x})) \le \mathrm{Card}(E) \le e^{C''\varepsilon T}\mathrm{Card} (F) \le C' e^{T [h+ (2+ C'')\varepsilon]}. \end{equation*} Given $\delta$, for each $T$ large enough, $N(x, T, \delta)\le N(x, T, \mathscr{D}_{T.,x})$. \end{proof} The stochastic entropy is related to the spectral information of $\widetilde{\mathcal{M}}$, the bottom of the spectrum $\lambda_{0}:=\inf \mathrm{Spec}(\Delta_{\widetilde{\mathcal{M}}})$ of the Laplacian on $\widetilde{\mathcal{M}}$. Note that $\lambda_{0}=(d-1)^{2}a^{2}/4$ if $\widetilde{\mathcal{M}}$ has constant negative curvature $-a^{2}$. It was proved in Proposition 3 of \cite{Led_1990} for co-compact manifolds. The proof is valid for pinched negative curvature and even the co-finiteness is not required. \begin{prop} $4\lambda_{0}\le\hs.$ \end{prop} \begin{proof} Since $\wp(t, x, y)$ is a solution of the heat equation, \begin{align*} \wp(t, x, y)\log \wp(t, x, y) &= \int_{0}^{t} \frac{\partial}{\partial s} \left( \wp(s, x, y)\log \wp(s, x, y) \right) ds\\ &= \int_{0}^{t} (1+\log \wp(s, x, y))\frac{\partial}{\partial s}\wp(s, x, y) ds\\ &=\int_{0}^{t} (1+\log \wp(s, x, y))\Delta_{y}\wp(s, x, y) ds. \end{align*} By applying this equation, \begin{align*} \hs &= \lim_{t\to \infty} -\frac{1}{t} \int_{\widetilde{\mathcal{M}}} \wp(t, x, y)\log \wp(t, x, y) d\mathrm{vol}(y)\\ &= \lim_{t\to\infty} \frac{1}{t} \int_{0}^{t} \int_{\widetilde{\mathcal{M}}} \langle \nabla \log \wp(s, x, y), \nabla \wp(s, x, y) \rangle_{g} d\mathrm{vol}(y) ds\\ &= \lim_{t\to\infty} \frac{4}{t} \int_{0}^{t} \int_{\widetilde{\mathcal{M}}} \left\| \nabla \sqrt{\wp(s, x, y)} \right\|^{2} d\mathrm{vol}(y) ds\\ &\ge \frac{4}{t}\int_{0}^{t}\lambda_{0} ds =4\lambda_{0}. \end{align*} The inequality is due to Rayleigh's theorem. \end{proof} \subsection{Thermodynamic formalisms in pinched negative curvature} We provide some general theory of thermodynamic formalisms for geodesic flows in pinched negative curvature. Notions and detailed arguments can be found in \cite{PPS}. A function $F$ on $\T^{1}\M$ is called a \textit{potential} on $\T^{1}\M$ if it is bounded and H\"older continuous. For a $\g^{t}$-invariant Borel probability measure $\mu$, if $h_{\mu}$ is the measure-theoretic entropy of the dynamical system $(\T^{1}\M, \g^{1}, \mu)$, we denote the pressure of $F$ for $\mu$ by $P(F, \mu)$: \begin{equation*} P(F, \mu)= h_{\mu} +\int_{\T^{1}\M} F d\mu. \end{equation*} An \textit{equilibrium state} $\mu_{F}$ for $F$ is a $\g^{t}$-invariant Borel probability measure of maximal pressure: \begin{equation*} P(F,\mu_{F})=\sup \, P(F, \mu) \end{equation*} where the supremum is taken among $\g^{t}$-invariant Borel probability measures $\mu$ s.t. $F_{-}:= \max\{ -F, 0\}$. We denote the supremum by $P_{F}$. Given a potential $F$ on $\T^{1}\M$, we denote the lift to $\T^{1}\widetilde{\mathcal{M}}$ by $\widetilde{F}$. We define a line integral of a potential by \begin{equation*} \int_{x}^{y} \widetilde{F} := \int_{0}^{\mathrm{d}(x,y)} \widetilde{F}(\g^{t}\mathrm{v}_{x}^{y})dt, \end{equation*} where $\mathrm{v}_{x}^{y}\in \T^{1}_{x}\widetilde{\mathcal{M}}$ is the unit vector at $x$ pointing $y$: $\gamma_{\mathrm{v}_{x}^{y}}(\mathrm{d}(x,y))= y$. A \textit{Patterson-Sullivan density} for $F$ of dimension $\delta$ is a family $(\mu_{x})_{x\in\widetilde{\mathcal{M}}}$ of finite Borel measures absolutely continuous to each other on $\partial\widetilde{\mathcal{M}}$ satisfying \begin{align*} \gamma_{*}\mu_{x} &=\mu_{\gamma x}, \\ d\mu_{y}(\xi) &=\exp\left( C_{F-\delta}(x, y, \xi)\right)d\mu_{x}(\xi), \end{align*} for each $x, y \in \widetilde{\mathcal{M}}$, $\gamma \in \Gamma$ where \begin{equation*} C_{F}(x, y, \xi) :=\lim_{z \to \xi} \int_{y}^{z}\widetilde{F} - \int_{x}^{z} \widetilde{F}. \end{equation*} We denote by $\mu^{\T}_{x}$ the \textit{spherical measure} at $x$, the push-forward measure of $\mu_{x}$ via the inverse of homeomorphism $\T^{1}_{x}\widetilde{\mathcal{M}} \to \partial\widetilde{\mathcal{M}}$ for each $x\in\widetilde{\mathcal{M}}$. Let $\mathrm{v}\in\T^{1}\M$ with a lift $\widetilde\mathrm{v}$ to a vector in $\T^{1}\widetilde{\mathcal{M}}$. Define the \textit{Bowen ball} around $\mathrm{v}$ by \begin{equation*} B(\mathrm{v}, T, T', r) := \{ \mathrm{w}\in \T^{1}\M: \sup_{t\in[-T', T]}\mathrm{d}( \gamma_{\widetilde\mathrm{v}}(t), \gamma_{\widetilde{\mathrm{w}}}(t))<r, \, \exists \,\textrm{a lift }\widetilde\mathrm{w}\in\T^{1}\widetilde{\mathcal{M}} \}. \end{equation*} One can construct a Gibbs measure from a Patterson-Sullivan density. That is, if a Patterson-Sullivan density $(\mu_{x})$ for $F$ of dimension $P_{F}$ is given, there is a $\g^{t}$-invariant Borel measure $\widetilde\mu$ on $\T^{1}\widetilde{\mathcal{M}}$ which is $\Gamma$-invariant and whose induced measure $\mu$ on $\T^{1}\M$ has a Gibbs property (see Section 3.8 of \cite{PPS}): For each compact set $K\in\T^{1}\widetilde{\mathcal{M}}$, there exist $r>0$ and $c_{K, r}>0$ such that for every $T, T'\ge0$ and for every $\mathrm{v}$, \begin{equation*} c_{K, r}^{-1}\exp\int_{-T'}^{T} \left(F(\g^{t}\mathrm{v})-P_{F}\right)dt \le \mu(B(\mathrm{v}, T, T', r) \le c_{K, r} \exp\int_{-T'}^{T} \left(F(\g^{t}\mathrm{v})-P_{F}\right) dt. \end{equation*} We call $\widetilde\mu$ the \textit{Gibbs measure} of $F$ and $(\mu_{x})$. The Gibbs measure determines whether an equilibrium state for $F$ exists or not. \begin{prop}\label{VP}(\cite{PPS}) $F$ is H\"older continous with $P_{F}<\infty$. \begin{enumerate} \item there is a Patterson-Sullivan density $(\mu_{x})$ for $F$ of dimension $P_{F}$ unique up to multiplicative constants. \item If the corresponding Gibbs measure $\widetilde\mu_{F}$ induces a finite measure $\mu_{F}$ on $\T^{1}\M$ then $\mu_{F}$ is the unique equilibrium state for $F$ and $\mu_{F}$ is ergodic. Otherwise, there is no equilibrium state for $F$. \end{enumerate} \end{prop} V. Pit and B. Schapira found a necessary and sufficient condition for the finiteness of Gibbs measure in \cite{PS_2018}. One can find the same statement also in \cite{PPS}. \begin{prop} \label{SPR} A H\"older continuous potential $F$ admits an equilibrium state if and only if for every maximal parabolic subgroup $\Pi$ of $\Gamma$, the following series converges: \begin{equation*} \sum_{\gamma\in\Pi} \mathrm{d}_{}(x, \gamma x) \exp \int_{x}^{\gamma x} (\widetilde{F}-P_{F}). \end{equation*} \end{prop} We have an ergodic theorem for the geodesic flow with respect to spherical measures. We also derive a Gibbs property for spherical measures (see \cite{Led_1988}). \begin{prop} \label{transversalerg} If a bounded H\"older continuous potential $F$ admits an equilibrium state $\mu$ then for every $\phi\in\C_{b}(\T^{1}\M)$, $x\in \M$ and for $\mu^{\T}_{x}$-a.e. $\mathrm{v}$ in $\T^{1}\M$, \begin{align} \label{transversalpterg} \frac{1}{t} \int_{0}^{t} \phi(\g^{s}\mathrm{v}) ds \to \int_{\T^{1}\M} \phi \, d \mu &\, \textrm{ as }t \to \infty, \\ \label{transversalmeasure} \lim_{t\to\infty} -\frac{1}{t} \log \mu^{\T}_{x}(B(\mathrm{v}, t, 0, \varepsilon)) = h_{\mu} &\, \textrm{ for some } \varepsilon>0. \end{align} \end{prop} \begin{proof} Since $\mu$ is ergodic, the set $G$ of the vectors for which the convergence (\ref{transversalpterg}) holds is a union of stable leaves with $\mu(G)=1$. Thus for any $x, y \in \M$, the projections $G_{x}^{+}:=\{\widetilde{\mathrm{v}}^{+}: \mathrm{v}\in\T^{1}_{x}\M\}$ and $G_{y}^{+}$ of fiber onto the boundary at infinity $\partial_{\infty}\widetilde{\mathcal{M}}$ are identical. Since $\mu_{x}^{\T}(G\cap \T^{1}_{x}\M) = \mu_{x}(G_{x}^{+})$, $G\cap \T^{1}_{x}\M$ is a $\mu_{x}^{\T}$-full set if and only if $G\cap\T^{1}_{y}\M$ is a $\mu_{y}^{\T}$-full set. Therefore $G\cap \T^{1}_{x}\M$ is a $\mu^{\T}_{x}$-full set for every $x\in\M$. From the $P_{F}$-Gibbs property of $\mu$, for $\mu^{\T}_{x}$-a.e. $\mathrm{v}\in G\cap \T^{1}_{x}\M$, \begin{align}\label{Gibbs} \lim_{t\to\infty} -\frac{1}{t} \log \mu (B(\mathrm{v}, t, 0, \varepsilon)) = P_{F} - \lim_{t\to\infty} \frac{1}{t}\int_{0}^{t} F(\g^{s}\mathrm{v})ds = P_{F} - \int Fd\mu = h_{\mu}. \end{align} A local stable manifold of $\mathrm{v} \in \T^{1}\M$ is \begin{equation*} W^{s}_{\varepsilon}(\mathrm{v}):=\{ \mathrm{w}: \mathrm{d}(\g^{t}\mathrm{v}, \g^{t}\mathrm{w})<\varepsilon, \, \forall t\ge0\}. \end{equation*} The spherical measure is a transversal measure, so it can be defined by local stable manifolds: \begin{equation*} \mu_{x}^{\T} (S) = \mu\left( \cup_{\mathrm{w}\in S} W^{s}_{\varepsilon}(\mathrm{w})\right). \end{equation*} Since the Bowen ball consists of local stable manifolds, (\ref{Gibbs}) holds when we replace $\mu$ by $\mu^{\T}_{x}$. \end{proof} There are two important potentials. The first is the zero potential, whose equilibrium state is the measure of maximal entropy, also called the \textit{Bowen-Margulis measure} if it admits an equilibrium state. The measure class of the Patterson-Sullivan density for the zero potential is called the \textit{visibility class}. The other is the \textit{geometric potential} $F^{su}$ induces from the $\Gamma$-invariant function \begin{equation*} \widetildeF^{su}(\mathrm{v})= -\left.\frac{d}{dt}\right|_{t=0} \log \det \T_{\mathrm{v}}\g^{t}|_{E^{su}(\mathrm{v})} \end{equation*} on $\widetilde{\mathcal{M}}$, where $\T_{\mathrm{v}}\g^{t}:\T_{\mathrm{v}}\T^{1}\widetilde{\mathcal{M}} \to \T_{\g^{t}\mathrm{v}}\T^{1}\widetilde{\mathcal{M}}$ is the tangent map of the flow map $\g^{t}$ at $\mathrm{v}$ and $E^{su}(\mathrm{v})= \T_{\mathrm{v}}\W^{su}(\mathrm{v})$ is the strong unstable distribution. Due to the pinched negative curvature and the uniform bound on the first derivatives of the sectional curvature, the angles between the stable leaves and the strong unstable leaves have positive lower bound and the foliations are H\"older continuous. Thus $F^{su}$ is H\"older continuous and the Liouville measure on $\T^{1}\M$ is the equilibrium state for $F^{su}$. The existence with an assumption on the pressure of $F^{su}$ is proved in Chapter 7 of \cite{PPS} and \cite{Riq_2018} proves that the assumption is true in our case. The measure class determined by the Patterson-Sullivan density for the geometric potential is called the \textit{Lebesgue class}. \section{Central limit theorem of Brownian motions} \subsection{Foliated Brownian motions} We shall introduce a Markov process on $\T^{1}\M$ called the foliated Brownian motion for the stable foliation of $\T^{1}\M$. The foliated Brownian motion was first introduced in the way to develop the ergodic theory of foliations (See \cite{CanCon}, \cite{Garnett_1983}). Fix a fundamental domain $\M_{0}\subset\widetilde{\mathcal{M}}$ of $\Gamma$. Identify $\M$, $\T^{1}\M$ with $\M_{0}$, $\M_{0}\times\partial\widetilde{\mathcal{M}}$, respectively. Note that $\widetilde\W^{s}(x, \xi)=\widetilde{\mathcal{M}}\times\{\xi\}$ is projected onto \begin{equation*} \W^{s}(x, \xi):=\{(y, \gamma^{-1} \xi) \in \T^{1}\M : y\in\M_{0}, \, \gamma \in \Gamma \}. \end{equation*} The stable foliation $\W^{s}=\{\W^{s}(\mathrm{v}):\mathrm{v}\in\T^{1}\M\}$ of $\T^{1}\M$ is the collection of the projected stable leaves. Similarly, we define the stable distribution $E^{s}$ of $\T^{1}\M$. The stable leaves of $\W^{s}$ inherit the Riemannian metric from $g_{s}$ on the leaves of $\widetilde\W^{s}$ which is also denoted by $g_{s}$. We denote the inherited differentials by $\mathrm{div}_{s}$ and $\Delta_{s}$. \begin{defn} Let $\mathcal{P}(\T^{1}\M)$ be the space of probability measures on $\T^{1}\M$. We define a transition semigroup $\mathbf{P}: (0, \infty)\times \T^{1}\M \to \mathcal{P}(\T^{1}\M)$ by \begin{equation*} d\,\mathbf{P}[t, \mathrm{v}](\mathrm{w}) = \sum_{\gamma\in\Gamma}\wp(t, x, \gamma y) \, d\delta_{\gamma^{-1}\xi}(\eta)\left. d\mathrm{vol}\right|_{\M_{0}}(y), \end{equation*} for $\mathrm{v}=(x,\xi), \mathrm{w}=(y,\eta)\in \T^{1}\M$. The transition semigroup defines a unique family $\{\mathbb{P}_{(x,\xi)}\}_{(x,\xi)\in\T^{1}\M}$ of Borel probability measures on the space $\T^{1}\Omega:=\C(\mathbb{R}_{+}, \T^{1}\M)$ of sample paths on $\T^{1}\M$. The canonical filtration is the collection of the smallest $\sigma$-algebras $\Fi_{t}=\Fi_{t}(\T^{1}\M):=\sigma\{\pi_{s}: 0 \le s \le t\}$ for which the projections $\pi_{s}(\omega)= \omega_{s}$ on $\T^{1}\Omega$ are measurable. The canonical process $Z_{t}(\omega)=\omega_{t}$ of the filtered space $( \T^{1}\Omega, \left\{ \Fi_{t} \right\}_{0\le t\le\infty})$ is a Markov process with respect to $\mathbb{P}_{(x, \xi)}$, which is called the foliated Brownian motion for $\W^{s}$ with initial distribution $\delta_{(x, \xi)}$, for each $(x, \xi)\in \T^{1}\M$. \end{defn} We define the Markov operator $\mathcal{Q}^{t}:\mathcal{C}_{b}(\T^{1}\M)\to\mathcal{C}_{b}(\T^{1}\M)$ on the space of bounded continuous functions on $\T^{1}\M$ by \begin{equation} \label{Markovint} \Q^{t}f (\mathrm{v}) := \int_{\T^{1}\M} f d\,\mathbf{P}[t, \mathrm{v}] = \sum_{\gamma\in\Gamma}\int_{\M_{0}} f(y, \gamma^{-1} \xi) \wp(t, x, \gamma y) d\mathrm{vol}(y). \end{equation} Note that the foliated Brownian motion for $\W^{s}$ is the projected process of a Markov process, called the foliated Brownian motion for $\widetilde\W^{s}$, with the transition semigroup \begin{equation} \label{diffusionint} d\, \widetilde{\mathbf{P}}[t, \mathrm{v}](\mathrm{w}) =\wp(t, x, y) d\delta_{\xi}(\eta)d\mathrm{vol}(y). \end{equation} Let $\widetilde\Q^{t}$ be the Markov operator on $\T^{1}\widetilde{\mathcal{M}}$. For any $f\in\C_{b}(\T^{1}\M)$ and for each $(x, \xi) \in \M_{0}\times\partial\widetilde{\mathcal{M}}$, \begin{equation*} \mathcal{Q}^{t}f(x, \xi) =\int_{\widetilde{\mathcal{M}}} \widetilde{f}(y, \xi)\wp(t, x, y)d\mathrm{vol}(y) =\widetilde\Q^{t}\widetilde{f}(x, \xi), \end{equation*} where $\widetilde{f}$ is the $\Gamma$-invariant lift of $f$ to $\T^{1}\widetilde{\mathcal{M}}$. Note that the infinitesimal generator of the Markov operator is the foliated Laplacian: \begin{equation*} \left. \frac{d}{dt}\right|_{t=0} \Q^{t} f = \Delta_{s} f. \end{equation*} L. Garnett proved in \cite{Garnett_1983} that the Markov operator $\widetilde\Q$ admits an invariant measure $\mathrm{m}^{\widetilde\Q}$ on $\T^{1}\widetilde{\mathcal{M}}$ of the form \begin{equation*} d\mathrm{m}^{\widetilde\Q}(x,\xi)= d\nu_{x}(\xi)d\widetilde\m(x)=\kk(x_{0}, x,\xi)d\widetilde\m(x) d\nu_{x_0}(\xi), \end{equation*} where $\widetilde{m}= \frac{1}{\mathrm{vol}(\M_{0})}\mathrm{vol}$ and $\nu_{x}$ is the harmonic measure. We have an induced probability measure $\hh := \mathrm{m}^{\widetilde\Q}|_{\M_{0}\times\partial\widetilde{\mathcal{M}}}$ on $\T^{1}\M$. By $\Gamma$-equivariance of $\nu_{x}$, \begin{align*} \int_{\T^{1}\M} \Q^{t}f d\hh &= \frac{1}{\mathrm{vol}(\M_{0})} \int_{\M_{0}}\int_{\partial\widetilde{\mathcal{M}}} \sum_{\gamma} \int_{\M_{0}} \widetilde{f}(y, \gamma^{-1}\xi)\wp(t, x, \gamma y) d\mathrm{vol}(y) d\nu_{x}(\xi) d\mathrm{vol}(x) \\ &= \frac{1}{\mathrm{vol}(\M_{0})} \int_{\M_{0}} \sum_{\gamma} \int_{\M_{0}}\int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(y, \xi)\wp(t, \gamma^{-1}x, y) d\nu_{\gamma^{-1}x}(\xi) d\mathrm{vol}(x) d\mathrm{vol}(y). \end{align*} Since we know $d\nu_{\gamma^{-1}x}(\xi) = \kk(y, \gamma^{-1} x, \xi)d\nu_{y}(\xi)$, the integrand in the right-handed side is: \begin{align*} &\sum_{\gamma} \int_{\M_{0}}\int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(y, \xi)\wp(t, \gamma^{-1}x, y) d\nu_{\gamma^{-1}x}(\xi) d\mathrm{vol}(x) \\ &= \int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(y, \xi) \sum_{\gamma} \int_{M_{0}}\wp(t, \gamma^{-1} x, y) \kk(y, \gamma^{-1} x, \xi) d\mathrm{vol}(x) d\nu_{y}(\xi)\\ &= \int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(y, \xi) \int_{\widetilde{\mathcal{M}}} \wp(t, x, y)\kk(y, x, \xi) d\mathrm{vol}(x) d\nu_{y}(\xi)\\ &=\int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(x, \xi) d\nu_{y}(\xi). \end{align*} We used the harmonicity of the Martin kernel in the last equality: \begin{equation*} \int_{\widetilde{\mathcal{M}}}\wp(t, x, y)\kk(y, x, \xi) d\mathrm{vol}(x)= \kk(y, y, \xi)=1. \end{equation*} Therefore, we have the $\Q^{t}$-invariance of $\hh$. The stationary measure of the foliated Brownian motion is $\mathbb{P}_{\hh}=\int_{\T^{1}\M} \mathbb{P}_{(x, \xi)}d\hh(x, \xi)$ and is ergodic for the shift map on $\T^{1}\Omega$. We have an integral expression of the linear drift and the stochastic entropy. Propsosition 2.9 and 2.16 in \cite{LedShu_2017} prove the same descriptions for the Brownian motion on co-compact negatively curved manifolds. The identities for co-finite manifolds follow in the same way. \begin{prop} \label{intcha} $\ld^{2}\le\hs$. Moreover, \begin{align*} \ld &= \int_{\M_{0}}\int_{\partial\widetilde{\mathcal{M}}} \Delta_{y} \bb(y, x, \xi) d\nu_{y}(\xi) d\widetilde\m(y) \\ &= \int_{\M_{0}} \int_{\partial\widetilde{\mathcal{M}}} \langle -\nabla_{y} \bb(y, x, \xi),\nabla_{y} \log \kk(x,y, \xi) \rangle_{g} \, d\nu_{y}(\xi) \, d\widetilde\m(y), \end{align*} \textrm{ and } \begin{align*} \hs &= \int_{\M_{0}} \int_{\partial\widetilde{\mathcal{M}}} |\nabla_{y} \log \kk(x,y, \xi)|^2 \, d\nu_{y}(\xi) \, d\widetilde\m(y) . \end{align*} \end{prop} \begin{proof} We only verify the second equality. The other equalities follow immediately from the same argument as in \cite{LedShu_2017}. \begin{align*} \ld &= \int_{\M_{0}}\int_{\partial\widetilde{\mathcal{M}}} \Delta_{y} \bb(y, x, \xi) d\nu_{y}(\xi) d\widetilde\m(y) \\ &= \int_{\partial\widetilde{\mathcal{M}}}\int_{\M_{0}} \Delta_{y}\bb(y, x, \xi) \kk(x, y, \xi) d\widetilde\m(y) d\nu_{x}(\xi) \\ &= \int_{\partial\widetilde{\mathcal{M}}}\int_{\M_{0}} \langle -\nabla_{y} \bb(y, x, \xi), \nabla_{y} \kk(x, y, \xi) \rangle_{g} d\widetilde\m(y) d\nu_{x}(\xi) \\ &= \int_{\M_{0}} \int_{\partial\widetilde{\mathcal{M}}} \langle -\nabla_{y} \bb(y, x, \xi), \nabla_{y}\log\kk(x, y, \xi)\rangle d\nu_{y}(\xi) d\widetilde\m(y). \end{align*} \end{proof} \subsection{Leafwise heat equation} We prove the contraction property on H\"older spaces of the foliated Brownian motion. Let $\tau>0$. We define a $\tau$-H\"older norm of $f$ in the space $\C_{b}(\T^{1}\M)$ of bounded continuous functions by \begin{equation*} \|f\|_{\mathcal{L}^{\tau}} =\|f\|_{\infty} +\sup_{x\in\M_{0}} \sup_{\xi,\eta\in\partial\widetilde{\mathcal{M}}} \frac{|\widetilde{f}(x,\xi)-\widetilde{f}(x,\eta)|}{d_{\infty}^{x,\tau}(\xi, \eta)}, \end{equation*} and we denote the corresponding H\"older space by \begin{equation*} \mathcal{L}^{\tau}=\{f\in\mathcal{C}_{b}( \T^{1}\M): \|f\|_{\mathcal{L}^{\tau}}<\infty\}. \end{equation*} The following statement corresponds to the uniqueness of a $\Q^{t}$-invariant measure for compact negatively curved manifolds (see \cite{Led_1995}). In \cite{Ham_1997}, it was shown that the uniqueness for the $(\Delta_{s}+Y)$-diffusion on compact negatively curved manifolds holds for a stably closed vector field $Y$ on $\T^{1}\M$ with positive pressure. \begin{prop} For every $\Q^{t}$-invariant measure $\eta$ on $\T^{1}\M$ and for each $f\in\mathcal{L}^{\tau}$, $$\int f d\eta = \int f d\hh.$$ \end{prop} \begin{proof} If $\eta$ is a $\Q^{t}$-invariant measure on $\T^{1}\M$, its $\Gamma$-invariant lift $\widetilde\eta$ to $\T^{1}\widetilde{\mathcal{M}}$ is disintegrated into $d\widetilde\eta(x,\xi)=d\widetilde\eta_{x}(\xi)d\widetilde\m(x)$ over the fibration $\T^{1}\widetilde{\mathcal{M}}=\widetilde{\mathcal{M}}\times\partial\widetilde{\mathcal{M}}$ where $\widetilde\eta_{x}$ are the conditional measures on the unit spheres $\T_{x}\widetilde{\mathcal{M}}=\{x\}\times\partial\widetilde{\mathcal{M}}$ of $\widetilde\eta$ (\cite{Garnett_1983}). As in the proof of Proposition \ref{transversalerg}, we can consider $\widetilde\eta_{x}$ as a probability measure of the union of local leaves; for some sufficiently small $\delta>0$, \begin{equation} \widetilde\eta_{x}(A):= \widetilde\eta( \cup_{\mathrm{w}\in A} W_{\delta}^{s}(\mathrm{w})) / \widetilde\eta(\cup_{\mathrm{v} \in \T^{1}_{x}\widetilde{\mathcal{M}}}W_{\delta}^{s}(\mathrm{v})). \end{equation} We denote by $\mathbb{E}_{x}$ the expectation with respect to $\mathbb{P}_{x}$. From the $\Q^{t}$-invariance, we have \begin{align*} \int_{\T^{1}\M}f d\eta &= \int_{\M_{0}} \int_{\partial\widetilde{\mathcal{M}}} \Q^{t} f(x,\xi) \, d\widetilde\eta_{x}(\xi) \, d\widetilde\m(x)\\ &= \int_{\M_{0}} \int_{\partial\widetilde{\mathcal{M}}} \int_{\widetilde{\mathcal{M}}} \wp(t,x,y) \widetilde{f}(y, \xi) \, d\mathrm{vol}(y) \, d\widetilde\eta_{x}(\xi) \, d\widetilde\m(x)\\ &= \int_{\M_{0}} \mathbb{E}_{x} \left[ \int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(\widetilde\omega_{t}, \xi) \, d\widetilde\eta_{x}(\xi) \right] \, d\widetilde\m(x). \end{align*} Note that given $\varepsilon>0$, $x\in\M_{0}$ and $f\in\mathcal{L}^{\tau}$, there is $\theta_{0}=\theta_{0}(\varepsilon)>0$ for every $y\in\widetilde{\mathcal{M}}$ and $\xi, \eta \in \partial\widetilde{\mathcal{M}}$ with $\angle_{y}(\xi, \eta)<\theta_{0}$, $ |\widetilde{f}(y, \xi)-\widetilde{f}(y, \eta)| < \varepsilon.$ Given $\xi \in \partial\widetilde{\mathcal{M}}$, we set for $T, \theta >0$, \begin{align*} \Upsilon(x, \xi, \theta) &:= \left\{ \widetilde\omega: \angle_{x}( \widetilde\omega_{\infty}, \xi ) <\frac{\theta}{3} \right\},\\ \Xi(x, T, \theta) &:= \left\{ \widetilde\omega: \mathrm{d}(x, \widetilde\omega_{t}) \ge \frac{\ld}{2}t, \angle_{x}( \theta (\widetilde\omega, t), \theta(\widetilde\omega, \infty) ) < \frac{\theta}{3}, \, \forall t \ge T \right\}. \end{align*} Then if $\theta\in(0,\theta_{0})$ is small enough, than for any $x$ and $\xi$, $\mathbb{P}_{x} (\Upsilon(x, \xi, \theta) ) <\frac{\varepsilon}{2\|f\|_{\infty}}$ (by \cite{BenHul_2019}). Choose such a small $\theta$. There is $T_{0}=T_{0}(x, \theta)$ such that if $t>T_{0}$, $|\widetilde{f}(\widetilde\omega_{t}, \xi) -\widetilde{f} (\mathrm{v}_{\widetilde\omega_{t}}^{x})|<\frac{\varepsilon}{2\|f\|_{\infty}}$ for each $\widetilde\omega\in\Xi(x, t, \theta) \setminus \Upsilon(x, \xi ,\theta)$ and $\mathbb{P}_{x} (\Xi(x, t, \theta) ) >1-\frac{\varepsilon}{2\|f\|_{\infty}}$. Hence if $t>T_{0}$, then \begin{align*} &\left| \mathbb{E}_{x} \left[ \int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(\widetilde\omega_{t}, \xi) - \widetilde{f} (\mathrm{v}_{\widetilde\omega_{t}}^{x}) \, d\widetilde\eta_{x}(\xi) \right] \right|\\ &\le \int_{\partial\widetilde{\mathcal{M}}} \mathbb{E}_{x} \left[ \left| \widetilde{f}(\widetilde\omega_{t}, \xi) -\widetilde{f} (\mathrm{v}_{\widetilde\omega_{t}}^{x}) \right| \left( \mathbf{1}_{\Xi(x, t, \delta) \setminus \Upsilon(x, \xi, \theta)} + \mathbf{1}_{\Xi(x, t, \delta)\cap \Upsilon(x, \xi, \theta)} +\mathbf{1}_{\Xi(x, t, \delta)^{c}} \right) \right] d \widetilde\eta_{x}(\xi) \\ &\le \varepsilon +2\|f\|_{\infty} \left( \int_{\partial\widetilde{\mathcal{M}}} \mathbb{P}_{x} \left[ \Upsilon(x, \xi, \theta) \right] d \widetilde\eta_{x}(\xi) + \mathbb{P}_{x} \left[ \Xi(x, t, \delta)^{c} \right] \right)\\ &< 3\varepsilon. \end{align*} Since $\phi_{t}(x):= \mathbb{E}_{x}\left[ \int \widetilde{f}(\widetilde\omega_{t}, \cdot)d\eta_{x}\right]$ and $\psi_{t}(x):=\mathbb{E}_{x}\left[\widetilde{f}(\mathrm{v}_{\widetilde\omega_{t}}^{x})\right]$ are bounded by $\|f\|_{\infty}$, it follows that $\phi_{t} - \psi_{t} \to 0$ as $t\to \infty$ in $L^{1}(\M_{0}, \widetilde\m)$ and hence $ \underset{t\to\infty}{\lim}\int \psi_{t} d\widetilde\m = \int f d\eta$. Thus we have \begin{align*} \int f d\eta =\lim_{t\to\infty} \int_{\M_{0}} \mathbb{E}_{x} \left[ \int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(\widetilde\omega_{t}, \xi) \, d\widetilde\eta_{x}(\xi) \right] \, d\widetilde\m(x) = \lim_{t\to\infty} \int_{\M_{0}} \mathbb{E}_{x} \left[ \widetilde{f}(\mathrm{v}_{\widetilde\omega_{t}}^{x}) \right] \, d\widetilde\m(x). \end{align*} From the $\Gamma$-invariance of $\widetilde{f}$ and the heat kernel, it follows that \begin{align*} \int_{\M_{0}} \mathbb{E}_{x} \left[ \widetilde{f}(\mathrm{v}_{\widetilde\omega_{t}}^{x}) \right] \, d\widetilde\m(x) &= \int_{\M_{0}} \int_{\M_{0}} \sum_{\gamma\in\Gamma} \wp(t,x,\gamma y) \widetilde{f}(\mathrm{v}_{\gamma y}^{x}) \, d\mathrm{vol}(y) \, d\widetilde\m(x)\\ &= \int_{\M_{0}} \int_{\M_{0}} \sum_{\gamma\in\Gamma} \wp(t,y, \gamma^{-1} x) \widetilde{f}(\mathrm{v}_{y}^{\gamma^{-1}x}) \, d\mathrm{vol}(x) \, d\widetilde\m(y)\\ &= \int_{\M_{0}} \mathbb{E}_{y} \left[ \widetilde{f}(\mathrm{v}_{y}^{\widetilde\omega_{t}}) \right] \, d\widetilde\m(y) . \end{align*} Letting $t$ tend to infinity, we have \begin{align*} \int f d\eta &= \int_{\M_{0}} \mathbb{E}_{y} \left[ \widetilde{f}(y, \widetilde\omega_{\infty}) \right] \, d\widetilde\m(y)\\ &= \int_{M_{0}} \int_{\partial\widetilde{\mathcal{M}}} \widetilde{f}(y, \xi) \, d\nu_{y}(\xi) \, d\widetilde\m(y). \end{align*} Therefore, $\int f d\eta = \int f d\hh$. \end{proof} We denote by $\N$ the integration operator on $\C_{b}(\T^{1}\M)$: \begin{equation*} \N(f) := \int_{\T^{1}\M} f \,d \hh. \end{equation*} The Markov operator $\Q^{t}$ converges to $\N$ on $\LL^{\tau}$. Furthermore the following theorem shows the rate of convergence is exponentially fast. We postpone the proof until Section \ref{ProofCtr}. \begin{thm}\label{Ctr} $\mathcal{Q}^{t}: \mathcal{L}^{\tau}\to\mathcal{L}^{\tau}$ defines a one-parameter semigroup of continuous operators for small enough $\tau>0$. Furthermore, there is $C=C(\tau)>0$ such that for every $t>0$, \begin{equation*} \| \mathcal{Q}^{t}-\mathcal{N}\|_{\mathcal{L}^{\tau}} \le e^{-Ct}. \end{equation*} \end{thm} Given $f\in \LL^{\tau}$, if $\int f d\hh=0$, then the $\LL^{\tau}$-limit of $\int_{0}^{T}\Q^{t}f dt$ exists by the contraction property. The limit $u:= \lim_{T\to\infty}\int_{0}^{T} \Q^{t}f dt$ is a weak solution of the leafwise heat equation $\Delta_{s} u = -f$, thus a strong solution in $\LL^{\tau}$. Since a leafwise harmonic $u$ is $\Q^{t}$-invariant, the uniqueness also follows from the contraction property (See \cite{Led_1995} for the detail). Therefore we obtain the following corollary. \begin{coro} \label{Heq} For small enough $\tau>0$ and every $f\in\LL^{\tau}$ with $\int fd\hh=0$, there exists a solution $u \in \LL^{\tau}$ to the leafwise heat equation $\Delta_{s} u=-f$ which is unique up to additive constants. In addition, $u$ is $\mathcal{C}^{2}$ along the stable leaves. \end{coro} Let $\alpha: \T^{1}\M\to E^{s*}$ be a continuous section of the dual bundle $E^{s*}$ of the stable distribution $E^{s}$ of $\T^{1}\M$ and $\widetilde\alpha : \T^{1}\widetilde{\mathcal{M}} \to \widetilde{E}^{s}$ be the lift of $\alpha$. The section $\alpha$ is called a \textit{leafwise closed 1-form} of class $\C^{1}$ if $\widetilde\alpha|_{\widetilde\W^{s}(\mathrm{v})}$ is a closed 1-form on $\widetilde\W^{s}(\mathrm{v})$ of class $\C^{1}$ for any $\mathrm{v}\in\T^{1}\widetilde{\mathcal{M}}$. For each $(x, \xi)\in \T^{1}\widetilde{\mathcal{M}}$, since $\widetilde\W^{s}(x,\xi)=\widetilde{\mathcal{M}}\times\{\xi\}$ is diffeomorphic to $\widetilde{\mathcal{M}}$, there is a 1-form $\widetilde\alpha^{\xi}$ on $\widetilde{\mathcal{M}}$ which agrees with the pull-back of $\widetilde\alpha|_{\widetilde\W^{s}(x,\xi)}$. Furthermore, if $\alpha$ is a leafwise closed 1-form of class $\C^{1}$, then there exists $A^{\xi}\in\C^{1}(\widetilde{\mathcal{M}})$ such that $dA^{\xi}=\widetilde\alpha^{\xi}$. Hence if $\alpha$ is a leafwise closed 1-form of class $\C^{1}$, we define for each foliated Brownian path $\omega \in \T^{1}\Omega$ starting from $(x,\xi)\in\T^{1}\M$, \begin{equation*} \int_{\omega_{0}}^{\omega_{t}}\alpha:=A^{\xi}(\widetilde\omega_{t})-A^{\xi}(\widetilde\omega_{0}) \end{equation*} for every $t\ge0$, where $\widetilde\omega$ is a Brownian path on $\T^{1}\widetilde{\mathcal{M}}$ such that $(\widetilde\omega_{t}, \xi)\in \T^{1}\widetilde{\mathcal{M}}$ is a lift of $\omega_{t}$. We denote by $\delta_{s}$ the leafwise codifferential $g_{s}$-dual to $-\mathrm{div}_{s}$, that is, $\delta_{s}\alpha = -\mathrm{div}_{s}\alpha^{\#}$ where $\alpha^{\#}:\T^{1}\M\to E^{s}$ is the continuous section $g_{s}$-dual to $\alpha$. Since \begin{equation*} \delta_{s}\widetilde\alpha(x,\xi) = -\mathrm{div}_{s}\widetilde\alpha^{\#}(x, \xi) = -\mathrm{div}\nabla A^{\xi}(x) = -\Delta A^{\xi}(x) = -\Delta_{s} A(x, {\xi}), \end{equation*} by It\^o's formula (see Chapter 3 of \cite{Hsu}), \begin{equation} \label{ito} \mathbf{X}_{t}(\omega) =\int_{\omega_{0}}^{\omega_{t}}\alpha+\int_{0}^{t}\delta_{s}\alpha(\omega_{r})dr =A^{\xi}(\widetilde\omega_{t})-A^{\xi}(\widetilde\omega_{0})-\int_{0}^{t} \Delta A^{\xi}(\widetilde\omega_{r})dr \end{equation} is a martingale on $(\T^{1}\Omega, \{\Fi_{t}(\T^{1}\M)\}_{0\le t \le\infty}, \mathbb{P}_{\hh})$ having the quadratic variation \begin{equation*} d\langle \mathbf{X}, \mathbf{X}\rangle_{t}(\omega)=(\Delta(A^{\xi})^{2}-2A^{\xi}\Delta A^{\xi})(\widetilde\omega_{t})dt=2\|\alpha^{\#}(\omega_{t})\|^{2}dt. \end{equation*} If $\beta$ is a leafwise closed 1-form of class $\C^{1}$ such that $\delta_{s}\beta$ is H\"older continuous on $\T^{1}\M$, applying Corollary \ref{Heq}, there is $u\in\mathcal{L}^{\tau}$ such that $\Delta_{s}u=\delta_{s}\beta-\int \delta_{s}\beta d\hh$. Hence, due to the equation (\ref{ito}) for $\alpha= \beta+du$, we have a martingale \begin{equation}\label{itohol} \mathbf{X}_{t} =\int_{\omega_{0}}^{\omega_{t}}(\beta+du)+\int_{0}^{t}\delta_{s}(\beta+du)(\omega_{r})dr =\int_{\omega_{0}}^{\omega_{t}}\beta +u(\omega_{t})-u(\omega_{0})+t\int\delta_{s}\beta d\hh \end{equation} with the quadratic variation $\langle \mathbf{X}, \mathbf{X} \rangle_{t}(\omega)=2\int_{0}^{t}\|\alpha^{\#}+\nabla u\|^{2}(\omega_{t})ds$. \subsection{Proof of Theorem \ref{CLT}} For $(x,\xi)\in\T^{1}\widetilde{\mathcal{M}}$, let $B(x, \xi):=\bb(x, x_{0}, \xi)$, $K(x, \xi):=\log \kk(x_{0}, x, \xi)$. Note that \begin{equation*} \Delta_{s}B(x, \xi) = \Delta_{x}\bb(x, x_{0}, \xi) \end{equation*} is H\"older continuous due to uniform bounds of the first derivatives of curvature. On the other hand, \begin{equation*} \Delta_{s} K(x, \xi) = -\| \nabla_{x} \log \kk(x_{0}, x, \xi)\|^{2} \end{equation*} is H\"older contunous due to \cite{AndSch_1985}, \cite{Hamenst_dt_1990}. By Corollary \ref{Heq} for $f= \Delta_{s}B$, $\Delta_{s}K$ there exist $u_{\bb}, u_{\kk} \in\mathcal{L}^{\tau}$ for which we obtain square-integrable martingales \begin{align*} \mathbf{B}_{t}(\omega) = \bb(\widetilde\omega_{t}, \widetilde\omega_{0}, \xi) - t\ld +u_{\bb}(\omega_{t}) -u_{\bb}(\omega_{0}), \, \, \mathbf{K}_{t}(\omega) = \log \kk(\widetilde\omega_{0}, \widetilde\omega_{t}, \xi) +t\hs +u_{\kk}(\omega_{t}) -u_{\kk}(\omega_{0}), \end{align*} for $\omega\in\T^{1}\Omega$ with a lift $(\widetilde\omega, \xi)\in \T^{1}\widetilde{\mathcal{M}}$, by the It\^o formula (\ref{itohol}) for $\beta_{(x, \xi)}=dB^{\xi}_{x}$, $dK^{\xi}_{x}$, respectively. Their quadratic variations are \begin{align} \label{quad} \langle \mathbf{B}, \mathbf{B} \rangle_{t}(\omega) = 2\int_{0}^{t}\| \nabla B +\nabla u_{\bb} \|^{2}(\omega_{s})ds, \,\, \langle\mathbf{K}, \mathbf{K}\rangle_{t}(\omega) = 2\int_{0}^{t}\| \nabla K + \nabla u_{\kk} \|^{2}(\omega_{s})ds. \end{align} We denote by $\mathbb{E}_{(x, \xi)}$ the expectation with respect to $\mathbb{P}_{(x, \xi)}$. From the equalities (\ref{quad}) of quadratic variations, \begin{align*} \mathbb{E}_{(x,\xi)} \left[ \frac{1}{t} \langle \mathbf{B}, \mathbf{B} \rangle_{t}(\omega) \right] =\frac{2}{t}\int_{0}^{t} \Q^{s}\| \nabla B +\nabla u_{\bb} \|^{2}(x, \xi) ds, \\ \mathbb{E}_{(x,\xi)} \left[ \frac{1}{t} \langle \mathbf{K}, \mathbf{K} \rangle_{t}(\omega) \right] =\frac{2}{t}\int_{0}^{t} \Q^{s}\| \nabla K +\nabla u_{\kk} \|^{2}(x, \xi) ds. \end{align*} Due to the ergodicity of $\hh$, for $\hh$-a.e. $(x, \xi)$, \begin{align} \label{QV1} \lim_{t\to\infty} \mathbb{E}_{(x, \xi)} \left[ \frac{1}{t} \langle \mathbf{B}, \mathbf{B} \rangle_{t}(\omega) \right] &= 2\int \| \nabla B + \nabla u_{\bb} \|^{2} d\hh, \\ \label{QV2} \lim_{t\to\infty} \mathbb{E}_{(x, \xi)} \left[ \frac{1}{t} \langle \mathbf{K}, \mathbf{K} \rangle_{t}(\omega) \right] &= 2\int \| \nabla K + \nabla u_{\kk} \|^{2} d\hh. \end{align} Using Markov property, we have \begin{align*} \mathbb{E}_{(x, \xi)}\left[ \frac{1}{t+1} \langle \mathbf{M}, \mathbf{M} \rangle_{t+1} \right] &=\mathbb{E}_{(x, \xi)} \left[ \frac{t}{t+1} \mathbb{E}_{\omega_{1}} \left[ \frac{1}{t} \langle \mathbf{M}, \mathbf{M} \rangle_{t} \right] \right]\\ &=\frac{t}{t+1} \mathbb{E}_{(x, \xi)} \left[ \frac{1}{t} \int_{0}^{t}\Q^{r}F (\omega) dr \right]. \end{align*} for $\mathbf{M} = \mathbf{B}$ or $\mathbf{K}$ and $F = 2\| \nabla B +\nabla u_{\bb} \|^{2}$ or $2\| \nabla K +\nabla u_{\kk} \|^{2}$, respectively. Given $x\in\widetilde{\mathcal{M}}$, for $\nu_{x}$-a.e. $\xi$ and $\mathbb{P}_{(x, \xi)}$-a.e. $\omega$, \begin{equation*} \lim_{t\to\infty} \frac{1}{t}\int_{0}^{t}\Q^{r}F(\omega) dr = \int F d\hh. \end{equation*} Hence for each $x$, there is $\xi$ for which we have the limits (\ref{QV1}) and (\ref{QV2}). We denote the square root of the limits by $\sigma_{\bb}$ and $\sigma_{\kk}$, respectively. Note that both of $\sigma_{\bb}, \sigma_{\kk}$ are positive since $B$ and $K$ are unbounded while $u_{\bb}$ and $u_{\kk}$ are bounded. We have $\sigma_{\bb}, \sigma_{\kk} <\infty $ since both of $2\| \nabla B +\nabla u_{\bb} \|^{2}$ or $2\| \nabla K +\nabla u_{\kk} \|^{2}$ are bounded. Thus for every $x$, there is $\xi$ such that the distributions of $\frac{\mathbf{B}_{t} }{\sigma_{\bb}\sqrt{t}}$ and $\frac{\mathbf{K}_{t}}{\sigma_{\kk}\sqrt{t}}$ under $\mathbb{P}_{(x, \xi)}$ converge to $N(0,1)$ as $t\to\infty$ due to the following lemma : \begin{lem} (\cite{Hel_1982}) Let $(M_{t})_{0\le t\le\infty}$ be a continuous, centered, square-integrable martingale on a filtered probability space with stationary increments. If $M_{0}=0$ and there is $\sigma>0$ such that $\lim_{t\to\infty}\mathbb{E}[ |\frac{1}{t}\langle M, M\rangle_{t}-\sigma^{2}| ] = 0$, then the distribution of $\frac{1}{\sigma\sqrt{t}}M_{t}$ is asymptotically normal. \end{lem} Let $W^{\ld}_{t}(\omega):=\mathrm{d} (\widetilde\omega_{0}, \widetilde\omega_{t})-t\ld$. Since the distribution of $W^{\ld}_{t}$ under $\mathbb{P}_{(x, \xi)}$ and the distribution of $Y^{\ld}_{t}$ under $\mathbb{P}_{x}$ coincide, it is enough to show that $W^{\ld}_{t}$ and $\mathbf{B}_{t}$ have the same $\mathbb{P}_{(x, \xi)}$-distribution. For $\mathbb{P}_{(x, \xi)}$-a.e. $\omega$ and a lift $(\widetilde\omega, \xi)$, since $B(\omega_{t})-B(\omega_{0}) - \mathrm{d}(\widetilde\omega_{0}, \widetilde\omega_{t}) = \bb(\widetilde\omega_{t}, \widetilde\omega_{0}, \xi) - \mathrm{d}(\widetilde\omega_{0}, \widetilde\omega_{t}) \to -2(\xi|\widetilde\omega_{\infty})_{\widetilde\omega_{0}}$ and $|(\xi|\widetilde\omega_{\infty})_{\widetilde\omega_{0}}|<\infty$, \begin{equation*} \lim_{t\to\infty} \frac{1}{\sigma_{\bb}\sqrt{t}} \left[ B(\omega_{t})-B(\omega_{0})-\mathrm{d}(\widetilde\omega_{0}, \widetilde\omega_{t}) \right] =0. \end{equation*} Hence the distribution of $\frac{1}{\sigma_{\bb}\sqrt{t}}W^{\ld}_{t}$ under $\mathbb{P}_{(x, \xi)}$ also converges to the normal distribution since \begin{align*} W^{\ld}_{t}(\omega) = & \left[ \mathrm{d}(\widetilde\omega_{0}, \widetilde\omega_{t}) -B(\omega_{t})+B(\omega_{0}) \right] - \left[ u_{\bb}(\omega_{t}) -u_{\bb}(\omega_{0}) \right] + \mathbf{B}_{t}(\omega), \end{align*} and \begin{equation*} \frac{1}{\sigma_{\bb}\sqrt{t}} \left| u_{\bb}(\omega_{t})-u_{\bb}(\omega_{0}) \right| \le \frac{2}{\sigma_{\bb}\sqrt{t}}\|u_{\bb}\|_{\infty} \to 0, \textrm{ as }t\to\infty. \end{equation*} Let $W^{\hs}_{t}(\omega):= \log \G(\widetilde\omega_{0}, \widetilde\omega_{t}) + t\hs$. Since the $\mathbb{P}_{(x, \xi)}$-distribution of $W^{\hs}_{t}$ and the $\mathbb{P}_{x}$-distribution of $Y^{\hs}_{t}$ are the same, to verify that $\frac{1}{\sigma_{\kk}\sqrt{t}} W^{\hs}_{t}$ is asymptotically normal, it is sufficient to show that for $\mathbb{P}_{(x, \xi)}$-a.e. $\omega$ with a lift $(\widetilde\omega, \xi)$ to $\T^{1}\widetilde{\mathcal{M}}$, \begin{equation} \label{GMeq} \limsup_{t\to\infty} \left| \log \G(\widetilde\omega_{0}, \widetilde\omega_{t}) -K(\omega_{t})+K(\omega_{0}) \right| <\infty. \end{equation} Note that for $\mathbb{P}_{(x, \xi)}$-a.e. $\omega$ with a lift $(\widetilde\omega, \xi)$, $K(\omega_{t})-K(\omega_{0})=\log \kk(\widetilde\omega_{0}, \widetilde\omega_{t}, \xi)$ and $\widetilde\omega_{\infty}\ne \xi$. We denote by $z_{t}$ the closest point to $\widetilde\omega_{0}$ on the geodesic ray $[\widetilde\omega_{t}, \xi)$ generated by $(\widetilde\omega_{t}, \xi)$, $z_{t}$ converges to a point $z_{\infty}\in\widetilde{\mathcal{M}}$ on the geodesic $(\widetilde\omega_{\infty}, \xi)$ joining two boundary points $\widetilde\omega_{\infty}$ and $\xi$. We have that for every $y$ on $[\widetilde\omega_{t}, \xi)$, \begin{align*} |\log\G(\widetilde\omega_{0}, \widetilde\omega_{t})-\log \kk(\widetilde\omega_{0}, \widetilde\omega_{t},\xi)| \le \left|\log \frac{\G(\widetilde\omega_{0},\widetilde\omega_{t})}{\G(z_{t}, \widetilde\omega_{t})}\right| + \left| \log \frac{\G(z_{t}, \widetilde\omega_{t})} {\left(\frac{\G(y, \widetilde\omega_{t})}{\G(y, z_{t})}\right)} \right| + \left|\log\frac {\left(\frac{\G(y, \widetilde\omega_{t})}{\G(y, z_{t})}\right)} {\kk(\widetilde\omega_{0}, \widetilde\omega_{t},\xi)}\right|. \end{align*} Applying the Harnack inequality to the first term in the right handed side, since $\{\mathrm{d}(\widetilde\omega_{0}, z_{t})\}_{t\ge0}$ is bounded, it follows that $\left|\log\frac{\G(\widetilde\omega_{0}, \widetilde\omega_{t})}{\G(z_{t},\widetilde\omega_{t})}\right| \le C_{1}$ for some constant $C_{1}=C_{1}(\widetilde\omega)>0$ dependent of $\widetilde\omega$ but not $t$. And by the Ancona inequality (\cite{Ancona_1987}), the second term in the right handed side is also bounded by $C_{2}(\widetilde\omega)$. Letting $y$ tend to $\xi$, we see that the last term converges to $\left|\log\frac{\kk(z_{t}, \widetilde\omega_{t}, \xi)}{\kk(\widetilde\omega_{0}, \widetilde\omega_{t}, \xi)}\right|$ which is also bounded by $C_{1}(\widetilde\omega)$ due to the Harnack inequality. Therefore we have (\ref{GMeq}) and this completes the proof of Theorem \ref{CLT}. \section{Proof of Theorem \ref{Ctr}}\label{ProofCtr} In this section, we prove the contraction property on H\"older spaces of the foliated Brownian motion. For the H\"older semi-norm, we prove a lower bound of the expectation of the Busemann functions at Brownian points which depends only on the dimension and the curvature bounds and linearly on time $T$. The lower bound follows from the fact that the Laplacian of the Busemann function has the same lower bound with the Laplacian of the distance function due to the Rauch comparison theorem. We also show the Doeblin property of the Brownian motion for the estimate of the uniform norm. \begin{prop} \label{CtrH} For sufficiently small $\tau$, there exists $C_{1}>0$ such that for each $t>0$, \begin{equation*} \sup_{x\in\M_{0}} \sup_{\xi,\eta\in\partial\widetilde{\mathcal{M}}} \frac{|\mathcal{Q}^{t}f(x,\xi)-\mathcal{Q}^{t}f(x,\eta)|}{d_{\infty}^{x,\tau}(\xi, \eta)} \le \|f\|_{\mathcal{L}^{\tau}} e^{-C_{1}t}. \end{equation*} \end{prop} \begin{proof} Since we have that \begin{align*} \frac{|\mathcal{Q}^{t}f(x,\xi)-\mathcal{Q}^{t}f(x,\eta)|}{d_{\infty}^{x,\tau}(\xi, \eta)} &<\int_{\widetilde{\mathcal{M}}} \wp(t,x, y) \frac{\left|\widetilde{f}(y, \xi)-\widetilde{f}(y, \eta)\right|}{d_{\infty}^{x,\tau}(\xi, \eta)} d\mathrm{vol}_{\widetilde{\mathcal{M}}}(y) \\ &<\|f\|_{\mathcal{L}^{\tau}} \int_{\widetilde{\mathcal{M}}} \wp(t,x,y) \frac{d_{\infty}^{y,\tau}(\xi, \eta)}{d_{\infty}^{x,\tau}(\xi, \eta)} d\mathrm{vol}_{\widetilde{\mathcal{M}}}(y)\\ &=\|f\|_{\mathcal{L}^{\tau}} \mathbb{E}_{x} \left[ \frac{d_{\infty}^{y,\tau}(\xi, \eta)}{d_{\infty}^{x,\tau}(\xi, \eta)} \right], \end{align*} it is sufficient to find $C_{1}>0$ such that \begin{equation*} \sup_{x, \xi, \eta} \mathbb{E}_{x} \left[ \frac{d_{\infty}^{\widetilde\omega_{t},\tau}(\xi, \eta)}{d_{\infty}^{x,\tau}(\xi, \eta)} \right] < e^{-C_{1}t}. \end{equation*} Due to the Markov property of the Brownian motion, \begin{align*} \sup_{x, \xi, \eta} \mathbb{E}_{x} \left[ \frac{d_{\infty}^{\widetilde\omega_{t+s},\tau}(\xi, \eta)}{d_{\infty}^{x,\tau}(\xi, \eta)} \right] &= \sup_{x, \xi, \eta} \mathbb{E}_{x} \left[ \frac{d_{\infty}^{\widetilde\omega_{s},\tau}(\xi, \eta)}{d_{\infty}^{x,\tau}(\xi, \eta)} \mathbb{E}_{x} \left[ \left. \frac{d_{\infty}^{\widetilde\omega_{t+s},\tau}(\xi, \eta)}{d_{\infty}^{\widetilde\omega_{s},\tau}(\xi, \eta)} \right| \mathscr{F}_{s}(\widetilde{\mathcal{M}}) \right] \right]\\ &\le \sup_{x, \xi, \eta} \mathbb{E}_{x} \left[ \frac{d_{\infty}^{\widetilde\omega_{t},\tau}(\xi, \eta)}{d_{\infty}^{x,\tau}(\xi, \eta)} \right] \sup_{x, \xi, \eta} \mathbb{E}_{x} \left[ \frac{d_{\infty}^{\widetilde\omega_{s},\tau}(\xi, \eta)}{d_{\infty}^{x,\tau}(\xi, \eta)} \right]. \end{align*} Let us write $g(\widetilde\omega_{t}):=(\xi|\eta)_{\widetilde\omega_{t}} -(\xi|\eta)_{x}$. Applying the Taylor theorem to the function $R\mapsto\exp(-\tau R)$ and substituting $g(\widetilde\omega_{t})$ for $R$, we have \begin{align*} \frac{d_{\infty}^{\widetilde\omega_{t},\tau}(\xi, \eta)}{d_{\infty}^{x,\tau}(\xi, \eta)} \le 1-\tau g(\widetilde\omega_{t}) +\tau^{2}\mathrm{d}_{}(x, \widetilde\omega_{t})^{2} e^{2\tau\mathrm{d}_{}(x,\widetilde\omega_{t})}. \end{align*} By Proposition \ref{HCom}, for some constant $C_{1}'>0$, \begin{equation}\label{SEM} \sup_{x} \mathbb{E}_{x} \left[ \mathrm{d}_{}(x, \widetilde\omega_{t})^{2} e^{2\tau\mathrm{d}_{}(x,\widetilde\omega_{t})} \right] < C_{1}'. \end{equation} Therefore, with (\ref{SEM}) and Lemma \ref{ctrlem} below, we have \begin{equation*} \sup_{0\le t \le T} \sup_{x, \xi,\eta} \mathbb{E}_{x} \left[ \frac{d_{\infty}^{\widetilde\omega_{t}, \tau}(\xi, \eta)} {d_{\infty}^{x,\tau}(\xi, \eta)} \right] \le 1- \tau (d-1)a + \tau^{2} C_{1}'. \end{equation*} Fix $T\ge1$ and sufficiently small $\tau$ such that $1- \tau (d-1)a + \tau^{2} C_{1}'<1$. For such small $\tau$, put $C_{1}=(1-a(d-1)\tau+C_{1}'\tau^{2})^{\frac{1}{T}}$ and the inequality follows. \end{proof} \begin{lem}\label{ctrlem} For every $T\ge0$, \begin{equation*} \inf_{x \in\M_{0}}\inf_{\xi\ne\eta} \mathbb{E}_{x}[(\xi|\eta)_{\widetilde\omega_{T}}-(\xi|\eta)_{x}] \ge(d-1)a T. \end{equation*} \end{lem} \begin{proof}[Proof of Lemma \ref{ctrlem}] Due to the equation $$(\xi|\eta)_{x}-(\xi|\eta)_{y}= \frac{1}{2} \bb(x, y, \xi)+\frac{1}{2}\bb(x,y,\eta),$$ it suffices to show that \begin{equation*} \mathbb{E}_{x}[\bb(\widetilde\omega_{T}, x, \eta)] \ge (d-1)a T. \end{equation*} Choose $z_{n}\in\widetilde{\mathcal{M}}$ such that $z_{n}\to \xi$ as $n\to\infty$ and write $$f_{n}(y):=\bb(y, x, z_{n})= \mathrm{d}_{}(y, z_{n})-\mathrm{d}_{}(x, z_{n}).$$ By the Rauch's comparison theorem (see \cite{Petersen_2016}, for instance), \begin{align}\label{RCom} \Delta f_{n} (y) &= \Delta_{y} \mathrm{d}_{}(y, z_{n}) \ge (d-1)\frac{ \textrm{sn}_{-a^2}'(\mathrm{d}_{}(y, z_{n}))}{\textrm{sn}_{-a^2}(\mathrm{d}_{}(y, z_{n}))}\\ &=a(d-1) \coth \left(a\,\mathrm{d}_{}(y, z_{n})\right) \end{align} where $\textrm{sn}_{-a^{2}}(t)=\frac{1}{a}\sinh (at)$. Let $f(y, \xi)=\bb(y, x,\xi)$. Then, since $\Delta_{s}$ is the generator of $\mathcal{Q}^{t}$ and $\Delta_{s}f(y,\xi) = \Delta_{y}f(y, \xi)$, $$ \mathbb{E}_{x}[\bb(\widetilde\omega_{T}, x, \xi)] =\mathcal{Q}^{T}f(x, \xi) =\int_{0}^{T} \mathcal{Q}^{t}\Delta_{s} f(x, \xi)dt =\int_{0}^{T} \mathbb{E}_{x}[\Delta\bb(\widetilde\omega_{t}, x, \xi)]dt. $$ Due to (\ref{RCom}) and Proposition \ref{C2conv}, \begin{equation*} \mathbb{E}_{x}[\bb(\widetilde\omega_{T}, x, \xi)]\ge (d-1)aT \end{equation*} for every $x\in\widetilde{\mathcal{M}}$ and every $\xi\in\partial\widetilde{\mathcal{M}}$. \end{proof} Write $P(t,x,y)=\sum_{\gamma\in\Gamma} \wp(t, x, \gamma y)$ for $x, y\in\M_{0}$. We have $\lim_{t\to\infty}P(t,x,y)=\frac{1}{\mathrm{vol}(\M)}$, in particular, $P(t, x, x)$ decreases as $t\to\infty$ (see \cite{ChaK_1991}). We also have that \begin{align} \label{CSineq} \left( \int_{\M} \left| P(t, x, y) -\frac{1}{\mathrm{vol}(\M)} \right| d\mathrm{vol}(y) \right)^{2} &\le \mathrm{vol}(\M) \int_{\M} \left| P(t, x, y) -\frac{1}{\mathrm{vol}(\M)} \right|^{2} d\mathrm{vol}(y)\\ &= \mathrm{vol}(\M) \left( P(2t, x, x)-\frac{1}{\mathrm{vol}(\M)} \right). \end{align} Hence the integral on the left-handed side decreases to zero as $t$ goes to infinity. Indeed, it decays exponentially fast (see \cite{Don_1987}). The following lemma shows that it has uniform exponential decay rate. \begin{lem} \label{Doeblin} There exists a constant $C_{2}=C_{2}(d, b)>0$ such that for each $x\in\M$, \begin{equation*} \int_{\M} \left| P(t, x, y) -\frac{1}{\mathrm{vol}(\M)} \right| d\mathrm{vol}(y) \le C_{2} e^{-\frac{\lambda_{1}}{2}t}, \end{equation*} where $\lambda_{1}=\inf\{ \lambda>0: \lambda \in \mathrm{Spec}(\Delta_{\M})\}$. \end{lem} \begin{rmk} Since the bottom of the ($L^{2}$-)esssential spectrum $\lambda_{ess}:=\inf \mathrm{Spec}_{ess}\left(\Delta_{\M}\right)$ of the Laplacian is positive (\cite{Dodziuk_1987}) and $\mathrm{Spec} \left(\Delta_{\M}\right)\cap [0, \lambda_{ess})$ is discrete (\cite{Don_1987}), the smallest nonzero the spectrum $\lambda_{1}$ is also positive. \end{rmk} \begin{proof} If we consider $\mathcal{P}^{t}f(x):=\int (P(t, x, y)-\mathrm{vol}(\M)^{-1})f(y)d\mathrm{vol}(y)$ as a self-adjoint operator acting on the space $L^{2}_{0}(\M)$ of square-integrable functions with zero integral, $\Delta|_{L^{2}_{0}(\M)}$ is the generator of $P^t$ with the bottom of the spectrum $\lambda_{1}$. Therefore the operator norm satisfies \begin{equation}\label{heatctr} \|\mathcal{P}^{t}\|\le e^{-\frac{\lambda_{1}t}{2}} \end{equation} for every $t>0$ (see the proof of Proposition V.1.2 in \cite{EngNag}). For every $x\in\M_{0}$, if we denote $f_{t}(y) = P(t, x, y)-\frac{1}{\mathrm{vol}(\M)}$, then $f_{t+t_{0}}(y) = \mathcal{P}^{t}f_{t_{0}}(y)$. It follows from (\ref{CSineq}) and (\ref{heatctr}) that \begin{align*} \int_{\M} \left| P(t+t_{0}, x, y) - \frac{1}{\mathrm{vol}(\M)} \right| d\mathrm{vol}(y) &\le \left( \mathrm{vol}(\M) \int_{\M} \left| f_{t+t_{0}}(y) \right|^{2} d\mathrm{vol}(y) \right)^{1/2} \\ &\le \|\mathcal{P}^{t}\| \left\| f_{t_{0}} \right\|_{2} \le e^{-\frac{\lambda_{1}t}{2}} |f_{2t_{0}}(x)|^{1/2} \\ &= e^{-\frac{\lambda_{1}t}{2}} \left| P(2t_{0}, x,x)-\frac{1}{\mathrm{vol}(\M)} \right|^{1/2}. \end{align*} Thus it suffices to prove that the diagonal supremum $\sup_{x\in \M} P(2 t_{0}, x, x)$ of the heat kernel on $\M$ is finite for some $t_{0}>0$. Recall that we identify the fundamental domain $\M_{0}$ with $\M$ and $P(t,x, y)= \sum_{\gamma\in\Gamma} \wp(t, x, \gamma y)$. In order to estimate the diagonal supremum $\sup_{x\in\M} P(t, x, x)$ of the heat kernel on $\M$, we shall use the Gaussian upper bound of the heat kernel on $\widetilde{\mathcal{M}}$ (Corollary 5 in \cite{Grigoryan_1994}): there is a constant $C=C(d, b)$ such that for each $t>1$, \begin{equation}\label{Gaussbdd} \wp(t,x,y) \le C \left(\frac{\mathrm{d}_{}(x,y)^{2}}{t}\right)^{1+\frac{d}{2}} \exp \left(-\frac{\mathrm{d}_{}(x,y)^{2}}{4t}-\lambda_{0} t \right). \end{equation} Fix $x_{0}\in\M_{0}$. For a cuspidal point $\xi \in \Pi(\M_{0}):=\partial\widetilde{\mathcal{M}}\cap \overline\M_{0}$, we denote the cuspidal region of level $n$ based at $\xi$ by \begin{equation*} \HH(\xi, n):= \{y \in \M_{0}: \bb(x_{0}, y, \xi)\ge n\}. \end{equation*} Let $x_{n}=x^{\xi}_n \in \HH(\xi, n)$ be the point in the geodesic ray joining $x_{0}$ and $\xi$ with $\bb(x_{0}, x_{n}, \xi)= n$. If $\gamma$ is in the stabilizer $\Gamma_{\xi}$ of $\xi$, then $x_{0}$ and $\gamma x_{0}$ are in the horosphere of the same level based at $\xi$. This implies that for every $\gamma \in \Gamma_{\xi}$, \begin{equation} \label{Rc} e^{-b(n+1)}\mathrm{d}_{}(x_{0}, \gamma x_{0}) \le \mathrm{d}_{}(x_{n}, \gamma x_{n}) \le e^{-an} \mathrm{d}_{}(x_{0}, \gamma x_{0}). \end{equation} Applying (\ref{Rc}) to the Gaussian bound (\ref{Gaussbdd}), for each $\gamma \in \Gamma_{\xi}$, \begin{equation}\label{Parabolic} \wp(t, x_{n}, \gamma x_{n}) \le C e^{-\lambda_{0} t} \mathrm{d}_{}(x_{0},\gamma x_{0})^{d+2} \exp \left( -\frac{\mathrm{d}_{}(x_{0}, \gamma x_{0})^{2}}{4t e^{2b(d-1)(n+1)}} -a(d+2)n \right). \end{equation} We want to show that given $\delta>0$, there is $t>0$ such that for every sufficiently large $n$, \begin{equation*} \textrm{the right-hand side of (\ref{Parabolic}) } \le e^{-\delta \mathrm{d}(x_{0}, \gamma x_{0})}. \end{equation*} To simplify the notation, we put \begin{equation*} f_{n,\xi}(R) := R^{d+2} \exp \left( -\frac{R^{2}}{4t e^{ 2b(d-1)(n+1)}} +\delta R \right). \end{equation*} Since its derivative is \begin{equation*} f_{n, \xi}'(R)= R^{d+1} \left( d+2 - \frac{R^{2}}{2t} e^{-2b(d+2)(n+1)} +\delta R \right) \exp \left( -\frac{R^{2}}{4t} e^{-2b(d+2)(n+1)}+\delta R\right), \end{equation*} the positive nonzero extreme point of $f_{n,\xi}$ is $R_{n}:=t\delta e^{2b(n+1)}+\sqrt{t^{2}\delta^{2}e^{4b(n+1)}+2t(d+2)e^{2b(n+1)}}$. Thus $f_{n, \xi}$ has the maximum on $\mathbb{R}_{+}$ at $R_{n}$: \begin{align*} f_{n, \xi}(R) &\le f_{n,\xi}(R_{n})\\ &= R_{n}^{d+2} \exp\left(-\frac{R_{n}^{2}}{4t} +\delta R_{n} -a(d+2)n\right) \\ & \le \left(3t\delta e^{2b(n+1)}\right)^{d+2} \exp \left( -\frac{t\delta^{2} e^{4b(n+1)}}{2} +3t\delta^{2} e^{2b(n+1)} -a(d+2)n \right) \\ & = \left[ (3t\delta) e^{2b(n+1)-an} \right]^{d+2} e^{-\frac{9\delta^{2}t}{2}} \exp \left( -\frac{\delta^{2}t}{2} ( e^{2b(n+1)} -3)^{2} \right). \end{align*} Therefore, there is $N_{\xi}(\delta, t)$ such that if $n>N_{\xi}(\delta, t)$, then $f_{n,\xi}(R)\le C^{-1} t^{-1-\frac{d}{2}}e^{\lambda_{0} t}$, hence $\wp(t, x_{n}, \gamma x_{n}) \le e^{-\delta \mathrm{d}( x_{0}, \gamma x_{0})}$. We conclude that \begin{align}\label{cuspidal} \sum_{\gamma\in \Gamma_{\xi}} \wp(t, x_{n}, \gamma x_{n}) \le \sum_{\gamma\in \Gamma_{\xi}} e^{-\delta \mathrm{d}(x_{0}, \gamma x_{0})} = Q_{\Gamma_{\xi}, x_{0}}(\delta), \end{align} where $Q_{G, x}(\delta):=\sum_{g\in G}e^{-\delta \mathrm{d}(x, g x)}$ denotes the Poincar\'e series of a discrete group $G$ of isometries on $\widetilde{\mathcal{M}}$. We denote the abscissa of convergence of $Q_{G, x}$, which is called the \textit{critical exponent} of $G$, by $\delta_{G}$. Put $N_{\xi}:= N_{\xi}\left(\delta_{\Gamma}+1, t\right)$ and choose $N$ larger than $\max_{\xi \in \Pi(\M_{0})} N_{\xi}$. We define a truncated domain in the fundamental domain $\M_{0}$ by \begin{equation*} \M_{N}:= \M_{0} \setminus \bigcup_{\xi\in\Pi(\M_{0})} \HH\left(\xi, N \right). \end{equation*} Note that $\M_{N}$ is a pre-compact domain. Take $x_{0}\in\M_{N}$ and $x \in \HH\left(\xi, N\right)$ for some $\xi\in \Pi(\M_{0})$. Then we can replace $x$ by $x_{n}=x^{\xi}_{n}$ for some $n\ge N$: there is $n\ge N$ such that $x\in \HH(\xi, n) \setminus \HH(\xi, n+1)$ and $d(x, x_{n})$ is bounded uniformly on $n\ge N$. We may assume that given $t>0$, $g(R)= \left(\frac{R^{2}}{t}\right)^{1+\frac{d}{2}}\exp\left(-\frac{R^2}{4t}\right)$ is decreasing and $g(R)\le \exp(-(\delta_{\Gamma}+1)R)$ for every $R>0$. Assume that $x_{n}$ is on the geodesic ray $[x_{0}, \xi)$ joining $x_{0}$ and $\xi$ and $d(x_{0}, x_{n})=n$. From $\mathrm{d}(x_{N}, \gamma x_{N}) -2(n-N) \le \mathrm{d} (x_{n}, \gamma x_{n})$, writing $R_{n}=\mathrm{d}(x_{n}, \gamma x_{n})$ for $n\ge N$, there exists $C'=C'(d)>1$ such that \begin{align*} g( R_{n} ) &\le g( R_{N} -2(n-N) )\\ &= \left(\frac{(R_{N}-2(n-N))^{2}}{t}\right)^{1+d/2} \exp\left(-\frac{(R_{N}-2(n-N))^{2}}{4t}\right)\\ &\le C' g(R_{N}) g_{N}( 2(n-N) ), \end{align*} where $g_{N}(T):= \left(\frac{T}{\sqrt{t}}\right)^{d+2}\exp\left(-\frac{T^2 - R_{N}T }{4t} \right)$. By the similar computation as in (\ref{cuspidal}), \begin{equation*} g_{N}(T) \le g_{N} (T_{N}) \le \left(\frac{R_{N}}{\sqrt{t}}\right)^{d+2} \exp\left(\frac{15}{64t}R_{N}^{2}\right), \end{equation*} where $T_{N}$ the critical value of $g_{N}$. Thus we have \begin{equation*} g(R_{n}) \le C' \left(\frac{R_{N}}{t}\right)^{2d+4} \exp\left( -\frac{1}{16t}R_{N}^2 \right) \le C'' e^{-(\delta_{\Gamma}+1)\mathrm{d}(x_{N}, \gamma x_{N})}, \end{equation*} for some $C''>1$ independent of $N$. Then it follows that \begin{align*} P(t, x, x) \le & C \sum_{\gamma\in\Gamma} \left( \frac{\mathrm{d}_{}(x, \gamma x)^{2}}{t} \right)^{1+\frac{d}{2}} \exp \left( -\frac{\mathrm{d}(x,\gamma x)^{2}}{4t} -\lambda t \right)\\ \le & C Q_{\Gamma_{\xi}, x_{0}}(\delta_{\Gamma}+1) + CC'' \sum_{\gamma\notin\Gamma_{\xi}} e^{-(\delta_{\Gamma}+1)\mathrm{d}(x_{N}, \gamma x_{N})}\\ \le & C(1+C'') \max \left\{Q_{\Gamma, x_{0}}(\delta_{\Gamma}+1), Q_{\Gamma, x_{N}^{\xi}}(\delta_{\Gamma})\right\}. \end{align*} Hence we have $\sup_{x\in \HH(\xi, N)} P(t, x, x,)<\infty$ for every $\xi \in \Pi(\M_{0})$. Therefore, since $\Pi(\M_{0})$ is a finite set, $\sup_{x\in\M}P(t, x, x) <\infty$. \end{proof} We are ready to verify the exponential decay of uniform norm and complete the proof of Theorem \ref{Ctr}. It is enough to show that the exponential decay of the supremum norm since we have already proved the exponential decay of H\"older norm in Proposition \ref{CtrH}. \begin{prop} There exists a constant $C_{2}>0$ such that for every $f\in\LL$, $t>0$ \begin{equation*} \| \Q^{t}f -\N f\|_{\infty} \le \|f\|_{\L_{\tau}} e^{-C_{2}t}. \end{equation*} \end{prop} \begin{proof} Denote $F_{t}(x):=\int \Q^{t}f(x,\xi) d\nu_{x}(\xi)$. \begin{align*} \left| \Q^{t}f(x,\xi) - \int f d\hh\right| &= \left| \Q_{t}f(x,\xi) -\int \Q^{\frac{t}{2}}f d\hh \right|\\ &\le \left| \Q^{t}f(x,\xi) - \Q^{\frac{t}{2}}F_{\frac{t}{2}}(x) \right| + \left| \Q^{\frac{t}{2}}F_{\frac{t}{2}}(x)-\int \Q^{\frac{t}{2}}f d\hh \right|\\ &\le \left| \Q^{\frac{t}{2}}\left(\Q^{\frac{t}{2}}f(x,\xi) - F_{\frac{t}{2}}(x)\right) \right| + \left| \Q^{\frac{t}{2}}F_{\frac{t}{2}}(x)-\int \Q^{\frac{t}{2}}f d\hh \right|. \end{align*} By Lemma \ref{Doeblin}, the last term of the last inequality decays exponentially: \begin{align*} \left| \Q^{\frac{t}{2}}F_{\frac{t}{2}}(x) -\int \Q^{\frac{t}{2}}f d\hh \right| &= \left| \int_{\M_{0}} P(t/2, x, y) F_{\frac{t}{2}}(y) d \mathrm{vol}(y) - \int_{\M_{0}} F_{\frac{t}{2}}(y) d\widetilde{m}(y) \right| \\ &\le \|F_{\frac{t}{2}}\|_{\infty} \int_{\M_{0}} \left| P(t/2, x, y) - \frac{1}{\mathrm{vol}(\M_{0})} \right| d\mathrm{vol}(y) \\ &\le \|f\|_{\tau} e^{-\frac{\lambda_{1}t}{4}}. \end{align*} For the first term, it follows from Proposition \ref{CtrH} that \begin{align*} \left| \Q^{\frac{t}{2}} \left( \Q^{\frac{t}{2}}f(x,\xi)-F_{\frac{t}{2}}(x,\xi) \right) \right| &\le \sup_{y\in \M_{0}} \left| \Q^{\frac{t}{2}}f(y,\xi)-F_{\frac{t}{2}}(y,\xi) \right|\\ &\le \sup_{y\in\M_{0}} \int \left| \Q^{\frac{t}{2}}f(y, \xi)-\Q^{\frac{t}{2}}f(y, \eta) \right| d \nu_{y}(\eta)\\ &\le \|f\|_{\tau}e^{-\frac{C_{1}t}{2}}. \end{align*} \end{proof} \section{Ergodic properties of Brownian motions} In this section, we discuss the thermodynamic formalisms for the harmonic potential, which arises from the Brownian motion and an equidistribution theorem of Brownian paths. Using such ergodic properties of the Brownian motion, we also provide a characterization of the asymptotic harmonicity as an application of the central limit theorem to the ergodic theory of the geodesic flow on $\M$. \subsection{Harmonic potentials} We introduce another natural potential $\K$ on $\T^{1}\M$ induced from the Brownian motion, which we call the \textit{harmonic potential}. Define a function $\widetilde{\K}$ on $\T^{1}\widetilde{\mathcal{M}}$ by \begin{equation*} \widetilde{\K}(\mathrm{v})=-\left.\frac{d}{dt}\right|_{t=0} \log \kk(\gamma_\mathrm{v}(0), \gamma_{\mathrm{v}}(t), \mathrm{v}_+) \end{equation*} where $\mathrm{v}_{+}$ denote the end point at infinity $\lim_{t\to\infty}\gamma_{\mathrm{v}}(t)$ of the geodesic $\gamma_{\mathrm{v}}$ generated by $\mathrm{v}$ and $\kk(x,y, \xi)$ is the Martin kernel of the Brownian motion on $\widetilde{\mathcal{M}}$. Since $\widetilde\K$ is a $\Gamma$-invariant H\"older continuous function on $\T^{1}\widetilde{\mathcal{M}}$ (\cite{Hamenst_dt_1990}), it induces a H\"older potential, which is denoted by $\K$, on $\T^{1}\M$. Note that $\K$ has the harmonic measure $(\nu_{x})_{x\in\widetilde{\mathcal{M}}}$ as a Patterson-Sullivan density of dimension $0$. Since the harmonic measure does not have atom (\cite{KifLed_1990}, \cite{BenHul_2019}), the set $\Pi_{\Gamma}$ of parabolic fixed points on $\Gamma$ in $\partial\widetilde{\mathcal{M}}$ has countably many points, $\Pi_{\Gamma}$ is a null set for the harmonic measure. As the set of conical fixed points $\Lambda_{c}\Gamma=\partial\widetilde{\mathcal{M}} \setminus \Pi_{\Gamma}$ has positive measure with respect to the harmonic measure, the topological pressure of $\K$ vanishes; $P_{\K}=0$ (Corollary 5.10 of \cite{PPS}). We denote by $\widetilde\nu$ the Gibbs measure on $\T^{1}\widetilde{\mathcal{M}}$ of $\K$ and $(\nu_{x})$. Proposition \ref{VP} for $\K$ demonstrates that $\K$ admits an equilibrium state $\nu$ on $\T^{1}\M$ for $\K$ if and only if $\widetilde\nu \left(\T^{1}\M_{0}\right)$ is finite and $\nu$ agrees with the induced measure on $\T^{1}\M$ by $\widetilde\nu$. From Proposition \ref{SPR} it follows that $\K$ admits an equilibrium state if and only if for every parabolic subgroup $\Pi$ of $\Gamma$, \begin{equation*} \sum_{\gamma\in\Pi} \frac {\mathrm{d}_{}(x, \gamma x)} {\kk(x, \gamma x, (\mathrm{v}_{x}^{\gamma x})_{+})} <\infty, \end{equation*} where $\mathrm{v}_{x}^{y}\in\T_{x}^{1}\widetilde{\mathcal{M}}$ such that $\g^{\mathrm{d}(x,y)}\mathrm{v}_{x}^{y}\in\T^{1}_{y}\widetilde{\mathcal{M}}$. We shall provide dynamical aspects of Brownian motions using the ergodic theory of $\nu$. Recall that given $x\in\widetilde{\mathcal{M}}$ we identify $(r, \mathrm{v})\in(0,\infty)\times\T^{1}_{x}\widetilde{\mathcal{M}}$ with $\exp_{x}(r\mathrm{v})\in\widetilde{\mathcal{M}}\setminus\{x\}$ and $g=dr^{2}+\lambda_{x}(r, \mathrm{v})g_{\mathbb{S}}$. Now we denote the density of volume at $z=(r, \mathrm{v})$ with respect to the polar coordinate at $x$ by $A_{x}(z)$: \begin{equation*} d \mathrm{vol} (z) = A_{x}(z) dr d\mathrm{vol}_{\mathbb{S}}(\mathrm{v}). \end{equation*} Note that $A_{x}(z)= \lambda_{x}^{d-1}(r, \mathrm{v})$. We denote by $\theta(\widetilde\omega, t)$ the unit vector in $\T^{1}_{x}\widetilde{\mathcal{M}}$ such that $\widetilde\omega_{t} =(r, \theta)(\widetilde\omega, t) :=(r(\widetilde\omega, t), \theta(\widetilde\omega, t))$. The following theorem demonstrates how dynamical invariants and stochastic invariants are related to each other. We follow the argument in \cite{Led_1988}, but we complete the proof by showing the inequality $\hs \le \ld h_{\nu}$ using the idea in \cite{Led_1987} \begin{thm}\label{SE} If $\K$ admits an equilibrium state $\nu$, then \begin{equation*} \hs = \ld h_{\nu}. \end{equation*} \end{thm} \begin{proof} Let $x\in\widetilde{\mathcal{M}}$, $\delta\in \left( 0, \frac{1}{2} \right)$ and $0<\varepsilon, \varepsilon'$. We denote for each $T>0$, \begin{align*} \mathscr{C}_{T} := \{\widetilde\omega: \ & \mathrm{d}(\widetilde\omega_{T}, (\ld T, \widetilde\omega_{\infty}))\le \varepsilon T \textrm{ and }\\ & \mu^{\T}_{x} \{ \mathrm{v}: d_{\ld T}(\mathrm{v}, \theta(\widetilde\omega, \infty))\le \varepsilon' \} \le e^{-(\ld h_{\nu}-\varepsilon)T} \}, \\ \mathscr{D}_{T} := \{\widetilde\omega: & \mathrm{d}(\widetilde\omega_{T}, (\ld T, \widetilde\omega_{\infty}))\le \varepsilon T \textrm{ and }\\ & \mu^{\T}_{x} \{ \mathrm{v}: \mathrm{d} \left(\gamma_{\mathrm{v}}(\ld T), \gamma_{\theta(\widetilde\omega, \infty)}(\ld T)\right)\le \varepsilon' \} \ge e^{-(\ld h_{\nu}+\varepsilon)T} \}. \end{align*} For every $T$ large enough, $\mathbb{P}_{x}(\mathscr{C}_{T})\ge 2\delta$ for some $\varepsilon'>0$ by (\ref{roughpath}). Thus if we fix a sufficiently large $T$ and choose $E\subset\widetilde{\mathcal{M}}$ with $\mathrm{Card} E = N(x, T, 1-\delta)$, \begin{equation*} \mathbb{P}_{x}\{ \mathrm{d}(\widetilde\omega_{T}, E)\le 1\}\ge 1-\delta. \end{equation*} We note that $E_{\infty}:= \{ \theta(\widetilde\omega,\infty): \widetilde\omega\in \mathscr{C}_{T}, \mathrm{d}( \widetilde\omega_{T}, E)\le 1 \}$ has the $\mu^{\T}_{x}$-measure greater than $\delta$ and $\{\gamma_{\theta(\widetilde\omega, \infty)}(\ld T): \widetilde\omega\in \mathscr{C}_{T}, \mathrm{d}( \widetilde\omega_{T}, E)\le 1 \}$ is covered by balls on the sphere of radius $\varepsilon'$ less than $N(x, T, 1-\delta) C^{\varepsilon T}$. ($C$ is the maximal cardinal of covers for the intersection of the sphere of radius $\ld T$ and $(\varepsilon+1) T$ balls by $\varepsilon'$ balls on the sphere.) Such ball $O$ in the sphere of radius $\varepsilon'$ is the set of base points of vectors in $\g^{\ld T}V$ where $V=\{\mathrm{v}: d_{\ld T} (\mathrm{v}, \mathrm{w})\le \varepsilon'\}$ for some $\mathrm{w}$. We conclude that since such $V$ has the $\mu_{\T^{1}\M}$-measure less than $e^{-(\ld h_{\nu}-\varepsilon)T}$, \begin{align*} \delta \le \mu^{\T}_{x} (E_{\infty}) \le N(x, T, 1-\delta) e^{ -T[\ld h_{\nu} -\varepsilon-\varepsilon\log C]}. \end{align*} Thus we have $\ld h_{\nu} \le \lim_{T\to\infty} \frac{1}{T} \log N(x, T, 1-\delta)$. Choose a smallest set $E\subset \widetilde{\mathcal{M}}$ such that $\mathrm{d}(\widetilde\omega_{T}, E)\le 1$ for each $\widetilde\omega\in\mathscr{D}_{T}$ and a maximal $\varepsilon'$-separeted set $F\subset \{\gamma_{\theta(\widetilde\omega, \infty)}(\ld T): \widetilde\omega \in \mathscr{D}_{T}\}$. Since $\mathscr{D}_{T}\subset\{ \widetilde\omega: \mathrm{d}(\widetilde\omega_{T}, E)\le 1\}$, $\mathrm{Card} (E) \ge N(x, T, \mathbb{P}_{x}(\mathscr{D}_{T}))$ and $\mathrm{Card} (F) \le C' e^{\ld h_{\nu}T+\varepsilon T}$. ($C'$ is the maximal number of overlappings.) For every $f\in F$ if we denote \begin{equation*} N(f):=\{ e \in E: \exists \, \widetilde\omega \in \mathscr{D}_{T} \textrm{ s.t. } \mathrm{d}(f, \gamma_{\theta(\widetilde\omega, \infty)}(\ld T))\le \varepsilon', \mathrm{d}(e, \widetilde\omega_{T}) \le 1\}. \end{equation*} Since $\cup_{e\in N(f)} B(e, 1) \subset B(f, \varepsilon T +\varepsilon'+1)$, there exists $C''>0$ such that \begin{equation*} \mathrm{Card}{N(f)} \le \sup_{e\in E, f \in F} \frac{\mathrm{vol}\left( B(f, \varepsilon T+\varepsilon'+1)\right)}{\mathrm{vol}( B(e, 1))} \le e^{C''\varepsilon T}. \end{equation*} Therefore, \begin{equation*} N(x, T, \mathbb{P}_{x}(\mathscr{D}_{T}) )\le \mathrm{Card} (E) \le \exp(C''\varepsilon T)\mathrm{Card} F\le e^{T[\ld h_{\nu} + (1+ C'')\varepsilon ]}. \end{equation*} \end{proof} The following proposition means that Brownian paths are equidistributed with respect to $\nu$; Geodesics which Brownian paths roughly follow are generic with respect to $\nu$. The proof follows the argument for compact manifolds (\cite{Led_1988}). \begin{prop}\label{eqd} Assume that $\K$ admits an equilibrium state $\nu$. For every $x\in\widetilde{\mathcal{M}}$, for each bounded continuous function $\phi \in \C_{b}(\T^{1}\M)$ and for $\mathbb{P}_{x}$-a.e. $\widetilde\omega$, \begin{equation*} \int \phi \, d\nu = \lim_{t\to\infty} \frac{1}{\ld t} \int_{0}^{r(\widetilde\omega, t)} \widetilde{\phi}( \g^{s}\theta(\widetilde\omega, t)) \,ds. \end{equation*} \end{prop} \begin{proof} For $\mathrm{v}, \mathrm{w}\in \T^{1}\widetilde{\mathcal{M}}$, let $d_{t}(\mathrm{v}, \mathrm{w})$ be the distance on the geodesic sphere $S(x, t)$ between $\g^{t} \mathrm{v}$ and $\g^{t} \mathrm{w}$. Then \begin{equation*} d_{t}(\mathrm{v}, \mathrm{w})\le d_{s}(\mathrm{v}, \mathrm{w})\frac{\sinh(at)}{\sinh(as)} \end{equation*} for every $0<t<s$ due to the curvature upper bound $\sec_{\widetilde{\mathcal{M}}}\le -a^{2}<0$. Since the Sasaki distance is H\"older equivalent to the distance $\mathrm{d}_{0}(\mathrm{v}, \mathrm{w}):= \sup_{0\le t\le1}\mathrm{d}\left( \gamma_{\mathrm{v}}(t), \gamma_{\mathrm{w}}(t)\right)$, \begin{align*} \left| \int_{0}^{t} \widetilde\phi\circ\g^{s}(\mathrm{v}) ds - \int_{0}^{t} \widetilde\phi\circ\g^{s}(\mathrm{w}) ds \right| \le C(a, \phi) \mathrm{d}( \gamma_{\mathrm{v}}(t), \gamma_{\mathrm{w}}(t) ). \end{align*} Hence the proposition follows from of Proposition \ref{transversalerg} and the limit (\ref{roughpath}): for $\mathbb{P}_{x}$-a.e. $\widetilde\omega$, \begin{align*} &\lim_{t\to\infty} \left| \frac{1}{\ld t} \int_{0}^{r(\widetilde\omega, t)} \widetilde\phi(\g^{s}\theta(\widetilde\omega, t)) ds -\int\phi d\nu \right|\\ &\le \lim_{t\to\infty} \frac{1}{\ld t} \left| \int_{0}^{r(\widetilde\omega, t)} \widetilde\phi(\g^{s}\theta(\widetilde\omega, t)) ds -\int_{0}^{\ld t} \widetilde\phi(\g^{s}\theta(\widetilde\omega, \infty)) ds \right| + \lim_{t\to\infty} \left| \frac{1}{\ld t} \int_{0}^{\ld t} \widetilde\phi(\g^{s}\theta(\widetilde\omega, \infty)) ds - \int\phi d\nu \right|\\ &= \lim_{t\to\infty} \frac{1}{\ld t} \left| \int_{0}^{r(\widetilde\omega, t)} \widetilde\phi(\g^{s}\theta(\widetilde\omega, t)) ds - \int_{0}^{r(\widetilde\omega, t)} \widetilde\phi(\g^{s}\theta(\widetilde\omega, \infty)) ds \right| +0 \\ &\le \lim_{t\to\infty} \frac{C(a, \phi)}{\ld t} \mathrm{d} (\widetilde\omega_{t}, (r, \theta)(\widetilde\omega, t)) =0. \end{align*} We used (\ref{transversalpterg}) of Proposition \ref{transversalerg} in the equation and (\ref{roughpath}) in the last inequality. \end{proof} The equidistribution of Brownian paths provides another stochastic invariant, the exponential growth along Brownian paths. It helps understanding the relation between the harmonic measure class and the Lebesgue measure class. The proof in \cite{Led_1988} extends to the finite-volume case. \begin{thm} For each $x\in\widetilde{\mathcal{M}}$ and for $\mathbb{P}_{x,}$-a.e. $\widetilde\omega$, if $\K$ admits an equilibrium state $\nu$, the following limit exists: \begin{equation*} \Upsilon=\lim_{t\to\infty} \frac{1}{t}\log {A}(x, \widetilde\omega_{t})=-\ld \int F^{su} d\nu. \end{equation*} \end{thm} \begin{proof} Let $\T_{\mathrm{v}}\g^{t}$ be the tangent map of the flow map $\g^{t}$ at $\mathrm{v}=(x, \xi) \in \T^{1}\widetilde{\mathcal{M}}$. Since the angle between stable distribution $E^s(\mathrm{v})$ and $\T_{\mathrm{v}}\g^{t}(\T_{\mathrm{v}}\T^{1}_{x}\widetilde{\mathcal{M}})$, where $\T_{\mathrm{v}}\T^{1}_{x}\widetilde{\mathcal{M}}$ is the tangent space of the sphere $\T^{1}_{x}\widetilde{\mathcal{M}}$, is bounded away from zero uniformly on $\mathrm{v}$ and $t>0$, \begin{align*} \lim_{t\to\infty} \frac{1}{t} \log A(\pi\mathrm{v}, \pi\g^{t}\mathrm{v}) &=\lim_{t\to\infty} \frac{1}{t} \log \det \T\g^{t}|_{\T_{\mathrm{v}}\T^{1}_{x}\widetilde{\mathcal{M}}}\\ &=\lim_{t\to\infty} \frac{1}{t} \log \det \T\g^{t}|_{E^{uu}(\mathrm{v})}\\ &=\lim_{t\to\infty}-\frac{1}{t}\int_{0}^{t} F^{su}(\g^{s} \mathrm{v}) ds. \end{align*} Therefore, by Proposition \ref{eqd}, for $\mathbb{P}_{x}$-a.e. $\widetilde\omega$, \begin{align*} \lim_{t\to\infty} \frac{1}{t} \log A(x, \widetilde\omega_{t}) = \lim_{t\to\infty} -\frac{1}{t} \int_{0}^{r(\widetilde\omega, t)} F^{su}(\g^{s} \theta(\widetilde\omega_{t})) ds = -\ld \int F^{su} d\nu. \end{align*} \end{proof} Since $h_{\nu} \le \htop$ and $h_{\nu}+\int F^{su} d\nu \le P_{F^{su}}=0$, from the previous theorems, we have the following theorem as a corollary. \begin{thm} Denote the topological entropy of $(\T^{1}\M, (\g^{t}))$ by $\htop$. \begin{enumerate} \item $\hs \le \ld \htop$. The equality holds if and only if the harmonic measure class and the visibility class coincide. \item $\hs \le \Upsilon = -\ld \int F^{su} \,d\nu$. The equality holds if and only if the harmonic measure class and the Lebesgue class agree. \end{enumerate} \end{thm} \begin{proof} Due to Theorem \ref{SE}, the equality $\Upsilon=\hs$ is equivalant to \begin{equation*} P(F^{su}, \nu)=h_{\nu}+\int F^{su} d\nu=0, \end{equation*} which holds if and only if $\nu$ is the equilibrium state for $F^{su}$. \end{proof} \subsection{Proof of Theorem \ref{AsympH}} We conclude this section with the proof of Theorem \ref{AsympH}. We begin with the proof of the integral equation for the foliated Laplacian (\cite{Yue_1991}): for every bounded function $\varphi$ uniformly $\C^{2}$ on stable leaves, \begin{equation} \label{yue} \int 2 \langle \nabla \log \kk, \nabla \varphi \rangle d\hh = -\int \Delta_{s} \varphi d\hh. \end{equation} Consider the function $\Phi(y) := \int_{\partial\widetilde{\mathcal{M}}} \varphi(y, \xi) d\nu_{y}(\xi) = \int_{\partial\widetilde{\mathcal{M}}} \varphi(y, \xi) \kk(x, y, \xi) d\nu_{x}(\xi)$. Applying the Laplacian, since $\Delta_{y} \kk(x, y, \xi)=0$ we have \begin{align*} \Delta \Phi (y) &= \int_{\partial\widetilde{\mathcal{M}}} \kk(x, y, \xi) \Delta_{s} \varphi (y, \xi) + 2 \langle \nabla_{y} \varphi(y, \xi) , \nabla_{y} \kk(x, y, \xi) \rangle d\nu_{x}(\xi)\\ &= \int_{\partial\widetilde{\mathcal{M}}} \Delta_{s}\varphi(y, \xi) + 2 \langle \nabla_{y} \varphi(y, \xi), \nabla_{y} \log \kk (x, y, \xi) \rangle d\nu_{y}(\xi). \end{align*} Thus integrating with respect to $\mathrm{vol}$ and using Green's formula, since $\Phi$ is uniformly $\C^{2}$, \begin{align*} &\int_{\T^{1}\M} \Delta_{s}\varphi (y, \xi) + 2 \langle \nabla_{y} \varphi(y, \xi) , \nabla_{y} \log \kk(x, y, \xi) \rangle d\hh (y, \xi)\\ &= \int_{\M} \Delta\Phi(x) d\mathrm{vol}(x)\\ &= \lim_{\varepsilon \to 0} \int_{\M_{\varepsilon}}\Delta\Phi(x) d\mathrm{vol}(x)\\ &=\lim_{\varepsilon \to 0} \int_{\partial \M_{\varepsilon}} \langle \nabla \Phi , \textbf{n}_{\varepsilon}\rangle =0, \end{align*} where $\M_{\varepsilon}=\{ x\in \M: inj(x) \ge \varepsilon\}$ and $\textbf{n}_{\varepsilon}$ is the unit normal vector on $\partial \M_{\epsilon}$. From the integral formula (\ref{yue}) for the foliated Laplacian, it follows that \begin{equation*} \sigma_{\kk}^{2} = 2\int_{\T^{1}\M} \|\nabla \log \kk(x, \cdot, \xi) + \nabla u_{\kk}\|^{2} d\hh = 2\int \| \nabla \log \kk\|^{2}+\| \nabla u_{\kk}\|^{2} d\hh, \end{equation*} since $\int \Delta_{s} u_{\kk} d\hh=0$. Since $\int \| \nabla \log \kk \|^{2} d\hh= \hs$ (Proposition \ref{intcha}), \begin{equation*} \sigma_{\kk}^{2}-2\hs=2\int \| \nabla u_{\kk}\|^{2} d\hh \ge 0, \end{equation*} and the equality holds if and only if $u_{\kk}$ is constant. Since $u_{\kk}$ is the solution of the leafwise heat equation $\Delta_{s}u=\|\nabla\log \kk\|^{2}-\hs$, $u_{\kk}$ is constant if and only if $\|\nabla \log \kk\|$ is constant and equal to $\hs$. First we verify that $\|\nabla\log\kk\|^{2}=\hs$ implies the asymptotic harmonicity of $\M$. \begin{align*} \hs &\le \ld \htop \le \sup_{\mu}-\ld \int \K d\mu \\ &=\sup_{\mu} \ld \int \langle X, \nabla\log\kk \rangle d\mu \le \sup_{\mu} \ld \left|\int \|\nabla\log\kk\|^{2} d\mu\right|^{1/2} = \ld\sqrt{\hs}\le \hs, \end{align*} where $X(x, \xi) := (x, \xi)$ and the supremum taken among invariant measures. Hence we have equalities and $\nu$ is the measure of maximal entropy. Replacing the supremum of integrations by integrations with respect to $\nu$, we have \begin{equation*} \int \langle X, \nabla\log\kk \rangle d\mu = \left|\int \|\nabla\log\kk\|^{2} d\mu\right|^{1/2}, \end{equation*} which occurs if and only if $X=\nabla\log\kk$. Therefore the mean curvature of the stable horosphere $\mathrm{div}X=\Delta \log \kk = -\|\nabla \log \kk\|^{2}= -h$ is constant. Conversely, if $\M$ is asymptotically harmonic, the geometric potential $F^{su}=\mathrm{div}X$ is constant. Since $P_{F^{su}}=0$, the Liouville measure is the measure of maximal entropy and $F^{su}=\htop$. For $x, y\in\widetilde{\mathcal{M}}$ and $\xi\in\partial\widetilde{\mathcal{M}}$, if we write \begin{equation*} \Psi_{x,\xi}(y) = \exp (-\htop\bb(y, x, \xi)), \end{equation*} then we have $\Delta \Psi_{x,\xi}=0$ and $\lim_{y\to \eta} \Psi_{x, \xi}=0$ if $\eta \ne \xi$ and $\infty$ if $\eta=\xi$. Hence $\kk(x, y, \xi) = \Psi_{x,\xi}(y)$ and $\| \nabla \log \kk\|^{2}=\htop^2$. It follows that the harmonic class and the visibility class coincide, which implies $\hs =\ld \htop$. Since $\ld = -\int \mathrm{div} X dm^{\Q}=\htop$, $\hs= \htop^{2}= \|\nabla \log \kk\|^{2}$. Therefore we have $\sigma_{\kk}^{2}=2\hs$. This completes the proof of Theorem \ref{AsympH}.

https://arxiv.org/abs/1912.10486

A Polynomial Time Algorithm for the $k$-Disjoint Shortest Paths Problem

The disjoint paths problem is a fundamental problem in algorithmic graph theory and combinatorial optimization. For a given graph $G$ and a set of $k$ pairs of terminals in $G$, it asks for the existence of $k$ vertex-disjoint paths connecting each pair of terminals. The proof of Robertson and Seymour [JCTB 1995] of the existence of an $n^3$ algorithm for any fixed $k$ is one of the highlights of their Graph Minors project. In this paper, we focus on the version of the problem where all the paths are required to be shortest paths. This problem, called the disjoint shortest paths problem, was introduced by Eilam-Tzoreff [DAM 1998] where she proved that the case $k = 2$ admits a polynomial time algorithm. This problem has received some attention lately, especially since the proof of the existence of a polynomial time algorithm in the directed case when $k = 2$ by Bérczi and Kobayashi [ESA 2017]. However, the existence of a polynomial algorithm when $k = 3$ in the undirected version remained open since 1998. In this paper we show that for any fixed $k$, the disjoint shortest paths problem admits a polynomial time algorithm. In fact for any fixed $C$, the algorithm can be extended to treat the case where each path connecting the pair $(s,t)$ has length at most $d(s,t) + C$.

\section{Introduction} Given a graph $G$ and a set of pairs of vertices $(s_1, t_1), \dots, (s_k, t_k)$, the \textit{Vertex-Disjoint Paths Problem} asks whether there exists a set of vertex-disjoint paths $P_1, \dots, P_k$ such that every $P_i$ is an $(s_i,t_i)$ path. This is a classical NP-hard \cite{karp1975} problem in graph theory with applications to VLSI-layouts and networks problem \cite{schrijver1990paths} which has been extensively studied. The proof of the existence of a $O(n^3)$ algorithm for any fixed $k$ by Robertson and Seymour \cite{Seymour95} is one of the highlights from the Graph Minors project, and the running time has later been improved to $O(n^2)$ \cite{KAWARABAYASHI2012424}. In the directed case, the problem is NP-hard even for $k =2$ \cite{FORTUNE1980111}, but some results are known for special classes of digraphs like acyclic digraphs \cite{FORTUNE1980111}, planar digraphs \cite{schrijver1994finding} or tournaments \cite{CHUDNOVSKY2015582}. One natural question is, given an instance of the Vertex-Disjoint Path Problem, to find a solution which minimises the sum of the lengths of the $P_i$. This problem appears to be much harder, as only the case $k = 2$ was recently solved by Björklund and Husfeldt \cite{Bjorklund}. In fact, even deciding if the problem admits an optimal solution, i.e where every $P_i$ is a shortest path between $s_i$ and $t_i$ is open for $k \geq 3$. This problem was first considered by Eilam-Tzoreff \cite{EILAMTZOREFF1998113} 20 years ago. In the same paper, she gave an algorithm for the case $k=2$ and conjectured that a polynomial algorithm exists for any fixed $k$, both in the directed and undirected setting. This problem has received some attention lately. In particular Bérczi and Kobayashi \cite{berczi2017directed} gave a nice proof that the directed version admits a polynomial time algorithm when $k = 2$. They also show that the algorithm of Schrijver \cite{schrijver1994finding} can be used to prove Eilam-Tzoreff's conjecture when the input (di)graph is planar. The goal of this paper is to solve this problem for any $k$ in the undirected case. \begin{theorem}\label{thm:SDP} For any fixed integer $k$, there exists an algorithm running in $n^{O(k^{5^k})}$ time that decides, given a graph $G$ and $k$ pairs of vertices $(s_1, t_1), \dots (s_k, t_k)$, of the existence of $k$ internally vertex-disjoint paths $P_1, \dots, P_k$ such that each $P_i$ is a shortest $(s_i, t_i)$-path. \end{theorem} We also show that this problem is $W[1]$-hard parameterized by $k$, which means that we cannot hope to remove the dependency in $k$ in the exponent. However, we can extend the previous result to the case where paths are not required to be shortest paths, but of length at most $d(s_i, t_i) + C$, where $C$ is a fixed constant. \begin{corollary}\label{thm:SDPC} For any fixed integers $k$ and $C$ , there exists an algorithm running in $n^{O((Ck)^{5^k})}$ that decides, given a graph $G$ and $k$ pairs of vertices $(s_1, t_1), \dots (s_k, t_k)$, of the existence of $k$ internally vertex-disjoint paths $P_1, \dots, P_k$ such that each $P_i$ is a path of length at most $d(s_i, t_i) +C$ between $s_i$ and $t_i$. \end{corollary} \section{Preliminaries} For any integer $k$, $[k]$ denotes the set of integers between $1$ and $k$, and for any integer $j \leq k$, $[j..k]$ denotes the set of integers between $j$ and $k$. The \textit{length} of a path $P$ correspond to the number of edges in the path, and given two vertices $x$ and $y$ in the same connected component, $d(x,y)$ denotes the minimal length of a path between $x$ and $y$. A graph $G$ is said to be a \textit{$k$-shortest} graph if there exists $k$ partitions of $G$ $(V^1_1, \dots, V^1_{l_1})$, $(V^2_1, \dots, V^2_{l_2})$, $\dots, (V^k_1, \dots, V^k_{l_k})$ such that $xy$ is an edge of $G$ implies that there exists $i \in [k]$ and $j \in [l_i - 1]$ such that $x \in V^i_{j}$ and $y \in V^i_{j+1}$. Moreover, if $x \in V^i_j$ and $y \in V^i_l$, for $i \in [k]$, $j,l \in [l_i]$ and $|j-l| > 1$ then $xy \not \in E(G)$. Intuitively, one way to obtain a $k$-shortest graph is to start from any graph, doing $k$ breadth-first searches from different vertices and removing all the edges which are not between two consecutive levels of at least one of the BFS. For a $k$-shortest graph $G$, we associate naturally $k$ colours to the edges of $G$ as follows: any edge between $x\in V^i_{j}$ and $y \in V^i_{j+1}$ is said to be of colour $i$. Note that the same edge can be of different colours. Moreover, each colour $i$ defines a partial order on the vertices of $G$ $\leq_i$ as follows: $x \leq_i y$ if $x \in V^i_j$ and $y \in V^i_r$ with $j \leq r$. This naturally defines an orientation of the edges of colour $i$. Note that the same edge can have two different orientations for two different colours. Let $G$ be a $k$-shortest graph and $r$ and $i$ be two indices. We say that a path $P_i = x_1, \dots, x_r$ is a \textit{path of colour $i$} if for every $j \in [r-1]$, $x_jx_{j+1}$ is an edge of colour $i$ and $x_j \leq_i x_{j+1}$. Note that, since whenever $u \in V^i_j$ and $v \in V^i_l$ with $|j-l|>1$ there is no edge between $u$ and $v$, any path of colour $i$ between $x$ and $y$ is also a shortest path in $G$. Moreover, concatenating two paths of colour $i$ also gives a path of colour $i$. By convention, we consider the paths of colour $i$ to be oriented from the endpoint which belongs to the part of $(V^i_1, \dots, V^i_{l_i})$ with the lowest index to the endpoint with the largest one. In particular, an \textit{$(x,y)$-path of colour $i$} is a path of colour $i$ between $x$ and $y$ oriented from $x$ to $y$. For a directed path $P$ and two vertices $x$ and $y$ belonging to this path, $P[x,y]$ denotes the subpath of $P$ from $x$ to $y$. By convention, if $y$ is before $x$ along $P$, then $P[x,y]$ will be the empty path. The \textit{length} of a path is its number of edges. A \textit{path-partition} of a path $P$ is a set of internally vertex-disjoint subpaths of $P$ such that concatenation of all the paths gives $P$. Let $Q_1, Q_2$ be two different path partitions of the same path $P$, the \textit{intersection} of $Q_1$ and $Q_2$ is the path partition of $P$ obtained as follows: If $S$ is the set of vertices which are endpoints of paths of either $Q_1$ or $Q_2$, then the intersection of $Q_1$ and $Q_2$ consists of all the subpaths of $P$ between vertices of $S$ which are consecutive along $P$. Note that the number of paths in the intersection of $Q_1$ and $Q_2$ is at most the sum of the number of paths in $Q_1$ and $Q_2$. Moreover, every path in the intersection is a subpath of some path in $Q_1$ and some path in $Q_2$. For an oriented edge $e=xy$, $x$ is called the \textit{tail} of $e$ and is denoted as $t(e)$ and $y$ the \textit{head}, denoted as $h(e)$. Let $G$ be a $k$-shortest graph, $(s_1, t_1), \dots (s_l, t_l)$ a set of $l$ pairs of vertices and $c$ a function from $[l]$ to $[k]$, the \textit{$k$-DSP} problem defined by $G$, the $(s_i, t_i)$ and $c$ is the problem of finding a set of internally vertex-disjoint paths $P_1, \dots, P_l$ such that for any $i \in [l]$, $P_i$ is a path of colour $c(i)$ between $s_i$ and $t_i$. The $(s_i,t_i)$ will be referred to as \textit{requests}. The following lemma shows that we can reduce Eilam-Tzoreff's question to solving an instance of $k$-DSP. \begin{lemma}\label{lem:from_DSP_to_shortest_graph} Let $G$ be a graph and $(s_1, t_1), \dots, (s_k, t_k)$ be a set of $k$ pairs of vertices in $G$. Let $G'$ be the $k$-shortest graph obtained from $G$ by taking for each $i$ $(V^i_1, \dots, V^i_{l_i})$ the partition obtained by doing a breath-first-search from $s_i$, and removing the edges which are not between two consecutive levels of some BFS. There exists a set of internally vertex-disjoint paths $P_1, \dots, P_k$ such that for every $i \in [k]$, $P_i$ is a shortest path between $s_i$ and $t_i$ if and only if the $k$-DSP problem defined by $G'$, the $(s_i, t_i)$ and the identity function $c:[k] \rightarrow [k]$ has a solution. \end{lemma} \begin{proof} The proof follows from the fact that, for every $i$ and $l$, the set $V^i_l$ corresponds to the set of vertices at distance $l$ from $s_i$. Therefore, a shortest path in $G$ between $s_i$ and some vertex $ x\in V^i_l$ is a path of colour $i$ from $s_i$ to $x$ in $G'$ and vice versa. \end{proof} The main contribution of this paper is to prove the following result, which implies Theorem \ref{thm:SDP} \begin{theorem}\label{thm:main} Let $G$ be a $k$-shortest graph, $(s_1, t_1), \dots, (s_l, t_l)$ a set of $l$ pairs of vertices and $c$ a function from $[l]$ to $[k]$. There exists an algorithm running in time $n^{O(l^{5^k})}$ deciding if the problem of $k$-DSP defined by $G$, the $(s_i,t_i)$ and $c$ has a solution. \end{theorem} In fact, by noting that a path of length at most $d(s_i,t_i) + C$ between $s_i$ and $t_i$ uses at most $C$ edges which are not edges between consecutive levels of the BFS starting in $s_i$, trying all $\binom{n^2}{kC}$ choices for these edges allows us to reduce the problem of Corollary \ref{thm:SDPC} to a $k$-DSP problem with at most $k \cdot C$ pairs. An interesting case is when $k=1$. The problem then reduces to the problem of directed disjoint paths in acyclic digraphs by orienting all the edges of $G$ from each set $V_i$ to $V_{i+1}$. Therefore, the algorithm of Fortune et al. \cite{FORTUNE1980111} gives a solution in $n^{O(l)}$. As noted in \cite{Bang}, we can also reduce the problem of directed-disjoint paths in acyclic digraphs to $1$-DSP, which implies the following theorem: \begin{theorem}\label{th:hardness} The $1$-DSP problem is $W[1]$-hard parameterized by the number of requests. \end{theorem} \begin{proof} Consider an instance $D$ and $(s_1, t_1), \dots, (s_k, t_k)$ of disjoint paths in acyclic digraphs and $v_1, \dots, v_n$ a topological ordering of the vertices of $D$. Let $D'$ be the digraph obtained from $D$ by subdividing each arc $(v_i, v_j)$ $j -i - 1$ times. By doing so, every $(v_i,v_j)$-path in $D'$ now has length exactly $i-j -i$, which means that every path is a shortest path. The underlying graph $G'$ obtained by forgetting the orientation of the arcs in $D'$ is a $1$-shortest graph, as all the edges appear in the BFS starting from $v_1$, and a shortest path between $v_i$ and $v_j$ still corresponds to a $(v_i,v_j)$-path in $D$. This means that solving the $1$-DSP problem defined by $G'$ and the $(s_i, t_i)$ gives a solution to the original instance of disjoint paths in $D$, which ends the proof as disjoint paths in acyclic digraphs is $W[1]$-hard \cite{Slivkins}\footnote{The hardness proof of Slivkins, while stated for the arc-disjoint version works also for the vertex-disjoint one.}. \end{proof} The rest of the paper is devoted to the proof of Theorem \ref{thm:main}. The main idea behind the proof is to reduce to a set of $O(l^{5^k})$ requests such that, roughly speaking, for each pair of requests of different colours, no pair of shortest paths solving these requests can intersect. Once we have achieved this, it means that the only potential conflicts arise for pairs of requests of the same colour. However, the edges of each colour class can be seen as an acyclic digraph, and we can adapt the algorithm of Fortune et al. to that case. The main difficulty lies in reducing to these $O(l^{5^k})$ requests. To achieve this, we need to look at a potential solution to the original $k$-DSP problem and say that, for each pair of paths of different colours in this solution, there is a way to partition each of these paths into a finite number of subpaths, such that the endpoints of each pair of subpaths now correspond to requests that can never intersect. The next two sections are devoted to this task. In particular, it is devoted to the understanding of the structure of bi-coloured edges. \section{Bi-coloured components} Let $G$ be a $k$-shortest graph and $i,j$ two integers in $[k]$. Consider $G_{i,j}^+$ (resp. $G_{i,j}^-$ ), the graph induced by the edges $xy$ of $G$ of colour both $i$ and $j$ such that $x \leq_i y$ and $x \leq_j y$ (resp. $x \leq_i y$ and $y \leq_j x$). A \textit{bi-coloured} component of colours $i,j$ is a connected component of $G_{i,j}^+$ or $G_{i,j}^-$. Note that $G_{i,j}^+$ and $G_{i,j}^-$ play identical roles, as reversing the order of the partition $(V^j_1, \dots, V^j_{l_j})$ transforms the $k$-shortest graph $G$ into a $k$-shortest graph $G'$ where every component of $G_{i,j}^-$ becomes a component of $G_{i,j}^+$ and vice-versa. Bi-coloured components will play an essential role in order to decompose a $k$-DSP problem into a set of $O(l^{5^k})$ requests such that the intersections between request of different colours behave nicely\footnote{We will define what we mean by nicely in the next section.}. The first property we need to prove is that for any path $P_i$ of colour $i$ and bi-coloured component $C$ of colour $i$ and $j$, then $P_i \cap C$ is a subpath of $P_i$. Let us first show some properties of the bi-coloured components. \begin{lemma}\label{lemma_diff_pos} Let $G$ be a $k$-shortest graph, $i,j$ two indices in $[k]$, and $S$ some component of $G_{i,j}^+$. There exists a constant $C_{S}$ such that for any vertex $x \in S$, if $x \in V^i_r$, then $x \in V^j_{r + C_{S}}$. \end{lemma} \begin{proof} Let $x$ be any vertex belonging to $S$. Let $r$ and $t$ be the constants such that $ x \in V^i_r$ and $x \in V^j_t$ and define $C_S = t-r$. Let $y$ be another vertex of $S$. By definition of $S$, there exists a path $P$ in $G_{i,j}^+$ between $x$ and $y$. Let $s_1$ be the number of edges of $P$ which are used positively for the order induced by the colour $i$ when going from $x$ to $y$, and $s_2$ the number of edges used negatively. By definition of this order, we have that $y \in V^i_{r + s_1 - s_2}$. Because the orders induced by the colours $i$ and $j$ are the same on $S$, we also have that $y \in V^j_{t + s_1 - s_2}$, which ends the proof. \end{proof} Let us now show the following properties of paths of colour $i$. \begin{proposition}\label{prop:shortest_path} Let $G$ be a $k$-shortest graph. Suppose $x \in V^i_r$ and $y \in V^i_t$ for some $i \in [k]$ and $r,t \in [l_i]$ with $r > t+1$. If there exists a path in $G$ of length $r-t$ between $x$ and $y$, then this path is a path of colour $i$ from $y$ to $x$. \end{proposition} \begin{proof} Let $P= x_1, \dots , x_s$ with $x_1 = y$, $x_s = x$ and $s = r-t +1$ be a path of length $r-t$ between $x$ and $y$. For every $j \in [s]$, let $i_j$ be the integer such that $x_j \in V^i_{i_j}$. We know that for any $j \in [2..s]$, $i_j \leq i_{j-1} + 1$ as $x_j$ and $x_{j-1}$ are adjacent. However, $i_1 = t$, $i_s = r$ and $s = r-t +1$. This means that $i_j = i_{j-1} + 1$ for every $j \in [2..s]$ and all the edges of $P$ are edges of colour $i$. \end{proof} \begin{proposition}\label{prop:two_path_colours} Let $G$ be a $k$-shortest graph, $i,j$ two indices of $[k]$ and $x$ and $y$ two vertices of $G$. If there exists a path $P_i$ of colour $i$ between $x$ and $y$ and a path $P_j$ of colour $j$ between $x$ and $y$, then $P_j$ is also a path of colour $i$ and $P_i$ is also a path of colour $j$. \end{proposition} \begin{proof} We know that $P_i$ and $P_j$ are shortest paths between $x$ and $y$, and in particular have the same length. The result follows by applying Proposition \ref{prop:shortest_path} to the paths $P_j$ and $P_i$. \end{proof} We are ready to prove the following lemma, which shows how paths of colour $i$ interact with $G_{i,j}^+$. \begin{lemma}\label{lem:compo_inter_subpath} Let $G$ be a $k$-shortest graph, $i,j$ two indices in $[k]$, $P_i$ a path of colour $i$ and $S$ some bi-coloured component of colours $i,j$. The intersection of $P_i$ and $S$ is a subpath of $P_i$. \end{lemma} \begin{proof} Let $S$ be a component of $G_{i,j}^+$ and suppose that $P_i$ does not intersect $S$ along a single subpath. This means that we can find a subpath $P'$ of $P_i$ of colour $i$ between two vertices $x$, $y$ of $S$ such that $P'$ uses no edge of $S$. Suppose $x$ is the first endpoint of this path, $y$ the last and $l$ denote the length of $P'$. Then $x \in V^i_r$ and $y \in V^i_{r+ l}$. However, by Lemma \ref{lemma_diff_pos}, we know that there exists a constant $C_S$ such that, since both $x$ and $y$ belong to $S$, $x \in V^j_{r + C_S}$ and $y \in V^j_{r+ l + C_S}$. By Proposition \ref{prop:shortest_path}, this implies that $P'$ is also a path of colour $j$, and thus $P' \in S$. The case where $S$ is a component of $G_{i,j}^-$ is very similar and thus omitted. \end{proof} Let $G$ be a $k$-shortest graph, $i,j$ two indices in $[k]$, $P_i$ a path of colour $i$ and $P_j$ a path of colour $j$. We say that $P_i$ and $P_j$ are in \textit{conflict} if there exists a bi-coloured component $S$ of colour $i,j$ such that the intersection of $P_i$ with $S$ is a $(s_1',t_1')$-path and the intersection of $P_j$ with $S$ is an $(s_2', t_2')$-path for some $s_1', s_2', t_1', t_2' \in S$, with the property that there exists a $(s_1',t_1')$-path $P_i'$ of colour $i$ and an $(s_2', t_2')$-path $P_j'$ of colour $j$ using at least one vertex outside of $s_1', s_2', t_1', t_2'$ in common. The component $S$ will be called a \textit{conflicting component} for $P_i$ and $P_j$. By convention, two paths of the same colour are never conflicting. The following lemma is the main ingredient of our Algorithm. It shows that for two paths of colour $i$ and $j$, there is at most one conflicting component. \begin{lemma}\label{lem:inter_main} Let $G$ be a $k$-shortest graph, $i,j$ two indices in $[k]$, $P_i$ a path of colour $i$ and $P_j$ a path of colour $j$. Suppose $S$ is a conflicting component for the paths $P_i$ and $P_j$, then $P_i$ and $P_j$ do not have any vertex in common outside $S$. \end{lemma} \begin{proof} Again, we can assume that $S$ is a component of $G_{i,j}^+$ by potentially reversing the order $i$. Suppose that the intersection of $P_i$ with $S$ is an $(s_1',t_1')$-path and the intersection of $P_j$ with $S$ is an $(s_2', t_2')$-path for some $s_1', s_2', t_1', t_2' \in S$. We prove the lemma by contradiction, distinguishing several cases depending on which part of $P_i$ and $P_j$ (compared to $S$) the intersection lies on. Suppose first that the intersection of $P_i$ and $P_j$ lies after $S$ for both paths. By definition of conflicting components, we know that there exists a vertex $x \in S$ such that there exists a path $P_1$ of colour $i$ from $x$ to $t_1'$ and a path $P_2$ of colour $j$ from $x$ to $t_2'$. Let $z$ be the first vertex belonging to the intersection of $P_i \cap P_j$ after $S$. By applying Proposition \ref{prop:two_path_colours} to the path obtained by concatenating $P_1$ and $P_i[t_1', z]$ and the one obtained by concatenating $P_2$ and $P_j[t_2',z]$, we get that these two paths are both of colour $i$ and $j$. In particular this implies that the edges of these paths belong to $G_{i,j}^+$ and $z \in S$, which is a contradiction. Suppose now that the intersection of $P_i$ and $P_j$ lies before $S$ on $P_i$ and after $S$ on $P_j$. By definition of conflicting components, we know that there exists a vertex $x \in S \setminus \{ s_1', s_2', t_1', t_2' \}$ such that there exists a path $P_1$ of colour $i$ from $s_1'$ to $x$ and a path $P_2$ of colour $j$ from $x$ to $t_2'$. Because $x \not \in \setminus \{ s_1', s_2', t_1', t_2' \}$, the two paths $P_1$ and $P_2$ have both length at least $1$. Moreover, they are both paths of colour $i$ and $j$, and with the same orientation associated to these colours. Let $z$ denote the first vertex in the intersection of $P_i$ and $P_j$ before $S$ on $P_i$ and after $S$ on $P_j$ and consider the path $H_1 = P_i[z, s_1']P_1P_2$. $H_1$ is a path of colour $i$ between $z$ and $t_2'$, and thus $|H_1| = |P_j[t_2',z]|$. Likewise, we can show that $|P_i[z,s_1']P_1| = |P_2P_j[t_2', z]|$, which gives us a contradiction. The other cases are symmetrical. \end{proof} Let us now explain how the previous lemma will be used. Remember that our goal is to reduce an original instance of $k$-DSP with $l$ requests to one with $O(l^{5^k})$ requests such that for every pair of requests of different colours, no pair of shortest paths solving these requests can intersect. Suppose $P_i$ and $P_j$ are two paths of different colours in a solution of the original $k$-DSP which are in conflict. Let $S$ denote the conflicting component for $P_i$ and $P_j$. Because of Lemma \ref{lem:compo_inter_subpath}, we know that the intersection of $P_i$ and $P_j$ with $S$ are subpaths. For every $a \in \{i, j\}$, consider the path partition $(P^1_a, P^2_a, P^3_a)$ of $P_a$, where $P^2_a$ is the subpath of $P_a$ on $S$, $P^1_a$ the part of $P_a$ before this component, and $P^3_a$ the part after. What Lemma \ref{lem:inter_main} roughly says is that the endpoints of $P^1_i, P^1_j, P^3_i$ and $P^3_j$ correspond to requests such that no pair of shortest paths solving these requests can intersect, which is exactly what we wanted. This is not true for the requests associated to $P^2_j$ and $P^2_i$, however since they both belong to a bi-coloured component $S$, these two requests can be considered of the same colour. Surprisingly, the case where $P_i$ and $P_j$ are not in conflict is harder to handle. This is the goal of the next section, but let us first show sufficient conditions to guarantee the existence of a conflicting component for a pair of paths. \begin{lemma}\label{lem:3_conflict} Let $G$ be a $[k]$-shortest graph and $i,j$ two different indices in $[k]$, $P_i$ a path of colour $i$ and $P_j$ a path of colour $j$. If $P_i$ and $P_j$ have three common vertices, then they have a conflicting component. \end{lemma} \begin{proof} Let $x_1$, $x_2$ and $x_3$ be three vertices in $P_i \cap P_j$. We claim that they belong to the same bi-coloured component. Indeed, consider the subpaths of $P_i$ and $P_j$ between $x_1$ and $x_2$. By Proposition \ref{prop:two_path_colours} they are both paths of colour $i$ and $j$ and belong to the same component $S$. Without loss of generality, suppose $S$ is a component of $G_{i,j}^+$ and $x_1 \leq_i x_2$. If $x_3$ belongs to $P_i[x_1, x_2]$ or $P_j[x_1, x_2]$, we have that $x_3 \in S$, which ends the proof of the claim. Assume now $x_3$ appears after $x_2$ on $P_i$, the other case being symmetrical. If it appears after $x_2$ on $P_j$, then the same argument shows that $P_i[x_2,x_3]$ is also a path of $S$. Suppose now that $x_3$ appears before $x_1$ on $P_j$. In that case we have that $P_j[x_3,x_2]$ and $P_i[x_2, x_3]$ are both shortest path, and thus have the same size. However, this implies that $P_j[x_3,x_1]$ is strictly shorter than $P_i[x_1,x_3]$, which is a contradiction. Now that we know that $x_1$, $x_2$ and $x_3$ belong to the same bi-coloured component, let us show that this component is a conflicting component for $P_i$ and $P_j$. Indeed, since $x_1, x_2$ and $x_3$ belong to the same component, they either appear in the same order on the paths $P_i$ and $P_j$ if $S$ is a component of $G_{i,j}^+$ or in reverse order if $S$ is a component of $G_{i,j}^-$. In both cases, the vertex in the middle is the same in both paths, and this implies that $P_i$ and $P_j$ are conflicting on this component. \end{proof} \section{Blind Paths} As explained earlier, if $P_i$ and $P_j$ are two paths of the solution of some $k$-DSP problem which are in conflict, then Lemma \ref{lem:inter_main} allows us to show that the paths $P_i$ and $P_j$ can be decomposed into a finite set of requests such that each pair of requests of different colours are without any possible intersection, meaning that for any shortest path $P$ solving the first request and $P'$ solving the second request, $P \cap P' = \emptyset$. Moreover, finding these decompositions only requires to guess the conflicting component and the intersection of $P_i$ and $P_j$. Unfortunately, it is not true that any positive instance of $k$-DSP has a solution where every pair of paths of different colours are in conflict. For this purpose we need the definition of blind pair of paths. Let $P_i$ be some $(s_i, t_i)$-path of colour $i$ and $P_j$ some $(s_j, t_j)$-path of colour $j$ which are internally vertex-disjoint. We say that $P_i$ \textit{sees} $P_j$ if there exists an internal vertex $x$ of $P_i$ such that there exists a path of colour $i$ from $x$ to $t_i$ which intersects $P_j \setminus \{s_j, t_j\}$. We say that the pair $P_i$ and $P_j$ is \textit{blind} if $P_j$ does not see $P_i$ and $P_i$ does not see $P_j$. Note that if $|P_i| = 2$, then $P_i$ does not see, or is not seen, by any other path $P_j$. Note also that, if $P_i$ and $P_j$ are blind, then it is possible to find a $(s_j, t_j)$-path of colour $j$ $P'_j$ and a $(s_i, t_i)$-path of colour $i$ $P'_i$ such that $P'_i \cap P'_j \not = \emptyset$. In that sense, blind paths is a weaker notion than what we could obtain from Lemma \ref{lem:inter_main} outside of the conflicting component. The following lemma shows how to use Lemma \ref{lem:inter_main} to find blind paths from conflicting paths. \begin{lemma}\label{lem:blind_1} Let $G$ be a $k$-shortest graph, $i,j$ two integers in $[k]$, $P_i$ a path of colour $i$ and $P_j$ a path of colour $j$ which are internally vertex-disjoint. There exists a path partition $L_i$ of $P_i$ and a path partition $L_j$ of $P_j$, both of size at most $9$ with the following properties: \begin{itemize} \item All the paths of $L_a$ are paths of colour $a$ for $a \in \{i, j\}$ \item For any pair of paths $H_i \in L_i$ and $H_j \in L_j$, then either $H_i$ and $H_j$ are paths of the same bi-coloured component of colour $i,j$ or $H_i$ does not see $H_j$. \end{itemize} \end{lemma} \begin{proof} Suppose $P_i$ is an $(s_i, t_i)$-path and $P_j$ is an $(s_j, t_j)$ path. If $P_i$ does not see $P_j$, then $L_i = \{ P_i\}$ and $L_j = \{P_j\}$ satisfy the properties of the lemma. Suppose now $P_i$ sees $P_j$ and let $x_1$ denote the last vertex of $P_i$ from which there exists a path $Q^1_i$ of colour $i$ to $t_i$ which uses some vertex of $P_j \setminus \{s_j, t_j\}$. Because $P_i$ and $P_j$ are internally vertex disjoint, $x_1 \not = t_i$. Let $x_1'$ denote the vertex just after $x_1$ on $P_i$. Note that $P_i[x_1',t_1]$ does not see $P_j$. Now let $x_2$ denote the last vertex of $P_i[s_i, x_1]$ from which there exists a path of colour $i$ to $x_1$ which uses some vertex of $P_j \setminus \{s_j, t_j\}$. Again, if this vertex does not exist, then $L_i = \{P_i[s_i, x_1], (x_1,x_1'), P_i[x_1', t_i] \}$ and $L_j = \{P_j\}$ satisfy the properties of the lemma. Suppose from now on that $x_2$ exists and let $x_2'$ be the vertex just after $x_2$ on $P_i$. Since $P_i$ and $P_j$ are internally vertex-disjoint, $x_2 \not = x_1$ and thus $x_2' \in P_i[s_i, x_1]$. Again, note that $P_i[x_2',x_1]$ does not see $P_j$. Let $x_3$ denote the last vertex of $P_i[s_i, x_2]$ from which there exists a path $Q^3_i$ of colour $i$ to $x_2$ which uses some vertex of $P_j \setminus \{s_j, t_j\}$. Again, we can assume that this vertex exists or $L_i = \{ P_i[s_i, x_2], (x_2, x_2'), P_i[x_2',x_1], (x_1, x_1'), P_i[x_1', t_i] \} $ and $L_j = \{P_j\}$ satisfy the properties of the lemma. Let $x_3' \in P_i[s_i, x_2]$ denote the vertex just after $x_3$ on $P_i$. Again, note that $P_i[x_3',x_2]$ does not see $P_j$. Note that for any internal vertex $x \in Q^1_i$ and $y \in Q^2_i$, $y <_i x$. This implies that the intersection of $ Q^1_i$ and $Q^2_i$ is equal to $x_1$, and the same argument applies for $Q^3_i \cap Q^2_i$ and $Q^3_i \cap Q^1_i$. This means that the paths $P_j$ and $P'_i = P_i[s_i, x_3]Q^3_iQ^2_iQ^1_i$ intersect on at least 3 vertices and thus are conflicting by Lemma \ref{lem:3_conflict}. Let $S$ denote the conflicting component of $P_i'$ and $P_j$. Suppose first that none of the $s_i, t_i, s_j, t_j$ belong to $S$ and denote by $e_i$ the last edge of $P'_i$ without both endpoints in $S$, $e_j$ the last edge of $P_j$ before $S$, $h_i$ the first edge of $P'_i$ after $S$ and $h_j$ the first edge of $P_j$ after $S$. \begin{claim} All the pairs of paths among $P'_i[s_i, t(e_i)]$,$e_i$, $P'_i[h(e_i), t(h_i)]$, $h_i$, $P'_i[h(h_i), t_i]$, $P_j[s_j, t(e_j)]$, $e_j$, $P_j[h(e_j), t(h_j)]$, $h_j$ and $P_j[h(h_j), t_j]$ are blind, except from $P_j[h(e_j), t(h_j)]$ and $P'_i[h(e_i), t(h_i)]$ which belong to the same bi-coloured component. \end{claim} \begin{proof} Since $e_i$ and $h_i$ are not edges of $S$ and there exists a path of colour $i$ in this component from $h(e_i)$ to $t(h_i)$, then by Lemma \ref{lem:compo_inter_subpath} no path of colour $i$ from $s_i$ to $t(e_i)$ or from $h(h_i)$ to $t_i$ can use any vertex of $S$. However, any path of colour $j$ from $h(e_j)$ to $t(h_j)$ is a path of $S$, so it cannot intersect any path of colour $i$ from $s_i$ to $t(e_i)$ or from $h(h_i)$ to $t_i$. This means that $(P'_i[s_i, t(e_i)],P_j[h(e_j), t(h_j)])$ and $(P'_i[h(h_i), t_i],P_j[h(e_j), t(h_j)])$ are blind pairs. By reversing the role of $i$ and $j$, it also means that $(P_j[s_j, t(e_j)],P_i'[h(e_i), t(h_i)])$ and $(P_j[h(h_j), t_j],P_i'[h(e_i), t(h_i)])$ are blind pairs. By the definition of conflicting and Lemma \ref{lem:inter_main}, we can show that no path of colour $i$ from $s_i$ to $t(e_i)$ or from $h(h_i)$ to $t_i$ can intersect a path of colour $j$ from $s_j$ to $t(e_j)$ or $h(h_j)$ to $t_j$. Indeed, suppose for example that there exists a path $H_i$ of colour $i$ from $s_i$ to $t(e_i)$ that intersects a path $H_j$ of colour $j$ from $s_j$ to $t(e_j)$. In that case the paths $H_iP_i'[t(e_i), h(h_i)]$ and $H_jP_j[t(e_j), h(h_j)]$ contradict Lemma \ref{lem:inter_main} as $S$ is a conflicting component for these two paths, but they also intersect outside of $S$. The other cases are symmetrical and thus all pairs of paths among $P'_i[s_i, t(e_i)]$, $P'_i[h(h_i), t_i]$, $P_j[s_j, t(e_j)]$ and $P_j[h(h_j), t_j]$ are blind. This ends the proof of the claim as the other pairs contain an edge and are blind by definition and $P_j[h(e_j), t(h_j)]$ and $P'_i[h(e_i), t(h_i)]$ are paths of $S$. \end{proof} Let $L_j = \{P_j[s_j, t(e_j)], e_j, P_j[h(e_j), t(h_j)], h_j, P_j[h(h_j), t_j] \}$. Suppose first that $t(e_i)$ appears after $x_3$ on $P'_i$. It means that $P_i[s_i, x_3]$ is a subpath of $P'_i[s_i, t(e_i)]$, and in particular $P_i[s_i, x_3]$ does not see $P_j$. Setting $L_i = \{P_i[s_i, x_3],(x_3, x_3'),P_i[x_3', x_2], (x_2, x_2'), P_i[x_2', x_1], (x_1,x_1'), P_i[x_1', t_i], \}$, we then have that no path of $L_i$ sees $P_j$ and thus any path of $L_j$. Suppose now that $t(e_i)$ appears before $x_3$ on $P'_i$. Note that $t(h_i)$ has to appear after $x_3$ or there is no path from $x_3$ to $t_i$ intersecting $P_j$, which contradicts the choice of $x_3$. In that case, setting $L_i = \{P_i[s_i, t(e_i)], e_i, P_i[h(e_i), x_3], (x_3, x_3'), P_i[x_3, x_2], (x_2, x_2') P_i[x_2', x_1], (x_1, x_1'), P_i[x_1, t_i] \}$, we also have that the only path of $L_i$ that sees a path of $L_j$ is $P_i[h(e_i), x_3]$. Moreover, it can only see $P_j[h(e_j), t(h_j)]$, but these paths belong to the same bi-coloured component $S$. The cases where some of the $s_i, t_i, s_j, t_j$ belong to $S$ are treated exactly the same, except that some of the $e_i, e_j, h_j, h_i$ might not exist, which means we have fewer paths to consider. \end{proof} By applying the previous lemma several times, we obtain the following: \begin{lemma}\label{lem:blind_main} There exists a constant $C$ such that if $G$ is a $k$-shortest graph, $i,j$ two integers in $[k]$, $P_i$ a path of colour $i$ and $P_j$ a path of colour $j$ which are internally vertex-disjoint, then there exists a path partition $L_i$ of $P_i$ and a path partition $L_j$ of $P_j$, both of size at most $C$ with the following properties: \begin{itemize} \item Each $L_a$ consists of at most $C$ paths of colour $a$. \item For any pair of path $H_i \in L_i$ and $H_j \in L_j$ which are not blind, then $H_i$ and $H_j$ are paths of the same bi-coloured component. \end{itemize} \end{lemma} \begin{proof} Let $Q_i$, $Q_j$ be the path partitions obtained by applying Lemma \ref{lem:blind_1} to $P_i$ and $P_j$. We know that for any pair of paths $H_i \in Q_i$ and $H_j \in Q_j$, then either $H_i$ and $H_j$ are paths of the same bi-coloured component of colours $i,j$, or $H_i$ does not see $H_j$. Now as long as there exists a path in $H_i \in Q_i$ such that there exists some path $H_j \in Q_j$, such that $H_j$ sees $H_i$ and is not a path of the same bi-coloured component as $H_i$, we do the following. Let $H_{j,1}, \dots, H_{j,r}$ denote all the paths of $Q_j$ which see $H_i$ and do not belong to the same bi-coloured component. For any $a \in [r]$, let $Q_{j,a}$ and $Q_{i,a}$ denote the set of path partitions obtained by applying Lemma \ref{lem:blind_1} to $H_{j,a}$ and $H_i$. Let $Q'_i$ be the intersection of all the partitions $Q_{i,a}$ of $P_i$. Because every path of $Q'_i$ is a subpath of some $Q_{i,a}$ for any $a \in [r]$, it means that this path is not seen by any path in $Q_{j,a}$ which is not a path of the same bi-coloured component. Let us update $Q_i$ by replacing $H_i$ by $Q'_i$ and update $Q_j$ by replacing each of the $H_{j,a}$ by $Q_{j,a}$. By doing that, the number of paths in $Q_i$ which is seen by some path $H_j \in Q_j$ which is not a path of the same bi-coloured component decreases strictly as none of the paths of $Q'_i$ satisfy these properties. At each step, we multiply the number of paths in $Q_j$ by at most 9 and the number of paths in $Q_i$ by at most $9|Q_j|$. However, we only have to do this $9$ steps as initially the sets $Q_i$ and $Q_j$ have size at most $9$. Finally, this means that after 9 steps, $|Q_j| \leq 9^9$ and $|Q_i|\leq 9(9|Q_j|)^{9} \leq 9^{91}$. This ends the proof for $C = 9^{91}$ \end{proof} \section{Proof of the main theorem} We are now ready to describe and prove our algorithm. First we will show, using Lemma \ref{lem:blind_1} and induction, that any solution $P_1, \dots, P_l$ of some $k$-DSP can be decomposed into some path partitions $L_1, \dots, L_l$ where each $L_i$ has size at most $C(k,l)$ for some function $C$ and any pair of paths in the union of the $L_i$ of different colours is blind. The algorithm then consists of guessing the endpoints and colours of the partitions $L_1, \dots, L_l$ (there is at most $n^{O(C(k,l))}$ possible choices) and then solve the $k$-DSP defined using the fact that all the pairs of paths of different colours are blind. Roughly speaking, each colour class is solved almost independently using Fortune et al. Algorithm. \subsection{Proof of the blind case} Let us first show the existence of the algorithm in the blind case. Note that we also have some list of forbidden components for each path in the partitions $L_1, \dots, L_l$. This is due to some technicalities in the proof of the existence of such decomposition. \begin{lemma}\label{lem:blind_case} Let $G$ be a $k$-shortest graph, $(s_1, t_1), \dots, (s_l, t_l)$ a set of pairs and $c$ a function from $[l]$ into $[k]$. Moreover, suppose that for every $i$, there is a list $F_i$ of bi-coloured components where one of the colours being $c(i)$. There exists an algorithm running in time $ n^{O(l)}$ that either returns a solution $P_1, \dots, P_l$ to the $k$-DSP defined by $G$, the $(s_i, t_i)$ and $c$ or shows that no solution is such that each $P_i$ does not use any vertex of any component in $F_i$ and moreover, for any indices $i$ and $j$, either $P_j$ and $P_i$ are blind or $P_i$ is a path of some component of $F_j$ or $P_j$ is a path of some component of $F_i$. \end{lemma} To prove this lemma, we will build an auxiliary digraph $D$ such that a solution satisfying the properties of the lemma exists if and only if there exists a directed path in $D$ between two specified vertices. First note that, by potentially replacing some vertices with an independent set with the same neighbourhood, we can assume that all the $s_i$ and $t_i$ are disjoint. The vertices of $D$ will correspond to $l$-tuples $(x_1, \dots, x_l)$ of vertices of $G$. Intuitively, we are trying to build the paths $P_i$ starting from $s_i$, and $x_i$ is the last vertex of a prefix of $P_i$ we are considering. For any pair of vertices $(x_1, \dots, x_l)$ and $(y_1, \dots, y_l)$, $D$ contains the arc from $(x_1, \dots, x_l)$ to $(y_1, \dots, y_l)$ if the following are satisfied: \begin{itemize} \item There exists $i \in [l]$ such that $x_j = y_j$ for all $j \in [l]$, $j \not = i$. \item $x_iy_i$ is an edge of colour $c(i)$ such that there exists a path of colour $c(i)$ from $y_i$ to $t_i$ avoiding the components in $F_i$. \item For all $j \in [l]$ different from $i$, $y_i \not = x_j$ and either there is no path of colour $c(j)$ from $x_j$ to $t_j$ that uses the vertex $x_i$, or $x_i$ is a vertex of a component of $F_j$. \end{itemize} Let $S = (s_1, \dots, s_l)$ and $T = (t_1, \dots t_l)$. The next two claims finishes the proof of Lemma \ref{lem:blind_case}. \begin{claim} If there exists a solution $P_1, \dots, P_l$ to the $k$-DSP defined by $G$, $(s_i, t_i)$ and $c$ such that each $P_i$ does not use a vertex of any component in $F_i$ and moreover, for any indices $i$ and $j$, either $P_j$ and $P_i$ are blind, $P_i$ is a path of some component of $F_j$ or $P_j$ is a path of some component of $F_i$, then there is a path in $D$ from $S$ to $T$. \end{claim} \begin{proof} Let $P_1, \dots P_l$ denote such a solution in $G$. Let $X$ be the set of vertices of $D$ corresponding to $l$-tuples obtained by taking one vertex per path $P_i$. We can define a natural order on $X$ by considering for each $P_i$ the order induced by the path and taking the lexicographic order. Note that $T$ is the maximal element of $X$. Consider now the largest element $A=(x_1, \dots, x_l)$ of $X$ which is reachable in $D$ from $S$ and suppose, in order to reach a contradiction, that this element is not $T$. Consider some colour $c_1$ and $I$ the set of indices of $i$ of $[l]$ such that for $c(i) = c_1$ and $x_i \not = t_i$. Because the edges of colour $c_1$ induce an acyclic digraph, there exists an index $i \in I$ such that for every $j \in I$ with $j \not = i$, there is no path of colour $c_1$ from $x_j$ to $x_i$. Now for any $j \in [l]$ such that $c(j) \not = c_1$, then either the path $P_i$ and $P_j$ are blind, in which case there is no path of colour $c(j)$ from $x_j$ to $t_j$ that uses $x_i$, or $P_i$ (and thus $x_i$) is in some component of $F_j$, or $P_j$ is a path of some component of $F_i$. Note that in the last case, any path from $x_j$ to $t_j$ is a path of some component of $F_i$, but $x_i$ cannot be a vertex of this component, and thus no such path can use $x_i$. Therefore, if we note $x_i'$ the vertex just after $x_i$ on $P_i$ and $A'$ the vertex of $D$ obtained from $A$ by only changing $x_i$ into $x_i'$, then there exists an arc from $A$ to $A'$. However, this means that $A'$ is reachable from $S$ in $D$, which contradicts the maximality of $A$. \end{proof} And the opposite direction. \begin{claim} If there exists a path from $S$ to $T$ in $D$, then there exists a solution $P_1, \dots, P_l$ to the DSP defined by $G'$, the $(s_i, t_i)$ and $c$ such that $P_i$ does not use any vertex of any component in $F_i$. \end{claim} \begin{proof} Suppose there exists a path $P = X_1, \dots, X_r$ from $S$ to $T$ in $D$. For every $j \in [r]$, note $X_j = (x^j_1, \dots, x^j_l)$. For every $i$ and $j$, consider the graph $P^j_i$ induced by the vertices $x^t_i$, for $t \leq j$. By definition of $D$, $P^j_i$ is a path of colour $c(i)$ from $s_i$ to $x^j_i$ avoiding the components of $F_i$. We will prove by induction on $j$, that the paths $P^j_1, \dots, P^j_l$ are such that \begin{itemize} \item All the $P^j_i$ are internally disjoint. \item For any $i$ and $r$, there is no path of colour $c(i)$ from $x^j_i$ to $t_j$ avoiding the components in $F_i$ that uses any vertex of $P^j_r$ outside of possibly $x^j_r$. \end{itemize} Since the path starts at $S$, all the properties are satisfied when $j = 1$. Suppose now that this is true for some $j \in [r-1]$ and let us show that the properties hold for $j+1$. By definition of the arcs of $D$, there exists an index $i$ such that $x^j_ix^{j+1}_i$ is an edge of colour $c(i)$ and for every other index $s$, $x^j_s = x^{j+1}_s$. This means that $P^{j+1}_i$ is the concatenation of $P^j_i$ with $x^{j+1}_i$ and all the $P^{j+1}_s$ are equal to $P^{j}_s$ for $s \not = i$. Moreover, by definition of $D$, we know that for any $s \not = i$, $x^{j+1}_i$ is disjoint from all the $x^j_s$ and by induction hypothesis $x^{j+1}_i$ does not belong to any of the $P^j_s$. This implies that the $P^j_s$ for $s \in [l]$, are disjoint. Any path of colour $c(i)$ from $x^{j+1}_i$ to $t_i$ avoiding the components of $F_i$ is a subpath of a path of colour $c(i)$ from $x^{j}_i$ to $t_i$ avoiding the components of $F_i$. This means that no such path can use any vertex of $P^j_s = P^{j+1}_s$ outside of possibly $x^j_s$ for all $s \in [l]$ different from $i$. Finally, for $s \in [l]$ different from $i$ we know that no path of colour $c(s)$ from $x^j_r = x^{j+1}_r$ to $t_s$ avoiding the components in $F_s$ can use any vertex of $P^j_i$ outside of $x^i_s$ by induction hypothesis. Moreover, these paths can also not use $x^{j+1}_i$ by definition of the arcs of $D$, which ends our induction. This means that each $P^r_i$ is a path of colour $c(i)$ avoiding the components in $F_i$ from $s_i$ to $t_i$, and all these paths are disjoints, which ends the proof. \end{proof} Therefore, the problem reduces to deciding the existence of a path in $D$. As $|D| = n^l$, this can be done in $n^{O(l)}$. \subsection{Reducing to the blind case} The next lemma shows how to reduce to the blind case with some forbidden lists. \begin{lemma}\label{lem:decomp_sol} Let $G$ be a $k$-shortest graph, $(s_1, t_1), \dots (s_l, t_l)$ a set of pairs and $c$ a function from $[l]$ to $[k]$. Let $P_1, \dots, P_l$ be a solution to the $k$-DSP defined by $G$, the $(s_i,t_i)$ and $c$. There exists a constant $C(k,l)$ depending only on $k$ and $l$, a set of path partitions $L_1, \dots, L_l$, a function $a$ that associates to each path of the $L_i$ a colour in $[k]$ and a function $b$ that associates to each path of the $L_i$ a set of bi-coloured components with the following properties: \begin{itemize} \item For every $i \in [l]$, $L_i$ is a path partition of $P_i$ of at most $C(k,l)$ paths. \item If $P$ is a path of some $L_i$ with $a(P) = c_1$, then $P$ is a path of colour $c_1$ and $b(P)$ consists of a set of at most $C(k,l)$ bi-coloured components where one of the colours is $c_1$ and such that $P$ does not use any vertex in these components. \item For any pair of paths $H_i \in L_i$ and $H_j \in L_j$ such that $a(H_i) \not = a(H_j)$, then either $H_i$ and $H_j$ are blind, $H_j$ is a path contained in one component of $b(H_i)$ or $H_i$ is a path contained in one component of $b(H_j)$. \end{itemize} \end{lemma} Note that if $P_i$ is a path of colour $j$, then any path partition of $P_i$ consists of paths of colour $j$. The function $a$ is there to reassign the colour of some paths belonging to bi-coloured components in order to achieve the last property of the lemma. \begin{proof} Let $C$ be the constant from Lemma \ref{lem:blind_main}. We will prove by induction on $k$ that the lemma is true with $C(k,l) \leq (7Cl)^{5^k}$. When $k =1$, there is only one colour and setting $L_i = \{P_i\}$ for all $i$ satisfies the properties, and thus $C(1,l) \leq l$. Suppose now that $k >1$, and let $I_1$ denote the set of indices $i \in [l]$ such that $c(i) = 1$. Note that we can assume $I_1$ to be non empty, or the induction step is trivial. For every pair of indices $i,j \in [l]$, let $Q_{i,j}, Q_{j,i}$ denote the path partitions of $P_i$, $P_j$ obtained by applying Lemma \ref{lem:blind_main} to this pair and for every $i \in [l]$, let $Q_i$ denote the intersection of all the $Q_{i,j}$. Note that, since every $Q_{i,j}$ has size $C$, this implies that the $Q_i$ have size at most $C l$. For any $i \in I_1$, let $Q'_i$ denote the set of paths of $Q_i$, for which there exists some other path among some $Q_j$ with $j \in [k] \setminus I_1$ such that the pair is not blind. By Lemma \ref{lem:blind_main}, all these paths belong to some bi-coloured component of colour $1$ and some other colour $t$. Let $B_i$ denote the set of all these bi-coloured components and let $R_i = Q_i \setminus Q'_i$. Let $R = \bigcup_{ i \in I_1} R_i$, $B = \bigcup_{ i \in I_1} B_i$ and note that $|B| \leq Cl^2$ and $|R| \leq Cl^2$. For any path $R_t \in R$, we know that the intersections of $R_t$ with any component $C_j \in B$ is a subpath by Lemma \ref{lem:compo_inter_subpath}. Let $e_j$ be the last edge of $R_t$ before $C_j$ and $h_j$ the first edge after. Let $a_t$ and $b_t$ denote the first and last vertex of $R_t$, and consider $R_{t,j} = \{R_t[a_t, t(e_j)], e_j, R_t[h(e_j), t(h_j)],$ $ h_j, R_t[h(h_j), b_t] \}$ a path partition of $R_t$. Note that, except the two edges $e_j$ and $h_j$, each path of $R_{t,j}$ is either disjoint from $C_j$ or a path of this component. Let $L(R_t)$ denote the path partition of $R_t$ obtained by taking the intersection of all the $R_{t,j}$. We know that, since $|B| \leq Cl^2$, $L(R_t) \leq 5Cl^2$. Moreover, we know that for any path $P'$ of $L(R_t)$, and any $C_j \in B$, there is a path $r_{t,j} \in R_{t,j}$ such that $P'$ is a subpath of $r_{t,j}$. In particular it means that $P'$ is either an edge, disjoint from $C_j$, or a path of $C_j$. Let $R^2_t$ be the set of paths of $L(R_t)$ which belong to one of the component of $B$, and $R^1_t = L(R_t) \setminus R^2_t$. Note that every path in $R^1_t$ is either an edge or a path disjoint from all the components of $B$. Let $H_1$ denote the set of paths in all the $Q'_i$ for $i \in I_1$ and all the $R^2_t$ for $R_t \in R$. Note that for every path $H' \in H_1$, $H'$ is path of a bi-coloured component of $B$. Denote by $c'(H')$ the colour of this component which is not $1$. We will now consider $H'$ as a path of colour $c'(H')$ (possibly reversing the endpoints if the component is a component of $G_{1,c'(H')}^-$). Let $G_1$ be the $(k-1)$-shortest graph obtained from $G$ be removing the partition associated to colour $1$ and removing all the edges which are edges of colour $1$ only. Consider now the $(k-1)$-DSP problem defined on $G_1$ by all the endpoints of the paths in $Q_i$ for $i \in [l]$, $i \not \in I_1$, considered as path of colour $c(i)$, and all the paths in $H' \in H_1$ considered as path of colour $c'(H')$. Note that the set of paths in $Q_i$ and $H_1$ is a solution to this problem and moreover, there is at most $7(Cl^2)^2$ requests, as each $Q_i$ is smaller than $Cl$ and $H_1$ is smaller than $6(Cl^2)^2$. By induction hypothesis, there exists a path partition of all the paths of the $Q_i$ for $i \in [l]$, $i \not \in I_1$ and $H_1$ as well as two functions $a'$ and $b'$ defined on these paths such that each of these path partitions consists of at most $C(k-1, 7(Cl^2)^2)$ paths, and $b'$ associates to each path at most $C(k-1, 7(Cl^2)^2)$ bi-coloured components. Let us define the path partitions $L_i$, as well as $a$ and $b$ as follows: For every $i \in I_1$, $L_i$ is the union of all the paths in $R^1_t$ for some $R_t \in R_i$ as well as all the paths in the path partitions of the paths in $R^2_t$ and $Q'_i$ obtained by applying induction. For every path $P$ of some $R^1_t$, let $a(P) = 1$ and $b(P) = B$. For all the other paths of $L_i$, $a$ and $b$ correspond to the value of $a'$ and $b'$ on these paths. Likewise, for every $i \in [l]$ such that $i \not \in I_1$, $L_i$ consists of the union of the path partitions for the paths in $Q_i$ obtained by applying induction, and the function $a$ and $b$ correspond to the $a'$ and $b'$ on these paths. Let us now show that the $L_i$ and functions $a$ and $b$ satisfy the required properties. First, it is clear that the $L_i$ thus defined are path partitions, as they are obtained by replacing paths of some path partitions by their own path partition. Moreover, $|L_i| \leq 7(Cl^2)^2 \cdot C(k-1, 7(Cl^2)^2) \leq 7(Cl^2)^2 \cdot (7^2(Cl^2)^2)^{5^{k-1}} \leq (7Cl)^{5^k} $. Likewise, for any paths $P$ in these partitions, $b(P)$ is smaller than $\max\{C(k-1, Cl^2), |B|\} \leq (7Cl)^{5^k} $. Now suppose $H_i$ is a path of $L_i$ and $H_j$ is a path of $L_j$ such that $a(H_i) \not = a(H_j)$. If none of these paths belong to some $R^1_t$ for $R_t \in R$, then the last property of the lemma is satisfied for $H_i$ and $H_j$ by induction and because $a$ and $b$ correspond to $a'$ and $b'$ on these paths. Suppose now that one of the paths, say $H_i$ belongs to $R^1_t$ for some $R_t \in R$. Because $R^1_t$ is a subpath of an element of $R$, it means that if $H_j$ is not a subpath of some path of $Q'_s$ for $s \in I_1$, then by definition of $R$, $H_i$ and $H_j$ are blind. However, if $H_j$ is a subpath of some path of $Q_s'$, then $H_j$ is a path belonging to some component of $B$. However, $b(H_i) = B$, which ends the proof. \end{proof} Finally, we can prove our main result. \begin{proof}[Proof of Theorem \ref{thm:main}] Suppose there exists a solution $P_1, \dots, P_l$ to the $k$-DSP problem defined by $G$, the $(s_i,t_i)$ and $c$. Let $L_1, \dots L_l$, $a$ and $b$ be the path partitions and functions obtained by applying Lemma \ref{lem:decomp_sol} to $P_1, \dots, P_l$. For every $i \in [l]$, let $P_{i,1}, \dots, P_{i,l_i}$ denote the paths of $L_i$ and $(s^i_1, t^i_1), \dots (s^i_{l_i}, t^i_{l_i}) $ the endpoints of these paths. Remember that by Lemma \ref{lem:decomp_sol}, $|L_i| \leq C(k,l)$ for all $i \in [l]$. Suppose we guess all the $(s^i_j, t^i_j)$, as well as the functions $a$ and $b$ for each of the $P_{i,j}$, and consider the $k$-DSP problem defined by all the remaining pairs $(s^i_j, t^i_j)$, then the set of paths $P_{i,j}$ is a solution to this problem such that, for any pair of paths $P_{i,j}$, $P_{i',j'}$ such that $a(P_{i,j}) \not = a(P_{i',j'})$, either $P_{i,j}$ and $P_{i,j}$ are blind, $P_{i,j}$ is a path contained in one component of $b(P_{i',j'})$ or $P_{i',j'}$ is a path contained in one component of $b(P_{i,j})$. This means that we can apply the algorithm of Lemma \ref{lem:blind_case} to find a solution of the $k$-DSP defined by $(s^i_j, t^i_j)$ in $n^{O(C(k,l))}$. By concatenating for each $i$ all the paths of this solution corresponding to the paths of $L_i$, we obtain a solution to the initial $k$-DSP problem. As there is at most $n^{O(C(k,l))}$ choices for the $(s^i_j, t^i_j)$, $a(P_{i,j})$ and $b(P_{i,j})$ this gives an algorithm running in time $n^{O(C(k,l))}$, which ends the proof. \end{proof} \section*{Acknowledgements} The author wishes to thank Frédéric Havet and Saket Saurabh for their useful comments on the manuscript.

https://arxiv.org/abs/1606.05299

Common Fixed Point Results for Family of Generalized Multivalued F-contraction Mappings in Ordered Metric Spaces

In this paper, we study the existence of common fixed points of family of multivalued mappings satisfying generalized F-contractive conditions in ordered metric spaces. These results establish some of the general common fixed point theorems for family of multivalued maps.

\section{COMMON\ FIXED POINT RESULTS} \noindent In this section, we obtain common fixed point theorems for family of multivalued mappings. We begin with the following result. \noindent \textbf{Theorem 2.1.} \ \ Let $(X,d,\preceq )$ be an ordered complete metric space and $\{T_{i}\}_{i=1}^{m}:X\rightarrow P_{cl}(X)$ be family of multivalued mappings. Suppose that for every $(x,y)\in \Delta _{1}$ and $u_{x}\in T_{i}(x),$ there exists $u_{y}\in T_{i+1}(y)$ for $i\in \{1,2,...,m\}$ (with $T_{m+1}=T_{1}$ by convention) such that, (u_{x},u_{y})\in \Delta _{2}$ implie \begin{equation} \tau +F\left( d(u_{x},u_{y})\right) \leq F(M(x,y;u_{x},u_{y})), \tag{2.1} \end{equation where $\tau $\ is a positive real number an \begin{equation*} M(x,y;u_{x},u_{y})=\max \{d(x,y),d(x,u_{x}),d(y,u_{y}),\dfrac{d\left( x,u_{y}\right) +d\left( y,u_{x}\right) }{2}\}. \end{equation* Then the following statements hold: \begin{enumerate} \item[(1)] $Fix(T_{i})\neq \emptyset $ for any $i\in \{1,2,...,m\}$ if and only if $Fix(T_{1})=Fix(T_{2})=...=Fix(T_{m})\neq \emptyset .$ \item[(2)] $Fix(T_{1})=Fix(T_{2})=...=Fix(T_{m})\neq \emptyset $ provided that any one $T_{i}$ for $i\in \{1,2,...,m\}$\ is upper semicontinuous. \item[(3)] $\cap _{i=1}^{m}Fix(T_{i})$ is well-ordered if and only if $\cap _{i=1}^{m}Fix(T_{i})$ is singleton set. \end{enumerate} \noindent \textbf{Proof.} \ \ To prove (1), let $x^{\ast }\in T_{k}(x^{\ast })$ for some $k\in \{1,2,...,m\}.$ Assume that $x^{\ast }\notin T_{k+1}\left( x^{\ast }\right) ,$ then there exists an $x\in T_{k+1}\left( x^{\ast }\right) $ with $\left( x^{\ast },x\right) \in \Delta _{2}\ $such tha \begin{equation*} \tau +F\left( d(x^{\ast },x)\right) \leq F(M(x^{\ast },x^{\ast };x^{\ast },x)), \end{equation* wher \begin{eqnarray*} M(x^{\ast },x^{\ast };x^{\ast },x) &=&\max \{d(x^{\ast },x^{\ast }),d(x^{\ast },x^{\ast }),d(x,x^{\ast }),\dfrac{d(x^{\ast },x)+d(x^{\ast },x^{\ast })}{2}\} \\ &=&d(x,x^{\ast }), \end{eqnarray* implies tha \begin{equation*} \tau +F\left( d(x^{\ast },x)\right) \leq F(d(x^{\ast },x)), \end{equation* a contradiction as $\tau >0$. Thus $x^{\ast }=x$. Thus $x^{\ast }\in T_{k+1}\left( x^{\ast }\right) $ and so $Fix(T_{k})\subseteq Fix(T_{k+1}).$ Similarly, we obtain that\ $Fix(T_{k+1})\subseteq Fix(T_{k+2})$ and continuing this way, we get $Fix(T_{1})=Fix(T_{2})=...Fix(T_{k}).$ The converse is straightforward.\smallskip \noindent To prove (2), suppose that $x_{0}$ is an arbitrary point of $X.$ If $x_{0}\in T_{k_{0}}\left( x_{0}\right) $ for any $k_{0}\in \{1,2,...,m\},$ then by using (1), the proof is finishes. So we assume that $x_{0}\notin T_{k_{0}}\left( x_{0}\right) $ for any $k_{0}\in \{1,2,...,m\}.$ Now for i\in \{1,2,...,m\}$, if $x_{1}\in T_{i}(x_{0}),$ then there exists $x_{2}\in T_{i+1}(x_{1})$ with $(x_{1},x_{2})\in \Delta _{2}$ such tha \begin{equation*} \tau +F\left( d(x_{1},x_{2})\right) \leq F(M(x_{0},x_{1};x_{1},x_{2})), \end{equation* wher \begin{eqnarray*} M(x_{0},x_{1};x_{1},x_{2}) &=&\max \{d(x_{0},x_{1}),d(x_{0},x_{1}),d(x_{1},x_{2}),\dfrac d(x_{0},x_{2})+d(x_{1},x_{1})}{2}\} \\ &=&\max \{d(x_{0},x_{1}),d(x_{1},x_{2}),\dfrac{d(x_{0},x_{2})}{2}\} \\ &=&\max \{d(x_{0},x_{1}),d(x_{1},x_{2})\}. \end{eqnarray* Now, if $M(x_{0},x_{1};x_{1},x_{2})=d(x_{1},x_{2})$ the \begin{equation*} \tau +F\left( d(x_{1},x_{2})\right) \leq F(d(x_{1},x_{2})), \end{equation* a contradiction as $\tau >0$. Therefore M(x_{0},x_{1};x_{1},x_{2})=d(x_{0},x_{1})$ and we hav \begin{equation*} \tau +F\left( d(x_{1},x_{2})\right) \leq F\left( d(x_{0},x_{1})\right) . \end{equation* Next for this $x_{2}\in T_{i+1}\left( x_{1}\right) ,$ there exists $x_{3}\in T_{i+2}(x_{2})$ with $\left( x_{2},x_{3}\right) \in \Delta _{2}$ such tha \begin{equation*} \tau +F\left( d(x_{2},x_{3})\right) \leq F(M(x_{1},x_{2};x_{2},x_{3})), \end{equation* wher \begin{eqnarray*} M(x_{1},x_{2};x_{2},x_{3}) &=&\max \{d(x_{1},x_{2}),d(x_{1},x_{2}),d(x_{2},x_{3}),\dfrac d(x_{1},x_{3})+d(x_{2},x_{2})}{2}\} \\ &=&\max \{d(x_{1},x_{2}),d(x_{2},x_{3})\}. \end{eqnarray* Now, if $M(x_{1},x_{2};x_{2},x_{3})=d(x_{2},x_{3})$ the \begin{equation*} \tau +F\left( d(x_{2},x_{3})\right) \leq F(d(x_{2},x_{3})), \end{equation* a contradiction as $\tau >0$. Therefore M(x_{1},x_{2};x_{2},x_{3})=d(x_{1},x_{2})$ and we hav \begin{equation*} \tau +F\left( d(x_{2},x_{3})\right) \leq F\left( d(x_{1},x_{2})\right) . \end{equation* Continuing this process, for $x_{2n}\in T_{i}(x_{2n-1})$, there exist x_{2n+1}\in T_{i+1}\left( x_{2n}\right) $ with $\left( x_{2n},x_{2n+1}\right) \in \Delta _{2}$ such tha \begin{equation*} \tau +F\left( d(x_{2n},x_{2n+1})\right) \leq F\left( M(x_{2n-1},x_{2n};x_{2n},x_{2n+1})\right) , \end{equation* wher \begin{eqnarray*} M(x_{2n-1},x_{2n};x_{2n},x_{2n+1}) &=&\max \{d(x_{2n-1},x_{2n}),d(x_{2n-1},x_{2n}),d(x_{2n},x_{2n+1}), \\ &&\dfrac{d(x_{2n-1},x_{2n+1})+d(x_{2n},x_{2n})}{2}\} \\ &=&\{d(x_{2n-1},x_{2n}),d(x_{2n},x_{2n+1}),\dfrac{d(x_{2n-1},x_{2n+1})}{2}\} \\ &\leq &d(x_{2n-1},x_{2n}), \end{eqnarray* that is \begin{equation*} \tau +F\left( d(x_{2n},x_{2n+1})\right) \leq F\left( d(x_{2n-1},x_{2n})\right) . \end{equation* Similarly, for $x_{2n+1}\in T_{i+1}(x_{2n})$, there exist $x_{2n+2}\in T_{i+2}\left( x_{2n+1}\right) $ such that for $\left( x_{2n+1},x_{2n+2}\right) \in \Delta _{2}$ implie \begin{equation*} \tau +F\left( d(x_{2n+1},x_{2n+2})\right) \leq F\left( d(x_{2n},x_{2n+1})\right) . \end{equation* Hence, we obtain a sequence $\{x_{n}\}$ in $X$ such that for $x_{n}\in T_{i}(x_{n-1})$, there exist $x_{n+1}\in T_{i+1}\left( x_{n}\right) $ with \left( x_{n},x_{n+1}\right) \in \Delta _{2}$ such tha \begin{equation*} \tau +F\left( d(x_{n},x_{n+1})\right) \leq F\left( d(x_{n-1},x_{n})\right) . \end{equation* Therefor \begin{eqnarray} F\left( d(x_{n},x_{n+1})\right) &\leq &F\left( d(x_{n-1},x_{n})\right) -\tau \leq F\left( d(x_{n-2},x_{n-1})\right) -2\tau \notag \\ &\leq &...\leq F\left( d(x_{0},x_{1})\right) -n\tau . \TCItag{2.2} \end{eqnarray From (2.2), we obtain $\lim\limits_{n\rightarrow \infty }F\left( d(x_{n},x_{n+1})\right) =-\infty $ that together with ($F_{2}$) give \begin{equation*} \lim\limits_{n\rightarrow \infty }d(x_{n},x_{n+1})=0. \end{equation* From ($F_{3}$), there exists $\lambda \in \left( 0,1\right) $ such tha \begin{equation*} \lim\limits_{n\rightarrow \infty }[d(x_{n},x_{n+1})]^{\lambda }F\left( d(x_{n},x_{n+1})\right) =0. \end{equation* From (2.2) we hav \begin{eqnarray*} &&[d(x_{n},x_{n+1})]^{\lambda }F\left( d(x_{n},x_{n+1})\right) -[d(x_{n},x_{n+1})]^{\lambda }F\left( d(x_{0},x_{n+1})\right) \\ &\leq &-n\tau \lbrack d(x_{n},x_{n+1})]^{\lambda }\leq 0. \end{eqnarray* On taking limit as $n\rightarrow \infty $ we obtai \begin{equation*} \lim\limits_{n\rightarrow \infty }n[d(x_{n},x_{n+1})]^{\lambda }=0. \end{equation* Hence $\lim\limits_{n\rightarrow \infty }n^{\frac{1}{\lambda }d(x_{n},x_{n+1})=0$ and there exists $n_{1}\in \mathbb{N} $ such that $n^{\frac{1}{\lambda }}d(x_{n},x_{n+1})\leq 1$ for all $n\geq n_{1}.$ So we hav \begin{equation*} d(x_{n},x_{n+1})\leq \frac{1}{n^{1/\lambda }} \end{equation* for all $n\geq n_{1}.$ Now consider $m,n\in \mathbb{N} $ such that $m>n\geq n_{1}$, we hav \begin{eqnarray*} d\left( x_{n},x_{m}\right) &\leq &d\left( x_{n},x_{n+1}\right) +d\left( x_{n+1},x_{n+2}\right) +...+d\left( x_{m-1},x_{m}\right) \\ &\leq &\sum_{i=n}^{\infty }\frac{1}{i^{1/\lambda }}. \end{eqnarray* By the convergence of the series $\sum_{i=1}^{\infty }\frac{1}{i^{1/\lambda },$ we get $d\left( x_{n},x_{m}\right) \rightarrow 0$ as $n,m\rightarrow \infty $. Therefore $\{x_{n}\}$ is a Cauchy sequence in $X.$ Since $X$ is complete, there exists an element $x^{\ast }\in X$ such that x_{n}\rightarrow x^{\ast }$ as $n\rightarrow \infty $.\smallskip Now, if $T_{i}$ is upper semicontinuous for any $i\in \{1,2,...,m\}$, then as $x_{2n}\in X,$ $x_{2n+1}\in T_{i}\left( x_{2n}\right) $ with x_{2n}\rightarrow x^{\ast }$ and $x_{2n+1}\rightarrow x^{\ast }$ as n\rightarrow \infty $ implies that $x^{\ast }\in T_{i}\left( x^{\ast }\right) .$ Thus from (2), we get $x^{\ast }\in T_{1}\left( x^{\ast }\right) =T_{2}\left( x^{\ast }\right) =...=T_{m}\left( x^{\ast }\right) $.\smallskip \noindent Finally to prove (3),\ suppose the set $\cap _{i=1}^{m}Fix\left( T_{i}\right) $ is a well-ordered.\ We are to show that $\cap _{i=1}^{m}Fix\left( T_{i}\right) \ $is singleton. Assume on contrary that there exist $u$ and $v$ such that $u,v\in \cap _{i=1}^{m}Fix\left( T_{i}\right) $ but $u\neq v$. As $(u,v)\in \Delta _{2}$, so for (u_{x},v_{y})\in \Delta _{2}$ implie \begin{eqnarray*} \tau +F\left( d(u,v)\right) &\leq &F(M(u,v;u,v)) \\ &=&F(\max \{d(u,v),d(u,u),d(v,v),\dfrac{d\left( u,v\right) +d\left( v,u\right) }{2}\}) \\ &=&F\left( d\left( u,v\right) \right) , \end{eqnarray* a contradiction as $\tau >0$.\ Hence $u=v$. Conversely, if $\cap _{i=1}^{m}Fix\left( T_{i}\right) $ is singleton, then it follows that $\cap _{i=1}^{m}Fix\left( T_{i}\right) $ is a well-ordered. $\square $\newline \newline \noindent The following corollary extends and generalizes Theorem 4.1 of \cite{LatifBeg} and Theorem 3.4 of \cite{Rus}. \noindent \textbf{Corollary 2.2.} \ \ Let $(X,d,\preceq )$ be an ordered complete metric space and $T_{1},T_{2}:X\rightarrow P_{cl}(X)$ be two multivalued mappings. Suppose that for every $(x,y)\in \Delta _{1}$ and u_{x}\in T_{i}(x),$ there exists $u_{y}\in T_{j}(y)$ for $i,j\in \{1,2\}$ with $i\neq j$ such that, $(u_{x},u_{y})\in \Delta _{2}$ implie \begin{equation*} \tau +F\left( d(u_{x},u_{y})\right) \leq F(M(x,y;u_{x},u_{y})), \end{equation* where $\tau $\ is a positive real number an \begin{equation*} M(x,y;u_{x},u_{y})=\max \{d(x,y),d(x,u_{x}),d(y,u_{y}),\dfrac{d\left( x,u_{y}\right) +d\left( y,u_{x}\right) }{2}\}. \end{equation* Then the following statements hold: \begin{enumerate} \item[(1)] $Fix(T_{i})\neq \emptyset $ for any $i\in \{1,2\}$ if and only if $Fix(T_{1})=Fix(T_{2})\neq \emptyset .$ \item[(2)] $Fix(T_{1})=Fix(T_{2})\neq \emptyset $ provided that $T_{1}$ or T_{2}$\ is upper semicontinuous. \item[(3)] $Fix(T_{1})\cap Fix(T_{2})$ is well-ordered if and only if Fix(T_{1})\cap Fix(T_{2})$ is singleton set. \end{enumerate} \noindent \textbf{Example 2.3.}\ \ \ Let $X=\{x_{n}=\dfrac{n(n+1)}{2}:n\in \{1,2,3,...\}\}$ endow with usual order $\leq .$ Le \begin{eqnarray*} \Delta _{1} &=&\{(x,y):x\leq y\text{ where }x,y\in X\}\text{ and} \\ \Delta _{2} &=&\{(x,y):x<y\text{ where }x,y\in X\}. \end{eqnarray* Define $T_{1}$, $T_{2}:X\rightarrow P_{cl}(X)$ as follows \begin{eqnarray*} T_{1}\left( x\right) &=&\{x_{1}\}\text{ for }x\in X,\text{ and } \\ T_{2}\left( x\right) &=&\left\{ \begin{array}{lll} \{x_{1}\} & , & \text{ }x=x_{1} \\ \{x_{1},x_{n-1}\} & , & \text{ }x=x_{n},\text{ for }n>1\text{. \end{array \right. \end{eqnarray* Take $F\left( \alpha \right) =\ln \alpha +\alpha ,$ $\alpha >0$ and $\tau =1. $ For a Euclidean metric $d$ on $X,$ and $\left( u_{x},u_{y}\right) \in \Delta _{2},$ we consider the following cases: \begin{enumerate} \item[(i)] If $x=x_{1},y=x_{m},$ for $m>1,$ then for $u_{x}=x_{1}\in T_{1}\left( x\right) ,$ there exists $u_{y}=x_{m-1}\in T_{2}\left( y\right) , $ such tha \begin{eqnarray*} d(u_{x},u_{y})e^{d(u_{x},u_{y})-M\left( x,y;u_{x},u_{y}\right) } &\leq &d(u_{x},u_{y})e^{d(u_{x},u_{y})-d\left( x,y\right) } \\ &=&\frac{m^{2}-m-2}{2}e^{-m} \\ &<&\frac{m^{2}+m-2}{2}e^{-1} \\ &=&e^{-1}d\left( x,y\right) \\ &\leq &e^{-1}M\left( x,y;u_{x},u_{y}\right) . \end{eqnarray*} \item[(ii)] If $x=x_{n},$ $y=x_{n+1}$ with $n>1,$ then for $u_{x}=x_{1}\in T_{1}\left( x\right) ,$ there exists $u_{y}=x_{n-1}\in T_{2}\left( y\right) , $ such tha \begin{eqnarray*} d(u_{x},u_{y})e^{d(u_{x},u_{y})-M\left( x,y;u_{x},u_{y}\right) } &\leq &d(u_{x},u_{y})e^{d(u_{x},u_{y})-[\tfrac{d\left( x,u_{y}\right) +d\left( y,u_{x}\right) }{2}]} \\ &=&\frac{n^{2}-n-2}{2}e^{\frac{-3n-2}{2}} \\ &<&\frac{n^{2}+4n}{2}e^{-1} \\ &=&e^{-1}[\dfrac{d\left( x,u_{y}\right) +d\left( y,u_{x}\right) }{2}] \\ &\leq &e^{-1}M_{1}\left( x,y;u_{x},u_{y}\right) . \end{eqnarray*} \item[(iii)] When $x=x_{n},$ $y=x_{m}$ with $m>n>1,$ then for u_{x}=x_{1}\in T_{1}\left( x\right) ,$ there exists $u_{y}=x_{n-1}\in T_{2}\left( y\right) ,$ such tha \begin{eqnarray*} d(u_{x},u_{y})e^{d(u_{x},u_{y})-M\left( x,y;u_{x},u_{y}\right) } &\leq &d(u_{x},u_{y})e^{d(u_{x},u_{y})-d\left( x,u_{x}\right) } \\ &=&\frac{n^{2}-n-2}{2}e^{-n} \\ &<&\frac{n^{2}+n-2}{2}e^{-1} \\ &=&e^{-1}d\left( x,u_{x}\right) \\ &\leq &e^{-1}M\left( x,y;u_{x},u_{y}\right) . \end{eqnarray*} \end{enumerate} \noindent Now we show that for $x,y\in X$, $u_{x}\in T_{2}\left( x\right) $; there exists $u_{y}\in T_{1}\left( y\right) $ such that $\left( u_{x},u_{y}\right) \in \Delta _{2}$ and (2.1) is satisfied. For this, we consider the following cases: \begin{enumerate} \item[(i)] If $x=x_{n},$ $y=x_{1}$ with $n>1,$ we have for $u_{x}=x_{n-1}\in T_{2}\left( x\right) ,$ there exists $u_{y}=x_{1}\in T_{1}\left( y\right) ,$ such tha \begin{eqnarray*} d(u_{x},u_{y})e^{d(u_{x},u_{y})-M\left( x,y;u_{x},u_{y}\right) } &\leq &d(u_{x},u_{y})e^{d(u_{x},u_{y})-d\left( x,y\right) } \\ &=&\frac{n^{2}-n-2}{2}e^{-n} \\ &<&\frac{n^{2}+n-2}{2}e^{-1} \\ &=&e^{-1}d\left( x,y\right) \\ &\leq &e^{-1}M\left( x,y;u_{x},u_{y}\right) . \end{eqnarray*} \item[(ii)] In case $x=x_{n},$ $y=x_{m}$ with $m>n>1,$ then for u_{x}=x_{n-1}\in T_{2}\left( x\right) ,$ there exists $u_{y}=x_{1}\in T_{2}\left( y\right) ,$ such tha \begin{eqnarray*} d(u_{x},u_{y})e^{d(u_{x},u_{y})-M\left( x,y;u_{x},u_{y}\right) } &\leq &d(u_{x},u_{y})e^{d(u_{x},u_{y})-d\left( y,u_{y}\right) } \\ &=&\frac{n^{2}-n-2}{2}e^{n^{2}-n-m^{2}-m} \\ &<&\frac{m^{2}+m-2}{2}e^{-1} \\ &=&e^{-1}d\left( y,u_{y}\right) \\ &\leq &e^{-1}M\left( x,y;u_{x},u_{y}\right) . \end{eqnarray*} \end{enumerate} \noindent Thus for all $x,y$ in $X$, (2.1) is satisfied. Hence all the conditions of Theorem 2.1 are satisfied. Moreover, $x_{1}=1$ is the unique common fixed point of $T_{1}$ and $T_{2}$ with $Fix(T_{1})=Fix(T_{2}).$ \square $\newline \noindent The following results generalizes Theorem 3.4 of \cite{Rus} and Theorem 3.4 of \cite{Sgr}. \noindent \textbf{Theorem 2.4.} \ \ Let $(X,d,\preceq )$ be an ordered complete metric space and $\{T_{i}\}_{i=1}^{m}:X\rightarrow P_{cl}(X)$ be family of multivalued mappings. Suppose that for every $(x,y)\in \Delta _{1}$ and $u_{x}\in T_{i}(x),$ there exists $u_{y}\in T_{i+1}(y)$ for $i\in \{1,2,...,m\}$ (with $T_{m+1}=T_{1}$ by convention) such that, (u_{x},u_{y})\in \Delta _{2}$ implie \begin{equation} \tau +F\left( d(u_{x},u_{y})\right) \leq F(M_{2}(x,y;u_{x},u_{y})), \tag{2.3} \end{equation where $\tau $\ is a positive real number an \begin{equation*} M_{2}(x,y;u_{x},u_{y})=\alpha d(x,y)+\beta d(x,u_{x})+\gamma d(y,u_{y})+\delta _{1}d\left( x,u_{y}\right) +\delta _{2}d\left( y,u_{x}\right) , \end{equation* and $\alpha ,\beta ,\gamma ,\delta _{1},\delta _{2}\geq 0,$ $\delta _{1}\leq \delta _{2}$ with $\alpha +\beta +\gamma +\delta _{1}+\delta _{2}\leq 1$. Then the following statements hold: \begin{enumerate} \item[(1)] $Fix(T_{i})\neq \emptyset $ for any $i\in \{1,2,...,m\}$ if and only if $Fix(T_{1})=Fix(T_{2})=...=Fix(T_{m})\neq \emptyset .$ \item[(2)] $Fix(T_{1})=Fix(T_{2})=...=Fix(T_{m})\neq \emptyset $ provided that any one $T_{i}$ for $i\in \{1,2,...,m\}$\ is upper semicontinuous. \item[(3)] $\cap _{i=1}^{m}Fix(T_{i})$ is well-ordered if and only if $\cap _{i=1}^{m}Fix(T_{i})$ is singleton set. \end{enumerate} \noindent \textbf{Proof.} \ \ To prove (1), let $x^{\ast }\in T_{k}(x^{\ast })$ for some $k\in \{1,2,...,m\}.$ Assume that $x^{\ast }\notin T_{k+1}\left( x^{\ast }\right) ,$ then there exists an $x\in T_{k+1}\left( x^{\ast }\right) $ with $\left( x^{\ast },x\right) \in \Delta _{2}\ $such tha \begin{equation*} \tau +F\left( d(x^{\ast },x)\right) \leq F(M_{2}(x^{\ast },x^{\ast };x^{\ast },x)), \end{equation* wher \begin{eqnarray*} M_{2}(x^{\ast },x^{\ast };x^{\ast },x) &=&\alpha d(x^{\ast },x^{\ast })+\beta d(x^{\ast },x^{\ast })+\gamma d(x,x^{\ast }) \\ &&+\delta _{1}d(x^{\ast },x)+\delta _{2}d(x^{\ast },x^{\ast }) \\ &=&(\gamma +\delta _{1})d(x,x^{\ast }), \end{eqnarray* implies tha \begin{eqnarray*} \tau +F\left( d(x^{\ast },x)\right) &\leq &F((\gamma +\delta _{1})d(x^{\ast },x)) \\ &\leq &F(d(x^{\ast },x)), \end{eqnarray* a contradiction as $\tau >0$. Thus $x^{\ast }=x$. Thus $x^{\ast }\in T_{k+1}\left( x^{\ast }\right) $ and so $Fix(T_{k})\subseteq Fix(T_{k+1}).$ Similarly, we obtain that\ $Fix(T_{k+1})\subseteq Fix(T_{k+2})$ and continuing this way, we get $Fix(T_{1})=Fix(T_{2})=...Fix(T_{k}).$ The converse is straightforward.\smallskip \noindent To prove (2), suppose that $x_{0}$ is an arbitrary point of $X.$ If $x_{0}\in T_{k_{0}}\left( x_{0}\right) $ for any $k_{0}\in \{1,2,...,m\},$ then by using (1), the proof is finishes. So we assume that $x_{0}\notin T_{k_{0}}\left( x_{0}\right) $ for any $k_{0}\in \{1,2,...,m\}.$ Now for i\in \{1,2,...,m\}$, if $x_{1}\in T_{i}(x_{0}),$ then there exists $x_{2}\in T_{i+1}(x_{1})$ with $(x_{1},x_{2})\in \Delta _{2}$ such tha \begin{equation*} \tau +F\left( d(x_{1},x_{2})\right) \leq F(M_{2}(x_{0},x_{1};x_{1},x_{2})), \end{equation* wher \begin{eqnarray*} M_{2}(x_{0},x_{1};x_{1},x_{2}) &=&\alpha d(x_{0},x_{1})+\beta d(x_{0},x_{1})+\gamma d(x_{1},x_{2}) \\ &&+\delta _{1}d(x_{0},x_{2})+\delta _{2}d(x_{1},x_{1}) \\ &\leq &(\alpha +\beta +\delta _{1})d(x_{0},x_{1})+(\gamma +\delta _{1})d(x_{1},x_{2}). \end{eqnarray* Now, if $d(x_{0},x_{1})\leq d(x_{1},x_{2}),$ then we hav \begin{eqnarray*} \tau +F\left( d(x_{1},x_{2})\right) &\leq &F((\alpha +\beta +\gamma +2\delta _{1})d(x_{1},x_{2})) \\ &\leq &F(d(x_{1},x_{2})), \end{eqnarray* a contradiction. Therefor \begin{equation*} \tau +F\left( d(x_{1},x_{2})\right) \leq F\left( d(x_{0},x_{1})\right) . \end{equation* Next for this $x_{2}\in T_{i+1}\left( x_{1}\right) ,$ there exists $x_{3}\in T_{i+2}(x_{2})$ with $\left( x_{2},x_{3}\right) \in \Delta _{2}$ such tha \begin{equation*} \tau +F\left( d(x_{2},x_{3})\right) \leq F(M_{2}(x_{1},x_{2};x_{2},x_{3})), \end{equation* wher \begin{eqnarray*} M_{2}(x_{1},x_{2};x_{2},x_{3}) &=&\alpha d(x_{1},x_{2})+\beta d(x_{1},x_{2})+\gamma d(x_{2},x_{3}) \\ &&+\delta _{1}d(x_{1},x_{3})+\delta _{2}d(x_{2},x_{2}) \\ &\leq &(\alpha +\beta +\delta _{1})d(x_{1},x_{2})+(\gamma +\delta _{1})d(x_{2},x_{3}). \end{eqnarray* Now, if $d(x_{1},x_{2})\leq d(x_{2},x_{3})$ the \begin{eqnarray*} \tau +F\left( d(x_{2},x_{3})\right) &\leq &F((\alpha +\beta +\gamma +2\delta _{1})d(x_{2},x_{3})) \\ &\leq &F\left( d\left( x_{2},x_{3}\right) \right) , \end{eqnarray* a contradiction as $\tau >0$. Therefor \begin{equation*} \tau +F\left( d(x_{2},x_{3})\right) \leq F\left( d(x_{1},x_{2})\right) . \end{equation* Continuing this process, for $x_{2n}\in T_{i}(x_{2n-1})$, there exist x_{2n+1}\in T_{i+1}\left( x_{2n}\right) $ with $\left( x_{2n},x_{2n+1}\right) \in \Delta _{2}$ such tha \begin{equation*} \tau +F\left( d(x_{2n},x_{2n+1})\right) \leq F\left( M_{2}(x_{2n-1},x_{2n};x_{2n},x_{2n+1})\right) , \end{equation* wher \begin{eqnarray*} M_{2}(x_{2n-1},x_{2n};x_{2n},x_{2n+1}) &=&\alpha d(x_{2n-1},x_{2n})+\beta d(x_{2n-1},x_{2n})+\gamma d(x_{2n},x_{2n+1}) \\ &&+\delta _{1}d(x_{2n-1},x_{2n+1})+\delta _{2}d(x_{2n},x_{2n}) \\ &\leq &\left( \alpha +\beta +\delta _{1}\right) d(x_{2n-1},x_{2n})+\left( \gamma +\delta _{1}\right) d(x_{2n},x_{2n+1}) \\ &\leq &d(x_{2n-1},x_{2n}), \end{eqnarray* that is \begin{equation*} \tau +F\left( d(x_{2n},x_{2n+1})\right) \leq F\left( d(x_{2n-1},x_{2n})\right) . \end{equation* Similarly, for $x_{2n+1}\in T_{i+1}(x_{2n})$, there exist $x_{2n+2}\in T_{i+2}\left( x_{2n+1}\right) $ such that for $\left( x_{2n+1},x_{2n+2}\right) \in \Delta _{2}$ implie \begin{equation*} \tau +F\left( d(x_{2n+1},x_{2n+2})\right) \leq F\left( d(x_{2n},x_{2n+1})\right) . \end{equation* Hence, we obtain a sequence $\{x_{n}\}$ in $X$ such that for $x_{n}\in T_{i}(x_{n-1})$, there exist $x_{n+1}\in T_{i+1}\left( x_{n}\right) $ with \left( x_{n},x_{n+1}\right) \in \Delta _{2}$ such tha \begin{equation*} \tau +F\left( d(x_{n},x_{n+1})\right) \leq F\left( d(x_{n-1},x_{n})\right) . \end{equation* Therefor \begin{eqnarray} F\left( d(x_{n},x_{n+1})\right) &\leq &F\left( d(x_{n-1},x_{n})\right) -\tau \leq F\left( d(x_{n-2},x_{n-1})\right) -2\tau \notag \\ &\leq &...\leq F\left( d(x_{0},x_{1})\right) -n\tau . \TCItag{2.4} \end{eqnarray From (2.4), we obtain $\lim\limits_{n\rightarrow \infty }F\left( d(x_{n},x_{n+1})\right) =-\infty $ that together with ($F_{2}$) give \begin{equation*} \lim\limits_{n\rightarrow \infty }d(x_{n},x_{n+1})=0. \end{equation* Follows the arguments those in proof of Theorem 2.1, $\{x_{n}\}$ is a Cauchy sequence in $X.$ Since $X$ is complete, there exists an element $x^{\ast }\in X$ such that $x_{n}\rightarrow x^{\ast }$ as $n\rightarrow \infty .\smallskip Now, if $T_{i}$ is upper semicontinuous for any $i\in \{1,2,...,m\}$, then as $x_{2n}\in X,$ $x_{2n+1}\in T_{i}\left( x_{2n}\right) $ with x_{2n}\rightarrow x^{\ast }$ and $x_{2n+1}\rightarrow x^{\ast }$ as n\rightarrow \infty $ implies that $x^{\ast }\in T_{i}\left( x^{\ast }\right) .$ Thus from (2), we get $x^{\ast }\in T_{1}\left( x^{\ast }\right) =T_{2}\left( x^{\ast }\right) =...=T_{m}\left( x^{\ast }\right) $.\smallskip \noindent Finally to prove (3),\ suppose the set $\cap _{i=1}^{m}Fix\left( T_{i}\right) $ is a well-ordered.\ We are to show that $\cap _{i=1}^{m}Fix\left( T_{i}\right) \ $is singleton. Assume on contrary that there exist $u$ and $v$ such that $u,v\in \cap _{i=1}^{m}Fix\left( T_{i}\right) $ but $u\neq v$. As $(u,v)\in \Delta _{2}$, so for (u_{x},v_{y})\in \Delta _{2}$ implie \begin{equation*} \tau +F\left( d(u,v)\right) \leq F(M_{2}(u,v;u,v)), \end{equation* wher \begin{eqnarray*} M_{2}(u,v;u,v) &=&\alpha d(u,v)+\beta d(u,u)+\gamma d(v,v) \\ &&+\delta _{1}d\left( u,v\right) +\delta _{2}d\left( v,u\right) \\ &=&\left( \alpha +\delta _{1}+\delta _{2}\right) d\left( x,y\right) , \end{eqnarray* that is \begin{eqnarray*} \tau +F\left( d(u,v)\right) &\leq &F\left( \left( \alpha +\delta _{1}+\delta _{2}\right) d\left( x,y\right) \right) \\ &\leq &F\left( d\left( u,v\right) \right) , \end{eqnarray* a contradiction as $\tau >0$.\ Hence $u=v$. Conversely, if $\cap _{i=1}^{m}Fix\left( T_{i}\right) $ is singleton, then it follows that $\cap _{i=1}^{m}Fix\left( T_{i}\right) $ is a well-ordered. $\square $\newline \noindent The following corollary extends Theorem 3.1 of \cite{Rus}, in the case of family of mappings in ordered metric space. \noindent \textbf{Corollary 2.5.} \ \ Let $(X,d,\preceq )$ be an ordered complete metric space and $\{T_{i}\}_{i=1}^{m}:X\rightarrow P_{cl}(X)$ be family of multivalued mappings. Suppose that for every $(x,y)\in \Delta _{1}$ and $u_{x}\in T_{i}(x),$ there exists $u_{y}\in T_{i+1}(y)$ for $i\in \{1,2,...,m\}$ (with $T_{m+1}=T_{1}$ by convention) such that, (u_{x},u_{y})\in \Delta _{2}$ implie \begin{equation} \tau +F\left( d(u_{x},u_{y})\right) \leq F(\alpha d\left( x,y\right) +\beta d(x,u_{x})+\gamma d(y,u_{y})]), \tag{2.5} \end{equation where $\tau $\ is a positive real number and $\alpha ,\beta ,\gamma \geq 0$ with $\alpha ,\beta ,\gamma \leq 1.$ Then the conclusions obtained in Theorem 2.3 remain true.\bigskip \noindent The following corollary extends Theorem 4.1 of \cite{LatifBeg}. \noindent \textbf{Corollary 2.6.} \ \ Let $(X,d,\preceq )$ be an ordered complete metric space and $\{T_{i}\}_{i=1}^{m}:X\rightarrow P_{cl}(X)$ be family of multivalued mappings. Suppose that for every $(x,y)\in \Delta _{1}$ and $u_{x}\in T_{i}(x),$ there exists $u_{y}\in T_{i+1}(y)$ for $i\in \{1,2,...,m\}$ (with $T_{m+1}=T_{1}$ by convention) such that, (u_{x},u_{y})\in \Delta _{2}$ implie \begin{equation} \tau +F\left( d(u_{x},u_{y})\right) \leq F(h[d(x,u_{x})+d(y,u_{y})]), \tag{2.6} \end{equation where $\tau $\ is a positive real number and $h\in \lbrack 0,\dfrac{1}{2}].$ Then the conclusions obtained in Theorem 2.3 remain true.\bigskip \noindent \textbf{Corollary 2.7.} \ \ Let $(X,d,\preceq )$ be an ordered complete metric space and $\{T_{i}\}_{i=1}^{m}:X\rightarrow P_{cl}(X)$ be family of multivalued mappings. Suppose that for every $(x,y)\in \Delta _{1}$ and $u_{x}\in T_{i}(x),$ there exists $u_{y}\in T_{i+1}(y)$ for $i\in \{1,2,...,m\}$ (with $T_{m+1}=T_{1}$ by convention) such that, (u_{x},u_{y})\in \Delta _{2}$ implie \begin{equation} \tau +F\left( d(u_{x},u_{y})\right) \leq F(d(x,y)), \tag{2.7} \end{equation where $\tau $\ is a positive real number. Then the conclusions obtained in Theorem 2.3 remain true.\bigskip \noindent The above corollary extends Theorem 4.1 of \cite{LatifBeg}.\bigskip \noindent \textbf{Conclusion.} Recently many results appeared in the literature giving the problems related to the common fixed point for multivalued maps. I this paper we obtained the results for existence of common fixed points of family of maps that satisfying generalized $F -contractions in ordered structured metric spaces. We presented some examples to show the validity of established results. \noindent \textbf{Acknowledgement.} The first author is grateful to the Erasmus Mundus project FUSION for the postdoctoral fellowship visiting to \"{a}lardalen University Sweden and to the Division of Applied Mathematics at the School of Education, Culture and Communication for creating excellent research environment.

https://arxiv.org/abs/1412.3841

Merging of Bézier curves with box constraints

In this paper, we present a novel approach to the problem of merging of Bézier curves with respect to the $L_2$-norm. We give illustrative examples to show that the solution of the conventional merging problem may not be suitable for further modification and applications. As in the case of the degree reduction problem, we apply the so-called restricted area approach -- proposed recently in (P. Gospodarczyk, Computer-Aided Design 62 (2015), 143--151) -- to avoid certain defects and make the resulting curve more useful. A method of solving the new problem is based on box-constrained quadratic programming approach.

\section{Introduction}\label{Sec:intro} Nowadays, people of various professions use different CAD systems. There are many ways to represent curves and surfaces, therefore, the exchange of geometric data between those systems often requires approximate conversion. As it was stated in~\cite{Hos87}, there are two main operations that should be considered: degree reduction and merging. In the past $30$ years, both problems have been extensively investigated. In this paper, we focus on the constrained merging of segments of a composite B\'ezier curve, i.e., we look for a single B\'ezier curve that approximates multiple adjacent B\'ezier curves and satisfies certain conditions. We propose the so-called \textit{box constraints}, which appear for the first time in the context of the merging problem. A conventional problem of merging is to approximate multiple adjacent B\'ezier curves with a single B\'ezier curve which minimizes a selected error function and satisfies some continuity constraints at the endpoints. Most of the papers deal with merging of only two B\'ezier curves (see \cite{HTJS01,Lu14,Lu13,THH03,ZW09}). Obviously, to merge more than two curves, one could use those algorithms repeatedly. However, such an approach increases the error of the approximation as well as the computational cost (see~\cite[\S1]{Lu15}). There are three methods that specialize in merging of more than two B\'ezier curves at the same time (see~\cite{CW08,Lu15,WGL}). Regardless of how many curves are merged, the most frequently used strategy is to solve a system of normal equations (see,~e.g.,~\cite{Lu15}). In~\cite{WGL}, one can observe a different approach which is based on the properties of the so-called constrained dual Bernstein basis polynomials (to our knowledge, this method is the fastest one available). The parametric (see,~e.g.,~\cite{CW08,Lu15,WGL}) or geometric (see,~e.g.,~\cite{Lu14,Lu15,ZW09}) continuity at the endpoints is preserved. In~\cite{Gos15}, one of us proposed a new approach to the problem of degree reduction of B\'ezier curves. The author noticed that as a result of the conventional degree reduction, the computed control points can be located far away from the plot of the curve. He also explained why this is a serious defect. Next, to eliminate this issue, he solved the degree reduction problem with constraints of a new type. In this paper, we show that the same observations may apply to the control points of the merged curve. Therefore, the main goal of this paper is to formulate a new problem of merging of B\'ezier curves. As in~\cite{Gos15}, the new approach requires completely different methods than in the case of the conventional one. The outline of the paper is as follows. Further on in this section, we give necessary definitions and notation. In Section~\ref{Sec:Pre}, we formulate the problem of merging of B\'{e}zier curves with box constraints. The example motivating the restricted area approach is given in Section~\ref{Sec:motiv}. Section~\ref{Sec:Mer} brings a solution of the new problem. Some illustrative examples are presented in Section~\ref{Sec:Ex}. For a brief summary of the paper, see Section~\ref{Sec:Conc}. Let $\Pi_m^d$ denote the space of all parametric polynomials in $\mathbb{R}^d$ of degree at most $m$; $\Pi_m:=\Pi^1_m$. Further on in the paper, we use $\mathbf{b}_{m,t} := \left[B_0^{m}(t),B_1^{m}(t),\ldots,B_m^{m}(t)\right]$, where \begin{equation*} B_i^{m}(t) := \binom{m}{i}t^i(1-t)^{m-i} \qquad (i = 0, 1,\ldots, m;\ m \in \mathbb{N}), \end{equation*} to denote the vector of \textit{Bernstein polynomial basis} in $\Pi_m$. We recall the well-known \textit{Gramian matrix} $\mathbf{G}_{m,n} := \left[g_{ij}\right] \in \mathbb{R}^{(m+1) \times (n+1)}$ of the Bernstein basis with the elements given by $$ g_{ij} := \frac{1}{m+n+1}\binom{m}{i}\binom{n}{j}\rbinom{m+n}{i+j} \qquad (i=0,1,\ldots,m;\ j=0,1,\ldots,n). $$ \textit{Forward difference operator} is defined by $$ \Delta^0q_i := q_i,\quad \Delta^jq_i := \Delta^{j-1}q_{i+1} - \Delta^{j-1}q_{i} \quad(j =1,2,\ldots). $$ Let $\mathbf{M} \in \mathbb{R}^{n \times m}$ be a matrix, and let $\mathcal{A} := \left\{i_1, i_2,\ldots, i_\alpha\right\} \subset \left[0,\,n-1\right]$, $\mathcal{B} := \left\{j_1, j_2,\ldots, j_\beta\right\} \subset \left[0,\,m-1\right]$ be the sets of natural numbers sorted in ascending order. Notation \begin{equation}\label{notMat} \mathbf{M}^{\mathcal{A}, \mathcal{B}} \end{equation} defines a matrix formed by rows $i_1+1, i_2+1,\ldots, i_\alpha+1$ and columns $j_1+1, j_2+1,\ldots, j_\beta+1$ of the matrix $\mathbf{M}$. Similarly, we use $\mathbf{v}^{\mathcal{A}}$, where $\mathbf{v}$ is a vector in $\mathbb{R}^{n}$. \section{Problem of merging of B\'{e}zier curves with box constraints}\label{Sec:Pre} In this section, we formulate the following new problem of merging of B\'{e}zier curves. \begin{prob} \,[\textsf{Merging of B\'ezier curves with box constraints}]\label{P:Problem}\\ Let $0=t_0<t_1<\ldots<t_s=1$ be a partition of the interval $[0,\,1]$. Let there be given a \emph{composite B\'ezier curve} $P(t)$ ($t\in[0,\,1]$) in $\R^d$, which in the interval $[t_{i-1},\,t_i]$ ($i=1,2,\ldots,s$) is exactly represented as a \emph{B\'ezier curve} $P^i(t) \in \Pi_{n_i}^d$, \[ P(t)=P^i(t):=\sum_{j=0}^{n_i}p^i_j\,B^{n_i}_j(u_i(t)) \equiv \mathbf{b}_{n_i,u_i(t)}\mathbf{p}^i \qquad (t_{i-1}\le t\le t_i), \] where $u_i(t) := \frac{t-t_{i-1}}{\Delta t_{i-1}}$, and $\mathbf{p}^i := \left[p^i_0, p^i_1,\ldots, p^i_{n_i}\right]^T$ with $p^i_j := \left(p^{i,1}_j,p^{i,2}_j,\ldots,p^{i,d}_j\right) \in \mathbb{R}^d$.\\ \noindent Find a B\'ezier curve $R(t) \in \Pi_{m}^d$, \[ R(t):=\sum_{j=0}^m r_j\,B^m_j(t) \equiv \mathbf{b}_{m,t}\mathbf{r} \qquad (0\le t\le 1), \] where $\mathbf{r} := \left[r_0, r_1,\ldots, r_m\right]^T$ with $r_j := \left(r_j^1,r_j^2,\ldots,r_j^d\right) \in \mathbb{R}^d$, satisfying the following conditions: \begin{itemize} \item[(i)] value of the squared \emph{$L_2$-error} \begin{equation}\label{min} E(\mathbf{r}) \equiv E^2_2 := \int_{0}^{1}\|P(t)-R(t)\|^2\dt, \end{equation} where $\|\cdot\|$ denotes the \emph{Euclidean vector norm} in $\mathbb{R}^d$, is minimized in the space $\Pi_m^d$; \item[(ii)] \emph{parametric continuity constraints at the endpoints} are satisfied, i.e., \begin{equation}\label{cont} \begin{array}{l} R^{(i)}(0)=P^{(i)}(0) \qquad (i=0,1,\ldots,k-1),\\[0.5ex] R^{(j)}(1)=P^{(j)}(1) \qquad (j=0,1,\ldots,l-1), \end{array} \end{equation} where $k \leq n_1+1$, $l \leq n_s+1$, and $k+l\leq m$; \item[(iii)] control points $r_j$ $(k \leq j \leq m-l)$ are located inside the specified $d$-dimensional cube including the edges, i.e., the following box constraints are fulfilled: \begin{equation}\label{box} c_h \leq r_{j}^h \leq C_h \qquad (j = k, k+1,\ldots, m-l;\ h= 1, 2,\ldots,d), \end{equation} where $c_h, C_h \in\mathbb R$. \end{itemize} \end{prob} Notice that in the case of degree reduction of B\'ezier curves, analogical problem was formulated (cf.~\cite[Problem~3.1]{Gos15}). \begin{remark}\label{R:Trad} Note that papers \cite{Lu15, WGL} deal with the minimization of~\eqref{min}, with the conditions~\eqref{cont}, but without the box constraints~\eqref{box}. In addition, a reasonable assumption that $m \ge \max_i{n_i}$ is made. Further on in this paper, such an approach is called the \textit{traditional merging} (cf.~\cite[Remark~3.2]{Gos15}). \end{remark} \section{Motivation of the paper}\label{Sec:motiv} As it turns out, the observations on the degree reduction problem (see~\cite[\S2]{Gos15}) also apply to the merging problem. In order to see the issue clearly, let us consider the following example. \begin{example}\label{Ex:1} We give the planar composite B\'{e}zier curve ``Ampersand'', with three fifth degree B\'{e}zier segments (see Figure~\ref{figure1a}), defined by the control points $\{(0.49, 0.07),$ $(0.43, 0.22),$ $(0.08, 0.67),$ $(0, 0.97),$ $(0.29, 0.98),$ $(0.36, 0.9)\}$, $\{(0.36, 0.9),$ $(0.43, 0.84),$ $(0.43, 0.68),$ $(0.25, 0.58),$ $(0.1, 0.36),$ $(0.09, 0.23)\}$, and $\{(0.09, 0.23),$ $(0.08, 0.13),$ $(0.14, 0.06),$ $(0.34, 0),$ $(0.52, 0.08),$ $(0.48,$ $0.23)\}$, respectively. Assuming that the partition of the interval $[0,\,1]$ is given by $t_0 = 0,\ t_1 \doteq 0.45,\ t_2 \doteq 0.76,\ t_3 = 1$; we look for a single B\'{e}zier curve being the result of the traditional merging for $m = 14$, $k = 3$, $l = 1$ (see Remark~\ref{R:Trad}). Figure~\ref{figure1b} shows the original composite curve and the merged curve. Clearly, the result of the approximation is very accurate. The errors are $E_2 = 5.49e{-}03$ and $E_{\infty} = 2.28e{-}02$, where $$ E_{\infty} := \max_{t \in D_M} \|P(t) - R(t)\| \approx \max_{0 \leq t \leq 1} \|P(t) - R(t)\| $$ with $D_M := \left\{0, 1/M, 2/M,\ldots, 1\right\}$ for $M := 500$. Observe also that the original control points are quite close to the plot of the curve (see Figure~\ref{figure1a}). In contrast, the resulting control points are located far away from the plot of the curve (see Figure~\ref{figure1c}). Note that we are unable to see the curve and its control points in one figure. Because of the non-intuitive location of the control points, further modeling of the merged curve is hard to imagine. A~designer that modifies the control points uses a convex hull property, which gives an intuition on shape and location of the curve. As it was stated in~\cite[\S2]{Gos15}, the size of the convex hull is a~measure of \textit{predictability} of the curve. Furthermore, let us recall that a small convex hull can be helpful while checking that two curves do not intersect, a curve and a surface do not intersect, a point does not lie on a curve. Observe that the convex hull of the resulting curve is huge, therefore, completely useless. Comparing this result with the ones from~\cite{Gos15}, we see that the defect seems to be even more significant (cf.~\cite[Figures 1b, 4b and 5b]{Gos15}). Now, let us impose some box constraints. We want the searched control points to be inside the specified rectangular area (including edges of the rectangle). Figure~\ref{figure2} presents the solution of Problem~\ref{P:Problem} for $m = 14$, $k = 3$, $l = 1$, $c_1 = -0.17$, $c_2 = 0$, $C_1 = 0.73$, $C_2 = 1.15$. Notice that the approximation is quite accurate (errors: $E_2 = 1.85e{-}02$ and $E_{\infty} = 6.10e{-}02$). Moreover, in this case, the computed control points are located much closer to the merged curve. As a result, the curve can be easily and intuitively modified by moving these points. What is more, we have obtained much smaller convex hull, which can be used to solve efficiently some important problems. More examples can be found in Section~\ref{Sec:Ex}. \begin{figure}[H] \captionsetup{margin=0pt, font={scriptsize}} \begin{center} \setlength{\tabcolsep}{0mm} \begin{tabular}{c} \subfloat[]{\label{figure1a}\includegraphics[width=0.32\textwidth]{ampersand1}} \subfloat[]{\label{figure1b}\includegraphics[width=0.302\textwidth]{ampersand2}} \subfloat[]{\label{figure1c}\includegraphics[width=0.281\textwidth]{ampersand3}} \end{tabular} \caption{Figure~(a) shows the original composite B\'ezier curve with its control points. Figure~(b) illustrates the original composite B\'ezier curve (blue solid line), and the merged B\'ezier curve (red dashed line) being the solution of the traditional merging problem. Figure~(c) presents the control points of the merged curve (red color).} \end{center} \end{figure} \end{example} \begin{figure}[H] \captionsetup{font=scriptsize} \centering \includegraphics[width=73mm]{ampersand4} \caption{The original composite B\'ezier curve (blue solid line with blue control points), and the merged B\'ezier curve (red dashed line with red control points) being the solution of Problem~\ref{P:Problem}. See the restricted area (black dotted-dashed frame).} \label{figure2} \end{figure} \section{Merging of B\'{e}zier curves with box constraints}\label{Sec:Mer} Now, we give the method of solving Problem~\ref{P:Problem}. First, we notice that some observations concerning the box-constrained degree reduction are also true in the case of the box-constrained merging. Clearly, a B\'ezier curve being the solution of Problem~\ref{P:Problem} can be obtained in a componentwise way (cf.~\cite[Remark~3.3]{Gos15}). \begin{comment} because $$ E(\mathbf{r}) = \sum_{h=1}^{d} E_h(\mathbf{r}_h), $$ where $\mathbf{r}_h := \left[r_0^h, r_1^h,\ldots, r_m^h\right]^T$ and $$ E_h(\mathbf{r}_h) := \int_{0}^{1}(P_h(t)-R_h(t))^2\dt $$\end{comment} Therefore, it is sufficient to describe our method in the case of $d = 1$. Further on in this section, we assume that $P^i \in \Pi_{n_i}$ $(i=1,2,\ldots,s)$, $R \in \Pi_m$, and $c, C \in \R$ are the lower and upper bounds for the box constraints~\eqref{box}. Next, we recall that the conditions~\eqref{cont} yield the following well-known formulas (see,~e.g.,~\cite[Theorem~3.1]{WGL}): \begin{align*} & \displaystyle r_j=\binom{n_1}{j}\binom{m}{j}^{\!-1}\,\Delta^jp^{1}_0- \sum_{h=0}^{j-1}(-1)^{j+h}\binom{j}{ h}r_{h}\qquad (j=0,1,\ldots,k-1),\\[1ex] & \displaystyle r_{m-j} =(-1)^j\binom{n_s}{j}\binom{m}{j}^{\!-1}\,\Delta^jp^{s}_{n_s-j}- \sum_{h=1}^{j}(-1)^h\binom{j}{ h}r_{m-j+h}\qquad (j=0,1,\ldots,l-1). \end{align*} What remains is to minimize $E(\mathbf{r})$ subject to the conditions~\eqref{box} for $d=1$. One can see clearly that $E(\mathbf{r})$ is a quadratic function. Therefore, in the following subsection, we are dealing with the so-called \textit{box-constrained quadratic programming problem}. \subsection{Quadratic programming with box constraints}\label{SubSubSec:QP} In this subsection, we use the quadratic programming approach to solve Problem~\ref{P:Problem}. Quadratic programming is an \textit{optimization problem} of minimizing or maximizing a quadratic \textit{objective function} of several variables subject to linear constraints on these variables. Taking into account the particular form of the restrictions~\eqref{box}, let us consider the following quadratic programming problem with box constraints: \begin{equation}\label{E:QP} \min_{}\;{\frac{1}{2}\mathbf{x}^T\mathbf{Q}\mathbf{x} + \mathbf{x}^T\mathbf{d}}, \qquad \mbox{s.t.} \qquad c \leq x_j \leq C, \end{equation} where $\mathbf{d}, \mathbf{x}:=\left[x_j\right] \in \mathbb{R}^{m-k-l+1}$ and $\mathbf{Q} \in \mathbb{R}^{(m-k-l+1) \times (m-k-l+1)}$. In our case, we set $\mathbf{x} := \mathbf{r}^{\mathcal{F}}$, where we define $\mathcal{F} := \left\{k, k+1,\ldots,m-l\right\}$ and use the notation of~\eqref{notMat}. Now, we will adjust $E(\mathbf{r})$ to the form~\eqref{E:QP}. First, taking into account that $P$ is a piecewise polynomial, we have to subdivide the searched polynomial $R$ as well. This can be done by applying the de Casteljau algorithm. In~\cite[\S2]{Lu15}, Lu gave the following formula: $$ R(t) = R^i(t) := \mathbf{b}_{m,u_i(t)}\mathbf{D}_{i}\mathbf{r} \qquad (t_{i-1}\le t\le t_i;\ i=1,2,\ldots,s), $$ where \begin{equation}\label{Eq:Di} \mathbf{D}_i := \mathbf{A}_1(t_{i-1}/t_i)\mathbf{A}_2(t_i) \end{equation} with $$ \mathbf{A}_1(\lambda) = \begin{bmatrix} B_0^m(\lambda) & B_1^m(\lambda) & \cdots & B_m^m(\lambda) \\ 0 & B_0^{m-1}(\lambda) & \cdots & B_{m-1}^{m-1}(\lambda) \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & 1 \end{bmatrix},\quad \mathbf{A}_2(\lambda) = \begin{bmatrix} 1 & 0 & \cdots & 0 \\ B_0^1(\lambda) & B_1^1(\lambda) & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ B_0^m(\lambda) & B_1^m(\lambda) & \cdots & B_m^m(\lambda) \end{bmatrix}. $$ \begin{remark}\label{R:d} According to~\cite[Lemmas~2.5,\;2.4]{WGL}, $\mathbf{D}_i = \left[d^{(i)}_{jh}\right] \in \mathbb{R}^{(m+1)\times(m+1)}$, where the entries $d^{(i)}_{jh}$ $(i = 1, 2,\ldots, s;\ j = 0, 1,\ldots, m;\ h = 0, 1,\ldots, m)$ satisfy the following recurrence relation: \begin{multline*} \Delta t_{i-1}\left[(m-j+1)d_{j-1,h}^{(i)}+(2j-m)d_{jh}^{(i)}-(j+1)d_{j+1,h}^{(i)}\right]\\ \qquad\quad = (m-h)d_{j,h+1}^{(i)} + (2h-m)d_{jh}^{(i)} - hd_{j,h-1}^{(i)}\\ (1 \leq j \leq m-1;\ 0 \leq h \leq m). \end{multline*} Therefore, one can avoid matrix multiplications and compute the matrices $\mathbf{D}_1,\mathbf{D}_2,\ldots,\mathbf{D}_s$ efficiently, with the complexity $O(sm^2)$, using~\cite[Algorithm~4.1]{WGL}. Observe that the direct use of~\eqref{Eq:Di} results in the complexity $O(sm^3)$. \end{remark} Next, assuming that $\mathcal{K} := \left\{0,1,\ldots,m\right\}$, $\mathcal{C} := \mathcal{K}\setminus \mathcal{F}$ and using the notation of~\eqref{notMat}, we write (cf.~\cite[(11)]{Lu15}) \begin{comment}we split $\mathbf{r}_x$, $\mathbf{b}_{m,u_i(t)}$ and $\mathbf{D}_i$ into $\mathbf{r}^{\mathcal{F}}_x, \mathbf{r}^{\mathcal{C}}_x$, $\mathbf{b}_{m,u_i(t)}^{\mathcal{F}}, \mathbf{b}_{m,u_i(t)}^{\mathcal{C}}$ and $\mathbf{D}_i^{\mathcal{K},\mathcal{F}}$, $\mathbf{D}_i^{\mathcal{K},\mathcal{C}}$, respectively. Then, \end{comment} \begin{align*} E(\mathbf{r}) &= \int_{0}^{1}(P(t)-R(t))^2\dt = \sum_{i=1}^s\int_{t_{i-1}}^{t_i}\left(P^i(t)-R^i(t)\right)^2\dt \\ &= \sum_{i=1}^s\Delta t_{i-1} \int_{0}^{1} \left(\mathbf{b}_{n_i,v}\mathbf{p}^{i} - \mathbf{b}_{m,v}\mathbf{D}_i^{\mathcal{K},\mathcal{C}}\mathbf{r}^{\mathcal{C}} - \mathbf{b}_{m,v}\mathbf{D}_i^{\mathcal{K},\mathcal{F}}\mathbf{r}^{\mathcal{F}}\right)^2\dv \\ &= \frac{1}{2}\left(\mathbf{r}^{\mathcal{F}}\right)^T\mathbf{Q}\mathbf{r}^{\mathcal{F}} + \left(\mathbf{r}^{\mathcal{F}}\right)^T\mathbf{d} + a \ =: \ g\left(\mathbf{r}^{\mathcal{F}}\right) + a, \end{align*} where \begin{align*} & \mathbf{Q} := 2\sum_{i=1}^s\Delta t_{i-1}\left(\mathbf{D}_i^{\mathcal{K},\mathcal{F}}\right)^T\mathbf{G}_{m,m}\mathbf{D}_i^{\mathcal{K},\mathcal{F}},\\ & \mathbf{d} := 2\sum_{i=1}^s\Delta t_{i-1}\left(\mathbf{D}_i^{\mathcal{K},\mathcal{F}}\right)^T \left(\mathbf{G}_{m,m}\mathbf{D}_i^{\mathcal{K},\mathcal{C}}\mathbf{r}^{\mathcal{C}} - \mathbf{G}_{m,n_i}\mathbf{p}^{i}\right), \end{align*} and $a \in \mathbb{R}$ is a certain constant term. Obviously, $a$ is meaningless in the minimization process, therefore, the \textit{significant terms} of $E(\mathbf{r})$ are given by $g(\mathbf{r}^{\mathcal{F}})$, which is written in the form~\eqref{E:QP}. \begin{remark}\label{R:Sol} Matrix $\mathbf{Q}$ is positive definite (see~\cite[\S3.1]{Lu15}), therefore, the objective function $g$ is strictly convex. Furthermore, the feasible set is nonempty, closed and convex. We conclude that the quadratic programming problem has a~unique solution (see, e.g.,~\cite[Proposition~2.5]{DOS}) and so does Problem~\ref{P:Problem}. In contrast, a solution of the analogical degree reduction problem may not be unique (cf.~\cite[Theorem~4.1]{Gos15}). The difference is that, in the present paper, we consider the \textit{continuous inner product} (see~\eqref{min}) instead of the \textit{discrete inner product} (see~\cite[(3.1)]{Gos15}). \end{remark} There are many papers dealing with the box-constrained quadratic programming problem. To solve it, one can use a variety of strategies, including \textit{active set methods} (see, e.g.,~\cite{Fer98}) and \textit{interior point algorithms} (see, e.g.,~\cite{Han90}). Some of the approaches combine the active set strategy with \textit{gradient projection method} (see, e.g.,~\cite{Mor89}). For extensive lists of references, see the mentioned papers. \section{Examples}\label{Sec:Ex} In this section, we apply our method to the composite B\'ezier curves in $\R^2$. As in~\cite{WGL}, we generalize the approach of \cite{Lu14} and obtain a partition of the interval $[t_0,\,t_s]=[0,\,1]$ according to the lengths of segments $P^i$: \begin{equation}\label{E:part} t_j :=L_j/L_s \qquad (j=1,2,\ldots,s-1), \end{equation} where \[ L_q:=\sum_{i=1}^{q}\int_{0}^{1}\left\Vert\frac{\der}{\dt}\sum_{h=0}^{n_i}p^i_hB^{n_i}_h(t)\right\Vert\dt. \] Integrals are evaluated using $\mbox{Maple}{\small \texttrademark}$ \texttt{int} procedure with the option \texttt{numeric}. A solution of the traditional merging problem (see Remark~\ref{R:Trad}) is computed using~\cite[Algorithm~4.2]{WGL}. The complexity of this algorithm is $O(sm^2)$ which, to our knowledge, is significantly less than cost of other methods of merging with the constraints~\eqref{cont} (cf.~\cite{CW08,Lu15}). To solve the quadratic programming problem with box constraints~\eqref{E:QP}, we use the matrix version of Maple{\small \texttrademark} \texttt{QPSolve} command. It is worth noting that this procedure implements an iterative active set method and it is suited for the box constraints, i.e., the vectors of lower and upper bounds can be passed using the optional parameter \texttt{bd}. According to the documentation provided by $\mbox{Maplesoft}{\small \texttrademark}$, in the case of the convex optimization, a global minimum is returned (cf. Remark~\ref{R:Sol}). For the initial point, we choose the lower bounds, i.e., $c_1$ and $c_2$. The results have been obtained on a computer with \texttt{Intel Core i5-3337U 1.8GHz} processor and \texttt{8GB} of \texttt{RAM}, using $24$-digit arithmetic. $\mbox{Maple}{\small \texttrademark}13$ worksheet containing programs and tests is available at \url{http://www.ii.uni.wroc.pl/~pgo/papers.html}. \begin{example}\label{Ex:2} We introduce the composite B\'ezier curve ``D'' (see Figure~\ref{figure3a}), formed by three cubic segments which are defined by the control points $\{(0.32, 0.81),$ $(0.26, 0.59),$ $(0.18, 0),$ $(0.06, 0.27)\}$, $\{(0.06, 0.27),$ $(0, 0.42),$ $(0.42, 0.08),$ $(0.57, 0.25)\}$ and $\{(0.57, 0.25),$ $(0.76, 0.46),$ $(0.8,$ $1),$ $(0.22, 0.85)\}$, respectively. Formula \eqref{E:part} implies $t_0 = 0,\ t_1 \doteq 0.32,\ t_2 \doteq 0.57,\ t_3 = 1$. Figure~\ref{figure3b} shows the result of the traditional merging for $m = 18$, $k = 1$, $l = 2$. The merged curve looks like a perfect approximation (errors: $E_2 = 3.35e{-}03$ and $E_{\infty}= 9.57e{-}03$), unfortunately, it suffers from the defect described in Section~\ref{Sec:motiv} (see Figure~\ref{figure3c}). To avoid this, we solve Problem~\ref{P:Problem} for $m = 18$, $k = 1$, $l = 2$, with the following box constraints: \begin{equation}\label{Eq:Res1} \begin{aligned} &c_1 := \min_{1 \leq i \leq s}\min_{0 \leq j \leq n_i}p_j^{i,1} - 0.2 = -0.2,\qquad &&C_1 := \max_{1 \leq i \leq s}\max_{0 \leq j \leq n_i}p_j^{i,1} = 0.8,\\ &c_2 := \min_{1 \leq i \leq s}\min_{0 \leq j \leq n_i}p_j^{i,2} - 0.3 = -0.3,\qquad &&C_2 := \max_{1 \leq i \leq s}\max_{0 \leq j \leq n_i}p_j^{i,2} = 1 \end{aligned} \end{equation} (cf.~\eqref{box}), and obtain the curve shown in Figure~\ref{figure3d} (errors: $E_2 = 1.38e{-}02$ and $E_{\infty}= 2.98e{-}02$). Compare Figure~\ref{figure3d} with Figure~\ref{figure3c} to see a big difference in the location of the resulting control points. Obviously, the curve in Figure~\ref{figure3d} is much more satisfying in this regard. \begin{figure}[H] \captionsetup{margin=0pt, font={scriptsize}} \begin{center} \setlength{\tabcolsep}{0mm} \begin{tabular}{c} \subfloat[]{\label{figure3a}\includegraphics[width=0.346\textwidth]{D1}} \subfloat[]{\label{figure3b}\includegraphics[width=0.388\textwidth]{D2}}\\ \subfloat[]{\label{figure3c}\includegraphics[width=0.65\textwidth]{D3}}\\ \subfloat[]{\label{figure3d}\includegraphics[width=0.346\textwidth]{D4}} \end{tabular} \caption{Merging of three segments of the composite B\'{e}zier curve. The original composite curve (blue solid line with blue control points) and the merged curve (red dashed line with red control points), parameters: $m = 18$, $k = 1$, $l = 2$. Figure~(a) shows the original composite curve with its control points. Figure~(b) illustrates the curve being the solution of the traditional merging problem. Figure~(c) presents the control points of the merged curve shown in Figure~(b). The curve being the solution of~Problem~\ref{P:Problem} with the resulting control points and the restricted area (black dotted-dashed frame) are shown in Figure~(d).} \end{center} \end{figure} \end{example} \begin{example}\label{Ex:3} Now, we consider the composite B\'ezier curve with four fifth degree B\'{e}zier segments (see Figure~\ref{figure4a}). For the original control points, see~\cite[Example 3]{Lu15}. To place the curve inside the unit box, we have divided each coordinate of the control points by $5.1$. According to~\eqref{E:part}, we get $t_0 = 0,\ t_1 \doteq 0.24,\ t_2 \doteq 0.49,\ t_3 \doteq 0.76,\ t_4 = 1$. As a result of the traditional merging ($m = 19$, $k = l = 1$), we obtain the B\'ezier curve which is illustrated in Figure~\ref{figure4b}. Once again, we get a good approximation (errors: $E_2 = 2.08e{-}03$ and $E_{\infty}= 5.65e{-}03$), however, the resulting control points are located far away from the plot of the curve (see Figure~\ref{figure4c}). Taking into account the axis scale in Figure~\ref{figure4c}, we conclude that this example seems to be extremely difficult. Nonetheless, the solution of Problem~\ref{P:Problem} for $m = 19$, $k = l = 1$, with the box constraints \begin{equation*} \begin{aligned} &c_1 := \min_{1 \leq i \leq s}\min_{0 \leq j \leq n_i}p_j^{i,1} - 0.2 = -0.2,\qquad &&C_1 := \max_{1 \leq i \leq s}\max_{0 \leq j \leq n_i}p_j^{i,1} + 0.2 \doteq 0.65,\\ &c_2 := \min_{1 \leq i \leq s}\min_{0 \leq j \leq n_i}p_j^{i,2} - 0.2 = -0.2,\qquad &&C_2 := \max_{1 \leq i \leq s}\max_{0 \leq j \leq n_i}p_j^{i,2} + 0.2 = 1.2 \end{aligned} \end{equation*} is quite decent (errors: $E_2 = 9.71e{-}03$ and $E_{\infty}= 1.90e{-}02$). See Figure~\ref{figure4d}. \end{example} \begin{figure}[H] \captionsetup{margin=0pt, font={scriptsize}} \begin{center} \setlength{\tabcolsep}{0mm} \begin{tabular}{c} \subfloat[]{\label{figure4a}\includegraphics[width=0.386\textwidth]{clef1}} \subfloat[]{\label{figure4b}\includegraphics[width=0.388\textwidth]{clef2}}\\ \subfloat[]{\label{figure4c}\includegraphics[width=0.439\textwidth]{clef3}} \subfloat[]{\label{figure4d}\includegraphics[width=0.388\textwidth]{clef4}} \end{tabular} \caption{Merging of four segments of the composite B\'{e}zier curve. The original composite curve (blue solid line with blue control points) and the merged curve (red dashed line with red control points), parameters: $m = 19$, $k = l = 1$. Figure~(a) shows the original composite curve with its control points. Figure~(b) illustrates the curve being the solution of the traditional merging problem. Figure~(c) presents the control points of the merged curve shown in Figure~(b). The curve being the solution of~Problem~\ref{P:Problem} with the resulting control points and the restricted area (black dotted-dashed frame) are shown in Figure~(d).} \end{center} \end{figure} \begin{remark}\label{R:Res} As stated in~\cite[Remark~6.3]{Gos15}, selection of the restricted area is a difficult issue. The choice always depends on the considered example and on the precision level that we accept as satisfactory. However, there is a strategy that seems to work quite well for the given examples. To explain this procedure, let us revisit Example~\ref{Ex:2}. At the beginning, we set \begin{equation}\label{Eq:DBox1} \begin{aligned} &c^{(1)}_1 := \min_{1 \leq i \leq s}\min_{0 \leq j \leq n_i}p_j^{i,1}=0,\qquad &&C^{(1)}_1 := \max_{1 \leq i \leq s}\max_{0 \leq j \leq n_i}p_j^{i,1}=0.8,\\ &c^{(1)}_2 := \min_{1 \leq i \leq s}\min_{0 \leq j \leq n_i}p_j^{i,2}=0,\qquad &&C^{(1)}_2 := \max_{1 \leq i \leq s}\max_{0 \leq j \leq n_i}p_j^{i,2}=1. \end{aligned} \end{equation} Consequently, the resulting control points will be bounded by the outermost control points of the original curves. Unfortunately, the obtained curve is unsatisfactory (see~Figure~\ref{figure5a}). Next, to improve this result, we must expand the restricted area. Intuition tells us that we should try to move the borders with the highest numbers of the control points. We consider \begin{equation}\label{Eq:DBox2} \begin{aligned} c^{(2)}_1 := c^{(1)}_1 - 0.04w_1 \doteq -0.05,\qquad C^{(2)}_1 := C^{(1)}_1,\\ c^{(2)}_2 := c^{(1)}_2 - 0.04w_1 \doteq -0.05,\qquad C^{(2)}_2 := C^{(1)}_2, \end{aligned} \end{equation} where $$ w_i := \sqrt{\left(C^{(i)}_1 - c^{(i)}_1\right)^2+\left(C^{(i)}_2 - c^{(i)}_2\right)^2} $$ is the diagonal length of $i$-th restricted area. Notice that the error is now lower (see~Figure~\ref{figure5b} and Table~\ref{tab:table1}). Therefore, we should try to make another step in the same direction. This time, the expansion is greater, i.e., we set \begin{equation}\label{Eq:DBox3} \begin{aligned} c^{(3)}_1 := c^{(2)}_1 - 0.08w_2 \doteq -0.16,\qquad C^{(3)}_1 := C^{(2)}_1,\\ c^{(3)}_2 := c^{(2)}_2 - 0.08w_2 \doteq -0.16,\qquad C^{(3)}_2 := C^{(2)}_2. \end{aligned} \end{equation} The result can be seen in Figure~\ref{figure5c}. See also Table~\ref{tab:table1}. Observe that, in Example~\ref{Ex:2}, the restricted area~\eqref{Eq:Res1} is even larger. \begin{comment}, i.e., \begin{equation*} \begin{aligned} &c^{(3)}_1 - 0.03w_3 \doteq c_1, &\qquad C^{(3)}_1 = C_1,\\ &c^{(3)}_2 - 0.095w_3 \doteq c_2, &\qquad C^{(3)}_2 = C_2. \end{aligned} \end{equation*} \end{comment} See Figure~\ref{figure3d}. Taking into account that \texttt{QPSolve} is an iterative method which we apply separately for each coordinate, pairs of numbers of iterations are also given in Table~\ref{tab:table1}. According to our experiments, if the control points of the optimal solution of the traditional merging are located very far away from the plot of the curve (see Figures~\ref{figure1c}, \ref{figure3c} and \ref{figure4c}), then it is difficult to find a satisfying solution of Problem~\ref{P:Problem}. For that reason, the examples given in this paper are much more demanding than the ones presented in~\cite{Gos15}. Moreover, note that in the case of the box-constrained merging, majority of the resulting control points are located on borders (see Figures~\ref{figure2}, \ref{figure3d} and \ref{figure4d}). Regardless of choice of the restricted area, one should realize that because of the additional constraints~\eqref{box}, approximation error must be inevitably larger than for the traditional approach. \begin{table}[H] \captionsetup{margin=0pt, font={scriptsize}} \ra{1.3} \centering \scalebox{1}{ \begin{tabular}{@{}ccccccc@{}} \toprule Box constraints & \phantom{ab} & $E_2$ & \phantom{ab} & $E_{\infty}$ &\phantom{ab} & Iterations \\ \midrule \eqref{Eq:DBox1} & & $2.25e{-}02$ & & $5.54e{-}02$ & & $(19,19)$\\ \eqref{Eq:DBox2} & & $1.86e{-}02$ & & $4.14e{-}02$ & & $(19,19)$\\ \eqref{Eq:DBox3} & & $1.51e{-}02$ & & $3.30e{-}02$ & & $(17,18)$\\ \eqref{Eq:Res1} & & $1.38e{-}02$ & & $2.98e{-}02$ & & $(22,29)$\\ \hline \end{tabular}} \caption{$L_2$-errors, maximum errors and numbers of iterations for merging of three segments of the composite B\'{e}zier curve ``D'' with box constraints. Parameters: $m = 18$, $k = 1$, $l = 2$.} \label{tab:table1} \end{table} \begin{figure}[H] \captionsetup{margin=0pt, font={scriptsize}} \begin{center} \setlength{\tabcolsep}{0mm} \begin{tabular}{c} \subfloat[]{\label{figure5a}\includegraphics[width=0.335\textwidth]{D5}} \subfloat[]{\label{figure5b}\includegraphics[width=0.335\textwidth]{D6}} \subfloat[]{\label{figure5c}\includegraphics[width=0.33\textwidth]{D7}} \end{tabular} \caption{Merging of three segments of the composite B\'{e}zier curve. The original composite curve (blue solid line with blue control points) and the merged curve (red dashed line with red control points) satisfying the following box constraints (black dotted-dashed frames): \eqref{Eq:DBox1} (see Figure~(a)), \eqref{Eq:DBox2} (see Figure~(b)) and \eqref{Eq:DBox3} (see Figure~(c)). Parameters: $m = 18$, $k = 1$, $l = 2$.} \end{center} \end{figure} \end{remark} \begin{remark}\label{R:Time} To solve the box-constrained quadratic programming problem~\eqref{E:QP}, one can choose a method provided by a software library of a selected programming language or implement one of the algorithms given in~\cite{Fer98,Han90,Mor89}. For that reason, the running times strongly depend on the implementation of the selected method. However, regardless of the choice, the box constraints make Problem~\ref{P:Problem} more difficult to solve. Therefore, the running times of methods dealing with the new problem must be longer than in the case of the traditional merging. See the comparison given in Table~\ref{tab:table2}. \begin{table}[H] \captionsetup{margin=0pt, font={scriptsize}} \centering \ra{1.3} \scalebox{1}{ \begin{tabular}{@{}clcccc@{}} \toprule & \phantom{a} & Traditional merging & \phantom{a} & \multicolumn{2}{c}{Problem~\ref{P:Problem}}\\ \cmidrule{3-3} \cmidrule{5-6} & & Running times [ms] & & Running times [ms] & Iterations \\ \midrule Example~\ref{Ex:1} & & $29$ & & $224$ & $(15,17)$\\ Example~\ref{Ex:2} & & $50$ & & $414$ & $(22,29)$\\ Example~\ref{Ex:3} & & $72$ & & $690$ & $(18,27)$\\ \hline \end{tabular}} \caption{Running times of the traditional and box-constrained merging of B\'ezier curves.} \label{tab:table2} \end{table} \end{remark} \section{Conclusions}\label{Sec:Conc} The new approach to the problem of merging of B\'ezier curves is introduced. We propose constraints of the new type and explain the purpose of those restrictions. A curve being the solution of Problem~\ref{P:Problem} is suitable for further modification and applications. Moreover, the resulting convex hull is much smaller than the one obtained using the traditional approach. Consequently, it can be helpful while solving some important problems. These positive attributes make the new problem worth of consideration, despite the inevitably longer running times and the larger approximation errors, which are also unavoidable. What is more, the comparison of the results with the ones from~\cite{Gos15}, leads to a conclusion that in the case of the traditional merging, the defect described in Section~\ref{Sec:motiv} is even more significant. In the near future, the authors intend to study a more general version of Problem~\ref{P:Problem} with the geometric continuity constraints instead of the conditions~\eqref{cont}. Furthermore, a different strategy of setting the restricted area could also improve the results. \section*{Acknowledgments} The authors are grateful to the referees for their remarks which helped to improve the paper.

https://arxiv.org/abs/1503.05402

Orthogonal polynomials on the disk in the absence of finite moments

We introduce a new family of orthogonal polynomials on the disk that has emerged in the context of wave propagation in layered media. Unlike known examples, the polynomials are orthogonal with respect to a measure all of whose even moments are infinite.

\section{Introduction} For each $\alpha>-1$ there is a corresponding family of disk polynomials that are orthogonal with respect to the measure $(1-x^2-y^2)^\alpha dxdy\,$ on the unit disk $\mathbb{D}$; these are sometimes referred to as generalized Zernike polynomials, named for the case $\alpha=0$ introduced in \cite{Ze:1934}. The well-established theory of disk polynomials is detailed in \cite{Ko:1975, Wu:2005, DuXu:2001}. The constraint $\alpha>-1$ stems from the requirement that the measure $(1-x^2-y^2)^\alpha dxdy\,$ have finite moments, which is necessary for meaningful evaluation of the corresponding scalar product \begin{equation}\label{scalar} \langle p,q\rangle_\alpha=\int_{\mathbb{D}}p(x,y)\overline{q(x,y)}(1-x^2-y^2)^\alpha dxdy \end{equation} on arbitrary polynomials $p$ and $q$. Recent work on the propagation of waves in layered media \cite{Gi:SIAP2014, Gi:Smooth2014} has brought to light a family of polynomials orthogonal with respect to $(1-x^2-y^2)^{-1}dxdy$. Since \begin{equation}\label{moments} \int_{\mathbb{D}}x^{2m}y^{2n}(1-x^2-y^2)^{-1} dxdy=\infty \end{equation} for every pair of nonnegative integers $m$ and $n$, the scalar product (\ref{scalar}) is not defined for arbitrary polynomials in the case $\alpha=-1$. Nevertheless, polynomials---which we term scattering polynomials---comprise an orthogonal basis for $L^2\bigl(\mathbb{D},dxdy/(1-x^2-y^2)\bigr)$. The purpose of the present paper is to present the details of this result. \section{Definition and properties of scattering polynomials\label{sec-main}} Referring to the notation $z=x+iy$ for points in the unit disk $\mathbb{D}$, one has the option of working with euclidean $x,y$-coordinates, or with complex coordinates $z$ and $\bar{z}$. As far as orthogonal polynomials are concerned these are essentially equivalent, as elaborated in \cite{Xu:2015}; the present paper uses whichever coordinates are most convenient for the task at hand. We define \emph{scattering polynomials} by a Rodrigues type formula, as follows. For every $(p,q)\in\mathbb{Z}^2$ with min$\{p,q\}\geq 1$, set \begin{equation}\label{defn-scattering} \varphi^{(p,q)}(z)=\frac{(-1)^p}{q(p+q-1)!}(1-z\bar{z})\frac{\partial^{p+q}}{\partial z^p\partial \bar{z}^q}(1-z\bar{z})^{p+q-1}. \end{equation} The chosen normalization simplifies the formulation of the boundary Green's function for scattering in layered media (see \cite{Gi:Smooth2014}) and so is physically natural, although not important for present considerations. Note that disk polynomials satisfy a Rodrigues formula similar to that of scattering polynomials, but there is a qualitative difference: it follows directly from (\ref{defn-scattering}) that $\varphi^{(p,q)}(z)=0$ for every $z$ on the unit circle $\mathbb{T}$, whereas all disk polynomials have constant non-zero modulus on $\mathbb{T}$, cf.~\cite{Ko:1975}. Our main result concerns completeness of scattering polynomials, as follows. \begin{thm}\label{thm-main} Scattering polynomials $\varphi^{(p,q)}$ defined by (\ref{defn-scattering}), where $(p,q)\in\mathbb{Z}^2$ and $\min\{p,q\}\geq 1$, comprise an orthogonal basis for $L^2\bigl(\mathbb{D},dxdy/(1-x^2-y^2)\bigr)$. \end{thm} In \S\ref{sec-eigenfunctions} and \S\ref{sec-Jacobi} below we show that scattering polynomials are eigenfunctions of a second order differential operator and may be expressed in terms of Jacobi polynomials; these results contribute to a proof of Theorem~\ref{thm-main} completed in \S\ref{sec-completeness}. \subsection{Eigenfunctions of $-(1-x^2-y^2)\Delta/4$\label{sec-eigenfunctions}} \pagestyle{fancyplain} Let $\widetilde{\Delta}$ denote the modified laplacian \begin{equation}\label{mlap} \widetilde{\Delta}=(1-z\bar{z})\frac{\partial^2}{\partial z\partial \bar{z}}=\frac{1-x^2-y^2}{4}\Delta, \end{equation} where $\Delta$ is the usual (euclidean) laplacian. Direct computation using (\ref{defn-scattering}) shows that for all integers $p,q\geq 1$, \begin{equation}\label{eigenfunctions} -\widetilde{\Delta}\varphi^{(p,q)}=pq\,\varphi^{(p,q)}. \end{equation} Letting $\sigma_0:\mathbb{Z}_+\rightarrow\mathbb{Z}_+$ denote the divisor function, there is thus a family of $\sigma_0(k)$ eigenfunctions of $-\widetilde{\Delta}$ of the form $\varphi^{(p,q)}$ corresponding to each positive integer eigenvalue $k$. We show in the next section that these eigenfunctions are linearly independent. \subsection{Representation in terms of {J}acobi polynomials\label{sec-Jacobi}} Like disk polynomials, scattering polynomials have a representation in terms of Jacobi polynomials, but again, there is a qualitative difference. The disk polynomials corresponding to parameter $\alpha>-1$ can be expressed in terms of Jacobi polynomials $P^{(\alpha,\beta)}_n$ for nonnegative integer values of $\beta$. Since there is no Jacobi polynomial corresponding to $\alpha=-1$, the same cannot be true for scattering polynomials. Indeed it turns out that scattering polynomials can be formulated in terms of $P^{(1,\beta)}_n$, where $\beta$ is a nonnegative integer, as follows. Expanding the binomial $(1-z\bar{z})^{p+q-1}$ in the formula (\ref{defn-scattering}), and then applying the derivative $\partial^{p+q}/\partial z^p\partial\bar{z}^q$, yields \begin{equation}\label{step-one} \varphi^{(p,q)}(z)= \frac{(-1)^{q+\nu+1}}{q}(1-z\bar{z})z^{m+\nu-p+1}\bar{z}^{m+\nu-q+1}\sum_{j=0}^{\nu}(-1)^j\frac{(j+\nu+m+1)!}{j!(j+m)!(\nu-j)!}(z\bar{z})^j, \end{equation} where $m=|p-q|$ and $\nu=\min\{p,q\}-1$; the latter notation will be used in the remainder of this section. Switching to polar form $z=re^{i\theta}$, it follows from (\ref{step-one}) that \begin{equation}\label{step-two} \varphi^{(p,q)}\bigl(re^{i\theta}\bigr)=e^{i(q-p)\theta}f^{(p,q)}(r), \end{equation} where \begin{equation}\label{fpq} f^{(p,q)}(r)=\frac{(-1)^{q+\nu+1}}{q}(1-r^2)r^m\sum_{j=0}^{\nu}(-1)^j\frac{(j+\nu+m+1)!}{j!(j+m)!(\nu-j)!}r^{2j}. \end{equation} The radial functions $f^{(p,q)}$ were first discovered in \cite{Gi:SIAP2014}, as was the following connection to Jacobi polynomials, valid for $\nu\geq 0$: \begin{equation}\label{Jacobi} f^{(p,q)}(r)=\frac{(-1)^{q+m+\nu+1}(m+\nu+1)}{q}(1-r^2)r^mP^{(1,m)}_{\nu}(2r^2-1). \end{equation} Combined with (\ref{step-two}) this yields the representation \begin{equation}\label{representation} \varphi^{(p,q)}\bigl(re^{i\theta}\bigr)=\frac{(-1)^{q+\max\{p,q\}}\max\{p,q\}}{q}(1-r^2)r^{|p-q|}P^{(1,|p-q|)}_{\min\{p,q\}-1}(2r^2-1)e^{i(q-p)\theta}. \end{equation} Note that the angular part of $\varphi^{(p,q)}(re^{i\theta})$, namely $e^{i(q-p)\theta}$, is a pure frequency. Therefore if $ q-p\neq q^\prime-p^\prime, $ then $\varphi^{(p,q)}$ and $\varphi^{(p^\prime,q^\prime)}$ are orthogonal in \[ L^2\Bigl(\mathbb{D},\frac{dxdy}{1-x^2-y^2}\Bigr)=L^2\Bigl(\mathbb{D},\frac{rdrd\theta}{1-r^2}\Bigr). \] In particular, if $pq=p^\prime q^\prime$ and $(p,q)\neq(p^\prime,q^\prime)$, then $\varphi^{(p,q)}$ and $\varphi^{(p^\prime,q^\prime)}$ are orthogonal, so the set of scattering polynomials corresponding to any fixed eigenvalue of $-\widetilde{\Delta}$ is linearly independent. \subsection{Completeness in $L^2\!\left(\mathbb{D},\frac{r\,drd\theta}{1-r^2}\right)$\label{sec-completeness}} In general, given a measure $\mu$ on a locally compact metric space $X$ and a positive measurable weight function $w:X\rightarrow\mathbb{R}_+$, \begin{equation}\label{L2} L^2(X,w\,d\mu)=\frac{1}{\sqrt{w}}L^2(X,d\mu), \end{equation} and a sequence $\{b_\nu\}_{\nu=0}^\infty$ is an orthogonal basis for $L^2(X,w\,d\mu)$ if and only if the corresponding sequence $\{\sqrt{w}b_\nu\}_{\nu=0}^\infty$ is an orthogonal basis for $L^2(X,d\mu)$. In particular, setting $d\mu=rdrd\theta/(1-r^2)$, \begin{equation}\label{L2-relation} L^2(\mathbb{D},d\mu)=\sqrt{1-r^2}L^2(\mathbb{D},rdrd\theta). \end{equation} Also, since for any nonnegative integer $m$, $\left\{P^{(1,m)}_\nu(u)\right\}_{\nu=0}^\infty$ is an orthogonal basis for \[ L^2\bigl([-1,1],(1-u)(1+u)^m\,du\bigr), \] it follows that the quasipolynomials \begin{equation}\label{Q} Q^{(1,m)}_\nu(u)=\left(\frac{1-u}{2}\right)^{\frac{1}{2}}\left(\frac{1+u}{2}\right)^{\frac{m}{2}}P^{(1,m)}_\nu(u) \end{equation} comprise an orthogonal basis for $L^2([-1,1],du)$; see \cite{Sz:1975}. In order to show that \begin{equation}\label{basis} \mathcal{B}=\left\{\varphi^{(p,q)}\,\left|\,(p,q)\in\mathbb{Z}^2\;\&\;\min\{p,q\}\geq 1\right.\right\} \end{equation} is an orthogonal basis of $L^2\bigl(\mathbb{D},rdrd\theta/(1-r^2)\bigr)$, we first argue that the functions $\varphi^{(p,q)}$ are orthogonal, and then that the span of $\mathcal{B}$ is dense. It was proven in \S\ref{sec-Jacobi} that $\varphi^{(p,q)}$ and $\varphi^{(p^\prime,q^\prime)}$ are orthogonal if $pq=p^\prime q^\prime$ and $(p,q)\neq(p^\prime,q^\prime)$. On the other hand, if $pq\neq p^\prime q^\prime$, then orthogonality of $\varphi^{(p,q)}$ and $\varphi^{(p^\prime,q^\prime)}$ follows from the fact that they are eigenfunctions, corresponding to distinct eigenvalues, of the self-adjoint operator $-\widetilde{\Delta}$; self-adjointness of $-\widetilde{\Delta}$ follows from that of $-\Delta$ by (\ref{mlap}). It remains to show that $\spn\mathcal{B}$ is dense in $L^2\bigl(\mathbb{D},rdrd\theta/(1-r^2)\bigr)$. Toward this end, suppose that $h\in L^2\bigl(\mathbb{D},rdrd\theta/(1-r^2)\bigr)$ is orthogonal to every member of $\mathcal{B}$. By (\ref{L2-relation}) there exists $g\in L^2\bigl(\mathbb{D},rdrd\theta\bigr)$ such that \begin{equation}\label{gh} h(r,\theta)=\sqrt{1-r^2}\,g(r,\theta). \end{equation} Let $\alpha_{p,q}=(-1)^{q+\max\{p,q\}}\max\{p,q\}/q$ denote the coefficient occurring on the right-hand side of (\ref{representation}). Then for each fixed $n=q-p\in\mathbb{Z}$, for every $\nu=\min\{p,q\}-1\geq 0$, \begin{equation}\nonumbe \begin{split} 0&=\int_\mathbb{D} h(r,\theta)\,\overline{\varphi^{(p,q)}(r,\theta)}\,\frac{rdrd\theta}{1-r^2}\\ &=\alpha_{p,q}\int_0^1\left(\int_0^{2\pi}h(r,\theta)e^{-in\theta}\,d\theta\right)(1-r^2)r^{|n|}P^{(1,|n|)}_{\nu}(2r^2-1)\,\frac{rdr}{1-r^2}\qquad(\mbox{ by (\ref{Jacobi})})\\ &=\alpha_{p,q}\int_0^1\left(\int_0^{2\pi}g(r,\theta)e^{-in\theta}\,d\theta\right)\sqrt{1-r^2}\,r^{|n|}P^{(1,|n|)}_{\nu}(2r^2-1)\,rdr\qquad(\mbox{by (\ref{gh})})\\ &=\frac{\alpha_{p,q}}{4}\int_{-1}^1\left(\int_0^{2\pi}g\left(\textstyle\sqrt{\frac{1+u}{2}},\theta\right)e^{-in\theta}\,d\theta\right)\sqrt{\frac{1-u}{2}}\,\sqrt{\frac{1+u}{2}}^{|n|}P^{(1,|n|)}_{\nu}(u)\,du\qquad(u=2r^2-1)\\ &=\frac{\alpha_{p,q}}{4}\int_{-1}^1\left(\int_0^{2\pi}g\left(\textstyle\sqrt{\frac{1+u}{2}},\theta\right)e^{-in\theta}\,d\theta\right)Q^{(1,|n|)}_{\nu}(u)\,du\qquad(\mbox{ as in (\ref{Q})}).\\ \end{split} \end{equation} Since the quasipolynomials $Q^{(1,|n|)}_\nu$ are an orthogonal basis for $L^2([-1,1],du)$, it follows that for each $n\in\mathbb{Z}$, \begin{equation}\label{g0} \int_0^{2\pi}g\left(\textstyle\sqrt{\frac{1+u}{2}},\theta\right)e^{-in\theta}\,d\theta=0 \end{equation} for every $u\in[-1,1]$ outside a set $E_n$ of measure zero. Since $\{e^{in\theta}\}_{n\in\mathbb{Z}}$ is an orthogonal basis of $L^2([0,2\pi],d\theta)$, it follows in turn that for $u\not\in\cup E_n$ \[ g\left(\textstyle\sqrt{\frac{1+u}{2}},\theta\right)=0, \] for almost every $\theta\in[0,2\pi]$. Thus $g(r,\theta)=0$ for almost every $(r,\theta)\in\mathbb{D}$ and $g=0$ as a function in $L^2(\mathbb{D},rdrd\theta)$, whence $h=0$ also. This proves that the orthogonal complement of $\mathcal{B}$ in $L^2\bigl(\mathbb{D},rdrd\theta/(1-r^2)\bigr)$ is empty, and hence that $\mathcal{B}$ is an orthogonal basis. \section{Conclusions} Since the vector space $L^2\bigl(\mathbb{D},rdrd\theta/(1-r^2)\bigr)=\sqrt{1-r^2}L^2\bigl(\mathbb{D},rdrd\theta\bigr)$ is dense in \[ L^2\bigl(\mathbb{D},rdrd\theta\bigr)=L^2\bigl(\mathbb{D},dxdy\bigr), \] and convergence in the former space implies convergence in the latter, scattering polynomials comprise a (non-orthogonal) basis for$L^2\bigl(\mathbb{D},dxdy\bigr)$ consistent with Dirichlet boundary values. From the perspective of analysis of functions on the disk, this provides an alternative to Zernike polynomials---and their generalizations the disk polynomials---which are non-zero on the boundary circle and so inconsistent with Dirichlet conditions. More generally, scattering polynomials illustrate that orthogonal polynomials can comprise an orthogonal basis for a function spaces $L^2(X,d\mu)$ in which not all polynomials are integrable. A natural question for further investigation is the existence and extent of other such examples.

https://arxiv.org/abs/0910.3607

Multigraded Factorial Rings and Fano varieties with torus action

In a first result, we describe all finitely generated factorial algebras over an algebraically closed field of characteristic zero that come with an effective multigrading of complexity one by means of generators and relations. This enables us to construct systematically varieties with free divisor class group and a complexity one torus action via their Cox rings. For the Fano varieties of this type that have a free divisor class group of rank one, we provide explicit bounds for the number of possible deformation types depending on the dimension and the index of the Picard group in the divisor class group. As a consequence, one can produce classification lists for fixed dimension and Picard index. We carry this out expemplarily in the following cases. There are 15 non-toric surfaces with Picard index at most six. Moreover, there are 116 non-toric threefolds with Picard index at most two; nine of them are locally factorial, i.e. of Picard index one, and among these one is smooth, six have canonical singularities and two have non-canonical singularities. Finally, there are 67 non-toric locally factorial fourfolds and two one-dimensional families of non-toric locally factorial fourfolds. In all cases, we list the Cox rings explicitly.

\section*{Introduction} Let ${\mathbb K}$ be an algebraically closed field of characteristic zero. A first aim of this paper is to determine all finitely generated factorial ${\mathbb K}$-algebras $R$ with an effective complexity one multigrading $R = \oplus_{u \in M} R_u$ satisfying $R_0 = {\mathbb K}$; here effective complexity one multigrading means that with $d := \dim \, R$ we have $M \cong {\mathbb Z}^{d-1}$ and the $u \in M$ with $R_u \ne 0$ generate $M$ as a ${\mathbb Z}$-module. Our result extends work by Mori~\cite{Mo} and Ishida~\cite{Is}, who settled the cases $d=2$ and $d=3$. An obvious class of multigraded factorial algebras as above is given by polynomial rings. A much larger class is obtained as follows. Take a sequence $A = (a_0, \ldots, a_r)$ of vectors $a_i \in {\mathbb K}^2$ such that $(a_i,a_k)$ is linearly independent whenever $k \ne i$, a sequence $\mathfrak{n} = (n_0, \ldots, n_r)$ of positive integers and a family $L = (l_{ij})$ of positive integers, where $0 \le i \le r$ and $1 \le j \le n_i$. For every $0 \le i \le r$, we define a monomial $$ f_i \ := \ T_{i1}^{l_{i1}} \cdots T_{in_i}^{l_{in_i}} \ \in \ {\mathbb K}[T_{ij}; \; 0 \le i \le r, \; 1 \le j \le n_i], $$ for any two indices $0 \le i,j \le r$, we set $\alpha_{ij} := \det(a_i,a_j)$, and for any three indices $0 \le i < j < k \le r$, we define a trinomial $$ g_{i,j,k} \ := \ \alpha_{jk}f_i \ + \ \alpha_{ki}f_j \ + \ \alpha_{ij}f_k \ \in \ {\mathbb K}[T_{ij}; \; 0 \le i \le r, \; 1 \le j \le n_i]. $$ Note that the coefficients of $g_{i,j,k}$ are all nonzero. The triple $(A,\mathfrak{n},L)$ then defines a ${\mathbb K}$-algebra $$ R(A,\mathfrak{n},L) \ := \ {\mathbb K}[T_{ij}; \; 0 \le i \le r, \; 1 \le j \le n_i] \ / \ \bangle{g_{i,i+1,i+2}; \; 0 \le i \le r-2}. $$ It turns out that $R(A,\mathfrak{n},L)$ is a normal complete intersection, see Proposition~\ref{prop:RAnLnormal}. In particular, it is of dimension \begin{eqnarray*} \dim \, R(A,\mathfrak{n},L) & = & n_0 + \ldots + n_r \ - \ r \ + \ 1. \end{eqnarray*} If the triple $(A,\mathfrak{n},L)$ is {\em admissible\/}, i.e., the numbers $\gcd(l_{i1}, \ldots, l_{in_i})$, where $0 \le i \le r$, are pairwise coprime, then $R(A,\mathfrak{n},L)$ admits a canonical effective complexity one grading by a lattice $K$, see Construction~\ref{constr:Kgrading}. Our first result is the following. \goodbreak \begin{introthm1} Up to isomorphy, the finitely generated factorial ${\mathbb K}$-algebras with an effective complexity one grading $R = \oplus_M R_u$ and $R_0 = {\mathbb K}$ are \begin{enumerate} \item the polynomial algebras ${\mathbb K}[T_1, \ldots, T_d]$ with a grading $\deg(T_i) = u_i \in {\mathbb Z}^{d-1}$ such that $u_1, \ldots, u_d$ generate ${\mathbb Z}^{d-1}$ as a lattice and the convex cone on ${\mathbb Q}^{d-1}$ generated by $u_1, \ldots, u_d$ is pointed, \item the $(K \times {\mathbb Z}^m)$-graded algebras $R(A,\mathfrak{n},L)[S_1,\ldots,S_m]$, where $R(A,\mathfrak{n},L)$ is the $K$-graded algebra defined by an admissible triple $(A,\mathfrak{n},L)$ and $\deg \, S_j \in {\mathbb Z}^m$ is the $j$-th canonical base vector. \end{enumerate} \end{introthm1} The further paper is devoted to normal (possibly singular) $d$-dimensional Fano varieties $X$ with an effective action of an algebraic torus $T$. In the case $\dim \, T = d$, we have the meanwhile extensively studied class of toric Fano varieties, see~\cite{Bat1}, \cite{WaWa} and~\cite{Bat2} for the initiating work. Our aim is to show that the above Theorem provides an approach to classification results for the case $\dim \, T = d-1$, that means Fano varieties with a complexity one torus action. Here, we treat the case of divisor class group $\operatorname{Cl}(X) \cong {\mathbb Z}$; note that in the toric setting this gives precisely the weighted projective spaces. The idea is to consider the Cox ring \begin{eqnarray*} \mathcal{R}(X) & = & \bigoplus_{D \in \operatorname{Cl}(X)} \Gamma(X, \mathcal{O}_X(D)). \end{eqnarray*} The ring $\mathcal{R}(X)$ is factorial, finitely generated as a ${\mathbb K}$-algebra and the $T$-action on $X$ gives rise to an effective complexity one multigrading of $\mathcal{R}(X)$ refining the $\operatorname{Cl}(X)$-grading, see~\cite{BeHa1} and~\cite{HaSu}. Consequently, $\mathcal{R}(X)$ is one of the rings listed in the first Theorem. Moreover, $X$ can be easily reconstructed from $\mathcal{R}(X)$; it is the homogeneous spectrum with respect to the $\operatorname{Cl}(X)$-grading of $\mathcal{R}(X)$. Thus, in order to construct Fano varieties, we firstly have to figure out the Cox rings among the rings occuring in the first Theorem and then find those, which belong to a Fano variety; this is done in Propositions~\ref{prop:coxchar} and~\ref{Prop:FanoPicard}. In order to produce classification results via this approach, we need explicit bounds on the number of deformation types of Fano varieties with prescribed discrete invariants. Besides the dimension, in our setting, a suitable invariant is the {\em Picard index\/} $[\operatorname{Cl}(X):\operatorname{Pic}(X)]$. Denoting by $\xi(\mu)$ the number of primes less or equal to $\mu$, we obtain the following bound, see Corollary~\ref{cor:finitefanos}: for any pair $(d,\mu) \in {\mathbb Z}^2_{>0}$, the number $\delta(d,\mu)$ of different deformation types of $d$-dimensional Fano varieties with a complexity one torus action such that $\operatorname{Cl}(X) \cong {\mathbb Z}$ and $\mu = [\operatorname{Cl}(X):\operatorname{Pic}(X)]$ hold is bounded by \begin{eqnarray*} \delta(d,\mu) & \le & (6d\mu)^{2\xi(3d\mu)+d-2}\mu^{\xi(\mu)^2 + 2\xi((d+2)\mu)+2d+2}. \end{eqnarray*} In particular, we conclude that for fixed $\mu \in {\mathbb Z}_{>0}$, the number $\delta(d)$ of different deformation types of $d$-dimensional Fano varieties with a complexity one torus action $\operatorname{Cl}(X) \cong {\mathbb Z}$ and Picard index $\mu$ is asymptotically bounded by $d^{Ad}$ with a constant~$A$ depending only on~$\mu$, see~Corollary~\ref{cor:asymptoticsmu}. In fact, in Theorem~\ref{Th:FiniteIndex} we even obtain explicit bounds for the discrete input data of the rings $R(A,\mathfrak{n},L)[S_1,\ldots,S_m]$. This allows us to construct all Fano varieties $X$ with prescribed dimension and Picard index that come with an effective complexity one torus action and have divisor class group ${\mathbb Z}$. Note that, by the approach, we get the Cox rings of the resulting Fano varieties $X$ for free. In Section~\ref{sec:tables}, we give some explicit classifications. We list all non-toric surfaces $X$ with Picard index at most six and the non-toric threefolds~$X$ with Picard index up at most two. They all have a Cox ring defined by a single relation; in fact, for surfaces the first Cox ring with more than one relation occurs for Picard index~29, and for the threefolds this happens with Picard index~3, see Proposition~\ref{prop:fano22rel} as well as Examples~\ref{ex:fanosurf2rel} and~\ref{ex:fano32rel}. Moreover, we determine all locally factorial fourfolds~$X$, i.e. those of Picard index one: 67 of them occur sporadic and there are two one-dimensional families. Here comes the result on the locally factorial threefolds; in the table, we denote by $w_i$ the $\operatorname{Cl}(X)$-degree of the variable $T_i$. \goodbreak \begin{introthm2} The following table lists the Cox rings $\mathcal{R}(X)$ of the three-dimensional locally factorial non-toric Fano varieties $X$ with an effective two torus action and $\operatorname{Cl}(X) = {\mathbb Z}$. \begin{center} \begin{longtable}[htbp]{llll} \toprule No. & $\mathcal{R}(X)$ & $(w_1,\ldots, w_5)$ & $(-K_X)^3$ \\ \midrule 1 \hspace{.5cm} & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^5 + T_3^3 + T_4^2} $ \hspace{.5cm} & $(1,1,2,3,1)$ \hspace{.5cm} & $8$ \\ \midrule 2 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2T_3^4 + T_4^3 + T_5^2} $ & $(1,1,1,2,3)$ & $8$ \\ \midrule 3 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^2T_3^3 + T_4^3 + T_5^2} $ & $(1,1,1,2,3)$ & $8$ \\ \midrule 4 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2 + T_3T_4 + T_5^2} $ & $(1,1,1,1,1)$ & $54$ \\ \midrule 5 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^2 + T_3T_4^2 + T_5^3} $ & $(1,1,1,1,1)$ & $24$ \\ \midrule 6 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^3 + T_3T_4^3 + T_5^4} $ & $(1,1,1,1,1)$ & $4$ \\ \midrule 7 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^3 + T_3T_4^3 + T_5^2} $ & $(1,1,1,1,2)$ & $16$ \\ \midrule 8 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^5 + T_3T_4^5 + T_5^2} $ & $(1,1,1,1,3)$ & $2$ \\ \midrule 9 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^5 + T_3^3T_4^3 + T_5^2} $ & $(1,1,1,1,3)$ & $2$ \\ \bottomrule \end{longtable} \end{center} \end{introthm2} Note that each of these varieties $X$ is a hypersurface in the respective weighted projective space ${\mathbb P}(w_1, \ldots, w_5)$. Except number~4, none of them is quasismooth in the sense that ${\rm Spec} \, \mathcal{R}(X)$ is singular at most in the origin; quasismooth hypersurfaces of weighted projective spaces were studied in~\cite{JoKo} and~\cite{CCC}. In Section~\ref{sec:geom3folds}, we take a closer look at the singularities of the threefolds listed above. It turns out that number~1,3,5,7 and 9 are singular with only canonical singularities and all of them admit a crepant resolution. Number~6 and 8 are singular with non-canonical singularities but admit a smooth relative minimal model. Number two is singular with only canonical singularities, one of them of type $\mathbf{cA_1}$, and it admits only a singular relative minimal model. Moreover, in all cases, we determine the Cox rings of the resolutions. The authors would like to thank Ivan Arzhantsev for helpful comments and discussions and also the referee for valuable remarks and many references. \section{UFDs with complexity one multigrading} \label{sec:factrings} As mentioned before, we work over an algebraically closed field ${\mathbb K}$ of characteristic zero. In Theorem~\ref{thm:factrings}, we describe all factorial finitely generated ${\mathbb K}$-algebras~$R$ with an effective complexity one grading and $R_0={\mathbb K}$. Moreover, we characterize the possible Cox rings among these algebras, see Proposition~\ref{prop:coxchar}. First we recall the construction sketched in the introduction. \begin{construction} \label{constr:triple2ring} Consider a sequence $A = (a_0, \ldots, a_r)$ of vectors $a_i = (b_i,c_i)$ in ${\mathbb K}^2$ such that any pair $(a_i,a_k)$ with $k \ne i$ is linearly independent, a sequence $\mathfrak{n} = (n_0, \ldots, n_r)$ of positive integers and a family $L = (l_{ij})$ of positive integers, where $0 \le i \le r$ and $1 \le j \le n_i$. For every $0 \le i \le r$, define a monomial $$ f_i \ := \ T_{i1}^{l_{i1}} \cdots T_{in_i}^{l_{in_i}} \ \in \ {\mathbb K}[T_{ij}; \; 0 \le i \le r, \; 1 \le j \le n_i], $$ for any two indices $0 \le i,j \le r$, set $\alpha_{ij} := \det(a_i,a_j) = b_ic_j-b_jc_i$ and for any three indices $0 \le i < j < k \le r$ define a trinomial $$ g_{i,j,k} \ := \ \alpha_{jk}f_i \ + \ \alpha_{ki}f_j \ + \ \alpha_{ij}f_k \ \in \ {\mathbb K}[T_{ij}; \; 0 \le i \le r, \; 1 \le j \le n_i]. $$ Note that the coefficients of this trinomial are all nonzero. The triple $(A,\mathfrak{n},L)$ then defines a ring $$ R(A,\mathfrak{n},L) \ := \ {\mathbb K}[T_{ij}; \; 0 \le i \le r, \; 1 \le j \le n_i] \ / \ \bangle{g_{i,i+1,i+2}; \; 0 \le i \le r-2}. $$ \end{construction} \begin{proposition} \label{prop:RAnLnormal} For every triple $(A,\mathfrak{n},L)$ as in~\ref{constr:triple2ring}, the ring $R(A,\mathfrak{n},L)$ is a normal complete intersection of dimension $$ \dim \, R(A,\mathfrak{n},L) \ = \ n-r+1, \qquad\qquad n \ := \ n_0 + \ldots + n_r. $$ \end{proposition} \begin{lemma} \label{lem:alltrins} In the setting of~\ref{constr:triple2ring}, one has for any $0 \le i < j < k < l \le r$ the identities $$ g_{i,k,l} \ = \ \alpha_{kl} \cdot g_{i,j,k} + \alpha_{ik} \cdot g_{j,k,l}, \qquad\qquad g_{i,j,l} \ = \ \alpha_{jl} \cdot g_{i,j,k} + \alpha_{ij} \cdot g_{j,k,l}. $$ In particular, every trinomial $g_{i,j,k}$, where $0 \le i < j < k \le r$ is contained in the ideal $\bangle{g_{i,i+1,i+2}; \; 0 \le i \le r-2}$. \end{lemma} \begin{proof} The identities are easily obtained by direct computation; note that for this one may assume $a_j = (1,0)$ and $a_k = (0,1)$. The supplement then follows by repeated application of the identities. \end{proof} \begin{lemma} \label{lem:twotrinszero} In the notation of~\ref{constr:triple2ring} and~\ref{prop:RAnLnormal}, set $X := V({\mathbb K}^n,g_0,\ldots, g_{r-2})$, and let $z \in X$. If we have $f_i(z) = f_j(z) = 0$ for two $0 \le i < j \le r$, then $f_k(z) = 0$ holds for all $0 \le k \le r$. \end{lemma} \begin{proof} If $i<k<j$ holds, then, according to Lemma~\ref{lem:alltrins}, we have $g_{i,k,j}(z)=0$, which implies $f_k(z)=0$. The cases $k<i$ and $j<k$ are obtained similarly. \end{proof} \begin{proof}[Proof of Proposition~\ref{prop:RAnLnormal}] Set $X := V({\mathbb K}^n; g_0, \ldots, g_{r-2})$, where $g_i := g_{i,i+1,i+2}$. Then we have to show that $X$ is a connected complete intersection with at most normal singularities. In order to see that $X$ is connected, set $\ell := \prod n_i \prod l_{ij}$ and $\zeta_{ij} : = \ell n_i^{-1}l_{ij}^{-1}$. Then $X \subseteq {\mathbb K}^n$ is invariant under the ${\mathbb K}^*$-action given by \begin{eqnarray*} t \cdot z & := & (t^{\zeta_{ij}} z_{ij}) \end{eqnarray*} and the point $0 \in {\mathbb K}^n$ lies in the closure of any orbit ${\mathbb K}^* \! \cdot \! x \subseteq X$, which implies connectedness. To proceed, consider the Jacobian $J_g$ of $g := (g_0, \ldots, g_{r-2})$. According to Serre's criterion, we have to show that the set of points of $z \in X$ with $J_g(z)$ not of full rank is of codimension at least two in $X$. Note that the Jacobian $J_g$ is of the shape \begin{eqnarray*} J_g & = & \left( \begin{array}{rrrrrcrrrrr} \delta_{0 \, 0} & \delta_{0 \, 1} & \delta_{0 \, 2} & 0 & & &&&&& 0 \\ 0 & \delta_{1 \, 1} & \delta_{1 \, 2} & \delta_{1 \, 3} & 0 & &&&&& \\ &&&&& \vdots &&&&& \\ \\ &&&&& & 0 & \delta_{r-3 \, r-3} & \delta_{r-3 \, r-2} & \delta_{r-3 \, r-1} & 0 \\ 0 &&&&& & & 0 & \delta_{r-2 \, r-2} & \delta_{r-2 \, r-1} & \delta_{r-2 \, r} \end{array} \right) \end{eqnarray*} where $\delta_{ti}$ is a nonzero multiple of the gradient $\delta_i := {\rm grad} \, f_i$. Consider $z \in X$ with $J_g(z)$ not of full rank. Then $\delta_i(z) = 0 = \delta_k(z)$ holds with some $0 \le i < k \le r$. This implies $z_{ij} = 0 = z_{kl}$ for some $1 \le j \le n_i$ and $1 \le l \le n_k$. Thus, we have $f_i(z) = 0 = f_k(z)$. Lemma~\ref{lem:twotrinszero} gives $f_s(z) = 0$, for all $0 \le s \le r$. Thus, some coordinate $z_{st}$ must vanish for every $0 \le s \le r$. This shows that $z$ belongs to a closed subset of $X$ having codimension at least two in $X$. \end{proof} \begin{lemma} \label{lem:tijprime} Notation as in~\ref{constr:triple2ring}. Then the variable $T_{ij}$ defines a prime ideal in $R(A,\mathfrak{n},L)$ if and only if the numbers $\gcd(l_{k1}, \ldots, l_{kn_k})$, where $k \ne i$, are pairwise coprime. \end{lemma} \begin{proof} We treat exemplarily $T_{01}$. Using Lemma~\ref{lem:alltrins}, we see that the ideal of relations of $R(A,\mathfrak{n},L)$ can be presented as follows \begin{eqnarray*} \bangle{g_{s,s+1,s+2}; \; 0 \le s \le r-2} & = & \bangle{g_{0,s,s+1}; \; 1 \le s \le r-1}. \end{eqnarray*} Thus, the ideal $\bangle{T_{01}} \subseteq R(A,\mathfrak{n},L)$ is prime if and only if the following binomial ideal is prime $$ \mathfrak{a} \ := \ \bangle{\alpha_{s+1 \, 0}f_s + \alpha_{0 s}f_{s+1}; \; 1 \le s \le r-1} \ \subseteq \ {\mathbb K}[T_{ij}; \; (i,j) \ne (0,1)]. $$ Set $l_i := (l_{i1}, \ldots, l_{in_i})$. Then the ideal $\mathfrak{a}$ is prime if and only if the following family can be complemented to a lattice basis $$ (l_1,-l_2,0,\ldots,0), \ \ldots, \ (0,\ldots,0,l_{r-1},-l_r). $$ This in turn is equivalent to the statement that the numbers $\gcd(l_{k1}, \ldots, l_{kn_k})$, where $1 \le k \le r$, are pairwise coprime. \end{proof} \begin{definition} We say that a triple $(A,\mathfrak{n},L)$ as in~\ref{constr:triple2ring} is {\em admissible\/} if the numbers $\gcd(l_{i1}, \ldots, l_{in_i})$, where $0 \le i \le r$, are pairwise coprime. \end{definition} \begin{construction} \label{constr:Kgrading} Let $(A,\mathfrak{n},L)$ be an admissible triple and consider the following free abelian groups $$ E \quad := \quad \bigoplus_{i=0}^r \bigoplus_{j=1}^{n_i} {\mathbb Z} \! \cdot \! e_{ij}, \qquad \qquad K \quad := \quad \bigoplus_{j=1}^{n_0} {\mathbb Z} \! \cdot \! u_{0j} \ \oplus \ \bigoplus_{i=1}^r \bigoplus_{j=1}^{n_i-1} {\mathbb Z} \! \cdot \! u_{ij} $$ and define vectors $u_{in_i} := u_{01} + \ldots + u_{0r} - u_{i1} - \ldots - u_{in_i-1} \in K$. Then there is an epimorphism $\lambda \colon E \to K$ fitting into a commutative diagram with exact rows $$ \xymatrix{ 0 \ar[rr] && E \ar[rr]_{\alpha}^{e_{ij} \mapsto l_{ij} e_{ij}} \ar[d]^{\eta}_{e_{ij} \mapsto u_{ij}} && E \ar[rr]^{e_{ij} \mapsto \b{e}_{ij} \qquad} \ar[d]^{\lambda} && {\bigoplus_{i,j} {\mathbb Z} / l_{ij} {\mathbb Z}} \ar[rr] \ar@{<->}[d]^{\cong} && 0 \\ 0 \ar[rr] && K \ar[rr]_{\beta} && K \ar[rr] && {\bigoplus_{i,j} {\mathbb Z} / l_{ij} {\mathbb Z}} \ar[rr] && 0 } $$ Define a $K$-grading of ${\mathbb K}[T_{ij}; \; 0 \le i \le r, \; 1 \le j \le n_i]$ by setting $\deg \, T_{ij} := \lambda(e_{ij})$. Then every $f_i = T_{i1}^{l_{i1}} \cdots T_{in_i}^{l_{in_i}}$ is $K$-homogeneous of degree $$ \deg \, f_i \ = \ l_{i1} \lambda(e_{i1}) + \ldots + l_{in_i}\lambda(e_{in_i}) \ = \ l_{01} \lambda(e_{01}) + \ldots + l_{0n_0}\lambda(e_{0n_0}) \ \in \ K. $$ Thus, the polynomials $g_{i,j,k}$ of~\ref{constr:triple2ring} are all $K$-homogeneous of the same degree and we obtain an effective $K$-grading of complexity one of $R(A,\mathfrak{n},L)$. \end{construction} \begin{proof} Only for the existence of the commutative diagram there is something to show. Write for short $l_i := (l_{i1}, \ldots, l_{in_i})$. By the admissibility condition, the vectors $v_i := (0, \ldots, 0,l_i,-l_{i+1},0,\ldots,0)$, where $0 \le i \le r-1$, can be completed to a lattice basis for $E$. Consequently, we find an epimorphism $\lambda \colon E \to K$ having precisely ${\rm lin}(v_0, \ldots, v_{r-1})$ as its kernel. By construction, $\ker(\lambda)$ equals $\alpha(\ker(\eta))$. Using this, we obtain the induced morphism $\beta \colon K \to K$ and the desired properties. \end{proof} \begin{lemma} \label{lem:pwnonassoc} Notation as in~\ref{constr:Kgrading}. Then $R(A,\mathfrak{n},L)_0 = {\mathbb K}$ and $R(A,\mathfrak{n},L)^* = {\mathbb K}^*$ hold. Moreover, the $T_{ij}$ define pairwise nonassociated prime elements in $R(A,\mathfrak{n},L)$. \end{lemma} \begin{proof} The fact that all elements of degree zero are constant is due to the fact that all degrees $\deg \, T_{ij} = u_{ij} \in K$ are non-zero and generate a pointed convex cone in $K_{{\mathbb Q}}$. As a consequence, we obtain that all units in $R(A,\mathfrak{n},L)$ are constant. The $T_{ij}$ are prime by the admissibility condition and Lemma~\ref{lem:tijprime}, and they are pairwise nonassociated because they have pairwise different degrees and all units are constant. \end{proof} \goodbreak \begin{theorem} \label{thm:factrings} Up to isomorphy, the finitely generated factorial ${\mathbb K}$-algebras with an effective complexity one grading $R = \oplus_M R_u$ and $R_0 = {\mathbb K}$ are \begin{enumerate} \item the polynomial algebras ${\mathbb K}[T_1, \ldots, T_d]$ with a grading $\deg(T_i) = u_i \in {\mathbb Z}^{d-1}$ such that $u_1, \ldots, u_d$ generate ${\mathbb Z}^{d-1}$ as a lattice and the convex cone on ${\mathbb Q}^{d-1}$ generated by $u_1, \ldots, u_d$ is pointed, \item the $(K \times {\mathbb Z}^m)$-graded algebras $R(A,\mathfrak{n},L)[S_1,\ldots,S_m]$, where $R(A,\mathfrak{n},L)$ is the $K$-graded algebra defined by an admissible triple $(A,\mathfrak{n},L)$ as in~\ref{constr:triple2ring} and~\ref{constr:Kgrading} and $\deg\, S_j \in {\mathbb Z}^m$ is the $j$-th canonical base vector. \end{enumerate} \end{theorem} \begin{proof} We first show that for any admissible triple $(A,\mathfrak{n},L)$ the ring $R(A,\mathfrak{n},L)$ is a unique factorization domain. If $l_{ij} = 1$ holds for any two $i,j$, then, by~\cite[Prop.~2.4]{HaSu}, the ring $R(A,\mathfrak{n},L)$ is the Cox ring of a space ${\mathbb P}_1(A,\mathfrak{n})$ and hence is a unique factorization domain. Now, let $(A,\mathfrak{n},L)$ be arbitrary admissible data and let $\lambda \colon E \to K$ be an epimorphism as in~\ref{constr:Kgrading}. Set $n := n_0 + \ldots + n_r$ and consider the diagonalizable groups $$ {\mathbb T}^n \ := \ {\rm Spec} \, {\mathbb K}[E], \qquad H \ := \ {\rm Spec} \, {\mathbb K}[K], \qquad H_0 \ := \ {\rm Spec} \, {\mathbb K}[\oplus_{i,j} {\mathbb Z} / l_{ij} {\mathbb Z}]. $$ Then ${\mathbb T}^n = ({\mathbb K}^*)^n$ is the standard $n$-torus and $H_0$ is the direct product of the cyclic subgroups $H_{ij} := {\rm Spec} \, {\mathbb K}[{\mathbb Z} / l_{ij} {\mathbb Z}]$. Moreover, the diagram in~\ref{constr:Kgrading} gives rise to a commutative diagram with exact rows $$ \xymatrix{ 0 && {{\mathbb T}^n} \ar[ll] && {{\mathbb T}^n} \ar[ll]_{(t_{ij}^{l_{ij}}) \mapsfrom (t_{ij})} && \ar[ll] H_0 && 0 \ar[ll] \\ 0 && {H} \ar[ll] \ar[u]^{\imath} && {H} \ar[ll] \ar[u]^{\jmath} && H_0 \ar[ll] \ar@{<->}[u]_{\cong} && 0 \ar[ll] } $$ where $t_{ij} = \chi^{e_{ij}}$ are the coordinates of ${\mathbb T}^n$ corresponding to the characters $e_{ij} \in E$ and the maps $\imath$, $\jmath$ are the closed embeddings corresponding to the epimorphisms $\eta$, $\lambda$ respectively. Setting $\deg \, T_{ij} := e_{ij}$ defines an action of ${\mathbb T}^n$ on ${\mathbb K}^n = {\rm Spec} \, {\mathbb K}[T_{ij}]$; in terms of the coordinates $z_{ij}$ corresponding to $T_{ij}$ this action is given by $t \! \cdot \! z = (t_{ij} z_{ij})$. The torus $H$ acts effectively on ${\mathbb K}^n$ via the embedding $\jmath \colon H \to {\mathbb T}^n$. The generic isotropy group of $H$ along $V({\mathbb K}^n,T_{ij})$ is the subgroup $H_{ij} \subseteq H$ corresponding to $K \to K/\lambda(E_{ij})$, where $E_{ij} \subseteq E$ denotes the sublattice generated by all $e_{kl}$ with $(k,l) \ne (i,j)$; recall that we have $K/\lambda(E_{ij}) \cong {\mathbb Z} / l_{ij}{\mathbb Z}$. Now, set $l_{ij}' := 1$ for any two $i,j$ and consider the spectra $X := {\rm Spec} \, R(A,\mathfrak{n},L)$ and $X' := {\rm Spec} \, R(A,\mathfrak{n},L')$. Then the canonical surjections ${\mathbb K}[T_{ij}] \to R(A,\mathfrak{n},L)$ and ${\mathbb K}[T_{ij}] \to R(A,\mathfrak{n},L')$ define embeddings $X \to {\mathbb K}^n$ and $X' \to {\mathbb K}^n$. These embeddings fit into the following commutative diagram $$ \xymatrix{ {{\mathbb K}^n} \ar@{<-}[rrr]_{\pi}^{(z_{ij}^{l_{ij}}) \mapsfrom (z_{ij})} & & & {{\mathbb K}^n} \\ X' \ar@{<-}[rrr] \ar[u] & & & X \ar[u] } $$ The action of $H$ leaves $X$ invariant and the induced $H$-action on $X$ is the one given by the $K$-grading of $R(A,\mathfrak{n},L)$. Moreover, $\pi \colon {\mathbb K}^n \to {\mathbb K}^n$ is the quotient map for the induced action of $H_0 \subseteq H$ on ${\mathbb K}^n$, we have $X = \pi^{-1}(X')$, and hence the restriction $\pi \colon X \to X'$ is a quotient map for the induced action of $H_0$ on $X$. Removing all subsets $V(X;T_{ij},T_{kl})$, where $(i,j) \ne (k,l)$ from $X$, we obtain an open subset $U \subseteq X$. By Lemma~\ref{lem:pwnonassoc}, the complement $X \setminus U$ is of codimension at least two and each $V(U,T_{ij})$ is irreducible. By construction, the only isotropy groups of the $H$-action on $U$ are the groups $H_{ij}$ of the points of $V(U,T_{ij})$. The image $U' := \pi(U)$ is open in $X'$, the complement $X' \setminus U'$ is as well of codimension at least two and $H/H_0$ acts freely on $U'$. According to~\cite[Cor.~5.3]{KKV}, we have two exact sequences fitting into the following diagram $$ \xymatrix{ & & 1 \ar[d] & \\ & & {\operatorname{Pic}}(U') \ar[d]^{\pi^*} & \\ 1 \ar[r] & {{\mathbb X}(H_0)} \ar[r]^{\alpha} & {\operatorname{Pic}_{H_0}}(U) \ar[r]^{\beta} \ar[d]^{\delta} & {\operatorname{Pic}}(U) \\ & & {\prod_{i,j}} {\mathbb X}(H_{ij}) & } $$ Since $X'$ is factorial, the Picard group $\operatorname{Pic}(U')$ is trivial and we obtain that $\delta$ is injective. Since $H_0$ is the direct product of the isotropy groups $H_{ij}$ of the Luna strata $V(U,T_{ij})$, we see that $\delta \circ \alpha$ is an isomorphism. It follows that $\delta$ is surjective and hence an isomorphism. This in turn shows that $\alpha$ is an isomorphism. Now, every bundle on $U$ is $H$-linearizable. Since $H_0$ acts as a subgroup of $H$, we obtain that every bundle is $H_0$-linearizable. It follows that $\beta$ is surjective and hence $\operatorname{Pic}(U)$ is trivial. We conclude $\operatorname{Cl}(X) = \operatorname{Pic}(U) = 0$, which means that $R(A,\mathfrak{n},L)$ admits unique factorization. The second thing we have to show is that any finitely generated factorial ${\mathbb K}$-algebra $R$ with an effective complexity one multigrading satisfying $R_0 = {\mathbb K}$ is as claimed. Consider the action of the torus $G$ on $X = {\rm Spec}\, R$ defined by the multigrading, and let $X_0 \subseteq X$ be the set of points having finite isotropy $G_x$. Then~\cite[Prop~3.3]{HaSu} provides a graded splitting \begin{eqnarray*} R & \cong & R'[S_1, \ldots, S_m], \end{eqnarray*} where the variables $S_j$ are identified with the homogeneous functions defining the prime divisors $E_j$ inside the boundary $X \setminus X_0$ and $R'$ is the ring of functions of $X_0$, which are invariant under the subtorus $G_0 \subseteq G$ generated by the generic isotropy groups $G_j$ of $E_j$. Since $R'_0 = R_0 = {\mathbb K}$ holds, the orbit space $X_0/G$ has only constant functions and thus is a space ${\mathbb P}_1(A,\mathfrak{n})$ as constructed in~\cite[Section~2]{HaSu}. This allows us to proceed exactly as in the proof of Theorem~\cite[Thm~1.3]{HaSu} and gives $R' = R(A,\mathfrak{n},L)$. The admissibility condition follows from Lemma~\ref{lem:tijprime} and the fact that each $T_{ij}$ defines a prime element in $R'$. \end{proof} \begin{remark} \label{rem:mori} Let $(A,\mathfrak{n},L)$ be an admissible triple with $\mathfrak{n} =(1,\ldots, 1)$. Then $K = {\mathbb Z}$ holds, the admissibility condition just means that the numbers $l_{ij}$ are pairwise coprime and we have $$ \dim \, R(A,\mathfrak{n},L) \ = \ n_0 + \ldots + n_r - r + 1 \ = \ 2. $$ Consequently, for two-dimensional rings, Theorem~\ref{thm:factrings} specializes to Mori's description of almost geometrically graded two-dimensional unique factorization domains provided in~\cite{Mo}. \end{remark} \begin{proposition} \label{prop:coxchar} Let $(A,\mathfrak{n},L)$ be an admissible triple, consider the associated $(K \times {\mathbb Z}^m)$-graded ring $R(A,\mathfrak{n},L)[S_1, \ldots, S_m]$ as in Theorem~\ref{thm:factrings} and let $\mu \colon K \times {\mathbb Z}^m \to K'$ be a surjection onto an abelian group $K'$. Then the following statements are equivalent. \begin{enumerate} \item The $K'$-graded ring $R(A,\mathfrak{n},L)[S_1, \ldots, S_m]$ is the Cox ring of a projective variety $X'$ with $\operatorname{Cl}(X') \cong K'$. \item For every pair $i,j$ with $0 \le i \le r$ and $1 \le j \le n_i$, the group $K'$ is generated by the elements $\mu(\lambda(e_{kl}))$ and $\mu(e_s)$, where $(i,j) \ne (k,l)$ and $1 \le s \le m$, for every $1 \le t \le m$, the group $K'$ is generated by the elements $\mu(\lambda(e_{ij}))$ and $\mu(e_s)$, where $0 \le i \le r$, $1 \le j \le n_i$ and $s \ne t$, and, finally the following cone is of full dimension in $K'_{{\mathbb Q}}$: $$ \bigcap_{(k,l)} {\rm cone}(\mu(\lambda(e_{ij})), \mu(e_s); \; (i,j) \ne (k,l)) \ \cap \ \bigcap_{t} {\rm cone}(\mu(\lambda(e_{ij})), \mu(e_s); \; s \ne t). $$ \end{enumerate} \end{proposition} \begin{proof} Suppose that~(i) holds, let $p \colon \rq{X}' \to X'$ denote the universal torsor and let $X'' \subseteq X'$ be the set of smooth points. According to~\cite[Prop.~2.2]{Ha2}, the group $H' = {\rm Spec} \, {\mathbb K}[K']$ acts freely on $p^{-1}(X'')$, which is a big open subset of the total coordinate space ${\rm Spec} \, R(A,\mathfrak{n},L)[S_1, \ldots, S_m]$. This implies the first condition of~(ii). Moreover, by~\cite[Prop.~4.1]{Ha2}, the displayed cone is the moving cone of $X'$ and hence of full dimension. Conversely, if~(ii) holds, then the $K'$-graded ring $R(A,\mathfrak{n},L)[S_1, \ldots, S_m]$ can be made into a bunched ring and hence is the Cox ring of a projective variety, use~\cite[Thm.~3.6]{Ha2}. \end{proof} \section{Bounds for Fano varieties} We consider $d$-dimensional Fano varieties $X$ that come with a complexity one torus action and have divisor class group $\operatorname{Cl}(X) \cong {\mathbb Z}$. Then the Cox ring $\mathcal{R}(X)$ of $X$ is factorial~\cite[Prop.~8.4]{BeHa1} and has an effective complexity one grading, which refines the $\operatorname{Cl}(X)$-grading, see~\cite[Prop.~2.6]{HaSu}. Thus, according to Theorem~\ref{thm:factrings}, it is of the form \begin{eqnarray*} \mathcal R(X) & \cong & {\mathbb K}[T_{ij}; \; 0 \le i \le r, \; 1 \le j \le n_i][S_1,\ldots, S_m] \ / \ \bangle{g_{i,i+1,i+2}; \; 0 \le i \le r-2}, \\ g_{i,j,k} & := & \alpha_{jk} T_{i1}^{l_{i1}} \cdots T_{in_i}^{l_{in_i}} \ + \ \alpha_{ki} T_{j1}^{l_{j1}} \cdots T_{jn_{j}}^{l_{jn_{j}}} \ + \ \alpha_{ij}T_{k1}^{l_{k1}} \cdots T_{kn_{k}}^{l_{kn_{k}}}. \end{eqnarray*} Here, we may (and will) assume $n_0 \ge \ldots \ge n_r \ge 1$. With $n := n_0 + \ldots + n_r$, we have $n + m = d + r$. For the degrees of the variables in $\operatorname{Cl}(X) \cong {\mathbb Z}$, we write $w_{ij} := \deg \, T_{ij}$ for $0 \le i \leq r$, $1 \le j \le n_i$ and $u_k = \deg \, S_k$ for $1 \le k \le m$. Moreover, for $\mu \in {\mathbb Z}_{>0}$, we denote by $\xi(\mu)$ the number of primes in $\{2, \ldots, \mu\}$. The following result provides bounds for the discrete data of the Cox ring. \begin{theorem} \label{Th:FiniteIndex} In the above situation, fix the dimension $d = \dim(X)$ and the Picard index $\mu = [\operatorname{Cl}(X):\operatorname{Pic}(X)]$. Then we have $$ u_k \ \le \ \mu \quad \text{for } 1 \le k \le m. $$ Moreover, for the degree $\gamma$ of the relations, the weights $w_{ij}$ and the exponents $l_{ij}$, where $0 \le i \le r$ and $1 \le j \le n_i$ one obtains the following. \begin{enumerate} \item Suppose that $r = 0,1$ holds. Then $n + m \le d+1$ holds and one has the bounds $$ w_{ij} \ \le \ \mu \quad\text{for } 0 \le i \le r \text{ and } 1 \le j \le n_i, $$ and the Picard index is given by $$ \mu \ = \ \mathrm{lcm}(w_{ij},u_k; \;0 \le i \le r, 1 \le j \le n_i, 1 \le k \le m ). $$ \item Suppose that $r \ge 2$ and $n_0=1$ hold. Then $r \le \xi(\mu)-1$ and $n=r+1$ and $m=d-1$ hold and one has $$ w_{i1} \ \le \ \mu^r \quad \text{for } 0 \le i \le r, \qquad l_{01} \cdots l_{r1} \ \mid \ \mu, \qquad l_{01} \cdots l_{r1} \ \mid \ \gamma \ \le \ \mu^{r+1}, $$ and the Picard index is given by $$ \mu \ = \ \mathrm{lcm}(\gcd(w_{j1}; \; j \neq i), u_k;\; 0 \le i \le r, 1\le k \le m). $$ \item Suppose that $r \ge 2$ and $n_0 > n_1=1$ hold. Then we may assume $l_{11} > \ldots > l_{r1} \ge 2$, we have $r \le \xi(3d\mu)-1$ and $n_0+m = d$ and the bounds $$ w_{01},\ldots,w_{0n_0} \ \le \ \mu, \qquad l_{01},\ldots,l_{0n_0} \ < \ 6d\mu, $$ $$ w_{11},l_{21} \ < \ 2d\mu, \qquad w_{21},l_{11} \ < \ 3d\mu, $$ $$ w_{i1} \ < \ 6d\mu, \quad l_{i1} \ < \ 2d\mu \quad \text{for } 2 \le i \le r, $$ $$ l_{11} \cdots l_{r1} \ \mid \ \gamma \ < \ 6d\mu, $$ and the Picard index is given by $$ \mu \ = \ \mathrm{lcm}(w_{0j}, \gcd(w_{11},\ldots,w_{r1}), u_k; \; 1 \le j \le n_0, 1 \le k \le m ). $$ \item Suppose that $n_1 > n_2 = 1$ holds. Then we may assume $l_{21} > \ldots > l_{r1} \ge 2$, we have $r \le \xi(2(d+1)\mu)-1$ and $n_0+n_1+m = d+1$ and the bounds $$ w_{ij} \ \le \ \mu \quad \text{for } i=0,1 \text{ and } 1 \le j \le n_i, \qquad w_{21} \ < \ (d+1)\mu, $$ $$ w_{ij}, l_{ij} \ < \ 2(d+1)\mu \quad \text{for } 0 \le i \le r \text{ and } 1 \le j \le n_i, $$ $$ l_{21} \cdots l_{r1} \ \mid \ \gamma \ < \ 2(d+1)\mu, $$ and the Picard index is given by $$ \mu \ = \ \mathrm{lcm}(w_{ij}, u_k; \; 0 \le i\le 1, 1 \le j \le n_i, 1 \le k \le m). $$ \item Suppose that $n_2 > 1$ holds and let $s$ be the maximal number with $n_{s}>1$. Then one may assume $l_{s+1,1} > \ldots > l_{r1} \ge 2$, we have $r \le \xi((d+2)\mu)-1$ and $n_0+ \ldots + n_s+m = d+s$ and the bounds $$ w_{ij} \ \le \ \mu, \quad \text{for } 0 \le i \le s, $$ $$ w_{ij}, l_{ij} \ < \ (d+2)\mu \quad \text{for } 0 \le i \le r \text{ and } 1 \le j \le n_i, $$ $$ l_{s+1,1} \cdots l_{r1} \ \mid \ \gamma \ < \ (d+2)\mu, $$ and the Picard index is given by $$ \mu \ = \ \mathrm{lcm}(w_{ij}, u_k; \; 0 \le i \le s, 1 \le j \le n_i, 1 \le k \le m). $$ \end{enumerate} \end{theorem} Putting all the bounds of the theorem together, we obtain the following (raw) bound for the number of deformation types. \begin{corollary} \label{cor:finitefanos} For any pair $(d,\mu) \in {\mathbb Z}^2_{>0}$, the number $\delta(d,\mu)$ of different deformation types of $d$-dimensional Fano varieties with a complexity one torus action such that $\operatorname{Cl}(X) \cong {\mathbb Z}$ and $[\operatorname{Cl}(X):\operatorname{Pic}(X)]=\mu$ hold is bounded by \begin{eqnarray*} \delta(d,\mu) & \le & (6d\mu)^{2\xi(3d\mu)+d-2}\mu^{\xi(\mu)^2 + 2\xi((d+2)\mu)+2d+2}. \end{eqnarray*} \end{corollary} \begin{proof} By Theorem~\ref{Th:FiniteIndex} the discrete data $r$, $\mathfrak{n}$, $L$ and $m$ occuring in $\mathcal{R}(X)$ are bounded as in the assertion. The continuous data in $\mathcal{R}(X)$ are the coefficients $\alpha_{ij}$; they stem from the family $A = (a_0, \ldots, a_r)$ of points $a_i \in {\mathbb K}^2$. Varying the $a_i$ provides flat families of Cox rings and hence, by passing to the homogeneous spectra, flat families of the resulting Fano varieties $X$. \end{proof} \begin{corollary} \label{cor:asymptoticsd} Fix $d \in {\mathbb Z}_{>0}$. Then the number $\delta(\mu)$ of different deformation types of $d$-dimensional Fano varieties with a complexity one torus action, $\operatorname{Cl}(X) \cong {\mathbb Z}$ and Picard index $\mu := [\operatorname{Cl}(X):\operatorname{Pic}(X)]$ is asymptotically bounded by $\mu^{A \mu^2 / \log^2 \mu}$ with a constant~$A$ depending only on~$d$. \end{corollary} \begin{corollary} \label{cor:asymptoticsmu} Fix $\mu \in {\mathbb Z}_{>0}$. Then the number $\delta(d)$ of different deformation types of $d$-dimensional Fano varieties with a complexity one torus action, $\operatorname{Cl}(X) \cong {\mathbb Z}$ and Picard index $\mu := [\operatorname{Cl}(X):\operatorname{Pic}(X)]$ is asymptotically bounded by $d^{Ad}$ with a constant~$A$ depending only on~$\mu$. \end{corollary} We first recall the necessary facts on Cox rings, for details, we refer to~\cite{Ha2}. Let $X$ be a complete $d$-dimensional variety with divisor class group $\operatorname{Cl}(X) \cong {\mathbb Z}$. Then the Cox ring $\mathcal{R}(X)$ is finitely generated and the total coordinate space $\b{X} := {\rm Spec} \, \mathcal{R}(X)$ is a factorial affine variety coming with an action of ${\mathbb K}^*$ defined by the $\operatorname{Cl}(X)$-grading of $\mathcal{R}(X)$. Choose a system $f_1, \ldots, f_\nu$ of homogeneous pairwise nonassociated prime generators for $\mathcal{R}(X)$. This provides an ${\mathbb K}^*$-equivariant embedding $$ \b{X} \ \to \ {\mathbb K}^{\nu}, \qquad \b{x} \ \mapsto \ (f_1(\b{x}), \ldots, f_{\nu}(\b{x})). $$ where ${\mathbb K}^*$ acts diagonally with the weights $w_i = \deg(f_i) \in \operatorname{Cl}(X) \cong {\mathbb Z}$ on ${\mathbb K}^{\nu}$. Moreover, $X$ is the geometric ${\mathbb K}^*$-quotient of $\rq{X} := \b{X} \setminus \{0\}$, and the quotient map $p \colon \rq{X} \to X$ is a universal torsor. By the local divisor class group $\operatorname{Cl}(X,x)$ of a point $x \in X$, we mean the group of Weil divisors $\operatorname{WDiv}(X)$ modulo those that are principal near~$x$. \begin{proposition} \label{Prop:FanoPicard} For any $\b{x} =(\b{x}_1,\ldots,\b{x}_{\nu}) \in \rq{X}$ the local divisor class group $\operatorname{Cl}(X,x)$ of $x := p(\b{x})$ is finite of order $\gcd(w_i; \; \b{x}_i \ne 0)$. The index of the Picard group $\operatorname{Pic}(X)$ in $\operatorname{Cl}(X)$ is given by \begin{eqnarray*} [\operatorname{Cl}(X):\operatorname{Pic}(X)] & = & \mathrm{lcm}_{x \in X}( |\operatorname{Cl}(X,x)| ). \end{eqnarray*} Suppose that the ideal of $\b{X} \subseteq {\mathbb K}^{\nu}$ is generated by $\operatorname{Cl}(X)$-homogeneous polynomials $g_1, \ldots, g_{\nu-d-1}$ of degree $\gamma_j := \deg(g_j)$. Then one obtains $$ -\mathcal{K}_X \ = \ \sum_{i=1}^{\nu} w_i - \sum_{j=1}^{\nu-d-1} \gamma_j, \qquad (-\mathcal{K}_X )^d \ = \ \left(\sum_{i=1}^{\nu} w_i - \sum_{j=1}^{\nu-d-1} \gamma_j\right)^d \frac{\gamma_1 \cdots \gamma_{\nu-d-1}}{w_1 \cdots w_\nu} $$ for the anticanonical class $-\mathcal{K}_X \in \operatorname{Cl}(X) \cong {\mathbb Z}$. In particular, $X$ is a Fano variety if and only if the following inequality holds \begin{eqnarray*} \sum_{j=1}^{\nu-d-1} \gamma_j & < & \sum_{i=1}^{\nu} w_i. \end{eqnarray*} \end{proposition} \begin{proof} Using~\cite[Prop.~2.2, Thm.~4.19]{Ha2}, we observe that $X$ arises from the bunched ring $(R,\mathfrak{F},\Phi)$, where $R = \mathcal{R}(X)$, $\mathfrak{F} = (f_1, \ldots, f_\nu)$ and $\Phi = \{{\mathbb Q}_{\ge 0}\}$. The descriptions of local class groups, the Picard index and the anticanonical class are then special cases of~\cite[Prop.~4.7, Cor.~4.9 and Cor.~4.16]{Ha2}. The anticanonical self-intersection number is easily computed in the ambient weighted projective space ${\mathbb P}(w_1, \ldots, w_\nu)$, use~\cite[Constr.~3.13, Cor.~4.13]{Ha2}. \end{proof} \begin{remark} If the ideal of $\b{X} \subseteq {\mathbb K}^{\nu}$ is generated by $\operatorname{Cl}(X)$-homogeneous polynomials $g_1, \ldots, g_{\nu-d-1}$, then~\cite[Constr.~3.13, Cor.~4.13]{Ha2} show that $X$ is a well formed complete intersection in the weighted projective space ${\mathbb P}(w_1, \ldots, w_\nu)$ in the sense of~\cite[Def.~6.9]{IaFl}. \end{remark} We turn back to the case that $X$ comes with a complexity one torus action as at the beginning of this section. We consider the case $n_0 = \ldots = n_r=1$, that means that each relation $g_{i,j,k}$ of the Cox ring $\mathcal{R}(X)$ depends only on three variables. Then we may write $T_i$ instead of $T_{i1}$ and $w_i$ instead of $w_{i1}$, etc.. In this setting, we obtain the following bounds for the numbers of possible varieties~$X$ (Fano or not). \begin{proposition} \label{prop:Finite3Var} For any pair $(d,\mu) \in {\mathbb Z}^2_{>0}$ there is, up to deformation, only a finite number of complete $d$-dimensional varieties with divisor class group ${\mathbb Z}$, Picard index $[\operatorname{Cl}(X):\operatorname{Pic}(X)] = \mu$ and Cox ring $$ {\mathbb K}[T_0,\ldots, T_r,S_1,\ldots, S_m] \ / \ \bangle{ \alpha_{i+1,i+2} T_i^{l_i} + \alpha_{i+2,i} T_{i+1}^{l_{i+1}} + \alpha_{i,i+1} T_{i+2}^{l_{i+2}}; \; 0 \le i \le r-2}. $$ In this situation we have $r \le \xi(\mu)-1$. Moreover, for the weights $w_i := \deg\, T_i$, where $0 \le i \le r$ and $u_k := \deg\, S_k$, where $1 \le k \le m$, the exponents $l_i$ and the degree $\gamma := l_0w_0$ of the relation one has $$ l_0 \cdots l_r \ \mid \ \gamma, \qquad l_0 \cdots l_r \ \mid \ \mu, \qquad w_i \ \le \ \mu^{\xi(\mu)-1}, \qquad u_k \ \le \ \mu. $$ \end{proposition} \begin{proof} Consider the total coordinate space $\b{X} \subseteq {\mathbb K}^{r+1+n}$ and the universal torsor $p \colon \rq{X} \to X$ as discussed before. For each $0 \le i \le r$ fix a point $\b{x}(i) = (\b{x}_0, \ldots, \b{x}_r, 0, \ldots, 0)$ in $\rq{X}$ such that $\b{x}_i = 0$ and $\b{x}_j \ne 0$ for $j \ne i$ hold. Then, denoting $x(i) := p(\b{x}(i))$, we obtain $$ \gcd(w_j; j \ne i) \ = \ \vert \operatorname{Cl}(X,x(i)) \vert \ \mid \ \mu. $$ Consider $i,j$ with $j \ne i$. Since all relations are homogeneous of the same degree, we have $l_iw_i = l_jw_j$. Moreover, by the admissibility condition, $l_i$ and $l_j$ are coprime. We conclude $l_i \vert w_j$ for all $j \ne i$ and hence $l_i \vert \gcd(w_j; \; j \ne i)$. This implies $$ l_0 \cdots l_r \ \mid \ l_0w_0 \ = \ \gamma, \qquad\qquad l_0 \cdots l_r \ \mid \ \mu. $$ We turn to the bounds for the $w_i$, and first verify $w_0 \le \mu^r$. Using the relation $l_iw_i = l_0w_0$, we obtain for every $l_i$ a presentation $$ l_i \ = \ l_0 \cdot \frac{w_0 \cdots w_{i-1}}{w_1 \cdots w_i} \ = \ \eta_i \cdot \frac{\gcd(w_0, \ldots ,w_{i-1})}{\gcd(w_0, \ldots, w_i)} $$ with suitable integers $1 \le \eta_i \le \mu$. In particular, the very last fraction is bounded by $\mu$. This gives the desired estimate: $$ w_0 = \frac{w_{0}}{\gcd(w_{0},w_{1})} \cdot \frac{\gcd(w_{0},w_1)}{\gcd(w_{0},w_{1},w_2)} \cdots \frac {\gcd(w_{0},\ldots,w_{r-2})} {\gcd(w_{0},\ldots,w_{r-1})} \cdot \gcd(w_{0},\ldots,w_{r-1}) \le \mu^r. $$ Similarly, we obtain $w_i \le \mu^r$ for $1 \le i \le r$. Then we only have to show that $r+1$ is bounded by $\xi(\mu)$, but this follows immediately from the fact that $l_0, \ldots, l_r$ are pairwise coprime. Finally, to estimate the $u_k$, consider the points $\b{x}(k) \in \rq{X}$ having the $(r+k)$-th coordinate one and all others zero. Set $x(k) : =p(\b{x}(k))$. Then $\operatorname{Cl}(X,x(k))$ is of order $u_k$, which implies $u_k \le \mu$. \end{proof} \goodbreak \begin{lemma} \label{Lem:1relation} Consider the ring ${\mathbb K}[T_{ij}; \; 0 \le i \le 2, \; 1 \le j \le n_i][S_1,\ldots,S_k] / \bangle{g} $ where $n_0 \ge n_1 \ge n_2 \ge 1$ holds. Suppose that $g$ is homogeneous with respect to a ${\mathbb Z}$-grading of ${\mathbb K}[T_{ij},S_k]$ given by $\deg \, T_{ij} = w_{ij} \in {\mathbb Z}_{>0}$ and $\deg \, S_k = u_k \in {\mathbb Z}_{>0}$, and assume \begin{eqnarray*} \deg \, g & < & \sum_{i=0}^2\sum_{j=1}^{n_i}w_{ij} \ + \ \sum_{i=1}^m u_i. \end{eqnarray*} Let $\mu \in {\mathbb Z}_{>1}$, assume $w_{ij} \le \mu$ whenever $n_i > 1$, $1 \le j \le n_i$ and $u_k \le \mu$ for $1 \le k \le m$ and set $d := n_0+n_1+n_2+m-2$. Depending on the shape of $g$, one obtains the following bounds. \begin{enumerate} \item Suppose that $g = \eta_0 T_{01}^{l_{01}} \cdots T_{0n_0}^{l_{0n_0}} + \eta_1 T_{11}^{l_{11}} + \eta_2 T_{21}^{l_{21}}$ with $n_0 > 1$ and coefficients $\eta_i \in {\mathbb K}^*$ holds, we have $l_{11} \ge l_{21} \ge 2$ and $l_{11}$, $l_{21}$ are coprime. Then, one has $$ \qquad\qquad w_{11}, l_{21} \ < \ 2d\mu, \qquad w_{21}, l_{11} \ < \ 3d\mu, \qquad \deg \, g \ < \ 6d\mu. $$ \item Suppose that $g = \eta_0 T_{01}^{l_{01}} \cdots T_{0n_0}^{l_{0n_0}} + \eta_1 T_{11}^{l_{11}} \cdots T_{1n_1}^{l_{1n_1}} + \eta_2 T_{21}^{l_{21}}$ with $n_1 > 1$ and coefficients $\eta_i \in {\mathbb K}^*$ holds and we have $l_{21} \ge 2$. Then one has $$ \qquad\qquad w_{21} \ < \ (d+1)\mu, \qquad\qquad \deg \, g \ < \ 2(d+1)\mu. $$ \end{enumerate} \end{lemma} \begin{proof} We prove~(i). Set for short $c := (n_0+m)\mu = d\mu$. Then, using homogeneity of $g$ and the assumed inequality, we obtain $$ l_{11}w_{11} \ = \ l_{21}w_{21} \ = \ \deg \, g \ < \ \sum_{i=0}^2\sum_{j=1}^{n_i}w_{ij} + \sum_{i=1}^m u_i \ \le \ c+w_{11}+w_{21}. $$ Since $l_{11}$ and $l_{21}$ are coprime, we have $l_{11} > l_{21} \ge 2$. Plugging this into the above inequalities, we arrive at $2 w_{11} < c + w_{21}$ and $w_{21} < c + w_{11}$. We conclude $w_{11} < 2c$ and $w_{21} < 3c$. Moreover, $l_{11}w_{11} = l_{21}w_{21}$ and $\gcd(l_{11},l_{21}) = 1$ imply $l_{11} \vert w_{21}$ and $l_{21} \vert w_{11}$. This shows $l_{11} < 3c$ and $l_{21} < 2c$. Finally, we obtain $$ \deg \, g \ < \ c + w_{11} + w_{21} \ < \ 6c. $$ We prove (ii). Here we set $c := (n_0+n_1+m)\mu = (d+1)\mu$. Then the assumed inequality gives $$ l_{21}w_{21} \ = \ \deg g \ < \ \sum_{i=0}^1\sum_{j=1}^{n_i}w_{ij}+ \sum_{i=1}^m u_i+ w_{21} \ \le \ c+w_{21}. $$ Since we assumed $l_{21} \geq 2$, we can conclude $w_{21} < c$. This in turn gives us $\deg \, g < 2c$ for the degree of the relation. \end{proof} \begin{proof} [Proof of Theorem~\ref{Th:FiniteIndex}] As before, we denote by $\b{X} \subseteq {\mathbb K}^{n+m}$ the total coordinate space and by $p \colon \rq{X} \to X$ the universal torsor. We first consider the case that $X$ is a toric variety. Then the Cox ring is a polynomial ring, $\mathcal R(X) = {\mathbb K}[S_1,\ldots,S_m]$. For each $1 \le k \le m$, consider the point $\overline x(k) \in \rq{X}$ having the $k$-th coordinate one and all others zero and set $x(k) := p(\b{x}(k))$. Then, by~Proposition~\ref{Prop:FanoPicard}, the local class group $\operatorname{Cl}(X,x(k))$ is of order $u_k$ where $u_k := \deg \, S_k$. This implies $u_k \le \mu$ for $1 \le k \leq m$ and settles Assertion~(i). Now we treat the non-toric case, which means $r \ge 2$. Note that we have $n \ge 3$. The case $n_0=1$ is done in Proposition~\ref{prop:Finite3Var}. So, we are left with $n_0>1$. For every $i$ with $n_i > 1$ and every $1 \le j \le n_i$, there is the point $\b{x}(i,j) \in \rq{X}$ with $ij$-coordinate $T_{ij}$ equal to one and all others equal to zero, and thus we have the point $x(i,j) := p(\b{x}(i,j)) \in X$. Moreover, for every $1 \le k \le m$, we have the point $\b{x}(k) \in \rq{X}$ having the $k$-coordinate $S_k$ equal to one and all others zero; we set $x(k):=p(\b{x}(k))$. Proposition~\ref{Prop:FanoPicard} provides the bounds $$ w_{ij} \ = \ \deg \, T_{ij} \ = \ \vert \operatorname{Cl}(X,x(i,j)) \vert \ \le \ \mu \qquad \text{for } n_i > 1, \, 1 \le j \le n_i, $$ $$ u_k \ = \ \deg \, S_k \ = \ \vert \operatorname{Cl}(X,x(k)) \vert \ \le \ \mu \qquad \text{for } 1 \le k \le m. $$ Let $0 \le s \le r$ be the maximal number with $n_{s} > 1$. Then $g_{s-2,s-1,s}$ is the last polynomial such that each of its three monomials depends on more than one variable. For any $t \ge s$, we have the ``cut ring'' \begin{eqnarray*} R_t & := & {\mathbb K}[T_{ij}; \; 0 \le i \le t, \; 1 \le j \le n_i] [S_1,\ldots,S_m] \ / \ \bangle{g_{i,i+1,i+2}; \; 0 \le i \le t-2} \end{eqnarray*} where the relations $g_{i,i+1,i+2}$ depend on only three variables as soon as $i > s$ holds. For the degree $\gamma$ of the relations we have \begin{eqnarray*} (r-1)\gamma & = & (t-1)\gamma \ + \ (r-t)\gamma \\ & = & (t-1)\gamma \ + \ l_{t+1,1}w_{t+1,1} + \ldots + l_{r1}w_{r1} \\ & < & \sum_{i=0}^r\sum_{j=1}^{n_i}w_{ij} \ + \ \sum_{i=1}^m u_i \\ & = & \sum_{i=0}^t \sum_{j=1}^{n_i}w_{ij} \ + \ w_{t+1,1}+ \ldots + w_{r1} \ + \ \sum_{i=1}^m u_i. \end{eqnarray*} Since $l_{i1}w_{i1} > w_{i1}$ holds in particular for $t+1 \le i \le r$, we derive from this the inequality \begin{eqnarray*} \gamma & < & \frac{1}{t-1} \left( \sum_{i=0}^t\sum_{j=1}^{n_i}w_{ij} \ + \ \sum_{i=1}^m u_i \right). \end{eqnarray*} To obtain the bounds in Assertions~(iii) and~(iv), we consider the cut ring $R_t$ with $t=2$ and apply Lemma~\ref{Lem:1relation}; note that we have $d = n_0+n_1+n_2+m-2$ for the dimension $d = \dim(X)$ and that $l_{22} \ge 0$ is due to the fact that $X$ is non-toric. The bounds $w_{ij}, l_{0j} < 6d\mu$ in Assertion~(iii) follow from $l_{ij}w_{ij} = \gamma < 6 d\mu$ and $l_{i1} < 2d\mu$ follows from $l_{i1} \mid w_{21}$ for $3 \le i \le r$. Moreover, $l_{i1} \mid w_{11}$ for $2 \le i \le r$ implies $l_{11} \cdots l_{r1} \mid \gamma = l_{11}w_{11}$. Similarly $w_{ij},l_{ij} < 2(d+1)\mu$ in Assertion~(iv) follow from $l_{ij}w_{ij} = \gamma < 2(d+1) d\mu$ and $l_{21} \cdots l_{r1} \mid \gamma = l_{21}w_{21}$ follows from $l_{i1} \mid w_{21}$ for $3 \le i \le r$. The bounds on $r$ in~(iii) in~(iv) are as well consequences of the admissibility condition. To obtain the bounds in Assertion~(v), we consider the cut ring $R_t$ with $t=s$. Using $n_i=1$ for $i \ge t+1$, we can estimate the degree of the relation as follows: $$ \gamma \ \le \ \frac{(n_0 + \ldots + n_t + m) \mu}{t-1} \ = \ \frac{(d + t) \mu}{t-1} \ \le \ (d + 2) \mu. $$ Since we have $w_{ij}l_{ij} \le \deg \, g_0$ for any $0 \le i \le r$ and any $1 \le j \le n_i$, we see that all $w_{ij}$ and $l_{ij}$ are bounded by $(d+2)\mu$. As before, $l_{s+1,1} \cdots l_{r1} \mid \gamma$ is a consequence of $l_{i1} \mid \gamma$ for $i = s+2, \ldots, r$ and also the bound on $r$ follows from the admissibility condition. Finally, we have to express the Picard index $\mu$ in terms of the weights $w_{ij}$ and $u_k$ as claimed in the Assertions. This is a direct application of the formula of Proposition~\ref{Prop:FanoPicard}. Observe that it suffices to work with the $p$-images of the following points: For every $0 \le i \le r$ with $n_i > 1$ take a point $\b{x}(i,j) \in \rq{X}$ with $ij$-coordinate $T_{ij}$ equal to one and all others equal to zero, for every $0 \le i \le r$ with $n_i = 1$ whenever $n_i=1$ take $\b{x}(i,j) \in \rq{X}$ with $ij$-coordinate $T_{ij}$ equal to zero, all other $T_{st}$ equal to one and coordinates $S_k$ equal to zero, and, for every $1 \le k \le m$, take a point $\b{x}(k) \in \rq{X}$ having the $k$-coordinate $S_k$ equal to one and all others zero. \end{proof} We conclude the section with discussing some aspects of the not necessarily Fano varieties of Proposition~\ref{prop:Finite3Var}. Recall that we considered admissible triples $(A,\mathfrak{n},L)$ with $n_0 = \ldots = n_r =1$ and thus rings $R$ of the form $$ {\mathbb K}[T_0,\ldots, T_r,S_1,\ldots, S_m] \ / \ \bangle{\alpha_{i+1,i+2} T_i^{l_i} + \alpha_{i+2,i} T_{i+1}^{l_{i+1}} + \alpha_{i,i+1} T_{i+2}^{l_{i+2}}; \; 0 \le i \le r-2}. $$ \begin{proposition} \label{prop:MoriCox} Suppose that the ring $R$ as above is the Cox ring of a non-toric variety $X$ with $\operatorname{Cl}(X) = {\mathbb Z}$. Then we have $m \ge 1$ and $\mu := [\operatorname{Cl}(X):\operatorname{Pic}(X)] \ge 30$. Moreover, if $X$ is a surface, then we have $m=1$ and $w_i= l_i^{-1} l_0 \cdots l_r$. \end{proposition} \begin{proof} The homogeneity condition $l_{i}w_{i}=l_{j}w_{j}$ together with the admissibility condition $\gcd(l_{i},l_{j})=1$ for $0 \le i \ne j\leq r$ gives us $l_{i} \mid \gcd(w_{j}; j \ne i)$. Moreover, by Proposition~\ref{prop:coxchar}, every set of $m+r$ weights $w_i$ has to generate the class group ${\mathbb Z}$, so they must have greatest common divisor one. Since $X$ is non-toric, $l_{i} \ge 2$ holds and we obtain $m \ge 1$. To proceed, we infer $l_0 \cdots l_r \mid \mu$ and $l_0 \cdots l_r \mid \deg g_{ijk}$ from Proposition~\ref{Prop:FanoPicard}. As a consequence, the minimal value for $\mu$ and $\deg g_{ijk}$ is obviously $2\cdot3\cdot5=30$. what really can be received as the following example shows. Note that if $X$ is a surface we have $m=1$ and $\gcd(w_{i}; 0\le i\le r )=1$. Thus, $l_iw_i=l_jw_j$ gives us $\deg g_{ijk}= l_0 \cdots l_r$ and $w_i= l_i^{-1} l_0 \cdots l_r$. \end{proof} The bound $[\operatorname{Cl}(X):\operatorname{Pic}(X)] \ge 30$ given in the above proposition is even sharp; the surface discussed below realizes it. \begin{example} Consider $X$ with $\mathcal R(X)= {\mathbb K}[T_{0},T_{1},T_{2},T_{3}]/ \langle g\rangle$ with $g=T_{0}^2+T_{1}^3+T_{2}^5$ and the grading $$ \deg \, T_0 \ = \ 15, \quad \deg \, T_1 \ = \ 10, \quad \deg \, T_2 \ = \ 6, \quad \deg \, T_3 \ = \ 1. $$ Then we have $\gcd(15,10)=5$, $\gcd(15,6)=3$ and $\gcd(10,6)=2$ and therefore $[\operatorname{Cl}(X):\operatorname{Pic}(X)]=30$. Further $X$ is Fano because of $$ \deg \, g \ = \ 30 \ < \ 32 \ = \ \deg \, T_0 + \ldots + \deg \, T_3. $$ \end{example} Let us have a look at the geometric meaning of the condition $n_0 = \ldots = n_r = 1$. For a variety $X$ with an action of a torus $T$, we denote by $X_0 \subseteq X$ the union of all orbits with at most finite isotropy. Then there is a possibly non-separated orbit space $X_0/T$; we call it the maximal orbit space. From~\cite{HaSu}, we infer that $n_0 = \ldots = n_r = 1$ holds if and only if $X_0/T$ is separated. Combining this with Propositions~\ref{prop:Finite3Var} and~\ref{prop:MoriCox} gives the following. \begin{corollary} For any pair $(d,\mu) \in {\mathbb Z}^2_{>0}$ there is, up to deformation, only a finite number of $d$-dimensional complete varieties $X$ with a complexity one torus action having divisor class group ${\mathbb Z}$, Picard index $[\operatorname{Cl}(X):\operatorname{Pic}(X)] = \mu$ and maximal orbit space ${\mathbb P}_1$ and for each of these varieties the complement $X \setminus X_0$ contains divisors. \end{corollary} Finally, we present a couple of examples showing that there are also non-Fano varieties with a complexity one torus action having divisor class group ${\mathbb Z}$ and maximal orbit space ${\mathbb P}_1$. \begin{example} Consider $X$ with $\mathcal R(X)= {\mathbb K}[T_{0},T_{1},T_{2},T_{3}]/ \langle g\rangle$ with $g=T_{0}^2+T_{1}^3+T_{2}^{7}$ and the grading $$ \deg \, T_0 \ = \ 21, \quad \deg \, T_1 \ = \ 14, \quad \deg \, T_2 \ = \ 6, \quad \deg \, T_3 \ = \ 1. $$ Then we have $\gcd(21,14)=7$, $\gcd(21,6)=3$ and $\gcd(14,6)=2$ and therefore $[\operatorname{Cl}(X):\operatorname{Pic}(X)]=42$. Moreover, $X$ is not Fano, because its canonical class $\mathcal{K}_X$ is trivial $$ \mathcal{K}_X \ = \ \deg \, g - \deg \, T_0 - \ldots -\deg \, T_3 \ = \ 0. $$ \end{example} \begin{example} Consider $X$ with $\mathcal R(X)= {\mathbb K}[T_{0},T_{1},T_{2},T_{3}]/ \langle g\rangle$ with $g=T_{0}^2+T_{1}^3+T_{2}^{11}$ and the grading $$ \deg \, T_0 \ = \ 33, \quad \deg \, T_1 \ = \ 22, \quad \deg \, T_2 \ = \ 6, \quad \deg \, T_3 \ = \ 1. $$ Then we have $\gcd(22,33)=11$, $\gcd(33,6)=3$ and $\gcd(22,6)=2$ and therefore $[\operatorname{Cl}(X):\operatorname{Pic}(X)]=66$. The canonical class $\mathcal{K}_X$ of $X$ is even ample: $$ \mathcal{K}_X \ = \ \deg \, g - \deg \, T_0 - \ldots - \deg \, T_3 \ = \ 4. $$ \end{example} The following example shows that the Fano assumption is essential for the finiteness results in Theorem~\ref{Th:FiniteIndex}. \begin{remark} For any pair $p,q$ of coprime positive integers, we obtain a locally factorial ${\mathbb K}^*$-surface $X(p,q)$ with $\operatorname{Cl}(X) = {\mathbb Z}$ and Cox ring $$ \mathcal{R}(X(p,q)) \ = \ {\mathbb K}[T_{01},T_{02},T_{11},T_{21}] \ / \ \bangle{g}, \qquad\qquad g \ = \ T_{01}T_{02}^{pq-1} +T_{11}^{q} +T_{21}^{p}; $$ the $\operatorname{Cl}(X)$-grading is given by $\deg \, T_{01} = \deg \, T_{02} =1 $, $\deg \, T_{11} = p$ and $\deg \, T_{21} = q$. Note that $\deg \, g =pq$ holds and for $p,q\geq 3$, the canonical class $\mathcal{K}_X$ satisfies $$ \mathcal{K}_X \ = \ \deg \, g - \deg \, T_{01} - \deg \, T_{02} - \deg \, T_{11} - \deg \, T_{21} \ = \ pq - 2 - p - q \ \ge \ 0. $$ \end{remark} \section{Classification results} \label{sec:tables} In this section, we give classification results for Fano varieties~$X$ with $\operatorname{Cl}(X) \cong {\mathbb Z}$ that come with a complexity one torus action; note that they are necessarily rational. The procedure to obtain classification lists for prescribed dimension $d = \dim \, X$ and Picard index $\mu = [\operatorname{Cl}(X) : \operatorname{Pic}(X)]$ is always the following. By Theorem~\ref{thm:factrings}, we know that their Cox rings are of the form $\mathcal{R}(X) \cong R(A,\mathfrak{n},L)[S_1,\ldots,S_m]$ with admissible triples $(A,\mathfrak{n},L)$. Note that for the family $A = (a_0, \ldots, a_r)$ of points $a_i \in {\mathbb K}^2$, we may assume $$ a_0 \ = \ (1,0), \qquad a_1 \ = \ (1,1), \qquad a_2 \ = \ (0,1). $$ The bounds on the input data of $(A,\mathfrak{n},L)$ provided by Theorem~\ref{Th:FiniteIndex} as well as the criteria of Propositions~\ref{prop:coxchar} and~\ref{Prop:FanoPicard} allow us to generate all the possible Cox rings $\mathcal{R}(X)$ of the Fano varieties $X$ in question for fixed dimension~$d$ and Picard index~$\mu$. Note that $X$ can be reconstructed from $\mathcal{R}(X)= R(A,\mathfrak{n},L)[S_1,\ldots,S_n]$ as the homogeneous spectrum with respect to the $\operatorname{Cl}(X)$-grading. Thus $X$ is classified by its Cox ring $\mathcal{R}(X)$. In the following tables, we present the Cox rings as ${\mathbb K}[T_1, \ldots, T_s]$ modulo relations and fix the ${\mathbb Z}$-gradings by giving the weight vector $(w_1, \ldots, w_s)$, where $w_i := \deg \, T_i$. The first classification result concerns surfaces. \begin{theorem} Let $X$ be a non-toric Fano surface with an effective ${\mathbb K}^*$-action such that $\operatorname{Cl}(X)={\mathbb Z}$ and $[\operatorname{Cl}(X):\operatorname{Pic}(X)]\leq 6$ hold. Then its Cox ring is precisely one of the following. \begin{center} \begin{longtable}[htbp]{llll} \multicolumn{4}{c}{\bf $[\operatorname{Cl}(X):\operatorname{Pic}(X)] = 1$} \\[1ex] \toprule No. & $\mathcal{R}(X)$ & $(w_1,\ldots,w_4)$ & $(-K_X)^2$ \\ \midrule 1 \hspace{.5cm} & ${\mathbb K}[{T_1,\ldots,T_4}]/ \bangle{T_1T_2^5+T_3^3+T_4^2}$ \hspace{.5cm} & $(1,1,2,3)$ \hspace{.5cm} & $1$ \\ \bottomrule \\[2ex] \multicolumn{4}{c}{\bf $[\operatorname{Cl}(X):\operatorname{Pic}(X)] = 2$} \\[1ex] \toprule No. & $\mathcal{R}(X)$ & $(w_1,\ldots,w_4)$ & $(-K_X)^2$ \\ \midrule 2 & ${\mathbb K}[{T_1,\ldots,T_4}]/ \bangle{T_1^4T_2+T_3^3+T_4^2}$ & $(1,2,2,3)$ & $2$ \\ \bottomrule \\[2ex] \multicolumn{4}{c}{\bf $[\operatorname{Cl}(X):\operatorname{Pic}(X)] = 3$} \\[1ex] \toprule No. & $\mathcal{R}(X)$ & $(w_1,\ldots,w_4)$ & $(-K_X)^2$ \\ \midrule 3 & ${\mathbb K}[{T_1,\ldots,T_4}]/ \bangle{T_1^3T_2+T_3^3+T_4^2}$ & $(1,3,2,3)$ & $3$ \\ \midrule 4 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1T_2^3+T_3^5+T_4^2}$ & $(1,3,2,5)$ & $1/3$ \\ \midrule 5 & ${\mathbb K}[{T_1,\ldots,T_4}]/ \bangle{T_1^7T_2+T_3^5+T_4^2}$ & $(1,3,2,5)$ & $1/3$ \\ \bottomrule \\[2ex] \multicolumn{4}{c}{\bf $[\operatorname{Cl}(X):\operatorname{Pic}(X)] = 4$} \\[1ex] \toprule No. & $\mathcal{R}(X)$ & $(w_1,\ldots,w_4)$ & $(-K_X)^2$ \\ \midrule 6 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^2T_2+T_3^3+T_4^2}$ & $(1,4,2,3)$ & $4$ \\ \midrule 7 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^6T_2+T_3^5+T_4^2}$ & $(1,4,2,5)$ & $1$ \\ \bottomrule \\[2ex] \multicolumn{4}{c}{\bf $[\operatorname{Cl}(X):\operatorname{Pic}(X)] = 5$} \\[1ex] \midrule No. & $\mathcal{R}(X)$ & $(w_1,\ldots,w_4)$ & $(-K_X)^2$ \\ \midrule 8 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1T_2+T_3^3+T_4^2}$ & $(1,5,2,3)$ & $5$ \\ \midrule 9 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^5T_2+T_3^5+T_4^2}$ & $(1,5,2,5)$ & $9/5$ \\ \midrule 10 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^9T_2+T_3^7+T_4^2}$ & $(1,5,2,7)$ & $1/5$ \\ \midrule 11 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^7T_2+T_3^4+T_4^3}$ & $(1,5,3,4)$ & $1/5$ \\ \bottomrule \\[2ex] \multicolumn{4}{c}{\bf $[\operatorname{Cl}(X):\operatorname{Pic}(X)] = 6$} \\[1ex] \toprule No. & $\mathcal{R}(X)$ & $(w_1,\ldots,w_4)$ & $(-K_X)^2$ \\ \midrule 12 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^4T_2+T_3^5+T_4^2}$ & $(1,6,2,5)$ & $8/3$ \\ \midrule 13 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^8T_2+T_3^7+T_4^2}$ & $(1,6,2,7)$ & $2/3$ \\ \midrule 14 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^6T_2+T_3^4+T_4^3}$ & $(1,6,3,4)$ & $2/3$ \\ \midrule 15 & ${\mathbb K}[{T_1,\ldots,T_4}] / \bangle{T_1^9T_2+T_3^3+T_4^2}$ & $(1,3,4,6)$ & $2/3$ \\ \bottomrule \end{longtable} \end{center} \end{theorem} \begin{proof} As mentioned, Theorems~\ref{thm:factrings}, \ref{Th:FiniteIndex} and Propositions~\ref{prop:coxchar}, \ref{Prop:FanoPicard} produce a list of all Cox rings of surfaces with the prescribed data. Doing this computation, we obtain the list of the assertion. Note that none of the Cox rings listed is a polynomial ring and hence none of the resulting surfaces $X$ is a toric variety. To show that different members of the list are not isomorphic to each other, we use the following two facts. Firstly, observe that any two minimal systems of homogeneous generators of the Cox ring have (up to reordering) the same list of degrees, and thus the list of generator degrees is invariant under isomorphism (up to reordering). Secondly, by Construction~\ref{constr:Kgrading}, the exponents $l_{ij} >1$ are precisely the orders of the non-trivial isotropy groups of one-codimensional orbits of the action of the torus $T$ on $X$. Using both principles and going through the list, we see that different members $X$ cannot be $T$-equivariantly isomorphic to each other. Since all listed $X$ are non-toric, the effective complexity one torus action on each $X$ corresponds to a maximal torus in the linear algebraic group ${\rm Aut}(X)$. Any two maximal tori in the automorphism group are conjugate, and thus we can conclude that two members are isomorphic if and only if they are $T$-equivariantly isomorphic. \end{proof} We remark that in~\cite[Section~4]{tfano}, log del Pezzo surfaces with an effective ${\mathbb K}^*$-action and Picard number 1 and Gorenstein index less than 4 were classified. The above list contains six such surfaces, namely no. 1-4, 6 and~8; these are exactly the ones where the maximal exponents of the monomials form a platonic triple, i.e., are of the form $(1,k,l)$, $(2,2,k)$, $(2,3,3)$, $(2,3,4)$ or $(2,3,5)$. The remaining ones, i.e., no. 5, 7, and~9-15 have non-log-terminal and thus non-rational singularities; to check this one may compute the resolutions via resolution of the ambient weighted projective space as in~\cite[Ex.~7.5]{Ha2}. With the same scheme of proof as in the surface case, one establishes the following classification results on Fano threefolds. \goodbreak \begin{theorem} \label{thm:3fano} Let $X$ be a three-dimensional locally factorial non-toric Fano variety with an effective two torus action such that $\operatorname{Cl}(X) = {\mathbb Z}$ holds. Then its Cox ring is precisely one of the following. \begin{center} \begin{longtable}[htbp]{llll} \toprule No. & $\mathcal{R}(X)$ & $(w_1,\ldots, w_5)$ & $(-K_X)^3$ \\ \midrule 1 \hspace{.5cm} & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^5 + T_3^3 + T_4^2} $ \hspace{.5cm} & $(1,1,2,3,1)$ \hspace{.5cm} & $8$ \\ \midrule 2 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2T_3^4 + T_4^3 + T_5^2} $ & $(1,1,1,2,3)$ & $8$ \\ \midrule 3 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^2T_3^3 + T_4^3 + T_5^2} $ & $(1,1,1,2,3)$ & $8$ \\ \midrule 4 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2 + T_3T_4 + T_5^2} $ & $(1,1,1,1,1)$ & $54$ \\ \midrule 5 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^2 + T_3T_4^2 + T_5^3} $ & $(1,1,1,1,1)$ & $24$ \\ \midrule 6 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^3 + T_3T_4^3 + T_5^4} $ & $(1,1,1,1,1)$ & $4$ \\ \midrule 7 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^3 + T_3T_4^3 + T_5^2} $ & $(1,1,1,1,2)$ & $16$ \\ \midrule 8 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^5 + T_3T_4^5 + T_5^2} $ & $(1,1,1,1,3)$ & $2$ \\ \midrule 9 & $ {\mathbb K}[T_1, \ldots, T_5] \ / \ \bangle{T_1T_2^5 + T_3^3T_4^3 + T_5^2} $ & $(1,1,1,1,3)$ & $2$ \\ \bottomrule \end{longtable} \end{center} \end{theorem} The singular threefolds listed in this theorem are rational degenerations of smooth Fano threefolds from~\cite{fano3}. The (smooth) general Fano threefolds of the corresponding families are non-rational see~\cite{Gri} for no.~1-3, \cite{CG} for no.~5, \cite{IM} for no.~6, \cite{voi,tim}~for no.~7 and \cite{Isk80} for no. 8-9. Even if one allows certain mild singularities, one still has non-rationality in some cases, see \cite{Gri2}, \cite{Co,Pu}, \cite{CM}, \cite{CP}. \begin{theorem} \label{thm:3fano2} Let $X$ be a three-dimensional non-toric Fano variety with an effective two torus action such that $\operatorname{Cl}(X)={\mathbb Z}$ and $[\operatorname{Cl}(X):\operatorname{Pic}(X)]=2$ hold. Then its Cox ring is precisely one of the following. \begin{center} \begin{longtable}[htbp]{llll} \toprule No. \hspace{.5cm} & $\mathcal{R}(X)$ \hspace{.5cm} & $(w_1,\ldots,w_5)$ \hspace{.5cm} & $(-K_X)^3$ \\ \midrule 1 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2+T_3^3+T_4^2 \rangle$ & $(1,2,2,3,1)$ & $27/2$ \\ \midrule 2 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2^3+T_3^5+T_4^2 \rangle$ & $(1,2,2,5,1)$ & $1/2$ \\ \midrule 3 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^8T_2+T_3^5+T_4^2 \rangle$ & $(1,2,2,5,1)$ & $1/2$ \\ \midrule 4 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2+T_3^3+T_4^2 \rangle$ & $(1,2,2,3,2)$ & $16$ \\ \midrule 5 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2^3+T_3^5+T_4^2 \rangle$ & $(1,2,2,5,2)$ & $2$ \\ \midrule 6 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^8T_2+T_3^5+T_4^2 \rangle$ & $(1,2,2,5,2)$ & $2$ \\ \midrule 7 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5+T_3^3+T_4^2 \rangle$ & $(1,1,2,3,2)$ & $27/2$ \\ \midrule 8 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^9+T_3^5+T_4^2 \rangle$ & $(1,1,2,5,2)$ & $1/2$ \\ \midrule 9 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^7+T_3^5+T_4^2 \rangle$ & $(1,1,2,5,2)$ & $1/2$ \\ \midrule 10 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^{11}+T_3^3+T_4^2 \rangle$ & $(1,1,4,6,1)$ & $1/2$ \\ \midrule 11 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^5T_2^7+T_3^3+T_4^2 \rangle$ & $(1,1,4,6,1)$ & $1/2$ \\ \midrule 12 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^{11}+T_3^3+T_4^2 \rangle$ & $(1,1,4,6,2)$ & $2$ \\ \midrule 13 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^5T_2^7+T_3^3+T_4^2 \rangle$ & $(1,1,4,6,2)$ & $2$ \\ \midrule 14 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^5+T_3^3+T_4^2 \rangle$ & $(1,2,4,6,1)$ & $2$ \\ \midrule 15 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^{10}T_2+T_3^3+T_4^2 \rangle$ & $(1,2,4,6,1)$ & $2$ \\ \midrule 16 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^2+T_3^3+T_4^2 \rangle$ & $(2,2,2,3,1)$ & $16$ \\ \midrule 17 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^4+T_3^5+T_4^2 \rangle$ & $(2,2,2,5,1)$ & $2$ \\ \midrule 18 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^3+T_3^5+T_4^2 \rangle$ & $(2,2,2,5,1)$ & $2$ \\ \midrule 19 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^2+T_3T_4+T_5^3 \rangle$ & $(1,1,1,2,1)$ & $81/2$ \\ \midrule 20 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^4+T_3T_4^2+T_5^5 \rangle$ & $(1,1,1,2,1)$ & $5/2$ \\ \midrule 21 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^3+T_3T_4^2+T_5^5 \rangle$ & $(1,1,1,2,1)$ & $5/2$ \\ \midrule 22 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^3+T_3^2T_4+T_5^4 \rangle$ & $(1,1,1,2,1)$ & $16$ \\ \midrule 23 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^4+T_3^3T_4+T_5^5 \rangle$ & $(1,1,1,2,1)$ & $5/2$ \\ \midrule 24 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^3+T_3^3T_4+T_5^5 \rangle$ & $(1,1,1,2,1)$ & $5/2$ \\ \midrule 25 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^3+T_3^2T_4+T_5^2 \rangle$ & $(1,1,1,2,2)$ & $27$ \\ \midrule 26 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5+T_3^2T_4^2+T_5^3 \rangle$ & $(1,1,1,2,2)$ & $3/2$ \\ \midrule 27 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5+T_3^4T_4+T_5^3 \rangle$ & $(1,1,1,2,2)$ & $3/2$ \\ \midrule 28 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^4+T_3^4T_4+T_5^3 \rangle$ & $(1,1,1,2,2)$ & $3/2$ \\ \midrule 29 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5+T_3^4T_4+T_5^2 \rangle$ & $(1,1,1,2,3)$ & $8$ \\ \midrule 30 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^3+T_3^4T_4+T_5^2 \rangle$ & $(1,1,1,2,3)$ & $8$ \\ \midrule 31 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^7+T_3^2T_4^3+T_5^2 \rangle$ & $(1,1,1,2,4)$ & $1$ \\ \midrule 32 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^5+T_3^2T_4^3+T_5^2 \rangle$ & $(1,1,1,2,4)$ & $1$ \\ \midrule 33 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^7+T_3^6T_4+T_5^2 \rangle$ & $(1,1,1,2,4)$ & $1$ \\ \midrule 34 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^5+T_3^6T_4+T_5^2 \rangle$ & $(1,1,1,2,4)$ & $1$ \\ \midrule 35 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^3+T_3T_4+T_5^4 \rangle$ & $(1,1,2,2,1)$ & $27$ \\ \midrule 36 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5+T_3T_4^2+T_5^6 \rangle$ & $(1,1,2,2,1)$ & $3/2$ \\ \midrule 37 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^3+T_3T_4+T_5^2 \rangle$ & $(1,1,2,2,2)$ & $16$ \\ \midrule 38 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5+T_3T_4^2+T_5^3 \rangle$ & $(1,1,2,2,2)$ & $6$ \\ \midrule 39 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^4+T_3T_4^2+T_5^3 \rangle$ & $(1,1,2,2,2)$ & $6$ \\ \midrule 40 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^3+T_3T_4^2+T_5^2 \rangle$ & $(1,1,2,2,2)$ & $27/2$ \\ \midrule 41 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^5+T_3T_4^3+T_5^2 \rangle$ & $(1,1,2,2,2)$ & $32$ \\ \midrule 42 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5+T_3T_4^2+T_5^2 \rangle$ & $(1,1,2,2,3)$ & $4$ \\ \midrule 43 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^7+T_3T_4^3+T_5^2 \rangle$ & $(1,1,2,2,4)$ & $32$ \\ \midrule 44 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^9+T_3T_4^4+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 45 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^9+T_3^2T_4^3+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 46 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^7+T_3T_4^4+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 47 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^7+T_3^2T_4^3+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 48 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^5T_2^5+T_3T_4^4+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 49 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^5T_2^5+T_3^2T_4^3+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 50 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2+T_3T_4+T_5^3 \rangle$ & $(1,2,1,2,1)$ & $48$ \\ \midrule 51 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2+T_3^2T_4+T_5^4 \rangle$ & $(1,2,1,2,1)$ & $27$ \\ \midrule 52 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^2+T_3T_4^2+T_5^5 \rangle$ & $(1,2,1,2,1)$ & $10$ \\ \midrule 53 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^2+T_3^3T_4+T_5^5 \rangle$ & $(1,2,1,2,1)$ & $10$ \\ \midrule 54 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2+T_3^3T_4+T_5^5 \rangle$ & $(1,2,1,2,1)$ & $10$ \\ \midrule 55 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2+T_3^4T_4+T_5^6 \rangle$ & $(1,2,1,2,1)$ & $3/2$ \\ \midrule 56 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2+T_3^2T_4+T_5^2 \rangle$ & $(1,2,1,2,2)$ & $32$ \\ \midrule 57 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^2+T_3^4T_4+T_5^3 \rangle$ & $(1,2,1,2,2)$ & $6$ \\ \midrule 58 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2+T_3^4T_4+T_5^3 \rangle$ & $(1,2,1,2,2)$ & $6$ \\ \midrule 59 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2+T_3^4T_4+T_5^2 \rangle$ & $(1,2,1,2,3)$ & $27/2$ \\ \midrule 60 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^3+T_3^2T_4^3+T_5^2 \rangle$ & $(1,2,1,2,4)$ & $4$ \\ \midrule 61 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^3+T_3^6T_4+T_5^2 \rangle$ & $(1,2,1,2,4)$ & $4$ \\ \midrule 62 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^6T_2+T_3^6T_4+T_5^2 \rangle$ & $(1,2,1,2,4)$ & $4$ \\ \midrule 63 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2^3+T_3^4T_4^3+T_5^2 \rangle$ & $(1,2,1,2,5)$ & $1/2$ \\ \midrule 64 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^8T_2+T_3^4T_4^3+T_5^2 \rangle$ & $(1,2,1,2,5)$ & $1/2$ \\ \midrule 65 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^8T_2+T_3^8T_4+T_5^2 \rangle$ & $(1,2,1,2,5)$ & $1/2$ \\ \midrule 66 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2+T_3T_4+T_5^4 \rangle$ & $(1,2,2,2,1)$ & $32$ \\ \midrule 67 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2+T_3T_4^2+T_5^6 \rangle$ & $(1,2,2,2,1)$ & $6$ \\ \midrule 68 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2+T_3T_4^2+T_5^2 \rangle$ & $(1,2,2,2,3)$ & $16$ \\ \midrule 69 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2^3+T_3T_4^4+T_5^2 \rangle$ & $(1,2,2,2,5)$ & $2$ \\ \midrule 70 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2^3+T_3^2T_4^3+T_5^2 \rangle$ & $(1,2,2,2,5)$ & $2$ \\ \midrule 71 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^8T_2+T_3T_4^4+T_5^2 \rangle$ & $(1,2,2,2,5)$ & $2$ \\ \midrule 72 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^8T_2+T_3^2T_4^3+T_5^2 \rangle$ & $(1,2,2,2,5)$ & $2$ \\ \midrule 73 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2T_3^{10}+T_4^3+T_5^2 \rangle$ & $(1,1,1,4,6)$ & $1/2$ \\ \midrule 74 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^2T_3^9+T_4^3+T_5^2 \rangle$ & $(1,1,1,4,6)$ & $1/2$ \\ \midrule 75 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^3T_3^8+T_4^3+T_5^2 \rangle$ & $(1,1,1,4,6)$ & $1/2$ \\ \midrule 76 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^4T_3^7+T_4^3+T_5^2 \rangle$ & $(1,1,1,4,6)$ & $1/2$ \\ \midrule 77 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5T_3^6+T_4^3+T_5^2 \rangle$ & $(1,1,1,4,6)$ & $1/2$ \\ \midrule 78 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^3T_3^7+T_4^3+T_5^2 \rangle$ & $(1,1,1,4,6)$ & $1/2$ \\ \midrule 79 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^5T_3^5+T_4^3+T_5^2 \rangle$ & $(1,1,1,4,6)$ & $1/2$ \\ \midrule 80 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^4T_3^5+T_4^3+T_5^2 \rangle$ & $(1,1,1,4,6)$ & $1/2$ \\ \midrule 81 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2T_3^2+T_4^3+T_5^2 \rangle$ & $(1,1,2,2,3)$ & $27/2$ \\ \midrule 82 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^3T_3+T_4^3+T_5^2 \rangle$ & $(1,1,2,2,3)$ & $27/2$ \\ \midrule 83 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^2T_3+T_4^3+T_5^2 \rangle$ & $(1,1,2,2,3)$ & $27/2$ \\ \midrule 84 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2T_3^4+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 85 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^3T_3^3+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 86 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5T_3^2+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 87 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^7T_3+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 88 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^2T_3^3+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 89 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^6T_3+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 90 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^3T_3^2+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 91 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^5T_3+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 92 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2^4T_3+T_4^5+T_5^2 \rangle$ & $(1,1,2,2,5)$ & $1/2$ \\ \midrule 93 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2T_3^5+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 94 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^3T_3^4+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 95 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^5T_3^3+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 96 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^7T_3^2+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 97 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1T_2^9T_3+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 98 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^4T_3^3+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 99 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2^8T_3+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 100 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^5T_3^2+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 101 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^3T_2^7T_3+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 102 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2^6T_3+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 103 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^5T_2^5T_3+T_4^3+T_5^2 \rangle$ & $(1,1,2,4,6)$ & $2$ \\ \midrule 104 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2T_3+T_4^3+T_5^2 \rangle$ & $(1,2,2,2,3)$ & $16$ \\ \midrule 105 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^2T_2T_3^3+T_4^5+T_5^2 \rangle$ & $(1,2,2,2,5)$ & $2$ \\ \midrule 106 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^4T_2T_3^2+T_4^5+T_5^2 \rangle$ & $(1,2,2,2,5)$ & $2$ \\ \midrule 107 & ${\mathbb K}[{T_1,\ldots,T_5}]/ \langle T_1^6T_2T_3+T_4^5+T_5^2 \rangle$ & $(1,2,2,2,5)$ & $2$ \\ \bottomrule \end{longtable} \end{center} \end{theorem} The varieties no. 2,3 and 25, 26 are rational degenerations of quasismooth varieties from the list in \cite{IaFl}. In \cite{CPR} the non-rationality of a general (quasismooth) element of the corresponding family was proved. The varieties listed so far might suggest that we always obtain only one relation in the Cox ring. We discuss now some examples, showing that for a Picard index big enough, we need in general more than one relation, where this refers always to a presentation as in Theorem~\ref{thm:factrings}~(ii). \begin{example} \label{ex:fanosurf2rel} A Fano ${\mathbb K}^*$-surface $X$ with $\operatorname{Cl}(X)={\mathbb Z}$ such that the Cox ring $\mathcal{R}(X)$ needs two relations. Consider the ${\mathbb Z}$-graded ring \begin{eqnarray*} R & = & {\mathbb K}[T_{01},T_{02},T_{11},T_{21},T_{31}]/\bangle{g_0,g_1}, \end{eqnarray*} where the degrees of $T_{01},T_{02},T_{11},T_{21},T_{31}$ are $29,1,6,10,15$, respectively, and the relations $g_0,g_1$ are given by $$ g_0 \ := \ T_{01}T_{02}+T_{11}^5+T_{21}^3, \qquad\qquad g_1 \ := \ \alpha_{23} T_{11}^5+\alpha_{31} T_{21}^3+\alpha_{12}T_{31}^2 $$ Then $R$ is the Cox ring of a Fano ${\mathbb K}^*$-surface. Note that the Picard index is given by $ [\operatorname{Cl}(X):\operatorname{Pic}(X)]= \mathrm{lcm}(29,1)=29. $ \end{example} \begin{proposition} \label{prop:fano22rel} Let $X$ be a non-toric Fano surface with an effective ${\mathbb K}^*$-action such that $\operatorname{Cl}(X) \cong {\mathbb Z}$ and $[\operatorname{Cl}(X):\operatorname{Pic}(X)] < 29$ hold. Then the Cox ring of $X$ is of the form \begin{eqnarray*} \mathcal{R}(X) & \cong & {\mathbb K}[T_1, \ldots, T_4]/\bangle{T_1^{l_1}T_2^{l_2} + T_3^{l_3} + T_4^{l_4}}. \end{eqnarray*} \end{proposition} \begin{proof} The Cox ring $\mathcal{R}(X)$ is as in Theorem~\ref{thm:factrings}, and, in the notation used there, we have $n_0 + \ldots + n_r + m = 2+r$. This leaves us with the possibilities $n_0=m=1$ and $n_0=2$, $m=0$. In the first case, Proposition~\ref{prop:MoriCox} tells us that the Picard index of $X$ is at least $30$. So, consider the case $n_0=2$ and $m=0$. Then, according to Theorem~\ref{thm:factrings}, the Cox ring $\mathcal{R}(X)$ is ${\mathbb K}[T_{01},T_{02},T_1 \ldots, T_r]$ divided by relations $$ g_{0,1,2}=T_{01}^{l_{01}}T_{02}^{l_{02}} + T_1^{l_1} + T_2^{l_2}, \quad g_{i,i+1,i+2}= \alpha_{i+1,i+2}T_i^{l_i} + \alpha_{i+2,i}T_{i+1}^{l_{i+1}} + \alpha_{i,i+1}T_{i+2}^{l_{i+2}}, $$ where $1 \le i \le r-2$. We have to show that $r=2$ holds. Set $\mu := [\operatorname{Cl}(X):\operatorname{Pic}(X)]$ and let $\gamma \in {\mathbb Z}$ denote the degree of the relations. Then we have $\gamma = w_il_i$ for $1 \le i \le r$, where $w_i := \deg \, T_i$. With $w_{0i} := \deg \, T_{0i}$, Proposition~\ref{Prop:FanoPicard} gives us \begin{eqnarray*} (r-1) \gamma & < & w_{01} + w_{02} + w_1 + \ldots + w_r. \end{eqnarray*} We claim that $w_{01}$ and $w_{02}$ are coprime. Otherwise they had a common prime divisor $p$. This $p$ divides $\gamma = l_iw_i$. Since $l_1,\ldots,l_r$ are pairwise coprime, $p$ divides at least $r-1$ of the weights $w_1,\ldots, w_r$. This contradicts the Cox ring condition that any $r+1$ of the $r+2$ weights generate the class group ${\mathbb Z}$. Thus, $w_{01}$ and $w_{02}$ are coprime and we obtain $$ \mu \ \ge \ \rm{lcm}(w_{01},w_{02}) \ = \ w_{01}\cdot w_{02} \ \ge \ w_{01}+w_{02}-1. $$ Now assume that $r \ge 3$ holds. Then we can conclude $$ 2 \gamma \ < \ w_{01} + w_{02} + w_1 + w_2 + w_3 \ \le \ \mu + 1 + \gamma \left( \frac{1}{l_1} + \frac{1}{l_2} + \frac{1}{l_3} \right) $$ Since the numbers $l_i$ are pairwise coprime, we obtain $l_1 \ge 5$, $l_2 \ge 3$ and $l_3 \ge 2$. Moreover, $l_iw_i = l_jw_j$ implies $l_i \mid w_j$ and hence $l_1l_2l_3 \mid \gamma$. Thus, we have $\gamma \ge 30$. Plugging this in the above inequality gives $$ \mu \ \ge \ \gamma\left(2- \frac{1}{l_1} - \frac{1}{l_2} - \frac{1}{l_3} \right)-1 \ = \ 29. $$ \end{proof} The Fano assumption is essential in this result; if we omit it, then we may even construct locally factorial surfaces with a Cox ring that needs more then one relation. \begin{example} A locally factorial ${\mathbb K}^*$-surface $X$ with $\operatorname{Cl}(X)={\mathbb Z}$ such that the Cox ring $\mathcal{R}(X)$ needs two relations. Consider the ${\mathbb Z}$-graded ring \begin{eqnarray*} R & = & {\mathbb K}[T_{01},T_{02},T_{11},T_{21},T_{31}]/\bangle{g_0,g_1}, \end{eqnarray*} where the degrees of $T_{01},T_{02},T_{11},T_{21},T_{31}$ are $1,1,6,10,15$, respectively, and the relations $g_0,g_1$ are given by $$ g_0 \ := \ T_{01}^7T_{02}^{23}+T_{11}^5+T_{21}^3, \qquad\qquad g_1 \ := \ \alpha_{23} T_{11}^5+\alpha_{31}T_{21}^3+\alpha_{12}T_{31}^2 $$ Then $R$ is the Cox ring of a non Fano ${\mathbb K}^*$-surface~$X$ of Picard index one, i.e, $X$ is locally factorial. \end{example} For non-toric Fano threefolds~$X$ with an effective 2-torus action $\operatorname{Cl}(X) \cong {\mathbb Z}$, the classifications~\ref{thm:3fano} and~\ref{thm:3fano2} show that for Picard indices one and two we only obtain hypersurfaces as Cox rings. The following example shows that this stops at Picard index three. \begin{example} \label{ex:fano32rel} A Fano threefold $X$ with $\operatorname{Cl}(X)={\mathbb Z}$ and a 2-torus action such that the Cox ring $\mathcal{R}(X)$ needs two relations. Consider \begin{eqnarray*} R & = & {\mathbb K}[T_{01},T_{02},T_{11},T_{12},T_{21},T_{31}]/ \bangle{g_0,g_1} \end{eqnarray*} where the degrees of $T_{01},T_{02},T_{11},T_{12},T_{21},T_{31}$ are $1,1,3,3,2,3$, respectively, and the relations are given by $$ g_0 \ = \ T_{01}^5T_{02}+T_{11}T_{12}+T_{21}^3, \qquad g_1 \ = \ \alpha_{23} T_{11}T_{12}+\alpha_{31}T_{21}^3+\alpha_{12}T_{31}^2. $$ Then $R$ is the Cox ring of a Fano threefold with a 2-torus action. Note that the Picard index is given by $$ [\operatorname{Cl}(X):\operatorname{Pic}(X)] \ = \ \mathrm{lcm}(1,1,3,3) \ = \ 3. $$ \end{example} Finally, we turn to locally factorial Fano fourfolds. Here we observe more than one relation in the Cox ring even in the locally factorial case. \begin{theorem} Let $X$ be a four-dimensional locally factorial non-toric Fano variety with an effective three torus action such that $\operatorname{Cl}(X)={\mathbb Z}$ holds. Then its Cox ring is precisely one of the following. \begin{center} \begin{longtable}[htbp]{llll} \toprule No. & $\mathcal{R}(X)$ & $(w_1,\ldots,w_6)$ & $(-K_X)^4$ \\ \midrule 1 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^5+T_3^3+T_4^2 \rangle$ & $(1,1,2,3,1,1)$ & $81$ \\ \midrule 2 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^9+T_3^2+T_4^5 \rangle$ & $(1,1,2,5,1,1)$ & $1$ \\ \midrule 3 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^3T_2^7+T_3^2+T_4^5 \rangle$ & $(1,1,2,5,1,1)$ & $1$ \\ \midrule 4 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^4+T_4^3+T_5^2 \rangle$ & $(1,1,1,2,3,1)$ & $81$ \\ \midrule 5 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^3+T_4^3+T_5^2 \rangle$ & $(1,1,1,2,3,1)$ & $81$ \\ \midrule 6 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^8+T_4^5+T_5^2 \rangle$ & $(1,1,1,2,5,1)$ & $1$ \\ \midrule 7 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^7+T_4^5+T_5^2 \rangle$ & $(1,1,1,2,5,1)$ & $1$ \\ \midrule 8 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^3T_3^6+T_4^5+T_5^2 \rangle$ & $(1,1,1,2,5,1)$ & $1$ \\ \midrule 9 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^4T_3^5+T_4^5+T_5^2 \rangle$ & $(1,1,1,2,5,1)$ & $1$ \\ \midrule 10 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^2T_2^3T_3^5+T_4^5+T_5^2 \rangle$ & $(1,1,1,2,5,1)$ & $1$ \\ \midrule 11 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^3T_2^3T_3^4+T_4^5+T_5^2 \rangle$ & $(1,1,1,2,5,1)$ & $1$ \\ \midrule 12 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2+T_3T_4+T_5^2 \rangle$ & $(1,1,1,1,1,1)$ & $512$ \\ \midrule 13 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2+T_3T_4^2+T_5^3 \rangle$ & $(1,1,1,1,1,1)$ & $243$ \\ \midrule 14 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^3+T_3T_4^3+T_5^4 \rangle$ & $(1,1,1,1,1,1)$ & $64$ \\ \midrule 15 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^4+T_3T_4^4+T_5^5 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 16 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^4+T_3^2T_4^3+T_5^5 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 17 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^2T_2^3+T_3^2T_4^3+T_5^5 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 18 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^3+T_3T_4^3+T_5^2 \rangle$ & $(1,1,1,1,2,1)$ & $162$ \\ \midrule 19 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^5+T_3T_4^5+T_5^3 \rangle$ & $(1,1,1,1,2,1)$ & $3$ \\ \midrule 20 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^5+T_3^2T_4^4+T_5^3 \rangle$ & $(1,1,1,1,2,1)$ & $3$ \\ \midrule 21 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^5+T_3T_4^5+T_5^2 \rangle$ & $(1,1,1,1,3,1)$ & $32$ \\ \midrule 22 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^5+T_3^3T_4^3+T_5^2 \rangle$ & $(1,1,1,1,3,1)$ & $32$ \\ \midrule 23 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^7+T_3T_4^7+T_5^2 \rangle$ & $(1,1,1,1,4,1)$ & $2$ \\ \midrule 24 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^7+T_3^3T_4^5+T_5^2 \rangle$ & $(1,1,1,1,4,1)$ & $2$ \\ \midrule 25 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^3T_2^5+T_3^3T_4^5+T_5^2 \rangle$ & $(1,1,1,1,4,1)$ & $2$ \\ \midrule 26 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3T_4^3+T_5^3+T_6^2 \rangle$ & $(1,1,1,1,2,3)$ & $81$ \\ \midrule 27 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^2T_4^2+T_5^3+T_6^2 \rangle$ & $(1,1,1,1,2,3)$ & $81$ \\ \midrule 28 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3T_4^7+T_5^5+T_6^2 \rangle$ & $(1,1,1,1,2,5)$ & $1$ \\ \midrule 29 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^2T_4^6+T_5^5+T_6^2 \rangle$ & $(1,1,1,1,2,5)$ & $1$ \\ \midrule 30 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^3T_4^5+T_5^5+T_6^2 \rangle$ & $(1,1,1,1,2,5)$ & $1$ \\ \midrule 31 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^4T_4^4+T_5^5+T_6^2 \rangle$ & $(1,1,1,1,2,5)$ & $1$ \\ \midrule 32 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^2T_4^5+T_5^5+T_6^2 \rangle$ & $(1,1,1,1,2,5)$ & $1$ \\ \midrule 33 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^3T_4^4+T_5^5+T_6^2 \rangle$ & $(1,1,1,1,2,5)$ & $1$ \\ \midrule 34 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^3T_3^3T_4^3+T_5^5+T_6^2 \rangle$ & $(1,1,1,1,2,5)$ & $1$ \\ \midrule 35 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^2T_2^2T_3^3T_4^3+T_5^5+T_6^2 \rangle$ & $(1,1,1,1,2,5)$ & $1$ \\ \midrule 36 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3+T_4T_5^2+T_6^3 \rangle$ & $(1,1,1,1,1,1)$ & $243$ \\ \midrule 37 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^2+T_4T_5^3+T_6^4 \rangle$ & $(1,1,1,1,1,1)$ & $64$ \\ \midrule 38 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^3+T_4T_5^4+T_6^5 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 39 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^3+T_4^2T_5^3+T_6^5 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 40 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^2+T_4T_5^4+T_6^5 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 41 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^2+T_4^2T_5^3+T_6^5 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 42 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^2+T_4T_5^3+T_6^2 \rangle$ & $(1,1,1,1,1,2)$ & $162$ \\ \midrule 43 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^4+T_4T_5^5+T_6^3 \rangle$ & $(1,1,1,1,1,2)$ & $3$ \\ \midrule 44 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^4+T_4^2T_5^4+T_6^3 \rangle$ & $(1,1,1,1,1,2)$ & $3$ \\ \midrule 45 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^3+T_4T_5^5+T_6^3 \rangle$ & $(1,1,1,1,1,2)$ & $3$ \\ \midrule 46 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^3+T_4^2T_5^4+T_6^3 \rangle$ & $(1,1,1,1,1,2)$ & $3$ \\ \midrule 47 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^2T_2^2T_3^2+T_4T_5^5+T_6^3 \rangle$ & $(1,1,1,1,1,2)$ & $3$ \\ \midrule 48 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^3+T_4^3T_5^3+T_6^2 \rangle$ & $(1,1,1,1,1,3)$ & $32$ \\ \midrule 49 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^3+T_4T_5^5+T_6^2 \rangle$ & $(1,1,1,1,1,3)$ & $32$ \\ \midrule 50 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^4+T_4^3T_5^3+T_6^2 \rangle$ & $(1,1,1,1,1,3)$ & $32$ \\ \midrule 51 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^4+T_4T_5^5+T_6^2 \rangle$ & $(1,1,1,1,1,3)$ & $32$ \\ \midrule 52 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^6+T_4T_5^7+T_6^2 \rangle$ & $(1,1,1,1,1,4)$ & $2$ \\ \midrule 53 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2T_3^6+T_4^3T_5^5+T_6^2 \rangle$ & $(1,1,1,1,1,4)$ & $2$ \\ \midrule 54 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^5+T_4T_5^7+T_6^2 \rangle$ & $(1,1,1,1,1,4)$ & $2$ \\ \midrule 55 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2T_3^5+T_4^3T_5^5+T_6^2 \rangle$ & $(1,1,1,1,1,4)$ & $2$ \\ \midrule 56 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^3T_3^4+T_4T_5^7+T_6^2 \rangle$ & $(1,1,1,1,1,4)$ & $2$ \\ \midrule 57 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^3T_3^4+T_4^3T_5^5+T_6^2 \rangle$ & $(1,1,1,1,1,4)$ & $2$ \\ \midrule 58 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^2T_2^3T_3^3+T_4T_5^7+T_6^2 \rangle$ & $(1,1,1,1,1,4)$ & $2$ \\ \midrule 59 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^2T_2^3T_3^3+T_4^3T_5^5+T_6^2 \rangle$ & $(1,1,1,1,1,4)$ & $2$ \\ \midrule 60 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2+T_3T_4+T_5T_6 \rangle$ & $(1,1,1,1,1,1)$ & $512$ \\ \midrule 61 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^2+T_3T_4^2+T_5T_6^2 \rangle$ & $(1,1,1,1,1,1)$ & $243$ \\ \midrule 62 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^3+T_3T_4^3+T_5T_6^3 \rangle$ & $(1,1,1,1,1,1)$ & $64$ \\ \midrule 63 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^3+T_3T_4^3+T_5^2T_6^2 \rangle$ & $(1,1,1,1,1,1)$ & $64$ \\ \midrule 64 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^4+T_3T_4^4+T_5T_6^4 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 65 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^4+T_3T_4^4+T_5^2T_6^3 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 66 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1T_2^4+T_3^2T_4^3+T_5^2T_6^3 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 67 & ${\mathbb K}[{T_1,\ldots,T_6}]/ \langle T_1^2T_2^3+T_3^2T_4^3+T_5^2T_6^3 \rangle$ & $(1,1,1,1,1,1)$ & $5$ \\ \midrule 68 & ${\mathbb K}[T_1,\ldots,T_7] / \left\langle \begin{smallmatrix} T_1T_2 + T_3T_4 + T_5T_6, \\ \alpha T_3T_4 + T_5T_6 + T_7^2 \end{smallmatrix} \right\rangle $ & $(1,1,1,1,1,1,1)$ & $324$ \\ \midrule 69 & ${\mathbb K}[T_1,\ldots,T_7] / \left\langle \begin{smallmatrix} T_1T_2^2 + T_3T_4^2 + T_5T_6^2, \\ \alpha T_3T_4^2 + T_5T_6^2 + T_7^3 \end{smallmatrix} \right\rangle $ & $(1,1,1,1,1,1,1)$ & $9$ \\ \bottomrule \end{longtable} \end{center} where in the last two rows of the table the parameter $\alpha$ can be any element from ${\mathbb K}^* \setminus \{1\}$. \end{theorem} By the result of \cite{Pu2}, the singular quintics of this list are rational degenerations of smooth non-rational Fano fourfolds. \section{Geometry of the locally factorial threefolds} \label{sec:geom3folds} In this section, we take a closer look at the (factorial) singularities of the Fano varieties~$X$ listed in Theorem~\ref{thm:3fano}. Recall that the discrepancies of a resolution $\varphi \colon \t{X} \to X$ of a singularity are the coefficients of $K_{\t{X}} - \varphi^* K_X$, where $K_X$ and $K_{\t{X}}$ are canonical divisors such that $K_{\t{X}} - \varphi^* K_X$ is supported on the exceptional locus of $\varphi$. A resolution is called crepant, if its discrepancies vanish and a singularity is called canonical (terminal), if it admits a resolution with nonnegative (positive) discrepancies. By a relative minimal model we mean a projective morphism $\t{X} \to X$ such that $\t{X}$ has at most terminal singularities and its relative canonical divisor is relatively nef. \begin{theorem} \label{prop:3foldsing} For the nine 3-dimensional Fano varieties listed in Theorem~\ref{thm:3fano}, we have the following statements. \begin{enumerate} \item No.~4 is a smooth quadric in ${\mathbb P}^4$. \item Nos.~1,3,5,7 and 9 are singular with only canonical singularities and all admit a crepant resolution. \item Nos.~6 and 8 are singular with non-canonical singularities but admit a smooth relative minimal model. \item No.~2 is singular with only canonical singularities, one of them of type $\mathbf{cA_1}$, and admits only a singular relative minimal model. \end{enumerate} The Cox ring of the relative minimal model $\widetilde{X}$ as well as the the Fano degree of $X$ itself are given in the following table. \begin{center} \begin{longtable}[htbp]{llc} \toprule No. & \hspace{3.5cm}$\mathcal{R}(\widetilde{X})$ & $(-K_X)^3$ \\ \midrule 1 & ${\mathbb K}[T_1,\ldots,T_{14}]/(T_1T_2T_3^2T_4^3T_5^4T_6^5+T_7^3T_8^2T_9+T_{10}^2T_{11}\rangle$ & $8$ \\ \cmidrule{1-3} 2 & ${\mathbb K}[T_1,\ldots,T_9]/\langle T_1T_2T_3^2T_4^4+T_5T_6^2T_7^3+T_8^2 \rangle$ & $8$ \\ \cmidrule{1-3} 3 & ${\mathbb K}[T_1,\ldots,T_8]/ \langle T_1T_2^2T_3^3+T_4T_5^3+T_6T_7^2\rangle$ & $8$ \\ \cmidrule{1-3} 4 & ${\mathbb K}[T_1, \ldots, T_5]/\bangle{T_1T_2 + T_3T_4 + T_5^2}$ & $54$ \\ \cmidrule{1-3} 5 & ${\mathbb K}[T_1,\ldots,T_6]/\langle T_1T_2^2+T_3T_4^2+T_5^3T_6\rangle$ & $24$ \\ \cmidrule{1-3} 6 & ${\mathbb K}[T_1,\ldots,T_6]/ \langle T_1T_2^3+T_3T_4^3+T_5^4T_6 \rangle$ & $4$ \\ \cmidrule{1-3} 7 & ${\mathbb K}[T_1, \ldots, T_7]/\bangle{T_1T_2^3 + T_3T_4^3 + T_5^2T_6}$ & $16$ \\ \cmidrule{1-3} 8 & ${\mathbb K}[T_1, \ldots, T_7]/\bangle{T_1T_2^5 + T_3T_4^5 + T_5^2T_6}$ & $2$ \\ \cmidrule{1-3} 9 & $\displaystyle {\mathbb K}[T_1,\ldots,T_{46}]/ \left\langle \begin{smallmatrix} T_1T_2T_3T_4^2T_5^2T_6^3T_7^3T_8^4T_9^4T_{10}^5 \;+\; \\ \;+\; T_{11} \cdots T_{18} T_{19}^2\cdots T_{24}^2 T_{25}^3 T_{26}^3 \;+\; T_{27}\cdots T_{32} T_{33}^2 \end{smallmatrix} \right\rangle$ & $2$\\ \bottomrule \end{longtable} \end{center} \end{theorem} For the proof, it is convenient to work in the language of polyhedral divisors introduced in~\cite{MR2207875} and~\cite{divfans}. As we are interested in rational varieties with a complexity one torus action, we only have to consider polyhedral divisors on the projective line $Y = {\mathbb P}^1$. This considerably simplifies the general definitions and allows us to give a short summary. In the sequel, $N \cong {\mathbb Z}^n$ denotes a lattice and $M = {\rm Hom}(N,{\mathbb Z})$ its dual. For the associated rational vector spaces we write $N_{\mathbb Q}$ and $M_{\mathbb Q}$. A {\em polyhedral divisor\/} on the projective line $Y := {\mathbb P}^1$ is a formal sum \begin{eqnarray*} \mathcal{D} & = & \sum_{y \in Y} \mathcal{D}_y \cdot y, \end{eqnarray*} where the coefficients $\mathcal{D}_y \subseteq N_{{\mathbb Q}}$ are (possibly empty) convex polyhedra all sharing the same tail (i.e.~recession) cone $\mathcal{D}_Y = \sigma \subseteq N_{\mathbb Q}$, and only finitely many $\mathcal{D}_y$ differ from $\sigma$. The {\em locus\/} of $\mathcal{D}$ is the open subset $Y(\mathcal{D}) \subseteq Y$ obtained by removing all points $y \subseteq Y$ with $\mathcal{D}_y = \emptyset$. For every $u \in \sigma^{\vee} \cap M$ we have the {\em evaluation\/} \begin{eqnarray*} \mathcal{D}(u) & := & \sum_{y \in Y} \min_{v \in \mathcal{D}_y} \bangle{u ,v} \! \cdot \! y, \end{eqnarray*} which is a usual rational divisor on $Y(\mathcal{D})$. We call the polyhedral divisor $\mathcal{D}$ on $Y$ {\em proper\/} if $\deg \, \mathcal{D} \subsetneq \sigma$ holds, where the {\em polyhedral degree\/} is defined by \begin{eqnarray*} \deg \, \mathcal{D} & := & \sum_{y \in Y} \mathcal{D}_y. \end{eqnarray*} Every proper polyhedral divisor $\mathcal{D}$ on $Y$ defines a normal affine variety $X(\mathcal{D})$ of dimension ${\rm rk}\,(N)+1$ coming with an effective action of the torus $T = {\rm Spec} \, {\mathbb K}[M]$: set $X(\mathcal{D}) := {\rm Spec} \, A(\mathcal{D})$, where $$ A(\mathcal{D}) \ := \ \bigoplus_{u \in \sigma^\vee \cap M} \Gamma(Y(\mathcal{D}),\mathcal{O}(\mathcal{D}(u))) \ \subseteq \ \bigoplus_{u \in M} {\mathbb K}(Y) \cdot \chi^u. $$ A {\em divisorial fan\/}, is a finite set $\Xi$ of polyhedral divisors $\mathcal{D}$ on $Y$, all having their polyhedral coefficients $\mathcal{D}_y$ in the same $N_{{\mathbb Q}}$ and fulfilling certain compatibility conditions, see~\cite{divfans}. In particular, for every point $y \in Y$, the {\em slice\/} \begin{eqnarray*} \Xi_y & := & \left\{\mathcal{D}_y; \; \mathcal{D} \in \Xi \right\} \end{eqnarray*} must be a polyhedral subdivision. The {\em tail fan\/} is the set $\Xi_Y$ of the tail cones $\mathcal{D}_Y$ of the $\mathcal{D} \in \Xi$; it is a fan in the usual sense. Given a divisorial fan $\Xi$, the affine varieties $X(\mathcal{D})$, where $\mathcal{D} \in \Xi$, glue equivariantly together to a normal variety $X(\Xi)$, and we obtain every rational normal variety with a complexity one torus action this way. Smoothness of $X = X(\Xi)$ is checked locally. For a proper polyhedral divisor $\mathcal{D}$ on $Y$, we infer the following from~\cite[Theorem~3.3]{tfano}. If $Y(\mathcal{D})$ is affine, then $X(\mathcal{D})$ is smooth if and only if ${\rm cone}(\{1\} \times \mathcal{D}_y) \subseteq {\mathbb Q} \times N_{{\mathbb Q}}$, the convex, polyhedral cone generated by $\{1\} \times \mathcal{D}_y$, is regular for every $y \in Y(\mathcal{D})$. If $Y(\mathcal{D}) = Y$ holds, then $X(\mathcal{D})$ is smooth if and only if there are $y,z \in Y$ such that $\mathcal{D} = \mathcal{D}_y y + \mathcal{D}_z z$ holds and ${\rm cone}(\{1\} \times \mathcal{D}_y) + {\rm cone}(\{-1\} \times \mathcal{D}_z)$ is a regular cone in ${\mathbb Q} \times N_{{\mathbb Q}}$. Similarly to toric geometry, singularities of $X(\mathcal{D})$ are resolved by means of subdividing~$\mathcal{D}$. This means to consider divisorial fans $\Xi$ such that for any $y \in Y$, the slice $\Xi_y$ is a subdivision of $\mathcal{D}_y$. Such a $\Xi$ defines a dominant morphism $X(\Xi) \rightarrow X(\mathcal{D})$ and a slight generalization of~\cite[Thm.~7.5.]{divfans} yields that this morphism is proper. \goodbreak \begin{proposition} \label{sec:prop-divfans} The 3-dimensional Fano varieties No. 1-8 listed in Theorem~\ref{thm:3fano} and their relative minimal models arise from divisorial fans having the following slices and tail cones. \myrule{1}{ \threefoldG } \myrule{2}{ \threefoldE } \myrule{3}{ \threefoldF } \myrule{4}{ \threefoldA } \myrule{5}{ \threefoldB } \myrule{6}{ \threefoldC } \myrule{7}{ \threefoldH } \myrule{8}{ \threefoldD } \noindent \end{proposition} The above table should be interpreted as follows. The first three pictures in each row are the slices at $0$, $1$ and $\infty$ and the last one is the tail fan. The divisorial fan of the fano variety itself is given by the solid polyhedra in the pictures. Here, all polyhedra of the same gray scale belong to the same polyhedral divisor. The subdivisions for the relative minimal models are sketched with dashed lines. In general, polyhedra with the same tail cone belong all to a unique polyhedral divisor with complete locus. For the white cones inside the tail fan we have another rule: for every polyhedron $\Delta \in \Xi_y$ with the given white cone as its tail there is a polyhedral divisor $\Delta \cdot y + \emptyset \cdot z \in \Xi$, with $z \in \{0,1,\infty\} \setminus \{y\}$. Here, different choices of $z$ lead to isomorphic varieties, only the affine covering given by the $X(\mathcal{D})$ changes. In order to prove Theorem~\ref{prop:3foldsing}, we also have to understand invariant divisors on $X = X(\Xi)$ in terms of $\Xi$, see~\cite[Prop.~4.11 and~4.12]{HaSu} for details. A first type of invariant prime divisors, is in bijection $D_{y,v} \leftrightarrow (y,v)$ with the vertices $(y,v)$, where $y \in Y$ and $v \in \Xi_y$ is of dimension zero. The order of the generic isotropy group along $D_{y,v}$ equals the minimal positive integer $\mu(v)$ with $\mu(v) v \in N$. A second type of invariant prime divisors, is in $D_{\varrho} \leftrightarrow \varrho$ with the extremal rays $\varrho \in \Xi_Y$, where a ray $\varrho \in \Xi_Y$ is called extremal if there is a $\mathcal{D} \in \Xi$ such that $\varrho \subseteq \mathcal{D}_Y$ and $\deg \, \mathcal{D} \cap \varrho = \emptyset$ holds. The set of extremal rays is denoted by $\Xi_Y^\times$. The divisor of a semi-invariant function $f \cdot \chi^u \in {\mathbb K}(X)$ is then given by \begin{eqnarray*} {\rm div}(f \cdot \chi^u) & = & - \sum_{y \in Y} \sum_{v \in \Xi_y^{(0)}} \mu(v) \cdot (\langle v, u \rangle + {\rm ord}_y f) \cdot D_{y,v} \ - \ \sum_{\varrho \in \Xi_Y^\times} \langle n_\varrho, u \rangle \cdot D_\varrho. \end{eqnarray*} Next we describe the canonical divisor. Choose a point $y_0 \in Y$ such that $\Xi_{y_0} = \Xi_Y$ holds. Then a canonical divisor on $X = X(\Xi)$ is given by \begin{eqnarray*} K_X & = & (s - 2) \cdot y_0 \ - \ \sum_{\Xi_y \ne \Xi_Y} \sum_{v \in \Xi_i^{(0)}} D_{y,v} \ - \ \sum_{\varrho \in \Xi_Y^\times} E_\varrho. \end{eqnarray*} \begin{proposition} \label{prop:discrepancies} Let $\mathcal{D}$ be a proper polyhedral divisor with $Y(\mathcal{D}) = {\mathbb P}_1$, let $\Xi$ be a refinement of $\mathcal{D}$ and denote by $y_1, \ldots, y_s \in Y$ the points with $\Xi_{y_i} \ne \Xi_Y$. Then the associated morphism $\varphi \colon X(\Xi) \to X(\mathcal{D})$ satisfies the following. \begin{enumerate} \item The prime divisors in the exceptional locus of $\varphi$ are the divisors $D_{y_i,v}$ and $D_{\varrho}$ corresponding to $v \in \Xi_{y_i}^{(0)} \setminus \mathcal{D}_{y_i}^{(0)}$ and $\varrho \in \Xi_Y^\times \setminus \mathcal{D}^\times$ respectively. \item Then the discrepancies along the prime divisors $D_{y_i,v}$ and $D_{\varrho}$ of~(i) are computed as \[ d_{y_i,v} \ = \ -\mu(v)\cdot (\langle v, u' \rangle + \alpha_y) - 1, \qquad\qquad d_{\varrho} \ = \ -\langle v_\varrho , u' \rangle - 1, \] where the numbers $\alpha_i$ are determined by \begin{eqnarray*} \begin{pmatrix} -1 & -1 & \ldots & -1& 0 \\ \hline \mu(v_{1}^1) & 0 & \ldots & 0 & \mu(v_{1}^1) v_{1}^1 \\ \vdots & \vdots & & \vdots &\vdots \\ \mu(v_{1}^{r_1})& 0 & \ldots & 0 & \mu(v_{1}^{r_1}) v_{1}^{r_1} \\ & & \ddots & & \\ 0 & 0 & \ldots & \mu(v_{s}^1) & \mu(v_{s}^1) v_{s}^{1} \\ \vdots & \vdots & & \vdots &\vdots \\ 0 & 0 & \ldots & \mu(v_{s}^{r_s}) & \mu(v_{s}^{r_s}) v_{s}^{r_s} \\ \hline 0 & 0 & \ldots & 0 & n_{\varrho_1} \\ \vdots & \vdots & & \vdots &\vdots \\ 0 & 0 & \ldots & 0 & n_{\varrho_{r}} \end{pmatrix} \ \cdot \ \begin{pmatrix} \alpha_{y_1}\\ \vdots\\ \alpha_{y_s}\\ u \end{pmatrix} & = & \begin{pmatrix} 2-s\\ 1 \\ \vdots \\ 1\\ \hline 1\\ \vdots\\ 1\\ \end{pmatrix} \end{eqnarray*} \end{enumerate} \end{proposition} \begin{proof} The first claim is obvious by the characterization of invariant prime divisors. For the second claim note that by~\cite[Theorem~3.1]{tidiv} every Cartier divisor on $X(\mathcal{D})$ is principal. Hence, we may assume $$ \ell\cdot K_X \ = \ {\rm div}(f \cdot \chi^{u}), \qquad\qquad {\rm div}(f) \ = \ \sum_y \alpha_y \cdot y. $$ Then our formul{\ae} for ${\rm div}(f \cdot \chi^{u})$ and $K_X$ provide a row for every vertex $v_{i}^j \in \Xi_{y_i}$, $i=0,\ldots,s$, and for every extremal ray $\varrho_i \in \Xi^\times$, and ${\ell}^{-1}(\alpha,u)$ is the (unique) solution of the above system. \end{proof} Note, that in the above Proposition, the variety $X(\mathcal{D})$ is ${\mathbb Q}$-Gorenstein if and only if the linear system of equations has a solution. \begin{proof}[Proof of Theorem~\ref{prop:3foldsing} and Proposition~\ref{sec:prop-divfans}] We exemplarily discuss variety number eight. Recall that its Cox ring is given as \begin{eqnarray*} \mathcal{R}(X) & = & \mathbb{K}[T_1,\ldots,T_5]/(T_1T_2^5+T_3T_4^5+T_5^2) \end{eqnarray*} with the degrees $1,1,1,1,3$. In particular, $X$ is a hypersurface of degree $6$ in ${\mathbb P}(1,1,1,1,3)$, and the self-intersection of the anti-canonical divisor can be calculated as $$ (-K_X^3) \ = \ 6 \cdot \frac{(1+1+1+1+3-6)^3}{1\cdot 1\cdot 1\cdot 1\cdot 3} \ = \ 2. $$ The embedding $X \subseteq {\mathbb P}(1,1,1,1,3)$ is equivariant, and thus we can use the technique described in~\cite[Sec.~11]{MR2207875} to calculate a divisorial fan $\Xi$ for $X$. The result is the following divisorial fan; we draw its slices and indicate the polyhedral divisors with affine locus by colouring their tail cones $\mathcal{D}_Y \in \Xi_Y$ white: \threefoldDplain \noindent One may also use~\cite[Cor.~4.9.]{HaSu} to verify that $\Xi$ is the right divisorial fan: it computes the Cox ring in terms of $\Xi$, and, indeed, we obtain again $\mathcal{R}(X)$. Now we subdivide and obtain a divisorial fan having the refined slices as indicated in the following picture. \threefoldD \noindent Here, the white ray ${\mathbb Q}_{\geq 0}\cdot (1,0)$ indicates that the polyhedral divisors with that tail have affine loci. According to~\cite[Cor.~4.9.]{HaSu}, the corresponding Cox ring is given by \begin{eqnarray*} \mathcal{R}(\widetilde{X}) & = & {\mathbb K}[T_1, \ldots, T_7]/\bangle{T_1T_2^5 + T_3T_4^5 + T_5^2T_6}. \end{eqnarray*} We have to check that $\widetilde{X}$ is smooth. Let us do this explicitly for the affine chart defined by the polyhedral divisor $\mathcal{D}$ with tail cone $\mathcal{D}_Y = {\rm cone}((1,2),(3,1))$. Then $\mathcal{D}$ is given by \begin{eqnarray*} \mathcal{D} & = & \left(\left(\frac{3}{5},\frac{1}{5}\right) + \sigma\right) \cdot \{0\} \ + \ \left(\left[-\frac{1}{2},0\right] \times 0 + \sigma\right)\cdot \{\infty\}. \end{eqnarray*} Thus, ${\rm cone}(\{1\} \times \mathcal{D}_0) + {\rm cone}(\{-1\} \times \mathcal{D}_\infty)$ is generated by $(5,3,1)$, $(-2,-1,0)$ and $(-1,0,0)$; in particular, it is a regular cone. This implies smoothness of the affine chart $X(\mathcal{D})$. Furthermore, we look at the affine charts defined by the polyhedral divisors $\mathcal{D}$ with tail cone $\mathcal{D}_Y = {\rm cone}(1,0)$. Since they have affine locus, we have to check ${\rm cone}(\{1\} \times \mathcal{D}_y)$, where $y \in Y$. For $y \neq 0, 1$, we have $\mathcal{D}_y = \mathcal{D}_Y$. In this case, ${\rm cone}(\{1\} \times \mathcal{D}_y)$ is generated by $(1,1,0)$, $(0,1,0)$ and thus is regular. For $y=0$, we obtain that ${\rm cone}(\{1\} \times \mathcal{D}_y)$ is generated by $(5,3,1)$, $(1,0,0)$, $(0,1,0)$ and this is regular. For $y=1$ we get the same result. Hence, the polyhedral divisors with tail cone $\mathcal{D}_y = {\rm cone}(1,0)$ give rise to smooth affine charts. Now we compute the discrepancies according to Proposition~\ref{prop:discrepancies}. The resolution has two exceptional divisors $D_{\infty, \mathbf{0}}$ and $E_{(1,0)}$. We work in the chart defined by the divisor $\mathcal{D} \in \Xi$ with tail cone $\mathcal{D}_Y = {\rm cone}((1,2),(1,0))$. The resulting system of linear equations and its unique solution are given by \[ \left(\begin{array}{ccccc|c} -1 & -1 & -1 & 0 & 0 & -1\\ 5 & 0 & 0 & 3 & 1 & 1\\ 0 & 1 & 0 & 0 & 0 & 1\\ 0 & 5 & 0 & 0 & -1 & 1\\ 0 & 0 & 2 & -1 & 0 & 1 \end{array}\right), \qquad\qquad \begin{pmatrix} \alpha_0\\ \alpha_1\\ \alpha_\infty\\ \hline u \end{pmatrix} \ = \ \begin{pmatrix} 0\\ 1\\ 0 \\ \hline -1\\ 4 \end{pmatrix}. \] The formula for the discrepancies yields $d_{\infty,\mathbf{0}}= -1$ and $d_{(1,0)}= -2$. In particular, $X$ has non-canonical singularities. By a criterion from~\cite[Sec.~3.4.]{tidiv}, we know that $D_{\infty, \mathbf{0}} + 2 \cdot E_{(1,0)}$ is a nef divisor. It follows that $\t{X}$ is a minimal model over $X$. \end{proof}

https://arxiv.org/abs/2103.06207

On the mean-field equations for ferromagnetic spin systems

We derive mean-field equations for a general class of ferromagnetic spin systems with an explicit error bound in finite volumes. The proof is based on a link between the mean-field equation and the free convolution formalism of random matrix theory, which we exploit in terms of a dynamical method. We present three sample applications of our results to Kać interactions, randomly diluted models, and models with an asymptotically vanishing external field.

\section{Introduction} The subject of this note is a ferromagnetic spin system with Hamiltonian $H\vcentcolon \{-1, 1\}^N \to {\mathbb{R}}$ defined by \beq{eq:firsthdef} H(\sigma) = -\frac{1}{2}\sum_{ij} J_{ij} \sigma_i \sigma_j - \sum_i h_i \sigma_i, \qquad J_{ij} \geq 0 \end{equation} and Gibbs expectation of $f\vcentcolon \{-1, 1\}^N \to {\mathbb{R}}$ given by \[\mean{f} = \frac{1}{Z} \sum_\sigma f(\sigma) e^{-H(\sigma)}, \qquad Z = \sum_\sigma e^{-H(\sigma)}. \] There is no loss of generality in assuming that $J_{ij} = J_{ji}$, which we will do from now on. The results below will be meaningful in the setting where $\max J_{ij} \to 0$ as $N \to \infty$. The prototypical example of such a system is the Curie-Weiss model, which corresponds to the choice $J_{ij} = \beta N^{-1}$ with $\beta \geq 0$ and $h_i = h \in {\mathbb{R}}$ (see~\cite{MR2189669} for a comprehensive discussion and bibliography). The phase diagram of the Curie-Weiss model can be obtained by studying the observable $M = N^{-1} \sum_i \sigma_i$, whose Gibbs expectation satisfies the mean-field equation \beq{eq:cwmf} \mean{M} = \tanh(h + \beta \mean{M})\end{equation} in the thermodynamic limit $N \to \infty$. To understand~\eqref{eq:cwmf} heuristically, we note that \[\mean{\sigma_i} = \Braket{\tanh\left(h + \frac{\beta}{N} \sum_{j \neq i} \sigma_j \right) }_i\] where $\mean{\cdot}_i$ is the Gibbs expectation with the spin $\sigma_i$ removed. Since one expects $M$ to concentrate around its expectation both at high temperature (spins are approximately independent) and at low temperature when $h \neq 0$ (almost all spins take on the same value), it should be possible to pull the Gibbs expectation inside the $\tanh$. Although this intuition should extend also to the more general case~\eqref{eq:firsthdef}, rigorous treatments of~\eqref{eq:cwmf} rely strongly on the symmetries of the Curie-Weiss model. The classical approach (see~\cite[Ch.\,2]{MR3752129}) is to compute the entropy using that $H = -\frac{\beta}{2} N M^2 - h N M$ essentially depends on only one degree of freedom. Other proofs exploit this symmetry using the Hubbard-Stratonovich transformation~\cite{PhysRevLett.3.77,1957SPhD....2..416S}, Varadhan's lemma~\cite{MR203230}, or the exchangeability of certain spin configurations~\cite{MR2707160, MR2288072}. We propose a simple dynamical method for establishing explicit finite-volume versions of the mean-field equations that are valid for general interactions. Setting $m_i = \mean{\sigma_i}$, the analogue of~\eqref{eq:cwmf} in the general setting is \beq{eq:genmf} {\mathbf{m}} = \tanh({\mathbf{h}} + {\mathbf{J}} {\mathbf{m}}) \end{equation} where ${\mathbf{m}} = (m_i)$, ${\mathbf{h}} = (h_i)$, ${\mathbf{J}} = (J_{ij})$, and the $\tanh$ of a vector is defined in an entrywise sense. To state our main result, we also introduce the notation \[\hat{{\mathbf{h}}} = \min h_i, \qquad \|{\mathbf{J}}\|_{\infty, \infty} = \max_j \sum_i J_{ij}, \qquad \|{\mathbf{J}}\|_{1, \infty} = \max_{ij} J_{ij}.\] The typical mean-field setting corresponds to $\|{\mathbf{J}}\|_{\infty, \infty} = {\mathcal{O}}(1)$ and $\|{\mathbf{J}}\|_{1, \infty} = {\mathcal{O}}(N^{-1})$ as $N \to \infty$. \begin{theorem} \label{thm:main} Let ${\mathbf{h}} \geq 0$. Then \[ \|{\mathbf{m}} - \tanh({\mathbf{h}} + {\mathbf{J}} {\mathbf{m}})\|_\infty \le \frac{\|{\mathbf{J}}\|_{1, \infty}}{\hat{{\mathbf{h}}}} \left(3 + \|{\mathbf{J}}\|_{\infty, \infty} + \log\left(1 + \frac{\|{\mathbf{J}}\|_{\infty, \infty}}{\hat{{\mathbf{h}}}} \right) \right).\] \end{theorem} Theorem~\ref{thm:main} is only informative when $\hat{{\mathbf{h}}} > 0$. However, when $\|{\mathbf{J}}\|_{1,\infty} \to 0$ as $N \to \infty$, the theorem is sufficient to study the physically relevant order of limits that first lets $N \to \infty$ and then lets ${\mathbf{h}} \to 0$. Moreover, one can arrange for both $\hat{{\mathbf{h}}}$ and the right hand side in Theorem~\ref{thm:main} to also vanish asymptotically if $\|{\mathbf{J}}\|_{1, \infty}$ does -- a point we will return to in Section~\ref{sec:applications}. To the best of our knowledge, quantitative bounds like Theorem~\ref{thm:main} are relatively scarce in the standard literature on the Curie-Weiss model. Stein's method for exchangeable pairs~\cite{MR2707160, MR2288072} yields concentration bounds for $M$ under the Gibbs measure with fluctuations of order $N^{-1/2}$. In particular, these bounds prove Theorem~\ref{thm:main} for the Curie-Weiss model with an error rate of $N^{-1/2}$. When ${\mathbf{h}} > 0$, our result improves the error rate to $N^{-1}$ while remaining valid for models with general interactions ${\mathbf{J}}$. There have also been recent developments concerning the fluctuations of general nonlinear functions of Bernoulli random variables whose gradients are close to a low-dimensional manifold \cite{MR3519474}. The application of these ideas to mean-field Gibbs measures was further explored in the works~\cite{MR3663625, pmlr-v75-jain18b, fluct-preprint}. The general mean-field equations~\eqref{eq:genmf} are very common in the physics literature~\cite[Sec.\,3.2]{MR1007980}. Nevertheless, it was noted in~\cite[Sec.\,3.4]{MR2707160} that completely general interactions ${\mathbf{J}}$ seem to pose significant challenges for the existing mathematical strategies. It was also shown in~\cite[Thm.\,3.5]{MR2707160} that methods based on exchangeable pairs can yield analogues of the mean-field equations for certain conditional averages with high probability under the Gibbs measure, from which~\eqref{eq:genmf} follows at sufficiently high temperatures. Exchangeable pairs have also been used to analyze the rank-one case ${\mathbf{J}} = \mathbf{w}\mathbf{w}^\intercal$ with a regular and non-negative $\mathbf{w} \in {\mathbb{R}}^N$~\cite{MR4046512}. A different perspective can be found in the works~\cite{MR1989669,MR2219531}, which prove that the magnetizations of sufficiently high-dimensional or long-range systems approximately minimize the free energy of the associated mean-field theory. These are very strong results that, among many other things, imply the approximate validity of the mean-field equations but also rely on the powerful input of infrared bounds derived from reflection positivity. There is a strong analogy between the mean-field equation and the subordination relations in random matrix theory and free probability. The central example of the latter is concerned with the resolvent $G(t, z) = (A_t - z)^{-1}$ of an $N \times N$ matrix $A_t = A_0 + \sqrt{t} \Phi$, where $A_0$ is a diagonal matrix and $\Phi$ is drawn from the Gaussian orthogonal ensemble. The main assertion, that the limiting empirical eigenvalue distribution of $A_t$ is given by the free convolution of the limiting empirical eigenvalue distribution of $A_0$ and a semicircular element, can be captured by the fact that \[s_0(z) = \frac{1}{N} \sum_{\lambda \in \sigma(A_0)} \frac{1}{\lambda-z}, \qquad s_t(z) = \frac{1}{N} \sum_{\lambda \in \sigma(A_t)} \frac{1}{\lambda - z}\] are analytic functions that map the half-plane ${\operatorname{Im}\,} z > 0$ into itself and satisfy \beq{eq:subordination} s_t(z) = s_0(z + t s_t(z))\end{equation} in the $N \to \infty$ limit~\cite{MR0475502}. To illustrate the connection to the mean-field equation, we consider the simplest case of~\eqref{eq:genmf} when $h_i = h > 0$ and $\sum_j J_{ij} = \beta$ for all $i$. In this case, one can construct a solution of~\eqref{eq:genmf} by letting each entry solve the scalar equation \beq{eq:tracemf} m = \tanh(h + \beta m). \end{equation} It is not hard to show that the positive solution $m = m(h)$ of~\eqref{eq:tracemf} extends to an analytic function of $h$ that maps the half-plane ${\operatorname{Re }\,} h > 0$ into itself. It follows that $\tilde{m}(z) = i m(-iz)$ is an analytic function of $\{{\operatorname{Im}\,} z > 0\}$ into itself with \[\tilde{m}(z) = \tan (z + \beta \tilde{m}(z)).\] It is a consequence of the Herglotz trick~\cite{MR3823190} that \[\tan(z) = \sum_{\lambda \in \Lambda} \frac{1}{\lambda - z}, \qquad \Lambda = \pi ({\mathbb{Z}} + 1/2),\] so $\tilde{m}$ is the Stieltjes transform of the free convolution of $\sum_{\lambda \in \Lambda} \delta_\lambda$ and a semicircular element. Expanding on this theme, the assertion of Theorem~\ref{thm:main} that the relation~\eqref{eq:tracemf} is not only valid asymptotically but that the error is small even in finite volumes when $\hat{{\mathbf{h}}} \gg \|{\mathbf{J}}\|_{1, \infty}$, is analogous to the local deformed semicircle law~\cite{MR3134604}. The proof of the local law depends crucially on the fact that \beq{eq:wardid} \left| \parder{}{z} G_{ii}(t, z) \right| \le \sum_{k} |G_{ik}(t, z)|^2 = \frac{{\operatorname{Im}\,} G_{ii}(t, z)}{{\operatorname{Im}\,} z}.\end{equation} Lemma~\ref{thm:corrbound}, which rests on the Lee-Yang theorem~\cite{PhysRev.87.410}, contains a similar inequality for $m_i$ that lies at the heart of Theorem~\ref{thm:main}. It was observed in~\cite{MR0475502} that the subordination relation~\eqref{eq:subordination} is equivalent to the partial differential equation \beq{eq:pde} \parder{}{t} s_t(z) = s_t(z) \parder{}{z} s_t(z).\end{equation} In finite volumes, $s_t$ exactly satisfies a perturbed version of this transport equation, which enables a simple proof the local law by combining the analytic structure contained in~\eqref{eq:wardid} with an approximate characteristic curve~\cite{MR3920502,MR4049087}. Our approach to Theorem~\ref{thm:main} is to generalize this analysis by exploiting the identity \beq{eq:corederid} \parder{}{J_{ij}} \mean{f} = m_i \parder{}{h_j} \mean{f} + m_j \parder{}{h_i} \mean{f} + \parder{^2}{h_i \partial h_j} \mean{f}. \end{equation} The relationship between~\eqref{eq:corederid} and~\eqref{eq:pde} is most apparent in the Curie-Weiss model, where~\eqref{eq:corederid} yields \[\parder{}{\beta} \mean{M} = \mean{M} \parder{}{h} \mean{M} + \frac{1}{2N} \parder{^2}{h^2} \mean{M},\] which is the evolution studied in~\cite{MR3920502} without the stochastic terms. We note that similar differential \textit{inequalities} have been studied in a variety of models that go beyond the mean-field setting~\cite{MR894398,MR874906}. This paper is structured as follows. In Section~\ref{sec:proof}, we use~\eqref{eq:corederid} to derive a transport equation in the general setting and prove the analogue of~\eqref{eq:wardid}. This allows us to extend the ideas of~\cite{MR3920502,MR4049087} concerning approximate characteristic curves to the present setting and prove Theorem~\ref{thm:main}. Then, in Section~\ref{sec:applications}, we present three sample applications of our method to Ka\'{c} interactions, randomly diluted models, and models with an asymptotically vanishing external field. \section{Proof of Theorem~\ref{thm:main}}\label{sec:proof} Rearranging indices shows that Theorem~\ref{thm:main} follows if we can prove that \beq{eq:toprove}|m_1 - \tanh(h_1 + {\mathbf{J}}_1^\intercal {\mathbf{m}})| \le \frac{\|{\mathbf{J}}_1\|_\infty}{\hat{{\mathbf{h}}}} \left(3 + \|{\mathbf{J}}_1\|_1 + \log\left(1 + \frac{\|{\mathbf{J}}_1\|_1}{\hat{{\mathbf{h}}}} \right) \right) \end{equation} where ${\mathbf{J}}_1 = (J_{1i})$ is the first column of ${\mathbf{J}}$. By symmetry, we can write $H(\sigma) = H(1, \sigma)$, where \[H(t, \sigma) = t \sum_i J_{1i} \sigma_1 \sigma_i + \sum_i h_i \sigma_i + H_1(\sigma_2, \dots, \sigma_N)\] and $H_1$ does not depend on $\sigma_1$. Moreover, if ${\mathbf{m}}(t, {\mathbf{h}})$ denotes the vector of magnetizations under $H(t, \cdot)$, the identity~\eqref{eq:corederid} yields \beq{eq:realpde} \parder{}{t} {\mathbf{m}}(t, {\mathbf{h}}) = m_1(t, {\mathbf{h}}) \parder{}{{\mathbf{J}}_1} {\mathbf{m}}(t, {\mathbf{h}}) + \parder{^2}{h_1 \partial {\mathbf{J}}_1} {\mathbf{m}}(t, {\mathbf{h}}) + ({\mathbf{J}}_1^\intercal {\mathbf{m}}) \parder{}{h_1} {\mathbf{m}}(t, {\mathbf{h}}),\end{equation} where $\parder{}{{\mathbf{J}}_1} = {\mathbf{J}}_1^\intercal \nabla_{{\mathbf{h}}}$ is the directional derivative with respect to ${\mathbf{h}}$ in direction ${\mathbf{J}}_1$. If we knew that the first two terms on the right hand side of~\eqref{eq:realpde} were negligible, then~\eqref{eq:realpde} would reduce to a transport equation that could be solved by varying $h_1$ along a suitable characteristic curve. We will prove the appropriate bounds for this purpose with the help of the following integral representation. Suppose that $f$ is a holomorphic function defined on a half-plane ${\operatorname{Re }\,} z > -\kappa$ such that ${\operatorname{Re }\,} f \geq 0$ and such that $f$ remains bounded as $z \to \infty$ along ${\mathbb{R}}$. Then, \[f(z) = a + \int_{{\mathbb{R}}} \! \frac{1}{z + \kappa - i\lambda } \, \mu(d\lambda)\] for some positive measure $\mu$ and $a \in {\mathbb{C}}$~\cite[Ch.\,5]{MR1307384}, which implies that \[\left|f^\prime(z) \right| \le \frac{{\operatorname{Re }\,} f(z)}{{\operatorname{Re }\,} z + \kappa}.\] We note that the use of integral representations of holomorphic functions to bound correlations is a classical idea (see for instance~\cite{MR2905800, MR432092}). \begin{lemma} \label{thm:corrbound} For every $k$ and every ${\mathbf{s}} \geq 0$, \[ \left| \parder{}{{\mathbf{s}}} m_k(t, {\mathbf{h}}) \right| \le \frac{\|{\mathbf{s}}\|_\infty}{\hat{{\mathbf{h}}}} \, m_k(t, {\mathbf{h}})\] and \[\left| \parder{^2}{h_1 \partial {\mathbf{s}}} m_k(t, {\mathbf{h}}) \right| \le \frac{\|{\mathbf{s}}\|_\infty}{\hat{{\mathbf{h}}}} \,\frac{m_k(t, {\mathbf{h}})}{h_1} \] for all $t \geq 0$ and ${\mathbf{h}} > 0$. \end{lemma} \begin{proof} If we fix $h_j$ with ${\operatorname{Re }\,} h_j > 0$ for $j \neq k$, the meromorphic function $f(h_k) = m_k(t, {\mathbf{h}})$ satisfies $f^\prime(h_k) = 1 - f^2(h_k)$ wherever it is analytic. Hence, combining the Picard-Lindel\"{o}f theorem with analytic continuation shows that there is some $\Gamma \in {\mathbb{C}}$ such that \[f(h_k) = \tanh(h_k + \Gamma)\] for all $h_k \in {\mathbb{C}}$. If ${\operatorname{Re }\,} {\mathbf{h}} > 0$, the Lee-Yang theorem asserts that the partition function $Z(t, {\mathbf{h}}) = \sum_{\sigma} e^{-H(t,\sigma)}$ cannot vanish and therefore $f$ cannot have a pole in this region. This is only possible when ${\operatorname{Re }\,} \Gamma \geq 0$ and therefore \[{\operatorname{Re }\,} m_k(t, {\mathbf{h}}) = {\operatorname{Re }\,} f(h_k) \geq 0\] whenever ${\operatorname{Re }\,} {\mathbf{h}} \geq 0$. To prove the first bound, we fix ${\mathbf{h}} \geq 0$ and consider the function \[f(z) = m_k(t, {\mathbf{h}} + z{\mathbf{s}}).\] Then ${\operatorname{Re }\,} f > 0$ on the half-plane ${\operatorname{Re }\,} z > -\kappa$ with $\kappa = \hat{{\mathbf{h}}} / \|{\mathbf{s}}\|_\infty$ so \[\left| \parder{}{{\mathbf{s}}} m_k(t, {\mathbf{h}}) \right| = \left| f^\prime(0) \right| \le \frac{{\operatorname{Re }\,} f(0)}{\kappa} = \frac{{\operatorname{Re }\,} m_k(t, {\mathbf{h}})}{\kappa}.\] Since $m_k(t, {\mathbf{h}})$ is real when ${\mathbf{h}}$ is real, this is the first assertion of the lemma. For the second bound, we use the auxiliary function \[g(z) = \partial_1 m_k(t, {\mathbf{h}} + z{\mathbf{s}}).\] Then $g$ is also holomorphic on ${\operatorname{Re }\,} z > -\kappa$ and considering the right hand side as a function of $h_1$ shows that \[|g(z)| \le \frac{{\operatorname{Re }\,} m_k(t, {\mathbf{h}} + z{\mathbf{s}})}{{\operatorname{Re }\,} h_1}.\] Letting $C$ be a positively oriented circle about $0$ of radius $\kappa$, we have \[ \parder{^2}{h_1 \partial {\mathbf{s}}} m_k(t, {\mathbf{h}}) = g^\prime(0) = \oint_C \! \frac{g(\xi)}{\xi^2} \, d\xi\] so \[ \left|\parder{^2}{h_1 \partial {\mathbf{s}}} m_k(t, {\mathbf{h}}) \right| \le \frac{1}{2\pi \kappa} \int_0^{2\pi} \frac{{\operatorname{Re }\,} m_k({\mathbf{h}} + e^{i\theta} {\mathbf{s}})}{{\operatorname{Re }\,} h_1} \, d\theta = \frac{{\operatorname{Re }\,} m_k(t,{\mathbf{h}})}{\kappa \, {\operatorname{Re }\,} h_1} \] using the mean value property of the harmonic function $z \to {\operatorname{Re }\,} m_k(t, {\mathbf{h}} + z{\mathbf{s}})$. \end{proof} With Lemma~\ref{thm:corrbound} in place, we now fix ${\mathbf{h}} \geq 0$ and define an approximate characteristic curve ${\mathbf{w}} = (w_i)$ for~\eqref{eq:realpde} by \[\parder{}{t} w_i(t) = \begin{cases}-{\mathbf{J}}_1^\intercal {\mathbf{m}}(t, {\mathbf{w}}(t)) & i = 1\\ 0 & \mbox{else} \end{cases}, \qquad {\mathbf{w}}(1) = {\mathbf{h}}. \] Since ${\mathbf{m}}$ is non-negative and uniformly Lipschitz continuous in ${\mathbf{h}} \geq 0$, such a curve exists and satisfies $\hat{{\mathbf{w}}}(t) \geq \hat{{\mathbf{h}}}$ for all $t \in [0, 1]$. The following lemma shows that any weighted average ${\mathbf{s}}^\intercal {\mathbf{m}}$ does not significantly change along the curve ${\mathbf{w}}(t)$, provided that $\|{\mathbf{s}}\|_\infty$ is small. \begin{lemma}\label{thm:avgbound} Let ${\mathbf{s}} \geq 0$. Then \[\sup_{t \in [0,1]} \left| {\mathbf{s}}^\intercal {\mathbf{m}}({\mathbf{h}}) - {\mathbf{s}}^\intercal {\mathbf{m}}(t, {\mathbf{w}}(t)) \right| \le \frac{\|{\mathbf{s}}\|_\infty}{\hat{{\mathbf{h}}}} \left(\|{\mathbf{J}}_1\|_1 + \log\left(1 + \frac{\|{\mathbf{J}}_1\|_1}{\hat{{\mathbf{h}}}} \right) \right).\] \end{lemma} \begin{proof} Inserting the characteristic curve into~\eqref{eq:realpde} and multiplying by ${\mathbf{s}}$, we obtain \beq{eq:jfluct} {\mathbf{s}}^\intercal {\mathbf{m}}({\mathbf{h}}) - {\mathbf{s}}^\intercal {\mathbf{m}}(t, {\mathbf{w}}(t)) = \int_t^1 \! m_1(r, {\mathbf{w}}(r)) \parder{}{{\mathbf{J}}_1} {\mathbf{s}}^\intercal {\mathbf{m}}(r, {\mathbf{w}}(r)) + \parder{^2}{h_1 \partial {\mathbf{J}}_1} {\mathbf{s}}^\intercal {\mathbf{m}}(r, {\mathbf{w}}(r)) \, dr.\end{equation} Combining the identity \[\parder{}{{\mathbf{J}}_1} {\mathbf{s}}^\intercal {\mathbf{m}} = \parder{}{{\mathbf{s}}} {\mathbf{J}}_1^\intercal {\mathbf{m}}\] with Lemma~\ref{thm:corrbound} shows that the first integral on the right hand side of~\eqref{eq:jfluct} is bounded by \[\int_t^1 \! \left| m_1(r, {\mathbf{w}}(r)) \parder{}{{\mathbf{J}}_1} {\mathbf{s}}^\intercal {\mathbf{m}}(r, {\mathbf{w}}(r)) \right| \, dr \le \frac{\|{\mathbf{s}}\|_\infty \|{\mathbf{J}}_1\|_1}{\hat{{\mathbf{h}}}},\] whereas the second integral is bounded by \beq{eq:secondint}\int_t^1 \! \left| \parder{^2}{h_1 \partial {\mathbf{J}}_1} {\mathbf{s}}^\intercal {\mathbf{m}}(r, {\mathbf{w}}(r)) \right| \, dr \le \frac{\|{\mathbf{s}}\|_\infty}{\hat{{\mathbf{h}}}} \int_t^1 \! \frac{{\mathbf{J}}_1^\intercal {\mathbf{m}}(r, {\mathbf{w}}(r))}{w_1(r)} \, dr.\end{equation} The right hand side of~\eqref{eq:secondint} can be calculated explicitly since $\parder{}{t} w_1(t) = - {\mathbf{J}}_1^\intercal {\mathbf{m}}(t, {\mathbf{w}}(t))$, which yields a final bound of \[ \int_t^1 \! \left| \parder{^2}{h_1 \partial {\mathbf{J}}_1} {\mathbf{s}}^\intercal {\mathbf{m}}(r, {\mathbf{w}}(r)) \right| \, dr \le \frac{\|{\mathbf{s}}\|_\infty}{\hat{{\mathbf{h}}}} \log\left( \frac{w_1(t)}{h_1} \right)\le \frac{\|{\mathbf{s}}\|_\infty}{\hat{{\mathbf{h}}}} \log\left(1 + \frac{\|{\mathbf{J}}_1\|_1}{\hat{{\mathbf{h}}}} \right) .\] \end{proof} The evolution of $m_1$ along the characteristic curve is given by \[\parder{}{t} m_1(t, {\mathbf{w}}(t)) = m_1(t, {\mathbf{w}}(t)) \parder{}{{\mathbf{J}}_1} m_1(t, {\mathbf{w}}(t)) + \parder{^2}{h_1 \partial {\mathbf{J}}_1} m_1(t, {\mathbf{w}}(t)).\] By Lemma~\ref{thm:corrbound}, \[\left| m_1(t, {\mathbf{w}}(t)) \parder{}{{\mathbf{J}}_1} m_1(t, {\mathbf{w}}(t)) \right| \le \frac{\|{\mathbf{J}}_1\|_\infty}{\hat{{\mathbf{h}}}}\] and \[\left| \parder{^2}{h_1 \partial {\mathbf{J}}_1} m_1(t, {\mathbf{w}}(t)) \right| = 2 \left| m_1(t, {\mathbf{w}}(t)) \parder{}{{\mathbf{J}}_1} m_1(t, {\mathbf{w}}(t)) \right| \le \frac{2\|{\mathbf{J}}_1\|_\infty}{\hat{{\mathbf{h}}}},\] so \beq{eq:outerbound}\left| m_1({\mathbf{h}}) - \tanh(w_1(0)) \right| = \left| m_1(1, {\mathbf{w}}(1)) - m_1(0, {\mathbf{w}}(0)) \right| \le \frac{3\|{\mathbf{J}}_1\|_\infty}{\hat{{\mathbf{h}}}}. \end{equation} Since \[w_1(0) = h_1 + \int_0^1 \! {\mathbf{J}}_1^\intercal {\mathbf{m}}(t, {\mathbf{w}}(t)) \, dt,\] Lemma~\ref{thm:avgbound} with ${\mathbf{s}} = {\mathbf{J}}_1$ implies that \[\left| w_1(0) - h_1 - {\mathbf{J}}_1^\intercal {\mathbf{m}}({\mathbf{h}}) \right| \le \frac{\|{\mathbf{J}}_1\|_\infty}{\hat{{\mathbf{h}}}} \left( \|{\mathbf{J}}_1\|_1 + \log\left(1 + \frac{\|{\mathbf{J}}_1\|_1}{\hat{{\mathbf{h}}}} \right) \right).\] Inserting this into~\eqref{eq:outerbound} and using the Lipschitz continuity of $\tanh$ completes the proof of~\eqref{eq:toprove}. \section{Three applications}\label{sec:applications} Our first two applications consist of showing that two classes of models have the same thermodynamic behavior as the Curie-Weiss model. In the first, we consider a system in a box $\Lambda \subset {\mathbb{Z}}^d$ with Ka\'{c} interactions \[J_{ij} = \beta \lambda^d f(\lambda(i-j)), \qquad h_i = h > 0, \qquad i, j \in \Lambda\] where $f$ is a bounded Riemann integrable probability density on ${\mathbb{R}}^d$ and $\beta, \lambda > 0$. Ka\'{c} interactions are ``physical'' in the sense that they yield a convex free energy, but this free energy still converges to the convex envelope of the Curie-Weiss free energy as $\lambda \to 0$~\cite{MR187835, MR148416}. This fact can be used to provide a justification of Maxwell's equal-area rule for the van der Waals isotherm (see also \cite[Ch.\,4]{MR3752129} for further details). There is also an extensive literature on Ka\'{c} interactions with fixed $\lambda > 0$ and related models -- we refer the reader to~\cite{MR1694123, MR1453742, MR1935654, MR1414119, MR2460018} and references therein. Writing ${\mathbf{m}}_\Lambda$ for the vector of magnetizations corresponding to the box $\Lambda$, Theorem~\ref{thm:main} asserts that \[\|{\mathbf{m}}_\Lambda - \tanh(h + {\mathbf{J}} {\mathbf{m}}_\Lambda)\|_{\infty} \le C \frac{\lambda^d}{h \log h} \] for some absolute constant $C < \infty$. Translation invariance and standard convexity arguments show that there is some $m \in {\mathbb{R}}$ such that for any fixed $i \in {\mathbb{Z}}^d$ we have $m_{\Lambda, i} \to m$ as $\Lambda \to {\mathbb{Z}}^d$. By the dominated convergence theorem the limit $m$ still satisfies \[\left| m - \tanh\left(h + \beta m \sum_{i} \lambda^d f(\lambda i) \right)\right| \le C \frac{\lambda^d}{h \log h}\] and therefore $m = \tanh(h + \beta m)$ in the $\lambda \to 0$ limit. The second model we consider is the randomly diluted model where \[J_{ij} = \frac{\beta}{Np} \epsilon_{ij}, \qquad h_i = h > 0\] and $\epsilon_{ij}$ are independent (up to symmetry) Bernoulli random variables with ${\mathbb{E} \,} \epsilon_{ij} = p$. In the case where $p = p(N)$ is chosen such that $Np \to \infty$ as $N \to \infty$, it was shown in~\cite{MR1239568} that the limiting magnetizations coincide with those of the Curie-Weiss model. Under this assumption, the variance of a weighted average is bounded by \beq{eq:bernoullivar} {\mathbb{E} \,} \left| {\mathbf{J}}_1^\intercal \mathbf{x} - \frac{\beta}{N} \sum_i x_i \right|^2 \le \frac{\beta^2 \|\mathbf{x}\|_\infty^2}{Np} \to 0.\end{equation} Applying the single-site bound~\eqref{eq:toprove}, it follows that \[m_1 - \tanh(h + {\mathbf{J}}_1^\intercal {\mathbf{m}}) \to 0\] in $L^2({\mathbb{P}})$ as $N \to \infty$. We write $m = N^{-1} \sum_i m_i$ and let $m_i^{(1)}$ denote the Gibbs mean of $\sigma_i$ with $\sigma_1$ removed. Using Lemma~\ref{thm:avgbound} and repeating the bound above yields \[{\mathbf{J}}_1^\intercal {\mathbf{m}}(h) - \sum_{i > 1} J_{1i} m_i^{(1)}(h) \to 0\] and \[m - \frac{1}{N} \sum_{i > 1} m_i^{(1)}(h) \to 0\] in $L^2({\mathbb{P}})$. Since $m_i^{(1)}$ is independent of ${\mathbf{J}}_1$, combining this with the bound~\eqref{eq:bernoullivar} shows that ${\mathbf{J}}_1^\intercal {\mathbf{m}} - \beta m \to 0$ in $L^2({\mathbb{P}})$. We conclude that both \[m_1 - \tanh(h + \beta m) \to 0, \qquad m - \tanh(h + \beta m) \to 0\] in $L^2({\mathbb{P}})$. Finally, we consider a generic model with an asymptotically vanishing external field \[\|{\mathbf{J}}\|_{1, \infty} = o(1), \qquad \hat{{\mathbf{h}}} = \|{\mathbf{J}}\|_{1, \infty}^{\frac{1}{2} - \delta}.\] We assume a low-temperature condition of the form that there exists $\alpha > 1$, independent of $N$, such that for all $\mathbf{x} \geq 0$ there is some index $i$ with \beq{eq:lowtemp}({\mathbf{J}} \mathbf{x})_i \geq \alpha \mathbf{x}_i.\end{equation} Provided that $\|{\mathbf{J}}\|_{\infty, \infty} = {\mathcal{O}}(1)$, Theorem~\ref{thm:main} implies that \[\tanh({\mathbf{h}} + {\mathbf{J}} {\mathbf{m}}) = {\mathbf{m}} + \boldsymbol \epsilon, \qquad \| \boldsymbol \epsilon\|_\infty = {\mathcal{O}}\left(-\|{\mathbf{J}}\|_{1, \infty}^{\frac{1}{2} + \delta} \log \|{\mathbf{J}}\|_{1, \infty} \right). \] There exists some $K > 0$ such that $\alpha \tanh(x) > x$ when $x \in (0, K)$. If it were true that ${\mathbf{m}} \le K$, combining this with the fact that ${\mathbf{h}}, \boldsymbol \epsilon \to 0$ and ${\mathbf{h}} \gg \boldsymbol \epsilon$ as $N \to \infty$ would imply that \[{\mathbf{J}} {\mathbf{m}} < \alpha {\mathbf{m}}\] for sufficiently large $N$, contradicting~\eqref{eq:lowtemp}. We conclude that an external field strength of $\hat{{\mathbf{h}}} = \|{\mathbf{J}}\|_{1, \infty}^{\frac{1}{2} - \delta}$ is sufficient to select a positive Gibbs state in the sense that \beq{eq:posgibbs} \liminf_{N \to \infty} \max_i m_i > 0.\end{equation} Degenerate examples like the case where $J_{1i} = 0$ for all $i$ demonstrate that, in general, one cannot hope for a stronger statement than~\eqref{eq:posgibbs}. For the Curie-Weiss model, the previous argument shows that an external field strength of $h = N^{-\frac{1}{2} + \delta}$ is sufficient to select the positive Gibbs state below the critical temperature of $\beta = 1$. The book~\cite[Sec.\,III.1]{MR1026102} mentions that actually an external field strength of $h = N^{-1 + \delta}$ already suffices, but we have not been able to locate a mathematical proof of this assertion in the literature. \bigskip \minisec{Acknowledgments} We thank M.\,Biskup, G.\,Genovese, and S.\,Warzel for their helpful comments. The work of P.\,S. is supported by the DFG grant SO 1724/1-1. \bibliographystyle{abbrv}

https://arxiv.org/abs/1807.04426

A likelihood-ratio type test for stochastic block models with bounded degrees

A fundamental problem in network data analysis is to test Erdös-Rényi model $\mathcal{G}\left(n,\frac{a+b}{2n}\right)$ versus a bisection stochastic block model $\mathcal{G}\left(n,\frac{a}{n},\frac{b}{n}\right)$, where $a,b>0$ are constants that represent the expected degrees of the graphs and $n$ denotes the number of nodes. This problem serves as the foundation of many other problems such as testing-based methods for determining the number of communities (\cite{BS16,L16}) and community detection (\cite{MS16}). Existing work has been focusing on growing-degree regime $a,b\to\infty$ (\cite{BS16,L16,MS16,BM17,B18,GL17a,GL17b}) while leaving the bounded-degree regime untreated. In this paper, we propose a likelihood-ratio (LR) type procedure based on regularization to test stochastic block models with bounded degrees. We derive the limit distributions as power Poisson laws under both null and alternative hypotheses, based on which the limit power of the test is carefully analyzed. We also examine a Monte-Carlo method that partly resolves the computational cost issue. The proposed procedures are examined by both simulated and real-world data. The proof depends on a contiguity theory developed by Janson \cite{J95}.

\section{Introduction} In recent years, stochastic block model (SBM) has attracted increasing attention in statistics and machine learning. It provides the researchers a ground to study many important problems that arise in network data such as community detection or clustering (\cite{ACBL13,AL18,NN14,SB15,BC09,ZLZ12}), goodness-of-fit of SBMs (\cite{BS16,L16,MS16,BM17,B18,GL17a,GL17b}) or various phase transition phenomena (\cite{MNS15,MNS17,AS17}). See \cite{A17} for a comprehensive review about recent development in this field. A key assumption in most of the literature is that the expected degree of every node tends to infinity along with the number of nodes $n$. For instance, in community detection (\cite{BC09,ZLZ12}), such a condition is needed for proving weak consistency of the detection methods; to prove strong consistency, the expected degree is further assumed to grow faster than $\log{n}$. For goodness-of-fit test, the growing-degree condition is needed to derive various asymptotic distributions for the test statistics (\cite{BS16,L16,BM17,B18,GL17a,GL17b}). Many real-world network data sets are highly sparse. For instance, the LinkedIn network, the real-world coauthorship networks, power transmission networks and web link networks all have small average degrees (see \cite{LLDM08, S01}). Therefore, it is reasonable to assume bounded degrees in such networks. There is a breakthrough recently made by \cite{MNS15,MNS17,AS17} about the possibility of successfully detecting the community structures when the expected degree of SBM is bounded. Specifically, the signal-to-noise ratio (SNR) of the multi-community SBM is used in these work as a phase transition parameter to indicate the possibility of successful detection. Motivated by such a groundbreaking result, it is natural to ask whether one can propose successful testing methods for SBMs with bounded degrees. Progress in this field may help researchers better understand the roles played by the expected degrees of SBMs in hypothesis testing, as well as provide a substantially broader scope of network models in which a successful test is possible. In this paper, we address this problem in the bisection SBM scenario. We propose a likelihood-ratio (LR) type test statistic to distinguish an Erd\"{o}s-R\'{e}nyi model versus a bisection SBM whose expected degrees are finite constants, and investigate its asymptotic properties. In what follows, we describe the models and our contributions more explicitly. \subsection{Models and Our Contributions.} Let us provide a brief review for Erd\"{o}s-R\'{e}nyi model and bisection SBM. Throughout the whole paper, assume that $a>b>0$ are \textit{fixed and known} constants unless otherwise indicated. For $n\in{\mathbb{N}}$, let $\mathcal{G}\left(n,\frac{a}{n},\frac{b}{n}\right)$ denote the bisection stochastic block model of random $\pm$-labeled graphs in which each vertex $u\in[n]:=\{1,2,\ldots,n\}$ is assigned, independently and uniformly at random, a label $\sigma_u\in\{\pm\}$, and then each possible edge $(u,v)$ is included with probability $a/n$ if $\sigma_u=\sigma_v$ and with probability $b/n$ if $\sigma_u\neq\sigma_v$. Let $A=[A_{uv}]_{u,v=1}^n\in\{0,1\}^{n\times n}$ denote the observed symmetric adjacency matrix in which $A_{uu}=0$ for all $1\le u\le n$, and for $1\le u<v\le n$, $A_{uv}=1$ indicates the inclusion of edge $(u,v)$ and $A_{uv}=0$ otherwise. Conditional on $\sigma=(\sigma_1,\ldots,\sigma_n)$, the variables $A_{uv}$, $1\le u<v\le n$, are assumed to be independent which follow \begin{eqnarray}\label{G:p:q:model} \textrm{$P(A_{uv}=1|\sigma)=p_{uv}(\sigma)$ and $P(A_{uv}=0|\sigma)=q_{uv}(\sigma)$,} \end{eqnarray} where \[ p_{uv}(\sigma)=\left\{\begin{array}{cc} \frac{a}{n},&\sigma_u=\sigma_v\\ \frac{b}{n},&\sigma_u\neq\sigma_v \end{array}\right., q_{uv}(\sigma)=1-p_{uv}(\sigma). \] The Erd\"{o}s-R\'{e}nyi model $\mathcal{G}\left(n,\frac{a+b}{2n}\right)$ has the same average degree as $\mathcal{G}\left(n,\frac{a}{n},\frac{b}{n}\right)$. It is interesting to decide which model an observed graph is generated from. Specifically, we are interested in the following hypothesis testing problem \begin{equation}\label{H0:H1} \textrm{$H_0$: $A\sim \mathcal{G}\left(n,\frac{a+b}{2n}\right)$\,\,\,\, vs.\,\,\,\, $H_1:$ $A\sim \mathcal{G}\left(n,\frac{a}{n},\frac{b}{n}\right)$.} \end{equation} To be more specific, we want to test whether the nodes on an observed random graph belong to the same community, or they belong to two equal-sized communities. Let $\kappa=\frac{(a-b)^2}{2(a+b)}$ denote the signal-to-noise ratio (SNR) associated with $\mathcal{G}\left(n,\frac{a}{n},\frac{b}{n}\right)$. It was conjectured by Decelle, Krzkala, Moore and Zdeborov\'{a} (\cite{DKMZ11}) that successful community detection is possible when $\kappa\ge1$, and impossible when $\kappa<1$. This conjecture was recently proved by Mossel, Neeman and Sly (\cite{MNS15}) through Janson's continuity theory (\cite{J95}). In the meantime, their result indicates that \textit{no test can be successful when $\kappa<1$} (see \cite{MNS15,MS16}), and so we primarily focus on the high SNR scenario $\kappa\ge1$. Classic likelihood-ratio (LR) tests for (\ref{H0:H1}) are not valid since the probability measures associated with $H_0$ and $H_1$ are asymptotically orthogonal as discovered by \cite{MNS15}. The result of \cite{MNS15} also implies that counting the cycles of length $\log^{1/4}{n}$ leads to an asymptotically valid test; see their Theorem 4. However, such test is unrealistic since $n$ should be at least $e^{81}$ to make the length at least 3. In Section \ref{sec:varepsilon:LR}, we propose a regularized LR-type test for (\ref{H0:H1}) to address these limitations. Our test does not suffer from the orthogonality issue of LR and is applicable for moderately large $n$. Our test involves a regularization parameter that can reduce the variability of the classic LR test so that it becomes valid. Based on a contiguity theory for random regular graphs developed by Janson \cite{J95}, we derive the asymptotic distributions as power Poisson laws under both $H_0$ and $H_1$, which turn out to be infinite products of power Poisson variables (see Section \ref{sec:ppl}). Based on power Poisson laws, we rigorously analyze the asymptotic power of our test. In Section \ref{sec:power:analysis}, we show that the test is powerful provided that $\kappa$ approaches infinity, and the limit power is not sensitive to the choice of regularization parameter. Our test is practically useful in that the parameters $a,b$ can be consistently estimated when $\kappa>1$, and so the regularization parameter can be empirically selected. Our procedure is based on averaged likelihood-ratios whose computational cost scales exponentially with $n$. This computational issue is partly resolved in Section \ref{sec:approx} via Monte Carlo approximations, with the number of experiments suggested to guarantee the success of such approximations. Simulation examples are provided in Section \ref{sec:sim} to demonstrate the finite sample performance of our methods. In particular, our method achieves desirable size and power, while the methods designed for denser graphs appear to be less powerful. \subsection{Related References.} The problem of testing (\ref{H0:H1}) has been recently considered by \cite{BS16,L16,MS16,BM17,B18,GL17a,GL17b} but only in the growing-degree regime, i.e., $a,b\to\infty$. Specifically, \cite{BS16,MS16,L16} proposed spectral algorithms; \cite{BM17,B18} proposed linear spectral statistics and LR test relating to signed cycles; \cite{GL17a,GL17b} proposed algorithms based subgraph counts. In particular, the LR test by \cite{BM17} was proposed under low SNR which may not be directly applicable here. The growing-degree condition is necessary to guarantee the validity of all these methods which also result in different asymptotic laws than ours. As far as we know, an effective testing procedure that distinguishes SBMs with bounded degrees is still missing. As a side remark, the power Poisson law is unique in sparse network models with bounded degrees as demonstrated in \cite{J95}. In the end, we mention a few papers addressing different models or testing problems than ours: \cite{FH15} proposed a test for examining dependence between network factors and nodal-level attributes; \cite{MPOW17} proposed a variant of multivariate t-test for model diagnosis based on a collection of network samples. \section{LR-Type Test and Asymptotic Properties}\label{sec:varepsilon:LR} The classic LR test requires the calculations of the marginal probability distributions of $A_{uv}$'s under both $H_0$ and $H_1$. By straightforward calculations, it can be shown that, under $H_1$, the marginal distribution of $A$ is \[ P_1(A)=\sum_{\sigma\in\{\pm\}^n}P(A|\sigma)P(\sigma)=2^{-n}\sum_{\sigma\in\{\pm\}^n}\prod_{u<v}p_{uv}(\sigma)^{A_{uv}}q_{uv}(\sigma)^{1-A_{uv}}; \] and under $H_0$, the marginal distribution of $A$ is \[ P_0(A)=\prod_{u<v}p_0^{A_{uv}}q_0^{1-A_{uv}}, \] where $p_0=1-q_0=\frac{a+b}{2n}$. The classic LR test for (\ref{H0:H1}) is then given as follows: \begin{equation}\label{lrt:case1} Y_n=\frac{P_1(A)}{P_0(A)}=2^{-n}\sum_{\sigma\in\{\pm\}^n}\prod_{u<v}\left(\frac{p_{uv}(\sigma)}{p_0}\right)^{A_{uv}} \left(\frac{q_{uv}(\sigma)}{q_0}\right)^{1-A_{uv}}, \end{equation} where $p_{uv}(\sigma)$ and $q_{uv}(\sigma)$ are defined in (\ref{G:p:q:model}). However, \cite{MNS15} shows that $P_0(\cdot)$ and $P_1(\cdot)$ are asymptotically orthogonal when $\kappa\ge1$. So with positive probability, $Y_n$ is asymptotically degenerate to either $0$ or $\infty$. Here we provide a more heuristic understanding for such degenerateness phenomenon. Note that the probability ratio $\frac{p_{uv}(\sigma)}{p_0}$ is equal to either $\frac{2a}{a+b}$ or $\frac{2b}{a+b}$, depending on whether $u,v$ belong to the same community. When $\kappa\ge1$, i.e., $a-b$ is large compared with $a+b$, the two probability ratios considerably differ from each other which brings too much uncertainty into $Y_n$. We propose a regularized LR test, called as $\varepsilon$-LR test, to resolve the degenerateness issue. The idea is quite natural: incorporate a regularization parameter $\varepsilon$ into $Y_n$ to reduce its uncertainty. Our $\varepsilon$-LR test is defined as follows. Let $\kappa_\varepsilon=\frac{(a_\varepsilon-b_\varepsilon)^2}{2(a+b)}$, where $a_\varepsilon=a-\varepsilon$, $b_\varepsilon=b+\varepsilon$. For any $\varepsilon$ satisfying \begin{equation}\label{vareps:condition} \textrm{$0<\varepsilon<\frac{a-b}{2}$ and $\kappa_\varepsilon<1$,} \end{equation} define \begin{equation}\label{lrt:case2} Y_n^\varepsilon=2^{-n}\sum_{\sigma\in\{\pm\}^n}\prod_{u<v}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}} \left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}, \end{equation} where \[ p_{uv}^\varepsilon(\sigma)=\left\{\begin{array}{cc} \frac{a_\varepsilon}{n},&\sigma_u=\sigma_v\\ \frac{b_\varepsilon}{n},&\sigma_u\neq\sigma_v \end{array}\right.,\,\,\,\,\,\, q_{uv}^\varepsilon(\sigma)=1-p_{uv}^\varepsilon(\sigma). \] In other words, we replace $p_{uv}(\sigma)$ and $q_{uv}(\sigma)$ in (\ref{lrt:case1}) by their counterparts $p_{uv}^\varepsilon(\sigma)$ and $q_{uv}^\varepsilon(\sigma)$. The new probability ratio $\frac{p_{uv}^\varepsilon(\sigma)}{p_0}$ is equal to either $\frac{2a_\varepsilon}{a+b}$ or $\frac{2b_\varepsilon}{a+b}$, which are closer to each other due to regularization. Such a trick will be proven to effectively reduce the variability of the classic LR test. Asymptotic distributions and power analysis of $Y_n^\varepsilon$ are provided in subsequent Sections \ref{sec:ppl} and \ref{sec:power:analysis}. \begin{Remark} A more naive approach is to reject $H_0$ if $Y_n>c$ with $c>0$ a predetermined constant. However, the choice of $c$ is a challenging issue. In particular, due to the degenerateness of $Y_n$, it is hard to determine the (asymptotic) probability of rejection given any value of $c$, which poses challenges in analyzing size and power of the test. Instead, our $\varepsilon$-LR test has valid asymptotic distributions which avoids the above issues. \end{Remark} \subsection{Power Poisson Laws.}\label{sec:ppl} Let us first present a power Poisson law for $Y_n^\varepsilon$ under $H_0$. \begin{Theorem}\label{testing:consistency:case2} If $\kappa\ge1$ and $\varepsilon$ satisfies (\ref{vareps:condition}), then under $H_0$, $Y_n^\varepsilon\overset{d}{\to}W_0^\varepsilon$ as $n\to\infty$, where \[ W_0^\varepsilon=\prod_{m=3}^\infty\left(1+\delta_m^\varepsilon\right)^{Z_m^0}\exp\left(-\lambda_m\delta_m^\varepsilon\right),\,\,\,\, Z_m^0\overset{ind}{\sim}\textrm{Poisson}\left(\lambda_m\right). \] Here, $\lambda_m=\frac{1}{2m}\left(\frac{a+b}{2}\right)^m$ and $\delta_m^\varepsilon=\left(\frac{a_\varepsilon-b_\varepsilon}{a+b}\right)^m$. \end{Theorem} Theorem \ref{testing:consistency:case2} shows that, under $H_0$, $Y_n^\varepsilon$ converges in distribution to an infinite product of power Poisson variables. Its proof is based on a contiguity theory for regular random graphs developed by \cite{J95}. Power Poisson law is unique in sparse network with bounded degree, e.g., the number of subgraphs, the number of perfect matchings and the number of edge colourings all follow such a law (see \cite{J95}). This decidedly differs from the growing-degree regime. For instance, when the average degree is growing along with $n$, \cite{BS16} proposed a spectral algorithm that follows Tracy-Widom law; \cite{BM17,B18} examined the classic LR statistics under $\kappa<1$ and linear spectral statistics relating to signed cycles that follow power Gaussian law; \cite{GL17a,GL17b} proposed subgraph-based algorithms that follow Gaussian distributions. According to Theorem \ref{testing:consistency:case2}, we test (\ref{H0:H1}) at significance level $\alpha$ based on the following rule: \[ \textrm{reject $H_0$ iff $Y_n^\varepsilon\ge w_\alpha^\varepsilon$,} \] where $w_\alpha^\varepsilon>0$ satisfies $P(W_0^\varepsilon\le w_\alpha^\varepsilon)=1-\alpha$. The following theorem shows that, under $H_1$, $Y_n^\varepsilon$ asymptotically follows another power Poisson law. \begin{Theorem}\label{power:case2} If $\kappa\ge1$, $\varepsilon$ satisfies (\ref{vareps:condition}) and $(a-b)(a_\varepsilon-b_\varepsilon)<\frac{2(a+b)}{3}$, then under $H_1$, $Y_n^\varepsilon\overset{d}{\to}W_1^\varepsilon$ as $n\to\infty$, where \[ W_1^\varepsilon=\prod_{m=3}^\infty\left(1+\delta_m^\varepsilon\right)^{Z_m^1}\exp\left(-\lambda_m\delta_m^\varepsilon\right),\,\,\,\, Z_m^1\overset{ind}{\sim}\textrm{Poisson}\left(\lambda_m(1+\delta_m)\right). \] Here, $\lambda_m$ and $\delta_m^\varepsilon$ are the same as in Theorem \ref{testing:consistency:case2} and $\delta_m=\left(\frac{a-b}{a+b}\right)^m$. \end{Theorem} We notice that $W_1^\varepsilon$ differs from $W_0^\varepsilon$ only in the Poisson powers, i.e., $Z_m^1$ has larger means than $Z_m^0$. Intuitively, the power of $Y_n^\varepsilon$ should increase when such differences become substantial. Based on Theorems \ref{testing:consistency:case2} and \ref{power:case2}, we can derive the asymptotic power of $Y_n^\varepsilon$ as stated in the corollary below. The power is an unexplicit function of $(a,b,\varepsilon)$. \begin{Corollary}\label{cor:power:case2} If $\kappa\ge1$, $\varepsilon$ satisfies (\ref{vareps:condition}) and $(a-b)(a_\varepsilon-b_\varepsilon)<\frac{2(a+b)}{3}$, then as $n\to\infty$, the power of $Y_n^\varepsilon$ satisfies $P(\textrm{reject $H_0$}|\textrm{under $H_1$})\to P(a,b,\varepsilon)$, where $P(a,b,\varepsilon):=P(W_1^\varepsilon\ge w_\alpha^\varepsilon)$. \end{Corollary} \begin{Remark} The value of $\varepsilon$ can be empirically selected. Specifically, choose $\varepsilon$ to satisfy (\ref{vareps:condition}) and $(a-b)(a_\varepsilon-b_\varepsilon)<\frac{2(a+b)}{3}$ with $a,b$ therein replaced by their consistent estimators. Existence of such consistent estimators is guaranteed by \cite{MNS15} when $\kappa>1$. \end{Remark} \subsection{Power Analysis.}\label{sec:power:analysis} Corollary \ref{cor:power:case2} derives an asymptotic power $P(a,b,\varepsilon)$ for $Y_n^\varepsilon$. In this section, we further examine this power and demonstrate whether and when it can approach one. It is challenging to directly analyze $P(a,b,\varepsilon)$ for fixed $a,b$ due to the lack of explicit expression. Instead, we will consider the relatively easier growing-degree regime ($a+b\to\infty$) and discuss its connection to existing work. Theorem \ref{lim:power:case:2} provides an explicit expression for the limit of $P(a,b,\varepsilon)$. Let $\Phi(\cdot)$ denote the cumulative distribution function of standard normal variable and $z_{1-\alpha}$ denote its $1-\alpha$ quantile, i.e., $\Phi(z_{1-\alpha})=1-\alpha$. \begin{Theorem}\label{lim:power:case:2} If $\kappa\ge1$ and $\varepsilon\in\left(0,\frac{a-b}{2}\right)$ satisfies, when $a+b\to\infty$, $\frac{(a_\varepsilon-b_\varepsilon)^2}{2(a+b)}\to k_1$ and $\frac{(a-b)(a_\varepsilon-b_\varepsilon)}{2(a+b)}\to k_2$ for constants $k_1,k_2\in(0,1)$, then $P(a,b,\varepsilon)\to\Phi\left(\frac{\sigma_2^2}{\sigma_1}-z_{1-\alpha}\right)$ as $a+b\to\infty$, where $\sigma_l^2=\sum_{m=3}^\infty\frac{1}{2m}k_l^m=-\frac{1}{2}\left(\log(1-k_l)+k_l+\frac{1}{2}k_l^2\right)$, $l=1,2$. \end{Theorem} We remark that the limit power $\Phi\left(\frac{\sigma_2^2}{\sigma_1}-z_{1-\alpha}\right)$ approaches one if $\kappa\to\infty$ (regardless of the choice of $\varepsilon$). To see this, note that \begin{equation}\label{eq:power:anal} \frac{\sigma_2^2}{\sigma_1}=-\frac{\log(1-k_2)+k_2+\frac{1}{2}k_2^2}{\sqrt{-2\left(\log(1-k_1)+k_1+\frac{1}{2}k_1^2\right)}} \asymp\kappa^{3/2}. \end{equation} The above (\ref{eq:power:anal}) holds uniformly for $\varepsilon$ satisfying the conditions of Theorem \ref{lim:power:case:2} and $\kappa^{3/2}$ on the right side is free of $\varepsilon$. If $\kappa\to\infty$, then $\frac{\sigma_2^2}{\sigma_1}\to\infty$, and so $\Phi\left(\frac{\sigma_2^2}{\sigma_1}-z_{1-\alpha}\right)$ approaches one. The power behavior merely relies on $\kappa$ while being free of $\varepsilon$. Our result is closely relating to \cite{BM17} who investigate the asymptotic power of the classic LR test which nonetheless requires $0<\kappa<1$. \cite{MS16} proposed an efficient method based on semidefinite program but their size and power are not explicitly quantifiable like ours. \subsection{Monte-Carlo Approximation.}\label{sec:approx} Despite its theoretically nice properties, the test statistic $Y_n^\varepsilon$ might be computationally infeasible. This can be easily seen from (\ref{lrt:case2}), i.e., $Y_n^\varepsilon$ can be viewed as the average of the quantity $g_n^\varepsilon(\sigma)$ over the entire space of configurations $\{\pm\}^n$, where \[ g_n^\varepsilon(\sigma)=\prod_{u<v}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}} \left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}. \] The computational effort for the direct averages scales exponentially with $n$. So an accurate and computationally efficient approximation of $Y_n^\varepsilon$ would be needed for practical use. In this section, we consider the classical Monte Carlo (MC) method which randomly chooses a set of $M$ configurations $\sigma[1],\ldots,\sigma[M]$ from $\{\pm\}^n$. The MC method works directly under the bounded-degree regime. The average $Y_n^\varepsilon$ is naturally approximated by the sample mean of $g_n^\varepsilon(\sigma[l])$'s, which may substantially reduce computational cost if $M\ll 2^n$. However, a small choice of $M$ may result in inaccurate approximation. An interesting question is how small $M$ can be to ensure valid approximation. More explicitly, we aim to find an order of $M$ such that the following approximation becomes valid: \begin{equation}\label{approx:lrt2} \widehat{Y}_n^\varepsilon\equiv\frac{1}{M}\sum_{l=1}^M g_n^\varepsilon(\sigma[l])=Y_n^\varepsilon+o_P(1). \end{equation} The following theorem shows that the validity of (\ref{approx:lrt2}) is possible. \begin{Theorem}\label{thm:approx2} Suppose $\kappa\ge1$, $\varepsilon$ satisfies (\ref{vareps:condition}), and $M\gg\exp\left(\frac{n\kappa_\varepsilon}{2}\right)$. Then (\ref{approx:lrt2}) holds under $H_0$. Moreover, if $(a_\varepsilon-b_\varepsilon)(a-b)<a+b$, then (\ref{approx:lrt2}) holds under $H_1$. \end{Theorem} According to Theorem \ref{thm:approx2}, (\ref{approx:lrt2}) becomes valid when only $M\gg\exp\left(\frac{n\kappa_\varepsilon}{2}\right)$ configurations are used, e.g., $M\asymp(\log{n})^c\exp\left(\frac{n\kappa_\varepsilon}{2}\right)$ for a constant $c>0$. When $\kappa_\varepsilon$ is close to zero, this may substantially reduce the computational cost from $O(2^n)$ to nearly $O([\exp(\kappa_\varepsilon/2)]^n)$. However, MC method still requires heavy computation. A more efficient method would be highly useful. \section{Numerical Studies} In this section, we examine the performance of the proposed testing procedure through simulation studies in Section \ref{sec:sim}, and through real-world data sets in Section \ref{sec:realdata}. \subsection{Simulation.}\label{sec:sim} The empirical performance of our test statistic $\widehat{Y}_n^\varepsilon$ is demonstrated through simulation studies. We also compared our method with the spectral method proposed by Bickel and Sarkar \cite{BS16} and the subgraph count method proposed by Gao and Lafferty \cite{GL17a}. Throughout we assume that both $a$ and $b$ are known. We evaluated the size and power of various methods at significance level 0.05. For size, data were generated from $\mathcal{G}\left(n,\frac{a+b}{2n}\right)$. For power, data were generated from $\mathcal{G}\left(n,\frac{a}{n},\frac{b}{n}\right)$. Both size and power were calculated as proportions of rejections based on 500 independent experiments. We examined various choices of $a,b,n,\varepsilon$ for $\widehat{Y}_n^\varepsilon$. For convenience, denote $a=2.5+c$, $b=2.5-c$. We chose $n=20,30,40,45$ and $(c,\varepsilon)=(2.10, 1.10), (2.15, 1.15), (2.25, 1.25), (2.35, 1.35)$. The $\varepsilon$ in each case was chosen to be approximately $c/2$. The corresponding values of SNR $\kappa=\frac{(a-b)^2}{2(a+b)}$ are $1.76, 1.85, 2.03, 2.21$. We chose $100n^3e^{\frac{n\kappa_\varepsilon}{2}}$ samples for MC approximations according to Theorem \ref{thm:approx2} for calculating $\widehat{Y}_n^\varepsilon$. Table \ref{taba2} summarizes the size and power of our test. For all cases, the sizes of the $\widehat{Y}_n^\varepsilon$ are close to the 0.05 nominal level indicating the validity of the test. For each choice of $(c,\varepsilon)$, the power increases along with $n$. For any fixed $n$, the power increases as $\kappa$ increases, consistent with Theorem \ref{lim:power:case:2} which states that the power should increase with $\kappa$. \begin{table}[h] \centering \begin{tabular}{ |c c |c |c|c|c| } \hline $(c,\epsilon)$& $\kappa$ & $n=20$ & $n=30$ & $n=40$ & $n=45$\\ \hline (2.10, 1.10) & 1.76 &0.572 (0.058) &0.654 (0.042) &0.730 (0.050)& 0.812 (0.048)\\ \hline (2.15, 1.15) & 1.85 &0.598 (0.054) &0.684 (0.052) &0.764 (0.050) & 0.852 (0.040)\\ \hline (2.25, 1.25) & 2.03 &0.626 (0.056) &0.704 (0.044) &0.802 (0.042) & 0.910 (0.044)\\ \hline (2.35, 1.35) & 2.21 &0.648 (0.044) &0.736 (0.042) &0.888 (0.058) & 1.000 (0.042)\\ \hline \end{tabular} \caption{Power (Size) of $\varepsilon$-LR test based on various choices of $c,\varepsilon,n$.}\label{taba2} \end{table} Tables \ref{tb::bs16} and \ref{tabgaoad} summarize the size and power of BS's spectral method and GL's subgraph count method. It is worth mentioning that the sizes of both methods are free of $c$ since the null models under various values of $c$ are equivalent and the sizes of both methods are uniquely determined by the common null model. Due to the high sparsity of the simulated networks, subgraph counts are generally small, and we obtained the critical values for GL's method based on resampling instead of using asymptotic distribution. It is observed that both methods achieve smaller power than $\varepsilon$-LR while maintaining the correct size. \begin{table}[h] \centering \begin{tabular}{ |c c |c |c|c|c| } \hline $c$& $\kappa$ & $n=20$ & $n=30$ & $n=40$ & $n=45$\\ \hline 2.10 & 1.76 &0.308 (0.070)&0.260 (0.048)&0.266 (0.058)&0.300 (0.054)\\ \hline 2.15 & 1.85 &0.330 (0.070)&0.280 (0.048)&0.288 (0.058)&0.314 (0.054)\\ \hline 2.25 & 2.03 &0.364 (0.070)&0.336 (0.048)&0.348 (0.058)&0.362 (0.054)\\ \hline 2.35 & 2.21 &0.400 (0.070)&0.376 (0.048)&0.402 (0.058)&0.402 (0.054)\\ \hline \end{tabular} \caption{Power (Size) of BS's spectral test based on various choices of $c,n$.}\label{tb::bs16} \end{table} \begin{table}[h] \centering \begin{tabular}{ |c c |c|c|c|c| } \hline $c$ & $\kappa$ & $n=20$ & $n=30$ & $n=40$ & $n=45$\\ \hline 2.10 & 1.76 & 0.156 (0.05) &0.216 (0.05) &0.196 (0.05) &0.168 (0.05) \\ \hline 2.15& 1.85 & 0.216 (0.05) &0.224 (0.05) &0.204 (0.05) &0.206 (0.05) \\ \hline 2.25 & 2.03 & 0.296 (0.05) &0.234 (0.05) &0.246 (0.05) &0.300 (0.05)\\ \hline 2.35 & 2.21 & 0.306 (0.05) &0.350 (0.05) &0.328 (0.05) &0.336 (0.05) \\ \hline \end{tabular} \caption{Power (Size) of GL's subgraph count test based on various choices of $c,n$.}\label{tabgaoad} \end{table} \subsection{Real Data Analysis.}\label{sec:realdata} In this section, we applied our procedure to analyze the political book data (\cite{Newman06}) which has 105 political books (nodes). Two books are connected if they were frequently co-purchased on Amazon. This data was analyzed by \cite{ZLZ11} who detected three communities. We used R package \textit{igraph} based on a spin-glass model and simulated annealing to recover their findings, and denote the three communities by $C_I, C_{II}, C_{III}$ which contain 20, 44, 41 nodes, respectively. Books within the same community are expected to demonstrate similar political tendencies. The aim of this study is to examine whether our method can detect the existence of the communities. Our $\varepsilon$-LR method was based on $M=10^7e^{n\delta}$ MC samples with $\delta\approx 0.01$ and $\varepsilon\approx\frac{\hat{a}+\hat{b}}{2}$, where $\widehat{a}=21.633$, $\widehat{b}=1.139$ are MLEs of $a,b$ under $H_1$. We first examined whether $H_0$ is rejected (at 0.05) over each community. Table \ref{tabrej:pbd:1} summarizes the results. We find that all three methods rejected $H_0$ over $C_I$. This incorrect decision might be due to the small size of the first community. Moreover, $\varepsilon$-LR failed to reject $H_0$ over communities $C_{II}$, $C_{III}$; BS rejected $H_0$ over $C_{II}$, $C_{III}$; GL rejected $H_0$ over $C_{III}$ while failed to reject $H_0$ over $C_{II}$. \begin{table}[h] \centering \begin{tabular}{ cccc } Method & \multicolumn{3}{c}{P-value}\\ \hline & $C_I$ & $C_{II}$ & $C_{III}$\\ \hline $\varepsilon$-LR&0.000 &1.000 &1.000 \\ BS & 0.000 & 0.000 & 0.000\\ GL & 0.000 & 0.465 & 0.039\\ \hline \end{tabular} \caption{P-values of three methods over communities $C_I, C_{II}, C_{III}$ for Political Book Data.}\label{tabrej:pbd:1} \end{table} We then examined whether $H_0$ is rejected for a subnetwork with nodes from two different communities. In particular, we uniformly sampled 20 nodes out of $C_{II}$ without replacement, and combined it with $C_I$. Therefore, the combined network has two communities of 20 nodes each. We also examined the combination regimes $C_I$ \& $C_{II}$ and $C_{II}$ \& $C_{III}$ with 20 nodes uniformly sampled from $C_{II}$ in the former and from both $C_{II}, C_{III}$ in the latter (so each combination has 40 nodes in total). We repeated each combination regime 100 times and calculated the proportions that $H_0$ was rejected. Results are summarized in Table \ref{tabrej:pbd:2}. It is observed that all three methods rejected $H_0$ 100 times over $C_I$ \& $C_{II}$ hence the success rates are 100\%. The rejection rates of $\varepsilon$-LR over $C_I$ \& $C_{III}$ and over $C_{II}$ \& $C_{III}$ are $92\%$ and $83\%$ respectively. The success rates of BS $100\%$, and the success rates of GL are 98\% and 100\%, for both combination regimes. \begin{table}[h] \centering \begin{tabular}{ cccc } Method & \multicolumn{3}{c}{Rejection Proportion}\\ \hline & $C_I$ \& $C_{II}$ & $C_I$ \& $C_{III}$ & $C_{II}$ \& $C_{III}$\\ \hline $\varepsilon$-LR&100\% &92\% &83\% \\ BS & 100\% & 100\% & 100\%\\ GL & 100\% & 98\% & 100\%\\ \hline \end{tabular} \caption{Rejection proportions by three methods in different combination regimes for Political Book Data.}\label{tabrej:pbd:2} \end{table} \section{Discussions} The work of \cite{DKMZ11} implies that extension of the current work to multi-community setting is highly important but nontrivial. As far as we know, only a few works address such settings but mostly in community detection. For instance, \cite{NN14} provides a sufficient condition for impossible detection; \cite{AS17} presents an information-theoretic phase transition for the SNR to yield successful detection which strengthens the work of \cite{NN14}. The test statistic $\widehat{Y}_n^\varepsilon$ can be viewed as a type of partition function over Gibbs field. Popular approximations of partition functions in statistical physics include MC approximations and mean-field approximations. This paper only considers the former while leaves the latter as a future topic. Mean-field approximation has proven to work well in dense magnetism such as Curie-Weiss model (see \cite{W1907}). Recently, validity of mean-field approximation was established by \cite{BM17-2} in the sparser settings which satisfy the so-called ``mean-field assumption,'' i.e., the trace of the squared adjacency matrix is $o_P(n)$. This assumption fails in our setting in that the trace becomes $O_P(n)$. Additional theory is needed to extend the results of \cite{BM17-2}. \section{Appendix: Proofs} In this section, we prove the main results of this paper. Our asymptotic results are derived based on the following Proposition \ref{basic:prop} which was proved by Janson in \cite{J95}. For arbitrary non-negative integer $x$, let $[x]_j$ denote the descending factorial $x(x-1)\cdots(x-j+1)$. \begin{Proposition}\label{basic:prop} Let $\lambda_i>0$, $i=1,2,\ldots$, be constants and suppose that for each $n$ there are random variables $X_{in}$, $i=1,2,\ldots$, and $Y_n$ (defined on the same probability space) such that $X_{in}$ is non-negative integer valued and $E\{Y_n\}\neq0$ (at least for large $n$), and furthermore the following conditions are satisfied: \begin{itemize} \item[(A1)] $X_{in}\overset{d}{\to}Z_i$ as $n\to\infty$, jointly for all $i$, where $Z_i\sim\textrm{Poisson}(\lambda_i)$ are independent Poisson random variables; \item[(A2)] ${\mathbb{E}}\{Y_n[X_{1n}]_{j_1}\cdots[X_{kn}]_{j_k}\}/{\mathbb{E}}\{Y_n\}\to\prod_{i=1}^k\mu_i^{j_i}$, as $n\to\infty$, for some $\mu_i\ge0$ and every finite sequence $j_1,\ldots,j_k$ of non-negative integers; \item[(A3)] $\sum_{i=1}^\infty\lambda_i\delta_i^2<\infty$, where $\delta_i=\mu_i/\lambda_i-1$; \item[(A4)] ${\mathbb{E}}\{Y_n^2\}/({\mathbb{E}}\{Y_n\})^2\to\exp\left(\sum_{i=1}^\infty\lambda_i\delta_i^2\right)$. \end{itemize} Then \[ \frac{Y_n}{{\mathbb{E}}\{Y_n\}}\overset{d}{\to}W\equiv\prod_{i=1}^\infty(1+\delta_i)^{Z_i}\exp(-\lambda_i\delta_i),\,\,\,\,\textrm{as $n\to\infty$}. \] \end{Proposition} \begin{Remark}\label{rem:J95} Janson (1995) \cite{J95} showed that the infinite product defining $W$ in Proposition \ref{basic:prop} converges in $L^2$ a.s. with ${\mathbb{E}} W=1$ and ${\mathbb{E}} W^2=\exp\left(\sum_{i\ge1}\lambda_i\delta_i^2\right)$. \end{Remark} Before proofs, we need the following lemma. \begin{Lemma}\label{lemma:short:cycles} For a random graph $G$ with vertex $1,\ldots,n$, let $X_{mn}$ be the number of $m$-cycles of $G$, for $m\ge3$. Let $\lambda_m=\frac{1}{2m}\left(\frac{a+b}{2}\right)^m$ and $\delta_m=\left(\frac{a-b}{a+b}\right)^m$. \begin{enumerate} \item\label{lemma:short:cycles:1} Under $G\sim\mathcal{G}(n,p_0)$, for any $k\ge3$, $\{X_{mn}\}_{m=3}^k$ jointly converge to independent Poisson variables with mean $\lambda_m$. \item\label{lemma:short:cycles:2} Under $G\sim\mathcal{G}\left(n,\frac{a}{n},\frac{b}{n}\right)$, for any $k\ge3$, $\{X_{mn}\}_{m=3}^k$ jointly converge to independent Poisson variables with mean $\lambda_m(1+\delta_m)$. \end{enumerate} \end{Lemma} \begin{proof}[Proof of Lemma \ref{lemma:short:cycles}] The first part was well known (see \cite{MNS15}). We only prove the second part. Denote ${\mathbb{E}}_1$ the expectation based on hypothesis $H_1$. Let $H$ be a graph on a subset of $[n]$ with vertex set $\mathcal{V}(H)$ and edge set $\mathcal{E}(H)$. Use $1_H$ to denote the 0-1 random variable that is 1 when $\mathcal{E}(H)\subseteq\mathcal{E}(G)$ and $P(H)$ for the probability that $1_H=1$. For $3\le m\le k$, let $H_{m1},\ldots,H_{mj_m}$ be a $j_m$-tuple of distinct $m$-cycles. Then \[ \prod_{m=3}^k[X_{mn}]_{j_m}=\sum_{(H_{mi})}\prod_{m=3}^k\prod_{i=1}^{j_m}1_{H_{mi}}, \] where the sum ranges over all tuples of distinct cycles $\{H_{mi}: 3\le m\le k, 1\le i\le j_m\}$; each $H_{mi}$ is an $m$-cycle and all cycles are distinct. Let $A$ be the set of all such tuples of cycles for which the cycles are vertex-disjoint and let $\bar{A}$ be its complement, i.e., any tuple of $\bar{A}$ contains two cycles with at least one common vertex. Then \begin{eqnarray}\label{proof:lemma:short:cycles:eqn1} {\mathbb{E}}_1\prod_{m=3}^k[X_{mn}]_{j_m}&=&\sum_{(H_{mi})}{\mathbb{E}}_1\prod_{m=3}^k\prod_{i=1}^{j_m}1_{H_{mi}}\nonumber\\ &=&\sum_{(H_{mi})\in A}{\mathbb{E}}_1\prod_{m=3}^k\prod_{i=1}^{j_m}1_{H_{mi}}+\sum_{(H_{mi})\in \bar{A}}{\mathbb{E}}_1\prod_{m=3}^k\prod_{i=1}^{j_m}1_{H_{mi}} \end{eqnarray} Since the number of $m$-cycles on a graph of $s$ vertexes is $\frac{s!}{(s-m)!2m}$ (two directions and $m$ distinct starting vertexes give us $2m$ the same $m$-cycles), one gets that $|A|=\frac{n!}{(n-M)!}\prod_{m=3}^k\left(\frac{1}{2m}\right)^{j_m}$ with $M=\sum_{m=3}^k mj_m$ (see also \cite[Chapter 4]{B01} for more complete derivation). Meanwhile, take $\tau$ uniformly from $\{\pm\}^n$ and define $\tau^{mi}$ be the restriction of $\tau$ on the vertexes of $H_{mi}$, and define $N_{mi}=\sum_{(u,v)\in\mathcal{E}(H_{mi})}1(\tau^{mi}_u\neq\tau^{mi}_v)$. The $\tau^{mi}$'s are independent thanks to the vertex disjointness of $H_{mi}$'s. Following \cite[Lemma 3.3]{MNS15} one can show that $P(N_{mi}=l)=2^{-m+1}{m\choose l}$ for even $l\in[0,m]$ and zero for odd $l$. Then one has \begin{eqnarray*} {\mathbb{E}}_1\prod_{m=3}^k\prod_{i=1}^{j_m}1_{H_{mi}}&=&{\mathbb{E}}_\tau{\mathbb{E}}_1\{\prod_{m=3}^k\prod_{i=1}^{j_m}1_{H_{mi}}|\tau\}\\ &=&{\mathbb{E}}_\tau\prod_{m=3}^k\prod_{i=1}^{j_m}\prod_{(u,v)\in\mathcal{E}(H_{mi})}\left(\frac{a}{n}\right)^{1(\tau_u=\tau_v)}\left(\frac{b}{n}\right)^{1(\tau_u\neq\tau_v)}\\ &=&{\mathbb{E}}_\tau\prod_{m=3}^k\prod_{i=1}^{j_m}\prod_{(u,v)\in\mathcal{E}(H_{mi})}\left(\frac{a}{n}\right)^{1(\tau^{mi}_u=\tau^{mi}_v)}\left(\frac{b}{n}\right)^{1(\tau^{mi}_u\neq\tau^{mi}_v)}\\ &=&{\mathbb{E}}_\tau\prod_{m=3}^k\prod_{i=1}^{j_m}\left(\frac{a}{n}\right)^{m-N_{mi}}\left(\frac{b}{n}\right)^{N_{mi}}. \end{eqnarray*} Since $\tau$ is broken into disjoint and independent $(\tau^{mi})_{3\le m\le k, 1\le i\le j_m}$, the above is equal to \begin{eqnarray*} &&\prod_{m=3}^k\prod_{i=1}^{j_m}{\mathbb{E}}_{\tau^{mi}}\left(\frac{a}{n}\right)^{m-N_{mi}}\left(\frac{b}{n}\right)^{N_{mi}}\\ &=&\prod_{m=3}^k\prod_{i=1}^{j_m}2^{-m}\left[\left(\frac{a+b}{n}\right)^m+\left(\frac{a-b}{n}\right)^m\right] =n^{-M}\prod_{m=3}^k\left[\left(\frac{a+b}{2}\right)^m\left(1+\delta_m\right)\right]. \end{eqnarray*} Then the first part of (\ref{proof:lemma:short:cycles:eqn1}) becomes \begin{eqnarray*} &&|A|\times n^{-M}\prod_{m=3}^k\left[\left(\frac{a+b}{2}\right)^m\left(1+\delta_m\right)\right] \\ &=&\frac{n!}{(n-M)!n^M}\prod_{m=3}^k (\lambda_m(1+\delta_m))^{j_m}\overset{n\to\infty}{\rightarrow}\prod_{m=3}^k (\lambda_m(1+\delta_m))^{j_m}. \end{eqnarray*} On the other hand, for any $(H_{mi})\in \bar{A}$, $H:=\cup H_{mi}$ has at most $M-1$ vertexes and $M$ edges, and $|\mathcal{E}(H)|>|\mathcal{V}(H)|$. Since \[ {\mathbb{E}}_1\{\prod_{m=3}^k\prod_{i=1}^{j_m}1_{H_{mi}}|\tau\}= \prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{a}{n}\right)^{1(\tau_u=\tau_v)}\left(\frac{b}{n}\right)^{1(\tau_u\neq\tau_v)}\le\left(\frac{\max\{a,b\}}{n}\right)^{|\mathcal{E}(H)|}, \] and there are ${n\choose |\mathcal{V}(H)|}|\mathcal{V}(H)|!$ graphs isomorphic to $H$, then \[ \sum_{\textrm{$H'$ is isomorphic to $H$}}{\mathbb{E}}_1\{1_{H'}|\tau\}\le \left(\frac{\max\{a,b\}}{n}\right)^{|\mathcal{E}(H)|}{n\choose |\mathcal{V}(H)|}|\mathcal{V}(H)|!\to0. \] Since there are a bounded number of isomorphism classes, the second part of (\ref{proof:lemma:short:cycles:eqn1}) tends to zero as $n\to\infty$. Hence, ${\mathbb{E}}_1\prod_{m=3}^k[X_{mn}]_{j_m}\to\prod_{m=3}^k (\lambda_m(1+\delta_m))^{j_m}.$ for any $k\ge3$ and integers $j_3,\ldots,j_k$. It follows by \cite[Lemma 2.8]{W99} that the desirable result holds. \end{proof} \subsection{Proofs in Section \ref{sec:ppl}} \begin{proof}[Proof of Theorem \ref{testing:consistency:case2}] Let ${\mathbb{E}}_0$ denote the expectations under hypotheses $H_0$. We will use Proposition \ref{basic:prop} to prove the result, for which we will check the Conditions A1 to A4 therein. Some of the details are rooted in \cite{MNS15}. To ease reading, we provide the detailed proofs. Obviously, ${\mathbb{E}}_0Y_n^\varepsilon=1$. Let $X_{mn}$ be the number of $m$-cycles of $G\sim\mathcal{G}(n,p_0)$, for $m\ge3$. Following Lemma \ref{lemma:short:cycles} Part \ref{lemma:short:cycles:1}, for any $k\ge3$, $\{X_{mn}\}_{m=3}^k$ jointly converge to independent Poisson variables with mean $\lambda_m=\frac{1}{2m}\left(\frac{a+b}{2}\right)^m$. This verifies Condition A1. To check Condition A2, let $H=(H_{mi})_{3\le m\le k, 1\le i\le j_m}$ be a tuple of short cycles of disjoint vertexes; each $H_{mi}$ is an $m$-cycle, $M=\sum_{m=3}^k mj_m$, and the vertexes of $H_{mi}$'s are disjoint. Let $\sigma^{1mi},\sigma^{2mi}$ be the restrictions of $\sigma$ over $\mathcal{V}(H_{mi})$ and $[n]\backslash\mathcal{V}(H_{mi})$, and $\sigma^1,\sigma^2$ be the restrictions of $\sigma$ over $\mathcal{V}(H)$ and $[n]\backslash\mathcal{V}(H)$. By direct examinations we have \begin{eqnarray} {\mathbb{E}}_0 Y_n^\varepsilon 1_H &=&2^{-n}\sum_{\sigma\in\{\pm\}^n}{\mathbb{E}}_0 1_H \prod_{u<v}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}}\left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}} \nonumber\\ &=&2^{-n}\sum_{\sigma\in\{\pm\}^n}{\mathbb{E}}_0 1_H\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}}\left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}\nonumber\\ &&\times\prod_{(u,v)\in\overline{\mathcal{E}(H)}}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}} \left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}\nonumber\\ &=&2^{-n}\sum_{\sigma\in\{\pm\}^n}{\mathbb{E}}_0 1_H\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}}\left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}.\label{testing:consistency:eqn1} \end{eqnarray} Since $\sigma$ is broken into $\sigma^1$ and $\sigma^2$ which are supported on $\mathcal{V}$ and its complement respectively, and $p_{uv}^\varepsilon(\sigma)$, $q_{uv}^\varepsilon(\sigma)$ only depend on $\sigma^1$ when $(u,v)\in\mathcal{E}(H)$, (\ref{testing:consistency:eqn1}) is equal to the following \begin{eqnarray} &&2^{-n}\sum_{\sigma^1\in\{\pm\}^{\mathcal{V}(H)}}\sum_{\sigma^2\in\{\pm\}^{[n]\backslash\mathcal{V}(H)}} {\mathbb{E}}_0 1_H\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma^1)}{p_0}\right)^{A_{uv}} \left(\frac{q_{uv}^\varepsilon(\sigma^1)}{q_0}\right)^{1-A_{uv}}\nonumber\\ &=&2^{-M}\sum_{\sigma^1\in\{\pm\}^{\mathcal{V}(H)}}{\mathbb{E}}_0 1_H\prod_{(u,v)\in\mathcal{E}(H)} \left(\frac{p_{uv}^\varepsilon(\sigma^1)}{p_0}\right)^{A_{uv}} \left(\frac{q_{uv}^\varepsilon(\sigma^1)}{q_0}\right)^{1-A_{uv}}.\label{testing:consistency:eqn1-1} \end{eqnarray} Since $1_H=1$ implies $\mathcal{E}(H)\subset\mathcal{E}(G)$, any $(u,v)\in\mathcal{V}(H)$ leads to $A_{uv}=1$. Meanwhile, ${\mathbb{E}}_0 1_H=p_0^M$, hence (\ref{testing:consistency:eqn1-1}) equals \begin{eqnarray*} &&2^{-M}p_0^M\sum_{\sigma^1\in\{\pm\}^{\mathcal{V}(H)}}\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma^1)}{p_0}\right)\nonumber\\ &=&{\mathbb{E}}_{\sigma^1}\prod_{(u,v)\in\mathcal{E}(H)}p_{uv}^\varepsilon(\sigma^1)=\prod_{m=3}^k\prod_{i=1}^{j_m}{\mathbb{E}}_{\sigma^{1mi}}\prod_{(u,v)\in\mathcal{E}(H_{mi})} p_{uv}^\varepsilon(\sigma^{1mi}) =n^{-M}\prod_{m=3}^k\prod_{i=1}^{j_m}{\mathbb{E}}_{\sigma^{1mi}}a_\varepsilon^{m-N_{mi}}b_\varepsilon^{N_{mi}},\nonumber \end{eqnarray*} where $N_{mi}=\sum_{(u,v)\in\mathcal{E}(H_{mi})}1(\sigma_u^{1mi}\neq\sigma^{1mi}_v)$, the number of edges over $H_{mi}$ with distinct end points. Following the proof of \cite[Lemma 3.3]{MNS15}, \begin{eqnarray*} {\mathbb{E}}_{\sigma^{1mi}}a_\varepsilon^{m-N_{mi}}b_\varepsilon^{N_{mi}}=2^{-m}\left[(a_\varepsilon+b_\varepsilon)^m+(a_\varepsilon-b_\varepsilon)^m\right]= \left(\frac{a+b}{2}\right)^m\left(1+\left(\frac{a_\varepsilon-b_\varepsilon}{a+b}\right)^m\right). \end{eqnarray*} Hence, \[ {\mathbb{E}}_0 Y_n^\varepsilon 1_H=n^{-M}\prod_{m=3}^k\left( \left(\frac{a+b}{2}\right)^m\left(1+\left(\frac{a_\varepsilon-b_\varepsilon}{a+b}\right)^m\right) \right)^{j_m}. \] Let $A$ be the set of tuples $(H_{mi})_{3\le m\le k, 1\le i\le j_m}$ for which the cycles are vertex-disjoint and let $\bar{A}$ be its complement. Using $|A|=\frac{n!}{(n-M)!}\prod_{m=3}^k\left(\frac{1}{2m}\right)^{j_m}$ (see proof of Lemma \ref{lemma:short:cycles}) we get that \begin{eqnarray*} \sum_{H\in A}{\mathbb{E}}_0 Y_n^\varepsilon 1_H&=&|A|n^{-M}\prod_{m=3}^k\left( \left(\frac{a+b}{2}\right)^m\left(1+\left(\frac{a_\varepsilon-b_\varepsilon}{a+b}\right)^m\right) \right)^{j_m}\\ &\overset{n\to\infty}{\rightarrow}&\prod_{m=3}^k \left(\frac{1}{2m}\left(\frac{a+b}{2}\right)^m\left(1+\left(\frac{a_\varepsilon-b_\varepsilon}{a+b}\right)^m\right)\right)^{j_m}\\ &=&\prod_{m=3}^k\left(\lambda_m(1+\delta_m^\varepsilon)\right)^{j_m}, \end{eqnarray*} where $\delta_m^\varepsilon=\left(\frac{a_\varepsilon-b_\varepsilon}{a+b}\right)^m$. Similar to (\ref{testing:consistency:eqn1}) one gets that, for $H\in\bar{A}$, \begin{eqnarray*} {\mathbb{E}}_0Y_n^\varepsilon 1_H&=&2^{-n}\sum_{\sigma\in\{\pm\}^n}{\mathbb{E}}_0 1_H\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}}\left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}\\ &=&2^{-n}\sum_{\sigma\in\{\pm\}^n}\prod_{(u,v)\in\mathcal{E}(H)}\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\times P_0(H)\\ &\le&a_\varepsilon^{|\mathcal{E}(H)|}n^{-|\mathcal{E}(H)|}, \end{eqnarray*} where the last inequality follows from $p_{uv}^\varepsilon(\sigma)\le a_\varepsilon/n$ and $P_0(H)=p_0^{|\mathcal{E}(H)|}$. So \[ \sum_{\textrm{$H'$ is isomorphic to $H$}}{\mathbb{E}}_0Y_n^\varepsilon 1_H \le a_\varepsilon^{|\mathcal{E}(H)|}n^{-|\mathcal{E}(H)|}{n\choose |\mathcal{V}(H)|}|\mathcal{V}(H)|!\to0, \] which leads to $\sum_{H\in\bar{A}}{\mathbb{E}}_0Y_n^\varepsilon 1_H\to0$ using a similar argument as the proof of Lemma \ref{lemma:short:cycles} Part \ref{lemma:short:cycles:2}. So as $n\to\infty$, \[ {\mathbb{E}}_0 Y_n^\varepsilon [X_{3n}]_{j_3}\cdots[X_{kn}]_{j_k}=\sum_{H\in A}{\mathbb{E}}_0Y_n^\varepsilon 1_H+\sum_{H\in\bar{A}}{\mathbb{E}}_0Y_n^\varepsilon 1_H\to \prod_{m=3}^k\left(\lambda_m(1+\delta_m^\varepsilon)\right)^{j_m}, \] which verifies Condition A2. Condition A3 holds due to the following trivial fact: \[ \sum_{m\ge3}\lambda_m(\delta_m^\varepsilon)^2=\sum_{m\ge3}\frac{1}{2m}\left(\frac{(a_\varepsilon-b_\varepsilon)^2}{2(a+b)}\right)^m= \sum_{m\ge3}\frac{1}{2m}\left(\frac{(a-b-2\varepsilon)^2}{2(a+b)}\right)^m <\infty. \] In the end let us check Conditions A4. Let $N_{uv}^{\sigma\tau}=1(\sigma_u=\sigma_v)+1(\tau_u=\tau_v)$. Note that \begin{eqnarray*} {\mathbb{E}}_0 (Y_n^\varepsilon)^2&=& 4^{-n}\sum_{\sigma,\tau\in\{\pm\}^n}\prod_{u<v}{\mathbb{E}}_0 \left(\frac{p_{uv}^\varepsilon(\sigma)p_{uv}^\varepsilon(\tau)}{p_0^2}\right)^{A_{uv}} \left(\frac{q_{uv}^\varepsilon(\sigma)q_{uv}^\varepsilon(\tau)}{q_0^2}\right)^{1-A_{uv}}\\ &=&4^{-n}\sum_{\sigma,\tau\in\{\pm\}^n}\prod_{u<v}\left( \frac{p_{uv}^\varepsilon(\sigma)p_{uv}^\varepsilon(\tau)}{p_0}+ \frac{q_{uv}^\varepsilon(\sigma)q_{uv}^\varepsilon(\tau)}{q_0} \right) \\ &=&4^{-n}\sum_{\sigma,\tau\in\{\pm\}^n}\prod_{u<v} \left(\frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right)^{N_{uv}^{\sigma\tau}} \left(\frac{b_\varepsilon}{n}\right)^{2-N_{uv}^{\sigma\tau}}+\frac{1}{q_0} \left(1-\frac{a_\varepsilon}{n}\right)^{N_{uv}^{\sigma\tau}} \left(1-\frac{b_\varepsilon}{n}\right)^{2-N_{uv}^{\sigma\tau}}\right)\\ &=&4^{-n}\sum_{\sigma,\tau\in\{\pm\}^n} \prod_{N_{uv}^{\sigma\tau}=0}\left(\frac{1}{p_0}\left(\frac{b_\varepsilon}{n}\right)^2 +\frac{1}{q_0}\left(1-\frac{b_\varepsilon}{n}\right)^2\right)\\ &&\times\prod_{N_{uv}^{\sigma\tau}=2}\left(\frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right)^2 +\frac{1}{q_0}\left(1-\frac{a_\varepsilon}{n}\right)^2\right)\\ &&\times\prod_{N_{uv}^{\sigma\tau}=1}\left(\frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right) \left(\frac{b_\varepsilon}{n}\right)+\frac{1}{q_0} \left(1-\frac{a_\varepsilon}{n}\right) \left(1-\frac{b_\varepsilon}{n}\right)\right). \end{eqnarray*} It is easy to check that \begin{eqnarray}\label{expression:gamman:varepsilon} \frac{1}{p_0}\left(\frac{b_\varepsilon}{n}\right)^2 +\frac{1}{q_0}\left(1-\frac{b_\varepsilon}{n}\right)^2&=&1+\gamma_n^\varepsilon+O(n^{-3})\nonumber\\ \frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right)^2 +\frac{1}{q_0}\left(1-\frac{a_\varepsilon}{n}\right)^2&=&1+\gamma_n^\varepsilon+O(n^{-3})\nonumber\\ \frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right) \left(\frac{b_\varepsilon}{n}\right)+\frac{1}{q_0} \left(1-\frac{a_\varepsilon}{n}\right) \left(1-\frac{b_\varepsilon}{n}\right)&=&1-\gamma_n^\varepsilon+O(n^{-3}), \end{eqnarray} where $\gamma_n^\varepsilon=\frac{\kappa_\varepsilon}{n}+\frac{(a_\varepsilon-b_\varepsilon)^2}{4n^2}$, $\kappa_\varepsilon=\frac{(a_\varepsilon-b_\varepsilon)^2}{2(a+b)}$. Let \[ s_+=\#\{(u,v): u<v, \sigma_u\sigma_v\tau_u\tau_v=+\}, s_-=\#\{(u,v): u<v, \sigma_u\sigma_v\tau_u\tau_v=-\}. \] Let $\rho=\frac{1}{n}\sum_{u=1}^n\sigma_u\tau_u$. Following \cite{MNS15}, we have $s_+=\frac{n^2}{4}(1+\rho^2)-\frac{n}{2}$ and $s_-=\frac{n^2}{4}(1-\rho^2)$. Then using the approximation technique in \cite{MNS15}, i.e., Lemmas 5.3, 5.4, 5.5 therein, it holds that \begin{eqnarray*} {\mathbb{E}}_0(Y_n^\varepsilon)^2&=&4^{-n}\sum_{\sigma,\tau}(1+\gamma_n^\varepsilon+O(n^{-3}))^{s_+}(1-\gamma_n^\varepsilon+O(n^{-3}))^{s_-}\\ &=&(1+o(1))4^{-n}\sum_{\sigma,\tau}(1+\gamma_n^\varepsilon)^{\frac{n^2}{4}(1+\rho^2)-\frac{n}{2}}(1-\gamma_n^\varepsilon)^{\frac{n^2}{4}(1-\gamma^2)}\\ &=&(1+o(1))\exp\left(-\kappa_\varepsilon^2/4-\kappa_\varepsilon/2\right)4^{-n}\sum_{\sigma,\tau}\exp\left( \frac{\rho^2}{2}\left(n\kappa_\varepsilon+\frac{(a_\varepsilon-b_\varepsilon)^2}{4}\right) \right)\\ &=&(1+o(1))\exp\left(-\kappa_\varepsilon^2/4-\kappa_\varepsilon/2\right){\mathbb{E}}_{\sigma\tau} \exp\left(\frac{\rho^2}{2}\left(n\kappa_\varepsilon+\frac{(a_\varepsilon-b_\varepsilon)^2}{4}\right)\right)\\ &\overset{n\to\infty}{\rightarrow}&\exp\left(-\kappa_\varepsilon^2/4-\kappa_\varepsilon/2\right)(1-\kappa_\varepsilon)^{-1/2}= \exp\left(\sum_{m=3}^\infty\lambda_m(\delta_m^\varepsilon)^2\right). \end{eqnarray*} This verifies Condition A4. The result of Theorem \ref{testing:consistency:case2} follows from Proposition \ref{basic:prop}. \end{proof} \begin{proof}[Proof of Theorem \ref{power:case2}] Let $X_{mn}$ be the number of $m$-cycles of $G$, for $m\ge3$. Let $\lambda_m=\frac{1}{2m}\left(\frac{a+b}{2}\right)^m$ and $\delta_m=\left(\frac{a-b}{a+b}\right)^m$. It follows by Lemma \ref{lemma:short:cycles} Part \ref{lemma:short:cycles:2} that, under $H_1$, $\{X_{mn}\}_{m=3}^k$ jointly converge to independent Poisson variables with mean $\lambda_m(1+\delta_m)$, verifying Condition A1 of Proposition \ref{basic:prop}. This leaves us to check Conditions A2 to A4. Let $M=\sum_{m=3}^k mj_m$ for integers $j_3,\ldots,j_k$ and $k\ge3$. \textbf{Check Condition A2}. Denote ${\mathbb{E}}_1$ the expectation based on hypothesis $H_1$. Let $X_{mn}$ be the number of $m$-cycles of $G$, for $m\ge3$ and $[x]_j$ be the descending factorial. Define $M=\sum_{m=3}^k mj_m$ for $k\ge 3$ and integers $j_3,\ldots,j_k$. To check A2, notice that \begin{eqnarray} {\mathbb{E}}_1Y_n^\varepsilon [X_{3n}]_{j_3}\cdots[X_{kn}]_{j_k}&=&\sum_{(H_{mi})_{3\le m\le k, 1\le i\le j_m}} {\mathbb{E}}_1Y_n^\varepsilon 1_{\cup H_{mi}}\label{thm:power:case2:eqn0}\\ &=&\sum_{(H_{mi})\in A} {\mathbb{E}}_1Y_n^\varepsilon 1_{\cup H_{mi}}+\sum_{(H_{mi})\in\bar{A}} {\mathbb{E}}_1Y_n^\varepsilon 1_{\cup H_{mi}},\label{thm:power:case2:eqn-1} \end{eqnarray} where the sum in (\ref{thm:power:case2:eqn0}) ranges over $\mathcal{H}$, the collection of all $M$-tuples of cycles $(H_{mi})_{3\le m\le k, 1\le i\le j_m}$ with each $H_{mi}$ an $m$-cycle, and $A$ in the sum of (\ref{thm:power:case2:eqn-1}) is the set of such tuples for which the cycles are vertex-disjoint and let $\bar{A}=\mathcal{H}\backslash A$, i.e., $\bar{A}$ contains $M$-tuples of cycles $(H_{mi})_{3\le m\le k, 1\le i\le j_m}$ with at least one common vertex among those cycles. Let us look at the first part of (\ref{thm:power:case2:eqn-1}). Take $\tau$ uniformly distributed from $\{\pm\}^n$. For any $H=(H_{mi})\in\mathcal{H}$, define $\tau^1,\tau^2$ to be the restrictions of $\tau$ over $\mathcal{V}(H)$ and $[n]\backslash\mathcal{V}(H)$ respectively. One can check that, for any $H\in\mathcal{H}$, \begin{eqnarray} &&{\mathbb{E}}_1\{Y_n^\varepsilon 1_H|\tau\}\nonumber\\ &=&{\mathbb{E}}_1\left\{1_H 2^{-n}\sum_{\sigma}\prod_{u<v}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}} \left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}\bigg|\tau\right\}\nonumber\\ &=&2^{-n}\sum_{\sigma}{\mathbb{E}}_1\left\{1_H\prod_{u<v}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}}\left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}\bigg|\tau\right\}\nonumber\\ &=&2^{-n}\sum_{\sigma}{\mathbb{E}}_1\left\{1_H\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}}\left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}} \bigg|\tau\right\}\nonumber\\ &&\times\prod_{(u,v)\in\overline{\mathcal{E}(H)}}{\mathbb{E}}_1\left\{\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)^{A_{uv}}\left(\frac{q_{uv}^\varepsilon(\sigma)}{q_0}\right)^{1-A_{uv}}\bigg|\tau\right\}\nonumber\\ &=&2^{-n}\sum_{\sigma}{\mathbb{E}}_1\{1_H|\tau\}\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma)}{p_0}\right)\prod_{(u,v)\in\overline{\mathcal{E}(H)}} \left(\frac{p_{uv}^\varepsilon(\sigma)p_{uv}(\tau)}{p_0}+\frac{q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)}{q_0}\right)\nonumber\\ &=&2^{-n}\sum_{\sigma}\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma^1)p_{uv}(\tau^1)}{p_0}\right) \prod_{(u,v)\in\overline{\mathcal{E}(H)}} \left(\frac{p_{uv}^\varepsilon(\sigma)p_{uv}(\tau)}{p_0}+\frac{q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)}{q_0}\right),\label{thm:power:case1:eqn24} \end{eqnarray} which leads to that \begin{eqnarray} &&{\mathbb{E}}_1 Y_n^\varepsilon 1_H\nonumber\\ &=&2^{-2n}\sum_{\tau}\sum_{\sigma}\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma^1)p_{uv}(\tau^1)}{p_0}\right) \prod_{(u,v)\in\overline{\mathcal{E}(H)}} \left(\frac{p_{uv}^\varepsilon(\sigma)p_{uv}(\tau)}{p_0}+\frac{q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)}{q_0}\right)\nonumber\\ &=&{\mathbb{E}}_{\sigma\tau}\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{p_{uv}^\varepsilon(\sigma^1)p_{uv}(\tau^1)}{p_0}\right) \prod_{(u,v)\in\overline{\mathcal{E}(H)}} \left(\frac{p_{uv}^\varepsilon(\sigma)p_{uv}(\tau)}{p_0}+\frac{q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)}{q_0}\right)\nonumber\\ &\equiv&{\mathbb{E}}_{\sigma\tau}X_H^\varepsilon(\sigma^1,\tau^1)W_H^\varepsilon(\sigma,\tau)Z_H^\varepsilon(\sigma^2,\tau^2),\label{thm:power:case2:eqn2} \end{eqnarray} where \[ X_H^\varepsilon(\sigma^1,\tau^1)=\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{1}{p_0}p_{uv}^\varepsilon(\sigma^1)p_{uv}(\tau^1)\right), \] \[ W_H^\varepsilon(\sigma,\tau)=\prod_{(u,v)\in S_1(H)} \left(\frac{1}{p_0}p_{uv}^\varepsilon(\sigma)p_{uv}(\tau)+\frac{1}{q_0}q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)\right) \] and \[ Z_H^\varepsilon(\sigma^2,\tau^2)=\prod_{(u,v)\in S_2(H)} \left(\frac{1}{p_0}p_{uv}^\varepsilon(\sigma^2)p_{uv}(\tau^2)+\frac{1}{q_0}q_{uv}^\varepsilon(\sigma^2)q_{uv}(\tau^2)\right). \] Here $S_1(H)=\{(u,v)\in\overline{\mathcal{E}(H)}:\textrm{$u\in\mathcal{V}(H)$ or $v\in\mathcal{V}(H)$}\}$ and $S_2(H)=\{(u,v)\in\overline{\mathcal{E}(H)}:u,v\notin\mathcal{V}(H)\}$. We will show that $W_H^\varepsilon(\sigma,\tau)$ is uniformly bounded over $\sigma,\tau,H$, and that \begin{equation}\label{thm:power:case2:eqn3} \sup_{H\in \mathcal{H}}|W_H^\varepsilon(\sigma,\tau)-1|\to0,\,\,a.s. \end{equation} To see this, observe that \begin{eqnarray}\label{thm:power:case2:eqn4} W_H^\varepsilon(\sigma,\tau)&=&\prod_{(u,v)\in\overline{\mathcal{E}(H)}, u,v\in\mathcal{V}(H)}\left(\frac{1}{p_0} p_{uv}^\varepsilon(\sigma)p_{uv}(\tau)+\frac{1}{q_0}q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)\right)\nonumber\\ &&\times\prod_{v\in\mathcal{V}(H)}\prod_{u\notin\mathcal{V}(H)} \left(\frac{1}{p_0}p_{uv}^\varepsilon(\sigma)p_{uv}(\tau)+\frac{1}{q_0}q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)\right). \end{eqnarray} We note that \[ \frac{1}{p_0} p_{uv}^\varepsilon(\sigma)p_{uv}(\tau)+\frac{1}{q_0}q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)= 1+O(n^{-1}), \] where the $O(n^{-1})$ term is uniform for $u,v,\sigma,\tau,H$. The first product in (\ref{thm:power:case2:eqn4}) is therefore equal to $(1+O(n^{-1}))^{{M\choose2}-M}=1+o(1)$. We turn to the second product in (\ref{thm:power:case2:eqn4}). For any $v\in\mathcal{V}$, let \begin{eqnarray*} S_v^1&=&\#\{u\notin\mathcal{V}(H): \sigma_u^2=\sigma_v^1, \tau_u^2=\tau_v^1\}\\ S_v^2&=&\#\{u\notin\mathcal{V}(H): \sigma_u^2=\sigma_v^1, \tau_u^2\neq\tau_v^1\}\\ S_v^3&=&\#\{u\notin\mathcal{V}(H): \sigma_u^2\neq\sigma_v^1, \tau_u^2=\tau_v^1\}\\ S_v^4&=&\#\{u\notin\mathcal{V}(H): \sigma_u^2\neq\sigma_v^1, \tau_u^2\neq\tau_v^1\}. \end{eqnarray*} Also let $S_{ll'}=\#\{u\notin\mathcal{V}(H): \sigma_u^2=l,\tau_u^2=l'\}$ and $N_{ll'}=\{v\in\mathcal{V}(H): \sigma_v^1=l, \tau_v^1=l'\}$ for $l,l'=\pm$. Then the second product in (\ref{thm:power:case2:eqn4}) equals to \begin{eqnarray*} &&\prod_{v\in\mathcal{V}(H)}\prod_{u\notin\mathcal{V}(H)} \left(\frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right)^{1(\sigma_u^2=\sigma_v^1)} \left(\frac{b_\varepsilon}{n}\right)^{1(\sigma_u^2\neq\sigma_v^1)} \left(\frac{a}{n}\right)^{1(\tau_u^2=\tau_v^1)}\left(\frac{b}{n}\right)^{1(\tau_u^2\neq\tau_v^1)}\right.\\ &&\left.+\frac{1}{q_0}\left(1-\frac{a_\varepsilon}{n}\right)^{1(\sigma_u^2=\sigma_v^1)} \left(1-\frac{b_\varepsilon}{n}\right)^{1(\sigma_u^2\neq\sigma_v^1)} \left(1-\frac{a}{n}\right)^{1(\tau_u^2=\tau_v^1)}\left(1-\frac{b}{n}\right)^{1(\tau_u^2\neq\tau_v^1)}\right)\\ &=& \prod_{v\in\mathcal{V}(H)} \left(\frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right) \left(\frac{a}{n}\right)+\frac{1}{q_0}\left(1-\frac{a_\varepsilon}{n}\right)\left(1-\frac{a}{n}\right)\right)^{S_v^1}\\ &&\times\left(\frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right)\left(\frac{b}{n}\right)+\frac{1}{q_0} \left(1-\frac{a_\varepsilon}{n}\right)\left(1-\frac{b}{n}\right)\right)^{S_v^2}\\ &&\times\left(\frac{1}{p_0}\left(\frac{b_\varepsilon}{n}\right)\left(\frac{a}{n}\right)+\frac{1}{q_0} \left(1-\frac{b_\varepsilon}{n}\right)\left(1-\frac{a}{n}\right)\right)^{S_v^3}\\ &&\times\left(\frac{1}{p_0}\left(\frac{b_\varepsilon}{n}\right)\left(\frac{b}{n}\right)+\frac{1}{q_0} \left(1-\frac{b_\varepsilon}{n}\right)\left(1-\frac{b}{n}\right)\right)^{S_v^4}\\ &=&(1+\widetilde{\gamma}_n^\varepsilon+O(n^{-3}))^{\sum_{v\in\mathcal{V}(H)}(S_v^1+S_v^4)} (1-\widetilde{\gamma}_n^\varepsilon+O(n^{-3}))^{\sum_{v\in\mathcal{V}(H)}(S_v^2+S_v^3)}. \end{eqnarray*} In the above we have used the following trivial facts: \begin{eqnarray*} \frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right) \left(\frac{a}{n}\right)+\frac{1}{q_0}\left(1-\frac{a_\varepsilon}{n}\right)\left(1-\frac{a}{n}\right)&=&1+\widetilde{\gamma}_n^\varepsilon+O(n^{-3})\\ \frac{1}{p_0}\left(\frac{a_\varepsilon}{n}\right)\left(\frac{b}{n}\right)+\frac{1}{q_0} \left(1-\frac{a_\varepsilon}{n}\right)\left(1-\frac{b}{n}\right)&=&1-\widetilde{\gamma}_n^\varepsilon+O(n^{-3})\\ \frac{1}{p_0}\left(\frac{b_\varepsilon}{n}\right)\left(\frac{a}{n}\right)+\frac{1}{q_0} \left(1-\frac{b_\varepsilon}{n}\right)\left(1-\frac{a}{n}\right)&=&1-\widetilde{\gamma}_n^\varepsilon+O(n^{-3})\\ \frac{1}{p_0}\left(\frac{b_\varepsilon}{n}\right)\left(\frac{b}{n}\right)+\frac{1}{q_0} \left(1-\frac{b_\varepsilon}{n}\right)\left(1-\frac{b}{n}\right)&=&1+\widetilde{\gamma}_n^\varepsilon+O(n^{-3}), \end{eqnarray*} where $\widetilde{\gamma}_n^\varepsilon=\frac{\widetilde{\kappa}_\varepsilon}{n}+\frac{(a-b)(a_\varepsilon-b_\varepsilon)}{4n^2}$ and $\widetilde{\kappa}_\varepsilon=\frac{(a-b)(a_\varepsilon-b_\varepsilon)}{2(a+b)}$. Note that \begin{eqnarray*} \sum_{v\in\mathcal{V}(H)}(S_v^1+S_v^4)&=&\sum_{\sigma_v^1=+,\tau_v^1=+}(S_{++}+S_{--}) +\sum_{\sigma_v^1=+,\tau_v^1=-}(S_{+-}+S_{-+})\\ &&+\sum_{\sigma_v^1=-,\tau_v^1=+}(S_{-+}+S_{+-}) +\sum_{\sigma_v^1=-,\tau_v^1=-}(S_{--}+S_{++})\\ &=&(S_{++}+S_{--})(N_{++}+N_{--})+(S_{+-}+S_{-+})(N_{+-}+N_{-+})\equiv N_1, \end{eqnarray*} similarly, \[ \sum_{v\in\mathcal{V}(H)}(S_v^2+S_v^3)=(S_{+-}+S_{-+})(N_{++}+N_{--})+(S_{++}+S_{--})(N_{+-}+N_{-+})\equiv N_2. \] So the second product in (\ref{thm:power:case2:eqn4}) equals to \begin{eqnarray*} (1+o(1))(1+\widetilde{\gamma}_n^\varepsilon)^{N_1}(1-\widetilde{\gamma}_n^\varepsilon)^{N_2}=(1+o(1))\exp\left(\frac{N_1-N_2}{n}\widetilde{\kappa}_\varepsilon\right), \end{eqnarray*} where the $o(1)$ term is uniform for $u,v,\sigma,\tau,H$, thanks to $N_1, N_2\le Mn$. By law of large number, $(N_1-N_2)/n\to0$, a.s., uniformly for $H\in\mathcal{H}$. Therefore (\ref{thm:power:case2:eqn3}) holds. The above analysis also shows that $W_H^\varepsilon(\sigma,\tau)$ is uniformly bounded over $\sigma,\tau,H$. Next let us analyze the term $Z_H^\varepsilon(\sigma^2,\tau^2)$. By Taylor expansions and direct examinations it can be checked that for $u,v\in[n]\backslash\mathcal{V}(H)$, \[ \frac{1}{p_0}p_{uv}^\varepsilon(\sigma^2)p_{uv}(\tau^2)+\frac{1}{q_0}q_{uv}^\varepsilon(\sigma^2)q_{uv}(\tau^2) =\left\{\begin{array}{cc} 1+\widetilde{\gamma}_n^\varepsilon+O(n^{-3}),&\textrm{if $\sigma_u^2\sigma_v^2\tau_u^2\tau_v^2=+$}\\ 1-\widetilde{\gamma}_n^\varepsilon+O(n^{-3}),&\textrm{if $\sigma_u^2\sigma_v^2\tau_u^2\tau_v^2=-$}. \end{array}\right. \] Let $s_+=\#\{(u,v): u,v\in[n]\backslash\mathcal{V}(H), u<v, \sigma_u^2\sigma_v^2\tau_u^2\tau_v^2=+\}$ and $s_-=\#\{(u,v): u,v\in[n]\backslash\mathcal{V}(H), u<v, \sigma_u^2\sigma_v^2\tau_u^2\tau_v^2=-\}$. Let $\rho=\rho(\sigma^2,\tau^2)=\frac{1}{n-M}\sum_{u\in [n]\backslash\mathcal{V}(H)}\sigma_u^2\tau_u^2$. By direct examinations we have \begin{eqnarray} Z_H^\varepsilon(\sigma^2,\tau^2) &=&(1+\widetilde{\gamma}_n^\varepsilon+O(n^{-3}))^{s_+}(1-\widetilde{\gamma}_n^\varepsilon+O(n^{-3}))^{s_-}\nonumber\\ &=&(1+o(1))(1+\widetilde{\gamma}_n^\varepsilon)^{s_+}(1-\widetilde{\gamma}_n^\varepsilon)^{s_-}\nonumber\\ &=&(1+o(1))\exp\left(-\frac{\widetilde{\kappa}_\varepsilon^2(n-M)^2}{4n^2}-\frac{\widetilde{\kappa}_\varepsilon(n-M)}{2n}\right)\nonumber\\ &&\times\exp\left(\frac{(\sqrt{n-M}\rho)^2}{2}\left(\frac{(a-b)(a_\varepsilon-b_\varepsilon)(n-M)}{4n^2}+\frac{\widetilde{\kappa}_\varepsilon(n-M)}{n}\right)\right).\label{thm:power:case1:eqn5} \end{eqnarray} By the condition $(a-b)(a_\varepsilon-b_\varepsilon)<2(a+b)/3$, $\widetilde{\kappa}_\varepsilon<1$. Let $Z_n=\sqrt{n-M}\rho$. Let $\kappa_n=\frac{(a-b)(a_\varepsilon-b_\varepsilon)(n-M)}{4n^2}+\frac{\widetilde{\kappa}_\varepsilon(n-M)}{n}$ which is nonrandom tending to $\widetilde{\kappa}_\varepsilon$. By Hoeffding's inequality: for any $C>0$, \begin{eqnarray}\label{thm:power:case1:eqn8} P\left(\exp(\kappa_n Z_n^2/2)\ge C\right)\le2C^{-1/\kappa_n}. \end{eqnarray} From (\ref{thm:power:case1:eqn5}) there exists a universal constant $C_0$ such that $Z_H^\varepsilon(\sigma^2,\tau^2)\le C_0\exp(\kappa_n Z_n^2/2)$, hence, it follows from (\ref{thm:power:case1:eqn8}) that for all $C>0$, \[ P\left(Z_H^\varepsilon(\sigma^2,\tau^2)\ge C\right)\le2(C/C_0)^{-1/\kappa_n}. \] Therefore, by (\ref{thm:power:case1:eqn8}) we have that \begin{eqnarray} &&{\mathbb{E}}_{\sigma^2\tau^2}Z_H^\varepsilon(\sigma^2,\tau^2)1(Z_H^\varepsilon(\sigma^2,\tau^2)\ge C)\nonumber\\ &=&\int_0^\infty P\left(Z_H^\varepsilon(\sigma^2,\tau^2)1(Z_H^\varepsilon(\sigma^2,\tau^2)\ge C)>t\right)dt\nonumber\\ &=&CP\left(Z_H^\varepsilon(\sigma^2,\tau^2)\ge C\right)+\int_C^\infty P\left(Z_H^\varepsilon(\sigma^2,\tau^2)>t\right)dt\nonumber\\ &\le&2C_0^{1/\kappa_n}C^{1-1/\kappa_n}/(1-\kappa_n).\label{thm:power:case1:eqn9} \end{eqnarray} We can also show that, as $n\to\infty$, \begin{equation}\label{thm:power:case2:eqn5} \sup_{H\in\mathcal{H}}|{\mathbb{E}}_{\sigma^2\tau^2}Z_H^\varepsilon(\sigma^2,\tau^2)-\exp\left(-\widetilde{\kappa}_\varepsilon^2/4-\widetilde{\kappa}_\varepsilon/2\right)(1-\widetilde{\kappa}_\varepsilon)^{-1/2}|\to0, n\to\infty. \end{equation} To see this, let $\rho_0=\frac{1}{n-M}\sum_{u\in[n]}\sigma_u\tau_u$ and $r_H=\frac{1}{n-M}\sum_{u\in\mathcal{V}(H)}\sigma_u\tau_u$, therefore, $\rho=\rho_0-r_H$. Let $Z_{0n}=\sqrt{n-M}\rho_0$. Then for any $H\in \mathcal{H}$, $|r_H|\le M/(n-M)$ which leads to \[ Z_{0n}^2-2M|\rho_0|\le Z_n^2\le Z_{0n}^2-2M|\rho_0|+M^2/(n-M). \] Both left and right hand sides in the above are free of $H$ and converge to $\chi_1^2$ thanks to $\rho_0\to0$, a.s. So \[ \sup_{H\in \mathcal{H}}|{\mathbb{E}}\exp(\kappa_n Z_n^2/2)-(1-\widetilde{\kappa}_{\varepsilon})^{-1/2}|\to0,\,\,n\to\infty. \] This, together with (\ref{thm:power:case1:eqn5}), prove (\ref{thm:power:case2:eqn5}). Next let us analyze $X_H^\varepsilon(\sigma^1,\tau^1)$. Assume $H=(H_{mi})_{3\le m\le k, 1\le i\le j_m}\in A$. For $3\le m\le k$ and $1\le i\le j_m$, let $\tau^{1mi}$, $\sigma^{1mi}$ be the restrictions of $\tau^1$, $\sigma^1$ over the vertexes of $H_{mi}$. Since $H_{mi}$ are vertex-disjoint, $\tau^{1mi}$'s, $\sigma^{1mi}$'s are all independent. Let $N_{mi}=\sum_{(u,v)\in\mathcal{E}(H_{mi})}1(\sigma_u^{1mi}\neq\sigma^{1mi}_v)$, the number of edges over $H_{mi}$ with distinct end points. Following the proof of \cite[Lemma 3.3]{MNS15}, we get that \begin{eqnarray} &&{\mathbb{E}}_{\sigma^1\tau^1}X_H^\varepsilon(\sigma^1,\tau^1)\nonumber\\ &=&{\mathbb{E}}_{\sigma^1\tau^1}\prod_{(u,v)\in\mathcal{E}(H)}\left(\frac{1}{p_0}p_{uv}^\varepsilon(\sigma^1)p_{uv}(\tau^1)\right)\nonumber\\ &=&p_0^{-M}\prod_{m=3}^k\prod_{i=1}^{j_m}{\mathbb{E}}_{\sigma^{1}}\prod_{(u,v)\in\mathcal{E}(H_{mi})}p_{uv}^\varepsilon(\sigma^{1mi})\times{\mathbb{E}}_{\tau^1} \prod_{(u,v)\in\mathcal{E}(H_{mi})}p_{uv}(\tau^{1mi})\nonumber\\ &=&p_0^{-M}n^{-2M}\prod_{m=3}^k\prod_{i=1}^{j_m}{\mathbb{E}}_{\sigma^{1mi}}a_\varepsilon^{m-N_{mi}}b_\varepsilon^{N_{mi}}\times {\mathbb{E}}_{\tau^{1mi}}a^{m-N_{mi}}b^{N_{mi}}\nonumber\\ &=&p_0^{-M}n^{-2M}\prod_{m=3}^k\prod_{i=1}^{j_m}2^{-m}\left[(a_\varepsilon+b_\varepsilon)^m+(a_\varepsilon-b_\varepsilon)^m\right] \times 2^{-m}\left[(a+b)^m+(a-b)^m\right]\nonumber\\ &=&n^{-M}\prod_{m=3}^k\left[\left(\frac{a+b}{2}\right)^m(1+\delta_m)(1+\delta_m^\varepsilon)\right]^{j_m},\label{case2:XHe} \end{eqnarray} recalling $\delta_m=\left(\frac{a-b}{a+b}\right)^m$ and $\delta_m^\varepsilon=\left(\frac{a_\varepsilon-b_\varepsilon}{a+b}\right)^m$. Meanwhile, it is easy to see that $n^MX_H^\varepsilon(\sigma^1,\tau^1)$ is almost surely bounded and the bound is unrelated to the vertexes of $H$, i.e., \begin{equation}\label{thm:power:case1:eqn11} n^M X_H^\varepsilon(\sigma^1,\tau^1)\le\left(\frac{2a_\varepsilon a}{a+b}\right)^M,\,\,\,\,\forall\sigma^1,\tau^1\in\{\pm\}^{\mathcal{V}(H)}. \end{equation} By (\ref{thm:power:case2:eqn3}), (\ref{thm:power:case1:eqn9}), (\ref{thm:power:case1:eqn11}), and bounded convergence theorem, we can show that \begin{equation}\label{thm:power:case1:eqn10} \sum_{H\in A}{\mathbb{E}}_{\sigma\tau}X_H^\varepsilon(\sigma^1,\tau^1)|W_H^\varepsilon(\sigma,\tau)-1|Z_H^\varepsilon(\sigma^2,\tau^2)\to0. \end{equation} More precisely, using $|A|=\frac{n!}{(n-M)!}\prod_{m=3}^k\left(\frac{1}{2m}\right)^{j_m}$ (see proof of Lemma \ref{lemma:short:cycles}), (\ref{thm:power:case1:eqn10}) follows from the following \begin{eqnarray*} &&\sum_{H\in A}{\mathbb{E}}_{\sigma\tau}X_H^\varepsilon(\sigma^1,\tau^1)|W_H^\varepsilon(\sigma,\tau)-1|Z_H^\varepsilon(\sigma^2,\tau^2)\\ &=&\sum_{H\in A}{\mathbb{E}}_{\sigma\tau}X_H^\varepsilon(\sigma^1,\tau^1)|W_H^\varepsilon(\sigma,\tau)-1|Z_H^\varepsilon(\sigma^2,\tau^2)1(Z_H^\varepsilon(\sigma^2,\tau^2)\le C)\\ &&+ \sum_{H\in A}{\mathbb{E}}_{\sigma\tau}X_H^\varepsilon(\sigma^1,\tau^1)|W_H^\varepsilon(\sigma,\tau)-1|Z_H^\varepsilon(\sigma^2,\tau^2)1(Z_H(\sigma^2,\tau^2)>C)\\ &\lesssim&Cn^{-M}|A|{\mathbb{E}}_{\sigma\tau}\sup_{H\in A}|W_H^\varepsilon(\sigma,\tau)-1|+n^{-M}|A|\sup_{H\in A}{\mathbb{E}}_{\sigma\tau}Z_H^\varepsilon(\sigma^2,\tau^2)1(Z_H^\varepsilon(\sigma^2,\tau^2)>C)\\ &\to&0, \end{eqnarray*} where the last limit follows by first taking $C\to\infty$ and then $n\to\infty$. By (\ref{thm:power:case2:eqn2}), (\ref{thm:power:case2:eqn5}) and (\ref{case2:XHe}), we have that \begin{eqnarray*} &&\sum_{H\in A}{\mathbb{E}}_1 Y_n^\varepsilon 1_H\\ &=&|A|n^{-M}\prod_{m=3}^k\left[\left(\frac{a+b}{2}\right)^m(1+\delta_m^\varepsilon)(1+\delta_m)\right]^{j_m}\exp(-\widetilde{\kappa}_\varepsilon^2/4-\widetilde{\kappa}_\varepsilon/2)/\sqrt{1-\widetilde{\kappa}_\varepsilon}+o(1)\\ &\overset{n\to\infty}{\rightarrow}& \prod_{m=3}^k(\lambda_m(1+\delta_m^\varepsilon)(1+\delta_m))^{j_m}\exp(-\widetilde{\kappa}_\varepsilon^2/4-\widetilde{\kappa}_\varepsilon/2)/\sqrt{1-\widetilde{\kappa}_\varepsilon}, \end{eqnarray*} recalling $\lambda_m=\frac{1}{2m}\left(\frac{a+b}{2}\right)^m$. From (\ref{thm:power:case2:eqn2}), the uniform boundedness of ${\mathbb{E}}_{\sigma^2\tau^2}Z_H^\varepsilon(\sigma^2,\tau^2)$ and the uniform boundedness of $W_H^\varepsilon(\sigma,\tau)$, and the independence of $\sigma^1,\tau^1,\sigma^2,\tau^2$ that, there exists a constant $C_1$ s.t. for any $H\in\bar{A}$, \begin{eqnarray*} {\mathbb{E}}_1Y_n^\varepsilon 1_H\le C_1{\mathbb{E}}_{\sigma^1\tau^1}X_H^\varepsilon(\sigma^1,\tau^1). \end{eqnarray*} Also notice from the definition of $X_H^\varepsilon$ that \begin{eqnarray*} X_H^\varepsilon(\sigma^1,\tau^1)&=&p_0^{-|\mathcal{E}(H)|}\prod_{(u,v)\in\mathcal{E}(H)} \left(\frac{a_\varepsilon}{n}\right)^{1(\sigma_u^1=\sigma_v^1)+1(\tau_u^1=\tau_v^1)}\left(\frac{b_\varepsilon}{n}\right)^{1(\sigma_u^1\neq\sigma_v^1)+1(\tau_u^1\neq\tau_v^1)}\\ &\le&n^{-2|\mathcal{E}(H)|}p_0^{-|\mathcal{E}(H)|}a_\varepsilon^{2|\mathcal{E}(H)|}=n^{-|\mathcal{E}(H)|}\left(\frac{2a_\varepsilon^2}{a+b}\right)^{|\mathcal{E}(H)|}. \end{eqnarray*} Since there are at most ${n\choose |\mathcal{V}(H)|}|\mathcal{V}(H)|!$ graphs isomorphic to $H$, and $|\mathcal{E}(H)|>|\mathcal{V}(H)|$ for $H\in\bar{A}$, we get that, as $n\to\infty$, \begin{eqnarray*} \sum_{\textrm{$H'$ is isomorphic to $H$}}{\mathbb{E}}_1 Y_n^\varepsilon 1_H\le C_1\left(\frac{2a_\varepsilon^2}{a+b}\right)^{|\mathcal{E}(H)|}n^{-|\mathcal{E}(H)|}{n\choose |\mathcal{V}(H)|}|\mathcal{V}(H)|!\to0. \end{eqnarray*} Since there is a bounded number of isomorphism classes, we get that the second part of (\ref{thm:power:case2:eqn-1}) tends to zero as $n\to\infty$. Hence, as $n\to\infty$, \[ \textrm{(\ref{thm:power:case2:eqn0}) $\to\prod_{m=3}^k(\lambda_m(1+\delta_m^\varepsilon)(1+\delta_m))^{j_m}\exp(-\widetilde{\kappa}_\varepsilon^2/4-\widetilde{\kappa}_\varepsilon/2)/\sqrt{1-\widetilde{\kappa}_\varepsilon}$.} \] As for ${\mathbb{E}}_1 Y_n^\varepsilon$, note that it is equal to \[ {\mathbb{E}}_1 Y_n^\varepsilon=4^{-n}\sum_{\sigma,\tau}\prod_{u<v}\left(\frac{1}{p_0}p_{uv}^\varepsilon(\sigma)p_{uv}(\tau) +\frac{1}{q_0}q_{uv}^\varepsilon(\sigma)q_{uv}(\tau)\right). \] Similar to (\ref{thm:power:case2:eqn5}), i.e., taking $H$ therein as empty graph, one gets that \begin{equation}\label{thm:power:case2:eqn7} {\mathbb{E}}_1 Y_n^\varepsilon\to\exp(-\widetilde{\kappa}_\varepsilon^2/4-\widetilde{\kappa}_\varepsilon/2)/\sqrt{1-\widetilde{\kappa}_\varepsilon}. \end{equation} Hence, \[ \frac{{\mathbb{E}}_1Y_n^\varepsilon [X_{3n}]_{j_3}\cdots[X_{kn}]_{j_k}}{{\mathbb{E}}_1 Y_n^\varepsilon}\overset{n\to\infty}{\rightarrow}\prod_{m=3}^k(\lambda_m(1+\delta_m^\varepsilon)(1+\delta_m))^{j_m}. \] This verifies Condition A2. \textbf{Check Condition A3}. Since $\frac{\lambda_m(1+\delta_m^\varepsilon)(1+\delta_m)}{\lambda_m(1+\delta_m)}-1=\delta_m^\varepsilon$, and by (\ref{vareps:condition}), we have \[ \sum_{m\ge3}\lambda_m(1+\delta_m)(\delta_m^\varepsilon)^2=\sum_{m=3}^\infty\frac{1}{2m}\left(\frac{(a_\varepsilon-b_\varepsilon)^2}{2(a+b)}\right)^m +\sum_{m=3}^\infty\frac{1}{2m}\left(\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{2(a+b)^2}\right)^m<\infty. \] \textbf{Check Condition A4}. By direct examinations it can be checked that \begin{eqnarray*} {\mathbb{E}}_1\{(Y_n^\varepsilon)^2|\tau\}&=&4^{-n}\sum_{\sigma}\sum_{\eta}{\mathbb{E}}_1\left\{\prod_{u<v} \left(\frac{p_{uv}^\varepsilon(\sigma)p_{uv}^\varepsilon(\eta)}{p_0^2}\right)^{A_{uv}} \left(\frac{q_{uv}^\varepsilon(\sigma)q_{uv}^\varepsilon(\eta)}{q_0^2}\right)^{1-A_{uv}}\bigg|\tau\right\}\\ &=&4^{-n}\sum_{\sigma}\sum_{\eta}\prod_{u<v}\left(\frac{1}{p_0^2}p_{uv}^\varepsilon(\sigma)p_{uv}^\varepsilon(\eta)p_{uv}(\tau)+\frac{1}{q_0^2}q_{uv}^\varepsilon(\sigma) q_{uv}^\varepsilon(\eta)q_{uv}(\tau)\right) \end{eqnarray*} So \begin{eqnarray}\label{thm:power:case1:eqn26} {\mathbb{E}}_1(Y_n^\varepsilon)^2&=&8^{-n}\sum_{\sigma}\sum_{\eta}\sum_{\tau} \prod_{u<v}\left(\frac{1}{p_0^2}p_{uv}^\varepsilon(\sigma)p_{uv}^\varepsilon(\eta)p_{uv}(\tau)+ \frac{1}{q_0^2}q_{uv}^\varepsilon(\sigma)q_{uv}^\varepsilon(\eta)q_{uv}(\tau)\right)\nonumber\\ &=&{\mathbb{E}}_{\sigma\eta\tau}\prod_{u<v}\left(\frac{1}{p_0^2}p_{uv}^\varepsilon(\sigma)p_{uv}^\varepsilon(\eta)p_{uv}(\tau)+ \frac{1}{q_0^2}q_{uv}^\varepsilon(\sigma)q_{uv}^\varepsilon(\eta)q_{uv}(\tau)\right), \end{eqnarray} where $\sigma,\eta,\tau$ in the above expectation are independent and uniformly distributed over $\{\pm\}^n$. By Taylor expansion and straightforward (but exhaustive) calculations, it can be shown that \begin{eqnarray}\label{def:gamma2+:gamma0-} \frac{1}{p_0^2}\left(\frac{a_\varepsilon}{n}\right)^2\left(\frac{a}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{a_\varepsilon}{n}\right)^2\left(1-\frac{a}{n}\right)&=&1+\gamma_{2+}+O(n^{-3})\nonumber\\ \frac{1}{p_0^2}\left(\frac{a_\varepsilon}{n}\right)^2\left(\frac{b}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{a_\varepsilon}{n}\right)^2\left(1-\frac{b}{n}\right)&=&1+\gamma_{2-}+O(n^{-3})\nonumber\\ \frac{1}{p_0^2}\left(\frac{a_\varepsilon}{n}\right)\left(\frac{b_\varepsilon}{n}\right)\left(\frac{a}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{a_\varepsilon}{n}\right)\left(1-\frac{b_\varepsilon}{n}\right)\left(1-\frac{a}{n}\right)&=&1+\gamma_{1+}+O(n^{-3})\nonumber\\ \frac{1}{p_0^2}\left(\frac{a_\varepsilon}{n}\right)\left(\frac{b_\varepsilon}{n}\right)\left(\frac{b}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{a_\varepsilon}{n}\right)\left(1-\frac{b_\varepsilon}{n}\right)\left(1-\frac{b}{n}\right)&=&1+\gamma_{1-}+O(n^{-3})\nonumber\\ \frac{1}{p_0^2}\left(\frac{b_\varepsilon}{n}\right)^2\left(\frac{a}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{b_\varepsilon}{n}\right)^2\left(1-\frac{a}{n}\right)&=&1+\gamma_{0+}+O(n^{-3})\nonumber\\ \frac{1}{p_0^2}\left(\frac{b_\varepsilon}{n}\right)^2\left(\frac{b}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{b_\varepsilon}{n}\right)^2\left(1-\frac{b}{n}\right)&=&1+\gamma_{0-}+O(n^{-3}) \end{eqnarray} where \begin{eqnarray*} \gamma_{2+}&=&\frac{(a_\varepsilon-b_\varepsilon)^2(2a_\varepsilon+b_\varepsilon)+x_\varepsilon}{n(a+b)^2}+ \frac{3(a_\varepsilon-b_\varepsilon)^2+y_\varepsilon}{4n^2}\\ \gamma_{2-}&=&-\frac{a_\varepsilon(a_\varepsilon-b_\varepsilon)^2+x_\varepsilon}{n(a+b)^2}-\frac{(a_\varepsilon-b_\varepsilon)^2+y_\varepsilon}{4n^2}\\ \gamma_{1+}&=&-\frac{a_\varepsilon(a_\varepsilon-b_\varepsilon)^2+z_\varepsilon}{n(a+b)^2}-\frac{(a_\varepsilon-b_\varepsilon)^2}{4n^2}\\ \gamma_{1-}&=&-\frac{b_\varepsilon(a_\varepsilon-b_\varepsilon)^2-z_\varepsilon}{n(a+b)^2}-\frac{(a_\varepsilon-b_\varepsilon)^2}{4n^2}\\ \gamma_{0+}&=&-\frac{b_\varepsilon(a_\varepsilon-b_\varepsilon)^2+w_\varepsilon}{n(a+b)^2}-\frac{(a_\varepsilon-b_\varepsilon)^2+y_\varepsilon}{4n^2}\\ \gamma_{0-}&=&\frac{(a_\varepsilon-b_\varepsilon)^2(a_\varepsilon+2b_\varepsilon)+w_\varepsilon}{n(a+b)^2}+\frac{3(a_\varepsilon-b_\varepsilon)^2+y_\varepsilon}{4n^2} \end{eqnarray*} with \[ x_\varepsilon=\varepsilon(a_\varepsilon-b_\varepsilon)(3a_\varepsilon+b_\varepsilon), y_\varepsilon=4\varepsilon(a_\varepsilon-b_\varepsilon), z_\varepsilon=\varepsilon(a_\varepsilon-b_\varepsilon)^2, w_\varepsilon=\varepsilon(a_\varepsilon-b_\varepsilon)(a_\varepsilon+3b_\varepsilon). \] Define $s_{r+}=\#\{(u,v): u<v, N_{uv}^{\sigma\tau}=r,\tau_u\tau_v=+\}$ and $s_{r-}=\#\{(u,v): u<v, N_{uv}^{\sigma\tau}=r,\tau_u\tau_v=-\}$, for $r=0,1,2$. Then it holds that \begin{eqnarray}\label{thm:power:case2:eqn6} {\mathbb{E}}_1(Y_n^\varepsilon)^2&=&{\mathbb{E}}_{\sigma\eta\tau}\prod_{r=0,1,2}(1+\gamma_{r+}+O(n^{-3}))^{s_{r+}} \times\prod_{r=0,1,2}(1+\gamma_{r-}+O(n^{-3}))^{s_{r-}}\nonumber\\ &=&(1+o(1)){\mathbb{E}}_{\sigma\eta\tau}\prod_{r=0,1,2}(1+\gamma_{r+})^{s_{r+}} \times\prod_{r=0,1,2}(1+\gamma_{r-})^{s_{r-}}. \end{eqnarray} Define \begin{eqnarray}\label{def:rhos} \rho_1&=&\frac{1}{\sqrt{n}}\sum_{u=1}^n\sigma_u,\rho_2=\frac{1}{\sqrt{n}}\sum_{u=1}^n\eta_u,\rho_3=\frac{1}{\sqrt{n}}\sum_{u=1}^n\tau_u,\nonumber\\ \rho_4&=&\frac{1}{\sqrt{n}}\sum_{u=1}^n\sigma_u\eta_u,\rho_5=\frac{1}{\sqrt{n}}\sum_{u=1}^n\sigma_u\tau_u, \rho_6=\frac{1}{\sqrt{n}}\sum_{u=1}^n\eta_u\tau_u, \rho_7=\frac{1}{\sqrt{n}}\sum_{u=1}^n\sigma_u\eta_u\tau_u. \end{eqnarray} Observe that \begin{eqnarray*} s_{2+}&=&\sum_{u<v}1(\sigma_u\sigma_v=+)1(\eta_u\eta_v=+)1(\tau_u\tau_v=+)\\ &=&\frac{n^2}{16}-\frac{n}{2}+\frac{n}{16}\left(\rho_1^2+\rho_2^2+\rho_3^2+\rho_4^2+\rho_5^2+\rho_6^2+\rho_7^2\right)\\ s_{2-}&=&\sum_{u<v}1(\sigma_u\sigma_v=+)1(\eta_u\eta_v=+)1(\tau_u\tau_v=-)\\ &=&\frac{n^2}{16}+\frac{n}{16}\left(\rho_1^2+\rho_2^2-\rho_3^2+\rho_4^2-\rho_5^2-\rho_6^2-\rho_7^2\right) \end{eqnarray*} \begin{eqnarray*} s_{1+}&=&\sum_{u<v}1(\sigma_u\sigma_v\eta_u\eta_v=-)1(\tau_u\tau_v=+) =\frac{n^2}{8}+\frac{n}{8}\left(\rho_3^2-\rho_4^2-\rho_7^2\right)\\ s_{1-}&=&\sum_{u<v}1(\sigma_u\sigma_v\eta_u\eta_v=-)1(\tau_u\tau_v=-) =\frac{n^2}{8}-\frac{n}{8}\left(\rho_3^2+\rho_4^2-\rho_7^2\right) \end{eqnarray*} \begin{eqnarray*} s_{0+}&=&\sum_{u<v}1(\sigma_u\sigma_v=-)1(\eta_u\eta_v=-)1(\tau_u\tau_v=+)\\ &=&\frac{n^2}{16}+\frac{n}{16}\left(-\rho_1^2-\rho_2^2+\rho_3^2+\rho_4^2-\rho_5^2-\rho_6^2+\rho_7^2\right)\\ s_{0-}&=&\sum_{u<v}1(\sigma_u\sigma_v=-)1(\eta_u\eta_v=-)1(\tau_u\tau_v=-)\\ &=&\frac{n^2}{16}+\frac{n}{16}\left(-\rho_1^2-\rho_2^2-\rho_3^2+\rho_4^2+\rho_5^2+\rho_6^2-\rho_7^2\right). \end{eqnarray*} Using the above notation $\gamma_{r\pm}$'s and $s_{r\pm}$'s we can write the right hand side of (\ref{thm:power:case2:eqn6}) as \[ \prod_{r=0,1,2}(1+\gamma_{r+})^{s_{r+}} \times\prod_{r=0,1,2}(1+\gamma_{r-})^{s_{r-}}\equiv T_1\times T_2, \] where \begin{eqnarray*} T_1&=&(1+\gamma_{2+})^{\frac{n^2}{16}-\frac{n}{2}} (1+\gamma_{2-})^{\frac{n^2}{16}}(1+\gamma_{1+})^{\frac{n^2}{8}} (1+\gamma_{1-})^{\frac{n^2}{8}}(1+\gamma_{0+})^{\frac{n^2}{16}}(1+\gamma_{0-})^{\frac{n^2}{16}}\\ &=&(1+o(1))\exp\left(-\frac{(a_\varepsilon-b_\varepsilon)^2}{4(a+b)^4} \left[(a_\varepsilon-b_\varepsilon)^2(a_\varepsilon^2+a_\varepsilon b_\varepsilon+b_\varepsilon^2) +\varepsilon(a_\varepsilon-b_\varepsilon)(3a_\varepsilon^2+2a_\varepsilon b_\varepsilon+3b_\varepsilon^2)\right.\right.\\ &&\left.\left.+\varepsilon^2(3a_\varepsilon^2+2a_\varepsilon b_\varepsilon+3b_\varepsilon^2)\right]-\frac{(a_\varepsilon-b_\varepsilon)^2(2a_\varepsilon+b_\varepsilon)+x_\varepsilon}{2(a+b)^2}\right)\\ &=&(1+o(1))\exp\left(-\frac{(a_\varepsilon-b_\varepsilon)^4}{16(a+b)^2}-\frac{(a_\varepsilon-b_\varepsilon)^4(a-b)^2}{16(a+b)^4}-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)^2}{8(a+b)^2}\right.\\ &&\left.-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{4(a+b)^2}-\frac{(a_\varepsilon-b_\varepsilon)^2}{4(a+b)} -\frac{(a-b)(a_\varepsilon-b_\varepsilon)}{2(a+b)}\right)\\ &=&(1+o(1))\exp\left(-\kappa_\varepsilon^2/4-\kappa_\varepsilon/2\right)\exp\left(-\widetilde{\kappa}_\varepsilon^2/2-\widetilde{\kappa}_\varepsilon\right)\exp\left(-\frac{(a_\varepsilon-b_\varepsilon)^4(a-b)^2}{16(a+b)^4}-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{4(a+b)^2}\right), \end{eqnarray*} and \begin{eqnarray*} T_2&=&(1+o(1))(1+\gamma_{2+})^{\frac{n}{16}\left(\rho_1^2+\rho_2^2+\rho_3^2+\rho_4^2+\rho_5^2+\rho_6^2+\rho_7^2\right)} (1+\gamma_{2-})^{\frac{n}{16}\left(\rho_1^2+\rho_2^2-\rho_3^2+\rho_4^2-\rho_5^2-\rho_6^2-\rho_7^2\right)}\\ &&\times(1+\gamma_{1+})^{\frac{n}{8}\left(\rho_3^2-\rho_4^2-\rho_7^2\right)} (1+\gamma_{1-})^{-\frac{n}{8}\left(\rho_3^2+\rho_4^2-\rho_7^2\right)}\\ &&\times (1+\gamma_{0+})^{\frac{n}{16}\left(-\rho_1^2-\rho_2^2+\rho_3^2+\rho_4^2-\rho_5^2-\rho_6^2+\rho_7^2\right)} (1+\gamma_{0-})^{\frac{n}{16}\left(-\rho_1^2-\rho_2^2-\rho_3^2+\rho_4^2+\rho_5^2+\rho_6^2-\rho_7^2\right)}\\ &=&(1+o(1))\exp\left(\frac{(a_\varepsilon-b_\varepsilon)^2(2a_\varepsilon+b_\varepsilon)+x_\varepsilon}{16(a+b)^2} \left(\rho_1^2+\rho_2^2+\rho_3^2+\rho_4^2+\rho_5^2+\rho_6^2+\rho_7^2\right)\right.\\ &&\left.-\frac{a_\varepsilon(a_\varepsilon-b_\varepsilon)^2+x_\varepsilon}{16(a+b)^2} \left(\rho_1^2+\rho_2^2-\rho_3^2+\rho_4^2-\rho_5^2-\rho_6^2-\rho_7^2\right)\right.\\ &&\left.-\frac{a_\varepsilon(a_\varepsilon-b_\varepsilon)^2+z_\varepsilon}{8(a+b)^2}\left(\rho_3^2-\rho_4^2-\rho_7^2\right)\right.\\ &&\left.+\frac{b_\varepsilon(a_\varepsilon-b_\varepsilon)^2-z_\varepsilon}{8(a+b)^2} \left(\rho_3^2+\rho_4^2-\rho_7^2\right)\right.\\ &&\left.+\frac{b_\varepsilon(a_\varepsilon-b_\varepsilon)^2+w_\varepsilon}{16(a+b)^2}\left(\rho_1^2+\rho_2^2-\rho_3^2-\rho_4^2+\rho_5^2+\rho_6^2-\rho_7^2\right)\right.\\ &&\left.-\frac{(a_\varepsilon-b_\varepsilon)^2(a_\varepsilon+2b_\varepsilon)+w_\varepsilon}{16(a+b)^2}\left(\rho_1^2+\rho_2^2+\rho_3^2-\rho_4^2-\rho_5^2-\rho_6^2+\rho_7^2\right) \right)\\ &=&(1+o(1))\exp\left(\frac{\kappa_\varepsilon}{2}\rho_4^2+\frac{\widetilde{\kappa}_\varepsilon}{2}\left(\rho_5^2+\rho_6^2\right)+\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{4(a+b)^2}\rho_7^2\right). \end{eqnarray*} We note that $\rho_7$ is independent of $(\rho_4,\rho_5,\rho_6)$, and the condition $\kappa_\varepsilon<\widetilde{\kappa}_\varepsilon\in(0,1/3)$ leads to uniform integrability of $\exp\left(\frac{\kappa_\varepsilon}{2}\rho_4^2+\frac{\widetilde{\kappa}_\varepsilon}{2}\left(\rho_5^2+\rho_6^2\right)\right)$, and $\rho_4,\rho_5,\rho_6,\rho_7$ jointly converge in distribution to independent standard normal variables. Therefore, we have that \begin{eqnarray*} &&{\mathbb{E}}_1(Y_n^\varepsilon)^2\\ &\overset{n\to\infty}{\rightarrow}& \exp\left(-\kappa_\varepsilon^2/4-\kappa_\varepsilon/2\right)\exp\left(-\widetilde{\kappa}_\varepsilon^2/2-\widetilde{\kappa}_\varepsilon\right)\exp\left(-\frac{(a_\varepsilon-b_\varepsilon)^4(a-b)^2}{16(a+b)^4}-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{4(a+b)^2}\right)\\ &&\times(1-\kappa_\varepsilon)^{-1/2}(1-\widetilde{\kappa}_\varepsilon)^{-1}\left(1-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{2(a+b)}\right)^{-1/2}. \end{eqnarray*} By (\ref{thm:power:case2:eqn7}) we get that \begin{eqnarray}\label{thm:power:case2:eqn8} &&\frac{{\mathbb{E}}_1(Y_n^\varepsilon)^2}{({\mathbb{E}}_1Y_n^\varepsilon)^2}\nonumber\\ &\overset{n\to\infty}{\rightarrow}&\exp\left(-\kappa_\varepsilon^2/4-\kappa_\varepsilon/2\right)(1-\kappa_\varepsilon)^{-1/2}\nonumber\\ &&\times\exp\left(-\frac{(a_\varepsilon-b_\varepsilon)^4(a-b)^2}{16(a+b)^4}-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{4(a+b)^2}\right)\left(1-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{2(a+b)}\right)^{-1/2}\nonumber\\ &=&\exp\left(\sum_{m=3}^\infty\frac{1}{2m}\left(\frac{(a_\varepsilon-b_\varepsilon)^2}{2(a+b)}\right)^m\right) \times\exp\left(\sum_{m=3}^\infty\frac{1}{2m}\left(\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{2(a+b)^2}\right)^m\right)\\ &=&\exp\left(\sum_{m=3}^\infty\lambda_m(1+\delta_m)(\delta_m^\varepsilon)^2\right),\nonumber \end{eqnarray} where (\ref{thm:power:case2:eqn8}) follows from the below trivial facts: \begin{eqnarray*} \exp\left(\sum_{m=3}^\infty\frac{1}{2m}\left(\frac{(a_\varepsilon-b_\varepsilon)^2}{2(a+b)}\right)^m\right)&=& \exp\left(-\kappa_\varepsilon^2/4-\kappa_\varepsilon/2\right)(1-\kappa_\varepsilon)^{-1/2}\\ \exp\left(\sum_{m=3}^\infty\frac{1}{2m}\left(\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{2(a+b)^2}\right)^m\right) &=&\exp\left(-\frac{(a_\varepsilon-b_\varepsilon)^4(a-b)^2}{16(a+b)^4}-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{4(a+b)^2}\right)\\ &&\times\left(1-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{2(a+b)}\right)^{-1/2}. \end{eqnarray*} This verifies Condition A4. In the end, notice that by (\ref{thm:power:case2:eqn7}), \begin{eqnarray*} {\mathbb{E}}_1Y_n^\varepsilon&\overset{n\to\infty}{\rightarrow}&\exp\left(-\widetilde{\kappa}_\varepsilon^2/4-\widetilde{\kappa}_\varepsilon/2\right)(1-\widetilde{\kappa}_\varepsilon)^{-1/2}\\ &=&\exp\left(\sum_{m=3}^\infty\frac{1}{2m}\widetilde{\kappa}_\varepsilon^m\right)\\ &=&\exp\left(\sum_{m=3}^\infty\frac{1}{2m}\left(\frac{(a_\varepsilon-b_\varepsilon)(a-b)}{2(a+b)}\right)^m\right) =\exp\left(\sum_{m=3}^\infty\lambda_m\delta_m\delta_m^\varepsilon\right). \end{eqnarray*} The it follows by Proposition \ref{basic:prop} that \[ Y_n^\varepsilon\overset{n\to\infty}{\rightarrow}\exp\left(\sum_{m=3}^\infty\lambda_m\delta_m\delta_m^\varepsilon\right) \prod_{m=3}^\infty(1+\delta_m^\varepsilon)^{Z_m^1}\exp\left(-\lambda_m(1+\delta_m)\delta_m^\varepsilon\right) =W_1^\varepsilon, \] where $Z_m^1$ are independent Poisson variable with mean $\lambda_m(1+\delta_m)$. \end{proof} \subsection{Proofs in Section \ref{sec:power:analysis}} Before proofs, we need the following technical lemma. \begin{Lemma}\label{lemma:series:convergence} Suppose that $\{c_{ml}\}_{m,l=1}^\infty$ is a real sequence satisfying (1) $\lim\limits_{M\to\infty}\lim\limits_{l\to\infty}\sum_{m=M}^\infty c_{ml}^2=0$, and (2) for any $m\ge 1$, $\lim\limits_{l\to\infty}c_{ml}=c_m$. Furthermore, for any $l\ge1$, $\{N_{ml}\}_{m=1}^\infty$ are independent random variables of zero mean and unit variance, and for any $m\ge1$, $N_{ml}\overset{d}{\to}N(0,1)$ as $l\to\infty$. Then, as $l\to\infty$, $\sum_{m=1}^\infty c_{ml}N_{ml}\overset{d}{\to}N(0,\sum_{m=1}^\infty c_m^2)$. \end{Lemma} \begin{proof}[Proof of Lemma \ref{lemma:series:convergence}] Notice that $c_m$ is a square summable sequence. To see this, note that for any $M<N$, \[ \sum_{m=M}^N c_m^2=\lim\limits_{l\to\infty}\sum_{m=M}^N c_{ml}^2\le\lim\limits_{l\to\infty}\sum_{m=M}^\infty c_{ml}^2, \] and hence, taking $N\to\infty$ on the left side we have, \[ \sum_{m=M}^\infty c_m^2\le\lim\limits_{l\to\infty}\sum_{m=M}^\infty c_{ml}^2, \] leading to $\lim_{M\to\infty}\sum_{m=M}^\infty c_m^2\le\lim\limits_{M\to\infty}\lim\limits_{l\to\infty}\sum_{m=M}^\infty c_{ml}^2=0$; see (1). Hence $\sum_{m=1}^\infty c_m^2<\infty$. For arbitrary $M$ and $\delta>0$, define an event $\mathcal{E}_{Ml}=\{|\sum_{m=M}^\infty c_{ml}N_{ml}|<\delta\}$. Since $E|\sum_{m=M}^\infty c_{ml}N_{ml}|^2=\sum_{m=M}^\infty c_{ml}^2$, by condition (1) we can choose $l$ and $M$ large so that $E|\sum_{m=M}^\infty c_{ml}N_{ml}|^2\le\delta^3$, and so $P(\mathcal{E}_{Ml})\ge 1-\delta$ by Chebyshev inequality. By independence and asymptotic normality of $N_{ml}$ for $1\le m\le M-1$, and condition (2), one has $\sum_{m=1}^{M-1}c_{ml}N_{ml}\overset{d}{\to}N(0,\sum_{m=1}^{M-1}c_m^2)$ as $l\to\infty$. Define $T_l=\sum_{m=1}^\infty c_{ml}N_{ml}$. Hence, for any $z\in\mathbb{R}$, \begin{eqnarray*} P\left(T_l\le z\right)&\le&P(T_l\le z,\mathcal{E}_{Ml})+\delta\\ &\le&P\left(\sum_{m=1}^{M-1}c_{ml}N_{ml}\le z+\delta\right)+\delta\overset{l\to\infty}{\to}\Phi\left(\frac{z+\delta}{\sqrt{\sum_{m=1}^{M-1}c_m^2}}\right)+\delta. \end{eqnarray*} Taking $\delta\to0$ and $M\to\infty$ in the above, we have $\limsup\limits_{l\to\infty}P(T_l\le z)\le \Phi\left(\frac{z}{\sqrt{\sum_{m=1}^\infty c_m^2}}\right)$. Likewise one can show that $\liminf\limits_{l\to\infty}P(T_l\le z)\ge \Phi\left(\frac{z}{\sqrt{\sum_{m=1}^\infty c_m^2}}\right)$. Then we have $\lim\limits_{l\to\infty}P(T_l\le z)=\Phi\left(\frac{z}{\sqrt{\sum_{m=1}^\infty c_m^2}}\right)$. Proof completed. \end{proof} \begin{proof}[Proof of Theorem \ref{lim:power:case:2}] The proof follows by Lemma \ref{lemma:series:convergence}. We will analyze the distributions of $W_0^\varepsilon$ and $W_1^\varepsilon$. Define $\Delta_\varepsilon=\sum_{m=3}^\infty\lambda_m\left(\log(1+\delta_m^\varepsilon)-\delta_m^\varepsilon\right)$. Since, as $a+b\to\infty$, \[ \sqrt{\lambda_m}\log(1+\delta_m^\varepsilon)\to\sqrt{\frac{1}{2m}k_1^m}, \sqrt{\lambda_m(1+\delta_m)}\log(1+\delta_m^\varepsilon)\to\sqrt{\frac{1}{2m}k_1^m}, \lambda_m\delta_m\log(1+\delta_m^\varepsilon)\to\frac{1}{2m}k_2^m, \] and \[ \frac{Z_m^0-\lambda_m}{\sqrt{\lambda_m}}\overset{d}{\to}N(0,1), \frac{Z_m^1-\lambda_m(1+\delta_m)}{\sqrt{\lambda_m(1+\delta_m)}}\overset{d}{\to}N(0,1). \] Therefore, by Lemma \ref{lemma:series:convergence} we have, as $a+b\to\infty$, \[ \log{W_0^\varepsilon}-\Delta_\varepsilon= \sum_{m=3}^\infty\frac{Z_m^0-\lambda_m}{\sqrt{\lambda_m}}\times \sqrt{\lambda_m}\log(1+\delta_m^\varepsilon)\overset{d}{\to}N(0,\sigma_1^2), \] and \begin{eqnarray*} \log{W_1^\varepsilon}-\Delta_\varepsilon&=&\sum_{m=3}^\infty\frac{Z_m^1-\lambda_m(1+\delta_m)}{\sqrt{\lambda_m(1+\delta_m)}}\times\sqrt{\lambda_m(1+\delta_m)} \log(1+\delta_m^\varepsilon)+\sum_{m=3}^\infty\lambda_m\delta_m\log(1+\delta_m^\varepsilon)\\ &\overset{d}{\to}&N(\sigma_2^2,\sigma_1^2). \end{eqnarray*} Therefore, as $a+b\to\infty$, \begin{eqnarray*} 1-\alpha=P(W_0^\varepsilon\le w_\alpha^\varepsilon)= P\left(\frac{\log{W_0^\varepsilon}-\Delta_\varepsilon}{\sigma_1}\le\frac{\log{w_\alpha^\varepsilon}-\Delta_\varepsilon}{\sigma_1}\right), \end{eqnarray*} which implies $\frac{\log{w_\alpha^\varepsilon}-\Delta_\varepsilon}{\sigma_1}\to z_{1-\alpha}$, and hence, \begin{eqnarray*} P(a,b,\varepsilon)=P(W_1^\varepsilon\ge w_\alpha^\varepsilon)= P\left(\frac{\log{W_1}^\varepsilon-\Delta_\varepsilon}{\sigma_1}\ge\frac{\log{w_\alpha^\varepsilon}-\Delta_\varepsilon}{\sigma_1}\right) \to\Phi\left(\frac{\sigma_2^2}{\sigma_1}-z_{1-\alpha}\right). \end{eqnarray*} Proof completed. \end{proof} \subsection{Proofs in Section \ref{sec:approx}} \begin{proof}[Proof of Theorem \ref{thm:approx2}] Observe that \[ Var\left(\frac{1}{M}\sum_{l=1}^M g_n^\varepsilon(\sigma[l])\bigg| A\right)=\frac{1}{M}\left[ {\mathbb{E}}_\sigma\left\{g_n^\varepsilon(\sigma)^2\big|A\right\}-{\mathbb{E}}_\sigma\left\{g_n^\varepsilon(\sigma)\big|A\right\}^2\right]\le\frac{1}{M} {\mathbb{E}}_\sigma\left\{g_n^\varepsilon(\sigma)^2\big|A\right\}, \] where the variance is taken w.r.t. $\sigma[l]$'s conditional on $A_{uv}$'s. So it is sufficient to deal with ${\mathbb{E}}_{A,\sigma}g_n^\varepsilon(\sigma)^2$. First, assume $H_0$ holds. Then it holds that \[ {\mathbb{E}}_{A,\sigma}g_n^\varepsilon(\sigma)^2={\mathbb{E}}_\sigma\prod_{u<v}\left(\frac{p_{uv}^\varepsilon(\sigma)^2}{p_0}+\frac{q_{uv}^\varepsilon(\sigma)^2}{q_0}\right)=(1+o(1)) (1+\gamma_n^\varepsilon)^{\frac{n(n-1)}{2}}, \] where $\gamma^\varepsilon=\frac{\kappa_\varepsilon}{n}+\frac{(a_\varepsilon-b_\varepsilon)^2}{4n^2}$, $\kappa_\varepsilon=\frac{(a_\varepsilon-b_\varepsilon)^2}{2(a+b)}$, and the last equality holds due to the following trivial fact: \[ \frac{p_{uv}^\varepsilon(\sigma)^2}{p_0}+\frac{q_{uv}^\varepsilon(\sigma)^2}{q_0}=1+\gamma_n^\varepsilon+O(n^{-3}),\,\,\textrm{uniformly for $\sigma\in\{\pm\}^n$.} \] Obviously, $(1+\gamma_n^\varepsilon)^{\frac{n(n-1)}{2}}=\exp\left(\frac{n\kappa_\varepsilon}{2}- \frac{\kappa_\varepsilon^2}{4}-\frac{\kappa_\varepsilon}{2}+\frac{(a_\varepsilon-b_\varepsilon)^2}{8}\right)$, hence, $\frac{1}{M}\sum_{l=1}^M g_n^\varepsilon(\sigma[l])=Y_n^\varepsilon+o_P(1)$ if $M\gg\exp\left(\frac{n\kappa_\varepsilon}{2}\right)$. Next assume $H_1$ holds. Let $N_{++}=\#\{(u,v): u<v, \sigma_u\sigma_v=+, \tau_u\tau_v=+\}$, $N_{+-}=\#\{(u,v): u<v, \sigma_u\sigma_v=+, \tau_u\tau_v=-\}$, $N_{++}=\#\{(u,v): u<v, \sigma_u\sigma_v=-, \tau_u\tau_v=+\}$, $N_{++}=\#\{(u,v): u<v, \sigma_u\sigma_v=-, \tau_u\tau_v=-\}$. Similar to the expressions of $s_{r\pm}$ for $r=0,1,2$ in the proof of Theorem \ref{power:case2}, one can derive that \begin{eqnarray*} N_{++}&=&\frac{n^2}{8}-\frac{n}{2}+\frac{n}{8}(\rho_1^2+\rho_3^2+\rho_5^2), N_{+-}=\frac{n^2}{8}+\frac{n}{8}(\rho_1^2-\rho_3^2-\rho_5^2),\\ N_{-+}&=&\frac{n^2}{8}-\frac{n}{8}(\rho_1^2-\rho_3^2+\rho_5^2), N_{--}=\frac{n^2}{8}-\frac{n}{8}(\rho_1^2+\rho_3^2-\rho_5^2). \end{eqnarray*} Following (\ref{def:gamma2+:gamma0-}), one can check that \begin{eqnarray*} &&{\mathbb{E}}_{A,\sigma}g_n^\varepsilon(\sigma)^2\\ &=&4^{-n}\sum_{\sigma,\tau}\prod_{u<v} \left(\frac{1}{p_0^2}p_{uv}^\varepsilon(\sigma)^2p_{uv}(\tau)+\frac{1}{q_0^2}q_{uv}^\varepsilon(\sigma)^2q_{uv}(\tau)\right)\\ &=&4^{-n}\sum_{\sigma,\tau}\left(\frac{1}{p_0^2}\left(\frac{a_\varepsilon}{n}\right)^2\left(\frac{a}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{a_\varepsilon}{n}\right)^2\left(1-\frac{a}{n}\right)\right)^{N_{++}}\\ &&\times\left(\frac{1}{p_0^2}\left(\frac{a_\varepsilon}{n}\right)^2\left(\frac{b}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{a_\varepsilon}{n}\right)^2\left(1-\frac{b}{n}\right)\right)^{N_{+-}}\\ &&\times\left(\frac{1}{p_0^2}\left(\frac{b_\varepsilon}{n}\right)^2\left(\frac{a}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{b_\varepsilon}{n}\right)^2\left(1-\frac{a}{n}\right)\right)^{N_{-+}}\\ &&\times\left(\frac{1}{p_0^2}\left(\frac{b_\varepsilon}{n}\right)^2\left(\frac{b}{n}\right) +\frac{1}{q_0^2}\left(1-\frac{b_\varepsilon}{n}\right)^2\left(1-\frac{b}{n}\right)\right)^{N_{--}}\\ &=&(1+o(1)){\mathbb{E}}_{\sigma\tau}(1+\gamma_{2+})^{N_{++}}(1+\gamma_{2-})^{N_{+-}}(1+\gamma_{0+})^{N_{-+}} (1+\gamma_{0-})^{N_{--}}. \end{eqnarray*} It follows from direct examinations that \[ (1+\gamma_{2+})^{\frac{n^2}{8}-\frac{n}{2}}(1+\gamma_{2-})^{\frac{n^2}{8}}(1+\gamma_{0+})^{\frac{n^2}{8}}(1+\gamma_{0-})^{\frac{n^2}{8}} \asymp\exp\left(\frac{n\kappa_\varepsilon}{2}\right), \] and \begin{eqnarray*} &&{\mathbb{E}}_{\sigma\tau}(1+\gamma_{2+})^{\frac{n}{8}(\rho_1^2+\rho_3^2+\rho_5^2)} (1+\gamma_{2-})^{\frac{n}{8}(\rho_1^2-\rho_3^2-\rho_5^2)} (1+\gamma_{0+})^{-\frac{n}{8}(\rho_1^2-\rho_3^2+\rho_5^2)} (1+\gamma_{0-})^{-\frac{n}{8}(\rho_1^2+\rho_3^2-\rho_5^2)}\\ &=&(1+o(1)){\mathbb{E}}_{\sigma\tau}\exp\left(\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{4(a+b)^2}\rho_3^2 +\frac{(a_\varepsilon-b_\varepsilon)(a-b)}{2(a+b)}\rho_5^2\right)\\ &\overset{n\to\infty}{\rightarrow}&\left(1-\frac{(a_\varepsilon-b_\varepsilon)^2(a-b)}{2(a+b)^2}\right)^{-1/2} \left(1-\frac{(a_\varepsilon-b_\varepsilon)(a-b)}{a+b}\right)^{-1/2}. \end{eqnarray*} The last limit follows by condition $(a_\varepsilon-b_\varepsilon)(a-b)<a+b$ and asymptotic independent standard normality of $\rho_3$ and $\rho_5$. Hence, ${\mathbb{E}}_{A,\sigma}g_n^\varepsilon(\sigma)^2\lesssim \exp\left(\frac{n\kappa_\varepsilon}{2}\right)$, leading to $\frac{1}{M}\sum_{l=1}^M g_n^\varepsilon(\sigma[l])=Y_n^\varepsilon+o_P(1)$ if $M\gg\exp\left(\frac{n\kappa_\varepsilon}{2}\right)$. \end{proof}

https://arxiv.org/abs/1209.5170

Identifying the successive Blumenthal-Getoor indices of a discretely observed process

This paper studies the identification of the Lévy jump measure of a discretely-sampled semimartingale. We define successive Blumenthal-Getoor indices of jump activity, and show that the leading index can always be identified, but that higher order indices are only identifiable if they are sufficiently close to the previous one, even if the path is fully observed. This result establishes a clear boundary on which aspects of the jump measure can be identified on the basis of discrete observations, and which cannot. We then propose an estimation procedure for the identifiable indices and compare the rates of convergence of these estimators with the optimal rates in a special parametric case, which we can compute explicitly.

\section{Introduction}\label{secintro} Let $X$ be a one-dimensional semimartingale defined on a finite time interval $[0,T]$. Our objective is to make some progress toward the identification of the jump measure of $X$ at high frequency. The motivation for what follows has its roots in a family of econometric problems, which can be stated as follows. We observe a single path of $X$, but not fully: although other observation schemes are possible, the most typical is one where we observe the variables $X_{i\Delta_{n}}$ for $i=0,1,\ldots,[T/\Delta_{n}]$, where $[x]$ denotes the integer part of the real $x$, over a fixed observation span $T$ and where $\Delta_{n}$ is small. Asymptotic results are derived in the high-frequency limit where the sequence $\Delta_{n}$ going to $0$. The overall objective is to find out what can be recovered, that is, identified, about the dynamics of $X$, in this setup where a single path, partially observed at a discrete time interval, is all that is available. For those parameters which can be identified, we also want asymptotically consistent estimators, with a rate whenever possible. For the dynamics of $X$, we restrict our attention to It\^{o} semimartingales, meaning that the characteristics $(B,C,\nu)$ of $X$ can be written as follows: \begin{eqnarray} \label{3 B_{t}(\omega)&=&\int_{0}^{t}b_{s}( \omega) \,ds,\nonumber\\[-8pt]\\[-8pt] C_{t}(\omega)&=&\int_{0}^{t}c_{s}(\omega)\,ds,\qquad \nu(\omega,dt,dx)=dt\otimes F_{t}(\omega,dx)\nonumber \end{eqnarray} for some adapted processes $b_{t}$ and $c_{t}$ and measure $F_{t}(\omega,dx)$. Recall that~$B$ is the drift, $C$ is the quadratic variation of the continuous martingale part and~$\nu$ is the compensator of the jump measure $\mu$ of $X$ [see \citet{jacodshiryaev2003} for more details on characteristics]. As is well known, these are the canonical models for arbitrage-free asset prices. A sizeable part of the paper, however, is concerned with the much-restrict\-ed class of L\'evy processes. A semimartingale $X$ is a L\'evy process if and only if~(\ref{3}) holds with $b_{t}(\omega)=b\in\mathbb{R}$ and $c_{t (\omega)=c\geq0$ and $F_{t}(\omega,dx)=F(dx)$ independent of $\omega$ and $t$. The measure $F$ is the \textit{L\'evy measure}, and it integrates $x^{2}\wedge 1$. The (deterministic) triple $(b,c,F)$ is then the characteristic triple coming in the L\'evy--Khintchine formula, providing the characteristic function of $X_{t}$, \begin{equation} \label{1} \mathbb{E}[ e^{iuX_{t}}] =\exp t\biggl( iub-\frac{cu^{2}}{2 +\int\bigl( e^{iux}-1-iux1_{\{|x|\leq1\}}\bigr) F(dx)\biggr). \end{equation} This completely characterizes the entire law of $X$. Ultimately, we would like to identify as much as we can of the characteristics $B$, $C$ and $\nu$, and give consistent estimators for the identifiable parameters. The situation is well understood for the first two characteristics, $B$ and $C$. When $X$ is fully observed on $[0,T]$, one knows the jumps (size and location) occurring within the interval, and the quadratic variation of $X$ on $[0,T]$, hence the function $t\mapsto C_{t}$ on $[0,T]$. On the other hand, and at least when $C$ is strictly increasing (which is the case in almost all models used in practice), nothing can be said about the drift~$B$. When the process is observed only at discrete times, $C_{t}$ is no longer exactly known, but there are well established methods to estimate it in a consistent way as the observation mesh goes to $0$, even in the presence of jumps. We focus on the remaining open question, which concerns identifiability and estimation for the third characteristic, $\nu$, or equivalently $F_{t}$, for a discretely sampled semimartingale. The measure $F_{t}$ in a sense describes the law of a jump occurring at time $t$, conditionally on the past before $t$. There is a~vast literature on identifying the L\'{e}vy measure when the time horizon $T$ is asymptotically infinite, and when $X$ is a L\'evy process; see, for example, \citet{basawabrockwell82}, \citet{figueroalopezhoudre06}, \citet{nishiyama09}, \citet{neumannreiss09} and \citet {comtegenoncatalot09}. But over a finite time horizon $T$, we cannot reconstruct $\nu$ fully because there are only finitely many jumps on $[0,T]$ with size bigger than any $\varepsilon>0$. The open question which we seek to address in this paper is: what can we and can we not identify about $\nu?$ High-frequency data analysis has proved a very fruitful area of research. As we will see, however, it is not able to achieve everything, and our objective in this paper is to pinpoint exactly the limitations, or frontier, involved in using high-frequency data over a fixed time span. We can say something about the concentration of $\nu$ around $0$. For example, we can decide for which $p\geq0$ we have $\int_{0}^{T}ds\int F_{t (\omega,dx)(|x|^{p}\wedge1)<\infty$, because outside a null set again these are exactly those $p$'s for which $\sum_{s\leq T}|\Delta X_{s}(\omega )|^{p}<\infty$,\vspace*{1pt} where $\Delta X_{s}=X_{s}-X_{s-}$ is the size of the jump at time~$s$, if any. The infimum of all such $p$'s is a generalization of the \textit{Blumenthal--Getoor} index (or BG index) of the process up to time $T$, and it is \textit{known} when $X$ is fully observed. Note that a priori it is random, and also increasing with $T$, and always with values in $[0,2]$. However, in the L\'{e}vy process case, it reduces to $\inf\{ p\dvtx\int F(dx)(|x|^{p}\wedge1)<\infty\}$, and is nonrandom and independent of time. It was originally introduced by \citet{blumenthalgetoor61}, and for a stable process the BG index is also the stability index of the process. The interest in identifying the BG index lies in the fact that the index allows for a classification of the processes from least active to most active: processes with BG index equal to $0$ are either finitely active or infinitely active but with slow, sub-polynomial, divergence of $\nu$ near $0;$ processes with BG index strictly positive are all infinitely active; processes with BG index less than $1$ have paths of finite variation; processes with BG index greater than $1$ have paths of infinite variation; and in the limit, processes with continuous paths have an ``activity index'' (the analog of the BG index which no longer exists) equal to $2$ when the volatility is not vanishing. In other words, jumps become more and more active as the BG index increases from $0$ to $2$, and we can think of this generalized BG index as an \textit{index of jump activity}. In the case of discrete observations at times $i\Delta_{n}$ with $\Delta_{n}$ going to $0$, recovering the random BG index in full generality seems out of reach, but \citet{yacjacod09b} constructed estimators of the nonrandom number $\beta$ that are consistent as $\Delta_{n}\rightarrow0$, under the main assumption that locally near $0$, we have the behavior \begin{equation} \label{I1 F_{t}(\omega,[-u,u]^{c})\sim\frac{a_{t}(\omega)}{u^{\beta}}\qquad \mbox{as }u\downarrow0 \end{equation} (plus a few technical hypotheses), where $a_{t}\geq0$ is a process: in this case, $\beta$ is the---deterministic---BG index at time $t$, on the set $\{ \int_{0}^{t}a_{s}\,ds>0\} $. We call this behavior ``proto-stable,'' since it is similar to that of a stable process but only near~$0$. Away from a neighborhood of $0$, the jump measure is completely unrestricted. We obtained the rate of convergence and a central limit theorem for the estimators, depending upon the rate in the approximations~(\ref{I1}). Related estimators or tests for $\beta$ include \citet{belomestny10}, \citet{contmancini11} and \citet{tauchentodorov10}. We can think of~(\ref{I1}) as providing the leading term, near $0$, of the jump measure of $X$. Given that this term is identifiable, but that the full measure $\nu$ is not, our aim is to examine where the boundary between what can versus what cannot be identified lies. Toward this aim, one direction to go is to view~(\ref{I1}) as giving the first term of the expansion of the ``tail'' $F_{t}(\omega,[-u,u]^{c})$ near $0$, and go further by assuming a series expansion such as \begin{equation} \label{I2 F_{t}(\omega,[-u,u]^{c})\sim\sum_{i\geq1}\frac{a^{i}_{t}(\omega )}{u^{\beta _{i}}}\qquad \mbox{as }u\downarrow0 \end{equation} (the precise assumption is given in Section~\ref{secSBG}), with successive powers $\beta_{1}=\beta>\beta_{2}>\beta_{3}>\cdots.$ Those $\beta _{i}$'s will be the ``successive BG indices.'' This series expansion can, for example, result from the superposition of processes with different BG indices, in a model consisting of a sum of such processes. The question then becomes one of identifying the successive terms in that expansion. The main theoretical result of the paper, which is somehow surprising, is as follows: the first index $\beta_{1}$ is always identifiable, as we already knew, but the subsequent indices $\beta_{i}$ which are bigger than $\beta_{1}/2$ are identifiable, whereas those smaller are not. An intuition for this particular value of the ``identifiability boundary'' is as follows: in view of~(\ref{I2}) the estimation of the $\beta_{i}$'s can only be based on preliminary estimations of $F_{t}(\omega,[-u,u]^{c})$, or of an integrated (in time) version of this, for a sequence $u_{n}\rightarrow0$. It turns out that, even in idealized circumstances, an estimation of $F_{t}(\omega,[-u_{n},u_{n}]^{c})$ or of its integrated version has a rate of convergence $u_{n}^{-\beta_{1}/2}$ (there is a central limit theorem for this), so that any term contributing to $F_{t}(\omega,[-u_{n},u_{n}]^{c})$ by an amount less than $u_{n}^{-\beta _{1}/2}$ is fundamentally unreachable: we can only hope to estimate a further coefficient $\beta_{i}$ if it leads to a number of increments greater than $u_{n}$ (which is of order $u_{n}^{-\beta_{i}}$) that is larger than the sampling error in the number of terms generated by the first coefficient, implying that any $\beta_{i}<\beta_{1}/2$ cannot be identified. This shows that there are limits to our ability to identify these successive terms, even in the unrealistic situation where the process is fully observed, and the behavior of $\nu$ around $0$ is only partly identifiable. When the identifiability conditions are satisfied, and when the process is observed at discrete times with mesh $\Delta_{n}$, we will construct estimators of the parameters which are consistent as $\Delta _{n}\rightarrow0$, and determine their rate of convergence, which we will see are slow. In the case we have only two indices $\beta_{1}>\beta_{2}$ with $\beta _{2}>\beta _{1}/2$, we will further compare the rates of the estimators we exhibit, which are semiparametric, to the optimal rate achievable in a~corresponding parametric sub-model (the sum of two stable processes, plus a drift and a Brownian motion) \begin{figure} \includegraphics{976f01.eps} \caption{Two BG component model: regions where the components are identified versus not identified, and optimal rate of convergence. \label{figtheorysummary} \end{figure} The main results of the paper are summarized in Figure~\ref{figtheorysummary} for the two-component situation. We already noted that $\beta_{2}$ can be identified only if it is bigger than $\beta_{1}/2;$ we will also see that the rate at which $\beta_{2}$ can be estimated increases as $\beta_{2}$ gets closer to $\beta_{1}$, and conversely decreases as $\beta_{2}$ gets closer to $\beta_{1}/2$, in the limit dropping to $0$ as $\beta_{2}$ approaches $\beta_{1}/2$, consistently with the loss of identification that occurs at that point. Beyond the two-component model, we will provide general identifiability conditions and rates of convergence for the leading and higher order BG indices. The paper is organized as follows. We first define the successive BG indices in Section~\ref{secSBG}. In Section~\ref{secLC}, we study the identifiability of the parameters appearing in the expansion, from a theoretical viewpoint and in the special case of L\'evy processes. Then we introduce consistent estimators for those parameters which we have found to be identifiable in the L\'evy case, hence proving de facto their identifiability. This is done according to a two-step procedure, with preliminary estimators given in Section~\ref{secD}, and final estimators with much faster rates in Section~\ref{secD2}. Unfortunately, although rates are given, we were not able to show a central limit theorem for these estimators, although such theorems ought to be available and would be crucial for obtaining confidence bounds. In principle, those estimators could be used on real data, but the rates of convergence for the higher order indices are, by necessity, quite slow. We show in Section~\ref{secFisher} that the slow nature of these rates of convergence is an inherent feature of the problem that cannot be improved upon. This is perhaps not too surprising since the range of values of the higher order indices that are identified is limited, and hence one would expect the rate of convergence to deteriorate all the way to zero as one approaches the region where identification disappears. We provide in Section \ref{secMC} a simulation study for a~model featuring a stochastic volatility plus two stable processes with different indices, the aim being to identify these two indices, especially the higher order one. A realistic application to high-frequency financial data, is out of the question for the typical sample sizes that are currently available, but may be useful in the future or in different fields of applications where semimartingales are used and where data are available in vast quantities, such as the study of Internet traffic or turbulence data in meteorology. The results do also present theoretical interest, especially as they set up bounds on what is asymptotically identifiable in the jump measure of a semimartingale, and consequently what is not.\vspace*{-1pt} \section{The successive Blumenthal--Getoor indices}\label{secSBG} Throughout the paper, $X$ is an It\^{o} semimartingale with characteristics given by~(\ref{3}), on a filtered probability space $(\Omega,\mathcal{F} ,(\mathcal{F} _{t})_{t\geq0},\mathbb{P})$. The time horizon for the observations is $T>0$, so the behavior of $X$ after time $T$ does not matter for us below. Our first aim is to give a precise meaning to an hypothesis like~(\ref{I2}). Instead of requiring an expansion like this for all times $t$, we rather use the ``integrated version'' which uses the following family of (adapted, continuous and increasing) processes: \begin{equation}\label{30 u>0\quad\Rightarrow\quad\overline{A}(u)_{t} =\int_{0}^{t}F_{s}([-u,u]^{c})\,ds. \end{equation} The basic assumption is as follows: \begin{assumption} \label{AA1} There are a \textit{nonrandom} integer $j$, a strictly decreasing sequence $(\beta_{i})_{1\leq i\leq j+1}$ of numbers in $[0,2)$ and a sequence $(A^{i})_{1\leq i\leq j+1}$ of processes such that \begin{equation} \label{31} t\in[0,T],\qquad 0<u\leq1\quad\Rightarrow\quad\Biggl\vert\overline{A}(u)_{t} -\sum_{i=1}^{j}\frac{A_{t}^{i}}{u^{\beta_{i}}}\Biggr\vert\leq\frac {A_{t}^{j+1}} {u^{\beta_{j+1}}}. \end{equation} Moreover, we have $A^{i}_{T}>0$ for $i=1,\ldots,j$. \end{assumption} If this assumption is satisfied with some $j\geq2$, it is also satisfied with any smaller integer. The processes $A^{i}$ and $A^{\prime}$ are nondecreasing nonnegative, and they can always be chosen to be predictable. Clearly, $\beta=\beta_{1}$ is the BG index, as introduced before, and the following definition comes naturally in: \begin{definition} \label{DD1} Under Assumption~\ref{AA1}, the numbers $\beta_{1},\beta _{2},\ldots, \beta_{j}$ are called the successive BG indices of the process $X$ over the time interval $[0,T]$, and the variables $A^{i}_{T}$ are called the associated integrated intensities. \end{definition} \begin{example} \label{E21} Let $Y^{1},\ldots,Y^{j}$ be independent stable processes with indices $\beta_{1}>\cdots>\beta_{j}$. Then $X=Y^{1}+\cdots+Y^{j}$ satisfies (\ref{31}) with $A^{j+1}=0$\vadjust{\goodbreak} and the successive indices and integrated intensities are $\beta_{i}$ and $Ta_{i}$, where $a_{i}=\lim_{u\to0 u^{\beta_{i}} F^{i}([-u,u]^{c})$, and $F^{i}$ is the L\'evy measure of $Y^{i}$. If the $Y^{i}$'s are tempered stable processes [see \citet{rosinski07}] the same is true, provided $\beta_{j}>\beta_{1}-1$.\vspace*{-2pt} \end{example} \begin{example} \label{E22} A semimartingale consisting of a continuous component and a jump part driven by a sum of such processes also satisfies~(\ref{31}). Let $X_{t}=X_{0}+Z_{t}+\sum_{i=1}^{j}\int_{0}^{t}H_{s}^{i}\,dY_{s}^{i}$, with $Z$ a continuous It\^{o} semimartingale and~$Y^{i}$ as in the previous example and $H^{i}$ locally bounded predictable processes with $\int_{0}^{T}|H^{i _{s}|^{\beta_{i}}\,ds>0$. The successive BG indices are again the $\beta _{i $'s, with the associated integrated intensities \[ A_{T}^{i}=a_{i}\int_{0}^{T}|H_{s}^{i}|^{\beta_{i}}\,ds.\vspace*{-2pt} \] \end{example} \begin{remark} \label{R19} We have taken a finite family of possible indices $\beta_{i}$. Nothing prevents us from taking an infinite sequence: we simply have to assume that Assumption~\ref{AA1} holds for all $j$, with additionally $\lim _{i\to\infty} \beta_{i}=0$. However, in view of the restriction imposed on the BG indices by our main theorems below about identifiability, this more general situation has no statistical interest.\vspace*{-2pt} \end{remark} \begin{remark} Assumption~\ref{DD1} imposes a certain structure on the behavior of the jump measure of the process near $0$. It is important to note that it does not restrict in any way the behavior of the jump measure away from~$0$. Although most models used in practice and with infinite activity jumps satisfy this assumption, the Gamma process does not: although it (barely) exhibits infinite activity, its BG index is $0$, and $\overline{A}(u)_{t}$ is of order $\log(1/u)$.\vspace*{-2pt} \end{remark} In Assumption~\ref{AA1}, expansion~(\ref{31}) is central, but one may wonder about the additional requirement $A^{i}_{T}>0$. So, we end this section with some comments and extensions, which may look complicated and are not necessary for the rest of the paper, but which we think are useful and somewhat enlightening.\vspace*{-2pt} \begin{Extension}\label{exten1} In Assumption~\ref{AA1} positive and negative jumps are treated in the same way. In practice, it might be useful for modeling purposes to establish the behavior of positive and negative jumps separately. Toward this end, one can replace~(\ref{30}) by \[ \overline{A}(u)_{t}^{(+)} =\int_{0}^{t}F_{s}((u,\infty))\,ds,\qquad \overline{A}(u)_{t}^{(+)} =\int_{0}^{t}F_{s}((-\infty,-u))\,ds. \] Then, if one is interested in positive jumps only, say, one replaces (\ref{31}) by a~similar expansion for $\overline{A}(u)_{t}^{(+)}$: all the content of the paper still holds, mutatis mutandis, under this modified assumption, for positive jumps. The same is true of negative jumps, and the ``positive'' and ``negative'' successive BG indices can of course be different.\vadjust{\goodbreak} \end{Extension} \begin{Extension}\label{exten2} Now we come to the requirement $A^{i}_{T}>0$, which in Assumption~\ref{AA1} is supposed to hold for all (or, almost all) $\omega$. This is of course unlikely to hold for the terminal time $T$, unless it holds for all $t>0$, and even unless the processes $A^{i}$ are strictly increasing. In Example~\ref{E22}, this amounts to suppose that none of processes $H^{i}$ vanishes. However, it might be relevant in practice to allow for each $H^{i}$ to vanish on some (possibly random) time intervals: we then can have different components of the model turned on and off at different times. Thus, let us examine what happens if we relax the requirements $A^{i}_{T}>0$. For any particular outcome $\omega$, the (first) BG index of the process $X$ is $\beta_{i}$, where $i$ is the smallest integer such that $A^{i}_{T}>0$, and if all of them vanish one only knows that the BG index is not bigger than $\beta_{j+1}$. The same applies to further indices. In other words, one can define a partition of $\Omega$ indexed by all subsets $D$ of $\{1,\ldots ,j\}$ as follows: \begin{equation} \label{37 \Omega_{T}(D)=\biggl( \bigcap_{i\in D}\{A_{T}^{i}>0\}\biggr) \cap\biggl( \bigcap_{i\in\{1,\ldots,j\}\setminus D}\{A_{T}^{i}=0\}\biggr). \end{equation} Then, for any $\omega$, the successive BG indices of $X$ over $[0,T]$ and the associated intensities are the numbers $\beta^{\prime}_{1}(\omega ),\ldots,\beta^{\prime}_{J}(\omega)$ and $\Gamma_{i}(\omega)$, defined as \begin{eqnarray} \label{36 J(\omega)&=&m,\qquad \beta^{\prime}_{i}(\omega)=\beta_{l_{i}},\nonumber\\[-8pt]\\[-8pt] \Gamma_{i}(\omega)&=&A_{t}^{l_{i}}(\omega) \qquad\mbox{if } \omega\in\Omega _{t}(\{l_{1},\ldots,l_{m}\}).\nonumber \end{eqnarray} On the set $\Omega_{T}(\varnothing)$, which is not necessarily empty, we have $J=0$ and no~$\beta^{\prime}_{i}$'s. All results of this paper are true if we relax $A^{i}_{T}>0$ in Assumption \ref{AA1}, provided we replace $j$ by $J$ and the $\beta_{i}$'s by the~$\beta^{\prime}_{i}$'s, \textit{in restriction to the set}~$\Omega_{T}(D)$: this is indeed very easy, because on this set the process $X$ coincides at all times $t\in[0,T]$ with a process $X^{\prime}$ with satisfies Assumption \ref{AA1} as stated above, with $(j,\beta_{1},\ldots,\beta_{j},\beta_{j+1})$ substituted with $(m,\beta_{l_{1}},\ldots,\beta_{j_{m}},\beta_{j+1})$, when $D=\{l_{1},\ldots,l_{m}\}$. \end{Extension} \section{\texorpdfstring{Identifiability in the L\'{e}vy case}{Identifiability in the Levy case}}\label{secLC} Loosely speaking, in an asymptotic statistical framework, identifiability of a parameter means the existence of a sequence of estimators which is (weakly) consistent. Identifiability can be ``proved'' by exhibiting such a sequence. It can be ``disproved'' by theoretical arguments, such as the fact that if the parameter is identifiable in our high-frequency observations setting, then, were the path $t\mapsto X_{t}$ fully observed on $[0,T]$, it would enjoy ``nonasymptotic'' identifiability in the sense that its value is almost surely known. For example, in the simple model $X_{t}=bt+W_{t}$ the parameter $b$ does not enjoy this nonasymptotic property because the laws of the process $X$ (restricted to $[0,T]$) are all equivalent when $b$ varies, and thus $b$ is even less identifiable in the asymptotic setting. Disproving identifiability is usually a hard task, especially in a nonparametric setting. However, if a parameter is not identifiable for a certain class of models, it is of course not identifiable for any wider class. These arguments lead us to consider the very special situation of a L\'evy processes $X$, with L\'{e}vy--Khintchine characteristics $(b,c,F)$ [see (\ref{1})] when the path $t\mapsto X_{t}$ is fully observed on $[0,T]$. In this section we are interested in nonasymptotic identifiability of those characteristics, or functions of them. Note that, were $T$ infinite, the triple $(b,c,F)$ would be identifiable because, for example, one would know the values of all the i.i.d. increments $X_{n+1}-X_{n}$, giving us almost surely the law of $X_{1}$, which in turn determines the triple $(b,c,F)$. This is no longer the case when, as in this paper, the time interval $[0,T]$ is finite. In this case, we give a formal definition of identifiability. We use $Q_{b,c,F}$ to denote the law of the process $X$, restricted to the interval $[0,T]$ ($T$ is kept fixed all throughout). So $Q_{b,c,F}$ is a probability measure on the Skorokhod space $\mathbb{D}=\mathbb{D} (|0,T],\mathbb{R})$. We also let $\mathcal{T}$ be some given subset of all possible triples $(b,c,F)$. \begin{definition} \label{DD2} A function $H$ is \textit{identifiable on the class $\mathcal{T}$} if, for any two $(b,c,F)$ and $(b^{\prime},c^{\prime},F^{\prime})$ in $\mathcal{T}$ such that $H(b^{\prime},c^{\prime},F^{\prime})\neq H(b,c,F)$, we have $Q_{b,c,F}\perp Q_{b^{\prime},c^{\prime},F^{\prime}}$ (\textit{i.e., the two measures $Q_{b,c,F}$ and $Q_{b^{\prime},c^{\prime},F^{\prime}}$ are mutually singular}). \end{definition} The rationale behind this definition is as follows: if $H$ is identifiable and $(b,c,F)\in\mathcal{T}$, and $X$ is drawn according to the law $Q_{b,c,F}$, then we can discard with probability $1$ any fixed $(b^{\prime },c^{\prime },F^{\prime})\in\mathcal{T}$ such that $H(b^{\prime},c^{\prime },F^{\prime })\neq H(b,c,F)$. Unfortunately, this does not mean that we can (almost surely) reject all $(b^{\prime},c^{\prime},F^{\prime})$ with $H(b^{\prime },c^{\prime},F^{\prime}) \neq H(b,c,F)$ simultaneously: this stronger property is (almost) never satisfied. There exists a criterion for mutual singularity of $Q_{b,c,F}$ and $Q_{b^{\prime},c^{\prime},F^{\prime}}$; see Remark IV.4.40 of \citet{jacodshiryaev2003}. We have a Lebesgue decomposition $F^{\prime }=f\bullet F+F^{\prime}{}^{\perp}$ of $F^{\prime}$ with respect to $F$, with $f$ a nonnegative Borel function and $F^{\prime}{}^{\perp}$ a measure supported by an $F$-null set. Then $Q_{b^{\prime},c^{\prime},F^{\prime }}\perp Q_{b,c,F}$ if and only if at least one of the following five properties is violated: \begin{equation}\label{LC1} \cases{ \displaystyle F^{\prime}{}^{\perp}(\mathbb{R})<\infty,\vspace*{2pt}\cr \displaystyle \alpha(F,F^{\prime})=\int\bigl( |f(x)-1|^{2}\wedge|f(x)-1|\bigr) F(dx)<\infty,\vspace*{2pt}\cr \displaystyle \alpha^{\prime}(F,F^{\prime})=\int_{\{|x|\leq1\}}|x| |f(x)-1| F(dx)<\infty, \vspace*{2pt}\cr \displaystyle c=0\quad\Rightarrow\quad b^{\prime}=b-\int_{\{|x|\leq1\}}x \bigl(f(x)-1\bigr) F(dx),\cr c^{\prime}=c.} \end{equation} It clearly follows that the function $H(b,c,F)=c$ is identifiable on any class~$\mathcal{T}$ (a well-known fact). The function $H(b,c,F)=b$ is not identifiable in general; however, on the class of all $(b,c,F)$ having $c=0$ and $\int_{\{|x|\leq1\}}|x|F(dx)<\infty$ the function $H(b,c,F)=\widehat {b}=b-\int_{\{|x|\leq1\}}xF(dx)$ (which is the ``real'' drift, in the sense that $X_{t}=\widehat{b}t+\sum_{s\leq t} \Delta X_{s}$) is identifiable. In the sequel we are not interested in $b$ or $c$, but in $F$ only. That is, we are looking at functions $H=H(F)$. This leads us to consider classes of the form\looseness=-1 \begin{equation}\label{LC2 \mathcal{T}=\mathbb{R}\times\mathbb{R}_{+}\times\mathcal{T}_{3} \qquad \mbox{where }\mathcal{T}_{3}\mbox{ is a set of L\'{e}vy measures.} \end{equation}\looseness=0 In words, we want no restriction on the parameters $b$ and $c$. Of course $\mathcal{T}_{3}$ should not be a singleton, and $H(F)$ should not be constant on $\mathcal{T}_{3}$, otherwise the identifiability problem is empty. The following example is clear: \begin{example} \label{ELC1} If $\mathcal{T}_{3}$ is a set of measures which coincide with some given~$F$ on a neighborhood of $0$, then by~(\ref{LC1}) no nontrivial $H(F)$ is identifiable on~$\mathcal{T}$. \end{example} This implies that, in the best-case scenario, a function $H(F)$ can be identifiable only if it depends on the ``behavior of the measure $F$ around $0$.'' Giving a necessary and sufficient condition for identifiability of such a function, other than saying that one of the properties in~(\ref{LC1}) fails when $H(F)\neq H(F^{\prime})$, seems out of reach. However, this is possible for some specific, but relatively large, classes of sets $\mathcal{T}_{3}$, with a priori relatively surprising results. Below we introduce such a class, in order to illustrate the nature of the available results. \begin{definition}[(The class $\mathcal{T}_{3}^{(1)}$ of L\'{e}vy measures)] We say that a L\'{e}vy measure $F$ belongs to this class if we have \begin{equation}\label{400}\quad \cases{ \displaystyle F(dx) = \widetilde{F}(dx)+\sum_{i=1}^{\infty}\frac{a_{i}\beta_{i} }{|x|^{1+\beta_{i}}} 1_{[-\eta,\eta]}(x)\,dx, \qquad \mbox{where $\eta>0$ and}\vspace*{2pt}\cr \displaystyle\qquad \mbox{\hphantom{ii}(i)\quad} 0\leq\beta_{i+1}\leq\beta_{i}<2,\qquad\beta_{i}>0\quad\Rightarrow\quad \beta_{i}>\beta_{i+1},\vspace*{2pt}\cr \displaystyle\qquad \mbox{\hphantom{(iii)}\quad}\lim_{i\rightarrow\infty}\beta_{i}=0,\vspace*{2pt}\cr \displaystyle\qquad \mbox{\hphantom{i}(ii)\quad} a_{i}>0\quad\Leftrightarrow\quad\beta_{i}>0,\vspace*{2pt}\cr \displaystyle\qquad \mbox{(iii)\quad} 0<\sum_{i=1}^{\infty}a_{i}<\infty,\vspace*{2pt}\cr \displaystyle\qquad \mbox{\hspace*{1pt}(iv)\quad} \widetilde{F} \mbox{ is a finite measure supported by }[-\eta,,\eta]^{c}.} \end{equation} \end{definition} Parts (i) and (ii) together ensure the uniqueness of the numbers $(a_{i},\beta_{i})$ in the representation of $F$, whereas if this representation holds for some $\eta>0$, it also holds\vadjust{\goodbreak} for all $\eta^{\prime} \in(0,\eta)$, with the same $(a_{i},\beta_{i})$. Part (iii) ensures that the infinite sum in the representation converges, without being zero (so equivalently, $a_{1}>0$, or $\beta_{1}>0$). The class $\mathcal{T}_{3}^{(1)}$ contains all sums of symmetric stable L\'evy measures. On the other hand, it is contained in the class of all L\'evy measures $F$ of a L\'evy process satisfying Assumption~\ref{AA1}: the latter is the class $\mathcal{T}_{3}^{(2)}$ of all $F$ such that \begin{equation} \label{4050} u\leq1 \quad\Rightarrow\quad\Biggl\vert F([-u,u]^{c}) -\sum_{i=1}^{j \frac{a_{i}}{u^{\beta_{i}}}\Biggr\vert\leq\frac{a^{\prime}} {u^{\beta _{j+1}}} \end{equation} for $2>\beta_{1}>\cdots>\beta_{j+1}\geq0$ and $a_{i}>0$ for $i=1,\ldots ,j$ and $a^{\prime}\geq0$, and those conditions are implied by~(\ref{400}), for any $j\leq\sup(i\dvtx\beta_{i}>0)$, with the same $\beta_{i}$ and~$a_{i}$. Considering $a_{i}$ and $\beta_{i}$ as functions on $\mathcal{T}_{3}^{(1)}$, the identifiability result goes as follows: \begin{theorem} \label{TLC1} In the previous setting, the following holds: \begin{longlist} \item The functions $\beta_{1}$ and $a_{1}$ are identifiable on the set $\mathcal{T}^{(1)}_{3}$. \item For any given $i\geq2$, the functions $\beta_{i}$ and $a_{i}$ are identifiable on the subset $\mathcal{T}^{(1)}_{3}(i)=\{F\in\mathcal{T} ^{(1)}_{3}\dvtx\beta_{i}(F)\geq\beta_{1}(F)/2\}$ of $\mathcal {T}^{(1)}_{3}$, and they are not on the complement $\mathcal{T}^{(1)}_{3}\setminus\mathcal {T ^{(1)}_{3}(i)$. \end{longlist} \end{theorem} \begin{remark} \label{R48} As mentioned in the ``first extension'' described in the previous section, a similar statement is true if we replace the first line of (\ref{400}) by \[ F(dx)=\widetilde{F}(dx)+\sum_{j=1}^{\infty}\biggl( \frac{a_{i}^{(+)}\beta_{i}^{(+)} }{|x|^{1+\beta_{i}}} 1_{(0,\eta] (x)+\frac{a_{i}^{(-)}\beta_{i}^{(-)} }{|x|^{1+\beta_{i}}} 1_{(-\eta ,0)}(x)\biggr)\,dx \] with both families $(\beta_{i}^{(\pm)},a_{i}^{(\pm)})$ satisfying (i)--(iii). Then the theorem above holds for both these families, with exactly the same proof. \end{remark} \begin{remark} \label{R50} As said before, any L\'{e}vy process $X$ whose L\'{e}vy measure~$F$ is in $\mathcal{T}_{3}^{(1)}$ satisfies Assumption~\ref{AA1}, but the converse is far from being true, so, even for L\'{e}vy processes, the identifiability question is not completely solved under Assumption~\ref{AA1}. More precisely, as the estimation results will show below,~(\ref{4050}) implies the ``positive'' identifiability results [(i) and the first part of (ii) of Theorem~\ref{TLC1}] for L\'{e}vy processes, but \textit{not} the ``negative'' results [second part of (ii)]. For example, consider the class $\mathcal{T}_{3}^{(3)}$ of all measure of the form \[ F(dx)=\frac{a_{1} \beta_{1}}{x^{1+\beta_{1}}} 1_{(0,1]}(x)\,dx+G(dx)\qquad \mbox{with }G = a_{2}\sum_{n\geq1}\varepsilon_{1/n^{1/\beta_{2}}}(dx) \] and $0<\beta_{2}<\beta_{1}<2$ and $a_{1},a_{2}>0$. Any such $F$ satisfies (\ref{405}), but not~(\ref{400}). On $\mathcal{T}_{3}^{(3)}$, all four parameters $\beta_{1},\beta_{2},a_{1},a_{2}$ are identifiable without the restriction $\beta_{2}\geq\beta_{1}/2$. This is of course due to the fact that the measure $G$ is singular, and any two measures $G$ and $G^{\prime}$ of the same type with $(\beta_{2},a_{2})\neq(\beta_{2}^{\prime},a_{2}^{\prime })$ have a Lebesgue decomposition $G^{\prime}=g\bullet G+G^{\prime\perp}$ with $G^{\prime\perp}(\mathbb{R})=\infty$ when $\beta_{2}\neq\beta _{2}^{\prime}$ and $\alpha(G,G^{\prime})=\infty$ when $\beta_{2}=\beta_{2}^{\prime}$ and $a_{2}\neq a_{2}^{\prime}$. We emphasize again that this example is quite singular, and verify here the fairly general principle that the less regular a statistical problem is, the easier it is to solve in the sense that more parameters can be estimated, and often with faster rates. \end{remark} \begin{remark} \label{R47} The class $\mathcal{T}_{3}^{(2)}$ may be bigger than $\mathcal{T}_{3}^{(1)}$, but it is \textit{very far} from containing all possible L\'evy measures. Indeed, any decreasing right-continuous function $f$ on $(0,\infty)$ with $f(x)\to0$ as $x\to\infty$ and $f(x)\leq K/x^{\alpha}$ for $x\in(0,1]$, for some constants $K>0$ and $\alpha\in(0,2)$, is the symmetrical tail $f(x)=F([-x,x]^{c})$ of a L\'evy measure, although of course it does not need to be equivalent to $a/x^{\beta}$ as $x\to0$ for some $\beta\in(0,2)$ and $a>0$: so~(\ref{31}) may fail even with $j=1$. \end{remark} \section{Discretely observed semimartingales: Preliminary estimators}\label{secD} Now we turn to the more general case of semimartingales. The process $X$ is observed at the times $i\Delta_{n}$ for $i=0,1,\ldots,[T/\Delta_{n}]$ (where $[x]$ denotes the integer part of the real~$x$). We thus observe the increments \begin{equation} \label{D1 \Delta_{i}^{n}X=X_{i\Delta_{n}}-X_{(i-1)\Delta_{n}}. \end{equation} The BG indices describes some properties of jumps, which are not observed. However, when an increment $\Delta_{i}^{n}X$ is relatively large, say bigger than $u_{n}$ with $u_{n}\gg\sqrt{\Delta_{n}}$, it is likely to be due to jumps because the drift plus the continuous martingale part have increments of order of magnitude $\sqrt{\Delta_{n}}$. Moreover it turns out that it is usually due to a single ``large'' jump of size bigger than $u_{n}$, although of course the observed value $\Delta^{n}_{i}X$ is not exactly the jump size. So one may expect the number of jumps with size bigger than $u_{n}=u$, over the time interval $[0,t]$, to be the following number, or be relatively close to it: \begin{equation}\label{D2 U(u,\Delta_{n})_{t}=\sum_{i=1}^{[t/\Delta_{n}]}1_{\{\Delta_{i}^{n}X>u\}}. \end{equation} In order for the previous statement to actually be true, we need some additional assumptions, though. Those are given in the following: \begin{assumption} \label{A2} The process $X$ is an It\^{o} semimartingale, and: \begin{longlist}[(a)] \item[(a)] The processes $b_{t}$, $c_{t}$ are locally bounded.\vadjust{\goodbreak} \item[(b)] We have Assumption~\ref{AA1} with $A_{t}^{i}=\int_{0}^{t} a_{s}^{i}\,ds$ for $i=1,\ldots,j+1$, where the processes $a^{i}$ are locally bounded. \item[(c)] We have $\beta_{j}>\beta_{1}/2$. \end{longlist} \end{assumption} Assumption~\ref{A2}(c) above may look strange, or too strong. However, in view of the identifiability results of the previous section, we cannot estimate consistently $\beta_{i}$ if it is strictly smaller than $\beta_{1}/2$, and as a matter of fact, the estimators described below are consistent only if $\beta_{i}>\beta_{1}/2$. Hence, since Assumption~\ref{AA1} for $j$ implies the same for all $j^{\prime}<j$, (c) above is really \textit{not} a restriction, but amounts to replacing $j$ in this assumption by $j\wedge\sup\{ i\dvtx\beta_{i}>\beta_{1}/2\} $. Apart from (c), this assumption is satisfied in Examples~\ref{E21} and \ref{E22}, and also by any L\'evy process satisfying Assumption~\ref{AA1}. The estimation procedure is a two-step procedure, and in this section we describe the first---preliminary---estimators. These estimators will be consistent, but with very slow rates of convergence. This is why, in the next subsection, we will derive final estimators which exhibit much faster (although still slow) rates. Those preliminary estimators require the knowledge of a number $\varepsilon>0$ which satisfies \begin{equation}\label{118 i=1,\ldots,j-1\quad\Rightarrow\quad\beta_{i}-\beta_{i+1}\geq\varepsilon. \end{equation} Such an $\varepsilon$ always exists, but here we suppose that it is known, somewhat in contradiction with the fact that the $\beta_{i}$ are unknown. It it is obviously quite difficult to estimate properly two contiguous indices $\beta_{i}$ and $\beta_{i+1}$ when they are very close to to one another. So from a statistical viewpoint, the assumption $\beta_{i}-\beta_{i+1 >\varepsilon$ for some fixed $\varepsilon>0$ is natural. Moreover, since we do not know a priori which $\omega$ is observed, this amounts to supposing that all possible values of the BG indices in the model satisfy this restriction. For models used in practice, this is not really a restriction since these models rely on at most a small number of indices that are separated from one another. The key ingredient for constructing the estimators is the counting process defined in~(\ref{D2}), evaluated at the terminal time $T$ and for suitable values of~$u$. In particular, we choose a sequence $u_{n}$ satisfying \begin{eqnarray}\label{D31} u_{n}\rightarrow0, \qquad\Delta_{n}^{\rho}\leq Ku_{n}\hspace*{40pt}\nonumber\\[-8pt]\\[-8pt] &&\eqntext{\mbox{with \displaystyle \rho<\frac{1}{2+\beta_{1}}\wedge\frac{2}{\beta_{1}(3+\beta_{1}) \wedge\frac{4}{\beta_{1}(5+3\beta_{1})}.} \end{eqnarray} Of course $\rho>0$ above (otherwise $u_{n}\rightarrow0$ would fail). The infimum of the upper bound for $\rho$ over all $\beta_{1}<2$ is $2/11$. Therefore, since we do not a~priori know the values of $\beta_{1}$, whereas as we will see the rates improve when the sequence $u_{n}$ becomes smaller (termwise), it is thus advisable to take $\rho =2/11$ above.\vadjust{\goodbreak} The first-step estimation is done by induction on $i$. We choose $\gamma>1$, and the estimators for $\beta_{1}$ and $A_{T}^{1}$ are \begin{eqnarray} \label{D3}\qquad \widetilde{\beta}_{1}^{n}&=&\cases{\displaystyle \frac{\log(U(u_{n},\Delta_{n})_{T}/U(\gamma u_{n},\Delta_{n})_{T})}{\log \gamma}, &\quad if $U(\gamma u_{n},\Delta_{n})_{T}>0$,\vspace*{2pt}\cr -1, &\quad otherwise,} \nonumber\\[-8pt]\\[-8pt] \widetilde{\Gamma}_{i}^{n}&=&(u_{n})^{\widetilde{\beta}_{1}^{n}} U(u_{n},\Delta_{n})_{T} \nonumber \end{eqnarray} For constructing the subsequent estimators, and with $\varepsilon$ in (\ref{118}), we set \begin{equation} \label{S10 u_{n,i}=u_{n}^{(\varepsilon/2)^{i-1}} \end{equation} (so $u_{n,1}=u_{n}$). We denote by $I(k,l)$ the set of all subsets of $\{1,\ldots,k\}$ having~$l$ elements. Assuming that we know $\widehat {\beta }_{i}^{n}$ and $\widehat{\Gamma}_{i}^{n}$ for $i=1,\ldots,k-1$, for some $k\in\{2,\ldots,j\}$, we set \begin{eqnarray}\label{D4} x&\geq&1\quad\Rightarrow\quad U^{n}(k,x)=\sum_{l=0}^{k-1}(-1)^{l} U(x \gamma ^{l} u_{n,k},\Delta_{n})_{T}\sum_{J\in I(k-1,l)}\gamma^{ \sum_{i\in J}\widetilde{\beta}_{i}^{n}},\hspace*{-35pt}\nonumber\\ \widetilde{\beta}_{k}^{n}&=&\cases{\displaystyle \frac{\log( U^{n}(k,1)/U^{n}(k,\gamma)) }{\log(\gamma)}, &\quad if $U^{n}(k,1)>0,U^{n}(k,\gamma)>0$,\vspace*{2pt}\cr -1, &\quad otherwise,}\hspace*{-35pt} \\ \widetilde{\Gamma}_{k}^{n}&=&u_{n,k}^{\widetilde{\beta}_{k}^{n}} \Biggl( U(u_{n,k},\Delta_{n})_{T}-\sum_{l=1}^{k-1}\widetilde{\Gamma}_{l}^{n} u_{n,k}^{-\widetilde{\beta}_{l}^{n}}\Biggr) .\nonumber\hspace*{-35pt} \end{eqnarray} Finally, in order to state the result, we need a further notation, for $i=1,\ldots,j-1$ \mbox{(so when $j=1$ the following is empty)}: \begin{equation} \label{S12 H_{i}=\frac{A_{T}^{i+1}}{A_{T}^{i} \log\gamma}\frac{\prod_{l=1} ^{i}( \gamma^{\beta_{l}-\beta_{i+1}}-1) }{\prod_{l=1} ^{i-1}( \gamma^{\beta_{l}-\beta_{i}}-1) }. \end{equation} \begin{theorem} \label{TD1} Under Assumption~\ref{A2} and~(\ref{118}), for all $i=1,\ldots ,j-1$ such that $\beta_{i+1}>\beta_{1}/2$, we have \begin{equation} \label{D5} \frac{\widetilde{\beta}_{i}^{n}-\beta_{i}}{u_{n,i}^{\beta_{i}-\beta_{i+1}} }\stackrel{\mathbb{P}}{\longrightarrow}-H_{i},\qquad \frac{\widetilde{\Gamma} _{i}^{n}-A_{T}^{i}}{u_{n,i}^{\beta_{i}-\beta_{i+1}} \log(1/u_{n,i} )}\stackrel{\mathbb{P}}{\longrightarrow}\Gamma_{i} H_{i}. \end{equation} Moreover if $\eta=\beta_{j}-\beta_{j+1}\vee\frac{\beta_{1}}2>0$, the following variables are bounded in probability: \begin{equation} \label{D6} \frac{\widetilde{\beta}_{j}^{n}-\beta_{i}}{u_{n,j}^{\eta}} ,\qquad \frac{\widetilde{\Gamma}_{j}^{n}-A_{T}^{j}}{u_{n,j}^{\eta} \log(1/u_{n,j})}. \end{equation} \end{theorem} The estimator $\widetilde{\beta}_{1}^{n}$ is exactly the estimator proposed in \citet{yacjacod09b} for the leading BG index $\beta_{1}$. So, not only does it satisfy~(\ref{D5}) when $j\geq2$ or the tightness of~(\ref{D6}) when $j=1$, but it also enjoys a central limit theorem centered at $\beta_{1}$ and with rate $u_{n}^{\beta_{1}/2}$ as soon as $\beta_{2}<\beta_{1}/2$ (this property implies $j=1$ here). Moreover, in this case one could prove that~$\widetilde{\Gamma}_{1}^{n}$ also satisfies a CLT with the rate $u_{n ^{\beta_{1}/2} \log(1/u_{n})$, although we will not prove it, since the emphasis here is on the case of several BG indices.\looseness=1 Some remarks are in order here: \begin{remark} \label{RS0} It is possible for the estimator $\widetilde{\Gamma }^{n}_{i}$ to be negative, in which case we may replace it by $0$, or by any other positive number. It may also happen that the sequence $\widetilde{\beta }^{n}_{i}$ is not decreasing, and we can then reorder the whole family as to obtain a decreasing family (we relabel the estimators of $A_{T}^{i}$ accordingly, of course). All these modifications are asymptotically immaterial. \end{remark} \begin{remark} \label{RS1} As mentioned in the Extension~\ref{exten2} at the end of Section \ref{secSBG}, we can relax $A^{i}_{T}>0$ in Assumption~\ref{AA1}. Then the above theorem is still valid, in restriction to the set $\Omega_{T (\{l_{1},\ldots,l_{m}\})$ of~(\ref{37}), as soon as $\beta_{l_{m} >\beta_{l_{1}}/2$. \end{remark} \begin{remark} \label{RS2} Suppose that $j\geq2$. The limits in~(\ref{D5}) are pure bias, hence precluding the existence of a proper\vspace*{1pt} central limit theorem. Note that $H_{i}>0$ if $i<j$, so the bias for $\widehat{\beta}_{i}^{n}$ and for $\widehat{\Gamma}_{i}^{n}$ are always negative and positive, respectively. Note also that the rate of convergence for estimating $\beta_{i}$ when $i\leq j-1$, say, is $u^{\beta_{i}-\beta_{i+1}}_{n,i}$, that is $u_{n}^{(\beta _{i}-\beta_{i+1})(\varepsilon/2)^{i-1}}$. This is exceedingly small, indeed. For example, suppose that we have three indices $\beta_{1}>\beta _{2}>\beta _{3}>\frac{\beta_{1}}2$. Then~(\ref{118}) implies necessarily $\varepsilon <\frac{\beta_{1}}2$, so the best\vspace*{1pt} possible rate for $i=2$ would be less than, but close to, $u_{n}^{(\beta_{2}-\beta_{3})\beta_{1}/4}$, upon taking $\varepsilon$ close to $\frac{\beta_{1}}2$, which is of course impossible because we do not know $\beta_{1}$ to start with. In the previous example, if we suspect that $\beta_{1}$ is bigger than $1$, say, it becomes (perhaps) not totally unreasonable to choose $\varepsilon =0.1$; the rates for $i=2$ and $i=3$ thus become $u_{n}^{(\beta _{2}-\beta _{3})/10}$ and $u_{n}^{(\beta_{3}-\beta_{1}/2)/100}$. This is of course on top of the fact that, because of~(\ref{D31}), $u_{n}$ is of order of magnitude $\Delta_{n}^{2/11}$, by a conservative choice of $\rho$. \end{remark} \textit{Practical considerations}. Letting aside the slow convergence rates, the previous result suffers from two main drawbacks: (1) It requires to know the number of indices to be estimated (this is implicit in Assumption~\ref{A2}). (2) It requires to know a number $\varepsilon>0$ satisfying~(\ref{118}).\vadjust{\goodbreak} About the first problem above, in real world one does not know the number of indices. On the other hand, if Assumption~\ref{AA1} holds, it seems reasonable to suppose that it holds for all $j$, whereas the estimation is made for those $\beta_{i}$ which are bigger than $\beta_{1}/2$ only. In connection with this, we assume $\beta_{i}-\beta_{i+1}\geq\varepsilon$ for all $i\leq j:= \sup(k\dvtx \beta_{k}>\beta_{1}/2)$, plus the property $\beta_{j >\beta_{1}/2+ \varepsilon$. Then the aim becomes to estimate $\beta _{i}$ and $A^{i}_{T}$ for all $i\leq j$, with $j$ unknown. Since the estimation procedure is done by induction on the successive indices, one can start the induction as described above, and stop it at the first $i$ such that $\widetilde{\beta}_{i}\leq\varepsilon+\widetilde{\beta}_{1}/2$. Asymptotically, this procedure will deliver the ``correct'' answer (the proof of this fact, not given below, is a simple extension of the proof of the second claim of the theorem). In practice, however, the solution to this stopping problem is not quite clear, since in particular the estimated sequence $\widetilde{\beta}_{i}$ is not necessarily decreasing, although it is so asymptotically. Problem 2 above is clearly more annoying. We have to admit that, in the setting presented here, we have no theoretical solution for solving it. A possible way out would be to make the estimation with several values of $\varepsilon$, going downward, until the estimated differences $\widetilde {\beta}_{i}-\widetilde{\beta}_{i-1}$ all become significantly bigger than the chosen~$\varepsilon$, but no mathematical result so far is available in this direction. In addition, since rates are very slow, the probability that such a difference is bigger than $\varepsilon$ when the true values satisfy the same inequality may be not close to $1$ (for finite, but even large, samples). Nonetheless, bad as it looks, this condition is probably relatively innocuous in practice: indeed, when two successive indices are very close to each other, they are obviously very difficult to tell apart. So the problem is practically meaningful only if the indices are a small number (as $2$, $3$ or perhaps~$4$) and reasonably well separated. Hence taking $\varepsilon =0.1$ for instance, as in Remark~\ref{R47}, seems to be safe enough. \section{Discretely observed semimartingales: An improved method}\label{secD2} The observation scheme is the same as in the previous section: $X$ is observed at the times $i\Delta_{n}$ smaller or equal to some fixed terminal time $T$. As already mentioned, the previous estimators converge at a \textit{very} slow rate, especially for higher order indices; see Remark~\ref{RS2}. So, in order to implement the estimation with any kind of reasonable accuracy, it is absolutely necessary to come up with better estimators. This is the aim of this section. Assuming Assumption~\ref{A2}, we also suppose that we can construct preliminary estimators, such as in the previous section. Exactly as there, we must know the number $j$ of BG indices that are to be estimated. The method consists in minimizing, at each stage $n$, a suitably chosen contrast function $\Phi_{n}$. First we take an integer $L\geq2j$ and numbers $1=v_{1}<v_{2}<\cdots<v_{L}$. We also choose positive weights $w_{k}$\vadjust{\goodbreak} (typically $w_{k}=1$, but any choice is indeed possible), and we pick truncation levels $u_{n}$ satisfying~(\ref{D31}). We also let $D$ be the set of all $(x_{i},\gamma_{i})_{1\leq i\leq j}$ with $0\leq x_{j}\leq x_{j-1} \leq\cdots\leq x_{1}\leq2$ and $\gamma_{i}\geq0$. Then the contrast function is defined on $D$ by \begin{equation} \label{D2-2 \Phi_{n}(x_{1},\gamma_{1},\ldots,x_{j},\gamma_{j})=\sum_{l=1}^{L} w_{l}\Biggl( U(v_{l}u_{n},\Delta_{n})_{T}-\sum_{i=1}^{j}\frac{\gamma_{i} }{(v_{l}u_{n})^{x_{i}}}\Biggr) ^{2}, \end{equation} where the sequence $u_{n}$ satisfies~(\ref{D31}). Then the estimation goes as follows: \begin{Step}\label{step1} We construct preliminary estimators $\widetilde{\beta}^{n}_{i}$ (decreasing in $i$) and~$\widetilde{\Gamma}^{n}_{i}$ (nonnegative) for $\beta_{i}$ and $A_{T}^{i}$ for $i=1,\ldots,j$, such that $(\widetilde {\beta }_{i}-\beta_{i}) /u_{n}^{\eta}$ and $(\widetilde{\Gamma}_{i}-A_{T}^{i} )/u_{n}^{\eta}$ go to $0$ in probability for some $\eta>0$. For example, we may choose those described in the previous section (see Remark~\ref{RS0}): the consistency requirement is fulfilled for any $\eta<(\varepsilon/2)^{j}$. \end{Step} \begin{Step}\label{step2} We denote\vspace*{1pt} by $D_{n}$ the (compact and nonempty) random subset of $D$ defined by $D_{n}=\{(x_{i},\gamma_{i})\in D\dvtx |x_{i}-\widetilde{\beta}^{n} _{i}|\leq\alpha u_{n}^{\eta},|\gamma_{i}-\widetilde{\Gamma}^{n}_{i} |\leq\alpha u_{n}^{\eta},\forall i=1,\ldots,j\}$, for some arbitrary (fixed) $\alpha>0$. Then the final estimators $\overline{\beta}{}^{n}_{i}$ and~$\overline{\Gamma}{}^{n}_{i}$ will be \begin{equation} \label{D2-3} (\overline{\beta}{}^{n}_{i},\overline{\Gamma}{}^{n}_{i})_{1\leq i\leq j}=\mathop{\arg\min}_{D_{n}}\Phi_{n}(x_{1},\gamma_{1},\ldots ,x_{j},\gamma_{j}). \end{equation} \end{Step} \begin{theorem} \label{TD2} Under Assumption~\ref{A2}, and for all choice of $v_{2} ,\ldots,v_{L}$ outside a $\lambda_{L-1}$-null set (depending on the $\beta _{i}$'s; $\lambda_{l}$ is the $l$-dimensional Lebesgue measure), the sequences \begin{equation} \label{D2-4} \frac{\overline{\beta}{}^{n}_{i}-\beta_{i}}{u_{n}^{\beta_{i -\beta_{1}/2-\mu}},\qquad \frac{\overline{\Gamma}{}^{n}_{i}-\Gamma_{i} {u_{n}^{\beta_{i}-\beta_{1}/2-\mu} \end{equation} are bounded in probability for all $i=1,\ldots,j$ and all $\mu>0$. \end{theorem} The rates obtained here are much faster than in Theorem~\ref{TD1}: we replace $u_{n,i}^{\beta_{i}-\beta_{i+1}\vee(\beta_{1}/2)}$ by $u_{n}^{\beta_{i -\beta_{1}/2}$, for two reasons: the exponent $\beta_{i}-\beta_{i+1}$ is bigger than $\beta_{i}-\beta_{i+1}\vee(\beta_{1}/2)$, unless $i=j$; more importantly, we replace the auxiliary truncation levels $u_{n,i}$ of (\ref{S10}) by the original sequence $u_{n}$, which is much smaller when $i\geq2$, and only subject to~(\ref{D31}). We will examine in the next section how far from optimality those rates are. \begin{remark} \label{R40} As stated, and as seen from the proof, we only need $L=2j$, and choosing $L>2j$ does not improve the asymptotic properties. However, from a practical viewpoint, it is probably wise to take $L$ bigger than $2j$ in order to smooth out the contrast function somehow, especially for (relatively) small samples. A~choice of the weights $w_{l}>0$ other than $w_{l}=1$, such as~$w_{l}$ decreasing in $l$, may serve to put less emphasis on the large truncation values $u_{n}v_{l}$ for which less data are effectively used.\vadjust{\goodbreak} \end{remark} \begin{remark} \label{R41} The result does not hold (or at least we could not prove it) for \textit{all} choices of the $v_{l}$'s, but only when $(v_{2},\ldots,v_{L})$ (recall $v_{1}=1$) does not belong to some Lebesgue-null set $G(\beta _{1},\ldots,\beta_{j})$. This seems a priori a~serious restriction, because $(\beta_{1},\ldots,\beta_{j})$ is unknown. In practice, we choose a priori $(v_{2},\ldots,v_{L})$, so we may have bad luck, just as we may have bad luck for the outcome $\omega$ which is drawn$\ldots.$ We may also do the estimation for a number of different choices for the weights and/or values of $L\geq2j$ and compare or average the results. This should contribute to weaken the numerical instability inherent to minimization problems such as~(\ref{D2-3}). This numerical instability is similar to the one occurring in nonlinear regression problems. We have to state, however, that these problems, just as those stated in the ``practical considerations'' of the previous section, are not fully addressed in this paper, and they are probably quite difficult to overcome. Our emphasis here is more on theoretical results, and on the possibility of performing the estimation with reasonable rates (see, however, Section~\ref{secMC} below, to see how the problem of finding a ``good'' $\varepsilon$ and doing preliminary estimation in our simulation study is skipped, without affecting the quality of the procedure in any noticeable way).\vspace*{-1pt} \end{remark} \section{Optimality in a special case}\label{secFisher}\vspace*{-1pt} \subsection{Why the convergence rates are necessarily slow} Intuitively, the fact that we are right at the boundary between identifiability and lack thereof suggests that we should expect the rate, as we approach the loss of identifiability boundary, to deteriorate all the way to zero. In order to quantify precisely how slow the rates of convergence for the estimators of the second (and higher) index must be, even in ideal circumstances, we study a simple parametric model of the following form. Let $W$ be a Brownian motion and $Y^{1},Y^{2}$ be two independent standard symmetric stable processes, and set \begin{equation} \label{eqFisher+2Stable X_{t}=bt+\sigma W_{t}+Y_{t}^{1}+Y_{t}^{2}. \end{equation} Each $Y^{i}$ depends on two parameters, the index $\beta_{i}$ and a scale parameter~$a_{i}$, the latter being characterized by the fact that the L\'evy measure of~$Y^{i}$ is \begin{equation} \label{LR4-7} F^{j}(dx)=\frac{a_{j} \beta_{j}}{|x|^{1+\beta_{j}}}\,dx. \end{equation} We have six parameters, \begin{equation} \label{IJ2} b\in\mathbb{R},\qquad c=\sigma^{2}>0,\qquad a_{1},a_{2 >0,\qquad 0<\beta_{2}<\beta_{1}<2, \end{equation} among which $b$ is not identifiable, and $c,\beta_{1},a_{1}$ are identifiable, and $(\beta_{2},a_{2})$ are identifiable if and only if $\beta_{2}\geq \beta_{1}/2$. In what follows, we restrict our attention to the four parameters $\beta_{1},\beta_{2},a_{1},a_{2}$. In order to find at which rate it is possible to estimate these four parameters, when $X$ is observed\vadjust{\goodbreak} at the discrete times $(i\Delta _{n}\dvtx i=0,1,\ldots,[T/\Delta_{n}])$ and $\Delta_{n} \rightarrow0$, we study the behavior of the Fisher information matrix. Due to the fact that $X$ is a L\'{e}vy process, the information matrix at stage $n$ is $[T/\Delta_{n}]$ times the information matrix obtained when we observe only the variable~$X_{\Delta_{n}}$; since the variable $X_{\Delta}$ admits a density $x\mapsto p(_{\Delta}(x|c,\beta_{1},a_{1},\allowbreak \beta_{2},a_{2})$ which is $C^{\infty}$ in $x$, and also in $(c,\beta_{1},a_{1},\beta_{2},a_{2})$ on the domain defined by~(\ref{IJ2}), it is no wonder that Fisher's information $I_{\Delta}$ for a single observation~$X_{\Delta}$ (recall $X_{0}=0$) exists, and we can study its behavior as $\Delta\rightarrow0$. Only the diagonal entries are important for the various rates of convergence, so we only need to focus on the following diagonal entries of this matrix: \[ I_{\Delta}^{\beta_{1}\beta_{1}},I_{\Delta}^{a_{1}a_{1}},I_{\Delta ^{\beta_{2}\beta_{2}}, I_{\Delta}^{a_{2}a_{2}}. \] The main result of this section follows, giving the asymptotic order of the relevant terms in Fisher's information: \begin{theorem} \label{theo-fisher}We have the following equivalences, as $\Delta \rightarrow 0$: \begin{eqnarray*} I_{\Delta}^{\beta_{1}\beta_{1}} & \sim & \frac{a_{1} }{2(2-\beta_{1 )^{\beta_{1}/2} c^{\beta_{1}/2}} \Delta^{1-\beta_{1}/2} (\log (1/\Delta))^{2-\beta_{1}/2},\\ I_{\Delta}^{a_{1}a_{1}} & \sim & \frac{2\beta_{1}c_{\beta_{1}} a_{1} ^{\beta_{1}}}{(2-\beta_{1})^{\beta_{1}/2} \sigma^{\beta_{1}} a_{1}^{2} } \frac{\Delta^{1-\beta_{1}/2}}{(\log(1/\Delta))^{\beta_{1}/2} \end{eqnarray*} and also, provided $\beta_{2}>\beta_{1}/2$, \begin{eqnarray*} I_{\Delta}^{\beta_{2}\beta_{2}} & \sim &\frac{a_{2}^{2} \beta_{2}^{2}} {2a_{1} \beta_{1}(2 \beta_{2}-\beta_{1})(2-\beta_{1})^{\beta_{2}-\beta_{1}/2} c^{\beta_{2}-\beta_{1}/2}}\\ &&{}\times \Delta^{1-\beta_{2}+\beta_{1}/2} (\log (1/\Delta))^{2-\beta_{2}+\beta_{1}/2},\\ I_{\Delta}^{a_{2}a_{2}} & \sim & \frac{2\beta_{2}^{2}} {a_{1} \beta _{1}(2\beta_{2}-\beta_{1})(2- \beta_{1})^{\beta_{2}-\beta_{1}/2} c^{\beta _{2}-\beta_{1}/2}} \frac{ \Delta^{1-\beta_{2}+\beta_{1}/2}}{(\log (1/\Delta))^{\beta_{2}-\beta_{1}/2}}. \end{eqnarray*} \end{theorem} \begin{remark} We are not concerned here with the identification and estimation of the volatility parameter $c$; the term $I_{\Delta}^{cc}$ in a simpler model has been studied in \citet{yacjacod08}, as well as $I_{\Delta}^{a_{1}a_{1}}$ when $a_{2}=0$ (i.e., when there is only one stable process on top of the Brownian motion). The asymptotic equivalent for the term $I_{\Delta}^{a_{1}a_{1}}$ of course reduces to (4.11) of that paper, with $\alpha=\beta_{1}$, $\beta=2$, $\theta=a_{1}$, up to a change of parametrization for $a_{1}$, since here we use the parametrization~(\ref{LR4-7}) which corresponds to the notation of Assumption~\ref{AA1}, which is fulfilled here. \end{remark} Coming back to the original problem, we deduce that it should be possible in principle to find estimators $\widehat{\beta}_{i}^{n}$ and $\widehat{a} _{i}^{n}$ having the following properties: \begin{eqnarray} \label{119} \frac{(\log(1/\Delta_{n}))^{1-\beta_{1}/4}}{\Delta_{n}^{\beta_{1}/4} } (\widehat{\beta}_{1}^{n}-\beta_{1}) & \stackrel{\mathcal {L}}{\longrightarrow } & \mathcal{N}(0,1/T\mathcal{I}^{\beta_{1}\beta_{1} }),\nonumber\\ \frac{1}{\Delta_{n}^{\beta_{1}/4} (\log(1/\Delta_{n}))^{\beta_{1}/4} } (\widehat{a}_{1}^{n}-a_{1}) & \stackrel{\mathcal{L}}{\longrightarrow} & \mathcal{N}(0,1/T\mathcal{I}^{a_{1}a_{1}}),\nonumber\\[-8pt]\\[-8pt] \frac{(\log(1/\Delta_{n}))^{1-\beta_{2}/2+\beta_{1}/4}}{\Delta _{n}^{\beta _{2}/2-\beta_{1}/4}} (\widehat{\beta}_{2}^{n}-\beta_{2}) & \stackrel {\mathcal{L}}{\longrightarrow} & \mathcal{N}(0,1/T\mathcal{I}^{\beta_{2} \beta_{2}}),\nonumber\\ \frac{1}{\Delta_{n}^{\beta_{2}/2-\beta_{1}/4} (\log(1/\Delta _{n}))^{\beta _{2}/2-\beta_{1}/4}} (\widehat{a}_{2}^{n}-a_{2}) & \stackrel{\mathcal{L} }{\longrightarrow} & \mathcal{N}(0,1/T\mathcal{I}^{a_{2}a_{2}}), \nonumber \end{eqnarray} where $\mathcal{I}^{\beta_{1}\beta_{1}}$, $\mathcal{I}^{a_{1}a_{1}}$, $\mathcal{I}^{\beta_{2}\beta_{2}}$ and $\mathcal{I}^{a_{2}a_{2}}$ are the constants in front of the term involving $\Delta$ in the equivalences above, for $I_{\Delta}^{\beta_{1}\beta_{1}}$, $I_{\Delta}^{a_{1}a_{1}}$, $I_{\Delta }^{\beta_{2}\beta_{2}}$ and $I_{\Delta}^{a_{2}a_{2}}$, respectively. Conversely, by the Cram\'er--Rao lower bound, Theorem~\ref{theo-fisher} also implies that it will be impossible to find consistent estimators with faster rates of convergence, or smaller asymptotic variance, that those exhibited in (\ref{119}). Note that these rates are consistent with the results of Theorem~\ref{TLC1}. The first two convergences above shows that it is always possible to estimate consistently $\beta_{1}$ and $a_{1}$, the third one implies consistency for $\beta_{2}$ only if $\beta_{2}\geq\beta_{1}/2$, and the last one implies consistency for $a_{2}$ only if $\beta_{2}>\beta_{1}/2$. The last statement seems contradictory with Theorem~\ref{TLC1} when $\beta_{2}=\beta _{1}/2$, but of course it is possible to have a (somewhat irregular) statistical model for which consistency holds even though the Fisher information does not go to infinity. \subsection{Comparison of rates} Now, we can compare these optimal rates with the rates obtained in Theorem \ref{TD2}. Doing as such, we compare a semiparametric model with a parametric sub-model. However, a minimax rate for a given parameter in a semiparametric model cannot be faster than the rate obtained for any parametric sub-model, hence the previous results are bounds for the rates in the general model considered in this paper. Neglecting the logarithmic terms, and considering only the estimation of~$\beta_{i}$ for $i=1,2$, the rates above are $\Delta_{n}^{\gamma_{i}}$, whereas in Theorem~\ref{TD2}, and upon choosing $u_{n}$ optimally [i.e., $\rho$ as large as possible in~(\ref{D31})], they are $\Delta_{n ^{\gamma^{\prime}_{i}}$, where \[ \gamma_{i}=\frac{2\beta_{i}-\beta_{1}}4,\qquad \gamma^{\prime}_{i}=\cases{ \displaystyle \gamma_{i} \frac{2}{2+\beta_{1}}-\varepsilon, &\quad if $\beta_{1 \leq\bigl(\sqrt{97}-1\bigr)/6\approx1.475$,\vspace*{2pt}\cr \displaystyle \gamma_{i} \frac{8}{5\beta_{1}+3\beta_{1}^{2}}-\varepsilon, &\quad if $\beta_{1}\geq\bigl(\sqrt{97}-1\bigr)/6$,} \] and $\varepsilon>0$ arbitrarily small (and if $\beta_{i}>\beta_{1}/2$ when $i=2$). As it should be, we have $\gamma_{i}\leq\gamma_{i}^{\prime}$, and if equality were holding we would conclude that our estimators achieve the minimax rate (up to $\Delta_{n}^{-\varepsilon}$, of course, but $\varepsilon$ is arbitrarily small). What one can say is that the actual minimax rate lies somewhere in between these two values, and the ratio $\gamma_{i}/\gamma _{i}^{\prime}$ is a kind of (imperfect) measure of the quality of the estimators proposed in Section~\ref{secD2}: the closest to $1$, the closest to optimality. Then we can conclude the following: (a) This ratio is the same for $j=1,2$, which is an a priori surprising result: the quality of our estimator for $\beta_{2}$, relative to the optimal estimators in the stable sub-model, is the same as for $\beta_{1}$. (b) This ratio is close to $1$ (near optimality) when $\beta_{1}$ is small, and decreases down to $4/11$ as $\beta_{1}$ increases up to $2$. The worst value is small, but not catastrophically such, especially in light of the fact that we are considering semiparametric estimators whereas the rates are optimal in the parametric context (i.e., assuming additional structure). \section{Simulation results}\label{secMC} We now provide some simulation evidence regarding the estimators in the case where $j=2$; we are attempting to estimate the first two jump activity indices of the process $\beta_{1}$ and $\beta_{2}$. The data generating process is a stochastic volatility model for $X_{t}$ with jumps driven by two stable processes $Y^{1}$ and $Y^{2}$, with $W,Y^{1},Y^{2}$ independent below: \begin{equation}\label{eqMCsde dX_{t}=\sigma_{t}\,dW_{t}+\theta_{1}\,dY_{t}^{1}+\theta_{2}\,dY_{t}^{2} \end{equation} with $\sigma_{t}=v_{t}^{1/2}$, $dv_{t}=\kappa(\eta-v_{t})\,dt+\gamma v_{t} ^{1/2}\,dB_{t}+dJ_{t}$, $\mathbb{E}[dW_{t}\,dB_{t}]=\rho \,dt$, $\eta^{1/2}=0.25$, $\gamma=0.5$, $\kappa=5$, $\rho=-0.5$, the volatility jump term $J$ is a compound Poisson jump process with jumps that are uniformly distributed on $[-0.3,0.3]$ and intensity $\lambda=10$ and $X_{0}=1$. Recall that the second component can be identified only if $\beta_{2}>\beta_{1}/2$. We consider the situation where $(\beta_{1},\beta_{2})=(1.00,0.75)$. Given $\eta$, each scale parameter $\theta_{i}$ (or equivalently $A_{T}^{i}$) of the stable process in simulations is calibrated to deliver different various values of the tail probability $P_{i}=\mathbb{P}(|\Delta Y_{t ^{i}|\geq4\eta^{1/2}\Delta_{n}^{1/2})$. In the various simulations' design, we hold $\eta$ fixed and consider the cases where $P_{1}=0.05$ and $P_{2}=0.005$. We sample the process $X$ over $T=21$ days ($6.5$ hours per day) every $\Delta_{n}=0.01$ second. This results of course in a number of observations (nearly $5\times10^{7}$) that is unrealistically high for most high-frequency financial data series, at least presently, but extremely large numbers of observations are needed if we are going to be able to see the component $\beta_{2}$ of the model ``behind'' the two components with indices of activity $2$ (the continuous component) and $\beta_{1}$ (the most active jump component). Of course, much smaller datasets would be sufficient in the absence of a continuous component Note that in general, and besides the preliminary estimators $\widetilde {\beta}^{n}_{i}$ and $\widetilde{\Gamma}^{n}_{i}$, we need to choose the number $\alpha>0$ coming in the definition of the set $D_{n}$. Since in practice $n$ (or $\Delta_{n}$) is given, we need to choose in fact the number $\alpha u_{n}^{\eta}$. So in concrete situations one probably can forget about the preliminary estimators and take a domain $D_{n}$ which is the set of all $(x_{i},\gamma_{i})$ in $D$ with $\gamma_{i}\leq A$ for some ``reasonably chosen'' $A$, or even $A=\infty$. This is what we do below, by taking the estimators to be \begin{eqnarray}\label{eqargmin4} && (\overline{\beta}{}^{\prime n}_{1},\overline{\beta}{}^{\prime n}_{2 ,\overline{\Gamma}{}^{\prime n}_{1},\overline{\Gamma}{}^{\prime n}_{2 )\nonumber\\[-8pt]\\[-8pt] &&\qquad=\mathop{\arg\min}_{( x_{1},\gamma_{1},x_{2},\gamma_{2}) }\sum_{l=1}^{L}\biggl( U(v_{l}u_{n},\Delta_{n})_{T}-\frac{\gamma_{1} {(v_{l}u_{n})^{x_{1}}}-\frac{\gamma_{2}}{(v_{l}u_{n})^{x_{2}}}\biggr) ^{2},\nonumber \end{eqnarray} where the cutoff levels $v_{l}u_{n}$ are chosen in terms of the number $\alpha_{l}$ of the long-term standard deviation $\sqrt{\eta\Delta_{n}}$ over a time lag $\Delta_{n}$ of the continuous martingale part of the process: we take $\alpha_{l}$ to be $\{ 7,10,15,20\} $ and multiples $\{ 2,4,6\} $ thereof (giving all together $L=10$ distinct values). Here we know $\eta$: we could also estimate for each path the average volatility, using truncated estimators for the integrated volatility [see, e.g., \citet{mancini04} and \citet{yacjacod09a}]. The optimization problem~(\ref{eqargmin4}) is a quadratic problem similar to classical nonlinear least squares minimization. In situations where the parameter space is high dimensional, the objective function can exhibit local extrema, which can make the search for the optimal solution time-consuming as many starting values must be employed to validate the solution. In the case of the application here, we are only including $4$ parameters, and for this small dimension, this is not causing many difficulties. In any case, it is unlikely, given the slow rates of convergence, that one would want to go beyond the second index $\beta_{2}$ in practice. \begin{figure} \includegraphics{976f02.eps} \caption{Monte Carlo simulation results: estimators $\widehat{\beta}_{1}^{n}$ (upper left graph), $\widehat{\beta}_{2}^{n}$ (upper right graph), $\widehat{A _{T}^{1,n}$ (lower left graph), $\widehat{A}_{T}^{2,n}$ (lower right graph). \label{figHistBeta1and2} \end{figure} The results in Figure~\ref{figHistBeta1and2} are obtained with $M=1000$ simulations: the estimators appear to be reasonably good, but then again this is for an unrealistically large number of observations, at least from the point of view of financial applications; it is perhaps feasible in other applications, such as Internet data traffic or wind measurement. \section{Conclusions}\label{conclusion} This paper determined theoretically what the successive BG indices are and how they are identified, including the perhaps surprising theoretical bound on the identification of the successive indices as a function of the previous ones. This result clarifies the border between the aspects of the jump measure which are identifiable from those which are not on the basis of discrete observations on a finite time horizon. Beyond the leading index, the identification requires in practice vast quantities of data which are out of reach of financial applications at present but may be relevant in other fields (such as the study of turbulence data, or Internet traffic). We showed through explicit calculations of Fisher's information that this limitation is a genuine, inescapable feature of the problem. There are a number of important questions that this paper does not touch upon: central limit theorems for the estimators, estimators that achieve the optimal rates of convergence, estimators that are robust to microstructure noise, estimators that are applicable with random sampling intervals, among others. The issue of the optimality of the rates in general remains an open question. \begin{appendix}\label{app} \section*{Appendix: Proofs} We use the following notation throughout the \hyperref[app]{Appendix}. First, $K$ denotes a constant which may change from line to line, and may depend on the characteristics or the law of the processes at hand. It never depends on $n$, and it is denoted as $K_{p}$ if it depends on an additional parameter $p$. Second, for any sequence $Z_{n}$ of variables and any sequence $v_{n}$ of positive numbers, \begin{equation}\label{S8} Z_{n}=\cases{ O_{P}(v_{n}), &\quad if $Z_{n}/v_{n}$ is bounded in probability,\cr o_{P}(v_{n}), &\quad if $Z_{n}/v_{n}\stackrel{\mathbb{P}}{\longrightarrow }0$.}\ \end{equation} \setcounter{section}{0} \section{\texorpdfstring{Proof of Theorem \protect\ref{TLC1}}{Proof of Theorem 1}} (1) We fix\vspace*{1pt} $F\in\mathcal{T}_{3}^{(1)}$, with $F$ given by~(\ref{400}). We also consider another $F^{\prime}\in\mathcal{T}_{3}^{(1)}$, with $F^{\prime }$ given by~(\ref{400}) with $\beta_{i}^{\prime}$, $a_{i}^{\prime}$ and $\widetilde {F}^{\prime}$. As said before, it is not a restriction to assume the representation~(\ref{400}) with the same $\eta>0$ for both~$F$ and $F^{\prime }$. Set \begin{equation}\label{405 j=\inf\bigl(1\leq1\dvtx (\beta_{i},a_{i})\neq(\beta_{i}^{\prime},a_{i}^{\prime})\bigr). \end{equation} The result amounts to proving the following two properties, with $j$ as above and $b,b^{\prime}\in\mathbb{R}$ and $c,c^{\prime}\geq0$: \begin{eqnarray} \label{402 \beta_{j}&\geq&\frac{\beta_{1}}{2} \quad\Rightarrow\quad Q_{b,c,F}\perp Q_{b^{\prime},c^{\prime},F^{\prime}},\hspace*{-35pt} \\ \label{403 \beta_{j}&<&\frac{\beta_{1}}{2}\quad\Rightarrow\quad\cases{ \exists b^{\prime\prime}\in\mathbb{R},\exists F^{\prime\prime}\in \mathcal{T}_{3}^{(1)}\cr \qquad\mbox{with }F^{\prime\prime}=F^{\prime}\mbox{ on }[-\eta,\eta)\mbox{ and Q_{b,c,F}\not\perp Q_{b^{\prime\prime},c,F^{\prime}}.}\hspace*{-35pt} \end{eqnarray} These conditions being symmetrical in $F$ and $F^{\prime}$, in both (\ref{402}) and~(\ref{403}) we may assume \begin{equation} \label{404 \mbox{either }\beta_{j}>\beta_{j}^{\prime}\quad\mbox{or}\quad\beta_{j}=\beta _{j}^{\prime}\quad\mbox{and}\quad a_{j}>a_{j}^{\prime}. \end{equation} (2) In this step we assume~(\ref{404}). We set \[ \widehat{F}(dx)=\sum_{i\geq1}\frac{a_{i}\beta_{i}}{|x|^{1+\beta_{i}} } 1_{[-\eta,\eta]}(x)\,dx,\qquad \widehat{F}^{\prime}(dx)= \sum_{i\geq1 \frac{a_{i}^{\prime}\beta_{i}^{\prime}} {|x|^{1+\beta_{i}^{\prime} } 1_{[-\eta,\eta]}(x)\,dx. \] Then $\widehat{F}^{\prime}=f\bullet\widehat{F}$, where $f=\frac {g^{\prime} {g}$ (with $\frac00=1$) and $g=H+G$ and $g^{\prime}=H+G^{\prime}$ and \begin{eqnarray*} H(x)&=&\sum_{i=1}^{j-1}\frac{a_{i}\beta_{i}}{|x|^{\beta_{i}}} 1_{[-\eta ,\eta ]}(x),\qquad G(x)=\sum_{i\geq j}\frac{a_{i}\beta_{i}}{|x|^{\beta_{i}} 1_{[-\eta,\eta]}(x),\\ G^{\prime}(x)&=&\sum_{i\geq j}\frac{a_{i}^{\prime \beta_{i}^{\prime}}{|x|^{\beta_{i}^{\prime}}} 1_{[-\eta,\eta]}(x). \end{eqnarray*} On $[-\eta,\eta]$ we have $f-1=\frac{G^{\prime}-G}{H+G}$ and \begin{eqnarray*} && G(x)-G^{\prime}(x)\\ &&\qquad=\frac{a_{j}\beta_{j}}{|x|^{1+\beta_{j}}}\biggl( 1-\frac{a_{j}^{\prime}\beta_{j}^{\prime}}{a_{j}\beta_{j}} |x|^{\beta_{j} -\beta_{j}^{\prime}}+\sum_{i\geq j+1}\frac{a_{i}\beta_{i}}{a_{j}\beta_{j} } x^{\beta_{j}-\beta_{i}}-\sum_{i\geq j+1}\frac{a_{i}^{\prime} \beta _{i}^{\prime}}{a_{j}\beta_{j}} |x|^{\beta_{j}-\beta_{i}^{\prime}}\biggr) . \end{eqnarray*} By virtue of (ii), (iii) and (iv) of~(\ref{400}), and of~(\ref{404}), we then deduce that \begin{equation}\label{406 x\in(-\varepsilon,\varepsilon) \quad\Rightarrow\quad\cases{ \displaystyle A_{-} |x|^{\beta_{1}-\beta_{j}}\leq|f(x)-1|\leq A_{+} |x|^{\beta_{1} -\beta_{j}},\vspace*{2pt}\cr \displaystyle \frac{A_{-}}{|x|^{1+\beta_{1}}} \leq g(x) \leq \frac{A_{+}}{|x|^{1+\beta_{1}}},} \end{equation} for three constants $A_{+}>A_{-}>0$ and $\varepsilon\in(0,\eta)$, depending on the two sequences $(\beta_{i},a_{i})$ and $(\beta_{i}^{\prime },a_{i}^{\prime })$. (3) Now we\vspace*{1pt} prove~(\ref{402}). Since $\alpha(F,F^{\prime})\geq\alpha (\widehat {F}, \widehat{F}^{\prime})$, it is enough to show that $\alpha(\widehat {F},\widehat{F}^{\prime}) =\infty$. By~(\ref{406}), $|f(x)-1|\leq1$ when $x\in(-\varepsilon^{\prime},\varepsilon^{\prime})$ for some $\varepsilon ^{\prime}\in(0,\varepsilon]$. Thus \[ \alpha(\widehat{F},\widehat{F}^{\prime}) \geq \int_{-\varepsilon }^{\varepsilon^{\prime} }|f(x)-1|^{2} g(x)\,dx \geq A_{-}^{3} \int _{-\varepsilon^{\prime}}^{\varepsilon^{\prime}} |x|^{\beta_{1}-2\beta _{j -1}\,dx. \] The last integral is infinite when $\beta_{j}\geq\beta_{1}/2$, and~(\ref{403}) follows by~(\ref{LC1}).\vspace*{1pt} (4) Finally we prove~(\ref{403}). Recall that $F=\widehat{F}+\widetilde {F}$ and $F^{\prime}=\widehat{F}^{\prime}+\widetilde{F}^{\prime}$. The measure $F^{\prime\prime}=\widehat{F}^{\prime}+\widetilde{F}$ is obviously in $\mathcal{T}^{(1)}_{3}$ and satisfies $F^{\prime\prime}=f\bullet F$. Since $f(x)=1$ outside $[-\eta,\eta]$, the quantity $\alpha^{\prime }(F,F^{\prime \prime})$ introduced in~(\ref{LC1}) is \begin{eqnarray*} \alpha^{\prime}(F,F^{\prime\prime}) & = & \int_{-\eta}^{\eta x\bigl(f(x)-1\bigr)g(x)\,dx\\ & \leq & A_{+}^{2}\int_{-\varepsilon}^{\varepsilon}|x|^{-\beta_{j}}\,dx +\biggl(\int_{\varepsilon}^{\eta}+\int_{-\eta}^{-\varepsilon \biggr) |x| |f(x)-1| g(x)\,dx, \end{eqnarray*} which is finite by~(\ref{406}) (because $\beta_{j}<\beta_{1}/2<1$) and because $f$ and $g$ are bounded on $[\varepsilon,\eta]\cup[-\eta,-\varepsilon]$. Therefore the number $b^{\prime\prime}=b-\int_{0}^{\eta\wedge1 x(f(x)-1)g(x)\,dx$ is well defined. Now we consider the two triples $(b,c,F)$ and $(b^{\prime\prime},c,F^{\prime\prime})$. From what precedes they satisfy the first and the last three properties in~(\ref{LC1}). We also have by~(\ref{406}) \begin{eqnarray*} \alpha(F,F^{\prime\prime}) & = & \int_{-\eta}^{\eta}\bigl(|f(x)-1|^{2 \wedge|f(x)-1|\bigr) g(x)\,dx\\ & \leq & A_{+}^{2}\int_{-\varepsilon}^{\varepsilon} \bigl( |x|^{-\beta _{j}-1}\wedge( A_{+}|x|^{\beta_{1}-2\beta_{j}-1}) \bigr) \,dx\\ &&{} +\biggl(\int_{\varepsilon}^{\eta}+\int_{-\eta}^{-\varepsilon }\biggr) \bigl(|f(x)-1|^{2}\wedge|f(x)-1|\bigr) g(x)\,dx. \end{eqnarray*} Since $\beta_{j}<\beta_{1}/2$ and that $f$ and $g$ are bounded on $[\varepsilon,\eta]\cup[-\eta,-\varepsilon]$, we deduce $\alpha (F,F^{\prime \prime})<\infty$. So all conditions in~(\ref{LC1}) are satisfied, and we have proved~(\ref{403}). \section{Comparing big jumps and big increments} Before starting, let us mention that for the proofs of Theorems \ref {TD1} and \ref{TD2} one may use a localization argument which allows us to replace Assumption~\ref{A2} by the so-called ``strengthened Assumption~\ref{A2},'' which is the same except that all processes $b_{t}$, $c_{t}$, $a^{i}_{t}$ are bounded, as well as the process $A^{j+1}_{t}$ and $X_{t}$ itself. In this section we compare the number of ``large'' increments of $X$ with the number of correspondingly large jumps, that is, the numbers \begin{equation} \label{S3} V(u)_{t} = \sum_{s\leq t} 1_{\{|\Delta X_{s}|>u\}} \end{equation} We will indeed show that the difference $U(u_{n},\Delta_{n})_{T}- V(u_{n )_{T}$ is negligible for our purposes, when the sequence $u_{n}$ satisfies (\ref{D31}). The reason for doing this is that the analysis of the processes $V(u_{n})$ is an easy task. Indeed, as soon as $u_{n}\to0$, \begin{equation} \label{KP} V(u_{n})_{T}-\overline{A}(u_{n})_{T} = O_{P}(u_{n}^{-\beta_{1}/2}). \end{equation} To see this, we observe that each process $M^{n}=u_{n}^{\beta_{1}/2}( V(u_{n})-\overline{A}(u_{n})) $ is a~quasi-left continuous, purely discontinuous, martingale with jumps smaller than $u_{n}^{\beta _{1}/2}$, which goes to $0$. Its predictable quadratic variation is $\langle M^{n} ,M^{n}\rangle=u_{n}^{\beta_{1}} \overline{A}(u_{n})$, which by~(\ref{31}) converges for each $t$ to $A_{t}^{1}$. Since further $A^{1}$ is a~continuous process, it follows from Theorem VI.4.13 of \citet{jacodshiryaev2003}, for example, that the sequence $M^{n}$ is $C$-tight (and even converges in law), so a fortiori,~(\ref{KP})~holds. The main result of this section is the next proposition: \begin{proposition} \label{PD1} Under the strengthened Assumption~\ref{A2} and if the sequence $u_{n}$ satisfies~(\ref{D31}), we have \begin{equation} \label{D30 U(u_{n},\Delta_{n})_{T}-V(u_{n})_{T} = \frac{1}{u_{n}^{\beta_{1}}} O_{P}(u_{n}^{\beta_{1}-\beta_{j+1}}+u_{n}^{\beta_{1}/2} ). \end{equation} \end{proposition} The proof is based on a series of lemmas. The constant $K$ may depend on an implicit way on the bounds in this strengthened assumption, but not on the two numbers $u,r\in(0,1)$ which are fixed in most of this section. With any c\`{a}dl\`{a}g process $Y$ and $u\in(0,1]$, we associate the process and the variables \begin{equation} \label{D21 Y(u)_{t} = \sum_{s\leq t}\Delta Y_{s}1_{\{|\Delta Y_{s}|>u\}},\qquad \zeta(Y,u)_{i}^{n} = 1_{\{|\Delta_{i}^{n}Y|>u\}}. \end{equation} For simpler notation, we denote by $\mathbb{E}_{i-1}^{n}$ and $\mathbb{P} _{i-1}^{n}$, respectively, the conditional expectation and conditional probability, with respect to $\mathcal{F}_{(i-1)\Delta_{n}}$. \begin{lemma} \label{LD1} For all $u,r\in(0,1]$ with $u^{r}<1/3$, all $w\in(0,1/3)$ and all $k\geq1$, we have \begin{eqnarray} \label{D23 &&\mathbb{P}_{i-1}^{n}\bigl(\Delta_{i}^{n}V(u)\geq k\bigr) \leq( K\Delta _{n} u^{-\beta_{1}}) ^{k},\hspace*{-35pt} \\ \label{D25} &&\mathbb{P}_{i-1}^{n}\bigl(u(1-w)<\Delta_{i}^{n}X(u^{1+r})\leq u(1+w)\bigr)\hspace*{-35pt}\nonumber\\[-8pt]\\[-8pt] &&\qquad\leq K\bigl( \Delta_{n}u^{-\beta_{1}}w+\Delta_{n}u^{- \beta_{j+1}} +\Delta_{n}^{2}u^{-\beta_{1}(2+r)}+\Delta_{n}^{3}u^{-\beta_{1}(3+3r)}\bigr). \nonumber\hspace*{-35pt} \end{eqnarray} Moreover there is a $\gamma>0$ such that, if \begin{equation} \label{D265} \Delta_{n} \leq\gamma u^{\beta_{1}(1+r)} \end{equation} we have for all $u\in(0,1]$ \begin{equation} \label{D24} \mathbb{E}_{i-1}^{n}\bigl( \vert\zeta(X(u^{1+r}),u)_{i}^{n}-\Delta _{i}^{n}V(u)\vert\bigr) \leq K\bigl( \Delta_{n}^{2} u^{-\beta _{1}(2+r)}+\Delta_{n}^{3} u^{-\beta_{1}(3+3r)}\bigr) .\hspace*{-35pt} \end{equation} \end{lemma} \begin{pf} If $D\subset\mathbb{R}$ the compensator of the process $N(D)_{t}= \sum _{s\leq t}1_{D}(\Delta X_{s})$ is $\int_{0}^{t}F_{s}(D)\,ds$. Our strengthened assumption implies the existence of a constant $\theta$ such that $F_{s}(D)\leq\phi(D)$, where \[ \phi(D)=\cases{ \theta u^{-\beta_{1}}, &\quad if $D\subset[-u,u]^{c}$,\vspace*{1pt}\cr \theta( u^{-\beta_{1}}w+u^{-\beta_{j+1}}), &\quad if $D=\bigl[-u(1+w),-u\bigr)\cup\bigl(u,u(1+w)\bigr]$,\vspace*{1pt}\cr &\quad $0<w\leq1$.} \] Then for any finite stopping time $S$ we have \[ \mathbb{E}\bigl(N(D)_{S+t}-N(D)_{S}\mid\mathcal{F}_{S}\bigr) \leq t \phi(D). \] Let $S(D)_{0}=(i-1)\Delta_{n}$ and $S(D)_{1},S(D)_{2},\ldots$ be the successive jump times of $N(D)$ after time $(i-1)\Delta_{n}$. What precedes implies that for $k\geq1$ and on the set $\{S(D)_{k-1}<i\Delta_{n}\}$, \[ \mathbb{P}\bigl(S(D)_{j}\!\leq\!i\Delta_{n}\mid\mathcal{F}_{S(D)_{k-1}}\bigr)\!\leq\! \mathbb{E}\bigl( N(D)_{i\Delta_{n}}\!-\!N(D)_{S(D)_{k-1}}\mid\mathcal{F} _{S(D)_{k-1}}\bigr)\!\leq\!\Delta_{n} \phi(D). \] An induction on $k$ yields the following, which gives us the first part of (\ref{D23}): \begin{equation}\label{D250 \mathbb{P}_{i-1}^{n}\bigl(\Delta_{i}^{n}N(D)\geq k\bigr) = \mathbb{P}_{i-1}^{n} \bigl(S(D)_{k}\leq i\Delta_{n}\bigr) \leq( \Delta_{n} \gamma(D)) ^{k}. \end{equation} In the same way, if $D\cap D^{\prime}=\varnothing$, the set $\{\Delta_{i} ^{n}N(D)\geq k,\Delta_{i}^{n}N(D^{\prime})\geq1\}$ is the union for $l=1,\ldots,k+1$ of the sets $\Gamma_{l}=\{S(D)_{l-1}<S(D^{\prime} )_{1}<S(D)_{l}\leq i\Delta_{n}\}$, whereas \begin{eqnarray*} \hspace*{-4pt}&& \mathbb{P}_{i-1}^{n}\bigl(S(D)_{l-1}<S(D^{\prime} )_{1}<S(D)_{l}\leq i\Delta _{n}\bigr)\\ \hspace*{-4pt}&&\quad=\mathbb{E}_{i-1}^{n}\bigl( 1_{S(D)_{l-1} <S(D^{\prime})_{1}<i\Delta _{n}}\mathbb{P}\bigl( N(D)_{i\Delta_{n} } - N(D)_{S(D^{\prime})_{1 }\geq k-l+1 \hspace*{-0.3pt}\mid\hspace*{-0.3pt}\mathcal{F}_{S(D)_{1}} \bigr) \bigr) \\ \hspace*{-4pt}&&\quad\leq(\Delta_{n}\phi(D))^{k-l+1}\mathbb {P}_{i-1}^{n}\bigl(S(D)_{l-1}<S(D^{\prime })_{1}<i\Delta_{n}\bigr)\\ \hspace*{-4pt}&&\quad=(\Delta_{n}\phi(D))^{k-l+1}\\ \hspace*{-4pt}&&\qquad{}\times\mathbb{E}_{i-1}^{n}\bigl( 1_{S(D)_{l-1} <i\Delta_{n}}\mathbb{P}\bigl( N(D^{\prime})_{i\Delta_{n}}- N(D^{\prime })_{S(D^{\prime\prime})_{l-1}}\geq1 \mid\mathcal{F}_{S(D)_{1}} \bigr) \bigr) \\ \hspace*{-4pt}&&\quad\leq(\Delta_{n}\phi(D))^{k-l+1}\Delta_{n}\phi(D^{\prime})\mathbb{P} _{i-1}^{n}\bigl(S(D)_{l-1}<i\Delta_{n}\bigr), \end{eqnarray*} where~(\ref{D250}) has been applied twice. Another application of the same then yields \begin{eqnarray} \label{D231 D\cap D^{\prime}&=&\varnothing\Rightarrow\mathbb{P}_{i-1}^{n}\bigl(\Delta_{i} ^{n}N(D)\geq k, \Delta_{i}^{n}N(D^{\prime})\geq1\bigr)\nonumber\\[-8pt]\\[-8pt] &\leq&(k+1)\Delta_{n} ^{k+1}\phi(D)^{k}\phi(D^{\prime}).\nonumber \end{eqnarray} Next, let $w\in(0,1/3]$. By convention $(a,b]=\varnothing$ when $a\geq b$ below. If $u(1-w)<\Delta_{i}^{n} X(u^{1+r})\leq u(1+w)$ we have four (nonexclusive) possibilities: either $\Delta_{i}^{n}N((u^{1+r},\infty))\geq3$, or $\Delta _{i}^{n}N((u^{1+r},u/3])=\Delta_{i} ^{n}N((u/3,\infty))=1$, or $\Delta_{i} ^{n}N((u/3,\infty))=2$, or $\Delta_{i}^{n}N((u(1-w),u(1+w)])=1$. We an analogous implication if $-u(1+w)<\Delta_{i}^{n} X(u^{1+r})\leq -u(1-w)$. Then (\ref{D25}) easily follows from~(\ref{D250}) applied with $D=[-u^{1+r ,u^{1+r}]^{c}$, with $D=[-u/3,u/3]^{c}$ and with $D= [-u(1+w),-u(1-w))\cup (u(1-w),u(1+w)]$, and from~(\ref{D231}) applied with $D=(-u/3,-u^{1+r )\cup(u^{1+r},u/3]$ and $D^{\prime}=[-u/3,u/3]^{c}$. Finally we prove~(\ref{D24}). Let $H=|\zeta(X(u^{1+r}),u)_{i}^{n}-\Delta _{i}^{n}V(u)|$ and $D=[-u/2$, $-u^{1+r})\cup(u^{1+r},u/2]$ and $D^{\prime }=[-u/2,u/2]^{c}$ and $D^{\prime\prime}=D\cup D^{\prime}$. From what precedes, we have \begin{eqnarray}\label{D240} \mathbb{P}_{i-1}^{n}\bigl(\Delta_{i}^{n}N(D^{\prime\prime})\geq k\bigr)&\leq&\bigl( \theta\Delta_{n}u^{-\beta_{1}(1+r)}\bigr) ^{k},\nonumber\\ \mathbb{P}_{i-1}^{n}\bigl(\Delta_{i}^{n}N(D^{\prime})=2\bigr)&\leq&\theta^{2}\Delta _{n}^{2}u^{-2\beta_{1}},\\ \mathbb{P}_{i-1}^{n}\bigl(\Delta_{i}^{n}N(D)=\Delta_{i}^{n}N(D^{\prime} )=1\bigr)&\leq&\theta^{2}\Delta_{n}^{2}u^{-\beta_{1}(2+r)}. \nonumber \end{eqnarray} We have $H=0$ on the sets $\{\Delta_{i}^{n}N(D^{\prime\prime})\leq1\}$ and $\{\Delta_{i}^{n}N(D^{\prime\prime})=\Delta_{i}^{n}N(D)=2\}$, and $H\leq k-1$ on the set $\{\Delta_{i}^{n}N(D^{\prime\prime})=k\}$, for all $k\geq2$. Thus if $v=\theta\Delta_{n}u^{-\beta_{1}(1+r)}$, \begin{eqnarray*} \mathbb{E}_{i-1}^{n}(H) & \leq & \sum_{k=3}^{\infty}k \mathbb{P}_{i-1} ^{n}\bigl(\Delta_{i}^{n}N(D^{\prime\prime})\geq k\bigr)+\mathbb {P}_{i-1}^{n}\bigl(\Delta _{i}^{n}N(D^{\prime})=2\bigr)\\ &&{} +\mathbb{P}_{i-1}^{n}\bigl(\Delta_{i}^{n}N(D)=\Delta_{i}^{n}N(D^{\prime })=1\bigr)\\ & \leq & \sum_{k=3}^{\infty}kv^{k}+\theta^{2}\Delta_{n}^{2}u^{-2\beta_{1} }+\theta^{2}\Delta_{n}^{2}u^{-\beta_{1}(2+r) \end{eqnarray*} by~(\ref{D240}). When $v\leq1/2$, that is, when $\Delta_{n}\leq\gamma u^{\beta_{1}(1+r)}$ for $\gamma=1/2\theta$, we have $\sum_{k=3}^{\infty} kv^{k}\leq Kv^{3}$, and the above is smaller than the right-hand side of~(\ref{D24}). \end{pf} \begin{lemma} \label{LD301} Let $q\geq2$ and $u,r\in(0,1)$. As soon as~(\ref{D265}) holds for some constant $\gamma>0$, we have \begin{equation} \label{D266 \mathbb{E}\bigl(\bigl|\Delta_{i}^{n}\bigl(X-X(u^{1+r})\bigr)\bigr|^{q}\bigr) \leq K_{\gamma,q} \Delta_{n}\bigl(\Delta_{n}^{q/2-1}+u^{(q-\beta_{1})(1+r)}\bigr). \end{equation} \end{lemma} \begin{pf} Letting $X^{c}$ and $\mu$ be the continuous martingale part and the jump measure of $X$, we have $X-X(u^{1+r})=B+B^{\prime}+X^{c}+M$, where \[ B_{t}^{\prime}=-\int_{0}^{t}ds\int_{\{u^{1+r}<|x|\leq1\}}xF_{s}(dx),\quad M_{t}=\int_{0}^{t}\int_{\{0<|x|\leq u^{1+r}\}}x (\mu-\nu)(ds,dx). \] By the strengthened Assumption~\ref{A2}, for any $y>0$ the integral $\int_{\{|x|>y\}}|x| F(dx)$ is smaller than $K$ when $\beta_{1}<1$, than $K\log\frac{1}{y}$ when $\beta_{1}=1$, and than $Ky^{1-\beta_{1}}$ when $\beta_{1}>1$. Therefore, since~(\ref{D265}) implies $2\beta_{1 (1+r)>(\beta_{1}-1)^{+}$ we have $|\Delta_{i}^{n}B^{\prime}|\leq K_{\gamma }\sqrt{\Delta_{n}}$. The strengthened Assumption~\ref{A2} also implies $|\Delta^{n}_{i}B|\leq K\Delta_{n}$ and, by well-known estimates about continuous and purely discontinuous martingales [see, e.g., \citet{yacjacod11}], we also deduce that \[ \mathbb{E}(|\Delta_{i}^{n}M^{\prime}|^{q})\leq K_{q}\Delta_{n} u^{(q-\beta _{1})(1+r)} ,\qquad \mathbb{E}(|\Delta_{i}^{n}X^{c}|^{q})\leq K_{q}\Delta _{n}^{q/2}. \] All these estimates readily give~(\ref{D266}). \end{pf} \begin{pf*}{Proof of Proposition~\ref{PD1}} (a) It follows from~(\ref{D31}) that $u_{n}^{2\beta_{1}}/\Delta_{n} \to\infty$, so for any $\gamma>0$~(\ref{D265}) is satisfied for all $r\in(0,1)$ and all $n$ large enough. Hence, both estimates~(\ref{D24}) and~(\ref{D266}) hold, with constants $K$ and $K_{\gamma,q}$ independent of $r$, for all $n$ large enough. The following inequality, where $u,w\in(0,1)$ and $x,y\in\mathbb{R}$, is elementary: \[ \bigl\vert1_{\{x+y>u\}}-1_{\{x>u\}}\bigr\vert \leq1_{\{|y|\geq uw\}}+1_{\{u(1-w)<|x|\leq u(1+w)\}}. \] We apply this with $x=\Delta_{i}^{n}X(u^{1+r})$ and $x+y=\Delta ^{n}_{i}X$ and $u=u_{n}$, and with $w\leq1/3$ to be chosen later.\vadjust{\goodbreak} In order to evaluate the probabilities for having $|y|\geq u_{n}w$, respectively, $u_{n}(1-w)<|x|\leq u_{n}(1+w)$, we use~(\ref{D266}) and Markov's inequality, respectively, (\ref{D25}). This gives that $\mathbb{E} (|\zeta(X(u_{n}^{1+r}),u_{n})^{n}_{i}- \zeta(X,u_{n})^{n}_{i}|)$ is smaller, for all $q\geq2$, than \[ \frac{K_{q}\Delta_{n}}{u_{n}^{\beta_{1}}}\biggl( \frac{\Delta_{n}^{q/2-1}}{w^{q} u_{n}^{q-\beta_{1}}}+ \frac{u_{n} ^{(q-\beta_{1})r}}{w^{q}}+w+\frac{\Delta_{n}} {u_{n}^{\beta_{1}(1+r)} +\frac{\Delta_{n}^{2}}{u_{n}^{\beta_{1}(2+3r)}} +u_{n}^{\beta_{1}-\beta _{j+1 }\biggr). \] Optimizing over $w$ leads to take $w=w_{n}$ such that $w_{n}^{q+1}= u_{n}^{(q-\beta_{1})r}+\Delta_{n}^{q/2-1}/\allowbreak u_{n}^{q-\beta_{1}}$, which is indeed smaller than $1/3$ for all $n$ large. Thus, putting the above together with~(\ref{D24}), and recalling that $\Delta_{n}\leq Ku_{n}^{1/\rho}$, we end up with \begin{equation}\label{D119 \mathbb{E}\bigl(|\zeta(X,u_{n})^{n}_{i}-\Delta^{n}_{i}V(u_{n})|\bigr) \leq \frac {K_{q}\Delta_{n}}{u_{n}^{\beta_{1}}}\sum_{k=1}^{5}u_{n}^{x_{k}}, \end{equation} where $x_{k}=x_{k}(q,r)$ are given by \begin{eqnarray*} x_{1}&=&\frac{qr-\beta_{1}r}{q+1} ,\qquad x_{2}=\frac{q(1-2\rho)-2+2\beta _{1}\rho }{2\rho(q+1)}, \\ x_{3}&=&\frac1{\rho}-\beta_{1}(1+r),\qquad x_{4}=\frac2{\rho}-\beta _{1}(2+3r),\qquad x_{5}=\beta_{1}-\beta_{j+1}. \end{eqnarray*} (b) Now, for proving~(\ref{D30}), it clearly follows from~(\ref{D119}) that it suffices to show that one can choose $q$ and $r$ in such a way that $x_{k} \geq\beta_{1}/2$ for $k=1,2,3,4$. When $q\rightarrow\infty$ we see that $x(1)\to x^{\prime}(1)=r$ and $x(2)\to x^{\prime}(2)=\frac{1-2\rho }{2\rho}$, so it remains to show that one can choose $r\in(0,1)$ such that $x^{\prime }(k)\geq\beta_{1}/2$ for $k=1,2$ and $x_{k}\geq\beta_{1}/2$ for $k=3,4$. Letting $r$ be bigger than but as close as possible to $\beta_{1}/2$, we deduce from~(\ref{D31}) that such a choice or $r$ is possible, and the proof is complete. \end{pf*} \section{\texorpdfstring{Proof of Theorem \protect\ref{TD1}}{Proof of Theorem 2}} (1) In addition to the strengthened Assumption~\ref{A2}, we assume~(\ref{118}) for some $\varepsilon>0$. Theorem~\ref{TD1} says something about the estimators of $\beta_{i}$ and $A^{i}_{T}$ only when $\beta_{i}>\frac {\beta _{1}}2$. Moreover, if~(\ref{31}) holds for the sequence $\beta_{1 ,\ldots,\beta_{j+1}$, it also holds for the sequence $\beta^{\prime _{1},\ldots,\beta^{\prime}_{j^{\prime}+1}$, where $j^{\prime}=j$ if $\beta_{j+1}\geq\frac{\beta_{1}}2$ and $j^{\prime}=\sup(i\dvtx\beta_{i >\frac{\beta_{1}}2)$ otherwise, and where $\beta^{\prime}_{i}=\beta _{i}$ when $i\leq j^{\prime}$ and $\beta^{\prime}_{j^{\prime}+1}=\beta_{j^{\prime}+1} \vee\frac{\beta_{1}}2$. Henceforth, upon discarding the indices such that $\beta_{i}\leq\frac{\beta_{1}}2$, we can assume without loss of generality that \begin{equation} \label{127} \beta_{1}>\cdots>\beta_{j}>\beta_{j+1}=\frac{\beta_{1}}2. \end{equation} Under this additional assumption, we have $\beta_{1}-\beta_{j}<1$, and (\ref{S10}) yields \begin{equation} \label{S1001} 1\leq i\leq k<j \quad\Rightarrow\quad u_{n,i}^{\beta_{i}-\beta_{i+1}} \log\frac {1}{u_{n,i}} = o( u_{n,k}^{\beta_{i}-\beta_{k+1}}).\vadjust{\goodbreak} \end{equation} Moreover, combining~(\ref{31}),~(\ref{KP}) and~(\ref{D30}), we deduce that \begin{equation} \label{S9 U(v_{n},\Delta_{n})_{T} = \sum_{i=1}^{j}\frac{A_{T}^{i}}{v_{n}^{\beta _{i} }+O_{P} (v_{n}^{-\beta/2}) \end{equation} for any sequence $v_{n}$ such that $v_{n}\leq u_{n}$, and in particular for the sequences $v_{n}=u_{n,i}$. All of the proof will rely on this, and below $H_{i}$ is always given by~(\ref{S12}). (2) We first consider the case $i=1$ when $j>1$. A simple calculation, based on (\ref{S9}) applied with $v_{n}=u_{n}$ and $v_{n}=\gamma u_{n}$, yields that in restriction to the set $\Omega_{T}$, \[ \log\bigl(U(v_{n},\Delta_{n})_{T}/U(\gamma v_{n},\Delta_{n})_{T}\bigr) = ( \beta_{1}-H_{1} u_{n}^{\beta_{1}-\beta_{2}}) \log\gamma+o_{P (u_{n}^{\beta_{1}-\beta_{2}}). \] This gives the first part of~(\ref{D5}). It also implies that \begin{eqnarray*} u_{n}^{\widetilde{\beta}_{1}^{n}}&=&u_{n}^{\beta_{1}}e^{-(\widetilde{\beta}_{1} ^{n}-\beta_{1})\log(1/u_{n})}\\ &=&u_{n}^{\beta_{1}}\bigl( 1+H_{1}u_{n} ^{\beta_{1}-\beta_{2}} \log(1/u_{n})+o_{P}\bigl(u_{n}^{\beta_{1}-\beta_{2} \log(1/u_{n})\bigr)\bigr) . \end{eqnarray*} This and~(\ref{S9}) yield the second part of~(\ref{D5}). (3) Now we suppose that~(\ref{D5}) holds for all $i\leq k-1$, for some $k\in\{2,\ldots,j-1\}$. We observe that we have the following identities, for all $y=(y_{1},\ldots,y_{k+1})$ and $r=1,\ldots,k+1$: \begin{eqnarray*} && \sum_{l=0}^{k-1}(-1)^{l}\gamma^{-ly_{r}}\sum_{J\in I(k-1,l)} \gamma ^{\sum_{j\in J}y_{j}}\\ &&\qquad=\prod_{l=1}^{k-1}( 1-\gamma^{y_{i}-y_{r}}) =\cases{ 0, &\quad if $r\leq k-1$,\cr G(k,y,\gamma), &\quad if $r=k$,\cr G^{\prime}(k,y,\gamma), &\quad if $r=k+1$,} \end{eqnarray*} where $G(k,y,\gamma)=\prod_{i=1}^{k-1}( 1-\gamma^{y_{i}-y_{k}}) $ and $G^{\prime}(k,y,\gamma)=\prod_{i=1}^{k-1}( 1-\gamma^{y_{i}-y_{k+1} }) $. Therefore,~(\ref{S9}) applied to $v_{n}=x\gamma^{l}u_{n,k}$ and the definition of $U^{n}(k,x)$ yield for all $x\geq1$ fixed, and with $\beta=(\beta_{1},\ldots,\beta_{k+1})$, \begin{eqnarray*} U^{n}(k,x) & = & \sum_{r=1}^{k-1}\frac{A_{T}^{r}}{x^{\beta_{r}} u_{n,k} ^{\beta_{r}}} \sum_{l=0}^{k-1}(-1)^{l}( \gamma^{-l\beta_{r}} -\gamma^{-l\widetilde{\beta}_{r}^{n}}) \sum_{J\in I(k-1,l) \gamma^{ \sum_{j\in J}\widehat{\beta}_{j}^{n}}\\ &&{} +\sum_{r=k}^{k+1}\frac{A_{T}^{r}}{x^{\beta_{r}} u_{n,k}^{\beta_{r}} \sum_{l=0}^{k-1}(-1)^{l}\gamma^{-l\beta_{r}}\sum_{J\in I(k-1,l)}( \gamma^{\sum_{j\in J}\widetilde{\beta}_{j}^{n}}-\gamma^{ \sum_{j\in J} \beta_{j}}) \\ &&{} +\frac{A_{T}^{k}}{x^{\beta_{k}} u_{n,k}^{\beta_{k}}} G(k,\beta ,\gamma)+\frac{A_{T}^{k+1}}{x^{\beta_{k+1}} u_{n,k}^{\beta_{k+1}} G^{\prime}(k,\beta,\gamma)+o_{P}(u_{n,k}^{-\beta_{k+1}}). \end{eqnarray*} The functions $z\mapsto\gamma^{-lz}$ are $C^{\infty}$. The induction hypothesis gives $\widetilde{\beta}_{i}^{n}-\beta _{i}=O_{P}(u_{n,i}^{\beta _{i}-\beta_{i+1}})$ for $i=1,\ldots,k-1$. Then~(\ref{S1001}) and $\beta _{i}-\beta_{i+1}>\varepsilon$ allow us to deduce \begin{eqnarray*} &\displaystyle 0\leq l\leq k-1, J\in I(k-1,l) \quad\Rightarrow\quad \gamma^{ \sum_{j\in J} \widetilde{\beta}_{j}^{n}}-\gamma^{ \sum_{j\in J}\beta_{j}} = o_{P} (u_{n,k}^{\beta_{k-1}-\beta_{k+1}}),&\\ &\displaystyle 1\leq r\leq k-1 \quad\Rightarrow \quad\gamma^{-l\beta_{r}} -\gamma ^{-l\widetilde {\beta}_{r}^{n}} = o_{P}(u_{n,i}^{\beta_{r}-\beta_{r+1}}) = o_{P}(u_{n,k} ^{\beta_{i}-\beta_{k+1}}).& \end{eqnarray*} Therefore we finally obtain \begin{eqnarray} \label{S16 U^{n}(k,x) &=& \frac{A_{T}^{k} G(k,\beta,\gamma)}{x^{\beta_{k}} u_{n,k}^{\beta_{k}}}+\frac{A_{T}^{k+1} G^{\prime}(k,\beta,\gamma) {x^{\beta_{k+1}} u_{n,k}^{\beta_{k+1}}}+o_{P}(u_{n,k}^{-\beta_{k+1} })\nonumber\\[-8pt]\\[-8pt] &=& \frac{A_{T}^{k} G(k,\beta,\gamma)}{x^{\beta_{k}} u_{n,k}^{\beta_{k} } \biggl( 1+\frac{H_{k} \log\gamma}{\gamma^{\beta_{k}-\beta_{k+1}} -1} (xu_{n,k})^{\beta_{k}-\beta_{k+1}}+o_{P}(u_{n,k}^{\beta_{k} -\beta _{k+1 })\biggr) ,\hspace*{-10pt}\nonumber \end{eqnarray} where the last equality comes from the definition of $H_{k}$ in~(\ref{S12}). Then exactly as in Step~\ref{step2}, a simple calculation shows the first half of (\ref{D5}) for $i=k$. For the second part of~(\ref{D5}), and as in Step~\ref{step2}, we first deduce from the above that \[ u_{n,k}^{\widetilde{\beta}_{k}^{n}} = u_{n,k}^{\beta_{k}}\bigl( 1+H_{k} u_{n,k}^{\beta_{k}-\beta_{k-1}} \log(1/u_{n,k})+o_{P}\bigl(u_{n,k}^{\beta _{k}-\beta_{k-1}} \log(1/u_{n,k})\bigr)\bigr) . \] Therefore it is enough to show that \[ u_{n,k}^{\beta_{k}}\Biggl( U(u_{n,k})_{T}-\sum_{i=1}^{k-1}\widetilde{\Gamma }_{i}^{n} u_{n,k}^{-\widetilde{\beta}_{i}^{n}}\Biggr) = A_{T}^{k} +o_{P}\bigl(u_{n,k}^{\beta_{k}-\beta_{k-1}} \log(1/u_{n,k})\bigr). \] In view of~(\ref{S9}) with $v_{n}=u_{n,k}$ this amounts to proving for $i=1,\ldots,k-1$, \begin{equation} \label{S17 \widetilde{\Gamma}_{i}^{n} u_{n,k}^{\beta_{k}-\widetilde{\beta}_{i}^{n}} -A_{T}^{i} u_{n,k}^{\beta_{k}-\beta_{i}} = o_{P}\bigl(u_{n,k}^{\beta_{k} -\beta_{k-1}} \log(1/u_{n,k})\bigr). \end{equation} The induction hypothesis yields that \begin{eqnarray*} u_{n,k}^{\beta_{k}-\widetilde{\beta}_{i}^{n}} &=& u_{n,k}^{\beta_{k}-\beta _{i }\bigl( 1+O_{P}\bigl(u_{n,i}^{\beta_{i}-\beta_{i+1}} \log(1/u_{n,k})\bigr)\bigr) , \\ \widetilde{\Gamma}_{i}^{n} &=& A_{T}^{i}+O_{P}\bigl(u_{n,i}^{\beta_{i}-\beta_{i+1} } \log(1/u_{n,i})\bigr). \end{eqnarray*} Then~(\ref{D5}) readily follows from~(\ref{S10}). (4) It remains to prove that the variables in~(\ref{D6}) are tight. The difference with the previous case is that~(\ref{S16}) no longer holds when $i=j=1$ or $k=j>1$, but it can be replaced by \[ U^{n}(j,x) = \frac{A_{T}^{j} G(j,\beta,\gamma)}{x^{\beta_{j}} u_{n,j ^{\beta_{j}}} \bigl( 1+O_{P}(u_{n,j}^{\beta_{j}-\beta/2})\bigr) . \] The rest of the proof goes unchanged [note that $\eta$ in~(\ref{D6}) is $\beta_{j}-\beta/2$ here]. \section{\texorpdfstring{Proof of Theorem \protect\ref{TD2}}{Proof of Theorem 3}} We use simplifying notation: a point in $D$ is $\theta=(x_{i},\gamma _{i})_{1\leq i\leq j}$, and we define the functions $F_{n,l}(\theta )=\sum_{i=1}^{j}\gamma_{i}/(v_{l}u_{n})^{x_{i}}$. The ``true value'' of the parameter is $\theta_{0}=(\beta _{i},\Gamma _{i})_{1\leq i\leq j}$, the preliminary estimators are $\widetilde {\theta} _{n}=(\widetilde{\beta}_{i}^{n},\widetilde{\Gamma}_{i}^{n})_{1\leq i\leq j}$, and the final estimators are $\overline{\theta}_{n}=(\overline{\beta}{}^{n}_{i} ,\overline{\Gamma}{}^{n}_{i})_{1\leq i\leq j}$. We set $h_{n} =\log (1/u_{n})$, and as in the previous proof we can assume~(\ref{127}). (1) We introduce some specific notation. For $m\geq2$ we set $G_{m =(1,\infty)^{m-1}$, a~point in $G_{m}$ being denoted as $\overline{v =(v_{2},\ldots,v_{m})$. For $1\leq k\leq j$ and $\overline{v}\in G_{2k}$, and with the convention $v_{1}=1$, we let $\Sigma(\overline{v})$ be the $2k\times2k$ matrix with entries \begin{equation} \label{D2-25}\Sigma(\overline{v})_{l,i} = \cases{ v_{l}^{-\beta_{i}}, &\quad if $1\leq i\leq k$,\cr v_{l}^{-\beta_{i-k}} \log v_{l}, &\quad if $k+1\leq i\leq2k$.} \end{equation} The aim of this step is to show that the set $Z_{k}$ of all $\overline {v}\in G_{2k}$ for which the matrix $\Sigma(\overline{v})$ is invertible satisfies $\lambda_{2k}((Z_{k})^{c})=0$, where $\lambda_{r}$ is the Lebesgue measure on~$G_{r}$. When $1\leq m\leq2k$ and $\overline{v}\in G_{2k}$, we denote by $\mathcal{M _{m}(\overline{v})$ the family of all $m\times m$ sub-matrices of the $m\times2k$ matrix $(\Sigma( \overline{v})_{l,r}\dvtx 1\leq l\leq m,1\leq r\leq 2k)$. A key fact is that $\mathcal{M}_{m}(\overline{v})=\mathcal{M _{m}(\overline{v}_{m})$ only depends on the restriction $\overline{v _{m}=(v_{2},\ldots, v_{m})$ of $\overline{v}$ to its first $m-1$ coordinates. Moreover, $\Sigma(\overline{v})_{1i}$ equals $1$ if $i\leq k$ and $0$ otherwise: so the entries of the first column of any $M\in\mathcal{M _{m}(\overline{v })$ are $0$ or $1$, and $\mathcal{M}^{\prime}_{m (\overline{v})$ denotes the subset of all $M\in\mathcal{M}_{m}(\overline{v})$ for which $M_{1,i}=1 $ for at least one value of $i$. Finally, $H_{m}$ stands for the set of all $\overline{v}_{m}\in G_{m}$ such that all $M\in \mathcal{M }^{\prime}_{m}(\overline{v}_{m})$ are invertible. Since $\mathcal{M} ^{\prime }_{2k}(\overline{v})$ is the singleton $\{\Sigma(\overline{ v})\}$, we have $Z_{k}=H_{2k}$. If $m\geq2$ and $\overline{v}_{m}=(v_{2},\ldots,v_{m})\in G_{m}$ and $M\in\mathcal{M}_{m}^{\prime}(\overline{v}_{m})$, by expanding along the last column, we see that \begin{equation} \label{B4-1201} \det(M) = \sum_{i=1}^{k}v_{m}^{\beta_{i}}(a_{i}+a_{k+i} \log v_{m}), \end{equation} where each $a_{r}$ is of the form: either (i) $a_{r}$ is plus or minus $\det(M_{r})$ for some $M_{r}\in\mathcal{M}_{m-1}(\overline{v}_{m})$ (for $m$ values of $r$) or (ii) $a_{r}=0$ (for the other $2k-m$ values of~$r$). Note that we can also have $a_{r}=0$ in case (i), and since $M\in\mathcal{M ^{\prime}_{m}(\overline{v}_{m})$ there is at least one $a_{r}$ of type (i) with $M_{r}\in\mathcal{M}^{\prime}_{m-1}(\overline{v}_{m})$. When at least one $a_{r}$ in~(\ref{B4-1201}) is not $0$, the right-hand side of this expression, as a function of $v_{m}$, has finitely many roots only, because all $\beta_{i}$'s are distinct. Observing that $\mathcal{M} ^{\prime }_{1}(\overline{v})$ is the $1\times1$ matrix equal to $1$, it follows that, with $(\overline{v}_{m-1},v_{m})= (v_{2},\ldots,v_{m-1},v_{m})$ when $\overline{v}_{m-1} =(v_{2},\ldots,v_{m-1})$, and recalling that with our standing notation $\lambda_{2}$ is the Lebesgue measure on $(1,\infty)$, \begin{eqnarray} \label{B4-12}\qquad m=2 &\quad\Rightarrow\quad& \lambda_{2}((H_{2})^{c}) = 0,\nonumber\\[-8pt]\\[-8pt] m\geq3, \qquad\overline{v}_{m-1}\in H_{m-1} &\quad\Rightarrow\quad& \lambda_{2}\bigl( v_{m \dvtx(\overline{va}_{m-1},v_{m})\notin H_{m}\bigr) = 0. \nonumber \end{eqnarray} Since \[ \lambda_{m}((H_{m})^{c}) = \int_{G_{m-1}} \lambda_{2}\bigl(v_{m}\dvtx ( \overline{v}_{m-1},v_{m})\notin H_{m}\bigr) \lambda_{m-1}(d \overline {v}_{m-1}), \] which equals $\int_{H_{m-1}} \lambda_{1}(v_{m}\dvtx (\overline{v} _{m-1},v_{m})\notin H_{m}) \lambda_{m-1} (d\overline{v}_{m-1})$ if $\lambda_{m-1}((H_{m-1})^{c})=0$, when $m\geq3$, we deduce from~(\ref{B4-12}), by induction on~$m$, that indeed $\lambda_{m}((H_{m})^{c})=0$ for all $m=2,\ldots,2k$. Recalling $Z_{k}=H_{2k}$, the result follows. Since the claim of the theorem holds for all $(v_{2},\ldots,v_{L})$ outside a $\lambda_{L}$-null set only, and $L\geq2k$, we thus can and will assume below that the numbers~$v_{l}$ are such that $\overline{v} _{2k}=(v_{2 ,\ldots,v_{2k}) \in Z_{k}$, hence $\Sigma(\overline{ v}_{2k})$ is invertible, for all $k=1,\ldots,j$. (2) Our assumptions on the preliminary estimators yield that the set $\Omega_{n}$ on which $\Vert\widetilde{\theta}_{i}^{n}-\theta_{0}\Vert \leq1/u_{n}^{\eta}$ satisfies $\mathbb{P}(\Omega_{n})\rightarrow1$. So below we argue on the set~$\Omega_{n}$, or equivalently (and more conveniently) we suppose $\Omega_{n}=\Omega$. Then~$\overline{\theta}_{n}$ converges pointwise to $\theta_{0}$, which belongs to all the sets $D_{n}$. Set \[ y_{i}^{n} = A_{T}^{i}(\overline{\beta}{}^{n}_{i}-\beta_{i}),\qquad z_{i} ^{n} = \overline{\Gamma}{}^{n}_{i}-A_{T}^{i}+y_{i}^{n}h_{n},\qquad a_{i ^{n} = |y_{i}^{n}|h_{n}+|z_{i}^{n}|. \] We have $a_{i}^{n}\leq2u_{n}^{-\eta}h_{n}$ because $\Omega_{n}=\Omega$. Then an expansion of $(x_{i},\gamma_{i})\mapsto\gamma _{i}/(v_{l}u_{n})^{x_{i} }$ around $(\beta_{i},A_{T}^{i})$ yields for all $l$, \begin{equation} \label{D2-5 \frac{\overline{\Gamma}_{i}}{(v_{l}u_{n})^{\overline{\beta}_{i}}}-\frac {A_{T}^{i}}{(v_{l}u_{n})^{\beta_{i}}}= \frac{1}{(v_{l}u_{n})^{\beta_{i}} }( z_{i}^{n}-y_{i}^{n}\log v_{l}+x_{i,l}^{n}), \end{equation} where \[ |x_{i,l}^{n}| \leq K|y_{i}^{n}|h_{n}(|z_{i}^{n}|+|y_{i}^{n} |) \leq K|y_{i}^{n}|h_{n}a_{i}^{n} \leq K(a_{i}^{n})^{2}. \] Combining~(\ref{31}),~(\ref{KP}) and~(\ref{D30}), we see that \[ U(v_{l}u_{n},\Delta_{n})_{T}-F_{n,l}(\theta_{0}) = O_{P}(u_{n}^{-\beta_{1} /2}). \] Since $\Phi_{n}(\theta)=\sum_{l=1}^{L}w_{l}( U(v_{l}u_{n},\Delta _{n})_{T}-F_{n,l}(\theta)) ^{2}$, we deduce \[ \Phi_{n}(\theta_{0} )=O_{P}(u_{n}^{-\beta_{1}}). \] Since $\theta_{0}\in D_{n}$ and $\overline {\theta}_{n}$ minimizes $\Phi_{n}$ over $D_{n}$, we also have $\Phi _{n}(\overline{\theta}_{n})=O_{P}(u_{n}^{-\beta_{1}})$, hence $F_{n,l} (\theta_{0})-F_{n,j}(\overline{\theta}_{n})=O_{P}(u_{n}^{-\beta _{1}/2})$ for all $l$. Using~(\ref{D2-5}), this can be rewritten as \begin{equation}\label{D2-6 \sum_{i=1}^{j}\frac{1}{(v_{p}u_{n})^{\beta_{i}}} ( z_{i}^{n}-y_{i} ^{n}\log v_{l}+x_{i,l}^{n}) = O_{P}(u_{n}^{-\beta_{1}/2}). \end{equation} (3) Taking $k$ between $1$ and $j$, we consider the $2k$-dimensional vectors $\zeta(k,n)$ and $\xi(n)$ with components (for $l=1,\ldots,2k$), \begin{eqnarray*} \zeta(k,n)_{l} &=& \sum_{i=1}^{k}\frac{1}{(v_{p}u_{n})^{\beta_{i}}} ( z_{i}^{n}-y_{i}^{n}\log v_{l}) ,\vadjust{\goodbreak}\\ \xi(k,n)_{i}&=&\cases{ z_{i}^{n} u_{n}^{-\beta_{i}}, &\quad if $1\leq i\leq k$,\cr -y_{i-k}^{n} u_{n}^{-\beta_{i-k}}, &\quad if $k+1\leq i\leq2k$.} \end{eqnarray*} With matrix notation, and~(\ref{D2-25}), we have $\zeta(k,n)= \Sigma (\overline{v}_{2k})\xi(k,n)$, hence \begin{equation} \label{D2-7 \xi(k,n) = \Sigma(\overline{v}_{2k})^{-1} \zeta(k,n). \end{equation} Next, we have \[ \frac{1}{(v_{l}u_{n})^{\beta_{i}}} \bigl|z_{i}^{n}+\bigl(v^{ \prime}_{g}(\beta _{i})+v_{g}(\beta_{i})\log\delta_{l}\bigr)y_{i} ^{n}+x_{i,l}^{n}\bigr| \leq \frac{Ka_{i}^{n}}{u_{n}^{\beta_{i}}},\qquad \frac{|x_{i,l}^{n}|}{(v_{l u_{n})^{\beta_{i}}}\leq\frac{K(a_{i}^{n})^{2}}{u_{n}^{\beta_{i}}}, \] and hence~(\ref{D2-6}) and $a_{i}^{n}\leq Ku_{n}^{\eta}h_{n}\leq K/h_{n}^{2} \leq K$ yield \[ |\zeta(k,n)_{l}| \leq K\Biggl( \sum_{i=1}^{k-1}(a_{i}^{n})^{2} u_{n} ^{-\beta_{i}}+\frac{a^{n}_{k}}{h_{n}^{2}} u_{n}^{-\beta_{k}}+\sum _{i=k+1 ^{j}a_{i}^{n} u_{n}^{-\beta_{i}}\Biggr) +O_{P}(u_{n ^{-\beta_{j+1}}). \] By~(\ref{D2-7}) the variables $\xi(k,n)_{l}$ satisfy the same estimate. Since $a_{k}^{n}\leq(|\xi(k,n)_{k}|+ |\xi(k,n)_{2k} |h_{n})u_{n}^{\beta_{k}}$, \[ a_{k}^{n} \leq Ch_{n}\Biggl( \sum_{i=1}^{k-1}(a_{i}^{n})^{2} u_{n}^{\beta _{k}-\beta_{i}}+\frac{a^{n}_{k}}{h_{n}^{2}} +\sum_{i=k+1}^{j}a_{i}^{n} u_{n}^{\beta_{k}-\beta_{i}}\Biggr) +O_{P}(h_{n} u_{n}^{ \beta_{k}-\beta_{j+1}}) \] for some constant $C$. When $n$ is large enough, $C/h_{n}\leq\frac12$, and we deduce \begin{equation} \label{B4-17}\qquad a_{k}^{n} \leq2Ch_{n}\Biggl( \sum_{i=1}^{k-1}(a_{i}^{n})^{2} u_{n}^{\beta_{k}-\beta_{i}} +\sum_{i=k+1}^{j}a_{i}^{n} u_{n}^{\beta _{k}-\beta_{i}}\Biggr) +O_{P}(h_{n} u_{n}^{\beta_{k}-\beta _{j+1}}). \end{equation} (4) In view of the definition of $y^{n}_{i}$ and $z^{n}_{i}$, to get the result, and recalling that we assume $\beta_{j+1}=\beta_{1}/2$, it is clearly enough to prove the existence of a~number $\nu>0$ such that, for all $i=1,\ldots,j$, we have \begin{equation} \label{B4-21} a^{n}_{i} = O_{P}(h_{n}^{\nu} u_{n}^{\beta_{i}-\beta_{j+1}}). \end{equation} To this aim, we introduce the following property, denoted ($P_{m,q,r}$), where~$r$ runs through $\{1,\ldots,j\}$ and $m,q\geq1$, and where we use the notation $\zeta_{r}=\beta_{r}-\beta_{r+1}$: \begin{equation} \label{B4-22} i=1,\ldots,r \quad\Rightarrow\quad a_{i}^{n} = O_{P}\bigl( h_{n ^{m} (u_{n}^{\beta_{i}-\beta_{r}+q\zeta_{r}} +u_{n}^{\beta_{i}-\beta_{r+1}})\bigr). \end{equation} Since $a_{i}^{n}\leq K$, applying~(\ref{B4-17}) with $k=1$ yields $a_{1} ^{n}=O_{P}(h_{n}u_{n}^{\beta_{1}-\beta_{2}})$, which is ($P_{1,1,1}$). Next, we suppose that ($P_{m,q,r}$) holds for some $r<j$, and for some $m,q\geq1$. Letting first $k=r+1$, we deduce from~(\ref{B4-17}) that, since again $a^{n}_{i}\leq K$, \begin{eqnarray} \label{B4-23 a^{n}_{k} & = & O_{P}\Biggl(h_{n}^{1+2m}\sum_{i=1}^{k-1}\bigl( u_{n}^{ \beta _{k}-\beta_{i}+2(\beta_{i}-\beta_{r}+q\zeta_{r})} +u_{n}^{\beta_{i -\beta_{r+1}} \bigr)\nonumber\\ & &\hspace*{70pt}{} +h_{n}\sum_{i=k+1}^{j} u_{n}^{\beta_{k}-\beta_{i}}+h_{n u_{n}^{\beta_{k}-\beta_{j+1}}\Biggr)\\ & = & O_{P}\bigl(h_{n}^{1+2m} ( u_{n}^{\beta_{k}-\beta _{r}+2q\zeta_{r}} +u_{n}^{\zeta_{r}}+u_{n}^{\beta_{k}-\beta_{r+2}})\bigr), \nonumber \end{eqnarray} where the last line holds because $k=r+1$ and $h_{n}>1$ for $n$ large enough and the sequence $\beta_{i}$ is decreasing. This in turn implies, for $k=r+1$ again, \begin{equation} \label{B4-2333} a^{n}_{k}=O_{P}\bigl(h_{n}^{r+2-k+2m} ( u_{n}^{\beta _{k}-\beta_{r}+2q \zeta_{r}}+u_{n}^{\beta_{k}-\beta_{r+1}})\bigr). \end{equation} Then, exactly as above, we apply~(\ref{B4-17}) with $k=r$, and~(\ref{B4-22}) and also~(\ref{B4-2333}) with $k=r+1$, to get that~(\ref{B4-2333}) holds for $k=r$ as well. Repeating the argument, a~downward induction yields that indeed (\ref{B4-2333}) holds for all $k$ between $1$ and $r+1$. Thus~(\ref{B4-22}) holds with $q$ and $m$ substituted with $2q$ and $r+1+2m$. Hence ($P_{m,q,r}$) implies ($P_{r+1+2m,2q,r}$). Since obviously ($P_{m,q,r}$) $\Rightarrow$ ($P_{m,q^{\prime},r}$) for any $q^{\prime}\in[1,q]$, by a repeated use of the previous argument we deduce that if ($P_{m,1,r}$) holds for some $m\geq1$, then for any $q^{\prime}\geq1$ we can find $m(q^{\prime})\geq1$ such that ($P_{m(q^{\prime}),q^{\prime},r}$) holds as well. Now, assuming ($P_{m,q,r}$) for some $m,q,r$, we take $q^{\prime}=\frac {\zeta_{r+1}}{2\zeta_{r}} \vee1$ and $m^{\prime}=m(q^{\prime})$. What precedes yields ($P_{m^{\prime},q^{\prime},r}$), hence~(\ref{B4-23}) holds for all $k\leq r+1$, with $q^{\prime}$ and $m^{\prime}$. In view of our choice of $q^{\prime}$, this implies that ($P_{r+1+m^{\prime},1,r+1}$) holds. Since ($P_{1,1,1}$) holds, we see by induction that for any $r\leq j$ there exists $m_{r}\geq1$ such that ($P_{m_{r},1,r}$) holds. It remains to apply~(\ref{B4-22}) with $r=j$ and $m=m_{r}$ and $q=1$, and we get~(\ref{B4-21}) with $\nu=m_{j}$. This completes the proof. \section{\texorpdfstring{Proof of Theorem \lowercase{\protect\ref{theo-fisher}}}{Proof of Theorem 4}} The proof of Theorem~\ref{theo-fisher} is contained in the supplemental article [\citet{yacjacod11asupp}] \end{appendix} \begin{supplement \stitle{Supplement to ``Identifying the successive Blumenthal--Getoor indices of a discretely observed process''} \slink[doi]{10.1214/12-AOS976SUPP} \sdatatype{.pdf} \sfilename{aos976\_supp.pdf} \sdescription{This supplement contains the proof of Theorem~\ref{theo-fisher}.} \end{supplement}

https://arxiv.org/abs/2009.04997

Hyperbolic 4-manifolds with perfect circle-valued Morse functions

We exhibit some (compact and cusped) finite-volume hyperbolic four-manifolds M with perfect circle-valued Morse functions, that is circle-valued Morse functions $f\colon M \to S^1$ with only index 2 critical points. We construct in particular one example where every generic circle-valued function is homotopic to a perfect one.An immediate consequence is the existence of infinitely many finite-volume (compact and cusped) hyperbolic 4-manifolds $M$ having a handle decomposition with bounded numbers of 1- and 3-handles, so with bounded Betti numbers $b_1(M)$, $b_3(M)$ and rank of $\pi_1(M)$.

\section*{Introduction} \label{introduction:section} One of the most intriguing phenomena in 3-dimensional topology is the existence of many finite-volume hyperbolic 3-manifolds $M$ that fiber over $S^1$. On such manifolds, the surface fiber $\Sigma$ generates a normal subgroup $\pi_1(\Sigma) \triangleleft \pi_1(M)$ and determines a geometrically infinite covering $\widetilde M \to M$ diffeomorphic to $\Sigma \times \matR$. Every finite-volume hyperbolic 3-manifold is finitely covered by a 3-manifold that fibers over $S^1$ by a celebrated theorem of Agol and Wise \cite{A, W}. All the hyperbolic manifolds and orbifolds in this paper are tacitly assumed to be complete. We could ask whether such fibrations occur also in higher dimension $n\geq 4$. The answer is certainly negative in even dimensions, for a simple reason: by the generalised Gauss -- Bonnet formula, the Euler characteristic of an even-dimensional finite-volume hyperbolic manifold does not vanish, while that of a fibration does. The following seems a more sensible question to ask. \begin{quest} Are there finite-volume hyperbolic $n$-manifolds with perfect circle-valued Morse functions in all dimensions $n$? \end{quest} We explain the terminology. A \emph{circle-valued Morse function} on a compact manifold $M$, possibly with boundary, is a smooth map $f \colon M\to S^1$ such that $f|_{\partial M}$ has no critical points and $f$ has finitely many critical points, all of non degenerate type. This implies in particular that $f|_{\partial M}$ is a fibration. Every finite-volume hyperbolic $n$-manifold $M$ is either closed or the interior of a compact manifold with boundary $M^*$. In the latter case, a circle-valued Morse function on $M$ is by definition the restriction of one on $M^*$. As it is customary in Morse theory, a circle-valued Morse function $f$ on a compact manifold $M$ yields a \emph{circular handle decomposition} for $M$, obtained by starting from any regular fiber, thickening it, adding a $i$-handle corresponding to each index-$i$ critical point, and then finally gluing the left and right vertical boundaries of the resulting manifold via some diffeomorphism. A circle-valued Morse function $f\colon M \to S^1$ is \emph{perfect} if it has exactly $|\chi(M)|$ critical points. In general we have $$\chi(M) = \sum (-1)^ic_i$$ where $c_i$ is the number of critical points of index $i$. Therefore $f$ is perfect if and only if it has the minimum possible number of critical points allowed by $\chi(M)$, and this holds precisely when all the indices of the critical points have the same parity. In odd dimensions we have $\chi(M)=0$ and hence a perfect circle-valued Morse function is just a fibration. In dimension 2 it is a simple exercise to prove that every closed orientable surface has a perfect circle-valued Morse function. In dimension 4 a finite-volume hyperbolic four-manifold $M$ has $\chi(M)>0$ and a circle-valued Morse function on $M$ is perfect if and only if the critical points have index 0, 2, or 4. Since $M$ is connected and not simply connected, we deduce easily (by looking at the corresponding circular handle decomposition) that in fact only the index 2 is allowed. Our main contribution is to provide some first examples, both in the cusped and in the compact setting. \begin{teo} \label{main:teo} There are some finite-volume (compact and cusped) hyperbolic 4-manifolds with perfect circle-valued Morse functions. \end{teo} \begin{table} \begin{center} \begin{tabular}{cccccccc} Name & cusps & Euler & $b_0$ & $b_1$ & $b_2$ & $b_3$ & $b_4$ \\ \hline $W$ & 5 & 2 & 1 & 5 & 10 & 4 & 0 \\ $X$ & 24 & 8 & 1 & 21 & 51 & 23 & 0 \\ $Y$ & 1 & 1/12 & & & & & \\ $Z$ & 0 & 272 & 1 & 115 & 500 & 115 & 1 \end{tabular} \vspace{.2 cm} \nota{The hyperbolic 4-manifolds $W, X, Z$ and 4-orbifold $Y$ that we consider here. Each has some perfect circle-valued Morse functions. For each manifold or orbifold we display the number of cusps, the Euler characteristic, and the Betti numbers over $\matR$ (only for manifolds).} \label{WXYZ:table} \end{center} \end{table} We construct in particular three examples: two cusped hyperbolic 4-manifolds $W, X$ and one closed hyperbolic 4-manifold $Z$, each equipped with some perfect circle-valued Morse function. We also build an additional hyperbolic 4-orbifold $Y$ that has a perfect circle-valued Morse function (in some natural sense) with particularly small fibers. Some basic information on the topology of these four objects is collected in Table \ref{WXYZ:table}. This produces an immediate corollary. If $M$ is a cusped hyperbolic manifold, a handle decomposition for $M$ is by definition one for the compactification $M^*$. \begin{cor} There are infinitely many finite-volume (compact and cusped) hyperbolic 4-manifolds $M$ with a handle decomposition with bounded numbers of 1- and 3-handles, hence with bounded Betti numbers $b_1(M)$ and $b_3(M)$ and rank of $\pi_1(M)$. \end{cor} \begin{proof} If $M$ is cusped, in what follows we work with the compactification $M^*$, still denoted by $M$ for simplicity. If $M$ has a perfect circle-valued Morse function $f$, it inherits a circular handle decomposition, which further decomposes into a handle decomposition as follows: fix a handle decomposition of a regular fiber $N$ of $f$, thicken it, add a 2-handle for each singular point of $f$, and then add a $(i+1)$-handle for each $i$-handle of $N$ to close everything up. For every $n \geq 2$ we can construct a cyclic covering $M_n \to M$ and a lift $f_n \colon M_n \to S^1$ by unwrapping $n$ times along $f$. The lifted $f_n$ is a circle-valued Morse function with the same regular fiber $N$ as above, so $M_n$ has a handle decomposition with a fixed number of $i$-handles for all $i\neq 2$. \end{proof} Recall that there are only finitely many hyperbolic 4-manifolds $M$ with bounded $b_2(M)$, since $\Vol(M) = \frac{4\pi^2}3 \chi(M)$ and $\chi(M) \leq 2 + b_2(M)$, and there are only finitely many finite-volume hyperbolic 4-manifolds $M$ with bounded volume \cite{Wang}. The main purpose of this work is to show that perfect circle-valued Morse functions are quite frequent, at least among the very limited types of manifolds that we are able to investigate at present, that is those that decompose into right-angled polytopes. We construct the manifolds $W,X,Z$ by colouring some well-known right-angled polytopes, and then we build some circle-valued Morse functions by combining the techniques of Bestvina -- Brady \cite{BB} and Jankiewicz -- Norin -- Wise \cite{JNW}, plus some additional arguments that are presented in this paper. For the manifold $W$ listed in Table \ref{WXYZ:table} we are able to determine precisely the integral cohomology classes that are represented by a perfect circle-valued Morse function. \begin{teo} \label{W:teo} The cusped hyperbolic 4-manifold $W$ has $H^1(W; \matR) = \matR^5$. The cohomology classes that are represented by some perfect circle-valued Morse function form the intersection $$U \cap H^1(W; \matZ)$$ where $U= \{x_i \neq 0\}$ is the complement of the coordinate hyperplanes. \end{teo} The open set $U$ is dense, so a homologically generic map $W \to S^1$ is homotopic to a perfect Morse function. Moreover, the set $U$ is \emph{polytopal}, that is it is the cone over some open facets of a polytope. The polytope here is the hyperoctahedron, convex hull of the vertices $\pm e_i$, with $2^5 = 32$ facets. So $U$ has 32 connected components. Of course we know from Thurston \cite{Th} that a similar theorem holds for any hyperbolic 3-manifold, with $U$ being the cone over some faces of the unit ball polytope of the Thurston norm. (Recall that a perfect circle-valued Morse function in dimension 3 is a fibration.) Here is another immediate corollary of Theorem \ref{main:teo}. \begin{cor} There are some geometrically infinite hyperbolic 4-manifolds $M$ that are diffeomorphic to $N \times [0,1]$ with infinitely many 2-handles attached on both sides, where $N$ is (the interior of) a 3-manifold that is either closed or bounded by tori. Here $\pi_1(M)$ is finitely generated but not finitely presented and $b_2(M) = \infty$. \end{cor} \begin{proof} Pick a perfect circle-valued Morse function $f\colon M_0 \to S^1$ on a closed or cusped hyperbolic 4-manifold $M_0$. It lifts to a Morse function $\tilde f \colon \widetilde M_0 \to \matR$ on the abelian covering $\widetilde M_0$ determined by $\ker f$. The manifold $M = \widetilde M_0$ is as stated. We have $b_2(M) = \infty$ because there are infinitely many 2-handles. If $N$ is a regular fiber, any finite set of generators for $\pi_1(N)$ generates also $\pi_1(M)$, which is hence finitely generated. If $\pi_1(M)$ were also finitely presented, then $H_2(\pi_1(M))$ would be finitely generated, but the asphericity of $M$ implies that $H_2(M) = H_2(\pi_1(M))$, a contradiction. \end{proof} In all the cases we could investigate here, we found strong numerical evidence that such a geometrically infinite $M$ should be infinitesimally and hence locally rigid. In one case we have a rigorous computer-assisted proof, see Theorem \ref{rigid:teo}. The theoretical and computational aspects of these results will be exposed in a forthcoming paper \cite{Bat}. Many results exposed in this paper were obtained by using both Regina \cite{BBP} and SnapPy \cite{Sna} inside a Sage environment. \subsection*{Summary of the paper} The hyperbolic 4-manifolds $W$, $X$, and $Z$ are all constructed by colouring appropriately the facets of three well-known right-angled polytopes, that are respectively $P_4$, the ideal 24-cell, and the compact right-angled 120-cell. We recall in Section \ref{colours:subsection} how a coloured right-angled polytope $P$ defines a manifold $M$ tessellated into copies of $P$. In Section \ref{cube:subsection} we build a cube complex $C$ dual to this tessellation. In Section \ref{states:subsection} we introduce the notion of \emph{real state}, that extends that of a state given in \cite{JNW}. A real state $s$ is simply a real number assigned to every facet of $P$. It produces a cohomology class $[s] \in H^1(M, \matR)$ and a piecewise-linear \emph{diagonal map} $f\colon \widetilde M \to \matR$ on the universal cover representing that class, as explained in Section \ref{diagonal:subsection}. We then use Bestvina -- Brady theory to analyse this map carefully in Section \ref{ad:subsection}. When $[s]$ is integral and certain conditions are fulfilled, the map $f$ descends to a map $f\colon M \to \matR/_\matZ = S^1$ that can be smoothened to a perfect circle-valued Morse function. The conditions are stated quite generally in Theorem \ref{ad:teo}. In Section \ref{manifolds:section} we use this machinery to build $W, X, Y, Z$ and their perfect circle-valued Morse functions. In Section \ref{W:subsection} we construct $W$, prove Theorem \ref{W:teo}, and analyse the \emph{singular fibers} of $f$ in one particular case. In Section \ref{X:subsection} we construct $X$ and analyse the singular fibers of 63 distinct maps $f$ that arise naturally from the combinatorics. We get 63 distinct hyperbolic 3-manifolds with as much as 28 cusps each, all distinguished by their hyperbolic volume, listed in Tables \ref{fibers:table} and \ref{fibers2:table}. The infinitesimal rigidity of all the abelian covers considered is verified numerically in all these cases. Among these 63 maps $f$, one is particularly interesting due to its many symmetries. In Section \ref{Y:subsection} we use this map to build a very small orbifold $Y$ with an appropriate kind of circle-valued Morse function, for which we can fully determine both the \emph{singular} and \emph{regular} fibers: they both belong to the first segment of the census of cusped hyperbolic 3-manifolds \cite{CHW}. Finally, we produce the compact example $Z$ in Section \ref{Z:subsection}. We make some comments and raise some open questions in Section \ref{questions:section}. \subsection*{Acknowledgements} We thank Marc Culler and Matthias Goerner for helpful discussions on SnapPy. \section{The construction} \label{colourings:section} Our aim is now to construct some compact and cusped hyperbolic 4-manifolds equipped with perfect circle-valued Morse functions. The manifolds are obtained by assembling some right-angled polytopes, and then applying some techniques of Bestvina -- Brady \cite{BB} and Jankiewicz -- Norin -- Wise \cite{JNW}, plus some additional arguments. \subsection{Colours} \label{colours:subsection} All the hyperbolic 4-manifolds we consider here are built by colouring the facets of a right-angled polytope. This is a simple and fruitful technique already considered in other contexts, see for instance \cite{IMM, KM}. We briefly recall how it works. Let $P\subset \ensuremath {\mathbb{X}}^n$ be a right-angled finite polytope in some space $\ensuremath {\mathbb{X}}^n = \matH^n, \matR^n$ or $\matS^n$. A \emph{$c$-colouring} of $P$ is the assignment of a colour (taken from some finite set of $c$ elements) to each facet of $P$, such that incident facets have distinct colours. We typically use $\{1,\ldots, c\}$ as a palette of colours and always suppose that every colour is painted on at least one facet. A colouring defines a manifold $M$, obtained by mirroring $P$ iteratively along its coloured facets (in any order). More precisely, consider $2^c$ disjoint copies of $P$ denoted as $P^v$ where $v\in \matZ_2^c$ varies. We identify each facet $F$ of $P^v$ via the identity map to the same facet of $P^{v+e_j}$ where $j$ is the colour of $F$. The result of these identifications is a $n$-manifold $M$ having the same geometry $\ensuremath {\mathbb{X}}^n$ of $P$, tessellated into $2^c$ copies of $P$. From a more algebraic point of view, the manifold $M$ can be constructed as follows. Let $\Gamma< {\rm Isom}(\ensuremath {\mathbb{X}}^n)$ be the Coxeter group generated by the reflections along the facets of $P$. We define a homomorphism $\Gamma \to \matZ_2^c$ by sending the reflection along a facet $F$ to $e_i$, where $i$ is the colour assigned to $F$. The kernel $\Gamma'$ of the homomorphism is torsion-free and we get $M=\ensuremath {\mathbb{X}}^n/\Gamma'$. We can generalise naturally this algebraic construction by assigning arbitrary vectors in $\matZ_2^c$ to the facets, requiring that vectors attached to any set of facets with non-empty intersection are independent, but we will not need this generalisation here: see \cite{KSla} for an introduction. If we permute the colours in the palette the resulting manifold $M$ is unaffected, so colourings are typically considered only up to such permutations: in other words, we should think of a colouring as a partition of the facets of $P$. We are interested in colourings with a small number of colours and many symmetries, since they produce manifolds that are reasonably small and with many isometries. Here are some examples: \begin{figure} \begin{center} \includegraphics[width = 3.2 cm]{chainlink} \end{center} \nota{The minimally twisted chain link with 6 components. } \label{chainlink:fig} \end{figure} \begin{itemize} \item The $n$-cube has a unique $n$-colouring, where opposite facets are coloured with the same colour. This colouring produces a flat torus. More generally, one can prove that any colouring on the $n$-cube produces a flat torus. \item The right-angled spherical $n$-simplex has a unique colouring, that produces the spherical manifold $S^n$. \item The ideal octahedron in $\matH^3$ has a unique 2-colouring. The colouring produces a cusped hyperbolic 3-manifold which is the complement of the minimally twisted chain link with 6 components shown in Figure \ref{chainlink:fig}, see \cite{KM}. \item The ideal 24-cell in $\matH^4$ has a unique 3-colouring. It produces a hyperbolic 4-manifold with 24 cusps with 3-torus sections, already considered in \cite{KM, MR}. \item There is a sequence of hyperbolic right-angle polytopes $P_3,\ldots, P_8$ with $P_n \subset \matH^n$. Some natural symmetric colourings were constructed in \cite{IMM} for each $P_n$. These produce some hyperbolic $n$-manifolds. \end{itemize} The right-angled polytopes that we use here are $P_4$, the ideal 24-cell, and the right-angled 120-cell. We will describe their colourings later in detail. \subsection{The dual cube complex} \label{cube:subsection} A coloured right-angled polytope $P$ produces a manifold $M$ tessellated into $2^c$ copies of $P$, denoted as $P^v$ with varying $v\in \matZ_2^c$. We now construct a cube complex $C$ dual to this tessellation. We work in the PL category. We consider $P$ as a compact Euclidean polytope (using the Klein model for $\matH^n$). The PL manifold $M$ is decomposed into $2^c$ identical copies $P^v$ of $P$, possibly with some ideal vertices: these are the pre-images of the ideal vertices of $P$, and they should be removed to get $M$. We call this an (ideal) polyhedral decomposition for $M$. A \emph{$k$-face} of this decomposition is by definition a $k$-face of some $P^v$. By convention each $P^v$ is a $n$-face, but the ideal vertices are not 0-faces (while the finite vertices are 0-faces). We fix a barycentric subdivision of $P$, which lifts to a barycentric subdivision of the polyhedral decomposition of $M$. We define $C$ as the dual decomposition to the polyhedral decomposition of $M$ inside its barycentric subdivision. More precisely, we do the following. Every $k$-face $F$ in the polyhedral decomposition has a barycenter $v$, and the dual $(n-k)$-cell in $C$ is defined as the union of all the simplexes of the barycentric subdivision that intersect $F$ in $v$. \begin{prop} The construction produces a cube complex $C\subset M$ onto which $M$ retracts. There is a natural 1-1 correspondence between the $k$-faces of the polyhedral decomposition of $M$ and the $(n-k)$-cubes of $C$, which reverses all containments. If there are no ideal vertices we have $C=M$. In general $M\setminus C$ consists of cusps. \end{prop} \begin{proof} The barycentric subdivision of the polyhedral decomposition of $M$ can be constructed abstractly via the following standard two-steps procedure: (i) consider the poset of the faces of the polyhedral decomposition ordered by inclusion, and (ii) from the poset, define an abstract simplicial complex by considering every element as a vertex and every ascending chain as a simplex. By construction the cell dual to a $k$-face $F$ is the simplicial subcomplex of the barycentric subdivision obtained from the chains in the poset formed by all the faces that contain $F$. Since $P$ is right-angled, we deduce easily that the poset is isomorphic to the poset of the faces of a $(n-k)$-cube. Therefore the cell is a barycentric subdivision of a $(n-k)$-cube. All these cubes intersect nicely to form a cube complex $C$ dual to the polyhedral decomposition. If $P$ has some ideal vertices, the complement $M\setminus C$ consists of the open stars of the ideal vertices in the simplicial subdivision, without the ideal vertices: these are the cusps of $M$, and $M$ deformation retracts onto $C$. \end{proof} If $P \subset \matH^n$ has some ideal vertices, some more work is needed to enlarge the cube complex $C$ to a bigger cube complex $C^*$ whose interior is homeomorphic to $M$. This technical part is explained below and is necessary only in the cusped case, so it may be skipped at first reading. \subsubsection{The cusped case} If $P\subset \matH^n$ has some ideal vertices, we consider it as a Euclidean polytope as above, and truncate all the ideal vertices to get a new Euclidean polytope $P^*$. The counterimage of $P^*$ in $M$ is a compact PL submanifold $M^*$ decomposed into copies of $P^*$, whose interior is homeomorphic to $M$. We get a polyhedral decomposition of $M^*$ into copies of $P^*$. The truncation of every ideal vertex of $P$ produces a small $(n-1)$-cubic facet of $P^*$. In particular the decomposition of $M^*$ into copies of $P^*$ induces a decomposition of $\partial M^*$ into $(n-1)$-cubes. As above, a \emph{$k$-face} of the decomposition of $M^*$ is a $k$-face of some copy of $P^*$. There are two types of $k$-faces: the \emph{boundary} ones, that are contained in $\partial M^*$, and the \emph{interior} ones, that are not. We fix a barycentric subdivision of $P^*$, which lifts to a barycentric subdivision of the polyhedral decomposition of $M^*$. We define $C^*$ to be the dual decomposition (defined as above) of the polyhedral decomposition of $M^*$ in this barycentric subdivision. \begin{prop} The construction produces a cube complex $C^*$ with $C^*=M^*$. There is a natural 1-1 correspondence between the $k$-faces of the polyhedral decomposition of $M^*$ and the $(n-k)$-cubes of $C^*$ that are not contained in $\partial M^*$, which reverses all containments. The original $C$ is naturally a subcomplex of $C^*$. \end{prop} \begin{proof} Let $F$ be a $k$-face of $M^*$. If $F$ is an interior face, its barycenter lies in the interior of $M^*$, and the dual face is a subdivided $(n-k)$-cube as above. If $F$ is a boundary face, its barycenter lies in $\partial M^*$. Similarly as above, we see that the dual cell is still naturally a $(n-k)$-cube, since it is a barycentrically subdivided $(n-k)$-cube cut in half. The decomposition $C^*$ is naturally a cube complex, dual to the polyhedral decomposition of $M^*$ as stated. By considering the faces that do not intersect the boundary we get the original subcomplex $C \subset C^*$. \end{proof} We call $C$ and $C^*$ the \emph{dual} and the \emph{extended dual} cube complexes. \begin{example} \label{triangle:example} Let $P\subset \matH^2$ be an ideal triangle. Its three edges are non-adjacent, so we can colour them with a single colour. The resulting manifold $M$ is obtained by mirroring $P$ along its whole boundary and is hence a thrice-punctured sphere. Here $C$ is a $\theta$-shaped graph, a spine of $M$. The compact polygon $P^*$ is a hexagon, $M^*$ is a compact pair-of-pants, and $C^*$ is a cube complex with 6 squares that contains $C$. Two squares are attached to each edge of $C$. \end{example} \subsection{The forgetful map} We introduce a forgetful map from the cube complex $C$ to the standard $c$-cube that is quite simple and useful. Here $c$ is as usual the number of colours. The $2^c$ vertices of $C$ are dual to the polytopes $P^v$ of the polyhedral decomposition of $M$, with $v \in \matZ_2^c$. We indicate the vertex of $C$ dual to $P^v$ with the same element $v \in \{0,1\}^c$. The forgetful map will send $v$ to $v$, considered as a vertex of $[0,1]^c$. By construction a $k$-cube in $C$ is dual to a $(n-k)$-face $F$ of the polyhedral decomposition, which is contained in $2^k$ polytopes $P^v$ whose vectors $v$ differ only in some components $v_{i_1},\ldots, v_{i_k}$. The numbers $i_1,\ldots, i_k$ are precisely the colours of the facets in $P$ that contain the image of $F$. This is a crucial fact that leads us to prove the following. \begin{prop} There is a canonical cube complex map from $C$ to the $c$-cube $[0,1]^c$ that sends $v$ to $v$. It sends $k$-cubes to $k$-cubes. \end{prop} \begin{proof} Every vertex $w$ of the barycentric subdivision of $M$ is the barycenter of some face $F$. As we just said, the vectors $v$ of the $P^v$ adjacent to $F$ differ only at some coordinates $i_1,\ldots, i_k$. These vectors span a $k$-face of $[0,1]^c$. We send $w$ to the barycenter of this face. We get a simplicial map from $C$ to the barycentric subdivision of $[0,1]^c$, which is in fact a cube complex map (it sends $k$-cubes to $k$-cubes). \end{proof} We call it the \emph{forgetful map} because all the $k$-cubes in $C$ sharing the same vertices are sent to the same $k$-cube in $[0,1]^c$. It is neither injective nor surjective in general. \begin{example} If $P$ is a 1-coloured ideal triangle as in Example \ref{triangle:example} the forgeftul map sends the $\theta$-shaped graph $C$ onto the segment $[0,1]$. If $P$ is the 2-coloured ideal octahedron $P$ the cube complex $C$ has four vertices indicated as $(0,0), (0,1), (1,0), (1,1)$, every vertex $(x,y)$ is connected to $(x,1-y)$ and $(1-x,y)$ with four edges each, and there are 12 squares in $C$, all with the same 4 vertices (but attached to different four-uples of edges). The forgetful map maps everything to a single square. \end{example} \subsection{Real states and cocycles} \label{states:subsection} Let again $P\subset \ensuremath {\mathbb{X}}^n$ be a coloured right-angled polytope, producing a manifold $M$. Let $C$ be the dual cube complex. We extend the notion of state given in \cite{JNW} by defining a \emph{real state} $s$ to be the assignment of a real number $s(F)$ to each facet $F$ of $P$. The states considered in \cite{JNW} correspond to the case where one uses only the numbers $1$ and $-1$. In the next lines we show that a real state $s$ on $P$ produces a cohomology class $[s] \in H^1(M, \matR)$ and a \emph{diagonal map} $f_s\colon \widetilde M \to \matR$ on the universal cover $\widetilde M$ of $M$. The whole construction is similar to \cite{JNW}. Consider the forgetful map $C \to [0,1]^c$. Every edge of $[0,1]^c$ is oriented canonically like the coordinate axis parallel to it. The pull-back of these orientations give a \emph{canonical orientation} on the edges of $C$. On every $k$-cube of $C$ parallel edges are oriented coherently as in Figure \ref{arrows:fig}. \begin{figure} \begin{center} \includegraphics[width = 7 cm]{arrows} \end{center} \nota{Parallel edges of every $k$-cube are cooriented.} \label{arrows:fig} \end{figure} Let $s$ be a real state for $P$. We assign to every oriented edge $e$ of $C$ the real number $s(e) = s(F)$ where $F$ is the facet of $P$ that is the image of the $(n-1)$-face of the polyhedral decomposition dual to $e$. By assigning a real number to every oriented edge $e$ of $C$ we have just defined a cellular 1-cochain, that we still denote by $s$. \begin{prop} \label{cocycle:prop} The cellular cochain $s$ is a cocycle. \end{prop} \begin{proof} We need to prove that the sum of the contributions on every square $Q$ of $C$ is zero. The edges of $Q$ are oriented as in Figure \ref{arrows:fig}. The square $Q$ is dual to a $(n-2)$-face $F$ of the polyhedral decomposition, which projects to a $(n-2)$-face of $P$, that is in turn the intersection of two facets $F_1$ and $F_2$ of $P$. The four edges of $Q$ are dual to the four $(n-1)$-faces of the decomposition containing $F$. By construction, these four faces project alternatively to $F_1$ and $F_2$. Two opposite edges $e,e'$ of $Q$ are dual to faces that project to the same facet $F_i$ and hence $s(e)=s(e')$. Their contributions to the sum is zero since they are oriented as in Figure \ref{arrows:fig} (with opposite directions with respect to the canonical orientation of $\partial Q$). So the overall contribution on $Q$ is zero and $s$ is a cocycle. \end{proof} Every real state $s$ determines a cohomology class $[s] \in H^1(C,\matR) = H^1(M,\matR)$. (Remember that $M$ deformation retracts onto $C$.) We can characterise easily which states yield the same classes: \begin{prop} Let $s,s'$ be two real states. We have $[s] = [s']$ if and only if there is a $\lambda = (\lambda_1,\ldots, \lambda_c)\in \matR^c$ such that $s'(F) = s(F) + \lambda_i$ where $i$ is the colour of $F$, for every facet $F$ of $P$. \end{prop} \begin{proof} We suppose that here is a $\lambda\in\matR^c$ that fulfills the requirement, and we show that $[s] = [s']$. It suffices to consider the case where $\lambda_j = 0$ for all $j$ distinct from some fixed $i$, the general case will follow by iterating. Consider the 0-cochain $\alpha$ that assigns $\lambda_{i}$ to all the vertices $v$ with $v_{i}=0$ and 0 to the others. One can verify that $s' = s + d\alpha$ and hence $[s] = [s']$. Conversely, we suppose that $[s] = [s']$ and prove that there is a $\lambda$ that fulfills the requirement. We identify $P^0$ with $P$. Two facets $F_1, F_2$ of $P^0=P$ sharing the same colour $i$ are dual to two edges of $C$ connecting $P^0$ to $P^{e_i}$. These two edges form a 1-cycle $c$ (after reversing the orientation of the second edge), and we have $s(c) = s(F_1)-s(F_2)$. Since $s(c)$ depends only on $[s]$, we deduce that $s(c) = s'(c)$. In particular we may define unambiguously $\lambda_i = s'(F) - s(F)$ by taking any facet $F$ with colour $i$, and the resulting $\lambda$ is as required. \end{proof} A real state $s$ is \emph{balanced} if the sum of all the $s(F)$ among the facets $F$ sharing the same colour is zero, for every colour. We can deduce easily the following. \begin{cor} For every state $s$ there is a unique balanced state $s'$ with $[s'] = [s]$. \end{cor} A balanced state is like a combinatorial harmonic representative. The balanced states form naturally a subspace of $H^1(C,\matR) = H^1(M,\matR)$. This subspace actually coincides with $H^1(M,\matR)$ on the three manifolds $W, X, Z$ that we construct in this paper, but this does not hold in general: there may be cohomological elements that are not represented by states. \begin{example} Consider the $n$-cube with its $n$-colouring, producing the $n$-torus $M$. A balanced state assigns two opposite numbers to each opposite facets. The balanced states form a vector space of dimension $n$, naturally identified with $H^1(M, \matR)$. If we use more than $n$ colours, we still get a flat $n$-torus, but the balanced states form only a proper subspace of $H^1(M, \matR)$. Consider the ideal regular hyperbolic octahedron with its 2-colouring, producing the complement $M$ of the chain link in Figure \ref{chainlink:fig}. A balanced state assigns two 4-uples of real numbers with sum zero. The balanced states form a space of dimension 6, naturally isomorphic to $H^1(M, \matR)$. \end{example} \subsubsection{The cusped case} If $P\subset \matH^n$ is cusped, we may need to extend the cocycle $s$ to the extended cubulation $C^*$, so that both $s$ and its extension represent the same element in $H^1(C,\matR) = H^1(C^*,\matR) = H^1(M,\matR)$. There is a canonical way to do it that we now explain. There are two types of vertices in $C^*$, the \emph{boundary vertices} that are contained in $\partial M^*$, and the \emph{interior vertices} that are not: the latter ones are precisely the vertices of $C$. Correspondingly, there are two types of edges in $C^* \setminus C$, those that connect an interior to a boundary vertex and those that connect two boundary vertices. We orient every edge $e$ of the first type towards the boundary, and we set $s(e)=1$. An edge $e$ of the second type is parallel to some edge of $C$ and hence inherits from it both the canonical orientation and the real number $s(e)$. It is easy to verify that this extension is a cocycle in $C^*$ that represents the same cohomology class in $H^1(M,\matR)$. \subsection{Diagonal maps} \label{diagonal:subsection} Let $s$ be a real state on $P$. We now suppose that $s$ is nowhere vanishing, that is $s(F)\neq 0$ for all the facets $F$ of $P$. This allows us to modify the orientation of the edges of $C$ and $C^*$, by reversing the canonical orientation of $e$ if and only if $s(e)<0$. We call the resulting orientation the \emph{orientation induced by $s$}. Parallel edges in every $k$-cube are still cooriented, because they were all assigned the same real number (see the proof of Proposition \ref{cocycle:prop}). We fix henceforth this orientation, and modify accordingly the signs of all the labels $s(e)$ so that $s(e)>0$ on every edge $e$. Orientations and labels have changed but of course $s$ identifies the same cohomology class as before. We identify inductively each $k$-cube of $C$ or $C^*$ with the standard Euclidean $k$-cube $[0,1]^k$ in an edge-orientation-preserving way: that is, the unique vertex from which all edges are departing is sent to $0$. The identification is meant to be fixed only up to permuting the $k$ coordinates. We lift the orientation of the edges, the cocycle $s$, and the identifications with the Euclidean cubes from the cube complex $C$ to its universal cover $\widetilde C$. The state $s$ induces a \emph{diagonal map} $f\colon \widetilde C \to \matR$ as follows. Choose a basepoint vertex $v_0$ in $\tilde C$ and set $f(v_0) = 0$. Extend this map to every $k$-cube via the diagonal map $$f(x_1, \ldots, x_k) = f(0) + s(e_1)x_1 + \cdots + s(e_k)x_k$$ where $e_i$ is the edge contained in the $i$-th axis. The constant $f(0)$ is chosen inductively on each $k$-cube so that this map matches with the previously assigned cubes. There is no ambiguity since $\tilde C$ is simply connected. The resulting $f\colon \widetilde C\to \matR$ is a Morse function, in the sense of \cite{BB}. Using the basepoint $v_0$ and the identification of $\pi_1(C)$ with the deck transformations group, the map $f$ induces a homomorphism $\pi_1(C) \to \matR$ and hence a class in $H^1(M; \matR)$ that is equal to $[s]$. If $[s] \in H^1(M;\matZ)$ the class has integral periods and therefore $f$ descends to a map $f\colon C \to \matR/_{\matZ} = S^1$. In the cusped case we actually work with $C^*$ instead of $C$. In both cases we get a map $f\colon \widetilde M \to \matR$ on the universal cover $\widetilde M$, called a \emph{diagonal map}, which descends to a map $f\colon M \to \matR/_\matZ = S^1$ if the class is integral. We sometimes denote it as $f_s$ to stress the dependence on $s$. \subsection{The ascending and descending links} \label{ad:subsection} As we have seen, a nowhere vanishing real state $s$ on $P$ defines a diagonal map $f\colon \widetilde M \to \matR$. We can now apply Bestvina -- Brady theory to study this map topologically. Let $v$ be a vertex of $C$ (or $C^*$ in the cusped case). Recall that the edges of $C$ (or $C^*$) are oriented by $s$. Let $\lk(v)$ be the link of $v$ in $C$ (or $C^*$). By construction $\lk(v)$ is an abstract simplicial complex homeomorphic to $S^{n-1}$ (or $D^{n-1}$ if $v$ is a boundary vertex of $C^*$). Every vertex of $\lk(v)$ indicates an oriented edge of $C$ (or $C^*$) incident to $v$, and we assign to it the \emph{status} I (In) or O (Out) according to whether the oriented edge points towards $v$ or away from $v$. Following \cite{BB}, we define the \emph{ascending link} $\lk_\uparrow(v)$ (respectively, \emph{descending link} $\lk_\downarrow(v)$) to be the subcomplex of $\lk(v)$ generated by all the vertices with status O (respectively, I). In the following we suppose that $\dim P \leq 4$, as it will certainly be the case here. The condition stated below must hold at every $v$ in $C$ (or $C^*$ in the cusped case). We refer to \cite{RS} for the PL notions of collapse and of $k$-dimensional polyhedron. \begin{teo} \label{ad:teo} Suppose that at every $v$ both the ascending and descending links collapse to a connected polyhedron of dimension $\leq 1$, and to a point if $v$ is a boundary vertex. Then $f\colon \widetilde M \to \matR$ can be smoothened to a Morse function with only critical points of index 2. If the ascending and descending links collapse to points at every $v$, then $f$ can be smoothened to a fibration. \end{teo} \begin{proof} For a subspace $J\subset \matR$, we define $\widetilde M_J = f^{-1}(J)$. We write $\widetilde M_t = M_{\{t\}}$. If an interval $J \subset \matR$ contains no image of a vertex of $\widetilde C$ (or $\widetilde C^*$), then $\widetilde M_J$ is a union of prisms and hence PL homeomorphic to a product submanifold $M_t \times J$, for any $t \in J$. Suppose that $t\in \matR$ is the image of some vertex $v$. We suppose for the sake of simplicity that no other vertex is sent to $[t-\varepsilon, t+\varepsilon]$. The following argument is of local nature and easily extends to the general case. As shown in \cite[Lemma 2.5]{BB}, the submanifold $\widetilde M_{(-\infty,t+\varepsilon]}$ collapses onto the union $\widetilde M_{(-\infty,t-\varepsilon]} \cup C_v$ where $C_v$ is the cone over the simplicial complex $\lk_\downarrow(v)$, considered inside $\widetilde M_{t-\varepsilon}$. By hypothesis $\lk_\downarrow(v)$ collapses either to a point or to a 1-dimensional polyhedron in $\widetilde M_{t-\varepsilon}$. In the first case the manifold $\widetilde M_{[a,t+\varepsilon]}$ collapses to $\widetilde M_{[a,t-\varepsilon]}$, and hence $\widetilde M_{[t-\varepsilon, t+ \varepsilon]}$ is PL homeomorphic to a product $M_t \times [-\varepsilon, + \varepsilon]$. In the second case $\widetilde M_{[a,t+\varepsilon]}$ collapses to $\widetilde M_{[a,t-\varepsilon]}$ with a cone over a 1-polyhedron in $\widetilde M_{t-\varepsilon}$. Therefore $\widetilde M_{[a,t+\varepsilon]}$ is PL homeomorphic to $\widetilde M_{[a,t-\varepsilon]}$ with some $2$-handles attached. Recall that this is allowed only when $v$ is not a boundary vertex. By what just proved, the function $f$ determines a PL handle decomposition for $\widetilde M$, where we start with any regular fiber and proceed both in the positive and in the negative direction by attaching 2-handles whenever we cross a vertex $v$ whose ascending or descending link collapses onto a non-simply connected 1-polyhedron. Since we are in low dimension $n\leq 4$, this handle decomposition can be smoothened \cite{HM, M}. The smooth handle decomposition can finally be transformed into a Morse function with only index 2 critical points. \end{proof} \begin{rem}[Integral class] If $[s]$ is integral, the smoothenings can be done equivariantly so that $f$ descends to a cicle-valued Morse function $f\colon M \to \matR/_\matZ = S^1$ with only critical points of index 2. \end{rem} \begin{rem} [The algorithm] \label{algorithm:rem} We describe a simple algorithm to compute the ascending and descending link at every vertex $v$ of $C$ when $P$ is compact. The link of $v$ in $C$ is a simplicial complex dual to $P^v$. Every vertex $w$ of the link corresponds dually to a facet $F$ of $P^v$. Recall that the vertices of $C$ are identified with $\matZ_2^c$. At $v=0$, the status of a vertex $w$ of the link is O and I depending on whether $s(F)$ is positive or negative. Whenever we pass from $v$ to $v+e_i$, the stati of the vertices dual to the facets $F$ coloured with $i$ are switched, while all the others stay the same. Via these simple moves we can determine the stati and hence the ascending and descending links of all the vertices $v\in \matZ_2^c$. This is the basis for the combinatorial game of \cite{JNW}. \end{rem} In what follows, a \emph{hyperoctahedron} is the $(n-1)$-dimensional simplicial complex dual to the boundary of a $n$-cube. It is of course homeomorphic to $S^{n-1}$. \begin{example}[$n$-cubes] \label{cubes:example} Let $P$ be a $n$-cube with some colouring. This produces a flat $n$-torus $M$. Let $s$ be a nowhere vanishing state for $P$. If there are two opposite facets $F_1, F_2$ of $P$ with the same colour and $s(F_1)$, $s(F_2)$ have opposite signs, then all the ascending and descending links collapse to points and $f_s$ is a fibration. To prove this, note that the link of a vertex $v$ of $C$ is a hyperoctahedron. Using the previous remark we deduce that at every $v$ we have two opposite vertices of the hyperoctahedron with opposite status I and O. This implies easily that both the ascending and the descending links collapse to these opposite vertices, that is to points. \end{example} \subsubsection{The cusped case} Let $P\subset \matH^n$ have some ideal vertices, equipped with a colouring and a nowhere vanishing real state $s$. We now expose a simple criterion that, when verified, allows us to forget about the boundary vertices, and to use $C$ instead of $C^*$ for the interior vertices. \begin{prop} \label{ideal:prop} If every ideal vertex of $P$ is adjacent to two facets $F_1, F_2$ with the same colour and with $s(F_1), s(F_2)$ having opposite signs, the hypothesis of Theorem \ref{ad:teo} is satisfied at every boundary vertex of $C^*$. Moreover, to check the hypothesis on each interior vertex $v$, it suffices to consider the link of $v$ in $C$ instead of $C^*$. \end{prop} \begin{proof} The link in $C^*$ of a boundary vertex is a cone over a hyperoctahedron. As in Example \ref{cubes:example}, the hypothesis implies that there are two opposite vertices of the hyperoctahedron with opposite status, and from this we can easily deduce that the ascending and descending links both collapse to points. The $v$ be an interior vertex of $C^*$. We now prove that it suffices to consider its link in $C$ instead of $C^*$. The link of $v$ in $C$ is dual to $P$, with a hole corresponding to every ideal vertex of $P$. Indeed the link is homeomorphic to $S^{n-1}$ minus some open balls (the holes). The boundary of each hole is a hyperoctahedron. The link of $v$ in $C^*$ is obtained by filling all these holes, thus getting a simplicial complex homeomorphic to $S^{n-1}$. Each hole is filled by adding a cone over the corresponding hyperoctahedron, with a new central vertex $w$ that indicates an edge connecting $v$ to a boundary vertex of $C^*$. The vertex $w$ has status O because we have oriented all these edges towards the boundary. The descending links of $v$ in $C$ and $C^*$ are the same. The ascending link in $C^*$ is obtained from that in $C$ by adding the new central vertices $w$ at each hyperoctahedron. By hypothesis there are two opposite vertices in each hyperoctahedron with opposite status. As already noticed, this implies that the link of $w$ in the ascending link collapses to a point. Therefore the ascending link in $C^*$ collapse to the ascending link in $C$. So it suffices to prove Theorem \ref{ad:teo} using the link of $v$ in $C$ instead of $C^*$. \end{proof} One may in fact prove that this condition is also necessary, but we will not need that. Summing up: if the stated condition near the ideal vertices of $P$ is verified (facets with the same colours have opposite signs), we can forget about $C^*$ and we need only to determine the ascending and descending links of all the vertices of $C$ inside $C$. This can then be done using Remark \ref{algorithm:rem}, as in the compact case. \section{The manifolds} \label{manifolds:section} We define here the manifolds $W, X, Z$, the orbifold $Y$, and some perfect circle-valued Morse functions on them. Many of the calculations exposed here have been carried out with a program written in Sage that is publicly available from \cite{codeM}. The program takes as an input the adjacency matrix of a right-angled polytope $P$, a colouring of its facets, and a state that assigns a value $\pm 1$ at each facet. It determines the Betti numbers and the cusps of the manifold $M$ obtained from $P$ and its colours, and verifies whether the state fulfills the hypothesis of Theorem \ref{ad:teo}. It also constructs a triangulation of the (singular) fibers, in a format that can be read by SnapPy \cite{Sna} and Regina \cite{BBP}. \subsection{The manifold $W$} \label{W:subsection} We now define $W$ and prove Theorem \ref{W:teo}. As already mentioned, there is a sequence of remarkable right-angled polytopes $P_3,\ldots, P_8$ already considered by various authors \cite{ALR, PV, ERT}. We are interested here in $P_4$. The polytope is clearly presented in \cite{RT4}. It has 10 facets, 5 real vertices, and 5 ideal vertices. Every facet is isometric to the polyhedron $P_3$, a right-angled bipyramid with two real vertices and three ideal ones. Its orbifold Euler characteristic is $\chi(P_4) = \frac 1{16}$. \begin{figure} \begin{center} \includegraphics[width = 10 cm]{P4} \end{center} \nota{The adjacency graph of the 10 facets of $P_4$. Some edges are superposed for simplicity, so some caution is needed here: to clarify this ambiguity, we have chosen a blue vertex and painted in red the 6 vertices adjacent to it and in black those that are not adjacent, in two cases (all the other cases are obtained by rotation).} \label{P4:fig} \end{figure} \begin{figure} \begin{center} \includegraphics[width = 5 cm]{P4_coloured} \end{center} \nota{A 5-colouring for $P_4$.} \label{P4_coloured:fig} \end{figure} The adjacency graph of its facets is shown in Figure \ref{P4:fig}. We assign to $P_4$ the 5-colouring shown in Figure \ref{P4_coloured:fig}. The colouring construction described in Section \ref{colours:subsection} produces a hyperbolic cusped orientable 4-manifold $W$, tessellated into $2^5$ copies of $P_4$ and hence with $\chi(W) = 2$. At every ideal vertex $v$ of $P_4$ one can verify that the Euclidean cube cross-section inherits from $P_4$ a colouring with all the 5 colours involved, one of which appears twice in opposite faces of the cube. Such a colouring yields a flat 3-torus that decomposes into $2^5$ cubes. The hyperbolic manifold $W$ has 5 cusps, one for each ideal vertex of $P_4$. Each cusp has a 3-torus section. Using Sage (or by hand) we find that $H^1(W; \matR) = \matR^5$, and the cohomology class $x\in \matR^5$ is described by the balanced real state $s$ that assigns $x_i/2$ and $-x_i/2$ to the two facets that share the $i$-th colour, as explained in Section \ref{states:subsection}. The factor $1/2$ is there only to ensure that the lattice $H^1(W; \matZ)$ corresponds to $\matZ^5$. This state $s$ defines a diagonal map $f_s$ on the universal cover of $W$, as explained in Section \ref{diagonal:subsection}. \begin{figure} \begin{center} \includegraphics[width = 12.5 cm]{P4_states} \end{center} \nota{The ascending and descending links for $P_4$, up to isomorphisms. A black (white) vertex has status I (O). The descending link, generated by the black vertices, is shown below. In the first case we get a circle, while in the other cases we always get a contractible complex made of 3 triangles, 2 triangles, and 1 tetrahedron and 1 triangle respectively. The ascending links are the same (ordered differently).} \label{P4_states:fig} \end{figure} \begin{prop} If $x_i \neq 0$ for all $i$, then $f_s$ can be smoothened to a perfect circle-valued Morse function with two index-2 singular points. \end{prop} \begin{proof} We show that the hypothesis of Theorem \ref{ad:teo} are fulfilled. The criterion of Proposition \ref{ideal:prop} is verified: every ideal vertex is adjacent to two facets with the same colour, and $s$ has opposite values on these by assumption. Hence we can use $C$ instead of $C^*$. To identify the ascending and descending links at the vertices $v\in \matZ_2^5$ of $C$ we follow Remark \ref{algorithm:rem}. We get the simplicial complexes shown in Figure \ref{P4_states:fig}. They all collapse either to a point or to a circle. There are two index-2 singular points: we find this either by inspection, or by recalling that $\chi(W)=2$. \end{proof} This completes the proof of the constructive part of Theorem \ref{W:teo}. To conclude we need to show that if $x_i=0$ then there exists no perfect circle-valued Morse function. The reason is quite simple: if $f\colon W \to S^1$ represents a cohomology class with $x_i=0$, we can prove that its restriction to the $i$-th cusp of $W$ is homotopically trivial. Therefore $f$ cannot be a fibration there, and it is not homotopic to any circle-valued Morse function of any kind. \subsubsection{The singular and regular fibers} \label{sr:subsubsection} We have proved the remarkable fact that each nowhere-vanishing $x\in H^1(W,\matZ) = \matZ^5$ is represented by a circle-valued Morse function $f\colon W \to S^1$ with two critical points of index 2. We now would like to understand these functions $f$ topologically. The isometries of $W$ act by switching the signs of each coordinate $x_i$ arbitrarily, so we can always suppose that $x$ lies in the first orthant $x_i>0$. Using Sage we have studied the case $x=(1,1,1,1,1)$. We now expose the topological information that we have found in this case. The function $f\colon W \to \matR/\matZ = S^1$ has two critical points, with values $0$ and $1/2$. We have two singular fibers $f^{-1}(0)$ and $f^{-1}(1/2)$ and two regular fibers $f^{-1}(1/4)$ and $f^{-1}(3/4)$. Each singular fiber is a 3-manifold with 6 boundary tori $W^{\rm sing}$, with one boundary torus coned to a point. The two nearby regular fibers $W^{\rm reg}$ have 5 boundary tori (one inside each cusp of $W$) and are obtained by substituting this cone point with a circle in two different ways: the two regular fibers are obtained one from the other via integral Dehn surgery along a knot. We have used Sage to identify all these fibers. Both singular fibers are the same hyperbolic manifold $W^{\rm sing}$ with 6 cusps (one of which is coned) and volume $\sim$ 65.318656269. It has a geometric triangulation with 72 tetrahedra. Both the regular fibers are the same hyperbolic manifold $W^{\rm reg}$ with 5 cusps and volume $\sim 54.991958042$. It has a geometric triangulation with 70 tetrahedra. The fact that we get twice the same manifolds is probably due to the symmetries of $W$. Using SnapPy \cite{Sna} we can see that indeed $W^{\rm reg}$ can be obtained from $W^{\rm sing}$ by filling one cusp along two different slopes at distance 1 -- this phenomenon is called \emph{cosmetic surgery} \cite{BHW}. There is an isometry $\psi$ of $W^{\rm sing}$ that sends the first slope to the second, and this explains why we get the same filled manifold. SnapPy also tells us a fact that seems relevant to us: the filled $W^{\rm reg}$ has an additional isometry $\varphi$ that is \emph{not} induced from $W^{\rm sing}$, and that acts homologically trivially on the cusps. \begin{rem}[Ping pong] \label{ping:pong:rem} Although we did not prove it (the combinatorics involved is non-trivial), we believe that the map $f$ should have the following dynamical behaviour, that may be described as a \emph{ping-pong} between the isometries $\psi$ and $\varphi$. We start with the regular fiber $W^{\rm reg}$ at some time $t$. We increase $t$ and when we cross a critical value the fiber is surgered along a knot, whose complement is $W^{\rm sing}$, to get a new manifold that is in fact diffeomorphic to the original $W^{\rm reg}$ via $\psi$. Now we act on $W^{\rm reg}$ via $\varphi$ and repeat the process from the beginning. The role of $\varphi$ should be fundamental to ``mix everything up'' like in the familiar pseudo-Anosov monodromies: since $\varphi$ is an additional isometry, not induced from one on $W^{\rm sing}$, we are not just surgerying all the time along the same knot -- that picture would be too simple and could not hold because it would produce some non-peripheral $\matZ \times \matZ$ inside $\pi_1(W)$, like with the reducible mapping classes in dimension $2+1=3$. In some vague sense, the two isometries $\psi$ and $\varphi$ look like the matrices $\matr 1101$ and $\matr 1011$, that when multiplied give rise to $\matr 2111$, the Anosov monodromy of the famous fibration of the figure-8 knot complement, fully guaranteed not to fix any non-peripheral closed curve on the punctured torus. \end{rem} \subsubsection{The abelian cover: strong numerical evidence for its rigidity}\label{abeliancover:subsection} We keep studying the case $x=(1,1,1,1,1)$. Consider the abelian cover $\widetilde W$ induced by $\ker f$. It is a geometrically infinite manifold with limit set $S^3$, obtained topologically from $W^{\rm reg} \times [0,1]$ by attaching infinitely many 2-handles on both sides. As noted in the introduction, the fundamental group of $\widetilde W$ is finitely generated but not finitely presented, since $b_2(\widetilde W) = \infty$. It is natural to ask whether the hyperbolic structure on the geometrically infinite $\widetilde W$ is rigid. One usually start by investigating whether the holonomy representation is infinitesimally rigid. Although this is a linear problem, it is not straightforward because $\dim {\rm Isom}(\matH^4)=10$ and $\pi_1(\widetilde W)$ has many generators and infinitely many relators. We have elaborated a general strategy that applies to abelian covers constructed from right-angled polytopes, colours, and states, and written a separate computer program. We found very strong numerical evidence for the infinitesimal rigidity of the holonomy of $\widetilde W$ and of any other similar abelian cover investigated in this paper. This will be explained in a forthcoming paper \cite{Bat}. We proved the infinitesimal rigidity rigorously in one case in Section \ref{very:symmetric:subsubsection} below. The strong numerical evidence for the infinitesimal rigidity of $\widetilde W$ and other abelian covers is in contrast to dimension three, where as a consequence of the (now proved \cite{BrBr, B, NS}) density conjecture we know that every geometrically infinite 3-manifold can be deformed into a geometrically finite one. This contrast is however not surprising: it is natural to experience more rigidity in dimension $n\geq 4$, especially for a manifold that has infinitely many 2-handles. Nevertheless, these seem to be the first examples of rigid geometrically infinite hyperbolic manifolds in any dimension. See \cite{KS} for a related rigidity result (that however does not apply in this case). \subsection{The manifold $X$} \label{X:subsection} We now construct one more cusped example, via another right-angled polytope, the ideal 24-cell $\calC \subset \matH^4$. This regular polytope has 24 facets, each being a right-angled ideal regular octahedron. It has a unique 3-colouring, shown in Figure \ref{colori_numeri:fig}. This produces a very symmetric cusped hyperbolic 4-manifold $X$ that was already considered in \cite{KM, MR}. We have $\chi(\calC) = 1$ and $\chi(X) = 8$. \begin{figure} \begin{center} \includegraphics[width = 11 cm]{colori_numeri} \end{center} \nota{The picture shows the dual $\calC^*$ of the 24-cell $\calC$, which is again a 24-cell. Not all the edges are drawn, for simplicity: every yellow vertex is the center of a cube and we should add 8 more edges connecting it with the vertices of the cube. The vertices of $\calC^*$ are 3-coloured. The picture is taken from similar figures in \cite{JNW}.} \label{colori_numeri:fig} \end{figure} At every ideal vertex of $\calC$ we have a Euclidean cube coloured with three colours, that gives rise to a 3-torus cusp section. The manifold $X$ has 24 such cusps, one for each ideal vertex of $\calC$. \begin{rem}[Quaternions] The colouring of $\calC$ can be described in the following way. Consider $\calC$ inside the quaternions space $\matH$. The dual of $\calC$ is another 24-cell $\calC^*\subset \matH$, whose vertices form the binary tetrahedral group $T_{24}^*$. Now $Q_8= \{\pm 1, \pm i, \pm j, \pm k\}$ is a normal subgroup $Q_8 \triangleleft T_{24}^*$ of index 3. The three lateral classes of $Q_8$ subdivide the 24 vertices of $\calC^*$ into three octets, and we assign a colour to each octet. \end{rem} The existence of various states for $\calC$ that give rise to connected ascending and descending links was first proved in \cite{JNW}. The original goal of our work was to better understand the topology of the resulting algebraically fibering maps. As in \cite{JNW}, we let a \emph{state} be a real state $s$ that assigns the number $\pm 1$ to each facet of $\calC$. (Actually, one should take $\pm 1/2$ to get a primitive integral class with our convention, but we ignore this point.) We have written a Sage program that classifies all the states $s$ for which the hypothesis of Theorem \ref{ad:teo} are satisfied, considered up to symmetries of $\calC$ and up to switching all the signs on facets with the same colours (both these operations produce the same maps $f_s$ up to an isometry of $X$). Among them, we find 63 states that are particularly interesting because the ascending and descending links all collapse to circles: they form a Hopf link at each vertex of the dual cubulation $C$. For such states $s$, Theorem \ref{ad:teo} furnishes a perfect circle-valued Morse function $f_s\colon M \to S^1$ that has precisely one critical point inside each 24-cell of the tessellation. This is coherent, since the total number of critical points is equal to $\chi(M)=8$, which is in turn equal to the number of 24-cells in the tessellation, since $\chi(\calC)=1$. These 63 states $s$ give rise to potentially topologically different types of perfect circle-valued Morse functions $f_s$ on the same hyperbolic manifold $X$. By construction, in all the cases the function $f_s$ has two singular values $0$ and $1/2$ in $\matR/_\matZ=S^1$, whose counterimages consist of 4 singular points each, like in Figure \ref{functions:fig}. There are two types of regular fibers $f^{-1}(1/4)$ and $f^{-1}(3/4)$ and two types of singular fibers $f^{-1}(0)$ and $f^{-1}(1/2)$. Each regular fiber $X^{\rm reg}$ is a 3-manifold with 24 torus boundary components, one inside each cusp of $X$. Each singular fiber is a 3-manifold $X^{\rm sing}$ with 28 toric boundary components, four of which have been coned. The two singular fibers are actually homeomorphic, because they are related by the isometry of $X$ that sends $\calC^{(v_1,v_2,v_3)}$ to $\calC^{(1-v_1,1-v_2,1-v_3)}$ identically. \begin{figure} \begin{center} \labellist \small\hair 2pt \pinlabel $f$ at 110 40 \pinlabel $X$ at 10 220 \pinlabel $X^{\rm reg}$ at 15 130 \pinlabel $X^{\rm sing}$ at 78 82 \pinlabel $X^{\rm reg}$ at 125 200 \pinlabel $X^{\rm sing}$ at 147 65 \pinlabel $X^{\rm reg}$ at 185 130 \pinlabel $0$ at 75 20 \pinlabel $1/2$ at 140 20 \pinlabel $\tilde f$ at 417 40 \pinlabel $\widetilde X$ at 310 220 \pinlabel $X^{\rm reg}$ at 290 193 \pinlabel $X^{\rm sing}$ at 260 65 \pinlabel $X^{\rm reg}$ at 373 200 \pinlabel $X^{\rm sing}$ at 328 82 \pinlabel $X^{\rm reg}$ at 420 193 \pinlabel $X^{\rm sing}$ at 395 65 \pinlabel $X^{\rm reg}$ at 503 200 \pinlabel $X^{\rm sing}$ at 458 82 \pinlabel $X^{\rm reg}$ at 550 193 \pinlabel $X^{\rm sing}$ at 525 65 \pinlabel $-1/2$ at 260 20 \pinlabel $0$ at 330 20 \pinlabel $1/2$ at 395 20 \pinlabel $1$ at 460 20 \pinlabel $3/2$ at 525 20 \endlabellist \includegraphics[width = 12.5 cm]{functions} \end{center} \nota{The circle-valued Morse function $f\colon X \to S^1$ (left) and its lift $\tilde f \colon \widetilde X \to \matR$ (right). There are two critical values 0 and $1/2$ for $f$, each containing 4 critical points in its counterimage, indicated with an X.} \label{functions:fig} \end{figure} We have determined $X^{\rm sing}$ for each state $s$ using a Sage code \cite{codeM} that uses both Regina \cite{BBP} and SnapPy \cite{Sna}. The manifold $X^{\rm sing}$ has 28 cusps and its first homology is $\matZ^{28}$ in all cases. The manifolds are listed in Tables \ref{fibers:table} and \ref{fibers2:table}. They are all hyperbolic and, quite remarkably, they can all be distinguished by their volumes. Without the great help of hyperbolic geometry it would have been very difficult to distinguish these 3-manifolds. Each manifold has an Epstein -- Penner canonical decomposition that is further cut into geometric hyperbolic ideal tetrahedra, whose number is shown in the last column. The data in the tables should be interpreted as numerical results and not as rigorous proofs: in case of need, one may wish to use the SnapPy verified computations to try to upgrade this discussion to a rigorous argument, but given the numbers of manifolds and tetrahedra involved we have refrained from doing this. Finally, each state $s$ gives rise to an abelian cover $\widetilde X_s$. Using the method and code mentioned in Subsection \ref{abeliancover:subsection} and described in \cite{Bat} we could get strong numerical evidence for the infinitesimal deformation of the holonomy representation of $\widetilde X_s$ in all the 63 cases. \begin{table} \begin{center} \begin{tabular}{cccc} N & Volume & Symmetry group & Tetrahedra \\ \hline 1 & 194.868788430654 & nonabelian group of order 192 & 192 \\ 2 & 189.47275083534 & D4 & 192 \\ 3 & 186.874353850537 & Z/2 + D4 & 192 \\ 4 & 186.340011593947 & Z/2 + Z/2 & 192 \\ 5 & 185.307321263554 & Z/2 & 192 \\ 6 & 185.035244912975 & D4 & 200 \\ 7 & 184.813075164402 & Z/2 + D4 & 192 \\ 8 & 184.301433361768 & nonabelian group of order 64 & 192 \\ 9 & 184.066725592654 & Z/2 + Z/2 & 196 \\ 10 & 183.87312700617 & Z/2 & 194 \\ 11 & 183.866948358501 & Z/2 & 193 \\ 12 & 183.544395602474 & D4 & 192 \\ 13 & 183.436816803216 & Z/2 + Z/2 + Z/2 & 192 \\ 14 & 183.392747517971 & Z/2 & 194 \\ 15 & 183.121509768255 & Z/2 + Z/2 & 196 \\ 16 & 182.360395141194 & Z/2 & 192 \\ 17 & 182.280935940832 & Z/2 + Z/2 & 192 \\ 18 & 182.171456556108 & 0 & 203 \\ 19 & 181.283359592032 & D4 & 192 \\ 20 & 181.127484303344 & Z/2 & 196 \\ 21 & 181.024645445741 & Z/2 & 196 \\ 22 & 180.934130108819 & Z/2 + Z/2 & 212 \\ 23 & 180.824987315671 & D4 & 200 \\ 24 & 180.660993296517 & Z/2 & 194 \\ 25 & 180.450694474362 & D4 & 200 \\ 26 & 180.387179283461 & Z/2 & 201 \\ 27 & 180.33108563585 & Z/2 & 190 \\ 28 & 180.248483292705 & Z/2 + Z/2 & 196 \\ 29 & 180.127608138661 & Z/2 + Z/2 & 198 \\ 30 & 179.869062465521 & Z/4 & 200 \\ 31 & 179.754141009737 & 0 & 197 \\ 32 & 179.656944120778 & Z/2 & 196 \\ 33 & 179.181472240992 & Z/2 & 194 \\ 34 & 178.902836506372 & Z/2 + D4 & 196 \\ 35 & 178.795804830626 & Z/2 + D4 & 192 \\ 36 & 178.709806437094 & 0 & 192 \\ 37 & 178.550276957958 & Z/2 + Z/2 & 202 \end{tabular} \vspace{.2 cm} \nota{The singular fibers that we obtained for $X$ by assigning different states to $\calC$. All these hyperbolic 3-manifolds have 28 cusps and $H_1=\matZ^{28}$. The last column shows the number of tetrahedra in the Epstein-Penner canonical decomposition (possibly after some subdivision).} \label{fibers:table} \end{center} \end{table} \begin{table} \begin{center} \begin{tabular}{cccc} N & Volume & Symmetry group & Tetrahedra \\ \hline 38 & 178.49844092298 & Z/2 + Z/2 + Z/2 & 196 \\ 39 & 178.355491610596 & nonabelian group of order 64 & 192 \\ 40 & 178.321815160374 & D4 & 204 \\ 41 & 177.898810518335 & D3 & 201 \\ 42 & 177.794415210729 & 0 & 194 \\ 43 & 177.551934214513 & Z/2 + D4 & 188 \\ 44 & 177.362796117873 & Z/2 & 198 \\ 45 & 177.250459763042 & nonabelian group of order 32 & 200 \\ 46 & 177.110609049916 & Z/2 + Z/2 & 196 \\ 47 & 176.982387058175 & Z/2 + Z/2 & 198 \\ 48 & 176.898729129159 & 0 & 198 \\ 49 & 175.421613694494 & 0 & 198 \\ 50 & 175.170011870291 & Z/2 + Z/2 & 192 \\ 51 & 175.085280565074 & Z/2 + Z/2 & 188 \\ 52 & 174.081891361605 & D4 & 200 \\ 53 & 173.80796075894 & Z/2 & 194 \\ 54 & 173.331415822106 & Z/2 + Z/2 + Z/4 & 192 \\ 55 & 173.211174252127 & D4 & 200 \\ 56 & 172.692650600288 & Z/2 & 192 \\ 57 & 172.581866456537 & D4 & 184 \\ 58 & 172.160507136111 & nonabelian group of order 32 & 192 \\ 59 & 171.484300906217 & Z/2 + D4 & 184 \\ 60 & 170.918046348271 & Z/2 + Z/4 & 192 \\ 61 & 166.465531880726 & Z/2 + Z/2 & 186 \\ 62 & 163.949780648344 & Z/4 & 184 \\ 63 & 154.990534348097 & Z/2 + Z/2 + D8 & 176 \end{tabular} \vspace{.2 cm} \nota{The singular fibers that we obtained for $X$ by assigning different states to $\calC$. All these hyperbolic 3-manifolds have 28 cusps and $H_1=\matZ^{28}$. The last column shows the number of tetrahedra in the Epstein-Penner canonical decomposition (possibly after some subdivision). (continue).} \label{fibers2:table} \end{center} \end{table} \subsubsection{A very symmetric case} \label{very:symmetric:subsubsection} The largest hyperbolic manifold in Table \ref{fibers:table} deserves more attention. By construction, every manifold $X^{\rm sing}$ among those listed in the tables has a topological triangulation with 192 tetrahedra, and hence volume $\leq 192 V$ where $V$ is the volume of the ideal regular tetrahedron (the canonical decomposition may contain less or more than 192 tetrahedra). The first manifold in Table \ref{fibers:table} decomposes precisely into 192 regular ideal tetrahedra, so it has the largest volume, and moreover also the largest symmetry group, yet of order 192. The manifold arises from a very symmetric state $s$ that we now describe. \begin{figure} \begin{center} \includegraphics[width = 12.5 cm]{colori} \end{center} \nota{The very symmetric state $s$. The black and white vertices of $\calC^*$ represent the status I and O respectively (right). In the left figure we show both the colouring and the state.} \label{colori:fig} \end{figure} This very symmetric state $s$ is shown in Figure \ref{colori:fig}. It has the following appealing algebraic description: each of the three lateral classes of $Q_8$ inside $T_{24}^*$ is preserved by the left multiplication by $i$, and the action decomposes it into two orbits of four vertices each. Assign the status $+1$ (also called O) to one orbit and $=-1$ (also called I) to the other. Up to symmetries, the resulting state $s$ is independent of the choice of the orbits; moreover, we would get the same $s$ (up to symmetries) also if we chose either left- or right-multiplication of any of $\pm i, \pm j, \pm k$. \begin{figure} \begin{center} \includegraphics[width = 7 cm]{nastri} \end{center} \nota{The descending link is a band decomposed into 12 triangles. The ascending link is isomorphic to it, and altogether they form two bands that collapse onto a Hopf link in $S^3$.} \label{nastri:fig} \end{figure} In this case, if we apply the algorithm of Remark \ref{algorithm:rem} to detect the ascending and descending links at every vertex $v\in \matZ_2^3$, we find the same state $s$ (up to symmetries of $\calC$) at every interior vertex $v$. At every $v$ the ascending and descending links form altogether a pair of bands in $S^3$ that collapse to a Hopf link. One band is pictured in Figure \ref{nastri:fig}. With some patience, we can prove by hand that in this very symmetric context the singular fiber $X^{\rm sing}$ decomposes into 192 ideal tetrahedra, and that every edge of the resulting triangulation is adjacent to precisely 6 of them. This implies immediately that $X^{\rm sing}$ has a hyperbolic structure obtained by assigning to each tetrahedron the structure of a regular ideal one. \begin{figure} \begin{center} \makebox[\textwidth][c]{\includegraphics[width = 16 cm]{Toro_9}} \end{center} \nota{The triangles in the 2-skeleton of the 24-cell that separate facets with opposite I/O status form a torus $T$, shown here from nine different viewpoints. In the pictures, we omit for the sake of clarity six triangles that are outside of the central cube and have a vertex at infinity. We may see that every vertex is adjacent to six triangles.} \label{torus:fig} \end{figure} The ideal triangulation of $X^{\rm sing}$ can be constructed by taking, for each $v\in \matZ_2^3$, the cone with vertex the center of $\calC$ over all the ideal triangles in $\calC$ that separate two facets having opposite status. These ideal triangles form a torus (with 24 punctures) shown in Figure \ref{torus:fig}. One can check that every vertex of the triangulated torus is adjacent to exactly six triangles. In this very symmetric case we were able to calculate also the regular fiber $X^{\rm reg}$, that is a hyperbolic 3-manifold with 24 cusps and volume $\sim 152.510077$. It may be identified as the double cover of the (quite messy) orbifold $O$ shown in Figure \ref{orbifold:fig}. The orbifold $O$ is tessellated into 24 (non regular!) ideal octahedra, precisely as the boundary of the 24-cell, and $X^{\rm reg}$ is tessellated into 48 of them. \begin{figure} \centering \makebox[\textwidth][c]{\includegraphics[width=16 cm, height= 3.5 cm]{Sing_3}} \nota{The hyperbolic orbifold $O$ has base space $S^3$ (with 24 punctures) and singular locus the (quite complicated) knotted 4-valent graph shown here with respect to three different viewpoints. The graph has 24 vertices that should be removed and indicate 24 cusps, each based on the flat orbifold $(S^2,2,2,2,2)$. For the sake of clarity, we omit the vertex at infinity and four edges that are connected to it. The graph is contained in the 1-skeleton of the 24-cell. It has many symmetries that are unfortunately not apparent from the figure.} \label{orbifold:fig} \end{figure} \begin{rem}[Fibers are pleated] We should mention that, like in the more familiar fibrations in dimension 3, the fibers $X^{\rm reg}$ and $X^{\rm sing}$ are embedded in $X$ in a very pleated way: they are decomposed respectively into octahedra and tetrahedra that form some angles along their 2-dimensional faces, in such a complicated way that any connected component $\bar X$ of the preimage of $X^{\rm reg}$ or $X^{\rm sing}$ in $\matH^4$ has the whole sphere at infinity $S^3 = \partial \matH^4$ as a limit set. This is a general fact when we analyse perfect circle-valued Morse functions on hyperbolic 4-manifolds. Remember that the fibers $X^{\rm reg}$ and $X^{\rm sing}$ are not $\pi_1$-injective in $X$, because the index-2 critical points annihilate some of the elements in their fundamental groups. This implies that the connected component $\bar X$ just mentioned is not simply connected. \end{rem} Finally, for this very symmetric state $s$ we could prove rigorously the following fact, see \cite{Bat} for details. \begin{teo} \label{rigid:teo} The holonomy representation of the abelian cover $\widetilde X$ is infinitesimally, and hence locally, rigid. \end{teo} \subsection{A very small orbifold example} \label{Y:subsection} Let $\widetilde X$ be the abelian cover of $X$ constructed from the very symmetric state of Section \ref{very:symmetric:subsubsection}. The manifolds $X$ and $\widetilde X$ have plenty of symmetries, many of which preserve the fibers of $f$. It is therefore tempting to try to quotient them by some symmetries, in order to find smaller examples. By quotienting $\widetilde X$ with an appropriate group of symmetries we have found a very small hyperbolic 4-orbifold $Y$ with a circle-valued Morse function $f\colon Y \to S^1$ whose singular and regular fibers appear in the very first segment of the SnapPea census \cite{CHW} as the manifolds ${\tt m036}$ and ${\tt m203}$. The construction of $Y$ may be of independent interest, so we briefly describe it. Let $\calC$ be the ideal regular 24-cell, with center $v$. Let $P$ be the cone on $v$ over a octahedral facet $O$ of $\calC$. Then $P$ is a 4-dimensional pyramid, with a octahedral base and eight tetrahedral lateral facets. It has 8 ideal vertices and one real one. Its dihedral angles are $2\pi/3$ at the triangles containing $v$ (the lateral ones), and $\pi/4$ at the triangles contained in $O$ (the base ones). It is not a Coxeter polytope since $2\pi/3$ does not divide $\pi$, but it may be used nevertheless to construct manifolds. \begin{figure} \begin{center} \labellist \small\hair 2pt \pinlabel $s$ at 160 20 \endlabellist \includegraphics[width = 10 cm]{new_lens} \end{center} \nota{There is only one way to pair the faces of the two octahedra so that all the symbols at the edges match. The result is the lens space $L(12,5)$. } \label{new_lens:fig} \end{figure} Figure \ref{new_lens:fig} shows two octahedra $O_1$ and $O_2$ and a face pairing between them giving the lens space $L(12,5)$. Note that every edge has valence 3: if we give each $O_i$ the structure of a spherical regular octahedron with dihedral angles $2\pi/3$, we get $L(12,5)$ with its spherical metric. We construct $Y$ by picking two copies $P_1$ and $P_2$ of $P$, and identifying their lateral facets by extending the pairing shown in Figure \ref{new_lens:fig} for their base octahedra $O_1$ and $O_2$. Then we glue $O_1$ to $O_2$ along the following isometry: first rotate $O_1$ of an angle $\pi$ along the axis $s$ shown in Figure \ref{new_lens:fig}, and then translate it to $O_2$. To verify that we get an orbifold $Y$ with $v$ as the only singular point, we just check that this identification produces two orbits of eight triangles, and since $8 \cdot \pi/4 = 2\pi$ we are done. The orbifold $Y$ has a single singular point with link $L(12,5)$, a single cusp, and $\chi(Y) = 1/12$. It can in fact be obtained by quotienting $\widetilde X$ by an appropriate group of isometries, and it inherits from it a circle-valued ``Morse function'' $f\colon Y \to S^1$, that has a single ``index-2 singular point'' at $v$. (A Morse function $f$ on a 4-manifold near an index-2 critical point is equivalent to $f(x_1,x_2,x_3,x_4) = x_1^2+x_2^2 -x_3^2-x_4^2$. Such a function $f$ is $\SO(2) \times \SO(2)$-invariant and hence descends to a function on the cone over any lens space.) \begin{figure} \begin{center} \includegraphics[width = 3 cm]{L6a2} \end{center} \nota{The link {\tt L6a2} from Thistelthwaite's link table. Its complement is the hyperbolic manifold ${\tt m 203}$.} \label{L6a2:intro:fig} \end{figure} \begin{figure} \begin{center} \includegraphics[width = 4.5 cm]{m036} \end{center} \nota{The hyperbolic manifold ${\tt m036}$ is obtained by pairing the faces of the octahedron with the unique pair of isometries that match the arrows on the edges. The manifold so obtained has three edges with valence 3, 3, and 6.} \label{m036:fig} \end{figure} The regular fiber of $f$ is the single-cusped manifold $Y^{\rm reg} = {\tt m036}$ from the census \cite{CHW}, with volume $\sim$ 3.177293 and homology $\matZ \times \matZ_3$. It decomposes as a single (non regular) ideal octahedron \cite{HPP} as in Figure \ref{m036:fig}. The singular fiber is the twice-cusped $Y^{\rm sing} ={\tt m203}$, that is the complement of the link in Figure \ref{L6a2:intro:fig}, with one cusp coned; it has volume $\sim$ 4.059766 and decomposes into four regular ideal tetrahedra \cite{CHW}. The manifolds $Y^{\rm reg}$ and $Y^{\rm sing}$ are covered by $X^{\rm reg}$ and $X^{\rm sing}$. Using SnapPy we discover that if we Dehn fill one cusp of {\tt m203} by filling the slope $\pm \frac 32$ we indeed get {\tt m036}. It is not necessary to specify which component of the link is filled since they are symmetric. SnapPy also says that there is an isometry $\psi$ of ${\tt m203}$ that sends the slope $\frac rs$ to $-\frac rs$, and this explains why we get the same manifold ${\tt m036}$ by filling the slope $\pm \frac 32$. The intersection between the two slopes is $\det \matr 3{-3}22 = 12$, coherently with the fact that $v$ is a cone over $L(12,5)$. As with the manifold $W$ considered in Section \ref{W:subsection}, using SnapPy we also discover that ${\tt m036}$ has an \emph{additional isometry} $\varphi$ that is not induced from ${\tt m203}$. A ping-pong dynamics between $\varphi$ and $\psi$ as in Remark \ref{ping:pong:rem} is very likely to hold also here. \subsubsection{A geometrically infinite Kleinian group in $\matH^4$ with 2 generators} The detour from $X$ to $Y$ also led us to find an explicit holonomy $\Phi$ for the regular fiber $Y^{\rm reg}$ of $Y$. The holonomy is interesting because its image in ${\rm Isom}(\matH^4)$ is discrete and has limit set $S^3$ and the group has only two generators. However, recall that the holonomy is not injective because $Y^{\rm reg}$ is not $\pi_1$-injective in $Y$. The holonomy is the following. A presentation for $\pi_1({\tt m036})$ is $$\langle\ a, b\ |\ a^{-2}b^{-2}ab^{-2}a^{-2}b \ \rangle.$$ We have $$\Phi(a) = \begin{pmatrix} 1/2 & -1/2 & 1/2 & 1/2 & 0 \\ 1/2 & 1/2 & -1/2 & 1/2 & 0 \\ 1/2 & -1/2 & -1/2 & -1/2 & 0 \\ 1/2 & 1/2 & 1/2 & -1/2 & 0 \\ 0 & 0 & 0 & 0 & 1 \end{pmatrix}$$ and $$ \Phi(b) = \begin{pmatrix} -7/2 & 3/2 & 3/2 & 3/2 & 3\sqrt 2 \\ -3/2 & 1/2 & 1/2 & -1/2 & \sqrt 2 \\ 3/2 & 1/2 & -1/2 & -1/2 & -\sqrt 2 \\ 3/2 & -1/2 & 1/2 & -1/2 & -\sqrt 2 \\ -3\sqrt 2 & \sqrt 2 & \sqrt 2 & \sqrt 2 & 5 \end{pmatrix}. $$ Both $\Phi(a)$ and $\Phi(b)$ are elliptic of order 12, while $a$ and $b$ have infinite order in $\pi_1({\tt m036})$. The representation $\Phi$ is \emph{type-preserving}, that is it sends parabolics to parabolics. The geometrically infinite discrete group $\Gamma = \Phi(\pi_1({\tt m036})) $ has only two generators and these are expressed explicitly above as some elliptic transformations. The limit set of $\Gamma$ is the whole sphere at infinity $S^3 = \partial \matH^4$. The group $\Gamma$ has infinitely many conjugacy classes of finite subgroups: the existence of such phenomena in dimension 4 is not a surprise \cite{Kf}. We can also descend one dimension further: we discover from \cite{Bell} that ${\tt m036}$ fibers over $S^1$ with fiber the surface $\Sigma$ with genus 2 and one puncture. Therefore there is a type-preserving representation $\Phi\colon \pi_1(\Sigma) \to \Isom(\matH^4)$ whose image is a discrete geometrically infinite group with limit set $S^3 = \partial \matH^4$. \begin{quest} Can we deform type-preservingly the representations $\Phi$ of $\pi_1({\tt m036})$ and $\pi_1(\Sigma)$? Can we find a path connecting them to their Fuchsian representations, or to any geometrically finite discrete representation? \end{quest} \subsection{A compact example} \label{Z:subsection} We finish by describing the compact example $Z$. Let $P$ be the compact right-angled 120-cell. It has a natural 5-colouring that may be described using quaternions as follows. The facets of $P$ are naturally identified with 120 elements of the binary icosahedral group $I^*_{120}$. This contains the binary tetrahedral group $T^*_{24}$ as an index-5 (not normal) subgroup. The five left lateral classes of $T^*_{24}$ in $I^*_{120}$ furnish a 5-colouring for $P$. Recall that $\chi(P) = 17/2$. The 5-colouring produces a compact hyperbolic 4-manifold $Z$ with $\chi(Z) = 272$. To build the state $s$, we try to mimic the algebraic construction that worked very well with the 24-cell: we pick the state for $T_{24}^*$ described in Section \ref{very:symmetric:subsubsection} and we extend it to the lateral classes by left-multiplying with some given elements. Using Sage we find that one $s$ constructed in this way fulfills the requirements of Theorem \ref{ad:teo} and hence produces a perfect circle-valued Morse function $f\colon Z \to S^1$ with as much as 272 singular points of index 2. An algebraic fibration on a manifold tessellated by copies of the 120-cell was already constructed in \cite{JNW}. \begin{rem}[Large Euler characteristic] With $P_4$ and $\calC$ the ascending and descending links collapsed either to a point or to a circle. This allowed us to conclude that each copy of $P_4$ and $\calC$ in the decomposition contains either 0 or 1 singular points. This was possible since $\chi(P_4) = 1/16$ and $\chi(\calC)=1$, so the average number of singular points in each polytope of the decomposition is $1/16$ and $1$ in these cases. Here we have $\chi(P) = 17/2$, and this is the average number of singular points in each copy of $P$. It is therefore impossible to find only ascending and descending links that collapse to points or circles: some of them must collapse to more complicated 1-complexes. As a consequence, the Morse function $f_s$ is not canonically determined by the combinatorics: we are coning along some Heegaard graphs in $S^3$ and there are many ways to perturb this to a disjoint simultaneous attachment of 2-handles. For this reason we did not make any attempt to determine the singular fibers. \end{rem} \section{Related results and open questions} \label{questions:section} The main inspiration of this work is the paper \cite{JNW} where the authors construct some algebraic fibrations on some hyperbolic 4-manifolds that cover the right-angled 24- and 120-cells. Our first contribution here is to promote these algebraic fibrations to perfect circle-valued Morse functions. Other related recent contributions to finding algebraic fibrations are \cite{AS, FV, Kie}. Like fibrations, perfect circle-valued Morse functions lift to any finite-sheeted cover. It therefore makes sense to ask the following rather far-reaching question. \begin{quest} Does every finite-volume hyperbolic $n$-manifold have a finite cover that has a perfect circle-valued Morse function? \end{quest} For the moment not a single example of perfect circle-valued Morse function is known in dimension $n\geq 5$. Given the lifting property of such functions, the question can be rephrased as follows: does every commensurability class contain some representative that has a perfect circle-valued Morse function? Concerning dimension $n=4$, the examples exhibited in this paper show that the answer is positive for the commensurability classes of $P_4$, the right-angled 24-cell, and the right-angled 120-cell. The first two commensurability classes are actually the same \cite{RT4}. Every Coxeter simplex in dimension four belongs to one of these two classes \cite{JKRT}, see \cite{M} for a survey on hyperbolic 4-manifolds. So in particular the Davis manifold \cite{Da} and the Conder -- Maclachlan manifold \cite{CM} also belong to the second commensurability class, and in particular they virtually have a perfect circle-valued Morse function. It is easy to construct a hyperbolic 4-manifold that does \emph{not} admit a perfect circle-valued Morse function: it suffices to pick a cusped one that has at least one cusp section homeomorphic to the Hantzsche-Wendt manifold. Since this flat 3-manifold does not fiber, the cusped 4-manifold has no circle-valued Morse function of any kind. All the 22 orientable manifolds in \cite{RT4} contain such a section. There is also a cusped hyperbolic 4-manifold whose cusp sections are all of this kind \cite{FKS}. Another obvious obstruction to having a perfect circle-valued Morse function is the vanishing of the first Betti number over the real numbers. Some of the cusped manifolds in \cite{RT4} have $b_1=0$, but no compact hyperbolic 4-manifold with $b_1=0$ seems known at present. In particular, we do not know a single example of a compact hyperbolic 4-manifold that is known not to have a perfect circle-valued Morse function. We have seen in Theorem \ref{W:teo} that the open set $U$ is \emph{polytopal}, that is it is the cone over the open facets of a polytope. \begin{quest} Let $M$ be a hyperbolic $n$-manifold with $n\geq 3$. Is there always an open polytopal subset $U\subset H^1(M;\matR)$ such that the classes represented by circle-valued Morse functions form the intersection $U\cap H^1(M;\matZ)$? \end{quest} This could be related to the fact that the BNS invariant is of polytopal type for many groups, see \cite{K}. The cohomology classes represented by perfect circle-valued Morse functions are contained in the BNS invariant of $\pi_1(M)$ since they have finitely generated kernel. \begin{quest} Let $M^{\rm reg}$ be a regular fiber of a perfect circle-valued Morse function on a finite-volume hyperbolic 4-manifold $M$. Is $M^{\rm reg}$ necessarily hyperbolic? Which hyperbolic 3-manifolds can arise in this way? \end{quest} If $M^{\rm reg}$ were not aspherical, we would get a counterexample to Whitehead's asphericity conjecture, since the abelian cover of $M$ is aspherical and obtained from $M^{\rm reg}$ by attaching infinitely many 2-handles.

https://arxiv.org/abs/1101.4262

A simple yet complex one-parameter family of generalized lorenz-like systems

This paper reports the finding of a simple one-parameter family of three-dimensional quadratic autonomous chaotic systems. By tuning the only parameter, this system can continuously generate a variety of cascading Lorenz-like attractors, which appears to be richer than the unified chaotic system that contains the Lorenz and the Chen systems as its two extremes. Although this new family of chaotic systems has very rich and complex dynamics, it has a very simple algebraic structure with only two quadratic terms (same as the Lorenz and the Chen systems) and all nonzero coefficients in the linear part being -1 except one -0.1 (thus, simpler than the Lorenz and Chen systems). Surprisingly, although this new system belongs to the family of Lorenz-type systems in some existing classifications such as the generalized Lorenz canonical form, it can generate not only Lorenz-like attractors but also Chen-like attractors. This suggests that there may exist some other unknown yet more essential algebraic characteristics for classifying general three-dimensional quadratic autonomous chaotic systems.

\section{Introduction} The scientific story of chaos dates back to earlier 1960s \cite{Alorenz}, when Lorenz studied the atmospheric convection phenomenon he found a chaotic attractor in a three-dimensional (3D) quadratic autonomous system \cite{Alorenz2}. The now-classic Lorenz system is described by \begin{equation} \left\{ \begin{array}{l} \dot{x}=\sigma \left( y-x\right) \\ \dot{y}=rx-y-xz \\ \dot{z}=-bz+xy \end{array \right. \end{equation} which is chaotic when $\sigma =10,\ r=28,\ b=\frac{8}{3}.$ Ever since then, many researchers were wondering and pondering whether the discovery of the Lorenz system was just a lucky incident or there actually exist other closely related systems around it? In 1999, from an engineering feedback anti-control approach, Chen provided a certain answer to this question by his finding of a new system \cite{Achen1999,Aueta2000}, lately referred to as the Chen system by others, described by \begin{equation} \left\{ \begin{array}{l} \dot{x}=a(y-x)\\ \dot{y}=(c-a)x-xz+cy \\ \dot{z}=-bz+xy \end{array \right. \end{equation} which is chaotic when $a=35,\ b=3,\ c=28$. Whereafter, it has been proved that Chen's attractor exists \cite{AChenexists} and the Chen system displays even more sophisticated dynamical behaviors than the Lorenz system \cite{Achendynamical2003}. It is also interesting to recall a unified chaotic system, which was constructed to encompass both the Lorenz system and the Chen system \cite{Alu2002b}. This unified chaotic system is by nature a convex combination of the two systems, and is described by \begin{equation}\label{unified} \left\{ \begin{array}{l} \dot{x}=(25\alpha+10)(y-x)\\ \dot{y}=(28-35\alpha)x-xz+(29\alpha-1)y \\ \dot{z}=-\frac{\alpha+8}{3}z+xy. \end{array} \right. \end{equation} When $\alpha=0$ it is the Lorenz system while with $\alpha=1$ it is the Chen system, and moreover for any $\alpha\in(0,1)$ the system remains to be chaotic. As a result of several years of continued research endeavor along the same line, a family of generalized Lorenz systems was found and characterized \cite{CFCV1994,CFCV2002a,CFCV2002b,CFCV2005}, which is defined through the so-called generalized Lorenz canonical form, as follows: \begin{eqnarray} \left( \begin{array}{c} \dot x \\ \dot y \\ \dot z \\ \end{array} \right)= \left( \begin{array}{ccc} a_{11} & a_{12} & 0 \\ a_{21} & a_{22} & 0 \\ 0 & 0 & a_{33} \\ \end{array} \right)\left( \begin{array}{c} x \\ y \\ z \\ \end{array} \right)+ \left( \begin{array}{c} 0 \\ -xz \\ xy \\ \end{array} \right), \end{eqnarray} where $x,\,y,\,z$ are the state variables of the system. This canonical form contains a family of chaotic systems with the same nonlinear terms, the same symmetry about the $z$-axis, the same stability of three equilibria, and similar attractors in shape with a two-scroll and butterfly-like structure According to \cite{CFCV1994} (see also \cite{CFvanecek1996}), the Lorenz system satisfies $a_{12}a_{21}>0$ while the Chen system was found to satisfy $a_{{12}}a_{{21}}<0$. In this sense, the Chen system is {\it dual\/} to the Lorenz system. In between, it was also found that there is a transition, the L\"{u} system, which satisfies $a_{{12}}a_{{21}}=0$ \cite{Alu2002}. Lately, another form of a unified Lorenz-type system and its canonical form were developed, which contain some generalized Lorenz-type systems and their corresponding conjugate Lorenz-type systems \cite{yang2006,yang2007}. Moreover, another classification was developed in \cite{yang2008}, where system (3) is classified into two groups, the Lorenz system group if $a_{11}a_{22}>0$ and the Chen system group if $a_{11}a_{22}<0$. This classification leads to the finding of a transition, the Yang-Chen chaotic system, which satisfies $a_{{11}}a_{{22}}=0$. Comparing these two classification, one can see that every such system is classified according to its algebraic structure. Specifically, it is determined either by $a_{12}a_{21}$ or by $a_{11}a_{22}$. Given the above discussions, at this point it is interesting to ask the following questions: \begin{enumerate} \item Concerning the conditions on the signs of $a_{12}a_{21}$, defined in \cite{CFCV1994} (see also \cite{CFvanecek1996}), are these signs essential to the system dynamics? For instance, can a system belonging to the Lorenz systems family generate Chen-like attractors? \item Concerning the conditions on the signs of $a_{11}a_{22}$, defined in \cite{yang2008}, are these signs essential to the system dynamics? For instance, can a system belonging to the Chen system group generate Lorenz-like attractors? \item Concerning the unified chaotic system (\ref{unified}), is there any other simpler chaotic system that can also generate both Lorenz-like attractor and Chen-like attractor? \end{enumerate} This paper aims to provide certain answers to the above questions via the finding of a new family of Lorenz-like systems. This new family has only one real parameter in a simple algebraic form but demonstrates very rich and complex dynamics in a way similar to the generalized Lorenz canonical form \cite{CFCV2002a,CFCV2002b,CFCV2005}. It belongs to the Lorenz-type of systems defined in \cite{CFCV1994} and it is also classified into the Chen system group defined in \cite{yang2008}. However, it has a very simple form and can generate both visually Lorenz-like attractors and Chen-like attractors by gradually changing the single parameter. This reveals that further study of the system algebraic structure and its effects on the system dynamics remains an important and interesting challenge. \section{The new family of chaotic systems} The new system under investigation is described by \begin{equation}\label{chenwangeq} \left\{ \begin{array}{l} \dot{x}=-x-y \\ \dot{y}=-x+ry-xz \\ \dot{z}=-0.1z+xy, \end{array} \right. \end{equation} where $r$ is a real parameter. This is a one-parameter family of chaotic systems in the sense that as the real parameter $r$ is gradually varied, a sequence of chaotic attractors can be continuously generated from the system, with very rich and complicated dynamics, as further demonstrated below. Before proceeding to chaotic dynamics, some basic dynamical properties of this new system is analyzed following a standard routine, which is necessary for understanding the nature of the new system. \subsection{Symmetry and dissipativity} First, it is apparent that system (\ref{chenwangeq}) has a natural symmetry about the $z$-axis under the following transformation: \[ S(x,y,z)\rightarrow(-x,-y,z). \] Second, one may construct the following Lyapunov function: \begin{equation} V(x,y,z)=\frac{1}{2}(x^2+y^2+z^2), \end{equation} which gives \begin{eqnarray} \dot V(x,y,z)&=&x\dot x+y\dot y+z\dot z\nonumber\\ &=&x(-x-y)+y(-x+ry-xz)+z(-0.1z+xy)\nonumber\\ &=&-(x+y)^2+(r+1)y^2-0.1z^2. \end{eqnarray} This implies that system (\ref{chenwangeq}) is globally uniformly asymptotically stable about its zero equilibrium if $r<-1$. Consequently, system (\ref{chenwangeq}) is not chaotic in the parameter region $r<-1$. Third, since \[ \nabla V=\frac{\partial{\dot x}}{\partial x}+\frac{\partial{\dot y}}{\partial y} +\frac{\partial{\dot z}}{\partial z}=r-1.1\,, \] the system is dissipative under the condition of $r<1.1$. More precisely, since $V(t)=V(t_0)e^{-(1.1-r)t}$, any initial volume $V_0$ containing the system trajectories shrinks to zero as $t\to +\infty$ at an exponential rate of $1.1-r$. \subsection{Equilibria and their stability} It is obvious that the origin is a trivial equilibrium of system (\ref{chenwangeq}). Other nonzero equilibria can be found by solving the following equations simultaneously: \[ -x-y=0;\qquad -x+ry-xz=0;\qquad -0.1z+xy=0. \] When $r>-1$, system (\ref{chenwangeq}) has three equilibria: $O(0,0,0)$, $E_1\left(\sqrt{\frac{1+r}{10}},-\sqrt{\frac{1+r}{10}},-1-r\right)$, $E_2\left(-\sqrt{\frac{1+r}{10}},\sqrt{\frac{1+r}{10}},-1-r\right)$. By linearizing system (\ref{chenwangeq}) at $O(0,0,0)$, one obtains the Jacobian \begin{eqnarray} J|_O =\left[ \begin{array}{ccc} -1 & -1 & 0 \\ -1-z & r & -x \\ y & x & -0.1 \\ \end{array} \right], \end{eqnarray} whose characteristic equation is \[ {\rm det}(\lambda I-J|_O) =\lambda^3+(1.1-r)\lambda^2-(1.1r+0.9)\lambda-0.1(r+1)=0, \] which yields \begin{eqnarray} \lambda_1&=&-0.1<0,\nonumber\\ \lambda_2&=&-0.5+0.5r+\frac{\sqrt{5+2r+r^2}}{2}>0,\nonumber\\ \lambda_3&=&-0.5+0.5r-\frac{\sqrt{5+2r+r^2}}{2}<0.\nonumber \end{eqnarray} Obviously, the equilibrium $O(0,0,0)$ is a saddle point, at which the stable manifold $W^s(O)$ is two-dimensional and the unstable manifold $W^u(O)$ is one-dimensional. Similarly, linearizing the system with respect to the other equilibria, $E_1$ and $E_2$, yields the following characteristic equations: \[ {\rm det}(\lambda I-J|_{E_{1,2}}) =\lambda^3+(1.1-r)\lambda^2+0.2\lambda+0.2(r+1)=0. \] According to the Routh-Hurwitz criterion, the following constraints are imposed: \begin{eqnarray} \Delta_1&=&(1.1-r)>0,\\ \Delta_2&=&\left| \begin{array}{cc} 1.1-r & 0.2(r+1) \\ 1 & 0.2 \\ \end{array} \right|=0.1-2r>0,\\ \Delta_3&=&0.2(r+1)\Delta_2>0. \end{eqnarray} This characteristic polynomial has three roots, all with negative real parts, under the condition of $-1<r<0.05$. Therefore, the two equilibria $E_1$ and $E_2$ are both stable nodes, or node-foci, if $-1<r<0.05$. However, if the condition $0.05<r<1.1$ holds, then the characteristic equation has one negative real root and one pair of complex conjugate roots with a positive real part. Therefore, the two equilibria $E_{1,2}$ are both saddle-foci, at each of which the stable manifold $W^s(E_{1,2})$ is one-dimensional and the unstable manifold $W^u(E_{1,2})$ is two-dimensional. Summarizing the above analysis and discussions establishes the following result: \begin{theorem} With parameter $-1<r<1.1$, system (\ref{chenwangeq}) has three equilibria: \[ O(0,0,0), \qquad E_1\left(\sqrt{\frac{1+r}{10}},-\sqrt{\frac{1+r}{10}},-1-r\right), \qquad E_2\left(-\sqrt{\frac{1+r}{10}},\sqrt{\frac{1+r}{10}},-1-r\right). \] Furthermore,\\ {\rm(i)}\ \ $O(0,0,0)$ is a saddle point, at which the stable manifold $W^s(O)$ is two-dimensional and the unstable manifold $W^u(O)$ is one-dimensional;\\ {\rm(ii)}\ if $0.05<r<1.1$, then the two equilibria $E_1$ and $E_2$ are both saddle-foci, at each of which the stable manifold $W^s(E_{1,2})$ is one-dimensional and the unstable manifold $W^u(E_{1,2})$ is two-dimensional;\\ {\rm(iii)} if $-1<r<0.05$, then the equilibria $E_1$ and $E_2$ are both stable nodes, or node-foci, at each of which the stable manifold $W^s(E_{1,2})$ is three-dimensional. \end{theorem} The Jacobian eigenvalues evaluated at the three equilibria of the unified chaotic system (\ref{unified}) and of the new system (\ref{chenwangeq}) are listed in Table 1 for comparison. \begin{table}[h] \centering \caption{Equilibria and eigenvalues of several typical systems.} {\begin{tabular}{c c c c}\\[-2pt] \hline Systems & Equations & Equilibria & Eigenvalues \\[6pt] \hline\\[-2pt] Unified system & $\dot x=10(y-x)$&$(0,0,0)$ &$-22.8277,-2.6667,11.8277$ \\[1pt] $\alpha=0$ &$\dot y=28x-y-xz$&{}&{}\\[2pt] {} &$\dot z=-\frac{8}{3}z+xy$ &$(\pm6\sqrt{2},\pm6\sqrt{2},27)$ &$-13.8546,\textbf{0.0940}\pm0.1945i$ \\\hline\\[-2pt] Unified system&$\dot x=\frac{45}{2}(y-x)$&$(0,0,0)$ &$-28.1696,-2.8333,19.1696$ \\[1pt] $\alpha=0.5$ &$\dot y=\frac{21}{2}x+\frac{27}{2}y-xz$&{}&{}\\[2pt] {} &$\dot z=-\frac{17}{6}z+xy$ &$(\pm\sqrt{68},\pm\sqrt{68},24)$ &$-16.9593,\textbf{2.5630}\pm13.1857i$ \\\hline\\[-2pt] Unified system&$\dot x=30(y-x)$&$(0,0,0)$ &$-30.000,-2.9333,22.200$ \\[1pt] $\alpha=0.8$ &$\dot y=\frac{111}{5}y-xz$&{}&{}\\[2pt] {} & $\dot z=-\frac{44}{15}z+xy$ &$(\pm\frac{2\sqrt{407}}{5},\pm\frac{2\sqrt{407}}{5},22.2)$ &$-17.9535,\textbf{3.6101}\pm14.3037i$ \\\hline\\[-2pt] Unified system&$\dot x=35(y-x)$&$(0,0,0)$ &$-30.8357,-3,23.8359$ \\[1pt] $\alpha=1$ &$\dot y=-7x+28y-xz$&{}&{}\\[2pt] {} &$\dot z=-3z+xy$ &$(\pm3\sqrt{7},\pm3\sqrt{7},21)$ &$-18.4288,\textbf{4.2140}\pm14.8846i$ \\\hline\hline\\[-2pt] New system&$\dot x=-x-y$&$(0,0,0)$ &$-0.1,0.6544,-1.6044$ \\[1pt] $r=0.05$ &$\dot y=-x+0.05y-xz$&{}&{}\\[2pt] {} &$\dot z=-0.1z+xy$ &$(\pm\sqrt{42}/20,\mp\sqrt{42}/20,-1.05)$ &$-1.05,\textbf{0}\pm0.4472i$ \\\hline\\[-2pt] New system&$\dot x=-x-y$&$(0,0,0)$ &$-0.1,0.6913,-1.5913$ \\[1pt] $r=0.1$ &$\dot y=-x+0.1y-xz$&{}&{}\\[2pt] {} &$\dot z=-0.1z+xy$ &$(\pm\sqrt{11}/10,\mp\sqrt{11}/10,-1.1)$ &$-1.0162, \textbf{0.0081}\pm0.4652i$ \\\hline\\[-2pt] New system&$\dot x=-x-y$&$(0,0,0)$ &$-0.1,1,-1.5$ \\[1pt] $r=0.5$ &$\dot y=-x+0.5y-xz$&{}&{}\\[2pt] {} &$\dot z=-0.1z+xy$ &$(\pm\sqrt{15}/10,\mp\sqrt{15}/10,-1.5)$ &$-0.8101, \textbf{0.1051}\pm0.5994i$ \\\hline\\[-2pt] New system&$\dot x=-x-y$&$(0,0,0)$ &$-0.1,1.1624,-1.4624$ \\[1pt] $r=0.7$ &$\dot y=-x+0.7y-xz$&{}&{}\\[2pt] {} &$\dot z=-0.1z+xy$ &$(\pm\sqrt{17}/10,\mp\sqrt{17}/10,-1.7)$ &$-0.7446, \textbf{0.1723}\pm0.6534i$ \\\hline\\[-2pt] New system&$\dot x=-x-y$&$(0,0,0)$ &$-0.1, 1.2038,-1.4538$ \\[1pt] $r=0.75$ &$\dot y=-x+0.75y-xz$&{}&{}\\[2pt] {} &$\dot z=-0.1z+xy$ &$(\pm\sqrt{70}/20,\mp\sqrt{70}/20,-1.75)$ &$-0.7312, \textbf{0.1906}\pm0.6651i$ \\[-2pt] \hline \end{tabular}} \end{table} Given the above analysis, system (\ref{chenwangeq}) will be discussed only within the parameter region of $0.05<r<1.1$ below. \subsection{Remarks on the classification of the new system} On one hand, it is noted that in system (\ref{chenwangeq}) one has $a_{12}a_{21}=1>0$, so it belongs to the Lorenz system family defined in \cite{CFCV1994}. On the other hand, by making the following transformation: $$ T: (x,y,z)\rightarrow (x,-y,-z), $$ system (\ref{chenwangeq}) becomes \begin{equation}\label{transformed} \left\{ \begin{array}{l} \dot{x}=-x+y \\ \dot{y}=x+ry-xz \\ \dot{z}=-0.1z+xy, \end{array} \right. \end{equation} which satisfies $a_{11}a_{22}<0$ and $a_{11}<0$, thus belong to the Chen system group according to the classification in \cite{yang2008} (see Tables 3 and 4 therein). It is therefore very interesting to see that the new system (\ref{chenwangeq}) belongs to the same class (either Lorenz-type of systems defined in \cite{CFCV1994} or Chen system group defined in \cite{yang2008}), yet can generate both Lorenz-like attractors and Chen-like attractors by gradually changing the single parameter $r$ from $0.05$ to $0.74$, as further detailed next. \section{Chaotic behaviors and other complex dynamics} A complete transition from Lorenz-like to Chen-like attractors in system (\ref{chenwangeq}) is shown in Fig.\ref{color}. The bifurcation diagram with respect to parameter $r$ is shown in Fig.\ref{fig:bifur}. From these figures, one can observe that system (\ref{chenwangeq}) evolves to a chaotic state and eventually to a limit cycle, as the parameter $r$ gradually increases. \begin{figure*} \centering \includegraphics[width=6in,height=8.2in]{lorenz-chen2} \caption{Colored figures showing the projected orbit transition from Lorenz-like to Chen-like attractors in system (\ref{chenwangeq}).} \label{color} \end{figure*} \begin{figure*} \centering \includegraphics[width=3.4in,height=2.8in]{bifur} \caption{Bifurcation diagram of system (\ref{chenwangeq}) with $r\in[-0.2,0.8]$.} \label{fig:bifur} \end{figure*} To verify the chaotic behaviors of system (\ref{chenwangeq}), its three Lyapunov exponents $L_1>L_2>L_3$ and the Lyapunov dimension are calculated, where the latter is defined by \[ D_L=j+\frac{1}{|L_{j+1}|}\sum_{i=1}^j{L_i}, \] in which $j$ is the largest integer satisfying $\sum_{i=1}^j{L_i}\ge0$ and $\sum_{i=1}^{j+1}{L_i}<0$. Note that system (\ref{chenwangeq}) is chaotic if $L_1>0,\,L_2=0,\,L_3<0$ with $|L_1|<|L_3|$. Fig.\ref{fig:newsyslle} shows the dependence of the largest Lyapunov exponent on the parameter $r$. In particular, for several values of $r$, the Lyapunov exponents and dimensions of the system (\ref{chenwangeq}) are summarized in Table 2. From Fig.\ref{fig:newsyslle}, it is clear that the largest Lyapunov exponent increases as the parameter $r$ increases from $0.05$ to $0.7$. \begin{figure*} \centering \includegraphics[width=2.9in,height=2.2in]{LLE} \caption{The largest Lyapunov exponent of system (\ref{chenwangeq}) with $r\in[0.05,0.75]$.} \label{fig:newsyslle} \end{figure*} \begin{table}[h] \centerin \caption{Numerical results for some values of the parameter $r$ with initial values $(1,1,1)$.} {\begin{tabular}{c c c c}\\[-2pt] \hline Parameters & Eigenvalues & Lyapunov Exponents & Fractal Dimensions \\[6pt] \hline\\[-2pt] $r=0.05$ & $\lambda_1=-1.05$& $L_1=0.03467$ &{}\\[1pt] {} & $\lambda_{2,3}=\pm0.4472i$ & $L_2=0$ & $D_L=2.0317$ \\[2pt] {} & {} & $L_3=-1.0843$ & {} \\\hline\\[-2pt] $r=0.1$ & $\lambda_1=-1.0162$& $L_1=0.03058$ &{}\\[1pt] {} & $\lambda_{2,3}=0.0081\pm0.4652i$ & $L_2=0$ & $D_L=2.029$ \\[2pt] {} & {} & $L_3=-1.0301$ & {} \\\hline\\[-2pt] $r=0.5$ & $\lambda_1=-0.8101$ & $L_1=0.10531$ & {} \\[1pt] {} & $\lambda_{2,3}=0.1051\pm0.5994i$ & $L_2=0$ & $D_L=2.1497$ \\[2pt] {} & {} & $L_3=-0.70561$ & {} \\\hline\\[-2pt] $r=0.7$ & $\lambda_1=-0.7446$ & $L_1=0.08953$ & {} \\[1pt] {} & $\lambda_{2,3}=-0.1723\pm0.6534i$ & $L_2=0$ & $D_L=2.1832$ \\[2pt] {} & {} & $L_3=-0.48973$ & {} \\\hline\\[-2pt] $r=0.75$ & $\lambda_1=-0.7312$ & $L_1=0$ & {} \\[1pt] {} & $\lambda_{2,3}=0.1906\pm0.6651i$ & $L_2=-0.04626$ & $D_L=1.0063$ \\[2pt] {} & {} & $L_3=-3.0403$ & {} \\[-2pt] \hline \end{tabular}} \end{table} \subsection*{Case 1: {\rm $r=0.05$}} When $r<0.05$, the corresponding system has one saddle and two stable node-foci (for reference, see \cite{yang2008}). The case of $r<0.05$ is subtle and somewhat complicated, which will be further studied elsewhere in the near future. Here, the discussion starts from $r=0.05$. When $r=0.05$, this new system has one saddle and two centers. The two centers have the same set of eigenvalues: one negative real eigenvalue and two conjugate complex eigenvalues with zero real parts. Nevertheless, the Lyapunov exponents are $L_1=0.03467$, $L_2=0$, and $L_3=-1.0843$, and the Lyapunov dimension is $D_L=2.0317$, for initial values $(1,1,1)$. This convincingly implies that the system is chaotic. \subsection*{Case 2: {\rm$r\in(0.05,0.7]$}} In this case, the new system has one saddle and two saddle-foci. The two saddle-foci have one negative real eigenvalue and two conjugate complex eigenvalues with a positive real part. Moreover, the largest Lyapunov exponent is larger than zero, which is increased as the parameter $r$ is increased from $0.05$ to $0.7$. At the same time, the attractor of this system is changed from Lorenz-like to Chen-like as the parameter $r$ is increased from $0.05$ to around $0.7$, as seen in Fig.\ref{fig:newsys}(a)-5(f). \begin{figure*} \centerin \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_20} {\small (a)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_30} {\small (b)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_40} {\small (c)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_50} {\small (d)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_60} {\small (e)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_70} {\small (f)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_72} {\small (g)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_75} {\small (h)} \end{minipage} \caption{Attractors of system (\ref{chenwangeq}) with parameter values: (a) $r=0.2$; (b) $r=0.3$; (c) $r=0.4$; (d) $r=0.50$; (e) $r=0.6$; (f) $r=0.7$; (g) $r=0.72$; (h) $r=0.75$.} \label{fig:newsys} \end{figure*} \subsection*{Case 3: {\rm$r\in(0.7,0.74)$}} When $r$ lies in this interval, system (\ref{chenwangeq}) is still chaotic, for Fig.\ref{fig:newsyslle} shows that the largest Lyapunov exponent is positive. However, the attractor is neither Lorenz-like nor Chen-like, as shown in Fig.\ref{fig:newsys}(g), when $r=0.72$. Moreover, the chaotic property becomes less significant as compared with Case 2. Indeed, the largest Lyapunov exponent decreases when the parameter is increased from $r=0.7$ to $r=0.74$, as shown in Fig.\ref{fig:newsyslle}. \subsection*{Case 4: {\rm$r\in[0.74,1]$}} In this case, although the new system has one saddle and two saddle-foci, associated with one negative real eigenvalue and two conjugate complex eigenvalues with a positive real part, the system is not chaotic. Instead, the system has a limit cycle, though different from Case 2 and Case 3. Fig.\ref{fig:newsys}(h) displays the limit cycle when $r=0.75$. \section{Comparison between the new family and the unified system} It is interesting to compare the new family of chaotic systems with the now-familiar generalized Lorenz systems family, particularly the corresponding Lorenz-type and Chen-type systems, as shown in Figs.\ref{fig:lorenz-old-new} and \ref{fig:chen-old-new}. Recall in particular the unified chaotic system (\ref{unified}), which is chaotic for all $\alpha\in[0,1]$, with a positive largest Lyapunov exponent as shown in Fig.\ref{fig:unifiedlle}. \begin{figure*} \centering \includegraphics[width=0.45\textwidth]{LLE-unified} \caption{The largest Lyapunov exponent of the unified chaotic system with $\alpha\in[0,1]$.} \label{fig:unifiedlle} \end{figure*} There are several common features between the new family of chaotic systems (\ref{chenwangeq}) and the unified chaotic system (\ref{unified}): \begin{enumerate} \item They are both one-parameter family of chaotic systems, representing infinitely many non-equivalent chaotic systems. \item They both demonstrate a continuously changing process generating from Lorenz-like to Chen-like attractors. \item They both have a similar simple algebraic structure with two same nonlinear terms (i.e., $-xz$ in the second equation and $xy$ in the third equation), juts like the Lorenz and the Chen systems. \item They both have the same $z$-axis rotational symmetry, therefore all their attractors are visually similar in shape with a two-scroll and butterfly-like structure. \item They both have three equilibria: one saddle and two saddle-foci. Moreover, the two saddle-foci have the same eigenvalues: one negative real number and two conjugate complex numbers with a positive real part. Specifically, for the unified chaotic system, the positive real part of the conjugate complex eigenvalues increases from the Lorenz system $(0.0940)$ to the Chen system $(4.2140)$, while for the new system, it increases as $r$ is increased from $0.05$ to $0.7$. \end{enumerate} The main difference between these two one-parameter families of chaotic systems are quite subtle. For system (\ref{chenwangeq}), the real part of its conjugate complex eigenvalues is precisely zero when $r=0.05$. This is a critical point when the stability of the equilibrium pairs is changed from stable to unstable. Therefore, one may define the value of $r=0.05$ as a boundary, or more precisely as a starting point of the Lorenz systems family. Since the parameter $r$ can take values on both sides of $r=0.05$ to generate chaos, the new family system (\ref{chenwangeq}) is richer and more interesting than the unified system system (\ref{unified}). \begin{figure*} \centerin \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_06} {\small (a)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF2D_06_xy} {\small (b)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF2D_06_xz} {\small (c)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF2D_06_yz} {\small (d)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{Lorenz3D} {\small (e)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{Lorenz2D_xy} {\small (f)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{Lorenz2D_xz} {\small (g)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{Lorenz2D_yz} {\small (h)} \end{minipage} \caption{Attractor of system (\ref{chenwangeq}) with $r=0.06$, (a) the 3D phase portrait; (b) projected orbit on $x$-$y$ plane; (c) projected orbit on $x$-$z$ plane; (d) projected orbit on $y$-$z$ plane. The Lorenz attractor corresponding to $\alpha=0$ in the system (\ref{unified}), (e) the 3D phase portrait; (f) projected orbit on $x$-$y$ plane; (g) projected orbit on $x$-$z$ plane; (h) projected orbit on $y$-$z$ plane.} \label{fig:lorenz-old-new} \end{figure*} \begin{figure*} \centerin \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF3D_70} {\small (a)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF2D_70_xy} {\small (b)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF2D_70_xz} {\small (c)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{SLF2D_70_yz} {\small (d)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{Chen3D} {\small (e)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{Chen2D_xy} {\small (f)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{Chen2D_xz} {\small (g)} \end{minipage} \begin{minipage}[b]{0.4\textwidth} \centering \includegraphics[width=5cm]{Chen2D_yz} {\small (h)} \end{minipage} \caption{Attractor of system (\ref{chenwangeq}) with $r=0.7$, (a) the 3D phase portrait; (b) projected orbit on $x$-$y$ plane; (c) projected orbit on $x$-$z$ plane; (d) projected orbit on $y$-$z$ plane. The Chen attractor corresponding to $\alpha=1$ in the system (\ref{unified}), (e) the 3D phase portrait; (f) projected orbit on $x$-$y$ plane; (g) projected orbit on $x$-$z$ plane; (h) projected orbit on $y$-$z$ plane.} \label{fig:chen-old-new} \end{figure*} \section{Concluding remarks and future research} \subsection{Concluding remarks} Regarding the few questions raised in the Introduction section, we now come out with some answers: \begin{enumerate} \item In this paper, we have reported the finding of a structurally simple yet dynamically complex one-parameter family of 3D autonomous systems with only two quadratic terms like the Lorenz system, with some key constant coefficients set as $1$ or $-1$, leaving only one tunable real parameter in the linear part of the system. \item By tuning the singe parameter, this new family of systems can generate a chain of chaotic attractors in a very similar gradual changing process like the one from Lorenz attractor to Chen attractor generated by the generalized Lorenz systems family and the one by the unified chaotic system. \item Unlike the typical Chen system, however, the new system that generates the Chen-like attractor does not satisfy the condition $a_{12}a_{21} < 0$ in the linear part of the generalized Lorenz canonical form defined in \cite{CFCV1994}. \item Unlike the classical Lorenz system, the new system that can generates a similar Lorenz-like attractor does not satisfy the condition $a_{11}a_{22}>0$ in the linear part of the generalzied Lorenz canonical form defined in \cite{yang2008}. \item This system is richer and more interesting than the unified chaotic system, in the sense that the new single parameter has a wider range to generate similar form of chaos. \item Similar to the generalized Lorenz systems family, the new system also has three equilibria, with one saddle and two saddle-foci, where the two saddle-foci have the same eigenvalues (one is negative real and two are complex conjugate with a positive real part). However, the in the new system the real part of the eigenvalues starts from zero, which can be used to define a starting point for the Lorenz-like systems family. In this sense, the new system has literally extended the unified chaotic system, even the generalized Lorenz systems family, to some extent. \end{enumerate} \subsection{Future research} Some related research issues may be further pursued in the near future: \begin{enumerate} \item There may exist other Lorenz-type systems with the same nonlinear terms as the classical Lorenz system, which can generate a variety of chaotic attractors. \item There may even exist other types of nonlinear terms that satisfy the expected $z$-axis rotational symmetry, for which the nonlinear term in the second equation must be like $xz$ or $yz$, and that in the third equation must be like $xy$, or $z^{2}$, or $x^{2}$, or $y^{2}$. So, if one wants to study 3D autonomous systems with two quadratic terms that can maintain the $z$-axis rotational symmetry, then there are totally $2\times4=8$ possible combinations to consider. It is therefore interesting to ask if there would be other chaotic systems with two nonlinear terms different from those studied so far around the Lorenz system. \item In the new family of chaotic systems studied in this paper, the real part of the complex conjugate eigenvalues at system equilibria starts from zero for the case with $r=0.05$, which can be defined as the starting point of the Lorenz-like systems family. This particularly interesting critical system with $r=0.05$ alone deserves further investigations. \item This paper has further revealed that some systems with different structures can literally generate similar-shaped chaotic attractors. Therefore, for 3D autonomous systems with two quadratic terms, the relation between the system algebraic structure and system chaotic dynamics is an important and interesting issue to be further revealed, understood and analyzed. \end{enumerate} \section*{Acknowledgements} The authors sincerely thank Professors Sergej \v{C}elikovsk\'{y}, Qigui Yang and Tianshou Zhou for their valuable comments and suggestions. This work was supported by the France/HongKong Joint Research Scheme under grant CityU 9052002/2011-12.

https://arxiv.org/abs/1902.02632

An Analysis of a Mathematical Model Describing Acid-mediated Tumor Invasion

We present a mathematical analysis of a reaction-diffusion model describing acid-mediated tumor invasion. The model describes the spatial distribution and temporal evolution of tumor cells, normal cells, and excess lactic acid concentration. The model assumes that tumor-induced alteration of microenvironmental pH provides a simple but complete mechanism for cancer invasion. We provide results on the existence and uniqueness of a solution considering Neumann and Dirichlet boundary conditions. We also provide numerical simulations to the solutions considering both boundary conditions.

\section*{This is an unnumbered first-level section head} \par \section{Introduction} The main objective of this work is to perform a rigorous mathematical analysis of a system of nonlinear partial differential equations corresponding to a generalization of a mathematical model describing the growth of a tumor proposed in \cite{Fassoni}. To describe the model, let $\Omega \subset \mathrm{I\!R\!}^{2}$, be an open and bounded set; let also $0< T< \infty$ be a given final time of interest and denote $t$ the times between $[0,T]$ and $Q=\Omega\times(0,T)$, the space-time cylinder and $\bar{\Gamma}=\partial \Omega\times(0,T)$, the space-time boundary. Then, the system of equations we are considering is the following: \begin{equation} \label{0riginalEquations} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial N}{\partial t} = r_N - \mu_N N - \beta_1 N A - \alpha_H\gamma_H N H, & \textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A + r_A A\left(1-\frac{A}{k_A}\right)-(\mu_A+\epsilon_A)A - \beta_3 NA , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial H}{\partial t} = \xi_H \Delta H + \nu A - \tau_H H - \gamma_H NH , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot) =\frac{\partial H}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle N(\cdot,0) = N_0(\cdot), A(\cdot,0) = A_{0}(\cdot), H(\cdot,0) = H_0(\cdot), &\textup{in}& \Omega. \end{array} \right. \end{equation} In \cite{Fassoni}, Fassoni studied a simplified model given by a system of ODE's describing the growth of a tumor and its effect in the normal tissue, together with the tissue response to the tumor and the effect of chemotherapeutic treatments. The aim of the authors was to provide some $insights$ on the description of cancer onset and treatment as transitions between alternative basins of attractions. The model studied in \cite{Fassoni} is given by the following ODE system: \begin{equation} \label{início1} \left\{ \begin{array}{lcl} \displaystyle \frac{d N}{d t} = r_N - \mu_N N - \beta_1 N A - \alpha_N\gamma_N D N, \vspace{0.2cm} \\ \displaystyle \frac{d A}{d t} = r_A A\left(1-\frac{A}{k_A}\right)-(\mu_A+\epsilon_A)A - \beta_3 NA - \alpha_A\gamma_A D A, \vspace{0.2cm} \\ \displaystyle \frac{d D}{d t} = \mu - \gamma_A D A - \gamma_N D N - \tau D, \end{array} \right. \end{equation} where $N$ represents the normal cells in a given tissue of the human body, $A$ represents the tumor cells in this tissue and $D$ represents the concentration of a drug used to treat such a tumor. In the model, parameter $r_N$ represents the total constant reproduction of normal cells and $\mu_N$ their natural mortality. A constant flow for normal cells is considered in the vital dynamics and does not depend on the density, as is the case of logistical growth generally assumed, see \cite{Jb}. The reason is that in a normal and already formed tissue, the imperative dynamics is not the intraspecific competition of cells for nutrients, but the maintenance of a homeostatic state, through the natural replenishment of old cells, see \cite{Simons}. In contrast, cancer cells have independence on growth signals and maintain their own growth program, as for example, a growing embryonic tissue, see \cite{Fedi}. Thus, a density-dependent growth is considered for tumor cells. Several growth laws could be used, such as Gompertz, generalized logistic, Von Bertanlanfy and others, see \cite{Sarapata}, but the logistic growth is chosen due to its simplicity. A natural mortality rate $\mu_A$ and an extra mortality rate $\epsilon_A$, due to apoptosis, see \cite{Danial}, are also included. Parameters $\beta_1$ and $\beta_3$ encompass in a simplified way the many negative interactions between tumor cells and normal. Parameter $\beta_3$ encompasses the effects imposed on cancer cells by normal cells, such as competition for nutrients and the release of anti-growth and death signals. In the same way, parameter $\beta_1$ encompasses mechanisms developed by tumor cells that damage the normal tissue, such as competition, release of death signals and increased local acidity. The hypotheses that lead to the third equation of (\ref{início1}), describing the pharmacokinetics of the chemotherapeutic drug are the following. The drug has a clearance rate $\tau$. The rates of absorption and deactivation of the drug by normal and cancer cells are described in terms of the law of mass action with rates $\gamma_N$ and $\gamma_A$. Following the linear hypothesis of \cite{Andre}, the amounts of drug absorbed by normal ($\gamma_N N D$) and cancer cells ($\gamma_A A D$) kill such cells at rates $\alpha_N$ and $\alpha_A$, respectively. Finally, parameter $\mu$ represents a constant infusion rate mimicking a metronomic dosage, i.e., a near continuous and long-term administration of the drug. System (\ref{início1}) is similar to the classic Lotka-Volterra competition model, commonly used in models for tumor growth and biological invasions, but there is a fundamental difference: the use of a constant flux instead of a logistic growth for normal cells breaks the symmetry observed in the classic Lotka-Volterra model, so that there will be no equilibrium for $N = 0$. Thus, normal cells will never be extinct, unlike these models. In \cite{Fassoni} the authors believe that this is not a problem, but on the contrary, it is a realistic result. In fact, roughly speaking, cancer does "don not win" by the fact that it kills all the cells in the tissue, but because it reaches a dangerous size that disrupts the proper functioning of the tissue and threatens the health of the individual. A constant flow term has already been taken in other well-known models of cancer, specifically, to describe the growth of immune cells, see \cite{Earn}. In this work, we are not interested in analyzing the dynamics of model (\ref{início1}), since in \cite{Fassoni} a study has already been made on the equilibrium points and questions about stability or instability of such points. Our objective is to study the existence and uniqueness of solutions to the modified system (\ref{0riginalEquations}). System (\ref{0riginalEquations}) considers the dynamics of normal and tumor cells under the same hypothesis as the system (\ref{início1}), and explicitly considers the production of lactic acid by cancer cells. Such acid production is well-known as a mediator of tissue invasion by tumor cells, and is a by-product of the altered metabolism of tumor cells, which exert glucose metabolism not by oxidative phosphorylation (as normal cells), but mainly by glycolysis. This altered metabolism is known as Warburg effect, see \cite{Jb}. The increased acid concentration in the tissue damages the normal cells, thereby destroying the ``barriers'' for tumor growth and contributing to tumor persistence and further tissue invasion. Model (\ref{0riginalEquations}) describes such phenomena by considering the original equations for the normal cells ($N(t)$) and cancer cells ($A(x,t)$), and an additional equation for the concentration of extracellular lactic acid in excess of normal tissue acid concentrations ($H(x,t)$), produced at a rate $\nu$ by tumor cells, cleared by tissue vasculature at a rate $\tau_H$ and absorbed by normal cells at a rate $\gamma_H$, following the mass-action law. The absorbed amount $\gamma_H NH$ causes a damage to normal cells with rate $\alpha_H$. The aspect of acid-mediated tumor invasion clearly needs to be considered in a spatial setting, thereby it is better represented by a PDE model instead of an ODE system, which describes a spatially homogeneous tumor. Therefore, we assume a diffusive behavior for both the tumor cells and the extracellular lactic acid excess, with diffusion coefficients $\xi_A$ and $\xi_H$ respectively, while normal cells do not move. This is in agreement with previous models, see \cite{Jb}, and regards the additional capability of tumor cells to move within a ``rigid'' tissue. Previous works approached similar models, but all considering a density dependent growth for normal cells (logistic growth), see (\cite{Jb}) and references therein. In this work, we are not interested in the treatment phase, but only on the invasion phase and the establishment of a tumor. In a future work, we may extend model (\ref{0riginalEquations}) by including a differential equation for the chemotherapeutic drug as given by system (\ref{início1}). This paper is organized as follows. In Section 2 we present the technical hypothesis state our main result. In Section 3 we study an auxiliary problem. Using its solution, we prove our main result in Section 4. In Section 5 we present numerical simulations illustrating model behavior. \section{Technical hypotheses and main result} Recall $\Omega \subset \mathrm{I\!R\!}^2$ be a domain with boundary $\partial\Omega$, $0\leq T <\infty$, and denote $Q = \Omega\times (0, T)$ and $\Gamma = \partial \Omega\times (0, T)$. We will use standard notations for Sobolev spaces, i.e., given $1~\leq~p~\leq~+\infty $ and $k \in\mathbb{N}$, we denote $$W_{p}^{k}(\Omega)=\left\{ f \\ \in L^{p}(\Omega) : D^{\alpha}f \in L^{p}(\Omega), |\alpha| \leq k \right\}; $$ when $p=2$, as usual we denote $W_{2}^{k}(\Omega) = H^k (\Omega)$; properties of these spaces can be found for instance in~Adams~\cite{Adams}. Problem~(\ref{0riginalEquations}) will be studied in the standard functional spaces denoted by \begin{eqnarray*} W_{q}^{2,1}(Q) & =& \left\{f\in L^{q}(Q):D^{\alpha}f\in L^{q}(Q), \, \forall 1\leq|\alpha|\leq 2, f_t \in L^{q}(Q)\right\}, \end{eqnarray*} \begin{eqnarray*} W &=& \left\{f\in L^\infty(Q): f_t \in L^\infty(Q)\right\} \end{eqnarray*} and \begin{eqnarray*} L^{p}(0,T;B) &=& \left\{f:(0,T)\rightarrow B: \|f(t)\|_{L^{p}(0,T;B)} <+\infty \right\}, \end{eqnarray*} where $B$ is suitable Banach space, and the norm is given by $\|f(t)\|_{L^{p}(0,T;B)} = \|\ \|f(t)\|_{B}\ \|_{L^{p}((0,T))}$. We remark that $L^{p}(Q) = L^{p}((0,T);L^{p}(\Omega))$. Results concerning these spaces can be found for instance in Ladyzhenskaya~\cite{Ladyzhenskaya} and Mikhaylov~\cite{Mikhaylov}. \vspace{0.1cm} Next, we state some hypotheses that will be assumed throughout this article. \subsection{Technical Hypotheses:} \label{MainHypotheses} \begin{itemize} \item[{\bf (i)}] $\Omega\subset\mathbb{R}^2$ is a bounded $C^2$-domain. \item[{\bf (ii)}] $0< T < \infty$, and $Q=\Omega\times(0,T)$. \item[{\bf (iii)}] $N_0 \in L^{\infty}(\Omega)$ and $A_0, H_0 \in W^{\frac{3}{2}}_{4}(\Omega)$, satisfying $\frac{\partial A_0}{\partial \eta} (\cdot) =\frac{\partial H_0}{\partial \eta} (\cdot) =0, \textup{ on } \partial\Omega$. \item[{\bf (iv)}] $0 \leq A_0 \leq k_A$ and $N_0, H_0 \ge 0$ a.e. on $\Omega$ \end{itemize} \begin{remark} The constraints imposed in~{\bf (iv)} on the initial conditions are natural biological requirements. \end{remark} \subsection{Main result:} \begin{theorem} \label{Teorema1} Assume that the Technical Hypotheses \ref{MainHypotheses} hold; then, there exists a unique nonnegative solution $(N,A,H) \in W \times W^{2,1}_4(Q) \times W^{2,1}_4(Q)$ of Problem (\ref{0riginalEquations}). Moreover, $N, A$ and $H$ are functions satisfying \begin{eqnarray*} N \leq ||N_0||_{L^\infty(Q)} + r_N T, \ A \leq k_A \ a.e. \ in \ Q \end{eqnarray*} and \begin{eqnarray*} ||N||_W + ||A||_{ W^{2,1}_4(Q)} + ||H||_{ W^{2,1}_4(Q)} \leq C, \end{eqnarray*} where $C$ is a constant depending on $r_N$, $\mu_N$, $\beta_1$, $\alpha_H$, $\gamma_H$, $k_A$, $\nu$, $T$, $\Omega$, $||N_0||_{L^\infty (\Omega)}$, $||A_0||_{W^{\frac{3}{2}}_4(\Omega)}$, $||H_0||_{W^{\frac{3}{2}}_4(\Omega)}$ and $||H||_{L^\infty(Q)}$. \end{theorem} \begin{remark} The explicit knowledge on how the constant $C$ appearing in the above estimates depends on the given data is important to application to related control problems. \end{remark} \subsection{Known technical results:} To ease the references, we also state some technical results to be used in this paper. The first one is consequence of Theorem~5.4, p. 97, in Adams~\cite{Adams}: \begin{lemma} \label{imersao_Sobolev} Suppose that $\Omega\subset\mathbb{R}^{n}$ satisfies the cone property and $1 \leq p <\infty$. The following continuous embeddings hold: \[ \begin{array}{l} \mbox{\bf (i)} \; \, W^{2}_{p}(\Omega)\hookrightarrow W^{2-\frac{2}{q}}_{q}(\Omega)$, for all $p\leq q\leq 5p/3$, if $ n\leq 3; \\ \mbox{\bf (ii)} $ If $ kp<n$, then $ \, W^{k}_{p}(\Omega)\hookrightarrow L^{q}(\Omega)$, for all $p \leq q \leq np/(n-kp); \\ \mbox{\bf (iii)} $ If $ kp=n$, then $ \, W^{k}_{p}(\Omega)\hookrightarrow L^{q}(\Omega)$, for all $p \leq q <\infty; \\ \mbox{\bf (iv)} $ If $ kp>n$, then $ \, W^{k}_{p}(\Omega)\hookrightarrow L^{\infty}(\Omega). \end{array} \] \end{lemma} The next result sometimes is called the Lions-Peetre embedding theorem (see Lions~\cite{Lions}, pp.15); it is also a particular case of Lemma~3.3, pp.80, in Ladyzhenskaya~\cite{Ladyzhenskaya}: (obtained by taking $l = 1$ and $r = s = 0$). \begin{lemma} \label{icontLp01} Let $\Omega$ be a domain of $\mathrm{I\!R\!}^n$ with boundary $\partial \Omega$ satisfying the cone property. Then, the functional space $W^{2,1}_p(Q)$ is continuously embedded in $u \in L^{q}(Q)$ for $q$ satisfying: {\bf (i)} $1 \leq q \leq \frac{p(n+2)}{n+2-2p}$, if $ p< \frac{n+2}{2}$; {\bf (ii)} $1 \leq q <\infty$, if $p= \frac{n+2}{2}$ and {\bf (iii)} $q=\infty$, if $p>\frac{n+2}{2}$. \noindent In particular, for such $q$ and any function $u \in W^{2,1}_p(Q)$ we have that \begin{eqnarray*} \label{i.01} \displaystyle \|u\|_{L^{q}(Q)} \leq C\|u\|_{W^{2,1}_p(Q)}, \end{eqnarray*} \noindent with a constant $C$ depending only on $\Omega$, $T$, $p$, $q$, $n$. In the cases {\bf (ii)}, {\bf (iii)} or in {\bf (i)} when $\displaystyle 1 \leq q < \frac{p(n+2)}{n+2-2p}$, the referred embedding is compact. \end{lemma} \vspace{0.1cm} Next, we consider the following general and simple parabolic initial-boundary value problem: \begin{equation} \label{P_Newmman} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial u}{\partial t} - \sum\limits_{i,j=1}^na_{ij}(x,t)\frac{\partial u^2}{\partial x_ix_j} + \sum\limits_{j=1}^na_i(x,t)\frac{\partial u}{\partial x_j} + a(x,t)u=f & \textup{in} & Q, \vspace{0.2cm} \\ \displaystyle \sum\limits_{i=1}^n b_i(x,t)\frac{\partial u}{\partial x_i} + b(x,t)u =0 & \textup{on} & \Gamma, \vspace{0.2cm} \\ \displaystyle u(\cdot,0)= u_0(\cdot) & \textup{in} & \Omega . \end{array} \right. \end{equation} Existence and uniqueness of solutions for this problem is a particular case of Theorem~$9.1$, pp.$341$, in Ladyzenskaya~\cite{Ladyzhenskaya} for the case of Neumann boundary condition, according to the remarks at the end Chapter IV, section 9, p. 351 in \cite{Ladyzhenskaya}. In the following, we state this particular result, stressing the dependencies certain norms of the coefficients, that will be important in our future arguments. \begin{proposition} \label{sol. Neumann} Let $\Omega$ be a bounded domain in $\mathbb{R}^n$, with a $C^{2}$ boundary $\partial \Omega$, $a_{ij}$ be bounded continuous functions in $Q$, and $q > 1$. Assume that \begin{enumerate} \item $a_{ij} \in C(\bar{Q})$, $i, j=1, \ldots, n$; $[a_{ij}]_{n \times n}$ is a real positive matrix such that for some positive constant $\beta$ we have $ \sum\limits_{i,j=1}^n a_{ij}(x,t)\xi_i\xi_j\geq \beta|\xi|^2$ for all $(x,t) \in Q$ and all $\xi \in R^n$, ; \item $\displaystyle f \in L^p(Q)$; \item $\displaystyle a_i \in L^r(Q)$ with either $r = \max\big(p, n + 2\big)$ if $p \neq n + 2$ or $r = n + 2 + \varepsilon$, for any $\varepsilon>0$, if $p = n + 2$; \item $\displaystyle a \in L^s(Q)$ with either $s = \max\big(p, (n + 2)/2\big)$ if $\displaystyle p \neq (n + 2)/2$ or $s = (n + 2)/2 + \varepsilon$, for any $\varepsilon>0$, if $\displaystyle p = (n + 2)/2$. \item $b_i, b \in C^2 (\bar{\Gamma})$, $i=1, \ldots, n$, and the coefficients $b_i(x,t)$ satisfy the condition $\left| \sum\limits_{i=1}^n b_i(x,t)\eta_i(x) \right|\geq \delta >0$ for $a.e.$ in $\partial\Omega \times (0,T)$, where $\eta_i(x)$ is the $i^{th}$-component of the unitary outer normal vector to $\partial\Omega$ in $x \in \partial\Omega$; \item $u_0 \in \ W^{2 - \frac{2}{p}}_p(\Omega)$ with $p\neq 3$ and satisfying the compatibility condition \\ $\displaystyle \sum\limits_{i=1}^n b_i \frac{\partial u_0}{\partial x_i} + b \ u_0 =0$ on $\partial \Omega$ when $p > 3$. \end{enumerate} Then, there exists a unique solution $u \in W^{2,1}_p(Q)$ of Problem~(\ref{P_Newmman}); moreover, there is a positive constant $C_p$ such that the solution satisfies \begin{equation} \label{BasicParabolicEstimate} \|u\|_{W^{2,1}_p(Q)} \leq C_{p} \left(\|f\|_{L^p(Q)} + \|u_0\|_{W^{2 - \frac{2}{p}}_p(\Omega)}\right). \end{equation} Such constant $C_{p}$ depends only on $\Omega$, $T$, $p$, $r$, $s$, $\beta$, $\delta$ and on the norms $\|b_i\|_{C^2 (\bar{\Gamma})}$, $\|b\|_{C^2 (\bar{\Gamma})}$, $\|a_{ij}\|_{C(\bar{Q})}$, $\|a_i\|_{L^r(Q)}$ and $\|a\|_{L^s(Q)}$. Moreover, we may assume that the dependencies of $C_{p}$ on stated the norms are non decreasing. \end{proposition} \begin{remark} The result set out in Proposition \ref{sol. Neumann} can be formulated for the parabolic problem with Dirichlet conditions (see Ladyzenskaya \cite[Theorem 9.1, pp.$341$]{Ladyzhenskaya}). In the problem with Dirichlet condition the compatibility condition in Proposition \ref{sol. Neumann}-($6$) can be replaced by $u_0=0$ on $\partial \Omega$ when $p > 3/2$. This way, all the results in this paper holds if we replaced the Neumann conditions by Dirichlet conditions. \end{remark} \section{An auxiliary problem} In this section we will prove an auxiliary result to be used in the proof of Theorem~\ref{Teorema1}. To cope with difficulties with the signs of certain terms during the derivation of the estimates, we firstly have to consider the following modified problem: \begin{equation} \label{P01} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \hat{N}}{\partial t} = r_N - \mu_N \hat{N} - \beta_1 \hat{N} |\hat{A}| - \alpha_H\gamma_H \hat{N} |\hat{H}|, &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \hat{A}}{\partial t} = \xi_A \Delta \hat{A} + r_A \hat{A}\left(1-\frac{\hat{A}}{k_A}\right)-(\mu_A+\epsilon_A)\hat{A} - \beta_3 \hat{N} \hat{A} , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \hat{H}}{\partial t} = \xi_H \Delta \hat{H} + \nu \hat{A} - \tau_H \hat{H} - \gamma_H \hat{N} \hat{H} , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \hat{A}}{\partial \eta} (\cdot) =\frac{\partial \hat{H}}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \hat{N}(\cdot,0) = N_0(\cdot), \hat{A}(\cdot,0) = A_{0}(\cdot), \hat{H}(\cdot,0) = H_0(\cdot), &\textup{in}& \Omega. \end{array} \right. \end{equation} Now we observe that, since the equation for $\hat{N}$ in this last problem is, for each $x \in \Omega$, an ordinary differential equation which is linear in $\hat{N}$, we can find an explicit expression for it in terms of $|\hat{A}|$ and $|\hat{H}|$. Using this observation, we introduce the operator $\Theta: L^{\infty}(Q) \times L^{\infty}(Q) \to L^{\infty}(Q)$, defined by \begin{equation} \label{P7} \begin{array}{c} \displaystyle \Theta(\phi, \varphi)(x,t) = \frac{N_0(x) + r_N \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi(x, \xi)| d\xi} ds}{e^{\mu_N t} e^{\alpha_H\gamma_H\int_{0}^{t}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{t}|\varphi(x,\xi)| d\xi}}, \end{array} \end{equation} \noindent where $0 \leq s \leq t \leq T$. \begin{remark} \label{obs1} Thus, $(\hat{N},\hat{A}, \hat{H})$ is solution of (\ref{P01}) if, and only if, $\hat{N} = \Theta (\hat{H}, \hat{A})$, $\hat{A}$ and $\hat{H}$ satisfies the following integro-differential system: \begin{equation} \label{P3} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \hat{A}}{\partial t} = \xi_A \Delta \hat{A} + r_A \hat{A}\left(1-\frac{\hat{A}}{k_A}\right)-(\mu_A+\epsilon_A)\hat{A} - \beta_3 \Theta(\hat{A}, \hat{H}) \hat{A} , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \hat{H}}{\partial t} = \xi_H \Delta \hat{H} + \nu \hat{A} - \tau_H \hat{H} - \gamma_H \Theta(\hat{A}, \hat{H}) \hat{H} , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \hat{A}}{\partial \eta} (\cdot)= \frac{\partial \hat{H}}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \hat{A}(\cdot,0) = A_{0}(\cdot), \hat{H}(\cdot,0) = H_0(\cdot), &\textup{in}& \Omega. \end{array} \right. \end{equation} \end{remark} \begin{remark} \label{obs2} Notice that, to guarantee that $(N, A, H)$, with $A= \hat{A}$, $H = \hat{H}$ and $N = \Theta(\hat{H}, \hat{A})$ is also a solution of system~(\ref{0riginalEquations}), it is enough to prove that the solution $(\hat{A}, \hat{H})$ of Problem~(\ref{P3}) is nonnegative. \end{remark} For the Problem \ref{P3}, we have the following existence result: \begin{proposition}\label{Prop1} Assuming that the Technical Hypotheses~\ref{MainHypotheses} hold, there exists at least one nonnegative solution $(\hat{A},\hat{H}) \in W^{2,1}_4(Q)\times W^{2,1}_4(Q)$ of Problem \eqref{P3}. Moreover, $\hat{A} \leq k_A$ a.e. in $Q$ and \begin{eqnarray*} ||\hat{A}||_{W^{2,1}_4(Q)} + ||\hat{H}||_{W^{2,1}_4(Q)} \leq C, \end{eqnarray*} where $C$ is a constant depending on $k_A$, $\nu$, $T$, $\Omega$, $||A_0||_{W^{\frac{3}{2}}_4(\Omega)}$ and $||H_0||_{W^{\frac{3}{2}}_4(\Omega)}$. \end{proposition} \begin{lemma} \label{base} Let $f:(0,T) \to \mathbb{R}$ differentiable such that $f(t) > 0$ and $f'(t) \ge 0$. If $g(t) = \frac{\int_{0}^{t} f(x) dx}{f(t)}$, then $g(t) \leq T$, for all $t \in (0, T)$. \end{lemma} \noindent {\bf Proof:} Since $f$ is continuous in $(0, T)$, it follows that \begin{eqnarray*} g'(t)=\frac{f(t)^2 - f'(t) \int_{0}^{t}f(x)dx}{f(t)^2} = 1 - \frac{f'(t)}{f(t)} g(t). \end{eqnarray*} As $f(t) > 0$ we have $g(t) \ge 0$ and using the fact that $f'(t) \ge 0$ we obtain $\frac{f'(t)}{f(t)} g(t) \ge 0$. Therefore, $g'(t) \leq 1$, which suggests $g(t) \leq t$, for all $t \in (0, T)$. Thus, $g(t) \leq T$, as intended. \hfill$\Box$ Since in the proof of existence of solutions of (\ref{P3}) the expression of $\Theta$ will play important roles, we state in the following some of their properties: \begin{lemma} \label{PropertiesEtcFirst} If $N_0 \in L^\infty(\Omega)$, then for any $\phi, \phi_1, \phi_2, \varphi, \varphi_1, \varphi_2 \in L^\infty(Q)$ and for almost everything $(x,t) \in Q$, there holds \[ \begin{array}{ll} \mbox{\bf (i)} & 0 \leq \Theta(\phi, \varphi)(x,t) \leq ||N_0||_{L^\infty(\Omega)} + r_N T; \vspace{0.2cm} \\ \mbox{\bf (ii)} & \|\Theta (\phi_1, \varphi_1) - \Theta (\phi_2, \varphi_2) \|_{L^\infty {(Q)}} \leq C_1 \|(\phi_1, \varphi_1) - (\phi_2, \varphi_2)\|_{L^\infty(Q)}, \\ & where \ C_1 \ is \ a \ constant \ depending \ on \ r_N, \mu_N, \beta_1, \alpha_H, \gamma_H, T, ||\phi_1||_{L^\infty(Q)}, \\ & ||\phi_2||_{L^\infty(Q)}, ||\varphi_1||_{L^\infty(Q)}, ||\varphi_2||_{L^\infty(Q)} \ and \ ||N_0||_{L^\infty(\Omega)}. \end{array} \] \end{lemma} \noindent {\bf Proof (i):} By expression (\ref{P7}) it is immediate that $\Theta(\phi, \varphi)(x,t) \ge 0$. To prove that $\Theta(\phi)(x,t) \leq ||N_0||_{L^\infty(\Omega)} + r_N T,$ we observe that $$ \begin{array}{rcl} \displaystyle \Theta(\phi, \varphi)(x,t) = \frac{N_0(x) + r_N \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi(x, \xi)| d\xi} ds}{e^{\mu_N t} e^{\alpha_H\gamma_H\int_{0}^{t}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{t}|\varphi(x,\xi)| d\xi}} \\ \\ \displaystyle \leq N_0(x) + r_N \frac{ \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi(x, \xi)| d\xi} ds}{e^{\mu_N t} e^{\alpha_H\gamma_H\int_{0}^{t}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{t}|\varphi(x,\xi)| d\xi}} . \end{array} $$ Fixed $x \in \Omega$, we define \begin{equation*} \displaystyle g(x,t) = \frac{\int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi(x, \xi)| d\xi} ds}{e^{\mu_N t} e^{\alpha_H\gamma_H\int_{0}^{t}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{t}|\varphi(x,\xi)| d\xi}}, \end{equation*} and using the Lemma \ref{base} with $f(x,t) = e^{\mu_N t} e^{\alpha_H\gamma_H\int_{0}^{t}|\phi(x,\xi)| d\xi}e^{\beta_1\int_{0}^{t}|\varphi(x,\xi)| d\xi} $, it follow that \begin{eqnarray*} \Theta(\phi, \varphi)(x,t) \leq N_0(x) + r_N T \\ \leq ||N_0||_{L^\infty(\Omega)} + r_N T. \end{eqnarray*} {\bf Proof (ii):} We firstly need to observe that, due to the mean value inequality, given any $z_1, z_2 \in \mathrm{I\!R\!}$, there is $\theta = \theta(z_1, z_2)$ such that $e^{z_2} - e^{z_1} = e^{(1-\theta) z_1 + \theta z_2 } (z_2 - z_1)$; in particular, for any $z_1, z_2 \leq 0$ we also have $(1-\theta) z_1 + \theta z_2 \leq 0$ and thus \begin{equation} \label{AlgebraicExponentialInequality} |e^{z_2} - e^{z_1} | \leq |z_2 - z_1|, \quad \forall z_1, z_2 \leq 0 . \end{equation} Secondly, we note that \begin{eqnarray*} \displaystyle \big|e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} - e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi}\big| &\leq& \\ \displaystyle e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}\big|e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} - e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi}\big| &+& \\ \displaystyle e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi} \big|e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi} - e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}\big|. \end{eqnarray*} By the inequality (\ref{AlgebraicExponentialInequality}) and by $\phi_i, \varphi_i \in L^\infty(Q)$, $i = 1,2$, we obtain \begin{equation}\label{i1} \begin{array}{rcl} \displaystyle \big|e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} - e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi}\big| &\leq& \\ \\ \displaystyle e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}\beta_1 \int_{0}^{t}|\varphi_2(x,\xi) - \varphi_1(x,\xi)|d\xi &+& \\ \\ \displaystyle e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi} \alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi) - \phi_1(x,\xi)|d\xi &\leq& \\ \\ \displaystyle e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}\beta_1 T ||\varphi_2 - \varphi_1||_{L^\infty(Q)} &+& \\ \\ \displaystyle e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi} \alpha_h\gamma_h T ||\phi_2 - \phi_1||_{L^\infty(Q)}. \end{array} \end{equation} Thirdly, we observe that \begin{eqnarray*} \bigg|e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_1(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_1(x, \xi)| d\xi} ds &-& \\ \displaystyle e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi} \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_2(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_2(x, \xi)| d\xi} ds \bigg| &\leq& \\ \displaystyle e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} &\times& \\ \displaystyle \int_{0}^{t} e^{\mu_N s}\big|e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_1(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_1(x, \xi)| d\xi} - e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_2(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_2(x, \xi)| d\xi}\big|ds &+& \\ \displaystyle \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_2(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_2(x, \xi)| d\xi} ds &\times& \\ \displaystyle \big|e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} - e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi}\big|. \end{eqnarray*} Study analogous to that done in (\ref{i1}), prove that \begin{eqnarray*} \bigg|e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_1(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_1(x, \xi)| d\xi} ds &-& \\ \displaystyle e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi} \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_2(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_2(x, \xi)| d\xi} ds \bigg| &\leq& \\ \displaystyle e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} &\times& \\ \displaystyle e^{\mu_N T} T \bigg(e^{\alpha_h\gamma_h T ||\phi_2||_{L^\infty(Q)}} \beta_1 T ||\varphi_2 - \varphi_1||_{L^\infty(Q)} + e^{\beta_1 T ||\varphi_1||_{L^\infty(Q)}} \alpha_h\gamma_h T ||\phi_2 - \phi_1||_{L^\infty(Q)} \bigg) &+& \\ \displaystyle T e^{\mu_N T} e^{\alpha_H\gamma_H T ||\phi_2||_{L^\infty(Q)}}e^{\beta_1 T ||\varphi_2||_{L^\infty(Q)}} &\times& \\ \displaystyle \bigg(e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}\beta_1 T ||\varphi_2 - \varphi_1||_{L^\infty(Q)} + e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi} \alpha_h\gamma_h T ||\phi_2 - \phi_1||_{L^\infty(Q)} \bigg). \end{eqnarray*} Finally, the expression in (\ref{P7}) suggests \begin{equation} \label{ineq1} \begin{array}{rcl} \displaystyle |\Theta(\phi_1, \varphi_1)(x,t) - \Theta(\phi_2, \varphi_2)(x,t)| & \leq & \\ \\ \displaystyle \frac{1}{e^{\mu_N t}e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi}e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi}} &\times& \\ \\ \displaystyle \bigg(N_0(x)e^{\mu_n t}\bigg|e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} - e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi}\bigg| &+& \\ \\ \displaystyle r_N e^{\mu_N t} \bigg|e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_2(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_2(x,\xi)|d\xi} \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_1(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_1(x, \xi)| d\xi} ds &-& \\ \displaystyle e^{\alpha_h\gamma_h \int_{0}^{t}|\phi_1(x,\xi)|d\xi}e^{\beta_1 \int_{0}^{t}|\varphi_1(x,\xi)|d\xi} \int_{0}^{t} e^{\mu_N s} e^{\alpha_H\gamma_H\int_{0}^{s}|\phi_2(x,\xi)| d\xi}e^{\beta_1\int_{0}^{s}|\varphi_2(x, \xi)| d\xi} ds \bigg|\bigg) \end{array} \end{equation} and using the estimates obtained in (\ref{i1}) and (\ref{ineq1}) and making the possible simplifications, we obtain $$ \begin{array}{rcl} \displaystyle |\Theta(\phi_1, \varphi_1)(x,t) - \Theta(\phi_2, \varphi_2)(x,t)| & \leq & \\ \\ \displaystyle ||N_0||_{L^\infty(\Omega)} \big( \beta_1 T ||\varphi_2 - \varphi_1||_{L^\infty(Q)} + \alpha_h\gamma_h T ||\phi_2 - \phi_1||_{L^\infty(Q)} \big) &+& \\ \\ \displaystyle r_N \bigg(e^{\mu_N T} T^2 e^{\alpha_h\gamma_h ||\phi_2||_{L^\infty(Q)}}\beta_1 ||\varphi_2 - \varphi_1||_{L^\infty(Q)} &+& \\ \\ \displaystyle e^{\mu_N T} T^2 e^{\beta_1 T ||\varphi_1||_{L^\infty(Q)}} \alpha_h\gamma_h ||\phi_2 - \phi_1||_{L^\infty(Q)} &+& \\ \\ \displaystyle \beta_1 T^2 e^{\mu_N T} e^{\alpha_H\gamma_H T ||\phi_2||_{L^\infty(Q)}} e^{\beta_1 T ||\varphi_2||_{L^\infty(Q)}} ||\varphi_2 - \varphi_1||_{L^\infty(Q)} &+& \\ \\ \displaystyle \alpha_h\gamma_h T^2 e^{\mu_N T} e^{\alpha_H\gamma_H T ||\phi_2||_{L^\infty(Q)}} e^{\beta_1 T ||\varphi_2||_{L^\infty(Q)}} ||\phi_2 - \phi_1||_{L^\infty(Q)}\bigg), \end{array} $$ for almost everything $(x,t) \in Q$, i.e., \begin{eqnarray*} \displaystyle ||\Theta(\phi_1, \varphi_1) - \Theta(\phi_2, \varphi_2)||_{L^\infty(Q)} \leq C_1 \big(||\phi_2 - \phi_1||_{L^\infty(Q)} + ||\varphi_2 - \varphi_1||_{L^\infty(Q)} \big). \end{eqnarray*} \hfill$\Box$ \subsection{Proof of Proposition \ref{Prop1}} To not overburden the notation, in this subsection we denote $(A,H)$ as a generic solution of the equations that follows. To get a solution of problem \eqref{P3}, we will apply the Leray-Schauder fixed point theorem to the mapping $\Psi$ defined as follows: \begin{equation} \label{oper} \begin{array}{rccl} \Psi: & [0,1]\times L^\infty(Q) \times L^\infty(Q) & \rightarrow & L^\infty(Q) \times L^\infty(Q) \\ &(l, \phi, \varphi) & \mapsto & (A,H) , \end{array} \end{equation} \noindent where $(A,H)$ is the unique solution of \begin{equation} \label{P66} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A + r_A A\left(1-\frac{A}{k_A}\right)-(\mu_A+\epsilon_A)A - l \beta_3 \Theta(\phi, \varphi) A , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial H}{\partial t} = \xi_H \Delta H + \nu A - \tau_H H - l \gamma_H \Theta(\phi, \varphi) H , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot) =\frac{\partial H}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A(\cdot,0) = A_{0}(\cdot), H(\cdot,0) = H_0(\cdot), &\textup{in}& \Omega, \end{array} \right. \end{equation} with $\Theta(\phi, \varphi)$ given by \eqref{P7}. We affirm that the operator defined in (\ref{oper}) is well defined if the problems \begin{equation} \label{separado1} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A + r_A A\left(1-\frac{A}{k_A}\right)-(\mu_A+\epsilon_A)A - l\beta_3 \Theta(\phi, \varphi) A , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A(\cdot,0) = A_{0}(\cdot), &\textup{in}& \Omega \end{array} \right. \end{equation} and \begin{equation} \label{separado2} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial H}{\partial t} = \xi_H \Delta H + \nu A - \tau_H H - l \gamma_H \Theta(\phi, \varphi) H , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial H}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle H(\cdot,0) = H_0(\cdot), &\textup{in}& \Omega, \end{array} \right. \end{equation} have a unique solution. \subsection{Existence and uniqueness of solution to problems (\ref{separado1}) and (\ref{separado2})} \label{sec32} \noindent\textbf{Step 1:} To determine a solution for the problem (\ref{separado1}), start by studying the following modified problem: \begin{equation} \label{separado31} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A + r_A A \bigg(1 - \frac{A^+}{k_A}\bigg) - (\mu_A + \epsilon_A)A - l \beta_3 \Theta(\phi, \varphi) A, &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A(\cdot,0) = A_{0}(\cdot), &\textup{in}& \Omega. \end{array} \right. \end{equation} In order to prove that problem (\ref{separado31}) admits solution, we define the operator \begin{equation} \label{oper1} \begin{array}{rcl} G: [0,1] \times L^\infty(Q) &\rightarrow& L^\infty(Q) \\ \\ (\lambda, g) &\mapsto& A, \end{array} \end{equation} where $A$ is the unique solution to the problem \begin{equation} \label{separado33} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A + \lambda r_A A \bigg(1 - \frac{g^{+}}{k_A}\bigg) - (\mu_A+\epsilon_A)A - l \beta_3 \Theta(\phi, \varphi) A, &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A(\cdot,0) = A_{0}(\cdot), &\textup{in}& \Omega. \end{array} \right. \end{equation} \vspace{0.1cm} To ensure that we can apply the Leray-Schauder fixed point theorem, we present next a sequence of lemmas: \begin{lemma} \label{lemaAi} Suppose $N_0 \in L^\infty(\Omega)$ and $A_0 \in W^{\frac{3}{2}}_4(\Omega)$. Then the mapping $G :[0,1] \times L^\infty(Q) \rightarrow L^\infty(Q)$ is well defined. \end{lemma} \noindent{\bf Proof:} Note that the coefficients of the problem \eqref{separado33} satisfy the hypotheses of the Proposition \ref{sol. Neumann}. For example, it is immediate that $\lambda r_A - \lambda \frac{r_A}{k_A}g^+ - (\mu_A+\epsilon_A) - l \beta_3 \Theta(\phi, \varphi) \in L^4(Q)$, because $g \in L^\infty(Q)$ and by Lemma \ref{PropertiesEtcFirst}, $\Theta(\phi, \varphi) \in L^\infty(Q)$. Thus, we conclude that there is a unique solution $A \in W^{2,1}_4(Q)$ of problem \eqref{separado33}. Moreover, $A$ satisfies the following estimate: \begin{equation}\label{AA} ||A||_{W^{2,1}_4(Q)} \leq C_p ||A_0||_{W^{\frac{3}{2}}_4(\Omega)}. \end{equation} Finally, from Lemma \ref{icontLp01}, we have $W^{2,1}_4(Q) \hookrightarrow L^\infty(Q)$, and we conclude that the operator $G$ in well defined. \hfill$\Box$ \\ \begin{lemma} \label{nãonegativa1} Suppose $A$ a solution of (\ref{separado33}) and $A_0 \ge 0$ a.e. in $\Omega$, then $A \ge 0$ a.e. in $Q$. \end{lemma} \noindent {\bf Proof:} Multiplying the first equation in $(\ref{separado33})$ by $A^-$ and integrating into $\Omega$, we get \begin{eqnarray*} -\frac{1}{2} \frac{d}{dt} \int_{\Omega} (A^{-})^2 dx &=& \xi_A \int_{\Omega} |\nabla A^-|^2 dx - \lambda r_A \int_{\Omega} (A^-)^2 dx + (\mu_A+\epsilon_A) \int_{\Omega}( A^-)^2 dx \\ &+& \lambda \frac{r_A}{k_A} \int_{\Omega} g_+ (A^-)^2 dx + l \beta_3 \int_{\Omega} \Theta(\phi, \varphi) (A^-)^2 dx. \end{eqnarray*} Thus, \begin{eqnarray*} \frac{d}{dt} \int_{\Omega} (A^{-})^2 dx \leq 2 r_A \int_{\Omega} (A^-)^2 dx \end{eqnarray*} and using the Gronwall's inequality and the fact that $A_0 \ge 0$ a.e. in $\Omega$, we obtain \begin{eqnarray*} \int_{\Omega} (A^{-})^2 dx \leq e^{2 r_A T} \int_{\Omega} ({A_0}^-)^2 dx = 0, \end{eqnarray*} that is, $||A^-(\cdot, t)||_{L^2(\Omega)} = 0$ for all $t \in (0,T)$, where we conclude that $A^{-} = 0$ a.e. in $Q$ and therefore $A \ge 0$ a.e. in $Q$. \hfill$\Box$ \\ \begin{lemma} \label{con1} For each fixed $\lambda\in[0,1]$, the mapping $G(\lambda, \cdot): L^\infty(Q) \rightarrow L^\infty(Q)$ is compact, i.e., it is continuous and maps bounded sets into relatively compacts sets. \end{lemma} \noindent{\bf Proof:} The functions $G(\lambda, g_1)= A_1$ and $G(\lambda, g_2) = A_2$ satisfy the problem \begin{equation*} \label{Aij} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A_i}{\partial t} = \xi_A \Delta A_i + \lambda r_A A_i\bigg(1 - \frac{g^{+}_i}{k_A} \bigg) - (\mu_A + \epsilon_A) A_i - l\beta_3\Theta(\phi, \varphi) A_i , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A_i}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A_i(\cdot,0) = A_0(\cdot), &\textup{in}& \Omega, \end{array} \right. \end{equation*} with $i=1,2$; letting $\tilde{A} := A_1 - A_2$, we have \begin{equation} \label{Ai1} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{A}}{\partial t} = \xi_A \Delta \tilde{A} + \bigg(\lambda r_A - (\mu_A + \epsilon_A) - l \beta_3 \Theta(\phi, \varphi) - \lambda \frac{r_A}{k_A}{g_2}^+ \bigg) \tilde{A} - \vspace{0.2cm} \\ \displaystyle \lambda \frac{r_A}{k_A} A_1 ({g_1}^+ - {g_2}^+), &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{A}}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{A}(\cdot,0) = \tilde{A}_0(\cdot) = 0, &\textup{in}& \Omega. \end{array} \right. \end{equation} Since $g_i, A_i \in L^\infty(Q)$, $i=1,2$, from Proposition \ref{sol. Neumann}, we get \begin{eqnarray*} ||\tilde{A}||_{W_{4}^{2, 1}(Q)} \leq C_p \lambda \frac{r_A}{k_A} || A_1 ({g_1}^+ - {g_2}^+) ||_{L^4(Q)} \\ \leq C_p \frac{r_A}{k_A} T^{\frac{1}{4}} |\Omega|^{\frac{1}{4}} ||A_1||_{L^\infty(Q)} ||g_1 - g_2||_{L^\infty(Q)}. \end{eqnarray*} Then, by Lemma \ref{icontLp01}, we finally have \begin{eqnarray*} ||G(\lambda, g_1) - G(\lambda, g_2)||_{L^\infty(Q)} \leq C ||g_1 - g_2||_{L^\infty(Q)}, \end{eqnarray*} where $C$ depends on $C_p$, $r_A$, $k_A$, $T$, $\Omega$, $||A_1||_{L^\infty(Q)}$ and the immersion constant. To show that $ G (\lambda, \cdot) $ is compact, we use the fact that immersion $W^{2,1}_4(Q) \hookrightarrow L^\infty(Q)$ is compact and that $G(\lambda, \cdot)$ is the composition between the inclusion operator and the solution operator, i.e., $G(\lambda, \cdot): L^\infty(Q) \rightarrow W^{2,1}_4(Q) \rightarrow L^\infty(Q)$. \hfill$\Box$ \\ \begin{lemma} \label{unifcon1} Given a bounded subset $B\subset\L^\infty(Q)$, for each $g \in B$, the mapping $G(\cdot, g): [0, 1] \rightarrow L^\infty(Q)$ is continuous, uniformly with respect to $B$. \end{lemma} \noindent{\bf Proof:} Since $B \in L^\infty(Q)$ is bounded, there is $r_B \ge 0$ such that, for any $g \in B$, we have $||g||_{L^\infty(Q)} \leq r_B$. Now, let us fix $g \in L^\infty(Q)$ and consider $\lambda_1, \lambda_2 \in [0,1]$ and denote $G(\lambda_1, g) = A_1$, $G(\lambda_2, g) = A_2$ and $\tilde{A} = A_1 - A_2$. Then, $\tilde{A}$ satisfies \begin{equation} \label{B} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{A}}{\partial t} = \xi_A \Delta \tilde{A} - \bigg(r_A \lambda_2 - (\mu_A + \epsilon_A) - l \beta_3 \Theta(\phi, \varphi) - \frac{r_A}{k_A} \lambda_2 g^+ \bigg) \tilde{A} - \vspace{0.2cm} \\ \displaystyle \bigg(r_A A_1 + \frac{r_A}{k_A} g^+ A_1\bigg) (\lambda_1 - \lambda_2), &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{A}}{\partial \eta} = 0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{A}(\cdot, 0) = \tilde{A}_0(\cdot) = 0, &\textup{in}& \Omega. \end{array} \right. \end{equation} Since $g, A_i \in L^\infty(Q)$, $i=1,2$, from Proposition \ref{sol. Neumann}, we get \begin{eqnarray*} ||\tilde{A}||_{W_{4}^{2,1}(Q)} \leq C_p |\lambda_1 - \lambda_2| \bigg|\bigg|\bigg(r_A A_1 + \frac{r_A}{k_A} g^+ A_1\bigg)\bigg|\bigg|_{L^4(Q)} \\ \leq C_p \bigg(r_A + \frac{r_A}{k_A} r_B \bigg) T^{\frac{1}{4}} |\Omega|^{\frac{1}{4}} ||A_1||_{L^\infty(Q)} |\lambda_1 - \lambda_2|. \end{eqnarray*} Then, by Lemma \ref{icontLp01}, we finally have \begin{eqnarray*} ||G(\lambda_1,g) - G(\lambda_2,g)||_{L^\infty(Q)} \leq C |\lambda_1 - \lambda_2|, \end{eqnarray*} where $C$ depends on $C_p$, $r_A$, $k_A$, $r_B$, $T$, $\Omega$, $||A_1||_{L^\infty(Q)}$ and the immersion constant. \hfill$\Box$ \\ \begin{lemma} \label{estimativa1} Suppose $A_0 \leq k_A$ a.e. in $\Omega$, then there exists a number $\rho_1 > 0$ such that, for any $\lambda \in [0,1]$ and any possible fixed point $A \in L^\infty(Q)$ of $G(\lambda, \cdot)$, there holds $\|A\|_{L^\infty(Q)} < \rho_1$. \end{lemma} \noindent {\bf Proof:} Let $A \in L^\infty(Q)$ be a possible fixed point of $G(l,\cdot)$ associated to a given $\lambda \in [0,1]$; then $A$ satisfies \begin{equation}\label{rhox} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A + \lambda r_A A \bigg(1 - \frac{A}{k_A}\bigg) - (\mu_A + \epsilon_A)A - l \beta_3 \Theta(\phi, \varphi) A, &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A(\cdot,0) = A_{0}(\cdot), &\textup{in}& \Omega, \end{array} \right. \end{equation} since by the Lemma \ref{nãonegativa1}, $A = A^+$. We observe that the first equation in (\ref{rhox}) can be rewritten as \begin{eqnarray*} \frac{\partial}{\partial t} (A - k_A) &=& \xi_A \Delta (A - k_A) - \lambda \frac{r_A}{k_A} A (A - k_A) \\ &-& (\mu_A+\epsilon_A)A - \beta_3 \Theta(\phi, \varphi) A. \end{eqnarray*} Now, multiplying by $(A - k_A)^+$ and integrating in $\Omega$, we obtain \begin{eqnarray*} \frac{1}{2} \frac{d}{dt} \int_{\Omega} \big((A - k_A)^+\big)^2 dx &=& - \xi_A \int_{\Omega} |\nabla(A - k_A)^{+}|^2 dx - \lambda \frac{r_A}{k_A} \int_{\Omega} A \big((A - k_A)^+\big)^2 dx \\ &-& (\mu_A+\epsilon_A) \int_{\Omega} A (A - k_A)^+ dx \\ &-& \beta_3 \int_{\Omega} \Theta(\phi, \varphi) A (A - k_A)^+ dx \\ &\leq& 0. \end{eqnarray*} Thus, using the Gronwall's inequality and the fact that $A_0 \leq k_A$ a.e. in $\Omega$, it follow that \begin{eqnarray*} \int_{\Omega} \big((A - k_A)^+\big)^2 dx &\leq & \int_{\Omega} \big((A_0 - k_A)^+\big)^2 dx = 0, \end{eqnarray*} that is, $||(A(\cdot, t) - k_A)^+||_{L^2(\Omega)} = 0$ for all $t \in (0,T)$ and therefore $(A - k_A)^+ = 0$ a.e. in $Q$, and we conclude that $A \leq k_A$ a.e. in $Q$. Thus, $||A||_{L^\infty(Q)} \leq k_A$, which suggests $\rho_1 = k_A + 1$. \hfill$\Box$ \\ \begin{lemma} \label{fix1} The mapping $G(0,\cdot): L^\infty(Q) \rightarrow L^\infty(Q)$ has a unique fixed point. \end{lemma} \noindent {\bf Proof:} Indeed, letting $\lambda =0$ in\eqref{separado33}, $A$ is a fixed point of $G(0, \cdot)$ if, and only if, $A$ is the unique solution to the problem \begin{equation*} \label{separado3} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A - \big( (\mu_A+\epsilon_A) + l\beta_3\Theta(\phi, \varphi)\big)A, &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A(\cdot,0) = A_{0}(\cdot), &\textup{in}& \Omega. \end{array} \right. \end{equation*} By Proposition {\ref{sol. Neumann}}, we have the existence of a unique solution $A \in W^{2,1}_{4}(Q)\hookrightarrow L^{\infty}(Q)$ of this problem; therefore $G(0, \cdot )$ has a unique fixed point in $L^\infty(Q)$. \hfill$\Box$ \\ \begin{proposition} \label{existenciaA} There is a nonnegative solution $A \in W^{2,1}_4(Q)$ of the problem (\ref{separado1}). \end{proposition} \noindent {\bf Proof:} From Lemmas \ref{lemaAi}, \ref{con1}, \ref{unifcon1}, \ref{estimativa1} and \ref{fix1}, we conclude that the mapping $G: [0,1] \times L^\infty(Q) \rightarrow L^\infty(Q)$ satisfies the hypotheses of the Leray-Schauder's fixed point theorem (see Friedman \cite[pp.~189, Theorem 3]{Friedman}). Thus, there exists $A \in L^\infty(Q)$ such that $G(1, A) = A$. Moreover, by Lemmas \ref{lemaAi} and \ref{nãonegativa1}, $A \in W^{2,1}_4(Q)$ is nonnegative and $A$ is the required solution of (\ref{separado1}). \hfill$\Box$ \\ \begin{lemma} \label{nãonegativa3} Suppose $A$ a solution of (\ref{separado1}) and $A_0 \leq k_A$ a.e. in $\Omega$, then $A \leq k_A$ a.e. in $Q$. \end{lemma} \noindent {\bf Proof:} Analogous to Lemma \ref{estimativa1}. \hfill$\Box$ \\ \begin{proposition} \label{unicidadeA} The solution $A$ of the problem (\ref{separado1}) is unique. \end{proposition} \noindent {\bf Proof:} Let $A_1$ and $A_2$ be solutions to the problem (\ref{separado1}); if $\tilde{A} = A_1 - A_2$, then $\tilde{A}$ satisfies the o problem \begin{equation} \label{unicidade11} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{A}}{\partial t} = \xi_A \Delta \tilde{A} + r_A \tilde{A} - \frac{r_A}{k_A}(A_1 + A_2) \tilde{A} - (\mu_A+\epsilon_A) \tilde{A} - l\beta_3 \Theta(\phi, \varphi) \tilde{A} , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{A}}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{A}(\cdot,0) = \tilde{A}_0(\cdot) = 0, &\textup{in}& \Omega. \end{array} \right. \end{equation} Multiplying the first equation of (\ref{unicidade11}) by $\tilde{A}$ and integrating in $\Omega$, we obtain \begin{eqnarray*} \frac{1}{2}\frac{d}{dt} \int_{\Omega} \tilde{A}^2 dx = - \xi_A \int_{\Omega} |\nabla \tilde{A}|^2 dx + r_A \int_{\Omega} \tilde{A}^2 dx - \frac{r_A}{k_A} \int_{\Omega} (A_1 + A_2) \tilde{A}^2 dx \\ - (\mu_A+\epsilon_A) \int_{\Omega} \tilde{A}^2 dx - l\beta_3 \int_{\Omega} \Theta(\phi, \varphi) \tilde{A}^2 dx. \end{eqnarray*} Therefore, \begin{eqnarray*} \frac{d}{dt} \int_{\Omega} \tilde{A}^2 dx &\leq& 2 r_A \int_{\Omega} \tilde{A}^2 dx, \end{eqnarray*} and by the Gronwall's inequality, we obtain \begin{eqnarray*} \int_{\Omega} \tilde{A}^2 dx &\leq& e^{2r_A T} \int_{\Omega} (\tilde{A}_0)^2 dx = 0, \end{eqnarray*} that is, $||\tilde{A}(\cdot,t)||_{L^2(\Omega)} = 0$ for all $t \in (0, T)$, which suggests $\tilde{A} = 0$ a.e. in $Q$ and therefore $A_1 = A_2$ a.e. in $Q$. \hfill$\Box$ \\ \noindent\textbf{Step 2:} The existence and uniqueness of solution for the problem (\ref{separado2}) is given by combining Propositions \ref{existenciaA} and \ref{unicidadeA} and the following result: \begin{proposition} \label{existenciaH} Suppose $N_0 \in L^{\infty}(\Omega)$ and $H_0 \in W_{4}^{\frac{3}{2}}(\Omega)$. Then the problem (\ref{separado2}) has a unique solution. \end{proposition} \noindent {\bf Proof:} It's immediate that the coefficients of the problem \eqref{separado2} satisfy the hypotheses of Proposition \ref{sol. Neumann}, and we conclude that there is a unique solution $H \in W^{2,1}_4(Q)$ of the problem $(\ref{separado2})$. Moreover, $H$ satisfies the following estimate: \begin{equation}\label{HH} ||H||_{W^{2,1}_4(Q)} \leq C_p \bigg( \nu k_A T^{\frac{1}{4}}|\Omega|^{\frac{1}{4}} + ||H_0||_{W^{\frac{3}{2}}_4(\Omega)}\bigg) \end{equation} \hfill$\Box$ \\ \begin{lemma} \label{nãonegativa2} Suppose $H$ a solution do problem (\ref{separado2}) and $H_0 \ge 0$ a.e. in $\Omega$, then $H \ge 0$ a.e. in $Q$. \end{lemma} \noindent {\bf Proof:} Multiplying the first equation in (\ref{separado2}) by $H^-$ and integrating in $\Omega$, we get \begin{eqnarray*} -\frac{1}{2} \frac{d}{dt} \int_{\Omega} (H^-)^2 dx &=& \xi_H \int_{\Omega} |\nabla H^-|^2 dx + \nu \int_{\Omega} A H^- dx \\ &+& \tau_H \int_{\Omega} (H^-)^2 + l \gamma_H \int_{\Omega} \Theta(\phi, \varphi) (H^-)^2 dx. \end{eqnarray*} Using the fact that $A \ge 0$, we obtain \begin{eqnarray*} \frac{d}{dt} \int_{\Omega} (H_-)^2 dx \leq 0. \end{eqnarray*} Therefore, using the Gronwall's inequality and the fact that $H_0 \ge 0$ a.e. on $\Omega$, it follow that $||H^-(\cdot, t)||_{L^2(\Omega)} = 0$ for all $t \in (0,T)$, that is, $H^- = 0$ a.e. in $Q$, which suggests $H \ge 0$ a.e. in $Q$. \hfill$\Box$ \\ \subsection{Continuation of Proof of Proposition \ref{Prop1}} To guarantee that we can apply the Leray-Schauder fixed point theorem, we present next a sequence of lemmas. \begin{lemma} \label{bemdefinido} The mapping $\Psi :[0,1]\times L^\infty(Q) \times L^\infty(Q) \rightarrow L^\infty(Q) \times L^\infty(Q)$ is well defined. \end{lemma} \noindent{\bf Proof:} Just combine the Propositions \ref{existenciaA}, \ref{unicidadeA} and \ref{existenciaH}. \hfill$\Box$ \\ \begin{lemma} \label{con} For each fixed $l \in[0,1]$, the mapping $\Psi(l, \cdot, \cdot): L^{\infty}(Q) \times L^{\infty}(Q) \rightarrow L^{\infty}(Q) \times L^{\infty}(Q)$ is compact, i.e., it is continuous and maps bounded sets into relatively compacts sets. \end{lemma} \noindent {\bf Proof:} Consider $(\phi_1, \varphi_1), (\phi_2, \varphi_2) \in L^{\infty}(Q) \times L^{\infty}(Q)$ such that $\Psi(l, \phi_1, \varphi_1) = (A_1, H_1), \Psi(l, \phi_2, \varphi_2) = (A_2, H_2)$; if $\tilde{A} = A_1 - A_2$ and $\tilde{H} = H_1 - H_2$, we have that $\tilde{A}$ and $\tilde{H}$ satisfy the following problems, respectively: \begin{equation} \label{separado1i} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{A}}{\partial t} = \xi_A \Delta \tilde{A} + \bigg(r_A - (\mu_A+\epsilon_A) - l \frac{r_A}{k_A}(A_1 + A_2) - l\beta_3\Theta(\phi_1, \varphi_1)\bigg) \tilde{A} - \vspace{0.2cm} \\ \displaystyle l\beta_3\big[\Theta(\phi_1, \varphi_1) - \Theta(\phi_1, \varphi_1)\big] A_2 , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{A}}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{A}(\cdot,0) = \tilde{A}_{0}(\cdot) = 0, &\textup{in}& \Omega, \end{array} \right. \end{equation} \begin{equation} \label{separado2i} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{H}}{\partial t} = \xi_H \Delta \tilde{H} - \big[ \tau_H + l \gamma_H \Theta(\phi_1, \varphi_1)\big] \tilde{H} + \nu \tilde{A} - \vspace{0.2cm} \\ \displaystyle l \gamma_H \big[\Theta(\phi_1, \varphi_1) - \Theta(\phi_2, \varphi_2)\big]H_2 , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{H}}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{H}(\cdot,0) = \tilde{H}_0(\cdot) = 0, &\textup{in}& \Omega. \end{array} \right. \end{equation} We observe that the system (\ref{separado1i}) satisfies the hypothesis of Proposition \ref{sol. Neumann}, therefore using the Lemmas \ref{PropertiesEtcFirst} and \ref{nãonegativa3}, $\tilde{A} \in W^{2,1}_4(Q)$ satisfies the following estimates \begin{eqnarray*} ||\tilde{A}||_{W_{4}^{2, 1}(Q)} \leq C_p || l\beta_3\big[\Theta(\phi_1, \varphi_1) - \Theta(\phi_2, \varphi_2)\big] A_2||_{L^4(Q)} \\ \leq C_p \beta_3 k_A T^{\frac{1}{4}} |\Omega|^{\frac{1}{4}} C_1 ||(\phi_1, \varphi_1) - (\phi_2, \varphi_2)||_{L^{\infty}(Q)}. \end{eqnarray*} Thus, by Proposition \ref{icontLp01}, we obtain \begin{equation}\label{ai} ||\tilde{A}||_{L^\infty(Q)} \leq C ||(\phi_1, \varphi_1) - (\phi_2, \varphi_2)||_{L^{\infty}(Q)}, \end{equation} where $C$ depends on $C_p, \beta_3, k_A, T, \Omega, C_1$ and the immersion constant. We observe that the system (\ref{separado2i}) satisfies the hypothesis of Proposition \ref{sol. Neumann}, therefore using the Lemma \ref{PropertiesEtcFirst}, $\tilde{H} \in W^{2,1}_4(Q)$ satisfies the following estimates \begin{eqnarray*} ||\tilde{H}||_{W_{4}^{2, 1}(Q)} \leq C_p || \nu \tilde{A} - l \gamma_H \big[\Theta(\phi_1, \varphi_1) - \Theta(\phi_2, \varphi_2)\big]H_2||_{L^4(Q)} \\ \leq C_p \nu ||\tilde{A}||_{L^4(Q)} + C_p \gamma_H T^{\frac{1}{4}} |\Omega|^{\frac{1}{4}} ||H_2||_{L^{\infty}(Q)} ||\Theta(\phi_1, \varphi_1) - \Theta(\phi_2, \varphi_2)||_{L^{\infty}(Q)} \\ \leq C_p \nu ||\tilde{A}||_{L^\infty(Q)} + C_p \gamma_H T^{\frac{1}{4}} |\Omega|^{\frac{1}{4}} C_1 ||H_2||_{L^{\infty}(Q)} ||(\phi_1, \varphi_1) - (\phi_2, \varphi_2)||_{L^{\infty}(Q)}, \end{eqnarray*} therefore, using the estimate (\ref{ai}) and Proposition \ref{icontLp01}, we obtain \begin{equation}\label{bx} ||\tilde{H}||_{L^\infty(Q)} \leq C ||(\phi_1, \varphi_1) - (\phi_2, \varphi_2)||_{L^{\infty}(Q)}, \end{equation} where $C$ depends on $C_p$, $\nu$, $\gamma_H$, $T$, $\Omega$, $C_1$, $||H_2||_{L^\infty(Q)}$ and immersion constants. Adding the inequalities obtained in (\ref{ai}) and (\ref{bx}), finally we get \begin{eqnarray*} ||\tilde{A}||_{L^\infty(Q)} + ||\tilde{H}||_{L^\infty(Q)} &\leq& C ||(\phi_1, \varphi_1) - (\phi_2, \varphi_2)||_{L^{\infty}(Q)}. \end{eqnarray*} To show that $\Psi(l, \cdot, \cdot) $ is compact, we use the fact that immersion $W^{2,1}_4(Q) \hookrightarrow L^\infty(Q)$ is compact and that $\Psi(l, \cdot, \cdot)$ is the composition between the inclusion operator and the solution operator, i.e., $\Psi(l, \cdot, \cdot): L^\infty(Q) \times L^\infty(Q) \rightarrow W^{2,1}_4(Q) \times W^{2,1}_4(Q) \rightarrow L^\infty(Q) \times L^\infty(Q)$. \hfill$\Box$ \\ \begin{lemma} \label{unifcon} Given a bounded subset $B \subset L^{\infty}(Q)\times L^{\infty}(Q)$, for each $(\phi, \varphi) \in B$, the mapping $\Psi(\cdot, \phi, \varphi): [0, 1] \rightarrow L^{\infty}(Q)\times L^{\infty}(Q)$ is continuous, uniformly with respect to $B$. \end{lemma} \noindent {\bf Proof:} Since $B \in L^\infty(Q) \times L^\infty(Q)$ is bounded, there is $r_B \ge 0$ such that, for any $(\phi, \varphi) \in B$ we have $||\phi||_{L^\infty(Q)} + ||\varphi||_{L^\infty(Q)} \leq r_B$. Now, let us fix $(\phi, \varphi) \in L^\infty(Q)$ and consider $l_1, l_2 \in [0,1]$ and denote $\Psi(\lambda_1, \phi, \varphi) = (A_1, H_1)$, $\Psi(l_2, \phi, \varphi) = (A_2, H_2)$; if $\tilde{A} = A_1 - A_2$ and $\tilde{H} = H_1 - H_2$, we have that $\tilde{A}$ and $\tilde{H}$ satisfy the following problems, respectively: \begin{equation} \label{B1} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{A}}{\partial t} = \xi_A \Delta \tilde{A} + \bigg(r_A - (\mu_A + \epsilon_A) - \frac{r_A}{k_A}l_1(A_1 + A_2) - \beta_3\Theta(\phi, \varphi)l_1 \bigg)\tilde{A} - \vspace{0.2cm} \\ \displaystyle \frac{r_A}{k_A}A_2^2(l_1 - l_2) - \beta_3\Theta(\phi, \varphi)A_2(l_1 - l_2), &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{A}}{\partial \eta} = 0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{A}(\cdot, 0) = \tilde{A}_0(\cdot) = 0, &\textup{in}& \Omega. \end{array} \right. \end{equation} \begin{equation} \label{B2} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{H}}{\partial t} = \xi_H \Delta \tilde{H} - (\tau_H + \gamma_H\Theta(\phi, \varphi)l_1)\tilde{H} + \vspace{0.2cm} \\ \nu\tilde{A} - \gamma_H\Theta(\phi, \varphi)H_2(l_1 - l_2), &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{H}}{\partial \eta} = 0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{H}(\cdot, 0) = \tilde{H}_0(\cdot) = 0, &\textup{in}& \Omega. \end{array} \right. \end{equation} We observe that the problem (\ref{B1}) satisfies the hypothesis of Proposition \ref{sol. Neumann}, therefore using the Lemmas \ref{PropertiesEtcFirst} and \ref{nãonegativa3}, $A \in W^{2,1}_4(Q)$ satisfies the following estimates \begin{eqnarray*} ||\tilde{A}||_{W_{4}^{2, 1}(Q)} \leq C_p \bigg|\bigg|\frac{r_A}{k_A}A_2^2(l_1 - l_2) - \beta_3\Theta(\phi, \varphi)A_2(l_1 - l_2)\bigg|\bigg|_{L^4(Q)} \\ \leq C_p k_A T^{\frac{1}{4}} |\Omega|^{\frac{1}{4}} (r_A + \beta_3 ||N_0||_{L^\infty(\Omega)} + r_N T) |l_1 - l_2|. \end{eqnarray*} Thus, by Proposition \ref{icontLp01}, we obtain \begin{equation}\label{aj} ||\tilde{A}||_{L^\infty(Q)} \leq C |l_1 - l_2|, \end{equation} where $C$ depends on $C_p$, $r_A$, $k_A$, $r_N$, $T$, $\Omega$, $\beta_3$, $||N_0||_{L^\infty(\Omega)}$ and the immersion constant. We observe that the system (\ref{B2}) satisfies the hypothesis of Proposition \ref{sol. Neumann}, therefore using the Lemma \ref{PropertiesEtcFirst}, $\tilde{H} \in W^{2,1}_4(Q)$ satisfies the following estimates \begin{eqnarray*} ||\tilde{H}||_{W_{4}^{2, 1}(Q)} \leq C_p ||\nu\tilde{A} - \gamma_H\Theta(\phi, \varphi)H_2(l_1 - l_2)||_{L^4(Q)} \\ \leq C_p \nu ||\tilde{A}||_{L^4(Q)} + C_p \gamma_H T^{\frac{1}{4}} |\Omega|^{\frac{1}{4}}||H_2||_{L^{\infty}(Q)} (||N_0||_{L^\infty(\Omega)} + r_N T) |l_1 - l_2| \\ \leq \bar{C}_p \nu ||\tilde{A}||_{L^\infty(Q)} + C_p \gamma_H T^{\frac{1}{4}} |\Omega|^{\frac{1}{4}} ||H_2||_{L^{\infty}(Q)} (||N_0||_{L^\infty(\Omega)} + r_N T) |l_1 - l_2|, \end{eqnarray*} therefore, using the estimate (\ref{aj}) and Proposition \ref{icontLp01}, we obtain \begin{equation}\label{Bii} ||\tilde{H}||_{L^\infty(Q)} \leq C |l_1 - l_2|. \end{equation} where $C$ depends on $C_p$, $r_A$, $k_A$, $r_N$, $\nu$, $\gamma_H$, $T$, $\Omega$, $\beta_3$, $||H_2||_{L^\infty(Q)}$, $||N_0||_{L^\infty(\Omega)}$ and immersion constants. Adding the inequalities obtained in (\ref{aj}) and (\ref{Bii}), finally we get \begin{eqnarray*} ||\tilde{A}||_{L^\infty(Q)} + ||\tilde{H}||_{L^\infty(Q)} &\leq& C |l_1 - l_2|. \end{eqnarray*} \hfill$\Box$ \\ \begin{lemma} \label{estimativa} There exists a number $\rho > 0$ such that, for any $l \in [0,1]$ and any possible fixed point $(A,H) \in L^\infty(Q) \times L^\infty(Q)$ of $\Psi(l, \cdot, \cdot)$, there holds $\|(A,H)\|_{L^\infty(Q)} < \rho$. \end{lemma} \noindent{\bf Proof:} Let $(A,H) \in L^\infty(Q) \times L^\infty(Q)$ be a possible fixed point of $\Psi(l, \cdot, \cdot)$ associated to a given $l \in [0,1]$; then $(A,H)$ satisfies \begin{equation} \label{conex12} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A + r_A A\left(1-\frac{A}{k_A}\right)-(\mu_A+\epsilon_A)A - l\beta_3 \Theta(\phi, \varphi) A , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial H}{\partial t} = \xi_H \Delta H + \nu A - \tau_H H - l \gamma_H \Theta(\phi, \varphi) H , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot) =\frac{\partial H}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A(\cdot,0) = A_{0}(\cdot), H(\cdot,0) = H_0(\cdot), &\textup{in}& \Omega. \end{array} \right. \end{equation} By Proposition \ref{existenciaH}, we know that $H$ satisfies the estimate (\ref{existenciaH}), that is, \begin{equation*}\label{HH} ||H||_{W^{2,1}_4(Q)} \leq C_p \bigg( \nu k_A T^{\frac{1}{4}}|\Omega|^{\frac{1}{4}} + ||H_0||_{W^{\frac{3}{2}}_4(\Omega)}\bigg). \end{equation*} Then, by Lemma \ref{icontLp01}, we have \begin{equation*}\label{X} ||H||_{L^\infty(Q)} \leq \bar{C}_p \bigg( \nu k_A T^{\frac{1}{4}}|\Omega|^{\frac{1}{4}} + ||H_0||_{W^{\frac{3}{2}}_4(\Omega)}\bigg). \end{equation*} Therefore, by Lemma \ref{estimativa1}, we have \begin{equation*}\label{X} ||A||_{L^\infty} + ||H||_{L^\infty(Q)} < \rho_1 + \bar{C}_p \bigg( \nu k_A T^{\frac{1}{4}}|\Omega|^{\frac{1}{4}} + ||H_0||_{W^{\frac{3}{2}}_4(\Omega)}\bigg), \end{equation*} which suggests we take \begin{eqnarray*} \rho = \rho_1 + \bar{C}_p \bigg( \nu k_A T^{\frac{1}{4}}|\Omega|^{\frac{1}{4}} + ||H_0||_{W^{\frac{3}{2}}_4(\Omega)}\bigg). \end{eqnarray*} \hfill$\Box$ \\ \begin{lemma} \label{fix} The mapping $\Psi(0, \cdot, \cdot): L^\infty(Q) \times L^\infty(Q) \rightarrow L^{\infty}(Q)\times L^{\infty}(Q)$ has a unique fixed point. \end{lemma} \noindent {\bf Proof:} Indeed, letting $l =0$ in (\ref{P66}), $(A,H)$ is a fixed point of $\Psi(0, \cdot, \cdot)$ if, and only if, $(A,H)$ is the unique solution to the problem \begin{equation*} \label{P6} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial A}{\partial t} = \xi_A \Delta A + r_A A\left(1-\frac{A}{k_A}\right)-(\mu_A+\epsilon_A) A, &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial H}{\partial t} = \xi_H \Delta H + \nu A - \tau_H H, &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial A}{\partial \eta} (\cdot)=\frac{\partial H}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle A(\cdot,0) = A_{0}(\cdot), H(\cdot,0) = H_0(\cdot), &\textup{in}& \Omega. \end{array} \right. \end{equation*} Analogous to what was done in Subsection \ref{sec32}, we guarantee the existence of a unique solution $(A,H) \in L^{\infty}(Q) \times L^\infty(Q)$ of this last problem; therefore $\Psi(0, \cdot, \cdot)$ has a unique fixed point in $L^\infty(Q) \times L^\infty(Q)$. \hfill$\Box$ \\ \begin{proposition} \label{h1} There is a nonnegative solution $(\hat{A}, \hat{H}) \in W^{2,1}_4(Q) \times W^{2,1}_4(Q)$ of problem (\ref{P3}). \end{proposition} \noindent {\bf Proof:} From Lemmas \ref{bemdefinido}, \ref{con}, \ref{unifcon}, \ref{estimativa} and \ref{fix}, we conclude that the mapping $\Psi: [0,1] \times L^\infty(Q) \times L^\infty(Q) \rightarrow L^\infty(Q) \times L^\infty(Q)$ satisfies the hipotheses of the Leray-Schauder's fixed point theorem (see Friedman \cite[pp.~189, Theorem 3]{Friedman}). Thus, there exists $(\hat{A}, \hat{H}) \in L^\infty(Q) \times L^\infty(Q)$ such that $\Psi(1, \hat{A}, \hat{H}) = (\hat{A}, \hat{H})$. Moreover, by Lemmas \ref{nãonegativa1}, \ref{existenciaA}, \ref{existenciaH} and \ref{nãonegativa2}, $(\hat{A}, \hat{H}) \in W^{2,1}_4(Q) \times W^{2,1}_4(Q)$ is nonnegative, meets the estimates (\ref{AA}), (\ref{existenciaH}) and $(\hat{A}, \hat{H})$ is the required solution of (\ref{P3}). \hfill$\Box$ \\ \section{Proof of Theorem \ref{Teorema1}} \begin{proposition} \label{h2} There is a nonnegative solution $(\hat{N}, \hat{A}, \hat{H}) \in L^\infty(Q) \times W^{2,1}_4(Q) \times W^{2,1}_4(Q)$ of the modified problem (\ref{P01}). \end{proposition} \noindent {\bf Proof:} Just combine the Proposition \ref{h1}, the Remark \ref{obs1} and the Lemma \ref{PropertiesEtcFirst}. \hfill$\Box$ \\ \begin{remark} \label{estimativaN} We affirm that $\hat{N} \in W$. Indeed, of the Lemma \ref{PropertiesEtcFirst} we know that $\hat{N} = \Theta(\hat{H}, \hat{A}) \in L^\infty(Q)$. Moreover, returning to the first equation of (\ref{P01}), using the Lemmas \ref{PropertiesEtcFirst} and \ref{nãonegativa3} and the fact that $\hat{H} \in W^{2,1}_4(Q) \subset L^\infty(Q)$, it follow that: \begin{equation}\label{Nt} \bigg|\frac{\partial \hat{N}}{\partial t}\bigg| \leq r_N + (\mu_N + \beta_1 k_A + \alpha_H\gamma_H ||\hat{H}||_{L^\infty(Q)}) (||N_0||_{L^\infty(\Omega)} + r_N T), \end{equation} a.e. in $Q$, i.e., $\hat{N}_t \in L^\infty(Q)$. \end{remark} \begin{proposition} \label{j1} There is a nonnegative solution $(N, A, H) \in W \times W^{2,1}_4(Q) \times W^{2,1}_4(Q)$ of problem (\ref{0riginalEquations}). \end{proposition} \noindent {\bf Proof:} Just combine the Proposition \ref{h2} and the Remarks \ref{obs2} and \ref{estimativaN}. \hfill$\Box$ \\ \begin{proposition} \label{j2} The solution $(N, A, H)$ of the problem (\ref{0riginalEquations}) is unique. \end{proposition} \noindent {\bf Proof:} Let $(N_1, A_1, D_1)$ and $(N_2, A_2, H_2)$ be solutions to the problem (\ref{0riginalEquations}); if $\tilde{N} = N_1 - N_2, \tilde{A} = A_1 - A_2$ and $\tilde{H} = H_1 - H_2$, then $\tilde{N}$, $\tilde{A}$ and $\tilde{H}$ satisfy the following problems, respectively: \begin{equation} \label{original1} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{N}}{\partial t} = - \mu_N \tilde{N} - \beta_1 N_1 \tilde{A} - \beta_1 A_2 \tilde{N} - \alpha_H\gamma_H N_1 \tilde{H} - \alpha_H\gamma_H H_2 \tilde{N} , & \textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \displaystyle \tilde{N}(\cdot,0) = \tilde{N}_0(\cdot) = 0, &\textup{in}& \Omega, \end{array} \right. \end{equation} \begin{equation} \label{original2} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{A}}{\partial t} = \xi_A \Delta \tilde{A} + r_A \tilde{A} - \frac{r_A}{k_A}(A_1 + A_2)\tilde{A} -(\mu_A+\epsilon_A)\tilde{A} - \beta_3 N_1 \tilde{A} - \beta_3 A_2\tilde{N}, &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{A}}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{A}(\cdot,0) = \tilde{A}_0(\cdot) = 0, &\textup{in}& \Omega, \end{array} \right. \end{equation} \begin{equation} \label{original3} \left\{ \begin{array}{lcl} \displaystyle \frac{\partial \tilde{H}}{\partial t} = \xi_H \Delta \tilde{H} + \nu \tilde{A} - \tau_H \tilde{H} - \gamma_H N_1 \tilde{H} - \gamma_H H_2 \tilde{N} , &\textup{in}& Q, \vspace{0.2cm} \\ \displaystyle \frac{\partial \tilde{H}}{\partial \eta} (\cdot) =0, &\textup{on}& \Gamma, \vspace{0.2cm} \\ \displaystyle \tilde{H}(\cdot,0) = \tilde{H}_0(\cdot) = 0, &\textup{in}& \Omega. \end{array} \right. \end{equation} Multiplying the first equation of (\ref{original1}) by $\tilde{N}$, integrating into $\Omega$, using the fact that $N_1 \leq ||N_0||_{L^\infty(\Omega)} + r_N T$ and the inequality of Young, we have \begin{eqnarray*} \frac{1}{2} \frac{d}{dt} \int_{\Omega} \tilde{N}^2 dx &=& - \mu_N \int_{\Omega} \tilde{N}^2 dx - \beta_1 \int_{\Omega} N_1 \tilde{A} \tilde{N} dx - \beta_1 \int_{\Omega} A_2 \tilde{N}^2 dx \\ &-& \alpha_H\gamma_H \int_{\Omega} N_1 \tilde{H} \tilde{N} dx - \alpha_H\gamma_H \int_{\Omega} H_2 \tilde{N}^2 dx \\ &\leq& (||N_0||_{L^\infty(\Omega)})\bigg(\beta_1 \int_{\Omega} |\tilde{A}||\tilde{N}| dx + \alpha_H\gamma_H \int_{\Omega} |\tilde{H}||\tilde{N}| dx\bigg) \\ &\leq& C \int_{\Omega} (\tilde{A}^2 + \tilde{N}^2 + \tilde{H}^2) dx, \end{eqnarray*} where $C$ depends on $\beta_1$, $\alpha_H$, $\gamma_H$, $r_N$, $T$ and $||N_0||_{L^\infty(\Omega)}$. Now, multiplying the first equation of (\ref{original2}) by $\tilde{A}$, integrating into $\Omega$, using the fact that $A \leq k_A$ and the inequality of Young, we obtain \begin{eqnarray*} \frac{1}{2} \frac{d}{dt} \int_{\Omega} \tilde{A}^2 dx &=& - \xi_A \int_{\Omega} |\nabla \tilde{A}|^2 dx + r_A \int_{\Omega} \tilde{A}^2 dx - \frac{r_A}{k_A} \int_{\Omega}(A_1 + A_2)\tilde{A}^2 dx \\ &-& (\mu_A+\epsilon_A)\int_{\Omega}\tilde{A}^2 dx - \beta_3 \int_{\Omega} N_1 \tilde{A}^2 dx - \beta_3 \int_{\Omega} A_2\tilde{N} \tilde{A} dx \\ &\leq& r_A \int_{\Omega} \tilde{A}^2 dx + \beta_3 k_A \int_{\Omega} |\tilde{N}| |\tilde{A}| dx \\ &\leq& C \int_{\Omega} (\tilde{A}^2 + \tilde{N}^2 + \tilde{H}^2) dx, \end{eqnarray*} where $C$ depends on $r_A$, $k_A$ and $\beta_3$. Lastly, multiplying the first equation of (\ref{original3}) by $\tilde{H}$, integrating into $\Omega$, using the fact that $A \leq k_A$ and the inequality of Young, we obtain \begin{eqnarray*} \frac{1}{2} \frac{d}{dt} \int_{\Omega} \tilde{H}^2 dx &=& - \xi_H \int_{\Omega} |\nabla \tilde{H}|^2 dx + \nu \int_{\Omega} \tilde{A} \tilde{H} dx - \tau_H \int_{\Omega} \tilde{H}^2 dx \\ &-& \gamma_H \int_{\Omega} N_1 \tilde{H}^2 dx - \gamma_H \int_{\Omega} H_2 \tilde{N} \tilde{H} dx \\ &\leq& \nu \int_{\Omega} |\tilde{A}| |\tilde{H}| dx + \gamma_H ||H_2||_{L^\infty(Q)} \int_{\Omega} |\tilde{N}| |\tilde{H}| dx \\ &\leq& C \int_{\Omega} (\tilde{A}^2 + \tilde{N}^2 + \tilde{H}^2) dx, \end{eqnarray*} where $C$ depends on $\nu$, $\gamma_H$ and $||H_2||_{L^\infty(Q)}$. Thus, \begin{eqnarray*} \frac{d}{dt}\bigg( \int_{\Omega} (|\tilde{N}|^2 + |\tilde{A}|^2 + |\tilde{H}|^2) dx \bigg) &\leq& C \int_{\Omega} (|\tilde{N}|^2 + |\tilde{A}|^2+ |\tilde{D}|^2) dx \end{eqnarray*} and using the Gronwall's inequality, we obtain \begin{equation*} \int_{\Omega} (|\tilde{N}|^2 + |\tilde{A}|^2 + |\tilde{H}|^2) dx \leq e^{CT} \int_{\Omega} (|\tilde{N}_0|^2 + |\tilde{A}_0|^2 + |\tilde{D}_0|^2) dx = 0, \end{equation*} that is, $||\tilde{N}(\cdot, t)||_{L^2(\Omega)}^2 + ||\tilde{A}(\cdot, t)||_{L^2(\Omega)}^2 + ||\tilde{D}(\cdot, t)||_{L^2(\Omega)}^2 = 0$, for all $t \in (0, T)$, where we conclude $\tilde{N} = \tilde{A} = \tilde{D} = 0$ a.e. in $Q$ and therefore $N_1 = N_2, A_1 = A_2$ and $D_1 = D_2$ a.e. in $Q$. \hfill$\Box$ \\ \section{Numerical simulations} In this section, we illustrate model behavior with numerical simulations. The simulations settings are the following. We consider the spatial domain as a square, $\Omega=[0,L] \times [0,L]$, with $L=1$, discritised with steps $\Delta x = \Delta y = 0.01$. The coupled ODEs arising from the spatial discretisation are solved with the method of lines in \textit{Mathematica}. The simulations run from time $t=0$ until $t=50$. To avoid large numbers and numerical instabilities, we rescale the populations with respect to their possible maximum values, setting $N \leftarrow N/(r_N/\mu_N)$ and $A\leftarrow A/K_A$. Therefore, the population sizes range from $0$ to $1$. The parameter values used to simulate the model were \[ r_N=1, \ \mu_N=1, \ \beta_3=1, \ r_A = 1, \ K_A = 1, \ \beta_1 = 1.5, \ \mu_A =0.05, \ \epsilon_A =0.05, \ \] \[ \nu= 2, \ \tau_H=0.9, \ \gamma_H=0.01, \ \alpha_H =2200, \ \xi_H=0.01, \xi_A=0.001. \] These values were chosen to describe: normal cells that reach the equilibrium $N=r_N/\mu_N=1$ at absence of tumor cells; a tumor with the same carrying capacity of normal cells ($K_A=r_N/\mu_N$), and that causes more damage to normal cells than the contrary ($\beta_1>\beta_3$) but is not able to invade the tissue without production of lactic acid (see simulation 1 below); a faster lactic acid diffusion in comparison with tumor cells ($\xi_H>\xi_A$); and a acid damage high enough to allow tumor invasion ($\alpha_H=2200$, see simulations below). The initial conditions for numerical simulations were $N(x,0)=r_N/\mu_N$, \[ A(x,0)=A_0 \exp \left(-\delta_A \left((x_1-L/2)^2+(x_2-L/2)^2\right)\right), \] and $H(x,0)=0$, with $x=(x_1,x_2)\in \Omega \subset \mathbb{R}^2$, $A_0=0.22$, $\delta_A=1000$. These initial condition describe the normal cells at the tissue normal homoeostatic state and the onset of a very small tumor in the middle of the tissue, with zero initial concentration of lactic acid (see Figure 1). We present the following simulation results. In the first simulation (Figure 1), we set the lactic acid production to zero ($\nu=0$), to observe the extinction of tumor cells with no ability to produce lactic acid. In the second simulation, we set the parameters as stated above and observe the invasion of a tumor and substantial reduction of normal cells (Figure 2). In the third simulation, we replace the Neuman homogeneus conditions on tumor cells and lactic acid by Dirichlet homogeneous conditions, representing a situation where the tissue boundary ($\partial \Omega$) is hostile to the tumor (Figure 3). We observe a similar behavior, with exception that in the boundary the normal cells remain at high numbers, due to the low presence of tumor cells and lactic acid. \begin{figure}[!htb] \centering \includegraphics[width=0.99\linewidth]{fig0.pdf} \caption{Simulation results of model (\ref{0riginalEquations}) with no lactic acid production. The plots correspond to transversal sections of model solutions $A(x,t)$ (cancer cells, red) and $N(x,t)$ (normal cells, blue) at $x=(x_1,x_2)=(L/2,x_2)$, $x_2 \in [0,L]$, at time points $t=0,5,30$. The initial tumor cells are extinct, due the negative effect exerted by the normal cells. } \end{figure} \begin{figure}[!htb] \centering \includegraphics[width=0.24\linewidth]{fig6_1.pdf} \includegraphics[width=0.24\linewidth]{fig6_2.pdf} \includegraphics[width=0.24\linewidth]{fig6_3.pdf} \includegraphics[width=0.24\linewidth]{fig6_4.pdf} \includegraphics[width=0.24\linewidth]{fig6_5.pdf} \includegraphics[width=0.24\linewidth]{fig6_6.pdf} \includegraphics[width=0.24\linewidth]{fig6_7.pdf} \includegraphics[width=0.24\linewidth]{fig6_8.pdf} \includegraphics[width=0.24\linewidth]{fig6_9.pdf} \includegraphics[width=0.24\linewidth]{fig6_10.pdf} \includegraphics[width=0.24\linewidth]{fig6_11.pdf} \includegraphics[width=0.24\linewidth]{fig6_12.pdf} \caption{Simulation results of model (\ref{0riginalEquations}). Plots of model solutions $A(x,t)$ (cancer cells, red, top row), $N(x,t)$ (normal cells, blue, middle row) and $H(x,t)$ (lactic acid concentration, green, bottom row) at time points $t=23, 27,31,35$, with $x\in \Omega=[0,L]\times [0,L]$. The initial tumor cells survive, produce lactic acid and invade the tissue, leading the drastic reduction on the number of normal cells. } \end{figure} \begin{figure}[!htb] \centering \includegraphics[width=0.24\linewidth]{fig5_1.pdf} \includegraphics[width=0.24\linewidth]{fig5_2.pdf} \includegraphics[width=0.24\linewidth]{fig5_3.pdf} \includegraphics[width=0.24\linewidth]{fig5_4.pdf} \includegraphics[width=0.24\linewidth]{fig5_5.pdf} \includegraphics[width=0.24\linewidth]{fig5_6.pdf} \includegraphics[width=0.24\linewidth]{fig5_7.pdf} \includegraphics[width=0.24\linewidth]{fig5_8.pdf} \includegraphics[width=0.24\linewidth]{fig5_9.pdf} \includegraphics[width=0.24\linewidth]{fig5_10.pdf} \includegraphics[width=0.24\linewidth]{fig5_11.pdf} \includegraphics[width=0.24\linewidth]{fig5_12.pdf} \caption{Simulation results of model (\ref{0riginalEquations}) with homogeneous Dirichlet boundary conditions. Plots of model solutions $A(x,t)$ (cancer cells, red, top row), $N(x,t)$ (normal cells, blue, middle row) and $H(x,t)$ (lactic acid concentration, green, bottom row) at time points $t=0,33,39,45$, with $x\in \Omega=[0,L]\times [0,L]$. The initial tumor cells survive, produce lactic acid and invade the tissue, leading the drastic reduction on the number of normal cells, except in the boundary, where the normal cells remain at a high number.} \end{figure} \newpage \bibliographystyle{amsplain}

https://arxiv.org/abs/2005.12733

Stein's method of exchangeable pairs in multivariate functional approximations

In this paper we develop a framework for multivariate functional approximation by a suitable Gaussian process via an exchangeable pairs coupling that satisfies a suitable approximate linear regression property, thereby building on work by Barbour (1990) and Kasprzak (2020). We demonstrate the applicability of our results by applying it to joint subgraph counts in an Erdős-Renyi random graph model on the one hand and to vectors of weighted, degenerate $U$-processes on the other hand. As a concrete instance of the latter class of examples, we provide a bound for the functional approximation of a vector of success runs of different lengths by a suitable Gaussian process which, even in the situation of just a single run, would be outside the scope of the existing theory.

\section{Introduction} In his seminal paper \cite{stein}, Charles Stein introduced a method for proving normal approximations and obtained a bound on the speed of convergence to the standard normal distribution. Later, Barbour \cite{barbour} and G\"otze \cite{gotze} developed the so-called generator approach to finding Stein's equation, which made it possible to study approximations by many other probability laws. As a result, in \cite{diffusion}, the method was adapted to approximations by the (infinite-dimensional) Wiener measure. Moreover, the exchangeable-pair approach, first developed by Stein in his monograph \cite{stein1} in the context of univariate normal approximations, has been at the heart of many results proved using Stein's method. It was extended by \cite{RiRo97} and used in the context of non-normal approximations in \cite{chatterjee, roellin1, chatterjee1, dobler, DP18b}. The publication of \cite{meckes, reinert_roellin, meckes09} brought a breakthrough in the understanding of the exchangeable-pair approach and made it available for applications to a wide array of multivariate normal approximation problems. The very recent paper \cite{Doe20} developed a functional analytic approach that provides a substantial extension of the method of exchangeable pairs and, in particular, makes it possible to dispense with the linear regression property in finite-dimensional settings. In \cite{kasprzak3} the method was applied to the study of functional limit results and approximations by univariate Gaussian processes, using the setup of \cite{stein1, RiRo97} and \cite{diffusion}. In this paper we combine the functional approximation of \cite{diffusion} and the multivariate exchangeable-pair method of \cite{reinert_roellin, meckes09}. We obtain an abstract approximation theorem, which is applied in the context of weighted degenerate U-statistics, a particularly interesting example of which are homogeneous sums. The strength of the abstract approximation result is also presented in a random-graph-theoretic application. \subsection{Motivation} We are motivated by examples of multivariate quantities whose distance from the normal distribution can be established using Stein's method of exchangeable pairs, and whose functional equivalents have not been studied yet. Functional limit results play an important role in applied fields. Scaling limits of discrete processes can be studied using stochastic analysis and are often more robust to changes in the local details than the discrete processes themselves. That is why researchers often choose to describe discrete phenomena with continuous models. The error they make by doing this is measured by rates of convergence in functional limit results. The current paper contributes to solving the problem of bounding those rates. The two main applications motivating the paper and considered therein are a continuous Gaussian-process approximation of a rescaled weighted U-statistic and the study of an Erd\H{o}s-Renyi random graph process. U-statistics are central objects in the field of mathematical statistics. Due to their appealing properties, they have found numerous applications to estimation, statistical testing and other problems. They appear in decompositions of more general statistics into sums of terms of a simpler form (see, e.g. \cite[Chapter 6]{serfling} or \cite{rubin_vitale} and \cite{vitale}) and play an important role in the study of random fields (see, e.g. \cite[Chapter 4]{christofides}). Moreover, functional limit theorems for rescaled U-statistics have found applications in the field of changepoint analysis (see e.g. \cite{CH88, Ferger94, GH95, CH_book, Ferger01, Gombay2004, HR_survey, RW19}), where it is particularly useful to know the functional limits of the related test statistics. On the other hand, the Erd\H{o}s-Renyi random graph model has found numerous applications in various fields (see \cite{models}), including epidemic modelling \cite{barbour_mollison} and modelling of evolutionary conflicts \cite{cannings}. The first application discussed in the paper deals with the approximation of so-called weighted $U$-processes, i.e. process analogues of the class of weighted $U$-statistics. This class of processes is very wide, containing the so-called homogeneous sum processes as well as symmetric, degenerate (complete or incomplete) $U$-processes. We derive a general result and successfully apply it to the case of homogeneous sum processes in Subsection \ref{homsums}. As a concrete example, in Subsection \ref{runs}, we provide a bound for a Gaussian approximation of a process that is defined as a vector of success runs of different lengths. For functional limit theorems involving the class of symmetric, degenerate $U$-processes, we refer the reader to the recent paper \cite{DKP}. Moreover, we remark that, even in the univariate case of weighted $U$-statistics, the literature about limit theory for these random quantities is quite restricted. Indeed, apart from the abundance of references on limit theorems for homogeneous sums, the majority of articles focus on the limiting behavior of so-called reduced or incomplete $U$-statistics, i.e. weighted $U$-statistics whose weights only assume the values $0$ and $1$ (see e.g. \cite{Blom, Ja, BrKil}). Limit theorems for general weighted $U$-statistics can be found in references \cite{RiRo97, OnRe, Major}. We stress, however, that the last two references focus on non-normal limiting distributions and that, in the degenerate case, \cite{RiRo97} only considers kernels of order $2$. Moreover, the literature about functional central limit theorems (FCLTs) for weighted $U$-statistics is even scarcer. Indeed, only for homogeneous sum processes \cite{Mik, Basa} have we been able to find comparable results in the literature. We defer a discussion and comparison with our findings to Subsection \ref{homsums}. The second example comes originally from \cite{Janson1991} and was studied using exchangeable pairs in a finite-dimensional context in \cite{reinert_roellin1}. We look at a (dynamic) Erd\H{o}s-Renyi random graph with $\lfloor nt\rfloor$ vertices, where $t$ denotes the time, and study the distance from the asymptotic distribution of the joint law of the number of edges and the number of two-stars. Our approach can, however, be also extended to cover the number of triangles. Those statistics are often used when approximating the clustering coefficient of a network and applied in conditionally uniform graph tests. \subsection{Contribution of the paper} The main achievements of the paper are the following: \begin{enumerate} \item An abstract approximation theorem (Theorem \ref{theorem1}), bounding the distance between a stochastic process $\mathbf{Y}_n$ valued in $\mathbbm{R}^d$, for a fixed positive integer $d$, and a Gaussian mixture process. The estimate is derived under the assumption that that the process $\mathbf{Y}_n$ satisfies the linear regression condition \begin{equation}\label{cond} Df(\mathbf{Y}_n)[\mathbf{Y}_n]=2\mathbbm{E}\left\lbrace \left.\vphantom{\sum}Df(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]\,\right|\,\mathbf{Y}_n\right\rbrace+R_f, \end{equation} for all $f:D\left([0,1],\mathbbm{R}^d\right)\to\mathbbm{R}$ in a certain class of test functions, a random process $\mathbf{Y}_n'$ such that $(\mathbf{Y}_n,\mathbf{Y}_n')$ is an exchangeable pair, some $\Lambda_n\in\mathbbm{R}^{d\times d}$ and some random variable $R_f=R_f(\mathbf{Y}_n)$. In \eqref{cond} (and in the entire paper) $Df$ denotes the Fr\'echet derivative of $f$. The class of test functions, with respect to which the bound in Theorem \ref{theorem1} is obtained, is so rich that the bound approaching zero fast enough implies weak convergence of the law of $\mathbf{Y}_n$ in the Skorokhod and uniform topologies on the Skorokhod space. The exact conditions under which this happens are stated in Proposition \ref{prop_m}. \item A novel framework for continuous Gaussian process approximations of vectors of weighted, degenerate $U$-processes, presented in Section \ref{weighted}. Apart from proving a general result about those, we show how it may be applied in examples involving non-degenerate $U$-processes. In order to study such examples using our theory, one may decompose the given $U$-process into the vector of its degenerate Hoeffding components and prove a multivariate Gaussian limit theorem for this vector. Then, by applying a linear functional, one obtains a Gaussian limit for the original process. This strategy, in a quantified fashion, is exemplified by the application to the $r$-runs process, discussed in Subsection \ref{runs}. We stress that, even in the case of just one $r$-run process, the results about univariate functional approximations via exchangeable pairs from \cite{kasprzak3} would not be sufficient to obtain a Gaussian approximation. Thus, in this example, the multidimensionality of our approach proves to be absolutely vital. Moreover, both the kernels and the coefficients of the weighted $U$-processes we study in our general result may (and will in most cases) depend on the sample size $n$, hence yielding Gaussian limits even in degenerate situations. At the same time, our methods are flexible enough in order to yield bounds for the classical results on asymptotic Gaussianity, in non-degenerate situations, when the kernels are fixed. \item A novel quantitative functional limit theorem for the edge counts and the number of two-stars in an Erd\H{o}s-Renyi random graph $G(n,p)$ on $n$ vertices with fixed edge probability $p$. Letting $I_{i,j}$, for $i,j=1,\cdots,n$ be the indicator that edge $(i,j)$ is present in the graph, we consider the following statistics: \[\mathbf{T}_n(t)=\frac{\lfloor nt\rfloor -2}{2n^2}\sum_{i,j=1}^{\lfloor nt\rfloor}I_{i,j},\quad\mathbf{V}_n(t)=\frac{1}{6n^2}\underset{i,j,k\text{ distinct}}{\sum_{1\leq i,j,k\leq \lfloor nt\rfloor}}I_{ij}I_{jk},\qquad t\in[0,1],\] corresponding to the number of edges and the number of two-stars, respectively. Theorem \ref{theorem_pre_limiting} provides a bound on the distance between the law of the process \begin{equation}\label{graph_ex} t\mapsto \left(\mathbf{T}_n(t)-\mathbbm{E}\mathbf{T}_n(t),\mathbf{V}_n(t)-\mathbbm{E}\mathbf{V}_n(t)\right)\quad t\in[0,1] \end{equation} and the law of a piecewise constant Gaussian process. Theorem \ref{theorem_continuous} estimates the distance between the law of (\ref{graph_ex}) and that of a continuous Gaussian process. These results extend the result of \cite{kasprzak3} bounding the distance between the distribution of the edge counts and a univariate Gaussian process. As a corollary to our results, we immediately obtain weak convergence of the law of (\ref{graph_ex}) in the Skorokhod and uniform topologies on the Skorokhod space to that of the continuous Gaussian process. \end{enumerate} \subsection{Stein's method in its generality} Stein's method in its generality is a powerful technique used to obtain bounds on quantities of the form $|\mathbbm{E}_{\nu}h-\mathbbm{E}_{\mu}h|$, where $\mu$ is the target (known) distribution, $\nu$ is an approximating measure and $h$ is a real-valued test function chosen from a suitable class $\mathcal{H}$. The method is composed out of three main steps. First, one needs to find an operator $\mathcal{A}$ acting on a class of real-valued functions, such that \[\left(\forall f\in\text{Domain}(\mathcal{A})\quad\mathbbm{E}_{\pi}\mathcal{A}f=0\right)\quad \Longleftrightarrow \quad\pi=\mu.\] Second, for a given function $h\in\mathcal{H}$, one solves the following Stein equation: \[\mathcal{A}f=h-\mathbbm{E}_{\mu}h.\] Finally, for $f=f_h$ solving the Stein equation, the following quantity: \begin{equation}\label{quantity} |\mathbbm{E}_{\nu}\mathcal{A}f_h| \end{equation} needs to be bounded. This is achieved using various mathematical tools (Taylor's expansions, Malliavin calculus, as described in \cite{nourdin}, coupling methods and others), applied in conjunction with smoothness properties of $f_h$. For an accessible account of the method we recommend the surveys \cite{reinert} and \cite{ross} as well as the books \cite{janson} and \cite{normal_approx}, which treat the cases of Poisson and normal approximation, respectively, in detail. A database of information and publications connected to Stein's method can also be found in \cite{swan}. \subsection{Stein's method of exchangeable pairs} The exchangeable-pair approach to Stein's method was first developed in \cite{stein1}. Therein, the author considered the setup in which, for a random variable $W$, one can construct another random variable $W'$ such that $(W,W')$ is an exchangeable pair and the following linear regression condition is satisfied \begin{equation}\label{regression_condition} \mathbbm{E}\left[W'-W|W\right]=-\lambda W \end{equation} for some $\lambda>0$. It follows from this assumption that \begin{align*} 0=&\mathbbm{E}\left[(f(W)+f(W'))(W-W')\right] =\mathbbm{E}\left[(f(W')-f(W))(W-W')\right]+2\lambda\mathbbm{E}[Wf(W)] \end{align*} and so \[\mathbbm{E}[Wf(W)]=\frac{1}{2\lambda}\mathbbm{E}\left[(f(W)-f(W'))(W-W')\right].\] Therefore, using Taylor's theorem, it can be proved that \begin{align*} &\left|\mathbbm{E}[f'(W)]-\mathbbm{E}[Wf(W)]\right| \leq\frac{\|f'\|_{\infty}}{2\lambda}\sqrt{\text{Var}\left[\mathbbm{E}\left[(W-W')^2|W\right]\right]}+\frac{\|f''\|_{\infty}}{2\lambda}\mathbbm{E}|W-W'|^3, \end{align*} which provides a bound on the quantity \eqref{quantity} for $\nu=\mathcal{L}(W)$ and $\mathcal{A}$ being the canonical Stein operator corresponding to the standard normal law. A multivariate version of the method was first described in \cite{meckes} and then in \cite{reinert_roellin}. In \cite{reinert_roellin}, for an exchangeable pair of $d$-dimensional vectors $(W,W')$, the following condition is used: \begin{equation}\label{lambda_matrix} \mathbbm{E}[W'-W|W]=-\Lambda W+R \end{equation} for some invertible matrix $\Lambda$ and a remainder term $R$. The approach of \cite{reinert_roellin} was further reinterpreted and combined with the approach of \cite{meckes} in \cite{meckes09}. Extending this multivariate version of the exchangeable-pair method to multivariate functional approximations, with the linear regression condition taking form similar to \eqref{lambda_matrix}, is the subject of the current paper. \subsection{Functional Stein's method} Approximations by laws of stochastic processes using Stein's method have been studied in \cite{diffusion, functional_combinatorial, shih, Coutin, decreusefond_higher} and recently in \cite{kasprzak1, kasprzak2, kasprzak3, decreusefond2, decreusefond_rough, campese}. These references can be divided into three groups. The ones belonging to the first group, containing \cite{diffusion, functional_combinatorial, kasprzak1, kasprzak2, kasprzak3}, all use, adapt and extend the setup of \cite{diffusion}. Therein, the author studied the rate of convergence in the celebrated functional central limit theorem, also called Donsker's theorem. Barbour considered test functions $g$ acting on the Skorokhod space $D \left( [0,1], \mathbbm{{R}} \right)$ of c\`adl\`ag real-valued maps on $[0,1]$, such that $g$ takes values in the reals, does not grow faster than a cubic, is twice Fr\'echet differentiable and its second derivative is Lipschitz. For each function $g$ belonging to this class he provided a bound on the absolute difference between the expectation of $g$ with respect to the law of a rescaled random walk and the expectation of $g$ with respect to the Wiener measure. Crucially, he also proved that this class of functions $g$ is so rich that his bounds imply weak convergence with respect to the Skorokhod topology of the considered rescaled random walk to Brownian Motion. This last property is vital for most applications of the limit theory for stochastic processes and may even be the main reason for the outstanding popularity of the Skorohod topology. Indeed, by means of the continuous mapping theorem, limit theorems for many natural, non-linear functionals such as the supremum over time, immediately follow from a weak limit theorem in the Skorokhod topology. On the other hand, the results of the second group of references, containing \cite{Coutin, decreusefond_higher, decreusefond2, decreusefond_rough}, develop Stein's theory on a Hilbert space using a Besov-type topology. The bounds obtained therein, however, do not imply weak convergence in the Skorokhod topology. Therefore, the continuous mapping theorem does not apply in their setting. For instance, as opposed to the results of the first group of references, one cannot study convergence of the supremum of a process using the analysis of the second group of papers. Finally, \cite{shih} develops approximations by abstract Wiener measures on a real separable Banach space and \cite{campese} proves bounds on measure-determining distances from Gaussian random variables valued in Hilbert spaces. As for the second group, despite the elegant abstract theory used and developed in these references, the results do not imply convergence in the Skorokhod topology on $D[0,1]$. In the current paper we shall follow the setup of the first group of references. We consider it more flexible than the one of the second group and more suited for applications to processes belonging to the widely-used (non-separable) Skorokhod space than the ones of the third group. In the context of these three groups of references and the present paper, we also mention the recent paper \cite{DKP} which, although not relying on functional approximation by Stein's method, provides functional limit theorems for the class of (degenerate and non-degenerate) symmetric $U$-processes with a kernel that may depend on the sample size $n$. Since it implicitly relies on a multivariate Gaussian limit theorem derived by Stein's method from \cite{DP19}, it is also naturally related to Stein's method. Moreover, since one main class of applications in the present paper involves weighted $U$-processes, it is worthwhile to compare our results and their applicability to those of \cite{DKP}. Firstly, as mentioned above, the paper \cite{DKP} focuses on Gaussian limit theorems for symmetric $U$-processes, which constitute a narrower class than the weighted $U$-processes considered in the present work. Moreover, thanks to the finite-dimensional convergence results from \cite{DP19}, the conditions for convergence from \cite{DKP} are phrased in term of $L^2$-norms of \textit{contraction kernels} and, as such, can be considered as fourth moment conditions. In contrast, as can be seen from the bounds and proofs of Section \ref{weighted}, the bounds and conditions in the present paper involve third moment quantities. This distinction is also clearly reflected in the respective applicability of the results proved in the present paper and those from \cite{DKP}. Indeed, whereas the symmetric $U$-processes considered in \cite{DKP} possess a global dependency structure, the results in Section \ref{weighted} are most useful whenever the dependence of the weighted $U$-process is local in the sense that the involved array of weighting coefficients $(a_J)_J$ is sparse in some sense. The runs example in Subsection \ref{runs} provides an instructive showcase for this observation. Moreover, the methods used in the proofs of the main results necessitate that the quantities in the bounds involve the absolute values of both the kernels and the coefficients. Hence, no cancellation effect, typically occuring under fourth moment conditions, may be relied on in this case. We therefore consider our theorems as rather complementary to the ones in \cite{DKP}. \subsection{Structure of the paper} Section \ref{section_notation} includes some introductory remarks about notation and the spaces of test functions with respect to which bounds on distances between probability laws in this paper will be derived. Section \ref{section_setup_stein} gives a general form of the pre-limiting process to which all the processes of interest will be compared using Stein's method. It also presents the corresponding Stein equation, its solution and the smoothness properties of the solution. Section \ref{section_abstract} contains the main abstract result of this paper providing a bound on the distance between a process valued in the Skorokhod space $D([0,1],\mathbbm{R}^d)$ and the pre-limiting process described in the previous section. Section \ref{weighted} discusses the application of the abstract theorem to weighted, degenerate U-statistics and presents a bound on their distance from a continuous Gaussian process. It furthermore explains how the bound simplifies in the context of homogeneous sums and applies it to the example of $r$-runs on the line. Section \ref{section6} discusses the example concerning an Erd\H{o}s-Renyi random graph process and the bound on the distance between the number of its edges and two-stars and a continuous Gaussian process. Technical details of some of the proofs in this paper are postponed to Section \ref{appendix}. \section{Notation and spaces $M$ and $M^0$}\label{section_notation} The following notation, similar to the one of \cite{diffusion} and \cite{kasprzak2}, is used throughout the paper. For a fixed positive integer $d$, let $D([0,1],\mathbbm{R}^d)$ be the Skorokhod space of c\`adl\`ag $\mathbbm{R}^d$-valued functions on $[0,1]$. For $i=1,\cdots,d$, by $e_i$ we denote the $i$th unit vector of the canonical basis of $\mathbbm{R}^d$. The $i$th component of any $x\in\mathbbm{R}^d$ will be denoted by $x^{(i)}$, so that $x=\left(x^{(1)},\cdots,x^{(d)}\right)$. For a function $w$ defined on $[0,1]$ and taking values in a Euclidean space, we will also write \[\|w\|=\sup_{t\in[0,1]}|w(t)|,\] where $|\cdot|$ denotes the Euclidean norm. Moreover, the notation $\mathbbm{E}^W[\,\cdot\,]$ will be used to represent $\mathbbm{E}[\,\cdot\,|W]$. Furthermore, we define \[\|f\|_L:=\sup_{w\in D([0,1],\mathbbm{R}^d)}\frac{|f(w)|}{1+\|w\|^3}\text{,}\] and let $L$ be the Banach space of continuous functions $f:D([0,1],\mathbbm{R}^d)\to\mathbbm{R}$ such that $\|f\|_L<\infty$. By $D^kf$ we will always mean the $k$-th Fr\'echet derivative of $f$. The norm $\|\cdot\|$ of a $k$-linear form $B$ on $L$ will be taken to be \[\|B\|=\sup_{\lbrace h:\|h_i\|\leq 1\,\forall i=1,\dots k\rbrace} |B[h_1,...,h_k]|,\] where $B[h_1,\dots,h_k]$ denotes $B$ applied to arguments $h_1,\dots,h_k\in L$. As in \cite{diffusion}, we define $M\subset L$ as a subspace of $L$ consisting of the twice Fr\'echet differentiable functions $f$, such that: \begin{equation}\label{space_m} \|D^2f(w+h)-D^2f(w)\|\leq k_f\|h\|\text{,} \end{equation} for some constant $k_f$, uniformly in $w,h\in D([0,1],\mathbbm{R}^d)$. We have following lemma (whose proof we omit), which may be proved in an analogous way to that used to show (2.6) and (2.7) of \cite{diffusion}: \begin{lemma}\label{first_der} For every $f\in M$, let: \begin{align*} \|f\|_M:=&\sup_{w\in D([0,1],\mathbbm{R}^d)}\frac{|f(w)|}{1+\|w\|^3}+\sup_{w\in D([0,1],\mathbbm{R}^d)}\frac{\|Df(w)\|}{1+\|w\|^2}+\sup_{w\in D([0,1],\mathbbm{R}^d)}\frac{\|D^2f(w)\|}{1+\|w\|}\\ &+\sup_{w,h\in D([0,1],\mathbbm{R}^d)}\frac{\|D^2f(w+h)-D^2f(w)\|}{\|h\|}. \end{align*} Then, for all $f\in M$, we have $\|f\|_M<\infty$. \end{lemma} We, furthermore, let $M^0$ be the class of functionals $g\in M$ such that: \begin{align} \|g\|_{M^0}:=&\sup_{w\in D([0,1],\mathbbm{R}^d)}|g(w)|+\sup_{w\in D([0,1],\mathbbm{R}^d)}\|Dg(w)\|+\sup_{w\in D([0,1],\mathbbm{R}^d)}\|D^2g(w)\|\nonumber\\ &+\sup_{w,h\in D([0,1],\mathbbm{R}^d)}\frac{\|D^2g(w+h)-D^2g(w)\|}{\|h\|}<\infty\nonumber \end{align} and note that $M^0\subset M$. Below, we present a $d$-dimensional version of \cite[Proposition 3.1]{functional_combinatorial} providing conditions, under which weak convergence of the approximating measure to the target one may be deduced from convergence of the corresponding expectations of functions $g\in M^0$. Its proof can be found in the appendix of \cite{kasprzak2}. \begin{proposition}\label{prop_m} Suppose that, for each $n\geq 1$, the random element $\mathbf{Y}_n$ of $D([0,1],\mathbbm{R}^d)$ is piecewise constant with intervals of constancy of length at least $r_n$. Let $\left(\mathbf{Z}_n\right)_{n\geq 1}$ be random elements of $D^p$ converging in distribution in $D([0,1],\mathbbm{R}^d)$, with respect to the Skorokhod topology, to a random element $\mathbf{\mathbf{Z}}\in C\left([0,1],\mathbbm{R}^d\right)$. If: \begin{equation}\label{assumption} |\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{\mathbf{Z}}_n)|\leq C\mathscr{T}_n\|g\|_{M^0} \end{equation} for each $g\in M^0$ and if $\mathscr{T}_n\log^2(1/r_n)\xrightarrow{n\to\infty}0$, then the law of $\mathbf{Y}_n$ converges weakly to that of $\mathbf{\mathbf{Z}}$ in $D([0,1],\mathbbm{R}^d)$, in both the uniform and the Skorokhod topologies. \end{proposition} \section{Setting up Stein's method for the pre-limiting approximation}\label{section_setup_stein} We set up Stein's method in a fashion similar to \cite{diffusion} and \cite{kasprzak2}. First, we define the process $\mathbf{D}_n$ whose distribution will be treated as the target measure. We then construct a process $\left(\mathbf{W}_n(\cdot,u):u\geq 0\right)$ for which the target measure is stationary. We subsequently calculate its infinitesimal generator $\mathcal{A}_n$ and take it as our Stein operator. Next, we solve the Stein equation $\mathcal{A}_nf=g$, using the analysis of \cite{kasprzak}, and prove several smoothness properties of the solution $f_n=\phi_n(g)$. \subsection{Target measure} Let \begin{equation}\label{d_n} \mathbf{D}_n(t)=\sum_{i_1,\cdots,i_m=1}^{n}\left(\tilde{Z}^{(1)}_{i_1,\cdots,i_m}J^{(1)}_{i_1,\cdots,i_m}(t),\cdots,\tilde{Z}^{(d)}_{i_1,\cdots,i_m}J^{(d)}_{i_1,\cdots,i_m}(t)\right),\quad t\in[0,1], \end{equation} where $\tilde{Z}^{(k)}_{i_1,\cdots,i_m}$'s for $k=1,\cdots,d$ are centred Gaussian and: \begin{enumerate} \item the covariance matrix $\Sigma_n\in\mathbbm{R}^{(n^md)\times(n^md)}$ of $\tilde{Z}$ is positive definite, for $\tilde{Z}\in \mathbbm{R}^{(n^md)}$ built out of the $\tilde{Z}^{(k)}_{i_1,\cdots,i_m}$'s in such a way that they appear in the lexicographic order with $\tilde{Z}^{(k)}_{i_1,\cdots,i_m}$ appearing before $\tilde{Z}^{(k+1)}_{j_1,\cdots,j_m}$'s for any $k=1,\cdots,d-1$ and $i_1,\cdots,i_m,j_1,\cdots,j_m=1,\cdots,n$; \item the collection of functions \[\left\lbrace J^{(k)}_{i_1,\cdots,i_m}\in D\left([0,1],\mathbbm{R}\right)\,:\, i_1,\cdots,i_m\in\lbrace 1,\cdots, n\rbrace, k\in\lbrace 1,\cdots,p\rbrace\right\rbrace\] is independent of the collection $\left\lbrace \tilde{Z}^{(k)}_{i_1,\cdots,i_m}\,:\, i_1,\cdots,i_m\in\lbrace 1,\cdots, n\rbrace, k\in\lbrace 1,\cdots,p\rbrace\right\rbrace$; a natural example of those would be $J^{(k)}_{i_1,\cdots,i_m}=\mathbbm{1}_{A^{(k)}_{i_1,\cdots,i_m}}$ for some measurable set $A^{(k)}_{i_1,\cdots,i_m}\subset[0,1]$. \end{enumerate} \begin{remark} It is worth noting that processes $\mathbf{D}_n$ of the form (\ref{d_n}) are often approximations of interesting continuous Gaussian processes. An example is $\mathbf{D}_n$ of (\ref{d_n}), where all the $\tilde{Z}^{(k)}_{i_1,\cdots,i_m}$'s are standard normal and independent, $m=1$ and $J_i^{(k)}=\mathbbm{1}_{[i/n,1]}$ for all $k=1,\cdots,d$ and $i=1,\cdots,n$. By Donsker's theorem, it approximates the standard Brownian motion. By Proposition \ref{prop_m}, under several assumptions, if a piecewise constant process $\mathbf{Y}_n$ is close enough to process $\mathbf{D}_n$, then the law of $\mathbf{Y}_n$ converges weakly to that of the continuous process that $\mathbf{D}_n$ approximates. \end{remark} Now consider an array of i.i.d. Ornstein-Uhlenbeck processes with stationary law $\mathcal{N}(0,1)$, independent of the $J^{(k)}_{i_1,\cdots,i_m}$'s, given by $\lbrace (\mathscr{X}^{(k)}_{i_1,\cdots,i_m}(u),u\geq 0):i_1,\cdots,i_m=1,...,n,\,k=1,...,d\rbrace$. Let $\tilde{\mathscr{U}}(u)=\left(\Sigma_n\right)^{1/2}\mathscr{X}(u)$, where $\Sigma_n$ is the covariance matrix of $\tilde{Z}$, as above, and $\mathscr{X}(u)\in\mathbbm{R}^{n^md}$ is a vector composed out of the $\mathscr{X}^{(k)}_{i_1,\cdots,i_m}(u)$'s in such a way that they are ordered exactly as $\tilde{Z}^{(k)}_{i_1,\cdots,i_m}$'s are ordered in $\tilde{Z}$. Write $\mathscr{U}_{i_1,\cdots,i_m}^{(k)}(u)=\left(\tilde{\mathscr{U}}(u)\right)_{I(k,i_1,\cdots,i_m)}$ using the bijection $I:\lbrace (k,i_1,\cdots,i_m):i_1,\cdots,i_m=1,\cdots, n, k=1,\cdots, d\rbrace\to\lbrace 1,\cdots,dn^m\rbrace$, given by: \begin{equation}\label{i} I(k,i_1,\cdots,i_m)=(k-1)n^m+(i_1-1)n^{m-1}+\cdots+(i_{m-1}-1)n+i_m. \end{equation} We will look at the process \[\mathbf{W}_n(t,u)=\left(\mathbf{W}_n^{(1)}(t,u),\cdots,\mathbf{W}_n^{(d)}(t,u)\right),\quad t\in[0,1],u\geq 0,\] where, for all $k=1,\cdots,d$: \[\mathbf{W}_n^{(k)}(t,u)=\sum_{i_1,\cdots,i_m=1}^{n}\mathscr{U}^{(k)}_{i_1,\cdots,i_m}(u)J^{(k)}_{i_1,\cdots,i_m}(t),\quad t\in[0,1],u\geq 0.\] It is easy to see that the stationary law of the process $\left(\mathbf{W}_n(\cdot,u)\right)_{u\geq 0}$ (which, for any fixed $u$, takes value in $D([0,1],\mathbbm{R}^d)$) is exactly the law of $\mathbf{D}_n$. \subsection{Stein equation} The following result follows immediately from \cite[Propositions 4.1 and 4.4]{kasprzak}: \begin{proposition}\label{prop12.7} The infinitesimal generator of the process $\left(\mathbf{W}_n(\cdot,u)\right)_{u\geq 0}$ acts on any $f\in M$ (for $M$ defined in Section \ref{section_notation}) in the following way: \begin{align*} &\mathcal{A}_nf(w)=-Df(w)[w]+\mathbbm{E}D^2f(w)\left[\mathbf{D}_n,\mathbf{D}_n\right]. \end{align*} Moreover, for any $g\in M$ such that $\mathbbm{E}g(\mathbf{D}_n)=0$, the Stein equation $\mathcal{A}_nf_n=g$ is solved by: \begin{equation}\label{phi} f_n=\phi_n(g)=-\int_0^{\infty}T_{n,u}gdu, \end{equation} where $(T_{n,u}f)(w)=\mathbbm{E}\left[f(we^{-u}+\sqrt{1-e^{-2u}}\mathbf{D}_n(\cdot)\right]$. Furthermore, for $g\in M$: \begin{align} \text{A)} \quad &\|D\phi_n(g)(w)\|\leq \|g\|_{ M}\left(1+\frac{2}{3}\|w\|^2+\frac{4}{3}\mathbbm{E}\|\mathbf{D}_n\|^2\right)\text{,}\nonumber\\ \text{B)} \quad &\|D^2\phi_n(g)(w)\|\leq \|g\|_{ M}\left(\frac{1}{2}+\frac{\|w\|}{3}+\frac{\mathbbm{E}\|\mathbf{D}_n\|}{3}\right)\text{,}\nonumber\\ \text{C)}\quad&\frac{\left\|D^2\phi_n(g)(w+h)-D^2\phi_n(g)(w)\right\|}{\|h\|}\nonumber\\ \leq&\sup_{w,h\in D^p}\frac{\|D^2(g+c)(w+h)-D^2(g+c)(w)\|}{3\|h\|},\label{m_bound} \end{align} for any constant function $c:D([0,1],\mathbbm{R}^d)\to\mathbbm{R}$ and for all $w,h\in D([0,1],\mathbbm{R}^d)$. \begin{remark} The fact that the process $\left(\mathbf{W}_n(\cdot,u)\right)_{u\geq 0}$ is built using Ornstein-Uhlenbeck processes and that the corresponding semigroup $T_{n,u}$ takes the convenient form, coming from Mehler's formula, plays an important role in the proof of Proposition \ref{prop12.7}. It is not clear to us whether this result can easily be extended beyond this context. \end{remark} \end{proposition} \section{An abstract approximation theorem}\label{section_abstract} The following result provides an expression for a bound on the distance between a process $\mathbf{Y}_n$ and $\mathbf{D}_n$, defined by (\ref{d_n}). It assumes that we can find some $\mathbf{Y}_n'$ such that $(\mathbf{Y}_n,\mathbf{Y}_n')$ is an exchangeable pair satisfying an appropriate condition. We explain in Remark \ref{remark_th_1} how our condition is similar to that of \cite[(1.7)]{reinert_roellin}. \begin{theorem}\label{theorem1} Assume that $(\mathbf{Y}_n,\mathbf{Y}_n')$ is an exchangeable pair of $D\left([0,1],\mathbbm{R}^d\right)$-valued random vectors such that: \begin{equation} Df(\mathbf{Y}_n)[\mathbf{Y}_n]=2\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]+R_f, \label{condition} \end{equation} where $\mathbbm{E}^{\mathbf{Y}_n}[\cdot]:=\mathbbm{E}\left[\cdot|\mathbf{Y}_n\right]$, for all $f\in M$, some $\Lambda_n\in\mathbbm{R}^{d\times d}$ and some random variable $R_f=R_f(\mathbf{Y}_n)$. Let $\mathbf{D}_n$ be defined by (\ref{d_n}). Then, for any $g\in M$: \begin{align*} \left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{D}_n)\right|\leq \epsilon_1+\epsilon_2+\epsilon_3, \end{align*} where \begin{align*} \epsilon_1&=\frac{\|g\|_M}{6}\mathbbm{E}\left[\|(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\|\|\mathbf{Y}_n-\mathbf{Y}_n'\|^2\right],\\ \epsilon_2&=\left|\mathbbm{E}D^2f(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]\right|,\\ \epsilon_3&=|\mathbbm{E}R_f|, \end{align*} and $f=\phi_n(g)$, as defined by (\ref{phi}). \end{theorem} \begin{remark}[Relevance of terms in the bound] Term $\epsilon_1$ measures how close $\mathbf{Y}_n$ and $\mathbf{Y}_n'$ are and how \textit{small} (in a certain sense) $\Lambda_n$ is. Term $\epsilon_2$ quantifies the difference between the covariance structures of $\mathbf{Y}_n-\mathbf{Y}_n'$ and $\mathbf{D}_n$. This term may be estimated in several applications (see Theorems \ref{theorem_weighted_pre} and \ref{theorem_pre_limiting} below), yet this often requires some effort. Term $\epsilon_3$ measures the error in the exchangeable-pair linear regression condition (\ref{condition}). \end{remark} \begin{remark} Condition (\ref{condition}) is always satisfied, for example with $\Lambda_n=0$ and $R_f=Df(\mathbf{Y}_n)[\mathbf{Y}_n]$ for all $f\in M$. However, for the bound in Theorem \ref{theorem1} to be small, we require the expectation of $R_f$ to be small in absolute value. \end{remark} \begin{remark} The term \[\left|\mathbbm{E}D^2f(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]\right|\] in the bound obtained in Theorem \ref{theorem1} is an analogue of the second condition in \cite[Theorem 3]{meckes09}. The main result of that paper provides a bound on approximation by $\mathcal{N}(0,\Sigma)$ of a $d$-dimensional vector $X$. This is achieved by constructing an exchangeable pair $(X,X')$ satisfying: \[\mathbbm{E}^X[X'-X]=\Lambda X+E\quad\text{and}\quad\mathbbm{E}^X[(X'-X)(X'-X)^T]=2\Lambda\Sigma+E'\] for some invertible matrix $\Lambda$ and some remainder terms $E$ and $E'$. In the same spirit, Theorem \ref{theorem1} could be rewritten to assume (\ref{condition}) and: \[ \mathbbm{E}^{\mathbf{Y}_n}D^2f(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]=D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]+R^1_f,\] for all $f\in M$. The bound would then take the form: \begin{align*} \left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{D}_n)\right|\leq&\frac{\|g\|_M}{6}\mathbbm{E}\left[\|(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\|\|\mathbf{Y}_n-\mathbf{Y}_n'\|^2\right]+|\mathbbm{E}R_f|+|\mathbbm{E}R^1_f|, \end{align*} for $f=\phi_n(g)$. \end{remark} \begin{remark}\label{remark_th_1} The role of $\Lambda_n$ in condition (\ref{condition}) is equivalent to that played by $\Lambda^{-1}$ in \cite{reinert_roellin} for $\Lambda$ defined by (1.7) therein. In the functional setting, condition (\ref{condition}) is more appropriate than a straightforward adaptation of the setup of \cite{reinert_roellin}. This is because, for general processes $\mathbf{Y}_n$, the properties of the Fr\'echet derivative do not allow us to treat evaluating the derivative in the direction of $\mathbf{Y}_n-\mathbf{Y}_n'$ as matrix multiplication. Indeed, multiplying both sides of the hypothetical condition: \[-Df(\mathbf{Y}_n)[\Lambda \mathbf{Y}_n]=\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)[\mathbf{Y}_n-\mathbf{Y}_n']\] by $\Lambda^{-1}$ does not yield: \[-Df(\mathbf{Y}_n)[\mathbf{Y}_n]=\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)[\Lambda^{-1}(\mathbf{Y}_n-\mathbf{Y}_n')].\] \end{remark} \begin{proof}[Proof of Theorem \ref{theorem1}] We will bound $\left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{D}_n)\right|$ by bounding $\left|\mathbbm{E}\mathcal{A}_nf(\mathbf{Y}_n)\right|$, where $f$ is the solution to the Stein equation: \[\mathcal{A}_nf=g-\mathbbm{E}g(\mathbf{D}_n),\] for $\mathcal{A}_n$ defined in Proposition \ref{prop12.7}. Note that, by exchangeability of $(\mathbf{Y}_n,\mathbf{Y}_n')$ and (\ref{condition}): \begin{align*} 0=&\mathbbm{E}\left(Df(\mathbf{Y}_n')+Df(\mathbf{Y}_n)\right)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]\\ =&\mathbbm{E}\left(Df(\mathbf{Y}_n')-Df(\mathbf{Y}_n)\right)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]+2\mathbbm{E}\left\lbrace\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]\right\rbrace\\ =&\mathbbm{E}\left(Df(\mathbf{Y}_n')-Df(\mathbf{Y}_n)\right)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]+\mathbbm{E}Df(\mathbf{Y}_n)[\mathbf{Y}_n]-\mathbbm{E}R_f \end{align*} and so: \begin{equation*} \mathbbm{E}Df(\mathbf{Y}_n)[\mathbf{Y}_n]=\mathbbm{E}\left(Df(\mathbf{Y}_n)-Df(\mathbf{Y}_n')\right)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]+\mathbbm{E}R_f. \end{equation*} Therefore: \begin{align} &\left|\mathbbm{E}\mathcal{A}_nf(\mathbf{Y}_n)\right|\notag\\ =&\left|\mathbbm{E}Df(\mathbf{Y}_n)[\mathbf{Y}_n]-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]\right|\nonumber\\ =&\left|\mathbbm{E}\left(Df(\mathbf{Y}_n)-Df(\mathbf{Y}_n')\right)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]+\mathbbm{E}R_f\right|\nonumber\\ \leq& \left|\mathbbm{E}\left(Df(\mathbf{Y}_n)-Df(\mathbf{Y}_n')\right)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right]-\mathbbm{E}D^2f(\mathbf{Y}_n')\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]\right|\nonumber\\ &+\left|\mathbbm{E}D^2f(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]\right|+|\mathbbm{E}R_f|\nonumber\\ \leq&\frac{\|g\|_M}{6}\mathbbm{E}\left[\|(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\|\|\mathbf{Y}_n-\mathbf{Y}_n'\|^2\right]+|\mathbbm{E}R_f|\nonumber\\ &+\left|\mathbbm{E}D^2f(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]\right|,\nonumber \end{align} where the last inequality follows by Taylor's theorem and Proposition \ref{prop12.7}. \end{proof} \section{Weighted, degenerate $U$-statistics}\label{weighted} In this Section we will apply Theorem \ref{theorem1} in order to prove bounds for the approximation of a vector of weighted, degenerate $U$-processes by suitable Gaussian processes. \subsection{Introduction}\label{intro_weighted} The setup will be the following. We fix positive integers $d,p_1,\dotsc,p_d$ and consider a sequence $(X_i)_{i\in\mathbbm{N}}$ of i.i.d. random variables with distribution $\mu$ on some measurable space $(E,\mathcal{E})$. Moreover, for $1\leq i\leq d$, we let $\psi(i)\in L^2(\mu^{p_i})$ be a symmetric kernel such that $\mathbbm{E}[\psi(i)^2(X_1,\dotsc,X_{p_i})]>0$. We assume that $\psi(i)$ is \textbf{(completely) degenerate} with respect to $\mu$, i.e. that \[ \mathbbm{E}[\left.\psi(i)(X_1,\dotsc,X_{p_i})\,\right|\,X_1,\dotsc,X_{p_i-1}]=0,\quad\text{a.s.} \] We denote by $\mathcal{D}_p(n)$ the collection of $p$-subsets of the set $[n]:=\{1,\dotsc,n\}$ (if $p>n$, we set $\mathcal{D}_p(n)=\emptyset$). Furthermore, we fix an integer $n\geq \max(p_1,\dotsc,p_d)$ and let $\lbrace a_J(i):\, 1\leq i\leq d,\, J\in\mathcal{D}_{p_i}(n)\rbrace$, be a (given) set of real numbers (weights). We further let $\lbrace\sigma_n(i):\,1\leq i\leq d\rbrace$ be a set of positive real numbers and, for $t\in[0,1]$, define \[\mathbf{Y}_n^{(i)}(t):=\frac{1}{\sigma_n(i)}\sum_{J\in\mathcal{D}_{p_i}(\lfloor nt\rfloor)} a_J(i)\psi(i)(X_j,j\in J)\,. \] In some applications it may be natural to take \[\sigma_n(i)^2=\mathbbm{E}[\psi(i)^2(X_1,\dotsc,X_{p_i})]\sum_{J\in\mathcal{D}_{p_i}(n)}a_J(i)^2\,,\quad 1\leq i\leq d,\] i.e. equal to the variance of the sum in the definition of $\mathbf{Y}_n^{(i)}(1)$. This is, however, not necessary for our results. For fixed $t$ (in particular for $t=1$), the quantity $\mathbf{Y}_n^{(i)}(t)$ is customarily referred to as a \textbf{degenerate, weighted $U$-statistic} based on $X_1,\dotsc, X_{\lfloor nt\rfloor}$ and, thus, we coin the whole random function $\mathbf{Y}_n^{(i)}$ a \textbf{degenerate, weighted $U$-process} . Limit theorems (not necessarily central) for such weighted $U$-statistics have been derived in \cite{OnRe, RiRo97, Major, RiUt} and in the (somehow) more special case of incomplete $U$-statistics in \cite{Ja, Blom, BrKil}. However, we have not been able to find FCLTs for \textbf{degenerate, weighted $U$-process} in the literature. With the above definitions, we let \[\mathbf{Y}_n:=(\mathbf{Y}_n^{(1)},\dotsc,\mathbf{Y}_n^{(d)})\,,\] which is, as one can easily observe, an element of $D([0,1],\mathbbm{R}^d)$. We will write $X:=(X_1,\dotsc,X_n)$ and construct an $X':=(X_1',\dotsc,X_n')$ such that the pair $(X,X')$ is exchangeable. Specifically, we let $X_0$ be another random variable with distribution $\mu$ and let $I$ be uniformly distributed on $[n]$ in such a way that $I,X_0,(X_j)_{j\in\mathbbm{N}}$ are jointly independent. For $1\leq j\leq n$, we let \begin{equation*} X_j':=\begin{cases} X_j\,,&\text{if }j\not=I\\ X_0\,,&\text{if }j=I\,. \end{cases} \end{equation*} Then, for $t\in[0,1]$ and $1\leq i\leq d$, we define \[(\mathbf{Y}_n^{(i)})'(t):=\frac{1}{\sigma_n(i)}\sum_{J\in\mathcal{D}_{p_i}(\lfloor nt\rfloor)} a_J(i)\psi(i)(X'_j,j\in J)\] and \[\mathbf{Y}'_n:=((\mathbf{Y}_n^{(1)})',\dotsc,(\mathbf{Y}_n^{(d)})')\,.\] The pair $(\mathbf{Y}_n,\mathbf{Y}'_n)$ is clearly exchangeable and, for $f\in M$, similarly as in the proof of \cite[Lemma 2.3]{DP16}, one can use degeneracy to show that \begin{equation*} Df(\mathbf{Y}_n)[\mathbf{Y}_n]=2\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right], \end{equation*} where \begin{equation}\label{lambda_weighted} \Lambda_n=\text{diag}\left(\frac{n}{2p_1},\dots,\frac{n}{2p_d}\right). \end{equation} Therefore condition (\ref{condition}) is satisfied for $\Lambda_n$ of (\ref{lambda_weighted}) and $R_f=0$. In what follows we will assume that $1\leq p_1\leq p_2\leq\cdots\leq p_d$. \subsection{A pre-limiting process}\label{pre_lim_weighted} We will construct a pre-limiting Gaussian process $\mathbf{D}_n$ of the form (\ref{d_n}) which has the same covariance structure as $\mathbf{Y}_n$. We take $\mathbf{D}_n=\left(\mathbf{D}_n^{(1)},\dots,\mathbf{D}_n^{(d)}\right)$ for \begin{equation*} \mathbf{D}_n^{(i)}(t)=\frac{1}{\sigma_n(i)}\sum_{J\in\mathcal{D}_{p_i}(\lfloor nt\rfloor)} a_J(i)Z_J(i), \end{equation*} where, for $i=1,\dots,d$ and $J\in\mathcal{D}_{p_i}(n)$, $Z_J(i)$ are jointly Gaussian random variables that are independent of $X$ and satisfy \begin{equation*} \mathbbm{E}\left[Z_J(i)Z_K(l)\right]=\begin{cases} \mathbbm{E}[\psi(i)(X_1,\dotsc,X_{p_i})\psi(l)(X_1,\dotsc,X_{p_l}) ],&\text{if }p_i=p_l\text{ and } K=J\\ 0,&\text{otherwise,}\end{cases} \end{equation*} for $i,l=1,\dots,d$, $J\in\mathcal{D}_{p_i}(n)$ and $K\in\mathcal{D}_{p_l}(n)$. \subsection{Distance from the pre-limiting process} Having established the setup and defined the pre-limiting process above, we prove the following result: \begin{theorem}\label{theorem_weighted_pre} Let $\mathbf{Y}_n$ be defined as in Section \ref{intro_weighted} and $\mathbf{D}_n$ be defined as in Section \ref{pre_lim_weighted}. Then, for any $g\in M$, \begin{align*} &\Bigl|\mathbbm{E}[g(\mathbf{Y_n})]-\mathbbm{E}[g(\mathbf{D}_n)]\Bigr| \leq\frac{2\sqrt{d}\|g\|_M}{3p_1}\sum_{i=1}^d\frac{\|\psi(i)\|_{L^3(\mu^{p_i})}^3 }{\sigma_n(i)^3}\sum_{l=1}^n \left(\sum_{\substack{J\in\mathcal{D}_{p_i}(n):\\ l\in J}} |a_J(i)|\right)^3\notag\\ &\;+ \|g\|_M \sum_{i,j,k=1}^d \frac{\|\psi(i)\|_{L^3(\mu^{p_i})}\|\psi(j)\|_{L^3(\mu^{p_j})}\|\psi(k)\|_{L^3(\mu^{p_k})}}{\sigma_n(i)\sigma_n(j)\sigma_n(k)}\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\K\in\mathcal{D}_{p_j}(n),\\L\in \mathcal{D}_{p_k}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_J(i)a_K(j)a_L(k)|. \end{align*} \end{theorem} \begin{proof}~\\ \textbf{Step 1.} First note that, for $\epsilon_1$ in Theorem \ref{theorem1}, \begin{equation}\label{s1_1} \left\|(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right\|\left\|\mathbf{Y}_n-\mathbf{Y}_n'\right\|^2\leq\frac{n}{2p_1}\left\|\mathbf{Y}_n-\mathbf{Y}_n'\right\|^3, \end{equation} which follows directly from the definition of $\Lambda_n$ in (\ref{lambda_weighted}) and our assumption that $p_1\leq\dots\leq p_d$. Now, note that \begin{align} \left\|\mathbf{Y}_n-\mathbf{Y}_n'\right\|^3=&\sup_{t\in[0,1]}\left[\left(\mathbf{Y}_n^{(1)}(t)-\left(\mathbf{Y}_n^{(1)}\right)'(t)\right)^2+\dots+\left(\mathbf{Y}_n^{(d)}(t)-\left(\mathbf{Y}_n^{(d)}\right)'(t)\right)^2\right]^{3/2}\nonumber\\ \leq&\sqrt{d}\sup_{t\in[0,1]}\left[\left|\mathbf{Y}_n^{(1)}(t)-\left(\mathbf{Y}_n^{(1)}\right)'(t)\right|^3+\dots+\left|\mathbf{Y}_n^{(d)}(t)-\left(\mathbf{Y}_n^{(d)}\right)'(t)\right|^3\right]\nonumber\\ \leq&\sqrt{d}\left[\left\|\mathbf{Y}_n^{(1)}-\left(\mathbf{Y}_n^{(1)}\right)'\right\|^3+\dots+\left\|\mathbf{Y}_n^{(d)}-\left(\mathbf{Y}_n^{(d)}\right)'\right\|^3\right].\label{s1_2} \end{align} Furthermore, for $\max(J):=\max\lbrace j\,:\,j\in J\rbrace$ and for all $i=1,\dots,d$: \begin{align} &\mathbbm{E}\left\|\mathbf{Y}_n^{(i)}-\left(\mathbf{Y}_n^{(i)}\right)'\right\|^3\nonumber\\ &=\frac{1}{\sigma_n(i)^3}\notag\\ &\cdot\mathbbm{E}\left\{\sup_{t\in[0,1]}\left| \sum_{\substack{J\in\mathcal{D}_{p_i}(\lfloor nt\rfloor):\\ I\in J}} a_J(i)\bigl(\psi(i)(X_j,j\in J)- \psi(i)(X_0,X_j,j\in J\setminus\{I\})\bigr) \mathbbm{1}_{[\frac{\max(J)}{n},1]}(t)\right|^3\right\}\notag\\ &\leq \frac{1}{\sigma_n(i)^3}\mathbbm{E}\left(\sum_{\substack{J\in\mathcal{D}_{p_i}(n):\\ I\in J}} |a_J(i)|\bigl|\psi(i)(X_j,j\in J)- \psi(i)(X_0,X_j,j\in J\setminus\{I\})\bigr|\right)^3\notag\\ &\leq\frac{1}{n\sigma_n(i)^3}\sum_{l=1}^n \sum_{\substack{J,K,L\in\mathcal{D}_{p_i}(n):\\ l\in J\cap K\cap L}} |a_J(i)a_K(i)a_L(i)| \mathbbm{E}\Biggl[\bigl|\psi(i)(X_j,j\in J)- \psi(i)(X_0,X_j,j\in J\setminus\{l\})\bigr|\notag\\ &\cdot\bigl|\psi(i)(X_j,j\in K)- \psi(i)(X_0,X_j,j\in K\setminus\{l\})\bigr|\bigl|\psi(i)(X_j,j\in L)- \psi(i)(X_0,X_j,j\in L\setminus\{l\})\bigr|\Biggr]\notag\\ &\leq \frac{\mathbbm{E}\bigl|\psi(i)(X_1,\dotsc,X_{p_i})-\psi(i)(X_2,\dotsc,X_{p_{i+1}})\bigr|^3 }{n\sigma_n(i)^3}\sum_{l=1}^n\sum_{\substack{J,K,L\in\mathcal{D}_{p_i}(n):\\ l\in J\cap K\cap L}} |a_J(i)a_K(i)a_L(i)|\label{s1_3a}\\ &\leq \frac{8\mathbbm{E}\bigl|\psi(i)(X_1,\dotsc,X_{p_i})\bigr|^3 }{n\sigma_n(i)^3}\sum_{l=1}^n \left(\sum_{\substack{J\in\mathcal{D}_{p_i}(n):\\ l\in J}} |a_J(i)|\right)^3\label{s1_3b}\,. \end{align} Combining \eqref{s1_1} -\eqref{s1_3b} we obtain \begin{align} \epsilon_1&\leq \frac{\sqrt{d}\|g\|_M}{12p_1}\sum_{i=1}^d\frac{\mathbbm{E}\bigl|\psi(i)(X_1,\dotsc,X_{p_i})-\psi(i)(X_2,\dotsc,X_{p_{i+1}})\bigr|^3 }{\sigma_n(i)^3}\notag\\ &\hspace{1.7cm}\cdot\sum_{l=1}^n\sum_{\substack{J,K,L\in\mathcal{D}_{p_i}(n):\\ l\in J\cap K\cap L}} |a_J(i)a_K(i)a_L(i)|\label{eps_1_weighted_a}\\ &\leq\frac{2\sqrt{d}\|g\|_M}{3p_1}\sum_{i=1}^d\frac{\mathbbm{E}\bigl|\psi(i)(X_1,\dotsc,X_{p_i})\bigr|^3 }{\sigma_n(i)^3}\sum_{l=1}^n \left(\sum_{\substack{J\in\mathcal{D}_{p_i}(n):\\ l\in J}} |a_J(i)|\right)^3.\label{eps_1_weighted_b} \end{align} \textbf{Step 2.} We will now bound $\epsilon_2$ of Theorem \ref{theorem1}. Denoting by $e_i$ the $i$th element of the canonical basis of $\mathbbm{R}^d$, for $i=1,\dots,d$, for any $f\in M$, we have \begin{align}\label{s2a} &D^2f(\mathbf{Y}_n)\left[\left(\mathbf{Y}_n-\mathbf{Y}_n'\right)\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]\nonumber\\ =&D^2f(\mathbf{Y}_n)\left[\sum_{i=1}^d\frac{n}{2p_i}\left(\mathbf{Y}^{(i)}_n-\left(\mathbf{Y}^{(i)}_n\right)'\right)e_i,\sum_{i=1}^d\left(\mathbf{Y}^{(i)}_n-\left(\mathbf{Y}^{(i)}_n\right)'\right)e_i\right]\nonumber\\ =&\sum_{i,j=1}^d\frac{n}{2p_i}D^2f(\mathbf{Y}_n)\left[\left(\mathbf{Y}^{(i)}_n-\left(\mathbf{Y}^{(i)}_n\right)'\right)e_i,\left(\mathbf{Y}^{(j)}_n-\left(\mathbf{Y}^{(j)}_n\right)'\right)e_j\right]. \end{align} We now let $f=\phi_n(g)$, as defined by \eqref{phi}, and fix some $i,j\in\lbrace 1,\dots,d\rbrace$. We have that \begin{align}\label{s2b} &\Bigg|\frac{n}{2p_i}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(\mathbf{Y}^{(i)}_n-\left(\mathbf{Y}^{(i)}_n\right)'\right)e_i,\left(\mathbf{Y}^{(j)}_n-\left(\mathbf{Y}^{(j)}_n\right)'\right)e_j\right]\notag\\ &\hspace{9cm}-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n^{(i)}e_i,\mathbf{D}_n^{(j)}e_j\right]\Bigg|\nonumber\\ &=\frac{1}{\sigma_n(i)\sigma_n(j)}\Biggl|\frac{n}{2p_i} \sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n)}}a_J(i)a_K(j) \mathbbm{E}\Bigl[ \bigl(\psi(i)(X_u,u\in J)-\psi(i)(X'_u,u\in J) \bigr)\notag\\ &\hspace{2cm}\cdot \bigl(\psi(j)(X_u,u\in K)-\psi(j)(X'_u,u\in K) \bigr)D^2f(\mathbf{Y}_n) \bigl[\mathbbm{1}_{[\frac{\max(J)}{n},1]}e_i,\mathbbm{1}_{[\frac{\max(K)}{n},1]}e_j\bigr]\Bigr]\notag\\ &\hspace{3cm}-\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n)}}a_J(i)a_K(j)\mathbbm{1}_{\{J=K\}} \mathbbm{E}\bigl[\psi(i)(X_1,\dotsc,X_{p_i})\psi(j)(X_1,\dotsc,X_{p_j})\bigr]\notag\\ &\hspace{7cm}\cdot\mathbbm{E}\Bigl[D^2f(\mathbf{Y}_n) \bigl[\mathbbm{1}_{[\frac{\max(J)}{n},1]}e_i,\mathbbm{1}_{[\frac{\max(K)}{n},1]}e_j\bigr]\Bigr]\Biggr|\notag\\ &=\frac{1}{2p_i\sigma_n(i)\sigma_n(j)}\Biggl|\sum_{l=1}^n\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n),\\l\in J\cap K }}a_J(i)a_K(j) \mathbbm{E}\Bigl[ \bigl(\psi(i)(X_u,u\in J)-\psi(i)(X_0,X_u,u\in J\setminus\{l\}) \bigr)\notag\\ &\hspace{0.5cm}\cdot \bigl(\psi(j)(X_u,u\in K)-\psi(j)(X_0,X_u,u\in K\setminus\{l\}) \bigr)D^2f(\mathbf{Y}_n) \bigl[\mathbbm{1}_{[\frac{\max(J)}{n},1]}e_i,\mathbbm{1}_{[\frac{\max(K)}{n},1]}e_j\bigr]\Bigr]\notag\\ &-2\sum_{l=1}^n\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n),\\ l\in J\cap K}}a_J(i)a_K(j)\mathbbm{1}_{\{J=K\}} \mathbbm{E}\bigl[\psi(i)(X_1,\dotsc,X_{p_i})\psi(j)(X_1,\dotsc,X_{p_j})\bigr]\notag\\ &\hspace{7cm}\cdot\mathbbm{E}\Bigl[D^2f(\mathbf{Y}_n) \bigl[\mathbbm{1}_{[\frac{\max(J)}{n},1]}e_i,\mathbbm{1}_{[\frac{\max(K)}{n},1]}e_j\bigr]\Bigr]\Biggr|\notag\\ &=\frac{1}{2p_i\sigma_n(i)\sigma_n(j)}\Biggl|\sum_{l=1}^n\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n),\\l\in J\cap K }}\hspace{-3mm}a_J(i)a_K(j) \mathbbm{E}\Biggl[\Biggl( \biggl(\psi(i)(X_u,u\in J)\notag\\ &\hspace{1.5cm}-\psi(i)(X_0,X_u,u\in J\setminus\{l\}) \biggr)\cdot \biggl(\psi(j)(X_u,u\in K)-\psi(j)(X_0,X_u,u\in K\setminus\{l\}) \biggr)\notag\\ &-2\mathbbm{1}_{\{J=K\}}\mathbbm{E}\bigl[\psi(i)(X_1,\dotsc,X_{p_i})\psi(j)(X_1,\dotsc,X_{p_j})\bigr]\Biggr)D^2f(\mathbf{Y}_n) \bigl[\mathbbm{1}_{[\frac{\max(J)}{n},1]}e_i,\mathbbm{1}_{[\frac{\max(K)}{n},1]}e_j\bigr]\Biggr]\Biggr|. \end{align} Now, we define \[\mathbf{Y}_n^{J,K}:=\Bigl(\bigl(\mathbf{Y}_n^{J,K}\bigr)^{(1)},\cdots,\bigl(\mathbf{Y}_n^{J,K}\bigr)^{(d)}\Bigr)\] via \begin{equation*} \bigl(\mathbf{Y}_n^{J,K}\bigr)^{(i)}(t):=\frac{1}{\sigma_n(i)}\sum_{\substack{L\in\mathcal{D}_{p_i}(\lfloor nt\rfloor):\\ L\cap(J\cup K)=\emptyset}} a_J(i)\psi(i)(X_j,j\in L)\,,\quad 1\leq i\leq d,\, t\in[0,1]\,. \end{equation*} Then, using independence, from \eqref{s2b} we obtain that \begin{align}\label{s2c} &\Bigg|\frac{n}{2p_i}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(\mathbf{Y}^{(i)}_n-\left(\mathbf{Y}^{(i)}_n\right)'\right)e_i,\left(\mathbf{Y}^{(j)}_n-\left(\mathbf{Y}^{(j)}_n\right)'\right)e_j\right]\notag\\ &\hspace{9cm}-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n^{(i)}e_i,\mathbf{D}_n^{(j)}e_j\right]\Bigg|\nonumber\\ &=\frac{1}{2p_i\sigma_n(i)\sigma_n(j)}\Biggl|\sum_{l=1}^n\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n),\\l\in J\cap K }}a_J(i)a_K(j) \mathbbm{E}\Biggl[\Biggl( \biggl(\psi(i)(X_u,u\in J)\notag\\ &\hspace{2cm}-\psi(i)(X_0,X_u,u\in J\setminus\{l\}) \biggr) \biggl(\psi(j)(X_u,u\in K)-\psi(j)(X_0,X_u,u\in K\setminus\{l\}) \biggr) \notag\\ &\hspace{5.5cm}-2\mathbbm{1}_{\{J=K\}}\mathbbm{E}\bigl[\psi(i)(X_1,\dotsc,X_{p_i})\psi(j)(X_1,\dotsc,X_{p_j})\bigr]\Biggr)\notag\\ &\hspace{5cm}\cdot\Bigl(D^2f(\mathbf{Y}_n)-D^2f(\mathbf{Y}_n^{J,K})\bigr) \bigl[\mathbbm{1}_{[\frac{\max(J)}{n},1]}e_i,\mathbbm{1}_{[\frac{\max(K)}{n},1]}e_j\bigr]\Biggr]\Biggr|\notag\\ &\stackrel{\eqref{m_bound}C}\leq \frac{\|g\|_M}{6p_i\sigma_n(i)\sigma_n(j)}\sum_{l=1}^n\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n),\\l\in J\cap K }}|a_J(i)a_K(j)| \mathbbm{E}\Biggl[\Biggl| \biggl(\psi(i)(X_u,u\in J)\notag\\ &\hspace{1.5cm}-\psi(i)(X_0,X_u,u\in J\setminus\{l\}) \biggr) \biggl(\psi(j)(X_u,u\in K)-\psi(j)(X_0,X_u,u\in K\setminus\{l\}) \biggr)\notag\\ &\hspace{3cm} -2\mathbbm{1}_{\{J=K\}}\mathbbm{E}\bigl[\psi(i)(X_1,\dotsc,X_{p_i})\psi(j)(X_1,\dotsc,X_{p_j})\bigr]\Biggr|\cdot\|\mathbf{Y}_n-\mathbf{Y}_n^{J,K}\| \Biggr]. \end{align} Now, we observe that \begin{align*} \|\mathbf{Y}_n-\mathbf{Y}_n^{J,K}\|&\leq \sum_{k=1}^d\frac{1}{\sigma_n(k)}\|\mathbf{Y}_n^{(k)}-(\mathbf{Y}_n^{J,K})^{(k)}\|\notag\\ & \leq\sum_{k=1}^d\frac{1}{\sigma_n(k)}\sum_{\substack{L\in\mathcal{D}_{p_k}(n):\\ L\cap(J\cup K)\not=\emptyset}} |a_J(k)| |\psi(k)(X_u,u\in L)|. \end{align*} Hence, \eqref{s2c} yields \begin{align}\label{s2d} &\Bigg|\frac{n}{2p_i}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(\mathbf{Y}^{(i)}_n-\left(\mathbf{Y}^{(i)}_n\right)'\right)e_i,\left(\mathbf{Y}^{(j)}_n-\left(\mathbf{Y}^{(j)}_n\right)'\right)e_j\right]\notag\\ &\hspace{9cm}-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n^{(i)}e_i,\mathbf{D}_n^{(j)}e_j\right]\Bigg|\nonumber\\ &\leq\sum_{k=1}^d \frac{\|g\|_M}{6p_i\sigma_n(i)\sigma_n(j)\sigma_n(k)}\sum_{l=1}^n\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n),\\l\in J\cap K }}\sum_{\substack{L\in\mathcal{D}_{p_k}(n):\\ L\cap(J\cup K)\not=\emptyset}}|a_J(i)a_K(j)a_L(k)|\mathbbm{E}\Biggl[\biggl| \Bigl(\psi(i)(X_u,u\in J)\notag\\ &\hspace{1.5cm}-\psi(i)(X_0,X_u,u\in J\setminus\{l\}) \Bigr)\cdot \Bigl(\psi(j)(X_u,u\in K)-\psi(j)(X_0,X_u,u\in K\setminus\{l\}) \Bigr)\notag\\ &\hspace{2.5cm}-2\mathbbm{1}_{\{J=K\}}\mathbbm{E}\bigl[\psi(i)(X_1,\dotsc,X_{p_i})\psi(j)(X_1,\dotsc,X_{p_j})\bigr]\biggr| \cdot|\psi(k)(X_u,u\in L)| \Biggr]\notag\\ &\leq \sum_{k=1}^d \frac{\|g\|_M}{p_i\sigma_n(i)\sigma_n(j)\sigma_n(k)}\notag\\ &\hspace{1cm}\cdot\sum_{l=1}^n\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\ K\in\mathcal{D}_{p_j}(n),\\l\in J\cap K }}\sum_{\substack{L\in\mathcal{D}_{p_k}(n):\\ L\cap(J\cup K)\not=\emptyset}}|a_J(i)a_K(j)a_L(k)| \|\psi(i)\|_{L^3(\mu^{p_i})}\|\psi(j)\|_{L^3(\mu^{p_j})}\|\psi(k)\|_{L^3(\mu^{p_k})}\notag\\ &\leq \sum_{k=1}^d \frac{\|g\|_M \|\psi(i)\|_{L^3(\mu^{p_i})}\|\psi(j)\|_{L^3(\mu^{p_j})}\|\psi(k)\|_{L^3(\mu^{p_k})}}{\sigma_n(i)\sigma_n(j)\sigma_n(k)}\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\K\in\mathcal{D}_{p_j}(n),\\L\in \mathcal{D}_{p_k}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_J(i)a_K(j)a_L(k)|. \end{align} Finally, \eqref{s2a} and \eqref{s2d} imply that \begin{align*} \epsilon_2\leq& \sum_{i,j=1}^d\Bigg|\frac{n}{2p_i}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(\mathbf{Y}^{(i)}_n-\left(\mathbf{Y}^{(i)}_n\right)'\right)e_i,\left(\mathbf{Y}^{(j)}_n-\left(\mathbf{Y}^{(j)}_n\right)'\right)e_j\right]\\ &\hspace{8cm}-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n^{(i)}e_i,\mathbf{D}_n^{(j)}e_j\right]\Bigg|\nonumber\\ \leq & \sum_{i,j,k=1}^d \frac{\|g\|_M \|\psi(i)\|_{L^3(\mu^{p_i})}\|\psi(j)\|_{L^3(\mu^{p_j})}\|\psi(k)\|_{L^3(\mu^{p_k})}}{\sigma_n(i)\sigma_n(j)\sigma_n(k)}\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\K\in\mathcal{D}_{p_j}(n),\\L\in \mathcal{D}_{p_k}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_J(i)a_K(j)a_L(k)|. \end{align*} \end{proof} \subsection{Distance from a continuous process} We now prove the following theorem, which bounds the distance between the law of $\mathbf{Y}_n$ and that of a continuous Gaussian process. Let us introduce some notation first. Let $\Sigma_n^{(m)}\in\mathbbm{R}^{d\times d}$ be given by \[\left(\Sigma_n^{(m)}\right)_{i,l}= \begin{cases} \frac{n}{\sigma_n{(i)}\sigma_n{(l)}}\underset{m=\max(J)}{\underset{{J\in\mathcal{D}_{p_i}(m):}}{\sum}}a_{J}{(i)}a_J{(l)}\mathbbm{E}\left[\psi(i)(X_1,\dots,X_{p_i})\psi(l)(X_1,\dots,X_{p_l})\right],&\text{if }p_i=p_l\\ 0,&\text{otherwise,} \end{cases}\] for $i,l=1,\dots,d$. For $i=1,\dots,d$, let \[\delta_n^{(i)}=\frac{1}{\left(\sigma_n{(i)}\right)^2}\sup_{m\in[n]}\underset{m=\max(J)}{\underset{{J\in\mathcal{D}_{p_i}(m):}}{\sum}}a_J(i)^2\mathbbm{E}\left[\psi(i)^2(X_1,\dots,X_{p_i})\right],\] where $[n]:=\lbrace 1,\dots, n\rbrace$, and \[T_n^{(i)}=\frac{1}{\left(\sigma_n{(i)}\right)^2}\sum_{J\in\mathcal{D}_{p_i}(n)}a_J(i)^2\mathbbm{E}\left[\psi(i)^2(X_1,\dots,X_{p_i})\right].\] Furthermore, let \[\varphi_n(s)=\sum_{m=p_1}^n\left(\Sigma_n^{(m)}\right)^{1/2}\mathbbm{1}_{\left(\frac{m-1}{n},\frac{m}{n}\right]}(s),\quad s\in[0,1]\] and suppose that $\varphi:[0,1]\to\mathbbm{R}^{d\times d}$ is a matrix of $L^2([0,1])$-functions such that, for all $i,j=1,\dots,d$, \[\lim_{n\to\infty}\int_0^1\left|\left(\varphi_n(s)-\varphi(s)\right)_{i,j}\right|^2\, ds=0.\] Let $\|\cdot\|_F$ denote the Frobenius norm. Suppose that $\mathbf{W}$ is a $d$-dimensional standard Brownian motion. Let \[\mathbf{Z}(t)=\int_0^t\varphi(s)d\mathbf{W}(s)\] and $\mathbf{Y}_n$ be defined as in Section \ref{intro_weighted}. \begin{theorem}\label{theorem_weighted_con} Under the above setup, for any $g\in M$, \[\left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{Z})\right|\leq \|g\|_{M}(\gamma_1+\gamma_2+\gamma_3+\gamma_4+\gamma_5),\] and, for any $g\in M^0$, \[\left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{Z})\right|\leq \|g\|_{M^0}(\gamma_1+\gamma_2+\gamma_3),\] where \begin{align*} \gamma_1&=\frac{2\sqrt{d}}{3p_1}\sum_{i=1}^d\frac{\|\psi(i)\|_{L^3(\mu^{p_i})}^3 }{\sigma_n(i)^3}\sum_{l=1}^n \left(\sum_{\substack{J\in\mathcal{D}_{p_i}(n):\\ l\in J}} |a_J(i)|\right)^3;\notag\\ \gamma_2&= \sum_{i,j,k=1}^d \frac{\|\psi(i)\|_{L^3(\mu^{p_i})}\|\psi(j)\|_{L^3(\mu^{p_j})}\|\psi(k)\|_{L^3(\mu^{p_k})}}{\sigma_n(i)\sigma_n(j)\sigma_n(k)}\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\K\in\mathcal{D}_{p_j}(n),\\L\in \mathcal{D}_{p_k}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_J(i)a_K(j)a_L(k)|;\\ \gamma_3&=2\sqrt{\int_0^1\left\|\varphi_n(s)-\varphi(s)\right\|_F^2ds}+12\sqrt{\sum_{i=1}^d\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)};\\ \gamma_4&=\sqrt{d}\sum_{i=1}^d\left[8447\left(\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)\right)^{3/2}+44\left(\sum_{j=1}^d\int_0^1\left[\left(\varphi_n(s)-\varphi(s)\right)_{i,j}\right]^2ds\right)^{3/2}\right];\\ \gamma_5&=\sqrt{d}\left(\int_0^1\left\|\varphi(s)\right\|_F^2ds\right)\sum_{i=1}^d\Bigg[50\sqrt{\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)}\\ &\hspace{7cm}+19\sqrt{\sum_{j=1}^d\int_0^1\left[\left(\varphi_n(s)-\varphi(s)\right)_{i,j}\right]^2ds}\;\Bigg]. \end{align*} \end{theorem} \begin{proof} Let us write $\mathbf{W}=\left(\mathbf{W}^{(1)},\dots,\mathbf{W}^{(d)}\right)$, where $\mathbf{W}^{(1)},\dots,\mathbf{W}^{(d)}$ are i.i.d. standard Brownian motions in $\mathbbm{R}$. \textbf{Step 1.} Consider process $\mathbf{D}_n$ defined in Section \ref{pre_lim_weighted}. Note that, for $i=1,\dots,d$, \begin{align*} \mathbf{D}_n^{(i)}(t)=&\frac{1}{\sigma_n^{(i)}}\sum_{J\in \mathcal{D}_{p_i}(\lfloor nt\rfloor)}a_J(i)Z_J(i)\\ =&\frac{1}{\sigma_n^{(i)}}\sum_{m=p_i}^{\lfloor nt\rfloor}\underset{m=\max(J)}{\sum_{J\in\mathcal{D}_{p_i}([m]):}}a_J(i)Z_J(i)\\ =&\frac{1}{\sigma_n^{(i)}}\sum_{m=p_i}^{\lfloor nt\rfloor}\tilde{Z}_m(i), \end{align*} where $\lbrace \tilde{Z}_{m}(i):m\in [n], i\in [d]\rbrace$ is a jointly Gaussian collection of centred random variables with the following covariance structure: \begin{align*} &\mathbbm{E}\left[\tilde{Z}_{m_1}(i)\tilde{Z}_{m_2}(l)\right]\\ =& \begin{cases} \underset{m_1=\max(J)}{\underset{{J\in\mathcal{D}_{p_i}([m_1]):}}{\sum}}a_{J}^{(i)}a_J^{(l)}\mathbbm{E}\left[\psi(i)(X_1,\dots,X_{p_i})\psi(l)(X_1,\dots,X_{p_l})\right],&\text{if }p_i=p_l\text{ and }m_1=m_2\\ 0,&\text{otherwise}. \end{cases} \end{align*} Using this observation, note that $\mathbf{D}_n$ has the same distribution as $\tilde{\mathbf{Z}}_n$ given by \begin{align*} \tilde{\mathbf{Z}}_n(t):=\frac{1}{\sqrt{n}}\sum_{m=p_1}^n\int_0^{\lfloor nt\rfloor}\left(\Sigma_n^{(m)}\right)^{1/2}\mathbbm{1}_{(m-1,m]}(s)d\mathbf{W}(s),\quad t\in[0,1], \end{align*} whose distribution, by a simple change of variables, is equal to that of \begin{align*} \mathbf{Z}_n(t):=\sum_{m=p_1}^n\int_0^{\lfloor nt\rfloor/n}\left(\Sigma_n^{(m)}\right)^{1/2}\mathbbm{1}_{\left(\frac{m-1}{n},\frac{m}{n}\right]}(s)d\mathbf{W}(s)=\int_0^{\lfloor nt\rfloor/n}\varphi_n(s)d\mathbf{W}(s),\quad t\in[0,1]. \end{align*} \textbf{Step 2.} By Doob's $L^2$ inequality and It\^o's isometry, we note that \begin{align} \mathbbm{E}\sup_{t\in[0,1]}\left|\int_0^{t}\left(\varphi_n(s)-\varphi(s)\right)d\mathbf{W}(s)\right|^2=&\mathbbm{E}\left[\sup_{t\in[0,1]}\sum_{i=1}^d\left(\sum_{j=1}^d\int_0^t\left(\varphi_n(s)-\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right)^2\right]\notag\\ \leq &4\sum_{i=1}^d\mathbbm{E}\left[\left(\sum_{j=1}^d\int_0^1\left(\varphi_n(s)-\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right)^2\right]\notag\\ =&4\sum_{i,j=1}^d\mathbbm{E}\left[\left(\int_0^1\left(\varphi_n(s)-\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right)^2\right]\notag\\ =&4\int_0^1\left\|\varphi_n(s)-\varphi(s)\right\|_F^2ds.\label{second1} \end{align} Similarly, by Doob's $L^{3}$ inequality, the formula for Gaussian moments and It\^o's isometry, \begin{align} &\mathbbm{E}\sup_{t\in[0,1]}\left|\int_0^{t}\left(\varphi_n(s)-\varphi(s)\right)d\mathbf{W}(s)\right|^3\\ =&\mathbbm{E}\left[\sup_{t\in[0,1]}\left(\sum_{i=1}^d\left(\sum_{j=1}^d\int_0^t\left(\varphi_n(s)-\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right)^2\right)^{3/2}\right]\notag\\ \leq &\frac{27\sqrt{d}}{8}\sum_{i=1}^d\mathbbm{E}\left[\left|\sum_{j=1}^d\int_0^1\left(\varphi_n(s)-\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right|^3\right]\notag\\ =&\frac{27\sqrt{d}}{2\sqrt{2\pi}}\sum_{i=1}^d\left(\mathbbm{E}\left[\left(\sum_{j=1}^d\int_0^1\left(\varphi_n(s)-\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right)^2\right]\right)^{3/2}\notag\\ =&\frac{27\sqrt{d}}{2\sqrt{2\pi}}\sum_{i=1}^d\left(\sum_{j=1}^d\int_0^1\left[\left(\varphi_n(s)-\varphi(s)\right)_{i,j}\right]^2ds\right)^{3/2}.\label{third1} \end{align} \textbf{Step 3.} We now apply an argument similar to that of \cite[Theorem 1]{ito_processes}. Note that \begin{align*} \mathbf{M}_n(t)=\int_0^{t\wedge 1}\varphi_n(s)d\mathbf{W}(s)+\left(\mathbf{W}(t)-\mathbf{W}(1)\right)\mathbbm{1}_{[t>1]} \end{align*} is a martingale vanishing at zero. In particular, so are the coordinate processes \[\mathbf{M}_n^{(i)}(t)=\int_0^{t\wedge 1} \sum_{j=1}^d\left(\varphi_n\right)_{i,j}d\mathbf{W}^{(j)}(s) +\left(\mathbf{W}^{(i)}(t)-\mathbf{W}^{(i)}(1)\right)\mathbbm{1}_{[t>1]}.\] Note that, by the Dambis-Dubins-Schwarz theorem, for each $i=1,\dots,d$, there exists a Wiener process $\tilde{\mathbf{W}}^{(i)}$, such that \begin{align*} \mathbf{M}_n^{(i)}(t)=\tilde{\mathbf{W}}^{(i)}\left(\left<\mathbf{M}_n^{(i)}\right>_t\right),\quad t\geq 0, \end{align*} where $\left<\mathbf{M}_n^{(i)}\right>_t$ is the quadratic variation of $\mathbf{M}_n^{(i)}$, i.e. \begin{align*} &\left<\mathbf{M}_n^{(i)}\right>_t=\sum_{j=1}^d\int_0^{t\wedge 1}\left((\varphi_n)_{i,j}\right)^2ds+(t-1)\vee 0. \end{align*} Note that \begin{align*} \left<\mathbf{M}_n^{(i)}\right>_1=&\sum_{m=p_1}^n\int_0^1\left(\Sigma_n^{(m)}\right)_{i,i}\mathbbm{1}_{\left(\frac{m-1}{n},\frac{m}{n}\right]}(s)ds=\frac{1}{n}\sum_{m=p_1}^n\left(\Sigma_n^{(m)}\right)_{i,i}=T_n^{(i)} \end{align*} and \begin{align*} \sup_{t\in[0,1]}\left(\left<\mathbf{M}_n^{(i)}\right>_t-\left<\mathbf{M}_n^{(i)}\right>_{\lfloor nt\rfloor/n}\right)=&\sup_{t\in[0,1]}\sum_{j=1}^d\int_{\lfloor nt\rfloor/n}^{t}\left((\varphi_n)_{i,j}(s)\right)^2ds\\ =&\sup_{t\in[0,1]}\sum_{j=1}^d\int_{\lfloor nt\rfloor/n}^{t}\left(\left(\Sigma_n^{\left((\lfloor nt\rfloor +1)\wedge n\right)}\right)^{1/2}\right)_{i,j}^2ds\\ =&\sup_{t\in[0,1]}\left(t-\frac{\lfloor nt\rfloor}{n}\right)\left(\Sigma_n^{\left((\lfloor nt\rfloor +1)\wedge n\right)}\right)_{i,i}\\ \leq&\frac{1}{\left(\sigma_n^{(i)}\right)^2}\sup_{m\in[n]}\underset{m=\max(J)}{\underset{{J\in\mathcal{D}_{p_i}(m):}}{\sum}}a_J(i)^2\mathbbm{E}\left[\psi(i)^2(X_1,\dots,X_{p_i})\right]\\ =&\delta_n^{(i)}. \end{align*} Therefore, using \cite[Lemma 3]{ito_processes}, we have that \begin{align*} &\mathbbm{E}\sup_{t\in[0,1]}\left|\left(\int_{\lfloor nt\rfloor/n}^t\varphi_n(s)d\mathbf{W}(s)\right)_i\right|^2\\ \leq &\mathbbm{E}\sup\Bigg\{ \left|\tilde{\mathbf{W}}^{(i)}(u)-\tilde{\mathbf{W}}^{(i)}(v)\right|^2:\\ &\hspace{3cm}u,v\in\left[0,\left<\mathbf{M}_n^{(i)}\right>_1\right],\,|u-v|\leq\sup_{t\in[0,1]}\left(\left<\mathbf{M}_n^{(i)}\right>_t-\left<\mathbf{M}_n^{(i)}\right>_{\lfloor nt\rfloor/n}\right)\Bigg\}\\ \leq &\mathbbm{E}\sup\left\lbrace \left|\tilde{\mathbf{W}}^{(i)}(u)-\tilde{\mathbf{W}}^{(i)}(v)\right|^2:\,u,v\in\left[0,T_n^{(i)}\right],\,|u-v|\leq\delta_n^{(i)} \right\rbrace\\ \leq& \frac{5\cdot 6^2}{2\log 2}\left(\delta_n^{(i)}\log\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right) \end{align*} and \begin{align*} &\mathbbm{E}\sup_{t\in[0,1]}\left|\left(\int_{\lfloor nt\rfloor/n}^t\varphi_n(s)d\mathbf{W}(s)\right)_i\right|^3\\ \leq &\mathbbm{E}\sup\Bigg\{ \left|\tilde{\mathbf{W}}^{(i)}(u)-\tilde{\mathbf{W}}^{(i)}(v)\right|^3:\\ &\hspace{3cm}u,v\in\left[0,\left<\mathbf{M}_n^{(i)}\right>_1\right],\,|u-v|\leq\sup_{t\in[0,1]}\left(\left<\mathbf{M}_n^{(i)}\right>_t-\left<\mathbf{M}_n^{(i)}\right>_{\lfloor nt\rfloor/n}\right)\Bigg\}\\ \leq &\mathbbm{E}\sup\left\lbrace \left|\tilde{\mathbf{W}}^{(i)}(u)-\tilde{\mathbf{W}}^{(i)}(v)\right|^3:\,u,v\in\left[0,T_n^{(i)}\right],\,|u-v|\leq\delta_n^{(i)} \right\rbrace\\ \leq& \frac{5\cdot 6^3}{\sqrt{\pi}(\log 2)^{3/2}}\left(\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)\right)^{3/2}. \end{align*} Finally, it follows that \begin{align} &\mathbbm{E}\sup_{t\in[0,1]}\left|\int_{\lfloor nt\rfloor/n}^t\varphi_n(s)d\mathbf{W}(s)\right|\leq \frac{6\sqrt{5}}{\sqrt{2\log 2}}\sqrt{\sum_{i=1}^d\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)};\label{second2}\\ &\mathbbm{E}\sup_{t\in[0,1]}\left|\int_{\lfloor nt\rfloor/n}^t\varphi_n(s)d\mathbf{W}(s)\right|^3\leq\sqrt{d}\sum_{i=1}^d\mathbbm{E}\sup_{t\in[0,1]}\left|\left(\int_{\lfloor nt\rfloor/n}^t\varphi_n(s)d\mathbf{W}(s)\right)_i\right|^3\notag\\ &\hspace{2cm}\leq \frac{5\cdot 6^3\sqrt{d}}{\sqrt{\pi}(\log 2)^{3/2}}\sum_{i=1}^d\left(\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)\right)^{3/2}.\label{third2} \end{align} \textbf{Step 3.} Using the calculations above, we note that \begin{align*} &\mathbbm{E}\|\textbf{Z}_n-\textbf{Z}\|\stackrel{\eqref{second1},\eqref{second2}}\leq 2\sqrt{\int_0^1\left\|\varphi_n(s)-\varphi(s)\right\|_F^2ds}+\frac{6\sqrt{5}}{\sqrt{2\log 2}}\sqrt{\sum_{i=1}^d\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)};\\ &\mathbbm{E}\|\textbf{Z}_n-\textbf{Z}\|^3\stackrel{\eqref{third1},\eqref{third2}}\leq\frac{20\cdot 6^3\sqrt{d}}{\sqrt{\pi}(\log 2)^{3/2}}\left(\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)\right)^{3/2}\\ &\hspace{6cm}+\frac{54\sqrt{d}}{\sqrt{2\pi}}\sum_{i=1}^d\left(\sum_{j=1}^d\int_0^1\left[\left(\varphi_n(s)-\varphi(s)\right)_{i,j}\right]^2ds\right)^{3/2}. \end{align*} We furthermore note that, using Doob's $L^3$ inequality, the formula for Gaussian moments and It\^o's isometry, \begin{align*} \mathbbm{E}\|\mathbf{Z}\|^3=&\mathbbm{E}\left[\sup_{t\in[0,1]}\left(\sum_{i=1}^d\left(\sum_{j=1}^d\int_0^t\left(\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right)^2\right)^{3/2}\right]\notag\\ \leq &\frac{27\sqrt{d}}{8}\sum_{i=1}^d\mathbbm{E}\left[\left|\sum_{j=1}^d\int_0^1\left(\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right|^3\right]\notag\\ =&\frac{27\sqrt{d}}{2\sqrt{2\pi}}\sum_{i=1}^d\left(\mathbbm{E}\left[\left(\sum_{j=1}^d\int_0^1\left(\varphi(s)\right)_{i,j}d\mathbf{W}^{(j)}(s)\right)^2\right]\right)^{3/2}\notag\\ =&\frac{27\sqrt{d}}{2\sqrt{2\pi}}\sum_{i=1}^d\left(\sum_{j=1}^d\int_0^1\left|\left(\varphi(s)\right)_{i,j}\right|^2ds\right)^{3/2}. \end{align*} Therefore, using the mean value theorem \begin{align*} \left|\mathbbm{E}g(\mathbf{D}_n)-\mathbbm{E}g(\mathbf{Z})\right| \leq& \mathbbm{E}\left[\sup_{c\in[0,1]}\|Dg(\mathbf{Z}+c(\mathbf{Z}_n-\mathbf{Z})\|\|\mathbf{Z}-\mathbf{Z}_n\|\right]\\ \leq&\|g\|_{M}\mathbbm{E}\left[\sup_{c\in[0,1]}\left(1+\|\mathbf{Z}+c(\mathbf{Z}_n-\mathbf{Z})\|^2\right)\|\mathbf{Z}-\mathbf{Z}_n\|\right]\\ \stackrel{\text{H\"older}}\leq&\|g\|_{M}\left\lbrace \mathbbm{E}\|\mathbf{Z}-\mathbf{Z}_n\|+2\mathbbm{E}\|\mathbf{Z}-\mathbf{Z}_n\|^3+2\left(\mathbbm{E}\|\mathbf{Z}\|^3\right)^{2/3}\left(\mathbbm{E}\|\mathbf{Z}-\mathbf{Z}_n\|^3\right)^{1/3}\right\rbrace\\ \leq &\|g\|_M(\gamma_3+\gamma_4+\gamma_5) \end{align*} and \begin{align*} \left|\mathbbm{E}g(\mathbf{D}_n)-\mathbbm{E}g(\mathbf{Z})\right|\leq& \|g\|_{M^0}\mathbbm{E}\left\|\mathbf{Z}_n-\mathbf{Z}\right\| \leq \|g\|_{M^0}\gamma_3. \end{align*} The result now follows by Theorem \ref{theorem_weighted_pre} and the triangle inequality. \end{proof} \begin{remark}\label{remark_weighted} The approximation results in this Section are merely stated for vectors of \textit{degenerate} weighted $U$-processes. In many applications, however, the given weighted $U$-process might involve non-degenerate kernels. If \[\mathbf{U}_n(t)=\sum_{J\in\mathcal{D}_p(\lfloor nt\rfloor)} a_J\psi(X_j,j\in J)\] is such a non-degenerate, weighted $U$-process, then it can be written in its \textit{Hoeffding decompoition} as a sum of degenerate, weighted $U$-processes as follows: \begin{align*} \mathbf{U}_n(t)&=\int_{E^p}\psi d\mu^p\sum_{J\in\mathcal{D}_p(\lfloor nt\rfloor)} a_J+\sum_{q=1}^p \sum_{K\in\mathcal{D}_q(\lfloor nt\rfloor)} \Bigl(\sum_{\substack{J\in \mathcal{D}_p(\lfloor nt\rfloor):\\ K\subseteq J}}a_J \Bigr)\psi_q(X_i,i\in K)\\ &=:\int_{E^p}\psi d\mu^p\sum_{J\in\mathcal{D}_p(\lfloor nt\rfloor)} a_J+\sum_{q=1}^p \mathbf{U}^{(q)}_n(t)\,, \end{align*} where the kernels $\psi_q$, $1\leq q\leq p$, are degenerate kernels which are expressible in terms of $\psi$. Hence, the results of this Section for the vector $(\mathbf{U}_n^{(1)},\dotsc,\mathbf{U}_n^{(p)})$ together with the application of a linear functional immediately yield bounds on the approximation of $\mathbf{U}_n$ by a suitable Gaussian process. For simplicity we do not state the resulting bounds explicitly but leave their derivation to the interested reader. In the very particular example of $d$-runs on the line, however, we will work out this procedure in full detail. \end{remark} \subsection{Homogeneous sum processes}\label{homsums} In this subsection we consider an important subclass of weighted, degenerate $U$-processess, namely the processes given as so-called \textbf{homogeneous sums} or \textbf{homogeneous sum processes}. In this case, the random variables $X_i,i\in\mathbbm{N}$, are real-valued such that $\mathbbm{E}|X_1|^3<\infty$, $\mathbbm{E}[X_1]=0$ and $\mathbbm{E}[X_1^2]=1$. Moreover, for each $1\leq i\leq d$, the kernel $\psi(i)$ is given by \begin{equation*} \psi(i)(x_1,\dotsc,x_{p_i})=\prod_{j=1}^{p_i} x_j\,. \end{equation*} In particular, $\psi(i)$ does not depend on $n$. Hence, for $1\leq i\leq d$ and $t\in[0,1]$ we have that \ \mathbf{Y}_n^{(i)}(t)=\frac{1}{\sigma_n(i)}\sum_{J\in\mathcal{D}_{p_i}(\lfloor nt\rfloor)} a_J(i)\prod_{j\in J} X_j\,,\] where the $\sigma_n(i)$ are positive reals and, in this special case, the random variables $Z_J(i)$ making up the processes $\mathbf{D}_n^{(i)}$, defined in Subsection \ref{pre_lim_weighted}, are standard normally distributed. In this situation we have the following results, which are direct consequences of Theorems \ref{theorem_weighted_pre} and \ref{theorem_weighted_con}, respectively. \begin{corollary}\label{corhumsums1} With the above definitions and notation we have that \begin{align*} &\Bigl|\mathbbm{E}g(\mathbf{Y_n})-\mathbbm{E}g(\mathbf{D}_n)\Bigr| \leq\frac{2\sqrt{d}\|g\|_M}{3p_1}\sum_{i=1}^d\frac{\bigl(\mathbbm{E}|X_1|^3\bigr)^{p_i}}{\sigma_n(i)^3}\sum_{l=1}^n \left(\sum_{\substack{J\in\mathcal{D}_{p_i}(n):\\ l\in J}} |a_J(i)|\right)^3\notag\\ &\;+ \|g\|_M \sum_{i,j,k=1}^d \frac{\bigl(\mathbbm{E}|X_1|^3\bigr)^{(p_i+p_j+p_k)/3}}{\sigma_n(i)\sigma_n(j)\sigma_n(k)}\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\K\in\mathcal{D}_{p_j}(n),\\L\in \mathcal{D}_{p_k}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_J(i)a_K(j)a_L(k)|. \end{align*} \end{corollary} \begin{corollary}\label{corhumsums2} Let $\Sigma_n^{(m)}\in\mathbbm{R}^{d\times d}$ be given by \[\left(\Sigma_n^{(m)}\right)_{i,l}= \begin{cases} \frac{n}{\sigma_n(i)\sigma_n(l)}\underset{m=\max(J)}{\underset{{J\in\mathcal{D}_{p_i}(m):}}{\sum}}a_{J}(i)a_J(l),&\text{if }p_i=p_l\\ 0,&\text{otherwise,} \end{cases}\] for $i,l=1,\dots,d$. For $i=1,\dots,d$, let \[\delta_n^{(i)}=\frac{1}{\left(\sigma_n{(i)}\right)^2}\sup_{m\in[n]}\underset{m=\max(J)}{\underset{{J\in\mathcal{D}_{p_i}(m):}}{\sum}}a_J(i)^2,\] where $[n]=\{1,\dots,n\}$, and \[T_n^{(i)}=\frac{1}{\left(\sigma_n{(i)}\right)^2}\sum_{J\in\mathcal{D}_{p_i}(n)}a_J(i)^2.\] Furthermore, let \[\varphi_n(s)=\sum_{m=p_1}^n\left(\Sigma_n^{(m)}\right)^{1/2}\mathbbm{1}_{\left(\frac{m-1}{n},\frac{m}{n}\right]}(s),\quad s\in[0,1]\] and suppose that $\varphi:[0,1]\to\mathbbm{R}^{d\times d}$ is matrix of $L^2([0,1])$-functions such that, for any $i,j=1,\dots,d$, \[\lim_{n\to\infty}\int_0^1\left|\left(\varphi_n(s)-\varphi(s)\right)_{i,j}\right|^2\, ds=0,\] Let $\mathbf{Y}_n$ be defined as in Section \ref{intro_weighted} and $\|\cdot\|_F$ denote the Frobenius norm. Suppose that $\mathbf{W}$ is a $d$-dimensional standard Brownian motion and \[\mathbf{Z}(t)=\int_0^t\varphi(s)d\mathbf{W}(s).\] Then, for any $g\in M$, \[\left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{Z})\right|\leq \|g\|_{M}(\gamma_1+\gamma_2+\gamma_3+\gamma_4+\gamma_5)\] and for any $g\in M^0$, \[\left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{Z})\right|\leq \|g\|_{M^0}(\gamma_1+\gamma_2+\gamma_3),\] where \begin{align*} \gamma_1&=\frac{2\sqrt{d}}{3p_1}\sum_{i=1}^d\frac{\left(\mathbbm{E}|X_1|^3\right)^{p_i} }{\sigma_n(i)^3}\sum_{l=1}^n \left(\sum_{\substack{J\in\mathcal{D}_{p_i}(n):\\ l\in J}} |a_J(i)|\right)^3;\notag\\ \gamma_2&= \sum_{i,j,k=1}^d \frac{\left(\mathbbm{E}|X_1|^3\right)^{(p_i+p_j+p_k)/3}}{\sigma_n(i)\sigma_n(j)\sigma_n(k)}\sum_{\substack{J\in\mathcal{D}_{p_i}(n),\\K\in\mathcal{D}_{p_j}(n),\\L\in \mathcal{D}_{p_k}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_J(i)a_K(j)a_L(k)|;\\ \gamma_3&=2\sqrt{\int_0^1\left\|\varphi_n(s)-\varphi(s)\right\|_F^2ds}+12\sqrt{\sum_{i=1}^d\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)};\\ \gamma_4&=\sqrt{d}\sum_{i=1}^d\left[8447\left(\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)\right)^{3/2}+44\left(\sum_{j=1}^d\int_0^1\left[\left(\varphi_n(s)-\varphi(s)\right)_{i,j}\right]^2ds\right)^{3/2}\right];\\ \gamma_5&=\sqrt{d}\Big(\int_0^1\left\|\varphi(s)\right\|_F^2ds\Big)\hspace{-1mm}\sum_{i=1}^d\left[50\sqrt{\delta_n^{(i)}\log\left(\frac{2T_n^{(i)}}{\delta_n^{(i)}}\right)} +19\sqrt{\sum_{j=1}^d\int_0^1\left[\left(\varphi_n(s)-\varphi(s)\right)_{i,j}\right]^2ds}\right]\hspace{-1mm}. \end{align*} \end{corollary} \begin{remark}~\\ \begin{enumerate} \item In the case $p=2$ the array $(a_J)_{J\in\mathcal{D}_2(n)}:=(a_J(1))_{J\in\mathcal{D}_2(n)}$ may be identified with the (symmetric) matrix $A=(a_{i,j})_{1\leq i,j\leq n}$, where $a_{i,i}=0$ and $a_{i,j}=a_{j,i}$ for all $1\leq i,j\leq n$. Many papers \cite{deJo87, GT, Mik, Rot1, NPR} have established sufficient conditions for the (univariate) CLT to hold for $Y_n:=\mathbf{Y}_n(1)$ in this case (with the choice of $\sigma_n^2=\sum_{1\leq i\neq j\leq n} a_{i,j}^2$). Remarkably, in \cite{NPR} the authors prove a universality principle for homogeneous sums of any order $p\geq 1$. In other words, they find necessary and sufficient conditions on the coefficient functions for the asymptotic normality of $Y_n$ to hold in the case when the $X_j$'s are i.i.d. standard Gaussian. They also show that these conditions imply asymptotic normality of $Y_n$ for any possible choice of the distribution of the $X_j$'s, as long as the $X_j$'s are independent and the usual moment assumptions hold. Now concentrating on $p=2$ and letting \[\lambda_n^*:=\max\{|\lambda|\,:\,\lambda \text{ eigenvalue of } A\}\,,\] for the matrix $A$ introduced above, a well-known sufficent condition (see, e.g. \cite[Theorem 1.1]{Mik}) for $Y_n$, $n\in\mathbbm{N}$, to be asymptotically normal is that $\lim_{n\to\infty} \lambda_n^*/\sigma_n=0$ (under our standing assumption that $\mathbbm{E}|X_1|^3<\infty$). The well-known inequalities (see e.g. \cite{GT}) \[\rho_n:=\sqrt{\max_{1\leq i\leq n}\sum_{j:j\not=i}a_{i,j}^2}\leq \lambda_n^*\leq\Gamma_n:=\max_{1\leq i\leq n}\sum_{j:j\not=i}|a_{i,j}|\] imply that this condition in particular implies the \textbf{Lindeberg type condition} $\lim_{n\to\infty}\rho_n^2/\sigma_n^2=0$, which roughly says that the \textbf{asymptotic influence} of every individual $X_i$ vanishes. On the other hand, it is implied by the stronger (and maybe easier to verify) condition that $\lim_{n\to\infty}\Gamma_n/\sigma_n=0$. We remark that the sufficient condition provided by \cite{NPR} for $d=2$ reduces to $\lim_{n\to\infty}\Tr(A^4)/\sigma_n^4=0$, which is easily seen to be equivalent to $\lim_{n\to\infty} \lambda_n^*/\sigma_n=0$. Here $\Tr(B)=\sum_{i=1}^n b_{i,i}$ denotes the trace of a matrix $B=(b_{i,j})_{1\leq i,j\leq n}$. From the easy to derive inequality \[\sigma_n^{-3}\sum_{i=1}^n\biggl(\sum_{j:j\not=i}|a_{i,j}|\Bigr)^3\geq \Bigl(\sigma_n^{-2}\rho_n^2\Bigr)^{3/2}\] we conclude that, \textbf{in the univariate case}, the condition $\gamma_1\to 0$ as $n\to\infty$, which follows from our bound in Corollary \ref{corhumsums2}, is also stronger than the Lindeberg condition. The Lindeberg condition is, however, neither necessary (consider e.g. $Y_n=(n-1)^{-1/2} X_1\sum_{j=2}^n X_j$ where the $X_j$ are i.i.d symmetric Rademacher random variables) nor sufficient for the asymptotic normality of the $Y_n$. Hence, by the above inequality, also the sufficient condition $\lim_{n\to\infty} \lambda_n^*/\sigma_n=0$ is not necessary for asymptotic normality to hold. We now provide upper bounds on the quantities $\gamma_1$ and $\gamma_2$ from our bound in this special case. First note that \begin{align*} &\sigma_n^{-3}\sum_{i=1}^n\biggl(\sum_{j:j\not=i}|a_{i,j}|\Bigr)^3=\sigma_n^{-3}\Biggl(\sum_{i=1}^n\sum_{j,k,l\not=i}|a_{i,j}||a_{i,k}|a_{i,l}| \Biggr)\\ &=\sum_{i=1}^n\sum_{j:j\not=i}|a_{i,j}|^3 +3\sum_{(i,j,k)\in[n]^3_{\not=}}|a_{i,j}||a_{i,k}|^2+\sum_{(i,j,k,l)\in[n]^4_{\not=}}|a_{i,j}||a_{i,k}|a_{i,l}|\\ &=:\sigma_n^{-3}\bigl(S_1+3S_2+S_3), \end{align*} where $[n]^p_{\not=}$ denotes the collection of all $(i_1,\dotsc,i_p)\in[n]^p$ such that $i_k\not=i_l$ whenever $k\not=l$. We have \begin{align*} S_1&\leq \left(\max_{k\not=l}|a_{k,l}|\right)\sum_{i=1}^n\sum_{j:j\not=i}|a_{i,j}|^2\leq \rho_n\sigma_n^2\,,\\ S_2&=\sum_{i\not=k}|a_{i,k}|^2\sum_{j:j\not=i,k} |a_{i,j}|\leq \Gamma_n\sum_{1\leq i\not=k\leq n}|a_{i,k}|^2=\Gamma_n\sigma_n^2,\\ S_3&=\sum_{i\not=j}|a_{i,j}|\sum_{k:k\not=i,j}|a_{i,k}|\sum_{l:l\not=i,k,j}|a_{i,l}|\leq\Gamma_n^2\sum_{1\leq i\not=j\leq n}|a_{i,j}|. \end{align*} Hence, there is an absolute constant $C_1$ such that \[\gamma_1\leq C_1\Biggl(\frac{\rho_n}{\sigma_n}+\frac{\Gamma_n}{\sigma_n}+\frac{\Gamma_n^2}{\sigma_n^2}\frac{\sum_{ i\not=j}|a_{i,j}|}{\sigma_n}\Biggr).\] The second term $\gamma_2$ in our bound in this case is of the same order as \begin{align*} \sigma_n^{-3}\sum_{\substack{J,K,L\in\mathcal{D}_{2}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_Ja_Ka_L| & \asymp \sigma_n^{-3}\Biggl(\sum_{i\not=j}|a_{i,j}|^3+\sum_{(i,j,k)\in[n]^3_{\not=}}|a_{i,j}|^2|a_{j,k}| \\ &\;+ \sum_{(i,j,k)\in[n]^3_{\not=}}|a_{i,j}||a_{j,k}||a_{k,i}|+ \sum_{(i,j,k,l)\in[n]^4_{\not=}}|a_{i,j}||a_{i,k}||a_{k,l}|\Biggr)\,, \end{align*} where, for positive sequences, we write $b_n\asymp d_n$ if there are $0<c<C<\infty$ such that $c b_n<d_n < Cb_n$ for all sufficiently large $n$. Note that we have \begin{align*} S_4&:=\sum_{(i,j,k)\in[n]^3_{\not=}}|a_{i,j}||a_{j,k}||a_{k,i}|=\sum_{i\not=j}|a_{i,j}|\sum_{k:k\not=i,j}|a_{j,k}||a_{k,i}|\\ &\leq \sum_{i\not=j}|a_{i,j}|\Bigl(\sum_{k:k\not=i,j}|a_{j,k}|^2\Bigr)^{1/2}\Bigl(\sum_{k:k\not=i,j}|a_{k,i}|^2\Bigr)^{1/2} \leq \rho_n^2\sum_{i\not=j}|a_{i,j}|\,,\\ S_5&:=\sum_{(i,j,k,l)\in[n]^4_{\not=}}|a_{i,j}||a_{i,k}||a_{k,l}|=\sum_{i\not=j}|a_{i,j}|\sum_{k:k\not=i,j}|a_{i,k}|\sum_{l:l\not=i,j,k}|a_{k,l}|\leq \Gamma_n^2\sum_{i\not=j}|a_{i,j}|\,. \end{align*} Thus, there is another absolute constant $C_2$ such that \[\gamma_2\leq C_2\Biggl(\frac{\rho_n}{\sigma_n}+\frac{\Gamma_n}{\sigma_n}+\frac{\Gamma_n^2}{\sigma_n^2}\frac{\sum_{ i\not=j}|a_{i,j}|}{\sigma_n}\Biggr).\] In particular, we obtain the asymptotic normality of $Y_n=\mathbf{Y}_n(1)$ under the assumption that \[\Gamma_n=o(\sigma_n) \quad\text{and}\quad \frac{\Gamma_n^2}{\sigma_n^2}=o\Biggl(\frac{\sigma_n}{\sum_{i\not=j}|a_{i,j}|}\Biggr)\,,\] which is a stronger condition than $\lambda_n^*=o(\sigma_n)$. However, if additionally the terms $\gamma_3, \gamma_4$ and $\gamma_5$ in Corollary \ref{corhumsums2} converge to zero, we can conclude the much stronger result that the whole process $\mathbf{Y}_n$ converges to a continuous Gaussian process on $[0,1]$. \item The literature around FCLTs for homogeneous sum processes is non-void but nevertheless extremely scarce. Indeed, the only references we have found, whose results might compare to ours (in the one-dimensional case) are \cite{Mik} and \cite{Basa}, of which \cite{Mik} only considers quadratic forms, i.e. the case $p=2$. It turns out that comparing our results to those in \cite{Mik} (for $p=2$) and to those in \cite{Basa} is complicated. Indeed, \cite[Theorem 1.6]{Mik} states the FCLT for the quadratic from $\mathbf{Y}_n$ under the (additional) assumption that $\|\tilde{A}\|^{-2}\| \tilde{A}^T\tilde{A}\|\to0$ as $n\to\infty$, where $\|\cdot\|$ denotes the Frobenius norm of a matrix and where $\tilde{A}=(\tilde{a}_{i,j})_{1\leq i,j\leq n}$ has entries $\tilde{a}_{i,j}=a_{i,j} 1_{\{i>j\}}$. Thus, the matrix $C:=\tilde{A}^T\tilde{A}$ has entries $c_{i,j}=\sum_{k=(i\vee j)+1}^{n} a_{i,k}a_{k,j}$ and, hence, its Frobenius norm is given by a quite complicated expression. Moreover, we have found that the argument leading to \cite[Theorem 1.1]{Basa} is flawed. Indeed, on page 187 therein, in the display below (2.9), one cannot simply drop the quantity $\tau_n^4$ (not even at the price of an enlarged absolute constant $C$) because the claimed inequality must hold for all fixed values of $n\in\mathbbm{N}$ (sufficiently large) and $t_1,t_2\in[0,1]$. Moreover, the application of \cite[Theorem 15.6]{Bill} on page 188 seems to be a bit rushed, since the almost sure left-continuity of the limiting Gaussian process $\xi_k$ is not verified. Moreover, the claimed limiting process $\xi_k$ appearing in \cite[Theorem 1.1]{Basa} is not even completely determined, since equation (1.4) theorof only specifies the one-dimensional distributions of $\xi_k$ but not its covariance function. \end{enumerate} \end{remark} \subsection{Example: runs on the line}\label{runs} Let $\xi_1,\dots,\xi_n$ be i.i.d. random variables, such that $\mathbbm{P}[\xi_1=1]=p=1-\mathbbm{P}[\xi_1=0]$, for $p\in(0,1)$. For any $1\leq r<n$ let $\sigma_n(r)=\sqrt{np^r(1-p)}$ and $V_r$ be the rescaled centred number of $r$-runs given by \[\mathbf{V}_n^{(r)}(t):=\frac{1}{\sigma_n(r)}\sum_{m=1}^{\lfloor nt\rfloor}\left(\xi_{m}\cdot\xi_{m+1}\cdot\ldots\cdot\xi_{m+r-1}-p^r\right),\quad t\in[0,1],\] where we adopt the torus convention, i.e. that $\xi_{n+1}=\xi_1, \xi_{n+2}=\xi_2$ and so on. A similar setup was considered in \cite{reinert_roellin}, where the authors studied the rate of the (finite-dimensional) weak convergence of the law of $\textbf{V}_n^{(r)}(1)$ to the normal distribution. The authors of \cite{reinert_roellin} note that the standard exchangeable-pair construction of \cite{RiRo97} does not lead to a bound going to zero as $n\to\infty$. In order to solve this problem, they apply their \textit{embedding method} and study the joint convergence of $\left(\textbf{V}_n^{(1)}(1),\dots,\textbf{V}_n^{(r)}(1)\right)$ to a multivariate normal law, using a slightly unusual construction of the exchangeable pair. Our propositions in this subsection provide bounds on the rate of the \textit{functional} convergence of $\left(\mathbf{V}_n^{(r_1)},\dots,\mathbf{V}_n^{(r_d)}\right)$ to a Gaussian process for \textit{any} collection $\lbrace r_1,\dots,r_d\rbrace$. They implicitly use the standard exchangeable-pair construction of Subsection \ref{intro_weighted}. Our bounds are of the same order as the bound on the rate of the (finite-dimensional) convergence provided in \cite{reinert_roellin}. We start with the following result on the pre-limiting approximation: \begin{proposition}\label{prop1runs} Adopt the notation from above. Let $d\geq 1$ and $\frac{n}{2}> r_1\geq r_2\geq\dots\geq r_d\geq 1$. Let \[\mathbf{V}_n=\left(\mathbf{V}_n^{(r_1)},\dots,\mathbf{V}_n^{(r_d)}\right).\] Let $\lbrace Z_J: J\in \mathcal{D}_j(n), \,j=1,\dots,r_1\rbrace$ be a collection of i.i.d. standard normal random variables. For $i=1,\dots,d$, let furthermore \[\mathbf{D}_n^{(r_i)}(t)=\frac{1}{\sigma_n(r_i)}\sum_{m=1}^{\lfloor nt\rfloor}\sum_{j=1}^{r_i}\sum_{0\leq i_1<\dots<i_j\leq r_i-1}p^{r-j}Z_{m+i_1,\dots,m+i_j},\quad t\in[0,1].\] and \[\mathbf{D}_n=\left(\mathbf{D}_n^{(r_1)},\dots,\mathbf{D}_n^{(r_d)}\right).\] Then, for any $g\in M^0$, \[\left|\mathbbm{E}g(\mathbf{V}_n)-\mathbbm{E}g(\mathbf{D}_n)\right|\leq \|g\|_{M^0}\left(\gamma_1+\gamma_2\right)n^{-1/2},\] where \begin{align*} \gamma_1=&\frac{2\sqrt{dr_1}\left(\sum_{i=1}^dr_i\right)^{3/2}}{3r_d}\sum_{i=1}^d\sum_{j=1}^{r_i}\frac{(1+p^3-2p^4)^jp^{3r_i/2-3j}}{(1-p)^{3/2}}{r_i-1\choose j-1}^3;\\ \gamma_2=&2\sqrt{dr_1}\left(\sum_{i=1}^dr_i\right)\sum_{u,v,w=1}^d\sum_{j_1=1}^{r_u} \sum_{j_2=1}^{r_v}\sum_{j_3=1}^{r_w}\frac{\bigl(1+p^3-2p^4\bigr)^{(j_1+j_2+j_3)/3}p^{(r_u+r_v+r_w)/2-j_1-j_2-j_3}}{(1-p)^{3/2}}\\ &\phantom{....................................................................}\cdot r_w(r_u\vee r_v)^2{r_u-1\choose j_1-1}{r_v-1\choose j_2-1}{r_w-1\choose j_3-1}. \end{align*} \end{proposition} \begin{proof}~\\ \textbf{Step 1.} For $i=1,2,\dots$, let $X_i=\xi_i-p$. It is easy to prove, by induction on $r$, that \begin{align}\label{decomposition} \mathbf{V}_n^{(r)}(t)=&\frac{1}{\sigma_n(r)}\sum_{m=1}^{\lfloor nt\rfloor}\sum_{j=1}^{r}\sum_{0\leq i_1<\dots<i_j\leq r-1}p^{r-j}X_{m+i_1}\dots X_{m+i_j},\quad t\in[0,1] \end{align} Indeed, for any $m=1,\dots,n$, \begin{align*} \xi_m-p=X_m \end{align*} and, assuming that \begin{align}\label{induc_hyp} \xi_m\xi_{m+1}\dots\xi_{m+r-1}-p^r=\sum_{j=1}^{r}\sum_{0\leq i_1<\dots<i_j\leq r-1}p^{r-j}X_{m+i_1}\dots X_{m+i_j}, \end{align} we have \begin{align*} &\xi_m\xi_{m+1}\dots\xi_{m+r}-p^{r+1}\\ =&\left(\xi_m\xi_{m+1}\dots\xi_{m+r-1}-p^r\right)\left(\xi_{m+r}-p\right)+p\left(\xi_m\xi_{m+1}\dots\xi_{m+r-1}-p^r\right)+p^r\left(\xi_{m+r}-p\right)\\ \stackrel{(\ref{induc_hyp})}{=}&\sum_{j=1}^{r}\sum_{0\leq i_1<\dots<i_j\leq r-1}p^{r-j}X_{m+i_1}\dots X_{m+i_j}X_{m+r}\\ &+\sum_{j=1}^{r}\sum_{0\leq i_1<\dots<i_j\leq r-1}p^{r+1-j}X_{m+i_1}\dots X_{m+i_j}+p^rX_{m+r}\\ =&\sum_{j=2}^{r+1}\sum_{0\leq i_1<\dots<i_j=r} p^{r+1-j}X_{m+i_1}\dots X_{m+i_j}\\ &+\sum_{j=1}^{r}\sum_{0\leq i_1<\dots<i_j\leq r-1}p^{r+1-j}X_{m+i_1}\dots X_{m+i_j}+p^rX_{m+r}\\ =&\sum_{j=1}^{r+1}\sum_{0\leq i_1<\dots<i_j\leq r}p^{r+1-j}X_{m+i_1}\dots X_{m+i_j}, \end{align*} as required. \textbf{Step 2.} Now, for any $r=1,2,\dots,r_1$ and $j=1,\dots,r$, note that \begin{align*} &\frac{p^{r-j}}{\sigma_n(r)}\sum_{m=1}^{\lfloor nt\rfloor}\sum_{0\leq i_1<\dots<i_j\leq r-1}X_{m+i_1}\dots X_{m+i_j}\\ =&\frac{p^{r-j}}{\sigma_n(r)}\sum_{m=1}^{\lfloor nt\rfloor}\sum_{m\leq i_1<\dots<i_j\leq m+r-1}X_{i_1}\dots X_{i_j}\\ =&\frac{p^{r-j}}{\sigma_n(r)}\sum_{1\leq i_1<\dots<i_j\leq \lfloor nt\rfloor+r-1}\left(\left(r-i_j+i_1\right)\vee 0\right)X_{i_1}\dots X_{i_j}\\ =&\frac{p^{r-j}}{\sigma_n(r)}\sum_{J\in\mathcal{D}_j \left((\lfloor nt\rfloor+r-1)\wedge n\right)}a_J(r)X_{i_1}\dots X_{i_j}, \end{align*} for \begin{align*} a_J(r):=&p^{r-j}\max\left(r-\max(J)+\min(J), 0\right)\\ &\hspace{-1cm}+p^{r-j}\max\left(r+\min(J\cap(n/2,n])-\max(J\cap[1,n/2))-n, 0\right)\mathbbm{1}_{\lbrace J\cap[1,n/2)\neq\emptyset\neq J\cap(n/2,n]\rbrace}. \end{align*} Furthermore, let \begin{align*} \mathbf{U}_n^{(r,j)}(t)=\frac{1}{\sigma_n(r)}\sum_{J\in D_{j}(\lfloor nt\rfloor)}a_J(r)\prod_{i\in J}X_i,\quad t\in\left[0,1\right] \end{align*} and define function $f:\left(D\left([0,1],\mathbbm{R}\right)\right)^{r_1+\dots+r_d}\to D\left([0,1],\mathbbm{R}^d\right)$, given by \begin{align*} &f\left(x_{1,1},\dots,x_{1,r_1},x_{2,2},\dots,x_{2,r_2},\dots,x_{d,1},\dots,x_{d,r_d}\right)\\ =&\left(\left(\sum_{j=1}^{r_1}x_{1,j}\left(\left(t+\frac{r_1-1}{n}\right)\wedge 1\right),\dots,\sum_{j=1}^{r_d}x_{d,j}\left(\left(t+\frac{r_d-1}{n}\right)\wedge 1\right)\right),t\in[0,1]\right). \end{align*} Hence, note that, by \eqref{decomposition}, \begin{align*} g\left(\mathbf{V}_n\right)=g\circ f\left(\mathbf{U}_n^{(r_1,1)},\dots,\mathbf{U}_n^{(r_1,r_1)},\dots,\mathbf{U}_n^{(r_d,1)},\dots,\mathbf{U}_n^{(r_d,r_d)}\right). \end{align*} It is proved in Lemma \ref{lem_prop1runs} in Section \ref{sec_prop1runs} of the Appendix that \begin{align}\label{m0_bound} \|g\circ f\|_{M^0}\leq \|g\|_{M^0}\sqrt{dr_1}\sum_{i=1}^dr_i. \end{align} \textbf{Step 3.} Now, note that, for $r,r_u,r_v,r_w\in\{1,2,\dots,r_1\}$, \begin{align} 1)\quad &\sum_{l=1}^n\left(\underset{l\in J}{\sum_{J\in\mathcal{D}_{j}(n):}}|a_J(r)|\right)^3\leq p^{3r-3j}\sum_{l=1}^n\left(\sum_{m=l-r+1}^{l}{r-1\choose j-1}\right)^3=p^{3r-3j}r^3{r-1\choose j-1}^3n\notag\\ 2)\quad & \sum_{\substack{J\in\mathcal{D}_{j_1}(n),\\K\in\mathcal{D}_{j_2}(n),\\L\in \mathcal{D}_{j_3}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_J(r_u)a_K(r_v)a_L(r_w)|\notag\\ \leq&\frac{p^{r_u+r_v+r_w-j_1-j_2-j_3}}{r_u\wedge r_v}\notag\\ &\hspace{1cm}\cdot\sum_{l=1}^n\sum_{m_1=l-r_u+1}^l\sum_{m_2=l-r_v+1}^l\sum_{k=l-r_u\vee r_v+1}^{l+r_u\vee r_v-1}\sum_{m_3=k-r_w+1}^k{r_u-1\choose j_1-1}{r_v-1\choose j_2-1}{r_w-1\choose j_3-1}\notag\\ \leq&2p^{r_u+r_v+r_w-j_1-j_2-j_3}r_w(r_u\vee r_v)^2{r_u-1\choose j_1-1}{r_v-1\choose j_2-1}{r_w-1\choose j_3-1}n \label{final_bounds} \end{align} and so, using \eqref{final_bounds} and \eqref{m0_bound}, for any $g\in M^0$, \begin{align*} A)\quad&\|g\circ f\|_{M^0}\frac{2\sqrt{\sum_{i=1}^dr_i}}{3r_d}\sum_{i=1}^d\sum_{j=1}^{r_i}\frac{\bigl(\mathbbm{E}|X_1|^3\bigr)^{j}}{\sigma_n(r_i)^3}\sum_{l=1}^n \left(\sum_{\substack{J\in\mathcal{D}_{j}(n):\\ l\in J}} |a_J(r_i)|\right)^3\\ \leq&\|g\|_{M^0}\frac{2\sqrt{dr_1}\left(\sum_{i=1}^dr_i\right)^{3/2}}{3r_d}\sum_{i=1}^d\sum_{j=1}^{r_i}\frac{(1+p^3-2p^4)^jp^{3r_i/2-3j}}{(1-p)^{3/2}}{r_i-1\choose j-1}^3n^{-1/2};\\ \leq&\|g\|_{M^0}\frac{2\sqrt{dr_1}\left(\sum_{i=1}^dr_i\right)^{3/2}}{3r_d}\sum_{i=1}^d\sum_{j=1}^{r_i}\frac{(1+p^3-2p^4)^jp^{3r_i/2-3j}}{(1-p)^{3/2}}{r_i-1\choose j-1}^3n^{-1/2};\\ B)\quad &\|g\circ f\|_{M^0}\sum_{u,v,w=1}^d\sum_{j_1=1}^{r_u} \sum_{j_2=1}^{r_v}\sum_{j_3=1}^{r_w}\frac{\bigl(\mathbbm{E}|X_1|^3\bigr)^{(j_1+j_2+j_3)/3}}{\sigma_n(r_u)\sigma_n(r_v)\sigma_n(r_w)}\sum_{\substack{J\in\mathcal{D}_{j_1}(n),\\K\in\mathcal{D}_{j_2}(n),\\L\in \mathcal{D}_{j_3}(n):\\J\cap K\not=\emptyset,\\L\cap(J\cup K)\not=\emptyset}}|a_J(r_u)a_K(r_v)a_L(r_w)|\\ \leq &2\sqrt{dr_1}\left(\sum_{i=1}^dr_i\right)\sum_{u,v,w=1}^d\sum_{j_1=1}^{r_u} \sum_{j_2=1}^{r_v}\sum_{j_3=1}^{r_w}\frac{\bigl(1+p^3-2p^4\bigr)^{(j_1+j_2+j_3)/3}p^{(r_u+r_v+r_w)/2-j_1-j_2-j_3}}{(1-p)^{3/2}}\\ &\phantom{...........................................................}\cdot r_w(r_u\vee r_v)^2 {r_u-1\choose j_1-1}{r_v-1\choose j_2-1}{r_w-1\choose j_3-1}n^{-1/2}. \end{align*} The result now follows by Corollary \ref{corhumsums1}. \end{proof} Next, we deal with the continuous process approximation as given in Corollary \ref{corhumsums2}. For this, we need to either compute or estimate the quantities $\delta_n^{(i)}$, $T_n^{(i)}$ and $\Sigma_n^{(m)}$. After rearranging the entries of the random vector according to their order as homogeneous sums, we can write $\Sigma_n^{(m)}$ as a block diagonal matrix. More precisely, for $1\leq q\leq r_1$ letting \begin{align}\label{n} N(q):=\max\{1\leq j\leq d\,:\, r_j\geq q\}\,, \end{align} we can write $\Sigma_n^{(m)}$ as a block diagonal matrix with blocks $\Sigma_n^{(m)}(1),\dotsc,\Sigma_n^{(m)}(r_1)$, where, for fixed $q=1,\dotsc,r_1$, $\Sigma_n^{(m)}(q)$ is an $N(q)\times N(q)$ matrix, namely the covariance matrix of the random vector \[\Bigl(\hspace{-1mm}\sqrt{n}\mathbf{U}_n^{(r_1,q)}(m/n)-\sqrt{n}\mathbf{U}_n^{(r_1,q)}((m-1)/n),\dotsc,\sqrt{n}\mathbf{U}_n^{(r_{N(q)},q)}(m/n)-\sqrt{n}\mathbf{U}_n^{(r_{N(q)},q)}((m-1)/n)\hspace{-1mm}\Bigr)^T\hspace{-1mm}.\] A simple computation shows that, for $q>1$ and $r_i\wedge r_l\leq m\leq n+1-r_i\wedge r_l$, \begin{align*} \Sigma_n^{(m)}(q)(i,l)&=\frac{n}{\sigma_n(r_i)\sigma_n(r_l)}\sum_{\substack{J\in\mathcal{D}_q(n):\\ \max (J)=m}} a_J(r_i)a_J(r_l)\\ &=\frac{p^{\frac{r_i+r_l}{2}-q}}{1-p} \sum_{k=q-1}^{r_i\wedge r_l -1}{k-1\choose q-2}(r_i-k)(r_l-k). \end{align*} Otherwise, for $q>1$ and $m\geq n+2-r_i\wedge r_l$, \begin{align*} \Sigma_n^{(m)}(q)(i,l)=&\frac{p^{\frac{r_i+r_l}{2}-q}}{1-p}\Bigg[\sum_{k=q-1}^{r_i\wedge r_l -1}{k-1\choose q-2}(r_i-k)(r_l-k)\\ &\hspace{3cm}+\sum_{u=n+2-r_i\wedge r_l}^m\sum_{k=(q-1)\vee (n-u-1)}^{r_i\wedge r_l-1}{k-1\choose q-2}(r_i-k)(r_l-k)\Bigg]. \end{align*} Moreover, for $q>1$ and $m\leq r_i\wedge r_l-1$, \begin{align*} \Sigma_n^{(m)}(q)(i,l)=\frac{p^{\frac{r_i+r_l}{2}-q}}{1-p}\sum_{k=q-1}^{m-1}{k-1\choose q-2}(r_i-k)(r_l-k), \end{align*} and, for all $1\leq m\leq n$ \[\Sigma_n^{(m)}(1)(i,l)=\frac{p^{\frac{r_i+r_l}{2}-1}}{1-p}r_ir_l.\] Hence, we let $\Sigma$ be a block diagonal matrix with blocks $\Sigma(1)\in\mathbbm{R}^{N(1)\times N(1)},\dots,\Sigma(r_1)\in\mathbbm{R}^{N(r_1)\times N(r_1)}$, where \begin{align}\label{sigma1} \Sigma(1)(i,l)=\frac{p^{\frac{r_i+r_l}{2}-1}}{1-p}r_ir_l \end{align} and for any $q=2,\dots,r_1$ and $i,l=1,\dots,N(q)$, \begin{align}\label{sigma2} \Sigma(q)(i,l)&=\frac{p^{\frac{r_i+r_l}{2}-q}}{1-p} \sum_{k=q-1}^{r_i\wedge r_l -1}{k-1\choose q-2}(r_i-k)(r_l-k). \end{align} Note that, for $\varphi(s)\equiv\Sigma^{1/2}$ and $\varphi_n(s)=\sum_{m=1}^n\left(\Sigma_n^{(m)}\right)^{1/2}\mathbbm{1}_{\left[\frac{m-1}{n},\frac{m}{n}\right]}(s)$, $s\in[0,1]$, \begin{align*} &\int_0^1\|\varphi_n(s)-\varphi(s)\|_F^2ds\\ \leq& \frac{2(r_1-1)}{n}\left[\sum_{m=1}^{r_1-1}\left\|\left(\Sigma_n^{(m)}\right)^{1/2}-\Sigma^{1/2}\right\|_F^2+\sum_{m=n+2-r_1}^{n}\left\|\left(\Sigma_n^{(m)}\right)^{1/2}-\Sigma^{1/2}\right\|_F^2\right]\\ \leq &\frac{4(r_1-1)}{n}\sum_{k=1}^d\sum_{i=1}^{r_k}\left[\sum_{m=1}^{r_1-1}\left(\left|\left(\Sigma_n^{(m)}\right)_{i,i}\right|+\left|\Sigma_{i,i}\right|\right)+\sum_{m=n+2-r_1}^{n}\left(\left|\left(\Sigma_n^{(m)}\right)_{i,i}\right|+\left|\Sigma_{i,i}\right|\right)\right]\\ \leq &\frac{24(r_1)^3}{n}\sum_{q=1}^{r_1}\sum_{i=1}^{N(q)}\sum_{k=q-1}^{r_i-1}\left({k-1\choose q-2}\mathbbm{1}_{[q>1]}+\mathbbm{1}_{[q=1]}\right)\frac{p^{r_i-q}}{1-p}(r_i-k)^2. \end{align*} Moreover, with obvious notation, \begin{align*} T_n^{(i)}(q)=\frac{1}{\left(\sigma_n{(r_i)}\right)^2}\sum_{J\in\mathcal{D}_{q}(n)}a_J(r_i)^2 =&\frac{1}{n}\sum_{m=1}^n\Sigma_n^{(m)}(q)(i,i)\\ =&\begin{cases} \frac{p^{r_i-1}}{1-p}r_i^2,\quad&\text{if }q=1,\\ \frac{p^{r_i-q}}{1-p}\sum_{k=q-1}^{r_i -1}{k-1\choose q-2}(r_i-k)^2,\quad &\text{if }q>1. \end{cases} \end{align*} Furthermore, for $q>1$, \begin{align*} &\delta_n^{(i)}(q)\\ =&\frac{1}{\left(\sigma_n{(r_i)}\right)^2}\sup_{m\in[n]}\underset{m=\max(J)}{\underset{{J\in\mathcal{D}_{q}(m):}}{\sum}}a_J(i)^2\\ =&\frac{p^{r_i-q}}{n(1-p)}\sum_{k=q-1}^{r_i -1}{k-1\choose q-2}(r_i-k)^2+\frac{p^{r_i-q}}{n(1-p)}\sum_{u=n+2-r_i}^n\sum_{k=(q-1)\vee (n-u-1)}^{r_i-1}{k-1\choose q-2}(r_i-k)^2 \end{align*} and \[\delta_n^{(i)}(1)=\frac{p^{r_i-1}}{n(1-p)}r_i^2.\] Therefore, for all $q=1,\dots,r_1$, \[\frac{1}{n}T_n^{(i)}(q)\leq\delta_n^{(i)}(q)\leq \frac{r_i}{n}T_n^{(i)}(q).\] Thus, taking \eqref{m0_bound} into account, we note that \begin{align*} &\|g\circ f\|_{M^0}\left[12\sqrt{\sum_{q=1}^{r_1}\sum_{i=1}^{N(q)}\delta_n^{(i)}(q)\log\left(\frac{2T_n^{(i)}(q)}{\delta_n^{(i)}(q)}\right)}+2\sqrt{\int_0^1\|\varphi_n(s)-\varphi(s)\|_F^2ds}\right]\\ \leq&\|g\|_{M^0}\sqrt{dr_1}\sum_{j=1}^dr_j\left[12\frac{\sqrt{\log n}}{\sqrt{n}}\Biggl(\sum_{q=2}^{r_1}\sum_{i=1}^{N(q)}\sum_{k=q-1}^{r_i -1}{k-1\choose q-2}\frac{p^{r_i-q}r_i}{(1-p)}(r_i-k)^2+\sum_{i=1}^d\frac{p^{r_i-1}}{1-p}r_i^3\Biggr)^{1/2}\right.\\ &\left.\phantom{...........}+\frac{4\sqrt{6}(r_1)^{3/2}}{\sqrt{n}}\left(\sum_{q=2}^{r_1}\sum_{i=1}^{N(q)}\sum_{k=q-1}^{r_i-1}{k-1\choose q-2}\frac{p^{r_i-q}}{1-p}(r_i-k)^2+\sum_{i=1}^d\frac{p^{r_i-1}}{1-p}r_i^2\right)^{1/2}\right]\\ \leq&\|g\|_{M^0}4\sqrt{d}r_1^2\left(\sum_{j=1}^dr_j\right)\left(\sum_{q=2}^{r_1}\sum_{i=1}^{N(q)}\sum_{k=q-1}^{r_i-1}{k-1\choose q-2}\frac{p^{r_i-q}}{1-p}(r_i-k)^2+\sum_{i=1}^d\frac{p^{r_i-1}}{1-p}r_i^2\right)^{1/2}\\ &\hspace{11cm}\cdot\frac{3\sqrt{\log n}+\sqrt{6}}{\sqrt{n}}. \end{align*} Hence, using Corollary \ref{corhumsums2} and Proposition \ref{prop1runs} (and noting that reordering the arguments of function $f$ does not change the bound on $\|g\circ f\|_{M^0}$ obtained in Lemma \ref{lem_prop1runs}), we obtain the following result: \begin{proposition}\label{prop2runs} Adopt the notation form above. In particular, let $N$ be as in \eqref{n}, $\mathbf{V}_n$ be defined as in Proposition \ref{prop1runs} and $\Sigma$ be the block diagonal matrix with blocks $\Sigma(1)\in\mathbbm{R}^{N(1)\times N(1)},\dots,\Sigma(r_1)\in\mathbbm{R}^{N(r_1)\times N(r_1)}$ defined by \eqref{sigma1} and \eqref{sigma2}. Let $\mathbf{Z}'=\Sigma^{1/2}\mathbf{W}$, where $\mathbf{W}$ is a $(\sum_{i=1}^d r_i)$-dimensional standard Brownian motion and write $\mathbf{Z}'=\left(\left(\mathbf{Z}'\right)^{(1)},\left(\mathbf{Z}'\right)^{(2)},\dots\right)$. Set $N(0)=0$. For $i=1,\dots,d$ and $t\in[0,1]$, define \begin{align*} \mathbf{Z}^{(i)}(t)=&\left(\left(\mathbf{Z}'\right)^{(i)}+\left(\mathbf{Z}'\right)^{(N(1)+i)}+\left(\mathbf{Z}'\right)^{(N(1)+N(2)+i)}+\dots+\left(\mathbf{Z}'\right)^{(N(1)+N(2)+\dots,N(r_i-1)+i)}\right)\\ &\hspace{9cm}\cdot\left(\left(t+\frac{r_i-1}{n}\right)\wedge 1\right) \end{align*} and let \[\mathbf{Z}=\left(\mathbf{Z}^{(1)},\dots,\mathbf{Z}^{(d)}\right).\] Then, for any $g\in M^0$, we have \[\left|\mathbbm{E}g(\mathbf{V}_n)-\mathbbm{E}g(\mathbf{Z})\right|\leq n^{-1/2}\|g\|_{M^0}\left(\gamma_1+\gamma_2+\gamma_3\sqrt{\log n}\right),\] where $\gamma_1$ and $\gamma_2$ are as in Proposition \ref{prop1runs} and \[\gamma_3=22\sqrt{d}r_1^2\left(\sum_{j=1}^dr_j\right)\left(\sum_{q=2}^{r_1}\sum_{i=1}^{N(q)}\sum_{k=q-1}^{r_i-1}{k-1\choose q-2}\frac{p^{r_i-q}}{1-p}(r_i-k)^2+\sum_{i=1}^d\frac{p^{r_i-1}}{1-p}r_i^2\right)^{1/2}\,.\] \end{proposition} \begin{remark} Assuming that $d, r_1,\dots,r_d$ are all fixed and do not depend on $n$, the bound in Proposition \ref{prop2runs} is of order $\sqrt{\frac{\log n}{n}}$. Therefore, by Proposition \ref{prop_m}, weak convergence of the law of $\mathbf{V}_n$ to that of $\mathbf{Z}$, in both the Skorokhod and the uniform topologies on the Skorokhod space, follows immediately from Proposition \ref{prop2runs} as a corollary. \end{remark} \begin{remark} It is possible to obtain bounds similar to those in Propositions \ref{prop1runs} and \ref{prop2runs} for the larger class of test functions $M$. It would, however, require some more involved computations, which would make the discussion of this example rather long. \end{remark} \section{Edge and two-star counts in Erd\H{o}s-Renyi random graphs}\label{section6} In this section we study an Erd\H{o}s-Renyi random graph with a fixed edge probability $p$ and $\lfloor nt\rfloor$ edges for $t\in[0,1]$. We analyse the asymptotic behaviour of the joint law of its (rescaled) number of edges and its (rescaled) number of two-stars (i.e. subgraphs which are trees with one internal node and $2$ leaves). Hence, we extend the result of \cite{kasprzak3}, where the univariate process convergence of the rescaled number of edges is studied. We also extend the analysis of \cite{reinert_roellin1}, whose authors provide a bound on the distance between the (three-dimensional) joint law of the (rescaled) number of edges, two-stars and triangles in a $G(n,p)$ graph and a Gaussian vector. In Theorem \ref{theorem_pre_limiting}, we establish a bound on the distance between our process and a pre-limiting Gaussian processes with paths in $D([0,1],\mathbbm{R}^2)$. Then, in Theorem \ref{theorem_continuous}, a bound on the quality of a continuous Gaussian process approximation is provided. It is worth noting that the analysis of a three-dimensional process representing the number of edges, triangles and two-stars in a $G(\lfloor nt\rfloor, p)$ graph does not pose any additional challenges except that it makes the algebraic computations more involved. The only reason we do not do it here is that it would make this section rather lengthy. \subsection{Introduction}\label{section1} Consider an Erd\H{o}s-Renyi random graph $G(\lfloor nt\rfloor,p)$ on $\lfloor nt\rfloor$ vertices, for $t\in[0,1]$, with a fixed edge probability $p$. Let $I_{i,j}=I_{j,i}$'s be i.i.d. Bernoulli$(p)$ random variables indicating that edge $(i,j)$ is present in this graph. We consider the following process, representing the re-scaled total number of edges \begin{align}\label{t_n} \mathbf{T}_n(t)=\frac{\lfloor nt\rfloor-2}{2n^2}\sum_{1\leq i\neq j\leq\lfloor nt\rfloor} I_{i,j}=\frac{\lfloor nt\rfloor-2}{n^2}\sum_{1\leq i<j\leq \lfloor nt\rfloor}I_{i,j}, \end{align} and a re-scaled statistic related to the number of two-stars \begin{align}\label{v_n} \mathbf{V}_n(t)=\frac{1}{6n^2}\underset{i,j,k\text{ distinct}}{\sum_{1\leq i,j,k\leq \lfloor nt\rfloor}}I_{ij}I_{jk}=\frac{1}{n^2}\sum_{1\leq i<j<k\leq \lfloor nt\rfloor}\left(I_{i,j}I_{j,k}+I_{i,j}I_{i,k}+I_{j,k}I_{i,k}\right). \end{align} Furthermore, let $\mathbf{Y}_n(t)=\left(\mathbf{T}_n(t)-\mathbbm{E}\mathbf{T}_n(t),\mathbf{V}_n(t)-\mathbbm{E}\mathbf{V}_n(t)\right)$ for $t\in\left[0,1\right]$. \begin{remark} Note that, for all $t\in[0,1]$, $\mathbbm{E}\mathbf{T}_n(t)=\frac{\lfloor nt\rfloor -2}{n^2}{\lfloor nt\rfloor \choose 2}p$ and $\mathbbm{E}\mathbf{V}_n(t)=\frac{3}{n^2}{\lfloor nt\rfloor \choose 3}p^2$ and, by an argument similar to that of \cite[Section 5]{reinert_roellin1}, the covariance matrix of $\left(\mathbf{T}_n(t)-\mathbbm{E}\mathbf{T}_n(t),\mathbf{V}_n(t)-\mathbbm{E}\mathbf{V}_n(t)\right)$ is given by \[3\frac{{\lfloor nt\rfloor \choose 3}}{n^4}p(1-p)\left(\begin{array}{cc} (\lfloor nt\rfloor -2)&2p(\lfloor nt\rfloor -2)\\ 2p(\lfloor nt\rfloor -2)&4p^2(\lfloor nt\rfloor -2)+p(1-p) \end{array}\right).\] The scaling therefore ensures that the covariances are of the same order in $n$. \end{remark} \subsection{Exchangeable pair setup}\label{section2} In order to construct a suitable exchangeable pair, following \cite{reinert_roellin1}, we pick $(I,J)$ according to $\mathbbm{P}[I=i,J=j]=\frac{1}{{n\choose 2}}$ for $1\leq i<j\leq n$. If $I=i,J=j$, we replace $I_{i,j}=I_{j,i}$ by an independent copy $I_{i,j}'=I_{j,i}'$ and set: \begin{align*} \mathbf{T}_n'(t)&=\mathbf{T}_n(t)-\frac{\lfloor nt\rfloor-2}{n^2}\left(I_{I,J}-I_{I,J}'\right)\mathbbm{1}_{[I/n,1]\cap[J/n,1]}(t)\\ \mathbf{V}_n'(t)&=\mathbf{V}_n(t)-\frac{1}{n^2}\sum_{k:k\neq I,J}\left(I_{I,J}-I_{I,J}'\right)\left(I_{J,k}+I_{I,k}\right)\mathbbm{1}_{[I/n,1]\cap[J/n,1]\cap[k/n,1]}(t). \end{align*} We, similarly, let $\mathbf{Y}_n'(t)=\left(\mathbf{T}_n'(t)-\mathbbm{E}\mathbf{T}_n(t),\mathbf{V}_n'(t)-\mathbbm{E}\mathbf{V}_n(t)\right)$ and note that, for $\mathbf{Y}_n=\left(\mathbf{Y}_n(t),t\in[0,1]\right)$ and $\mathbf{Y}_n'=\left(\mathbf{Y}_n'(t),t\in[0,1]\right)$, $(\mathbf{Y}_n,\mathbf{Y}_n')$ forms an exchangeable pair. Note that, for any $m=1,2$, any $f\in M$, as defined in Section \ref{section_notation}, and $e_1,e_2$ denoting the canonical basis vectors $(1,0)$ and $(0,1)$, respectively, we have \begin{align*} &\mathbbm{E}^{\mathbf{Y}_n}\left\lbrace Df(\mathbf{Y}_n)\left[\left(\mathbf{T}_n'-\mathbf{T}_n\right)e_m\right]\right\rbrace\\ =&\mathbbm{E}^{\mathbf{Y}_n}\left\lbrace Df(\mathbf{Y}_n)\left[\frac{\lfloor n\cdot\rfloor-2}{n^2}\left(I_{I,J}'-I_{I,J}\right)\mathbbm{1}_{[I/n,1]\cap[J/n,1]}e_m\right]\right\rbrace\\ =&\frac{2}{n^3(n-1)}\sum_{i<j}\mathbbm{E}^{\mathbf{Y}_n}\left\lbrace Df(\mathbf{Y}_n)\left[(\lfloor n\cdot\rfloor-2)\left(I_{i,j}'-I_{i,j}\right)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}e_m\right]|I=i,J=j\right\rbrace\\ =&-\frac{1}{{n\choose 2}}Df(\mathbf{Y}_n)[\mathbf{T}_ne_m]+\frac{2}{n^3(n-1)}p\sum_{i<j}Df(\mathbf{Y}_n)\left[(\lfloor n\cdot\rfloor-2)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}e_m\right]\\ =&-\frac{1}{{n\choose 2}}Df(\mathbf{Y}_n)[\left(\mathbf{T}_n(\cdot)-\mathbbm{E}\mathbf{T}_n(\cdot)\right)e_m].\\ \end{align*} Also: \begin{align*} &\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)[(\mathbf{V}_n-\mathbf{V}_n')e_m]\\ =&\frac{1}{n^2{n\choose 2}}\sum_{i<j}\mathbbm{E}^{\mathbf{Y}_n}\Bigg\{\sum_{k:k\neq i,j}Df(\mathbf{Y}_n)\bigg[\left(I_{i,j}-I_{i,j}'\right)\left(I_{j,k}+I_{i,k}\right)\\ &\hspace{7cm}\cdot\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}e_m\bigg]\Bigg|\,I=i,J=j\bigg\}\\ =&\frac{2}{{n\choose 2}}Df(\mathbf{Y}_n)[\mathbf{V}_ne_m]\\ &\hspace{2cm}-\frac{p}{n^2{n\choose 2}}\sum_{i<j}\sum_{k:k\neq i,j}\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)\left[\left(I_{j,k}+I_{i,k}\right)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}e_m\right]\\ =&\frac{2}{{n\choose 2}}Df(\mathbf{Y}_n)[\mathbf{V}_ne_m]-\frac{p}{n^2{n\choose 2}}\underset{i,j,k\text{ distinct}}{\sum_{1\leq i,j,k\leq n}}\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)\left[I_{i,j}\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}e_m\right]\\ =&\frac{2}{{n\choose 2}}Df(\mathbf{Y}_n)[\left(\mathbf{V}_n-\mathbbm{E}\mathbf{V}_n(\cdot)\right)e_m]\\ &\hspace{3cm}-\frac{p}{n^2{n\choose 2}}\underset{i,j,k\text{ distinct}}{\sum_{1\leq i,j,k\leq n}}\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)\left[(I_{i,j}-p)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}e_m\right]\\ =&\frac{2}{{n\choose 2}}Df(\mathbf{Y}_n)[\left(\mathbf{V}_n-\mathbbm{E}\mathbf{V}_n(\cdot)\right)e_m]\\ &\hspace{3cm}-\frac{2p}{{n\choose 2}}Df(\mathbf{Y}_n)\left[\frac{1}{\lfloor n\cdot\rfloor-2}\left(\mathbf{T}_n-\mathbbm{E}\mathbf{T}_n(\cdot)\right)e_m\left(\sum_{k=1}^n\mathbbm{1}_{[k/n,1]}-2\right)\right]\\ =&\frac{2}{{n\choose 2}}Df(\mathbf{Y}_n)[\left(\mathbf{V}_n-\mathbbm{E}\mathbf{V}_n(\cdot)\right)e_m]-\frac{2p}{{n\choose 2}}Df(\mathbf{Y}_n)\left[(\mathbf{T}_n-\mathbbm{E}\mathbf{T}_n(\cdot))e_m\right]. \end{align*} Therefore, for any $m=1,2$: \begin{align*} \text{A)}\quad Df(\mathbf{Y}_n)\left[\left(\mathbf{T}_n-\mathbbm{E}\mathbf{T}_n\right)e_m\right]=&\frac{n(n-1)}{2}\mathbbm{E}^{\mathbf{Y}_n}\left\lbrace Df(\mathbf{Y}_n)\left[(\mathbf{T}_n-\mathbf{T}_n')e_m\right]\right\rbrace\\ \text{B)}\quad Df(\mathbf{Y}_n)\left[\left(\mathbf{V}_n-\mathbbm{E}\mathbf{V}_n\right)e_m\right] =&\frac{n(n-1)}{4}\mathbbm{E}^{\mathbf{Y}_n}\bigg\{ Df(\mathbf{Y}_n)\left[(\mathbf{V}_n-\mathbf{V}_n')e_m\right]\\ &+pDf(\mathbf{Y}_n)\left[\left(\mathbf{T}_n-\mathbbm{E}\mathbf{T}_n\right)e_m\right]\bigg\}\\ =&\frac{n(n-1)}{4}\mathbbm{E}^{\mathbf{Y}_n}\left\lbrace Df(\mathbf{Y}_n)\left[\left(2p(\mathbf{T}_n-\mathbf{T}_n')+\mathbf{V}_n-\mathbf{V}_n'\right)e_m\right]\right\rbrace \end{align*} and so \[Df(\mathbf{Y}_n)[\mathbf{Y}_n]=2\mathbbm{E}^{\mathbf{Y}_n}Df(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\right],\] where \begin{equation}\label{eq_lambda} \Lambda_n=\frac{n(n-1)}{8}\left(\begin{array}{ccc} 2&2p\\ 0&1 \end{array}\right). \end{equation} Therefore, condition (\ref{condition}) is satisfied with $\Lambda_n$ of (\ref{eq_lambda}) and $R_f=0$. \subsection{A pre-limiting process}\label{section3} Suppose that the collection $\lbrace Z^{(1)}_{i,j}: i,j\in [n], i<j\rbrace\cup\lbrace Z^{(2)}_{i,j,k}: i,j,k\in [n], i<j<k\rbrace$ is jointly centred Gaussian with the following covariance structure: \begin{align*} &\mathbbm{E}Z_{ij}^{(1)}Z_{kl}^{(1)}=\begin{cases} \frac{p(1-p)}{n^4},&\text{if }(i,j)=(k,l),\\ 0,&\text{otherwise,}\end{cases}\\ &\mathbbm{E}Z_{i,j,k}^{(2)}Z_{l,m}^{(1)}=\begin{cases} \frac{2p^2(1-p)}{n^4},&\text{if }\lbrace l,m\rbrace\subset\lbrace i,j,k\rbrace,\\ 0,&\text{otherwise,} \end{cases}\\ &\mathbbm{E}Z_{i,j,k}^{(2)}Z_{r,s,t}^{(2)}=\begin{cases} \frac{3p^2(1+2p-3p^2)}{n^4},&\text{if }(i,j,k)=(r,s,t),\\ \frac{4p^3(1-p)}{n^4},& \text{if }\left|\lbrace i,j,k\rbrace\cap\lbrace r,s,t\rbrace\right|=2.\\ 0,&\text{otherwise.} \end{cases} \end{align*} Let $\mathbf{D}_n=(\mathbf{D}_n^{(1)},\mathbf{D}_n^{(2)})$ be defined in the following way: \begin{align*} &\mathbf{D}_n^{(1)}(t)=\left(\lfloor nt\rfloor-2\right)\sum_{1\leq i<j\leq\lfloor nt\rfloor}Z_{i,j}^{(1)},\quad t\in[0,1]\\ &\mathbf{D}_n^{(2)}(t)=\sum_{1\leq i<j<k\leq\lfloor nt\rfloor}Z_{i,j,k}^{(2)},\quad t\in[0,1]. \end{align*} Note that the covariance structure of the collection $\lbrace Z^{(1)}_{i,j}: i,j\in [n], i<j\rbrace\cup\lbrace Z^{(2)}_{i,j,k}: i,j,k\in [n], i<j<k\rbrace$ is the same as the covariance structure of the summands in the formulas \eqref{t_n} and \eqref{v_n}. \subsection{Distance from the pre-limiting process} We provide an estimate of the distance between $\mathbf{Y}_n$ and the pre-limiting piecewise constant Gaussian process. \begin{theorem}\label{theorem_pre_limiting} Let $\mathbf{Y}_n$ be defined as in Section \ref{section1} and $\mathbf{D}_n$ be defined as in Section \ref{section3}. Then, for any $g\in M$, \[\left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{D}_n)\right|\leq 23\|g\|_{M}n^{-1}.\] \end{theorem} \begin{remark} Our bound in Theorem \ref{theorem_pre_limiting} is of the same order as the analogous bound obtained in \cite{reinert_roellin1} on the distance between the (finite-dimensional) distributions of $\textbf{Y}_n(1)$ and $\textbf{D}_n(1)$. \end{remark} The proof is based on Theorem \ref{theorem1}. In \textbf{Step 1} we estimate term $\epsilon_1$, which involves bounding $\|\Lambda_n\|_2$ of (\ref{eq_lambda}) and the third moment of $\|\mathbf{Y}_n-\mathbf{Y}_n'\|$. In \textbf{Step 2} we treat $\epsilon_2$, using involved computations, which are, to a large extent, postponed to the appendix. Term $\epsilon_3$ is equal to zero as $R_f$ of Section \ref{section2} is equal to zero. \begin{proof}[Proof of Theorem \ref{theorem_pre_limiting}] We adopt the notation of sections \ref{section1}, \ref{section2}, \ref{section3} and apply Theorem \ref{theorem1}. \textbf{Step 1.} First note that, for $\epsilon_1$ in Theorem \ref{theorem1}, \begin{align*} |(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n|\leq \|\Lambda_n\|_2|\mathbf{Y}_n-\mathbf{Y}_n'|, \end{align*} where $|\cdot|$ denotes the Euclidean norm in $\mathbbm{R}^2$ and $\|\cdot\|_2$ is the induced operator $2$-norm. Furthermore, \begin{align*} \|\Lambda_n\|_2\leq \|\Lambda_n\|_F=\frac{n(n-1)}{8}\sqrt{2^2+(2p)^2+0^2+1^2}\leq \frac{3n(n-1)}{8}, \end{align*} for $\|\cdot\|_F$ denoting the Frobenius norm (which, for $\Theta\in\mathbbm{R}^{d_1\times d_2}$ is defined by $\|\Theta\|_F=\sqrt{\sum_{i=1}^{d_1}\sum_{j=1}^{d_2}|\Theta_{i,j}|}$). Therefore: \begin{align} &\mathbbm{E}\left[\|(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n\|\|\mathbf{Y}_n-\mathbf{Y}_n'\|^2\right]\nonumber\\ \leq&\frac{3n(n-1)}{8} \mathbbm{E}\|\mathbf{Y}_n-\mathbf{Y}_n'\|^3\nonumber\\ \leq& \frac{3n(n-1)}{8}\mathbbm{E}\left[\frac{(n-2)^2}{n^4}\left(I_{I,J}-I'_{I,J}\right)^2+\frac{1}{n^4}\left(\sum_{k:k\neq I,J}(I_{I,J}-I_{I,J}')\left(I_{J,k}+I_{I,k}\right)\right)^2\right]^{3/2}\nonumber\\ \leq&\frac{3n(n-1)}{8}\left[\frac{(n-2)^2}{n^4}+\frac{\left(2(n-2)\right)^2}{n^4}\right]^{3/2}\nonumber\\ \leq&\frac{5}{n},\nonumber \end{align} where the third inequality follows because $|I_{I,J}-I_{I,J}'|\leq 1$ and $|I_{J,k}+I_{I,k}|\leq 2$ for all $k$ and \begin{align} \epsilon_1\leq \frac{5\|g\|_{M}}{6n}.\label{4.1.1.1} \end{align} \textbf{Step 2.} In order to deal with $\epsilon_2$ in Theorem \ref{theorem1}, we need to bound \begin{align} &\left|\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(\mathbf{Y}_n-\mathbf{Y}_n'\right)\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]\right|\nonumber\\ =&\left|\frac{n(n-1)}{8}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(2(\mathbf{T}_n-\mathbf{T}_n'),2p(\mathbf{T}_n-\mathbf{T}_n')+(\mathbf{V}_n-\mathbf{V}_n')\right),\left(\mathbf{T}_n-\mathbf{T}_n',\mathbf{V}_n-\mathbf{V}_n'\right)\right]\right.\nonumber\\ &\left.-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]\right|\nonumber\\ \leq&S_1+S_2+S_3+S_4,\label{4_int} \end{align} where: \begin{align} S_1&=\Bigg|\frac{n(n-1)}{8}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(2(\mathbf{T}_n-\mathbf{T}_n'),0\right),(\mathbf{T}_n-\mathbf{T}_n',0)\right]\notag\\ &\hspace{7cm}-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(\mathbf{D}_n^{(1)},0\right),\left(\mathbf{D}_n^{(1)},0\right)\right]\Bigg|\nonumber\\ S_2&=\left|\frac{n(n-1)}{8}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(0,2p(\mathbf{T}_n-\mathbf{T}_n')+\mathbf{V}_n-\mathbf{V}_n'\right),(\mathbf{T}_n-\mathbf{T}_n',0)\right]\right.\nonumber\\ &\left.\hspace{7cm}-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(0,\mathbf{D}_n^{(2)}\right),\left(\mathbf{D}_n^{(1)},0\right)\right]\right|\nonumber\\ S_3&=\Bigg|\frac{n(n-1)}{8}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(2(\mathbf{T}_n-\mathbf{T}_n'),0\right),(0,\mathbf{V}_n-\mathbf{V}_n')\right]\notag\\ &\hspace{7cm}-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(\mathbf{D}_n^{(1)},0\right),\left(0,\mathbf{D}_n^{(2)}\right)\right]\Bigg|\nonumber\\ S_4&=\left|\frac{n(n-1)}{8}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(0,2p(\mathbf{T}_n-\mathbf{T}_n')+\mathbf{V}_n-\mathbf{V}_n'\right),(0,\mathbf{V}_n-\mathbf{V}_n')\right]\right.\nonumber\\ &\left.\phantom{............................................................................}-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\left(0,\mathbf{D}_n^{(2)}\right),\left(0,\mathbf{D}_n^{(2)}\right)\right]\right|\label{s1s2}. \end{align} In Lemma \ref{lemma8_app}, in the appendix, we obtain the following estimates: \begin{align} &S_1\leq \frac{\sqrt{5}\|g\|_{M}}{12n}, \quad S_2\leq\frac{\sqrt{178}\|g\|_{M}}{6n}, \quad S_3\leq \frac{\sqrt{178}\|g\|_{M}}{6n},\quad S_4\leq\frac{(\sqrt{612}+\sqrt{178})\|g\|_{M}}{3n}. \label{s_est} \end{align} Note that, therefore, by (\ref{4_int}) and (\ref{s_est}), \begin{align} &\epsilon_2=\left|\mathbbm{E}D^2f(\mathbf{Y}_n)\left[(\mathbf{Y}_n-\mathbf{Y}_n')\Lambda_n,\mathbf{Y}_n-\mathbf{Y}_n'\right]-\mathbbm{E}D^2f(\mathbf{Y}_n)\left[\mathbf{D}_n,\mathbf{D}_n\right]\right| \leq 18\|g\|_{M}n^{-1}.\label{4.8} \end{align} Using Theorem \ref{theorem1} together with (\ref{4.8}) and (\ref{4.1.1.1}) gives the desired result. \end{proof} \subsection{Distance from the continuous process} We now study the approximation of $\mathbf{Y}_n$ by a continuous Gaussian process with covariance equal to the limit of the covariance of $\mathbf{D}_n$. We obtain a bound on the quality of this approximation. This is achieved by applying Theorem \ref{theorem_pre_limiting} and by bounding the distance between $\mathbf{D}_n$ and the continuous process via the Brownian modulus of continuity. \begin{theorem}\label{theorem_continuous} Let $\mathbf{Y}_n$ be defined as in Subsection \ref{section1} and let $\mathbf{Z}=(\mathbf{Z}^{(1)},\mathbf{Z}^{(2)})$ be defined by: \[\begin{cases} \mathbf{Z}^{(1)}(t)=\frac{\sqrt{p(1-p)}}{\sqrt{2+8p^2}}t\mathbf{B}_1(t^2)+\frac{p\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}t\mathbf{B}_2(t^2),\\ \mathbf{Z}^{(2)}(t)=\frac{p\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}t\mathbf{B}_1(t^2)+\frac{2p^2\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}t\mathbf{B}_2(t^2) \end{cases},\] where $\mathbf{B}_1,\mathbf{B}_2$ are independent standard Brownian Motions. Then, for any $g\in M$: \[\left|\mathbbm{E}g(\mathbf{Y}_n)-\mathbbm{E}g(\mathbf{Z})\right|\leq \|g\|_{M}\left(16422n^{-1/2}\sqrt{\log n}+138n^{-1/2}\right).\] \end{theorem} \begin{remark} \item Theorem \ref{theorem_continuous}, together with Proposition \ref{prop_m}, implies that $\mathbf{Y}_n$ converges to $\mathbf{Z}$ in distribution with respect to the Skorokhod and uniform topologies. \end{remark} \begin{remark} Theorem \ref{theorem_continuous} can be adapted to situations in which $p=p_n$ varies with $n$. More precisely, as indicated by the necessary and sufficient conditions for approximate normality of the marginal distributions given in \cite{Ruc}, Theorem \ref{theorem_continuous} can be modified to yield a quantitative functional CLT in the case that $n^3p_n^2\to\infty$ and $n^2(1-p_n)\to\infty$. \end{remark} In \textbf{Step 1} of the proof of Theorem \ref{theorem_continuous}, we use i.i.d standard Brownian Motions to construct a process $\mathbf{Z}_n$ having the same distribution as $\mathbf{D}_n$. In \textbf{Step 2} we couple $\mathbf{Z}_n$ and $\mathbf{Z}$ and use the Brownian modulus of continuity to bound moments of the supremum distance between them. In \textbf{Step 3} we combine those bounds with the mean value theorem to obtain the desired final estimate. \begin{proof}[Proof of Theorem \ref{theorem_continuous}]~\\ \textbf{Step 1.} Let $\textbf{B}_3$ be another standard Brownian Motion, mutually independent with $\textbf{B}_1$ and $\textbf{B}_2$. Let $\mathbf{Z}_n=\left(\mathbf{Z}_n^{(1)},\mathbf{Z}_n^{(2)}\right)$ be defined by: \begin{align*} \text{A)}\quad \mathbf{Z}_n^{(1)}(t)=&\frac{(\lfloor nt\rfloor -2)\sqrt{p(1-p)}}{n^2\sqrt{2+8p^2}}\mathbf{B}_1\left(\lfloor nt\rfloor(\lfloor nt\rfloor -1)\right)\\ &+\frac{(\lfloor nt\rfloor -2)p\sqrt{2p(1-p)}}{n^2\sqrt{1+4p^2}}\mathbf{B}_2\left(\lfloor nt\rfloor(\lfloor nt\rfloor -1)\right);\\ \text{B)}\quad \mathbf{Z}_n^{(2)}(t)=&\frac{(\lfloor nt\rfloor -2)p\sqrt{2p(1-p)}}{n^2\sqrt{1+4p^2}}\mathbf{B}_1\left(\lfloor nt\rfloor(\lfloor nt\rfloor -1)\right)\\ &+\frac{(\lfloor nt\rfloor -2)2p^2\sqrt{2p(1-p)}}{n^2\sqrt{1+4p^2}}\mathbf{B}_2\left(\lfloor nt\rfloor(\lfloor nt\rfloor -1)\right)\\ &+\frac{p(1-p)}{n^{2}\sqrt{2}}\mathbf{B}_3\left(\lfloor nt\rfloor(\lfloor nt\rfloor-1)(\lfloor nt\rfloor-2)\right). \end{align*} Now, note that $\left(\mathbf{D}_n^{(1)},\mathbf{D}_n^{(2)}\right)\stackrel{\mathcal{D}}=\left(\mathbf{Z}_n^{(1)},\mathbf{Z}_n^{(2)}\right)$. To see this, observe that for all $u,t\in[0,1]$, \begin{align} \text{A)}\quad&\mathbbm{E}\mathbf{D}_n^{(1)}(t)\mathbf{D}_n^{(1)}(u)\notag\\ =&(\lfloor nt\rfloor -2)(\lfloor nu\rfloor -2)\lfloor n(t\wedge u)\rfloor(\lfloor n(t\wedge u)\rfloor-1)\frac{p(1-p)}{2n^4}\nonumber\\ =&\mathbbm{E}\mathbf{Z}_n^{(1)}(t)\mathbf{Z}_n^{(1)}(u)\nonumber;\\ \text{B)}\quad& \mathbbm{E}\textbf{D}_n^{(2)}(t)\textbf{D}_n^{(2)}(u)\notag\\ =&{\lfloor n(t\wedge u)\rfloor\choose 3}\frac{3p^2(1+2p-3p^2)}{n^4}\notag\\ &+{\lfloor n(t\wedge u)\rfloor\choose 2}\left[(\lfloor nt\rfloor -2)(\lfloor nu\rfloor-2)-\left(\lfloor n(t\wedge u)\rfloor -2\right)\right]\frac{4p^3(1-p)}{n^4}\notag\\ =&\lfloor n(t\wedge u)\rfloor(\lfloor n(t\wedge u)\rfloor-1)\frac{4p^3(1-p)(\lfloor nt\rfloor -2)(\lfloor nu\rfloor-2)+(\lfloor n(t\wedge u)\rfloor-2)p^2(1-p)^2}{2n^4}\notag\\ =&\mathbbm{E}\textbf{Z}^{(2)}(t)\textbf{Z}_{n}^{(2)};\notag\\ \text{C)}\quad&\mathbbm{E}\mathbf{D}_n^{(1)}(t)\mathbf{D}_n^{(2)}(u)\notag\\ =&(\lfloor nt\rfloor -2)(\lfloor nu\rfloor -2)\lfloor n(t\wedge u)\rfloor(\lfloor n(t\wedge u)\rfloor-1)\frac{p^2(1-p)}{n^4}\nonumber\\ =&\mathbbm{E}\mathbf{Z}_n^{(1)}(t)\mathbf{Z}_n^{(2)}(u).\label{cov_structure} \end{align} \textbf{Step 2.} We now let $\mathbf{Z}$ be constructed as in Theorem \ref{theorem_continuous}, using the same Brownian Motions $\mathbf{B}_1,\mathbf{B}_2$, as the ones used in the construction of $\mathbf{Z}_n$. In Lemma \ref{lemma10_app}, proved in the appendix, we obtain the following bounds: \begin{align} &\mathbbm{E}\left\|\mathbf{Z}_n-\mathbf{Z}\right\|\leq\frac{8}{\sqrt{n}}+\frac{39\sqrt{\log n}}{\sqrt{n}}\notag\\ &\mathbbm{E}\left\|\mathbf{Z}_n-\mathbf{Z}\right\|^3\leq\frac{49}{n^{3/2}}+\frac{8167(\log n)^{3/2}}{n^{3/2}}\notag\\ &\mathbbm{E}\|\mathbf{Z}\|^2\leq \frac{4}{3}.\label{4.1.2.1} \end{align} \textbf{Step 3.} We note that, by (\ref{4.1.2.1}): \begin{align*} \left|\mathbbm{E}g(\mathbf{Z})-\mathbbm{E}g(\mathbf{D}_n)\right|\stackrel{\text{MVT}}\leq&\mathbbm{E}\left[\sup_{c\in[0,1]}\left\|Dg(\mathbf{Z}+c(\mathbf{Z}_n-\mathbf{Z}))\right\|\|\mathbf{Z}-\mathbf{Z}_n\|\right]\\ \leq&\|g\|_{M}\mathbbm{E}\left[\sup_{c\in[0,1]}\left(1+\|\mathbf{Z}+c(\mathbf{Z}_n-\mathbf{Z})\|^2\right)\|\mathbf{Z}-\mathbf{Z}_n\|\right]\\ \leq&\|g\|_{M}\mathbbm{E}\left[\|\mathbf{Z}-\mathbf{Z}_n\|+\|\mathbf{Z}\|\|\mathbf{Z}-\mathbf{Z}_n\|+\|\mathbf{Z}-\mathbf{Z}_n\|^2\right]\\ \leq&\|g\|_{M}\left[\mathbbm{E}\|\mathbf{Z}-\mathbf{Z}_n\|+2\mathbbm{E}\|\mathbf{Z}-\mathbf{Z}_n\|^3+2\left(\mathbbm{E}\|\mathbf{Z}\|^3\right)^{2/3}\left(\mathbbm{E}\|\mathbf{Z}-\mathbf{Z}_n\|^3\right)^{1/3}\right]\\ \leq&\|g\|_{M}\left(\frac{115}{\sqrt{n}}+\frac{16422\sqrt{\log n}}{\sqrt{n}}\right), \end{align*} which, together with Theorem \ref{theorem_pre_limiting} gives the desired estimate. \end{proof} \begin{remark} The representation of $\mathbf{Z}$ in terms of two independent Brownian Motions comes from a careful analysis of the limiting covariance of $\mathbf{D}_n$, which may be derived using (\ref{cov_structure}). \end{remark} \section{Appendix - technical details of the proofs of Proposition \ref{prop1runs} and Theorems \ref{theorem_pre_limiting} and \ref{theorem_continuous}}\label{appendix} \subsection{Technical details of the proof of Proposition \ref{prop1runs} }\label{sec_prop1runs} \begin{lemma}\label{lem_prop1runs} Let $n,d\in\mathbbm{N}$ and $r_1\geq r_2\geq\dots\geq r_d\geq 1$. Define function \\ $f:\left(D\left([0,1],\mathbbm{R}\right)\right)^{r_1+\dots+r_d}\to D\left([0,1],\mathbbm{R}^d\right)$, given by \begin{align*} &f\left(x_{1,1},\dots,x_{1,r_1},x_{2,2},\dots,x_{2,r_2},\dots,x_{d,1},\dots,x_{d,r_d}\right)\\ =&\left(\left(\sum_{j=1}^{r_1}x_{1,j}\left(\left(t+\frac{r_1-1}{n}\right)\wedge 1\right),\dots,\sum_{j=1}^{r_d}x_{d,j}\left(\left(t+\frac{r_d-1}{n}\right)\wedge 1\right)\right),t\in[0,1]\right). \end{align*} Then , for any $g\in M^0$, \[\|g\circ f\|_{M^0}\leq \|g\|_{M^0}\sqrt{dr_1}\sum_{i=1}^dr_i.\] \end{lemma} \begin{proof} Note that function $f$ is twice Fr\'echet differentiable with \begin{align*} (A)\quad &Df(w)\left[\left(x_{1,1},\dots,x_{1,r_1},x_{2,1},\dots,x_{2,r_2},\dots,x_{d,1},\dots,x_{d,r_d}\right)\right]\\ =&\left(\left(\sum_{j=1}^{r_1}x_{1,j}\left(\left(t+\frac{r_1-1}{n}\right)\wedge 1\right),\dots,\sum_{j=1}^{r_d}x_{d,j}\left(\left(t+\frac{r_d-1}{n}\right)\wedge 1\right)\right),t\in[0,1]\right)\\ (B)\quad &D^2f(w)[x^{(1)},x^{(2)}]=0 \end{align*} for all $w,x^{(1)},x^{(2)},\left(x_{1,1},\dots,x_{1,r_1},x_{2,1},\dots,x_{2,r_2},\dots,x_{d,1},\dots,x_{d,r_d}\right)\in \left(D\left([0,1],\mathbbm{R}\right)\right)^{r_1+\dots+r_d}$. Furthermore, for any $w\in\left(D\left([0,1],\mathbbm{R}\right)\right)^{r_1+\dots+r_d}$, \begin{align*} a)&\quad \|f(w)\|\\ \leq&\sqrt{\sup_{t\in[0,1]}\left(\sum_{j=1}^{r_1}w_{1,j}\left(\left(t+\frac{r_1-1}{n}\right)\wedge 1\right)\right)^2+\dots+\sup_{t\in[0,1]}\left(\sum_{j=1}^{r_d}w_{d,j}\left(\left(t+\frac{r_d-1}{n}\right)\wedge 1\right)\right)^2}\\ \leq &\sqrt{\sum_{i=1}^d\sup_{t\in[0,1]}\left|\sum_{j=1}^{r_i}w_{i,j}(t)\right|^2}\\ b)&\quad \|Df(w)\|\leq \sqrt{\sum_{i=1}^dr_i}. \end{align*} Therefore, for any $w,h\in\left(D\left([0,1],\mathbbm{R}\right)\right)^{r_1+\dots+r_d}$, \begin{align} A)\quad &|g\circ f(w)|\leq\|g\|_{M^0};\notag\\ B)\quad&\left\|D(g\circ f)(w)\right\|=\left\|Dg(f(w))[Df(w)[\cdot]]\right\|\leq \|g\|_{M^0}\|Df(w)\| \leq\|g\|_{M^0}\sqrt{\sum_{i=1}^dr_i};\notag\\ C)\quad&\left\|D^2(g\circ f)(w)\right\|=\left\|D^2g(f(w))\left[Df(w),Df(w)\right]\right\|\leq \|g\|_{M^0}\left\|Df(w)\right\|^2\leq \|g\|_{M^0}\sum_{i=1}^dr_i;\notag\\ D)\quad&\left\|D^2(g\circ f)(w+h)-D^2(g\circ f)(w)\right\|\notag\\ =&\left\|D^2g(f(w+h))\left[Df(w+h),Df(w+h)\right]-D^2g(f(w))[Df(w),Df(w)]\right\|\notag\\ \leq &\left\|D^2g(f(w+h))\left[Df(w+h),Df(w+h)\right]-D^2g(f(w))\left[Df(w+h),Df(w+h)\right]\right\|\notag\\ &+\left\|D^2g(f(w))\left[Df(w+h),Df(w+h)\right]-D^2g(f(w))[Df(w),Df(w)]\right\|\notag\\ \leq &\|g\|_{M^0}\|f(w+h)-f(w)\|\|Df(w+h)\|^2\notag\\ \leq &\|g\|_{M^0}\left(\sum_{i=1}^d\sup_{t\in[0,1]}\left|\sum_{j=1}^{r_i}h_{i,j}(t)\right|^2\right)^{1/2}\sum_{i=1}^d r_i,\label{m01} \end{align} where $D)$ follows from the fact that $Df(w)=Df(w+h)$. Moreover, \begin{align} \frac{\left(\sum_{i=1}^d\sup_{t\in[0,1]}\left|\sum_{j=1}^{r_i}h_{i,j}(t)\right|^2\right)^{1/2}}{\sup_{t\in[0,1]}\left(\sum_{i=1}^d\sum_{j=1}^{r_i}h_{i,j}^2(t)\right)^{1/2}}\leq \frac{\sup_{t\in[0,1]}\left(dr_1\sum_{i=1}^d\sum_{j=1}^{r_i}h_{i,j}^2(t)\right)^{1/2}}{\sup_{t\in[0,1]}\left(\sum_{i=1}^d\sum_{j=1}^{r_i}h_{i,j}^2(t)\right)^{1/2}}=\sqrt{dr_1}.\label{m02} \end{align} Therefore, using \eqref{m01} and \eqref{m02}, \begin{align*} \|g\circ f\|_{M^0}\leq \|g\|_{M^0}\sqrt{dr_1}\sum_{i=1}^dr_i. \end{align*} \end{proof} \subsection{Technical details of the proof of Theorem \ref{theorem_pre_limiting}} \begin{lemma}\label{lemma8_app} For $S_i,i=1,2,3,4$ of (\ref{s1s2}), we have the following estimates: \begin{align*} &S_1\leq \frac{\sqrt{5}\|g\|_{M}}{12n}, \quad S_2\leq\frac{\sqrt{178}\|g\|_{M}}{6n}, \quad S_3\leq \frac{\sqrt{178}\|g\|_{M}}{6n},\quad S_4\leq\frac{(\sqrt{612}+\sqrt{178})\|g\|_{M}}{3n}. \end{align*} \end{lemma} \begin{proof} For $S_1$, for fixed $i,j\in\lbrace 1,\cdots, n\rbrace$, let $\mathbf{Y}_n^{ij}$ be equal to $\mathbf{Y}_n$ except for the fact that $I_{ij}$ is replaced by an independent copy, i.e. for all $t\in [0,1]$ let: \begin{align*} \mathbf{T}_n^{ij}(t)&=\mathbf{T}_n(t)-\frac{\lfloor nt\rfloor-2}{n^2}\left(I_{ij}-I_{ij}'\right)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}(t)\\ \mathbf{V}_n^{ij}(t)&=\mathbf{V}_n(t)-\frac{1}{n^2}\sum_{k:k\neq i,j}\left(I_{ij}-I_{ij}'\right)\left(I_{jk}+I_{ik}\right)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}(t) \end{align*} and let $\mathbf{Y}_n^{ij}(t)=\left(\mathbf{T}_n^{ij}(t)-\mathbbm{E}\mathbf{T}_n(t),\mathbf{V}_n^{ij}(t)-\mathbbm{E}\mathbf{V}_n(t)\right)$. By noting that the mean zero $Z_{i,k}^{(1)}$ and $Z_{i',j}^{(1)}$ are independent for $i\neq i'$, we obtain: \begin{align} S_1=&\Bigg|\vphantom{\sum_1^1}\frac{n(n-1)}{8}\mathbbm{E}D^2f(\mathbf{Y}_n)\left[(\mathbf{T}_n-\mathbf{T}_n')(2,0),(\mathbf{T}_n-\mathbf{T}_n')(1,0)\right]\nonumber\\ &-\mathbbm{E}D^2f(\mathbf{Y}_n)\bigg[\sum_{1\leq i< j\leq n}Z_{i,j}^{(1)}(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]},\notag\\ &\hspace{5cm}\sum_{1\leq i< j\leq n}Z_{i,j}^{(1)}(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\bigg]\Bigg|\nonumber\\ =&\left|\frac{1}{2n^4}\sum_{1\leq i< j\leq n}\mathbbm{E}\bigg\{\left(I_{i,j}-2pI_{i,j}+p\right)\right.\nonumber\\ &\hspace{1cm}\cdot D^2f(\mathbf{Y}_n)\left[(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]},(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\right]\bigg\}\nonumber\\ &-\sum_{1\leq i<j\leq n}\Bigg\{\mathbbm{E}\left(Z_{i,j}^{(1)}\right)^2\nonumber\\ &\left.\cdot\mathbbm{E}D^2f(\mathbf{Y}_n)\left[(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]},(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\right]\vphantom{\left(Z^{1}_2\right)^2}\bigg\}\right|\nonumber\\ =&\left|\sum_{1\leq i< j\leq n}\mathbbm{E}\left\lbrace \left(\frac{1}{2n^4}(I_{i,j}-2pI_{i,j}+p)-\mathbbm{E}\left(Z_{i,j}^{(1)}\right)^2\right)\right.\right.\nonumber\\ &\left.\left.\hphantom{\sum_{1\leq i\neq j\leq n}\mathbbm{E}}\cdot D^2f(\mathbf{Y}_n)\left[(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]},(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\right]\vphantom{\frac{1}{4n^2}}\right\rbrace\vphantom{\sum_1^2}\right|\nonumber\\ =&\Bigg|\sum_{1\leq i< j\leq n}\mathbbm{E}\bigg\{ \frac{1}{2n^4}(I_{i,j}-2pI_{i,j}+p)\left(D^2f(\mathbf{Y}_n)-D^2f(\mathbf{Y}_n^{ij})\right)\nonumber\\ &\hspace{3.5cm}\Big[(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]},(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\Big]\vphantom{\frac{1}{n^2}}\bigg\}\Bigg|\nonumber\\ \stackrel{\eqref{m_bound}}\leq&\frac{\|g\|_{M}}{6n^2}\sum_{1\leq i< j\leq n}\mathbbm{E}\left|(I_{i,j}-2pI_{i,j}+p)\right|\left\|\mathbf{Y}_n-\mathbf{Y}_n^{ij}\right\|\label{eq_star}, \end{align} where (\ref{eq_star}) follows from Proposition \ref{prop12.7}. Now, \[\left\|\mathbf{Y}_n-\mathbf{Y}_n^{ij}\right\|\leq\frac{1}{n^2}\sqrt{(\lfloor n\cdot\rfloor -2)^2(I_{ij}-I_{ij}')^2+\left(\sum_{k:k\neq i,j}|I_{ij}-I_{ij}'|(I_{jk}+I_{ik})\right)^2}\] and so, by (\ref{eq_star}), \begin{align} S_1\leq&\frac{\|g\|_{M}}{6n^4}\sum_{1\leq i< j\leq n}\mathbbm{E}\Bigg\{\left|I_{i,j}-2pI_{i,j}+p\right|\notag\\ &\hspace{4cm}\cdot\sqrt{(n-2)^2(I_{ij}-I_{ij}')^2+\bigg(\sum_{k\neq i,j}|I_{ij}-I_{ij}'|(I_{jk}+I_{ik})\bigg)^2}\Bigg\}\nonumber\\ \leq&\frac{\|g\|_{M}}{6n^3}\sum_{1\leq i<j\leq n}\mathbbm{E}\left\lbrace\left|I_{i,j}-2pI_{i,j}+p\right|\cdot\sqrt{(I_{ij}-I_{ij}')^2+\left(|I_{ij}-I_{ij}'|(I_{jk}+I_{ik})\right)^2}\right\rbrace\nonumber\\ \leq&\frac{\sqrt{5}\|g\|_{M}}{12n},\label{4.1} \end{align} where the last inequality holds because $|I_{ij}-2pI_{ij}+p|\leq 1$, $|I_{ij}-I_{ij}'|\leq 1$ and $I_{jk}+I_{ik}\leq 2$ for all $k\in\lbrace 1,\cdots,n\rbrace$. For $S_2$, let $\mathbf{Y}_n^{ijk}$ equal to $\mathbf{Y}_n$ except that $I_{ij},I_{jk},I_{ik}$ are replaced by $I_{ij}'$, $I_{jk}'$, $I_{ik}'$, i.e. for all $t\in[0,1]$ let \begin{align} \mathbf{T}_n^{ijk}(t)=&\mathbf{T}_n(t)-\frac{\lfloor nt\rfloor -2}{n^2}\left[(I_{ij}-I_{ij}')\mathbbm{1}_{[i/n,1]\cap[j/n,1]}(t)\right.\nonumber\\ &\left.+(I_{jk}-I_{jk}')\mathbbm{1}_{[j/n,1]\cap[k/n,1]}(t)+(I_{ik}-I_{ik}')\mathbbm{1}_{[i/n,1]\cap[k/n,1]}(t)\right]\nonumber\\ \mathbf{V}_n^{ijk}(t)=&\mathbf{V}_n(t)-\frac{1}{n^2}\sum_{l:l\neq i,j,k}\bigg[\left(I_{ij}-I_{ij}'\right)\left(I_{jl}+I_{il}\right)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[l/n,1]}(t)\nonumber\\ &\hspace{5cm}+\left(I_{jk}-I_{jk}'\right)\left(I_{jl}+I_{kl}\right)\mathbbm{1}_{[k/n,1]\cap[j/n,1]\cap[l/n,1]}(t)\notag\\ &\hspace{5cm}+\left(I_{ik}-I_{ik}'\right)\left(I_{jl}+I_{il}\right)\mathbbm{1}_{[i/n,1]\cap[k/n,1]\cap[l/n,1]}(t)\bigg]\nonumber\\ &\hspace{-1cm}-\frac{1}{n^2}\left[(I_{ij}I_{jk}-I_{ij}'I_{jk}')+(I_{ij}I_{ik}-I_{ij}'I_{ik}')+(I_{ik}I_{jk}-I_{ik}'I_{jk}')\right]\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}(t).\label{eq_ijk} \end{align} Let $\mathbf{Y}_n^{ijk}(t)=\left(\mathbf{T}_n^{ijk}(t)-\mathbbm{E}\mathbf{T}_n(t),\mathbf{V}_n^{ijk}(t)-\mathbbm{E}\mathbf{V}_n(t)\right)$ for all $t\in[0,1]$. Then \begin{align} S_2=&\Bigg|\vphantom{\sum_{1\leq i<j<k\leq n}}\frac{n(n-1)}{8}\mathbbm{E}\left\lbrace D^2f(\mathbf{Y}_n)\left[(\mathbf{T}_n-\mathbf{T}_n')(0,2p)+(\mathbf{V}_n-\mathbf{V}_n')(0,1),(\mathbf{T}_n-\mathbf{T}_n')(1,0)\right]\right\rbrace\nonumber\\ &-\mathbbm{E}D^2f(\mathbf{Y}_n)\bigg[\sum_{1\leq i<j<k\leq n}Z_{i,j,k}^{(2)}(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]},\notag\\ &\hspace{6cm}\sum_{1\leq i< j\leq n}Z_{i,j}^{(1)}(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\bigg]\Bigg|\nonumber\\ =&\left|\frac{1}{4n^4}\hspace{-1mm}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\left\lbrace \left[2p\left(I_{ij}-2pI_{ij}+p\right)+\left(I_{ij}-2pI_{ij}+p\right)(I_{jk}+I_{ik})-8p^2(1-p)\right]\right.\right.\notag\\ &\left.\left.\cdot \left(D^2f(\mathbf{Y}_n)-D^2f(\mathbf{Y}^{ijk}_n)\right)\left[(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]},(\lfloor n\cdot\rfloor-2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\right]\right\rbrace\vphantom{\underset{j}{\sum_{1\leq i<j<k\leq n}}}\right|\nonumber\\ \stackrel{(\ref{m_bound})}\leq &\frac{\|g\|_{M}}{12n^3}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\bigg\{\Big|2p\left(I_{ij}-2pI_{ij}+p\right)\notag\\ &\hspace{5cm}+\left(I_{ij}-2pI_{ij}+p\right)(I_{jk}+I_{ik})\Big|\|\mathbf{Y}_n-\mathbf{Y}_n^{ijk}\|\bigg\}\notag\\ \leq &\frac{\|g\|_{M}}{3n^3}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\|\mathbf{Y}_n-\mathbf{Y}_n^{ijk}\|. \label{4.2} \end{align} Now, by (\ref{eq_ijk}), we note that: \begin{align} &\|\mathbf{Y}_n-\mathbf{Y}_n^{ijk}\|\notag\\ \leq&\frac{1}{n^2}\left\lbrace(n-2)^2(|I_{ij}-I_{ij}'|+|I_{jk}-I_{jk}'|+|I_{ik}-I_{ik}'|)^2\vphantom{\left[\sum_{k\neq l}^l\right]^2}\right.\notag\\ &+\left[\sum_{l:l\neq i,j,k}\Big(|I_{ij}-I'_{ij}|(I_{jl}+I_{il})+|I_{jk}-I_{jk}'|(I_{jl}+I_{kl})+|I_{ik}-I_{ik}'|(I_{jl}+I_{il})\right.\notag\\ &\left.\left.+|I_{ik}-I_{ik}'|(I_{jl}+I_{il})\Big)+|I_{ij}I_{jk}-I_{ij}'I_{jk}'|+|I_{ij}I_{ik}-I_{ij}'I_{ik}'|+|I_{ij}I_{jk}-I_{ij}'I_{jk}'| \vphantom{\sum_{k\neq i}l}\right]^2\right\rbrace^{1/2}\notag\\ \leq&\frac{1}{n^2}\sqrt{ 9(n-2)^2+(8(n-3)+3)^2}\notag\\ =&\frac{\sqrt{73n^2-372n+477}}{n^2},\label{4.55} \end{align} where the second inequality follows from the fact that for all $a,b,c\in\lbrace 1,\cdots, n\rbrace$, $|I_{ab}-I_{ab}'|\leq 1$, $(I_{ab}+I_{bc})\leq 2$ and $|I_{ab}I_{bc}-I_{ab}'I_{bc}'|\leq 1$. Therefore, by (\ref{4.2}): \begin{align} S_2\leq&\frac{\|g\|_Mn(n-1)(n-2)\sqrt{73n^2-372n+477}}{6n^5}\leq \frac{\sqrt{178}\|g\|_{M}}{6n}.\label{4.3} \end{align} Similarly, for $S_3$, \begin{align} S_3=&\left|\vphantom{\sum_{1\leq i<j<k\leq n}}\frac{n(n-1)}{8}\mathbbm{E}\left\lbrace D^2f(\mathbf{Y}_n)\left[(\mathbf{T}_n-\mathbf{T}_n')(2,0),(\mathbf{V}_n-\mathbf{V}_n')(0,1)\right]\right\rbrace\right.\nonumber\\ &-\mathbbm{E}D^2f(\mathbf{Y}_n)\bigg[\sum_{1\leq i<j<k\leq n}Z_{i,j,k}^{(2)}(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]},\notag\\ &\hspace{6.5cm}\left.\sum_{1\leq i< j\leq n}Z_{i,j}^{(1)}(\lfloor n\cdot \rfloor -2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\bigg]\right|\nonumber\\ =&\left|\frac{1}{4n^4}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\left\lbrace \left[2\left(I_{ij}-2pI_{ij}+p\right)(I_{jk}+I_{ik})-8p^2(1-p)\right]\right.\right.\notag\\ &\left.\left.\cdot \left(D^2f(\mathbf{Y}_n)-D^2f(\mathbf{Y}^{ijk}_n)\right)\left[(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]},(\lfloor n\cdot\rfloor-2)(1,0)\mathbbm{1}_{[i/n,1]\cap[j/n,1]}\right]\right\rbrace\vphantom{\underset{j}{\sum_{1\leq i<j<k\leq n}}}\right|\nonumber\\ \stackrel{(\ref{m_bound})}\leq &\frac{\|g\|_{M}}{12n^3}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\left\lbrace\left|2\left(I_{ij}-2pI_{ij}+p\right)(I_{jk}+I_{ik})\right|\|\mathbf{Y}_n-\mathbf{Y}_n^{ijk}\|\right\rbrace\notag\\ \leq &\frac{\|g\|_{M}}{3n^3}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\|\mathbf{Y}_n-\mathbf{Y}_n^{ijk}\|\notag\\ \stackrel{\eqref{4.55}}\leq &\frac{\sqrt{178}\|g\|_{M}}{6n}. \label{s_3} \end{align} Now, for $S_4$, let $\mathbf{Y}_n^{ijkl}$ be equal to $\mathbf{Y}_n$ except that $I_{ij},I_{ik},I_{il},I_{jk},I_{jl},I_{kl}$ are replaced with independent copies $I_{ij}',I_{ik}',I_{il}',I_{jk}',I_{jl}',I_{kl}'$, i.e. for all $t\in[0,1]$ let \begin{align} \mathbf{T}_n^{ijkl}(t)=&\mathbf{T}_n(t)-\frac{\lfloor nt\rfloor -2}{n^2}\left[(I_{ij}-I_{ij}')\mathbbm{1}_{[i/n,1]\cap[j/n,1]}(t)+(I_{ik}-I_{ik}')\mathbbm{1}_{[i/n,1]\cap[k/n,1]}(t)\right.\nonumber\\ &+(I_{il}-I_{il}')\mathbbm{1}_{[i/n,1]\cap[l/n,1]}(t)+(I_{jk}-I_{jk}')\mathbbm{1}_{[j/n,1]\cap[k/n,1]}(t)\nonumber\\ &\left.+(I_{jl}-I_{jl}')\mathbbm{1}_{[j/n,1]\cap[l/n,1]}(t)+(I_{kl}-I_{kl}')\mathbbm{1}_{[k/n,1]\cap[l/n,1]}(t)\right]\nonumber\\ \mathbf{V}_n^{ijkl}(t)=&\mathbf{V}_n(t)-\frac{1}{n^2}\sum_{m:m\neq i,j,k,l}\left[\left(I_{ij}-I_{ij}'\right)\left(I_{im}+I_{jm}\right)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[m/n,1]}(t)\right.\nonumber\\ &\hspace{-1cm}+\left(I_{ik}-I_{ik}'\right)\left(I_{im}+I_{km}\right)\mathbbm{1}_{[i/n,1]\cap[k/n,1]\cap[m/n,1]}(t)\nonumber\\ &\hspace{-1cm}+\left(I_{il}-I_{il}'\right)\left(I_{im}+I_{lm}\right)\mathbbm{1}_{[i/n,1]\cap[l/n,1]\cap[m/n,1]}(t)\nonumber\\ &\hspace{-1cm}+\left(I_{jk}-I_{jk}'\right)\left(I_{jm}+I_{km}\right)\mathbbm{1}_{[j/n,1]\cap[k/n,1]\cap[m/n,1]}(t)\nonumber\\ &\hspace{-1cm}+\left(I_{jl}-I_{jl}'\right)\left(I_{jm}+I_{lm}\right)\mathbbm{1}_{[j/n,1]\cap[l/n,1]\cap[m/n,1]}(t)\nonumber\\ &\left.\hspace{-1cm}+\left(I_{kl}-I_{ll}'\right)\left(I_{km}+I_{lm}\right)\mathbbm{1}_{[k/n,1]\cap[l/n,1]\cap[m/n,1]}(t)\right]\nonumber\\ &\hspace{-1cm}-\frac{1}{n^2}\left[(I_{ij}I_{jk}-I_{ij}'I_{jk}')+(I_{ij}I_{ik}-I_{ij}'I_{ik}')+(I_{ik}I_{jk}-I_{ij}'I_{jk}')\right]\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}(t)\nonumber\\ &\hspace{-1cm}-\frac{1}{n^2}\left[(I_{ij}I_{jl}-I_{ij}'I_{jl}')+(I_{ij}I_{il}-I_{ij}'I_{il}')+(I_{il}I_{jl}-I_{ij}'I_{jl}')\right]\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[l/n,1]}(t)\nonumber\\ &\hspace{-1cm}-\frac{1}{n^2}\left[(I_{ik}I_{kl}-I_{ik}'I_{kl}')+(I_{ik}I_{il}-I_{ik}'I_{il}')+(I_{il}I_{kl}-I_{ik}'I_{kl}')\right]\mathbbm{1}_{[i/n,1]\cap[k/n,1]\cap[l/n,1]}(t)\nonumber\\ &\hspace{-1cm}-\frac{1}{n^2}\left[(I_{jk}I_{jl}-I_{jk}'I_{jl}')+(I_{jl}I_{kl}-I_{jl}'I_{kl}')+(I_{kl}I_{jk}-I_{kl}'I_{jk}')\right]\mathbbm{1}_{[j/n,1]\cap[k/n,1]\cap[l/n,1]}(t) \label{eq_ijkl} \end{align} and for all $t\in[0,1]$ let $\mathbf{Y}_n^{ijkl}(t)=\left(\mathbf{T}_n^{ijkl}(t)-\mathbbm{E}\mathbf{T}_n,\mathbf{V}_n^{ijkl}(t)-\mathbbm{E}\mathbf{V}_n(t)\right).$ Note that: \begin{align} S_4=&\Bigg|\frac{n(n-1)}{8}\mathbbm{E}\left\lbrace D^2f(\mathbf{Y}_n)\left[(\mathbf{T}_n-\mathbf{T}_n')(0,2p)+(\mathbf{V}_n-\mathbf{V}_n')(0,1),(\mathbf{V}_n-\mathbf{V}_n')(0,1)\right]\right\rbrace\nonumber\\ &-\mathbbm{E}D^2f(\mathbf{Y}_n)\bigg[\sum_{1\leq i<j<k\leq n}Z_{i,j,k}^{(2)}(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]},\notag\\ &\hspace{6.5cm}\sum_{1\leq i<j<k\leq n}Z_{i,j,k}^{(2)}(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}\bigg]\Bigg|\nonumber\\ \leq&\Bigg|\frac{1}{4n^4}\sum_{1\leq i<j\leq n}\underset{\lbrace k,l\rbrace\cap\lbrace i,j\rbrace=\emptyset}{\sum_{1\leq k\neq l\leq n}}\mathbbm{E}\bigg\{ \Big[2p\left(I_{ij}-2pI_{ij}+p\right)(I_{jk}+I_{ik})\notag\\ &\hspace{3cm}+(I_{ij}-2pI_{ij}+p)\left(I_{ik}I_{il}+I_{ik}I_{jl}+I_{jk}I_{il}+I_{jk}I_{jl}\right)-16p^3(1-p)\big]\notag\\ &\hspace{0.5cm}\cdot \left(D^2f(\mathbf{Y}_n)-D^2f(\mathbf{Y}^{ijkl}_n)\right)\left[(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[l/n,1]},(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}\right]\bigg\}\Bigg|\nonumber\\ &+\Bigg|\frac{1}{4n^4}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\bigg\{ \Big[2p\left(I_{ij}-2pI_{ij}+p\right)(I_{jk}+I_{ik})\notag\\ &\hspace{4cm}+(I_{ij}-2pI_{ij}+p)\left(I_{ik}+2I_{ik}I_{jk}+I_{jk}\right)-4p^2(1+2p-3p^2)\Big]\notag\\ &\cdot \left(D^2f(\mathbf{Y}_n)-D^2f(\mathbf{Y}^{ijk}_n)\right)\left[(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]},(0,1)\mathbbm{1}_{[i/n,1]\cap[j/n,1]\cap[k/n,1]}\right]\Big\}\Bigg|\nonumber\\ \leq &\frac{\|g\|_{M}}{12n^4}\sum_{1\leq i<j\leq n}\underset{\lbrace k,l\rbrace\cap\lbrace i,j\rbrace=\emptyset}{\sum_{1\leq k\neq l\leq n}}\mathbbm{E}\bigg\{\Big|\left(I_{ij}-2pI_{ij}+p\right)\notag\\ &\hspace{2.5cm}\cdot(2pI_{jk}+2pI_{ik}+I_{ik}I_{il}+I_{ik}I_{jk}+I_{jk}I_{il}+I_{jk}I_{jl})\Big|\cdot\|\mathbf{Y}_n-\mathbf{Y}_n^{ijkl}\|\bigg\}\notag\\ &+\frac{\|g\|_{M}}{12n^4}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\Big\{\left|\left(I_{ij}-2pI_{ij}+p\right)(2pI_{jk}+2pI_{ik}+I_{ik}+2I_{ik}I_{jk}+I_{jk})\right|\notag\\ &\hspace{10cm}\cdot\|\mathbf{Y}_n-\mathbf{Y}_n^{ijk}\|\Big\} \notag\\ \leq &\frac{2\|g\|_{M}}{3n^4}\sum_{1\leq i<j\leq n}\underset{\lbrace k,l\rbrace\cap\lbrace i,j\rbrace=\emptyset}{\sum_{1\leq k\neq l\leq n}}\mathbbm{E}\|\mathbf{Y}_n-\mathbf{Y}_n^{ijkl}\|+\frac{2\|g\|_{M}}{3n^4}\sum_{1\leq i<j\leq n}\underset{k\not\in\lbrace i,j\rbrace}{\sum_{1\leq k\leq n}}\mathbbm{E}\|\mathbf{Y}_n-\mathbf{Y}_n^{ijk}\|. \label{4.5} \end{align} Now, by (\ref{eq_ijkl}), note that: \begin{align*} &\|\mathbf{Y}_n-\mathbf{Y}_n^{ijkl}\|\\ \leq&\frac{1}{n^2}\Bigg\{\vphantom{\sum_{m:m\neq i,j,k,l}}(n-2)^2\left(|I_{ij}-I_{ij}'|+|I_{ik}-I_{ik}'|+|I_{il}-I_{i}'|+|I_{jk}-I_{jk}'|+|I_{jl}-I_{jl}'|+|I_{kl}-I_{kl}'|\right)^2\\ &+\Bigg[\sum_{m:m\neq i,j,k,l}\left[\left|I_{ij}-I_{ij}'\right|\left(I_{im}+I_{jm}\right)+\left|I_{ik}-I_{ik}'\right|\left(I_{im}+I_{km}\right)+\left|I_{il}-I_{il}'\right|\left(I_{im}+I_{lm}\right)\right.\\ &\left.\hphantom{\sum_{m:m\neq i,j,k,l}}+\left|I_{jk}-I_{jk}'\right|\left(I_{jm}+I_{km}\right)+\left|I_{jl}-I_{jl}'\right|\left(I_{jm}+I_{lm}\right)+\left|I_{kl}-I_{ll}'\right|\left(I_{km}+I_{lm}\right)\right]\\ &+|I_{ij}I_{jk}-I_{ij}'I_{jk}'|+|I_{ij}I_{ik}-I_{ij}'I_{ik}'|+|I_{ik}I_{jk}-I_{ij}'I_{jk}'|+|I_{ij}I_{jl}-I_{ij}'I_{jl}'|\\ &+|I_{ij}I_{il}-I_{ij}'I_{il}'|+|I_{il}I_{jl}-I_{ij}'I_{jl}'|+|I_{ik}I_{kl}-I_{ik}'I_{kl}'|+|I_{ik}I_{il}-I_{ik}'I_{il}'|\\ &+|I_{il}I_{kl}-I_{ik}'I_{kl}'|+|I_{jk}I_{jl}-I_{jk}'I_{jl}'|+|I_{jl}I_{kl}-I_{jl}'I_{kl}'|+|I_{kl}I_{jk}-I_{kl}'I_{jk}'|\Bigg]^2 \Bigg\}^{1/2}\\ \leq&\frac{\sqrt{36(n-2)^2+\left(12(n-4)+12\right)^2}}{n^2}\\ =&\frac{\sqrt{180n^2-1008n+1440}}{n^2}. \end{align*} Therefore, by (\ref{4.5}) and \eqref{4.55}, \begin{align} S_4\leq&\frac{\|g\|_{M}\cdot \sqrt{180n^2-1008n+1440}+\sqrt{73n^2-372n+477}}{3n^2}\leq \frac{\left(\sqrt{612}+\sqrt{178}\right)\|g\|_{M}}{3n}.\label{4.6} \end{align} The result now follows by (\ref{4.2}), (\ref{4.3}), (\ref{s_3}), (\ref{4.6}). \end{proof} \subsection{Technical details of the proof of Theorem \ref{theorem_continuous}} \begin{lemma}\label{lemma10_app} Using the notation of \textbf{Step 2} of the proof of Theorem \ref{theorem_continuous}, \begin{align*} &\mathbbm{E}\left\|\mathbf{Z}_n-\mathbf{Z}\right\|\leq\frac{8}{\sqrt{n}}+\frac{39\sqrt{\log n}}{\sqrt{n}}\\ &\mathbbm{E}\left\|\mathbf{Z}_n-\mathbf{Z}\right\|^3\leq\frac{49}{n^{3/2}}+\frac{8167(\log n)^{3/2}}{n^{3/2}}\\ &\mathbbm{E}\|\mathbf{Z}\|^2\leq \frac{4}{3}. \end{align*} \end{lemma} \begin{proof} Note the following: \begin{enumerate} \item By Doob's $L^2$ and $L^3$ inequalities, \begin{align} \hspace{-0.7cm}\text{A)}\,&\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\mathbf{B}_3\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)(\lfloor nt\rfloor -2)}{n^3}\right)\right|\right]\leq 2\sqrt{\mathbbm{E}\left[\left|\mathbf{B}_3\left(\frac{n(n -1)(n-2)}{n^3}\right)\right|^2\right]}\leq 2;\notag\\ \hspace{-0.7cm}\text{B)}\,&\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\mathbf{B}_3\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)(\lfloor nt\rfloor -2)}{n^3}\right)\right|^3\right]\leq \frac{27}{8}\mathbbm{E}\left[\left|\mathbf{B}_3\left(\frac{n(n -1)(n-2)}{n^3}\right)\right|^3\right]\leq \frac{27}{8}. \label{second_in} \end{align} \item By Doob's $L^2$ and $L^3$ inequality, for all $t\in[0,1]$, \begin{equation}\label{fourth_in_1} \mathbbm{E}\left[\sup_{t\in[0,1]}|\mathbf{B}_1(t^2)|\right]\leq 2,\quad \mathbbm{E}\left[\sup_{t\in[0,1]}|\mathbf{B}_1(t^2)|^3\right]\leq \frac{27}{8} \quad\text{and}\quad\left|\frac{\lfloor nt\rfloor-2}{n}-t\right|\leq \frac{3}{n}. \end{equation} \item Using \cite[Lemma 3]{ito_processes} and the fact that \[\left|\frac{\lfloor nt\rfloor(\lfloor nt\rfloor-1)}{n^2}-t^2\right|\leq \left|\frac{(nt-\lfloor nt\rfloor)(nt+\lfloor nt\rfloor)}{n^2}\right|+\frac{1}{n^2}\leq \frac{3}{n},\] we obtain \begin{align} &\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-\mathbf{B}_1(t^2)\right|\right]\leq \frac{30\sqrt{3\log\left(\frac{2n}{3}\right)}}{n^{1/2}\sqrt{\pi\log(2)}};\notag\\ &\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-\mathbf{B}_1(t^2)\right|^3\right]\leq\frac{1080\left(3\log\left(\frac{2n}{3}\right)\right)^{3/2}}{n^{3/2}\left(\pi\log(2)\right)^{3/2}}\label{fourth_in_2}. \end{align} \end{enumerate} Now, we can bound $\mathbbm{E}\left\|\mathbf{Z}_n-\mathbf{Z}\right\|$ in the following way: \begin{align*} &\mathbbm{E}\left\|\mathbf{Z}_n-\mathbf{Z}\right\|\nonumber\\ \leq&\frac{\sqrt{p(1-p)}}{\sqrt{2+8p^2}}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_1(t^2)\right|\right]\nonumber\\ &+\frac{p\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_2\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_2(t^2)\right|\right]\nonumber\\ &+\frac{p\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_1(t^2)\right|\right]\nonumber\\ &+\frac{2p^2\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_2\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_2(t^2)\right|\right]\nonumber\\ &+\frac{p(1-p)}{n^{1/2}}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\mathbf{B}_3\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)(\lfloor nt\rfloor -2)}{n^3}\right)\right|\right]\nonumber\\ \stackrel{(\ref{second_in})}\leq& \frac{(1+4p+4p^2)\sqrt{p(1-p)}}{\sqrt{2+8p^2}}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_1(t^2)\right|\right]\nonumber\\ &+\frac{2p(1-p)}{n^{1/2}}\nonumber\\ \leq&\frac{(1+4p+4p^2)\sqrt{p(1-p)}}{\sqrt{2+8p^2}}\left(\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\left(\frac{\lfloor nt\rfloor -2}{n}-t\right)\mathbf{B}_1(t^2)\right|\right]\right.\nonumber\\ &\left.+\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-\mathbf{B}_1(t^2)\right|\right]\right)+\frac{2p(1-p)}{n^{1/2}}\nonumber\\ \stackrel{(\ref{fourth_in_1}),(\ref{fourth_in_2})}\leq&\frac{(1+4p+4p^2)\sqrt{p(1-p)}}{\sqrt{2+8p^2}}\left(\frac{6}{n}+\frac{30\sqrt{3\log n}}{n^{1/2}\sqrt{\pi\log(2)}}\right)+\frac{2p(1-p)}{n^{1/2}}\nonumber\\ \leq&\frac{8}{\sqrt{n}}+\frac{39\sqrt{\log n}}{\sqrt{n}}. \end{align*} Similarly, \begin{align*} &\mathbbm{E}\|\mathbf{Z}_n-\mathbf{Z}\|^3\nonumber\\ \leq&4\sqrt{3}\left(\frac{p(1-p)}{2+8p^2}\right)^{3/2}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_1(t^2)\right|^3\right]\nonumber\\ &+4\sqrt{3}\left(\frac{2p^3(1-p)}{1+4p^2}\right)^{3/2}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_2\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_2(t^2)\right|^3\right]\nonumber\\ &+9\sqrt{3}\left(\frac{2p^3(1-p)}{1+4p^2}\right)^{3/2}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_1(t^2)\right|^3\right]\nonumber\\ &+9\sqrt{3}\left(\frac{8p^5(1-p)}{1+4p^2}\right)^{3/2}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_2\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_2(t^2)\right|^3\right]\nonumber\\ &+9\sqrt{3}\frac{p^3(1-p)^3}{n^{3/2}}\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\mathbf{B}_3\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)(\lfloor nt\rfloor -2)}{n^2}\right)\right|^3\right]\nonumber\\ \stackrel{(\ref{second_in})}\leq& \frac{\sqrt{6}p^{3/2}(1-p)^{3/2}(1+26p^3+126p^6)}{(1+4p^2)^{3/2}}\notag\\ &\hspace{1.5cm}\cdot\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\frac{\lfloor nt\rfloor -2}{n}\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-t\mathbf{B}_1(t^2)\right|^3\right]+\frac{243\sqrt{3}}{512n^{3/2}}\nonumber\\ \leq& \frac{4\sqrt{6}p^{3/2}(1-p)^{3/2}(1+26p^3+126p^6)}{(1+4p^2)^{3/2}}\left(\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\left(\frac{\lfloor nt\rfloor -2}{n}-t\right)\mathbf{B}_1(t^2)\right|^3\right]\right.\nonumber\\ &\left.+\mathbbm{E}\left[\sup_{t\in[0,1]}\left|\mathbf{B}_1\left(\frac{\lfloor nt\rfloor(\lfloor nt\rfloor -1)}{n^2}\right)-\mathbf{B}_1(t^2)\right|^3\right]\right)+\frac{243\sqrt{3}}{512n^{3/2}}\nonumber\\ \stackrel{(\ref{fourth_in_1}),(\ref{fourth_in_2})}\leq&\frac{4\sqrt{6}p^{3/2}(1-p)^{3/2}(1+26p^3+126p^6)}{(1+4p^2)^{3/2}}\left(\frac{81}{8n^3}+\frac{1080\left(3\log n\right)^{3/2}}{n^{3/2}\left(\pi\log(2)\right)^{3/2}}\right)+\frac{243\sqrt{3}}{512n^{3/2}}\nonumber\\ \leq&\frac{49}{n^{3/2}}+\frac{8167(\log n)^{3/2}}{n^{3/2}}. \end{align*} Furthermore, \begin{align*} \mathbbm{E}\|\mathbf{Z}\|^3\leq&\sqrt{2}\mathbbm{E}\left[\sup_{t\in[0,1]}\left(\frac{\sqrt{p(1-p)}}{\sqrt{2+8p^2}}t\mathbf{B}_1(t^2)+\frac{p\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}t\mathbf{B}_2(t^2)\right)^3\right]\nonumber\\ &+\sqrt{2}\mathbbm{E}\left[\sup_{t\in[0,1]}\left(\frac{p\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}t\mathbf{B}_1(t^2)+\frac{2p^2\sqrt{2p(1-p)}}{\sqrt{1+4p^2}}t\mathbf{B}_2(t^2)\right)^3\right]\nonumber\\ \leq&\frac{2p^{3/2}(1-p)^{3/2}(1+2^{7/2}p^3+2^{11/2}p^6)}{(1+4p^2)^{3/2}}\mathbbm{E}\left[\sup_{t\in[0,1]}|\mathbf{B}_1(t^2)|^3\right]\nonumber\\ \leq&\frac{27p^{3/2}(1-p)^{3/2}(1+2^{7/2}p^3+2^{11/2}p^6)}{4(1+4p^2)^{3/2}}\leq \frac{4}{3}. \end{align*} This finishes the proof. \end{proof}

https://arxiv.org/abs/2212.12591

Regular Hom-Lie structures on incidence algebras

We fully characterize regular Hom-Lie structures on the incidence algebra $I(X,K)$ of a finite connected poset $X$ over a field $K$. We prove that such a structure is the sum of a central-valued linear map annihilating the Jacobson radical of $I(X,K)$ with the composition of certain inner and multiplicative automorphisms of $I(X,K)$.

\section*{Introduction} A \textit{Hom-Lie algebra}~\cite{Makhlouf-Silvestrov08} over a field $K$ is a triple $(L,[\cdot,\cdot],\varphi)$, where $L$ is a $K$-vector space, $[\cdot,\cdot]$ is an anti-commutative bilinear multiplication on $L$ and $\varphi:L\to L$ is a linear map satisfying the so-called \textit{Hom-Jacobi identity} \begin{align}\label{hom-jacobi} [[a,b],\varphi(c)]+[[b,c],\varphi(a)]+[[c,a],\varphi(b)]=0 \end{align} for all $a,b,c\in L$. The notion of a Hom-Lie algebra first appeared in~\cite{Hartwig-Larsson-Silvestrov06} (under a slightly different definition), where the authors gave a construction of such an algebra based on a commutative associative algebra with a $\sigma$-derivation. A Hom-Lie algebra $(L,[\cdot,\cdot],\varphi)$, in which $\varphi$ is a homomorphism (resp. isomorphism) of $L$, is called \textit{multiplicative} (resp.~\textit{regular})~\cite{Makhlouf-Zusmanovich18}. If $(L,[\cdot,\cdot])$ is itself a (usual) Lie algebra, then by a (multiplicative, regular) \textit{Hom-Lie structure} on $L$ we mean a linear map $\varphi:L\to L$ making $(L,[\cdot,\cdot],\varphi)$ a (multiplicative, regular) Hom-Lie algebra. For example, $\varphi=\lambda\cdot\mathrm{id}$, where $\lambda\in K$, is always a Hom-Lie structure called \textit{trivial}. It is multiplicative only when $\lambda\in\{0,1\}$. Jin and Li~\cite{Jin-Li08} proved that all multiplicative Hom-Lie structures on a finite-dimensional simple Lie algebra are trivial. Xie, Jin and Liu improved the previous result and described arbitrary (not necessarily multiplicative) Hom-Lie structures on finite-dimensional simple Lie algebras in~\cite{Xie-Jin-Liu15}. The superalgebra case was investigated in \cite{Cao-Luo13,Yuan-Liu15}. Xie and Liu characterized Hom-Lie structures on the Virasoro algebra and on the loop and Cartan Lie algebras, completing thus a description of Hom-Lie structures on graded simple Lie algebras of finite growth~\cite{Xie-Liu17}. Makhlouf and Zusmanovich~\cite{Makhlouf-Zusmanovich18} showed that the space of Hom-Lie structures on an affine Kac–Moody algebra is linearly spanned by central Hom-Lie structures and the identity map. Benayadi and Makhlouf~\cite{Benayadi-Makhlouf14} established a correspondence between the class of involutive quadratic Hom–Lie algebras and the class of quadratic simple Lie algebras admitting involutive automorphisms. Chen and Yu~\cite{Chen-Yu20} proved that any regular Hom-Lie structure on a Borel subalgebra of a finite-dimensional simple Lie algebra over an algebraically closed field of characteristic zero is an extremal inner automorphism. Another paper by Chen and Yu~\cite{Chen-Yu22} is of special interest to us. In~\cite{Chen-Yu22} the authors describe regular Hom–Lie structures on the Lie algebra of \textit{strictly} upper triangular matrices over a field. Recall that the algebra $T_n(K)$ of \textit{all} upper triangular matrices of order $n$ over $K$ is a particular case of the incidence algebra $I(X,K)$ of a finite connected poset $X$. Lie automorphisms of $I(X,K)$ were fully described in~\cite{FKS}, so it is natural to use this description to characterize those Lie automorphisms which are (regular) Hom-Lie structures on $I(X,K)$ (endowed with the commutator Lie bracket $[\cdot,\cdot]$). This is done in the present paper. In \cref{sec-prelim} we give the necessary background on Lie automorphisms~\cite{FKS,FKS2} and usual automorphisms~\cite{Baclawski72,St} of $I(X,K)$. In \cref{0=identity} we obtain the conditions that a Lie automorphism $\varphi$ of $I(X,K)$ must satisfy to be a regular Hom-Lie structure. These conditions turn out to be sufficient as well, whenever $\varphi$ is elementary, as shown in \cref{0=id-and=sg=1=>vf-Hom-Lie}. As a consequence, a characterization of the \textit{elementary} Lie automorphisms of $I(X,K)$ that are regular Hom-Lie structures is summarized in \cref{vf-Hom_lie<=>0=id-and-sg=1}. After that, we are left to deal with a class of inner automorphisms of $I(X,K)$, which is done in \cref{xi-Hom-Lie<=>rho-in-Z}. The next \cref{vf-Hom-Lie<=>wtl-vf-and-xi-Hom-Lie} permits us to combine the results of \cref{0=id-and=sg=1=>vf-Hom-Lie,xi-Hom-Lie<=>rho-in-Z} into \cref{main-theorem}, which describes \textit{all} regular Hom-Lie structures on $I(X,K)$ in terms of simpler maps coming from~\cite{FKS}. A more explicit description is given in \cref{vf=xi_bt-circ-M_sg+nu}, specified to the case of posets of length one in \cref{dropping-bt_D=id}. As an application, we characterize regular Hom-Lie structures on the algebra of upper triangular matrices (seen as an incidence algebra) in \cref{Hom-Lie-T_n(K)}. We end the work with \cref{Hom-Lie-normal-subgroup} which is an analog of~\cite[Corollary 3.5]{Chen-Yu22} proving that the set of regular Hom-Lie structures on $I(X,K)$ is a normal subgroup of the group of all Lie automorphisms of $I(X,K)$. \section{Preliminaries}\label{sec-prelim} \subsection{Posets and incidence algebras} Let $(X,\le)$ be a finite poset. A \textit{chain} in $X$ is a linearly ordered subposet of $X$. The \textit{length} of a chain $C\subseteq X$ is defined to be $|C|-1$. The \textit{length} of $X$, denoted by $l(X)$, is the maximum length of all chains $C\subseteq X$. For simplicity, we write $l(x,y)$ for $l(\{z\in X : x\leq z\leq y\})$. The poset $X$ is \emph{connected} if for any pair of $x,y\in X$ there is a sequence $x=x_0,\dots,x_m=y$ in $X$ such that for all $0\le i\le m-1$ either $x_i\le x_{i+1}$ or $x_i\ge x_{i+1}$. We will denote by $\Min(X)$ (resp.~$\Max(X)$) the set of minimal (resp.~maximal) elements of $X$ and by $X^2_<$ (resp.~$X^2_{\leq}$) the set of pairs $(x,y)\in X^2$ such that $x<y$ (resp.~$x\leq y$). Let $X$ be a finite poset and $K$ a field. The \emph{incidence algebra} $I(X,K)$ of $X$ over $K$ (see~\cite{Rota64}) is the $K$-space with basis $\{e_{xy} : x\leq y\}$ and multiplication given by $$ e_{xy}e_{uv}=\delta_{yu}e_{xv}, $$ for any $x\leq y$, $u\leq v$ in $X$ (here $\delta_{yu}$ is the Kronecker delta). Given $f\in I(X,K)$, we write $f=\sum_{x\le y}f(x,y)e_{xy}$, where $f(x,y)\in K$. Let us denote $e_x := e_{xx}$. Then $I(X,K)$ is an associative $K$-algebra with identity element $\delta=\sum_{x\in X}e_x$. Throughout the rest of the paper $X$ will be a connected finite poset with $|X|>1$. Then $I(X,K)$ is a central algebra, by \cite[Corollary~1.3.15]{SpDo}. By \cite[Theorem~1.2.3]{SpDo}, the group of units of $I(X,K)$, denoted by $I(X,K)^{\ast}$, consists of all $f\in I(X,K)$ such that $f(x,x)\neq 0$ for all $x\in X$. It is well-known (see~\cite[Theorem~4.2.5]{SpDo}) that the Jacobson radical $J(I(X,K))$ of $I(X,K)$ is spanned by $\mathcal{B}:=\{e_{xy} : x<y\}$. \emph{Diagonal elements} of $I(X,K)$ are those $f\in I(X,K)$ satisfying $f(x,y)=0$ for $x\neq y$. They form a commutative subalgebra $D(X,K)$ of $I(X,K)$ with basis $\{e_{x} : x \in X\}$. Hence, each $f\in I(X,K)$ can be uniquely written as $f=f_D+f_J$ with $f_D\in D(X,K)$ and $f_J\in J(I(X,K))$. For each positive integer $m$, $J(I(X,K))^m=\Span_K\{e_{xy} : l(x,y)\geq m\}$, by \cite[Proposition~2.4]{FKS}. We will use the notation $J(I(X,K))^{0}=I(X,K)$. By \cite[Proposition 2.5]{FKS}, the center $Z(J(I(X,K)))$ of $J(I(X,K))$ is $\Span_K\{e_{xy}\in \mathcal{B} : x\in \Min(X) \text{ and } y\in\Max(X)\}$. It also coincides with the (bilateral) annihilator of $J(I(X,K))$. \subsection{Automorphisms and Lie automorphisms of $I(X,K)$} We are going to deal with two classes of automorphisms of $I(X,K)$, whose definitions are given below. An element $\sigma\in I(X,K)$ is \emph{multiplicative} if $\sigma(x,y)\neq 0$, for all $x\le y$, and $\sigma(x,z)=\sigma(x,y)\sigma(y,z)$ whenever $x\le y\le z$. Every such $\sigma$ determines the corresponding \emph{multiplicative automorphism} $M_{\sigma}$ of $I(X,K)$ by $M_{\sigma}(f)=\sigma\ast f$, for all $f\in I(X,K)$, where $\sigma\ast f$ is the Hadamard product given by $(\sigma\ast f)(x,y)=\sigma(x,y)f(x,y)$ for all $x,y\in X$. If $\sigma, \tau \in I(X,K)$ are multiplicative, then so is $\sigma\ast\tau$ and $M_\sigma\circ M_\tau=M_{\sigma*\tau}$ (see~\cite[7.3]{SpDo}). For each $\beta\in I(X,K)^{\ast}$ we denote by $\xi_{\beta}$ the inner automorphism of $I(X,K)$ given by $\xi_{\beta}(f)=\beta f \beta{}^{-1}$ for all $f\in I(X,K)$. Regarding the Lie automorphisms of $I(X,K)$, i.e. linear bijections of $I(X,K)$ preserving the commutator product $[f,g]=fg-gf$, we are going to use definitions and results from \cite{FKS}. Let $C: u_1<u_2<\dots<u_m$ be a maximal chain in $X$ and $\theta:\mathcal{B}\to \mathcal{B}$ a bijection. Then $\theta$ is called \textit{increasing (resp.~decreasing) on $C$} if there exists a maximal chain $D: v_1<v_2<\dots<v_m$ in $X$ such that $\theta(e_{u_iu_j})=e_{v_iv_j}$ for all $1\le i<j\le m$ (resp.~$\theta(e_{u_iu_j})=e_{v_{m-j+1}v_{m-i+1}}$ for all $1\le i<j\le m$). Moreover, we say that $\theta$ \textit{is monotone on maximal chains in $X$} if, for any maximal chain $C$ in $X$, $\theta$ is increasing or decreasing on $C$. Let $\theta:\mathcal{B}\to \mathcal{B}$ be a bijection. A map $\sigma:X^2_<\to K^*$ is \textit{compatible} with $\theta$ if, for all $x<y<z$, \begin{align*} \sigma(x,z)= \begin{cases} \sigma(x,y)\sigma(y,z), &\text{if }\theta(e_{xz})=\theta(e_{xy})\theta(e_{yz}),\\ -\sigma(x,y)\sigma(y,z), &\text{if }\theta(e_{xz})=\theta(e_{yz})\theta(e_{xy}). \end{cases} \end{align*} In particular, if $\theta$ is increasing on \textit{all} the maximal chains in $X$, then any such $\sigma$ extends to a multiplicative element of $I(X,K)$ by setting $\sigma(x,x)=1$ for all $x\in X$. There is also the notion of an \textit{admissible} bijection $\mathcal{B}\to \mathcal{B}$ whose definition is quite technical and will not be used explicitly in the present work (just let us mention that $\mathrm{id}_\mathcal{B}$ is always admissible). We refer the interested reader to~\cite[Definition 5.11]{FKS}. Let $X=\{x_1,\dots, x_n\}$. Given an admissible monotone bijection $\theta: \mathcal{B}\to \mathcal{B}$, a map $\sigma:X_<^2\to K^*$ compatible with $\theta$ and a sequence $c=(c_1,\dots,c_n)\in K^n$ such that $\sum_{i=1}^nc_i\in K^*$, we define in \cite[Definition~5.17]{FKS} the following Lie automorphism $\tau=\tau_{\theta,\sigma,c}$ of $I(X,K)$ where, for any $e_{xy}\in \mathcal{B}$, $$ \tau(e_{xy})=\sigma(x,y)\theta(e_{xy}) $$ and $\tau|_{D(X,K)}$ is uniquely determined by $$ \tau(e_{x_i})(x_1,x_1)=c_i, $$ $i=1,\dots,n$, as in Lemmas~5.8 and 5.16 from \cite{FKS} (the admissibility of $\theta$ guarantees that $\tau|_{D(X,K)}$ is well-defined). The Lie automorphisms of the form $\tau_{\theta,\sigma,c}$ are exactly the so-called \textit{elementary Lie automorphisms}. Moreover, the triple $(\theta,\sigma,c)$ is uniquely determined by $\tau$ (see~\cite[Definitions 4.1, 4.7 and Theorem 5.18]{FKS}). As in \cite{FKS}, we denote by $\laut(I(X,K))$ the group of Lie automorphisms of $I(X,K)$ and by $\widetilde\laut(I(X,K))$ its subgroup of elementary Lie automorphisms. We will also use the notation $\inn_1(I(X,K))$ for the subgroup of inner automorphisms $\xi_{\beta}$ with $\beta_D=\delta$. Observe that such $\beta$ is uniquely determined by $\xi_\beta$, so in fact $\inn_1(I(X,K))$ is isomorphic to the subgroup of all $\beta\in I(X,K)^*$ with $\beta_D=\delta$. \begin{theorem}\cite[Theorem~4.15]{FKS}\label{semidireto} The group $\laut(I(X,K))$ is isomorphic to the semidirect product $\inn_1(I(X,K))\rtimes\widetilde\laut(I(X,K))$. \end{theorem} Let $\varphi\in\laut(I(X,K))$. By \cref{semidireto}, $\varphi=\xi_\beta\circ \tau_{\theta,\sigma,c}$ for a unique quadruple $(\beta,\theta,\sigma,c)$ with $\beta_D=\delta$. In this case we write $\widetilde\varphi:=\tau_{\theta,\sigma,c}$, $\theta_{\varphi}:=\theta$ and $\sigma_{\varphi}:=\sigma$. For any $e_{xy}\in J(I(X,K))^i-J(I(X,K))^{i+1}$ we have \begin{align*} \varphi(e_{xy})=\widetilde\varphi(e_{xy})+\rho_{xy}, \end{align*} with $\widetilde\varphi(e_{xy})\in J(I(X,K))^i-J(I(X,K))^{i+1}$ and $\rho_{xy}\in J(I(X,K))^{i+1}$ (see \cite[Definition~4.1]{FKS}). \section{Regular Hom-Lie structures on $I(X,K)$}\label{sec-reg-Hom-Lie} \begin{remark}\label{psi-vf-psi-inv-Hom-Lie} Let $(L,[\cdot,\cdot])$ be a Lie algebra. \begin{enumerate} \item A linear map $\varphi: L\to L$ is a Hom-Lie structure if and only if \cref{hom-jacobi} holds for all $a,b,c$ from a basis of $L$. \item If $\varphi$ is a regular Hom-Lie structure on $L$ and $\psi$ is an arbitrary automorphism of $L$, then $\psi\circ\varphi\circ\psi{}^{-1}$ is also a regular Hom-Lie structure on $L$. Indeed, this follows by taking $a=\psi{}^{-1}(a')$, $b=\psi{}^{-1}(b')$ and $c=\psi{}^{-1}(c')$ in \cref{hom-jacobi} and applying $\psi{}^{-1}$ to both sides of \cref{hom-jacobi}. \end{enumerate} \end{remark} \begin{lemma}\label{basic-formulas} Let $\varphi\in\laut(I(X,K))$ be a regular Hom-Lie structure on $I(X,K)$. If $x<y<z$, then $[\varphi(e_{xy}),e_{yz}]=[e_{xz},\varphi(e_z)]$ and $[e_{xy},\varphi(e_{yz})]=[\varphi(e_x),e_{xz}]$. \end{lemma} \begin{proof} The equalities follow from \cref{hom-jacobi} applied to the triples $e_{xy}$, $e_{yz}$, $e_z$ and $e_x$, $e_{xy}$, $e_{yz}$, respectively. \end{proof} \begin{lemma}\label{not3non0comm} Let $\varphi$ be a regular Hom-Lie structure on $I(X,K)$. If $e_{xy},e_{zw},e_{uv}\in \mathcal{B}$, then at least one of the commutators $[e_{xy},e_{zw}]$, $[e_{zw},e_{uv}]$ and $[e_{uv},e_{xy}]$ is zero. \end{lemma} \begin{proof} Suppose that $[e_{xy},e_{zw}]\neq 0\neq [e_{zw},e_{uv}]$. Then $[e_{xy},e_{zw}]\neq 0$ implies either $x=w$ and $y\neq z$ or $y=z$ and $x\neq w$. Similarly, $[e_{zw},e_{uv}]\neq 0$ implies either $z=v$ and $w\neq u$ or $w=u$ and $z\neq v$. So we have to analyze four cases. If $x=w=u$ and $y\neq z\neq v$ or if $y=z=v$ and $x\neq w\neq u$, we obtain $e_{uv}=e_{xv}$ or $e_{uv}=e_{uy}$ and, therefore, $[e_{xy},e_{uv}]=0$. If $x=w\neq u$ and $y\neq z=v$, then $u<v=z<w=x<y$. Consequently $u\neq y$ and $v\neq x$ which implies $[e_{xy},e_{uv}]=0.$ The last possibility is $x\neq w=u$ and $y=z\neq v$. In this case $x<y=z<w=u<v$ and again $u\neq y$ and $v\neq x$, also implying $[e_{xy},e_{uv}]=0$. \end{proof} \begin{proposition}\label{0=identity} Let $\varphi\in\laut(I(X,K))$ be a regular Hom-Lie structure on $I(X,K)$. Then \begin{enumerate} \item $\theta_\varphi=\mathrm{id}_\mathcal{B}$;\label{0_vf=id} \item $\sigma_\varphi(x,y)=1$ unless $x\in\Min(X)$ and $y\in\Max(X)$.\label{sg(xy)=1} \end{enumerate} \end{proposition} \begin{proof} Write, for simplicity, $\theta=\theta_\varphi$ and $\sigma=\sigma_\varphi$. \cref{0_vf=id}. Let $x<y$ and $u<v$ such that $\theta(e_{xy})=e_{uv}$. Assume that $x\not\in\{u,v\}$. Then applying \cref{hom-jacobi} to the triple $e_u,e_{uv},e_x$ we get $[e_{uv},\varphi(e_x)]=0$. Write $\varphi(e_x)=\widetilde\varphi(e_x)+\rho_x$, where $\rho_x\in J(I(X,K))$. Then \begin{align*} [e_{uv},\widetilde\varphi(e_x)]=-[e_{uv},\rho_x]. \end{align*} However, $[e_{uv},\widetilde\varphi(e_x)]$ is a scalar multiple of $e_{uv}$ because $\widetilde\varphi(e_x)\in D(X,K)$, while $[e_{uv},\rho_x]\in J^{k+1}$, where $k=l(u,v)$. Hence, $[e_{uv},\widetilde\varphi(e_x)]=0$. On the other hand, \begin{align*} [e_{uv},\widetilde\varphi(e_x)]=\sigma(x,y){}^{-1}[\widetilde\varphi(e_{xy}),\widetilde\varphi(e_x)]=-\sigma(x,y){}^{-1}\widetilde\varphi(e_{xy}), \end{align*} a contradiction. Similarly, assuming $y\not\in\{u,v\}$ and applying \cref{hom-jacobi} to the triple $e_u,e_{uv},e_y$ we get $[e_{uv},\widetilde\varphi(e_y)]=0$, which contradicts \begin{align*} [e_{uv},\widetilde\varphi(e_y)]=\sigma(x,y){}^{-1}[\widetilde\varphi(e_{xy}),\widetilde\varphi(e_y)]=\sigma(x,y){}^{-1}\widetilde\varphi(e_{xy}). \end{align*} Thus, $e_{xy}=e_{uv}$. \cref{sg(xy)=1}. Assume that $x\not\in\Min(X)$, so there exists $u<x$. Using \cref{basic-formulas} we have \begin{align}\label{[e_ux_vf(e_xy)]=[vf(e_u)_e_uy]} [e_{ux},\varphi(e_{xy})]=[\varphi(e_u),e_{uy}]. \end{align} Since $\theta=\mathrm{id}_\mathcal{B}$ by \cref{0_vf=id}, then $\varphi(e_{xy})=\sigma(x,y)e_{xy}+\rho_{xy}$, where $\rho_{xy}=\sum\rho_{xy}(a,b)e_{ab}$ with $a\le x<y\le b$ and $l(x,y)<l(a,b)$. Write also $\varphi(e_u)=\widetilde\varphi(e_u)+\rho_u$, where $\rho_u\in J(I(X,K))$. Then \cref{[e_ux_vf(e_xy)]=[vf(e_u)_e_uy]} gives \begin{align}\label{[e_ux_sg(x_y)e_xy]-[wtl-vf(e_u)_e_uy]=-[e_ux_rho_xy]+[rho_u_e_uy]} [e_{ux},\sigma(x,y)e_{xy}]-[\widetilde\varphi(e_u),e_{uy}]=-[e_{ux},\rho_{xy}]+[\rho_u,e_{uy}]. \end{align} Now, \begin{align*} [e_{ux},\sigma(x,y)e_{xy}]&=\sigma(x,y)[e_{ux},e_{xy}]=\sigma(x,y)e_{uy},\\ [\widetilde\varphi(e_u),e_{uy}]&=[\widetilde\varphi(e_u),\sigma(u,y){}^{-1}\widetilde\varphi(e_{uy})]=\sigma(u,y){}^{-1}\widetilde\varphi([e_u,e_{uy}])\\ &=\sigma(u,y){}^{-1}\widetilde\varphi(e_{uy})=\sigma(u,y){}^{-1} \sigma(u,y)e_{uy}=e_{uy}. \end{align*} Hence, the left-hand side of \cref{[e_ux_sg(x_y)e_xy]-[wtl-vf(e_u)_e_uy]=-[e_ux_rho_xy]+[rho_u_e_uy]} equals $(\sigma(x,y)-1)e_{uy}$. We are going to show that the right-hand side of \cref{[e_ux_sg(x_y)e_xy]-[wtl-vf(e_u)_e_uy]=-[e_ux_rho_xy]+[rho_u_e_uy]} is zero at $(u,y)$. Let $l(u,y)=k$. Then $[\rho_u,e_{uy}]\in J^{k+1}$, so it is zero at $(u,y)$. Regarding $[e_{ux},\rho_{xy}]$, we have $e_{ux}\rho_{xy}=\sum_{y<b}\rho_{xy}(x,b)e_{ub}$, which is also zero at $(u,y)$. It follows that $\sigma(x,y)=1$. Similarly, if $y\not\in\Max(X)$, then there exists $v>y$, and $[\varphi(e_{xy}),e_{yv}]=[e_{xv},\varphi(e_v)]$ gives $\sigma(x,y)=1$. \end{proof} \begin{proposition}\label{0=id-and=sg=1=>vf-Hom-Lie} Let $\varphi\in\widetilde\laut(I(X,K))$ satisfying \cref{0_vf=id,sg(xy)=1} of \cref{0=identity}. Then $\varphi$ is a regular Hom-Lie structure on $I(X,K)$. \end{proposition} \begin{proof} Assume \cref{0_vf=id,sg(xy)=1} and let $\theta=\theta_\varphi$ and $\sigma=\sigma_\varphi$. We will first prove \cref{hom-jacobi} for $a=e_{xy}$, $b=e_{zw}$ and $c=e_{uv}$, where $x<y$, $z<w$ and $u<v$. \textit{Case 1.} Exactly one of the products $[e_{xy},e_{zw}]$, $[e_{zw},e_{uv}]$ and $[e_{uv},e_{xy}]$ is non-zero, say $[e_{xy},e_{zw}]\ne 0$ and $[e_{zw},e_{uv}]=[e_{uv},e_{xy}]=0$. We thus need to prove that $[[e_{xy},e_{zw}],\varphi(e_{uv})]=0$. Since $\varphi(e_{uv})=\sigma(u,v)e_{uv}$, the result follows from the usual Jacobi identity for the triple $e_{xy}$, $e_{zw}$, $e_{uv}$. \textit{Case 2.} Exactly two of the products $[e_{xy},e_{zw}]$, $[e_{zw},e_{uv}]$ and $[e_{uv},e_{xy}]$ are non-zero, say $[e_{xy},e_{zw}]\ne 0$, $[e_{zw},e_{uv}]\ne 0$ and $[e_{uv},e_{xy}]=0$. Since $\varphi(e_{xy})=\sigma(x,y)e_{xy}$ and $\varphi(e_{uv})=\sigma(u,v)e_{uv}$, we need to prove that \begin{align*} \sigma(u,v)[[e_{xy},e_{zw}],e_{uv}]+\sigma(x,y)[[e_{zw},e_{uv}],e_{xy}]=0. \end{align*} Observe that $[[e_{xy},e_{zw}],e_{uv}]+[[e_{zw},e_{uv}],e_{xy}]=0$ by the usual Jacobi identity. It is thus enough to prove that $\sigma(u,v)=\sigma(x,y)$. It follows from $[e_{xy},e_{zw}]\ne 0$ that either $y=z<w$, in which case $y\not\in\Max(X)$, or $z<w=x$, in which case $x\not\in\Min(X)$. Hence, $\sigma(x,y)=1$. Similarly, $[e_{zw},e_{uv}]\ne 0$ yields $z<w=u$, in which case $u\not\in\Min(X)$, or $v=z<w$, in which case $v\not\in\Max(X)$. Hence, $\sigma(u,v)=1$ as well. Now let us prove \cref{hom-jacobi} for $a=e_{xy}$, $b=e_{zw}$ and $c=e_{uv}$, where some of the elements $a,b,c$ belong to $D(X,K)$. \textit{Case 1.} Exactly one of the elements $a,b,c$ belongs to $D(X,K)$, say $x=y$, $z<w$ and $u<v$. \textit{Case 1.1.} $x\not\in\{z,w\}\cup\{u,v\}$. Then we need to prove that $[[e_{zw},e_{uv}],\varphi(e_x)]=0$. If $[e_{zw},e_{uv}]=0$, then there is nothing to prove. If $w=u$, then \begin{align*} [[e_{zw},e_{uv}],\varphi(e_x)]=[e_{zv},\varphi(e_x)]=\sigma(z,v){}^{-1}\varphi([e_{zv},e_x]), \end{align*} where $[e_{zv},e_x]=0$ because $x\not\in\{z,v\}$. And if $z=v$, then \begin{align*} [[e_{zw},e_{uv}],\varphi(e_x)]=-[e_{uw},\varphi(e_x)]=-\sigma(u,w){}^{-1}\varphi([e_{uw},e_x]), \end{align*} where $[e_{uw},e_x]=0$ because $x\not\in\{u,w\}$. \textit{Case 1.2.} $x\in\{z,w\}\setminus\{u,v\}$. Then we need to prove that \begin{align}\label{[[e_x_e_zw]_vf(e_uv)]+[[e_zw_e_uv]_vf(e_x)]=0} [[e_x,e_{zw}],\varphi(e_{uv})]+[[e_{zw},e_{uv}],\varphi(e_x)]=0. \end{align} \textit{Case 1.2.1.} $x=z\not\in\{u,v\}$. Then \begin{align*} [[e_x,e_{zw}],\varphi(e_{uv})]=[e_{xw},\varphi(e_{uv})]=\sigma(u,v)[e_{xw},e_{uv}]. \end{align*} If $[e_{xw},e_{uv}]=0$, there is nothing to prove. Otherwise $[e_{xw},e_{uv}]=e_{xv}$ because $x\ne v$. Since $x<w=u$ in this case, we have $u\not\in\Min(X)$ and consequently $\sigma(u,v)=1$. Thus, $[[e_x,e_{zw}],\varphi(e_{uv})]=e_{xv}$. On the other hand, \begin{align*} [[e_{zw},e_{uv}],\varphi(e_x)]&=[e_{xv},\varphi(e_x)]=\sigma(x,v){}^{-1}[\varphi(e_{xv}),\varphi(e_x)]=\sigma(x,v){}^{-1}\varphi([e_{xv},e_x])\\ &=-\sigma(x,v){}^{-1}\varphi(e_{xv})=-\sigma(x,v){}^{-1}\sigma(x,v)e_{xv}=-e_{xv}. \end{align*} Thus, \cref{[[e_x_e_zw]_vf(e_uv)]+[[e_zw_e_uv]_vf(e_x)]=0} holds. \textit{Case 1.2.2.} $x=w\not\in\{u,v\}$. Then \begin{align*} [[e_x,e_{zw}],\varphi(e_{uv})]=-[e_{zx},\varphi(e_{uv})]=-\sigma(u,v)[e_{zx},e_{uv}]. \end{align*} If $[e_{zx},e_{uv}]=0$, there is nothing to prove. Otherwise $[e_{zx},e_{uv}]=-e_{ux}$ because $x\ne u$. Since $v=z<x$ in this case, we have $v\not\in\Max(X)$ and consequently $\sigma(u,v)=1$. Thus, $[[e_x,e_{zw}],\varphi(e_{uv})]=e_{ux}$. On the other hand, \begin{align*} [[e_{zw},e_{uv}],\varphi(e_x)]&=-[e_{ux},\varphi(e_x)]=-\sigma(u,x){}^{-1}[\varphi(e_{ux}),\varphi(e_x)]=-\sigma(u,x){}^{-1}\varphi([e_{ux},e_x])\\ &=-\sigma(u,x){}^{-1}\varphi(e_{ux})=-\sigma(u,x){}^{-1}\sigma(u,x)e_{ux}=-e_{ux}. \end{align*} Thus, \cref{[[e_x_e_zw]_vf(e_uv)]+[[e_zw_e_uv]_vf(e_x)]=0} holds. \textit{Case 1.3.} $x\in\{u,v\}\setminus\{z,w\}$. Then we need to prove that \begin{align*} [[e_{zw},e_{uv}],\varphi(e_x)]+[[e_{uv},e_x],\varphi(e_{zw})]=0. \end{align*} This case is the same as Case 1.2 up to the interchange between $e_{uv}$ and $e_{zw}$. \textit{Case 1.4.} $x\in\{u,v\}\cap\{z,w\}$. \textit{Case 1.4.1.} $x=u=z$. Then we need to prove that \begin{align*} 0&=[[e_x,e_{xw}],\varphi(e_{xv})]+[[e_{xv},e_x],\varphi(e_{xw})]=\sigma(x,v)[e_{xw},e_{xv}]-\sigma(x,w)[e_{xv},e_{xw}], \end{align*} which is clearly satisfied because $[e_{xw},e_{xv}]=[e_{xv},e_{xw}]=0$. \textit{Case 1.4.2.} $x=u=w$. Then we need to prove that \begin{align*} 0&=[[e_x,e_{zx}],\varphi(e_{xv})]+[[e_{zx},e_{xv}],\varphi(e_x)]+[[e_{xv},e_x],\varphi(e_{zx})]\\ &=-[e_{zx},\varphi(e_{xv})]+[e_{zv},\varphi(e_x)]-[e_{xv},\varphi(e_{zx})]\\ &=[e_{zv},\varphi(e_x)]+(\sigma(z,x)-\sigma(x,v))e_{zv}. \end{align*} Since $z<x<v$, we have $x\not\in\Min(X)\cup\Max(X)$, whence $\sigma(z,x)=\sigma(x,v)=1$. Moreover, \begin{align*} [e_{zv},\varphi(e_x)]=\sigma(z,v){}^{-1}[\varphi(e_{zv}),\varphi(e_x)]=\sigma(z,v){}^{-1}\varphi([e_{zv},e_x])=0, \end{align*} because $[e_{zv},e_x]=0$. \textit{Case 1.4.3.} $x=v=z$. Then we need to prove that \begin{align*} 0&=[[e_x,e_{xw}],\varphi(e_{ux})]+[[e_{xw},e_{ux}],\varphi(e_x)]+[[e_{ux},e_x],\varphi(e_{xw})]\\ &=[e_{xw},\varphi(e_{ux})]-[e_{uw},\varphi(e_x)]+[e_{ux},\varphi(e_{xw})]\\ &=-[e_{uw},\varphi(e_x)]+(\sigma(x,w)-\sigma(u,x))e_{uw}. \end{align*} So, this case is the same as Case 1.4.2. \textit{Case 1.4.4.} $x=v=w$. Then we need to prove that \begin{align*} 0&=[[e_x,e_{zx}],\varphi(e_{ux})]+[[e_{ux},e_x],\varphi(e_{zx})]=-\sigma(u,x)[e_{zx},e_{ux}]+\sigma(z,x)[e_{ux},e_{zx}], \end{align*} which is clearly satisfied because $[e_{zx},e_{ux}]=[e_{ux},e_{zx}]=0$. \textit{Case 2.} Exactly two of the elements $a,b,c$ belong to $D(X,K)$, say $x=y$, $z=w$ and $u<v$. We assume that $x\ne z$, since the case $x=z$ is obvious. \textit{Case 2.1.} $\{x,z\}\cap\{u,v\}=\emptyset$. Then \cref{hom-jacobi} is trivial. \textit{Case 2.2.} $x=u$, $z\not\in\{u,v\}$. Then we need to prove that $0=[[e_{xv},e_x],\varphi(e_z)]=-[e_{xv},\varphi(e_z)]$, which is true because $[e_{xv},\varphi(e_z)]=\sigma(x,v){}^{-1}\varphi([e_{xv},e_z])$. \textit{Case 2.3.} $x=v$, $z\not\in\{u,v\}$. Then we need to prove that $0=[[e_{ux},e_x],\varphi(e_z)]=[e_{ux},\varphi(e_z)]$, which is true because $[e_{ux},\varphi(e_z)]=\sigma(u,x){}^{-1}\varphi([e_{ux},e_z])$. \textit{Case 2.4.} $z=u$, $x\not\in\{u,v\}$. Then we need to prove that $0=[[e_z,e_{zv}],\varphi(e_x)]=[e_{zv},\varphi(e_x)]$, which is Case 2.2. \textit{Case 2.5.} $z=v$, $x\not\in\{u,v\}$. Then we need to prove that $0=[[e_z,e_{uz}],\varphi(e_x)]=-[e_{uz},\varphi(e_x)]$, which is Case 2.3. \textit{Case 2.6.} $x=u$ and $z=v$. Then we need to prove that \begin{align*} 0=[[e_z,e_{xz}],\varphi(e_x)]+[[e_{xz},e_x],\varphi(e_z)]=-[e_{xz},\varphi(e_x)]-[e_{xz},\varphi(e_z)]. \end{align*} But \begin{align*} [e_{xz},\varphi(e_x)]&=\sigma(x,z){}^{-1}\varphi([e_{xz},e_x])=-\sigma(x,z){}^{-1}\varphi(e_{xz})=-e_{xz},\\ [e_{xz},\varphi(e_z)]&=\sigma(x,z){}^{-1}\varphi([e_{xz},e_z])=\sigma(x,z){}^{-1}\varphi(e_{xz})=e_{xz}. \end{align*} \textit{Case 2.7.} $x=v$ and $z=u$. Then we need to prove that \begin{align*} 0=[[e_z,e_{zx}],\varphi(e_x)]+[[e_{zx},e_x],\varphi(e_z)]=[e_{zx},\varphi(e_x)]+[e_{zx},\varphi(e_z)], \end{align*} which is Case 2.6. \textit{Case 3.} All the elements $a,b,c$ belong to $D(X,K)$. This case is trivial, since $D(X,K)$ is commutative. \end{proof} The following two corollaries are immediate consequences of \cref{0=identity,0=id-and=sg=1=>vf-Hom-Lie}. \begin{corollary}\label{vf-Hom_lie<=>0=id-and-sg=1} Let $\varphi\in\widetilde\laut(I(X,K))$. Then $\varphi$ is a regular Hom-Lie structure on $I(X,K)$ if and only if $\varphi$ satisfies \cref{0_vf=id,sg(xy)=1} of \cref{0=identity}. \end{corollary} \begin{corollary}\label{vf-Hom-Lie=>wtl-vf-Hom-Lie} If $\varphi\in\laut(I(X,K))$ is a regular Hom-Lie structure on $I(X,K)$, then so is $\widetilde\varphi$. \end{corollary} The converse of \cref{vf-Hom-Lie=>wtl-vf-Hom-Lie} does not always hold, as the next remark shows. \begin{remark} Let $l(X)>1$ and fix $x<y$ with $x\not\in\Min(X)$. Define $\varphi=\xi_{\beta}$, where $\beta=\delta+e_{xy}$. Then $\varphi\in\inn_1(I(X,K))$ and for all $u<x$ we have \begin{align*} &[[e_{y},e_{u}],\varphi(e_{ux})]+[[e_{u},e_{ux}],\varphi(e_{y})]+[[e_{ux},e_{y}],\varphi(e_{u})]\\ &\quad=[e_{ux},(\delta+e_{xy})e_{y}(\delta-e_{xy})]=[e_{ux},(e_{y}+e_{xy})(\delta-e_{xy})]\\ &\quad=[e_{ux},e_{y}+e_{xy}]=e_{uy}, \end{align*} so $\varphi$ is not a Hom-Lie structure, but $\widetilde\varphi=\mathrm{id}$ is. \end{remark} \begin{proposition}\label{xi-Hom-Lie<=>rho-in-Z} Let $\beta=\delta+\rho$, where $\rho\in J(I(X,K))$. Then $\xi_\beta$ is a regular Hom-Lie structure on $I(X,K)$ if and only if $\rho\in Z(J(I(X,K)))$. \end{proposition} \begin{proof} \textit{The ``only if'' part.} Let $\xi_\beta$ be a regular Hom-Lie structure on $I(X,K)$. Take $x<y$ such that $\rho(x,y)\ne 0$ and assume that $x\not\in\Min(X)$. Then there exists $u<x$. Applying \cref{hom-jacobi} to the triple $e_u,e_{ux},e_y$ we get \begin{align*} 0&=[e_{ux},(\delta+\rho)e_y\beta{}^{-1}]=e_{ux}(\delta+\rho)e_y\beta{}^{-1}=\rho(x,y)e_{uy}\beta{}^{-1}, \end{align*} whence $e_{uy}=0$, a contradiction. Consequently, $\rho$ is a linear combination of $e_{xy}$, where $x\in\Min(X)$. It follows that $\rho^2=0$, because $e_{xy}e_{uv}=0$ for all $x<y$ and $u<v$ with $x,u\in\Min(X)$. Therefore, $\beta{}^{-1}=\delta-\rho$. Now if $\rho(x,y)\ne 0$ and $y\not\in\Max(X)$, then taking $v>y$ and applying \cref{hom-jacobi} to the triple $e_{yv},e_v,e_x$ we get \begin{align*} 0&=[e_{yv},\beta e_x(\delta-\rho)]=-\beta e_x(\delta-\rho)e_{yv}=\rho(x,y)\beta e_{xv}, \end{align*} a contradiction. Thus, $\rho$ is a linear combination of $e_{xy}$, where $x\in\Min(X)$ and $y\in\Max(X)$, that is, $\rho\in Z(J(I(X,K)))$. \textit{The ``if'' part.} Assume that $\rho\in Z(J(I(X,K)))$. Then $\rho^2=0$, so $\beta{}^{-1}=\delta-\rho$ and $\xi_\beta(e_{xy})=e_{xy}+\rho'_{xy}$ for all $x\le y$, where $\rho'_{xy}\in J(I(X,K))$. In fact, $\rho'_{xy}\in Z(J(I(X,K)))$, because $\rho\in Z(J(I(X,K)))$, which is an ideal of $I(X,K)$. Therefore, \begin{align*} [[f,g],\xi_\beta(e_{xy})]=[[f,g],e_{xy}+\rho'_{xy}]=[[f,g],e_{xy}] \end{align*} for all $x\le y$ and $f,g\in I(X,K)$, because $[f,g]\in J(I(X,K))$. Hence, the Hom-Jacobi identity for $\xi_\beta$ reduces to the usual Jacobi identity in $I(X,K)$. \end{proof} \begin{proposition}\label{vf-Hom-Lie<=>wtl-vf-and-xi-Hom-Lie} Let $\varphi=\xi_\beta\circ\widetilde\varphi\in\laut(I(X,K))$, where $\xi_\beta\in\inn_1(I(X,K))$. Then $\varphi$ is a regular Hom-Lie structure on $I(X,K)$ if and only if so are $\widetilde\varphi$ and $\xi_\beta$. \end{proposition} \begin{proof} \textit{The ``only if'' part.} Assume that $\varphi$ is a regular Hom-Lie structure on $I(X,K)$. Then so is $\widetilde\varphi$ by \cref{vf-Hom-Lie=>wtl-vf-Hom-Lie}. Hence, $\widetilde\varphi$ satisfies \cref{0_vf=id,sg(xy)=1} of \cref{0=identity}. Since $\theta_\varphi=\mathrm{id}$, then $\widetilde\varphi$ is proper by \cite[Lemma 2.6]{FKS2}. Moreover, \begin{align}\label{wtl-vf(e_xy)-cases} \widetilde\varphi(e_{xy})= \begin{cases} e_{xy}, & x<y,\ x\not\in\Min(X)\text{ or } y\not\in\Max(X),\\ \sigma_\varphi(x,y)e_{xy}, & x<y,\ x\in\Min(X)\text{ and } y\in\Max(X),\\ e_x+\alpha_x\delta, & x=y\text{ (by \cite[the proof of Lemma 2.6]{FKS2})}, \end{cases} \end{align} where $\{\alpha_x\}_{x\in X}\subseteq K$. Let $x<y$. If $x\not\in\Min(X)$ or $y\not\in\Max(X)$, then $\varphi(e_{xy})=\xi_\beta(e_{xy})$ by \cref{wtl-vf(e_xy)-cases}, and consequently \begin{align}\label{[[f_g]_vf(e_xy)]=[[f_g]_xi_bt(e_xy)]} [[f,g],\varphi(e_{xy})]=[[f,g],\xi_\beta(e_{xy})] \end{align} for all $f,g\in I(X,K)$. Otherwise, $\widetilde\varphi(e_{xy})=\sigma_\varphi(x,y)e_{xy}$ by \cref{wtl-vf(e_xy)-cases} and $\xi_\beta(e_{xy})=e_{xy}$, so $\varphi(e_{xy})=\sigma_\varphi(x,y)e_{xy}$. Consequently, \begin{align*} [[f,g],\varphi(e_{xy})]=\sigma_\varphi(x,y)[[f,g],e_{xy}]=0=[[f,g],e_{xy}]=[[f,g],\xi_\beta(e_{xy})] \end{align*} for all $f,g\in I(X,K)$, where the second and the third equalities are due to the fact that $e_{xy}\in Z(J(I(X,K)))$ and $[f,g]\in J(I(X,K))$. Now let $x=y$. Then $\widetilde\varphi(e_{xy})=e_{xy}+\alpha_x\delta$, so $\varphi(e_{xy})=\xi_\beta(e_{xy})+\alpha_x\delta$, and again we have \cref{[[f_g]_vf(e_xy)]=[[f_g]_xi_bt(e_xy)]}. Thus, the left-hand side of \cref{hom-jacobi} for $\varphi$ is the same as for $\xi_\beta$, and $\xi_\beta$ is also a regular Hom-Lie structure on $I(X,K)$. \textit{The ``if'' part.} Assume that $\widetilde\varphi$ and $\xi_\beta$ are regular Hom-Lie structures on $I(X,K)$. Then $\widetilde\varphi$ is of the form \cref{wtl-vf(e_xy)-cases}. The proof of the ``if'' part shows that the left-hand sides of \cref{hom-jacobi} for $\varphi$ and $\xi_\beta$ are the same. Since $\xi_\beta$ is a regular Hom-Lie structure on $I(X,K)$, then so is $\varphi$. \end{proof} We gather \cref{vf-Hom-Lie<=>wtl-vf-and-xi-Hom-Lie,xi-Hom-Lie<=>rho-in-Z,0=identity,vf-Hom-Lie=>wtl-vf-Hom-Lie,0=id-and=sg=1=>vf-Hom-Lie} into the following theorem which is the main result of our work. \begin{theorem}\label{main-theorem} Let $\varphi\in\laut(I(X,K))$. Then $\varphi$ is a regular Hom-Lie structure on $I(X,K)$ if and only if $\varphi=\xi_\beta\circ\tau_{\theta,\sigma,c}$, where \begin{enumerate} \item $\theta=\mathrm{id}_\mathcal{B}$;\label{0-identity} \item $\sigma(x,y)=1$ unless $x\in\Min(X)$ and $y\in\Max(X)$;\label{sg=1-unles-x-min-and-y-max} \item $\beta_D=\delta$ and $\beta_J\in Z(J(I(X,K)))$.\label{bt=dl+rho-rho-in-Z} \end{enumerate} Moreover, such $\beta$, $\theta$, $\sigma$ and $c$ are unique. \end{theorem} \begin{corollary}\label{vf=xi_bt-circ-M_sg+nu} Let $\varphi\in\laut(I(X,K))$. Then $\varphi$ is a regular Hom-Lie structure on $I(X,K)$ if and only if $\varphi=\xi_\beta\circ M_\sigma+\nu$, where $\beta$ and $\sigma$ satisfy \cref{bt=dl+rho-rho-in-Z,sg=1-unles-x-min-and-y-max} of \cref{main-theorem} and $\nu$ is a central-valued linear map annihilating $J(I(X,K))$. Moreover, such $\beta$, $\sigma$ and $\nu$ are unique. \end{corollary} \begin{proof} By \cref{main-theorem}, $\varphi$ is a regular Hom-Lie structure on $I(X,K)$ if and only if $\varphi=\xi_\beta\circ\tau_{\theta,\sigma,c}$ for unique $\theta,\sigma$ and $\beta$ satisfying \cref{0-identity,bt=dl+rho-rho-in-Z,sg=1-unles-x-min-and-y-max}. In this case $\widetilde\varphi=\tau_{\theta,\sigma,c}$, whose action on the standard basis is given by \cref{wtl-vf(e_xy)-cases}. Define $\nu(e_{xy})=0$ for $x<y$ and $\nu(e_x)=\alpha_x\delta$. Then $\nu$ is a unique central-valued linear map annihilating $J(I(X,K))$ and such that $\widetilde\varphi=M_\sigma+\nu$, where $\sigma$ is the extension of $\sigma_{\varphi}$ to $X^2_\le$ by $\sigma(x,x)=1$. It follows that $\varphi=\xi_\beta\circ\widetilde\varphi=\xi_\beta\circ M_\sigma+\nu$, and such a representation is unique. \end{proof} In the case $l(X)=1$ the conditions \cref{sg=1-unles-x-min-and-y-max} on $\sigma$ and $\beta_J\in Z(J(I(X,K)))$ on $\beta$ become redundant, while the condition $\beta_D=\delta$ on $\beta$ can be dropped at the cost of the uniqueness of $\beta$. \begin{corollary}\label{dropping-bt_D=id} Let $l(X)=1$ and $\varphi\in\laut(I(X,K))$. Then $\varphi$ is a regular Hom-Lie structure on $I(X,K)$ if and only if $\varphi=\xi_\beta\circ M_\sigma+\nu$, where $\nu$ is a central-valued linear map annihilating $J(I(X,K))$. Moreover, such $\sigma$ and $\nu$ are unique. \end{corollary} \begin{proof} Item \cref{sg=1-unles-x-min-and-y-max} of \cref{main-theorem} is trivially satisfied, because for all $x<y$ we have $x\in\Min(X)$ and $y\in\Max(X)$. It is also clear that $\beta_J\in Z(J(I(X,K)))$, since $Z(J(I(X,K)))=J(I(X,K))$. So we only need to prove that $\beta_D=\delta$ can be dropped from \cref{main-theorem}\cref{bt=dl+rho-rho-in-Z}. This will be done via a suitable modification of $\sigma$ that does not change $\xi_\beta\circ M_\sigma$. Namely, we are going to show that for any $\beta\in I(X,K)^*$ and $\sigma:X^2_<\to K^*$ there are $\beta'\in I(X,K)^*$ and $\sigma':X^2_<\to K^*$ such that $\beta'_D=\delta$ and $\xi_\beta\circ M_\sigma=\xi_{\beta'}\circ M_{\sigma'}$. Write $\beta_D=\varepsilon$ and $\beta_J=\rho$. Then $\beta{}^{-1}=\varepsilon{}^{-1}-\varepsilon{}^{-1}\rho\varepsilon{}^{-1}$. Observe that \begin{align*} (\xi_\beta\circ M_\sigma)(e_{xy})&=\beta\sigma(x,y)e_{xy}\beta{}^{-1}=\varepsilon(x,x)\varepsilon(y,y){}^{-1}\sigma(x,y)e_{xy},\ x<y,\\ (\xi_\beta\circ M_\sigma)(e_x)&=\beta e_x\beta{}^{-1}=(\varepsilon+\rho)e_x(\varepsilon{}^{-1}-\varepsilon{}^{-1}\rho\varepsilon{}^{-1})=e_x+\rho\varepsilon{}^{-1} e_x-e_x\rho\varepsilon{}^{-1}. \end{align*} Thus, it suffices to take $\sigma'(x,y)=\varepsilon(x,x)\varepsilon(y,y){}^{-1}\sigma(x,y)$ (observe that we may define $\sigma'(x,y)$ as we wish, since any map $X^2_<\to K^*$ is compatible with $\theta$) and $\beta'=\delta+\rho\varepsilon{}^{-1}$. \end{proof} \begin{corollary}\label{Hom-Lie-T_n(K)} Let $X$ be a finite chain and $\varphi\in\laut(I(X,K))$. Then $\varphi$ is a regular Hom-Lie structure on $I(X,K)$ if and only if $\varphi=\xi_\beta+\nu$, where $\nu$ is a unique central-valued linear map annihilating $J(I(X,K))$ and $\beta$ is either a unique element of $I(X,K)^*$ satisfying \cref{bt=dl+rho-rho-in-Z} of \cref{main-theorem} (if $|X|>2$) or an arbitrary element of $I(X,K)^*$ (if $|X|=2$). \end{corollary} \begin{proof} Let $\varphi=\xi_\beta\circ M_\sigma+\nu$ as in \cref{vf=xi_bt-circ-M_sg+nu}. Write $X=\{x_1,\dots,x_n\}$, where $x_1<\dots<x_n$. Observe that $\sigma(x_i,x_j)=1$ unless $i=1$ and $j=n$. If $n>2$, then $\sigma(x_1,x_n)=\sigma(x_1,x_2)\sigma(x_2,x_n)=1$. Hence, $M_\sigma=\mathrm{id}$. Otherwise, $M_\sigma=\xi_\eta$, where $\eta(x_1,x_1)=\sigma(x_1,x_2)$, $\eta(x_2,x_2)=1$ and $\eta(x_1,x_2)=0$. Therefore, $\xi_\beta\circ M_\sigma=\xi_{\beta\eta}$. Since $l(X)=1$ in this case, $\beta$ is an arbitrary element of $I(X,K)^*$ by \cref{dropping-bt_D=id}. Hence, so is $\beta\eta$. \end{proof} \begin{corollary}\label{Hom-Lie-normal-subgroup} The regular Hom-Lie structures on $I(X,K)$ form a normal subgroup of $\laut(I(X,K))$. \end{corollary} \begin{proof} Let $\varphi,\psi\in\laut(I(X,K))$ be regular Hom-Lie structures on $I(X,K)$. Write $\varphi=\xi_\beta\circ M_{\sigma_\varphi}+\nu_\varphi$ and $\psi=\xi_\gamma\circ M_{\sigma_\psi}+\nu_\psi$ as in \cref{vf=xi_bt-circ-M_sg+nu}. Then $\varphi|_{J(I(X,K))}=(M_{\sigma_\varphi})|_{J(I(X,K))}$ and $\psi|_{J(I(X,K))}=(M_{\sigma_\psi})|_{J(I(X,K))}$ because \begin{align*} (\nu_\varphi)|_{J(I(X,K))}&=(\nu_\psi)|_{J(I(X,K))}=0,\\ (\xi_\beta)|_{J(I(X,K))}&=(\xi_\gamma)|_{J(I(X,K))}=\mathrm{id}_{J(I(X,K))}, \end{align*} so \begin{align}\label{(vf-circ-psi)|_J=M_sg_vf_sg_psi} (\varphi\circ\psi)|_{J(I(X,K))}=(M_{\sigma_\varphi*\sigma_\psi})|_{J(I(X,K))}. \end{align} Take $x\in X$ and observe that \begin{align}\label{vf-circ-psi(e_x)=(xi_bt-M_sg_vf-xi_gm)(e_x)+mu(e_x)} (\varphi\circ\psi)(e_x)=\varphi(\xi_\gamma(e_x)+\nu_\psi(e_x))&=(\xi_\beta\circ M_{\sigma_\varphi}\circ\xi_\gamma)(e_x)+\mu(e_x), \end{align} where $\mu:=\nu_\varphi+\nu_\psi+\nu_\varphi\circ\nu_\psi$ is a central-valued linear map that annihilates $J(I(X,K))$. Now, \begin{align*} \xi_\gamma(e_x)=(\delta+\gamma_J)e_x(\delta-\gamma_J)=e_x+\gamma_Je_x-e_x\gamma_J. \end{align*} Hence, \begin{align*} (\xi_\beta\circ M_{\sigma_\varphi}\circ\xi_\gamma)(e_x)&=(\xi_\beta\circ M_{\sigma_\varphi})(e_x+\gamma_Je_x-e_x\gamma_J)\\ &=\xi_\beta(e_x+M_{\sigma_\varphi}(\gamma_J)e_x-e_xM_{\sigma_\varphi}(\gamma_J))\\ &=e_x+\beta_Je_x-e_x\beta_J+M_{\sigma_\varphi}(\gamma_J)e_x-e_xM_{\sigma_\varphi}(\gamma_J). \end{align*} Thus, defining $\eta:=\delta+\beta_J+M_{\sigma_\varphi}(\gamma_J)$, we have $\eta_D=\delta$, $\eta_J\in Z(J(I(X,K)))$ and $(\xi_\beta\circ M_{\sigma_\varphi}\circ\xi_\gamma)(e_x)=\xi_\eta(e_x)$. Together with \cref{vf-circ-psi(e_x)=(xi_bt-M_sg_vf-xi_gm)(e_x)+mu(e_x)} this gives $(\varphi\circ\psi)|_{D(X,K)}=(\xi_\eta\circ M_{\sigma_\varphi*\sigma_\psi}+\mu)|_{D(X,K)}$. Since, moreover, $(\varphi\circ\psi)|_{J(I(X,K))}=(\xi_\eta\circ M_{\sigma_\varphi*\sigma_\psi}+\mu)|_{J(I(X,K))}$ by \cref{(vf-circ-psi)|_J=M_sg_vf_sg_psi}, we conclude that $\varphi\circ\psi=\xi_\eta\circ M_{\sigma_\varphi*\sigma_\psi}+\mu$, so $\varphi\circ\psi$ is a regular Hom-Lie structure on $I(X,K)$ by \cref{vf=xi_bt-circ-M_sg+nu}. Let $\varphi$ be a regular Hom-Lie structure on $I(X,K)$. Write $\varphi=\xi_\beta\circ\widetilde\varphi$ and $\varphi{}^{-1}=\xi_\gamma\circ\widetilde{\varphi{}^{-1}}$, where $\beta_D=\gamma_D=\delta$ and $\beta_J\in Z(J(I(X,K)))$. Since $\widetilde{\varphi{}^{-1}}=(\widetilde\varphi){}^{-1}$ by \cite[Proposition 4.2]{FKS}, then $\theta_{\varphi{}^{-1}}=(\theta_\varphi){}^{-1}$ and $\sigma_{\varphi{}^{-1}}(x,y)=\sigma_\varphi(x,y){}^{-1}$ for all $x<y$. It follows from \cref{vf-Hom_lie<=>0=id-and-sg=1} that $\widetilde{\varphi{}^{-1}}$ is also a regular Hom-Lie structure on $I(X,K)$. Now observe that $\varphi(e_{xy})=\widetilde\varphi(e_{xy})$ for all $x<y$. Then for all $x<y$ we have \begin{align*} e_{xy}=(\varphi{}^{-1}\circ\varphi)(e_{xy})=(\xi_\gamma\circ\widetilde{\varphi{}^{-1}}\circ\widetilde\varphi)(e_{xy})=\xi_\gamma(e_{xy}). \end{align*} Hence, $\gamma e_{xy}=e_{xy}\gamma$ for all $x<y$. But $\gamma_D=\delta$, so $\gamma_J e_{xy}=e_{xy}\gamma_J$ for all $x<y$. Thus, $\gamma_J\in Z(J(I(X,K)))$, and $\xi_\gamma$ is a regular Hom-Lie structure on $I(X,K)$ by \cref{xi-Hom-Lie<=>rho-in-Z}. Then so is $\varphi{}^{-1}$ by \cref{vf-Hom-Lie<=>wtl-vf-and-xi-Hom-Lie}. We have proved that regular Hom-Lie structures on $I(X,K)$ form a subgroup of $\laut(I(X,K))$. It is normal by \cref{psi-vf-psi-inv-Hom-Lie}. \end{proof} \section*{Acknowledgements} The second author was partially supported by CMUP, member of LASI, which is financed by national funds through FCT --- Fundação para a Ciência e a Tecnologia, I.P., under the project with reference UIDB/00144/2020.

https://arxiv.org/abs/1910.01927

An energy gap phenomenon for the Whitney sphere

For an immersed Lagrangian submanifold, let $\check{A}$ be the Lagrangian trace-free second fundamental form. In this note we consider the equation $\nabla^*T=0$ on Lagrangian surfaces immersed in $\mathbb{C}^2$, where $T=-2\nabla^*(\check{A}\lrcorner\omega)$, and we prove a gap theorem for the Whitney sphere as a solution to this equation.

\section{Introduction} Gap phenomena form an interesting topic in differential geometry, with many related results to be found. Sacks-Uhlenbeck's well-known energy gap lemma for harmonic maps (see \cite{S81}) is such an example. The following result by Kuwert-Sch\"{a}tzle for the Willmore surfaces immersed in Euclidean space is one of our motivations: \begin{theo*}[Gap lemma for Willmore surfaces, \text{\cite[Th.~1.1]{BR} or \cite[Th.~2.7]{KS01}}] Let $f:\Sigma\to\mathbb{R}^n$ be a properly immersed (compact or non-compact) Willmore surface, and let $\Sigma_\varrho (0):=f^{-1}(B_\varrho (0))$. Then there exists $\epsilon_0(n)>0$ such that if \begin{align*} &\liminf_{\rho\to\infty}\frac{1}{\varrho^4}\int_{\Sigma_\varrho (0)}|A|^2d\mu = 0\\ &\quad and \quad\int_{\Sigma}|\AA{}|^2 d\mu<\epsilon_0(n), \end{align*} where \AA{} is the trace-free second fundamental form of $f(\Sigma)$, $f$ is an embedded plane or sphere. \end{theo*} The small energy condition above is natural in the variational sense since it can be geometrically interpreted as how different of an immersion from being the simplest geometric models such as planes and standard spheres. Our personal interest is on Lagrangian submanifolds which often play an important role in symplectic geometry where objects often have a natural presentation as Lagrangian manifolds. And they also arise in many problems of mechanics and physics. Models such as the Lagrangian planes, the Clifford torus and Whitney spheres are the simplest objects to study in Lagrangian geometry. Followed by Kuwert-Sch\"{a}tzle's idea, Luo-Wang proved a similar result under Lagrangian settings: \begin{theo*}[Gap lemma for HW surfaces, \text{\cite[Th.~4.3]{LW15}}] Let $f:\Sigma\to\mathbb{C}^2$ be a properly immersed HW surface, then there exists $\epsilon_0(n)>0$ such that if the norm of the second fundamental form $\|A\|_{L^2}<\epsilon_0(n)$, it must be a Lagrangian plane. \end{theo*} According to their paper, it still remains open if there is a similar gap phenomenon for the Whitney sphere in the class of HW surfaces (see \cite{LW15} for the definition). To reformulate, we introduce a (0,2)-tensor $T:=\nabla(H\lrcorner\omega)-\frac{1}{2}\operatorname{div}JH\cdot g$ and consider the equation $\nabla^*T=0$ in this paper. From a geometric point of view, the tensor $T$ measures the deviation of the mean curvature vector field from a conformal field and one can easily check that Whitney spheres satisfy the equation. Therefore it is natural to ask if the Whitney sphere is unique under some small energy conditions. Instead of considering an energy condition on \AA{}, we introduce a Lagrangian trace-free second fundamental form $\check{A}$ (see definition in section 2) for its close relationship with Whitney spheres. The following is our main result: \begin{theo}\label{Theorem 1.1} Assume $f:\Sigma\to\mathbb{C}^2$ is a properly immersed Lagrangian surface (compact or non-compact) such that $\nabla^*T=0$, given $\gamma\in C^1_c(\Sigma)$ a positive function that satisfies $|\nabla\gamma|\leq\frac{C_0}{R}$ for any $R>0$, there exists a constant $\epsilon_0>0$ such that if \begin{equation*} \int_{\{\gamma>0\}}|\check{A}|^2 d\mu\leq\epsilon_0, \end{equation*} we have \begin{equation*} \int _{\Sigma}(|\nabla\check{A}|^2+|H|^2|\check{A}|^2)\gamma^2d\mu\leq \frac{C}{R^2}\int_{\{\gamma>0\}}|A|^2 d\mu. \end{equation*} \end{theo} Combining the previous result with a classification theorem from Lagrangian geometry (see \cite{CU93} or \cite{CU01}), we obtain a gap theorem which answers the above question: \begin{theo}\label{Theorem 1.2} Assume $f:\Sigma\to\mathbb{C}^2$ is a properly immersed Lagrangian surface (compact or non-compact) that satisfies $\nabla^*T=0$, and let ${\Sigma_{\varrho}(0)}=f^{-1}(B_{\varrho}(0))$, then there exists $\epsilon_0>0$ such that if \begin{equation*} \int_{\Sigma}|\check{A}|^2 d\mu\leq\epsilon_0\quad and\quad \liminf_{\varrho\to\infty}\frac{1}{\varrho^2}\int_{\Sigma_{\varrho}(0)}|A|^2 d\mu=0, \end{equation*} then $f$ is either a Lagrangian plane in $\mathbb{C}^2$ or a Whitney immersion. \end{theo} Our method to prove these therorems is establishing a Bochner type identity for the Lagrangian trace free curvature as Kuwert-Sch\"{a}tzle. But the difficulty here is that our condition on $\check{A}$ does not imply a control on $H$ as in their case. Therefore we write the Bochner identity in terms of the tensor $T$ so that we can make good use of its relationship with $\check{A}$. We organize this paper as follows: in Section 2 we introduce some elementary notions on Lagrangian submanifolds as well as the Willmore functional. Section 3 is devoted to a curvature estimate for Lagrangian surfaces which is essential for us to get the main gap theorem. In section 4, we will give a connection between our problem and the case of studying gap phenomena for the HW surfaces. \section{Preliminary on the Lagrangian geometry} In this section, let's recall some elementary notions in the Lagrangian geometry. Let $\mathbb{C}^2=\mathbb{R}^4$ be the 2-dimensional complex plane with the standard metric $ds^2=dx_i^2+dy_i^2$ (also denoted as $\langle\quad,\quad\rangle$) and $\omega=dx_i\wedge dy_i$ be the standard symplectic structure associated with it. Let $J$ be the standard complex structure of $\mathbb{C}^2$ such that $J^2=-id_{\mathbb{C}^2}$. These structures above satisfy the relationship: $\langle V,W\rangle=\omega(V,JW)$ for any vectors $V$ and $W$. Here we order the coordinates as $(x_1,y_1,x_2,y_2)$. We denote the connection on $\mathbb{C}^2$ induced by the Euclidean metric as $D$. Now for an immersion $f:\Sigma\to\mathbb{C}^2$, we define the second fundamental form $A:=(D^2 f)^\bot$, i.e. the normal part of the second order covariant derivative of $f$, the mean curvature $H=trA$, and the trace-free second fundamental form $\text{\AA{}}:=A-\frac{1}{2}g\otimes H$ as usual. \begin{definition} Let $\Sigma$ be a surface in $\mathbb{C}^2$, with tangent and normal bundles, $T\Sigma$ and $N\Sigma$, respectively. Then $\Sigma$ is Lagrangian if and only if one of the following equivalent conditions holds: \begin{itemize} \item[(1)] $\omega$ restricted to $\Sigma$ is zero, \item[(2)] $JT\Sigma=N\Sigma$, where $J$ is the standard complex structure on $\mathbb{C}^2$, \end{itemize} \end{definition} We usually treat $f(\Sigma)$ and $\Sigma$ as the same if there is no confusion. Hence the first condition has an alternative version: \begin{definition} An immersion $f$ from a surface $\Sigma$ into $\mathbb{C}^2$ is called a Lagrangian immersion if $f^*\omega =0$. \end{definition} The second condition allows us to choose frames properly which will be helpful when we are doing geometric calculations. The Lagrangian subspaces or planes are the simplest Lagrangian surfaces in $\mathbb{C}^2$, and there are also some well-known but non-trivial examples such as the Clifford torus and Whitney spheres. \begin{example}[Lagrangian planes in $\mathbb{C}^2$] All 2 dimensional subspaces of $\mathbb{R}^4$ whose restriction of symplectic form $\omega$ to this subspace is identically equal to zero is called Lagrangian planes. \end{example} \begin{example}[Whitney immersions in $\mathbb{C}^2$] \begin{equation} \begin{split} \Phi: \qquad\qquad\mathbb{S}^2&\longrightarrow \mathbb{C}^2 \\ (x_1,x_2,x_3) &\longmapsto\frac{r}{1+x^2_3}(x_1,x_1 x_3,x_2,x_2 x_3)+\overrightarrow{C} \end{split} \end{equation} is a family of Lagrangian immersion. Here we embed $\mathbb{S}^2$ into $\mathbb{R}^3$ with center at the origin to get its local coordinates at first. The image of $\Phi$ in $\mathbb{C}^2$ is called a Whitney sphere and denoted as $\mathbb{S}_W$, and the constants $r$ and $\overrightarrow{C}$ will be referred as the radius and the center respectively. \end{example} Topologically, it is well-known that there is no embedded sphere in $\mathbb{C}^2$ as a Lagrangian submanifold. Whitney spheres have possibly the simplest behaviour in this case because they only have one double point. Castro-Urbano \cite{CU93} (or Ros-Urbano \cite{RU98} for higher dimensional case) proved the following famous classification theorem: \begin{theo*}[\text{\cite[Th.~2]{RU98}}~] Let $\Psi: M\to\mathbb{C}^n$ be a Lagrangian immersion of an n-dimensional submanifold M, then \begin{equation*} A(v,w)=\frac{1}{4}\{\langle v,w\rangle H+\langle Jv, H\rangle Jw+\langle Jw, H\rangle Jv\} \end{equation*} holds for any vectors v and w tangent to M if and only if $\Psi(M)$ is either an open set of the Whitney sphere or is totally geodesic. \end{theo*} This reminds us of a classical theorem in $\mathbb{R}^n$ which states that if $\text{\AA{}}=0$, the immersion is either a plane or a standard sphere. Catro-Urbano's result is much more complicated to prove than this classical theorem. Motivated by this, we can define a similar quantity for Lagrangian surfaces: \begin{equation} \check{A}(v,w):=A(v,w)-\frac{1}{4}\{\langle v,w\rangle H+\langle Jv, H\rangle Jw+\langle Jw, H\rangle Jv\} \label{check A}. \end{equation} We call $\check{A}$ the Lagrangian trace-free second fundamental form and an immersion is Lagrangian umbilical if it satisfies $\check{A}=0$. \section{Estimates for Lagrangian surfaces with locally small $L^2$-norm of $\check{A}$} \subsection{Preparations} Let $\{e_i\}$ be a local orthonormal frame for $\Sigma$ and we denote $h_i=\langle H,Je_i\rangle=H\lrcorner\omega(e_i)$ and $A_{ijk}=\langle A(e_i.e_j), Je_k \rangle=A\lrcorner\omega(e_i,e_j,e_k)$ in local orthonormal coordinates. We would like to point out that $A$ is fully symmetric: \begin{equation} A_{ijk}=\langle D_{e_i}e_j,Je_k\rangle=-\langle e_j,JD_{e_i}e_k\rangle=\langle Je_j,D_{e_i}e_k\rangle=A_{ikj},\label{fully symmetric of A} \end{equation} where $D$ is the connection of the ambient space $\mathbb{C}^2$. Hence it doesn't matter which two of its three indices are contracted. We denote $g$ as the induced metric on $\Sigma$ by $f$, $\nabla$ as its connection and $d\mu$ as the induced area form. For any tensor fields $S\in\Gamma(\underbrace{T\Sigma\otimes\cdots\otimes T\Sigma}_{r}\otimes N\Sigma)$ on $\Sigma$, we define its covirant derivative as $\nabla ^{\bot} S=(D_X S)^\bot$ and its adjoint covariant derivative as $\nabla ^{\bot *}S=- e_i\lrcorner\nabla^{\bot}_{e_i}S$. We can verify that they have the relationship \begin{equation*} \int_{\Sigma}\langle\nabla^\bot S,T\rangle d\mu=\int_{\Sigma}\langle S,\nabla^{^\bot*} T\rangle d\mu \end{equation*} for any tensor fields $S\in \Gamma(\underbrace{T\Sigma\otimes\cdots\otimes T\Sigma}_{r}\otimes N\Sigma)$ and $T\in\Gamma(\underbrace{T\Sigma\otimes\cdots\otimes T\Sigma}_{r+1}\otimes N\Sigma)$. In the following, we will omit the superscript of $\nabla^\bot$ if there is no confusion and hence the rough Laplace operator on the normal bundle can be written as $\varDelta S=-\nabla^* \nabla S$. The fundamental curvature functions of submanifold geometry can be expressed as \begin{align} &K=\frac{1}{2}(|H|^2-|A|^2)=\frac{1}{4}|H|^2-\frac{1}{2}|\text{\AA{}}|^2,\label{Gauss equ.}\\ &(\nabla_X A)(Y,Z)=(\nabla_Y A)(X,Z), \quad(\nabla_X H)(Y,Z)=(\nabla_Y H)(X,Z),\label{Codazzi equ.}\\ &R^{\bot}(X,Y)\phi=A(e_i,X)\langle A(e_i,Y),\phi\rangle-A(e_i,Y)\langle A(e_i,X),\phi\rangle.\label{Ricci equ.} \end{align} for any tangential vector fields $X$, $Y$ and normal vector fields $\phi$. The following proposition links those geometric quantities together: \begin{prop} Assume $f:\Sigma\to\mathbb{C}^2$ is a immersed Lagrangian surface (compact or non-compact), then its Lagrangian second fundamental form $\check{A}$ satisfies \begin{align} |\check{A}|^2&=|A|^2-\frac{3}{4}|H|^2 \label{(2.1)},\\ -2\nabla^*(\check{A}\lrcorner\omega)&=\nabla(H\lrcorner\omega)-\frac{1}{2}\operatorname{div}JH\cdot g. \end{align} \end{prop} \begin{proof} In local orthonormal coordinates, we denote $\check{A}_{ijk}=\langle \check{A}(e_i.e_j), Je_k \rangle=\check{A}\lrcorner\omega(e_i,e_j,e_k)$. We use either property (\ref{fully symmetric of A}) or definition (\ref{check A}) to see that $\check{A}_{ijk}$ is fully symmetric. Hence we have \begin{align*} |\check{A}|^2&=\check{A}_{ijk}\check{A}^{ijk}\\ &=[A_{ijk}-\frac{1}{4}(\delta_{ij} h_k+\delta_{ik} h_j+\delta_{jk} h_i)]^2\\ &=A_{ijk} A^{ijk}-\frac{3}{4}h_i h^i. \end{align*} For the second identity, we use Codazzi equation (\ref{Codazzi equ.}) to get \begin{align*} -\nabla^*(\check{A}\lrcorner\omega)(e_i,e_j)&=\check{A}_{ijl,l}\\ &=A_{ijl,l}-\frac{1}{4}(\delta_{ij}\operatorname{div}JH+h_{i,j}+h_{j,i})\\ &=A_{ill,j}-\frac{1}{4}(\delta_{ij}\operatorname{div}JH+2h_{i,j})\\ &=h_{i,j}-\frac{1}{4}(\delta_{ij}\operatorname{div}JH+2h_{i,j}). \end{align*} \end{proof} \begin{definition} We define a (0,2)-tensor \begin{equation*} T:=-2\nabla^*(\check{A}\lrcorner\omega)=\nabla(H\lrcorner\omega)-\frac{1}{2}\operatorname{div}JH\cdot g, \end{equation*} or in local orthonormal basis: \begin{equation*} T_{ij}:=-2\check{A}_{ijl,l}=h_{i,j}-\frac{1}{2}h_{l,l}g_{ij}, \end{equation*} where we have used Einstein's summation convention. \end{definition} One can see that $T$ is symmetric and actually the trace-free part of $\nabla(H\lrcorner\omega)$. To proceed we need a Bochner type identity which allows us to work globally. To make it clearer, instead of presenting it with coordinate free form straightly, we will introduce its local form at first and switch to its global form later in subsection 3.2. \begin{prop}[Bochner identity] If $f:\Sigma\to\mathbb{C}^2$ is a properly immersed Lagrangian surface, in local coordinates we have \begin{multline}\label{Laplacian of check A} \check{A}_{ijk,mm}=3K\check{A}_{ijk}\\ +\frac{T_{ij,k}-\frac{1}{2}\delta_{ij}T_{km,m}}{3}+\frac{T_{jk,i}-\frac{1}{2}\delta_{jk}T_{im,m}}{3}+\frac{T_{ik,j}-\frac{1}{2}\delta_{ik}T_{jm,m}}{3}. \end{multline} \end{prop} \begin{proof} Writing $\check{A}$ under local coordinates: \begin{align*} \check{A}_{ijk,mm}&=\check{A}_{ijm,km}+\frac{1}{4}(\delta_{ij}h_{m,km}-\delta_{ij}h_{k,mm}+\delta_{im}h_{j,km}-\delta_{ik}h_{j,mm}\\&\quad+\delta_{jm}h_{i,km}-\delta_{jk}h_{i,mm}), \end{align*} we commute the second order derivative of $\check{A}$ by Ricci's identity: \begin{equation*} \check{A}_{ijm,km}=\check{A}_{ijm,mk}+\check{A}_{ljm}R_{ikm}^l+\check{A}_{lim}R_{jkm}^l+\check{A}_{lij}R_{mkm}^l. \end{equation*} Hence \begin{align*} & R.H.S.=\check{A}_{ijm,mk}+\check{A}_{ljm}R_{ikm}^l+\check{A}_{lim}R_{jkm}^l+\check{A}_{lij}R_{mkm}^l+\frac{1}{4}(h_{j,ki}+h_{i,kj}\\&\quad-\delta_{ik}h_{j,mm}-\delta_{jk}h_{i,mm}). \end{align*} Now by the definition of $\check{A}$ again, we may commute the order of $j$ and $m$: \begin{equation*} \check{A}_{ijm,mk}=\check{A}_{imm,jk}+\frac{1}{4}(\delta_{im}h_{m,jk}-\delta_{ij}h_{m,mk}+\delta_{im}h_{m,jk}-\delta_{im}h_{j,mk}+\delta_{mm}h_{i,jk}-\delta_{jm}h_{i,mk}). \end{equation*} So \begin{align*} &R.H.S.=\check{A}_{imm,jk}+\check{A}_{ljm}R_{ikm}^l+\check{A}_{lim}R_{jkm}^l+\check{A}_{lij}R_{mkm}^l+\frac{1}{4}(h_{j,ki}+h_{i,kj}\\&\quad-\delta_{ik}h_{j,mm}-\delta_{jk}h_{i,mm})+\frac{1}{4}(\delta_{im}h_{m,jk}-\delta_{ij}h_{m,mk}+\delta_{im}h_{m,jk}\\&\quad-\delta_{im}h_{j,mk}+\delta_{mm}h_{i,jk}-\delta_{jm}h_{i,mk}), \end{align*} where $\check{A}_{imm,jk}=0$ because $\check{A}$ is trace-free. For surfaces, we have $R_{ijkl}=K(g_{ik} g_{jl}-g_{il} g_{jk})$ by definition. It holds \begin{align*} &R.H.S.=\check{A}_{ljm}K(\delta_{lk}\delta_{im}-\delta_{lm}\delta_{ik})+\check{A}_{lim}K(\delta_{lk}\delta_{jm}-\delta_{lm}\delta_{jk})+\check{A}_{lij}K(\delta_{lk}\delta_{mm}\\ &\quad-\delta_{lm}\delta_{km})+\frac{1}{4}(h_{j,ki}+h_{i,kj}-\delta_{ik}h_{j,mm}-\delta_{jk}h_{i,mm})+\frac{1}{4}(\delta_{im}h_{m,jk}\\ &\quad-\delta_{ij}h_{m,mk}+\delta_{im}h_{m,jk}-\delta_{im}h_{j,mk}+\delta_{mm}h_{i,jk}-\delta_{jm}h_{i,mk})\\ &=3K\check{A}_{ijk}+\frac{1}{4}(2h_{i,jk}+h_{j,ki}+h_{i,kj})-\frac{1}{4}(\delta_{ik}h_{j,mm}+\delta_{jk}h_{i,mm}+\delta_{ij}h_{m,mk})\\ &=3K\check{A}_{ijk}+\frac{2h_{i,jk}+h_{j,ki}+h_{i,kj}-\delta_{ik}h_{j,mm}-\delta_{jk}h_{i,mm}-\delta_{ij}h_{m,mk}}{4}. \end{align*} Substituting terms like $h_{i,jk}$ with $(T_{ij,k}+\frac{1}{2}\delta_{ij}h_{l,lk})$ above, we get \begin{align*} &R.H.S.=3K\check{A}_{ijk}+\frac{2T_{ij,k}+T_{jk,i}+T_{ik,j}}{4}-\frac{\delta_{jk}h_i+\delta_{ik}h_j}{8}K-\frac{\delta_{ik}h_{j,mm}+\delta_{jk}h_{i,mm}}{8}\\ &=3K\check{A}_{ijk}+\frac{2T_{ij,k}+T_{jk,i}+T_{ik,j}-\delta_{ik}T_{jm,m}-\delta_{jk}T_{im,m}}{4}, \end{align*} Since $\check{A}$ is fully symmetric, we can apply the method of symmetrization to get (\ref{Laplacian of check A}) as desired. \end{proof} \subsection{Curvature estimates} The methods we use in this part are similar to Kuwert - Sch\"{a}tzle's work in \cite{KS01}. \begin{lem} Assume $f:\Sigma\to\mathbb{C}^2$ is a properly immersed Lagrangian surface (compact or non-compact), and let $\gamma$ be a cut-off function with $\|\nabla\gamma\|_{L^{\infty}}=\Gamma$, then we have: \begin{multline} \int_{\Sigma} (|\nabla\check{A}|^2+|H|^2|\check{A}|^2)\gamma^2d\mu\\ \leq C\int_{\Sigma}\langle\nabla^*T,H\rangle\gamma^2 d\mu+ C\int_{\Sigma}|\check{A}|^4\gamma^2 d\mu+C\Gamma^2\int_{\{\gamma>0\}}|A|^2 d\mu. \end{multline}\label{estimates for nabla check A} where $C$ is a positive constant independent of $f$. \end{lem} \begin{proof} Multiplying (\ref{Laplacian of check A}) by $\check{A}$ we get \begin{multline}\label{Laplacian of check A} \check{A}_{ijk,mm}\check{A}^{ijk}=3K\check{A}_{ijk}\check{A}^{ijk}\\ +(\frac{T_{ij,k}-\frac{1}{2}\delta_{ij}T_{km,m}}{3}+\frac{T_{jk,i}-\frac{1}{2}\delta_{jk}T_{im,m}}{3}+\frac{T_{ik,j}-\frac{1}{2}\delta_{ik}T_{jm,m}}{3})\check{A}^{ijk}. \end{multline} Since $\check{A}$ is trace-free, terms like $\delta_{ij}T_{km,m}\check{A}^{ijk}$ will vanish. Now (\ref{Laplacian of check A}) becomes \begin{equation} \langle \varDelta\check{A},\check{A}\rangle=\langle\nabla T,\check{A} \rangle+3K|\check{A}|^2. \end{equation} Using the Gau\ss{} equation (\ref{Gauss equ.}) and proposition 3.1, we have $K=\frac{|H|^2}{8}-\frac{|\check{A}|^2}{2}$. Multiplying (3.10) by $\gamma^2$ and integrating it on the surface, we have \begin{equation} \int_{\Sigma}\langle \varDelta\check{A},\gamma^2\check{A}\rangle d\mu=\int_{\Sigma}\langle\nabla T,\gamma^2\check{A} \rangle d\mu+\frac{3}{8}\int_{\Sigma}|H|^2|\check{A}|^2\gamma^2 d\mu-\frac{3}{2}\int_{\Sigma} |\check{A}|^4\gamma^2 d\mu. \end{equation} As for the L.H.S., we obtain \begin{align*} L.H.S.&=-\int_{\Sigma}\langle\nabla^*\nabla\check{A},\gamma^2\check{A}\rangle d\mu\\ &=-\int_{\Sigma}|\nabla\check{A}|^2\gamma^2d\mu-2\int_{\Sigma}\langle\nabla\check{A},\gamma\nabla\gamma\otimes\check{A}\rangle d\mu. \end{align*} Now for the first term of R.H.S., we use the definition of $T$ and integrate it by parts \begin{align*} \int_{\Sigma}\langle \nabla T,\gamma^2\check{A}\rangle d\mu&=\int_{\Sigma}\langle T,\gamma^2\nabla^*\check{A}\rangle d\mu-2\int_{\Sigma}\langle T\otimes\nabla\gamma,\gamma\check{A}\rangle d\mu\\ &=-\frac{1}{2}\int_{\Sigma}\langle T,T\rangle\gamma^2 d\mu-2\int_{\Sigma}\langle T\otimes\nabla\gamma,\gamma\check{A}\rangle d\mu\\ &=-\frac{1}{2}\int_{\Sigma}\langle T,\nabla H-\frac{1}{2}\operatorname{div}JH\cdot g\rangle\gamma^2 d\mu-2\int_{\Sigma}\langle T\otimes\nabla\gamma,\gamma\check{A}\rangle d\mu\\ &=-\frac{1}{2}\int_{\Sigma}\langle T,\nabla H\rangle\gamma^2 d\mu-2\int_{\Sigma}\langle T\otimes\nabla\gamma,\gamma\check{A}\rangle d\mu\\ &=-\frac{1}{2}\int_{\Sigma}\langle\nabla^*T,H\rangle\gamma^2 d\mu+\int_{\Sigma}\langle T,H\otimes\nabla\gamma\rangle\gamma d\mu-2\int_{\Sigma}\langle T\otimes\nabla\gamma,\gamma\check{A}\rangle d\mu. \end{align*} Hence with the help of $|T|\leq c|\nabla\check{A}|$ and $|\nabla\gamma|\leq C_0\Gamma^2$ we have \begin{align*} \int_{\Sigma}(&|\nabla\check{A}|^2+\frac{3}{8}|H|^2|\check{A}|^2)\gamma^2 d\mu\\ &=\frac{1}{2}\int_{\Sigma}\langle\nabla^*T,H\rangle\gamma^2 d\mu+\frac{3}{2}\int_{\Sigma}|\check{A}|^4\gamma^2 d\mu-2\int_{\Sigma}\langle\nabla\check{A},\gamma\nabla\gamma\otimes\check{A}\rangle d\mu\\ &\quad-\int_{\Sigma}\langle T,H\otimes\nabla\gamma\rangle\gamma d\mu+2\int_{\Sigma}\langle T\otimes\nabla\gamma,\gamma\check{A}\rangle d\mu\\ &\leq\frac{1}{2}\int_{\Sigma}\langle\nabla^*T,H\rangle\gamma^2 d\mu+\frac{3}{2}\int_{\Sigma}|\check{A}|^4\gamma^2 d\mu+C_0\Gamma^2\int_{\{\gamma>0\}}|\check{A}|^2 d\mu+\frac{1}{2}\int_{\Sigma}|\nabla\check{A}|^2\gamma^2 d\mu\\ &\quad+C_0\Gamma^2\int_{\{\gamma>0\}}|H|^2 d\mu\\ &\leq\frac{1}{2}\int_{\Sigma}\langle\nabla^*T,H\rangle\gamma^2 d\mu+\frac{3}{2}\int_{\Sigma}|\check{A}|^4\gamma^2 d\mu+\frac{1}{2}\int_{\Sigma}|\nabla\check{A}|^2\gamma^2 d+C\Gamma^2\int_{\{\gamma>0\}}|A|^2 d\mu. \end{align*} \end{proof} We now need the general Sobolev inequality of Michael-Simon to absorb $\int_{\Sigma}|\check{A}|^4\gamma^2 d\mu$. \begin{theo*}[Sobolev inequality with $m=2$, \text{\cite[Th.~2.1]{MS73}}] Let $f:\Sigma\to\mathbb{C}^2$ be an immersion and $v$ be a non-negative $C^1_c(U)$ function on $\Sigma$, where $U\subseteq \mathbb{C}^2$ is a domain contains $f(\Sigma)$. Then \begin{equation} \int_\Sigma v^2d\mu\leq c\Bigg(\int_\Sigma|\nabla v|d\mu\Bigg)^2+c\Bigg(\int_\Sigma v|H|d\mu\Bigg)^2\label{M-S Sobolev}, \end{equation} where $H$ is the mean curvature vector and c is a constant independent of $f$ . \end{theo*} \begin{lem} Under the same assumption as in Lemma 3.1, \begin{equation} \int_{\Sigma}|\check{A}|^4\gamma^2 d\mu\leq C\int_{\{\gamma>0\}}|\check{A}|^2 d\mu\int_{\Sigma}(|\nabla\check{A}|^2+|H|^2|\check{A}|^2)\gamma^2 d\mu+C\Gamma^2\Bigg(\int_{\{\gamma>0\}}|\check{A}|^2d\mu\Bigg)^2 \end{equation} holds. \end{lem} \begin{proof} Substituting $v=|\check{A}|^2\gamma$ in (\ref{M-S Sobolev}), we have \begin{align*} \int_{\Sigma}&|\check{A}|^4\gamma^2 d\mu\\ &\leq c\Bigg(\int_{\Sigma}|\check{A}||\nabla\check{A}|\gamma d\mu\Bigg)^2+c\Bigg(\int_{\Sigma}|\check{A}|^2|\nabla\gamma| d\mu\Bigg)^2+c\Bigg(\int_{\Sigma} |\check{A}|^2|H|\gamma d\mu\Bigg)^2\\ &\leq c'\int_{\{\gamma>0\}}|\check{A}|^2 d\mu\int_{\Sigma}(|\nabla\check{A}|^2+|H|^2|\check{A}|^2)\gamma^2 d\mu+c\Gamma^2\Bigg(\int_{\{\gamma>0\}}|\check{A}|^2d\mu\Bigg)^2. \end{align*} \end{proof} \begin{proof}[Proof of the Theorem \ref{Theorem 1.1}] By combining Lemma 3.1 and Lemma 3.2, one can straightforwardly get \begin{align*} \int _{\Sigma}(|\nabla\check{A}|^2+|H|^2|\check{A}|^2)\gamma^2d\mu&\leq C'\int_{\{\gamma>0\}}|\check{A}|^2 d\mu\int_{\Sigma}(|\nabla\check{A}|^2+|H|^2|\check{A}|^2)\gamma^2 d\mu\\ &+C'\Gamma^2\int_{\{\gamma>0\}}|A|^2 d\mu+C'\Gamma^2\Bigg(\int_{\{\gamma>0\}}|\check{A}|^2d\mu\Bigg)^2. \end{align*} If we choose $\epsilon_0<min\{1,\frac{1}{2C'}\}$ and assume $\int_{\{\gamma>0\}}|\check{A}|^2 d\mu\leq\epsilon_0$, it holds \begin{equation*} \int _{\Sigma}(|\nabla\check{A}|^2+|H|^2|\check{A}|^2)\gamma^2d\mu\leq\frac{C''\Gamma^2}{1-C'\int_{\{\gamma>0\}}|\check{A}|^2 d\mu}\int_{\{\gamma>0\}}|A|^2 d\mu\leq \tilde{C}\Gamma^2\int_{\{\gamma>0\}}|A|^2 d\mu, \end{equation*} let $\psi\in C^1(\mathbb{R})$ be a non-negative function which satisfies: \begin{equation*} \psi(t)= \begin{cases} 1,& t\leq \frac{1}{2},\\ 0,& t\geq 1. \end{cases} \end{equation*} and $\gamma(p)=\psi(\frac{1}{R}|f(p)|)\in C^1_c(\Sigma)$ be the cut-off function for any $p\in\Sigma$ and radius $R>0$, then $\Gamma\leq\frac{C_0}{R}$ holds and we obtain \begin{equation*} \int _{\Sigma}(|\nabla\check{A}|^2+|H|^2|\check{A}|^2)\gamma^2d\mu\leq \frac{C}{R^2}\int_{\{\gamma>0\}}|A|^2 d\mu \end{equation*} as dersired. \end{proof} \begin{proof}[Proof of the Theorem \ref{Theorem 1.2}] Letting $R\to\infty$ in (3.14), we have $\check{A}=0$ by the arbitrariness of cut off function $\gamma$. Then using the classification theorem for Lagrangian umblical surfaces (see \cite{CU93,RU98}), we see that $f$ is either a Lagrangian plane or a Whitney sphere. \end{proof} \section{On the relationship with Willmore immersion} For an immersed surface $f:\Sigma \to \mathbb{R}^n$, the Willmore functional is defined by \begin{equation} \mathcal{W}(f)=\int_{\Sigma}|\text{\AA{}}|^2 d\mu \label{Willmore Functional}, \end{equation} where we denote $d\mu$ as the induced area element on $\Sigma$ by $f$. The Euler-Lagrange operator of (\ref{Willmore Functional}) is \begin{equation} \operatorname{W}=\varDelta H+Q(\text{\AA{}})H, \label{E-L equ-1} \end{equation} where \begin{equation} Q(\text{\AA{}})\phi=\Sigma_{i,j=1}^{2}\text{\AA{}}(e_{i},e_{j})\langle\text{\AA{}}(e_i,e_j),\phi\rangle \label{definition of Q(A)H} \end{equation} under an orthonormal basis $\{e_{1},e_2{}\}$. The following lemma reveals the relationship between $T$ and the Euler-Lagrange equation of the Willmore functional: \begin{lem} If $f:\Sigma\to\mathbb{C}^2$ is a properly immersed Lagrangian surface, the Euler-Lagrange equation of the Willmore functional can be reformulated as: \begin{equation*} \operatorname{W}=\varDelta H+\frac{1}{8}|H|^2 H+Q(\check{A})H+\check{A}(JH,JH). \end{equation*} or in the dual form associated with the symplectic form $\omega$: \begin{equation} \operatorname{W}\lrcorner\omega=-2\nabla^* T+\frac{1}{2}|\check{A}|^2 H\lrcorner\omega+Q(\check{A})H+\check{A}\lrcorner\omega(JH,JH). \label{E-L equ.2} \end{equation} \end{lem} \begin{proof} By (\ref{E-L equ-1}) and (\ref{definition of Q(A)H}), we have \begin{align*} &Q(A)\\ &=A(e_i,e_j)\langle A(e_i,e_j),H\rangle \\ &=[\check{A}_{ijh}+\frac{1}{4}(\delta_{ij}h_h+\delta_{ih}h_j+\delta_{jh}h_i)][\check{A}_{ijs}+\frac{1}{4}(\delta_{ij}h_s+\delta_{is}h_j+\delta_{js}h_i)]\langle Je_s,H\rangle Je_h \\ &=\check{A}_{ijs}\langle \check{A}_{ijh},H\rangle\langle Je_s,H\rangle Je_h+\check{A}_{ijh} h_i h_j Je_h+\frac{5}{8}|H|^2H\\ &=Q(\check{A})H+\check{A}(JH,JH)+\frac{5}{8}|H|^2H. \end{align*} Using the Ricci identity (\ref{Ricci equ.}), the Gau\ss{} equation (\ref{Gauss equ.}) and the Codazzi equation (\ref{Codazzi equ.}), we calculate \begin{align*} \nabla^*T(e_i)&=-T_{ki,k}\\ &=-h_{i,kk}+\frac{1}{2}\delta_{ki}h_{l,lk} \\ &=-h_{i,kk}+\frac{1}{2}(h_{i,ll}+h_s R_{lli}^s) \\ &=-\frac{1}{2}\varDelta(H\lrcorner\omega)(e_i)-\frac{K}{2}h_i\\ &=-\frac{1}{2}\varDelta (H\lrcorner\omega)(e_i)-\frac{1}{2}(\frac{|H|^2}{8}-\frac{|\check{A}|^2}{2})H\lrcorner\omega(e_i), \end{align*} where $K$ is the Gaussian curvature of the surface. Substituting the above two formulas into (\ref{E-L equ-1}) we complete the proof. \end{proof} The Willmore Lagrangian surface is a Lagrangian surface satisfying the Euler-Lagrange equation (\ref{E-L equ-1}) Now let's recall Urbano and Castro's classification theorem as follows: \begin{theo*}[Uniqueness of Willmore Lagrangian spheres,\text{\cite[Cor.~1]{CU01}}] The Whitney sphere is the only Willmore Lagrangian surface of genus 0 in $\mathbb{C}^2$. \end{theo*} Then we obtain a gap-lemma for Willmore Lagrangian surfaces. \begin{cor}[Gap theorem for Willmore Lagrangian surfaces] Let $f:\Sigma\to\mathbb{C}^2$ be a closed Lagrangian surface satisfies $\operatorname{W}=0$, then there exists a universal constant $\epsilon_0$ such that if \begin{equation*} \int_{\Sigma}|\check{A}|^2 d\mu\leq\epsilon_0, \end{equation*} $f$ is a Whitney immersion. \end{cor} \begin{proof} Using the Gau\ss{}-Bonnet formula and (\ref{(2.1)}), we see that if $f$ is Lagrangian immersion, we have \begin{equation*} \int_{\Sigma}|\check{A}|^2d\mu=\frac{1}{2}\int_{\Sigma}|\text{\AA{}}|^2 d\mu-2\pi\chi(\Sigma). \end{equation*} If $\int_{\Sigma}|\check{A}|^2d\mu$ is sufficiently small, $\chi(\Sigma)$ must be non-negative, which means that it is either a sphere or a torus. However the result of \cite{KS01} eliminates the torus case given the Willmore energy is sufficiently small. \end{proof} \section*{Acknowledgement} The author would like to thank professor Guofang Wang for his many inspiring discussions. The project is supported by "Willmore Funktional und Langrangsche Fl\"{a}chen" in SPP 2026 of DFG.

https://arxiv.org/abs/1209.1517

Some monotonicity results for minimizers in the calculus of variations

We obtain monotonicity properties for minima and stable solutions of general energy functionals of the type $$ \int F(\nabla u, u, x) dx $$ under the assumption that a certain integral grows at most quadratically at infinity. As a consequence we obtain several rigidity results of global solutions in low dimensions.

\section{Introduction} In this paper we deal with monotonicity properties of minima for quite general energy functionals of the type \begin{equation}\label{functional} \int_\Omega F(\nabla u, u, x) \, dx \end{equation} where $\Omega$ is a domain of ${\mathbb R} ^n$. These monotonicity properties are often used for the classification of global minimizers, and therefore play a key role in the regularity theory in the calculus of variations, see for example the case of minimal surfaces theory \cite{giusti}, free boundary problems \cite{AlC}, \cite{CJK}, \cite{DS} phase transitions \cite{AC} etc. Some applications of our results are given in Section \ref{s2}. We consider the case when the domain $\Omega$ and the functional $F$ are invariant under translations in a number of directions $e_k$,..,$e_n$, and we are interested in monotonicity properties of energy minimizers (or stable solutions) in the class of functions obtained by piecewise Lipschitz domain deformations in these $e_k,..,e_n$ directions. Our main result states that, under rather mild assumptions on $F$, if there exists a constant $C>0$ such that for all large $R$, a minimizer $u$ satisfies $$\int_{\Omega \cap B_R} |D^2_pF \, (\nabla u,u,x)|\, |\nabla u|^2\,dx\le CR^2,$$ then $u$ is one-dimensional in each subspace generated by $e_k,..,e_n$ (see Theorem \ref{t3} for the precise statement). The general approach to obtain such rigidity results (see for instance the case of minimal surfaces) is to apply the stability inequality (see \eqref{stable eq}) to a suitable cutoff function. However, this approach becomes often difficult to implement. An example occurs when the functional $F$ becomes singular near $\partial \Omega$ as in the case of {\it $s$-nonlocal minimal surfaces}, $s \in (0,1)$ (see \cite{CRS}) when the energy functional has the form $$F(\nabla u,u,x)=|\nabla u|^2 x_1^{1-s}, \quad \Omega=\{x_1>0\}.$$ Then the stability inequality does not have a simple form due to the fact that integrations by parts are difficult to handle. We prove our results inspired by the simple method developed in \cite{CS} where we studied global nonlocal minimal surfaces in two dimensions. The main idea is to avoid the precise form of the stability inequality and just compare the energies of $u$ and a translation of itself. In this way we can deal with rather general situations of energy functionals and also consider minimizers directly in the natural class of functions obtained by domain deformations. We describe briefly the strategy below. We compare the energies of $u$ and $$\max\{u(x), u(x+te_n)\},$$ and for this we need to modify this comparison function at infinity so that it becomes a compact perturbation of $u$. The growth condition in the integral above guarantees that the difference between the energy of the perturbed function and the energy of~$u$ can be made arbitrarily small. On the other hand, if~$u$ is not monotone in the $e_n$ direction then we can modify locally the comparison function above and decrease its energy by a small fixed amount, and this contradicts the minimality of~$u$. The paper is organized as follows. In Section \ref{s2} we state our main theorems and in Section \ref{9ss} we provide some concrete applications of our results. The main ingredients of the proofs are given in Section \ref{0055} where we perform the local analysis, and in Section \ref{s3} where we estimate the energy of the perturbations at infinity. The proofs will be completed in Sections \ref{33s} and \ref{44s}. In Section \ref{55s} we discuss an explicit 1D example to illustrate better the notion of minimizer and stability in the class of piecewise Lipschitz deformations. Finally in Section \ref{details} we prove several remarks pointed out throughout the paper. \section{Main results}\label{s2} We consider energy functionals as in \eqref{functional} in the case when the domain $\Omega$ and the functional $F$ are invariant under translations in the $e_n$-direction, that is $$\Omega= \mathcal{U} \times {\mathbb R} , \quad \quad \mathcal {U} \subseteq {\mathbb R} ^{n-1},$$ and $F$ does not depend on the $x_n$-coordinate. Points in $\Omega$ are denoted by $x=(x',x_n)\in \mathcal{U} \times {\mathbb R} $. We assume that the functional $F$ is convex in with respect to the first variable. Precisely, we suppose that \begin{equation}\label{H1} F=F(p,z,x')\in C({\mathbb R} ^n\times {\mathbb R} \times \mathcal{U}), \end{equation} and for any $(z,x')\in{\mathbb R} \times {\mathcal{U}}$, $F$ is $C^2$ and uniformly convex in $p$ at all $p$ with $p_n \ne 0$. Furthermore, we assume that~$F_{pp}=D^2_pF$ satisfies the natural growth condition \begin{equation}\label{H2} |F_{pp}(p+q,z,x')|\le C \, |F_{pp} (p,z,x')| \end{equation} for some $C>0$, and any $p$, $q\in {\mathbb R} ^n$ with $|q|\le |p_n|/2$. For any $R>0$, we introduce the energy functional~${\mathcal E} _R$ defined by $$ {\mathcal E} _R (u):=\int_{\Omega \cap B_R} F(\nabla u(x),u(x),x')\,dx.$$ We study monotonicity properties of suitable minimal or stable solutions for the energy ${\mathcal E} $ among perturbations which are obtained by piecewise domain deformations in the $e_n$-direction. For this we introduce the following notation: \begin{definition}\label{d1} We say that $v$ is an $e_n$-Lipschitz deformation of $u$ in $B_R$ if there exists a Lipschitz function $\psi$ with compact support in $B_R$, and $\|\psi_n\|_{L^\infty({\mathbb R} ^n)} <1$ such that $$v(x)=u(x+\psi(x)e_n)\qquad\forall x\in\Omega.$$ \end{definition} In the notation of Definition~\ref{d1}, we have that if $u$ is (locally) Lipschitz then $v$ is (locally) Lipschitz as well. \begin{definition}\label{d2} Let $u\in C^{0,1}(\Omega)$. We say that $v\in C^{0,1}(\Omega)$ is a piecewise $e_n$-Lipschitz deformation of $u$ in $B_R$ and write $$v \in D_R(u)$$ if there exist a finite number $v^{(1)}$,..., $v^{(m)}$ of $e_n$-Lipschitz deformations of $u$ in $B_R$ such that $$v(x)=v^{(i)}(x) \quad \quad \mbox{for some $i$ (depending on $x\in\Omega$).}$$ Also, if all $v^{(i)}$ satisfy $$v^{(i)}(x)=u(x+\psi^{(i)}(x)e_n) \quad \mbox{with} \quad \|\psi^{(i)}\|_{C^{0,1}(\Omega)} \le \delta$$ for some $\delta>0$, we write $$v \in D^\delta_R(u).$$ \end{definition} We list some elementary properties that follow easily from Definition~\ref{d2}: \begin{equation}\begin{split}\label{765ty} &v, w \in D^\delta_R(u) \quad\Rightarrow \quad\min \{v,w\}, \max \{v,w \} \in D^\delta_R(u);\\ & v\in D^\delta_R(u), \quad w\in D^\delta_R(v)\quad\Rightarrow \quad w\in D^{3\delta}_R(u);\\ & v \in D^\delta_R(u)\quad\Rightarrow\quad \|v-u\|_{L^\infty(\Omega)} \le C \delta \|u\|_{C^{0,1}(\Omega)};\\ & v \in D^\delta_R(u), \quad u \in C^{1,1}(\Omega) \quad\Rightarrow\quad \|v-u\|_{C^{0,1}(\Omega)} \le C \delta \|u\|_{C^{1,1}(\Omega)}. \end{split}\end{equation} \begin{definition}\label{d3} We say that~$u\in C^{0,1}(\Omega)$ is an $e_n$-minimizer for~${\mathcal E} $ if for any $R>0$ we have that~${\mathcal E} _R(u)$ is finite and $${\mathcal E} _R(u) \le {\mathcal E} _R(v), \quad \quad \forall v \in D_{R}(u).$$ \end{definition} \begin{remark}\label{we} {\rm The standard definition in the calculus of variation consists in saying that $u$ is a classical minimizer for ${\mathcal E} $ if it minimizes the energy with respect to compact deformations of the graph of $u$ in the vertical direction ($e_{n+1}$-direction) that is: $${\mathcal E} _R(u) \le {\mathcal E} _R(u+\varphi) $$ for any Lipschitz~$\varphi$ with compact support in $\Omega \cap B_R$. We observe that when $\Omega={\mathbb R} ^n$, $e_n$-minimality is a weaker condition than classical minimality. For example any function which is constant in the $e_n$-direction is always an $e_n$-minimizer, but not necessarily a classical minimizer. }\end{remark} Our first general monotonicity result is the following: \begin{theorem}\label{GENERAL} Let~$u \in C^1(\Omega)$ be an $e_n$-minimizer for the energy ${\mathcal E} $ with $F$ satisfying~\eqref{H1} and~\eqref{H2}. If there exists $C>0$ such that for all large $R$ \begin{equation}\label{EE} \int_{\Omega \cap B_R} |F_{pp} \, (\nabla u,u,x')|\, |\nabla u|^2\,dx\le CR^2, \end{equation} then~$u$ is monotone on each line in the $e_n$-direction, i.e., for any~$\bar x\in\Omega$, either~$u_n(\bar x+te_n)\ge0$ or~$u_n(\bar x+te_n)\le 0$ for any~$t\in{\mathbb R} $. \end{theorem} \begin{remark}\label{R1} {\rm If a continuous function $u$ is monotone on each line in ${\mathbb R} ^n$ then it is one-dimensional, that is $u=f(x \cdot \xi)$ for some function $f:{\mathbb R} \to {\mathbb R} $ and some unit direction $\xi.$ See Section~\ref{details} for a proof.}\end{remark} Our second theorem is a version of Theorem~\ref{GENERAL} for stable critical points of the energy instead of $e_n$-minimizers. The stability condition we use involves the second variation of ${\mathcal E} $ for deformations of $u$ in the $e_n$-direction as well as in the vertical $e_{n+1}$-direction. The precise definition is the following: \begin{definition} \label{d4} We say that $w$ is a piecewise Lipschitz deformation of $u$ in the $\{ e_n, e_{n+1}\}$-directions and write $$w\in \mathcal D_R^\delta(u)$$ if $$w=v+\varphi \quad \mbox {with} \quad v \in D^\delta_R(u) \quad \mbox {and} \quad |\varphi|_{C^{0,1}(\Omega)} \le \delta$$ for some Lipschitz function $\varphi$ with compact support in $\Omega \cap B_R$. \end{definition} We remark that here the vertical perturbations $\varphi$ have compact support in $\Omega \cap B_R$ whereas the $e_n$-deformations $\psi^{(i)}$ in Definitions~\ref{d1} and~\ref{d2} have compact support in $B_R$ (i.e., if~$x\in B_R$ with~$x'\in \partial \mathcal{U}$ then~$\varphi(x)=0$ but~$\psi^{(i)}(x)$ may be different from~$0$). \begin{definition}\label{d5} We say that $u$ is a $\{e_n, e_{n+1}\}$-stable solution for ${\mathcal E} $ if for any $R>0$ and $\epsilon>0$ there exists $\delta>0$ depending on $R$, $\epsilon$ and $u$ such that for all $t\in(0,\delta)$ we have that~${\mathcal E} _R(u)$ is finite and \begin{equation}\label{55} {\mathcal E} _R(w)-{\mathcal E} _R(u) \ge -\epsilon t^2, \qquad \forall w \in \mathcal D^t_R(u).\end{equation} \end{definition} We point out that classical minimality (see Remark~\ref{we}) implies $\{e_n, e_{n+1}\}$-stablity (on the other hand, $e_n$-minimality and $\{e_n, e_{n+1}\}$-stablity do not imply each other in general). Also, since we allow perturbations in the $e_{n+1}$-direction in Definition~\ref{d5}, then any $\{e_n, e_{n+1}\}$-stable solution is a critical point of the energy functional. \begin{remark}\label{stable} {\rm In the calculus of variation, it is customary to consider stable solutions of partial differential equations. Classically, a solution (i.e., a critical point of the energy functional) is said to be stable if \begin{equation}\label{stable eq} \liminf_{t\to0} \frac{{\mathcal E} _R (u+t \phi)-{\mathcal E} _R(u) }{ t^2} \ge 0 \end{equation} for any Lipschitz function~$\phi$ supported in~$B_R$. If~$\Omega={\mathbb R} ^n$, $F\in C^2$ and~$u\in C^2$, this classical notion of stability is equivalent to the notion of~$\{e_n,e_{n+1}\}$-stable solution (for the proof of this, see Section~\ref{details}). }\end{remark} In the framework given by Definition~\ref{d5} we prove the following result. \begin{theorem}\label{GENERAL2} Let~$u\in C^{0,1}(\Omega)$ be a $\{e_n,e_{n+1}\}$-stable solution and assume $F \in C^3({\mathbb R} ^2\times{\mathbb R} \times{\mathcal{U}})$ satisfies \eqref{H2}. If the growth condition \eqref{EE} holds then~$u$ is monotone in the~$e_n$-direction, i.e. either~$u_n\ge0$ or~$u_n\le0$ in $\Omega$. \end{theorem} We observe that the hypotheses in the two theorems above are slightly different and the thesis of Theorem~\ref{GENERAL} is weaker than the one of Theorem~\ref{GENERAL2} since in Theorem~\ref{GENERAL} we do not say that $u_n(x)$ has the same sign for all $x$, but only that, fixed $x$, $u_n(x+te_n)$ has the same sign for any $t$. Our last theorem deals with $\{e_k,..., e_n\}$-stable solutions, that is, in the definition of stability we allow small piecewise Lipschitz deformations in the $e_k, .., e_n$-directions rather than only the $e_n$-direction or $\{e_n,e_{n+1}\}$-direction (see Definition \ref{d5.5} for a precise statement). \begin{theorem}\label{t3} Assume that $$\Omega= \mathcal{U} \times {\mathbb R} ^{n-k+1}, \quad \quad \mathcal {U} \subseteq {\mathbb R} ^{k-1},$$ $F$ does not depend on the $x_k$, ..., $x_n$ coordinates, ~$F$ satisfies \eqref{H2} and that~$F \in C^3$ at all $p$ with $(p_k,...,p_n) \ne(0,..,0)$. If $u \in C^1(\Omega)$ is $\{e_k,..., e_n\}$-stable and the growth condition~\eqref{EE} holds, then $u$ is one-dimensional in any subspace generated by $\{e_k,...,e_n\}$. \end{theorem} The theorem concludes that for each $(x_1,...,x_{k-1})\in\mathcal{U}$, $u(x)$ is one-dimensional (see Remark \ref{R1}) in the remaining variables $(x_k,\ldots,x_n)$. Of course, when $k=n$ the statement becomes trivial. We point out that the hypothesis above on $\{e_k,..., e_n\}$-stability for $u$ is in general easily satisfied by critical points of ${\mathcal E} $ which are monotone in the $e_n$ direction, and in fact such critical points are $\{e_k,..., e_n\}$-minimizers. We conclude this section with several remarks on the theorems above. \begin{remark}{\rm The results provided in this paper are in fact even more general: we did not attempt to give the most general conditions possible but rather to emphasize the method of proof (further generalizations will be outlined in subsequent remarks and some of these generalizations turn out to be important in the concrete applications). For instance, we observe that the functional in~\eqref{functional} may be generalized to \begin{equation}\label{functional2} \int_\Omega F(\nabla u, u, x') \, dx+\int_{\partial \Omega} G(u,x')\,d\mathcal{H}^{n-1}, \end{equation} where~$G$ satisfies the same regularity assumptions as~$F$. The proofs in this case are affected only by minor, obvious modifications. }\end{remark} \begin{remark}\label{log}{\rm Condition \eqref{EE} may be weakened by allowing logaritmic corrections too. For instance, the right hand side of \eqref{EE} may be replaced by $$ CR^2 \log R$$ or by $$ CR^2 (\log R)(\log\log R).$$ More generally, one can define $\ell_0(R):=R$ and recursively $$ \ell_k(R):= \log (\ell_{k-1}(R))= \underbrace{\log\circ\dots\circ\log}_{\text{$k$ times}}R $$ for any $k\in{\mathbb N} $, $k\ge 1$. Let also \begin{equation}\label{PI} \pi_k(R):=\prod_{j=0}^k \ell_j(R).\end{equation} Then, instead of \eqref{EE}, one may take the weaker condition \begin{equation}\label{EE weak} \int_{\Omega\cap B_R}|F_{pp}(\nabla u,u,x')|\,|\nabla u|^2\,dx \le CR \,\pi_k(R), \end{equation} for a given $k\in{\mathbb N} $. For the proof of this fact, see Section \ref{details} (notice that \eqref{EE weak} boils down to \eqref{EE} if $k=0$). An energy growth with a logaritmic correction of the type~$CR^2\log R$ was also considered in \cite{mos} in the case of semilinear equations. }\end{remark} \begin{remark} \label{ABS} {\rm At first glance, Definition \ref{d2} may look unnecessarily complicated, since one may think that Definition \ref{d1} suffices for Theorem \ref{GENERAL}. That is, one may think that if $u$ minimizes the energy with respect to any $e_n$-Lipschitz deformation and \eqref{EE} is satisfied, then $u$ must possess some kind of monotonicity. However this is not the case, as we show by an example in Section \ref{details}. }\end{remark} \section{Applications}\label{9ss} Below we present some direct applications of our results and obtain several rigidity results of global solutions in low dimensions. We remark however that our theorems do not give in general the optimal dimension for these rigidity results. \subsection{De Giorgi's conjecture} As a first application, we obtain a classical one-dimensional symmetry property related to a conjecture of De Giorgi (see~\cite{DG}): \begin{theorem}[\cite{GG, BCN, AC, AAC}] \label{DG} Let~$f\in C({\mathbb R} )$ and~$u\in C^2({\mathbb R} ^n)\cap L^\infty({\mathbb R} ^n)$ be a solution of~$-\Delta u+f(u)=0$ in the whole of~${\mathbb R} ^n$. Suppose that: \begin{itemize} \item either $n=2$ and $u$ is stable (according to the notation recalled in Remark~\ref{stable}), \item or $n=3$ and~$u_3>0$. \end{itemize} Then~$u$ is one-dimensional. \end{theorem} \begin{proof} We let~$\tilde F$ be a primitive of~$f$ and we define $$ F(p,z,x):=\frac12|p|^2+\tilde F(z).$$ Then,~$F$ is clearly convex in~$p$ and it satisfies~\eqref{H2}. It also satisfies~\eqref{EE}: when~$n=2$ this simply follows from the fact that~$|B_R|\le CR^2$, and when~$n=3$ it is a consequence of Theorem~5.2 in~\cite{AAC}. Now we apply Theorem~\ref{GENERAL2} and obtain that~$u$ is one-dimensional. \end{proof} We stress that the proof of Theorem~\ref{DG} that we give here is based on domain perturbations and it does not use some of the basic ingredients exploited in the existing literature: e.g., differently from \cite{BCN, AC, AAC}, it does not use any Liouville type result, differently from \cite{GG} it does not use the Ekeland's variational principle, differently from \cite{Far} it makes no use of any complex structure, differently from \cite{S} no costruction of barriers is needed, and differently from \cite{FarH, FSV} no geometric Poincar\'e inequality is exploited. \begin{remark}\label{6fsv}{\rm The one-dimensional results related to the Conjecture of De Giorgi in dimensions~$2$ and~$3$ may be extended to a very broad class of operators and nonlinearities: see Theorems~1.1 and~1.2 in~\cite{FSV}. We remark that our Theorem~\ref{GENERAL2} also implies Theorems~1.1 and~1.2 in~\cite{FSV} (at least in case of smooth nonlinearities; for a proof of this fact see Section~\ref{details}). }\end{remark} \subsection{Fractional De Giorgi conjecture} The one-dimensional symmetry of Theorem~\ref{DG} has a counterpart in the fractional Laplace framework, that may be also obtained as a consequence of the results of this paper: \begin{theorem}[\cite{Sola, CSire, CCinti, CCinti2}] \label{DGs} Let~$s\in(0,1)$, $f\in C({\mathbb R} )$ and~$u\in C^2({\mathbb R} ^n)\cap L^\infty({\mathbb R} ^n)$ be a solution of~$(-\Delta)^s u+f(u)=0$ in the whole of~${\mathbb R} ^n$. Suppose that: \begin{itemize} \item either $n=2$ and $u$ is stable (according to the notation recalled in Remark~\ref{stable}), \item or $n=3$, $s\in[1/2,1)$ and~$u_3>0$. \end{itemize} Then~$u$ is one-dimensional. \end{theorem} \begin{proof} We use the extension result in~\cite{Silv} and therefore we reduce this problem to an energy functional in~$(0,+\infty)\times{\mathbb R} ^n$ as the one in~\eqref{functional2} with~$${\mathcal{U}}:=(0,+\infty)\times {\mathbb R} ^{n-1}, \quad F(p,z,x):=x_{1}^{1-2s} |p|^2, \quad \quad G(z,x):=\tilde F(z),$$ where~$\tilde F$ is a primitive of~$f$ (notice that~$n$ in Theorem~\ref{GENERAL2} must be replaced by~$n+1$ for this application). Then, the desired energy growth follows from~\cite{CCinti, CCinti2}, according to which $$ {\mathcal E} _R(u)\le \left\{\begin{matrix} CR^{n-\min\{2s,1\}} & {\mbox{ if }}s\ne 1/2,\\ CR^{n-1}\log R & {\mbox{ if }}s=1/2. \end{matrix} \right.$$ Therefore, \eqref{EE} is satisfied when~$n=2$ and also when~$n=3$ and~$s\in( 1/2,1)$ (on the other hand, when~$n=3$ and~$s=1/2$, \eqref{EE weak} is satisfied and one has to make use of Remark~\ref{log}). Thus we obtain Theorem~\ref{DGs} as a consequence of Theorem~\ref{GENERAL2} and Remark~\ref{R1}. \end{proof} \subsection{Minimal surfaces} Minimal surfaces in ${\mathbb R} ^n$ can be thought as boundaries of sets $E \subset {\mathbb R} ^n$ that minimize the BV-norm or the perimiter (see \cite{giusti}) $$ Per(\chi_E):=\int |D\chi_E|\,dx.$$ Although the functional $F$ does not satisfy precisely the conditions of our theorems, the methods of proof of the next two sections easily apply to this case as well. Then condition \eqref{EE} reads $$Per(\chi_E, B_R)=\int_{B_R} |D\chi_E|\,dx \le C R^2.$$ On the other hand the perimeter of a minimal surface in $B_R$ is bounded by the surface area of $\partial B_R$, that is $C R^{n-1}$, hence the only global minimal surfaces in ${\mathbb R} ^3$ are the hyperplanes (one-dimensional). \subsection{Nonlocal minimal surfaces} As mentioned in the Introduction we discuss a result on nonlocal perimeters which was the original motivation for the techniques developed in this paper, see~\cite{CS}. Given a bounded domain~$\Omega\subset{\mathbb R} ^n$, the minimization of the following functional was introduced in~\cite{CRS}: $$ {\rm Per}_s(E,\Omega):= L(E\cap\Omega,{\mathbb R} ^n \setminus E) + L(E \setminus\Omega,\Omega\setminus E),$$ where $s\in(0,1)$ and for any disjoint measurable sets~$A$ and~$B$, $$ L(A,B) :=\int_A\int_B\frac{dx\,dy}{|x-y|^{n+s}}.$$ The regularity of $s$-minimal surfaces (i.e. of the boundary of a set~$E$ which minimizes~$ {\rm Per}_s(\cdot,\Omega)$ among all the measurable sets that agree with~$E$ outside $\Omega$) and of $s$-minimal cones (i.e. of $s$-minimal surfaces~$E$ such that are invariant under dilations) has been studied in some recent papers, such as~\cite{CRS, CV2, CS, Bego}. In particular, a complete regularity theory holds in the plane, according to the following result, that may also be obtained as a byproduct of the results in this paper: \begin{theorem}[\cite{CS}]\label{SVs} If $E$ is an $s$-minimal cone in~${\mathbb R} ^2$, then $E$ is a half-plane. \end{theorem} \begin{proof} By the extension result in Section~7 of~\cite{CRS}, we reduce the problem to a variational energy in~$(0,+\infty)\times{\mathbb R} ^2$, with~$$F(p,z,x):=x_1^{1-s} |p|^2,$$ for a minimizer homogenous of degree $0$. Then, \eqref{EE} easily follows in dimension $n=2$, and so we may use again Theorem~\ref{GENERAL2}. \end{proof} \subsection{Two-phase free boundary problem} This classical free boundary problem (see \cite{AC}, \cite{ACF}) consists in minimizing the energy $$\int |\nabla u|^2 dx + |\{u >0\}|.$$ In this case condition \eqref{EE} becomes $$\int_{B_R} |\nabla u|^2 dx \le C R^2,$$ which is clearly satisfied by a Lipschitz minimizer in dimension $n=2$. In conclusion, in ${\mathbb R} ^2$ any Lipschitz minimizer for the two-phase problem must be one-dimensional. \subsection{Thin one-phase problem} In this free boundary problem we minimize the following energy in ${\mathbb R} ^{n+1}_+$ (see \cite{DS}) $$\int_{{\mathbb R} ^{n+1}_+} |\nabla u|^2 dX + \mathcal H^n(\{ u(x,0)>0\}),$$ where we denote the points in ${\mathbb R} ^{n+1}$ by $X=(x,x_{n+1})$. Our results imply that in dimension $n=2$, any homogenous minimizer must be one-dimensional in the $x$ variable. This follows easily from \eqref{EE} since, due to the scaling of the energy, any homogeneous minimizer must be homogenous of degree $1/2$. \section{Local perturbations}\label{0055} In this section we show that in general we can perturb locally $$\max\{ u(x), u(x+te_n)\}$$ into a function with lower energy. The first lemma states that the maximum of two functions that form an angle at an intersection point cannot be an $e_n$-minimizer for ${\mathcal E} $ (this fact uses the strict convexity of~$F$ in the $p$ variable). \begin{lemma}\label{l1} Assume $0 \in \Omega$ and $u$, $v$ are $C^1$-functions such that \begin{equation}\label{angle} {\mbox{$u(0)=v(0)$ and $v_n(0)<0<u_n(0)$.}} \end{equation} Then $g:=\max\{u,v\}$ is not an $e_n$-minimizer for ${\mathcal E} $ in any ball $B_\eta$. \end{lemma} \begin{remark}\label{F5} {\rm In our setting, the transversal intersection described analytically by~\eqref{angle} can be obtained whenever $u$ is not monotone on each line along the $e_n$-direction. In this case we may reduce to the case in which~$u(\bar x+a_1 e_n)<u(\bar x+a_2e_n)$ and~$u(\bar x+a_2e_n)>u(\bar x+a_3e_n)$, with~$a_1<a_2<a_3$. Let~$c_i:=u(\bar x+a_i e_n)$. Then, by Sard's theorem we can find a regular value ~$c\in \big( \max\{c_1,c_3\},\, c_2\big)$ of $u$, thus we may find~$\alpha_c\in(a_1,a_2)$ and~$\beta_c\in (a_2,a_3)$ such that~$u(\bar x+\alpha_c e_n)=c= u(\bar x+\beta_c e_n)$ and ~$u_n (\bar x+\alpha_c e_n) > 0 > u_n (\bar x+\beta_c e_n)$. Then, the setting of~\eqref{angle} is fulfilled by supposing, up to translations, that~$ \bar x+\alpha_c e_n=0$ and by taking~$v(x):=u(x+(\beta_c-\alpha_c)e_n)$. }\end{remark} \begin{proof}[Proof of Lemma~\ref{l1}] Assume by contradiction that $g$ is an $e_n$-minimizer in some small ball $B_\eta$. We define~$F_0(p):=F(p,0,0)$, and we claim that we may reduce to the case in which \begin{equation*}\label{reduce} F_0(\nabla u(0))=F_0(\nabla v(0)). \end{equation*} To see this we notice that the property of minimality is not affected after subtracting a linear functional from $F$. Precisely if $$\tilde F(p,z,x):=F(p,z,x)-p_0\cdot p,$$ and $ \tilde{\mathcal E} _R$ is the associated energy functional for $\tilde F$ in $B_R$ then $${\mathcal E} _R(f)- \tilde {\mathcal E} _R(f)= \int_{B_R}p_0\cdot \nabla f \,dx =\int_{\partial B_R} f \, p_0\cdot\nu. $$ That is,~$\tilde {\mathcal E} _R(f)$ and~${\mathcal E} _R(f)$ only differ by a term depending on the boundary values of~$f$. Consequently, if~$f$ is an~$e_n$-minimizer for~${\mathcal E} $, it is also an~$e_n$-minimizer for~$\tilde{\mathcal E} $. Also, by possibly translating~$F$ in the~$z$-variable, we may assume that $u(0)=v(0)=0$. Now, for small~$r>0$, we consider the rescalings $$u_r(x):=r^{-1}u(rx), \quad v_r(x):=r^{-1}v(rx)$$ and we define~$g_r(x):=\max\{ u_r(x),v_r(x)\}$. Then,~$g_r$ is an $e_n$-minimizer for the rescaled functional $$F_r(p,z,x):=F(p, rz,rx)$$ in $B_{\eta/r}$. As $r \to 0^+$ then the following limits hold uniformly on compact sets: \begin{equation}\label{G3}\begin{split} & F_r \to F_0(p), \\ & u_r (x)\to u_0(x):=\nabla u(0) \cdot x, \quad \quad \nabla u_r \to \nabla u_0,\\ &v_r(x) \to v_0(x):= \nabla v(0) \cdot x, \quad \quad \nabla v_r \to \nabla v_0.\end{split}\end{equation} So we let $$g_0=\max\{u_0,v_0\}.$$ From the strict convexity of $F$ in the $p$ variable we see that~$g_0$ is not a minimizer for $F_0$. Indeed we first construct $h_0$, $$h_0:=1+ \alpha u_0 +(1-\alpha) v_0-\rho_R(x'), \quad \quad \rho_R(x'):=\max\{0,|x'|-R\}$$ for some $\alpha \in (0,1)$ small and $R$ large. Then \begin{equation*}\label{G4} \max\{g_0, h_0\}, \end{equation*} coincides with $g_0$ outside $B_{R+C}$ and notice that in $B_R$ we are cutting the graphs of two transversal linear functions by a single one. This function has lower energy for~$F_0$ than the one of~$g_0$ provided that we choose $R$ sufficiently large. By using the uniform convergence in~\eqref{G3}, we see that $$h_r:=\max\{ g_r, h_0 \},$$ has lower energy for $F_r$ than the one of~$g_r$. Scaling back, we have that~$h_\star(x):=r h_r(x/r)$ has less energy for~$F$ in~$B_{r(R+C)}\subseteq B_\eta$ than the one of~$g$. To reach a contradiction, it remains to check that~$h_\star$ is indeed an allowed perturbation according to Definition~\ref{d2}. This is equivalent to say that $h_r$ is a piecewise Lipschitz domain deformation of $g_r$ with the Lipschitz norm bounded by $\delta$. To obtain this, we use our hypothesis $\nabla u_0 \cdot e_n >0 >\nabla v_0 \cdot e_n$ and the uniform convergence (in $C^1$) of $u_r$ and $v_r$ to $u_0$ respectively $v_0$. Then, by the Implicit Function Theorem, the part of the graph of $h_r$ where $h_0 > g_r$ is obtained from $u_r$ by a Lipschitz domain deformation with Lipschitz norm less than $\delta$, provided that $\alpha$ is chosen sufficiently small. \end{proof} \begin{remark}\label{F6} {\rm In the proof we also showed that if $u$, $v$ are $C^1$ functions with $u(0)=v(0)$ and $\nabla u(0) \ne \nabla v(0)$ then $g:=\max\{u,v\}$ is not a classical minimizer for ${\mathcal E} $ in $B_\eta$. } \end{remark} The second lemma deals with perturbations for $\max\{u(x), u(x+te_n) \}$ (for small $t$) near a non-degenerate point on $\{u_n=0\}$. \begin{lemma}\label{l2} Assume that $u \in C^2(\Omega)$ is a critical point for the energy ${\mathcal E} $ in a neighborhood of the origin and the functional $F \in C^2$ in a neighborhood of $(\nabla u(0),u(0),0)$. Assume that \begin{equation}\label{NoG} u_n (0) =0, \quad \nabla u_n(0) \ne 0\end{equation} and let \begin{equation}\label{3.7a} w(x):=\max\{u(x), u(x+t e_n) \}.\end{equation} Then, for any $\eta>0$, there exists a Lipschitz function $\varphi$ with compact support in $B_\eta$ such that $${\mathcal E} _\eta(w+t \varphi)-{\mathcal E} _\eta(w) \le - c t^2 \quad \quad \mbox{for all $t$ small,}$$ for some small $c>0$ depending on $u$, $F$ and $\eta$. \end{lemma} \begin{proof} Let \begin{equation}\label{3.7bis} v(x):=\frac{u(x+te_n)-u(x)}{t}\end{equation} and notice that \begin{equation}\label{v-u} \|v-u_n\|_{C^{0,1}(B_\eta)}=o(1) \quad \mbox{as $t \to 0$}. \end{equation} Given a Lipschitz function $g$ we use that $F \in C^2$ in the $(p,z)$ variables and obtain $${\mathcal E} _\eta(u+t g)={\mathcal E} _\eta(u) + t L(g) + t^2 Q(g) + o(t^2)$$ with $$L(g):=\int_{B_\eta} F_p \cdot \nabla g + F_z \, g \, dx , $$ $$Q(g):= \int_{B_\eta} G(\nabla g, g,x) \, dx = \int_{B_\eta} (\nabla g)^TF_{pp} \nabla g + 2 g F_{pz} \cdot \nabla g + F_{zz} g^2 \, \, dx.$$ In the integrals above the function $F$ and its derivatives are evaluated at $(\nabla u,u,x)$ and the constant in the error term $o(t^2)$ depends on $u$, $F$ and $\|g\|_{C^{0,1}(B_\eta)}$. Since $u$ is a critical point for ${\mathcal E} $ we see that if $\varphi$ has compact support in $B_\eta$ then \begin{equation}\label{08} {\mathcal E} _\eta (u+t v^+ + t \varphi) - {\mathcal E} _\eta (u + t v^+)= t^2(Q(v^++ \varphi)-Q(v^+)) +o(t^2).\end{equation} {F}rom \eqref{3.7a} and \eqref{3.7bis}, we see that \begin{equation}\label{080} w=u+tv^+.\end{equation} Also, we claim that, if~$\eta$ is sufficiently small, \begin{equation}\label{09} Q(v^+)-Q(u_n^+)=o(1) \ {\mbox{ and }} \ Q(v^+ +\varphi)-Q(u_n^+ +\varphi)=o(1). \end{equation} We prove the first relation, the second being analogous. For this, we fix~$\mu>0$ and we define~${\mathcal{A}}_\mu:= B_\eta\cap \{|u_n|\le\mu\}$ and~${\mathcal{B}}_\mu:= B_\eta\cap \{|u_n|>\mu\}$. {F}rom~\eqref{v-u}, we have that \begin{equation}\label{090} \lim_{t\to0} \|v^+-u_n^+\|_{C^{0,1}({\mathcal{B}}_\mu)}= \lim_{t\to0}\|v-u_n\|_{C^{0,1}({\mathcal{B}}_\mu)}=0.\end{equation} On the other hand, since~$\nabla u_n(0) \ne 0$, for small~$\eta$ we have that the measure of~${\mathcal{A}}_\mu$ is (at most) of the order of~$\mu$. This and~\eqref{v-u} yield that $$ \lim_{t\to0}|Q(v^+)-Q(u_n^+)|\le C\mu$$ and so~\eqref{09} follows since~$\mu$ can be taken arbitrarily small. {F}rom~\eqref{09} we see that if $\eta$ is sufficiently small we can replace $v^+$ by $u_n^+$ in the right hand side of~\eqref{08}: accordingly, recalling also~\eqref{080}, we obtain \begin{equation}\label{u7} {\mathcal E} _\eta(w+t\varphi)-{\mathcal E} _\eta(w)=t^2(Q(u_n^+ +\varphi)-Q(u_n^+))+o(t^2).\end{equation} On the other hand $u_n$, $0$ and $G$ satisfy the hypotheses of Remark~\ref{F6}, hence $u_n^+$ is not a minimizer of $Q$. Thus we can choose $\varphi$ such that $$ Q(u_n^+ +\varphi)\le Q(u_n^+)-c$$ for some small~$c>0$, possibly depending on $u$, $F$ and $\eta$. So, by~\eqref{u7}, $$ {\mathcal E} _\eta(u+t v^+ + t \varphi) - {\mathcal E} _\eta(u + t v^+) \le -\frac{c}2 t^2$$ for all small $t$. \end{proof} \begin{remark}\label{r2} {\rm If $\nabla u(0) \ne 0$ then the function $w+t \varphi$ can be interpreted (via the Implicit Function Theorem) as a Lipschitz domain deformation of $w$ in the $\nabla u(0)$-direction (see Definition \ref{d1}) and the $C^{0,1}$-norm of the deformation is bounded by $C t$. Notice that, in general, the $\nabla u(0)$-direction and the $e_n$-direction are different. }\end{remark} The non-degeneracy hypothesis $\nabla u_n \ne 0$ of Lemma~\ref{l2} can be checked easily from Hopf lemma if $F\in C^3$ in a neighborhood of $\nabla u(0)$, as next result points out. \begin{lemma}\label{L3} Assume that $u\in C^1(\Omega)$ is a critical point for ${\mathcal E} $ and $F\in C^3$ in a neighborhood of $(\nabla u(0),u(0),0)$. If $u_n(0)=0$ and $u_n$ does not vanish identically in a neighborhood of $0$ then there exists a point $x_0$ close to $0$ such that $u_n(x_0)=0$, $\nabla u_n(x_0) \ne 0$. \end{lemma} \begin{proof} Since $u$ is a critical function for ${\mathcal E} $ then it satisfies the elliptic equation \begin{equation*} G(D^2u,Du,u,x'):=div \, F_p (\nabla u,u,x') - F_z(\nabla u,u,x')=0.\end{equation*} From the De Giorgi-Nash-Moser theorem and the Schauder estimates (see~\cite{GT}) it follows that if ~$u$ is locally Lipschitz and $F \in C^{2,\alpha}$ then ~$u \in C^{2,\alpha}$ and the equation above is satisfied there in the classical sense. If $F \in C^3$ then $G \in C^1$ hence by differentiating the equation in the $e_n$-direction we see that $v=u_n$ satisfies the linearized equation (in the viscosity sense) $$Lv:=G_{ij} v_{ij} + G_{p_i} v_i + G_z v =0,$$ where the derivatives of $G$ are evaluated at $(D^2u,D u ,u,x')$. Since $v$ does not vanish identically we can apply Hopf lemma to $v$ at a point $x_0\in \{v=0\}$ which admits a tangent ball from either $\{v>0\}$ or $\{v<0\}$. \end{proof} \section{Perturbations at infinity}\label{s3} For all $R$ large we define the Lipschitz continuous function $\psi_R$ with compact support in ${\mathbb R} $ given by \begin{equation}\label{12.1} \psi_R(s) := \begin{cases}1, \quad 0 \leq s \leq \sqrt R, \\ \ \\2-\dfrac{2\log s}{\log R}, \quad \quad \sqrt R < s \leq R, \\ \ \\ 0, \quad s > R.\end{cases}\end{equation} Notice that \begin{equation}\label{12.2} \psi'_R(s) = \begin{cases}0, \quad s \in (0,\sqrt R) \cup (R,\infty), \\ \ \\ \dfrac{-2}{s\log R}, \quad s\in(\sqrt R,R).\end{cases}\end{equation} For $0< t \le \sqrt R/4$, we define a bi-Lipschitz change of coordinates: $$x \mapsto y(x):= x+ t \psi_R (|x|) e_n$$ and let $$u_{R,t}^+(y) = u(x).$$ Notice that $u^+_{R,t}(x)$ coincides with $u(x-te_n)$ in $B_{\sqrt R/2}$ and with $u(x)$ outside $B_{R}$. Next we estimate ${\mathcal E} _{R}(u^+_R)$ in terms of ${\mathcal E} _{R}(u)$. We have $$D_x y = I+A,$$ with $$A(x) = t \, \psi_R'(|x|)\begin{pmatrix} 0 & 0 & \cdots & 0 \\ 0 & 0 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ \frac{x_1}{|x|} & \frac{x_2}{|X|} & \cdots & \frac{x_n}{|x|} \end{pmatrix} $$ and $$\|A\| \leq t |\psi'_R(|x|)| \ll 1.$$ Notice that $$D_y x = (I+A)^{-1} = I - \frac{1}{1+trA} A.$$ We have, $$\nabla_y u^+_R = \nabla_x u \;D_yx, \quad dy = (1+tr A) dx,$$ thus \begin{eqnarray*} && \int_{\Omega \cap B_R} F(\nabla_y u^+_{R,t},u^+_{R,t},y') dy \\ &&\quad= \int_{\Omega \cap B_R} F \left (\nabla_x u \, \, \left (I- \frac{1}{1+trA} A\right ), u, x'\right) (1+tr A) dx.\end{eqnarray*} We bound the right hand side from above by using that $|(pA)| \le |p \cdot e_n| /4$ which together with hypothesis \eqref{H2} for $F$ gives that $$F \left (p \, \, \left (I- \frac{1}{1+trA} A\right ), z, x'\right) (1+tr A) $$ is bounded above by \begin{equation}\label{ab} F(p,z,x') (1+ tr A) - F_p(p,z,t) \cdot (pA) + C |F_{pp} (p,z,t)| |pA|^2.\end{equation} By writing the same inequality for $u^-_{R,t}$ which is defined as $u_{R,t}^+$ with $t$ replaced by $-t$, thus $A$ is replaced by $-A$ in the formulas above, we obtain \begin{align}\label{AB} \nonumber & {\mathcal E} _{R}(u^+_{R,t})+{\mathcal E} _{R}(u^-_{R,t})-2 {\mathcal E} _{R} (u) \\ & \qquad\le C \int_{\Omega \cap B_R} |F_{pp}(\nabla u,u,x')|| \nabla u|^2 |A|^2 dx \\ & \qquad \le C \frac{t^2}{(\log R)^2} \int_{\Omega \cap \big( B_R \setminus B_{\sqrt R} \big)}\frac {|F_{pp}(\nabla u,u,x')|| \nabla u|^2}{|x|^2} dx. \nonumber \end{align} We denote by \begin{equation} \label{5.2a} a(r):= \int_{\Omega \cap B_r} |F_{pp}(\nabla u,u,x')|| \nabla u|^2 dx\end{equation} and by hypothesis \eqref{EE} we know that~$a(r) \le Cr^2$. Then the last integral in~\eqref{AB} is controlled, in polar coordinates, by \begin{equation}\label{5.2b} \int_{\sqrt R}^R a'(r) r^{-2} dr \le a(R)R^{-2} + 2 \int_{\sqrt R}^R a(r) r^{-3} \le C \log R.\end{equation} {F}rom \eqref{AB} and~\eqref{5.2b} we conclude that \begin{equation}\label{MAJOR} \limsup_{R\rightarrow+\infty}\sup_{t\in(0,\sqrt R/4)} t^{-2}\Big( {\mathcal E} _R(u^+_{R,t})+{\mathcal E} _R(u^-_{R,t})-2{\mathcal E} _R(u)\Big)\le0. \end{equation} \section{Proofs of Theorems \ref{GENERAL} and \ref{GENERAL2}}\label{33s} \begin{proof}[Proof of Theorem \ref{GENERAL}] Since $u$ is an $e_n$-minimizer we know that $${\mathcal E} _{R}(u^+_{R,t}) \ge {\mathcal E} _{R}(u).$$ This and~\eqref{MAJOR} imply that, for any fixed $t$, we have \begin{equation}\label{3.0} \lim_{R \to +\infty} {\mathcal E} _{R} (u^-_{R,t}) - {\mathcal E} _{R} (u) =0.\end{equation} Now we recall the integral formula \begin{equation}\label{3.1} {\mathcal E} _{R} (\max\{u^-_{R,t}, u \}) + {\mathcal E} _{R} (\min\{u^-_{R,t}, u \}) = {\mathcal E} _{R} (u^-_{R,t}) + {\mathcal E} _{R} (u) ,\end{equation} and we make use of the minimality of~$u$, which implies that \begin{equation}\label{3.2} {\mathcal E} _{R}(\min \{u^-_{R,t}, u\}) \ge {\mathcal E} _{R}(u). \end{equation} By~\eqref{3.0}, \eqref{3.1} and~\eqref{3.2} we find \begin{equation}\label{8.1} \lim_{R \to +\infty} {\mathcal E} _{R} (v_{R,t}) - {\mathcal E} _{R} (u) =0,\end{equation} with \begin{equation}\label{5.4a}v_{R,t}:=\max\{ u_{R,t}^-,u\}.\end{equation} Notice that $$v_{R,t}=\max\{ u(x), u(x+te_n)\} \quad \mbox{in $B_{\sqrt R /4}$},$$ and $ v_{R,t} \in D_R^t(u)$. Now assume by contradiction that $u\in C^1(\Omega)$ is not monotone on a line in the $e_n$-direction. Then we can find $t>0$ so that $u(x)$, $u(x+te_n)$ satisfy the hypotheses of Lemma~\ref{l1} (say, at some point $x_0 \in \Omega$, recall Remark~\ref{F5}). Thus we can perturb $v_{R,t}$ locally near $x_0$ into $\tilde v_{R,t}$ such that \begin{equation}\label{8.2} {\mathcal E} _R (\tilde v_{R,t}) \le {\mathcal E} _R (v_{R,t}) - c\end{equation} for some fixed $c>0$ depending only on $u$. {F}rom~\eqref{8.1} and~\eqref{8.2} we contradict the minimality of $u$ as $R \to +\infty$. \end{proof} \begin{proof}[Proof of Theorem \ref{GENERAL2}] We argue as above and use Lemma~\ref{l2} instead. Given $\epsilon>0$ we choose $R$ large such that $${\mathcal E} _R(u^+_{R,t})+{\mathcal E} _R(u^-_{R,t})-2{\mathcal E} _R(u) \le \epsilon t^2.$$ Since $u$ is $\{e_n,e_{n+1}\}$-stable we have $${\mathcal E} _R(w) \ge {\mathcal E} _R(u) - \epsilon t^2 \quad \forall w \in \mathcal D^t_R(u),$$ for all $t$ small enough (the first relation above comes from~\eqref{MAJOR} and the second one from Definition~\ref{d5}). Then, using also~\eqref{3.1} and~\eqref{5.4a}, we obtain $${\mathcal E} _{R} (v_{R,t}) - {\mathcal E} _{R} (u) \le 3 \epsilon t^2.$$ If $u_n$ changes sign in $\Omega$ then from Lemma \ref{L3} we can find a point $x_0 \in \Omega$ such that $u$ satisfies the hypothesis of Lemma~\ref{l2} at $x_0$. Thus we can perturb $v_{R,t}$ locally near $x_0$ into $\tilde v_{R,t}$ such that $${\mathcal E} _R (\tilde v_{R,t}) \le {\mathcal E} _R (v_{R,t}) - ct^2, \quad \quad \tilde v_{R,t} \in \mathcal D_R^{Ct}(u),$$ for some $c, C >0$ depending only on $u$. In conclusion $${\mathcal E} _R (\tilde v_{R,t}) \le {\mathcal E} _R (u) + (3 \epsilon - c)t^2, $$ and we contradict the stability inequality if we choose $\epsilon \ll c$. \end{proof} \section{Proof of Theorem \ref{t3}}\label{44s} In this section we assume that the domain $\Omega$ and the functional $F$ are invariant under translations in the $e_k$,..., $e_n$-directions. We define the notion of $u$ to be stable with respect to piecewise Lipschitz deformations in all directions generated by $\{e_k,...,e_n\}$ (but not with respect to vertical $e_{n+1}$ deformations as in Definition \ref{d4}). Below to give a precise definition of $\{e_k,..., e_n\}$-stability, we modify Definitions~\ref{d1} and~\ref{d2} according to the following notation: \begin{definition} We say that $v$ is an $\{e_k,..., e_n\}$-Lipschitz deformation of $u$ in $B_R$ if there exist Lipschitz functions $\psi^{(k)},...,\psi^{(n)}$ with compact support in $B_R$, and \begin{equation}\label{555} \sum_{k\le i,j\le n} \|\psi^{(i)}_j\|_{L^\infty({\mathbb R} ^n)}^2 <1 \end{equation} such that $$v(x)=u(x+\psi^{(k)}(x)e_k+...+\psi^{(n)}(x)e_n).$$ \end{definition} We remark that, under condition~\eqref{555}, the map $$x\mapsto x+\psi^{(k)}(x)e_k+...+\psi^{(n)}(x)e_n$$ is a diffeomorphism. \begin{definition}\label{DD} Let $u\in C^{0,1}(\Omega)$. We say that $v\in C^{0,1}(\Omega)$ is a piecewise $\{e_k,..., e_n\}$-Lipschitz deformation of $u$ in $B_R$ and write $$v \in D_{R,k}(u)$$ if there exist a finite number $v^{(1)}$,..., $v^{(m)}$ of $\{e_k,..., e_n\}$-Lipschitz deformations of $u$ in $B_R$ such that $$v(x)=v^{(i)}(x) \quad \quad \mbox{for some $i$ (depending on $x$).}$$ Also, if all $v^{(i)}$ satisfy $$v^{(i)}(x)=u(x+\psi^{(i,k)}(x)e_k+...+\psi^{(i,n)}(x)e_n) \quad \mbox{with} \quad \|\psi^{(i,j)}\|_{C^{0,1}(\Omega)} \le \delta$$ for some $\delta>0$, we write $$v \in D^\delta_{R,k}(u).$$ \end{definition} \begin{definition}\label{d5.5} We say that $u$ is $\{e_k,..., e_n\}$-stable for ${\mathcal E} $ if for any $R>0$ and $\epsilon>0$ there exists $\delta>0$ depending on $R$, $\epsilon$ and $u$ such that for all $t\in(0,\delta)$ we have that~${\mathcal E} _R(u)$ is finite and $$ {\mathcal E} _R(v)-{\mathcal E} _R(u) \ge -\epsilon t^2, \qquad \forall v \in D^t_{R,k}(u).$$ \end{definition} Notice that Definitions~\ref{d5} and~\ref{d5.5} are quite different, since vertical perturbations are allowed in Definition~\ref{d5} but not in Definition~\ref{d5.5}. On the other hand, Definition~\ref{d5.5} allows for horizontal perturbations in $(n-k+1)$-horizontal directions, while only one horizontal direction may be perturbed in Definition~\ref{d5}. \begin{remark}\label{97} {\rm We point out that if $u \in C^1(\Omega)$ is $\{e_k,..., e_n\}$-stable and $u_k$,...,$u_n$ do not vanish all at some point then $u$ is a critical point for ${\mathcal E} $ in a neighborhood of that point (because any vertical perturbation~$u+\epsilon\psi$ may be written in this case as a horizontal perturbation in the span of~$\{e_k,...,e_n\}$, due to the Implicit Function Theorem). }\end{remark} \begin{proof}[Proof of Theorem~\ref{t3}] The proof of Theorem~\ref{t3} follows as before from Lemma \ref{l2}, Remark \ref{r2} and Lemma \ref{L3}. First we may suppose that $k<n$, otherwise the statement is trivial. Let $Y_0$ be a point in $\mathcal U \subset {\mathbb R} ^{k-1}$ and we want to show that $\tilde u$ is one-dimensional where $$\tilde u(x_k,..,x_n):=u(Y_0,x_k,..,x_n).$$ Assume that $0 \in {\mathbb R} ^{n-k+1}$ is such that $\nabla \tilde u(0) $ is nonzero and it points in the $e_k$ direction. Then we may apply Lemma \ref{l2} and Remark \ref{r2} in the $e_{k+1}$,.., $e_n$ directions and conclude that $\tilde u$ is constant in a neighborhood of $0$ in all these directions. Then the set $$\big\{ (x_{k+1},..,x_n) {\mbox{ s.t. }} \tilde u(0,x_{k+1},..,x_n)= \tilde u(0), \quad \nabla \tilde u(0,x_{k+1},..,x_n)=\nabla \tilde u(0) \big\} $$ is both open and closed, hence the level set $\{ \tilde u=\tilde u(0) \}$ contains the hyperplane $0 \times {\mathbb R} ^{n-k}$. This argument shows that at all points where $\nabla \tilde u$ is nonzero, the gradient must point in the $e_k$ direction, thus $\tilde u$ depends only on the $x_k$ variable. \end{proof} \begin{remark} {\rm We point out that condition \eqref{H2} on $F$ can be weakened in Theorems \ref{GENERAL2} and \ref{t3}. Since we only need~\eqref{MAJOR} as $t \to 0$ we see from Section \ref{s3} that it suffices to have that $$x\,\mapsto\, \sup_{|p-\nabla u| \le |\nabla u|/2} \quad |F_{pp} (p,u,x')| \, \, |\nabla u|^2 $$ is a locally integrable function.}\end{remark} We conclude this section with a version of Theorem~\ref{GENERAL} for $e_n$-minimiziers with respect to piecewise Lipschitz perturbations with norm bounded by~$\delta$. \begin{definition}\label{6188} We say that~$u\in C^{0,1}(\Omega)$ is a $\{\delta,e_n\}$-minimizer for~${\mathcal E} $ if for any $R>0$ we have that~${\mathcal E} _R(u)$ is finite and $${\mathcal E} _R(u) \le {\mathcal E} _R(v), \quad \quad \forall v \in D_{R}^\delta(u).$$ \end{definition} \begin{theorem}\label{GENERAL delta} Let~$\delta>0$ and~$u \in C^1(\Omega)$ be a $\{\delta,e_n\}$-minimizer for the energy ${\mathcal E} $ with $F$ satisfying~\eqref{H1} and~\eqref{H2}. If \eqref{EE} is satisfied, then~$u$ is monotone on each segment in the $e_n$-direction of length less than~$2\delta$, i.e., for any~$\bar x\in\Omega$, either~$u_n(\bar x+te_n)\ge0$ or~$u_n(\bar x+te_n)\le 0$ for any~$t\in (-\delta,\delta)$. \end{theorem} The proof of Theorem~\ref{GENERAL delta} is identical to the one of Theorem~\ref{GENERAL}, we just need to choose~$|t|<\delta$. \section{A one-dimensional example}\label{55s} In this section, we briefly discuss a one-dimensional example, to clarify some of the notions of $e_n$-minimality and $e_n$-stability. We consider \begin{equation}\label{EXA}\begin{split} & F(p,z):=p^2-z^2,\qquad\quad\Omega:={\mathbb R} \\ {\mbox{and }}& u(s) := \begin{cases}\cos(s+\pi/2), \quad {\mbox{ if }} s \leq -\pi/2, \\ \ \\ 1, \quad \quad \quad \quad \quad {\mbox{ if }} -\pi/2 < s < \pi/2, \\ \ \\ \cos(s-\pi/2), \quad {\mbox{ if }} s\geq\pi/2.\end{cases} \end{split}\end{equation} \begin{proposition}\label{P.EXA} The function~$u$ in~\eqref{EXA} is~$e_1$-stable in~$(-\pi,\pi)$, (see Definition~\ref{d5.5}). \end{proposition} \begin{proof} Notice that~$u \in C^{1,1}({\mathbb R} )\cap C^\infty({\mathbb R} \setminus\{-\pi/2,\pi/2\})$. We prove that \begin{equation}\label{6.2} \begin{split} &{\mbox{ for any $R\in(0,\pi)$, any~$\delta\in(0,\pi/2)$}}\\ &{\mbox{ and any Lipschitz function~$\phi$ supported in~$(-R,R)$}}\\ &{\mbox{ with $\phi\le0$ in $[-\pi/2,\pi/2]$ and~$\phi=0$ in $[\delta-(\pi/2),(\pi/2)-\delta]$}}\\ &{\mbox{ we have that }} {\mathcal E} _R(u+\phi)\ge {\mathcal E} _R(u). \end{split} \end{equation} To prove it, we may suppose~$R\in(\pi/2,\pi)$, and we define~$I:=(-R,R)\setminus [-\pi/2,\pi/2]$, $J_-:=(-R,\,\delta-(\pi/2))$, $J_+:=((\pi/2)-\delta,\,R)$ and~$J:=J_-\cup J_+$. Given~$\ell>0$, we also denote by~$\lambda_\ell=\pi^2/\ell^2$ the first Dirichlet eigenvalue in the interval of length~$\ell$. By taking~$\ell:=R-(\pi/2)+\delta \in (0,\pi)$, we obtain that $$ \int_{J_\pm}\dot\phi^2\,ds\ge \lambda_\ell \int_{J_\pm}\phi^2\,ds\ge \int_{J_\pm}\phi^2\,ds$$ therefore $$ \int_J \phi^2-\dot\phi^2\,ds\le 0.$$ So, we compute: \begin{eqnarray*} && {\mathcal E} _R(u)-{\mathcal E} _R(u+\phi) \\&&\qquad=2\int_{-R}^R u\phi-\dot u\dot\phi\,ds+\int_{-R}^R \phi^2-\dot\phi^2\,ds \\ &&\qquad= 2\int_{-\pi/2}^{\pi/2} \phi\,ds+ 2\int_{I} u\phi-\dot u\dot\phi\,ds+ \int_J \phi^2-\dot\phi^2\,ds\\ &&\qquad\le 0 - 2\int_{I} \ddot u\phi+\dot u\dot\phi\,ds+0\\&&\qquad = -2\int_{I} \frac{d}{ds}(\dot u\phi)\,ds\\ &&\qquad= (\dot u\phi)(-\pi/2) -(\dot u\phi)(\pi/2) \\&&\qquad=0, \end{eqnarray*} which establishes~\eqref{6.2}. Now let~$v\in D^t_{R,1}$, with~$0<t<\delta$. Then we define~$\phi(s):=v(s)-u(s)$. Notice that~$\phi$ is Lipschitz, with~$\|\phi\|_{C^{0,1}({\mathbb R} )}\le C\|u\|_{C^{1,1}({\mathbb R} )}t$, and supported inside~$(-R,R)$. Also, $v\le 1$, since~$v$ is a deformation of~$u$ and~$u\le1$. Therefore, for any~$s\in [-\pi/2,\pi/2]$, we see that~$\phi(s)=v(s)-1\le0$. Finally, since~$v$ is a horizontal deformation of~$u$ of size~$t$, we have that $$ \inf_{[\delta-(\pi/2),(\pi/2)-\delta]} v \ge \inf_{[\delta-(\pi/2)-t,(\pi/2)-\delta+t]} u \ge \inf_{[-(\pi/2),(\pi/2)]} u =1.$$ Consequently, if~$s\in [\delta-(\pi/2),(\pi/2)-\delta]$ we have that~$v(s)=1$ and~$\phi(s)=0$. So we can apply~\eqref{6.2} and obtain~${\mathcal E} _R(v)={\mathcal E} _R(u+\phi)\ge{\mathcal E} _R(u)$. \end{proof} As a consequence of Proposition~\ref{P.EXA}, we have that $e_n$-minimizers are not necessarily critical for the energy ${\mathcal E} $ at the points where the gradient vanishes. We recall that the situation for~$\{e_n, e_{n+1}\}$-stable solutions was different, since in that case the criticality of the energy functional was granted by the vertical perturbations. The example in~\eqref{EXA} may be modified in order to obtain~$\{\delta,e_1\}$-minimality in the whole of~${\mathbb R} $. For instance one may consider: \begin{equation}\label{EXA2}\begin{split} & F(p,z):=p^2-\max\{ z,0\}^2,\qquad\quad\Omega:={\mathbb R} \\ {\mbox{and }}& u(s) := \begin{cases} s+\pi, \quad {\mbox{ if }} s<-\pi \\ \ \\ \cos(s+\pi/2), \quad {\mbox{ if }} s \in[-\pi, -\pi/2], \\ \ \\ 1, \quad \quad \quad \quad \quad {\mbox{ if }} -\pi/2 < s < \pi/2, \\ \ \\ \cos(s-\pi/2), \quad {\mbox{ if }} s\in[\pi/2,\pi] \\ \ \\ \pi-s, \quad {\mbox{ if }} s>\pi .\end{cases} \end{split}\end{equation} Then the proof of Proposition~\ref{P.EXA} may be easily modified to obtain: \begin{proposition}\label{P.EXA2} The function~$u$ in~\eqref{EXA2} is a~$\{\delta,e_1\}$-minimizer, for any~$\delta \in (0,\pi/2)$, according to Definition~\ref{6188}. \end{proposition} This shows that the statement of Theorem~\ref{GENERAL delta} is optimal, since~\eqref{EXA2} provides an example of~$\{\delta,e_1\}$-minimizer which is monotone on intervals of length~$2\delta<\pi$ but not on intervals of larger length. \section{Proofs of some remarks}\label{details} \begin{proof}[Proof of Remark~\ref{R1}] We consider a continuous function $u$ which is monotone on each line in ${\mathbb R} ^n$. We show that \begin{equation}\label{62} {\mbox{for any~$t\in{\mathbb R} $, the sublevel~$\{u<t\}$ is a half-space}}\end{equation} (unless it is empty). {F}rom this, it follows that, for different values of~$t$, $\partial\{u<t\}$ gives a collection of hyperplanes (which are parallel, since the level sets~$\{u=t\}$ cannot intersect for different values of~$t$), and so~$u$ is one-dimensional. To prove~\eqref{62}, first we remark that, from the monotonicity on each line of~$u$, it follows that \begin{equation}\label{p 62} {\mbox{both $\{u<t\}$ and $\{u\ge t\}$ are convex sets.}} \end{equation} Then, we take~$p\in \{u<t\}$. Since~$u$ is continuous, there exists~$\varrho>0$ such that \begin{equation}\label{631} B_\varrho(p)\subseteq \{u<t\}. \end{equation} We enlarge~$\varrho$ till there exists a point \begin{equation}\label{633} q\in\{u=t\}\cap \partial B_\varrho(p). \end{equation} We denote by~$\Pi_-$ the open halfspace tangent to $B_\varrho(p)$ at~$q$ that contains~$p$, and by~$\Pi_+$ the closed halfspace tangent to $B_\varrho(p)$ at~$q$ that does not contain~$p$. By looking at all the lines passing through~$q$, we deduce from~\eqref{p 62}, \eqref{631} and~\eqref{633} that \begin{equation}\label{934} \Pi_-\subseteq \{u<t\}\end{equation} \begin{comment} Indeed, suppose, by contradiction, that there exists a point \begin{equation}\label{634} a_-\in\Pi_-\cap \{ u\ge t\},\end{equation} and let~$\ell_-$ be the segment~$a_-$ and~$q$. By~\eqref{633}, \eqref{634} and~\eqref{p 62}, we have that \begin{equation}\label{6631} \ell_- \subseteq \{u\ge t\}. \end{equation} On the other hand,~$\ell_-$ must intersect~$B_\varrho(p)$, i.e. there exists~$b_-\in \ell_-\cap B_\varrho(p)$. {F}rom~\eqref{631} we obtain that~$b_-\in \{u<t\}$, while from~\eqref{6631} that~$b_-\in\{u\ge t\}$. This contradiction proves~\eqref{934}. \end{comment} and \begin{equation}\label{935} \Pi_+\subseteq \{u\ge t\}.\end{equation} \begin{comment} Suppose, by contradiction, that there exists a point \begin{equation}\label{2634} a_+\in\Pi_+\cap \{ u<t\}.\end{equation} By continuity, a neighborhood of~$a_+$ also lies in~$\{ u<t\}$, therefore, without loss of generality, we may suppose that \begin{equation}\label{i} {\mbox{$a_+$ must lie in the interior of~$\Pi_+$. }}\end{equation} Now, let~$b_+$ be the symmetric point of~$a_+$ with respect to~$q$. That is, we consider the segment~$\ell_+$ which joins~$a_+$ and~$b_+$ and has~$q$ as middle point. {F}rom~\eqref{i}, we see that~$b_+\in\Pi_-$, and therefore, by~\eqref{934}, we obtain that~$b_+\in \{u<t\}$. This,~\eqref{2634} and~\eqref{p 62} imply that~$\ell_+ \subseteq \{u<t\}$. Since~$\ell_+\ni q$, we conclude that~$q\in \{u<t\}$, which is in contradiction with~\eqref{633}. This establishes~\eqref{935}.\end{comment} By taking the complementary sets in~\eqref{935} and noticing that~$\Pi_+$ is the complement of~$\Pi_-$, we conclude that $$ \Pi_-\supseteq \{u<t\}.$$ This and~\eqref{934} give that~$\Pi_-=\{u<t\}$, proving~\eqref{62}. \end{proof} \begin{proof}[Proof of Remark~\ref{stable}] Suppose that~$u\in C^2({\mathbb R} ^n)$ is a classical stable solution, i.e. a critical point of the energy functional satisfying~\eqref{stable eq}. Since $F\in C^2$, we have that for any Lipschitz function~$\phi$ supported in a given ball~$B_R$, \begin{equation*}\label{922}\begin{split} &F(\nabla u+t\nabla\phi,u+t\phi,x)-F(\nabla u,u,x) \\ =\, &t\Big( F_{p_i}\phi_i+F_z\phi\Big)+\frac{t^2}{2} \Big( F_{p_i p_j} \phi_i \phi_j+F_{zz}\phi^2+2F_{p_i z}\phi\phi_i\Big) +o(t^2), \end{split}\end{equation*} where the derivatives of $F$ are evaluated at~$(\nabla u,u,x)$. Notice that~$o(t^2)$ above only depends on the Lipschitz norms of~$\phi$ and~$ u$ in~$B_R$, and on the $C^2$-norm of~$F$ in a bounded set (depending on~$R$ as well). When we integrate the equality above over~$B_R$, the term of order~$t$ disappears since~$u$ is a critical point, therefore we obtain \begin{equation}\label{92}\begin{split} &{\mathcal E} _R(u+t\phi)-{\mathcal E} _R(u)\\&\qquad=\frac{t^2}2\int_{B_R} \Big( F_{p_i p_j} (\zeta) \phi_i \phi_j+F_{zz}(\zeta)\phi^2+2F_{p_i z}(\zeta)\phi\phi_i\Big)\,dx +o(t^2).\end{split}\end{equation} Dividing by~$t^2$ and recalling~\eqref{stable eq}, we conclude that $$ \int_{B_R} \Big( F_{p_i p_j} (\zeta) \phi_i \phi_j+F_{zz}(\zeta)\phi^2+2F_{p_i z}(\zeta)\phi\phi_i\Big)\,dx\ge 0.$$ Hence, going back to~\eqref{92}, we obtain that \begin{equation}\label{1555} {\mathcal E} _R(u+t\phi)-{\mathcal E} _R(u)\ge o(t^2).\end{equation} Now, given~$w \in\mathcal D^t_R(u)$, we take~$\phi:=(w-u)/t$. Notice that the Lipschitz norm of~$\phi$ is bounded uniformly in~$t$, therefore~\eqref{1555} implies~\eqref{55} and so~$u$ is $\{e_n,e_{n+1}\}$-stable. Viceversa, suppose that~$u$ is $\{e_n,e_{n+1}\}$-stable. Then $u$ is a critical point and~\eqref{55} implies~\eqref{stable eq} by choosing~$w:=u+t\phi$ and taking~$\epsilon$ arbitrarily small. This shows that $u$ is a stable solution. \end{proof} \begin{proof}[Proof of Remark \ref{log}] We define $e_0(s):=s$ and then recursively $$ e_k(s):=\exp (e_{k-1}(s))= \underbrace{\exp\circ\dots\circ\exp}_{\text{$k$ times}} s$$ for any $k\in{\mathbb N} $, $k\ge1$. Let $\theta_k(R):=e_{k}(\sqrt{\ell_k(R)})$. Notice that \begin{equation}\label{21} \ell_{k+1}(\theta_k(R))=\log\sqrt{\ell_k(R)}=\frac12 \log(\ell_k(R))= \frac{\ell_{k+1}(R)}{2}.\end{equation} By induction over $k$, one sees that \begin{equation}\label{56} \ell_k'(r)=\big( \pi_{k-1}(r)\big)^{-1},\end{equation} where the notation in \eqref{PI} was used together with the setting $\pi_{-1}(r):= 1$ (in this way, $\pi_{k}(r)=\ell_k(r) \pi_{k-1}(r)$ for any $k\in{\mathbb N} $). We obtain that \begin{eqnarray*} \pi_k'(r)&=& \sum_{m=0}^k \;\prod_{ \genfrac{}{}{0pt}{}{0\le j\le k}{j\neq m} }\ell_j(r)\ell_m'(r) \\ &=& \sum_{m=0}^k \;\prod_{j=m+1}^k \ell_j(r)\\ &\le& (k+1) \prod_{j=1}^k \ell_j(r) \\ &=&(k+1) r^{-1}\pi_k(r)\end{eqnarray*} for large $r$, and so \begin{equation}\label{7.8} -\frac{d}{dr} \big( \pi_k(r)\big)^{-2}= 2\big( \pi_k(r)\big)^{-3}\pi_k'(r) \le 2(k+1) r^{-1}\big( \pi_k(r)\big)^{-2} \end{equation} for large $r$. Now, recalling~\eqref{21}, we modify \eqref{12.1} as follows: \begin{equation}\label{12.1a} \psi_R(s):=\left\{ \begin{matrix} 1, & {\mbox{ if }}0\le s\le \theta_k(R), \\ \, \\ 2-\displaystyle\frac{2\ell_{k+1}(s)}{\ell_{k+1}(R)}, & {\mbox{ if }} \theta_k(R)<s\le R,\\ \, \\ 0, & {\mbox{ if }}s>R. \end{matrix} \right.\end{equation} {F}rom \eqref{12.1a} and \eqref{56} we see that \begin{equation}\label{12.2a} \psi'_R(s)=\left\{ \begin{matrix} -\displaystyle\frac{2}{\ell_{k+1}(R) \pi_k(s)}, & {\mbox{ if }} \theta_k(R)<s\le R,\\ \, \\ 0, & {\mbox{ otherwise}}. \end{matrix}\right. \end{equation} Notice that \eqref{12.1a} and \eqref{12.2a} reduce to \eqref{12.1} and \eqref{12.2} respectively when $k=0$. Then, we can argue as in Section \ref{s3}. In this case, \eqref{AB} gets replaced by \begin{equation}\label{7.7}\begin{split} & {\mathcal E} _R(u^+_{R,t})+{\mathcal E} _R(u^-_{R,t})-2{\mathcal E} _R(u)\\ &\qquad \le C\frac{t^2}{\big(\ell_{k+1}(R)\big)^2} \int_{\Omega\cap(B_R\setminus B_{\theta_k(R)})}\frac{|F_{pp}(\nabla u,u,x')|\, |\nabla u|^2}{\big( \pi_{k}(|x|)\big)^2}\,dx\\ &\qquad = C\frac{t^2}{\big(\ell_{k+1}(R)\big)^2} \int_{\Omega\cap(B_R\setminus B_{\theta_k(R)})} |F_{pp}(\nabla u,u,x')|\, |\nabla u|^2\, \sigma(|x|)\,dx, \end{split}\end{equation} where $$ \sigma(r):= \big( \pi_{k}(r)\big)^{-2}.$$ Therefore, we recall \eqref{5.2a} and we notice that, in this case, $a(r)\le Cr \pi_k(r)$ for large $r$, thanks to \eqref{EE weak}. So we use \eqref{7.8} and~\eqref{56}, and, instead of \eqref{5.2b}, in this case we bound the last integral on right hand side of \eqref{7.7} in polar coordinates by \begin{eqnarray*} && \int_{\theta_n(R)}^R a'(r)\sigma(r)\,dr\le a(R)\sigma(R)- \int_{\theta_n(R)}^R a(r)\sigma'(r)\,dr \\ &&\qquad\le CR \big(\pi_k(R)\big)^{-1}+C\int_{\theta_n(R)}^R \big(\pi_k(r)\big)^{-1} \,dr\\ &&\qquad= CR \big(\pi_k(R)\big)^{-1}+C\int_{\theta_n(R)}^R \ell_{k+1}'(r)\,dr\\ &&\qquad\le C+C\ell_{k+1}(R). \end{eqnarray*} Therefore, \eqref{7.7} gives in this case \begin{eqnarray*}&& \limsup_{R\rightarrow+\infty}\sup_{t\in(0,\theta_k(R)/4)} t^{-2}\Big( {\mathcal E} _R(u^+_{R,t})+{\mathcal E} _R(u^-_{R,t})-2{\mathcal E} _R(u)\Big)\\ &&\qquad\le \limsup_{R\rightarrow+\infty}\sup_{t\in(0,\theta_k(R)/4)} \frac{C}{\big( \ell_{k+1}(R)\big)^2} \,\big( 1+\ell_{k+1}(R)\big) =0,\end{eqnarray*} which replaces~\eqref{MAJOR} in this case. \end{proof} \begin{proof}[Proof of Remark~\ref{ABS}] Here we construct a one-dimensional example of a Lipschitz function $u:{\mathbb R} \rightarrow{\mathbb R} $ that satisfies \eqref{EE} and that minimizes the energy with respect to any $e_n$-Lipschitz deformation, without being monotone. For this we take $u(t):=|t|$, $\Omega:={\mathbb R} $ and $F:=|p|^2$. Then, \eqref{EE} is obvious, and clearly $u$ is not monotone. Let us check that it is minimal with respect to any $e_n$-Lipschitz deformation, as described in Definition \ref{d1}: for this let $\psi$ be Lipschitz and supported in $(-R,R)$, with $|\psi'|<1$, and $v(t)=u(t+\psi(t))= |t+\psi(t)|$. We have $$ |v'(t)|^2-|u'(t)|^2=(1+\psi'(t))^2-1= 2\psi'(t)+(\psi'(t))^2$$ for almost any $t\in(-R,R)$. Therefore, if we integrate over $(-R,R)$ and we use that $\psi(-R)=0=\psi(R)$, we obtain $$ {\mathcal E} _R(v)-{\mathcal E} _R(u)=\int_{-R}^R (\psi'(t))^2\,dt\ge 0,$$ which is the minimality with respect to $e_n$-Lipschitz deformations. It is worth noticing that $u$ is not an $e_n$-minimizer, since piecewise $e_n$-Lipschitz deformations may decrease the energy (this justifies the importance of Definition \ref{d2}). To show this, we take $R:=2$, \begin{eqnarray*} &&\psi^{(1)}(t):=\left\{ \begin{matrix} -\displaystyle\frac{2+t}{3}, & {\mbox{ if }} t\in [-2,1],\\ \,\\ t-2 , & {\mbox{ if }} t\in (1,2], \end{matrix}\right.\\ {\mbox{ and }}&& \psi^{(2)}(t):=\left\{ \begin{matrix} t+2 , & {\mbox{ if }} t\in [-2,-1],\\ \, \\ \displaystyle\frac{2-t}{3}, & {\mbox{ if }} t\in (-1,2]. \end{matrix}\right.\end{eqnarray*} Let also $v^{(i)}(t):=u(t+\psi^{(i)}(t))$ and $$ v(t):=\left\{ \begin{matrix} v^{(1)}(t), & {\mbox{ if }} t\in [-2,0],\\ v^{(2)}(t), & {\mbox{ if }} t\in (0,2]. \end{matrix}\right.$$ Then $v$ is a piecewise $e_n$-Lipschitz deformation of $u$ according to Definition \ref{d2} and one may explicitly compute that $$ v(t)=\frac{2(|t|+1)}{3}.$$ In particular, ${\mathcal E} _R(v)=(8/9)R<2R={\mathcal E} _R(u)$, which shows that $u$ is not an $e_n$-minimizer. \end{proof} \begin{proof}[Proof of Remark~\ref{6fsv}] We define~$a(t)$, $\lambda_i(t)$, $\Lambda_i(t)$ and~$A_{ij}(p)$ as in~\cite{FSV} (see, in particular, formulas~(1.4)--(1.6) there). To avoid confusion with the notation here, the function~$F$ introduced below~(1.6) in~\cite{FSV} will be denoted by~$\tilde F$. The goal is to apply Theorem~\ref{GENERAL2} with~$F(p,z,x):=\Lambda_2(|p|)+\tilde F(z)$ (since this and Remark~\ref{R1} here plainly imply Theorems~1.1 and~1.2 in~\cite{FSV}). For this, we need to check the convexity of~$F$ in~$p$ and conditions~\eqref{H2} and~\eqref{EE}. We may focus on the case~$n=3$, i.e. on the case of Theorem~1.2 of~\cite{FSV} (this allows us to take also assumptions~(B1) and~(B2) in~\cite{FSV}). {F}rom~(1.6) and~(1.5) in~\cite{FSV}, we see that $$ F_{p_i}(p,z,x)=\lambda_2(|p|)\,p_i=a(|p|)\,p_i$$ and so $$ F_{p_ip_j}(p,z,x)=a(|p|)\,\delta_{ij} +a'(|p|)\,|p|^{-1}p_i p_j=A_{ij}(p).$$ Therefore, Lemma~2.1 in~\cite{FSV} gives the desired convexity of~$F$ and it implies that \begin{equation}\label{FF1} {\mbox{$ |F_{pp}(p,z,x)|$ is bounded from above and below by $ C\big(\lambda_1(|p|)+\lambda_2(|p|)\big)$}}.\end{equation} On the other hand, by Lemma~4.2 of~\cite{FSV}, we have that, if~$|p|\le M$, then \begin{equation}\label{FF2} \lambda_1(|p|)\le C_M \lambda_2(|p|)\end{equation} and \begin{equation}\label{FF3} \lambda_2(|p+q|)\le C_M \lambda_2(|p|)\end{equation} if~$|q|\le |p|/2$, for suitable~$C_M>0$ (possibly varying line after line). By plugging~\eqref{FF2} into~\eqref{FF1} we obtain that, if~$|p|\le M$, \begin{equation}\label{FF5} |F_{pp}(p,z,x)|\le C_M \lambda_2(|p|).\end{equation} Using~\eqref{FF5} and~\eqref{FF3} we see that if~$2|q|\le |p|\le M$, $$ |F_{pp}(p+q,z,x)|\le C_M \lambda_2(|p+q|)\le C_M \lambda_2(|p|) \le C_M|F_{pp}(p,z,x)|,$$ which gives~\eqref{H2} (notice that we may suppose~$|p|=|\nabla u|\le M$ in this case). Moreover, using~\eqref{FF5} here and~(4.3) of~\cite{FSV}, we obtain $$ |F_{pp}(p,z,x)|\,|p|^2\le C_M \lambda_2(|p|)\,|p|^2= C_M a(|p|)\,|p|^2\le \Lambda_2(|p|).$$ This and~(5.16) in~\cite{FSV} imply $$ \int_{B_R} |F_{pp}(\nabla u,u,x)|\,dx\le C_M \int_{B_R}\Lambda_2(|\nabla u|)\,dx\le C_M R^2.$$ This shows that~\eqref{EE} holds true in this case: so we may use Theorem~\ref{GENERAL2}, then recall Remark~\ref{R1}, and obtain the one-dimensional results of Theorems~1.1 and~1.2 of~\cite{FSV}. \end{proof}

https://arxiv.org/abs/1903.06263

Odometers of Divisible Sandpile Models: Scaling Limits, iDLA and Obstacle Problems. A Survey

The divisible sandpile model is a fixed-energy continuous counterpart of the Abelian sandpile model. We start with a random initial configuration and redistribute mass deterministically. Under certain conditions the sandpile will stabilize. The associated odometer function describes the amount of mass emitted from each vertex during stabilization. In this survey we describe recent scaling limit results of the odometer function depending on different initial configurations and redistribution rules. Moreover we review connections to the obstacle problem from potential theory, including the connection between odometers and limiting shapes of growth models such as iDLA. Finally we state some open problems.

\section{Introduction} The divisible sandpile model is a continuous height and fixed energy counterpart of the Abelian sandpile model (ASM) \citet{BTW}. It was introduced by \citet{LevPer} to study scaling limits of two growth models: the rotor-router model and internal diffusion limited aggregation. The model is interesting by itself and has further many nice connections to free boundary PDE, potential theory, algebraic geometry or computational complexity, see \citet{LapGrowth} and references therein. Consider a locally finite and undirected graph $G=(V,E)$. A divisible sandpile configuration $s:V\rightarrow \R$ assigns to each $x\in V$ a \textit{height} $s(x)$ representing some \textit{mass} in case $s(x)>0$ and a \textit{hole} for $s(x)<0$. At each time step, if $s(x)>1$, then redistribute the excess $(s(x)-1)^+$ among the neighbours. Repeat for all such unstable vertices $x$. We will make two different choices on how to do that and capture the choice by $\Delta_V^{\star}$. In the case where the excess is redistributed equally to nearest neighbours, we have $\Delta_V^{\star}$, where $\Delta_V^{\star}=\Delta$ the usual graph Laplacian. For the other case, consider a discrete fractional Laplacian $\Delta_V^{\star}=\Delta^{\alpha/2}$ capturing the redistribution to far away neighbours proportionally to the transition probabilities of a long-rang random walk. Note that once $s$ is fixed the evolution is deterministic. This model is Abelian in the sense that the final configuration does not depend on the toppling order (we can choose parallel toppling for simplicity). Call \[ u_n(x) = \text{total amount of mass emitted before time $n$ from $x$ to the neighbours} \] It is easy to see that $u_n$ is increasing in $n$ and hence will converge to some \[ u:V\rightarrow [0,\infty]. \] Define the following dichotomy: either for all $x\in V$, $u(x)<\infty$ so the divisible sandpile configuration will stabilize or $u(x)=\infty$ for all $x$ which means that $s$ explodes. The first question one might ask : \begin{center} Under which conditions on $s$ and $G$ does the sandpile stabilize? \end{center} \noindent \textbf{Dichotomy between stabilization and explosion:} When $|V|<\infty$ then it was shown in Lemma 7.1 of \citet{LMPU} that for any initial configuration $s$ satisfying $\sum_{x\in V}s(x)\leq |V|$ the sandpile will stabilize and explode otherwise. In the infinite volume case assume that $(s(x))_{x\in V}$ is a collection of i.i.d. random variables. Then it turns out that the mean or density in the physical sense (mass per unit volume) $\rho=\mathbb{E}(s(o))$ is the main parameter to determine stabilization. In Lemma 4.2. of \citet{LMPU} it was proven that indeed for $\rho<1$, the divisible sandpile stabilizes and in Lemma 4.1 for $\rho >1$ explodes almost surely and on the boundary case $\rho=1$ (Theorem 1.1.) the authors in \citet{LMPU} proved under a finite variance condition that $s$ does not stabilize a.s. This result was further extended to initial configurations with possibly no mean (Lemma 1) or variance (Theorem 2+3) in \citet{CHRheavy}. Note the interesting fact that in the Abelian sandpile model, the density alone is not enough to determine whether the sandpile stabilizes or explodes. In fact, for $V=\Z^d$ there exist infinite volume measures $\mu$ with mean $\rho \in (\rho_c, 2d)$ such that $\mu$ is not stabilizable (has not probability one on configurations which will stabilize), compare Theorem 5.1 in \citet{AnneFrank}. Recall that the ASM is a toy for model for self-organized criticality (SOC). A SOC system drives itself into a critical state (typically having power-law decay of correlations) without fine-tuning any parameters (temperature pressure, etc.). Let us point out the debate about whether self-organized criticality intrinsically involves tuning a parameter, namely the density. It was claimed in a series of papers, see Section III of \citet{Vesp} and subsequent papers. In fact, the authors in \citet{VespDenCon} in Section II introduced the concept of a fixed-energy Abelian sandpile (FES) where mass is not lost at the boundary, contrary to the ASM. They conjecture that the critical densities of a FES and ASM are the same. This would imply that the power-laws in ASM are coming from a real phase transition and not reminiscent of a SOC state. Although the critical values they found are close, it could be shown in large scale simulations that they are not the same, compare Theorem 1 of \citet{AnneDrive}. Given that we know the sandpile will stabilize on a finite graph, e.g. under the condition $\sum_{x\in V}s(x)\leq |V|$, another question is: \begin{center} What is the law of $u$ and it's scaling limit? \end{center} \noindent \textbf{Law of the odometer and it's scaling limit:} Consider the initial configuration $s$ given by \[ s(x) = 1+ \sigma(x) - \frac{1}{n^d}\sum_{z\in V}\sigma(z). \] Then we have trivially that $\sum_{x\in V} s(x)= |V|$. The collection of random variables $(\sigma(x))_{x\in V}$ will be either i.i.d. and satisfy a finite second moment assumption, correlated Gaussian random variables or i.i.d. and in the domain of attraction of $\alpha$-stable random variables. Using the redistribution $\Delta_V^{\star}$ defined above note that $u$ satisfies \[ \Delta_V^{\star} u = 1- s. \] We see that roughly \[ \begin{split} u &= (-\Delta_V^{\star})^{-1}\left (\sum_{x\in V} \sigma(x) - \frac{1}{|V|} \sum_{z\in V}\sigma(z) \right) \approx (\Delta_V^{\star})^{-1}\left (\sum_{x\in V} \sigma(x) \right ). \end{split} \] Intuitively we can guess the scaling limit using the following considerations. Let $\sigma(x) \sim N(0,1)$ for all $x\in V$, independently, and $\alpha \in (0,2]$ then the authors in \citet{CHR17} in Theorem 1+2 and Theorem 3.4 of \citet{LongRange} prove that for $V=\Z^d_n$ the discrete torus of length $n$, \[ (\Delta_V^{\star})^{-1}\left (\sum_{x\in V} \sigma(x) \right ) \longrightarrow \Delta^{- \alpha/2} W \] in law to the continuum fractional field on $\T^d$. $W$ denotes spatial white noise and the notation $\Delta^{- \alpha/2} W$ is adapted from \citet{LSSW} equation 1.1. Special cases include: the Gaussian free field ($\alpha=1$) and the membrane model or bi-Laplacian field ($\alpha=2$). This convergence was first conjectured for a sub-case ($\sigma$'s having finite variance and $\Delta_V^{\star}=\Delta$ the usual graph Laplacian) in \citet{LMPU} after Proposition 1.3. It is notable that we cannot go beyond membrane models by playing with the long-range parameter $\alpha$. In fact, to obtain fractional Gaussian fields with regularity $\Delta^{-(1+\delta)} W$, for $\delta>0$ consider correlated initial Gaussian variables $(\sigma(x))_{x\in V}$ such that roughly \[ (\Delta_V^{\star})^{-1}\left (\sum_{x\in V} \Delta^{-\delta} \sigma(x) \right ) \longrightarrow \Delta^{- (1+\delta)} W. \] This is a special case of the results in, Theorem 1 in \citet{JanFourier}, where the authors constructed a family of fractional fields specified by Fourier multipliers $\widehat{K}$ via defining an appropriate inverse covariance kernel for the Gaussian correlated random variables $(\sigma(x))_{x\in V}$. Can we go beyond Gaussianity? Yes, by imposing that $\sigma$'s will not satisfy a second moment condition but rather being in the domain of attraction of a $\alpha$-stable distribution. Then the authors proved in Theorem 5 of \citet{CHRheavy} that the limiting odometer field is in fact a $\alpha$-stable random field on $\T^d$. (Note that the parameter $\alpha$ from $\alpha$-stable distributions is not the same $\alpha$ which plays a role in the definition of a discrete fractional Laplacian and long-range random walk. It will be clear from the context that they are not the same parameters.)\\ \noindent \textbf{Obstacle problem and iDLA:} An equivalent description of the odometer function can be achieved via solutions of an obstacle problem. Given an obstacle $\gamma$ satisfying $\Delta_V^{\star} \gamma = 1-s$ find a function $v$ such that $v(x)=\inf \{f(x)|f\geq \gamma, (-\Delta_V)^{\star}f \leq 0 \}$. Then the odometer is equal to \[ u_{\infty} = v - \gamma. \] Note that the scaling limits obtained in Section \ref{sec:scale} are, by abuse of notation, scaling limits of the \textit{obstacle} and not the full odometer function. Define the formal field on the discretized torus: \[ \Xi^u_n = \sum_{z\in \T^d_n} u^{\star}(z\cdot n) {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,1/2n)}(x) \] for $x\in \T^d$. However, the formal field given $\Xi^u_n$ whose convergence we study involves to take an inverse Laplacian resp. fractional Laplacian with zero eigenvalue. We need to remove it by imposing zero mean test functions $f$. Hence in law: \[ \langle \Xi^u_n , f \rangle = \langle \Xi^{\gamma}_n , f \rangle \] where $\Xi_n^{\gamma}$ is the discretized field involving $\gamma$ and not $v-\gamma$. $v$ turns out to be related to the maximum of $\gamma$. One open question is what is the scaling limit of this maximum. Knowing the full scaling limit of the odometer allows to answer questions related to limiting shapes of growths models such as iDLA, rotor-router models and divisible sandpiles starting from a pile of $n$ chips at the origin. As we will explain in Section \ref{sec:iDLA}, the limiting shape will be related to determining the so-called non-coincidence set $\{x\in \R^d| v(x)>\gamma(x)\}$ for the obstacle problem.\\ The paper is organized as follows. In Section \ref{sec:not} we set the notation and preliminaries. We will introduce abstract Wiener spaces, sandpile models, stabilization and the reformulation of the odometer function in terms of solutions of an obstacle problem. The subsequent Section \ref{sec:law} will give an overview over results determining the law of the odometer on the discrete torus subjected to different redistribution rules and initial configurations. Section \ref{sec:scale} deals with scaling limit results for the Gaussian case and gives the main ideas of the proofs. In the following Section \ref{sec:stable} we state scaling limit results for the non-Gaussian case, when dealing with initial configurations being in the domain of attraction of $\alpha$-stable random variables. Section \ref{sec:iDLA} makes the connection between limiting shapes for growth models and non-coincidence sets for the odometer function. The last Section \ref{sec:dis} discusses open problems to explore. \section{Notation and Preliminaries}\label{sec:not} \subsection{Fourier analysis and abstract Wiener spaces} \subsubsection{Fourier analysis on the torus and continuum fractional Laplacians} \label{subsec-fourier-torus} We will from now on assume that $V=\Z^d_n$. Let $\T^d$ be the $d$-dimensional torus, alternatively $\frac{\R^d}{\Z^d}$ or as $[-\frac12,\,\frac12)^d\subset\R^d$. The discrete torus with side length $n$ is denoted by $\Z_n^d:=[-\frac{n}{2},\,\frac{n}{2}]^d\cap \Z^d$. Finally $\T_n^d:=[-\frac12,\,\frac12]^d\cap (n^{-1}\Z)^d$ is the discretization of $\T^d$. Moreover call $B(z,\,r)$ a ball centered at $z$ of radius $r>0$ in the $\ell^\infty$-metric. Define the inner product for $\ell^2(\mathbb{Z}^d_n)$: \[ \langle {f,g} \rangle = \frac{1}{n^d} \sum_{z \in \mathbb{Z}^d_n} f(z) \overline{g(z)}. \] We will discuss two cases involving the classical graph Laplacian $\Delta_g$ and discrete fractional Laplacian $-(-\Delta_n)^{\alpha/2}$ for $\alpha \in (0,\infty)$. $\Delta_g$ is defined by \[ \Delta_g f(x) = \frac{1}{2d}\sum_{\|y - x\|=1}\left (f(y) - f(x) \right ) \] and \begin{equation}\label{fracLap} -(-\Delta_n)^{\alpha/2} f(x) = \left (\sum_{y\in \Z^d_n} f(y) p^{\alpha}_n(x-y) \right) - f(x), \end{equation} where $p_n^{\alpha}:\Z^d_n \times \Z^d_n \rightarrow [0,1]$ is the transition kernel for a long-range random walk defined by \begin{equation}\label{def-random-walk-in-zdn} p^{(\alpha)}_n(0,x): = c^{(\alpha)} \sum_{\substack{z \in \mathbb{Z}^d\backslash \{0\} \\ z \equiv x \!\!\!\!\! \mod \mathbb{Z}^d_n }} \frac{1}{\|z\|^{d+\alpha}} , \end{equation} where $c^{(\alpha)}=(\sum_{z \in \mathbb{Z}^d \backslash \{0\}} \frac{1}{\|z\|^{d+\alpha}})^{-1}$ is the constant such that $\sum_{x \in \mathbb{Z}^d_n} p^{(\alpha)}_n(0,x) = 1$ and $x \equiv z \mod \mathbb{Z}^d_n$ denotes that $x_j \equiv z_j \mod n$ for all $j \in \{1,\dots,d\}$. With abuse of notation we will from now on write $p^{(\alpha)}_n(x):=p^{(\alpha)}_n(0,x)$. Consider the Fourier basis given by the eigenfunctions of both Laplacians $\{\psi_w\}_{w\in \mathbb{Z}^d_n}$ with \begin{equation}\label{def-fourier-basis-discrete} \psi_w(z)=\psi^{(n)}_w(z):=\exp \bigg(- 2\pi i z\cdot\frac{w}{n}\bigg). \end{equation} Given $f \in \ell^2(\mathbb{Z}^d_n)$, we define its discrete Fourier transform by \[ \widehat{f}(w) = \langle f, \psi_{w}\rangle = \frac{1}{n^d}\sum_{z \in \mathbb{Z}^d_n} f(z) \exp \bigg(-2\pi i z\cdot\frac{w}{n}\bigg) \] for $w\in \mathbb{Z}^d_n$. The eigenvalues $\{\lambda_w\}_{w\in \mathbb{Z}^d_n}$ corresponding to the classical graph Laplacian are given by \begin{equation}\label{eq:eigenvalues} \lambda_w:=-4\sum_{i=1}^d\sin^2\biggl(\frac{\pi w_i}{n}\biggr ), \end{equation} whereas for the discrete fractional Laplacian, see equation 4.2 in \citet{LongRange}, \[ \lambda^{(\alpha,n)}_w = -c^{(\alpha)} \frac{1}{n^{d+\alpha}} \sum_{x \in \frac{1}{n} \mathbb{Z}^d \setminus \{0\}} \frac{\sin^2(\pi x \cdot w )}{\|x\|^{d+\alpha}}, \] where $c^{(\alpha)}$ is just the normalising constant of the associated long range-random walk in $\mathbb{Z}^d$. In equation 4.6 of \citet{JanFourier} we consider the following definition of discrete fractional Laplacian \[ (-\Delta_g)^{-a} f(\cdot) = \sum_{v\in \Z^d \setminus \{o\}} (-\lambda_v)^{-s} \widehat{f}(v)\psi_v(\cdot) \] for $f\in \ell^2(\Z^d_n)$. Note that in contrast to \eqref{fracLap} there is no random walk representation. Similarly, if $f, g \in L^2(\mathbb{T}^d)$ denote by \[ (f,g)_{L^2(\mathbb{T}^d)}:= \int_{\mathbb{T}^d} f(z) \overline{g(z)} \text{d}z \] the inner product. Consider the Fourier basis $\{\phi_{\xi}\}_{\xi \in \mathbb{Z}^d}$ of $L^2(\T^d)$ given by $$\phi_{\xi}(x):=\exp(-2\pi i \xi \cdot x)$$ so \[ \widehat{f}(\xi): =(f,\phi_{\xi})_{L^2(\mathbb{T}^d)} = \int_{\mathbb{T}^d} f(z) \mathrm{e}^{- 2 \pi i \xi \cdot z} \text{d}z. \] It is important to notice that for $f \in C^\infty(\mathbb{T}^d)$, if we define $f_n: \mathbb{Z}^d_n \to \mathbb{R}$ by $f_n(z) := f( {z}/{n})$, then for all $\xi \in \mathbb{Z}^d$, $\widehat{f_n}(\xi) \to \widehat{f}(\xi)$ as $n\rightarrow \infty$. Finally, we write $C^\infty(\T^d)/\sim$ for the space of smooth functions modulo the equivalence relation of differing by a constant. \begin{definition} Let $a\in \R$ we define the fractional operator $(-\Delta)^a$ as \[ (-\Delta)^a f(\cdot) = \sum_{v\in \Z^d \setminus \{o\}} \|v\|^{2a} \widehat{f}(v) \phi_v(\cdot) \] for $f\in L^2(\T^d)$. \end{definition} Remark that \[ (-\Delta)^a \phi_v(\cdot) = \| v\|^{4a} \phi_v(\cdot). \] Note that most results about continuum fractional Laplacians are obtained for $a \in (0,2)$, compare also different representations in Theorem 1 of \citet{Kwasnicki15}. The reason is that for $\alpha \in (0,2)$ and $f \in C^{\infty}(\mathbb{R}^d)$, $f(\cdot + e_i)=f(\cdot)$ for all $i =1, \dots,d$ and $\{e_i\}_{i=1}^d$ the canonical basis of $\mathbb{R}^d $ we can write \begin{equation}\label{def-fractional-laplacian-1} -(-\Delta)^{\frac{\alpha}{2}}f(x) :=\frac {2^{\alpha} \Gamma (\frac{d+\alpha}{2})}{\pi^{d/2}|\Gamma(-\frac{\alpha}{2})|} \int_{\mathbb{R}^d}\frac{f(x+y)+f(x-y)-2f(y)}{\|y\|^{d+\alpha}} \text{d} y, \end{equation} where the integral above is defined in the sense of principal value. The constant in front of the integral is chosen to guarantee that for $\alpha,\beta \in (0,2)$ such that $\alpha +\beta <2$, we have $(-\Delta)^{\frac{\alpha}{2}}(-\Delta)^{\frac{\beta}{2}}f = (-\Delta)^{\frac{(\alpha+\beta)}{2}}f$ for all $f \in C^\infty(\mathbb{T}^d)$. \subsubsection{Abstract Wiener Spaces and continuum fractional Laplacians}\label{subsec:AWS} Given a Fourier multiplier function $\widehat{K}:\mathbb{Z}^d\to\mathbb{R}_{>0}$ consider \begin{equation}\label{norm} (f,g)_{K,a} := \sum_{\xi\in\mathbb{Z}^d\setminus\{0\}} \widehat{K}(\xi) \|\xi\|^{4a} \widehat{f}(\xi) \overline{\widehat{g}(\xi)}. \end{equation} We will make the following assumptions on $\widehat{K}$: $\widehat{K}$ is positive, even and real-valued. It is straightforward to see that \eqref{norm} defines a proper inner product on $C^\infty(\T^d)/\sim$. Define $H^a_K(\T^d)$ to be the Hilbert space completion of $C^\infty(\T^d)/\sim$ with respect to the norm $(\cdot,\cdot)_{K,a}=\|\cdot\|^2_{K,a}$. Recall the definition of abstract Wiener space (see~\S 8.2 in \citet{Str08}). \begin{definition}\label{aws2} A triple $(H,B,\mu)$ is called an abstract Wiener space (from now on abbreviated {AWS}) if \begin{enumerate} \item $H$ is a Hilbert space with inner product $(\cdot,\cdot)_H$. \item $B$ is the Banach space completion of $H$ with respect to the measurable norm $\|\cdot\|_B$. Furthermore $B$ is supplied with the Borel $\sigma$-algebra $\mathcal{B}$ induced by $\|\cdot\|_B$. \item $\mu$ is the unique probability measure on $B$ such that for all $\phi\in B^*$ we have $\mu\cdot\phi^{-1} = \mathcal{N}(0,\|\widetilde{\phi}\|_H^2)$, where $\widetilde{\phi}$ is the unique element of $H$ such that $\phi(h)=(\widetilde{\phi},h)_H$ for all $h\in H$. \end{enumerate} \end{definition} In Lemma 3 and 4 in \citet{JanFourier} we proved the following: \begin{lemma} Let $\epsilon > d/4 - a$, $\widehat{K}$ a general positive Fourier multiplier and the Sobolev spaces $H^a_K$ w.r.t the norm \eqref{norm} defined above. Then the triple $(H^{a}_K,H^{-\epsilon}_K,\mu_{-\epsilon})$ is an AWS. \end{lemma} The measure $\mu_{-\epsilon}$ is the unique Gaussian law on $H^{-\epsilon}_K$ such that the characteristic functional is given by \begin{equation}\label{chara} \Phi(f) = \exp \left ( -\frac{1}{2} \| f \|^2_{K,a}\right ) \end{equation} for a test function $f \in \mathcal{F} = \left\{C^{\infty}(\T^d); \int_{\T^d} f(x) dx =0 \right \}$. \begin{definition} A Gaussian random field $\Xi^K$ in $H^a_K(\T^d)$ is a collection $\{\langle \Xi^K, f \rangle; f\in \mathcal{F}\}$ of random variables such that $\mathbb{E}(\langle \Xi^K, f \rangle^2 ) = \| f||^2_{K,a}$. \end{definition} Note that we wrote $\Xi^K$ to indicate the dependence of a possible Fourier multiplier and the corresponding Sobolev space $H^{a}_{K}$. We will sometimes just write $\Xi$ or $\Xi^{\star}$ for $\Xi^K$, it will clear from the context what we will mean by that. \subsection{Divisible sandpiles, stabilization and obstacle problem} In this section we would like to define the divisible sandpile model, state the dichotomy between stabilization and explosion and finally make a connection between the odometer function and solutions to obstacle problems. \subsubsection{Divisible sandpiles and odometers} \begin{definition} A \textit{divisible sandpile configuration} $s$ is a function $s:\mathbb{Z}^d_n \rightarrow \mathbb{R}$. \end{definition} For $x \in \mathbb{Z}^d_n$, if $s(x)\geq 0$, we can interpret $s(x)$ as a \textit{mass} ($s(x)>0$) or \textit{hole} ($s(x)<0$) on the site $x$. If $s(x) > 1$, we call it \textit{unstable} and otherwise \textit{stable}. We then evolve the sandpile according to the following dynamics: unstable vertices will \text{topple} by keeping mass $1$ and redistributing the excess $(s(x)-1)^+$ over the other neighbouring vertices. We will distinguish between two different distribution rules:\\ \noindent \textbf{Distribution rules:} \begin{itemize} \item[(n.n.):] the excess is redistributed according to the transition probabilities of a simple random walk (equally to all nearest neighbours) \item[(l.r.):] the excess is redistributed according to the transition probabilities $p^{(\alpha)}_n(\cdot)$, see \eqref{def-random-walk-in-zdn}, of a long-range random walk at each discrete time step. \end{itemize} Note that unstable sites in long-range divisible sandpile models distribute mass to \textit{all} vertices (including itself). One could generate a divisible sandpile on a graph from any random walk defined on it, where on each time step the mass which in each vertex sent to its neighbours is proportional to the transition probabilities. Let us write $(-\Delta)^{\star} \in \left \{ (-\Delta_g), (-\Delta)_n^{\alpha/2} \right \}$ for denoting a general graph Laplacian which defines the redistribution. We will specify, when needed, the redistribution rule. Let $s_t=(s_t(x))_{x\in \mathbb{Z}^d_n}$ denote the sandpile configuration after $t\in \mathbb{N}$ discrete time steps (set $s_0:=s$ the initial configuration). Most of the times we will use parallel toppling, which we can define via an algorithm in the following way: \begin{algorithm}\label{alg-divisible-long-range} Set $t=1$ then run the following loop: \begin{enumerate} \item if $\max_{x \in \mathbb{Z}^d_n} s_{t-1}(x) \le 1$, stop the algorithm; \item for all $x \in \mathbb{Z}^d_n$, set $e_{t-1}(x):=(s_{t-1}(x)-1)^+$; \item set $s_t(x):=s_{t-1} - (-\Delta)^{\star}_n e_{t-1}(x) $; \item increase the value of $t$ by $1$ and go back to step $1$. \end{enumerate} \end{algorithm} Let us illustrate the algorithm in the nearest neighbour case on a simple example. \begin{figure}[htb] \includegraphics[scale=1]{exampDSP.pdf} \caption{Example of a toppling sequence } \label{fig:example_top} \end{figure} We zoom into a local neighbourhood of an initial configuration with the displayed height configuration at $t=0$ in Figure \ref{fig:example_top}. In the next time step we topple $s(x)=5$ and redistributed the excess equally among the nearest neighbours. At $t=2$ we topple in parallel the 4 unstable neighbours $y$ such that $s(y)=2$ of $x$. What we observe is that the holes (``negative heights" ) will slowly fill up and the heights with large mass decrease. \begin{definition} The odometer function is equal to $u^{\star}_t: \mathbb{Z}^d_n \rightarrow [0,\infty]$, where $u^{\star}_t(x)$ denotes the total mass emitted from $x$ up to time $t$, that is $u^\star_t(x):= \sum_{i=0}^{t-1} e_{i}(x)$. The superscript $\star \in \{ \text{n.n., l.r.}\}$ will combine the nearest neighbour and long-range case. \end{definition} Note that analogously to (Section 2 in) \citet{LMPU} we have pointwise: \begin{equation}\label{equilibrium-equation-long-range-divisible-sandpile} s_{t}(x) = s(x) - (-\Delta)^{\star} u_{t}(x). \end{equation} By construction, the odometer is an increasing function in $t$, call it $u^{\star}_{\infty}:=\lim_{t\rightarrow \infty} u^{\star}_t$. The model is Abelian in the sense that the final configuration will not depend on the toppling order. \subsubsection{Stabilization} As in Section 1 of \citet{LMPU} and Section 2 of \citet{LongRange} we define the following dichotomy: either for all $x\in \mathbb{Z}^d_n$ we have \textit{stabilisation}, i.e. $u^{\star}_{\infty}(x)<\infty$ or \textit{explosion}, i.e. for all $x\in \mathbb{Z}^d_n:$ $u^{\star}_{\infty}(x)=\infty$. Let us first consider more general toppling procedures, see Definition 2.2 \citet{LongRange} which was adapted from Definition 2.1 of \citet{LMPU}. \begin{definition}\label{def-top-procedure} Let $T \subset [0,\infty)$ be a well-ordered set of toppling times such that $0 \in T$, $T$ is closed subset of $[0,\infty)$. A \textit{toppling procedure} is a function \begin{align*} T \times \mathbb{Z}^d_n &\longrightarrow [0,\infty)\\ (t,x)& \longmapsto u^{\star}_t(x) \end{align*} such that for all $x \in \mathbb{Z}^d_n$ \begin{enumerate} \item $u^{\star}_0(x) = 0$; \item $u^{\star}_{t_1}(x)\le u^{\star}_{t_2}(x)$ for all $t_1 \le t_2$; and \item if $t_n \uparrow t$, then $u^{\star}_{t_n}(x)\uparrow u^{\star}_t(x)$. \end{enumerate} \end{definition} Given a toppling procedure $u^{\star}$, we say that it is \textit{legal} for the initial configuration $s$ if \[ u^{\star}_{t}(x)-u^{\star}_{t^-}(x) \le (s_{t^-}(x)-1)^+, \] where $t^-:= \sup \{ r \in T: r < T \} \in T$ (as $T$ is closed), and \textit{finite} if \[ u^{\star}_{\infty}(x):= \lim_{t \to \sup T} u^{\star}_t(x) <\infty. \] The limit is always well defined due the monotonicity of $u^{\star}$. If $u^{\star}$ is finite, then we have that \[ s_\infty := \lim_{t \to \sup T} s_t = s - \lim_{t \to \sup T} (-\Delta)^{\star}u^{\star}_\infty \] is well defined and it is equal to $s -(-\Delta)^{\star}u^\star_\infty$. \begin{definition}\label{def-stab-top-procedure} Given a toppling procedure $u^{\star}$, we say that it is \textit{stabilising} for $s$ if $u^{\star}$ is finite and $s_\infty \le 1$ pointwise. We say that $s$ \textit{stabilises} if there exists a stabilising toppling procedure $u^{\star}$ for $s$. \end{definition} It is clear that, if $s$ does not stabilize, then $u^{\star}=\infty$. When $s$ stabilizes, then its odometer $u^{\star}_{\infty}(x)$ is the total amount of mass sent from $x$ to its neighbours in any legal stabilizing toppling procedure for $s$. The following proposition is also true for general finite connected graphs $V$, we will state it here for the particular case when $V=\Z^d_n$. \begin{proposition}\label{prop:stab} Let $s:\Z^d_n\rightarrow \R$ be any initial sandpile configuration satisfying $\sum_{x\in \Z^d_n} s(x)=n^d$ (*). Then $s$ stabilizes to the all 1 configuration and its odometer $u^{\star}$ is the unique function satisfying \[ \begin{cases} & s-(-\Delta)^{\star} u^{\star} = 1\\ & \min_{x\in \Z^d_n} u^{\star}(x) =0. \end{cases} \] \end{proposition} We will sketch the proof which can be found in Lemma 7.1. in \citet{LMPU} or Proposition 3.2 of \citet{ LongRange}. Note that we do not assume anything on the initial configuration, besides (*), and on the redistribution specified by $\Delta^{\star}$. \begin{proof}(Sketch.) First we notice that the the rank of the operator $\Delta^{\star}$ is $rank(\Delta^{\star})=n^d-1$. Then there exists $h:\Z^d_n \rightarrow \R$ such that \[ -(-\Delta)^{\star} h =s-1. \] Set $f=h-\min h$ then $f\geq 0$ and $s-(-\Delta)^{\star} f=1$ so $f$ stabilizes. By Proposition \ref{prop-least-action} (see next section) the smallest of such solutions is the odometer with minimum equal to 0. \end{proof} \subsubsection{Least action principle and the obstacle problem} The next Proposition 2.5. from \citet{LMPU} is a variational characterization of the odometer function. We rewrote it in terms of our general Laplacian $\Delta^{\star}$. \begin{proposition}[Least action principle]\label{prop-least-action} Let $s \in \mathbb{R}^{\mathbb{Z}^d_n}$, consider \[ \mathcal{F}_{s} := \Big\{ f: \mathbb{Z}^d_n \longrightarrow \mathbb{R}: f \geq 0, s -(-\Delta)^{\star}f \le 1 \Big\}. \] Let $\ell$ be any legal toppling procedure for $s$. We have \begin{enumerate} \item For all $f \in \mathcal{F}_{s}$ and $x \in \mathbb{Z}^d_n$ \[ \ell_\infty \le f; \] \item for all $u^{\star}$ stabilising toppling procedure for $s$ and $x \in \mathbb{Z}^d_n$ \[ \ell_\infty \le u^{\star}_\infty; \text{ and} \] \item for all $u^{\star}$ legal stabilising toppling procedure for $s$, then for all $x \in \mathbb{Z}^d_n$ \begin{equation}\label{odoLast} u^{\star}_\infty(x) = \inf \{f(x): f \in \mathcal{F}_{s} \} , \end{equation} therefore, both $u^{\star}_\infty$ and $s_\infty$ do note depend on the legal choice of the legal stabilising toppling procedure. \end{enumerate} \end{proposition} The variational characterisation \eqref{odoLast} has an equivalent formulation summarized in the following lemma: \begin{lemma}[Lemma 2.2. in \citet{LapGrowth}]\label{odo_lev} Let $\gamma:\Z^d \rightarrow \R$ satisfy $(-\Delta)^{\star} \gamma = s_0-1$. Then the odometer of $u$ in \eqref{odoLast} is given by \[ u=v-\gamma \] where $v(x) = \inf \{ f(x)|f\geq \gamma; (-\Delta)^{\star}f \leq 0 \}$. \end{lemma} The function $\gamma$ is also called the \textit{obstacle} and $v$ (defined above) the \textit{solution to the obstacle problem}. For a fixed \textit{obstacle surface} $\gamma$ we want to find a graph $f$ which stays above the obstacle and is superharmonic at the same time, meaning in particular that $f$ has no local minima. The solution will be the lowest possible $f$ satisfying the conditions. \section{Law of the odometer on the discrete torus and discrete fractional Gaussian fields}\label{sec:law} \subsection{Law of the odometer on the discrete torus} From Proposition \ref{prop:stab} it is clear that as long as the initial sandpile configuration $s$ satisfies the constraint (*), the sandpile will stabilize to the all 1 configuration. The simplest choice of an initial configuration $(\sigma(x))_{x\in \Z^d_n}$ satisfying (*) is if we take for all $x\in \Z^d_n$: \begin{equation}\label{ini_s} s(x) = 1 + \sigma(x) -\frac{1}{n^d} \sum_{y\in \Z^d_n} \sigma(y). \end{equation} Note that at this point we do not make any assumptions on the law of $(\sigma(x))_{x\in \Z^d_n}$. We will discuss two different cases, the i.i.d. case and non-i.i.d. case.\\ \noindent \textbf{Assumptions on the $\sigma$'s:} \begin{enumerate} \item[(A-ind):] $(\sigma(x))_{x\in \Z^d_n}$ are i.i.d. random variables with $\mathbb{E}(\sigma(x))=0$ and $\mathsf{Var}(\sigma(x))=1$. \item[(A-cor):] $(\sigma(x))_{x\in \Z^d_n}$ are multivariate centered Gaussian variables with stationary covariance $\mathbb{E}(\sigma(x)\sigma(y)) = K_n(x-y)$. \end{enumerate} In inverse Fourier transforms of positive, real Fourier multipliers $\widehat{K}_n$ on $\Z^d_n$ correspond to a positive definite function $(x,y) \mapsto K_n (x-y)$ on $\Z^d_n\times \Z^d_n$. In Lemma 5 in \citet{JanFourier} we proved an analog of Bochner's theorem, which we will recall for completeness. \begin{lemma} The map $(x,y) \mapsto K_n(x,y)=K_n (x-y)$ on $\Z^d_n \times \Z^d_n$ is positive definite, hence a well-defined covariance matrix if and only if the corresponding Fourier multipliers $\widehat{K}_n$ are real-valued, even and positive. \end{lemma} Define the Green's function on the torus as \begin{equation} g^{\star}(x,y) = \frac{1}{n^d} \sum_{z \in \mathbb{Z}^d_n }g^{\star}_z(x,y) \end{equation} and $g^{\star}_z(x,y)=\mathbb{E}_x[\sum_{t=0}^{\tau_z-1} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\{X^{\star}_t=y \}}]$ is the expected number of visits to $y$ by a random walk $(X^{\star}_t)_{t\in \N}$ starting from $x$ before being killed at $z$. Recall, we write $(X^{\star}_t)_{t\in \N}$ for $\star \in \{ n.n., l.r. \}$ so if $\star=n.n.$ then we mean the simple random random walk and if $\star=l.r.$ we will consider long-range random walk with transition probabilities specified by \eqref{def-random-walk-in-zdn}. The following Proposition \ref{prop:dislaw} summarizes the laws of the odometers for different distribution rules and initial conditions. It is a combination of Proposition 1.3 from \citet{LMPU}, Proposition 4 in \citet{CHR17}, Proposition 3.2 in \citet{LongRange} and Lemma 6 in \citet{JanFourier}. \begin{proposition}\label{prop:dislaw} Let $s$ be an initial divisible sandpile configuration satisfying \eqref{ini_s}, namely \[ s(x) = 1 + \sigma(x) - \frac{1}{n^d}\sum_{y\in \Z^d_n} \sigma(y) \] and the collection of random variables $(\sigma(x))_{x\in \Z^d_n}$ satisfying (A-ind) or (A-cor). Then the law of the odometer $u^{\star}$ is equal to \begin{equation}\label{char-field-long-range-divisible-eq-1} \{u^{\star}_\infty(x)\}_{x \in \mathbb{Z}^d_n} \sim \Big\{ \eta^{\star}(x)- \min_{z \in \mathbb{Z}^d_n}\eta^{\star}(z) \Big\}_{x \in \mathbb{Z}^d_n }, \end{equation} where $\eta^{\star}(x)=\sum_{y\in \Z^d_n} g^{\star}(x,y)(s(y)-1)$ and in particular has covariances given by: \begin{equation}\label{cov} \mathbb{E}(\eta^{\star}(x)\eta^{\star}(y)) = \begin{cases} \sum_{z\in \Z^d_n} g^{n.n.}(x,z)g^{n.n.}(z,y) & \text{ for } \star=n.n., (A-ind) \\ \sum_{z\in \Z^d_n} g^{l.r.}(x,z)g^{l.r.}(z,y) & \text{ for } \star=l.r., (A-ind) \\ \sum_{z,z'\in \Z^d_n} K_n(z-z')g^{n.n.}(x,z)g^{n.n.}(z',y) & \text{ for } \star=n.n., (A-cor) \end{cases} \end{equation} for all $x,y\in \Z^d_n$. \end{proposition} The proof is a direct consequence of Proposition \ref{prop:stab}. Note that the covariance function $f(x,y):=\mathbb{E}(\eta^{\star}(x)\eta^{\star}(y))$ solves the following discrete PDE (up to constants) \[ (-\Delta^{\star}_n)^2 f(x,y) = \delta_x(y) -\frac{1}{n^d}. \] Of particular interest are initial $\sigma$'s which are Gaussian. Then the distribution of $u^{\star}$ is completely determined by the covariance $(f(x,y))_{x,y\in \Z^d_n}$. In fact, assume that $\star=n.n$ and $\sigma$'s are standard normals, then we retrieve the discrete bi-Laplacian field \[ (\eta(x))_{x\in \Z^d_n} \sim N(0, (f(x,y))_{x,y}) \] where $f(x,y)$ solves the discrete bi-Laplacian equation $(-\Delta_g)^2 f(x,y) = \delta_x(y) -\frac{1}{n^d}$. Let us look at some examples for different laws of the odometer $u^{\star}$ depending on the diverse assumptions.\\ \noindent \textbf{Examples:} \textbf{Case 1: nearest neighbour distribution and $\sigma$'s are i.i.d. standard normals}\\ For the nearest neighbour case and under assumption (A-ind) when the $\sigma$'s are independent standard normals, the odometer interface is depicted in Figure \ref{uncor}. \begin{figure}[htb] \includegraphics[scale=0.3]{uncor.jpg} \caption{Realization of the odometer for $u^{n.n.}$ and $\sigma$'s are i.i.d. $\sigma \sim N(0,1)$} \label{uncor} \end{figure} \textbf{Case 2: nearest neighbour distribution and $\sigma$'s are correlated}\\ Let $\sigma$'s is $(\sigma(x))_{x\in \mathbb{Z}^d_n} \sim \mathcal N(0,\, K_n)$ when the covariance matrix $K_n$ is polynomially decaying. Consider for example the summable kernel ($K\in \ell^1(\Z^d)$): \[ K_n^{\pm}(x,y) = \begin{cases} 7 & \text{ if } x=y \\ \pm \| x-y\|^{-3} & \text{ otherwise }.\end{cases} \] The corresponding realizations of the odometer function are indicated in Figures~\ref{Fig:PosK}-\ref{Fig:NegK}. \begin{figure}[h] \centering \centering \begin{minipage}{.5\textwidth} \includegraphics[scale=1]{PosK.jpg} \caption{Realization of the odometer associated to $K_n^+$.}\label{Fig:PosK} \end{minipage}% \begin{minipage}{.5\textwidth} \centering \includegraphics[scale=1]{NegK.jpg} \caption{Realization of the odometer associated to $K^-_n$.}\label{Fig:NegK} \end{minipage} \end{figure} A combination of the two correlated interface is presented in Figure \ref{fig_comp}. \begin{figure}[htb] \includegraphics[scale=0.5]{CompK.jpg} \caption{Combination of the odometer interfaces associated to $K^+_n$ and $K^-_n$.} \label{fig_comp} \end{figure} Observe that the realizations of the odometer function w.r.t. correlated initial $\sigma's$ look similar to the uncorrelated case. The positive or negative correlations play a role for maxima of the interface and how much it fluctuates. Positive correlations increase the position of the maxima and negative decrease them. The last Figure \ref{smoth} corresponds to choose $K_n(x,y)=C\|x- y\|^{-1}$, so $K_n\notin \ell^1(\Z^d)$. We get a smooth interface corresponding to the field $\Delta^{-(1+1/4)}W$ in the continuum. \begin{figure}[htb] \includegraphics[scale=0.3]{smoth.jpg} \caption{Odometer for non-summable correlations} \label{smoth} \end{figure} \textbf{Case 3: long-range distribution and $\sigma$'s are i.i.d. standard normals}\\ The third example concerns initial $\sigma$'s which are i.i.d. standard Gaussians and a long-range type redistribution w.r.t. a parameter $\alpha$. \begin{figure}[h] \centering \centering \begin{minipage}{.5\textwidth} \includegraphics[width=0.8\linewidth]{Fig1.jpg} \caption{Realization of the odometer $u^{l.r.}$ associated to $\alpha=0.5$.} \end{minipage}% \begin{minipage}{.5\textwidth} \centering \includegraphics[width=0.8\linewidth]{Fig2.jpg} \caption{Realization of the odometer $u^{l.r.}$ associated to $\alpha=1$.} \end{minipage} \end{figure} \begin{figure}[h] \centering \centering \begin{minipage}{.5\textwidth} \includegraphics[width=0.8\linewidth]{Fig3.jpg} \caption{Realization of the odometer $u^{l.r.}$ associated to $\alpha=1.5$.} \end{minipage}% \begin{minipage}{.5\textwidth} \centering \includegraphics[width=0.8\linewidth]{Fig4.jpg} \caption{Realization of the odometer $u^{l.r.}$ associated to $\alpha=2$.} \end{minipage} \end{figure} Observe that the lower $\alpha$ the ``rougher" the interface looks like. \subsection{Mean behaviour of the odometer, extrema and fluctuations} As described in Section 1.3 of \citet{LMPU}, the expected odometer can be seen as a \textit{measure of how difficult it is to stabilize}. In other words, it gives the information on how much mass is emitted in average from each site depending on dimension. Let $\sigma(x) \sim N(0,1)$ being i.i.d. and satisfying $\sum_{x\in \Z^d_n}s(x)=n^d$ and we will now consider general $\Delta^{\star}$. Note that from Proposition \ref{prop:dislaw} we obtain by invariance: \[ \mathbb{E}\left(u^{\star}_{\infty}(o) \right) = \mathbb{E}\left(\max_{x\in \Z^d_n} \eta^{\star}(x) \right). \] The order of the mean odometer will give the order of the maxima of the Gaussian field. We combine Theorem 1.2 and Proposition 8.3, 8.8 from \citet{LMPU} in the following Theorem \ref{max_odo_nn}. The proof replies on the fact that in law the odometer is equal to a discrete bi-Laplacian field shifted to have minimum value 0. \begin{theorem}\label{max_odo_nn} There exists a constant $C_d>0$ such that \[ C_d^{-1} \phi_d(n) \leq \mathbb{E} (u_{\infty}^{n.n.}(x)) \leq C_d \phi_d(n) \] where for $n\geq 2$, the function $\phi_d$ is defined as \[ \phi_d(n) =\begin{cases} n^{2- d/2}, & d<4 \\ \log(n), & d=4 \\ (\log(n))^{1/2}, & d\geq 5. \end{cases} \] Moreover $\mathbb{E}\left(\max_{x\in \Z^d_n} \eta^{n.n.}(x) \right) \asymp \phi_d(n)$. \end{theorem} A combination of Proposition 8.2 and 8.7 in \citet{LMPU} yields the following proposition for the variance of the $\eta$- configuration, (see also Proposition \ref{prop:dislaw} for the definition of $\eta$) . \begin{proposition} There exist $c_d,C_d>0$ such that, if $r=\|x\|_2$, \[ c_d \psi_d(n,r) \leq \mathbb{E}((\eta^{n.n.}(x)-\eta^{n.n.}(o))^2) \leq C_d \psi_d(n,r) \] where \[ \psi_d(n,r) = \begin{cases} n r^2, & d=1\\ r^2 \log(n/r), & d=2\\ r, & d=3 \\ \log(1+r) & d=4 \\ 1, & d\geq 5. \end{cases} \] \end{proposition} If we now allow to spread mass relative to the distance of the points, hence according to a long-range random walk, how fast does the sandpile stabilize? It is intuitively clear that, if mass will spread to far at each toppling, then the stabilization should be faster than in the previous case. Indeed it is the case, namely let us state Theorem 3.4. from \citet{LongRange} \begin{theorem}\label{max_odo} Let $\alpha \in (0,\infty)\setminus \{2\}$, $\sigma(x) \sim N(0,1)$ i.i.d., $\star=l.r.$ and $\gamma=\min\{2,\alpha \}$. Then there exist $C_d$ such that \[ C^{-1}_d \Phi_d(n) \leq \mathbb{E}(u_{\infty}^{l.r.}(x)) \leq C_d \Phi_d(n) \] where \[ \Phi_d(n) = \begin{cases} n^{\gamma - d/2}, & \gamma > d/2\\ \log(n), & \gamma=d/2\\ (\log(n))^{1/2}, & \gamma>d/2. \\ \end{cases} \] \end{theorem} Note the interesting point that $\gamma\leq 2$ so we will never have the correspondence to models with critical dimension higher than 4. The next Proposition 4.6 from \citet{LongRange} will give us the order of the variance of the $\eta$'s. It is the crucial computation to get Theorem \ref{max_odo}. \begin{proposition}\label{prop-upper-bounds-gaussian-distance} For $\alpha \in (0,2)$ there is a constant $C = C_{d,\alpha}>0$ such that for all $n \geq 1$ and all $x \in \mathbb{Z}^d_n$, \[ \mathbb{E}[(\eta^{l.r.}(0) - \eta^{l.r.}(x))^2] \le C \Psi_{d,\alpha(n,\|x\|)}, \] where \begin{equation}\label{def-asymp-mean-u} \Psi_{d,\alpha}(n,r):= \begin{cases} n^{2\alpha - d-2} r^2 ,& \text{ if } \alpha > \frac{d}{2}+1\\ \log(\frac{n}{r}) r^2 ,& \text{ if } \alpha = \frac{d}{2}+1\\ r^{2\alpha -d} ,& \text{ if } \alpha \in (\frac{d}{2},\frac{d}{2}+1)\\ \log(r) ,& \text{ if } \alpha = \frac{d}{2}\\ 1 ,& \text{ if } \alpha > \frac{d}{2}\\ \end{cases}. \end{equation} \end{proposition} \begin{proof} The proof goes via noticing that \begin{align*} \mathbb{E}[(\eta^{l.r.}(0) - \eta^{l.r.}(x))^2] &= \sum_{ w \in \mathbb{Z}^d_n} (g^\alpha(w,0)-g^\alpha(w,x))^2\\ & = \frac{1}{n^d}\sum_{ w \in \mathbb{Z}^d_n\backslash \{0\}} \frac{\sin^2(\frac{\pi x\cdot w}{n})}{(\lambda^{(\alpha,n)}_w)^2} \\ & = \int_{\R^d} G_{n,d,\alpha,x}(y) dy \\ & = \int_{\R^d} \sum_{ w \in \mathbb{Z}^d_n\backslash \{0\}} \frac{\sin^2(\frac{\pi x\cdot w}{n})}{(\lambda^{(\alpha,n)}_w)^2} 1_{B(\frac{w}{n},\frac{1}{2n})}(y) dy. \end{align*} Lemma 4.7 from \citet{LongRange} roughly states that $\lambda_{w}^{(\alpha, n)} \asymp \| w/n \|^{\gamma}$ where recall that $\gamma=\min \{2,\alpha \}$ and extra logarithmic corrections for $\gamma=2$. The proof goes via showing that for some $C>0$ \[ C^{-1} \sum_{ w \in \mathbb{Z}^d_n\backslash \{0\}} \frac{\sin^2(\frac{\pi x\cdot w}{n})}{ \|w\|^{2\gamma}} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(\frac{w}{n},\frac{1}{2n})}(y) \leq G_{n,d,\alpha,x}(y) \leq C \sum_{ w \in \mathbb{Z}^d_n\backslash \{0\}} \frac{\sin^2(\frac{\pi x\cdot w}{n})}{ \|w\|^{2\gamma}} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(\frac{w}{n},\frac{1}{2n})}(y). \] The most involved technical part is to get the proper upper and lower bounds for the terms \[ H_{n,d,\alpha,x}(y)=\sum_{ w \in \mathbb{Z}^d_n\backslash \{0\}} \frac{\sin^2(\frac{\pi x\cdot w}{n})}{ \|w\|^{2\gamma}} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(\frac{w}{n},\frac{1}{2n})}(y) \] depending on the dimension and parameter $\alpha$ Lemmas 4.8-4.15 in \citet{LongRange}. Comparisons with proper Riemann intergrals are used for that. \end{proof} \begin{proof}[Proof of Theorem \ref{max_odo}] The main ingredients are (similarly to the proof of Theorem \ref{max_odo_nn} in \citet{LMPU}) Dudley's bound Proposition 1.2.1 in \citet{Tal05} and the Majorizing Measure Theorem 2.1.1. in \citet{Tal05}. The idea is to study the mean of the extremes of a centered Gaussian field $\{\eta(x)\}_{x \in T}$ for some set of indexes through the metric in $T$ induced by \begin{align}\label{def-gaussian-metric} \nonumber d_\eta : T \times T &\longrightarrow \mathbb{R} \\ (x,y)& \longmapsto \mathbb{E}[(\eta(x) -\eta(y))^2]^{\frac{1}{2} }. \end{align} ``Good bounds" for $d_{\eta}(x,y)$ (which will come from Propostition \ref{prop-upper-bounds-gaussian-distance}) will imply ``good bounds" for $\mathbb{E}\left (\max_{x \in T} (\eta^{l.r.}(x) )\right )$ using similar computations as in \citet{LMPU}. \end{proof} \subsection{Discrete fractional Gaussian fields} Discrete Gaussian Free Fields (DGFF) on the torus were studied for example in \citet{ABDGFF} and \citet{AleMani}. In \citet{ABDGFF} the author studied level-sets percolation of the Gaussian Free Field (GFF) on the discretized $d$-dimensional torus $\Z^d_n$ for $d\geq 3$. He approximates the GFF on the torus by a GFF on the lattice $\Z^d$. The authors in \citet{AleMani} construct a DGFF on a compact manifold. To the author's knowledge the first time the discrete fractional Gaussian field via long-range random walks on the torus was defined, is in \citet{LongRange}. It is consistent with the definition on regular bounded domains $D\subset \R^d$ which can be found in \citet{LSSW}. Note that the authors in \citet{LSSW} (see section 12.2) define it for $\alpha \in (0,2)$ whereas in \citet{LongRange} it is defined for all $\alpha >0$ in equation 2.2. \begin{definition} For $\delta >0$ define $V^{\delta}:=\delta \Z^d \cap D$. Then the zero-boundary discrete fractional Gaussian field (DFGF) $h^{\delta}$ is a multivariate Gaussian function with density at $f\in \R^{V^{\delta}}$ proportional to \[ \exp \left ( -\frac{1}{2}\sum_{(x,y)\in (\delta \Z)^2, x\neq y } \frac{|f(x)-f(y)|}{|x-y|^{d+2\alpha}} \delta^d\right ). \] \end{definition} Then $h^{\delta}$ can be interpreted as a linear functional on $C^{\infty}_c(D)$ by setting \[ (h^{\delta}, f) =\sum_{x\in V^{\delta}} h^{\delta}(x) f(x) \delta^d \] for all test functions $f \in C^{\infty}_c(D)$. \section{Scaling limits of the odometer and continuum fractional Gaussian fields}\label{sec:scale} \subsection{Continuum fractional fields} Let $\alpha \in \R$, $D\subset \R^d$ some regular domain and $W$ spatial white noise on $\R^d$. Recall that $W$ is a spatial white noise on $\R^d$ if for all Schwartz functions $f\in \mathcal{S}(\R^d)$ we have that $\langle W,f\rangle \sim N(0,\|f\|^2_{L^2(\R^d)})$. We adapt the formal notation of \citet{LSSW} to describe the family of fractional Gaussian fields in the continuum by \begin{equation}\label{FGF} FGF_{\alpha}(D) :=\Delta^{-\alpha/2} W. \end{equation} Special cases include the GFF for $\alpha=1$ or the bi-Laplacian model for $\alpha=2$. Depending on the parameters $\alpha,\,d$ the field $\Delta^{-\alpha/2} W$ will be either a random distribution or a random continuous function. A random distribution (contrary to a function) cannot be defined pointwise but instead tested against suitable test functions. More precisely, if the Hurst parameter $H$ of the $FGF_{\alpha}(D)$ field \[ H:=\alpha-\frac{d}{2} \] is strictly negative, then the limit is a random distribution while for $H\in (k,\,k+1)$, $k\in \mathbb{N}\cup\{0\}$, the field is a $(k-1)$-differentiable function (\citet{LSSW} presented the results for domains with zero boundary conditions). \begin{definition} For $d>2\alpha$ and $\alpha \in (0,2)$ we define the a fractional Gaussian field $\Xi^{\alpha}$ as the unique distribution on $(C^{\infty}_c(\R^d))^*$ such that for all test functions $f\in C^{\infty}_c(\R^d)$, the random variable $\langle \Xi^{\alpha}, f\rangle$ is a centered Gaussian variable with variance \[ \mathbb{E}(\langle \Xi^{\alpha}, f\rangle^2) = \int_{\R^d}\int_{\R^d} f(x)f(y)\|x-y\|^{2\alpha -d} dxdy. \] \end{definition} \subsection{Scaling limits of the odometer} Let us define the formal field for $x\in \T^d:$ \begin{equation}\label{xi} \Xi^{\star}_n(x) := \sum_{z\in \T^d_n} u^{\star}_{\infty} (z\cdot n) {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,1/2n)}(x). \end{equation} It means that for every point in the continuum torus $x\in \T^d$ we assign the value of the odometer on the discretized torus which is in a ball of radius 1/2n around $z$ containing $x$. The scaling $a_n^{\star}$ of the formal field will depend on the initial configuration and redistribution rule. For pedagogical reasons we will write the results in 2 different theorems. The first Theorem \ref{theo_nn} is proven in \citet{CHR17} (see Theorem 1+2) and \citet{JanFourier} (refer to Theorem 1) and the second Theorem \ref{theo_lr} is proven in \citet{LongRange} in Theorem 3.5. \begin{theorem}[Scaling limit for $\star=n.n$]\label{theo_nn} Let the initial sandpile configuration $s$ be defined as in \eqref{ini_s} with $\sigma$'s satisfying (A-ind) or (A-cor). Define the scaling $a_n^{n.n.}$ as \[ a_n^{n.n.} =\begin{cases} 4\pi^2 n^{\frac{d-4}{2}} & \text{ if } (A-ind) \\ 4\pi^2 n^{-2} & \text{ if } (A-cor) \end{cases} \] Then the formal field defined in \eqref{xi} converges in law towards \[ a^{n.n.}_n \Xi^{n.n}_n \rightarrow \begin{cases} \Delta^{-1}W & \text{ if } (A-ind)\\ \Delta^{- (1+\delta)}W & \text{ if } (A-cor), \delta >0\\ \end{cases} \] on $\T^d$. The convergence holds in the Sobolev space $H^K_{-\epsilon}(\T^d)$ with topology induced by the norm $\|\cdot \|_{-\epsilon,K}$ for $\epsilon > \max\{d/4+1,d/2\}$. \end{theorem} Remark that for the (A-ind) case the Fourier multiplier $\widehat{K}$ is constant equal to 1 and recall that we use the formal notation \eqref{FGF} for the limiting fields $\Delta^{-1}W$ resp. $\Delta^{- (1+\delta)}W$. The fields obtained in the previous theorem have regularity properties of bi-Laplacian models or smoother interface models for $\Delta^{- (1+\delta)}W $. Let us remark that the main Theorem 1 proved in \citet{JanFourier} is even stronger, more precisely, the Gaussian limiting fields have variance given by \[ \mathbb{E}(\langle \Xi , f \rangle^2) = \sum_{z\in \Z^d \setminus \{o\}} \widehat{K}(z)\|z\|^{-4} |\widehat{f}(z)|^2 \] and depending on a Fourier multiplier which is related to the covariance kernel of the initial distribution of $\sigma$. Choosing $\mathbb{E}(\sigma(x)\sigma(y))=K_n(x-y)$ such that $\widehat{K}(z) = \|z\|^{-4\delta}$ on $\Z^d\setminus 0$ provides the field $\Delta^{- (1+\delta)}W $ for $\delta>0$. If the initial covariance kernel is summable on $\Z^d$ then $\widehat{K}(z)=Const$ and we obtain the bi-Laplacian model. The following theorem states the scaling limit result for the long-range case. Note that in particular we can obtain a whole class of limiting fields with different regularity properties including the GFF on the torus. The odometer approach presents an alternative way to construct for example a Gaussian Free Field on the torus. \begin{theorem}[Scaling limit for $\star=l.r$]\label{theo_lr} Let the initial sandpile configuration $s$ be defined as in \eqref{ini_s} with $\sigma$'s satisfying (A-ind). Define the scaling $a_n^{l.r.}$ as \[ a_n^{l.r.} = \begin{cases} n^{\frac{d-2\alpha}{2}} & \text{ if } \alpha < 2 \\ n^{\frac{d-2\alpha}{2}}\log(n) & \text{ if } \alpha = 2 \\ n^{\frac{d-4}{2}} & \text{ if } \alpha > 2. \end{cases} \] Then the formal field defined in \eqref{xi} converges in law to \[ a^{l.r.}_n \Xi^{r.r}_n \rightarrow \Delta^{-\gamma/2}W. \] The convergence holds in the Sobolev space $H_{-\epsilon}(\T^d)$ with topology induced by the norm $\|\cdot \|_{-\epsilon}$ for $\epsilon > \max\{d/4+ \gamma/2,d/2\}$ and $\gamma=\min\{2,\alpha\}$. \end{theorem} The proofs involve showing finite dimensional convergence and tightness. For ensuring tightness we will need to assume that $\epsilon > d/2$. On the other side for the AWS to be well-defined we will need that for the nearest neighbour case we have that $\epsilon > 1+d/4$ and in the long-range case $\epsilon > \gamma/2 + d/4$. The kernels of the operators will be specified in Theorem \ref{kernel} below. It is a combination of Corollary 1 in \citet{LongRange} and Theorem 3 of \citet{CHR17} which follows from the proofs of the scaling limit results in Theorem's \ref{theo_lr}-\ref{theo_nn}. We added a result for the kernel of the operator in the correlated case (A-cor) which is not specified in \citet{JanFourier} but follows from analogous computations as before in \citet{CHR17}. We will distinguish between two cases, below the critical dimension $d^{\star}_c$ which is equal to $d_c^{nn}=4$ for $\star=n.n.$ and $d^{lr}_c=2\gamma$ for $\star=l.r.$ and above. Note that the kernel at the critical dimension $d=4$ for the nearest neighbour case and $d=2\gamma$ for the long-range case is still an open question. We expect some logarithmic behaviour to play a role. \begin{theorem}[Kernel of the fractional operator for $d>d_c^{\star}$]\label{kernel} \begin{enumerate} \item[Case $\star=n.n.$, (A-ind):] The kernel $\mathbb{K}$ of the bi-Laplacian operator on the torus is a $L^1(\R^d)$-function and is equal to \[ \mathbb{K}(x,y) = c_d \|x-y\|^{4-d} + h_d(x-y) \] where $h_d \in C^{\infty}(\R^d)$ and $c_d>0$ some constant depending on $d$. \item[Case $\star=n.n.$, (A-cor):] Let us consider initial correlated Gaussian's $\sigma$ such that $\widehat{K}(\cdot)=\| \cdot \|^{-4\delta}$ with $\delta>0$. The kernel $\mathbb{K}$ of the fractional field on the torus is a $L^1(\R^d)$-function such that \[ \mathbb{K}(x,y) = c_{d,\delta} \|x-y\|^{4(1+\delta)-d} + h_{d,\delta}(x-y) \] for $h_{d,\delta} \in C^{\infty}(\R^d)$ and $c_{d,\delta}>0$ some constant depending on $d$ and $\delta$. \item[Case $\star=l.r.$, (A-ind):] Let $\alpha \in (0,2)$. The kernel $\mathbb{K}$ of the fractional field on the torus is a $L^1(\R^d)$ function such that \[ \mathbb{K}(x,y) = c_{d,\delta} \|x-y\|^{2\alpha-d} + h_{d,\alpha}(x-y) \] for $h_{d,\alpha} \in C^{\infty}(\R^d)$ and $c_{d,\alpha}>0$ some constant depending on $d$ and $\alpha$. \end{enumerate} \end{theorem} \begin{theorem}[Kernel of the fractional operator for $d<d_c^{\star}$] \begin{enumerate} \item[Case $\star=n.n.$, (A-ind):] For $x, y \in \T^d:$ $$\mathbb{K}(x,y) = \sum_{v\in \Z^d \setminus \{o\}} \frac{e^{2\pi i (x-y)v}}{\|v\|^4}$$. \item[Case $\star=n.n.$, (A-cor):] Let us consider initial correlated Gaussian's $\sigma$ such that $\widehat{K}(\cdot)=\| \cdot \|^{-4\delta}$ with $\delta>0$. Then \[ \mathbb{K}(x,y) = \sum_{v\in \Z^d \setminus \{o\}} \frac{e^{2\pi i (x-y)v}}{\|v\|^{4(1+\delta)}}. \] \item[Case $\star=l.r.$, (A-ind):] Let $\alpha \in (0,2)$. Then \[ \mathbb{K}(x,y) = \sum_{v\in \Z^d \setminus \{o\}} \frac{e^{2\pi i (x-y)v}}{\|v\|^{2\alpha}}. \] \end{enumerate} \end{theorem} The proofs are based on rewriting the variance of the field as an double integral and identifying the kernel. A priori the kernel can be written as a sum in any dimension (compare with the previous theorem). The fact that in the supercritical dimensions $d>d^{\star}_c$ power-law decay of correlations appears is due to the existence of infinite volume Gibbs measures in $\R^d$ of the corresponding bi-Laplacian field or fractional field. In the sub-critical dimensions there is only an implicit formula for the kernel. \subsection{Main ideas of the scaling limit proofs} Let us describe the main ingredients and ideas of the proofs of Theorems \ref{theo_lr}-\ref{theo_nn}, based on \citet{CHR17, LongRange} and \citet{JanFourier}. Recall that we proved in Proposition \ref{prop:dislaw} that the odometer on the discrete torus is distributed as \begin{equation*} \{u^{\star}_\infty(x)\}_{x \in \mathbb{Z}^d_n} \sim \Big\{ \eta^{\star}(x)- \min_{z \in \mathbb{Z}^d_n}\eta^{\star}(z) \Big\}_{x \in \mathbb{Z}^d_n }, \end{equation*} where $\eta^{\star}(x)=\sum_{y\in \Z^d_n} g^{\star}(x,y)(s(y)-1)$ has mean 0 and covariance specified in \eqref{cov}. Since we have mean 0 test functions $f$, note the important simplification: \[ \begin{split} \langle \Xi^{\star}_n , f\rangle & = \int_{\T^d} \left (\sum_{z\in \T^d_n} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,1/2n) }(x)u^{\star}_{\infty}(nz) \right ) f(x) dx \\ & = \int_{\T^d} \left (\sum_{z\in \T^d_n} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,1/2n) }(x)\eta^{\star}_{\infty}(nz) \right ) f(x) dx - [\min_{y\in \Z^d_n} \eta^{\star}(y)] \int_{\T^d} \left (\sum_{z\in \T^d_n} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,1/2n) }(x) \right ) f(x) dx\\ & = \int_{\T^d} \left (\sum_{z\in \T^d_n} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,1/2n) }(x)\eta^{\star}_{\infty}(nz) \right ) f(x) dx \\ & = \int_{\T^d} \left (\sum_{z\in \T^d_n} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,1/2n) }(x) \left [ \sum_{w\in \Z^d_n} g^{\star}(nz,w)\sigma(w) \right] \right ) f(x) dx \end{split} \] We need to impose mean zero on the test function to get rid of the 0 eigenvalue of the (fractional) Laplacian(s) in order to be able to invert them. Indeed, \[ \mathbb{E}(\eta_{\infty}^{\star}(x)\eta_{\infty}^{\star}(y)) = const + \sum_{z\in \mathbb{Z}^d_n\setminus \{o\}} \widehat{g}_x(z) \overline{\widehat{g}_y(z)} \] and $const$ is a dimensional constant which does not depend on $x,y$. By abuse of notation we will write \[ \langle \Xi^{\star}_n , f\rangle = \int_{\T^d} \left (\sum_{z\in \T^d_n} {\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,1/2n) }(x) \left [ \sum_{w\in \Z^d_n} g^{\star}(nz,w)\sigma(w) \right] \right ) f(x) dx . \] \noindent The proofs consist of two main steps.\\ \noindent \textbf{Step 1:} Tightness of $(\mathcal{L}(a_n^{\star} \Xi_n^{\star}))_{n\in \N}$ in $H_{-\epsilon}(\T^d)$ for $-\epsilon < -d/2$. \\ \noindent \textbf{Step 2:} From the first step we argue that there exists $(n_k)_{k\in \N}$ such that \[ a^{\star}_{n_k} \Xi^{\star}_{n_k} \overset{\mathcal{L}} \longrightarrow \Xi^{\star} \] as $k\rightarrow \infty$ in $H_{-\epsilon}(\T^d)$. It remains to show that the limit is unique. It suffices by Section 2.1 of \citet{ledoux:talagrand} to prove that for $f\in \mathcal{F}$ we have that the characteristic functional \begin{equation}\label{lim_chara} \lim_{n\rightarrow \infty} \mathbb{E} (\exp(i \langle \Xi^{\star}_n, f \rangle)) = \Phi(f) \end{equation} is the correct one with $\Phi$ is defined in \eqref{chara}. Let us give more details in the sequel.\\ \noindent \textbf{Tightness:} Let us give a sketch of the arguments used in order to prove tightness. Compare Section 4.2. in \citet{CHR17}, Section 3.1. from \citet{JanFourier} and Section 4 in \citet{LongRange}. We will use Rellich's theorem (\citet{Roe}) for showing tightness. \begin{theorem}[Rellich's theorem]\label{thm-rellich} If $k_1<k_2$ the inclusion operator $H^{k_2}(\mathbb{T}^d)\hookrightarrow H^{k_1}(\mathbb{T}^d)$ is a compact linear operator. In particular for any radius $R>0$, the closed ball $\overline{B_{ H_{-\frac{\varepsilon}{2}}}(0,\,R)}$ is compact in $H_{-\varepsilon}$. \end{theorem} Pick $k_1=-\epsilon$ and $k_2=-\frac{\epsilon}{2}$ and choose $-\epsilon > -\frac{d}{2}$. Observe that \[ \| a^{\star}_n{\Xi^{\star}_{n}}\|_{L^2(\mathbb{T}^d)}^2 = (a^{\star}_n)^2 \sum_{x,\,y\in \mathbb{Z}^d_n} g^{(\star)}(x,y)\sigma(x) \sum_{x',\,y'\in \mathbb{Z}^d_n }g^{(\star)}(x',y')\sigma(x') \] is a.s. finite, as, for any fixed $n$, it is a finite combination of essentially bounded random variables. Therefore $\Xi^{\star}_{n}\in L^2(\mathbb{T}^d) \subset H_{-\varepsilon}(\mathbb{T}^d)$ a.s. It will be enough to show that, for all $\delta>0$, there exists a $R=R(\delta)>0$ such that \[ \sup_{n\in \mathbb{N}} \mathbb{P} \Big(\| a^{\star}_n\Xi^{\star}_{n}\|_{H_{-\frac{\varepsilon}{2}}}\geq R\Big)\le \delta. \] Using Markov's inequality and exploiting the representation of the field $\Xi_n^{\star}$ in Fourier space, we can bound after some longer computations \[ \sup_{n\in \mathbb{N}} \mathbb{E} \Big[\| a^{\star}_n \Xi^{\star}_{n}\|_{ H_{-\frac{\varepsilon}{2}}}^2\Big] \le C. \] \noindent \textbf{Finite-dimensional convergence.} The first question is how do we guess the right scaling? To determine $a_n$ we do the following rough computation for (A-ind) and general $\star$: pick $x,y\in \T^d_n$ and compute the covariance of the odometer \[ \begin{split} \mathbb{E}(u^{\star}_{\infty}(n x) u^{\star}_{\infty}(n y)) &= \sum_{w \in \Z^d_n} g^{\star}(xn,w) g^{\star}(w,yn) \\ & \approx n^d \sum_{w\in \Z^d_n \setminus o} \widehat{g^{\star}}_{xn}(w) \overline{ \widehat{g^{\star}}_{yn}(w)} \\ & \approx n^{-d} \sum_{w\in \Z^d_n \setminus o} \frac{e^{2\pi in w(x-y)}}{|\lambda^{\star}_w|^2} \end{split} \] Now using roughly that \begin{equation}\label{ev} \lambda^{\star}_w \approx \begin{cases} \left \| \frac{w}{n}\right \|^2 & \text{ if } \star=n.n \\ \left \| \frac{w}{n}\right \|^{2}\log(n) & \text{ if } \star=l.r, \gamma= 2 \\ \left \| \frac{w}{n}\right \|^{\gamma} & \text{ if } \star=l.r, \gamma=\min\{2,\alpha\}, \gamma\neq 2\\ \end{cases} \end{equation} we argue \[ \mathbb{E}(u^{\star}_{\infty}(n x) u^{\star}_{\infty}(n y)) \approx \begin{cases} n^{4-d} \sum_{w\in \Z^d_n \setminus o}\frac{e^{2\pi in w(x-y)}}{\|w\|^4} & \text{ if } \star=n.n \\ n^{2-d}\log(n) \sum_{w\in \Z^d_n \setminus o} \frac{e^{2\pi in w(x-y)}}{\|w\|^{2}} & \text{ if } \star=l.r, \gamma=2 \\ n^{\gamma-d} \sum_{w\in \Z^d_n \setminus o} \frac{e^{2\pi in w(x-y)}}{\|w\|^{\gamma}} & \text{ if } \star=l.r, \gamma=\min\{2,\alpha\}, \gamma\neq 2. \end{cases} \] For proving convergence to the characteristic functional we will first prove moment convergence for initial $\sigma$'s whose moments exist. We state the following Proposition 4.16 from \citet{LongRange}. \begin{proposition}\label{prop-convergence-moments-bounded-case} Assume $\mathbb{E}[\sigma]=0, \mathbb{E}[\sigma^2]=1$ and that $ \mathbb{E}|\sigma|^k < \infty$ for all $k \in \mathbb{N}$. Then for all $m \geq 1$ and for all $f \in C^\infty (\mathbb{T}^d)/\sim$ with zero average, the following limits hold: \begin{equation} \lim_{n \to \infty} \mathbb{E}[\langle a^{\star}_n\Xi^{\star}_{n}, f \rangle^m ] = \begin{cases} (2m-1)!! [\mathbb{E}(\langle \Xi^{\star}, f \rangle^2)]^{m/2}, & m \in 2\mathbb{N} \\ 0, & m \in 2\mathbb{N} +1. \end{cases} \end{equation} \end{proposition} Note that it is more general than the corresponding Proposition 13 in \citet{CHR17} which assumes bounded $\sigma$'s almost surely instead of finite moments. Let us focus on the second moment and give a sketch for the proof for higher moments. For the second moment observe that for $f\in \mathcal{F}$, \begin{equation}\label{sec_mom} \begin{split} \mathbb{E}\left[\left\langle\Xi^{\star}_n,\,f \right\rangle^2 \right]&= n^{-2d}(a^{\star}_n)^2 \sum_{z,\,z'\in \T_n^d} f(z)f(z') \mathbb{E}[ \eta^{\star}(nz) \eta^{\star}(nz')]+\mathbb{E}\left[R_n(f)^2\right]\\ &+\mathbb{E}\left[n^{-d}a_n^{\star} \sum_{z\in \T_n^d} f(z)\eta^{\star}(nz)R_n(f)\right ] \end{split} \end{equation} where the reminder is equal to \[ R_n(f):= a_n^{\star} n^{-d} \sum_{z\in \T_n^d} \eta^{\star}(nz)\left(\int_{B(z,\,\frac{1}{2n})}n^d f(x) dx-f(z)\right). \] The key observation here is that the first term on the R.H.S. of \eqref{sec_mom} gives the desired variance and the remaining terms converge to 0 in $L^2(\T^d)$. Indeed we have, \begin{proposition} For any $f\in \mathcal{F}$, \[ \lim_{n\rightarrow \infty}\mathbb{E}\left[R^2_n(f)\right ] =0, \, \, \lim_{n\rightarrow \infty}\mathbb{E}\left[n^{-d}a_n^{\star} \sum_{z\in \T_n^d} f(z)\eta^{\star}(nz)R_n(f)\right ]=0. \] \end{proposition} The previous proposition corresponds to Propositions 6 in \citet{CHR17}, 4.16 in \citet{LongRange} and 9 in \citet{JanFourier}. The proof uses Cauchy-Schwartz inequality. It remains to show that (Proposition 5 in \citet{CHR17}, 4.17 in \citet{LongRange} and Proposition 8 in \citet{JanFourier}). \begin{proposition}\label{prop:var} For $f\in \mathcal{F}$, \[ \lim_{n\rightarrow \infty} n^{-2d} (a^{\star}_n)^2 \sum_{z,\,z'\in \T_n^d} f(z)f(z') \mathbb{E}[ \eta^{\star}(nz)\eta^{\star}(nz')] = \begin{cases} \| f \|^2_{-1}& \text{ if } \star=n.n., (A-ind) \\ \| f \|^2_{K, -1}& \text{ if } \star=n.n., (A-cor) \\ \| f \|^2_{-\gamma/2}& \text{ if } \star=l.r., (A-ind), \gamma=\{2,\alpha\}. \end{cases} \] \end{proposition} \begin{proof} To prove Proposition \ref{prop:var} we write: \begin{equation}\label{var_comp} \begin{split} n^{-2d} (a^{\star}_n)^2 \sum_{z,\,z'\in \T_n^d} f(z)f(z') \mathbb{E}[ \eta^{\star}(nz)\eta^{\star}(nz')] & = n^{-2d} (a^{\star}_n)^2 \sum_{z,\,z'\in \T_n^d} f(z)f(z') \sum_{w\in \Z^d_n} \frac{e^{2\pi i(z-z')\cdot w}}{|\lambda_w^{\star}|^2}. \end{split} \end{equation} Observe that \[ n^{-2d}\sum_{z,z'\in \T^d_n} f(z) f(z') \exp \left ( 2\pi i (z-z')w\right ) = | \widehat{f}(w) |^2. \] Under the assumption (A-ind) we can use the asymptotics of the eigenvalues \eqref{ev} which will be of order $a^{\star}_n$ to make the convergence work. We need to adjust our proofs to the case where \[ \sum_{w\in \Z^d\setminus \{o\}} \frac{1}{\|w\|^{4}} \, \, \, \text{ resp. } \sum_{w\in \Z^d\setminus \{o\}} \frac{1}{\|w\|^{2\gamma}} \] are converging or not. For $d<d_c^{\star}$ which was $d_c^{n.n.}=4$ when $\star=n.n.$ and $d_c^{l.r.}=2\gamma$ in the long-range case the sum converges and the argument is straightforward. In the other case we need a mollifying procedure. Let $\rho \in \mathcal{S}(\R^d)$ be a normalized function in the Schwartz space with support in $\T^d$. Let $\rho_k(x) =\frac{1}{k^p}\rho(x/k)$ then for all $m\in \N$ there exists $C>0$ such that \[ |\widehat{\rho}_k(w)| \leq \frac{c}{(1+\|w\|)^m}. \] It remains to prove that the mollified version converges to the desired variance \begin{equation}\label{eq-mollified-goes-to-norm} \lim_{\kappa \to 0^+} \lim_{ n \to \infty} n^{-2d} \sum_{z,z^\prime \in \mathbb{T}^d_n}f(z)f(z^\prime) \sum_{w \in \mathbb{Z}^d_n \backslash \{0\}} \widehat{\rho_\kappa}(w) \frac{\exp(2\pi i (z-z^\prime)\cdot w )}{\|w\|^{\star}}= \mathbb{E}(\langle \Xi^{\star},f \rangle^2), \end{equation} and the mollifier correction disappears, \begin{equation}\label{eq-error-in-mollification} \lim_{\kappa \to 0^+} \overline{\lim_{ n \to \infty}} \Big| n^{-2d} \!\!\! \sum_{z,z^\prime \in \mathbb{T}^d_n} \!\!\!\! f(z)f(z^\prime) \!\!\!\!\! \sum_{w \in \mathbb{Z}^d_n \backslash \{0\}} \!\!\!\!\! \Big(1-\widehat{\rho_\kappa}(w)\Big) \frac{\exp(2\pi i (z-z^\prime)\cdot w )}{\|w\|^{\star}} \Big| =0. \end{equation} \end{proof} Let us remark that in the correlated case (A-cor) we need only to worry about the second moment since the initial configuration is Gaussian. The variance will depend additionally on the Fourier multiplier $\widehat{K}$. The most involved part in the proofs is to control that the error introduced in \eqref{var_comp}, by replacing the eigenvalues $\lambda_w^{\star}$ by the leading order \eqref{ev}, is vanishing. In the long-range case this poses an extra difficulty since it requires a fine analysis to get to the right leading order and additionally to make the convergence work. Higher moments follow from the following combinatorial argument. Take a fixed $f \in \mathcal{F}$ and define a map $T_n: \mathbb{T}^d \longrightarrow \mathbb{R}$ by \begin{equation}\label{map_T} z \longmapsto \int_{B(z, \frac{1}{2n})} f(y) dy. \end{equation} For $m\geq 3$ denote by $\mathcal{P}(m)$ the set of all partitions of $\{1,2,...,m\}$ and $\Pi$ the elements of a partition $P\in \mathcal{P}(m)$. Then we write \[ \mathbb{E}(\langle \Xi_n^{\star},f\rangle^m) = \sum_{P \in \mathcal{P}(m) } \prod_{\Pi \in P} (a_n^{\star})^{|\Pi|} \mathbb{E}[\sigma^{|\Pi|}] \sum_{x \in \mathbb{Z}^d_n}\Big( \sum_{z \in \mathbb{T}^d_n} g^{(\alpha)}(x,nz)T_n(z) \Big)^{|\Pi|} \] and use again properties of the eigenvalues, independence of the $\sigma$'s to prove that it converges to the right moment of the corresponding Gaussian model, see Proposition \ref{prop-convergence-moments-bounded-case}. We stress that independence of the $\sigma$ variables is a key property here. In fact, if we want to consider correlated random variables, then the higher moment proof is much more involved.We expect that if the correlations are small then the proof should still work.\\ \noindent \textbf{Truncation method:} The proof will be completed once we show the truncation method. Fix an arbitrarily large (but finite) constant $R>0$ and set \begin{align} w_n^{<R}(x)&:=\frac{1}{2d}\sum_{y\in \Z_n^d}g^{\star}(x,\,y)\sigma(y){\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\{|\sigma(y)|< R\}},\\ w_n^{\geq R}(x)&:=\frac{1}{2d}\sum_{y\in \Z_n^d}g^{\star}(x,\,y)\sigma(y){\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\{|\sigma(y)|\geq R\}}. \end{align} Clearly $w_n(\cdot)=w_n^{<R}(\cdot)+w_n^{\geq R}(\cdot)$. To prove our result, we will use \begin{theorem}[Theorem~4.2 \citet{Bi68}] \label{thm:billy} Let $S$ be a metric space with metric $\rho$. Suppose that $(X_{n,\,u},\,X_n)$ are elements of $S\times S.$ If \[ \lim_{u\to+\infty} \limsup_{n\to+\infty} \mathbb{P}\left(\rho(X_{n,\,u},\,X_n)\geq \tau\right)=0 \] for all $\tau>0$, and $X_{n,\,u}\Rightarrow_{n}Z_u\Rightarrow_u X$, where $``\Rightarrow_x''$ indicates convergence in law as $x\to +\infty$, then $X_n\Rightarrow_n X$. \end{theorem} Following this Theorem, we need to show two steps: \begin{itemize} \item[(Step 1)] $\lim_{R\to+\infty} \limsup_{n\to+\infty} \mathbb{P}\left(\left\| a^{\star}_n\Xi^{\star}_{n}- a^{\star}_n\Xi^{\star}_{w_n^{<R}}\right\|_{ H_{-\epsilon}}\geq \tau\right)=0$ for all $\tau>0$. \item[(Step 2)] For a constant $v_R>0$, we have $a^{\star}_n \Xi^{\star}_{w_n^{<R}}\Rightarrow_n \sqrt{v_R}\, \Xi^{\star} \Rightarrow_{R}\Xi$ in the topology of $H_{-\epsilon}$. \end{itemize} As a consequence we will obtain that $a^{\star}_n\Xi^{\star}_{n}$ converges to $\Xi^{\star}$ in law in the topology of $H_{-\epsilon}.$ To prove (Step 1) notice that \[ \left\| a^{\star}_n\Xi_{n}- a^{\star}_n\Xi_{w_n^{<R}}\right\|_{H_{-\epsilon}}=\left\| a^{\star}_n \Xi_{w_n^{\geq R}}\right\|_{H_{-\epsilon}} \] by definition, for every realization of $(\sigma(x))_{x\in \Z_n^d}$. Since, for every $\tau>0$, \[ \mathbb{P}\left(\left\| a^{\star}_n \Xi_{w_n^{\geq R}}\right\|_{ H_{-\epsilon}}\geq\tau\right)\le\frac{\mathbb{E}\left[\left\| a^{\star}_n \Xi_{w_n^{\geq R}}\right\|_{ H_{-\epsilon}}^2\right]}{\tau^2} \] it will suffice to show that the numerator on the right-hand side goes to zero to show (Step 1). The second step follows from the next considerations. Set $v_R = Var(\sigma{\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\{|\sigma|<R\}})$ and \[ \sigma^R(x):=\sigma(x){\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{\{|\sigma(x)|<R\}}-m_R. \] Let us look at the following fields, $c>0$ some constant, \[ \Xi^{\star}_{n,\,R}(x):= c \cdot a^{\star}_n \sum_{z\in \T^d_n}\sum_{w\in\Z_n^d}g^{\star}(w,\,nz)\sigma^R(w){\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l_{B(z,\,\frac{1}{2n})}(x),\quad x\in \T^d. \] Since $\sum_{y\in \Z_n^d}g^{\star}(\cdot,\,y)$ is a constant function on $\Z_n^d$ it follows that \[ \left\langle \Xi^{\star}_{n,\,R},\,f \right\rangle=\left\langle \Xi^{\star}_{w_n^{<R}},\,f\right\rangle \] for all smooth functions $f$ with zero average. Therefore the field $\Xi^{\star}_{n,\,R}$ has the same law of $\Xi^{\star}_{w_n^{<R}}$. If we multiply and divide the former expression by $\sqrt{v_R}$ and take the limits we get the claim. \section{Alpha-Stable Distributions}\label{sec:stable} Which assumptions on the initial distribution or redistribution of the mass are responsible for the Gaussian nature of the limiting odometer interface? We have seen that the redistribution (nearest neighbours versus long-range) of the mass will influence the regularity of the limiting field but not the Gaussianity. In fact, the finite second moment assumption of the $\sigma$'s is responsible for the CLT-type behaviour. In \citet{CHRheavy} the authors proved that for $\sigma$'s which are in the domain of attraction of stable distributions the odometer scales to an $\alpha$-stable continuum field on $\T^d$. To the authors knowledge it is the first time that a continuum $\alpha$-stable field was constructed. \subsection{Stabilization on infinite graphs} In \citet{CHRheavy} the authors investigate the dichotomy between stabilization and explosion for infinite vertex transitive graphs $G$ when $(s(x))_{x\in V}$ is i.i.d. and has not necessarily a finite mean or variance. In Lemma 1 it is shown that for $\mathbb{E}(s(o))=\infty$ the sandpile will not stabilize almost surely while for $\mathbb{E}(s(o))=-\infty$ it does. Assuming finite mean of the initial configuration and infinite variance one has to be more careful. As in \citet{LMPU} one has to distinguish two cases, when $\sum_{y\in V}g(0,y)^{\alpha}=\infty$ and when it is finite. The first case is called the \textit{doubly transient case} and corresponds to e.g. $\Z^3,\Z^4$ and the latter the \textit{singly transient case} corresponding for example to $\Z^d$ for $d\geq 5$. In both cases it turns out that for $\mathbb{E}(s(o))=1$ and $\sigma$'s regularly varying, the probability of stabilization is 0 (compare Theorems 2 and 3 from \citet{CHRheavy}). The same result holds true in the finite variance case, proven already by \citet{LMPU} in Theorem 1.1. \subsection{Scaling limit of the odometer for heavy-tail inital distributions} Let us first state assumptions on the initial configuration $s$. We will again assume that $s$ satisfies \[ s(x) = 1 + \sigma(x) - \frac{1}{n^d}\sum_{z\in \Z^d_n}\sigma(z). \] A non-negative random variable $X$ is called {\em regularly varying} of index $\alpha\ge 0$, and we write $X\in RV_{-\alpha}$, if \[ \mathbb{P}(X>x) \sim x^{-\alpha} L(x) \quad \text{as }x \rightarrow +\infty \] where $L$ is a slowly varying function, i. e., \[ \lim_{x\rightarrow +\infty}\frac{L(tx)}{L(x)}=1\quad\text{for all }t>0. \] We recall the definition of random variables which fall in the domain of attraction of $\alpha$-stable distributions: \begin{definition}[Domain of normal attraction of stable variables] \label{firstassumption} Let $\alpha\in (0,\,2]$. Let $V$ be a countably infinite index set and $(W(x))_{x\in V}$ be i.i.d. symmetric random variables with common distribution function in the domain of normal attraction of an $\alpha$-stable distribution. This means that, for $V_1\subset V_2\subset\ldots$ such that $\cup_{k\geq 1}V_k=V$, we have the following limit: \begin{equation}\label{eq:doa:stable} \lim_{k\rightarrow +\infty}{|V_k|^{-\frac{1}{\alpha}}}\sum_{x\in V_k} W(x)\stackrel{d}{=} \rho_\alpha, \end{equation} where $\rho_\alpha$ has a symmetric $\alpha$-stable law which we denote as $S\alpha S(\mathfrak c)$, that is, $\mathbb{E}[ \exp(i \theta \rho_\alpha)]= \exp({-\mathfrak{c}^\alpha|\theta|^\alpha})$ for some $\mathfrak{c}\in \mathbb{R}$. \end{definition} The scaling limit result (Theorem 5) in \citet{CHRheavy} is proven in the setting $V_n:=\Z_n^d$ and $V= \Z^d$. If the scale parameter of the $\alpha$-stable law is $1$, we will write $\sigma(x)\stackrel{d}{=} S\alpha S(1)$. In this case, it is well known that $|\sigma(x)|$ has a regularly varying tail with index $-\alpha$, for $\alpha\in (0,2]$. The results we are going to prove can be extended to a more general set-up assuming further necessary and sufficient conditions for the $(\sigma(x))_{x\in V}$ to be in the domain of attraction of stable variables, see also \citet{ST}. Proposition \ref{prop:dislaw} is valid also for finite connected graphs $G=(V,E)$, compare Proposition 4 in \citet{CHRheavy}.\\ \noindent \textbf{Example:} For $\sigma$'s chosen from a Cauchy distribution or Pareto with parameter 1.5 (Figure \ref{Cauchy} resp. \ref{Pareto}) we see that the odometer interface has a completely different shape. Indeed, the limiting interface will not be Gaussian any more. \begin{figure}[h] \centering \centering \begin{minipage}{.5\textwidth} \includegraphics[scale=0.5]{Cauchy.jpg} \caption{Realization of the odometer interfaces for Cauchy initial $\sigma$'s} \label{Cauchy} \end{minipage}% \begin{minipage}{.5\textwidth} \centering \includegraphics[scale=0.5]{Pareto15.jpg} \caption{Realization of the odometer interfaces for Pareto $\sigma$'s with parameter 1.5} \label{Pareto} \end{minipage} \end{figure} In the following we define the continuum $\alpha$-stable field (compare Definition 2.1. in \citet{KumMan}). \begin{definition} A random distribution (field) $\Xi_{\alpha}$ on the dual space of smooth functions on $\T^d$, $(C^{\infty}(\T^d)/\sim)^*$, is called $\alpha$-stable if, given two independent copies $\Xi_{\alpha,\,1}$ and $\Xi_{\alpha,\,2}$ of $\Xi_\alpha$, then for any $a,\, b>0$ and $f \in (C^{\infty}(\T^d)/\sim)^*$ \[ \mathbb{E}[\exp(i \left\langle \Xi_{\alpha,\,1}, af\right\rangle )]\mathbb{E}[\exp(i \left\langle \Xi_{\alpha,\,2}, bf\right\rangle )]= \mathbb{E}\left[\exp\left(i \left\langle \Xi_{\alpha}, (a^\alpha+b^\alpha )^{\frac{1}{\alpha}} f \right\rangle \right) \right]. \] \end{definition} Alternatively we can represent the above characteristic functional of the field as \[ \mathbb{E}(\exp(i \langle \Xi_{\alpha},f\rangle)) = \exp \left ( -\|\Delta^{-1} f\|^{\alpha}_{L^{\alpha}(\T^d)} \right ). \] The following Theorem 5 from \citet{CHRheavy} states the finite-dimensional convergence of the odometer field. \begin{theorem}\label{thm:scaling_RV} Let $d\ge 1$, $(\sigma(x))_{x\in \Z_n^d}$ be i.i.d. and satisfy Definition \ref{firstassumption}, and furthermore let $(s(x))_{x\in \Z_n^d}$ satisfy \eqref{ini_s}. Consider $a_n=4\pi^2 n^{d-d/\alpha -2}$. There exists a random distribution $\Xi_\alpha$ on $(C^\infty(\T^d))^\ast$ such that: for all $m\in \N$ and $f_1, f_2,\ldots,\, f_m\in C^\infty(\T^d)$ with mean zero, the random variables $\left\langle a_n\Xi_n, f_j \right\rangle$ converge jointly in distribution to a random variable $\left\langle a_n \Xi_\alpha, f_j\right\rangle$. Moreover, the characteristic functional of $\Xi_\alpha$ is given by \begin{equation}\label{eq:limitfield} \mathbb{E}[\exp(i \left\langle \Xi_{\alpha}, f\right\rangle )]=\exp\left(-\int_{\T^d} \left| \sum_{z\in \Z^d\setminus \{0\}} \frac{\exp(-2\pi i z\cdot x)}{\|z\|^2} \widehat{f}(z)\right|^{\alpha} dx \right). \end{equation} \end{theorem} \begin{proof} Let us present a sketch of the proof and the main key steps. We will again rely on Fourier analysis since we need to determine a characteristic functional. Note that we are not working any more on AWS and Sobolev spaces but rather dual spaces of smooth functions on the torus. Therefore we do not have Rellich's theorem at hand and cannot prove tightness as in the previsous cases. However, since we are dealing now with random fields with possibly no moments, we cannot use any more a moment convergence argument. What will be helpful is to consider a simpler case. We will first consider $\sigma\sim S\alpha S(1)$ and then look at the general case. The proof is decomposed into 5 steps and they are related to each other by: \begin{center} Step~\ref{step:5}$\Rightarrow$ Step~\ref{step:4}$\Rightarrow$ Step~\ref{step:3}$\Rightarrow$ Step~\ref{step:2}$\Rightarrow$ Step~\ref{step:1}. \end{center} Similar as before, due to mean zero test functions, we can write \begin{equation}\label{Xi_alpha} \langle \Xi_{n}, f \rangle = \sum_{x\in \Z^d_n} \left ( \sum_{z\in \T^d_n} g(x,nz)T_n(z)\right )\sigma(x)=:\sum_{x\in \Z^d_n} k_n(x)\sigma(x) \end{equation} where $T_n$ was defined in \eqref{map_T}. Then we get \[ \mathbb{E}(\exp(i\langle a_n \Xi_n,f \rangle)) =\exp \left(-\sum_{x\in \Z^d_n} |a_n k_n(x)|^{\alpha} \right). \] Defining a reminder term \[ R_n(w):= \int_{B\left(w,\,\frac{1}{2n}\right)} (f(u)-f(w)) d u \] which will be small in the limit we can rewrite $k_n$: \begin{align} k_n(x)&=-\frac{1}{n^d}\sum_{w\in\T_n^d}\left( n^{-d} f(w)+R_n(w)\right) \sum_{z\in \Z_n^d\setminus\{0\}} \frac{ e^{2\pi i z(x-nw)} }{\lambda_z}\nonumber\\ &=-\frac{1}{n^{2d}}\sum_{w\in\T_n^d} f(w)\!\sum_{z\in \Z_n^d\setminus\{0\}} \frac{ e^{2\pi i z(x-nw)}}{\lambda_z}-\frac{1}{n^d}\sum_{w\in\T_n^d}R_n(w)\!\sum_{z\in \Z_n^d\setminus\{0\}} \frac{ e^{2\pi i z(x-nw) }}{\lambda_z}\nonumber\\ &:= l_n(x)+ C_n(x)\label{eq:def_lC}. \end{align} First one shows that the convergence of $\exp\left(-\sum_{x\in \Z_n^d} |a_n k_n(x)|^\alpha\right)$ can be given in terms of the same quantity where $k_n(\cdot)$ is replaced by $l_n(\cdot)$: \begin{step}\label{step:1} \[ \lim_{n\rightarrow +\infty}\left|\exp\left(-\sum_{x\in \Z_n^d} |a_n k_n(x)|^\alpha\right)- \exp\left(-\sum_{x\in \Z_n^d} |a_n l_n(x)|^\alpha\right)\right|=0. \] \end{step} In the next steps one proves that $l_n$ yields the correct characteristic function. In Step~\ref{step:2} we are introducing a mollifier which will help to extend sums from $\Z_n^d$ to the whole lattice. \begin{step}\label{step:2} {Let $\phi\in \mathcal S(\R^d)$, the Schwartz space, with $\int_{\R^d}\phi(x) d x=1$. Let $\epsilon>0$ and let $\phi_\epsilon(x):= \epsilon^{-d}\phi\left({x}{\epsilon}^{-1}\right)$ for $\epsilon>0$.} Then \begin{align*} \lim_{\epsilon\downarrow 0}\lim_{{n\rightarrow +\infty}}\left|\sum_{x\in \Z_n^d} |a_n l_n(x)|^{\alpha}- \frac{a_n^\alpha}{n^{d\alpha}} \sum_{x\in \T_n^d}\left| \sum_{z\in \Z_n^d\setminus \{0\}} \frac{\widehat \phi_\epsilon(z)\exp(-2\pi \imath z\cdot x) }{\lambda_z} \widehat{f_n}(z)\right|^{\alpha}\right|=0 \end{align*} where $\widehat{f_n}(z)= n^{-d}\sum_{w\in \T_n^d} f(w)\exp(2\pi i w\cdot z)$. \end{step} The goal of the third step is to approximate each eigenvalue $\lambda_z$ of the Laplacian with the norm of the point $z$, namely \begin{step}\label{step:3} Uniformly in $\epsilon>0$ \begin{align*} \lim_{n\to+\infty}\frac{a_n^\alpha}{n^{d\alpha}} &\left|\sum_{x\in \T_n^d}\left[\left| \sum_{z\in \Z_n^d\setminus \{0\}} \frac{\widehat \phi_\epsilon(z)\e^{-2\pi \imath z\cdot x} }{\lambda_z} \widehat{f_n}(z)\right|^{\alpha} \right.\right.\\ &\left.\left.-\frac{ n^{2\alpha}}{4^\alpha {\pi^{2\alpha}} } \left| \sum_{z\in \Z_n^d\setminus \{0\}} \frac{\widehat \phi_\epsilon(z)\e^{-2\pi \imath z\cdot x} }{\|z\|^2} \widehat{f_n}(z)\right|^{\alpha}\right]\right|=0 \end{align*} \end{step} In the next step we extend the sums in Step \ref{step:3} over $\Z^d$ using the decay of the mollifier. \begin{step}\label{step:4} Uniformly in $\epsilon>0$ \begin{align*} \lim_{n\rightarrow +\infty}\frac{a_n^\alpha n^{2\alpha}}{n^{d\alpha}{4^\alpha\pi^{2\alpha}}}&\left| \sum_{x\in \T_n^d}\left| \sum_{z\in \Z_n^d\setminus \{0\}} \frac{\widehat \phi_\epsilon(z)\exp(-2\pi \imath z\cdot x) }{\|z\|^2} \widehat{f_n}(z)\right|^{\alpha}\right.\\ & - \left. \sum_{x\in \T_n^d}\left| \sum_{z\in \Z^d\setminus \{0\}} \frac{\widehat \phi_\epsilon(z)\exp(-2\pi \imath z\cdot x) }{\|z\|^2} \widehat{f_n}(z)\right|^{\alpha}\right|=0. \end{align*} \end{step} At last, we can finally show the convergence of the sum to the required integral. \begin{step}\label{step:5} \eq{}\label{eq:ice}\lim_{\epsilon\downarrow 0}\lim_{{n\to+\infty}}\frac{1}{n^d}\sum_{x\in \T_n^d}\left| \sum_{z\in \Z^d\setminus \{0\}} \frac{\widehat \phi_\epsilon(z)\e^{-2\pi \imath z\cdot x} }{\|z\|^2} \widehat{f_n}(z)\right|^{\alpha} = \int_{\T^d} \left| \sum_{z\in \Z^d\setminus \{0\}} \frac{\e^{-2\pi \imath z\cdot x}}{\|z\|^2} \widehat{f}(z)\right|^{\alpha} \mathrm{d} x.\eeq{} \end{step} Those steps are quite involved to prove since we cannot rely on moment assumptions and need to perform fine analysis. It remains to prove Theorem \ref{thm:scaling_RV} for regularly varying initial distributions. This will be implied by Proposition 12 of \citet{CHRheavy} stating that for all $f\in \mathcal{F}$, $\epsilon>0$ \[ \lim_{n\rightarrow \infty} \mathbb{P} \left ( | \langle \Xi_n,f \rangle - \langle \Xi'_n,f \rangle| > \epsilon \right )=0 \] where \[ \langle \Xi'_{n}, f \rangle = \sum_{x\in \Z^d_n} \left ( \sum_{z\in \T^d_n} g(x,nz)T_n(z)\right )\sigma(x) \] and $\sigma \sim \rho_{\alpha}$ from Definition \ref{firstassumption}. \end{proof} \section{Internal Diffusion Limited Aggregation}\label{sec:iDLA} In this section we want to discuss the connection between odometer functions of divisible sandpile models and internal diffusion limited aggregation models (iDLA). It was introduced for the first time by \citet{DiaFul}. The iDLA model describes the growth of a random aggregate of particles from the inside out. Start initially with $n$ particles at the origin $o\in \Z^d$ and let each particle perform a random walk until it reaches an unoccupied site. Consider the random set of occupied sites $D(n)$. Bramson et al. in \citet{Brams} were the first ones to identify that $D(n^d)$ is very close to the Euclidean ball of radius $n$, which was later sharpened by \citet{Gaudi1,Gaudi2}. Levine and Peres in \citet{LevPerDLA} identified that the scaling limit of three growth models, namely the iDLA, rotor-router and divisible sandpile models on $\Z^d$ to be the same and are related to solutions to certain PDE free boundary problems in $\R^d$. To identify the limiting shape, the authors in \citet{LevPerDLA} consider the odometer function of the divisible sandpile. Recall that it satisfies the following relation (see Proposition \ref{prop:stab}): \[ \Delta u(x) = s_{\infty}(x) - s(x) \] where $s_{\infty}$ is the final divisible sandpile configuration and $s$ the initial one. Due to the constraint, $\sum_{x\in \Z_n^d} s(z) =n^d$, we obtain that $s_{\infty}\equiv 1$. In general for $\sum_{x\in \Z^d_n} s(x)< n^d$ the limiting configuration $s_{\infty}$ is a random configuration depending on $s$. Remember that an alternative way to find the odometer function is to solve the obstacle problem, see Lemma \ref{odo_lev}, where one has to find a proper obstacle $\gamma$ satisfying the Laplacian equation $(-\Delta) \gamma = s-1$ and a solution to the obstacle problem $v$ satisfying $v(x) = \inf \{ f(x)|f\geq \gamma; (-\Delta)f \leq 0 \}$. Then the odometer is equal to \begin{equation}\label{obs_u} u=v-\gamma. \end{equation} In fact, this was the approach chosen in \citet{frometa:jara} to determine the scaling limit of the odometer of a truncated long-range divisible sandpile in $\Z^2$. Let us point out again a crucial fact that the scaling limit for the divisible sandpile, obtained in the previous sections, corresponds to the scaling limit of the \textit{obstacle} in this context and not of \eqref{obs_u}. This is due to the fact that the mean 0 test functions cancel out the contribution of $v$ which is the solution of the obstacle problem. For finding the limiting shape of the odometer, one has first to find an appropriate obstacle $\gamma$. A good candidate will turn out to be \begin{equation}\label{gamma} \gamma(x) = -|x|^2 - \sum_{y\in \Z^d} G(x,y)s(y) \end{equation} where $G(x,y)$ is the Green's function in $\Z^d$ for $d\geq 3$ and the recurrent potential kernel in $\Z^2$. Chosen $\gamma$ in this way we get that $\Delta (u+\gamma) \leq 0$ hence $u+\gamma$ is a superharmonic function on $\Z^d$. For any superharmonic function $f\geq \gamma$ \[ \Delta (f-\gamma - u) \leq 0 \] on the domain $\overline{D}=\{ x\in \Z^d| s_{\infty}(x)=1 \}$ and non-negative outside $D$ hence non-negative everywhere. Hence we get that $u$ is equal to \eqref{obs_u}. The description in the continuum $\R^d$ is then immediate. Given an initial mass $s\in \R^d$ and obstacle \begin{equation}\label{gamma_r} \gamma(x) = -|x|^2 - \int_{\R^d} G(x,y)s(y)dy \end{equation} where \[ G(x,y) = \begin{cases} -\log(\|x-y \|) & \text{ if } d =2\\ \|x-y\|^{2-d} & \text{ if } d \geq 3 \end{cases} \] the odometer $u=v-\gamma$ where $v(x)=\{ f(x)| f \text{ is continuous, superharmonic }, f\geq \gamma\}$. \begin{definition} The non-coincidence set for the obstacle problem with obstacle $\gamma$ is the domain of occupied sites given by \[ D=\{ x\in \R^d| v(x) > \gamma(x)\}. \] \end{definition} For a lattice spacing $\delta_n \rightarrow 0$ as $n \rightarrow \infty$ and domains $A_n \subset \delta_n \Z^d, D\subset \R^d$ write $A_n\rightarrow D$ if for any $\epsilon >0$, $D_{\epsilon} \cap \delta_n \Z^d \subset A_n \subset D^{\epsilon}$ for $n$ large enough and $D^{\epsilon}=\{x\in \R^d|B(x,\epsilon) \not\subset D^c\}$ resp. $D_{\epsilon}= \{ x\in D| B(x,\epsilon) \subset D\}$. Let us state Theorem 1.2 from \citet{LevPerDLA} which identifies the non-coincidence set for the obstacle problem with the occupied sites of the iDLA, rotor-router and divisible sandpile model in the scaling limit. \begin{theorem}\label{theo_shape} Let $B\subset \R^d$, $d\geq 2$ be a bounded open set. Put $s:\R^d \rightarrow \N\cup \{0\}$ bounded, continuous almost everywhere satisfying $\{\sigma \geq 1\} = \overline{B}$. Consider the initial configuration $s_{(n)} : \delta_n \Z^d \rightarrow \N\cup \{0\}$ with density \[ s_{(n)}(x) = \frac{1}{\delta_n^d} \int_{[x-\frac{\delta_n}{2}, x+\frac{\delta_n}{2}]^d} s(y) dy. \] Denote by $D^{\star}_n$ the domain of occupied sites for $\star\in \{ \text{div. sand.}, \text{iDLA}, \text{ ro.-rou.}\}$ respectively. Then \[ D^{\star}_n\rightarrow D \cup B \] for all $\star\in \{ \text{div. sand.}, \text{iDLA}, \text{ ro.-rou.}\}$ as $n\rightarrow \infty$ and $\delta_n \log(n)\rightarrow 0$ for $\star=div.sand.$ The convergence holds in the sense described above. \end{theorem} As a consequence all three models have rotational invariant scaling limits hence the limiting shape for the iDLA and divisible sandpile when starting with a pile of $n$ particles at the origin is a ball. \section{Discussion and open problems}\label{sec:dis} In this last section we would like to give some perspectives and open questions.\\ \noindent \textbf{Scaling limit of the odometer in the super- and subcritical case:} Recall that the odometer $u$ is a solution of \[ \Delta^{\star} u(x) = s_{\infty}(x) - s(x). \] The assumption on the initial configuration $\sum_{x\in \Z^d_n} s(x)=n^d$ ensured that $s_{\infty}\equiv 1$ and therefore \[ u = (\Delta^{\star})^{-1} (1-s) \] depends only on the initial distribution of $s$. In the subcritical case $\sum_{x\in \Z^d_n} s(x)<n^d$ what is the distribution of $s_{\infty}$? How are $s_{\infty}$ and $s$ correlated? What is the scaling limit of $u$? For the supercritical case we know that the sandpile will not stabilize, but can we still control it and find a proper rescaling to make the odometer converge to a particular field? \\ \noindent \textbf{Scaling limit of maxima:} Concerning the scaling limit of the maxima note that the asymptotics of the mean odometer or expected maximum in Theorem \ref{max_odo_nn} agree with the asymptotics of the expected maximum of a bi-Laplacian model in $\Z^d$ when defining the model on a box $\Lambda_n\subset \Z^d$ of size $n$. It was proven in Theorem 1 of \citet{CCH2015} that the scaling limit of extrema of bi-Laplacian models are Gumbel distributed for $d\geq 5$ and in \citet{Biltu} that the maximum coincides with the one of the GFF for $d=2,3$, see Corollary 2.2. The case $d=4$ was very recently solved by \citet{mem4d} where the author showed in Theorem 1.1. that the scaling limit is a randomly shifted Gumbel distribution. We conjecture that for $d>2\alpha$ the scaling limit is a Fr\'echet type distribution and at the critical point $d=2\alpha$ it is a randomly shifted Fr\'echet.\\ \noindent \textbf{Limiting shapes for long-range divisible sandpiles:} In \citet{frometa:jara} the authors consider a truncated long-range divisible sandpile models in $\Z^2$. Start with an initial configuration on $\Z^2$ with finite support and consider a divisible sandpile model. Topple when the mass $s(x)\geq 1$ and redistribute not only to nearest neighbours but also to further away neighbours according to a long-range random walk with $\alpha \in (1,2)$ truncated to stay within a fixed range $M$. The authors obtain the scaling limit of the odometer function by solving a corresponding obstacle problem. Can one prove a similar result as Theorem \ref{theo_shape} for the (truncated) long-range divisible sandpile? What is the limiting shape? We conjecture that there exist a range of parameters $\alpha \in (\alpha_1, \alpha_2)$ where the limiting shape is not a ball.\\ \noindent \textbf{Divisible sandpiles on different graphs:} What are scaling limits of divisible sandpile models on different graphs? We know from \citet{LMPU} that for any finite graph as long as $\sum_{x\in \Z^d_n} s(x)\leq n^d$ then the divisible sandpile configuration will stabilize. In the infinite graph case for an initial i.i.d. configuration $s$, the divisible sandpile will stabilize if $\mathbb{E}(s(o))<1$. What about $\Z^d$ or other lattices or supercritical percolation (see also \citet{IDLA_perco})? We conjecture that the results will still hold true but require a much finer analysis and control over the convergence of Green's functions. \bibliographystyle{plainnat}

https://arxiv.org/abs/1612.07177

Network coding and spherical buildings

We develop a network coding technique based on flags of subspaces and a corresponding network channel model. To define error correcting codes we introduce a new distance on the flag variety, the Grassmann distance on flags and compare it to the commonly used gallery distance for full flags.

\section{Introduction} \label{intro} To transmit information over network channels, the currently used method consists of routing, i.e. simply forwarding the packets through each node. Network coding assumes that the packets that are sent through the network are elements of a vector space and the nodes in the network forward linear combinations of the received packets with randomly selected coefficients. It is well known (see \cite{Ahlswede}) that network coding allows multicast capacity achievability, which is not possible with packet forwarding only. When errors exist, the design of error correction codes for network-coded communication substantially differs from bit-level coding design. This is not only because of the richer algebraic structure but also because networking protocols exist between the physical transmission channel and the packet-level communication. Such protocols motivate packet-level error models. In the seminal work \cite{KK} a novel solution to the problem of error and erasure correction is tackled over a linear network-coded packet flow. The randomly selected coding coefficients are assumed unknown (i.e. incoherent transmission) and a novel framework of subspace coding is proposed. If the packets are elements of the vector space $V$ then the solution by K\"otter and Kschischang stems from the observation that information that is preserved after being linearly transformed by the network is the subspace generated by the input vectors, which is an element of the Grassmannian space ${\mathcal G}(V) $, the set of all subspaces of $V$. In this work, we extend the applicability of such framework under the assumption that in-network nodes can keep track of packet sequence numbering, as it is the case on the Internet. Under such assumption, we propose encoding information over flags, i.e. chains of subspaces of $V$, that are network coded by the in-network nodes with the stabilizers of the flags as they traverse the network \cite{A-ITW}. The set of all flags in $V$ forms a simplicial complex, known as the spherical building of the general linear group of $V$. We modify the well known geometry of this spherical building to develop a minimum distance decoding scheme in the new geometry. As the geometry of spherical buildings is governed by the symmetric group $\Sym _n$, Section 2 is mostly devoted to summarize the relevant facts about symmetric groups. In Section 3 we introduce the basics of the flag variety of $V$ and the associated spherical building. The major goal of the next section is to define a new distance on flags, the Grassmann distance (see Definition \ref{Grassdist}), which is more appropriate to measure errors and erasures in the transmission of flags through the network. The Grassmann distance seems to be much easier to compute than the commonly used gallery distance. In the last section, Section 5, we present a model for network transmission including errors and erasures which allows for the derivation of conditions for code constructions based on the Grassmann distance on flags. To set up a benchmark for comparing new flag codes to the classical situation of subspace codes we transfer and generalize the rank metric codes from \cite{SKK} to our situation. Other examples for flag codes of smaller minimum distance are given which allow for easy decoding. \section{Symmetric groups} The symmetric group $\Sym _n$ is the group of all bijective mappings from $\{ 1,\ldots , n \}$ to itself. As this group will govern the geometry of the flag variety we collect some relevant facts about symmetric groups in this section. \subsection{The length and the depth of a permutation} \begin{definition} \label{lengthdef} (see for instance \cite[Section 1.6]{Humphreys}) \\ The {\bf length} of a permutation $\pi \in \Sym _n$ is $$ \ell (\pi ) := \sum _{i=1}^{n} |\{ k \in \{ 1,\ldots , i\} \mid \pi(k) > \pi(i) \} | . $$ \end{definition} Then the identity is the unique element of $\Sym _n$ having length 0 and the elements $\pi \in \Sym _n$ of length 1 are exactly the transpositions $t_i= (i,i+1) (1\leq i \leq n-1) $ interchanging $i$ and $i+1$ and leaving the other points fixed. It is well known that any element of $\Sym _n$ is a product of these $t_i$. The length of $\pi \in \Sym _n$ is the number of factors in such a minimal expression of $\pi $ as a product of the $t_i$ (\cite[Section 1.7]{Humphreys}). As all the $t_i$ have order 2, this shows that $\ell (\pi ) = \ell (\pi ^{-1} )$. \begin{remark}\label{longest} There is a unique {\bf longest element} $w_0 \in \Sym _n$ with maximal $\ell (w_0) $. This element is $w_0 = (1,n)(2,n-1)(3,n-2) \cdots $ and has length $$ \frac{n(n-1)}{2} = {{n}\choose{2}} = | \{ (i,j) \in \{1,\ldots , n\} ^2 \mid i< j \} | $$ \end{remark} The most commonly used distance on the flag variety is the gallery distance (see Definition \ref{galery}). This distance is defined using the length of a permutation. For our purposes it seems to be more appropriate to work with the Grassmann distance on flags defined in Definition \ref{Grassdist} below. Here we replace the length function by a slightly different function called the depth function (see \cite[Theorem 1.1]{Petersen}). \begin{definition}\label{depth} For $\pi \in \Sym _n$ we define $$ \depth (\pi ) := \sum _{i=1}^{n-1} |\{ k\in \{ 1,\ldots , i \} \mid \pi (k) > i \} | .$$ Then $\depth : \Sym _n \to {\mathbb{N}}_0$ is called the {\bf depth function}. \end{definition} It is easy to see that $\depth (\pi ) = 0$ if and only if $\pi = \id $. Also $\depth (\pi ) = 1$ if and only if $\ell (\pi ) = 1$ if and only if $\pi = t_i = (i,i+1)$ for some $1\leq i \leq n-1$. More generally we get \begin{theorem}\label{lesym} (see Observation 2.2 in \cite{Petersen}) For any permutation $\pi \in \Sym _n $ we have $$\frac{\ell(\pi) + \ell_{tr}(\pi)}{2} \leq \depth(\pi) \leq \ell(\pi)$$ where $\ell_{tr}(\pi)$ is the smallest number of transpositions needed to write $\pi$. \end{theorem} \begin{remark} \label{scoopdist} \begin{itemize} \item[(a)] It is easy to see that $$\depth(\pi ) = \sum _{k=1,k<\pi(k)}^n \pi (k) - k .$$ \item[(b)] For the longest element $w_0$ from Remark \ref{longest} we compute $$\depth(w_0) = \sum _{k=1}^{\lfloor \frac{n}{2} \rfloor} (n-2k+1) = \left\{ \begin{array}{ll} (n/2)^2 & n \mbox{ even } \\ (n-1)(n+1)/4 & n \mbox{ odd } \end{array} \right. $$ (see Sequence A002620 in \cite{OEIS}) \item[(c)] The function $$s(\pi) := \depth(\pi ) + \depth(\pi^{-1}) = \sum _{k=1}^n |\pi (k) - k | $$ is known as the sum of distances function. As we will see in Corollary \ref{esym} $\depth(\pi ) = \depth(\pi ^{-1}) $ so $s(\pi ) = 2 \depth(\pi )$. \item[(d)] The number of permutations $\pi \in \Sym _n$ such that $\depth (\pi ) = k$ is denoted by $T(n,k) $ in the Sequence A062869 in \cite{OEIS}. Sequence A062870 in \cite{OEIS} gives $$T(n,k_0) = \left\{ \begin{array}{ll} (\frac{n}{2} !)^2 & n \mbox{ even } \\ (\frac{n(n-1)}{2} !)^2 & n \mbox{ odd } \end{array} \right. $$ where $k_0 = \depth(w_0) = \max \{ \depth(\pi ) \mid \pi \in \Sym _n \} .$ \end{itemize} \end{remark} \subsection{Young subgroups} \begin{defn} Let $T' := (k_1,\ldots, k_{m+1}) $ be a sequence of $m+1$ natural numbers $k_i \geq 1$ with $\sum _{i=1}^{m+1} k_i = n$. The {\bf Young subgroup} ${\mathcal Y}_{T'} \cong \Sym _{k_1} \times \Sym _{k_2} \times \ldots \times \Sym _{k_{m+1}}$ is the stabilizer in $\Sym _n$ of the sequence $$( \{ 1,\ldots , k_1 \} , \{ k_1+1,\ldots , k_1+k_2\} , \ldots , \{ k_1+\ldots + k_m +1,\ldots , n \} ) .$$ \end{defn} Clearly $|{\mathcal Y}_{T'} | = \prod _{i=1}^{m+1} k_i! $. For $m=0$ we get ${\mathcal Y}_{T'} = {\mathcal Y}_{(n)} = \Sym _n $. Also if $k_i=1$ for all $i$ then ${\mathcal Y}_{T'} = {\mathcal Y}_{(1,\ldots , 1 ) } = \{ \id \}$. It is well known (see for instance \cite{Jones}) that any double coset ${\mathcal Y}_{T'} \pi {\mathcal Y}_{T'}$ contains a unique element of minimal length. So these double cosets have canonical representatives which we collect in the set $\Sigma _{T'}$: \begin{definition}\label{doublecosetYoung} Let $\Sigma _{T'} $ denote the set of representatives of minimal length such that $$\Sym _n = \stackrel{.}{\cup} _{\pi \in \Sigma _{T'} } {\mathcal Y}_{T'} \pi {\mathcal Y}_{T'} .$$ \end{definition} \section{Spherical buildings} This section provides a constructive approach to the relevant facts about the spherical building of the general linear group of a finite dimensional vector space. For most of the proofs and more details we refer to the textbooks \cite{Abramenko}, \cite{Humphreys}, and \cite{Taylor}. \subsection{The flag variety}\label{flag} Let \emph{$K$} be a field and $V$ an $n$-dimensional vector space over $K$. The general linear group of $V$, $\GL(V)$, is the group of all linear automorphisms of $V$ (invertible linear maps from $V$ to itself). A {\bf flag} is a set of subspaces $\Lambda := \{ W_i \mid 1\leq i \leq m \}$ of $V $ with $$\{ 0 \} <W_1< \ldots < W_m < V .$$ The {\bf type} of $\Lambda $ is the set of dimensions $$\Type (\Lambda ):= \{ \dim(W_i) \mid W_i \in \Lambda \} \subseteq \{ 1,\ldots , n-1 \} .$$ Let $${\mathcal F} (V) := \{ \Lambda \mid \Lambda \mbox{ is a flag in } V \} $$ denote the set of all flags in $V$ and for $T\subset \{ 1,\ldots, n-1\}$ let $${\mathcal F} _T(V) := \{ \Lambda \in {\mathcal F} (V) \mid \Type (\Lambda ) = T \} $$ be the set of all flags in $V$ of type $T$. Note that the intersection of two flags is again a flag. The empty set is the unique minimal flag, its type is $\emptyset $. The second minimal flags $\{ W_1 \}$ are the proper subspaces $W_1$ of $V$. So the Grassmannian of all $k $-dimensional subspaces $${\mathcal G} _{k } (V) =\{ 0 < W_1 < V \mid \dim (W_1 ) = k \} $$ is in bijection with the set of flags ${\mathcal F} _{\{ k \} } (V)$ of type $\{ k \}$. A flag is called {\bf full}, if its type is $\{1,\ldots , n-1\}$. The set of full flags in $V$ is denoted by ${\mathcal F} _f(V)$. To construct a set of canonical representatives of the orbits of $\GL(V)$ on ${\mathcal F} (V)$ we choose and fix once and for all a full flag $$\Delta _0 = \{ V_1 , \ldots , V_{n-1} \} \in {\mathcal F} _f(V) $$ such that $\dim (V_i) = i$ and call $$B:= \{ g\in \GL(V) \mid V_i g = V_i \mbox{ for all } 1\leq i \leq n-1 \} = \Stab_{\GL(V)} (\Delta _0) $$ the standard {\bf Borel subgroup}. For $T =\{ d_1,\ldots , d_m \} \subseteq \{ 1,\ldots , n-1 \} $ define $$\Delta _T := \{ V_i \mid i \in T \} \in {\mathcal F} _T(V) \mbox{ and } P_T := \Stab _{\GL(V) } (\Delta _T ) .$$ The groups $P_T$ are called the standard {\bf parabolic subgroups} of $\GL(V)$. \begin{remark} If $T_1\subseteq T_2$, then $P_{T_2} \subseteq P_{T_1} $. We have $P_{\emptyset} = \GL(V) $ and $P_{\{ 1,\ldots ,n-1\} } = B$. \end{remark} We summarize the situation by listing some important points: \begin{fact}\label{orbits} \begin{itemize} \item[(a)] The group $\GL (V) $ acts on the set ${\mathcal F} (V)$. \item[(b)] The $\GL(V) $-orbits are separated by the type, so the partition $${\mathcal F} (V) = \bigcup _{T\subseteq \{ 1,\ldots , n-1 \} } {\mathcal F} _T(V) $$ is a partition of ${\mathcal F} (V) $ into $\GL(V)$-orbits. In particular $${\mathcal F}_T(V) = \{ \Delta _T g\mid g\in \GL(V) \}. $$ \item[(c)] For a given type $T\subseteq \{ 1,\ldots , n-1 \}$, the map $$ {\mathcal F}_T(V) \to P_T \backslash \GL(V), \Delta _T g \mapsto P_Tg $$ is a bijection between the set of all flags of type $T$ and the set $$P_T \backslash \GL(V) = \{ P_T g \mid g\in \GL(V) \} $$ of right cosets of $P_T$ in $\GL(V) $. \end{itemize} \end{fact} To define a geometry on the flag variety we want to study $\GL(V)$-invariant distance functions on ${\mathcal F}_T(V) $. \begin{rem}\label{distconst} Let $T$ be some type and $P_T = \Stab _{\GL(V)} (\Delta _T)$ the standard parabolic subgroup of $\GL(V)$. If $d$ is some $\GL(V)$-invariant function on ${{\mathcal F} }_T(V)\times {{\mathcal F} }_T(V)$, (so $d(\Lambda g, \Lambda ' g) = d(\Lambda,\Lambda ') $ for all $g\in \GL(V)$ $\Lambda , \Lambda '\in {{\mathcal F} }_T(V)$), then $d(\Delta_Tg,\Delta_Th) = \overline{d}(hg^{-1}) $ where $$\overline{d} (g) =d(\Delta _T ,\Delta _T g) \mbox{ for all } g\in \GL(V).$$ Moreover $\overline{d}$ is constant on the double coset $P_T gP_T$. \end{rem} \noindent\underline{Proof.}\ As $d$ is $\GL(V)$-invariant we see that $$d(\Delta _T g,\Delta _T h) = d(\Delta _T , \Delta _T hg^{-1} ) = \overline{d} (hg^{-1} ).$$ To see the second assertion let $b_1,b_2\in P_T$, $g\in \GL(V)$. Then $$\overline{d} (b_1gb_2) = d(\Delta _T , \Delta _T (b_1 g b_2) ) = d(\Delta _T b_2^{-1} , \Delta _T b_1 g) = d(\Delta _T, \Delta _T g ) = \overline{d} (g) .$$ \phantom{zzz}\hfill{$\square $}\smallskip As different double cosets are disjoint, we obtain a partition $$\GL(V) = \stackrel{.}{\cup} P_T g P_T $$ of the group $\GL(V)$ into a disjoint union of double cosets. The number of these double cosets does not depend on the field $K$. For $T= \{1,\ldots , n-1\}$ this number is always $n!$ and there is a canonical bijection between these double cosets and the group $\Sym _n$ of permutations of $\{1,\ldots, n\}$ where $n=\dim(V)$. Here we embed $\Sym _n$ as the set of permutation matrices into $\GL(V)$. This is captured by the Gau\ss-Bruhat decomposition. For any type $T$, the $P_T$ double cosets in $\GL (V)$ are in bijection with the double cosets of the Young subgroup ${\mathcal Y}_{T'} $ in the symmetric group $\Sym _n$. \begin{theorem}\label{GBgen} (see \cite{Humphreys}, \cite{Jones}) $$\GL(V) = \stackrel{.}{\cup} _{\pi \in \Sym_n} B \pi B .$$ More generally for a given type $T = \{ d_1,\ldots , d_m \} \subseteq \{ 1,\ldots , n-1 \} $ with $0<d_1<\ldots < d_m < n $ we define $T':= (k_1,\ldots , k_{m}, k_{m+1} )$ with $$k_1:= d_1, k_i = d_i - d_{i-1} \mbox{ for } 2\leq i \leq m \mbox{ and } k_{m+1} := n-d_m .$$ Then $$ P_T = \stackrel{.}{\cup} _{\pi \in {\mathcal Y} _{T'}} B\pi B $$ for the Young subgroup ${\mathcal Y}_{T'}$ and $$\GL (V) = \stackrel{.}{\cup} _{\pi \in \Sigma _{T'}} P_T \pi P_T $$ where $\Sigma _{T'} $ is defined in Definition \ref{doublecosetYoung} \end{theorem} The Gau\ss-Bruhat decomposition has a very nice property, as described in \cite[Theorem 5.10]{Taylor}: For each $\pi \in \Sym _n$ there is a subgroup $U_{\pi} \leq B$ such that any element in $B\pi B$ has a unique expression as $b\pi u$ with $b\in B$ and $u\in U_{\pi }$. If $K$ is a finite field, then the order of $U_{\pi }$ is $|K|^{\ell(\pi )} $ where $\ell $ is the length function on $\Sym _n$ (see Definition \ref{lengthdef}). \subsection{Buildings and apartments} To get a more precise model of the geometry of all flags, the so called spherical building of the group $\GL(V)$, we fix a basis $E:=\{ e_1,\ldots , e_n\} $ of $V$ and put $$\Delta _0 = \{ V_1,\ldots , V_{n-1} \} \mbox{ with } V_i = \langle e_1,e_2,\ldots , e_{i} \rangle .$$ We identify $\GL(V) $ with the group of invertible $n\times n$-matrices $\GL_n(K)$ using coordinate rows with respect to this basis. Then $B = \Stab_{\GL(V)} (\Delta _0)$ is identified with the group of all lower triangular matrices in $\GL_n(K)$ and the parabolic subgroup $P_T$ with all lower block triangular matrices $$\left( \begin{array}{ccccc} A_{11} & 0 & 0 & 0 & 0 \\ A_{21} & A_{22} & 0 & \ldots & \vdots \\ \vdots & \vdots & \ddots & \vdots & \vdots \\ A_{m1} & \ldots & \ldots & A_{mm} & 0 \\ A_{m+1,1} & \ldots & \ldots & A_{m+1,m} & A_{m+1,m+1} \end{array} \right) $$ where $ A_{ij} \in K^{k_i \times k_j } \mbox{ and } A_{ii} \in \GL_{k_i} (K) \mbox{ for } 1\leq j\leq i \leq m+1 $ if $T' = (k_1,\ldots , k_{m+1} ) $. For a permutation $\pi \in \Sym_n $ we denote by $$\Delta _{\pi } := \{ \langle e_{\pi(1)} \rangle , \langle e_{\pi(1)},e_{\pi(2)} \rangle , \ldots \langle e_{\pi(1)},e_{\pi(2)},\ldots , e_{{\pi(n-1)}} \rangle \} $$ the full flag constructed by reordering the basis vectors in $E$ according to $\pi$. For $\pi \in \Sym_n$ let $\tilde{\pi }\in \GL_n(K)$ denote the corresponding permutation matrix so that $\Delta _{\pi } = \Delta _0 \tilde{\pi } $. Then the set of all full flags that can be constructed by reordering the basis vectors in $E$ is $${\mathcal A} := \{ \Delta _{\pi } \mid \pi \in \Sym_n \} = \{ \Delta _0 \tilde{\pi } \mid \pi \in \Sym_n \} .$$ \begin{definition} \label{apart} The set ${\mathcal A}$ is called the {\bf standard apartment}. \end{definition} The standard apartment ${\mathcal A}$ has a very nice property that it contains a system of representatives of the $B$-orbits on ${{\mathcal F} }_f(V)$. This follows directly from the Gau\ss-Bruhat decomposition: \begin{cor}\label{GB} For all $\Delta \in {{\mathcal F} }_f(V) $ there is a unique $\pi (\Delta ) =: \pi \in \Sym _n$ such that $$\Delta b = \Delta _{\pi } \in {\mathcal A} $$ for some $b\in B$. \end{cor} The next lemma expresses the well known fact that any two flags have a compatible basis. \begin{lemma}\label{compbas} For any two $\Delta , \Delta ' \in {{\mathcal F} }_f(V)$ there is some $g\in \GL (V)$ such that $\Delta g = \Delta _0 $ and $\Delta ' g = \Delta _{\pi } \in {\mathcal A} $ for some $\pi \in \Sym _n$, uniquely determined by $\Delta $ and $\Delta '$. In particular any $\GL(V)$-invariant distance function $d$ on ${{\mathcal F}}_f(V)$ satisfies $d(\Delta, \Delta ') = d(\Delta _0 ,\Delta _{\pi })$. \end{lemma} \noindent\underline{Proof.}\ As the action of $\GL(V)$ on ${{\mathcal F} } _f (V)$ is transitive, there is some $h\in \GL(V)$ such that $\Delta h = \Delta _0$. By Corollary \ref{GB} there is some $b\in B$ such that $(\Delta ' h ) b = \Delta _{\pi }$ (with $\pi = \pi (\Delta ' h) \in \Sym _n$). Then $g:=hb$ satisfies $\Delta g = \Delta _0 b = \Delta _0$ and $\Delta' g = \Delta _{\pi} $ as desired. \phantom{zzz}\hfill{$\square $}\smallskip As any flag can be refined to be a full flag, Lemma \ref{compbas} holds equally for non full flags. \begin{corollary} \label{compbasgen} For any two flags $\Lambda , \Lambda ' \in {\mathcal F} (V) $ (not necessarily of the same type) there is some $g\in \GL(V)$ such that $\Lambda g$ and $\Lambda ' g$ are contained in full flags lying in ${\mathcal A}$. \end{corollary} \section{Distance functions on spherical buildings} \subsection{The $\Sym _n$-valued distance function} In this section we want to study $\GL (V)$-invariant distance functions on ${\mathcal F} _T(V)$. We have seen in Remark \ref{distconst} that such functions are constant on the double cosets. In particular for the full flags ${\mathcal F} _f(V)$ they factor through the $\Sym _n$-valued distance function: \begin{definition} The $\Sym _n$-valued distance function of ${\mathcal F} _f(V)$ is defined as $$d_W(\Delta _0 g , \Delta _0 h) = \pi \in \Sym _n \mbox{ if } hg^{-1} \in B\pi B .$$ \end{definition} For two flags $\Delta _{\pi }$ and $\Delta _{\sigma } $ in the standard apartment ${\mathcal A}$ we have $$d_W(\Delta _{\pi } , \Delta _{\sigma } ) = \sigma \pi ^{-1 } .$$ \begin{remark} For $\pi \in \Sym _n$ the $\pi $-circle around $\Delta _0$ is defined as $$C_{\pi }(\Delta _0) := \{ \Delta \in {\mathcal F}(V) \mid d_W(\Delta _0 ,\Delta ) = {\pi } \} .$$ Then $$C_{\pi }(\Delta _0) = \{ \Delta _0 \tilde{\pi } b \mid b\in B \} =\Delta _{\pi } B .$$ The $\pi$-circle $C_{\pi }(\Delta _0)$ is in bijection with the subgroup $U_{\pi }$ mentioned in the end of Section \ref{flags}, i.e. for each $\Delta \in C_{\pi }(\Delta _0)$ there is a unique $u\in U_{\pi }$ such that $$\Delta = \Delta _0 {\pi } u .$$ In particular if \emph{$K$} is a finite field with $q$ elements, then $C_{\pi }(\Delta _0)$ contains exactly $q^{\ell (\pi )}$ elements. For each ${\pi }\in \Sym _n$ the intersection of the $\pi $-circle around $\Delta _0$ with the standard apartment ${\mathcal A} $ defined in Definition \ref{apart} is $$C_{\pi }(\Delta _0) \cap {\mathcal A} = \{ \Delta _{\pi } \} .$$ \end{remark} \subsection{The Grassmann distance of flags} In this section we define a new distance on the set of all flags of a given type in $V$, which we call the {\bf Grassmann distance of flags}, because it is a direct generalization of the Grassmann distance on ${{\mathcal G}}_{k} (V)$ the set of subspaces of $V$ of dimension $k$. \begin{defn}\label{Grassdist} Let $\Lambda =\{ W_1,\ldots , W_{m} \}$ and $\Lambda ' =\{ W'_1,\ldots , W'_{m} \}$ be two flags in $V$ of the same type $T=\{d_1,\ldots , d_m\}$ with $d_i = \dim(W_i) = \dim (W'_i) \in \{ 1,\ldots, n-1 \}$ for all $i$. Then the Grassmann distance is defined as $$\E(\Lambda, \Lambda ') := \sum _{i=1}^{m} (d_i-\dim (W_i\cap W'_i) ) .$$ \end{defn} \begin{theorem}\label{GrassdistTheorem} For any type $T$, the Grassmann distance $\E$ is a $\GL(V)$-invariant distance function on the set ${\mathcal F} _T(V)$ of all flags of type $T$. \end{theorem} \noindent\underline{Proof.}\ Let $\Lambda =\{ W_1,\ldots , W_{m} \}$, $\Lambda ' =\{ W'_1,\ldots , W'_{m} \}$ and $\Lambda '' =\{ W''_1,\ldots , W''_{m} \}$ be flags in $V$ of the same type $\{ d_1,\ldots , d_m\}$ with $d_i = \dim(W_i) = \dim (W'_i) = \dim(W''_i) \in \{ 1,\ldots, n-1 \}$ for all $i$. \\ We clearly have that $\E(\Lambda , \Lambda ') = 0 $ if and only if $\Lambda = \Lambda' $. Also the symmetry $\E(\Lambda ,\Lambda') = \E(\Lambda' , \Lambda )$ is clear. The triangle inequality $$\E(\Lambda , \Lambda '') \leq \E(\Lambda , \Lambda ') + \E(\Lambda' , \Lambda '')$$ follows from the well known triangle inequality of the Grassmann distance on subspaces: For all $1\leq i \leq m$ we have $$ d_i-\dim (W_i\cap W''_i) \leq (d_i-\dim (W_i\cap W'_i)) + (d_i-\dim (W'_i\cap W''_i)) $$ so this also holds for the sum. That the function $\E$ is $\GL(V)$-invariant follows directly from the definition. \phantom{zzz}\hfill{$\square $}\smallskip For two subspaces $W_i,W'_i $ of dimension $i$ we have $$\dim (W_i\cap W'_i) + \dim (W_i + W'_i ) = \dim (W_i ) + \dim (W'_i) = 2i .$$ In particular $\dim (W_i \cap W'_i) \geq 2i -n $ so we have \begin{remark} (cf. Remark \ref{scoopdist}) For two full flags $\Delta, \Delta '$ we have that $$\E (\Delta ,\Delta ') \leq \left\{ \begin{array}{ll} (n/2)^2 & n \mbox{ even } \\ (n-1)(n+1)/4 & n \mbox{ odd } \end{array} \right. $$ \end{remark} The Gau\ss -Bruhat decomposition shows that every $\GL (V)$-invariant distance on the set of all full flags in $V$ factors through $d_W$. This also holds for the Grassmann distance $\E$, where $d_W$ is composed with the depth function $\depth$ from Definition \ref{depth}. \begin{corollary} For any pair $\Delta , \Delta ' \in {\mathcal F} _f(V)$ of full flags in $V$, we have $$\E(\Delta, \Delta ' ) = \depth(d_W(\Delta, \Delta ')) .$$ \end{corollary} \noindent\underline{Proof.}\ By Lemma \ref{compbas} it is enough to consider the standard apartment ${\mathcal A}$ and to show that for all $\pi \in \Sym _n$ $$\E(\Delta _0,\Delta _{\pi } ) = \depth(\pi ) .$$ So let $V_i= \langle e_1,\ldots ,e_i \rangle $ and $V'_i = \langle e_{\pi(1)} ,\ldots , e_{\pi(i)} \rangle $ for all $1\leq i \leq n-1$. Then $$V_i\cap V'_i = \langle e_j \mid j \leq i \mbox{ and } \pi ^{-1} (j) \leq i \rangle = \langle e_{\pi(k)} \mid \pi(k) \leq i \mbox{ and } k \leq i \rangle $$ in particular $$ i -\dim (V_i\cap V'_i) = | \{ k\in \{ 1,\ldots , i \} \mid \pi (k) > i \} | .$$ \phantom{zzz}\hfill{$\square $}\smallskip \begin{corollary}\label{esym} As $d_W(\Delta',\Delta ) = d_W(\Delta , \Delta ') ^{-1}$ and the function $\E $ is symmetric we obtain that $\depth(\pi ) = \depth(\pi ^{-1} )$ for all $\pi \in \Sym _n$. \end{corollary} \subsection{The gallery distance} In the theory of spherical buildings, the most commonly used distance function is the the gallery distance. This section compares the Grassmann distance to the gallery distance. \begin{definition}\label{galery} Two full flags $\Delta $ and $\Delta '$ are said to have gallery distance 1 $$d_G(\Delta,\Delta ') = 1 $$ if and only if their intersection $\Delta \cap \Delta $ has cardinality $n-2$, $$\Delta \cap \Delta ' = \Delta \setminus \{W_k \} = \Delta ' \setminus \{ W'_{k } \} $$ for some $W_k \in \Delta $, $W'_{k}\in \Delta '$. A {\bf gallery} is a sequence ${\mathcal G} = (\Delta _1,\Delta _2,\ldots , \Delta _m)$ of full flags $\Delta _i$ such that $d(\Delta _i,\Delta _{i+1}) = 1$ for all $1\leq i < m$. The {\bf length} of the gallery ${\mathcal G} $ is $m-1$. It is well known (and follows from elementary linear algebra) that any two flags $\Delta $ and $\Delta '$ can be joined by some gallery ${\mathcal G} = (\Delta, \Delta_1,\ldots , \Delta_{m -1 } , \Delta ')$. Then the {\bf gallery distance} $d_G(\Delta,\Delta ')$ is the minimal length $m $ of such a gallery. \end{definition} \begin{theorem}(\cite[Section 4.8]{Abramenko}) For all $\Delta \in {{\mathcal F} }(V)$ we have $d_G(\Delta _0 , \Delta ) = \ell (\pi (\Delta )) $. In particular if $\Delta ,\Delta '\in {\mathcal F} _f(V) $ then $$d_G(\Delta , \Delta ') = \ell (d_W(\Delta , \Delta ')).$$ \end{theorem} From Theorem \ref{lesym} we now immediately obtain the following Corollary. \begin{corollary}\label{dE} If $\Delta \neq \Delta ' \in {{\mathcal F} }_f(V)$ are full flags in $V$, then $$ 2\E(\Delta, \Delta ') > d_G(\Delta, \Delta') \geq \E(\Delta , \Delta ') .$$ \end{corollary} \section{The channel model} Throughout this section we will work with a fixed type $T = \{ d_1,d_2,\ldots, d_{m}\}$ with $0<d_1<\ldots < d_m < n$, and put $k_i := d_i - d_{i-1}$, $i=2,\ldots ,m$, $k_1:=d_1$. We will model our network as a finite directed, acyclic multigraph with a single source and possibly multiple receivers. Every edge gets a capacity of $1$, but we allow multiple edges between nodes to model different capacities. The source and the receivers agree on a set ${\mathcal C}\subset {\mathcal F}_T(V)$ of flags of type $T$, the error correcting code. Information is encoded as a flag $\Lambda \in {\mathcal C}$. \\ Assume now that the source has a flag $\Lambda = \{ V_1 , V_2 , \ldots , V_m \} \in {\mathcal C} $ with $d_i = \dim (V_i)$. Fixing a basis of $V$ and therewith identifying $V$ with the space of rows, $K^n$, we may think of $\Lambda $ as a sequence of row vectors $x_1,x_2, \ldots x_{d_m}\in K^n$ such that $x_1,x_2,\ldots, x_{d_i}$ form a basis of $V_i$. For $1\leq j\leq m$ let $X_j \in K^{d_j \times n}$ be the matrix whose $i$-th row is $x_i$. Of course $X_j$ is a submatrix of $X_{j+1}$ and so all the information is contained in the matrix $X_m$. \\ At every time step $1 \leq i \leq m$ and for every outgoing edge the source chooses random coefficients $y\in K^{1\times d_i}$ and sends $y \cdot X_i \in V$ through that edge. \\ Furthermore every intermediate node forms a random linear combination of everything received up to this point for every edge. \\ So at time $i$ the receiver receives many (say $a_i$) random linear combinations of the rows $x_1,\ldots , x_{d_i} $, i.e. $Z_i= Y_i \cdot X_i$ with $Y_i\in K^{a_i\times d_i }$. The receiver defines spaces $$W_i := \langle \text{ rows of } Z_j \mid 1\leq j\leq i \rangle. $$ Then by definition $W_i \leq W_{i+1}$ for all $i$. Put $\Gamma := (W_1 , W_2 , \ldots , W_m)$. \begin{remark} If $Z_i = Y_i \cdot X_i$ and the rank of the matrix formed by the last $k_i$ columns of $Y_i$ equals $k_i$ for all $i$, then $W_i = V_i$ for all $i$ and $\Lambda = \{ W_1,\ldots , W_m \}$. This is the case if there are no errors or erasures in the transmission. Note that a necessary condition is that each $Y_i$ has at least $k_i$ rows, so all the $k_i$ need to be at most the capacity of the network. \end{remark} Note however that due to erasures or errors the receiver gets some matrix $\tilde{Z}_i = Y_i X_i + E_i $ where the rank of $Y_i$ is smaller than $d_i$ (due to erasures) and $E_i \neq 0$ (due to errors). We then might have that $W_i\neq V_i$, and $\Gamma $ might not even be a flag of length $m$, but only a stuttering flag in the sense of the following definition. \begin{definition} \begin{itemize} \item[(a)] A {\em stuttering flag} of length $m$ is a sequence $\Gamma := (W_1 , W_2 , \ldots , W_m)$ of subspaces of $V$ such that $W_1\leq W_2 \leq \ldots \leq W_m \leq V$. \item[(b)] Let $\Lambda =\{V_1,\ldots , V_m\} \in {\mathcal C} $ be the sent flag and $\Gamma = (W_1 , W_2 , \ldots , W_m) $ the received stuttering flag. In analogy to \cite[Definition 1]{KK} we define $$\rho _i = \rho_i(\Lambda , \Gamma ) := \dim (V_i) - \dim (W_i\cap V_i)$$ to be the number of erasures in step $i$ and $$f_i = f_i(\Lambda , \Gamma ) := \dim (W_i) - \dim (W_i\cap V_i)$$ the number of errors in step $i$. \item[(c)] The final error count between $\Lambda $ and $\Gamma $ is $$E(\Lambda,\Gamma ) := \sum_{i=1}^m \dim(V_i + W_i) - \dim(V_i \cap W_i).$$ \end{itemize} \end{definition} \begin{remark} The final error count satisfies $E(\Lambda , \Gamma ) = \sum_{i=1}^m (\rho_i(\Lambda ,\Gamma ) + f_i (\Lambda, \Gamma ))$. \end{remark} \noindent\underline{Proof.}\ This follows from the famous Grassmann identity: \\ $\dim (V_i+W_i) + \dim (V_i \cap W_i) = \dim(V_i) + \dim(W_i) .$ \phantom{zzz}\hfill{$\square $}\smallskip Note that if $\Lambda $ and $\Gamma $ are both flags of type $T$ then $E(\Lambda , \Gamma ) = 2 \E (\Lambda ,\Gamma)$. The error count also originates from the Grassmannian distance on subspaces and is thus a metric satisfying the triangle inequality. Hence in analogy to \cite[Theorem 2]{KK} we get the following corollary. \begin{corollary} Let ${\mathcal C}$ be a set of flags of type $T$ and $$d({\mathcal C}) := \min \{\E(\Lambda', \Lambda) \mid \Lambda ' \neq \Lambda \in {\mathcal C}\}$$ be the minimum distance of ${\mathcal C} $. Using the code ${\mathcal C} $ for transmission through the network we can correct all errors as long as the error count satisfies $E(\Lambda , \Gamma) < d({\mathcal C}) ,$ meaning that in this case $\Lambda $ is the unique element of ${\mathcal C} $ such that $E(\Lambda,\Gamma )$ is minimal. \end{corollary} \noindent\underline{Proof.}\ Let $e := E(\Lambda, \Gamma)$. For another flag $\Lambda \neq \Delta \in {\mathcal C}$ set $f := E(\Delta, \Gamma).$ Then the triangle inequality gives us $$E(\Lambda, \Delta) \leq E(\Lambda, \Gamma) + E(\Gamma, \Delta) = e + f.$$ On the other hand $\Lambda$ and $\Delta$ are elements of ${\mathcal C}$ and thus we can use the observation from above to get $$d({\mathcal C}) \leq \E(\Lambda, \Delta) = \frac{E(\Lambda, \Delta)}{2}.$$ Putting these together we get $2d({\mathcal C}) \leq e+f$. But as we assumed that $e < d({\mathcal C})$ we thus have $f > d({\mathcal C}) > e$, hence $\Lambda$ is indeed the unique element of ${\mathcal C}$ having minimal distance to $\Gamma$. \phantom{zzz}\hfill{$\square $}\smallskip \subsection{Error correcting codes} For good error correcting codes, as in the classical situation, $|{\mathcal C} |$ and $d({\mathcal C} )$ both should be large. One idea is to construct ${\mathcal C} $ as an orbit $\Delta _T S$ of some subgroup $S\leq \GL(V) $ with $S\cap P_T = \{ 1 \}$. Then, using Remark \ref{distconst} we can compute the minimum distance on ${\mathcal C}= \Delta _T S $ as follows: \begin{remark} \label{distmat1} For $g,h\in \GL(V)$ we have $$\E(\Delta _T g , \Delta _T h ) =\E(\Delta _T (gh^{-1}), \Delta _T) =: \overline{\E}_T(gh^{-1}) .$$ In particular if $S\leq \GL(V) $ with $S\cap P_T = \{ 1\}$, then $$d(\Delta _T S) = d_T(S) := \min \{ \overline{\E}_T(g)\mid 1\neq g\in S \} .$$ As usually we abbreviate $\overline{\E}_{\{ 1,\ldots , n-1 \} } $ by $\overline{\E}_f$. \end{remark} \begin{lemma}\label{distmat} Let $T = \{ d_1,\ldots , d_m \}$, $k_i:=d_i-d_{i-1}$ ($2\leq i\leq m$), $k_1:=1$, $k_{m+1} :=n-d_m$, and $$g = \left(\begin{array}{ccccc} I_{k_1} & A_{12} & \ldots & \ldots & A_{1m} \\ 0 & I_{k_2} & A_{23} & \ldots & A_{2m} \\ \vdots & \ddots & \ddots & \ddots & \vdots \\ 0 & \ldots & 0 & I_{k_{m}} & A_{mm} \\ 0 & \ldots & \ldots & 0 & I_{k_{m+1}} \end{array} \right) $$ where $A_{ij} \in K^{k_i \times k_j }$. For $i=1,\ldots , m$ put $$g_i := \left(\begin{array}{ccc} A_{1,i+1} & \ldots & A_{1m} \\ \vdots & \vdots & \vdots \\ A_{i,i+1} & \ldots & A_{im} \end{array} \right) $$ the upper right $d_i \times (n-d_i)$ submatrix of $g$ and $r_i = \rk(g_i)$. Then $\overline{\E}_T(g) = \sum _{i=1}^m r_i .$ \end{lemma} \noindent\underline{Proof.}\ Let $\Delta_T = (V_1,V_2,\ldots, V_m)$ and $\Lambda = \Delta_T g = (W_1,W_2,\ldots, W_m)$ where $\dim(V_i) = \dim(W_i) = d_i$ for all $i$. We claim that $r_i = d_i - \dim(V_i\cap W_i)$, again for all $i$. The proof of the lemma follows directly from that claim by writing a sum on both sides. \\ To prove the claim fix one $i$ and consider the matrix $$M = \begin{pmatrix} \begin{matrix} I_{k_1} & A_{12} & \ldots & \ldots \\ 0 & I_{k_2} & A_{23} & \ldots \\ \vdots & \ddots & \ddots & \ddots \\ 0 & \ldots & 0 & I_{k_{i}} \end{matrix} & \begin{matrix} A_{1,i+1} & \ldots & A_{1m} \\ A_{2,i+1} & \ldots & A_{2m} \\ \vdots & \vdots & \vdots \\ A_{i,i+1} & \ldots & A_{im} \end{matrix} \\ \Scale[2]{I_{d_i}} & \Scale[2]{0}\end{pmatrix}. $$ Then the row space of $M$ equals $V_i+W_i$ and thus the rank of $M$ equals $$\rk(M) = \dim(V_i+W_i) = 2d_i-\dim(V_i\cap W_i).$$ To compute the rank of $M$ we use Gau\ss{}ian elemination. As we have a big identity matrix on the bottom we can use that to reduce $M$ to the matrix $$\begin{pmatrix}0 & g_i \\ I_{d_i} & 0\end{pmatrix}.$$ Now we see $\rk(M) = r_i+d_i$ and this gives us $$r_i+d_i = \rk(M) = 2d_i-\dim(V_i\cap W_i) \,\,\,\, \Rightarrow\,\,\,\, r_i = d_i-\dim(V_i\cap W_i).$$ \phantom{zzz}\hfill{$\square $}\smallskip In fact we retrieve the idea of \cite{SKK} here: \begin{example}\label{codes0} Let $T:= \{ k \} $, so ${\mathcal F} _T(V)$ correspond to the Grassmannian ${\mathcal G} _k(V)$. Choose some subspace $C\leq K^{k\times n-k} $ and put $$U_C := \{ u(c) := \left( \begin{array}{cc} I_k & c \\ 0 & I_{n-k} \end{array} \right) \mid c\in C \} .$$ Then $U_C$ is a subgroup of $\GL (V)$ isomorphic to the additive group of $C$. In particular for $c,c' \in C$ we compute $u(c) u(c')^{-1} = u(c-c') $. So Lemma \ref{distmat} is a direct generalization of \cite[Proposition 4]{SKK} that the Grassmann distance on the subspaces with basis matrix $(I_k|c)$ is the rank metric on $K^{k\times n-k }$. \end{example} To compare such commonly used subspace codes with our new flag codes assume for convenience that $n=4m$ and $C \leq K^{2m\times 2m}$ is an MRD code of dimension $2m$ with rank metric distance $$d(C) := \min \{ \rk (c) \mid 0\neq c\in C \} = 2m $$ (see for instance \cite[Section C]{SKK}). Then $$\dim (C) = 2m = \dim (U_C) \mbox{ and } d(C) = d_{\{2m \} } (U_C) = 2m .$$ Using flags of type $T= \{ m,2m,3m \} $ we can improve on the dimension of the flag code (we get dimension $3m$) keeping the minimum distance to be $2m$: \begin{prop}\label{codes1} Assume that $n=4m$ and put $T:= \{ m,2m,3m \} $. Given two MRD codes $C_m$ and $C_{2m}$ with $$C_i \leq K^{i\times i}, d(C_i) = i, \dim (C_i ) = i $$ we put $${\mathcal C} (C_{m},C_{2m}):= \{ \Delta _{T} u(x,y) \mid x\in C_m, y\in C_{2m} \} \subseteq {\mathcal F} _T(V)$$ where $$ u(x,y) := \left(\begin{array}{cc} \begin{array}{cc} I_m & x \\ 0 & I_m \end{array} & y \\ 0 & \begin{array}{cc} I_m & x \\ 0 & I_m \end{array} \end{array} \right) \in \GL_{4m} (K) $$ Then $\dim ({\mathcal C}(C_{m},C_{2m})) = 3m $ and $d({\mathcal C}(C_{m},C_{2m})) = 2m .$ \end{prop} For the proof we need the following elementary fact about multiplication of block triangular matrices. \begin{lemma}\label{multmat} Let $A \in \GL_m(K)$, $B,D\in K^{m\times m} $. Then $$\left( \begin{array}{cc} A & B \\ 0 & A \end{array} \right) ^{-1 } = \left( \begin{array}{cc} A^{-1} & -A^{-1}BA^{-1} \\ 0 & A^{-1} \end{array} \right) $$ and $$\left( \begin{array}{cc} A & D \\ 0 & A \end{array} \right) \left( \begin{array}{cc} A & B \\ 0 & A \end{array} \right) ^{-1 }= \left( \begin{array}{cc} I_m & (D-B)A^{-1} \\ 0 & I_m \end{array} \right) .$$ \end{lemma} \noindent\underline{Proof.}\ (of Proposition \ref{codes1}) For $0\neq x\in C_m$, the rank of $x\in K^{m\times m}$ is $m$ and hence $\overline{\E}_T(u(x,y)) \geq 2m $ for any $y\in K^{2m\times 2m}$ by Lemma \ref{distmat}. Now $u(x,y)^{-1} = u(-x,y'' )$ for some $y'' \in K^{2m\times 2m} $, hence $u(x',y') u(x,y)^{-1} = u(x'-x,y''') $ for some $y'''$ so by Remark \ref{distmat1}. $$\E(\Delta _T u(x,y) , \Delta _T u(x',y') ) = \overline{\E}_T(u(x'-x,y''')) \geq 2m \mbox{ if } x\neq x'.$$ Assume that $x=x'$ then by Lemma \ref{multmat} $ u(x,y') u(x,y)^{-1} = u(0,y'') $ with $$y'' = (y'-y) \left( \begin{array}{cc} 1 & -x \\ 0 & 1 \end{array} \right) $$ in particular the rank of $y'' $ is the same as the one of $y-y'$. If $y\neq y'$ then this is a non zero element of $C_{2m}$ so it has rank $2m$. Using Remark \ref{distmat1} we again find $\E (\Delta _T u(x,y) , \Delta _T u(x,y') ) = 2m $. \phantom{zzz}\hfill{$\square $}\smallskip \subsection{Checkerboard codes} We now want to iterate the idea from Proposition \ref{codes1}. Assume that we have a sequence of MRD codes $C_i \leq K^{2^i\times 2^i} $ $(i=0,\ldots , t)$ such that $$\dim _K(C_i ) = 2^i,\ \ \rk (c) = 2^i \mbox{ for all } 0\neq c\in C_i .$$ \begin{definition} \label{codes2} For $x_i \in C_i $ ($0\leq i \leq t$) we define $$u(x_0,x_1, \ldots ,x_t) \in \GL_{2^{t+1}} (K) $$ recursively as $$u(x_0) = \left(\begin{array}{cc} 1 & x_0 \\ 0 & 1 \end{array}\right) , \ u(x_0,\ldots , x_t) = \left( \begin{array}{cc} u(x_0,\ldots , x_{t-1}) & x_t \\ 0 & u(x_0,\ldots , x_{t-1}) \end{array}\right) .$$ Then the {\bf checkerboard code} associated to the MRD codes $C_i$ is $${\mathcal C} (C_0,C_1,\ldots , C_t) = \{ \Delta _0 u(x_0,x_1, \ldots ,x_t) \mid x_i \in C_i \mbox{ for } 0\leq i \leq t \} \subset {{\mathcal F} }_f (V).$$ \end{definition} Note that the dimension of ${\mathcal C} (C_0,C_1,\ldots , C_t) $ is $ \sum _{i=0}^t 2^i = 2^{t+1} -1 .$ \begin{prop} Let ${\mathcal C} := {\mathcal C} (C_0,C_1,\ldots , C_t) $. \\ Then $\dim ({\mathcal C}) = 2^{t+1}-1 $ and $d ({\mathcal C} ) = 2^{t} $. \end{prop} \noindent\underline{Proof.}\ We show by induction on $t$ that $d ({\mathcal C} (C_0,C_1,\ldots , C_t)) \geq 2^{t} $. For $t=0$ there is nothing to show. So let $$ g:=\left( \begin{array}{cc} A & B \\ 0 & A \end{array} \right), h:=\left( \begin{array}{cc} A' & B' \\ 0 & A' \end{array} \right) \in \GL _{2^{t+1}} (K) $$ where $A,A' \in \{ u(x_0,\ldots , x_{t-1}) \mid x_i \in C_i \}$, $B,B' \in C_t$. By Remark \ref{distmat1} we need to show that $\overline{\E} _f (hg^{-1}) \geq 2^t$, if $g\neq h$. By Lemma \ref{multmat} $$hg^{-1} = \left( \begin{array}{cc} A'A^{-1} & B'' \\ 0 & A'A^{-1 } \end{array} \right) $$ with $B'' = (B'-B) A^{-1} $ if $A=A'$. If $A\neq A'$ then $$\overline{\E}_f(hg^{-1} ) \geq 2 \overline{\E}_f (A'A^{-1}) \geq 2 \cdot 2^{t-1} = 2^t $$ by induction. If $A=A' $ then $B'\neq B \in C_{t}$ (because $g\neq h$) and hence $\rk (B'-B) = 2^t$ so $\rk (B'') = 2^t$ and $\overline{\E}_f(hg^{-1} ) \geq 2^t $ by Lemma \ref{distmat}. Note that $\overline{\E}_f(u(x_0,0,\ldots ,0) ) = 2^t $ by Lemma \ref{distmat} so $d ({\mathcal C} (C_0,C_1,\ldots , C_t)) \leq 2^{t} $ and hence we get the equality as claimed. \phantom{zzz}\hfill{$\square $}\smallskip \subsection{Derived subgroup codes} Take $D$ to be the subgroup of all upper uni-triangular matrices in $\GL_n(K)$: $$D = \left\{ \left( \begin{array}{cccc} 1 & * & * & * \\ 0 & 1 & * & * \\ 0 & 0 & \ddots & \vdots \\ 0 & 0 & 0 & 1\end{array} \right) \mid * \in K \right\} \leq \GL_n(K).$$ Then the derived subgroups of $D$ are of the form $$D^{(k)} = \{g \in D \mid g_{ij} = 0 \text{ for } 0 < j-i \leq k \},$$ thus having exactly $k$ secondary diagonals filled with zeros. \begin{prop}\label{derived} The code ${\mathcal C} (n,k) := \Delta_0 D^{(k)}$ consists of fine flags and has parameters $$d({\mathcal C} (n,k)) = k+1,\ \dim({\mathcal C} (n,k)) = \frac{(n-k)(n-k-1)}{2} .$$ \end{prop} \noindent\underline{Proof.}\ To compute the dimension of ${\mathcal C} (n,k) $ we just count the number of free parameters to be $$\sum _{j=1}^{n-k-1} j = \frac{(n-k)(n-k-1)}{2} .$$ For computing the minimal distance we use the fact that $D^{(k)}$ is a group. So it suffices to compute $\overline{\E}_f(g)$ for $1\neq g \in D^{(k)}$. Then $g$ has at least one non-zero entry at a position $(i,j)$ with $j \geq i+k+1$. Using Lemma \ref{distmat} this gives us $j-i \geq k+1$ matrices with rank at least one, hence $\overline{\E}_f(g) \geq k+1$. \\ On the other hand taking a $g \in D^{(k)}$ that only has one non-zero entry at position $(i,i+k+1)$ for some $i$ yields a matrix with $\overline{\E}_f(g) = k+1$. \phantom{zzz}\hfill{$\square $}\smallskip \begin{rem} The code ${\mathcal C}(n,k)$ allows for decoding erasures using only Gau\ss{}ian elimination. If we receive a stuttering flag $$\Lambda = (U_1,\ldots, U_{n-1})$$ we can uniquely recompute the corresponding matrix $g \in D^{(k)}$ as long as the longest subchain $U_i,U_{i+1},\ldots,$ of subspaces such that $\dim(U_{i+j}) < i+j$ has length at most $k$. \end{rem} \noindent\underline{Proof.}\ Let $g_j$ be the submatrix of $g$ consisting of the first $j \leq n$ rows. Then due to the zeros on the secondary diagonals the last $k+1$ rows of $g_j$ are not changed when computing the reduced row echelon form of $g_j$. If we receive a subspace $U_j$ with $\dim(U_j) = j$ we can hence compute a reduced row echelon form of a generator matrix of $U_j$ and by the uniqueness of that form we get the $i$-th row of $g$ for all $j-k \leq i \leq j$. Thus we can recompute $g$ as long as we have that at least every $k-$th space in $\Lambda$ has the correct dimension. \phantom{zzz}\hfill{$\square $}\smallskip {\small

https://arxiv.org/abs/1808.02251

Automorphisms on the ring of symmetric functions and stable and dual stable Grothendieck polynomials

The dual stable Grothendieck polynomials $g_\lambda$ and their sums $\sum_{\mu\subset\lambda} g_\mu$ (which represent $K$-homology classes of boundary ideal sheaves and structure sheaves of Schubert varieties in the Grassmannians) have the same product structure constants. In this paper we first explain that the ring automorphism $g_\lambda\mapsto\sum_{\mu\subset\lambda} g_\mu$ on the ring of symmetric functions is described as the operator $F^\perp$, the adjoint of the multiplication $(F\cdot)$, by a "group-like" element $F=\sum_{i} h_i$ where $h_i$ is the complete symmetric function. Next we give a generalization: starting with another "group-like" elements $\sum_{i} t^i h_i$, we obtain a deformation with a parameter $t$ of the ring automorphism above, as well as identities involving stable and dual stable Grothendieck polynomials.

\section{Introduction} The dual stable Grothendieck polynomials $g_\lambda$ are a certain family of inhomogeneous symmetric functions parametrized by interger partitions $\lambda$. These functions are a $K$-theoretic deformation of the Schur functions, and dual to another deformation called stable Grothendieck polynomials $G_\lambda$. Historically the stable Grothendieck polynomials (parametrized by permutations) were introduced by Fomin and Kirillov \cite{MR1394950} as a stable limit of the Grothendieck polynomials of Lascoux--Sch\"utzenberger \cite{MR686357}. In \cite{MR1946917} Buch gave a combinatorial formula for the stable Grothendieck polynomials $G_\lambda$ for partitions using so-called set-valued tableaux, and showed that their span $\bigoplus_\lambda \mathbb{Z} G_\lambda$ is a bialgebra and that the $K$-theory of a Grassmannian is isomorphic to a quotient of it. The dual stable Grothendieck polynomials $g_\lambda$ were introduced by Lam and Pylyavskyy \cite{MR2377012} as generating functions of reverse plane partitions (Definition \ref{defi:gla}), and shown to be the dual basis for $G_\lambda$ via the Hall inner product and represent the $K$-homology classes of the ideal sheaves of boundaries of Schubert varieties in Grassmannians. As seen in Remark \ref{rema:Gg_geom}, the sums $\sum_{\mu\subset\lambda}g_\mu$ represent the classes in $K$-homology of structure sheaves of Schubert varieties. In \cite{dualstablesum} the author showed that the bases $\{g_\lambda\}_\lambda$ and $\{\sum_{\mu\subset\lambda}g_\mu\}_\lambda$ have the same product structure constants, i.e.\,the linear map $g_\lambda\mapsto\sum_{\mu\subset\lambda}g_\mu$, which we denote by $I$, is a ring automorphism on the ring of symmetric functions $\Lambda$, by showing the Pieri rules with respect to these bases have the same coefficients. In this paper we first give alternative descriptions for this map $I$: it is realized as the substitution $f(x_1,x_2,\cdots)\mapsto f(1,x_1,x_2,\cdots)$ \eqref{eq:I=i*id}. The key observation used there is that the substitution $f\mapsto f(1,0,0,\cdots)$ maps $g_{\lambda/\mu}$ to $1$ for any skew shape $\lambda/\mu$ (Proposition \ref{theo:i}); then $I$ is a certain composition of this map and the coproduct on $\Lambda$. We also give two formulas for the image of $g_{\lambda/\mu}$ under $I$ \eqref{eq:I_skew}, which generalizes $I(g_\lambda)=\sum_{\nu\subset\lambda}g_\nu$ and implies some identities involving skew dual stable Grothendieck polynomials. Next we give another description using the Hopf structure on $\Lambda$ more explicitly. Due to the coalgebra structure of $\Lambda$ and the identification $\Lambda^*\simeq\widehat\Lambda$ (completion of $\Lambda$) via the Hall inner product, for any $F\in\widehat\Lambda$ there corresponds a linear map $F^\perp\colon\Lambda\longrightarrow\Lambda$, the adjoint of the multiplication map $(F\cdot)\colon\widehat\Lambda\longrightarrow\widehat\Lambda$. We explain in Lemma \ref{theo:group-like} that $F^\perp$ is a ring automorphism if and only if $F$ has ``group-like'' property. Then we show that the map $I$ is realized as $H(1)^\perp$ where $H(1):=\sum_{i}h_i=\sum_{\lambda} G_\lambda$ is group-like (Theorem \ref{theo:H(1)_overall}), while $I^{-1}=E(-1)^\perp$ where $E(-1):=\sum_{i}(-1)^i e_i=1- G_1$. Besides we explain that from presentations of the map $(H(1)\cdot)$ with respect to the basis $\{G_{\lambda}\}$ we can obtain identities involving $G_\lambda$, \eqref{eq:I*(1)} and \eqref{eq:D*(1)}. Finally we give a generalization: starting with another group-like elements $H(t)=\sum_{i\ge 0} t^i h_i$ and $E(t)=\sum_{i\ge 0} t^i e_i$, we obtain a deformation with a parameter $t$ of the automorphism $I$ and hence the identities obtained above involving $G_\lambda$ and $g_\lambda$. \subsection*{Organization} In Section \ref{sect:Prel} we recall Hopf algebras, symmetric functions and stable Grothendieck and dual stable Grothendieck polynomials. In Section \ref{sect:algebraic} we see connection between group-like elements in $\widehat\Lambda$ and automorphisms on $\Lambda$, as well as our main examples of group-like elements $H(t)$ and $E(t)$. Section \ref{sect:I} contains the first main results: after proving $I(f(x))=f(1,x)$ in Section \ref{sect:I:f(1,x)}, we show $I=H(1)^\perp$ and $I^{-1}=E(-1)^\perp$ in Section \ref{sect:I:H1_perp}. In following two sections we give a generalization. In Section \ref{sect:Ht} we give descriptions for the automorphism $H(t)^\perp$ and the multiplication map $(H(t)\cdot)$ using the bases $\{g_{\lambda}\}$ and $\{G_\lambda\}$, respectively, while in Section \ref{sect:Et} we treat $E(t)^\perp$ and $(E(t)\cdot)$. In Section \ref{sect:eg} we give some examples. Note that, although the results in Section \ref{sect:I} are special cases of that of Section \ref{sect:Ht} and \ref{sect:Et}, they need to be shown first according to the proof given in this paper. \Todo{check: $\mathrm{char}(K)=?$} \Todo{involution $\omega$} \subsection*{Acknowledgement} \AcknowledgementContent \section{Preliminaries}\label{sect:Prel} Throughout this paper, we fix a commutative ring $K$ and assume that any modules, algebras, morphisms etc.\ are over $K$. \subsection{Hopf algebra}\label{sect:Prel::Hopf} We recall some generalities on the Hopf algebra. For more details we refer the reader to \cite{MR0252485,MR594432,1409.8356} for example. An {\em algebra} $A$ is a $K$-module equipped with a {\em product} (or {\em multiplication}) $m=m_A\colon A\otimes A\longrightarrow A$ and a {\em unit} $u=u_A\colon K\longrightarrow A$ satisfying $m\circ(m\otimes\mathrm{id}) = m\circ(\mathrm{id}\otimes m)$ and $m\circ(\mathrm{id}\otimes u)=\mathrm{id}=m\circ(u\otimes\mathrm{id})$. A {\em coalgebra} $C$ is a $K$-module equipped with a {\em coproduct} (or {\em comultiplication}) $\Delta=\Delta_C\colon C\longrightarrow C\otimes C$ and a {\em counit} $\epsilon=\epsilon_C\colon C\longrightarrow K$ satisfying $(\Delta\otimes\mathrm{id})\circ\Delta = (\mathrm{id}\otimes \Delta)\circ\Delta$ and $(\mathrm{id}\otimes \epsilon)\circ\Delta=\mathrm{id}=(\epsilon\otimes\mathrm{id})\circ\Delta$. A $K$-linear map $\varphi\colon A\longrightarrow B$ between algebras is an {\em algebra morphism} if $\varphi\circ m_A = m_B\circ(\varphi\otimes\varphi)$ and $\varphi\circ u_A = u_B$. A $K$-linear map $\varphi\colon C\longrightarrow D$ between coalgebras is a {\em coalgebra morphism} if $(\varphi\otimes\varphi)\circ\Delta_C = \Delta_D\circ\varphi$ and $\epsilon_C = \epsilon_D\circ\varphi$. A $K$-module $A$ equipped with $m,u,\Delta,\epsilon$ is a {\em bialgebra} if $(A,m,u)$ is an algebra, $(A,\Delta,\epsilon)$ is a coalgebra, and the following equivalent conditions hold: (a) $\Delta,\epsilon$ are algebra morphisms; (b) $m, u$ are coalgebra morphisms. A bialgebra $A$ equipped with an \emph{antipode} map $S\colon A\longrightarrow A$ satisfying $m\circ(S\otimes\mathrm{id})\circ\Delta=u\circ\epsilon$ is called a \emph{Hopf algebra}. \subsubsection{duals} For a $K$-module $A$, let $A^*=\mathrm{Hom}(A,K)=\{f\colon A\longrightarrow K\colon K\text{-linear}\}$ and $(\,,)=(\,,)_A\colon A^*\times A\longrightarrow K\,;\,(f,a)=f(a)$. For a graded $K$-module $A=\bigoplus_{n\ge 0}A_n$, we denote by $A^o$ the \emph{graded dual} $\bigoplus_{n}A_n^*$, and $A$ is called {\em of finite type} if every $A_n$ is a finite free $K$-module. For any coalgebra $C$, its dual $C^*$ is an algebra by \begin{equation}\label{eq:dual_copro} (m_{C^*}(f\otimes g), a)_C = (f\otimes g, \Delta_C(a))_{C\otimes C} \end{equation} for $f,g\in C^*$ and $a\in C$. If an algebra $A$ is a finite free $K$-module (resp.\,$A$ is a graded algebra of finite type), then its dual $A^*$ (resp.\,graded dual $A^o$) is a coalgebra by \begin{equation}\label{eq:dual_pro} (\Delta_{A^*}(f), a\otimes b)_{A\otimes A}=(f,ab)_{A} \end{equation} for $f\in A^*$ (resp.\,$A^o$) and $a,b\in A$. \Todo{ Assume an algebra $A$ is finite free or of finite type and let $\{x_i\}_{i\in I}$ be a basis of $A$ and $\{f_i\}_{i\in I}$ be the dual basis of the coalgebra $A^*$ (or $A^o$). Then the product structure constants for $\{x_i\}$ coincide with the coproduct structure constants for $\{f_i\}$: if $x_i x_j=\sum_{k} a_{ij}^k x_k$ then $\Delta(f_k)=\sum_{i,j} a_{ij}^k f_i\otimes f_j$.} For a coalgebra $C$ and an algebra $A$, the space of linear maps $\mathrm{Hom}(C,A)$ becomes an associative algebra by the \emph{convolution product} $*$ defined by $f*g = m_A\circ(f\otimes g)\circ\Delta_C$. Then $u_A\circ\epsilon_C$ is the identity for $*$, and the convolution product on $C^*=\mathrm{Hom}(C,K)$ coincides with the product given in \eqref{eq:dual_copro}. \subsubsection{Module and comodule morphisms}\label{sect:Prel::Hopf::comod} For a coalgebra $C$, a linear map $\phi\colon C\longrightarrow C$ is \emph{$C$-comodule morphism} if $\Delta\circ\phi=(\phi\otimes\mathrm{id})\circ\Delta$. For an algebra $A$, a linear map $\psi\colon A\longrightarrow A$ is \emph{$A$-module morphism} if $\psi\circ m=m\circ(\psi\otimes\mathrm{id})$. \begin{lemm}\label{theo:comodule_hom} Let $C$ be a coalgebra and $C^*$ its dual algebra. For a linear map $\phi\colon C\longrightarrow C$, the following are equivalent: $(1)$ $\phi\colon C\longrightarrow C$ is a $C$-comodule morphism. $(2)$ $\phi^*\colon C^*\longrightarrow C^*$ is a $C^*$-module morphism. \end{lemm} \begin{proof} It easily follows from $(f,\phi(a)) = (\phi^*(f),a)$ and $(m(f\otimes g), a) = (f\otimes g, \Delta(a))$ (for $f,g\in C^*$ and $a\in C$). \end{proof} For a coalgebra $C$ and $f\in C^*$, the map $f^\perp\colon C\longrightarrow C$ is defined by $f^\perp(c) = \sum_{(c)} (f, c_1) c_2$ where we write $\Delta(c)=\sum_{(c)}c_1\otimes c_2$ by the Sweedler notation. In other words $f^\perp=(f\otimes\mathrm{id})\circ\Delta$. By \eqref{eq:dual_copro}, $f^\perp$ is the adjoint of $(f\cdot)\colon C^*\longrightarrow C^*\,;\,g\mapsto fg$, i.e.\,$(fg,c) = (g,f^\perp(c))$. Since the multiplication map $(f\cdot)$ is a $C^*$-module morphism, by Lemma \ref{theo:comodule_hom} we see that $f^\perp$ is a $C$-comodule morphism. Conversely, any $C$-comodule endomorphism on $C$ has the form $f^\perp$: \begin{lemm} \Enumerate{ \item For an algebra $A$, if $\psi\colon A\longrightarrow A$ is an $A$-module morphism then $\psi$ is the multiplication by $\psi(1_A)$. \item For a coalgebra $C$, if $\phi\colon C\longrightarrow C$ is a $C$-comodule morphism then $\phi = (\phi^*(1_{C^*}))^\perp$. } \end{lemm} \begin{proof} (1) is clear. (2) follows from (1), Lemma \ref{theo:comodule_hom} and the adjointness of $(\phi^*(1)\cdot)$ and $\phi^*(1)^\perp$. \end{proof} \subsection{Symmetric functions} For basic definitions for symmetric functions, see for instance \cite[Chapter I]{MR1354144}. Let $\Lambda$ ($=\Lambda(x)=\Lambda_{K}=\Lambda_K(x)$) be the ring of symmetric functions, namely the set of all symmetric formal power series of bounded degree in variable $x=(x_1,x_2,\dots)$ with coefficients in $K$. We omit the variable $x$ when no confusion arise. Let $\widehat\Lambda$ be its completion, consisting of all symmetric formal power series (with possibly unbounded degree). The Schur functions $s_\lambda$ ($\lambda\in\mathcal{P}$) are a family of homogeneous symmetric functions satisfying $\Lambda=\bigoplus_{\lambda\in\mathcal{P}}K s_\lambda$ and $\widehat\Lambda=\prod_{\lambda\in\mathcal{P}}K s_\lambda$. The {\em Hall inner product} $(\,,)$ is a bilinear form on $\Lambda$ for which the Schur functions form an orthonormal basis, i.e.\,$(s_\lambda,s_\mu)=\delta_{\lambda\mu}$. This is naturally extended to $(\,,)\colon\widehat\Lambda\times\Lambda\longrightarrow K$, whence we can identify $\widehat\Lambda$ with $\Lambda^*$ and $\Lambda$ with $\Lambda^o=\bigoplus_{n\ge 0} \Lambda_n^*$. Here $\Lambda_n$ denotes the homogeneous component of $\Lambda$ with degree $n$. The ring $\Lambda$ is a Hopf algebra with a product $m\colon\Lambda\otimes\Lambda\longrightarrow\Lambda\,;\,f\otimes g\mapsto fg$, a unit $u\colon K\longrightarrow\Lambda\,;\,1\mapsto 1$, a coproduct $\Delta\colon\Lambda=\Lambda(x)\longrightarrow\Lambda(x,y)\hookrightarrow\Lambda(x)\otimes\Lambda(y)\,;\,f(x)\mapsto f(x,y)$, a counit $\epsilon\colon\Lambda\longrightarrow K\,;\,f\mapsto f(0,0,\dots)$, i.e.\,$\epsilon(s_\lambda)=\delta_{\lambda\emptyset}$, and an antipode $S\colon\Lambda\longrightarrow\Lambda\,;\,s_\lambda\mapsto(-1)^{|\lambda|}s_{\lambda'}$. Here $\lambda'$ denotes the transpose of $\lambda\in\mathcal{P}$. The coincidence between the coefficients in the Littlewood-Richardson rules $s_\mu s_\nu=\sum_{\lambda} c^\lambda_{\mu\nu}s_\lambda$ and $\Delta(s_\lambda)=\sum_{\mu,\nu} c^\lambda_{\mu\nu}s_\mu\otimes s_\nu$ implies that $\Lambda$ is self-dual, i.e.\,the Hopf structure on $\Lambda^o$ via \eqref{eq:dual_copro} and \eqref{eq:dual_pro} coincides with one coming from the identification $\Lambda\simeq\Lambda^o$. Note that $\widehat\Lambda\simeq\Lambda^*$ is an algebra but not a coalgebra, since if $f\in\widehat\Lambda$ has unbounded degree then $f(x,y)$ may be unable to be written as a finite sum of $f_1(x)f_2(y)$ for $f_1,f_2\in\widehat\Lambda$. For $F\in\widehat\Lambda$, we have linear maps \Itemize{ \item $(F,-)\colon\Lambda\longrightarrow K\,;\,f\mapsto (F,f)$, and \item $F^\perp\colon\Lambda\longrightarrow\Lambda\,;\,f\mapsto\sum (F,f_1)f_2$ } where we put $\Delta(f)=\sum f_1\otimes f_2$ for $f\in\Lambda$ by the Sweedler notation. By the identification $\widehat\Lambda\simeq\Lambda^*$ this notation is the same as given in Section \ref{sect:Prel::Hopf}. Note that \begin{equation}\label{eq:Fperp} F^\perp = ((F,-)\otimes\mathrm{id})\circ\Delta = (\mathrm{id}\otimes(F,-))\circ\Delta \end{equation} where the second equality is by cocommutativity. We also have \begin{equation}\label{eq:F-} (F,-) = \epsilon\circ F^\perp \end{equation} since $\epsilon\circ F^\perp = \epsilon\circ((F,-)\otimes\mathrm{id})\circ\Delta = ((F,-)\otimes\epsilon)\circ\Delta = (F,-)*\epsilon = (F,-)$. The following lemma is standard: \begin{lemm}\label{theo:FG} For $F,G\in\widehat\Lambda$, \Enumerate{ \item $(FG,-) = (F,-)*(G,-)$ where $*$ denotes the convolution product on $\mathrm{Hom}(\Lambda,K)$. \item $(FG)^\perp=G^\perp\circ F^\perp$ $(=F^\perp\circ G^\perp)$. } \end{lemm} By arguments in Section \ref{sect:Prel::Hopf::comod} we have the following lemmas. \begin{lemm}\label{theo:Fperp} For $F\in\widehat\Lambda$, \Enumerate{ \item $F^\perp\colon\Lambda\longrightarrow\Lambda$ is a $\Lambda$-comodule morphism. \item $(F^\perp)^*=(F\cdot)$. Namely $(FG,f) = (G,F^\perp(f))$ for $G\in\widehat\Lambda$ and $f\in\Lambda$. } \end{lemm} \begin{lemm}\label{theo:cohom_mult} Let $\phi\colon\Lambda\longrightarrow\Lambda$ be a $\Lambda$-comodule morphism. Then the dual map $\phi^*\colon\widehat\Lambda\longrightarrow\widehat\Lambda$ is the multiplication by $\phi^*(1)$. Moreover $\phi = (\phi^*(1))^\perp$. \end{lemm} \subsubsection{Stable Grothendieck polynomials} In \cite[Theorem 3.1]{MR1946917} Buch gave a combinatorial description of the {\em stable Grothendieck polynomial} $G_\lambda$ as a generating function of so-called {\em set-valued tableaux}. We do not review the detail here and just recall some of its properties: $G_\lambda\in\widehat\Lambda$ (although $G_\lambda\notin\Lambda$ if $\lambda\neq\emptyset$), $G_\lambda$ is an infinite linear combination of $\{s_\mu\}_{\mu\in\mathcal{P}}$ whose lowest degree component is $s_\lambda$. Hence $\widehat\Lambda=\prod_{\lambda\in\mathcal{P}}K G_\lambda$, i.e.\,every element in $\widehat\Lambda$ is uniquely written as an infinite linear combination of $G_\lambda$. Moreover the span $\bigoplus_\lambda K G_\lambda$ ($\subset\widehat\Lambda$) is a bialgebra, in particular the expansion of the product $G_\mu G_\nu = \sum_{\lambda} c^{\lambda}_{\mu\nu} G_\lambda$ and the coproduct $\Delta(G_\lambda) = \sum_{\mu,\nu} d^{\lambda}_{\mu\nu} G_\mu\otimes G_\nu$ are finite. \subsubsection{Dual stable Grothendieck polynomials} Next we recall the dual stable Grothendieck polynomial $g_{\lambda/\mu}$. For a skew shape $\lambda/\mu$, a \textit{reverse plane partition} of shape $\lambda/\mu$ is a filling of the boxes in $\lambda/\mu$ with positive integers such that the numbers are weakly increasing in every row and column. \begin{defi}[{\cite{MR2377012}}]\label{defi:gla} For a skew shape $\lambda/\mu$, the {\em dual stable Grothendieck polynomial} $g_{\lambda/\mu}$ is defined by \begin{equation}\label{eq:gdef} g_{\lambda/\mu} = \sum_{T} x^T, \end{equation} summed over reverse plane partitions $T$ of shape $\lambda/\mu$, where $x^T=\prod_{i} x_i^{T(i)}$ where $T(i)$ is the number of columns of $T$ that contain $i$. % \end{defi} When $\mu=\emptyset$ we write $g_\lambda=g_{\lambda/\emptyset}$. It is shown in \cite{MR2377012} that $g_{\lambda/\mu}\in\Lambda$ and $g_\lambda$ has the highest degree component $s_{\lambda}$ and forms a basis of $\Lambda$ that is dual to $G_\lambda$ via the Hall inner product: $(G_\lambda,g_\mu)=\delta_{\lambda\mu}$. Hence the product (resp.\,coproduct) structure constants for $\{G_\lambda\}$ coincide with the coproduct (resp.\,product) structure constants for $\{g_\lambda\}$: $g_\mu g_\nu = \sum_{\lambda} d^{\lambda}_{\mu\nu} g_\lambda$ and $\Delta(g_\lambda) = \sum_{\mu,\nu} c^{\lambda}_{\mu\nu} g_\mu\otimes g_\nu$. After \cite{dualstablesum}, we denote by $I$ the linear map $\Lambda\longrightarrow\Lambda$ defined by $I(g_\lambda)=\sum_{\mu\subset\lambda}g_\mu$. In \cite{dualstablesum} it is shown that $I$ is a ring automorphism, but we need not use it as we rediscover it in \eqref{eq:I=i*id}. The inverse map is given by $I^{-1}(g_\lambda)=\sum_{\text{$\lambda/\mu$: rook strip}} (-1)^{|\lambda/\mu|} g_\mu$. Here $\lambda/\mu$ is called a \emph{rook strip} if any cell of $\lambda/\mu$ is removable corner of $\lambda$. \begin{rema}\label{rema:Gg_geom} We recall geometric interpretations of $G_\lambda$ and $g_\lambda$. Let $\mathrm{Gr}(k,n)$ be the Grassmannian of $k$-dimensional subspaces of $\mathbb{C}^n$, $R=(n-k)^k$ the rectangle of shape $(n-k)\times k$, and $\mc{O}_\lambda$ ($\lambda\subset R$) the structure sheaves of Schubert varieties of $\mathrm{Gr}(k,n)$. % % The $K$-theory $\Kth$, the Grothendieck group of algebraic vector bundles on $\mathrm{Gr}(k,n)$, has a basis $\{[\mc{O}_\lambda]\}_{\lambda\subset R}$, and the surjection $\bigoplus_{\lambda\in\mathcal{P}} \mathbb{Z} G_\lambda \longrightarrow \Kth = \bigoplus_{\lambda\subset R} \mathbb{Z} [\mc{O}_\lambda]$ that maps $G_\lambda$ to $[\mc{O}_\lambda]$ (which is considered as $0$ if $\lambda\not\subset R$) is an algebra homomorphism \cite{MR1946917}. % There is another basis of $\Kth$ consisting of the classes $[\mc{I}_\lambda]$ of ideal sheaves of boundaries of Schubert varieties. In \cite[Section 8]{MR1946917} it is shown that the bases $\{[\mc{O}_\lambda]\}_{\lambda\subset R}$ and $\{[\mc{I}_\lambda]\}_{\lambda\subset R}$ relates to each other by $[\mc{O}_\lambda] = \sum_{\lambda\subset\mu\subset R} [\mc{I}_\mu]$ and that they are dual: more precisely $([\mc{O}_\lambda],[\mc{I}_{\tilde\mu}])=\delta_{\lambda\mu}$ where $\tilde\mu=(n-k-\mu_k,\cdots,n-k-\mu_1)$ is the rotated complement of $\mu\subset R$ and the pairing $(\,,)$ is defined by $(\alpha,\beta) = \rho_*(\alpha\otimes\beta)$ where $\rho_*$ is the pushforward to a point. % The $K$-homology $\Kho$, the Grothendieck group of coherent sheaves, is naturally isomorphic to $\Kth$. % % Lam and Pylyavskyy proved in \cite[Theorem 9.16]{MR2377012} that the surjection $\Lambda=\bigoplus_{\lambda\in\mathcal{P}} \mathbb{Z} g_\lambda \longrightarrow \Kho = \bigoplus_{\mu\subset R} \mathbb{Z} [\mc{I}_\mu]$ that maps $g_\lambda$ to $[\mc{I}_{\tilde\lambda}]$ (which is considered as $0$ if $\lambda\subset R$) identifies the coproduct and product on $\Lambda$ with the pushforwards of the diagonal embedding map and the direct sum map. Since $\mu\subset\lambda\iff\tilde\mu\supset\tilde\lambda$, under this identification we see that $\sum_{\mu\subset\lambda}g_\mu\in\Lambda$ corresponds to $[\mc{O}_{\tilde\lambda}]\in\Kho$. \end{rema} \section{Group-like elements in $\widehat\Lambda$ and automorphisms on $\Lambda$}\label{sect:algebraic} An element $a$ of a coalgebra $(A,\Delta,\epsilon)$ is called \emph{group-like} if $\Delta(a)=a\otimes a$ and $\epsilon(a)=1$. The set of group-like elements in a Hopf algebra forms a group; if $a$ and $b$ are group-like then so are $ab$ and $S(a)=a^{-1}$. If $A$ is a bialgebra and $a\in A$ is group-like then the multiplication map $L_a\colon A\longrightarrow A\,;\,b\mapsto ab$ is a coalgebra morphism since $\Delta\circ L_a(b) = \Delta(ab) = \Delta(a)\Delta(b) = (a\otimes a)\Delta(b) = (L_a\otimes L_a)\circ\Delta(b)$, and hence the dual map $L_a^*\colon A^*\longrightarrow A^*$ is an algebra morphism. Although $\widehat\Lambda$ is not a coalgebra, we shall also say $F\in\widehat\Lambda$ is \emph{group-like} if $F(x,y)=F(x)F(y)$ and its constant term $F(0)$ is $1$. Again, here we mean $F(x)=F(x_1,x_2,\cdots)$, $F(y)=F(y_1,y_2,\cdots)$, $F(x,y)=F(x_1,x_2,\cdots, y_1, y_2,\cdots)$ and $F(0) = F(0,0,\cdots)$. Then these elements satisfy expected properties seen above: \begin{lemm} For group-like elements $F,F'\in\widehat\Lambda$, \Enumerate{ \item $FF'$ is group-like. \item $S(F)=F^{-1}$ is group-like. Here we extend the antipode $S\colon\Lambda\longrightarrow\Lambda\,;\,s_\lambda\mapsto(-1)^{|\lambda|}s_{\lambda'}$ to $\widehat\Lambda\longrightarrow\widehat\Lambda$. } \end{lemm} \begin{proof} \noindent (1) By $FF'(x,y)=F(x,y)F'(x,y)=F(x)F(y)F'(x)F'(y)=FF'(x)\cdot FF'(y)$ and $FF'(0) = F(0)F'(0)=1$. \noindent (2) Write $F=\sum_{\lambda}A_\lambda s_\lambda$ with $A_\lambda\in K$ (possibly an infinite sum). Since $F(x,y) = \sum_{\lambda} A_\lambda s_\lambda(x,y) = \sum_{\lambda} A_\lambda \sum_{\mu,\nu} c^{\lambda}_{\mu\nu} s_\mu(x)s_\nu(y) = \sum_{\mu,\nu} \big(\sum_{\lambda} A_\lambda c^{\lambda}_{\mu\nu}\big) s_\mu(x)s_\nu(y)$ and $F(x)F(y) = \big(\sum_{\mu} A_\mu s_\mu(x)\big) \big(\sum_{\nu} A_\nu s_\nu(y)\big) = \sum_{\mu,\nu} A_\mu A_\nu s_\mu(x)s_\nu(y)$, it follows that \begin{equation}\label{eq:grouplike_cond} \text{$F=\sum_{\lambda}A_\lambda s_\lambda$ is group-like} \iff A_\emptyset = 1 \text{ and } A_\mu A_\nu = \sum_{\lambda} A_\lambda c^{\lambda}_{\mu\nu} \text{ for } \forall \mu, \nu. \end{equation} Let $F':=S(F)=\sum A_\lambda (-1)^{|\lambda|}s_{\lambda'}$. Similarly we can see that $F'$ is group-like if and only if $A_\emptyset=1$ and $A_\mu A_\nu = \sum_{\lambda} A_\lambda c^{\lambda'}_{\mu'\nu'}$ for any $\mu$, $\nu$. Since $c^{\lambda}_{\mu\nu}=c^{\lambda'}_{\mu'\nu'}$ it follows that $F'$ is group-like. Since $\Delta(s_\lambda) = \sum_{\mu,\nu} c^{\lambda}_{\mu\nu} s_\mu\otimes s_\nu$, by applying $m\circ(\mathrm{id}\otimes S)$ we have $\sum_{\mu,\nu} (-1)^{|\nu|} c^{\lambda}_{\mu\nu} s_\mu s_{\nu'} = m\circ(\mathrm{id}\otimes S)\circ\Delta(s_\lambda) = u\circ\epsilon(s_\lambda) = \delta_{\lambda,\emptyset}$. From this and \eqref{eq:grouplike_cond} we have $FF' = \big(\sum_{\mu} A_\mu s_\mu\big) \big(\sum_\nu (-1)^{|\nu|} A_\nu s_{\nu'}\big) = \sum_{\mu,\nu} (-1)^{|\nu|} A_\mu A_\nu s_\mu s_{\nu'} = \sum_{\mu,\nu} \sum_{\lambda} (-1)^{|\nu|} c^{\lambda}_{\mu\nu} A_\lambda s_\mu s_{\nu'} = \sum_{\lambda} A_\lambda \big(\sum_{\mu,\nu} (-1)^{|\nu|} c^{\lambda}_{\mu\nu} s_\mu s_{\nu'}\big) = \sum_{\lambda} A_\lambda \delta_{\lambda,\emptyset} = A_\emptyset = 1$. Hence $F' = F^{-1}$. \end{proof} \begin{lemm}\label{theo:group-like} For $F\in\widehat\Lambda$, the followings are equivalent. \Enumerate{ \item $F\in\widehat\Lambda$ is group-like. \item $(F,-)\colon\Lambda\longrightarrow K$ is an algebra homomorphism. \item $F^\perp\colon\Lambda\longrightarrow\Lambda$ is an algebra automorphism. } \end{lemm} \begin{proof} \noindent\underline{$(1)\iff(2)$} Again we write $F=\sum_{\lambda}A_\lambda s_\lambda$ with $A_\lambda\in K$. (2) is equivalent to $(F,1)=1$ and $(F,s_\mu) (F,s_\nu) = (F,s_\mu s_\nu)$ for any $\mu$, $\nu$, which is equivalent to \eqref{eq:grouplike_cond} since $s_\mu s_\nu = \sum_{\lambda} c^{\lambda}_{\mu\nu} s_\lambda$. \noindent\underline{$(2)\implies(3)$} Since $(F,-)\colon\Lambda\longrightarrow K$ and $\mathrm{id}_\Lambda\colon\Lambda\longrightarrow\Lambda$ are algebra morphisms, $(F,-)\otimes\mathrm{id}_\Lambda\colon\Lambda\otimes\Lambda\longrightarrow\Lambda$ is an algebra morphism. Since $\Delta\colon\Lambda\longrightarrow\Lambda\otimes\Lambda$ is an algebra morphism by the axiom of bialgebras, it follows that $F^\perp=((F,-)\otimes\mathrm{id})\circ\Delta$ is an algebra morphism. By $S(F) = F^{-1}$ and Lemma \ref{theo:FG} (2) we have $S(F)^\perp = (F^\perp)^{-1}$. Hence $F^\perp$ is invertible. \noindent\underline{$(3)\implies(2)$} By \eqref{eq:F-} and the axiom of bialgebras that $\epsilon\colon\Lambda\longrightarrow K$ is an algebra morphism, it follows that $(F,-)=\epsilon\circ F^\perp$ is an algebra morphism. \end{proof} \begin{rema} There is no group-like element in $\Lambda$ except $1$ since $f(x,y)=f(x)f(y)$ implies $\deg(f) = \deg(f) + \deg(f)$. \end{rema} \subsection{Group-like elements $H(t)$, $E(t)$} There are well-known generating functions \begin{equation*} H(t) = \sum_{i\ge 0} t^i h_i, \qquad E(t) = \sum_{i\ge 0} t^i e_i \end{equation*} where $t\in K$. Note that \begin{equation}\label{eq:HE} H(t)E(-t)=1. \end{equation} \begin{lemm}\label{theo:HE:group-like} The elements $H(t), E(t) \in \widehat\Lambda$ are group-like. \end{lemm} \begin{proof} By $\Delta(h_k)=\sum_{i+j=k}h_i\otimes h_j$, we have $H(t)(x,y) = \sum_{k\ge 0} t^k h_k(x,y) = \sum_{k\ge 0} t^k \sum_{i+j=k} h_i(x) h_j(y) = \big(\sum_{i\ge 0} t^i h_i(x)\big) \big(\sum_{j\ge 0} t^j h_j(y)\big) = \big(H(t)(x)\big) \big(H(t)(y)\big)$. The proof for $E(t)$ is similar. \end{proof} Define an algebra homomorphism \[ \phi_t\colon\Lambda\longrightarrow\Lambda\,;\,f(x)\mapsto f(tx), \] where $x=(x_1,x_2,\dots)$ and $tx=(tx_1,tx_2,\dots)$. Clearly $\phi_t$ extends to an algebra homomorphism on $\widehat\Lambda$. If $t$ is invertible then $\phi_t\circ\phi_{t^{-1}}=\mathrm{id}$, whence $\phi_t$ is an automorphism. Since $(\phi_t(s_\lambda),s_\mu) = t^{|\lambda|} \delta_{\lambda\mu} = (s_\lambda, \phi_t(s_\mu))$, the map $\phi_t$ is self-adjoint. Note that \begin{equation}\label{eq:Ht} H(t)=\phi_t(H(1))\quad\text{ and }\quad E(t)=\phi_t(E(1)). \end{equation} \subsection{corresponding algebra morphisms} By Lemma \ref{theo:group-like} and \ref{theo:HE:group-like} we have algebra homomorphisms \[ (H(t),-)\colon\Lambda\longrightarrow \qquad\text{and}\qquad (E(t),-)\colon\Lambda\longrightarrow K \] and algebra automorphisms \[ H(t)^\perp\colon\Lambda\longrightarrow\L \qquad\text{and}\qquad E(t)^\perp\colon\Lambda\longrightarrow\Lambda \] \begin{lemm}\label{theo:(Ht,-)} $(H(t),-) = (H(1),-)\circ\phi_t$ and $(E(t),-) = (E(1),-)\circ\phi_t$. \end{lemm} \begin{proof} Since $\phi_t$ is self-adjoint, for $f\in\Lambda$ it follows that $(H(t),f) = (\phi_t(H(1)),f) = (H(1),\phi_t(f))$ and similarly $(E(t),f) = (E(1),\phi_t(f))$. \end{proof} \begin{lemm}\label{theo:Ht_perp} $\phi_t\circ H(t)^\perp=H(1)^\perp\circ \phi_t$ and $\phi_t\circ E(t)^\perp=E(1)^\perp\circ \phi_t$. Hence $H(t)^\perp=\phi_t^{-1}\circ H(1)^\perp\circ \phi_t$ and $E(t)^\perp=\phi_t^{-1}\circ E(1)^\perp\circ \phi_t$ when $t$ is invertible. \end{lemm} \begin{proof} Since $\phi_t$ is a self-adjoint algebra automorphism, for $F,G\in\widehat\Lambda$ and $f\in\Lambda$ we have $(F, \phi_t\circ(\phi_t(G))^\perp(f)) = (\phi_t(F), (\phi_t(G))^\perp(f)) = (\phi_t(G)\phi_t(F), f) = (\phi_t(GF), f) = (G F, \phi_t(f)) = (F, G^\perp\circ\phi_t(f))$, whence $\phi_t\circ(\phi_t(G))^\perp = G^\perp\circ\phi_t$. In particular $\phi_t\circ H(t)^\perp = H(1)^\perp\circ\phi_t$ and $\phi_t\circ E(t)^\perp = E(1)^\perp\circ\phi_t$ by \eqref{eq:Ht}. \end{proof} \begin{lemm}\label{theo:HE_Hall_perp} \Enumerate{ \item $(H(t),-)*(E(-t),-) = \epsilon$, where $\epsilon\colon\Lambda\longrightarrow K$ is the counit. \item $H(t)^\perp \circ E(-t)^\perp = \mathrm{id}_\Lambda$. } \end{lemm} \begin{proof} \noindent (1) By \eqref{eq:HE}, Lemma \ref{theo:FG}, and the fact that the counit is the identity element with respect to the convolution product $*$. (2) By \eqref{eq:HE} and Lemma \ref{theo:FG}. \end{proof} \section{Proof of $I=H(1)^\perp$ and $I^{-1}=E(-1)^\perp$}\label{sect:I} Define a ring homomorphism $i\colon \Lambda\longrightarrow K$ by the substitution \[ i\colon f\mapsto f(1) = f(1,0,0,\dots). \] We shall show in Theorem \ref{theo:H(1)_overall} that $i=(H(1),-)$ and $I=(i\otimes\mathrm{id})\circ\Delta=H(1)^\perp$. We start with the following observation. \begin{prop}\label{theo:i} $i(g_{\lambda/\mu})=1$ for any skew shape $\lambda/\mu$, and in particular $i(g_{\lambda})=1$ for any $\lambda\in\mathcal{P}$. \end{prop} \begin{proof} By the definition \eqref{eq:gdef} of $g_{\lambda/\mu}$, $i(g_{\lambda/\mu})$ is the number of reverse plane partitions on $\lambda/\mu$ filled with $1$. Clearly there is exactly one such filling. \end{proof} \begin{rema}\label{theo:i:hep} It is straightforward to check that $i(h_k)=1$ for any $k\ge 0$, $i(e_k)=0$ for any $k\ge 2$, and $i(p_k)=1$ for any $k\ge 1$. \end{rema} Since $\{g_\lambda\}_\lambda$ is a basis of $\Lambda$ and $\{G_\lambda\}_\lambda$ is their dual, from Proposition \ref{theo:i} we have \begin{equation}\label{eq:i=(G,-)} i=\Big(\sum_{\lambda\in\mathcal{P}}G_\lambda, -\Big). \end{equation} Another corollary of the proposition above is formulas on the structure constants in $g_\mu g_\nu = \sum_\lambda d^{\lambda}_{\mu\nu} g_\lambda$ and $g_{\lambda/\mu} = \sum_{\nu} c^{\lambda}_{\mu\nu} g_\nu$: \begin{coro}\label{theo:cd} \Enumerate{ \item For any $\mu,\nu\in\mathcal{P}$ we have $\sum_\lambda d^{\lambda}_{\mu\nu} = 1$. \item For any $\lambda,\mu\in\mathcal{P}$ we have $\sum_\nu c^{\lambda}_{\mu\nu} = 1$. } \end{coro} \subsection{Description of $I$ as a substitution}\label{sect:I:f(1,x)} Next we give another description of the map $I\colon g_\lambda \mapsto \sum_{\mu\subset\lambda}g_\mu$. For a skew shape $\lambda/\mu$ and a totally ordered set $X$ called {\em alphabets} (most commonly $\{1,2,3,\dots\}$), we shall denote by $\mathrm{RPP}(\lambda/\mu,X)$ the set of reverse plane partition of shape $\lambda/\mu$ where each box is filled with an element of $X$. The expression \eqref{eq:gdef} of $g_{\lambda/\mu}$ as a generating function of reverse plane partitions implies \begin{equation}\label{eq:copro} \Delta(g_{\lambda/\mu}) = \sum_{\mu\subset\nu\subset\lambda} g_{\lambda/\nu} \otimes g_{\nu/\mu}, \end{equation} since we have a natural bijection between $\mathrm{RPP}(\lambda/\mu, \{1,2,\cdots,1',2',\dots\})$ and $\bigsqcup_{\mu\subset\nu\subset\lambda} \mathrm{RPP}(\nu/\mu, \{1,2,\cdots\})\times\mathrm{RPP}(\lambda/\nu, \{1',2',\cdots\})$ where $1<2<\cdots<1'<2'<\cdots$. Since $(i\otimes\mathrm{id})\circ\Delta\colon f(x)\mapsto f(1,x)$ where $f(x)=f(x_1,x_2,\dots)$ and $f(1,x)=f(1,x_1,x_2,\dots)$, by applying $i\otimes\mathrm{id}$ to \eqref{eq:copro} we have \begin{equation}\label{eq:g1x} g_{\lambda/\mu}(1,x) = \sum_{\mu\subset\nu\subset\lambda} i(g_{\lambda/\nu}) g_{\nu/\mu}(x) = \sum_{\mu\subset\nu\subset\lambda} g_{\nu/\mu}(x), \end{equation} where the last equation is by Proposition \ref{theo:i}. Similarly, by applying $(\mathrm{id}\otimes i)\circ\Delta$ to \eqref{eq:copro} we have \begin{equation}\label{eq:gx1} g_{\lambda/\mu}(x,1) = \sum_{\mu\subset\nu\subset\lambda} g_{\lambda/\nu}(x) i(g_{\nu/\mu}) = \sum_{\mu\subset\nu\subset\lambda} g_{\lambda/\nu}(x). \end{equation} Setting $\mu=\emptyset$ in \eqref{eq:g1x}, for any $\lambda\in\mathcal{P}$ we have \[ g_{\lambda}(1,x) = \sum_{\nu\subset\lambda} g_{\nu}(x) \ \Big(=I(g_\lambda(x))\Big). \] Since $\{g_\lambda\}_\lambda$ form a basis of $\Lambda$, this implies \begin{prop}\label{theo:I(f)} For any $f\in\Lambda$ we have \begin{equation}\label{eq:I=i*id} I(f(x)) = f(1,x), \end{equation} or equivalently \begin{equation}\label{eq:I_i} I=(i\otimes\mathrm{id})\circ\Delta. \end{equation} \end{prop} In particular \eqref{eq:I=i*id} recovers that $I\colon\Lambda\longrightarrow\Lambda$ is a ring homomorphism, and the bijectivity follows from the fact that the transition matrix between $g_\lambda$ and $I(g_\lambda)=\sum_{\mu\subset\lambda}g_\mu$ is unitriangular. Moreover, \eqref{eq:g1x}, \eqref{eq:gx1} and \eqref{eq:I=i*id} imply that for any skew shape $\lambda/\mu$ \begin{equation}\label{eq:I_skew} I(g_{\lambda/\mu}) = \sum_{\mu\subset\nu\subset\lambda} g_{\nu/\mu} = \sum_{\mu\subset\nu\subset\lambda} g_{\lambda/\nu}. \end{equation} Besides, by \eqref{eq:Fperp}, \eqref{eq:I_i} and Corollary \ref{eq:i=(G,-)} we have \begin{equation}\label{eq:Iasperp} I=\Big(\sum_{\lambda\in\mathcal{P}} G_\lambda\Big)^\perp. \end{equation} By \eqref{eq:Iasperp} and Lemma \ref{theo:Fperp}, $I$ is a $\Lambda$-comodule morphism. Since $I$ is bijective we have \begin{prop}\label{theo:copro} $I\colon\Lambda\longrightarrow\Lambda$ is a $\Lambda$-comodule automorphism. \end{prop} \begin{rema} \noindent (1) We can prove Proposition \ref{theo:copro} by directly showing the equality \begin{equation}\label{eq:DeltaI} \Delta\circ I = (I\otimes \mathrm{id}_{\Lambda})\circ\Delta = (\mathrm{id}_{\Lambda}\otimes I)\circ\Delta. \end{equation} Indeed, by \eqref{eq:I=i*id} each side of \eqref{eq:DeltaI} maps $f(x)\in\Lambda(x)$ to $f(1,x,y)\in\Lambda(x)\otimes\Lambda(y)$. \noindent (2) Since $(G_\mu,g_\nu)=\delta_{\mu\nu}$ and $\Delta(g_\lambda) = \sum_{\nu} g_{\nu}\otimes g_{\lambda/\nu}$ (by \eqref{eq:copro}), we have $G_\mu^\perp(g_\lambda) = g_{\lambda/\mu}$. Here we consider $g_{\lambda/\mu}=0$ if $\mu\not\subset\lambda$. Since $F^\perp$ (for $F\in\widehat\Lambda$) commute each other, $I$ commutes with $G_\mu^\perp$. Hence, by applying $G_\mu^\perp$ to the equation $I(g_\lambda)=\sum_{\nu\subset\lambda}g_\nu$ we have \[ I(g_{\lambda/\mu}) = I(G_\mu^\perp(g_\lambda)) = G_\mu^\perp(I(g_\lambda)) = G_\mu^\perp\Big(\sum_{\nu\subset\lambda} g_\nu\Big) = \sum_{\nu\subset\lambda} G_\mu^\perp(g_\nu) = \sum_{\nu\subset\lambda} g_{\nu/\mu}, \] re-proving the first equation in \eqref{eq:I_skew}. % Similarly, by applying $I=\sum_{\nu}G_\nu^\perp$ to $g_\lambda$ we get a special case of the second equation of \eqref{eq:I_skew} \[ I(g_\lambda) = \sum_{\nu}G_\nu^\perp(g_\lambda) = \sum_{\nu} g_{\lambda/\nu}. \] \end{rema} \subsection{Dual map $I^*$ and proof of $I^*(1)=H(1)$}\label{sect:I:H1_perp} By Lemma \ref{theo:cohom_mult} and Proposition \ref{theo:copro} \Itemize{ \item $I^*$ is the multiplication by $I^*(1)$, \item $I=I^*(1)^\perp$, } and similarly \Itemize{ \item $(I^{-1})^*$ is the multiplication by $(I^{-1})^*(1)$, \item $I^{-1}=(I^{-1})^*(1)^\perp$. } Since $(G_\lambda,g_\mu)=\delta_{\lambda\mu}$, the maps $I\colon\Lambda\longrightarrow\Lambda\,;\,g_\lambda\mapsto \sum_{\mu\subset\lambda}g_\mu$ and $I^{-1}\colon g_\lambda\mapsto \sum_{\text{$\lambda/\mu$: rook strip}} (-1)^{|\lambda/\mu|} g_\mu$ induce the dual maps \begin{align} I^*&\colon\widehat\Lambda\longrightarrow\widehat\Lambda\,;\,G_\lambda\mapsto \sum_{\lambda\subset\mu} G_\mu \label{eq:I*} \\ (I^{-1})^*&\colon\widehat\Lambda\longrightarrow\widehat\Lambda\,;\,G_\lambda\mapsto\sum_{\text{$\mu/\lambda$: rook strip}} (-1)^{|\mu/\lambda|} G_\mu. \label{eq:I*-1} \end{align} By \eqref{eq:I*} we have $I^*(1)=\sum_{\mu\in\mathcal{P}}G_\mu$. Similarly $(I^{-1})^*(1) = 1-G_1$ by \eqref{eq:I*-1}. In particular $(1-G_1)\sum_{\mu}G_\mu = 1$. Since $G_{1}=e_1-e_2+e_3-\cdots$ it follows that $E(-1)=1-G_1$, whence \begin{equation}\label{eq:I-1:detail} (I^{-1})^* = (E(-1)\cdot) \qquad\text{and}\qquad I^{-1} = E(-1)^\perp. \end{equation} Hence, it follows from $H(1)E(-1)=1$ that \begin{equation}\label{eq:I:detail} I^* = (H(1)\cdot), \qquad \sum_{\lambda\in\mathcal{P}}G_\lambda \,\big(= I^*(1)\,\big)\,= H(1), \qquad\text{and}\qquad I = H(1)^\perp. \end{equation} We have proved the following theorems so far: \begin{theo}\label{theo:H(1)_overall} \Enumerate{ \item $H(1)=\sum_{i\ge 0} h_i = \sum_{\lambda\in\mathcal{P}} G_\lambda$. \item \label{item:H(1):i} The ring homomorphism $i\colon\Lambda\longrightarrow K$ defined by $i(f) = f(1,0,0,\dots)$ satisfies \Enumerate{ \item $i = (H(1),-)$. \item $i(g_{\lambda/\mu}) = 1$ for any skew shape $\lambda/\mu$. } \item The linear map $I\colon\Lambda\longrightarrow\Lambda$ defined by $I(g_\lambda) = \sum_{\mu\subset\lambda} g_\mu$ is a ring automorphism that satisfies \[I = H(1)^\perp = \sum_{\lambda\in\mathcal{P}} G_\lambda^\perp = \sum_{i\ge 0} h_i^\perp = \big(f(x_1,x_2,\dots)\mapsto f(1,x_1,x_2,\dots)\big) \] and \[ I(g_{\lambda/\mu}) = \sum_{\mu\subset\nu\subset\lambda} g_{\nu/\mu} = \sum_{\mu\subset\nu\subset\lambda} g_{\lambda/\nu} \] for any skew shape $\lambda/\mu$. In particular for any $\lambda\in\mathcal{P}$ \[ I(g_{\lambda}) = \sum_{\nu\subset\lambda} g_{\nu} = \sum_{\nu\subset\lambda} g_{\lambda/\nu}. \] } \end{theo} \begin{theo}\label{theo:E1:overall} \Enumerate{ \item $E(-1) = \sum_{i\ge 0}(-1)^i e_i = 1-G_1$. \item The map $I^{-1}$ satisfies \begin{equation* I^{-1} = E(-1)^\perp = \mathrm{id}_\Lambda - G_{1}^\perp = \sum_{i\ge 0}(-1)^i e_i^\perp. \end{equation*} } \end{theo} We take a closer look at $(E(-1),-)$ and $E(-1)^\perp$ in Section \ref{sect:E(-1)}. In Example \ref{exam:I_skew} we see an example for the identity \eqref{eq:I_skew}. \begin{rema} By \eqref{eq:I*}, \eqref{eq:I*-1}, \eqref{eq:I-1:detail} and \eqref{eq:I:detail} we rediscover the formulas \begin{align} \bigg(I^*(G_\lambda)=\bigg)\quad \big(\sum_{\mu\in\mathcal{P}}G_\mu\big) G_\lambda &= \sum_{\lambda\subset\mu} G_\mu \label{eq:I*(1)} \\ \bigg((I^{-1})^*(G_\lambda)=\bigg)\quad (1 - G_{1}) G_\lambda &= \sum_{\text{$\mu/\lambda$: rook strip}} (-1)^{|\mu/\lambda|} G_\mu, \label{eq:D*(1)} \end{align} which appeared in \cite[Section 8]{MR1946917}. These two are generalized in \eqref{eq:Ht:G} and \eqref{eq:Et:G}. \end{rema} \subsection{Description of $(E(-1),-)$ and $I^{-1} = E(-1)^\perp$} \label{sect:E(-1)} \begin{prop}\label{theo:(E1,-)_detail} The ring homomorphism $(E(-1),-)\colon\Lambda\longrightarrow K$ satisfies $(E(-1),g_{\lambda/\mu}) = (-1)^{|\lambda/\mu|}$ if $\lambda/\mu$ is a rook strip, and $(E(-1),g_{\lambda/\mu}) = 0$ otherwise. In particular $(E(-1),g_{\emptyset}) = 1$, $(E(-1),g_{(1)}) = -1$, and $(E(-1),g_{\lambda}) = 0$ for any $\lambda\in\mathcal{P}$ with $|\lambda|>1$. \end{prop} \begin{proof} Since $(H(1),-)*(E(-1),-)=\epsilon$ (Lemma \ref{theo:HE_Hall_perp}), applying $(H(1),-)\otimes(E(-1),-)$ to \eqref{eq:copro} we have for any skew shape $\lambda/\mu$ that \[ \delta_{\lambda\mu} = \epsilon(g_{\lambda/\mu}) = \sum_{\mu\subset\nu\subset\lambda} (H(1),g_{\lambda/\nu}) (E(-1),g_{\nu/\mu}) = \sum_{\mu\subset\nu\subset\lambda} (E(-1),g_{\nu/\mu}). \] Here the last equality is by Theorem \ref{theo:H(1)_overall} \ref{item:H(1):i}. Hence $(E(-1),g_{\nu/\mu})$ is equal to the value of the M\"obius function $\mu_{\mathcal{P}}(\mu,\nu)$, which is $(-1)^{|\nu/\mu|}$ if $\nu/\mu$ is a rook strip and $0$ otherwise. \end{proof} \begin{rema}\label{theo:E1:hep} It is easy to check that $(E(-1), e_i) = (-1)^i$ for $i\ge 0$, $(E(-1), h_i) = 0$ for $i \ge 2$, and $(E(-1), p_i) = -1$ for $i\ge 1$, from Remark \ref{theo:i:hep} and the fact that $(H(1),-)*(E(-1),-)=\epsilon$. \end{rema} \begin{rema} Unlike Proposition \ref{theo:H(1)_overall} \ref{item:H(1):i}, there is no $a_1,a_2,\dots\in\mathbb{R}$ such that $(E(-1),f) = f(a_1,a_2,\dots)$, since such numbers should satisfy $-1 = (E(-1),p_2) = a_1^2+a_2^2+\cdots$. \end{rema} Now we give a description of $E(-1)^\perp=I^{-1}$. \begin{prop} The ring automorphism $E(-1)^\perp=I^{-1}\colon\Lambda\longrightarrow\Lambda$ satisfies \[ I^{-1}(g_{\lambda/\mu}) = \sum_{\substack{\mu\subset\nu\subset\lambda \\ \lambda/\nu\text{: rook strip}}} (-1)^{|\lambda/\nu|} g_{\nu/\mu} = \sum_{\substack{\mu\subset\nu\subset\lambda \\ \nu/\mu\text{: rook strip}}} (-1)^{|\nu/\mu|} g_{\lambda/\mu}. \] In particular, when $\mu=\emptyset$ we have \begin{equation}\label{eq:Iinv:g} I^{-1}(g_{\lambda}) = \sum_{\lambda/\nu\text{: rook strip}} (-1)^{|\lambda/\nu|} g_{\nu} = \begin{cases} g_{\lambda} - g_{\lambda/(1)} & \text{if $\lambda\neq\emptyset$}, \\ 1 & \text{if $\lambda=\emptyset$}. \end{cases} \end{equation} \end{prop} \begin{proof} By \eqref{eq:Fperp}, \eqref{eq:copro} and Proposition \ref{theo:(E1,-)_detail}. \end{proof} \section{Description of $H(t)$, $(H(t),-)$ and $H(t)^\perp$}\label{sect:Ht} Now we give a description for the map $(H(t),-)$. Let $c(\lambda/\mu)$ denote the number of columns in the skew shape $\lambda/\mu$. \begin{prop}\label{theo:(Ht,-)_detail} The algebra homomorphism $(H(t),-)\colon\Lambda\longrightarrow K$ satisfies \begin{flalign*} &\mathrm{(1)} & (H(t), f) &= f(t,0,0,\cdots) & &\text{for any $f\in\Lambda$}, \\ &\mathrm{(2)} & (H(t),g_{\lambda/\mu}) &= t^{c(\lambda/\mu)} & &\text{for any skew shape $\lambda/\mu$,}\\ &\phantom{\mathrm{(b)}}\text{and in particular} & (H(t),g_{\lambda}) &= t^{c(\lambda)} & &\text{for any $\lambda\in\mathcal{P}$.} \end{flalign*} \end{prop} \begin{proof} (1) follows from Lemma \ref{theo:(Ht,-)} and Theorem \ref{theo:H(1)_overall} \ref{item:H(1):i}. \noindent (2) As we argued in the proof of Proposition \ref{theo:i} (2), there is exactly one reverse plane partition of shape $\lambda/\mu$ filled with the alphabet $1$, whose weight is $x_1^{c(\lambda/\mu)}$. Hence $(H(t),g_{\lambda/\mu}) = g_{\lambda/\mu}(t,0,0,\cdots) = t^{c(\lambda/\mu)}$. \end{proof} \begin{rema} It is straightforward to check from $(H(t),f) = f(t,0,0,\dots)$ that $(H(t), h_i) = t^i$ for $i\ge 0$, $(H(t), e_i) = 0$ for $i \ge 2$, and $(H(t), p_i) = t^i$ for $i\ge 1$. \end{rema} With the proposition above, Corollary \ref{theo:cd} is generalized as follows: \begin{coro} \Enumerate{ \item For any $\mu,\nu\in\mathcal{P}$, we have $t^{c(\mu)+c(\nu)} = \sum_\lambda d^{\lambda}_{\mu\nu} t^{c(\lambda)}$. \item For any $\lambda,\mu\in\mathcal{P}$, we have $t^{c(\lambda/\mu)} = \sum_\nu c^{\lambda}_{\mu\nu} t^{c(\nu)}$. } \end{coro} \begin{prop}\label{theo:Ht_perp:detail} The algebra automorphism $H(t)^\perp\colon\Lambda\longrightarrow\Lambda$ satisfies \begin{flalign*} &\mathrm{(1)} & H(t)^\perp(f(x_1,x_2,\dots)) &= f(t,x_1,x_2,\dots) & &\text{for any $f\in\Lambda$}, \\ &\mathrm{(2)} & H(t)^\perp(g_{\lambda/\mu}) &= \sum_{\mu\subset\nu\subset\lambda} t^{c(\lambda/\nu)} g_{\nu/\mu} = \sum_{\mu\subset\nu\subset\lambda} t^{c(\nu/\mu)} g_{\lambda/\nu} & &\text{for any $\mu\subset\lambda$,} \\ &\phantom{\mathrm{(2)}}\text{and in particular} & H(t)^\perp(g_{\lambda}) &= \sum_{\nu\subset\lambda} t^{c(\lambda/\nu)} g_{\nu} = \sum_{\nu\subset\lambda} t^{c(\nu)} g_{\lambda/\nu} & &\text{for any $\lambda\in\mathcal{P}$.} \end{flalign*} \end{prop} \begin{proof Since $H(t)^\perp = ((H(t),-)\otimes\mathrm{id})\circ\Delta = (\mathrm{id}\otimes(H(t),-))\circ\Delta$, (1) follows from Proposition \ref{theo:(Ht,-)_detail} (1), and (2) is obtained by applying $(H(t),-)\otimes\mathrm{id})$ and $\mathrm{id}\otimes(H(t),-))$ to \eqref{eq:copro} and using Proposition \ref{theo:(Ht,-)_detail} (2). \end{proof} We also give the expansion of $H(t)$ using $G_\lambda$. \begin{prop}\label{theo:Ht} The element $H(t)=\sum_{i\ge 0}t^i h_i\in\widehat\Lambda$ satisfies \begin{align*} H(t) G_\lambda &= \sum_{\lambda\subset\mu} t^{c(\mu/\lambda)} G_\mu. \end{align*} In particular, setting $\lambda=\emptyset$ we have \begin{equation}\label{eq:Ht:G} H(t) = \sum_{\mu\in\mathcal{P}} t^{c(\mu)} G_\mu \qquad\text{and hence}\qquad \Big(\sum_{\mu\in\mathcal{P}} t^{c(\mu)} G_\mu\Big) G_\lambda = \sum_{\lambda\subset\mu} t^{c(\mu/\lambda)} G_\mu. \end{equation} \end{prop} \begin{proof} In Proposition \ref{theo:Ht_perp:detail} (2) it is shown that $H(t)^\perp(g_{\lambda}) = \sum_{\mu\subset\lambda} t^{c(\lambda/\mu)} g_{\mu}$, from which we have $(H(t)^\perp)^*(G_\lambda) = \sum_{\lambda\subset\mu} t^{c(\mu/\lambda)} G_\mu$. By Lemma \ref{theo:Fperp} (2) and \ref{theo:cohom_mult} we see that $(H(t)^\perp)^*$ is the multiplication by $(H(t)^\perp)^*(1)=H(t)$. Hence the proof is done. \end{proof} \section{Description of $E(t)$, $(E(t),-)$ and $E(t)^\perp$}\label{sect:Et} \begin{prop}\label{theo:(Et,-)_detail} The ring homomorphism $(E(t),-)\colon\Lambda\longrightarrow K$ satisfies \[ (E(t),g_{\lambda/\mu}) = \begin{cases} t^{c(\lambda/\mu)} (t+1)^{|\lambda/\mu|-c(\lambda/\mu)} & \text{if $\lambda/\mu$ is a vertical strip}, \\ 0 & \text{otherwise} \end{cases} \] for any skew shape $\lambda/\mu$. In particular, for any $\lambda\in\mathcal{P}$, \[ (E(t),g_{\lambda}) = \begin{cases} 1 & \text{if $\lambda=\emptyset$}, \\ t (t+1)^{n-1} & \text{if $\lambda=(1^n)$ $(n\ge 1)$}, \\ 0 & \text{otherwise}. \end{cases} \] \end{prop} \begin{rema} By Lemma \ref{theo:(Ht,-)} and Proposition \ref{theo:E1:hep} it follows that $(E(t), e_i) = t^i$ for $i\ge 0$, $(E(t), h_i) = 0$ for $i \ge 2$, and $(E(t), p_i) = (-1)^{i-1} t^i$ for $i\ge 1$. \end{rema} Before proving Proposition \ref{theo:(Et,-)_detail}, we give as its corollaries descriptions for the element $E(t)$ and the map $E(t)^\perp$. \begin{prop}\label{theo:Et_perp:detail} The ring automorphism $E(t)^\perp\colon\Lambda\longrightarrow\Lambda$ satisfies \begin{align*} E(t)^\perp(g_{\lambda/\mu}) &= \sum_{\substack{\mu\subset\nu\subset\lambda \\ \lambda/\nu\text{: vertical strip}}} t^{c(\lambda/\nu)} (t+1)^{|\lambda/\nu|-c(\lambda/\nu)} g_{\nu/\mu} \\ &= \sum_{\substack{\mu\subset\nu\subset\lambda \\ \nu/\mu\text{: vertical strip}}} t^{c(\nu/\mu)} (t+1)^{|\nu/\mu|-c(\nu/\mu)} g_{\lambda/\nu} \end{align*} for any skew shape $\lambda/\mu$. % In particular, for any $\lambda\in\mathcal{P}$, \begin{align} E(t)^\perp(g_{\lambda}) &= \sum_{\substack{\nu\subset\lambda \\ \lambda/\nu\text{: vertical strip}}} t^{c(\lambda/\nu)} (t+1)^{|\lambda/\nu|-c(\lambda/\nu)} g_{\nu} \label{eq:Et_perp:g} \\ &= \begin{cases} g_\lambda + \sum_{k=1}^{l(\lambda)} t (t+1)^{k-1} g_{\lambda/(1^k)} & \text{if $\lambda\neq\emptyset$}, \\ g_\emptyset & \text{if $\lambda=\emptyset$}. \end{cases} \notag \end{align} \end{prop} \begin{proof} Proved similarly to Proposition \ref{theo:Ht_perp:detail} (2), with Proposition \ref{theo:(Et,-)_detail} in hand. \end{proof} \begin{prop}\label{theo:Et} The element $E(t)=\sum_{i\ge 0}t^i e_i\in\widehat\Lambda$ satisfies \begin{align*} E(t) G_\lambda &= \sum_{\mu/\lambda\text{: vertical strip}} t^{c(\mu/\lambda)} (t+1)^{|\mu/\lambda|-c(\mu/\lambda)} G_\mu. \end{align*} In particular, setting $\lambda=\emptyset$ we have \[ E(t) = 1 + \sum_{n\ge 1} t(t+1)^{n-1} G_{(1^n)}, \] and hence \begin{equation}\label{eq:Et:G} \Big(1 + \sum_{n\ge 1} t(t+1)^{n-1} G_{(1^n)}\Big) G_\lambda = \sum_{\mu/\lambda\text{: vertical strip}} t^{c(\mu/\lambda)} (t+1)^{|\mu/\lambda|-c(\mu/\lambda)} G_\mu. \end{equation} \end{prop} \begin{proof By \eqref{eq:Et_perp:g} we have \[ (E(t)^\perp)^*(G_{\nu}) = \sum_{\lambda/\nu\text{: vertical strip}} t^{c(\lambda/\nu)} (t+1)^{|\lambda/\nu|-c(\lambda/\nu)} G_{\lambda}. \] By Lemma \ref{theo:Fperp} (2) and \ref{theo:cohom_mult} we see that $(E(t)^\perp)^*$ is the multiplication by $(E(t)^\perp)^*(1)=E(t)$. Hence the proof is done. \end{proof} \subsection{Proof of Proposition \ref{theo:(Et,-)_detail}} We recall the \emph{incidence algebras} (see \cite[Chapter 3.6]{MR2868112} for details). Let $\mathrm{Int}(\mathcal{P}) = \{(\mu,\lambda)\in\mathcal{P}\times\mathcal{P}\mid\mu\subset\lambda\}$, consisting of all comparable (ordered) pairs in $\mathcal{P}$ (or equivalently all skew shapes, by identifying $(\mu,\lambda)$ with $\lambda/\mu$). The \emph{incidence algebra} $I(\mathcal{P})=I(\mathcal{P},K)$ is the algebra of all functions $f\colon\mathrm{Int}(\mathcal{P})\longrightarrow K$ where multiplication is defined by the convolution \begin{equation}\label{eq:IP_fg} (fg)(\mu,\lambda) = \sum_{\mu\subset\nu\subset\lambda} f(\mu,\nu) g(\nu,\lambda). \end{equation} % Then $I(\mathcal{P},K)$ is an associative algebra with two-sided identity $\delta := ((\mu,\lambda)\mapsto \delta_{\mu\lambda})$. A linear function $f\colon\Lambda\longrightarrow K$ can be considered as an element of $I(\mathcal{P},K)$ by setting $f(\mu,\lambda)=f(g_{\lambda/\mu})$. Then the convolution product $*$ on $\mathrm{Hom}(\Lambda,K)$ coincides with the multiplication on $I(\mathcal{P})$ due to \eqref{eq:copro}, i.e.\,this inclusion $\mathrm{Hom}(\Lambda,K)\longrightarrow I(\mathcal{P})$ is as algebras.% \footnote{ Since $\Delta(s_{\lambda/\mu}) = \sum_{\nu}s_{\lambda/\nu}\otimes s_{\nu/\mu}$, by setting $f(\mu,\lambda)=f(s_{\lambda/\mu})$ we can obtain another algebra inclusion, although we do not need it. } Note that the counit $\epsilon\in\mathrm{Hom}(\Lambda,K)$ is mapped to $\delta\in I(\mathcal{P})$. \begin{proof}[Proof of Proposition \ref{theo:(Et,-)_detail}] Define $i_t,j_t\in I(\mathcal{P})$ by \[ i_t(\mu,\lambda) = t^{c(\lambda/\mu)} \] and \[ j_t (\mu,\lambda) = \begin{cases} (-1)^{|\lambda/\mu|} t^{c(\lambda/\mu)} (t-1)^{|\lambda/\mu|-c(\lambda/\mu)} & \text{if $\lambda/\mu$ is a vertical strip}, \\ 0 & \text{otherwise}. \end{cases} \] By Proposition \ref{theo:(Ht,-)_detail} (2) $(H(t),-)\in\mathrm{Hom}(\Lambda,K)$ corresponds to $i_t\in I(\mathcal{P})$. Since $(H(t),-)*(E(-t),-)=\epsilon$, it suffices to show that $i_t j_t = \delta$ in order to prove that $(E(-t),-)$ corresponds to $j_t$, whence Proposition \ref{theo:(Et,-)_detail} follows by replacing $t$ with $-t$. By the definitions of $i_t$ and $j_t$ and \eqref{eq:IP_fg} \begin{equation}\label{eq:itjt} (i_t j_t)(\mu,\lambda) = \sum_{\substack{\mu\subset\nu\subset\lambda \\ \lambda/\nu\text{: vertical strip}}} t^{c(\nu/\mu)} (-1)^{|\lambda/\nu|} t^{c(\lambda/\nu)} (t-1)^{|\lambda/\nu|-c(\lambda/\nu)}. \end{equation} We need to show that the value of the right-hand side of \eqref{eq:itjt} is $\delta_{\mu\lambda}$. It is clear that if $\mu=\lambda$ then the value of \eqref{eq:itjt} is $1$. % Assume $\mu\subsetneq\lambda$. Since the value of \eqref{eq:itjt} is invariant under removal of empty rows in the skew shape $\lambda/\mu$, we can assume there is a box in the first row of $\lambda/\mu$, i.e.\,$\lambda_1 > \mu_1$. % Let $p$ be the index of the rightmost column of $\lambda$, i.e.\,$\lambda_1=p$. Note that $\lambda'_p>0=\mu'_p$. Let $\tilde\lambda$ be the partition obtained by removing the $p$-th (rightmost) column of $\lambda$, i.e.\,$\tilde\lambda_i = \min(\lambda_i,p-1)$. % (Note: in the figure below and hereafter we display Young diagrams in the French notation.) \[ \tikzitem[0.4]{ \draw (5,0) -| (9,3) -| (7,5) -| (4,7) -| (0,4) -| (2,2) -| (5,0); \draw[thick, loosely dotted] (5,0) -| (0,4); \draw (9,0) to [out=70, in=-70] node [right]{$\lambda'_p$} +(0,3) ; \draw [thick, dotted] (8.7,-0.2) -- +(0,-1) node [below]{$p=\lambda_1$}; \node at (4.5,3.5) {$\lambda/\mu$}; \node at (2.5,1) {$\mu$}; } \] For a vertical strip $\lambda/\nu$ with $\mu\subset\nu$, by removing the $p$-th column of $\nu$ (let $\tilde\nu$ denote the resulting shape) we get a vertical strip $\tilde\lambda/\tilde\nu$ that satisfies $\mu\subset\tilde\nu$ and $\tilde\nu'_{p-1}\ge \lambda'_p$. Conversely, for any vertical strip $\tilde\lambda/\kappa$ with $\mu\subset\kappa$ and $\kappa'_{p-1}\ge\lambda'_p$ and any integer $0\le i\le\lambda'_p$, by adding $i$ boxes in the $p$-th column of $\kappa$ we get the shape $\kappa+(1^i)$, for which $\lambda/(\kappa+(1^i))$ is a vertical strip. Therefore we have a bijection \begin{equation}\label{eq:vs_set} \{\nu\mid\mu\subset\nu\subset\lambda,\ \lambda/\nu\text{: vertical strip}\} \simeq \{\kappa\mid\mu\subset\kappa\subset\tilde\lambda, \ \tilde\lambda/\kappa\text{: vertical strip}, \ \kappa'_{p-1}\ge\lambda'_p \} \times \{0,1,\dots,\lambda'_p\} \end{equation} in which $\nu$ corresponds to $(\tilde\nu, \nu'_p)$, where $\tilde\nu$ is $\nu$ with its $p$-th column removed. % For $\nu$ in the left-hand side of \eqref{eq:vs_set}, it is easy to see that \begin{align*} c(\nu/\mu) &= c(\tilde\nu/\mu) + \DE{\nu'_p>0}, \\ |\lambda/\nu| &= |\tilde\lambda/\tilde\nu| + \lambda'_p-\nu'_p, \\ c(\lambda/\nu) &= c(\tilde\lambda/\tilde\nu) + \DE{\nu'_p < \lambda'_p}, \end{align*} where we use the notation $\DE{P}=1$ if $P$ is true and $\DE{P}=0$ if $P$ is false for a condition $P$. % Write simply $A = \{\kappa\mid\mu\subset\kappa\subset\tilde\lambda, \ \tilde\lambda/\kappa\text{: vertical strip}, \ \kappa'_{p-1}\ge\lambda'_p \}$. Substituting these to \eqref{eq:itjt}, we have \begin{align*} \text{(RHS of \eqref{eq:itjt})} &= \sum_{\kappa\in A} \sum_{i=0}^{\lambda'_p} t^{c(\kappa/\mu) + \DE{i>0}} (-1)^{|\tilde\lambda/\kappa| + \lambda'_p-i} t^{c(\tilde\lambda/\kappa) + \DE{i < \lambda'_p}} (t-1)^{|\tilde\lambda/\kappa| + \lambda'_p-i-(c(\tilde\lambda/\kappa) + \DE{i < \lambda'_p})} \\ &= \sum_{\kappa\in A} (-1)^{|\tilde\lambda/\kappa|} t^{c(\kappa/\mu) + c(\tilde\lambda/\kappa)} (t-1)^{|\tilde\lambda/\kappa| - c(\tilde\lambda/\kappa)} \underbrace{ \sum_{i=0}^{\lambda'_p} (-1)^{\lambda'_p-i} t^{\DE{i > 0} + \DE{i < \lambda'_p}} (t-1)^{\lambda'_p-i-\DE{i < \lambda'_p}} }_{(X)}. \end{align*} % We shall show $(X)=0$. Letting $q=\lambda'_p$ ($>0$) and $j=q-i$ we rewrite $(X)$ as \begin{equation}\label{eq:X} (X) = \sum_{j=0}^{q} (-1)^j t^{\DE{j > 0} + \DE{j < q}} (t-1)^{j-\DE{j > 0}}. \end{equation} % It is easy to check $\eqref{eq:X}=0$ when $q=1$. When $q\ge 2$, by checking \[ (-1)^q t (t-1)^{q-1} + (-1)^{q-1} t^{2} (t-1)^{q-2} = (-1)^{q-1} t (t-1)^{q-2}, \] we can carry induction on $q$ to obtain $\eqref{eq:X}=0$. Therefore we conclude $\eqref{eq:itjt}=0$ if $\lambda/\mu\neq\emptyset$, finishing the proof of Proposition \ref{theo:(Et,-)_detail}. \end{proof} \section{Example}\label{sect:eg} We display Young diagrams in the French notation. \begin{exam}\label{exam:I_skew} Let $\lambda/\mu=(3,2,1)/(1)=\sk{3,2,1}{1}$. We shall verify \eqref{eq:I_skew} for this $\lambda/\mu$ by expanding each term into a linear combination of $\{g_\nu\}$. We can check \begin{equation}\label{eq:321/1} \dgs{3,2,1}{1} = \dg{3,2} + \dg{3,1,1} + \dg{2,2,1} - \dg{3,1} - \dg{2,2} - \dg{2,1,1} + \dg{2,1}, \end{equation} using recursively the Pieri formula for skew dual stable Grothendieck polynomials \cite[Theorem 7.1]{1711.09544} \begin{equation}\label{eq:skewPieri} h_k g_{\mu/\nu} = \sum_{\substack{\lambda/\mu\text{: horizontal strip} \\ \nu/\eta\text{: vertical strip}}} (-1)^{k-|\lambda/\mu|} \binom{a(\lambda\para\mu) - a(\nu'\para\eta') - |\nu/\eta|}{k-|\lambda/\mu|-|\nu/\eta|} g_{\lambda/\eta}, \end{equation} where $a(\alpha\para\beta)$ is the number of $i\ge 1$ satisfying $\beta_i > \alpha_{i+1}$ and $\beta_{i}>\beta_{i+1}$, and the binomial coefficient $\binom{m}{n}$ is considered as $0$ when $n<0$. (Note: another way to check \eqref{eq:321/1} is to use \eqref{eq:Iinv:g}.) Applying $I$ to \eqref{eq:321/1} and using $I(g_\kappa)=\sum_{\alpha\subset\kappa}g_\alpha$, we compute the first term of \eqref{eq:I_skew} as \begin{align} I(\dgs{3,2,1}{1}) &= I(\dg{3,2} + \dg{3,1,1} + \dg{2,2,1} - \dg{3,1} - \dg{2,2} - \dg{2,1,1} + \dg{2,1}) \notag \\ &= \sum_{\kappa\subset\yd{3,2}} g_{\kappa} + \sum_{\kappa\subset\yd{3,1,1}} g_{\kappa} + \sum_{\kappa\subset\yd{2,2,1}} g_{\kappa} - \sum_{\kappa\subset\yd{3,1}} g_{\kappa} - \sum_{\kappa\subset\yd{2,2}} g_{\kappa} - \sum_{\kappa\subset\yd{2,1,1}} g_{\kappa} + \sum_{\kappa\subset\yd{2,1}} g_{\kappa} \notag \\ &= \sum_{\kappa\in [\emptyset, \yd{3,2}]\cup[\emptyset, \yd{3,1,1}]\cup[\emptyset, \yd{2,2,1}]} g_\kappa. \label{eq:I321/1} \end{align} Next we compute the second term of \eqref{eq:I_skew}, $\sum_{(1)\subset\nu\subset(3,2,1)}g_{\nu/(1)}$. Again using \eqref{eq:skewPieri} we have \begin{gather*} \dgs{3,2,1}{1} = \dg{3,2} + \dg{3,1,1} + \dg{2,2,1} - \dg{3,1} - \dg{2,2} - \dg{2,1,1} + \dg{2,1}, \\ \dgs{3,2}{1} = \dg{3,1} + \dg{2,2} - \dg{2,1}, \qquad \dgs{3,1,1}{1} = \dg{3,1} + \dg{2,1,1} - \dg{2,1}, \qquad \dgs{2,2,1}{1} = \dg{2,2} + \dg{2,1,1} - \dg{2,1}, \\ \dgs{3,1}{1} = \dg{3} + \dg{2,1} - \dg{2}, \qquad \dgs{2,2}{1} = \dg{2,1}, \qquad \dgs{2,1,1}{1} = \dg{2,1} + \dg{1,1,1} - \dg{1,1}, \\ \dgs{3}{1} = \dg{2}, \qquad \dgs{2,1}{1} = \dg{2} + \dg{1,1} - \dg{1}, \qquad \dgs{1,1,1}{1} = \dg{1,1}, \\ \dgs{2}{1} = \dg{1}, \qquad \dgs{1,1}{1} = \dg{1}, \qquad \dgs{1}{1} = \dg{}. \end{gather*} Summing them up, we have \begin{equation}\label{eq:sum_nu/1} \sum_{\yd{1}\subset\nu\subset\yd{3,2,1}}g_{\nu/\yd{1}} = \sum_{\kappa\in [\yd{}, \yd{3,2}]\cup[\yd{}, \yd{3,1,1}]\cup[\yd{}, \yd{2,2,1}]} g_\kappa. \end{equation} Finally we compute the last term of \eqref{eq:I_skew}, $\sum_{(1)\subset\nu\subset(3,2,1)}g_{(3,2,1)/\nu}$. We can check \begin{gather*} \dgs{3,2,1}{3,2,1} = \dg{}, \qquad \dgs{3,2,1}{3,2} = \dg{1}, \qquad \dgs{3,2,1}{3,1,1} = \dg{1}, \qquad \dgs{3,2,1}{2,2,1} = \dg{1}, \\ \dgs{3,2,1}{3,1} = \dg{2} + \dg{1,1} - \dg{1}, \qquad \dgs{3,2,1}{2,2} = \dg{2} + \dg{1,1} - \dg{1}, \qquad \dgs{3,2,1}{2,1,1} = \dg{2} + \dg{1,1} - \dg{1}, \\ \dgs{3,2,1}{3} = \dg{2,1}, \qquad \dgs{3,2,1}{2,1} = \dg{3} + 2\dg{2,1} + \dg{1,1,1} - 2\dg{2} - 2\dg{1,1} + \dg{1}, \qquad \dgs{3,2,1}{1,1,1} = \dg{2,1}, \\ \dgs{3,2,1}{2} = \dg{3,1} + \dg{2,2} + \dg{2,1,1} - 2\dg{2,1}, \qquad \dgs{3,2,1}{1,1} = \dg{3,1} + \dg{2,2} + \dg{2,1,1} - 2\dg{2,1}, \\ \dgs{3,2,1}{1} = \dg{3,2} + \dg{3,1,1} + \dg{2,2,1} - \dg{3,1} - \dg{2,2} - \dg{2,1,1} + \dg{2,1}. \end{gather*} Summing them up, we have \begin{equation}\label{eq:sum_321/nu} \sum_{\yd{1}\subset\nu\subset\yd{3,2,1}}g_{\yd{3,2,1}/\nu} = \sum_{\kappa\in [\yd{}, \yd{3,2}]\cup[\yd{}, \yd{3,1,1}]\cup[\yd{}, \yd{2,2,1}]} g_\kappa. \end{equation} Hence we see $\eqref{eq:I321/1}=\eqref{eq:sum_nu/1}=\eqref{eq:sum_321/nu}$, verifying \eqref{eq:I_skew}. \end{exam} \begin{rema} From \eqref{eq:I321/1} in the example above and a Pieri-type formula given in \cite{dualstablesum} ($I(g_\lambda)I(g_{(k)}) = \sum_{\mu} g_\mu$, summed over $\mu$ satisfying the set difference $\mu\setminus\lambda$ is a horizontal strip of size $\le k$), one may expect positivity in the expansions $I(g_{\lambda/\mu}) = \sum_{\nu} \tilde{c}^{\lambda}_{\mu\nu} g_\nu$ and $I(g_\mu)I(g_\nu)=\sum_{\lambda}\tilde{d}^{\lambda}_{\mu\nu} g_\lambda$ % (note that $\tilde{c}^{\lambda}_{\mu\nu} = \sum_{\mu\subset\kappa\subset\lambda} c^{\lambda}_{\kappa,\nu} = \sum_{\mu\subset\kappa\subset\lambda} c^{\kappa}_{\mu,\nu}$ and $\tilde{d}^{\lambda}_{\mu\nu} = \sum_{\substack{\alpha\subset\mu \\ \beta\subset\nu}} d^{\lambda}_{\alpha,\beta}$). % However, neither hold in general; a counterexample for the former is $\tilde{c}^{(53221)}_{(321),(321)}=-1$, and one for the latter is $\tilde{d}^{(5321)}_{(321),(321)}=-1$. \end{rema} \bibliographystyle{abbrv} \input{main.bbl} \end{document}

https://arxiv.org/abs/2010.11216

Null Kähler geometry and isomonodromic deformations

We construct the normal forms of null-Kähler metrics: pseudo-Riemannian metrics admitting a compatible parallel nilpotent endomorphism of the tangent bundle. Such metrics are examples of non-Riemannian holonomy reduction, and (in the complexified setting) appear in the Bridgeland stability conditions of the moduli spaces of Calabi-Yau three-folds.Using twistor methods we show that, in dimension four - where there is a connection with dispersionless integrability - the cohomogeneity-one anti-self-dual null-Kähler metrics are generically characterised by solutions to Painlevé I or Painlevé II ODEs.

\section{Introduction} A null--K\"ahler structure on a manifold $\mathcal{X}$ of real dimension $4n$ consists of a pseudo--Riemannian metric $g$ of signature $(2n, 2n)$, together with a rank $2n$ endomorphism of $T\mathcal{X}$ which, for all vector fields $X, Y$, satisfies \[ g(X, NY)+g(NX, Y)=0, \quad N^2=0 \] and is parallel with respect to the Levi--Civita connection of $g$. There are three reasons for considering such structures. Firstly, they provide an example of a pseudo-Riemannian holonomy reduction with no Riemannan analogue \cite{kath, BI, bryant, GL} (in the sense that a null--K\"ahler metric can not be analyticaly continued to Riemannian signature). Secondly, there exists a natural fibration of $\mathcal{X}$ over a symplectic manifold of dimension $2n$, such that the pull--back of the symplectic form to $\mathcal{X}$ agrees with the fundamental form $\Omega$ defined by $\Omega(X, Y)=g(NX, Y)$. This structure, albeit in the complexified setup and under the additional condition that $g$ is hyper--K\"ahler, underlies the Bridgeland approach to the stability conditions of moduli spaces of Calabi--Yau three--folds \cite{Bridgeland1, Bridgeland2, Bridgeland3}. Finally, in dimension four and under the additional anti--self--duality assumption, null--K\"ahler structures are characterised by solutions to a dispersionless integrable system \cite{D02}. In the next section we shall introduce null Hermitian structures on vector spaces, and in \S\ref{section3} we put these structures on manifolds. Our main result in \S\ref{section3} is the local normal form of the null--K\"ahler condition \begin{theo} \label{theorem1} Let $(\mathcal{X}, g, N)$ be a $4n$--dimensional null--K\"ahler manifold. There exist a local coordinate system $(x^i, y^i), i=1, \dots, 2n$ and a function $\Theta:\mathcal{X}\rightarrow \mathbb{R}$ such that \begin{eqnarray} \label{nk_form} g&=&\frac{1}{2}\sum_{i, j} \omega_{ij}(dy^i\otimes dx^j+dx^j\otimes dy^i) +\frac{\partial^2 \Theta}{\partial y^i \partial y^j}(dx^i \otimes dx^j +dx^j \otimes dx^i),\\ \quad N&=&\sum_{i} dx^i\otimes \frac{\partial}{\partial y^i}, \quad\mbox{where}\quad \omega_{ij}=\left(\begin{array}{cc} 0&\mathbb{I}_n\\ -\mathbb{I}_n&0 \end{array} \right).\nonumber \end{eqnarray} Conversely, the structure (\ref{nk_form}) is null--K\"ahler for any function $\Theta=\Theta(x^i, y^i)$. \end{theo} Thus, in the real analytic category, a null--K\"ahler manifold depends on one arbitrary function of $4n$ variables. In \S\ref{section32} and \S\ref{section33} we shall list systems of PDEs satisfied by this function if the metric is Einstein or (complexified) hyper--K\"ahler together with examples. In \S\ref{section4} we focus on oriented four--dimensional null--K\"ahler structures, with the natural choice of orientation which makes the fundamental form $\Omega$ self--dual. While null--K\"ahler metrics with self--dual Weyl tensor can be found explicitly, the anti--self--duality condition on Weyl tensor corresponds to solutions to a 4th order dispersionless integrable system \cite{D02}. A problem of finding Ricci--flat null reduces to a non--integrable hyper--heavenly equation of Plebanski and Robinson \cite{pleban_robinson}. In \S\ref{section5} we establish the main result of the paper. Imposing the invariance under the isometric action of $SL(2)$ on anti--self--dual null--K\"ahler structures leads to an ODE. By exploring the connection between the twistor distribution and the isomonodromic Lax pair we shall show that this ODE is either completely solvable, or transforms to Painlev\'e I or Painlev\'e II. \begin{theo} \label{theorem2} Let $(\mathcal{X}, g, N)$ be an anti--self--dual null--K\"ahler four--manifold with an isometric action of $SL(2)$ with three--dimensional orbits, and preserving the endomorphism $N$. Then either the metric $g$ is conformal to a Ricci--flat metric, or it can be put in the form \begin{equation} \label{metric_form} g=\sum_{\alpha, \beta=1}^3\gamma_{\alpha\beta}(t) \sigma^{\alpha}\otimes\sigma^{\beta} +\sum_{\alpha=1}^3 n_\alpha(t) (\sigma^{\alpha} \otimes dt+dt \otimes \sigma^{\alpha}), \end{equation} where the function $t:\mathcal{X}\rightarrow \mathbb{R}$ parametrises the orbits of $SL(2)$, and is constant on each orbit, $(\sigma^1, \sigma^2, \sigma^3)$ are left--invariant one--forms on $SL(2)$ which satisfy (\ref{one_formss}), $\gamma$ is a symmetric 3 by 3 matrix and $n$ is a vector with components given by (\ref{solv_form}), or depending on solutions of Painlev\'e I or Painlev\'e II as in (\ref{piform}) and (\ref{piiform}). \end{theo} Anti--self--dual $SL(2)$ invariant Ricci--flat metrics in neutral signature are all known \cite{DT_k}, so the novelty in Theorem \ref{theorem2} lies in the apperance of Painlev\'e equations in the conformal structures with no Ricci--flat representatives. The connection between the first two Painlev\'e transcendents, and anti--self--dual null--K\"ahler structures has twistorial origins: the additional structure on the twistor space ${\mathcal Y}$ of $\mathcal{X}$ which corresponds to the endomorophism $N$ is a holomorphic section of $\kappa^{-1/4}$, where $\kappa$ is the holomorphic canonical bundle of ${\mathcal Y}$. The isometric $SL(2)$ action on ${\mathcal Y}$ also gives rise to such section if the holomorphic vector fields generating this action are linearly dependent at one point, to order four, on each twistor lines. This corresponds to the isomonodromic problem underlying Painlev\'e I and II. \subsection*{Acknowledgments} I am grateful to Centro de Investigacion y de Estudios Avanzados in Mexico for the hospitality 19 years ago, when most of the results in \S\ref{section3} were obtained. My interest in the subject has been revived after I came across the works of Bridgeland \cite{Bridgeland1, Bridgeland2}, and Bridgeland and Strachan \cite{Bridgeland3}, where a null--K\"ahler structure implicitly arises on the space of stability conditions of a Calabi--Yau three--fold tirangulated category. My work has been partially supported by STFC consolidated grants ST/P000681/1, ST/T000694/1. \section{Algebraic preliminaries} \label{section2} Let $V$ be a vector space over field $\mathbb{F}$, where $\mathbb{F}$ is $\mathbb{R}$ or $\mathbb{C}$. \begin{defi} A null structure on an even--dimensional vector space $V$ is an endomorphism $N$ of $V$ such that \begin{equation} \label{eq1} N^2=0, \end{equation} and $\mbox{rank}(N)=\frac{1}{2}\mbox{dim}(V)$. \end{defi} For reasons to become clear later (see \S\ref{section3}) we shall chose $\mbox{dim}{(V)}=4n$, where $n$ is an integer. The kernel of $N$ is a $2n$--dimensional subspace of $V$, and any basis of this kernel can be extended to a basis of $V$. A convenient way to do it is to pick $2n$ linearly independent vectors $(X_1, X_2, \dots, X_{2n})$ not in $\mbox{Ker}(N)$, and use \[ X_1, \dots, X_{2n}, N(X_1), \dots, N(X_{2n}) \] as a basis of $V$. We shall call this basis adapted to $N$. The matrix of $N$ with respect to an adapted basis is \begin{equation} \label{eq2} N_0=\left(\begin{array}{cc} 0&\mathbb{I}_{2n}\\ 0&0 \end{array} \right), \end{equation} where $\mathbb{I}_{2n}$ is the $2n\times 2n $ identity matrix. Let $N$ and $N'$ be two null structures on vector spaces $V$ and $V'$ respectively. A linear map $\phi:V\rightarrow V'$ is called null linear if \[ N'\circ \phi=\phi \circ N. \] The sub--group ${\mathcal N} (V)\subset GL(V)$ of null--linear maps consists of matrices which commute with the matrix $N_0$. These matrices are of the form \[ \left(\begin{array}{cc} A&B\\ 0&A \end{array} \right), \] where $A$ and $B$ is an arbitrary $2n\times 2n$ matrix over $\mathbb{F}$, and $A$ is invertible. \begin{prop} There is a one-to-one correspondence between the set of null structures on $V$, and the homogeneous space $GL(V)/ {\mathcal N}(V)$. \end{prop} \noindent {\bf Proof.} Consider a $GL(V)$ action of a space of null structures given by \begin{equation} \label{Saction} N\longrightarrow \phi N\phi^{-1}, \qquad \phi\in GL(V). \end{equation} Let $N$ and $N'$ be two null structures of $V$ with adapted bases $(X_i, N(X_i)$ and $({X'}_i, N({X'}_i))$ respectively, where $i=1,\dots, 2n.$ Define the element $\phi\in GL(V)$ by \[ \phi(X_i)={X'}_i, \quad \phi(N(X_i))=N'({X'}_i). \] Therefore $N'=\phi N\phi^{-1}$ and the action (\ref{Saction}) is transitive. The isotropy subgroup is of this action is ${\mathcal N}(V)$, as $N_0=\phi N_0\phi^{-1}$ iff $\phi\in {\mathcal N}(V)$. \begin{flushright} $\Box $ \end{flushright} Our elementary discussion has so far followed the standard treatment of complex structures \cite{Kob_N}, except the endomorphism $N$ squares to $0$ rather than $-\mathbb{I}_{4n}$. The next step is to introduce the analog of Hermitian inner products\footnote{ The natural next step in the theory of complex structures is to introduce a complexification, where multiplication by complex numbers is given in terms of the complex structure $J$ by $(a+ib)X=a+b J(X)$, where $X\in V, a, b \in \mathbb{R}$ and $i^2=-1$. Pursuing this analogy for null structures leads to dual numbers in place of complex numbers. The algebra $\mathbb{D}$ of dual numbers consists of elements of the form \[ a+\epsilon\; b, \quad\mbox{where}\quad a, b\in \mathbb{R}\, \quad\mbox{and}\quad \epsilon^2=0. \] The dual numbers can be added, and multiplied according to \[ (a_1+\epsilon\; b_1)(a_2+\epsilon\; b_2)=a_1 a_2+\epsilon\;(a_1b_2+b_1 a_2), \quad (a_1+\epsilon \;b_1)+(a_2+\epsilon \;b_2)=a_1+a_2+\epsilon\; (b_1+b_2), \] but $\mathbb{D}$ is not a division algebra, as elements of the form $\epsilon\; b$ do not have inverses. The {\it infinitesimal} dual number $\epsilon$ underlies non--standard analysis, as it gives a framework to distinguish between real numbers like $1$, and $0.999\dots$ which are regarded as equal in ordinary analysis. A real vector space $V$ with a null structure $N$ can be turned into a vector space $V^{\mathbb{D}}$ over $\mathbb{D}$ by defining \[ (a+\epsilon b) X=aX+b N(X), \quad\mbox{where}\quad X\in V\quad \mbox{and}\quad a, b \in \mathbb{R}. \] In what follows we shall not explore any further connection between null structures and dual numbers, but will instead focus on null structures on curved manifolds.} \begin{defi} \label{defi2} A non--degenerate symmetric bi--linear inner product $g:V\times V\rightarrow \mathbb{F}$ is called null--Hermitian if \begin{equation} \label{compat} g(NX, Y)=-g(X, NY), \end{equation} for all $X, Y \in V$, where $N$ is a null structure of $V$. \end{defi} If $\mathbb{F}=\mathbb{R}$, then the signature of a null--Hermitian inner product is $(2n, 2n)$, also called split, neutral or Kleinian. The definition \ref{defi2} also implies $ g(X, NX)=0 $ and that $\mbox{Ker}(N)$ is a totally isotropic subspace of $V$. To each null--Hermitian inner product we associate a skew--symmetric bi--linear map $\Omega\in \Lambda^2(V^*)$ defined by \begin{equation} \label{hermitean_form} \Omega(X, Y)=g(NX, Y). \end{equation} Therefore $\Omega$ vanishes on $\mbox{Ker}(N)$, and it equips the $2n$--dimensional quotient vector space $V/\mbox{Ker}(N)$ with a symplectic structure. In a basis adapted to $N$ the inner product $g$ and the skew--form $\Omega$ are represented by \[ g_0=\left(\begin{array}{cc} 0&\omega\\ \omega^T&0 \end{array} \right),\qquad \Omega_0= \left(\begin{array}{cc} \omega&0\\ 0&0 \end{array} \right),\qquad\mbox{where}\quad\omega= \left(\begin{array}{cc} 0&\mathbb{I}_n\\ -\mathbb{I}_n&0 \end{array} \right). \] \subsubsection{Example} \label{example1} Let $W$ be a $2n$--dimensional symplectic vector space with a symplectic form $\omega$. The $4n$--dimensional space $V=W\oplus W$ carries a null--Hermitian structure defined by \[ N(x, y)=(0, x), \quad g((x, y), (x', y'))= \omega(x, y')-\omega(y, x') \] where $(x, y, x', y')\in W$. This inner product has signature $(2n, 2n)$ and it indeed satisfies (\ref{compat}) as \[ g((x, y), N(x', y'))=g((0, x'), (x, y))=\omega(x', x)=- g(N(x, y), (x', y')). \] \subsubsection{Example} \label{example2} Let $V$ be a $4n$--dimensional vector space over $\mathbb{F}=\mathbb{R}$ with two null structures $N_1, N_2$, such that \[ N_1N_2+N_2N_1=-\mbox{Id}, \] where $\mbox{Id}$ is the identity endomorphism on $V$. Then the endomorphisms \[ I:=N_1+N_2,\qquad S:=N_1-N_2,\qquad T:=[N_1, N_2] \] equip $V$ with a pseudo--quaternionic structure. Indeed \begin{eqnarray*} I^2&=&N_1N_2+N_2N_1=-\mbox{Id},\\ S^2&=&-N_1N_2-N_2N_1=\mbox{Id},\\ T^2&=&N_2(-\mbox{Id}-N_2N_1)N_1+N_1(-\mbox{Id}-N_1N_2)N_2=\mbox{Id},\\ IS&=&-T=-SI, \qquad IT=S=-TI, \qquad ST=I=-TS. \end{eqnarray*} If we instead consider $V$ over $\mathbb{F}=\mathbb{C}$ then $I, J:=iS, K:=-iT$ form a complexified hyper--complex structure on $V$. Then $\frac{1}{2}(I\pm iJ)$ are null structures on $V$. \section{Null K\"ahler structures} \label{section3} Let $\mathcal{X}$ be a smooth manifold of real dimension $4n$. We shall equip $\mathcal{X}$ with a null structure by smoothly extending such structure from each tangent space, and imposing integrability conditions. \begin{defi} A null structure $N$ on $\mathcal{X}$ is an endomorphism $N: T\mathcal{X}\rightarrow T\mathcal{X}$ such that $N^2=0$, and the sub--bundle $\mathcal{D}\subset T\mathcal{X}$ consisting of vectors fields annihilated by $N$ has rank $2n$, and is Frobenius--integrable, i. e. $ [\mathcal{D}, \mathcal{D}]\subset \mathcal{D}. $ \end{defi}\noindent The integrability condition holds if $N[NX, NY]=0$ or equivalently if $ T(NX, Y)=0 $ for all vector fields $X,Y$, where \[ T(X, Y):=[NX, NY]-N[NX, Y]-N[X, NY] \] is the Nijenhuis tensor of $N$. Null--structures are also called almost--tangent structures \cite{yano1}, and the following example shows why \subsubsection{Example} \label{example3} Let $\mathcal{X}=TM$ be the total space of the tangent bundle to a $2n$ dimensional manifold $M$. Let $U\in \Gamma(TM)$ be a vector field on $M$. Recall \cite{yano2} that the vertical lift $U^V$ of $U$ to $TM$ is a section of $T(TM)$ defined by \[ U^V(f)=\frac{d}{d\epsilon}\bigg\rvert_{\epsilon=0} f(m, u+\epsilon U) \] where $f:TM\rightarrow \mathbb{R}$ is an arbitrary function, $m\in M$ and $u\in T_m M$. The canonical null structure on $TM$ is the endomorphism $N:T(TM)\rightarrow T(TM)$ defined by \begin{equation} \label{eq5} N(X)=[\pi_*(X)]^V, \end{equation} where $\pi_*$ is the tangent map to the bundle projection $\pi:TM\rightarrow M$. Let $(x^1, \dots, x^{2n})$ be local coordinates on $M$ covering a neighbourhood of $m\in M$, and $(y^1, \dots, y^{2n})$ be the natural coordinates on $T_mM$ obtained by writing a tangent--vector as $U=\sum_{i} y^i\frac{\partial}{\partial x^i}$. Then $(x^i, y^i)$ are local coordinates on $TM$. If \[ X=\sum_{i} A^i\frac{\partial}{\partial x^i}+B^i\frac{\partial}{\partial y^i} \] is a general vector field on $TM$, then (\ref{eq5}) implies \[ N(X)=\sum_i A^i\frac{\partial}{\partial y^i}. \] Thus, in the natural coordinate system $(x^i, y^i)$ on $TM$ the null structure is given by a $(1, 1)$ tensor\footnote{There is another, equivalent definition of this canonical null structure which we shall now describe. We have defined vertical lifts of vector fields to the tangent bundle. We can also define vertical lifts of functions: if $h:M\rightarrow \mathbb{R}$, then $h^V=h\circ\pi$ is a function on $TM$. Vertical lifts of all tensor fields are defined by $(P\otimes Q)^V=P^V\otimes Q^V$. In particular the vertical lift of the identity endomorphism of $TM$ to $T(TM)$ is given by $N$.} \[ N=\sum_{i} dx^i\otimes\frac{\partial}{\partial y^i}. \] \begin{defi} \label{defi3} A signature $(2n, 2n)$ pseudo--Riemannian metric $g$ on a manifold $\mathcal{X}$ with a null structure $N$ is called null--K\"ahler if \begin{equation} \label{defi_Ng} g(NX, Y)=-g(X, NY)\qquad\mbox{and}\qquad \nabla N=0, \end{equation} where $\nabla$ is the Levi--Civita connection of $g$. \end{defi}\noindent The fundamental two--form $\Omega\in \Lambda^2(T^*\mathcal{X})$ defined by \[ \Omega(X, Y)=g(NX, Y) \] is covariantly--constant, and therefore closed. It satisfies \[ \Omega^{\wedge n}:=\underbrace{\Omega\wedge\dots\wedge\Omega}_n\neq 0, \quad \Omega^{\wedge (n+1)}=0. \] This should be contrasted with the K\"ahler condition, where $\Omega^{\wedge 2n}\neq 0$, and justifies the terminology. \subsubsection{Example} \label{example4} Let $\mathcal{X}_\mathbb{C}$ be a complexified hyper--K\"ahler manifold, i.e. a complex manifold of complex dimension $4n$ with three holomorphic parallel complex structures $I, J, K$ satisfying the algebra of quaternions and Hermitian with respect to a holomorphic metric $g$ on $\mathcal{X}_\mathbb{C}$. Then $N=\frac{1}{2}(I+iJ)$ is a (one of many) null--K\"ahler structure on $\mathcal{X}_\mathbb{C}$. This example underlies the occurrence of null structures in the geometric approach to Donaldson--Thomas invariants \cite{Bridgeland1, Bridgeland3}. \vskip5pt In the \S\ref{section31} we shall present a canonical normal--form of null--K\"ahler metrics. In the rest of this section we list properties of such metrics which do not refer to any choices of coordinates. \begin{prop} The Riemann curvature $R$, the Ricci curvature $r$ and the Ricci scalar $S$ of a null--K\"ahler metric satisfy \begin{subequations} \label{curvatures} \begin{align} R(X, Y)\circ N&=N\circ R(X,Y), \label{cu1}\\ R(NX, Y)&=-R(X, NY), \label{cu2} \\ r(NX, Y)&=0, \label{cu3} \\ S&=0 \label{cu4}. \end{align} \end{subequations} \end{prop} \noindent {\bf Proof.} Formula (\ref{cu1}) follows directly from $\nabla N=0$, and the definition of the curvature \[ R(X, Y)V=[\nabla_X, \nabla_Y]V-\nabla_{[X, Y]}V. \] To prove (\ref{cu2}) we use the (\ref{cu1}) together with the symmetry properties of the Riemannian curvature: \begin{eqnarray*} g(R(NX, Y)V, U)&=&g(R(U, V)Y, NX)= -g(N R(U, V)Y, X)=-g(R(U, V)NY, X)\\ &=&-g(R(X, NY)V, U). \end{eqnarray*} From its definition \[ r(X, Y)=\mbox{Tr}(V\rightarrow R(V, X)Y). \] The third formula (\ref{cu3}) then follows if we take $V\in \mathcal{D}$ as, setting $V=NU$, we have \begin{eqnarray*} r(NX, Y)&=&\mbox{Tr}(NU\rightarrow R(NU, NX)Y)\\ &=&-\mbox{Tr}(NU\rightarrow R(U, N^2X)Y)=0. \end{eqnarray*} Finally, to prove (\ref{cu4}) we shall compute $S=\mbox{Tr}_g(r)$ in the basis adapted to $N$, and regard $r$ and $g^{-1}$ as linear maps. The property (\ref{cu3}) implies that in this basis the matrix of $r$ is of the form \[ \left(\begin{array}{cc} *&0\\ 0&0 \end{array} \right).\] The property (\ref{compat}) implies that the matrix of $g^{-1}$ is of the form \[ \left(\begin{array}{cc} 0&\omega\\ \omega^T&\Theta \end{array} \right). \] for some block $2n$ by $2n$ matrices $\Theta$ and $\omega$ such that $\omega$ is skew and non--degenerate. Therefore $S=\mbox{Tr}(g^{-1}\cdot r)=0$. \begin{flushright} $\Box $ \end{flushright} The next result relates null--K\"ahler structures to special holonomy, and manifolds with parallel spinors \cite{bryant, kath} \begin{prop} \label{prop_spinor} A null--K\"ahler manifold admits a canonical parallel pure spinor. \end{prop}\noindent {\bf Proof.} This is really a result in linear algebra which builds on a bijection between the set of pure semi-spinors in $V=\mathbb{R}^{2n, 2n}$ and the Grasmannian of totally null $2n$-dimensional subspaces of $V$. Let ${\mathcal{C}}l(2n, 2n)$ be a real $2^{4n}$--dimensional Clifford algebra (see, e. g. \cite{ML}) of a null-Hermitian space $(V=\mathbb{R}^{2n, 2n}, N)$. This algebra is generated by $2^{2n}\times 2^{2n}$ matrices $\gamma(X)$ subject to the relations \[ \gamma(X)\gamma(Y)+\gamma(Y)\gamma(X)=2g(X, Y){\bf 1}, \quad\mbox{where}\quad X, Y\in V. \] The multiplicative group $\mbox{Spin}(2n, 2n)$ is generated by all elements $\gamma(X)\gamma(Y)$, where $X, Y$ are vectors of squared norm $\pm 1$. The {\em spin space} $\mathbb{S}$ is a reducible representation space of $\mbox{Spin}(2n, 2n)$. It can be decomposed as \[ \mathbb{S}=\mathbb{S}_+\oplus\mathbb{S}_-\cong\mathbb{R}^{2^{2n-1}}\oplus\mathbb{R}^{2^{2n-1}}, \] where $\mathbb{S}_{\pm}$ are irreducible spaces of {\em semi-spinors}. It is a simple algebraic fact that any totally isotropic subspace of $V$ has dimension at most $2n$. A semi-spinor $\iota$ is called {\em pure} iff \begin{equation} \label{pures} \mbox{dim}\{X\in V, \gamma(X)\iota=0\}=2n. \end{equation} The system of equations underlying (\ref{pures}) is of rank $2n$, and defines $2n$-dimensional plane. This plane is totally isotropic as \[ 0=\gamma(X)\gamma(X)\iota=g(X,X)\iota \] so $g(X, X)=0$. The space of totally isotropic planes in $\mathbb{R}^{2n, 2n}$ has two components defined by a pure element of $\mathbb{S}_+$ and $\mathbb{S}_-$ respectively. A pure semi-spinor $\iota$ is annihilated by $\gamma(X_1)\gamma(X_2)\cdots\gamma(X_{2n})\in {\mathcal Cl}(2n, 2n)$ where $X_1, \cdots, X_{2n}$ span a totally null plane. Therefore $\iota$ corresponds to an element of the Grassmann algebra $\xi\in \Lambda^{2n}(V^*)$ such that $\xi\wedge\xi=0$, and the assertion of the Proposition follows because $\xi$ is defined by a null-Hermitian structure $ \xi=\Omega^{\wedge n}, $ and therefore is parallel. \begin{flushright} $\Box $ \end{flushright} If $n=1$ then all semi-spinors are pure. The first non-trivial case corresponds to 8-dimensional null-Hermitian structures. \subsection{Null K\"ahler potential} \label{section31} In this section we shall find a canonical normal form of a null--K\"ahler metric, and express it in terms of second derivatives of one arbitrary function on $\mathcal{X}$. \noindent {\bf Proof of Theorem \ref{theorem1}.} Let \[ M=\mathcal{X}/\mbox{Ker}(N) \] be a $2n$--dimensional quotient manifold by the kernel of the $2n$--dimensional integrable distribution $\mathcal{D}$ of vector fields annihilated by $N$. Locally we regard $\mathcal{X}$ is the tangent bundle to $M$, and in the coordinate system of Example \ref{example3} the endomorphism $N$ is \begin{equation} \label{latestN} N=\sum_{i} dx^i\otimes\frac{\partial}{\partial y^i}, \end{equation} where $(x^1, \dots, x^{2n})$ are local coordinates on $M$, and $(y^1, \dots, y^{2n})$ are linear coordinates on fibres of $TM\rightarrow M$. The distribution $\mathcal{D}=\mbox{span}\{\partial/\partial y^1, \dots, \partial/\partial y^{2n}\}$ is totally isotropic, and therefore there exists a collection of functions on $\mathcal{X}$ \[ \omega_{ij}=\omega_{ij}(x, y), \quad \Theta_{ij}=\Theta_{ij}(x, y) \] such that $\Theta_{ij}=\Theta_{ji}$ and \begin{equation} \label{metric_meantime} g=\frac{1}{2}\sum_{i, j}\omega_{ij}(dy^i\otimes dx^j+dx^j\otimes dy^i)+ \frac{1}{2}\sum_{i, j}\Theta_{ij}(dx^i\otimes dx^j+dx^j\otimes dx^i). \end{equation} Evaluating the null--K\"ahler condition (\ref{defi_Ng}) on coordinate vector fields shows that \[ \omega_{ij}=-\omega_{ji}. \] We now impose the parallel condition $\nabla N=0$, where $N$ is given by (\ref{latestN}). The $dy^i\otimes dx^j$ components of $\nabla N$ vanish iff \[ \frac{\partial \omega_{ij}}{\partial y^k}=0 \] so that $\omega_{ij}=\omega_{ij}(x)$, and \[ \omega=\frac{1}{2}\sum_{i, j}\omega_{ij} dx^i\wedge dx^j \] is a symplectic form on $M$. Locally there exists a diffeomorphism of $M$ to a Darboux coordinate system $\tilde{x}=\tilde{x}(x)$ such that \[ \Omega=d\tilde{x}^1\wedge d\tilde{x}^{n+1}+\dots+ d\tilde{x}^n\wedge d\tilde{x}^{2n}. \] The transformation of $\mathcal{X}=TM$ induced by this diffeomorphism is \[ \tilde{x}^i=\tilde{x}^i(x^j),\quad \tilde{y}^i=\sum_{j}\frac{\partial\tilde{x}^i}{\partial x^j}y^j, \quad i, j=1, \dots, 2n, \] and it preserves the form of $N$, as $N=\sum_i d\tilde{x}^i\otimes\partial/\partial \tilde{y}^i$. We shall use this Darboux coordinate system from now on, and drop tildes. This has an effect of reducing $\omega_{ij}$ to a constant symplectic matrix as in (\ref{nk_form}). We now move on the vanishing of $dx^i\otimes dx^j$ components in $\nabla N$. This is equivalent to \[ \frac{\partial \Theta_{ij}}{\partial y^k}=\frac{\partial \Theta_{ik}}{\partial y^j} \] which gives the integrability conditions for the existence of $2n$ function $\Theta_1, \dots, \Theta_{2n}$ on $\mathcal{X}$ such that \[ \Theta_{ij}=\frac{\partial\Theta_i}{\partial x^j}. \] The symmetry condition $\Theta_{ij}=\Theta_{ji}$ implies the existence of a single function $\Theta=\Theta(x, y)$ such that \[ \Theta_i=\frac{\partial\Theta}{\partial y^i}. \] This puts the metric (\ref{metric_meantime}) in the canonical form (\ref{nk_form}). \begin{flushright} $\Box $ \end{flushright} The local normal form (\ref{nk_form}) is not invariant under general diffeomorphisms of $\mathcal{X}$. The subgroup of the pseudogroup of all diffeomorphisms changing coordinates $(x^i, y^i)$ as well as $\Theta$, while preseving (\ref{nk_form}) is a semi--direct product of $\mbox{SDiff(M)}$ and $\Gamma(M)$, where the symplectomorphisms $\mbox{SDiff}(M)$ of $M$ act on $\mathcal{X}=TM$ by a Lie lift, and $\Gamma(M)$ acts on the fibres of $TM$ by translations. The details are as follows: Let $Y$ be a vector field on $\mathcal{X}$ generating a one--parameter group of diffeomorphisms. The conditions \[ {\mathcal L}_Y\Omega=0, \quad {\mathcal L}_YN=0 \] imply \[ Y=\sum_{i, j} \omega^{ij}\frac{\partial H}{\partial x^j}\frac{\partial}{\partial x^i}+ \sum_{i, j, k} \Big(y^k\omega^{ij}\frac{\partial^2 H}{\partial x^k \partial x^j}\Big) +\sum_{i}T^i\frac{\partial}{\partial y^i}, \] where $\omega^{ij}$ is the inverse matrix of $\omega_{ij}$, i. e. $\omega^{ik}\omega_{kj}={\delta^i}_j$, and $(H, T^1, \dots, T^{2n})$ are arbitrary functions of $(x^1, \dots, x^{2n})$. Set \[ \tilde{x}^i=x^i+\epsilon {\mathcal L}_Y(x^i) ,\quad \tilde{y}^i=y^i+\epsilon {\mathcal L}_Y(y^i) , \quad \widetilde{\Theta}=\Theta+\epsilon\delta\Theta. \] Using \[ \frac{\partial}{\partial \tilde{y}^i}=\frac{\partial}{\partial y^i}-\sum_{j, k}\epsilon\omega^{jk}\frac{\partial^2 H}{\partial y^i\partial y^k}\frac{\partial}{\partial y^j} \] we find that the action generated by $Y$ preserves the form of $g$, i. e. \[ {g}= \frac{1}{2}\sum_{i, j} \omega_{ij}(d\tilde{y}^i\otimes d\tilde{x}^j+d\tilde{x}^j\otimes d\tilde{y}^i) +\frac{\partial^2 \widetilde{\Theta}}{\partial \tilde{y}^i \partial \tilde{y}^j}(d\tilde{x}^i \otimes d\tilde{x}^j +d\tilde{x}^j \otimes d\tilde{x}^i)+O(\epsilon^2) \] if \begin{equation} \label{freedom} \delta \Theta=\sum_{i, j, k} \Big(\frac{1}{6}y^iy^jy^k\frac{\partial^3 H}{\partial x^i\partial x^j \partial x^k}-\frac{1}{2}y^jy^k\omega_{ij} \frac{\partial T^i}{\partial x^k}\Big)+\sum_i y^i Q_i+R, \end{equation} where $(Q_1, \dots, Q_{2n}, R)$ are arbitrary functions of $x^i$. \subsection{Null--K\"ahler Einstein metrics} \label{section32} Computing the Ricci tensor of a null--K\"ahler metric in the form (\ref{nk_form}) we find \[ r=\sum_{i, j}\frac{\partial^2 f}{\partial y^i\partial y^j} (dx^i\otimes dx^j+dx^j\otimes dx^i), \] where \begin{equation} \label{form_of_f} f\equiv\sum_{i, j}\omega^{ij}\frac{\partial^2 \Theta}{\partial y^i\partial x^j}+\sum_{i, j, k, l}\frac{1}{2}\omega^{ik}\omega^{jl} \frac{\partial^2 \Theta}{\partial y^i\partial y^j} \frac{\partial^2 \Theta}{\partial y^k\partial y^l}. \end{equation} The Ricci--flat condition on $g$ therefore reduces to a system of fourth order PDEs on $\Theta$ which can be integrated twice to give a single second orde PDE on $\Theta$ \begin{equation} \label{einstein_eq} f=G+y^iF_i, \end{equation} where $(G, F_1, \dots, F_{2n})$ are arbitrary functions of $x^i$. Applying the Cauchy--Kovalevskaya theorem shows that in the real--analytic category the general Ricci--flat null K\"ahler metric depends on two arbitrary functions of $4n-1$ variables and some number of functions of $2n$ variables. \subsubsection{Example} It can be explicitly verified that for \begin{equation} \Theta=\frac{c}{\rho^{2n-1}}\quad\mbox{where}\quad \rho= \sum_{i, j} \omega_{ij}y^i x^j, \quad c=\mbox{const} \end{equation} the linear and non--linear terms in (\ref{form_of_f}) vanish separately resulting in $f=0$. The resulting metric \[ g=\frac{1}{2}\sum_{i, j} \omega_{ij}(dy^i\otimes dx^j+dx^j\otimes dy^i) +\frac{2c n(2n-1)}{\rho^{2n+1}}\Big(\sum_{k, l}\omega_{kl} x^k dx^l\Big)^{\otimes 2} \] is therefore null--K\"ahler, and Ricci--flat. This metric with $n=1$ is the Sparling--Tod $H$--space \cite{ST}. \subsection{Complex hyper--K\"ahler metrics with affine symplectic fibrations} \label{section33} In the complexified setting the coordinates $(x^i, y^i)$ are holomorphic on the complex manifold $\mathcal{X}_\mathbb{C}$ of complex dimension $4n$. If $\Theta=\Theta(x, y)$ satisfies the system of PDEs \begin{eqnarray} \label{BS} H_{ij}&=&0,\quad i, j=1, \dots, 2n \quad\mbox{where}\nonumber\\ H_{ij}&\equiv&\frac{\partial^2 \Theta}{\partial y^i \partial x^j}-\frac{\partial^2 \Theta}{\partial y^j \partial x^i}+ \sum_{k, l}\omega^{kl}\frac{\partial^2 \Theta}{\partial y^i \partial y^l} \frac{\partial^2 \Theta}{\partial y^j \partial y^k}=0, \end{eqnarray} then the metric (\ref{nk_form}) is complexified hyper--K\"ahler. In \cite{Bridgeland3} it was shown that if $(M_\mathbb{C}, \omega)$ is a complexifed symplectic manifold of complex dimension $2n$, and $g$ is a complexified hyper--K\"ahler metric on $\mathcal{X}_\mathbb{C}=TM_\mathbb{C}$ such that the null--K\"ahler two--form \[ \pi^*(\omega)=\Omega_I+i\Omega_J, \] then $g$ is locally of the form (\ref{nk_form}), where $\Theta$ satisfies the system (\ref{BS}). The system (\ref{BS}) consists of some of the flows of the hyper--K\"ahler hierarchy \cite{Tak, DM}. It implies the Frobenius integrability \begin{equation} \label{thels} [l_i, l_j]=0, \quad i, j=1,\dots, 2n \end{equation} for the rank--$2n$ distribution spanned by \[ l_i=\frac{\partial}{\partial y^i}+\lambda\Big(\frac{\partial}{\partial x^i} +\sum_{j, k}\omega^{jk}\frac{\partial^2 \Theta}{\partial y^i\partial y^j}\frac{\partial}{\partial y^k} \Big) \] on $\mathcal{X}_\mathbb{C}\times \mathbb{CP}^1$, where $\lambda$ is the affine coordinate on $\mathbb{CP}^1$. Vanishing of the Lie brackets (\ref{thels}) gives a weaker set of conditions\footnote{Note that $\sum_{i, j}\omega^{ij}H_{ij}=2f$. Therefore (\ref{weaker_form}) implies Ricci--flatness, but the converse is not true.} \begin{equation} \label{weaker_form} \frac{\partial H_{ij}}{\partial y^k}=0, \quad \mbox{so that} \quad H_{ij}=C_{ij}(x), \end{equation} for some skew $C_{ij}$. If $n=1$ then a transformation $\Theta\rightarrow \Theta+\sum_{i}y^iQ_i(x)$ can be used to set $C_{ij}$ to zero. For general $n$ this only seems possible if the two--form $\sum_{i, j}C_{ij}dx^i\wedge dx^j$ is closed. In \cite{Bridgeland3} it has been argued that for any $n$ the conditions (\ref{weaker_form}) together with the additional assumption that the function $\Theta$ is odd in the fibre variables $(y^1, \dots, y^{2n})$ imply (\ref{BS}). Geometrically, $\lambda$ labels the $2n$--dimensional surfaces (the $\alpha$--surfaces in the twistor approach \cite{DM}) through each point of $\mathcal{X}_\mathbb{C}$. The twistor space of $(\mathcal{X}_\mathbb{C}, g)$ is the space of these $\alpha$--surfaces. It is a complex manifold of complex dimension $2n+1$, which arises as the quotient of $\mathcal{X}_\mathbb{C}\times \mathbb{CP}^1$ by the distribution spanned by $l_i$. The points in $\mathcal{X}_\mathbb{C}$ correspond to rational curves in ${\mathcal Y}$ with normal bundle $\mathbb{C}^{2n}\otimes \mathcal{O}(1)$, where $\mathcal{O}(1)$ is the line bundle with Chern class 1 on $\mathbb{CP}^1$. In \cite{DM} this twistor correspondence has been extended to the full hyper--K\"ahler hierarchy. \subsubsection{Example} A strong Joyce structure has, in \cite{Bridgeland2}, been defined to be a solution $\Theta$ of the system (\ref{BS}) subject to three additional conditions: \begin{enumerate} \item $\Theta$ is odd in the variables $y^i$. \item $Z\equiv \sum_ix^i\frac{\partial}{\partial x^i}$ is a homothetic Killing vector field such that \[ {\mathcal L}_Z g=g, \quad {\mathcal L}_Z \Theta=-\Theta. \] \item The metric is invariant under the lattice transformations \[ y^i\rightarrow y^i+2\pi\sqrt{-1}, \quad i=1, \dots, 2n. \] \end{enumerate} An example of a solution to (\ref{BS}) which also satisfies the three conditions above is \begin{equation} \Theta=\frac{\sinh{y^1}}{x^1}. \end{equation} The resulting metric is non--flat, and is an example of a Ricci--flat plane wave. \subsection{Conformal invariance} In four dimension the restricted conformal transformations, where the conformal factor is constant along the distribution ${\mathcal D}=\mbox{Ker}(N)$, preserve the null--K\"ahler condition: If $F=F(x^1, x^2)$, and \[ \hat{g}=F^2g, \quad\hat{\Omega}=F^3\Omega, \quad\mbox{then}\quad \hat{\nabla}\hat{\Omega}=0. \] This conformal invariance is not present in other dimensions: for $F^k\Omega$ to be closed we need $k=0$, and then $\hat{\nabla}\Omega=0$ implies $F=\mbox{const}$. \subsection{Walker structures} Recall that a distribution ${\mathcal D}$ on a pseudo--Riemannian manifold $\mathcal{X}$ is called parallel if $\nabla_Y X\in \Gamma({\mathcal D})$ for all $X\in \Gamma({\mathcal D})$ and $Y\in \Gamma(T\mathcal{X})$. The pseudo--Riemannian manifolds admitting a parallel distribution of rank equal to half of the manifold dimension are called the Walker manifolds \cite{Gilkey}, and it was shown by Walker \cite{Walker_1}, that locally a Walker metric is of the form (\ref{metric_meantime}) for some functions $\Theta_{ij}$. The null--K\"ahler manifolds form a subclass of the Walker manifolds, where $\Theta_{ij}$ is a Hessian of one function. Indeed, any vector field in ${\mathcal D}$ is of the form $N(X)$ for some $X\in \Gamma(T\mathcal{X})$, and we have \[ \nabla_Y N(X)=N \nabla_Y X\in\mbox{Ker}(N). \] \section{Four dimensions} \label{section4} The author has first came across the null--K\"ahler structures when investigating twistor theory and integrability of a certain fourth order PDE in four dimensions \cite{D02}. In four dimensions the existence of a maximal rank parallel endomorphism $N$ with $N^2=0$ is equivalent to existence of a parallel semi--spinor, i. e. a parallel section of a rank--two symplectic vector bundle which we chose to be $\mathbb{S}_+\rightarrow \mathcal{X}$ where \begin{equation} \label{can_bun_iso} T \mathcal{X}\cong {\mathbb{S}_+}\otimes {\mathbb{S}_-} \end{equation} is a canonical bundle isomorphism, and $\mathbb{S}_-$ is another rank--2 symplectic vector bundle. This isomorphism is related to the metric on $\mathcal{X}$ by \[ g(v_1\otimes w_1,v_2\otimes w_2) =\varepsilon_+(v_1,v_2)\varepsilon_-(w_1, w_2) \] where $v_1, v_2\in \Gamma(\mathbb{S}_+)$ and $w_1, w_2\in \Gamma(\mathbb{S}_-)$, and $\varepsilon_{\pm}$ are symplectic structures on $\mathbb{S}_\pm$ which are parallel with respect to $\nabla$. The Hodge $\ast$ operator is an involution on two-forms, and induces a decomposition \begin{equation} \label{splitting} \Lambda^{2}(T^*\mathcal{X}) = \Lambda_{+}^{2}(T^*\mathcal{X}) \oplus \Lambda_{-}^{2}(T^*\mathcal{X}) \end{equation} of two-forms into self-dual (SD) and anti-self-dual (ASD) components. Given a parallel section of $\mathbb{S}_+$, another isomorphism \begin{equation} \label{another_iso} \Lambda_{+}^{2}\cong \mbox{Sym}^2({\mathbb{S}_+}^{*}) \end{equation} implies the existence of a parallel self--dual two--form $\Omega$ such that $\Omega\wedge\Omega=0$. This two--form, together with $g$ define the nilpotent endomorphism $N$ by (\ref{hermitean_form}). In four dimensions there are three non--linear systems of PDEs, one of them completely solvable, one integrable, and one not-integrable, which can be imposed on the null structure. Before writing these systems down in coordinates of Theorem (\ref{theorem1}) recall \cite{AHS} that in four dimensions the Riemann tensor of $g$ can be regarded as a map $\mathcal{R}: \Lambda^{2}(T^*\mathcal{X}) \rightarrow \Lambda^{2}(T^*\mathcal{X})$ which admits a decomposition under the splitting (\ref{splitting}): \begin{equation} \label{decomp} {\mathcal R}= \left( \mbox{ \begin{tabular}{c|c} &\\ $C_+ +\frac{1}{12}S$&$r_0$\\ &\\ \cline{1-2}&\\ $r_0$ & $C_-+\frac{1}{12}S$\\&\\ \end{tabular} } \right) . \end{equation} Here $C_{\pm}$ are the SD and ASD parts of the Weyl tensor, $r_0$ is the trace-free Ricci curvature, and $S$ is the scalar curvature which acts by scalar multiplication. We are now ready to present the three systems of PDEs \subsection{Self--dual null--K\"ahler} The condition $C_-=0$ is equivalent to \[ \frac{\partial^4\Theta}{\partial y^i\partial y^j\partial y^k \partial y^l}=0. \] Therefore the most general self--dual null--K\"ahler metric in four dimensions is of the form (\ref{nk_form}) with \[ \Theta=\sum_{i, j, k}\Gamma_{ijk}y^iy^jy^k \] where the functions $\Gamma_{ijk}$ depend only on $(x^1, x^2)$, as the coordinate freedom (\ref{freedom}) can be used to remove the quadratic and linear terms from $\Theta$. The resulting metric is a Walker's projective extension \cite{DT_k} of a projective structures on the surface $M=\mathcal{X}/\mathcal{D}$. \subsection{ Anti--self--dual null--K\"ahler} The condition $C_+=0$ is equivalent to a 4th order PDE for $\Theta$: \begin{eqnarray} \label{eqmd} f&=& \Theta_{x^1y^2}-\Theta_{x^2y^1}+\Theta_{y^1y^1}\Theta_{y^2y^2}-{(\Theta_{y^1y^2})}^2\\ \Delta_g f&:=&f_{x^1y^2}-f_{x^2y^1}+\Theta_{y^2y^2}f_{y^1y^1}+\Theta_{y^1y^1}f_{y^2y^2} -2\Theta_{y^1y^2}f_{y^1y^2}=0\nonumber \end{eqnarray} where subscripts denote partial derivatives, i. e. $\Theta_{x^1}=\partial\Theta/\partial x^1$ etc. Note that $\Delta_g$ is the Laplace--Beltrami operator of the metric $g$, and the expression for $f$ agrees with the general formula (\ref{form_of_f}). This equation is integrable by twistor transform \cite{D02}, the dressing method \cite{Bogdanov}, and the Manakov--Santini inverse scattering transform \cite{Yi}. The general solution depends on 4 functions of 3 variables. It has recently been shown to arise from a second-order integrable Lagrangian \cite{Ferapontov}. \subsection{Null--K\"ahler Einstein} As the scalar curvature of null--K\"ahler manifolds always vanishes, the Einstein condition is equivalent to the vanishing of the Ricci tensor of $g$. The resulting second order PDE (\ref{einstein_eq}) on $\Theta$ is the hyper–heavenly equation of Pleba\'nski and Robinson \cite{pleban_robinson} for non--expanding metrics with self--dual Weyl tensor $C_+$ of type ${\bf N}$. (Recall that $(\mathcal{X}, g)$ is called hyper–heavenly if the self-dual Weyl tensor is algebraically special, i. e. has a repeated root when regarded as a binary quartic. Type ${\bf N}$ corresponds to a repeated root of order $4$). \subsection{Heavenly equation} Imposing the Einstein condition together with the anti--self--duality of the Weyl tensor reduces the 4th order equation (\ref{eqmd}) to a second order PDE \[ f=0. \] This is Pleba\'nski's second heavenly equation \cite{pleban}. The resulting metric is pseudo--hyper--K\"ahler. \section{Anti--self--duality and isomonodromy} \label{section5} In this section we shall assume that $(\mathcal{X}, N, g)$ is a null--K\"ahler four--manifold with anti--self--dual Weyl curvature which is cohomogeneity-one, i. e. there exists an isometry group $G$ acting transitively on three--dimensional surfaces in $\mathcal{X}$. The four-dimensional cohomogeneity-one metrics can be classified according to the Bianchi type of the three-dimensional real Lie algebra of $G$. Locally $\mathcal{X}=\mathbb{R}\times G$ and the ASD cohomogeneity--one null-K\"ahler condition reduces to solving a system of ODEs. To write this system down, and recognise it as the isomonodromy problem for Painlev\'e I and II if $G=SL(2)$, we shall use the twistor methods \cite{penrose_NG, D02}. We shall therefore work in the holomorphic category, and assume that $\mathcal{X}_\mathbb{C}$ is a complex oriented four--manifold, and $(N, g)$ are holomorphic. \begin{defi} \label{alpha_defi} An $\alpha$-surface is two-dimensional surface $\zeta\subset \mathcal{X}_\mathbb{C}$ such that for all $p \in \mathcal{X}_\mathbb{C}$ the tangent space $T_p\xi$ is a totally null plane with self--dual tangent bi-vector. \end{defi} The Nonlinear Graviton theorem of Penrose \cite{penrose_NG} states that there locally exist a three--parameter family of $\alpha$--surfaces iff the self-dual part of the Weyl tensor of $g$ vanishes. The twistor space ${\mathcal Y}$ of an ASD four-manifold is defined to be the space of $\alpha$--surfaces. It is a three--dimensional complex manifold with a four--parameter family of rational curves with normal bundle $\mathcal{O}(1)\oplus\mathcal{O}(1)$. The points of the three--dimensional twistor space ${\mathcal Y}$ are $\alpha$--surfaces in $\mathcal{X}_\mathbb{C}$: There is a rational curve $L_p\cong\mathbb{CP}^1$ worth of such surfaces through each point $p\in \mathcal{X}_\mathbb{C}$, and therefore points in $\mathcal{X}_\mathbb{C}$ correspond to rational curves in ${\mathcal Y}$. The conformal structure on $\mathcal{X}_\mathbb{C}$ is defined by declaring two points $p_1, p_2 \in \mathcal{X}_\mathbb{C}$ to be null--separated iff the corresponding rational curves in ${\mathcal Y}$ intersect at one point. The correspondence between $\mathcal{X}_\mathbb{C}$ and ${\mathcal Y}$ can be expressed in the the double fibration picture (see e.g. \cite{mason}). \begin{equation} \label{doublefib} {\mathcal{X}_\mathbb{C}}\stackrel{r}\longleftarrow {\mathcal F}\stackrel{q}\longrightarrow {\mathcal Y}, \end{equation} where the five--complex--dimensional correspondence is defined by \[ {\mathcal F}={\mathcal{Y}}\times {\mathcal{X}_\mathbb{C}}|_{\zeta\in L_p}= {\mathcal{X}_\mathbb{C}}\times\mathbb{CP}^1 \] where $L_p$ is the rational curve in ${\mathcal Y}$ that corresponds to $p\in {X_\mathbb{C}}$, and $\zeta\in{\mathcal{Y}}$ lies on $L_p$. The twistor space arises as a quotient of ${\mathcal F}$ by a two--dimensional integrable distribution spanned by the vector fields \begin{equation} \label{tetradlax} l_1 = E_{11} - \lambda E_{12} + f_1 \frac{\partial}{\partial \lambda}, \quad l_2 = E_{21} - \lambda E_{22} + f_2 \frac{\partial}{\partial \lambda}, \end{equation} where $\lambda$ is an affine coordinate on $\mathbb{CP}^1$, the functions $f_1, f_2$ on ${\mathcal F}$ are cubic in $\lambda$, and $E_{ij}$ are four independent holomorphic vector fields on $\mathcal{X}_\mathbb{C}$ such that the conformal structure defined by the contravariant metric \begin{equation} \label{tetrad_0} g=\frac{1}{2}(E_{11}\otimes E_{22} +E_{22}\otimes E_{11} - E_{12}\otimes E_{21}-E_{21}\otimes E_{12}). \end{equation} The Frobenius integrability condition \begin{equation} \label{intlm} [l_1, l_2]=0 \quad (\mbox{mod}\; l_1, l_2) \end{equation} is equivalent to the anti--self--duality condition $C_+=0$ on $\mathcal{X}_\mathbb{C}$. If the integrability condition holds then there is a $\mathbb{CP}^1$--worth of $\alpha$-surfaces spanned by $\{ E_{11} - \lambda E_{12}, E_{21} - \lambda E_{22}\}$ through any point in $\mathcal{X}_\mathbb{C}$. If all vectors $E_{ij}$ are real then the signature of $g$ is $(2, 2)$, and there exists an $\mathbb{RP}^1$--worth of real $\alpha$-surfaces through each point of a real four--manifold $\mathcal{X}$. The null--K\"ahler condition on top of anti--self--duality gives rise to an additional structure on the twistor space: \begin{theo}\cite{D02} \label{mytheo} Let ${\mathcal Y}$ be a three-dimensional complex manifold with \begin{enumerate} \item A four-parameter family of rational curves with normal bundle $\mathcal{O}(1)\oplus \mathcal{O}(1)$. \item A preferred section of $\kappa^{-1/4}$ where $\kappa$ is the holomorphic canonical bundle of ${\mathcal Y}$. \item An anti-holomorphic involution $\rho: {\mathcal Y}\rightarrow {\mathcal Y}$ fixing a real equator of each rational curve, and leaving the section of $\kappa$ above invariant. \end{enumerate} Then the real moduli space $\mathcal{X}$ of the $\rho$--invariant curves is equipped with a restricted conformal class $[g]$ of anti--self--dual null-K\"ahler metrics: if $g\in [g]$, and $\Omega$ is a null-K\"ahler two-form, then $(\hat{g}=F^2g, \hat{\Omega}=F^3\Omega)$ is also null--K\"ahler for any function $F$ such that $dF\wedge\Omega=0$. Conversely, given a real analytic ASD null--K\"ahler metric, there exists a corresponding twistor space ${\mathcal Y}$ with the above structures. \end{theo} If one is only interested in the complexified picture, where $g$ and $N$ are holomorphic on $\mathcal{X}_\mathbb{C}$, then condition (3) in Theorem \ref{mytheo} can be dropped. From now on we shall additionally assume that there exists a three--dimensional complex Lie group $G$ acting on $\mathcal{X}_\mathbb{C}=\mathbb{C}\times G$ by isometries with generically three--dimensional orbits. We shall make a choice for $G$, and take it to be $SL(2, \mathbb{C})$ (or $SL(2, \mathbb{R})$ if $\mathcal{X}$ is a real four--manifold with a $(2, 2)$ metric). Its Lie algebra is generated by the left invariant vector fields $L_1, L_2, L_3$ on $G$ which satisfy \begin{equation} \label{bianchi_2} [L_1, L_2]=L_2, \quad [L_1, L_3]=-L_3, \quad [L_2, L_3]=2L_1. \end{equation} The conformal isometries are generated by the right-invariant vector fields $R_\alpha, \alpha=1, 2, 3$ on $G$. The metric on $\mathcal{X}_\mathbb{C}$ will be expressed in terms of the left--invariant one--forms $\sigma^1, \sigma^2, \sigma^3$ on $SL(2)$ such that \[ {\mathcal L}_{R_{\alpha}}\sigma^\beta=0, \quad L_\alpha\hook\sigma^\beta=\delta^{\beta}_{\alpha} \] and \begin{equation} \label{one_formss} d\sigma^1=2\sigma^3\wedge \sigma^2, \quad d\sigma^2=\sigma^2\wedge \sigma^1,\quad d \sigma^3=\sigma^1\wedge\sigma^3. \end{equation} The $G$--action on $\mathcal{X}_\mathbb{C}$ maps $\alpha$--surfaces to $\alpha$--surfaces and thus gives rise to a holomorphic group action of $G$ on the twistor space ${\mathcal Y}$. Let the $\widetilde{R}_\alpha, \alpha=1, 2, 3$ be holomorphic vector fields on ${\mathcal Y}$ generating this action and corresponding to $R_\alpha$. Consider a quartic \begin{equation} \label{quartic} s = \mbox{vol}_{\mathcal Y} ({\widetilde{R}}_1, {\widetilde{R}}_2, {\widetilde{R}}_3), \end{equation} where $\mbox{vol}_{\mathcal Y}$ is a holomorphic volume form on the twistor space with values in $\mathcal{O}(4)$. This quartic vanishes at each twistor line at four points, where the holomorphic vector fields corresponding to the isometries become linearly dependent. We shall, form now on assume that the four zeros of the quartic coincide, and so $s$ gives a preferred section of $\kappa^{-1/4}$. The corresponding conformal structure therefore contains a null-K\"ahler structure by Theorem \ref{mytheo}. We shall first need to establish two technical results about the quartic (\ref{quartic}), as the canonical form of the metric depends on whether $s$ vanishes identically, or not. \begin{prop} \label{quartic_lemma} If the quartic (\ref{quartic}) vanishes identically then the conformal class containing $g$ is hyper--complex, or equivalently if there exists a holomorphic fibration of ${\mathcal Y}$ over $\mathbb{CP}^1$ such that the twistor curves are sections of this fibration. \end{prop} \noindent {\bf Proof.} We shall first introduce some notation. Let $\pi^i=[\pi^1, \pi^2]$ be homogeneous coordinates on $\mathbb{CP}^1$--fibres of the bundle $\mathbb{P}(\mathbb{S}_+)$ such that $\lambda=-\pi^2/\pi^1$ in the patch where $\pi^1\neq 0$, and let $\pi_i=\sum_{j}\varepsilon_{ji}\pi^j$. Assemble the frame in (\ref{tetradlax}) into a vector--valued two by two matrix $E$ with components $E_{ij}$ so that the twistor distribution ${\mathcal D}_{\mathcal Y}\equiv\mathcal{O}(-1)\otimes\mathbb{C}^2$ given by (\ref{tetradlax}) takes the form \[ l_i=\sum_j \pi^j E_{ij}+f_i\frac{\partial}{\partial \lambda}, \quad i=1, 2. \] For this to be homogeneous of degree $1$ in $\pi$ the functions $(f_1, f_2)$ need to be sections of $\mathcal{O}(3)$. Let $e^{ij}$ be a frame of one--forms dual to $E_{ij}$ so that $E_{ij}\hook e^{mn}={\delta_i}^m{\delta_j}^n$, and the metric is given by \[ g=\frac{1}{2}(e^{11}\otimes e^{22}+e^{22}\otimes e^{11}-e^{12}\otimes e^{21}-e^{21}\otimes e^{12}). \] In the double fibration picture (\ref{doublefib}) the quartic $s$ pulls back to a quartic on ${\mathcal F}$ given by \begin{equation} \label{quartic1} q^*(s) = (d \lambda \wedge \mbox{vol}) (l_1, l_2, {R}_1, {R}_2, {R}_3), \end{equation} where $\mbox{vol}$ is the holomorphic volume form on $\mathcal{X}_\mathbb{C}$ such that $\mbox{vol}( E_{11}, E_{21}, E_{12}, E_{22})=1$, and we have chosen to work in an invariant frame, where the lifts of $R_\alpha$s to the correspondence space ${\mathcal F}$ are given by $R_\alpha$s. Such a frame always exists, as given a cohomogeneity--one metric of the form (\ref{tetrad_0}) we can choose a frame of one--forms $e^{ij}$ which are linear combinations of the left invariant one--forms on on $G$, and $dt$. Here $t$ is a function on such that the surfaces of constant $t$ are the orbits of $SL(2)$, and so $dt$is normal to the surfaces of homogeneity. The coefficients of this combination only depend on $t$, so the self--dual two--forms \[ e^{11}\wedge e^{21},\quad e^{11}\wedge e^{22}-e^{21}\wedge e^{12}, \quad e^{12}\wedge e^{22} \] constructed from the frame $e^{ij}$ are also $G$--invariant. Therefore the lift of the $SL(2)$ action to the bundle $\mathbb{S}_+$ is trivial, but the correspondence space is the projectivisation of this bundle. Define $T_{ij}$ by \begin{equation} \label{formtt} dt=\sum_{ij} T_{ij} e^{ij}, \quad \mbox{so that}\quad T_{ij}=E_{ij}(t). \end{equation} where we have used $d=\sum_{i, j} e^{ij}\otimes E_{ij}$. Let $\mbox{vol}=dt\wedge \mbox{vol}_{SL(2)}$, so that \[\mbox{vol}_{SL(2)}(R_1, R_2, R_3)=1.\] Therefore, using (\ref{formtt}), \[ q^*(s)=\frac{1}{2}\sum_{i, j, m, n}{\varepsilon_-}^{mn}T_{ij} d\lambda\wedge e^{ij}(l_m, l_n) \] where $\varepsilon_-$ is the symplectic structure on ${\mathbb{S}_-}^*$. Using $e^{ij}(l_m, \cdot)={\delta^i}_m\pi^j$ gives \begin{eqnarray} \label{new_form} q^*(s)&=& \sum_{i, j, k, m, n}T_{ij}f_m\pi^k{\delta^i}_n{\delta^j}_k {\varepsilon_-}^{mn}\nonumber\\ &=& \sum_{i, j, k} f_i T_{jk}\pi^k{\varepsilon_-}^{ij}. \end{eqnarray} If the invariant frame $E_{ij}$ is also such that $f_1=f_2=0$, then $q^*(s)$ given by (\ref{new_form}) is identically zero. In \cite{D99} it was shown that a frame with $f_1=f_2=0$ (and therefore a holomorphic fibration ${\mathcal Y}\rightarrow \mathbb{CP}^1$ \cite{Boyer, Hitchi_frame, Calderbank}) exists iff $g$ is hyper--complex. Therefore the hyper--complex condition is necessary for the vanishing of $s$. \begin{flushright} $\Box $ \end{flushright} \newpage {\bf Remarks} \subsection{} A complexified hyper--Hermitian (which in four dimensions is equivalent to complexified hyper--complex) structure on $\mathcal{X}_\mathbb{C}$ is a triple of holomoprhic Hermitian endomorphisms $I, J, K$ of $T\mathcal{X}_\mathbb{C}$ which satisfy the algebra of quaternions. If $J=iS, K=-iT$, and $I, S, T$ are all real, then they form a pseudo--hyper--Hermitian structure on split--signature real four--manifold $\mathcal{X}$. The endomorphism $I$ endows $\mathcal{X}$ with the structure of a two--dimensional complex K\"ahler manifold, and so does every other complex structure $aI+bS+cT$ parametrised by the points of the hyperboloid \begin{equation} \label{hyperboloid} a^2-b^2-c^2=1. \end{equation} \subsection{} \label{sub51} The converse of Proposition \ref{quartic_lemma} does not hold: if $g$ is ASD and Ricci--flat (and therefore hyper--complex) but the $SL(2)$ action rotates the covariantly constant self--dual two forms, then $s$ does not vanish. That is to say the covariantly constant frame does not have to be invariant. The basis of two--forms in an invariant frame (which, as we have argued, always exists) is not covariantly constant and so $f_1$ and $f_2$ will not vanish. ASD Taub--NUT or the Atiyah--Hitchin metrics are both examples illustrating this phenomenon. We can however say more if the isometric group action preserves the null--K\"ahler structure: \begin{lemma} If $(\mathcal{X}_\mathbb{C}, g, N)$ is a cohomogeneity--one $SL(2)$ invariant null--K\"ahler structure which is Ricci flat, and such that $N$ is preserved by the group action, then the quartic $s$ vanishes identically. \end{lemma} \noindent {\bf Proof.} Let $\iota\in \Gamma(\mathbb{S}_+)$ be the covariantly constant spinor defining $N$. Then $\iota$ must be in a linear combination of the covariantly constant basis of $\mathbb{S}_+$ (which exists for ASD metrics iff they are Ricci flat) with constant coefficients (or it can not be parallel). Therefore the null structure $N$ belongs to the hyperboloid (\ref{hyperboloid}) of complex structures defined by the covariantly constant basis. The group $SL(2)$ acts on this hyperboloid, and we require that it fixes $N$. But this implies that it must fix all other points of the hyperboloid (as otherwise the Lie algebra relations would be violated). Therefore the covariantly constant frame is also invariant and $s=0$. \begin{flushright} $\Box $ \end{flushright} \subsection{}Proposition \ref{quartic_lemma} was established by Hitchin who used representation--theoretic arguments \cite{Hitchinis} under an additional assumption that the twistor space admits a real structure which singles out a Riemannian real section of $\mathcal{X}_\mathbb{C}$. In these circumstances the quartic $s$ either vanishes identically, or it admits two repeated roots, or all four roots are distinct. This assumption is not valid in the context of null-K\"ahler structures and split signature metrics. \begin{lemma} If $(\mathcal{X}_\mathbb{C}, g)$ is hyper--Hermitian, and null--K\"ahler, then the metric $g$ is conformal to a Ricci--flat metric. \end{lemma} \noindent {\bf Proof.} We shall use the formulation of the hyper--Hermitian condition due to Boyer \cite{Boyer}, which is also applicable in the complexified setting \cite{D99}: a metric on $\mathcal{X}_\mathbb{C}$ is hyper--Hermitian if and only if there exists a basis $(\Sigma^1, \Sigma^2, \Sigma^2)$ of ${\Lambda^2}_+$ such that \begin{equation} \label{boyer_hc} d\Sigma^{\alpha}=2A\wedge\Sigma^{\alpha}, \quad \alpha=1, 2, 3 \end{equation} for some one--form $A$. Moreover a hyper--Hermitian $g$ is locally conformal to Ricci--flat iff $A$ closed. The formula (\ref{boyer_hc}) together with the isomorphism (\ref{another_iso}) imply the existence of a basis $(o, \rho)$ of $\Gamma(\mathbb{S}_+)$ such that \[ \nabla o=A\otimes o, \quad \nabla \rho=A\otimes \rho. \] Let $\iota\in \Gamma(\mathbb{S}_+)$ be the covariantly constant spinor defining $N$. Then $\iota= h_1 o+h_2 \rho$ for some functions $h_1, h_2$ on $\mathcal{X}_\mathbb{C}$. But then \[ \nabla (h_1 o+h_2\rho)=0 \] gives $h_1=\mbox{const}\cdot h_2$ and $A=-d\ln{(h_1)}$ so $g$ is conformal to a Ricci--flat metric. \begin{flushright} $\Box $ \end{flushright} \vskip 5pt \begin{prop} \label{two_divisors} Let $g$ be an $SL(2)$--invariant cohomogeneity--one metric with anti--self--dual Weyl tensor on $\mathcal{X}_{\mathbb{C}}$, such that the quartic (\ref{quartic}) does not identically vanish. Then the quartic vanishes on each twistor line at one point to order $4$ if and only if there exists a null--K\"ahler structure $N$ (compatible with some metric in the conformal class of $g$) which is Lie--derived by the $SL(2)$ action. \end{prop} \noindent {\bf Proof.} First assume that the quartic $s$ vanishes to order $4$. Therefore its pull-back (\ref{quartic1}) to ${\mathcal F}$ factories as \begin{equation} \label{quartic2} q^*(s)=\Big(\sum_i \iota_i\pi^i\Big)^4, \end{equation} where $[\pi]$ are homogeneous coordinates on the fibres of $\mathbb{P}(\mathbb{S}_+)$ and $\iota$ is a section of $\mathbb{S}_+$. This is a pull--back from ${\mathcal Y}$, so it is constant along the twistor distribution (\ref{tetradlax}), i. e. \begin{equation} \label{laxs} l_i(q^*(s))=0, \quad i=1, 2 \end{equation} which implies that $\iota$ satisfies the rank--one conformally invariant twistor equation \[ \nabla_{i(j}\iota_{k)}=0, \] where $\nabla$ is the spin connection on $\mathbb{S}_+$ induced by the Levi--Civita connection of $g$. In \cite{D02} it was shown that in this case there exists a conformal rescaling \[ \hat{g}=F^2 g, \quad \hat{\iota}= F\iota, \quad \hat{\varepsilon}_+= F\varepsilon_+ \] such that $\hat{\iota}$ is parallel with respect to the Levi--Civita connection of $\hat{g}$. This section defines the null--K\"ahler structure via the isomorphism (\ref{another_iso}). Conversely, let us assume that $N$ is an $SL(2)$--invariant null--K\"ahler structure. Then, by Theorem \ref{mytheo}, $N$ gives rise to a divisor line bundle over ${\mathcal Y}$ given by a canonical section of $\kappa^{-1/4}$. The zero--set of this section pulls back to the hypersurface \begin{equation} \label{omega_zero} \Omega(l_1, l_2)^{1/2}=0 \end{equation} in ${\mathcal F}$, where $\Omega$ is the fundamental two--form of $N$ given by (\ref{hermitean_form}), and $l_1, l_2$ span the twistor distribution. As $N$ is invariant under the $SL(2)$ action, the holomorphic vector fields $\widetilde{R}_\alpha, \alpha=1, 2, 3$ on ${\mathcal Y}$ preserve the canonical section of $\kappa^{-1/4}$, and so are tangent to the surface of vanishing (\ref{omega_zero}), or equivalently \[ \sum_i\pi_i \iota^i=0, \] where $\iota$ is the section of $\mathbb{S}_+$ corresponding, via (\ref{another_iso}), to $N$. Every twistor line intersects this surface at one point given, in homogeneous coordinates, by $[\pi]=[\iota]$. Therefore the quartic $s$ (which by assumption does not vanish identically) vanishes at this point to order 4. \begin{flushright} $\Box $ \end{flushright} \vskip5pt We are now going to use the structure of the twistor distribution (\ref{tetradlax}) to establish Theorem \ref{theorem2} from the introduction. \noindent {\bf Proof of Theorem \ref{theorem2}} Let $t:\mathcal{X}_\mathbb{C}\rightarrow \mathbb{C}$ be a function parametrising the orbits of $G=SL(2)$ in $\mathcal{X}_\mathbb{C}$ such that \[ {\mathcal L}_{R_\alpha} t=0, \quad \alpha=1, 2, 3. \] We can choose coordinates on $G$ such that\footnote{This form is general, but is different from the one usually used (see \cite{tod1, tod2, tod3, Hitchinis}), where the vector field $\partial/\partial t$ is not null, and orthogonal to the orbits of $G$. We shall explain the connection between the two forms (which are equivalent) in \S\ref{sectionequiv}.} \[ g=\sum_{\alpha, \beta}\gamma_{\alpha\beta}(t) \sigma^{\alpha}\otimes\sigma^{\beta} +\sum_{\alpha} n_\alpha(t) (\sigma^{\alpha} \otimes dt+dt \otimes \sigma^{\alpha}), \] where $\gamma$ is a symmetric 3 by 3 matrix and $n$ is a vector with components depending on $t$. We can express the frame $E_{ij}$ in the distribution (\ref{tetradlax}) in terms of the vector field $\partial_t$, and three linearly independent vector fields $P, Q, R$ tangent to $G$ which are $t$--dependent and invariant under left translations. A convenient choice which gives rise to the general metric of the form (\ref{metric_form}) is \begin{equation} \label{tetrad} E_{11} = Q,\quad E_{22} = P, \quad E_{12}= -2\frac{\partial}{\partial t}, \quad E_{21}=2\frac{\partial}{\partial t}-R. \end{equation} The invariance condition implies that the functions $f_1$ and $f_2$ in (\ref{tetradlax}) are constant on $G,$ and so depend only on $\lambda$ and $t$. The quartic $s$ is proportional to $(f_1-\lambda f_2)$. By Proposition (\ref{two_divisors}) it must have a quadrupole zero which can be moved to $\lambda=\infty$ by a M\"obius transformation. Using the freedom in the frame rotations of the frame we set $(f_1=-1, f_2=0)$ so that \[ l_1=Q+2\lambda\frac{\partial}{\partial t}-\frac{\partial}{\partial \lambda}, \quad l_2=2\frac{\partial}{\partial t}-R-\lambda P. \] Now consider a pair of linear combinations of $l_1$ and $l_2$ (\ref{tetradlax}) given by \begin{eqnarray} \label{tetradisolaxL} m_1 &:=& \frac{l_1 -\lambda l_2}{f_1 + \lambda f_2} = \frac{\partial}{\partial \lambda} - Q- \lambda R-\lambda^2 P,\\ \nonumber m_2 &:=& \frac{1}{2}\frac{f_2 l_1 - f_1 l_2}{f_1 + \lambda f_2} = \frac{\partial}{\partial t} -\frac{1}{2}R-\frac{1}{2}\lambda P. \end{eqnarray} Since the conformal class is anti--self--dual, the integrability condition (\ref{intlm}) implies that $[m_1, m_2] =0,$ modulo $m_1$ and $m_2.$ As the Lie bracket $[m_1, m_2]$ does not contain $\partial_\lambda$ or $\partial_t,$ it must be identically zero which yields \begin{equation} \label{isom_airy} \dot{P}=0, \quad \dot{Q}=\frac{1}{2}[R, Q]+\frac{1}{2}P, \quad \dot{R}=\frac{1}{2}[P, Q], \end{equation} where $\dot{P}=dP/dt$ etc. The system (\ref{isom_airy}) underlies the isomonodromic problem with irregular singularity of order four. To make this transparent we shall use the representation of $\mathfrak{sl}(2)$ by 2 by 2 matrices rather than vector fields, and make the replacements \[ L_1\rightarrow \left(\begin{array}{cc} 1/2&0\\ 0&-1/2 \end{array} \right),\quad L_2\rightarrow \left(\begin{array}{cc} 0&1\\ 0&0 \end{array} \right),\quad L_3\rightarrow \left(\begin{array}{cc} 0&0\\ 1&0 \end{array} \right). \] The associated the Lax pair (\ref{tetradisolaxL}) is the isomonodromic Lax pair for Painlev\'e II if $P$ is diagonalisable, and is gauge equivalent to the isomonodromic Lax pair for Painlev\'e I or leads to a solvable equation if $P$ is nilpotent. The system (\ref{isom_airy}) underlies the isomonodromic problem with irregular singularity of order four. To set this problem up consider a $2\times 2$ matrix \[ {\mathcal A}(t, \lambda)=Q+ \lambda R+\lambda^2 P, \] where $\lambda\in\mathbb{CP}^1$, and $P, Q, R$ are elements of a matrix Lie algebra $\mathfrak{g}=\mathfrak{sl}(2)$ which also depend on a parameter $t$. When $t$ is allowed to vary on the complex plane, the matrix fundamental solution $\Psi$ of the ODE \[ \frac{d \Psi}{d \lambda}={\mathcal A} \Psi \] depends on $\lambda$ and $t$. The monodromy around the pole of order four at $\lambda=\infty$ does not depend on $t$ if $\Psi$ satisfies \cite{Jimbo} \begin{equation} \label{over_det} \frac{\partial \Psi}{\partial \lambda}-{\mathcal A}\Psi=0, \quad \frac{\partial \Psi}{\partial t}-{\mathcal B}\Psi=0,\quad \mbox{where}\quad {\mathcal B}:=\frac{1}{2}R+\frac{1}{2}\lambda P. \end{equation} The compatibility conditions \begin{equation} \label{comcon} \partial_t {\mathcal A}-\partial_{\lambda}{\mathcal B}+[{\mathcal A}, {\mathcal B}]=0 \end{equation} for the overdetermined linear system (\ref{over_det}) reduce to system of nonlinear matrix ODEs (\ref{isom_airy}) for $(P, Q, R)$. We shall follow the seminal work of \cite{mason0, mason} - but make different gauge choices - to reduce this system further to a single ODE. There are three gauge equivalence classes to consider. The first two lead to Painlev\'e ODEs and the last one is completely solvable. \begin{itemize} \item If $P$ is diagonalisable, and $\mathfrak{g}=\mathfrak{sl}(2, \mathbb{C})$, then without loss of generality we can take \cite{Jimbo, mason} \[ P=2L_1, \quad R= uL_2-2\frac{z}{u} L_3, \quad Q=(2z+t)L_1-uy L_2 -\frac{2yz+\frac{1}{2}-\alpha}{u} L_3, \] where $u, y$ and $z$ are functions of $t$. Equations (\ref{isom_airy}) become \[ \dot{u}=-yu, \quad \dot{z}=-2yz+\alpha-\frac{1}{2}, \quad\dot{y}=z+y^2+ \frac{t}{2}, \] which imply \begin{equation} \ddot{y}=2y^3+ty+\alpha, \end{equation} where $\alpha$ is a constant parameter. This is the Painlev\'e II equation. \item If $P$ is nilpotent, then (as it is also constant) we can chose it to be $L_2$. Assume that $\mbox{Tr}(PR)\neq 0$, and perform a gauge transformation \[ {\mathcal A}\rightarrow \gamma {\mathcal A}\gamma^{-1}+\partial_\lambda\gamma \cdot \gamma^{-1}, \quad {\mathcal B}\rightarrow \gamma {\mathcal B} \gamma^{-1}+\partial_t\gamma \cdot \gamma^{-1} \] with the group element $\gamma=\gamma(t)$ such that \[ \partial_t \gamma\cdot \gamma^{-1}= yP, \quad\mbox{where}\quad y\equiv\frac{1}{8}\mbox{Tr}(R^2). \] Then \[ P=L_2, \quad R= yL_2+4L_3, \quad Q=-2z L_1+\Big(y^2+\frac{t}{2}\Big)L_2-4yL_3 \] and \[ {\mathcal B}=\frac{\partial}{\partial t}-\frac{1}{2}(R+yL_2)-\frac{1}{2}\lambda P. \] The compatibility conditions (\ref{comcon}) give \begin{equation} \label{PI} \dot{y}=z, \quad \dot{z}=6y^2+t\quad\mbox{so that}\quad \ddot{y}=6y^2+t \end{equation} which is the Painlev\'e I equation. \item Finally let us consider the case where $P$ is nilpotent, and $\mbox{Tr}(PR)$=0. We shall set $P=L_2$, as in the case leading to (\ref{PI}), and consider \[ R=\sum_{\alpha} r_{\alpha}(t) L_{\alpha}, \quad Q=\sum_{\alpha} q_{\alpha}(t) L_{\alpha}, \quad r_3=\mbox{Tr}(PR)=0 \] for some functions $r_\alpha, q_\alpha, \alpha=1, 2, 3$ of $t$. The 3rd equation in (\ref{isom_airy}) gives \[ q_1=-2\dot{r}_2, \quad \dot{r}_1=q_3. \] The 2nd equation in (\ref{isom_airy}) gives \begin{equation} \label{casel3} 2\ddot{r}_1+\dot{r}_1r_1=0, \quad 2\ddot{r}_2+\dot{r}_1r_2=0, \quad 2\dot{q}_2-2r_2\dot{r}_2-q_2r_1-1=0. \end{equation} The first of these equations has a singular solution $r_1=\mbox{const}$ which eventually leads to a degenerate tetrad (\ref{tetrad}). We therefore focus on the regular solution, and absorb two constants of integrations in $r_1$ into an affine transformation of $t$ which results in a constant rescaling of the metric. The remaining two equations can also be solved: \begin{eqnarray*} r_1&=&4\tanh{(t)}, \quad r_2=(a+bt)r_1-4b,\\ q_2&=&\frac{1}{4}\sinh{(2t)}-\frac{d}{dt}\Big((a+bt)r_2\Big)+c\cdot\cosh{(t)}^2, \end{eqnarray*} where $a, b, c$ are the remaining constants of integration. \end{itemize} Now we shall construct the conformal classes corresponding to Painlev\'e I and Painlev\'e II equations, and in each case find a null--K\"ahler metric in the conformal class. These structures will be expressed in terms of left--invariant one--forms (\ref{one_formss}). Each conformal class is represented by a covariant metric dual to (\ref{tetrad_0}) \[ \frac{1}{2}(e^{11}\otimes e^{22}+e^{22}\otimes e^{11}- e^{12}\otimes e^{21}-e^{21}\otimes e^{12}). \] The null--K\"ahler two--form $\Omega$ can be read off from the divisor quartic (\ref{quartic}). In the spinor--form \[ \Omega=\iota\otimes\iota\otimes \varepsilon_-, \] where $\iota\in\Gamma({\mathbb{S}_+}^*)$ is the parallel spinor. When regarded as a section of $\mathbb{P}({\mathbb{S}_+}^*)$ it gives a point on $\mathbb{CP}^1$ which is the quadruple root of the quartic (\ref{quartic}). In our case this gives $\Omega$ proportional to $e^ {11}\wedge e^{21}$. The proportionality factor will be found together with the conformal factor for the metric which makes $\Omega$ parallel. \subsection{} For Painlev\'e II the one--forms dual to the tetrad (\ref{tetrad}) are \begin{eqnarray*} e^{22}&=&\frac{1}{2} \sigma^1+\frac{z(2z+t)}{u(4yz+1-2\alpha)}\sigma^2 +\frac{u\Big(z+\frac{t}{2}\Big)}{4yz+1-2\alpha} \sigma^3,\\ e^{11}&=&-\frac{2z}{u(4yz+1-2\alpha)}\sigma^2- \frac{u}{4yz+1-2\alpha}\sigma^3,\\ e^{21}&=&-\frac{2yz+1-2\alpha}{u(4yz+1-2\alpha)}\sigma^2+ \frac{yu}{4yz+1-2\alpha}\sigma^3,\\ e^{12}&=&-\frac{1}{2}dt-\frac{2yz+1-2\alpha}{u(4yz+1-2\alpha)}\sigma^2+ \frac{yu}{4yz+1-2\alpha}\sigma^3. \end{eqnarray*} The unique conformal factor which makes the null--K\"ahler two--from parallel is \[ k=4yz+1-2\alpha \] so that \begin{equation} \label{piiform1} \Omega=2\sigma^3\wedge\sigma^2 \end{equation} and the metric is given by (\ref{metric_form}) with \begin{eqnarray} \label{piiform} \gamma&=&\left(\begin{array}{ccc} 0&\frac{z}{u}&\frac{u}{2}\\ \frac{z}{u}& \frac{8z^3+(8y^2+4t)z^2+(8-16\alpha)yz+8(\alpha-\frac{1}{2})^3}{ku^2} &-2\frac{2y^2z+(1-2\alpha)y-z(2z+t)}{k} \\ \frac{u}{2}&-2\frac{2y^2z+(1-2\alpha)y-z(2z+t)}{k}&\frac{u^2(2y^2+2z+t)}{k} \end{array} \right),\nonumber\\ n&=&(0, \frac{2yz-2\alpha+1}{2u}, -\frac{yu}{2}). \end{eqnarray} The Weyl tensor is ASD if Painlev\'e II holds. \subsection{} In the case of Painlev\'e I, we first read--off the tetrad form the form of ${\mathcal A}$ and ${\mathcal B}$ to be \[ E_{11}=Q, \quad E_{22}=P, \quad E_{12}=-2\frac{\partial}{\partial t}+yL_2, \quad E_{21}=2\frac{\partial}{\partial t}-(R+yL_2). \] Computing the dual tetrad gives \begin{eqnarray*} e^{22}&=&\frac{y^2+\frac{t}{4}}{z}\sigma^1+\sigma^2-\frac{y}{4}\sigma^3+\frac{y}{2}dt,\\ e^{11}&=&-\frac{1}{2z}\sigma^1,\\ e^{21}&=&\frac{y}{2z}\sigma^1-\frac{1}{4}\sigma^3,\\ e^{12}&=&\frac{y}{2z}\sigma^1-\frac{1}{4}\sigma^3-\frac{1}{2}dt \end{eqnarray*} and rescalling the resulting metric and two--form by $k=16z(t)$ gives \begin{equation} \label{piform1} \Omega=2\sigma^3\wedge\sigma^1. \end{equation} This is the only scaling factor which makes $\Omega$ closed. Using the same conformal factor for the metric yields and $g$ in the form (\ref{metric_form}) with \begin{equation} \label{piform} \gamma=-\left(\begin{array}{ccc} \frac{12y^2+2t}{z}&4&-3y\\ 4&0&0\\ -3y&0&z \end{array} \right),\quad n=(0, 0, -z). \end{equation} The two--form $\Omega$ is now parallel, as \[ \nabla\Omega=\Big(\frac{6y^2+t-\dot{z}}{z}\sigma^1+ \frac{3(z-\dot{y})}{z}\sigma^2\Big)\otimes\Omega=0 \] where we used (\ref{PI}). The null--K\"ahler structure is given by \[ N=\frac{1}{2}\Big(\sigma^3\otimes L_2-\frac{2}{z}\sigma^1\otimes\frac{\partial}{\partial t}\Big). \] Computing the self--dual part of the Weyl tensor we find that it vanishes as a consequence of Painlev\'e I \subsection{} Computing the dual tetrad in the case (\ref{casel3}) gives \begin{eqnarray*} e^{22}&=& (b\coth{(t)}-(a+bt))\sigma^1+\sigma^2\\ &&+ \Big(2b(a+bt)\coth{(t)}+b^2 \cosh{(t)}^2 -\frac{1}{8}\sinh{(t)}\cosh{(t)}^3-(a+bt)^2-\frac{c}{4}\cosh{(t)}^2 \Big)\sigma^3 ,\\ e^{11}&=&\frac{1}{4}\cosh{(t)}^2\sigma^3 ,\\ e^{21}&=&-\frac{1}{4} \coth{(t)}\sigma^1-\frac{1}{2}\Big(b\cosh{(t)}^2+(a+bt)\coth{(t)}\Big)\sigma^3 ,\\ e^{12}&=&-\frac{1}{2}dt -\frac{1}{4}\coth{(t)}\sigma^1-\frac{1}{2}\Big(b\cosh{(t)}^2+(a+bt)\coth{(t)}\Big)\sigma^3 \end{eqnarray*} and rescaling the resulting metric and two--form by $k=8\sinh{(t)}/\cosh{(t)}^3$ gives \[ \Omega=\sigma^3\wedge\sigma^1. \] This is the only scaling factor which makes $\Omega$ closed. Using the same conformal factor for the metric yields $g$ in the form (\ref{metric_form}) with \begin{eqnarray} \label{solv_form} \gamma&=&\left(\begin{array}{ccc} \frac{2}{\sinh{2t}} &0&(a+bt)\coth{(2t)}\\ 0&0&-\tanh{(2t)}\\ (a+bt)\coth{(2t)} & -\tanh{(2t)} & 4(a+bt)^2\coth{(t)}+\frac{1}{8}\sinh{(2t)}^2(1+2c\cdot \coth{(t)}) \end{array} \right),\nonumber\\ n&=&(\cosh{(t)}^{-2}, 0, 2(a+bt)\cosh{(t)}^{-2}+2b\tanh{(t)}). \end{eqnarray} The two--form $\Omega$ is parallel and the Weyl tensor is anti--self--dual. The Ricci--tensor is \[ r=\frac{1}{2}\sinh{(t)}\cosh{(t)}^3\sigma^3\otimes\sigma^3. \] \begin{flushright} $\Box $ \end{flushright} {\bf Remarks} \subsection{} All solutions to PI are transcendental, and PII admits solutions expressible in terms of know functions only for integer and half--integer values of the parameter $\alpha$. These will lead to explicit metrics. On the other hand all metrics arising from (\ref{solv_form}) are explicit. The simplest is obtained by setting $a=b=c=0$. Setting $\tau=\tanh{(t)}$ gives \[ g=\frac{1-\tau^2}{\tau}\sigma^1\otimes\sigma^1 +\sigma^1\otimes d\tau+d\tau\otimes\sigma^1- 2\tau(\sigma^2\otimes\sigma^3+\sigma^3\otimes\sigma^2)+ \frac{\tau^2}{2(1-\tau^2)^2}\sigma^3\otimes\sigma^3. \] \subsection{} The null--Kahler structures arising from PI and and PII can be distinguished by examining the restriction of kernel of the endomorphism $N$ to the orbits of $SL(2)$. In the PI case, this kernel - when regarded as the element of the Lie algebra of $SL(2)$ is nilpotent, but in the PII case it is not. This can be seen directly from (\ref{piform1}) and (\ref{piiform1}). \subsection{} \label{sectionequiv} The usual form of cohomogeneity--one metrics \cite{tod1, tod2, tod3, Hitchinis} is \[ g=\frac{1}{4}dt^2+\sum_{\alpha, \beta}h_{\alpha\beta}(t)(\sigma^\alpha\otimes\sigma^\beta+\sigma^\beta\otimes \sigma^\alpha). \] This arises from a frame of the form \[ E_{11}=Q,\quad E_{22}=P,\quad E_{12}=-2\frac{\partial}{\partial t}-\frac{1}{2}R,\quad E_{21}=2\frac{\partial}{\partial t}-\frac{1}{2}R. \] Following the argument above which lead to the Lax pair (\ref{tetradisolaxL}) gives the system of ODEs \cite{CD} \[ \dot{Q}=\frac{1}{4}[R, Q]+\frac{1}{2}P, \quad \dot{R}=\frac{1}{2}[P, Q], \quad \dot{P}=\frac{1}{4}[P, R] \] together with the isomonodromic Lax pair $[m_1, m_2]=0$, where \begin{equation} \label{system_new} m_1=-l_1+\lambda l_2=\partial_\lambda-{\mathcal A}, \quad m_2=\frac{1}{2}l_2=\partial_t-{\mathcal B} \end{equation} where now \[ {\mathcal A}=Q+\lambda R+\lambda^2 P, \quad {\mathcal B}=\frac{1}{4}R+\frac{1}{2}\lambda P. \] The gauge transformation \[ {\mathcal A}\rightarrow \gamma {\mathcal A} \gamma^{-1}+\partial_\lambda\gamma\cdot\gamma^{-1}, \quad {\mathcal B}\rightarrow \gamma {\mathcal B} \gamma^{-1}+\partial_t\gamma\cdot\gamma^{-1} \] with $\gamma=\gamma(t)$ such that $\gamma^{-1}\cdot\dot{\gamma}=\frac{1}{4}R$ brings this Lax pair to the form (\ref{tetradisolaxL}), and the system (\ref{system_new}) to the form (\ref{isom_airy}). Indeed, we can verify that $\widetilde{P}\equiv \gamma \cdot P\cdot\gamma^{-1}$ is constant, and the other two equations also hold with $\widetilde{Q}\equiv \gamma \cdot Q\cdot\gamma^{-1}$ and $\widetilde{R}\equiv \gamma \cdot R\cdot\gamma^{-1}$. Thus the two forms of the metric are equivalent by a diffeomorphism. \subsection{} The isomonodromic Lax pair (\ref{tetradisolaxL}) arising as the combination of $(l_1, l_2)$ can be constructed invariantly using the notation from the proof of Proposition \ref{quartic_lemma} as follows: Let $T$ be the vector field dual to $dt$ with respect to the isomorphism between tangent and cotangent bundle given by the metric. This vector is normal to the orbits of $SL(2)$, and is given by $T=\sum_{i, j} T^{ij} E_{ij}$ in the frame $E_{ij}$. First note that \[ l_i(t)=\sum_j \pi^j T_{ij}, \quad\mbox{so that}\quad \sum_{i, j} T^{ij}\pi_j l_i(t)=\sum_{i, j, k, m} \pi_i\pi_j\varepsilon_{km}T^{ki}T^{mj}=0 \] by symmetry. This implies that \[ \sum_{i, j} T^{ij}\pi_jl_i=\sum_{i, j} T^{ij}\pi_j f_i\frac{\partial}{\partial\lambda}+\sum_{i, j, k} T^{ij}\pi_j\pi^k E_{ik}, \] where the second term on the RHS is tangent to the orbits of $SL(2)$, so is in the span of the left--invariant vector fields and does not contain $\partial/\partial t$. Using (\ref{new_form}) we identify the multiple of $\partial/\partial \lambda$ as the quartic $s$. Assuming that this quartic is not identically zero we define an $\mathcal{O}(-2)$--valued vector field \[ m_1=\frac{\sum_{i, j}T^{ij}\pi_j l_i}{s}. \] The second vector field $m_2$ does not contain the $\partial/\partial\lambda$ terms, and is defined by \[ m_2=\frac{1}{2}\frac{\sum_{i, j} \varepsilon^{ij}f_i l_j}{s}. \] This agrees with (\ref{tetradisolaxL}). If the metric is given by (\ref{metric_form}), then \[ T=-\frac{1}{\sum_{\alpha, \beta}\gamma^{\alpha\beta} n_\alpha n_\beta}\Big(\frac{\partial}{\partial t}-\sum_{\alpha, \beta}\gamma^{\alpha\beta} n_\alpha L_\beta\Big), \] where $\gamma^{\alpha\beta}(t)$ is the inverse matrix of $\gamma_{\alpha\beta}(t)$. \subsection{} Selecting a one--parameter family of transformations $\mathbb{R}^{*}$ in $SL(2, \mathbb{R})$ generated by a non--null Killing vector $K$ reduces the null--K\"ahler ASD conditions to an Einstein--Weyl structure in 2+1 dimension which additionally admits a parallel weighted null vector field. Such structures correspond to solutions of the dispersionless Kadomtsev-Petviashvili equation, and in \cite{DTT}, such solutions were constructed and shown to be constant on central quadrics and expressed in terms of solutions to Painlev\'e I or Painlev\'e II. \subsection{} If the Lie algebra $\mathfrak{g}$ underlying (\ref{isom_airy}) is instead the Bianchi II algebra then the isomonodromic condition is the (derivative of ) the Airy equation (see \cite{CD}, where a class of ASD null--K\"ahler four manifolds has been constructed). In \cite{Hitchinis, mmw} it was instead assumed that the quartic $s$ has four distinct zeros, and that $G=SL(2, \mathbb{C})$ which lead to the isomonodromic Lax pair \cite{Jimbo} for Painlev\'e VI. If $s$ has two double zeroes then the conformal class contains an Einstein metric \cite{tod2, Hitchinis}, and the isomonodromic Lax pair leads to Painlev\'e III.

https://arxiv.org/abs/0706.2348

Analytic linearization of nonlinear perturbations of Fuchsian systems

Nonlinear perturbation of Fuchsian systems are studied in regions including two singularities. Such systems are not necessarily analytically equivalent to their linear part (they are not linearizable). Nevertheless, it is shown that in the case when the linear part has commuting monodromy, and the eigenvalues have positive real parts, there exists a unique correction function of the nonlinear part so that the corrected system becomes analytically linearizable.

\section{Introduction}\label{Introduction} \subsection{Setting.}\label{Setting} The paper studies linearization criteria of nonlinear perturbations of Fuchsian systems of the form \begin{equation}\label{perFuchs} \mathcal{E}[\mathbf{f}]:\ \ \ \ \ \ \ \ \frac{d\mathbf{u}}{dx}=M(x)\mathbf{u}+\frac{1}{x^2-1}\mathbf{f}(x,\mathbf{u})\ \ \ \ \ \ \ \ (\mathbf{w}\in\mathbb{C}^d,\ x\in\mathbb{C}) \end{equation} (Systems (\ref{perFuchs}) are denoted by $\mathcal{E}[\mathbf{f}]$ to distinguish among them by their nonlinear part.) The linear part \begin{equation}\label{genLinFuchSys} \mathcal{E}[0]:\ \ \ \ \ \ \ \ \ \ \ \frac{d\mathbf{w}}{dx}=M(x)\mathbf{w}\ \ \ \ \ \ \ \ (\mathbf{w}\in\mathbb{C}^d,\ x\in\mathbb{C}) \end{equation} is assumed to be a Fuchsian system with three singularities in the extended complex domain. Their location can be arbitrarily placed using a rational linear transformation, and in the present paper we assume they are located at $x=1,x=-1$ and $\infty$, therefore \begin{equation}\label{formM} M(x)=\frac{1}{x-1}\, A\, +\, \frac{1}{x+1}\, B\ \ \ \ \ \ \ \,A,B\in\mathcal{M}_d(\mathbb{C}) \end{equation} It is assumed that the matrices $A$ and $B$ commute, and formal linearization results are obtained; convergence is proved under the supplementary assumption that the eigenvalues of $A,B$ have positive real parts. The function $\mathbf{f}(x,\cdot)$ collects the nonlinear terms in $\mathbf{u}$: it has a zero of order two at $\mathbf{u}=0$ and $\mathbf{f}(x,\mathbf{u})$ is analytic for $x$ in a simply connected domain $D\ni\{\pm 1\}$ (which will be assumed large enough) and for $\mathbf{u}\in\mathbb{C}^d$, $|\mathbf{u}|<R$ (for some $R>0$). Note that the nonlinear term, $\frac{1}{x^2-1}\mathbf{f}(x,\mathbf{u})$, is therefore allowed to have at most simple poles at the singularities $x=\pm 1$ of $M(x)$. Denote also \begin{equation}\label{defQ} Q(x)=x^2-1 \end{equation} \subsection{Motivation.}\label{Motivation} The problem of linearization and the more general, of equivalence, are fundamental in the theory of differential equations, and for many applications and many classes of problems can be reduced to systems of the form (\ref{perFuchs}) - see \cite{Norm_Form} for references. The case when the domain $D$ contains just one singular point of $M(x)$ was studied in \cite{Nonln}: generic such systems are linearizable (see \S\ref{rev1s} for details). When the domain contains two singularities conditions for formal linearizability of (\ref{perFuchs}), (\ref{formM}) were found in \cite{Norm_Form}, under polynomiality assumptions on the nonlinear part, and not requiring commuting monodromy. The present paper shows analytic linearization of canonically corrected systems for general analytic nonlinear part, in the case of commuting monodromy, and positive real parts of eigenvalues of $A$ and $B$. \subsection{Existence of formal corrections and formal normal forms: results proved in \cite{Norm_Form}.}\label{Prior_results} \subsubsection{Assumptions.}\label{eq} \ (a) The function $\mathbf{f}(x,\mathbf{u})$, which collects the nonlinear terms, has a zero of order two at $\mathbf{u}=0$, and is analytic for $x$ in the simply connected domain $D\ni\{\pm 1\}$ and for $\mathbf{u}\in\mathbb{C}^d$, $|\mathbf{u}|<R$. (b) The eigenvalues $\boldsymbol{a}=(a_1,\ldots ,a_d)$ of $A$, and respectively $\boldsymbol{b}=(b_1,\ldots ,b_d)$ of $B$, satisfy the Diophantine condition: there exist $C,\nu>0$ so that \begin{equation}\label{DioCond} \Big| \mathbf{n}\cdot\boldsymbol{a} +l-a_s\Big|>C\left( |\mathbf{n}|+|l|\right)^{-\nu}\ \ \ {\mbox{and}}\ \ \ \Big| \mathbf{n}\cdot\boldsymbol{b} +l-b_s\Big|>C\left( |\mathbf{n}|+|l|\right)^{-\nu} \end{equation} for all $l\in\mathbb{N}$, $s\in\{1,...,d\}$, and $\mathbf{n}\in\mathbb{N}^d$ with $ |\mathbf{n}|\geq 2$, with the notation $$\mathbf{n}=(n_1\ldots n_d),\ \ \ \ \ \mathbf{n}\cdot\boldsymbol{a}=n_1a_1+\ldots+n_da_d,\ \ \ \ \ |\mathbf{n}|=n_1+\ldots+n_d$$ (c) No eigenvalue of $A$, or of $B$, is an integer: $a_i,b_i\not\in\mathbb{Z},\ i=1,\ldots ,d$. (d) The eigenvalues $\lambda_1,\ldots ,\lambda_d$ of the matrix $A+B$ satisfy the following nonresonance condition: \begin{equation}\label{non_res} k+\mathbf{n}\cdot {\boldsymbol{\lambda}}-\lambda_j \not=0\ \ \ {\mbox{for\ all}}\ \ \mathbf{n}\in\mathbb{N}^d\ ,\ k\in\mathbb{N}\ ,\ \ j=1,\ldots ,d \end{equation} \subsubsection{Obstructions to linearization and formal correction.}\label{formal} The following result was proved in \cite{Norm_Form}. Note that there are no commutativity assumptions on $A$ and $B$. \begin{Theorem}\label{Obstructions} Consider the system (\ref{perFuchs}) under the assumptions of \S\ref{eq}. Assume that the nonlinear part is of polynomial type in $x$, in the sense that \begin{equation}\label{pol_type} {\mathbf{f}}(x,\mathbf{u})=\sum_{|\mathbf{m}|\geq 2}\mathbf{f}_\mathbf{m}(x)\mathbf{u}^\mathbf{m}\ \ \ \ \ {\mbox{with}}\ \mathbf{f}_\mathbf{m}(x)\ {\mbox{polynomials}} \end{equation} with the usual notation $$\mathbf{u}^\bfm=u_1^{m_1}u_2^{m_2}\ldots u_d^{m_d}\ \ \ {\mbox{for\ }} \mathbf{u}=(u_1,u_2,\ldots ,u_d)\in\mathbb{C}^d,\ \bfm\in\mathbb{N}^d$$ Then {\em{there exists a unique}} correction $\boldsymbol{\phi}(\mathbf{u})$ of $\mathbf{f}(x,\mathbf{u})$ as a formal series \begin{equation}\label{serphi} \boldsymbol{\phi}(\mathbf{u})=\sum_{|\mathbf{n}|\geq 2} \boldsymbol{\phi}_\mathbf{n} \mathbf{u}^\mathbf{n} \end{equation} so that the correction $\mathcal{E}[\mathbf{f}-\boldsymbol{\phi}]$ of $\mathcal{E}[\mathbf{f}]$: \begin{equation}\label{corperFuchs} \mathcal{E}[\mathbf{f}-\boldsymbol{\phi}]:\ \ \ \ \ \frac{d\mathbf{u}}{dx}=\left(\frac{1}{x-1}\, A\, +\, \frac{1}{x+1}\, B\right)\mathbf{u}+\frac{1}{x^2-1}\left[\mathbf{f}(x,\mathbf{u})-\boldsymbol{\phi}(\mathbf{u})\right] \end{equation} is linearizable by a formal series transformation \begin{equation}\label{serh} {\mathbf{u}=\mathbf{H}(x,\mathbf{w})=\mathbf{w}+\sum_{|\mathbf{m}|\geq 2}\mathbf{h}_\bfm(x)\mathbf{w}^\mathbf{m}} \end{equation} where $\mathbf{h}_\bfm(x)$ are functions analytic on $D$ (in fact, they are polynomials). \end{Theorem} Note that the coefficients $\boldsymbol{\phi}_\mathbf{n}$, $\mathbf{n}\in\mathbb{N}^d, |\mathbf{n}|\geq 2$ represent the obstructions to linearization, in the sense that the equation $\mathcal{E}[\mathbf{f}]$ is linearizable if and only if all $\boldsymbol{\phi}_\mathbf{n}$ are zero. \subsubsection{Remark.} Obviously, formal linearization is a necessary condition for analytic linearization. \section{Main Results.}\label{Main_res} While the assumptions of \S\ref{eq} are essential for the results of Theorem\,\ref{Obstructions} to hold, it is natural to expect that the condition (\ref{pol_type}) that $\mathbf{f}$ be of polynomial type in $x$ to be generalizable to holomorphic functions of $x$. It is also to be expected that the series (\ref{serphi}) and (\ref{serh}) converge if the eigenvalues of $A+B$ are not "too close to resonance" (see (\ref{non_res})). The present paper considers the case when the matrices $A$ and $B$ commute, and when their eigenvalues have positive real parts. Under these assumptions there are no polynomiality assumptions on the nonlinear part $\mathbf{f}$, which can be any holomorphic function (on a domain large enough). It is shown that a unique formal correction $\boldsymbol{\phi}(\mathbf{u})$ still exist {{for functions $\mathbf{f}(x,\mathbf{u})$ analytic in $x$}} - Theorem \ref{anf}, and furthermore, that the linearization series (\ref{serh}), and the series (\ref{serphi}), converge, hence we have {\em{analytic linearization }}- Theorem \ref{ConvSer}. \subsection{Setting.}\label{eq1} Consider the equation $\mathcal{E}[\mathbf{f}]$ given by (\ref{perFuchs}),(\ref{formM}) under the assumptions of \S\ref{eq}. Furthermore, it is assumed that $A$ and $B$ are simultaneously diagonalizable: \begin{equation}\label{ass_diag} A={\mbox{diag}} (a_1,\ldots ,a_d),\ \ \ \ \ B={\mbox{diag}} (b_1,\ldots ,b_d) \end{equation} Theorem\,\ref{ConvSer} is proved under the additional assumption \begin{equation}\label{rhp} \Re a_j>0,\ \ \ \ \ \Re b_j>0\ \ \ \ \ {\mbox{for\ all\ }}j=1,\ldots, d \end{equation} \subsection{Obstructions to linearization - equations with analytic nonlinear terms.}\label{SecObs} \begin{Theorem}\label{anf} Consider equation $\mathcal{E}[\mathbf{f}]$ given by (\ref{perFuchs}),(\ref{formM}) under the assumptions of \S\ref{eq1}. Let $\mathbf{f}$ be holomorphic for $x\in D$ and $|\mathbf{u}|<R$. Then{\em{ there exists a unique}} "correction" $\boldsymbol{\phi}(\mathbf{u})$ of $\mathbf{f}(x,\mathbf{u})$ as a formal series (\ref{serphi}) so that the "corrected" equation $\mathcal{E}[\mathbf{f}-\boldsymbol{\phi}]$ - see (\ref{corperFuchs}) - is linearizable by a formal series (\ref{serh}) with coefficients $\mathbf{h}_\bfm(x)$ holomorphic on $D$. \end{Theorem} \subsection{Analytic linearization.} Theorem\,\ref{ConvSer} shows that the series in Theorem\,\ref{anf} converge, therefore any nonlinear term $\mathbf{f}$ can be corrected to an analytic function so that the resulting equation is analytically linearizable. \subsubsection{{The numbers $c$ and $\rho_{min}$.}}\label{rho_min} These quantities are used in the statement of Theorem\,\ref{ConvSer} to define a domain $D$ (where the nonlinear term will be required to be analytic) and are defined as follows. Denote \begin{equation}\label{cm} c_\bfm=\frac{(\bfm/n)\cdot(\mathbf{b}-\mathbf{a})}{(\bfm/n)\cdot(\mathbf{b}+\mathbf{a})}\ \ \ \ \ \ \ {\mbox{for}}\ \ \ \bfm\in\mathbb{N}^d,\ n=|\bfm|\geq 2 \end{equation} The set of all points $c_\bfm$ given by (\ref{cm}) belong to a compact set $K$: $$K=\left\{\, g(\mathbf{t})=\frac{\mathbf{t}\cdot(\mathbf{b}-\mathbf{a})}{\mathbf{t}\cdot(\mathbf{b}+\mathbf{a})}\, ;\, \mathbf{t}\in [0,1]^d,\, |\mathbf{t}|=1\,\right\}$$ (In fact the boundary of $K$ consists of arcs of the circles $g(\mathbf{t})$ obtained for $\mathbf{t}$ having only two nonzero coordinates.) Let $c\in K$ and let $\rho_{min}$ positive and large enough so that the disk $|x-c|<\rho_{min}$ contains the points $-1$, $1$ and the set $K$. \subsubsection{Analytic linearization theorem} \begin{Theorem}\label{ConvSer} Consider the system (\ref{perFuchs}),(\ref{formM}) under the assumptions of \S\ref{eq1}. Let $c$ and $\rho_{min}$ as in \S\ref{rho_min}. Assume that the nonlinear part $\mathbf{f}(x,\mathbf{u})$ is holomorphic for $|\mathbf{u}|<R_0$ and $x$ in a disk $|x-c|<\rho_0$ where $\rho_0>\rho_{min}$. Then the series (\ref{serphi}) and (\ref{serh}) converge in a subdomain $|\mathbf{u}|<R_e$ and $|x-c|<\rho_e$ ($\rho_{min}<\rho_e<\rho_0$, $0<R_e<R_0$). Furthermore, $\rho_e$ can be made arbitrarily close to $\rho_0$ if $R_e$ is small enough. \end{Theorem} \subsubsection{Comments regarding the proofs.} {\em{(i)}} The proof of Theorem\,\ref{Obstructions} in \cite{Norm_Form}, which is done under polynomiality assumptions of the nonlinear terms, relies on an algebraic structure: the terms are expanded in a base of polynomials in $x$ satisfying a generalized Rodrigues formula. Under the additional commutativity assumption of the present paper, these special polynomials are Jacobi polynomials (see also \cite{Gen_Jacobi}). The natural approach to proving {Theorem}\,\ref{ConvSer} seems to be by expansions in Jacobi series in the variable $x$, on an elliptic domain $D$ with foci at $x=\pm 1$, see \cite{Carlson}, and the condition (\ref{rhp}) could be replaced by a Diophantine condition in (\ref{non_res}). However, some technical results needed for estimates are not available yet, or at least, are not known to the author. The proof of convergence is done here using estimates on a disk $D$ by saddle-point and Laplace's method, \S\ref{PhiJ}. To avoid small denominator problems a rapidly convergent iteration is used in the proof. The estimates of \S\ref{homeq}-\S\ref{convseqem} follow the line in \cite{Arnold} \S 28. Additional refinements were needed in the proof of convergence of the correction (\ref{serphi}) in \S\ref{convcorr}. {\em{(ii)}} The fact that a formally corrected system which is formally linearizable turns out to be analytically linearizable under the positivity assumption (\ref{rhp}) is a result in the spirit of the Poincar\'e-Dulac theorem (see e.g. \cite{Arnold}). It should be noted that this theorem cannot be used in the present context because after subtracting the formal correction $\boldsymbol{\phi}(\mathbf{u})$ the equation is only presented as a formal series, hence not known to be analytic. The convergence of both the linearization map and of the correction series is proved here simultaneously. \section{Proofs.}\label{Proofs} \subsection{Proof of Theorem \ref{anf}.}\label{PfOb} \subsubsection{Obstructions to linearization.}\label{recsys} Section \S\ref{recsys} contains the first steps in the proof. They follow \cite{Norm_Form} and are reproduced here for completeness. If a map (\ref{serh}), $\mathbf{u}=\mathbf{H}(x,\mathbf{w})=\mathbf{w}+\mathbf{h}(x,\mathbf{w})$, is a linearization map of (\ref{perFuchs}) then then it satisfies the nonlinear partial differential equation \begin{equation}\label{nonlinPDE} \partial_x\mathbf{h}+d_\mathbf{w}\mathbf{h}\, M\mathbf{w}-M\mathbf{h}=\frac{1}{x^2-1}\, \left[\mathbf{f}(x,\mathbf{w}+\mathbf{h})-\boldsymbol{\phi}(\mathbf{w}+\mathbf{h})\right] \end{equation} where $M$ is given by (\ref{formM}). Denote by $\mathbf{h}_n$ the homogeneous part of degree $n$ of the function $\mathbf{h}(x,\mathbf{w})$ \begin{equation} \mathbf{h}_n(x,\mathbf{w})=\sum_{|\mathbf{m}|=n}\mathbf{h}_\mathbf{m}(x)\mathbf{w}^\mathbf{m} \end{equation} and use a similar notation for other functions ($\mathbf{f}(x,\mathbf{w})$ etc.). Expanding in power series in $\mathbf{w}$ equation (\ref{nonlinPDE}) yields a recursive system for $\mathbf{h}_n$: \begin{equation}\label{eqhn} \partial_x\mathbf{h}_n+d_\mathbf{w}\mathbf{h}_n\, M\mathbf{w}-M\mathbf{h}_n=\frac{1}{x^2-1}\, \mathbf{R}_n(x,\mathbf{w})\ \ \ \ \ (n\geq 2) \end{equation} where \begin{equation}\label{formRn} \mathbf{R}_n=\mathbf{f}_n-\boldsymbol{\phi}_n+\tilde{\mathbf{R}}_n \end{equation} with $\tilde{\mathbf{R}}_n$ a polynomial in $\boldsymbol{\phi}_\mathbf{m}$, $\mathbf{h}_\bfm$, $\mathbf{f}_\mathbf{m}$ with $|\mathbf{m}|<n$. Denote by $Y(x)$ a fundamental matrix for the linear part: $Y'=MY$. Using the variation of parameters formula for a linear nonhomogeneous equation, and choosing the solution which is not branched at $x=-1$ we obtain \begin{equation}\label{hnat-1} \mathbf{h}_n(x,\mathbf{w})=Y(x)\int_{-1}^xQ(t)^{-1}Y(t)^{-1}\mathbf{R}_n(t,Y(t)Y(x)^{-1}\mathbf{w})dt \end{equation} is a particular solution of (\ref{eqhn}), and this solution is analytic at $x=-1$ (see the Appendix, \S\ref{details} for details). Rewriting (\ref{hnat-1}) as \begin{equation}\label{splithn} \mathbf{h}_n(x,\mathbf{w})=Y(x)\int_{-1}^1Q^{-1}Y^{-1}\mathbf{R}_ndt+Y(x)\int_{1}^xQ^{-1}Y^{-1}\mathbf{R}_ndt \end{equation} the last term of (\ref{splithn}) is the unique linearization map of (\ref{perFuchs}) which is analytic at $x=1$. Then the linearization map $\mathbf{h}$ is analytic at both $x=1$ and $x=-1$ if and only if the first term in right hand side of (\ref{splithn}) vanishes: \begin{equation}\label{obcond} \int_{-1}^1Q(t)^{-1}Y(t)^{-1}\mathbf{R}_n(t,Y(t)Y(x)^{-1}\mathbf{w})\, dt\, =\, 0\ \ \ {\mbox{for\ all\ }}\mathbf{w}\in\mathbb{C}^d,\ n\geq 2 \end{equation} Formulas (\ref{obcond}) show the obstructions to linearization: there is, for every $\mathbf{m}\in\mathbb{N}^d, |\mathbf{m}|\geq 2$, one numerical condition (vector-valued in $\mathbb{C}^d$). \subsubsection{Existence of the correction $\boldsymbol{\phi}(\mathbf{u})$.}\label{extphi} Using (\ref{formRn}) in (\ref{obcond}) we obtain recursively for $n\geq 2$ equations for $\boldsymbol{\phi}_n(\mathbf{w})$ of the form \begin{equation}\label{sysphi1} \int_{-1}^1Q(t)^{-1}Y(t)^{-1}\boldsymbol{\phi}_n(Y(t)Y(x)^{-1}\mathbf{w})\, dt\, =\mathbf{F}_n(x,\mathbf{w}) \end{equation} where $\mathbf{F}_n(x,\mathbf{w})$ are homogeneous polynomials in $\mathbf{w}$ degree $n$, with coefficients vector-valued functions, analytic in $x$ on $D$. The matrices $A$ and $B$ being given by (\ref{ass_diag}), then $Y(x)$ is the diagonal matrix: $$Y(x)={\mbox{diag}}\, \left[ y_1(x), \ldots , y_d(x)\right]\ \ \ \ {\mbox{where}}\ y_j(x)=(x-1)^{a_j}(x+1)^{b_j}$$ and (\ref{sysphi1}) becomes \begin{equation}\label{matsys} \sum_{|\bfm|=n}\, \frac{1}{\mathbf{y} (x)^\bfm}\, \int_{-1}^1\, \, (t-1)^{\bfm\cdot\mathbf{a} -1}(t+1)^{\bfm\cdot\mathbf{b} -1}Y(t)^{-1}\, dt\, \boldsymbol{\phi}_\bfm\mathbf{w}^\bfm\, =\, \sum_{|\bfm|=n}\, \mathbf{F}_\bfm(x)\mathbf{w}^\bfm \end{equation} Equation (\ref{matsys}) is a linear system for $\{\boldsymbol{\phi}_\bfm\}_{|\bfm|=n}$. The left-hand-side of (\ref{matsys}) is a diagonal linear operator having as entries Eulerian integrals of the first kind $$\frac{1}{\mathbf{y} (x)^\bfm}\, \int_{-1}^1\, \, (t-1)^{\alpha_{\bfm,j} -1}(t+1)^{\beta_{\bfm,j} -1}\, dt\, =\, \frac{(-1)^{\alpha_{\bfm,j} -1}2^{\alpha_{\bfm,j}+\beta_{\bfm,j}-1}B(\alpha_{\bfm,j},\beta_{\bfm,j})}{\mathbf{y} (x)^\bfm}$$ where $\alpha_{\bfm,j}= \bfm\cdot\mathbf{a}-a_j$, $\beta_{\bfm,j}= \bfm\cdot\mathbf{b}-b_j$, and $B(p,q)$ is the beta function \begin{equation}\label{Bfun} B(p,q)=\int_{0}^1s^{p-1}(1-s)^{q-1}ds=\frac{\Gamma(p)\Gamma(q)}{\Gamma(p+q)} \end{equation} Therefore equation (\ref{matsys}) has a unique solution if the nonresonance condition (\ref{non_res}) is satisfied, and Theorem \ref{anf} is proved. \qed \subsection{Proof of {Theorem} \ref{ConvSer}.} The proof is written in dimension one, for simplicity of notation. The only (minor) difference for dimension $d>1$ is mentioned in \S\ref{ddim}. \subsubsection{Initial steps.} Note that in view of Theorem\,\ref{anf} it can be assumed that $f_n=0$ for $2\leq n< N$ for any chosen $N>2$. It follows that also $\phi_n=0$ and $h_n=0$ for $n<N$. \subsubsection{The functionals $\Phi_n$ and the operators $J_n$.}\label{PhiJ} We will consider $\rho$ in a closed interval $I=[{\rho}_{min},\rho_{0}] $ and $\delta\in (0, 1/2)$. Let $c=\frac{b-a}{b+a}$. Denote $D_\rho=\{x\in\mathbb{C}\, ;\, |x-c|<\rho\}$. Let $\mathcal{B}_\rho$ be the Banach space of functions analytic on $D_\rho$, continuous on $\overline{D_\rho}$, with the sup norm: $$\|F\|=\sup_{x\in D_\rho}|F(x)| $$ Denote for simplicity the sup norm on $\mathcal{B}_{\rho\rm{e}^{-\delta}}$ by $ \|\cdot \|_\delta$: $$ \|F\|_\delta=\sup_{x\in D_{\rho\rm{e}^{-\delta}}}|F(x)|$$ \noindent Note that using a Cauchy integral formula (see also Remark\,\ref{R6}) we have \begin{equation}\label{estimF'} \|F'\|_\delta \leq \frac{\|F\|}{\rho\left( 1-\rm{e}^{-\delta}\right)}\leq \frac{2}{\rho\delta}\, \|F\| \end{equation} Let $n\geq 1$. For $F\in\mathcal{B}_\rho$ define the linear functionals \begin{equation}\label{defPhi} \Phi_{n+1}[F]=\frac{2^{1-n(a+b)}}{B(na,nb)}\, \int_{-1}^1 (1-t)^{na-1}(1+t)^{nb-1}\, F(t)\, dt \end{equation} \noindent (see (\ref{Bfun})) and the linear operators \begin{equation}\label{opJ} J_{n+1}[F](z)=(1-z)^{-na}(1+z)^{-nb}\, \int_{-1}^z\, (1-t)^{na-1}(1+t)^{nb-1}\, \left( F(t)-\Phi_{n+1}[F]\right)\, dt \end{equation} where $\Phi_{n+1}[F]$ is given by (\ref{defPhi}) and therefore $J_{n+1}[F]$ is a function analytic at both endpoints $z=\pm 1$ (see \S\ref{recsys} and \S\ref{extphi}) Note that \begin{equation}\label{invalPJ} \Phi_n[1]=1\ ,\ J_n[1]=0\ ,\ \Phi_n[x-c]=0\ ,\ J_{n+1}[x-c]=-\frac{1}{n(a+b)} \end{equation} Denote \begin{equation}\label{tildenot} F(x)=\sum_{k\geq 0}F_k(x-c)^k=F_0+(x-c)F_1+\tilde{F}(x) \end{equation} \begin{Lemma}\label{LemmaPhi} If $F\in\mathcal{B}_\rho$ then $$\Phi_{n+1}[F]=F(c)+\frac{1}{n+1}\tilde{\Phi}_{n+1}[\tilde{F}]$$ with \begin{equation}\label{estimRn} | \tilde{\Phi}_{n+1}[\tilde{F}] | \leq \, {\mbox{const}}\ \max \left\{\| F\|, \|F'\|\right\} \end{equation} where the constant is independent of $\rho\in I$. \end{Lemma} \begin{Lemma}\label{LemmaJ} If $F\in\mathcal{B}_\rho$ then $J_{n+1}[F]\in\mathcal{B}_{\rho{\rm{e}}^{-\delta}}$ and $$J_{n+1}[F]=-\frac{1}{n(a+b)}\frac{F(x)-F(c)}{x-c}+\frac{1}{n^2}\tilde{J}_{n+1}[\tilde{F}]$$ with \begin{equation}\label{estimSn} \Big\| \tilde{J}_n[\tilde{F}](x) \Big\|_{\delta/2}\, \leq \, {\mbox{const}}\ \ \max \left\{ \|F\|,\|F'\|\right\} \end{equation} where the constant is independent of $\rho\in I$ and $\delta\in (0,\frac{1}{2})$. In particular, $$\| \tilde{J}_n[\tilde{F}] \|_\delta \, \leq \, {\mbox{const}}\ \ \delta^{-1}\| F\|$$ \end{Lemma} \begin{Corollary}\label{Estim1} The operators $J_n$ satisfy the estimates \begin{equation}\label{Estim1} \big\|J_{n+1}[F]\big\|_\delta\leq \ \frac{ {\mbox{const}}}{n}\ \delta^{-1}\|F\| \end{equation} with the constant not depending on $\delta$ or $\rho$ for all $\rho\in I$ and $\delta\in (0,\frac{1}{2})$. \end{Corollary} \noindent {\bf{Proofs}} The proofs of Lemma\,\ref{LemmaPhi} and Lemma\,\ref{LemmaJ}, respectively, are done by finding the asymptotic behavior of the integrals for large $n$. \noindent{\em{I. The lines of steepest ascent.}} Consider the function \begin{equation}\label{fung} g(x)=a\ln(1-x)+b\ln(1+x),\ \ \ \ \ x\in D_\rho \end{equation} The function $\Re g(x)$ has a saddle-point at $x=c$. The stable manifold intersects the real line at $c_0=\frac{\Re (b-a)}{\Re(b+a)}$. To the left of the stable manifold steepest descent lines wind toward $x=-1$, while to the right of this manifold they wind towards $x=1$. A straightforward calculation shows that, for any $z$, $\Re g(x)$ is increasing on the segment $[-1,z]$ (steep ascent) if this segment does not intersect the disk $\mathcal{C}_-$ bounded by the circle passing through the points $1$, $c_0$ and $c$. Similarly, $\Re g(x)$ increasing on the segment $[1,z]$ (steep ascent) if this segment does not intersect the disk $\mathcal{C}_+$ bounded by the circle through the points $-1$, $c_0$ and $c$. \noindent{\em{II. Proof of Lemma \ref{LemmaPhi}.}} To obtain the behavior of the integral for large $n$ the path of integration in (\ref{defPhi}) must be taken to pass through $c$, along the unstable manifold. The saddle-point method (see, e.g., \cite{Bender-Orszag}) gives $$\Phi_{n+1}[F]\sim \frac{2^{1-n(a+b)}}{B(na,nb)}\, {\rm{e}}^{ng(c)}\, F(c)\, \int_{-\infty}^{\infty} {\rm{e}}^{ng''(c)/2 \, s}\, ds\ \ \ \ {\mbox{as}} \ n\to\infty$$ and Lemma \ref{LemmaPhi} follows by direct calculation and standard estimates of the remainder. \noindent{\em{III. Proof of Lemma \ref{LemmaJ}.}} It was shown that the maximum modulus of $(1-x)^a(1+x)^b$ is attained for $x=z$ for any $z\not\in \mathcal{C}_-\cap\mathcal{C}_+$. (Note that if $c_0=c$ these disks are tangent. This happens for, e.g. $a=b$, or $a,b>0$.) For $z\in \mathcal{C}_-\cap\mathcal{C}_+$ consider the following path of integration in (\ref{opJ}), described here in opposite direction: the line of steepest descent through $z$ until this line exits $\mathcal{C}_-\cap\mathcal{C}_+$, followed by a segment to $-1$ (steep descent) if $z$ is on the left-side of the unstable manifold, or on a segment to $1$, if $z$ lies to the right of this manifold. Then, also in this case the maximum modulus of $(1-x)^a(1+x)^b$ is attained at $x=z$. Therefore in all cases the Laplace method yields $$J_{n+1}[F](z)\, \sim\, {\rm{e}}^{-ng(z)}\, \int_\cdot^z\, {\rm{e}}^{ng(z)+ng'(z)(t-z)}\, \frac{F(z)-F(c)}{1-z^2}\, dt\ \ \ {\mbox{as\ }}n\to\infty$$ and Lemma \ref{LemmaJ} follows by direct calculation and standard estimates of the remainder. \qed \subsubsection{The homological equation.}\label{homeq} The dominant part of (\ref{nonlinPDE}) is \begin{equation}\label{homologeq} \partial_x h+d_w h\, Mw=Mh+\frac{1}{1-x^2}\, \left( f(x,w)-\Phi[f](w)\right) \end{equation} where $\Phi[f](w)$ is the unique power series for which equation (\ref{homologeq}) has a solution $h$ as a power series in $w$ with coefficients analytic in $x$: $$\Phi[f](w)=\sum_{n\geq 2}\, \Phi[f]_n\, w^n\ \ \ {\mbox{where}}\ \ \ \Phi[f]_n=\Phi_n[f_n]$$ If $h$ satisfies (\ref{homologeq}) then the map $u=H(x,w)=w+h(x,w)$ is an equivalence map between the equation $$u'=M(x)u+\frac{f(x,u)-\Phi[f](u)}{1-x^2}$$ and $$w'=M(x)u+\frac{\mathcal{R}[f](x,w)}{1-x^2}$$ where \begin{equation}\label{defR} \mathcal{R}[f](x,w)=\left( {1+d_wh} \right)^{-1}\, {{\mathcal{R}_0}[f](x,w)} \end{equation} with \begin{equation}\label{defD} \mathcal{R}_0[f](x,w)=f(x,w+h)-f(x,w)-\Phi[f](w+h)+\Phi[f](w) \end{equation} Expanding in power series in $w$ equation (\ref{homologeq}) gives \begin{equation} (1-x^2)h_n'(x)+(n-1)\left[(b-a)-(b+a)x\right]h_n(x)=f_n-\Phi[f]_n\ ,\ \ \ n\geq 2 \end{equation} with the solution \begin{equation}\label{solhomeq} h_n=J_n[f_n]\ \ \ {\mbox{and\ }}\ \ \Phi[f]_n=\Phi_{n}[f_{n}] \end{equation} (see (\ref{opJ}), (\ref{defPhi})). \subsubsection{Estimates on the solution of the homological equation.} Let $\rho\in I$, $R>0$, $\delta\in (0,\frac{1}{2})$. Denote by $\Pi$ the Banach space of functions $f(x,w)$ analytic on the polydisk $$\Delta_{\rho,R}=\{ (x,w)\, ;\, |x-c|<\rho,\ |w|<R\}$$ and continuous up to the boundary, with the sup-norm, denoted $\|\ \|$. Let $\Delta_\delta=\Delta_{\rho{\rm{e}}^{-\delta},R{\rm{e}}^{-\delta}}$ and let $\Pi_\delta$ be the space of functions analytic on the polydisk $\Delta_\delta$, continuous up to the boundary; the sup-norm on this space is denoted by $\|\ \|_\delta$. For $f\in\Pi$ having a zero of order 2 at $w=0$ we have \begin{equation}\label{serf} f(x,w)=\sum_{n\geq 2}f_n(x)w^n=\sum_{n\geq 2,k\geq 0}f_{n,k}w^n(x-c)^k \end{equation} with \begin{equation}\label{estfnk} |f_{n,k}|\leq \|f\|R^{-n}\rho^{-k}\ ,\ \ \ \ \ \sup_{D\rho}|f_n(x)|\leq\|f\|R^{-n} \end{equation} By (\ref{solhomeq}) we have \begin{equation}\label{ser} \Phi[f](w)=\sum_{n\geq 2}\Phi_n[f_n]w^n\ \ \ \ \ \ \ \ \ \ \ h(x,w)=\sum_{n\geq 2}J_n[f_n](x)w^n \end{equation} \begin{Lemma}\label{L3} If $f\in\Pi$ the series (\ref{ser}) converge absolutely on $\Pi_\delta$ and the sums satisfy the estimates \begin{equation}\label{L3i} \|\Phi[f](w)\|_\delta\leq c_1\delta^{-2}\|f\| \end{equation} and \begin{equation}\label{L3ii} \|h\|_\delta\leq c_2\delta^{-2}\|f\| \end{equation} with the constants $c_1,c_2$ independent of $\rho\in I$ and $\delta\in (0,\frac{1}{2})$. \end{Lemma} {\bf{Proof of Lemma \ref{L3}}} By Lemma \ref{LemmaPhi} and (\ref{estfnk}) we have $$\|\Phi[f](w)\|_\delta=\big\| \sum_{n\geq 2}\left(f_{n}(c)+\frac{1}{n}\tilde{\Phi}_n[\tilde{f}_n]\right)w^n\big\|_\delta$$ $$\leq \left[\sum_{n\geq 2}\left(1+\frac{c_5}{n}\delta^{-1}\right){\rm{e}}^{-n\delta}\right]\, \|f\|<c_1 \delta^{-2}\|f\|$$ The estimate (\ref{L3ii}) follows directly from (\ref{Estim1}), (\ref{estfnk}), (\ref{solhomeq}).\qed \ Denote by $\Pi^0$ the functions $f\in\Pi$ with $f(x,0)=0$ and let \begin{equation}\label{norm0} \|f\|^0=\sup_{\Delta_{\rho,R}}\, \frac{|f(x,w)|}{|w|} \end{equation} Note that: \begin{equation}\label{Note} \frac{1}{R}\|f\|\leq \|f\|^0,\ \ |f_{n,k}|\leq \|f\|^0\rho^{-k}R^{-n+1},\ \ \sup_{D_\rho}|f_n(x)|\leq \|f\|^0R^{-n+1} \end{equation} \begin{Remark}\label{R4} For $f\in\Pi^0$ we have $$\|f\|_\delta\leq \delta^{-2}R\|f\|^0$$ \end{Remark} This estimate is obtained using (\ref{estfnk}) and (\ref{Note}). \begin{Remark}\label{R5} For $f\in\Pi^0$ Lemma \ref{L3} holds in $\|\ \|^0$, namely $$\|h\|_\delta^0\leq c_2\delta^{-2}\|f\|^0$$ \end{Remark} The proof follows the same steps as the proof of Lemma \ref{L3}. \begin{Remark}\label{R6} Let $0<R_1<R_2$. A standard Cauchy estimate and (\ref{Note}) shows that if $\phi(w)$ is a function analytic for $|w|<R_2$, continuous for $|w|\leq R_2$, with a zero of order two at $w=0$ then $$\max_{|w|\leq R_1}|d_w\phi|\leq \frac{1}{1-R_1/R_2}\, \max_{|w|\leq R_2}\frac{|\phi(w)|}{|w|}$$ \end{Remark} \begin{Lemma}\label{L4} There exists a constant $\kappa\geq 2$ such that for any numbers $\beta\geq 4$ and $\delta\in(0,\kappa^{-1})$ the following estimate holds: if $\|f\|^0\leq\delta^\beta$ then $$\| d_wh\|_\delta\leq c_4\delta\ \ \ {\mbox{and}}\ \ \ \|(1+d_wh)^{-1}\|_\delta\leq c_3$$ \end{Lemma} {\bf{Proof}} \noindent Using Remarks\,\ref{R5} and \ref{R6} we have $$\|d_wh\|_\delta\leq\, \frac{1}{1-{\rm{e}}^{-\delta/2}}\, \|h\|_{\delta/2}^0\, \leq\, \frac{c_2\delta^{-2}}{1-{\rm{e}}^{-\delta/2}}\, \|f\|^0$$ $$\leq\, c'_2\delta^{-3}\, \|f\|^0\, \leq\, c'_2\delta^{\beta-3}\, \leq c_4\delta$$ Let $\kappa\geq 2$ be large enough so that $c'_3\equiv c_4\kappa^{-1}<1$. Then for $\delta\in(0,\kappa^{-1})$ we have $$ \|(1+d_wh)^{-1}\|_\delta\leq \left(1-\|d_wh\|\right)^{-1}<(1-c'_3)^{-1}=c_3$$ which completes the proof of Lemma\,\ref{L4}.\qed \begin{Lemma}\label{LD} There exists a number $\beta_2$ depending only on $c_2$ and $\kappa$ such that for any $\beta\geq\max\{\beta_2,2\}$ and for any $\delta\in(0,\kappa^{-1})$ we have: if $\|f\|^0\leq\delta^\beta$ then $$\|w+h\|_\delta\leq\, R\, {\rm{e}}^{-\delta/2}$$ In particular, $f\left( x,w+h(x,w)\right)\in\Pi_\delta^0$. \end{Lemma} {\bf{Proof}} \noindent Using Remark\,\ref{R5} we have $$\|w+h\|_\delta\leq \, R\, {\rm{e}}^{-\delta}\,\left(1+\|h\|^0_\delta\right)\,\leq\, R\, {\rm{e}}^{-\delta}\,\left(1+c_2\delta^{-2}\|f\|^0\right)\,\leq\, R\, {\rm{e}}^{-\delta}\,\left(1+c_2\delta^{\beta-2}\right)$$ which is less than $R\, {\rm{e}}^{-\delta/2}$ if $\beta\geq 3+{\ln(2c_2)}/{\ln\kappa}\equiv\beta_2$.\qed \begin{Corollary}\label{Conseq} If $(x,w)\in\Delta_\delta$ then $$\sup_{[w,w+h]}\big| d_wf(x,\cdot )-d_w\Phi[f]\big|\leq\frac{{\rm{e}}^{\delta/4}}{1-{\rm{e}}^{-\delta/4}}\, \max_{|w|\leq R{\rm{e}}^{-\delta/4}}\,|f(x,\cdot)-\Phi[f]\big|$$ \end{Corollary} \begin{Lemma}\label{LE} Under the assumptions of Lemma\,\ref{LD} we have $$\max_{|x-c|\leq\rho{\rm{e}}^{-\delta}}\ \max_{|w|\leq R{\rm{e}}^{-\delta/4}}\ \frac{|f(x,w)-\Phi[f](w)|}{|w|}\,\leq\, c_6\delta^{-2}\, \|f\|^0$$ \end{Lemma} {\bf{Proof}} \noindent A Taylor series expansion in $w$ and Lemma\,\ref{LemmaPhi} give $$\frac{|f(x,w)-\Phi[f](w)|}{|w|}\,\leq\, \sum_{n\geq 2}|f_n(x)-f_n(c)-\frac{1}{n}T_n[f_n]|\, |w|^{n-1}$$ $$\leq\, \sum_{n\geq 2}\,\left[ |x-c|\, \sup\, |f_n'|\, +\, \frac{1}{n}|T_n[f_n]|\right]\, |w|^{n-1}$$ and Lemma\,\ref{LE} follows from (\ref{L3ii}) and Remark\,\ref{R4}. \qed \begin{Lemma}\label{L5} Under the assumptions of Lemma\,\ref{LD} the difference (\ref{defD}) satisfies $$\|\mathcal{R}_0[f]\|^0_\delta\, \leq \, c_7\, \delta^{-5}\left(\|f\|^0\right)^2$$ \end{Lemma} {\bf{Proof}} \noindent We have $$\mathcal{R}_0[f](x,w)\leq\ |h(x,w)|\ \max_{z\in[w,w+h]}\,|d_z\left(f(x,z)-\Phi[f](z)\right) |$$ $$\|\mathcal{R}_0[f]\|^0_\delta\leq\ \|h\|^0_\delta\, \max_{|x-c|\leq \rho{\rm{e}}^{-\delta}}\ \max_{|z|\leq R{\rm{e}}^{-\delta/2}}\,|d_z\left(f(x,z)-\Phi[f](z)\right) |$$ Remark\,\ref{R6} and Lemma\,{\ref{LE} give $$\leq\, \frac{\|h\|^0_\delta}{1-{\rm{e}}^{-\delta/4}}\, \max_{|x-c|\leq \rho{\rm{e}}^{-\delta}}\ \max_{|z|\leq R{\rm{e}}^{-\delta/4}}\, \frac{|f(x,z)-\Phi[f](z)|}{|z|}\leq\, \frac{\|h\|^0_\delta}{1-{\rm{e}}^{-\delta/4}}\, c_6\delta^{-2}\|f\|^0$$ Remark\,\ref{R5} completes the proof.\qed \begin{Lemma}\label{L6} Let $\kappa$ be as in Lemma\,\ref{L4} and assume $\beta>\max\{\beta_2,5\}$. Then there exists $\kappa_0>\kappa$ so that if $\delta\in (0,\kappa_0^{-1})$ then the remainder (\ref{defR}) satisfies $\|\mathcal{R}\|^0_\delta\leq \delta^{-6}\left(\|f\|^0\right)^2$. \end{Lemma} The {{proof}} is immediate using Lemmas\,\ref{LD}, \ref{L4}, and \ref{L5}. \qed \subsubsection{The sequence of parameters.}\label{param} \ \noindent 1. Choose $\beta>\max\{\beta_2,5\}$. Fix $\rho_1\in (\rho_{min},\rho_0)\subset I$. \noindent 2. Choose $\delta_1>0$ small enough, in the following way. Define recursively the sequence $\delta_{k+1}=\delta_k^{3/2}$. Therefore $\delta_{k+1}=\delta_1^{(3/2)^{k}}$. Define $R_{k+1}=R_k{\rm{e}}^{-\delta_k}$ and $\rho_{k+1}=\rho_k{\rm{e}}^{-\delta_k}$. Therefore $R_{k+1}=R_1{\rm{e}}^{-\eta_k(\delta_1)}$ where $\eta_k(\delta_1)=\delta_1+\delta_1^{3/2}+\delta_1^{(3/2)^2}+\ldots+\delta_1^{(3/2)^k}$ The sequence of continuous functions $\eta_k$ converges uniformly on the interval $[0,\kappa_0^{-1}]$ to a continuous function $\eta$, which satisfies $\eta(0)=0$ and $\eta(\delta_1)>0$ for all $\delta_1>0$. Choose $\delta_1$ small enough so that the following three inequalities hold: \noindent(i) $\rho_1{\rm{e}}^{-\eta(\delta_1)}>\rho_{min}$.\newline (ii) ${\rm{e}}^{-\delta_1}+c_2\delta_1^4<1$\newline (iii) $c_4\delta_1<1$. \noindent 3. Choose $\nu\geq\max\{\beta,6\}$. {\em{Notation.}} Denote $\Delta_k=\Delta_{\rho_k, R_k}$, and the sup-norm on the polydisk $\Delta_k$ by $\|\cdot\|_k$. \noindent 4. Choose $R_1$ small enough, so that the nonlinear term $f$ of the given equation (\ref{perFuchs}) satisfies $\|f\|^0_1\leq\delta_1^\nu$ (possible since $f(x,\cdot)$ has a zero of order two at $w=0$). \subsubsection{The iteration.} We construct recursively a sequence of analytic maps. Let $f^{[1]}=f$ analytic on $\Delta_1$. We determine $\phi=\phi^{[1]}$ so that equation (\ref{homologeq}) with $f=f^{[1]}$ has a solution $h^{[1]}$ analytic at both $x=\pm1$. By \S\ref{homeq} the map $H^{[1]}(x,w)=w+h^{[1]}(x,w)$ is an analytic equivalence map between equation $\mathcal{E}[f^{[1]}-\phi^{[1]}]$ (see (\ref{corperFuchs})) and equation $\mathcal{E}[f^{[2]}]$ where $f^{[2]}=\mathcal{R}[f^{[1]}]$ (see (\ref{defR})). Note that $f^{[2]}_2(x)=0$, in other words, $f^{[2]}(x,\cdot)$ has a zero of order three at $w=0$. The construction is continued inductively. The second step produces a function $\phi^{[2]}$ and a map $h^{[2]}$ so that $\mathcal{E}[f^{[2]}-\phi^{[2]}]$ is equivalent to $\mathcal{E}[f^{[3]}]$ by the analytic map $H^{[2]}(x,w)=w+h^{[2]}(x,w)$ and so on. \subsubsection{Convergence of the sequence of equivalence maps.}\label{convseqem} Consider the map $H_k=H^{[1]}\circ H^{[2]}\circ\ldots\circ H^{[k]}$ (where composition is in the $w$-variable only). \begin{Lemma}\label{R7} (i) $H^{[k]}(\Delta_{k+1})$ is relatively compact in $\Delta_k$. (ii) $f^{[k]}$, $\phi^{[k]}$ are holomorphic on $\Delta_{k+1}$ and continuous on $\overline{\Delta_{k+1}}$. (iii) $H^{[k]}$ is a biholomorphism from $ \Delta_{k+1}$ onto its image. \end{Lemma} {\bf{Proof}} \ (i) is proved by induction on $k$. For $k=1$, using (in order) Lemma\,\ref{L3}, \S\ref{param} points 4 and 3, Remark\,\ref{Note}, and \S\ref{param} point 2 we obtain that $|w+h^{[1]}(x,w)|<R_1$ on $\Delta_2$. The other steps are similar. (ii) follows by Lemmas\,\ref{L3}, \ref{LD} and \ref {L6}. (iii) $H^{[k]}$ is one-to-one since assuming $w'+h^{[k]}(x,w')=w''+h^{[k]}(x,w'')$ it follows that $$|w'-w''|=|h^{[k]}(x,w')-h^{[k]}(x,w'')|\leq |w'-w''|\, \sup_{\Delta_k}|d_wh^{[k]}|$$ which, by Lemma \ref{L4} is less or equal than $|w'-w''|c_4\delta_k$. Since $c_4\delta_k<c_4\delta_0<c_4\delta_1$ by \S\ref{param} point 2, it follows that $|w'-w''|=0$. \qed \ {\bf{Note}} that all the domains $\Delta_k$ contain the polydisk $\Delta_e$ of radii $(\rho_1{\rm{e}}^{-\eta(\delta_1)},R_1{\rm{e}}^{-\eta(\delta_1)})$. \begin{Lemma}\label{serc} (i) The sequence $\{H_k\}_k$ is Cauchy in the norm $\|\cdot\|^0_e$ on $\Pi^0_e$. (ii) The limit $H=\lim H_k$ is one-to-one and analytic on $\Delta_e$. \end{Lemma} {\bf{Proof}} Since $\|f^{[k]}\|^0_k\leq \delta_k^\beta$ (by \S\ref{param} point 3) we may apply Lemma \ref{L4} and obtain that $\|d_wh^{[k]}\|_k\leq c_4\delta_k$ and therefore \begin{equation}\label{cstar} \|d_wH_k\|_k\leq\|1+d_wh^{[1]}\|_2\ldots \|1+d_wh^{[k]}\|_{k+1}\leq (1+c_4\delta_1)\ldots(1+c_4\delta_k)<{\rm{e}}^{c_4\eta(\delta_1)} \end{equation} Then $$\|H_k-H_{k+1}\|_e^0=\max\frac{|H^{[k]}(x,w)-H^{[k]}(x,w+h^{[k+1]}(x,w))|}{|w|}$$ $$\leq\|d_wH^{[k]}\|_{k+1}\ \|h^{[k+1]}\|^0_e\leq \, {\rm{e}}^{c_4\eta(\delta_1)}\, c_2\, \delta_{k+1}^{-2}\, \|f^{[k+1]}\|^0_{k+1}\leq\, {\rm{e}}^{c_4\eta(\delta_1)}\, c_2\, \delta_{k+1}^{\beta-2}$$ by Remark \ref{R5} and (\ref{cstar}). Since the series $\sum_k\delta_{k+1}^{\beta-2}$ converges, using Remark\,\ref{R4}, and noting that all the functions $H_k(x,\cdot)$ are perturbations of the identity the result follows. \qed \subsubsection{Convergence of the sequence of correction maps.}\label{convcorr} \begin{Remark}\label{const2n} Suppose the function $f\in\Pi$ has a zero of order (at least) $p$ at w=0: $f_n(x)=0$ for all $n\leq p-1$ (see (\ref{serf})). Then $\phi[f]$ has a zero of order $p$ as well (see (\ref{ser})), while $h$ and $R[f]$ (see (\ref{defR})) have a zero of order $2p-1$. As a consequence, $f^{[k]}$ and $\Phi^{[k]}$ have a zero of order $2^{k-1}+1$ for all $k\geq 1$. \end{Remark} A sequence of holomorphic functions $\Psi^{[k]}(x,u)$ converging uniformly to the correction $\phi(u)$ of Theorem\,\ref{anf} is now constructed as follows. Denote by $K^{[n]}$ the $w$-inverse of $H^{[n]}$: $K^{[n]}(x,H^{[n]}(x,w))=w$. Let $\Psi^{[1]}$ be the first correction of $f$: $\Psi^{[1]}=\phi^{[1]}$. Note that $\phi^{[1]}(u)=\phi_2u^2+O(u^3)$. The next approximation, $\Psi^{[2]}$, of $\phi$ is obtained as the second correction of $f$: the image of $f^{[2]}-\phi^{[2]}$ through the $w$-the inverse of the map $u=H_1(x,w)$ is $f(x,u)-\Psi^{[2]}(x,u)$ where $$\Psi^{[2]}(x,u)=\Phi^{[1]}(x,u)+\frac{1}{\partial_uK^{[1]}(x,u)}\Phi^{[2]}\left(x,K^{[1]}(x,u)\right)$$ $$\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =\Phi^{[1]}(x,u)+\partial_wH^{[1]}\left(x,K^{[1]}(x,u)\right)\Phi^{[2]}\left(x,K^{[1]}(x,u)\right)$$ In general, the inverse of $u=H_n(x,w)$ takes $f^{[n+1]}-\phi^{[n+1]}$ into $f-\Psi^{[n+1]}$ where $$\Psi^{[n+1]}(x,u)=\Phi^{[1]}(x,u)+\partial_wH^{[1]}\left(x,K_1(x,u)\right)\Phi^{[2]}\left(x,K_1(x,u)\right)$$ $$+\partial_wH^{[2]}\left(x,K_2(x,u)\right)\Phi^{[2]}\left(x,K_2(x,u)\right)+\ldots+\partial_wH^{[n]}\left(x,K_n(x,u)\right)\Phi^{[n+1]}\left(x,K_n(x,u)\right)$$ where the following notation was used: $$K_1=K^{[1]},\ {\mbox{and\ }} K_{j+1}(x,u)=K^{[j+1]}\left(x,K_j(x,u)\right)\ \ {\mbox{for\ }}j\geq 1$$ \begin{Lemma}\label{lem2ton} Consider the Taylor expansion $\Psi^{[n]}(x,u)=\sum_k\Psi^{[n]}_k(x)u^k$. (i) We have that $\Psi^{[n]}_k(x)\equiv \Psi^{[n]}_k$ do not depend on $x$ for all $k\leq 2^{n-1}+1$. (ii) Furthermore, $\Psi^{[n]}_k=\phi_k$ for all $k\leq 2^{n-1}+1$. (iii) The Taylor coefficients of the expansions in powers of $w$ of map $H_{n+1}$ and of the map (\ref{serh}) coincide for all powers of $w$ less or equal to $2^{n-1}+1$. \end{Lemma} {\bf{Proof}} \noindent (i) Follows by induction using Remark\,\ref{const2n}. \noindent (ii) Follows from the fact that the equation $\mathcal{E}[f-\Psi^{[n]}]$ is analytically equivalent (through the map $u=H_n(x,w)$) to an equation with the nonlinear term $f^{[n]}$ having a zero of order $2^n+1$, and the correction is unique. \noindent (iii) Is shown directly by induction. \qed As a consequence of Lemmas\,\ref{lem2ton} and \ref{serc} the series (\ref{serh}) converges, and therefore so does (\ref{serphi}). \subsubsection{The multidimensional case.}\label{ddim} In the vector-valued case the series are split by degree of homogeneity, then the proof is similar to the one-dimensional case. A slight change is that in Lemmas\,\ref{LemmaPhi} and \ref{LemmaJ}, which estimate $\Phi_\bfm[\mathbf{F}]$ and $J_\bfm[\mathbf{F}]$, the saddle points (for large $n$) are given by $c_\bfm$, see (\ref{cm}). \section{Acknowledgments} The author would like to express warmest gratitude towards D. Lubinsky for his kind help regarding estimates of Jacobi polynomials. \section{Appendices} \subsection{Region containing one regular singularity.}\label{rev1s} Consider a nonlinear perturbation of a system with a regular singular point \begin{equation}\label{gen1sing} \frac{d\mathbf{u}}{dx}=\frac{1}{x}\, L(x)\mathbf{u}+\frac{1}{x}\, \tilde{\mathbf{f}}(x,\mathbf{u})\ \ \ \ \ \ \mathbf{u}\in\mathbb{C}^d,\ x\in\mathbb{C} \end{equation} where ${L}(x)$ is a matrix analytic at $x=0$ and the nonlinear term $\frac{1}{x}\, \tilde{\mathbf{f}}(x,\mathbf{u})$ can be allowed to have a first order pole at the singular point $x=0$. If the eigenvalues of $L(0)$ are nonresonant, after an analytic change of coordinates it can be assumed that $L(x)$ is constant (see \cite{Norm_Form} for more details), hence consider \begin{equation}\label{1sing} \frac{d\mathbf{u}}{dx}=\frac{1}{x}\, L\mathbf{u}+\frac{1}{x}\, {\mathbf{f}}(x,\mathbf{u})\ \ \ \ \ \ \mathbf{u}\in\mathbb{C}^d,\ x\in\mathbb{C} \end{equation} It is assumed that $\mathbf{f}(x,\mathbf{u})$ is analytic for $|\mathbf{u} |<\rho$ ($\rho>0$) and $x$ in a domain $D_r$ which is either a disk centered at the origin: $|x|<r'$, or an annulus $r''<|x|<r'$.\footnote{Allowing the nonlinear part to be singular at $x=0$ accommodates systems corresponding to higher order equations.} Such systems are generically analytically linearizable \cite{Nonln}: \begin{Theorem}\label{Th1sing} Assume that the eigenvalues $\mu_1,...,\mu_d$ of the matrix $L$ satisfy the Diophantine condition (\ref{DioCond}) for all $l\in\mathbb{Z}$ if $D_r$ is an annulus (respectively $l\in\mathbb{N}$ if $D_r$ is a disk), and for all $s\in\{1,...,d\}$, and $\mathbf{k}\in\mathbb{N}^d$ with $ |\mathbf{k}|\geq 2$. \noindent Then {{the system (\ref{1sing}) is analytically equivalent to its linear part}} $\mathbf{w}'=\frac{1}{x}L\mathbf{w}$ for $x\in{D}_ {\tilde{r}} \subset D_r$ and $|\mathbf{u} |<{\tilde{\rho}}<\rho$. \end{Theorem} {\bf{Remarks}} {\bf{1.}} The domain ${D}_ {\tilde{r}}$ can be made arbitrarily close to $D_r$ if $\tilde{\rho}$ is small enough. {\bf{2.}} The analytic equivalence map which is close to identity is unique if no eigenvalue $\mu_j$ is integer. \subsection{Analyticity of (\ref{hnat-1}) at $x=-1$.}\label{details} Let {$G$} be the {monodromy matrix} of $Y'=MY$ at $x=-1$: after analytic continuation along a closed loop around $x=-1$ the matrix $Y(x)$ becomes {$AC_{(-1)}Y(x)=Y(x)G$}. Then the analytic continuation of (\ref{hnat-1}) on a closed loop around $x=-1$ yields \begin{equation} AC_{(-1)}{\mathbf{h}_n(x,w)=Y(x)G\int_{-1}^xQ(t)^{-1}G^{-1}Y(t)^{-1}\mathbf{R}_n(t,Y(t)GG^{-1}Y(x)^{-1}\mathbf{w})dt} \end{equation} \begin{equation}\nonumber =\mathbf{h}_n(x,w) \end{equation} \noindent which means that (\ref{hnat-1}) are the coefficients of the unique linearization map of (\ref{perFuchs}) which is analytic at $x=-1$.

https://arxiv.org/abs/1509.04701

Subdivisions in the Robber Locating Game

We consider a game in which a cop searches for a moving robber on a graph using distance probes, which is a slight variation on one introduced by Seager. Carragher, Choi, Delcourt, Erickson and West showed that for any n-vertex graph $G$ there is a winning strategy for the cop on the graph $G^{1/m}$ obtained by replacing each edge of $G$ by a path of length $m$, if $m \geqslant n$. They conjectured that this bound was best possible for complete graphs, but the present authors showed that in fact the cop wins on $K^{1/m}$ if and only if $m \geqslant n/2$, for all but a few small values of $n$. In this paper we extend this result to general graphs by proving that the cop has a winning strategy on $G^{1/m}$ provided $m \geqslant n/2$ for all but a few small values of $n$; this bound is best possible. We also consider replacing the edges of $G$ with paths of varying lengths.

\section{Introduction} Pursuit and evasion games on graphs have been widely studied, beginning with the introduction by Parsons \cite{Par76} of a game where a fixed number of searchers try to find a lost spelunker in a dark cave. The searchers cannot tell where the target is, and aim to move around the vertices and edges of the graph in such a way that one of them must eventually encounter him. The spelunker may move around the graph in an arbitrary fashion, and in the worst case may be regarded as an antagonist who knows the searchers' positions and is trying to escape them. The best-known variant is the classical Cops and Robbers game, introduced by Quillot \cite{Qui78}, and independently in a paper of Nowakowski and Winkler \cite{NW83} (where it is attributed to G. Gabor). Unlike the Lost Spelunker game, Cops and Robbers is played with perfect information, so that at any time each of the agents knows the location of all others. A fixed number of cops take up positions on vertices of a connected graph and a robber then starts on any unoccupied vertex. The cops and the robber take turns: at his turn the robber may move to any adjacent vertex or remain where he is, and at their turn all cops simultaneously make moves of this form. The cops win if at any point one of them reaches the robber's location. On a particular graph $G$ the question is whether a given number of cops have a strategy which is guaranteed to win, or whether there is a strategy for the robber which will allow him to evade capture indefinitely. The \textit{cop number} of a graph is the minimum number of cops that can guarantee to catch the robber. Early results on this game include those obtained by Nowakowski and Winkler \cite{NW83}, who categorised the graphs of cop number 1, and Aigner and Fromme \cite{AF84}, who showed that every planar graph has cop number at most 3. An important open problem is Meyniel's conjecture, published by Frankl \cite{Fra87}, that the cop number of any $n$-vertex connected graph is at most $O(\sqrt{n})$ -- this has been shown to be true up to a $\log(n)$ factor for random graphs by Bollob{\'a}s, Kun and Leader \cite{BLK13}, following which \L uczak and Pra\l at improved the error term \cite{LP10}. More recently several variations on the game have been analysed by Clarke and Nowakowski (e.g. \cite{CN00}). In this paper we consider the Robber Locating game, introduced in a slightly different form by Seager \cite{Sea12} and further studied by Carragher, Choi, Delcourt, Erickson and West \cite{CCDEW}. Like the Lost Spelunker game, the focus is on locating a hidden target, but, like Cops and Robbers, the target is a robber who moves around the vertices in discrete steps. There is a single cop, who is not on the graph but can probe vertices and receive information about how far away the robber is (in terms of the normal graph distance) from the vertex probed. For ease of reading we shall refer to the cop as female and the robber as male. The robber initially occupies any vertex. Each round consists of a move for the robber, in which he may move to an adjacent vertex or stay where he is, followed by a probe of any vertex by the cop. The cop then wins immediately if she is able to determine the robber's current location from the results of that probe and previous ones. This game may be viewed as a variant of the Sequential Locating game, also studied by Seager \cite{Sea13}, with the difference between the two being that in the Robber Locating game the target can move about the graph; in both games the choice of probe made may depend on the results of previous probes. If we instead require all the probes to be chosen at once (with a stationary target), we recover the Graph Locating game, independently introduced by Slater \cite{Sla75} and by Harary and Melter \cite{HM76}. In games with a stationary target, the searcher can guarantee to win eventually, simply by probing every vertex, and the natural question is the minimum number of probes required to guarantee victory on a given graph $G$. For the Graph Locating game, this is the \textit{metric dimension} of $G$, written $\mu(G)$. In the Robber Locating game, by contrast, it is not necessarily true that the cop can guarantee to win in any number of probes. Consequently the primary question in this setting is whether, for a given graph $G$, the cop can guarantee victory in bounded time on $G$, or equivalently whether she can catch a robber who has full knowledge of her strategy. We say that a graph is \textit{locatable} if she can do this and \textit{non-locatable} otherwise. In the game as introduced by Seager there was an additional rule that the robber cannot move to the vertex probed in the previous round (the \textit{no-backtrack condition}). Carragher et al.\ considered the game without this restriction, as do we, and Seager also considered the version without the no-backtrack rule for trees \cite{Sea14}. A similar game in which the searcher wins only if she probes the current location of the target and receives no information otherwise, but the target must move at each turn, was recently analysed by one of the authors \cite{Has13}. The main result of Carragher et al.\ \cite{CCDEW} is that for any graph $G$ a sufficiently large equal-length subdivision of $G$ is locatable. Formally, write $G^{1/m}$ for the graph obtained by replacing each edge of $G$ by a path of length $m$ through new vertices. Carragher et al.\ proved that $G^{1/m}$ is locatable whenever $m\geqslant \min\{\abs{V(G)},1+\max\{\mu(G)+2^{\mu(G)},\Delta(G)\}\}$. In most graphs this bound is simply $\abs{V(G)}$, and they conjectured that this was best possible for complete graphs, i.e.\ that $K_n^{1/m}$ is locatable if and only if $m\geqslant n$. The present authors \cite{HJK} showed that in fact $K_n^{1/m}$ is locatable if and only if $m\geqslant n/2$, for every $n\geqslant 11$. In this paper we show that the same improvement may be obtained in general: provided $\abs{V(G)}\geqslant 23$, $G^{1/m}$ is locatable whenever $m\geqslant\abs{V(G)}/2$. This bound is best possible, since $K_n^{1/m}$ is not locatable if $m=(n-1)/2$, and some lower bound on $\abs{V(G)}$ is required for it to hold, since $K_{10}^{1/5}$ is not locatable \cite{HJK}. These results, and those of Carragher et al., fundamentally depend on taking equal-length subdivisions; in the final section of this paper we show that an unequal subdivision is also locatable provided every edge is subdivided by at least a certain amount. \section{Subdivisions and maximal matchings} Recall that $G^{1/m}$ is the graph obtained by replacing each edge of $G$ with a path of length $m$ through new vertices. Each such path is called a \textit{thread}, and an \textit{branch vertex} in $G^{1/m}$ is a vertex that corresponds to a vertex of $G$. We write $u\thrd v$ for the thread between branch vertices $u$ and $v$. We use ``a vertex on $u\cdots v$'' to mean any of the $m+1$ vertices of the thread, but ``a vertex inside $u\cdots v$'' excludes $u$ and $v$. Our basic strategy to locate the robber on sufficiently large equal-length subdivisions of $G$ is to ensure the following. \begin{enumerate} \item Whenever the robber is at a branch vertex, the probe we make reveals that fact. \item If the robber ever spends $r$ turns without visiting a branch vertex, we establish which thread he is inside and then can win on the next turn, where $r$ depends only on $G$. \item If, when the robber visits a branch vertex, there is more than one possibility for his location, we can ensure that by the next time he is at a branch vertex we reduce the set of possibilities to a simpler set. \end{enumerate} 1 and 2 above are sufficient to ensure that $G^{1/m}$ is locatable for sufficiently large $m$, since if $m>r$ the robber can only ever visit one branch vertex without being caught, and we can eventually find which one it is. For smaller $m$ the robber may be able to visit several branch vertices, so to get a better bound we need a way to progressively reduce the possibilities as in 3. This reduction is not necessarily to a smaller set but we ensure that only a bounded number of reductions can occur (and, by 2, each takes bounded time) before we reach a singleton set. Note that the reduction occurs by the next time the robber is at a branch vertex, even if this is because he stays at his current branch vertex until the next turn. To get a bound of close to $\abs{V(G)}/2$, roughly speaking, our strategy is that in between the robber's visits to branch vertices we aim to eliminate possible destinations in pairs. To do this we must probe inside threads, aiming to eliminate both ends of the thread. This approach is simplified in the case of complete graphs, analysed extensively in \cite{HJK}, by the knowledge that any pair of branch vertices has a thread between them. In the general case we need some knowledge of the structure of $G$. To that end we will, so far as this is possible, divide the vertices of $G$ into adjacent pairs. So we take a maximal matching in $G$, and we will get a bound which depends on the size of that maximal matching. Fix a connected graph $G$ with $n$ vertices, and let $M$ be a maximal matching. Write $k$ for $\abs{M}$ and $X$ for $V(G)\setminus V(M)$; since $M$ is maximal, $X$ is an independent set. Note that $M$ must be maximal, i.e.\ inextensible, but need not be a maximum-size matching. \begin{thm}\label{match}If $m\geqslant k+1$ and $m\geqslant 12$, $G^{1/m}$ is locatable. \end{thm} The bound $m\geqslant k+1$ is best possible: $K_{2k+1}$ has a maximal matching of size $k$, but $K_{2k+1}^{1/k}$ is not locatable (\cite{HJK}, Theorem 3). If $m$ is odd, let $t=(m-1)/2$. We use the word ``midpoint'' to refer to the vertex at distance $m/2$ from each end of a thread if $m$ is even, and either of the two vertices at distance $t$ from one end if $m$ is odd. If $m$ is odd, we also use the term ``off-midpoint'' to refer to either vertex at distance $t-1$ from one end and $t+2$ from the other. Note that probing a branch vertex will establish the robber's exact distance to his nearest branch vertex (by considering the result mod $m$). Probing a midpoint will also establish the robber's exact distance to the nearest branch vertex when $m$ is even, but when $m$ is odd it will only give this distance within two consecutive possibilities. This uncertainty makes the odd case significantly more complicated. In particular, note that point 1 above can be assured by probing branch vertices and midpoints when $m$ is even, but only by probing branch vertices when $m$ is odd. For this reason our strategy reverts to probing branch vertices whenever it is possible, given the results of previous probes, that the robber is at a branch vertex. (This is not necessary when $m$ is even, but we present a strategy which will work in either case rather than considering the two cases separately.) For the remainder of this section we assume that $m\geqslant k+1$ and $m\geqslant 12$. Before giving the proof of Theorem~\ref{match} we prove some preparatory lemmas. \begin{lem}\label{astar}Suppose that, immediately following some probe, we know that the robber is on a thread between $u$ and a vertex in $Z$, where $u$ is any specified vertex and $Z\subseteq X$. Then there is a strategy that will, within $\abs{Z}$ steps and by the next time the robber reaches a branch vertex, either win or identify that he is in $Z$. \end{lem} \begin{proof}We may neglect any vertices in $Z$ which do not have threads leading to $a$. Simply probe the remaining vertices of $Z$ in order. If the robber reaches a branch vertex, we will identify whether it is the one just probed (distance 0), $a$ (distance $m$), or some other vertex in $Z$ (distance a higher multiple of $m$, since no threads go between vertices in $X$). If he does not reach a branch vertex within $\abs{Z}$ turns then we will have probed one end of the thread he is on (identified by getting a result between 0 and $m$) and so located him (since the other end is known to be $a$). \end{proof} \begin{lem}\label{noturnback}Suppose that, immediately following some probe, we know that the robber is adjacent to a branch vertex. Write $x$ for the unknown branch vertex he is adjacent to. Let $v$ be any vertex and $w_1w_1',\ldots,w_{k-2}w_{k-2}'$ be a set of $k-2$ edges of $G$. Then it is possible to make at most $2k-3$ probes in such a way that \begin{enumerate}[(i)] \item if the robber is at a branch vertex for one of these probes, then the cop identifies this fact and that branch vertex must be $x$; \item otherwise, for each of the sets $\{v\},\{w_1,w'_1\},\ldots,\{w_{k-2},w_{k-2}'\}$ in turn, the cop identifies how many endpoints of the robber's current thread are in that set. \end{enumerate} \end{lem} \begin{proof} We start by probing $v$. This establishes whether the robber is at $x$, distance 1 from $x$ or distance $2$ from $x$. If the robber is at $x$ then we are done by (i). Otherwise he is inside a thread and the first probe also establishes whether that thread meets $v$. We now consider $w_1w_1',\ldots,w_{k-2}w_{k-2}'$ in order. For each thread $w_i\thrd w_i'$, we either probe a vertex near the middle or probe both endpoints in turn. We do the former unless it is possible (given previous probe results) that the robber is at a branch vertex by the time of this probe. When probing a vertex near the middle, we use a midpoint unless $m$ is odd, $m=k+1$, $i=t$, and the result of the previous probe was consistent with the robber being at distance $t+1$ from $x$ and with him being at distance $t-1$ from $x$ at that time; in this specific case we use an off-midpoint. We refer to the probe or probes on $w_i\thrd w'_i$ as ``stage $i$''. First we analyse the case where we probe an off-midpoint. If we do this, $m$ must be odd and the previous probe must have been at a midpoint, since if it were at a branch vertex there would be no uncertainty about the robber's distance to the nearest branch vertex. In order for the result of the previous probe to meet the conditions for us to probe an off-midpoint, it must be $\pm 1$ (mod $m$), since if it were a multiple of $m$ then the robber being $t-1$ steps from a branch vertex would be impossible and if it were anything else then the robber being $t+1$ steps from a branch vertex would be impossible. Consequently, at the time of the previous probe the robber was at a midpoint or one step away from a midpoint. We now probe the off-midpoint of $w_t\thrd w'_t$, and at this point the robber can be at most two steps from a midpoint. So if he is on $w_t\thrd w'_t$ the distance returned is at most 4, if he is on an adjacent thread it is in $\{2t-3,\ldots,2t+5\}$, i.e.\ in $\{m-4,\ldots,m+4\}$, and if he is anywhere else it is at least $2m-4$. Since $m$ is odd we must have $m\geqslant 13$, and we will successfully distinguish these possibilities. Likewise, if we probe a midpoint of a thread and get a result $r$ then, since we know the robber cannot be at a branch vertex, $r\leqslant t$ if he is inside the probed thread, $t+1\leqslant r \leqslant t+m$ if he is inside an adjacent thread, and $r>t+m$ otherwise. Therefore, by this method we establish in stage $i$ whether the robber has reached a branch vertex and, if not, whether he is on a thread ending in $w_i$ or $w_i'$ (and whether he is on $w_i\thrd w'_i$). It remains to prove that if the robber is at a branch vertex for one of these probes then that vertex is $x$. We claim that at each probe in stage $i$ the robber's distance from $x$ is at most $i+2$, and that equality is possible for stage $i$ only if $i\leqslant 2$ or we had equality for the final probe in stage $i-1$; this is sufficient since $i+2\leqslant k<m$. If $i=1$ then the previous probe was at $v$ and established the robber's exact distance from $x$; if this was 2 then we need only one additional probe for $w_1\thrd w'_1$ and if it was 1 we need two additional probes. So the claim is true for $i=1$. For $i\in\{2,\ldots k-3\}$ we proceed by induction. The robber's distance from $x$ was at most $i+1$ at the time of the previous probe. If the previous probe was consistent with him being at distance 1 from a branch vertex, then he must have been distance at most 2 from a branch vertex; since $i+1\leqslant k-2<m-2$, his distance from $x$ was at most 2, and so for each of the two probes required for this stage it is at most $4\leqslant i+2$; if $i>2$ the inequality is strict. If the previous probe was not consistent with him being next to a branch vertex, we only require one probe for this stage so the robber's distance from $x$ when we make this probe is at most $i+2$, and it is less unless there was equality for the previous probe. Finally, for $i=k-2$, we again know that the robber's distance from $x$ was at most $k-1$ at the time of the previous probe. Again, if the result of the previous probe was consistent with him being next to a branch vertex he must in fact have been within two steps from a branch vertex (and, if exactly two steps away, $m$ must be odd). Suppose he was two steps from a branch vertex other than $x$, i.e.\ at distance $k-1$ from $x$. In this case, the previous probes must be consistent with him being at distance $i+2$ at the end of stage $i$ for every $i$. If the result of stage $t-1$ was inconsistent with him being $t-1$ away from $x$ then (since it was consistent with him being $t+1$ away) the robber was at least $t$ away from $x$ at that point, and we therefore know he is not adjacent to a branch vertex at the end of stage $k-3$. Conversely, if the result of stage $t-1$ was consistent with the robber being $t-1$ from $x$ at the end of that stage, then we probed an off-midpoint at stage $t$. When we probed the off-midpoint, the robber must have been at distance $t+2$ from $x$ and so the result of that probe (mod $m$) would be in $\{0,\pm 3\}$. If he had been at distance $t-2$ from $x$, the result would have been in $\{\pm 1,\pm 4\}$, so we can eliminate this possibility. Consequently, if the robber is at distance $k-1$ at the end of stage $k-3$ then we will know that he is not adjacent to a branch vertex, and so we only use one probe for stage $k-2$ and his distance from $x$ is at most $k$ when we make it, as required. Otherwise, even if we use two probes for stage $k-2$, his distance at each of them is at most $k$. This proves the claim. \end{proof} \begin{lem}\label{gamma}Suppose that, immediately following some probe, we know that the robber is at the vertex at distance 1 from $Z$ on a thread between a vertex in $V(M)\setminus\{a\}$ and a vertex in $Z$, where $Z\subseteq X$. Then there is a strategy that will, within $2k-2+\abs{Z}$ steps and by the next time the robber reaches a branch vertex, either win or identify that he is in $Z$. \end{lem} \begin{proof} If $k=1$ then $V(M)\setminus \{a\}=\{a'\}$, and so the result is true by Lemma~\ref{astar}. Otherwise, choose any $bb'\in M$ not equal to $aa'$, and take $v=a'$ and $w_1w_1',\ldots,$ $w_{k-2}w_{k-2}'$ to be the edges in $M$ other than $aa'$ and $bb'$. Now we probe as in Lemma~\ref{noturnback}, stopping if we identify that the robber is at a branch vertex or if any probe indicates that he is on an adjacent thread. By Lemma~\ref{noturnback}, this takes at most $2k-3$ probes, and if he reaches a branch vertex we have identified that he is in $Z$. If none of the probes indicate that he is in $Z$ or on an adjacent thread to the one probed then he must be on a thread adjacent to $b\thrd b'$ (since one end of his thread is outside $X$). Consequently, within $2k-3$ probes, we have identified, at the time of the last probe, either that the robber was in $Z$ or that he was inside a thread leading from $Z$ to either $c$ or $c'$ for some $cc'\in M$ (perhaps only one of these is actually possible). In the latter case the result of the last probe either told us that the robber was distance at most 2 from the nearest branch vertex or that he was distance at least 2 from the nearest branch vertex. If we know he was not adjacent to a branch vertex then we may probe $c$ to establish whether he is inside a thread leading to $c$ or to $c'$, and then we are done in at most $\abs{Z}$ additional steps by Lemma~\ref{astar}. If not then probe the vertex at distance 4 from $c$ along $c\thrd c'$. If the robber is now in $Z$, this will return $m+4$ or $2m-4$. If he is at $c$ or $c'$ then it will return 4 or $m-4$ respectively. If he is inside a thread leading to $c$ then it will return a result in $\{5,6,7,m+1,m+2,m+3\}$ and if he is inside a thread leading to $c'$ then it will return a result in $\{m-3,m-2,m-1,m+5,m+6,m+7,2m-7,2m-6,2m-5\}$. Since $m\geqslant 12$ these five possibilities are all distinguished, so either we are done immediately or within $\abs{Z}$ additional steps by Lemma~\ref{astar}. \end{proof} \begin{lem}\label{delta}Suppose that, immediately following some probe, we know that the robber is at the vertex at distance 1 from $a$ on a thread not leading to $a'$, where $aa'\in M$. Then there is a strategy that will, within $2k-2+\abs{X}$ steps and by the next time the robber reaches a branch vertex, either win or identify that he is in $X$. \end{lem} \begin{proof} The strategy here proceeds in much the same way as for Lemma~\ref{gamma}. If $k=1$ then the robber is on a thread between $a$ and $X$ and we are done by Lemma~\ref{astar}. Otherwise, choose any $bb'\in M$ not equal to $aa'$, take $v=b$ and $w_1w_1',\ldots,w_{k-2}w_{k-2}'$ to be the edges in $M$ other than $aa'$ and $bb'$, and proceed as in Lemma~\ref{noturnback}. If a probe reveals that the robber is on an adjacent thread, interrupt the process. If we know which thread he is on, we can win by probing either end; if not we know he is on $a\thrd w_i$ or $a\thrd w'_i$, and from the result of the last probe we also know either that he not adjacent to $a$ or that he is not adjacent to the other end. Consequently, probing $w_i$ will either win or identify that he is inside $a\thrd w'_i$, and in the latter case we can now win by probing either end. If this does not happen, and we do not establish that the robber has returned to $a$, then at the time of the last probe the robber was inside a thread leading to $b'$ or to some vertex in $X$. If there is no thread $a\thrd b'$ then we are done by Lemma~\ref{astar}, so we assume that there is. We know from the result of the last probe either that his distance to the nearest branch vertex was at least 2 or that it was at most 2 (and at most 1 if $m$ is even). In the former case we probe $b'$; this wins if he is on $a\thrd b'$ and we are done by Lemma~\ref{astar} if not. In the latter case we probe the vertex at distance 3 from $a$ along the thread $a\thrd b'$. If the robber is at $a$ the result will be 3, if he is at $b'$ it will be $m-3$, and if he is in $X$ it will be $m+3$. If he is inside $a\thrd b'$, the result will be in $\{0,1,2,m-6,m-5,m-4\}$ ($\{1,2,m-5,m-4\}$ if $m$ is even) and if he is inside a thread between $a$ and $X$ it will be in $\{4,5,6,m,m+1,m+2\}$ ($\{4,5,m+1,m+2\}$ if $m$ is even). Since $m\geqslant 12$ these possibilities are distinguished, and we have either won, established that he is in $X$, or are done by Lemma~\ref{astar}. \end{proof} We are now ready to combine these elements into a complete strategy to locate the robber. \begin{proof}[Proof of Theorem~\ref{match}] We describe a winning strategy for the cop. First, probe branch vertices in turn until the answer reveals that the robber is at a branch vertex. Either this eventually happens or he remains inside a single thread, in which case we eventually identify both ends and win. Suppose that we know, on receiving the distance from some probe, that the robber's current location is in a set $A\cup Y$ of branch vertices, where $A\subseteq V(M)$ and $Y\subseteq X$. We show that, by the time the robber next reaches a branch vertex and within time $n+2$, \begin{enumerate}[(i)] \item\label{inx} if $A=\varnothing$, we either locate him or reach a point where he is known to be in a smaller set $Y'\subset Y$; \item\label{pair} if $A=\{a\}$ or $A=\{a,a'\}$, where $aa'\in M$, we either locate him or reduce to \eqref{inx}; \item\label{rest} otherwise we either locate him, reduce to \eqref{inx} or \eqref{pair}, or reach a point where he is known to be in a smaller set $A'\cup Y$ with $A'\subset A$. \end{enumerate} Note that when we reduce to an earlier case the number of candidate vertices in $X$ may increase; going from knowing he is in $\{a,a'\}$ to knowing he is in $X$ is a valid reduction. In case \eqref{inx}, pick a vertex $y\in Y$ and probe a neighbour of $y$, i.e.\ a vertex on $y\thrd a$ for some $a\in V(M)$. If the response is 0 or 1 the robber is caught. If it is 2 we know the robber is one step away from $y$ heading for some vertex in $V(M)\setminus\{a\}$. Setting $Z=\{y\}$ in Lemma~\ref{gamma}, we can win in time $2k-1$ (and by the time he next reaches a branch vertex). If the response is larger, we will take $Y'=Y\setminus\{y\}$. If the response is $2m-2$ we know the robber is one step away from some vertex in $Y'$ on a thread leading to $a$, and so we can win or reach a position where the robber is known to be in $Y'$ in the required time by Lemma~\ref{astar}. If the response is $\pm 1$ mod $m$ (but not 1), the robber is known to be in $Y'$ and we are done. The only other possibility is that the answer is 0 or $\pm 2$ mod $m$, but larger than $2m-2$, in which case the robber is known to be one step from a vertex in $Y'$, on a thread which does not lead to $a$. Now, by Lemma~\ref{gamma}, we can win or reach a point where he is known to be in $Y'$ within the required time. In case \eqref{pair}, probe the vertex at distance 2 from $a$ on $a\thrd a'$. If the response is 1, 2, $m-3$ or $m-2$, we have won. If it is 3 then we know the robber is one step from $a$ on a thread not leading to $a'$, and if it is $m-1$ then we know he is one step from $a'$ on a thread not leading to $a$; in either of these cases we are done by Lemma~\ref{delta}. If the response is $m+1$ then the robber is on a thread between $Y$ and $a$, and we are done by Lemma~\ref{astar}. If it is greater than $m+1$ and $\pm 2$ mod $m$, the robber is in $Y$ and so we are done. Finally, if the response is greater than $m+1$ and not $\pm 2$ mod $m$ then the robber must be one step from a vertex in $Y$ on a thread not leading to $a$, so we are done by Lemma~\ref{gamma}. In case \eqref{rest}, first choose some $aa'\in M$ such that $a\in A$ and probe $a$. If the result is 0 we have won and if it is any higher multiple of $m$ we know the robber is in $Y\cup A\setminus\{a\}$, so are done. Otherwise we either know that the robber is adjacent to $a$ or that he is adjacent to a vertex in $Y\cup A\setminus\{a\}$. Now choose any $bb'\in M$ not equal to $aa'$, take $v=a'$ and $w_1\thrd w_1',\ldots,w_{k-2}\thrd w_{k-2}'$ to be the threads corresponding to edges in $M$ other than $aa'$ and $bb'$, and proceed as in Lemma~\ref{noturnback}. If the robber returns to a branch vertex, either we know it is $a$ and have won, or we know it is in $Y\cup A\setminus\{a\}$ and are done. If he does not, we establish whether the robber is on a thread containing $a$ or $a'$, and for each $w_i\thrd w'_i$ in turn we establish whether he is on that thread, on an adjacent thread, or elsewhere. If one of these probes locates the robber's thread exactly then we interrupt the process and probe either end of that thread to win. If we establish that one end of the robber's thread is $x$ and the other is $w_i$ or $w'_i$, for some $x$ and $i$, then, since the previous probe was either at $x$ or a midpoint or off-midpoint of $w_i\thrd w'_i$, the result of that probe can be consistent with the robber being adjacent to $x$ or with him being adjacent to the other end of his thread, but not both. So we interrupt the process and probe $w_i$. If the result is $m$ we know whether the robber is at $x$ or $w'_i$, so have won; if it is less than $m$ we have located him on $x\thrd w_i$ and if it is more than $m$ he is on $x\thrd w'_i$ and we can locate him by probing $x$. If we establish that one end of the robber's thread is $w_i$ or $w'_i$ and the other is $w_j$ or $w'_j$, for some $i<j$, then again we interrupt the process. Since the last probe was at a midpoint or off-midpoint of $w_j\thrd w'_j$, the result of that probe can be consistent with the robber being within distance 2 of $\{w_i, w'_i\}$ or with him being within distance 2 of $\{w_j,w'_j\}$ but not both. Assume without loss of generality that we know he is not that close to $\{w_j,w'_j\}$. Now probe $w_i$. If the result is 0 or $m$ the robber is at $w_i$ or $w'_i$ respectively and we win; otherwise we have established which of these vertices is an end-point of his thread, and can win by probing $w_j$ and then (if necessary) $w'_j$. If none of these occur, we continue to the end of the $k-2$ stages of Lemma~\ref{noturnback}. The final probe will either establish that the robber is distance at most 2 from a branch vertex or establish that he is distance at least 2 from every branch vertex. Further, we will have established one of the following. \begin{enumerate}[(a)] \item Both ends of the robber's thread are in $\{b,b'\}\cup X$. \item One end of the robber's thread is in $\{c,c'\}$ and the other is in $\{b,b'\}\cup X$, for some $cc'\in M$ (this includes the case where we further know which of $c$ and $c'$ it is; we will not use this information). \end{enumerate} In case (a), by probing vertices in $\{b,b'\}\cup X$ in turn, we will either win or establish that he has reached a branch vertex in $\{b,b'\}\cup X$, reducing to \eqref{pair} as desired. We therefore assume case (b). If the robber was known not to be adjacent to a branch vertex at the last probe, then probe a midpoint of $b\thrd b'$. This will reveal if the robber is on a thread ending in $b$ or $b'$. If so, it will also establish either that he is distance at least 2 from $\{b,b'\}$ or that he is distance at least 2 from $\{c,c'\}$, without loss of generality the former. We can then win by probing $c$ followed by $b$ and $b'$. If the probe reveals that the robber is not in a thread meeting $b$ or $b'$, both ends of his thread are in $\{c,c'\}\cup X$, and by probing vertices in this set in turn we will win or reduce to \eqref{pair}. Alternatively, if the robber was known to be within 2 of a branch vertex at the end of the $k-2$ stages of Lemma~\ref{noturnback}, probe the vertex at distance 3 from $b$ along $b\thrd b'$. If the result is 3 or $m-3$ the robber is at $b$ or $b'$ respectively and is caught. If the result is larger and $\pm 3$ mod $m$, he is in $\{c,c'\}\cup X$. If the result is 4, 5 or 6 he is on $c\thrd b$ or $c'\thrd b$ near $b$, if it is $m+1$ or $m+2$ he is on one of these threads near the other end, and if it is $m-2$ or $m-1$ he is on $c\thrd b'$ or $c'\thrd b'$ near $b'$; in any of these cases probing $w_i$ will locate him. If the result is $m$ then the robber may be on $c\thrd b'$ or $c'\thrd b'$ near $b'$ or on $c\thrd b$ or $c'\thrd b$ near the other end, but he is not adjacent to a branch vertex so probing $b$ will determine which of these two possibilities is the case and then probing $c$ will locate him. If none of these apply then he is not on a thread meeting $b$ or close to $b'$, so probing $b'$ will determine whether he is on $c\thrd b'$ or $c'\thrd b'$. If he is we can locate him as before; if not both ends of his thread are in $\{c,c'\}\cup X$ and we can probe vertices in this set until we locate him or reduce to \eqref{pair}. This completes the analysis of cases, and so from any point where we know that the robber is in a set of branch vertices, we can reduce that set using \eqref{inx}, \eqref{pair} or \eqref{rest}. It is simple to check that in each case at most $2k+2+\abs{X}=n+2$ probes are required. At most $n-1$ reductions can occur before the set is a singleton and the robber is located, and so this strategy guarantees to catch the robber in bounded time. \end{proof} \section{Imperfect maximal matchings} Theorem~\ref{match} gives a bound in terms of the size of a maximal matching, and where there are maximal matchings of different sizes we are free to choose that which gives the best bound, i.e.\ the smallest one. This bound is therefore weakest when all maximal matchings are perfect matchings. We next show that, since $G$ is connected, there are only two possibilities for such a graph. Write $\mmm(G)$ for the minimum size of a maximum matching of $G$. \begin{lem}\label{mmm}If $G$ is a connected graph with $2r$ vertices such that $\mmm(G)=r$, then either $G\cong K_{2r}$ or $G\cong K_{r,r}$. \end{lem} \begin{proof}If $r=1$ then this is trivial. If $r=2$ and $G$ contains a vertex of degree 3, the each of those three edges can be extended to a matching of size 2, so every other pair of vertices is adjacent and $G\cong K_4$. If $r=2$ and $G$ has no vertex of degree 3 then, since $G$ is connected, $G\cong P_4$ or $G\cong K_{2,2}$, but $\mmm(P_4)=1$ by taking the middle edge. So the result is true for $r=2$. Let $G$ be such a graph for some $r>2$, and suppose the result is true for $r-1$. First note that if $G$ had a cutvertex, $v$, then $G-v$ would have at least one odd component, $C_1$ say, and another component $C_2$; letting $w$ be a neighbour of $v$ in $C_2$, $G-\{v,w\}$ would have an odd component, so the largest matching containing $vw$ would have size less than $r$, contradicting $\mmm(G)=r$. So $G$ is 2-connected. \begin{clm}If $vw$ is any edge of $G$ then $G-\{v,w\}$ is connected. \end{clm} \begin{poc}Suppose not. If any component of $G-\{v,w\}$ is odd, then $vw$ cannot be extended to a matching of size $r$, contradicting $\mmm(G)=r$. If all components are even then let $C_1$ and $C_2$ be any two components. Since $G$ is 2-connected, $v$ and $w$ each have neighbours in every component. Let $x$ be a neighbour of $v$ in $C_1$ and $y$ be a neighbour of $w$ in $C_2$. Now $C_1-x$ has an odd number of vertices, so at least one odd component, which is also an odd component of $G-\{v,w,x,y\}$. Consequently $\{vx,wy\}$ cannot be extended to a matching of size $r$, contradicting $\mmm(G)=r$. So the claim is proved. \end{poc} Since $G-\{v,w\}$ is connected for every $vw\in E(G)$, by the induction hypothesis it is isomorphic to $K_{2(r-1)}$ or $K_{r-1,r-1}$, since otherwise $G$ has a maximal matching of size $1+\mmm(G-\{v,w\})<r$. Next we show that the graph so obtained is isomorphic to the same one of these for every edge. \begin{clm}Either $G-\{v,w\}\cong K_{2(r-1)}$ for every $vw\in E(G)$, or $G-\{v,w\}\cong K_{r-1,r-1}$ for every $vw\in E(G)$. \end{clm} \begin{poc}Suppose $uv,vw\in E(G)$. Since $r>2$, there are vertices $x,y,z\not\in\{u,v,w\}$. If these vertices form a triangle in $G$ then $G-\{u,v\}\not\cong K_{r-1,r-1}$, so $G-\{u,v\}\cong K_{2(r-1)}$, and similarly $G-\{v,w\}\cong K_{2(r-1)}$. If they do not then neither graph is complete, so each must be isomorphic to $K_{r-1,r-1}$. So any pair of edges which share a vertex produce the same graph, and since $G$ is connected the same is true for any pair of edges. \end{poc} If $G$ is not complete then it has non-adjacent vertices $x$ and $y$. If every edge meets $x$ or $y$ then $\mmm(G)\leqslant 2<r$. So some edge does not, and then removing the endpoints of that edge gives a non-complete graph. So either $G$ is complete or $G-\{v,w\}\cong K_{r-1,r-1}$ for every $vw\in E(G)$. Suppose $G-\{v,w\}\cong K_{r-1,r-1}$ for every $vw\in E(G)$. Take disjoint edges $uv,wx,yz$ (this is possible since $G-\{u,v\}$ has $r-1$ disjoint edges). It is not possible for two vertices to appear in the same part of one of the graphs $G-\{u,v\},G-\{w,x\},G-\{y,z\}$ and different parts of another, so we can 2-colour the vertices of $G$ consistently with each of the colourings of these three graphs. Now any two vertices both appear in at least one of the three graphs, so are adjacent if and only if they are opposite colours. Since this means that $u$ and $v$ are opposite colours, and $G-\{u,v\}\cong K_{r-1,r-1}$, there are equal numbers overall and $G\cong K_{r,r}$. \end{proof} This allows us to complete the proof of our main theorem. \begin{thm}Let $G$ be a connected graph of order $n$, where $n\geqslant 23$. Then $G^{1/m}$ is locatable for any $m\geqslant n/2$. \end{thm} \begin{proof}If $m\geqslant n/2$ then also $m\geqslant 12$, since $n\geqslant 23$. If $G=K_n$ then $G^{1/m}$ is locatable by Theorem 4 of \cite{HJK}. If $G=K_{n/2,n/2}$ then $G^{1/m}$ is locatable by Theorem 5 of \cite{HJK}. Otherwise, by Lemma~\ref{mmm}, $G$ has a maximal matching $M$ of size $k$, where $k<n/2\leqslant m$. Since $k$ and $m$ are integers, $m\geqslant k+1$ and so, by Theorem~\ref{match}, $G^{1/m}$ is locatable. \end{proof} \section{Unequal subdivisions} So far all our results, like those of \cite{CCDEW}, have assumed equal-length subdivisions. Carraher et al.\ conjectured that this restriction was not necessary, and that further subdividing a locatable graph always gives another locatable graph, but the present authors \cite{HJK} and Seager \cite{Sea14} independently gave a counterexample to this conjecture. Here we present a proof that a subdivision where the edges of $G$ are replaced by paths of arbitrary length is locatable, provided that the minimum length of path used is sufficiently large. For a given graph $G$ and function $l:E(G)\to\mathbb{Z}^+$, define $G^{1/l}$ to be the graph obtained from $G$ by replacing each edge $e\in E(G)$ by a path of length $l(e)$. \begin{thm}For any graph $G$ with $n$ vertices, if $l(e)\geqslant 2n$ for every $e\in E(G)$ then $G^{1/l}$ is locatable. \end{thm} \begin{proof}We give a winning strategy on $G^{1/l}$. First probe each branch vertex in turn, and write $d_x$ for the distance obtained from probing $x$. Let $(u,v)$ be a pair for which $(d_u,d_v)$ is as small as possible (lexicographically). Now consider the robber's possible positions at the time of the next probe. Write $x$ for the nearest branch vertex and $d$ for the robber's distance from that branch vertex. If the robber is not on a thread containing $u$ then $d_u\geqslant 2n+d-n$ and $d_x\leqslant d+n$. We cannot have equality in both cases, since the first would require $u$ to have been the first vertex probed, and the second would require $x$ to have been the first vertex probed. Thus $d_u>d_x$, contradicting the definition of $u$. So the robber must be on some thread containing $u$; let the other end of the thread be $w$. The shortest path from $u$ to the robber's current location is along $u\thrd w$, for if not $d_u\geqslant 2n+d-n$ and $d_w\leqslant d+n$ (and again equality cannot occur in both cases simultaneously). Suppose $x\neq u$, i.e. the robber is nearer to $w$ than to $u$. Then $d_w\geqslant d+n$, but if $v\neq w$ then $d_v\leq 2n+d-n$. Again, we can have equality in at most one of these, so $v\neq w$ would imply $d_v>d_w$, contradicting the definition of $v$. Consequently either $x=u$ or $x=v$. Now probe $v$. If the distance returned is at most $n$, the robber must be on $u\thrd v$ and we have won. Otherwise, probe $u$, and write $d'$ for the distance returned. If $d'=0$ we have won, so assume $d'>0$. If $w=v$ then we know from the previous probe that the robber's distance from $w$ (at the time of probing $u$) is at least $n-1$, whereas if $w\neq v$ then we know the robber's distance from $w$ (at the time of probing $u$) is at least $l(uw)/2-1$, which is again at least $n-1$ (since otherwise the robber was nearer to $w$ than $u$ on the previous turn, so $w=v$). For every branch vertex $y$ with $uy\in E(G)$ and $d'+n-1\leqslant l(uy)$, select the $(d'+n-1)$th vertex on $u\thrd y$ (counting $u$ as the $0$th). Note that every possible candidate for $w$ is covered, and there are at most $n-1$ vertices selected. Probe these vertices in turn until the distance returned is at most $\min(2(n-1),d'+n-1)$. This happens if and only if the robber is on the thread being probed (including the case where the robber has returned to $u$), so it must eventually occur. When it does we have identified the robber's location uniquely. \end{proof} \section{Acknowledgements} The first author acknowledges support from the European Union through funding under FP7--ICT--2011--8 project HIERATIC (316705), and is grateful to Douglas B. West for drawing his attention to this problem. The second author acknowledges support through funding from NSF grant DMS~1301614 and MULTIPLEX grant no.\ 317532, and is grateful to the organisers of the 8th Graduate Student Combinatorics Conference at the University of Illinois at Urbana-Champaign for drawing his attention to the problem. The third author acknowledges support through funding from the European Union under grant EP/J500380/1 as well as from the Studienstiftung des Deutschen Volkes.

https://arxiv.org/abs/2110.07214

Confining integro-differential equations originating from evolutionary biology: ground states and long time dynamics

We consider nonlinear mutation selection models, known as replicator-mutator equations in evolutionary biology. They involve a nonlocal mutation kernel and a confining fitness potential. We prove that the long time behaviour of the Cauchy problem is determined by the principal eigenelement of the underlying linear operator. The novelties compared to the literature on these models are about the case of symmetric mutations: we propose a new milder sufficient condition for the existence of a principal eigenfunction, and we provide what is to our knowledge the first quantification of the spectral gap. We also recover existing results in the non-symmetric case, through a new approach.

\section{Introduction}\label{sec:intro} The starting point of this work is the nonlinear integro-differential equation \begin{equation} \label{eq:rep-mut} \partial_{t}u = \sigma ^2(J*u-u)-\left(W(x)-\overline W[u](t)\right)u, \quad t>0,\; x\in \mathbb R^N, \end{equation} where $\sigma>0$ is a given parameter, $J : x \mapsto J(x)$ a probability density on $\mathbb R^N$, and $W : x \mapsto W(x)$ a so-called confining potential (meaning that $W(x)\to +\infty$ as $|x| \to +\infty$), while ${\overline W}[u](t)$ is a nonlocal term defined by $$ {\overline W}[u](t) := \langle W, u(t,\cdot)\rangle := \int_{\mathbb R^{N}} W(y)u(t,y)dy,$$ and can be seen, at least formally, as a Lagrange multiplier: it ensures that the solution to \eqref{eq:rep-mut} starting at $t=0$ from a probability density $u_{0}(x) = u(0,x)$ remains a probability density for $t>0$. This model is known as a {\it replicator-mutator} model in evolutionary biology. In this context, $u(t,x)$ stands for the density of a population, at time $t$, per unit of phenotypic trait $x$ on the multi-dimensional phenotypic trait space $\mathbb R^N$. The function $x \mapsto - W(x)$ represents the {\it fitness} of the phenotype $x\in \mathbb R^N$ and models the individual reproductive success, and $t \mapsto -\overline W[u](t)$ stands for the mean fitness of the population at time $t$. Our main goal in this paper is to understand, under appropriate assumptions, the long time behaviour of the Cauchy problem associated with \eqref{eq:rep-mut}. To do so, we prove two intermediate results on the underlying linear problem, which we believe are interesting in their own right. First, concerning the linear eigenvalue problem \begin{equation} \label{eq:eigenvalue} -\sigma^2(J*u-u)+W(x)u=\lambda u \quad \text{ in } \mathbb R^N, \end{equation} we prove that there exists a unique nonnegative principal eigenfunction $\phi$ with total mass one, which is also called a Perron eigenfunction or {\it ground state} for such problems, see Theorem~\ref{thm:eigen}. Next, concerning the long time dynamics of the linear Cauchy problem \begin{equation} \label{eq:parabolique} \partial_{t}u = \sigma ^2(J*u-u)-W(x)u, \quad t>0,\; x\in \mathbb R^N, \end{equation} we show that it is well-posed in a family of relevant Banach spaces, and we prove that its long time dynamics is determined by the principal eigenfunction $\phi$, see Theorem~\ref{thm:asympto}. From this point on, we are in a position to establish our result on the nonlinear Cauchy problem~\eqref{eq:rep-mut}: we prove that any solution converges, as the time $t\to+\infty$, to a multiple of the ground state, see Theorem~\ref{thm:rep-mut}. \medskip Replicator-mutator models aim at describing Darwinian evolutionary processes, whose fundamental principles are mutations and selection. Under the constraint of constant mass $\textstyle \int_\mathbb{R} u(t,x)\, dx=1$, the replicator dynamics is given by \begin{equation*}\label{nodiffusion} \partial_{t}u = -\left(W(x) - {\overline W[u](t)}\right)u. \end{equation*} As an attempt to take into account evolutionary phenomena, mutations are typically modelled by integral operators, thus yielding models such as \eqref{eq:rep-mut}; we refer to the influential work of M. Kimura \cite{Kimura1965}, as well as R. Lande \cite{Lan-75}, W.H. Fleming \cite{Fle-79} and R. B\"urger \cite{Bur-86, Bur-88}. In some situations, for instance those discussed in R. B\"urger \cite[Chapter VI, subsection 6.4]{Bur-00-book}, these integral operators can be approximated by a local diffusion operator, as in \begin{equation} \label{eq:rep-mut-laplacien} \partial_{t}u = \sigma ^2\Delta u -\left(W(x)-\overline W[u](t)\right)u. \end{equation} We refer to the work of N. Champagnat, R. Ferrière \& S. Méléard~\cite{Champagnat2008} or the recent paper by J.Y. Wakano, T. Funaki \& S. Yokoyama \cite{Wak-Fun-Yok-17} for a rigorous derivation of the replicator-mutator problem from individual based models. The case of a linear fitness function, that is when $W(x) := -x$ (say $N=1$), was completely investigated in the works by M. Alfaro \& R. Carles~\cite{Alf-Car-14} (Laplacian case, Equation \eqref{eq:rep-mut-laplacien}) and R. Bürger~\cite{Bur-91} and M.E. Gil \& al. \cite{Gil-17} (mutation kernel case, Equation \eqref{eq:rep-mut}), whereas the quadratic case, that is $W(x) := -x^{2}$ (say $N=1$), was studied in M. Alfaro \& R. Carles~\cite{Alf-Car-17} for Equation \eqref{eq:rep-mut-laplacien}. These two cases share the property consisting in the fact that the function $W$ is unbounded from below, meaning that some phenotypes are infinitely well-adapted. These two cases yield rich mathematical behaviours (acceleration, extinction) but, unless introducing some {\it context-dependent} mutation kernels as performed by M.E. Gil \& al. in \cite{Gil-Ham-Mar-Roq-18}, are not well-suited models for biological applications. On the other hand, the confining prototype case $W(x) := x^{2}$ (say $N=1$), considered in the pioneering work by M. Kimura~\cite{Kimura1965} and analyzed by R. Bürger~\cite{Bur-86} for Equation~\eqref{eq:rep-mut} and by M. Alfaro and R. Carles~\cite{Alf-Car-17} for Equation \eqref{eq:rep-mut-laplacien}, prevents extinction phenomena and leads to convergence to the underlying principal eigenfunction. In a related but different setting, let us mention the recent work of F. Hamel \& al. \cite{Hamel2020} where a \lq\lq two-patches environment'' is considered. However, the case $W(x)=x^2$ does not suffice to take into account more realistic cases for which fitness functions are defined by a linear combination of two components (e.g. birth and death rates), each maximized by different optimal values of the underlying trait, a typical case being $W(x) := x^4 - x^2$. In this setting, let us mention the recent work of M. Alfaro \& M. Veruete \cite{Alf-Ver-18} which provides a rigorous treatment of the Cauchy problem \eqref{eq:rep-mut-laplacien} when the fitness function $W$ is confining, and also raises the issue of \emph{evolutionary branching}, consisting of the spontaneous splitting from uni-modal to multi-modal distribution of the trait. Our main goal is to extend these results to the mutation kernel model \eqref{eq:rep-mut}, revisiting the work of R. B\"urger~\cite{Bur-88} and refining it in the case when the convolution kernel $J$ is an even function (that is $J(x) = J(-x)$ for all $x\in {\Bbb R}^N$). This requires to perform first a detailed analysis of the integral eigenproblem \eqref{eq:eigenvalue}. The local counterpart to \eqref{eq:eigenvalue}, namely $$ -\sigma^2\Delta u + W(x)u = \lambda u, $$ is well understood, see for instance M. Reed \& B. Simon \cite{ReedSimonVol4}, L.A. Takhtajan \cite{Takhtajan}. In the quantum mechanics terminology, the principal eigenfunction $\phi$ is called the ground state and corresponds to the bound-state having minimal energy, and the principal eigenvalue $\lambda_{1}$ is characterized by a classical variational formulation. In the non-local case \eqref{eq:eigenvalue}, the principal eigenvalue is not necessarily associated with a principal eigenfunction. The lack of regularizing effect of the integral operator compared to the Laplace operator may result in situations where the ground state is a singular measure (containing atoms). It happens when the fitness function $W$ is confining and has a cusp at its minimum, as noticed first by R. Bürger in~\cite{Bur-88} and then further investigated by R. Bürger \& I.M. Bomze in~\cite{Bur-Bom}. More recently, this problem was revisited by J. Coville and co-authors in the case of a bounded domain \cite{Bon-Cov-Leg,Cov-10,Cov-13,Li-Cov-Wan}, see also Q. Griette \cite{Gri-19}. In the study of these concentration phenomena (formation of Dirac masses), a first fundamental question is to identify sharp conditions on $J$ and $W$ that ensure the existence (or non-existence) of a principal eigenfunction. Such conditions can be found in~\cite{Bur-88,Bur-Bom,Cov-10,Li-Cov-Wan}, and the first contribution of the present paper is to give a new milder criterion (see Assuption~\ref{ass:linkJW}) in the case of an even convolution kernel $J$. We also recover, through a different method, the existing criteria in the non-even case, see Section~\ref{sec:non-sym}. The second novel result is a quantification of the spectral gap in $L^2$, and accordingly of the rate of convergence to the equilibrium for Equation~\eqref{eq:rep-mut} in this space, under some stronger assumptions, see Theorem~\ref{thm:spectral-gap}. Notice that in the works of R. Bürger or J. Coville, more general mutation kernels than convolution are usually considered, {\it i.e.} with $k(x,y)$ in place of $J(x-y)$. To simplify the statement of the assumptions and limit technicalities in the proofs, we choose to treat only the convolution case, which is biologically relevant, see~\cite{Kimura1965}. However, the proofs can be adapted to encompass more general kernels, replacing the evenness of $J$ by the symmetry of $k$ (namely $k(x,y)=k(y,x)$). \section{Main results}\label{sec:main} Before stating our main results, we give the notations and assumptions used throughout the paper. For an almost everywhere positive measurable weight function $\rho$ we denote by $L^2(\rho)$ the weighted Lebesgue space $$L^2(\rho) := \left\{u : {\Bbb R}^N \longrightarrow {\Bbb R} \text{ such that } u \text{ is measurable and }\, \int_{{\Bbb R}^N}|u(x)|^2\rho(x) dx < \infty \right\}.$$ We shall denote by $\langle \cdot\,,\cdot\rangle$ the scalar product of $L^2({\Bbb R}^N)$ and by $\|\cdot\|_{L^2}$ its associated norm. For any choice of space $E$ among either of the Lebesgue spaces $L^{p}(\mathbb R^{N})$ with $1 \leq p <\infty$, the space of continuous functions converging to zero at infinity $C_{0}(\mathbb R^{N})$, or $\MM(\mathbb R^N)$ the space of bounded measures on ${\Bbb R}^N$, we denote by $\|\cdot\|_{E}$ the standard associated norm, namely the $L^p$ norm, the $L^\infty$ norm, or the total variation norm, respectively. If $X,Y$ are two Banach spaces and $A : X \longrightarrow Y$ is a linear bounded operator we denote $\|A\|_{X \to Y}$ its norm in the space $\LL(X,Y)$. Our precise assumptions on the mutation kernel and on the potential are the following. \begin{assumption}[On the kernel $J$]\label{ass:kernel} The kernel $J:\mathbb R^{N}\longrightarrow \mathbb R$ satisfies \begin{enumerate} \item $J(x) \geq 0$ for almost all $x \in \mathbb R^N$. \item There exists $r_0>0$ such that $J(x) > 0$ for almost all $x$ in the ball $B(0,r_{0})$. \item $J\in L^{1}(\mathbb R ^{N})\cap L^{2}(\mathbb R ^{N})$ and $\displaystyle \int_{\mathbb R^N}J(x)dx = 1$. \end{enumerate} \end{assumption} \begin{assumption}[Symmetry of the kernel $J$]\label{ass:J-symmetric} The kernel $J:\mathbb R^{N}\longrightarrow \mathbb R$ satisfies $J(-x) = J(x)$ for almost all $x \in \mathbb R^N$. \end{assumption} \begin{assumption}[On the potential $W$]\label{ass:potential} The potential $W:\mathbb R^{N}\longrightarrow \mathbb R$ is a continuous function which satisfies \begin{enumerate} \item $W(x)\to +\infty$ as $|x| \to +\infty$ (confining assumption). \end{enumerate} Equation~\eqref{eq:rep-mut} is left unchanged by adding a constant to $W$, so we assume w.l.o.g. that \begin{enumerate}[resume] \item $W\geq 0$. \end{enumerate} \end{assumption} In the sequel, for a function $f : \mathbb R^N \longrightarrow \mathbb R$ and a constant $\alpha \in \mathbb R$, the set of points $x \in \mathbb R^N$ such that $f(x) \geq \alpha$ is denoted by $[f \geq \alpha]$. \begin{assumption}[Linking $J$, $\sigma$, and $W$]\label{ass:linkJW} There exist $\epsilon > 0$ and a Borel set $B\subset\mathbb R^N$ such that \begin{equation} \label{hyp:non-int} \sigma^2\iint_{B_\epsilon\times B_\epsilon}\frac{J(x-y)}{W(x)W(y)}\,dxdy>\int_{B_\epsilon}\frac{1}{W(x)}\,dx \end{equation} where $B_\epsilon:=B\cap[W\geq\epsilon]$. \end{assumption} \medskip We define the linear operator $(L,D(L))$ by setting \begin{equation}\label{eq:Def-L} Lu := - K*u + W(x) u, \quad\text{for }\, u \in D(L) := L^2(1+W), \qquad\text{where }\, K := \sigma^2J. \end{equation} It can be easily seen that $(L,D(L))$ is an unbounded operator on $L^2({\Bbb R}^N)$, and that when $J$ is an even function, that is when Assumption \ref{ass:J-symmetric} is satisfied, the operator $L$ is self-adjoint. Also, since Young's inequality yields $\|K*u\| _{L^2}\leq \|K\|_{L^1}\|u\|_{L^2} = \sigma^2\|u\|_{L^2}$ for all $u \in D(L) = L^2(1+W)$, we have $$\langle Lu,u\rangle \geq - \sigma^2\|u\|_{L^2}^2,$$ so that when Assumption \ref{ass:J-symmetric} is satisfied, that is when $L$ is self-adjoint, the numerical range of $L$, and thus its spectrum, is contained in $[-\sigma^2,+\infty)$. In general neither $L$ nor its resolvent are compact, so in order to show that it has an eigenvalue, when $L$ is self-adjoint, we use the variational characterization of the bottom of its spectrum. Namely, setting \begin{equation} \label{def:S} S := \left\{u\in L^2(1+W)\text{ such that } \int_{\mathbb R^N} u^2(x)dx = 1\right\}, \end{equation} we define the energy functional \begin{equation} \label{def:energy} \mathscr E(u):= \langle Lu,u\rangle = -\int_{\mathbb R ^N} (K*u)(x)u(x)dx + \int_{\mathbb R^N}W(x)u^2(x)dx, \end{equation} and its infimum \begin{equation} \label{def:lambda-un} \lambda_{1} := \inf_{u\in S} \mathscr E(u), \end{equation} and we show that under our assumptions $\lambda_{1}$ is achieved. Our first main result concerns the eigenvalue problem \eqref{eq:eigenvalue}, namely existence and uniqueness (up to a multiplicative constant) of a principal eigenfunction, or ground state. \begin{theorem}[Principal eigenpair, symmetric case]\label{thm:eigen} Let Assumptions \ref{ass:kernel}, \ref{ass:J-symmetric}, \ref{ass:potential} and \ref{ass:linkJW} hold. Then $\lambda_{1}$, defined by \eqref{def:lambda-un}, is achieved by a unique $\phi \in S \cap C_{0}(\mathbb R^N)$ such that $\phi > 0$ and \begin{equation}\label{eq-vp} -K*\phi + W(x)\phi = \lambda_{1} \phi\quad \text{ in }\, \mathbb R ^N. \end{equation} Moreover, \begin{enumerate}[label=\roman*)] \item $-\sigma^2 < \lambda_{1}\leq -b_\epsilon$, where \begin{equation}\label{eq:b-eps} b_\epsilon:=\bigg(\sigma^2\iint_{B_\epsilon\times B_\epsilon}\frac{J(x-y)}{W(x)W(y)}\,dxdy-\int_{B_\epsilon}\frac{dx}{W(x)}\bigg)\bigg(\int_{B_\epsilon}\frac{dx}{W(x)}\bigg)^{-2}>0, \end{equation} \item $\displaystyle\forall x\in\mathbb R^N,\ 0 < \phi(x) \leq \frac{\|K\|_{L^2}}{W(x) - \lambda_{1}}$, \item $\lambda_{1}$ is an eigenvalue of multiplicity one, that is if $\psi\in L^2(\mathbb R^N)$, or $\psi \in C_{0}(\mathbb R^N)$, is such that \begin{equation*}\label{psi-ae} -K*\psi + W(x)\psi = \lambda_{1} \psi \quad \text{ a.e. in }\, \mathbb R^N, \end{equation*} then there is $\alpha \in \mathbb R$ such that $\psi = \alpha \phi$. \end{enumerate} \end{theorem} \medskip We now turn to the long time dynamics of the linear equation \eqref{eq:parabolique}. Clearly, under the assumptions of Theorem \ref{thm:eigen} the pair $(\lambda_{*} ,\phi)$ with \begin{equation}\label{eq:Def-lambda-*} \lambda_{*} := \lambda_1 + \sigma^2 > 0 \end{equation} is solution to~\eqref{eq:eigenvalue}, and thus the function $u_{*}$ defined by \begin{equation*}\label{eq:Def-u*} u_{*}(t,x) := {\rm e}^{-\lambda_{*}t}\phi(x), \end{equation*} is solution to~\eqref{eq:parabolique}. In fact the large time behaviour of any nonnegative solution to \eqref{eq:parabolique} is given by this particular solution (recall that we denote by $E$ either of the spaces $L^p({\Bbb R}^N)$ for $1\leq p < \infty$, or $C_{0}({\Bbb R}^N)$ or $\MM({\Bbb R}^N)$). \begin{theorem}[Long time dynamics, linear case]\label{thm:asympto} Let Assumptions \ref{ass:kernel}, \ref{ass:J-symmetric}, \ref{ass:potential} and \ref{ass:linkJW} hold, and for any $u_{0}\in E$ denote by $\langle u_{0},\phi \rangle = \int_{\mathbb R^N}u_{0}(x)\phi(x)dx$ its weighted mass\footnote{As will be proved in Corollary \ref{villoise}, $\varphi \in L^1(\mathbb R^N)$ and thus, since $\varphi \in C_0(\mathbb R^N)$, $\varphi \in L^q(\mathbb R^N)$ for any $1\leq q \leq \infty$. In particular the weighted mass $\langle u_{0},\phi \rangle$ is finite as soon as $u_0\in E$.}. Then there exist two constants $C > 0$ and $a > 0$ such that, for any $u_{0}\in E$, the solution $u=u(t,x)$ of \eqref{eq:parabolique} starting from $u_{0}=u_0(x)$ (see Section \ref{sec:asympto} for the precise definition) satisfies, for all $t > 0$, $$ {\rm e}^{\lambda_{*}t} \big\|u(t,\cdot) - \langle u_{0},\phi \rangle u_{*}(t,\cdot)\big\|_{E} = \big\| {\rm e}^{\lambda_{*} t}u(t,\cdot)-\langle u_{0},\phi \rangle \phi\big\|_{E} \leq C\, {\rm e}^{-at}\left\|u_{0} - \langle u_{0},\phi \rangle\phi\right\|_{E} . $$ \end{theorem} \medskip The next result deals with the nonlinear replicator mutator model \eqref{eq:rep-mut}. This equation is concerned with the evolution of probability distributions, so that we only consider an initial data $u_{0}\in E$ which is nonnegative and such that $\langle u_{0},\mathbf 1\rangle=1$, where $\mathbf 1$ stands for the constant function equal to $1$ over $\mathbb R^N$. \begin{theorem}[Long time dynamics, replicator-mutator model]\label{thm:rep-mut} Let Assumptions \ref{ass:kernel}, \ref{ass:J-symmetric}, \ref{ass:potential} and \ref{ass:linkJW} hold. Then there exists $a>0$ such that, for any $u_{0}\in E$ with $u_{0} \geq 0$ and $\langle u_{0},\mathbf 1\rangle=1$, there is a constant $C(u_{0})>0$ for which the solution $u=u(t,x)$ of \eqref{eq:rep-mut} starting from $u_{0} = u_{0}(x)$ (see Section \ref{sec:rep-mut} for the precise definition) satisfies, for all $t > 0$, $$ \left\| u(t,\cdot)-\frac{\phi}{\langle\phi,\mathbf 1\rangle}\right\|_{E} \leq C(u_{0}) \, {\rm e}^{-at}. $$ Moreover, in the case $E = L^1(\mathbb R^N)$ or $E = \MM(\mathbb R^N)$, there exists a constant $c_{0} > 0$ such that for any $u_{0} \in E$ with $u_{0} \geq 0$ and $\langle u_{0},\mathbf 1\rangle=1$, we can take $$C(u_{0}) = c_{0}\left\|\frac{u_{0}}{\langle u_{0},\phi \rangle} - \phi\right \|_{E}.$$ \end{theorem} \medskip Our last main result provides a quantitative estimate of the rate of convergence $a$ appearing in Theorems \ref{thm:asympto} and \ref{thm:rep-mut}. To do so we need to further assume that $W$ admits a unique global minimum, which we can choose w.l.o.g. to be the origin. \begin{assumption}\label{ass:W-fort} The potential $W:\mathbb R^N\longrightarrow\mathbb R$ is a continuous function which satisfies \begin{enumerate} \item $W(x)\to +\infty$ as $|x| \to +\infty$. \item $W(0)=0$. \item For all $x\neq0$, $W(x)>0$. \end{enumerate} \end{assumption} Then we consider a convex, open and bounded set $\Omega\subset\mathbb R^N$ such that $0\in\Omega$, and we define \[\eta:=\esssup_{2\Omega}K-\essinf_{2\Omega}K,\] where $2\Omega=\{x\in\mathbb R^N:x/2\in\Omega\}$. We also define the function $\Phi:[0,+\infty)\to\mathbb R\cup\{+\infty\}$ by \begin{equation}\label{def:Phi} \Phi(\xi):=\min\Big\{\sigma^2-\xi,\eta\meas(\Omega)+a_1\xi+a_2\sqrt{\xi}\Big\}, \end{equation} where \[a_1:=\Big(\essinf_{2\Omega}K\Big)\bigg(\int_{\Omega^c}\frac1W\bigg)\quad\text{and}\quad a_2:=2\sigma\sqrt{\sup_{x\in\mathbb R^N}\int_{\Omega^c}\frac{K(x-y)}{W(y)}dy},\] and we denote by $\bar\Phi\in(0,+\infty]$ the largest value of $\Phi$ \[\bar\Phi:=\sup_{\xi\geq0}\Phi(\xi).\] \begin{theorem}[Spectral gap and rate of convergence]\label{thm:spectral-gap} Let Assumptions~\ref{ass:kernel}, \ref{ass:J-symmetric}, \ref{ass:linkJW} and \ref{ass:W-fort} hold. Then, for any eigenvalue $\lambda$ of the operator $L$ in $L^2(\mathbb R^N)$, we have \[\lambda\neq\lambda_1\quad\implies\quad \lambda\geq-\bar\Phi.\] As a consequence, if the sets $\Omega$ and $B_\epsilon$ are such that \[a_*:=b_\epsilon-\bar\Phi>0\] then the convergences in Theorems~\ref{thm:asympto} and~\ref{thm:rep-mut} hold in $L^2(\mathbb R^N)$ for any $a<a_*$. \end{theorem} \medskip Let us now comment on the above assumptions and results. \begin{enumerate} \item Assumption~\ref{ass:linkJW} is the crucial condition which ensures, when $J$ is even, that the ground state is a function and not a singular measure. It relaxes the existing criteria, the mildest one in the literature being, to the best of our knowledge, the one proposed by F. Li, J. Coville \& X. Wang in~\cite{Li-Cov-Wan} (see Condition~(2.3) in Theorem~2.1). In the case when $\meas[W=0]=0$, their condition can be written in the following form (see Appendix~\ref{app:ex1} for the proof of the equivalence). \begin{assumption}\label{ass:Coville} There exist $\epsilon>0$ and a Borel set $B\subset\mathbb R^N$ such that \begin{equation*}\label{hyp:Coville} \sigma^2\essinf_{x\in B_\epsilon}\int_{B_\epsilon}\frac{J(x-y)}{W(y)}\,dy>1 \end{equation*} where we recall the notation $B_\epsilon=B\cap[W\geq\epsilon]$. \end{assumption} Note that this condition can also be proved to be sufficient from the paper \cite{Bur-88} by R. Bürger, see \cite[p.\,250, {\it Note added in proof.}]{Bur-Bom}. Clearly, Assumption~\ref{ass:Coville} implies Assumption~\ref{ass:linkJW}, but the converse is false as shown by the following example. \begin{example}\label{ex:1} Consider the one dimensional space $\mathbb R^N=\mathbb R$ and \[J(z)=\frac12\mathbf 1_{[-1,1]}(z)\qquad\text{and}\qquad W(x)=\sqrt{|x|}.\] Then Assumption~\ref{ass:Coville} is satisfied if and only if $\sigma^2>\frac 1{\sqrt 2}$, while Assumption~\ref{ass:linkJW} is verified as soon as $\sigma^2>\frac{4}{4+\pi}$. We refer to Appendix~\ref{app:ex1} for a proof of this claim. \end{example} The question whether Assumption~\ref{ass:linkJW} is enough for ensuring the existence of a first eigenfunction without the evenness condition on $J$ is still an open question. Finally, let us also mention that the condition \[\sigma^2\esssup_{x\in \mathbb R^N}\int_{\mathbb R^N}\frac{J(x-y)}{W(y)}\,dy<1\] guarantees that the ground state is a singular measure (with atoms), and consequently no principal eigenfunction exists, see~\cite{Bur-Bom}. \item Since the work of S. Mischler and J. Scher~\cite{Mis-Sch} in 2016, quantifying the spectral gap of non-local and non-conservative linear equations is an active field of research, see~\cite{Ban-Clo-Gab-Mar,Can-Gab-Yol,Clo-Gab-21,Clo-Gab-20,Gab-Mar}. To our knowledge, the result in Theorem~\ref{thm:spectral-gap} is the first quantified spectral gap result in the literature for Equation~\eqref{eq:parabolique}. For some particular choices of coefficients, as the one in the following example, it provides an estimate of the spectral gap. \begin{example}\label{ex:2} Consider the one dimensional space $\mathbb R^N=\mathbb R$ and \[J(z)=\frac14\mathbf 1_{[-2,2]}(z)\qquad\text{and}\qquad W(x)=|x|^m\ (m>1).\] Then the lower bound $a_*$ on the spectral gap satisfies \[a_*\geq\sigma^2\left(\frac14-\frac{\sigma^2}{(m-1)2^m}-2\frac{\sigma}{\sqrt{(m-1)2^m}}\right)\] and is thus positive for $\sigma^2$ small enough or $m$ large enough. We refer to Appendix~\ref{app:ex2} for a proof of this estimate. \end{example} \end{enumerate} \medskip Assumptions~\ref{ass:kernel} and \ref{ass:potential} are supposed to be verified throughout the paper, while the symmetry of $J$, that is Assumption~\ref{ass:J-symmetric}, is assumed only in Sections~\ref{sec:L}, \ref{sec:asympto} and \ref{sec:rep-mut}, Section~\ref{sec:non-sym} being devoted to the non-symmetric case. The remainder of the paper is organized as follows. In Section \ref{sec:L} we gather some results concerning the operator $L$: we give a strong maximum principle for this operator, we prove Theorem~\ref{thm:eigen} on the eigenvalue problem \eqref{eq:eigenvalue}, and we establish a functional inequality which yields the first part of Theorem~\ref{thm:spectral-gap} about the quantification of the spectral gap. The long time behaviour of the linear problem \eqref{eq:parabolique}, see Theorem \ref{thm:asympto} and the second part of Theorem~\ref{thm:spectral-gap}, is studied in Section \ref{sec:asympto}. Theorem \ref{thm:rep-mut} on the replicator-mutator model \eqref{eq:rep-mut} is proved in Section \ref{sec:rep-mut}. Finally, in Section~\ref{sec:non-sym}, we give a new proof of known results when $J$ is not assumed to be even. \section{The operator $L$}\label{sec:L} In this section, we investigate some remarkable properties of the operator $L$, defined in~\eqref{eq:Def-L}. \subsection{A strong maximum principle}\label{sec:MaxPrinciple} In the context of an elliptic second order equation such as $$-\sum_{i,j=1}^N \partial_{i}(a_{ij}(x)\partial_{j}u) + (c(x) + \lambda) u = f \qquad \text{in }\, \mathbb R^N,$$ where the matrix $(a_{ij})_{1\leq i,j \leq N}$ is uniformly coercive and $a_{i,j},c \in L^\infty(\mathbb R^N)$, while for instance $f \in C_{0}(\mathbb R^N)\cap L^2(\mathbb R^N)$, it is well known that if $f \geq 0$ and $f \not\equiv 0$ then $u > 0$ in $\mathbb R^N$, provided $c(x) +\lambda > 0$ a.e. in $\mathbb R^N$. The nonlocal operator $L$ defined above in \eqref{eq:Def-L} satisfies an analogous strong maximum principle. \begin{lemma}\label{lem:Max-principle} Assume that Assumptions \ref{ass:kernel} and \ref{ass:potential} are satisfied. Let $\lambda > \sigma^2$ and $f \in L^p(\mathbb R^N)$ such that $ f \geq 0$ and $f \not\equiv 0$, with $1 \leq p \leq \infty$. Let $u \in L^p(\mathbb R^N)$ satisfy $$L u + \lambda u = -K*u + (W(x)+ \lambda)u = f.$$ Then $u > 0$ a.e. on $\mathbb R^N$. \end{lemma} \begin{proof} First, assuming that $1 \leq p < \infty$, we show that $u \geq 0$. Writing $u = u^+ - u^-$ with $u^+ := \max(u,0)$, we have $$K*u^- + (W(x)+ \lambda)u = f + K*u^+ \geq 0.$$ Multiplying this by $(u^-)^{p-1}\mathbf 1_{[u < 0]}$ and integrating we get $$\int_{\mathbb R^N}(K*u^-)(x)(u^-)^{p-1}\mathbf 1_{[u < 0]}(x)dx - \int_{\mathbb R^N}(W(x) + \lambda)|u^-(x)|^{p}dx \geq 0.$$ However, using H\"older's inequality with $K*u^- \in L^p(\mathbb R^N)$ and $(u^-)^{p-1}\mathbf 1_{[u < 0]} \in L^{p'}(\mathbb R^N)$ where we denote $p' := p/(p-1)$, we have \begin{eqnarray*} \int_{\mathbb R^N}(K*u^-)(x)(u^-)^{p-1}\mathbf 1_{[u < 0]}(x)dx & \leq \|K*u^-\|_{L^p}\|(u^-)^{p-1}\mathbf 1_{[u < 0]}\|_{L^{p'}} \\ &\leq \|K\|_{L^1}\|u^-\|_{L^p}^p = \sigma^2\|u^-\|_{L^p}^p, \end{eqnarray*} and thus from the previous inequality we infer that $$\int_{\mathbb R^N} (W(x) + \lambda - \sigma^2)|u^-(x)|^pdx \leq 0,$$ that is, since $W \geq 0$ and $\lambda > \sigma^2$, we have $u^- \equiv 0$, and thus $u = u^+ \geq 0$. Next assume that $p = \infty$, and denote by $m$ the essential infimum of $u$, that is $$m := \essinf\limits_{x \in \mathbb R^N} u(x).$$ Since $K \geq 0$, we deduce that, for a.e. $x \in \mathbb R^N$, we have $$\int_{\mathbb R^N}K(x-y)u(y)dy \geq m \int_{\mathbb R^N}K(x-y)dy = m\,\sigma^2,$$ and thus $$(W(x) + \lambda)u(x) \geq f(x) + \int_{\mathbb R^N}K(x-y)u(y)dy \geq m\,\sigma^2.$$ Since $W \geq 0$ and $\lambda > \sigma^2$, this inequality implies that $m\geq 0$. Indeed, if $m < 0$, taking a sequence $(x_{n})_{n \geq 1}$ such that $u(x_{n}) \to m$ as $n \to \infty$ and $m \leq u(x_{n+1}) \leq u(x_{n})$, then for $n$ large enough so that $u(x_{n}) < 0$ and $$0 \leq u(x_{n}) - m \leq \frac{1}{2\lambda}(\lambda - \sigma^2)|m|,$$ we would have $$0 > (\lambda - \sigma^2)m + \lambda (u(x_{n}) - m) \geq - W(x_{n})u(x_{n}) \geq 0,$$ which is a contradiction. Thus $m \geq 0$, that is $u \geq 0$ in $\mathbb R^N$. \medskip Now, in order to show that $u > 0$ almost everywhere, we introduce the continuous function $U:\mathbb R^N\to\mathbb R$ defined by \[U(x)=\int_{B(0,r_0)}u(x-y)\,dy\] where $r_0$ is defined in item 2 of Assumption \ref{ass:kernel}, and we consider the closed set $[U = 0]$. If this set were not empty, then we may take $x_{0} \in \mathbb R^N$ such that $U(x_{0}) = 0$. Since $u \geq 0$, we would have $u(x)=0$ for almost very $x\in B(x_0,r_0)$ and accordingly, by using Tonelli's theorem, \begin{align*} 0 = \int_{B(x_0,r_0)}(W(x) + \lambda)u(x)\,dx &= \int_{B(x_0,r_0)}(f(x) + K*u(x))\,dx\\ & \geq \int_{\mathbb R^N} K(y)U(x_0•-y)\,dy \geq \int_{B(0,r_{0})} K(y)U(x_0-y)\,dy\geq0. \end{align*} Since $K>0$ a.e. on $B(0,r_0)$, we deduce that $U(x)=0$ for all $x\in B(x_0,r_0)$, that is whenever $x_{0} \in [U = 0]$ we have also $B(x_{0},r_{0}) \subset [U = 0]$. This means that the closed set $[U = 0]$ is also open, and $\mathbb R^N$ being a connected set, we infer that either the set $[U = 0] $ is empty, or it is all of $\mathbb R^N$. The latter would imply $u\equiv0$, which is ruled out since $f\not\equiv 0$. Thus $[U = 0] = \emptyset$, that is $U > 0$ on $\mathbb R^N$. Since $K>0$ a.e. on $B(0,r_0)$, this necessarily implies that $\int_{B(0,r_0)}K(y)u(x-y)\,dy>0$ for all $x\in\mathbb R^N$, and consequently $$u(x)\geq\frac{1}{\lambda+W(x)}\int_{B(0,r_0)}K(y)u(x-y)\,dy>0,$$ and the strong maximum principle is proved. \end{proof} \subsection{The eigenvalue problem}\label{sec:eigen} In this section we prove existence and uniqueness of a principal eigenfunction to \eqref{eq:eigenvalue}, that is we prove Theorem \ref{thm:eigen}. Before going further, recall the definitions of the function $K$ in \eqref{eq:Def-L}, the energy $\mathscr E(u)$ in \eqref{def:energy}, the eigenvalue candidate $\lambda_{1}$ in \eqref{def:lambda-un}, and that of the set $S$ in \eqref{def:S}. Observe that, for $u\in S$ we have $$ \int_{\mathbb R^N} (K*u)u\leq \left\Vert K*u\right\Vert_{L^2}\left\Vert u\right\Vert_{L^2}\leq \left\Vert K\right\Vert_{L^1}\left\Vert u\right\Vert_{L^2}\left\Vert u\right\Vert_{L^2}=\left\Vert K\right\Vert_{L^{1}}, $$ and thus $\lambda_{1} \geq -\left\Vert K\right\Vert_{L^{1}} = -\sigma^2$. We first take advantage of the condition \eqref{hyp:non-int} to prove the following. \begin{lemma}\label{lem:energy-neg} There is $\phi_{*} \in S$ such that $\mathscr E(\phi_{*}) < 0$. \end{lemma} \begin{proof} It is a direct consequence of Assumption~\ref{ass:linkJW}. Indeed, due to the monotone convergence theorem, we can assume w.l.o.g. that the set $B$ in Assumption~\ref{ass:linkJW} is essentially bounded. So for any $C>0$ the function $$ \phi_*(x) := C\frac{1}{W(x)}{\mathbf 1}_{B_{\epsilon}}(x) $$ belongs to $L^2(1+W)$, and we can choose $C$ such that $\int_{\mathbb R^N}\phi_*^2(x)dx = 1$, {\it i.e.} $\varphi_*\in S$. Then, recalling the definition of $b_\epsilon>0$ in~\eqref{eq:b-eps}, we have $ \mathscr E(\phi_*) = -b_\epsilon<0$. \end{proof} The following is a sort of compactness result, or rather a weak sequential continuity, concerning the quadratic mapping $$u \mapsto \int_{\mathbb R^N}(K*u)(x)u(x)dx.$$ \begin{lemma}\label{lem:convergence} If $(u_n)_{n\geq0}\subset S$ verifies \begin{equation}\label{cv-faible} u_{n} \rightharpoonup u \quad \text{in }\, L^2(1+W), \end{equation} then $$ \int_{\mathbb R^{N}}(K*u_{n})u_{n} \to \int_{\mathbb R^{N}} (K*u)u. $$ \end{lemma} \begin{proof} Since $(u_{n})_{n}$ is weakly convergent in $L^2(1+W)$, we can set $$M := \sup_{n \geq 1}\|u_{n}\|_{L^2(1+W)} < \infty.$$ Due to the Cauchy-Schwarz inequality we have \begin{equation} \label{est-pointwise} \vert (K* u_{n})(x)\vert\leq \Vert K\Vert_{L^2}\Vert u_{n}\Vert_{L^2}=\Vert K\Vert_{L^2}. \end{equation} Next note that since $(u_{n})_{n}$ converges weakly to $u$ in $L^2(1+W)$, we have also $u_{n} \rightharpoonup u$ in $L^{2}$. Now, for a given $x\in \mathbb R ^{N}$, we have $K(x - \cdot) \in L^2(\mathbb R^N)$ and therefore $$ (K* u_{n})(x) = \int_{\mathbb R^N} K(x-y) u_{n}(y)dy \to \int_{\mathbb R^N} K(x-y) u (y)dy = (K* u)(x). $$ From this and \eqref{est-pointwise}, using the Lebesgue dominated convergence theorem we deduce that \begin{equation} \label{L2loc} K* u_{n} \to K* u \quad \text{ in } L^{2}_{\rm loc}(\mathbb R^{N}). \end{equation} Now, let $\epsilon > 0$ be given. We may choose $R > 0$ large enough so that $(1+W(x))^{-1/2}\leq \epsilon$ when $|x| > R$. Then \begin{eqnarray*} \left\vert \int_{\vert x\vert >R} (K* u_{n}) u_{n}\right\vert&\leq &{\varepsilon} \int_{\vert x\vert >R}(K*\vert u_{n}\vert) (1+W)^{\frac 12}\vert u_{n}\vert\\ &\leq &{\varepsilon}\, \Vert K*\vert u_{n}\vert \,\Vert_{L^2}\;\Vert (1+W)^{\frac 12}\vert u_n\vert\, \Vert_{L^2}\\ &\leq &{\varepsilon}\, \Vert K\Vert_{L^{1}}\;\Vert u_{n} \Vert_{L^2}\;\Vert (1+W)^{\frac 12} u_{n} \Vert_{L^2}\\ &\leq & \epsilon M^2\|K\|_{L^1}, \end{eqnarray*} and it is clear that the same estimate holds for $\left\vert \int_{\vert x\vert >R} (K* u) u\right\vert$. As a result \begin{multline*} \left\vert \int_{\mathbb R^{N}} (K* u_{n}) u_{n}-\int_{\mathbb R^{N}} (K* u) u \right\vert\\ \leq \left\vert\int_{\vert x\vert \leq R} (K*( u_{n}- u)) u_{n}\right\vert+\left\vert\int_{\vert x\vert \leq R} (K* u)( u_{n}- u)\right\vert + 2\epsilon M^2\|K\|_{L^1}. \end{multline*} As $n\to+\infty$, the first and second terms in the right hand side tend to zero due to \eqref{L2loc} and \eqref{cv-faible} respectively. This concludes the proof of the lemma. \end{proof} We are now in a position to prove our main result concerning the eigenvalue problem~\eqref{eq:eigenvalue}, namely the existence of a unique (up to normalization) principal eigenfunction. \begin{proof}[Proof of Theorem \ref{thm:eigen}] We consider a sequence $\phi_{n} \in S$ such that $$ \lambda_{1}\leq \mathscr E(\phi_{n})\leq \lambda_{1} + \frac 1n. $$ Since $\mathscr E(|u|)\leq \mathscr E(u)$, up to replacing $\phi_{n}$ by $|\phi_{n}|$ we can assume $\phi_{n}\geq 0$. We have also \begin{eqnarray*} \int_{\mathbb R^N}W(x)\phi_{n}^2(x)dx &=& \displaystyle \mathscr E(\phi_{n}) + \int_{\mathbb R^N}(K*\phi_{n})(x)\phi_{n}(x)dx \\ & \leq &\lambda_{1} + 1 + \|K\|_{L^1}\|\phi_{n}\|_{L^2}^2 = \lambda_{1} + 1 + \sigma^2, \end{eqnarray*} so that $(\phi_{n})_{n}$ is bounded in $L^2(1+W)$, and thus there exists $\phi \in L^2(1+W)$ and a subsequence, denoted again by $(\phi_{n})_{n}$, such that \begin{equation}\label{cv-faible-bis} \phi_{n} \rightharpoonup \phi \quad\text{and}\quad \sqrt W\phi_{n} \rightharpoonup \sqrt W\phi\quad \text{ in } L^2. \end{equation} Next, using Lemma \ref{lem:convergence}, we have \begin{equation}\label{ineg} 0\leq \int_{\mathbb R^N} W(x)\phi_{n}^2(x)dx = \mathscr E(\phi_{n}) + \int_{\mathbb R^N} (K*\phi_{n})\phi_{n} \to \lambda_{1}+\int_{\mathbb R^N} (K*\phi)\phi. \end{equation} Since by Lemma \ref{lem:energy-neg} we have $\lambda_{1} < 0$, the above inequality implies that $\phi\not\equiv 0$ and $\phi \geq 0$. On the other hand, thanks to the weak convergences given in \eqref{cv-faible-bis} we have $$ \int_{\mathbb R^N} W(x)\phi^2(x)dx \leq \liminf_{n\to +\infty}\int_{\mathbb R^N}W(x)\phi_{n}^2(x) dx = \lambda_{1} + \int_{\mathbb R^N}(K*\phi)\phi , $$ and also $$\int_{\mathbb R^N}\phi^2(x)dx \leq \liminf_{n\to +\infty} \int_{\mathbb R^N}\phi_{n}^2(x)dx = 1.$$ Thus $\mathscr E(\phi)\leq \lambda_{1}$ and $\theta^2 := \int_{\mathbb R^N}\phi^2(x)dx \leq 1$. Since $\phi\not\equiv 0$, setting $ {\widetilde \phi} := \theta^{-1}\phi$ we have ${\widetilde \phi} \in S$ and $$ \lambda_{1} \leq \mathscr E(\widetilde \phi) = \theta^{-2}\mathscr E(\phi)\leq \theta^{-2}\lambda_{1} \leq \lambda_{1}, $$ where in the last inequality we use the fact that $\lambda_{1} < 0$. Clearly this implies that $\theta^2 = 1$ and thus ${\widetilde \phi} = \phi \in S$: this means that $\|\phi_{n}\|_{L^2} \to \|\phi\|_{L^2}$ while $\phi_{n} \rightharpoonup \phi$ in $L^2(\mathbb R^N)$, yielding that the convergence of $(\phi_{n})_{n}$ to $\phi$ is strong. The same above inequalities imply also that $\mathscr E(\phi) = \lambda_{1}$, while from \eqref{ineg} we infer that $$ \int_{\mathbb R^N}W(x)\phi_{n}^2(x)dx \to \int_{\mathbb R^N}W(x)\phi^2(x)dx, $$ that is $\|\phi_{n}\|_{L^2(1+W)} \to \|\phi\|_{L^2(1+W)}$, again yielding that $\phi_{n} \to \phi$ in $L^2(1+W)$. Finally we have $\phi \in S$ and $\mathscr E (\phi) = \lambda_{1}$. Since here Assumption \ref{ass:J-symmetric} is satisfied, $L$ is self-adjoint and therefore there exists a Lagrange multiplier $\lambda\in \mathbb R$ such that $$L\phi = -K*\phi+W(x)\phi = \lambda \phi \quad \text{ a.e. in }\, \mathbb R ^N,$$ and, obviously upon multiplying this equation by $\phi$, one sees that $\lambda = \lambda_{1}$. We thus have \begin{equation}\label{egalite} (W(x)-\lambda_{1})\phi(x) = (K*\phi)(x) \quad\text{that is}\quad \phi(x) = \frac{(K*\phi)(x)}{W(x) - \lambda_{1}}. \end{equation} Since $K\in L^2(\mathbb R^N)$ and $\phi \in L^2(\mathbb R^N)$, we know that $K*\phi \in C_{0}(\mathbb R^N)$. As a result, from \eqref{egalite}, the continuity of $W$ and $W(x)-\lambda_{1} \geq -\lambda_{1} > 0$, we also have $\phi\in C_{0}(\mathbb R^{N})$, and \eqref{eq-vp} holds. Now, in order to see that $\phi > 0$, recalling that $\phi \in C_{0}(\mathbb R^N) \cap L^2(\mathbb R^N)$ and $\phi \geq 0$ satisfies $$-K*\phi + (W(x) - \lambda_{1} + 1)\phi = \phi,$$ by the strong maximum principle, see Lemma \ref{lem:Max-principle}, we have $\phi > 0$. Also, using \eqref{est-pointwise} and \eqref{egalite}, we deduce the pointwise estimate \begin{equation*} 0 < \phi(x)\leq \frac{\Vert K\Vert_{L^2}}{W(x)-\lambda_{1}}, \quad \forall x \in \mathbb R^N. \end{equation*} Once we know that $\phi > 0$ on $\mathbb R^N$, we can show $\lambda_{1} > -\sigma^2$. Indeed, multiplying equality $$(-\lambda_{1} + W(x))\phi = K*\phi$$ by $\phi$ and integrating, since $\phi > 0$ and $W \not\equiv 0$ is nonnegative, we get $$-\lambda_{1} < \int_{\mathbb R^N}(-\lambda_{1} + W(x))\phi^2(x)dx = \int_{\mathbb R^N}(K*\phi)(x)\phi(x)dx \leq \|K\|_{L^1} = \sigma^2,$$ where we have used H\"older's inequality on the right hand side together with Young's inequality $\|K*\phi\|_{L^2} \leq \|K\|_{L^1}\|\phi\|_{L^2}$, and the fact that $\|\phi\|_{L^2} = 1$, while $\|K\|_{L^1} = \sigma^2$. It remains to prove the uniqueness of $\phi$, or in other terms the fact that the eigenspace corresponding to $\lambda_{1}$ has dimension one: that is if $\psi \in L^{2}(\mathbb R^{N})$ satisfies $$\psi\not\equiv 0, \qquad -K*\psi + W(x)\psi = \lambda_{1} \psi\qquad \text{in }\, \mathbb R^N, $$ then for a constant $\alpha \in \mathbb R$ we have $\alpha\psi = \phi$. Arguing as above, we conclude first that $\psi \in C_{0}(\mathbb R^{N})$ and, without loss of generality we may assume that there exists $x^{*} \in \mathbb R^N$ such that $\psi(x^{*}) > 0$, at the cost of replacing $\psi$ by $-\psi$, if necessary. Next, let $R > 0$ be large enough so that $W(x)-\lambda_{1} - \sigma^2 > 0$ for $|x| > R$, where we recall that $\sigma^2 = \|K\|_{L^1}$ (this is possible thanks to the fact that $W$ is confining). Since $\phi > 0$, we can choose $\epsilon > 0$ small enough so that $u_{\epsilon} := \phi - \epsilon \psi >0$ on $B(0,R)$. Let us now prove that \begin{equation} \label{eq:claim-partout} u_{\epsilon} \geq 0 \quad \text{ on the whole of }\, \mathbb R^{N}. \end{equation} If this were not true, then using the fact that $u_{\epsilon} \in C_{0}(\mathbb R^{N})$, we infer that $u_{\epsilon}$ achieves its global negative minimum at some $x_{0} \in \mathbb R^N$, and we necessarily have $|x_{0}| > R$. Since on the one hand $$K*u_{\epsilon}(x_{0}) - \sigma^2 u_{\varepsilon}(x_{0}) = \int_{\mathbb R^N} K(y)\left(u_{\epsilon}(x_{0} - y) - u_{\epsilon}(x_{0})\right) dy \geq 0,$$ and on the other hand, using the linear equations satisfied by $\phi$ and $\psi$, we have $$ K*u_{\epsilon}(x_{0}) - \sigma^2 u_{\epsilon}(x_{0}) = (W(x_{0}) - \lambda_{1} - \sigma^2) u_{\epsilon}(x_{0}) < 0, $$ we have a contradiction, which implies that \eqref{eq:claim-partout} holds. Now, since $\psi(x^{*}) > 0$, we point out that if $u_{\epsilon}(x) \geq 0$ in $\mathbb R^N$, in particular $u_{\epsilon}(x_{*}) \geq 0$ and thus $\epsilon \leq \phi(x_{*})/\psi(x_{*})$. Hence we can define the real number \begin{equation}\label{eq:Def-alpha-A} \alpha := \sup A, \quad \text{where}\quad A := \left\{\epsilon > 0: u_{\epsilon} := \phi - \epsilon \psi \geq 0 \,\text{ on }\, \mathbb R^{N} \right\} , \end{equation} and we know that $0 < \alpha \leq \phi(x_{*})/\psi(x_{*})$. In particular we infer that if we set $u_{\alpha} := \phi - \alpha \psi$ then $u_{\alpha} \geq 0$ and satisfies $$u_{\alpha} \in C_{0}(\mathbb R^N), \qquad -K*u_{\alpha} + (W(x) - \lambda_{1} + 1)u_{\alpha} = u_{\alpha} \geq 0.$$ However, if we had $u_{\alpha} \not\equiv 0$, thanks to Lemma \ref{lem:Max-principle} we would have $u_{\alpha} > 0$ in $\mathbb R^N$ and there would exist $\epsilon_{0} > 0$ small enough such that $u_{\alpha + \epsilon_{0}} := u_{\alpha} - \epsilon_{0} \psi > 0$ on the ball $B(0,R)$. Proceeding as in the proof of \eqref{eq:claim-partout}, we would deduce that $u_{\alpha + \epsilon_{0}} \geq 0$ in $\mathbb R^N$ and thus $\alpha + \epsilon_{0} \in A$, the set defined in \eqref{eq:Def-alpha-A}, contradicting the definition of $\alpha$. Therefore we must have $u_{\alpha} \equiv 0$, that is $\phi = \alpha \psi$. \end{proof} \subsection{A quantified spectral gap result in $L^2(\mathbb R^N)$} In this section we suppose that Assumption~\ref{ass:W-fort} is verified, and we consider a convex, open and bounded set $\Omega\subset\mathbb R^N$ that contains the origin. Then we have the following results. \begin{lemma}\label{lm:fct-ineq} For all $u\in D(L)=L^2(1+W)$ such that $\int_{\mathbb R^N}u=0$ and $u\not\equiv0$ we have \[\frac{\langle-Lu,u\rangle}{\|u\|_{L^2}^2}\leq \Phi\bigg(\frac{\int_{\Omega^c}Wu^2}{\|u\|_{L^2}^2}\bigg),\] where the function $\Phi$ is defined in~\eqref{def:Phi}. \end{lemma} \begin{proof} First, we clearly have, for all $u\in L^2(1+W)$, \begin{equation} \label{qqch0} \langle-Lu,u\rangle=\langle K*u,u\rangle-\langle Wu,u\rangle\leq\|K\|_{L^1}\|u\|_{L^2}^2-\int_{\Omega^c}Wu^2=\sigma^2 \|u\|_{L^2}^2-\int_{\Omega^c}Wu^2. \end{equation} The second part of the minimum defining the function $\Phi$ in \eqref{def:Phi} deserves more attention and is valid only under the condition $\int_{\mathbb R^N}u=0$. Due to the non-negativity of $W$ we have \begin{equation}\label{qqch} \langle-Lu,u\rangle\leq\langle K*u,u\rangle=\iint_{\Omega\times\Omega}K(x-y)u(x)u(y)\,dxdy+\iint_{(\Omega\times\Omega)^c}K(x-y)u(x)u(y)\,dxdy. \end{equation} We start by estimating the first term. Using that $K$ is symmetric we have \begin{align*} \iint_{\Omega\times\Omega}&K(x-y)u(x)u(y)\,dxdy=\iint_{\Omega\times\Omega}K(x-y)u^+(x)u^+(y)\,dxdy\\ &\qquad+\iint_{\Omega\times\Omega}K(x-y)u^-(x)u^-(y)\,dxdy-2\iint_{\Omega\times\Omega}K(x-y)u^+(x)u^-(y)\,dxdy\\ &\leq\big(\esssup_{2\Omega}K\big)\bigg(\Big(\int_\Omega u^+\Big)^2+\Big(\int_\Omega u^-\Big)^2\bigg)-2\big(\essinf_{2\Omega}K\big)\Big(\int_\Omega u^+\Big)\Big(\int_\Omega u^-\Big)\\ &\leq\big(\essinf_{2\Omega}K\big)\bigg(\int_\Omega u^+-\int_\Omega u^-\bigg)^2+\eta\bigg(\Big(\int_\Omega u^+\Big)^2+\Big(\int_\Omega u^-\Big)^2\bigg)\\ &\leq\big(\essinf_{2\Omega}K\big)\bigg(\int_\Omega u\bigg)^2+\eta\meas(\Omega)\,\|u\|_{L^2}^2\,. \end{align*} Since $\int_{\mathbb R^N}u=0$, using the Cauchy-Schwarz inequality, \[\bigg(\int_\Omega u\bigg)^2=\bigg(\int_{\Omega^c} u\bigg)^2\leq\bigg(\int_{\Omega^c}\frac1W\bigg)\bigg(\int_{\Omega^c}Wu^2\bigg).\] As a result \begin{equation} \label{qqch2} \iint_{\Omega\times\Omega}K(x-y)u(x)u(y)\,dxdy\leq \eta\meas(\Omega)\,\|u\|_{L^2}^2+(\essinf_{2\Omega}K\big)\bigg(\int_{\Omega^c}\frac1W\bigg)\bigg(\int_{\Omega^c}Wu^2\bigg). \end{equation} For the second term, using again the symmetry of $K$, we have \begin{align*} \iint_{(\Omega\times\Omega)^c}K(x-y)u(x)u(y)\,dxdy&=\iint_{\Omega^c\times\mathbb R^N}K(x-y)u(x)u(y)\,dxdy\\ &\qquad+\iint_{\Omega\times\Omega^c}K(x-y)u(x)u(y)\,dxdy\\ &\leq2\int_{\Omega^c}(K*u)(x)u(x)\,dx\\ &\leq2\sqrt{\int_{\Omega^c}\frac{|(K*u)(x)|^2}{W(x)}dx}\,\sqrt{\int_{\Omega^c}W(x)u^2(x)\,dx}\,. \end{align*} Since \[|(K*u)(x)|^2\leq\|K\|_{L^1}\int_{\mathbb R^N}K(x-y)u^2(y)\,dy,\] we get \begin{align} \iint_{(\Omega\times\Omega)^c}K(x-y)u(x)u(y)\,dxdy&\leq 2\sqrt{\|K\|_{L^1}\int_{\mathbb R^N}\Big(\int_{\Omega^c}\frac{K(x-y)}{W(x)}dx\Big)u^2(y)dy}\,\sqrt{\int_{\Omega^c}Wu^2}\nonumber \\ &\leq2\sigma\|u\|_{L^2}\,\sqrt{\sup_{y\in\mathbb R^N}\int_{\Omega^c}\frac{K(x-y)}{W(x)}dx}\,\sqrt{\int_{\Omega^c}Wu^2}.\label{qqch3} \end{align} In view of \eqref{def:Phi}, it now suffices to combine \eqref{qqch0}, \eqref{qqch}, \eqref{qqch2} and \eqref{qqch3} to prove the result. \end{proof} \begin{corollary}\label{cor:spectral-gap} If $\lambda\in\mathbb R$ is an eigenvalue of $L$ such that $\lambda\neq\lambda_1$, then $\lambda\geq -\bar\Phi:=-\sup_{[0,\infty)}\Phi$. \end{corollary} \begin{proof} Let $\lambda\in\mathbb R$ (recall that since $L$ is self-adjoint it has a real spectrum) and $\psi\in D(L)$ such that $L\psi=\lambda\psi$ with $\lambda\neq\lambda_1$. Then necessarily $\langle\psi,\varphi\rangle=0$, where $\varphi>0$ is the principal eigenfunction, so that $\psi$ cannot be of constant sign. If $\int_{\mathbb R^N}\psi=0$ then Lemma~\ref{lm:fct-ineq} applied to $u=\psi$ immediately ensures that $-\lambda\leq\bar\Phi$. If $\int_{\mathbb R^N}\psi\neq0$, there exists $\alpha\in\mathbb R$ such that $\tilde \psi:=\psi+\alpha\varphi$ verifies $\int_{\mathbb R^N}\tilde\psi=0$ and Lemma~\ref{lm:fct-ineq} applied to $u=\tilde\psi$ yields, using that $\lambda\geq\lambda_1$, $\|\varphi\|_{L^2}=1$ and $\langle\psi,\varphi\rangle=0$, \[-\lambda(\|\psi\|_{L^2}^2+\alpha^2)\leq-\lambda\|\psi\|_{L^2}^2-\lambda_1\alpha^2=\langle-L\tilde\psi,\tilde\psi\rangle\leq\bar\Phi\|\tilde\psi\|_{L^2}^2=\bar\Phi(\|\psi\|_{L^2}^2+\alpha^2),\] which concludes the proof. \end{proof} This result provides a quantified estimate of the distance between $\lambda_1$ and the other eigenvalues of $L$ provided that an upper bound smaller than $-\bar\Phi$ is known for $\lambda_1$. \section{Long time asymptotics of the linear problem} \label{sec:asympto} This section is devoted to the linear evolution equation~\eqref{eq:parabolique}. Recalling that the eigenpair $(\lambda_{*},\phi)$ satisfies \eqref{eq:eigenvalue}, with $\lambda_{*} := \lambda_{1} + \sigma^2 > 0$, defined in \eqref{eq:Def-lambda-*}, and $K := \sigma^2 J$, we readily observe that the solutions $u=u(t,x)$ to \eqref{eq:parabolique} are related to the solutions $v=v(t,x)$ of the abstract Cauchy problem \begin{equation}\label{eq:evolution-A} \left\{\begin{array}{l} \displaystyle\frac {dv(t)}{dt} = \mathcal A v(t) \qquad\text{for }\, t > 0, \vspace{2mm}\\ v(0) = u_{0}, \end{array}\right. \end{equation} where the operator $(\mathcal A,D(\mathcal A))$ and the function $v(t)$ are defined by ($L$ being as in \eqref{eq:Def-L}): \begin{equation}\label{eq:Def-A-u-v} D(\mathcal A) := D(L), \qquad \mathcal A v := -Lv + \lambda_{1}v, \qquad\text{and}\quad u(t,\cdot) = {\rm e}^{-\lambda_{*}t} v(t,\cdot). \end{equation} Recall that a function $v(t)$ is called a {\it classical solution} of Equation~\eqref{eq:evolution-A} if it lies in $D(A)$, is continuously differentiable, and~\eqref{eq:evolution-A} holds. It is called a {\it mild solution} if $\int_0^tv(s)ds\in D(A)$ for all $t\geq0$ and \[v(t)=u_0+A\int_0^tv(s)\,ds.\] From Theorem \ref{thm:eigen}, we know that $\phi$ is the unique positive steady state (up to normalization) of Equation~\eqref{eq:evolution-A} in $C_{0}(\mathbb R^{N})$ and in $L^2(\mathbb R^{N})$, and thus proving Theorem \ref{thm:asympto} is tantamount to showing that positive solutions of \eqref{eq:evolution-A} converge to (a multiple of) this stationary solution. That is why in this section we shall work with the modified equation \eqref{eq:evolution-A}. \medskip To analyse the long time behaviour of Equation~\eqref{eq:evolution-A}, we take advantage of the theory of strongly continuous semigroups, also called $C_{0}$-semigroups, of positive linear operators. There is a large literature on this field, but the standard references K.~Yosida~\cite{YosidaK}, W.~Arendt \& al.~\cite{Nagel86} and K.J.~Engel \& R.~Nagel~\cite{EN} will be enough here. Recall that we study Equation~\eqref{eq:evolution-A} in one of the following Banach lattices: $E = L^p(\mathbb R^N)$ with $1\leq p<\infty$, or $E = C_{0}(\mathbb R^N)$, or $E = \MM(\mathbb R^N) = (C_{0}(\mathbb R^N))'$, equipped with the norm $\Vert \cdot \Vert_E$ denoting the $L^p$ norm, or the $L^\infty$ norm, or the total variation norm, respectively. \subsection{Analysis in the space $L^2(\mathbb R^N)$} To begin with, let us study Equation~\eqref{eq:evolution-A} in the Lebesgue space $L^2(\mathbb R^N)$. Since $(L,D(L))$ is a self-adjoint operator acting in $L^2(\mathbb R^N)$, and since for $v \in D(L)$ we have $\langle Lv,v\rangle \geq \lambda_{1}\|v\|^2$, by the very definition of $\mathcal A$ by \eqref{eq:Def-A-u-v} we conclude that $(\mathcal A,D(\mathcal A))$ is self-adjoint and $\langle \mathcal A v,v\rangle \leq 0$, that is $\mathcal A$ is an $m$-dissipative operator. Also since $A\phi = 0$, this means that zero is the principal eigenvalue of the operator $(\mathcal A,D(\mathcal A))$, and that its spectrum $\sigma(A)$ is contained in $(-\infty,0]$. Thus by the Hille-Yosida theorem (see for instance K.~Yosida \cite[Chapter IX, Section 8]{YosidaK}) $\mathcal A$ generates a $C_{0}$-semigroup of contractions which we shall denote by $(T_{t})_{t\geq 0}$, or sometimes by $T_{t} = \exp(t\mathcal A) = {\rm e}^{t\mathcal A}$. Moreover, since $(\mathcal A,D(\mathcal A))$ is self-adjoint, $T_{t}$ is also self-adjoint on $L^2(\mathbb R^N)$ and the semigroup $(T_{t})_{t \geq 0}$ is analytic, that is for any $u_{0} \in L^2(\mathbb R^N)$ we have $T_{t}u_{0} \in D(\mathcal A)$ for $t > 0$. In particular for any $u_{0} \in L^2(\mathbb R^N)$ the function $v(t) := T_{t}u_{0}$ is the unique solution of equation \eqref{eq:evolution-A} in the classical sense on the interval $(0,\infty)$. Note that since $(\mathcal A,D(\mathcal A))$ is $m$-dissipative, we have $\|T_{t}\|_{L^2(\mathbb R^N) \to L^2(\mathbb R^N)} \leq 1$, but since $A\phi = 0$ we have $T_{t}\phi = \phi$ for all $t \geq 0$ and thus $$\|T_{t}\|_{L^2(\mathbb R^N) \to L^2(\mathbb R^N)} = 1.$$ This implies that the growth bound of the semigroup $(T_{t})_{t\geq0}$, that is the real number $$\omega_{0}(\mathcal A) := \inf\left\{w \in\mathbb R: \exists M >0,\; \forall t \geq0,\; \|T_{t}\|_{L^2(\mathbb R^N) \to L^2(\mathbb R^N)}\leq M{\rm e}^{w t}\right\},$$ is equal to zero. Besides, since $(T_{t})_{t\geq0}$ is analytic, the spectral bound of the operator $A$, that is $$\mathrm s(\mathcal A) := \sup\{\Re\lambda : \lambda\in\sigma(\mathcal A)\}$$ is equal to the growth bound of the semigroup $(T_{t})_{t\geq0}$ generated by $A$ (see for instance K.J.~Engel \& R. Nagel~\cite[Corollary IV.3.12]{EN}). We conclude that \begin{equation*}\label{eq:s-equal-omega0} \mathrm s(\mathcal A) = \omega_{0}(\mathcal A) = 0. \end{equation*} As it is customary in the study of large time behaviour of solutions to linear evolution equations, we wish to show that there is a {\it gap} in the spectrum of $\mathcal A$, in the sense that there exists a number $a > 0$ such that $$\sigma(\mathcal A) \setminus \{0\} \subset (-\infty,-a).$$ Once this is shown, then it is not difficult to see that, if $v(t)$ is the solution of \eqref{eq:evolution-A} its orthogonal projection on the space $(\mathbb R\phi)^\perp$ converges to zero at least as fast as ${\rm e}^{-at}$. Indeed the restriction $A_{|(\mathbb R\varphi)^\perp}$ of $A$ to the invariant subspace $(\mathbb R\phi)^\perp$ verifies in this case $$\omega_0(A_{|(\mathbb R\varphi)^\perp})=\mathrm s(A_{|(\mathbb R\varphi)^\perp})<-a.$$ For proving the existence of a spectral gap, we use the notion of {\it essential growth bound}, which is defined similarly as the growth bound. First, we define the essential norm of a bounded linear operator $T$ in a Banach space $E$ by \[{\|T\|}_\mathrm{ess}:=\inf\big\{{\|T-\mathcal K\|}_{E\to E}\,:\, \mathcal K:E\longrightarrow E\ \text{is compact}\big\}.\] Then we define the essential growth bound of a semigroup $(T_t)_{t\geq0}$ in $E$ by $$\omega_\mathrm{ess}(\mathcal A) := \inf\left\{w \in\mathbb R: \exists M >0,\; \forall t \geq0,\; {\|T_{t}\|}_\mathrm{ess}\leq M{\rm e}^{w t}\right\}.$$ Clearly, $\omega_\mathrm{ess}(A)\leq\omega_0(A)$, and a semigroup $(T_t)_{t\geq0}$ is said to be quasi-compact if $\omega_\mathrm{ess}(A)<0$. The usefulness of the essential growth bound lies in the following result (see for instance K.J.~Engel \& R. Nagel~\cite[Corollary IV.2.11]{EN}): \begin{center}\begin{minipage}{.9\linewidth} For every $w>\omega_\mathrm{ess}(A)$ the set $\sigma(A)\cap\{\lambda\in\mathbb C: \Re\lambda\geq w\}$ is composed of a finite number of eigenvalues with finite algebraic multiplicity. \end{minipage}\end{center} As a consequence, for our self-adjoint semigroup in $L^2(\mathbb R^N)$, if we can prove that $\omega_\mathrm{ess}(A)<0$, that is $(T_t)_{t\geq0}$ is quasi-compact, then we immediately get the existence of $a\in(0,-\omega_\mathrm{ess}(A))$ such that $\sigma(A)\setminus\{0\}\subset(-\infty,-a)$. \medskip In order to prove that $\omega_\mathrm{ess}(A)<0$, we split the operator $\mathcal A$ defined in~\eqref{eq:Def-A-u-v} as the sum of a local unbounded operator, namely $$\mathcal A_{0} u:=\lambda_{1} u - W(x) u,\qquad D(\mathcal A_{0}) = D(L) = \{u\in E: (1+W) u \in E\},$$ and a nonlocal bounded perturbation, given by $$\BB u := K*u = \sigma^2 J*u,$$ where we have $\|\BB\|_{L^2(\mathbb R^N) \to L^2(\mathbb R^N)}\leq\sigma^2$. It is straightforward to see that the operator $(\mathcal A_{0},D(\mathcal A_{0}))$ generates a $C_{0}$-semigroup of contractions which we shall denote by $(S_{t})_{t \geq 0}$, and as a matter of fact it can be written explicitly, not only in the space $L^2(\mathbb R^N)$ but in any of the spaces $E$ defined above. \begin{lemma}\label{lem:St-analytique} The unbounded operator $\big(\mathcal A_{0},D(\mathcal A_{0})\big)$ generates a positive $C_{0}$-semigroup $(S_{t})_{t\geq0}$ in $E$, explicitly given by $$(S_{t} u_{0})(x)={\rm e}^{(\lambda_{1}-W(x))t}u_{0}(x).$$ For any $u_{0} \in E$ and $t > 0$ we have $(1 + W)S_{t}u_{0} \in E$, that is $S_{t}u_{0} \in D(\mathcal A_{0})$. In particular $(S_{t})_{t \geq 0}$ is an analytic semigroup on $E$ and $\|S_{t}\|_{L^2(\mathbb R^N) \to L^2(\mathbb R^N)} \leq {\rm e}^{\lambda_{1} t}$. \end{lemma} On the other hand, since $\BB$ is a bounded operator, we readily deduce the following expression of the semigroup $T_{t}$ in terms of the semigroup $S_{t}$ (see for instance K.J.~Engel \& R.~Nagel~\cite[Chapter III.1]{EN}). Indeed, noting that \eqref{eq:evolution-A} can be written as $$\frac{dv}{dt} = \mathcal A_{0}v(t) + Bv(t), \qquad v(0) = u_{0},$$ the solution $v(t)$ is given by the Duhamel formula $$T_{t}u_{0} = v(t) = S_{t}u_{0} + \int_{0}^t S_{t - \tau}Bv(\tau)d\tau = S_{t}u_{0} + \int_{0}^t S_{t - \tau}BT_{\tau}u_{0}d\tau .$$ Analogously, the solution of $$\frac{dz}{dt} = \mathcal A_{0}z(t) = Az(t) - Bz(t), \qquad z(0) = u_{0},$$ is given by $$S_{t}u_{0} = z(t) = T_{t}u_{0} - \int_{0}^t T_{t - \tau}Bz(\tau) d\tau = T_{t}u_{0} - \int_{0}^t T_{t - \tau}B S_{\tau}u_{0} d\tau, $$ so that we have also $$T_{t}u_{0} = S_{t}u_{0} + \int_{0}^t T_{t - \tau}B S_{\tau}u_{0} d\tau. $$ We can thus state the following. \begin{proposition}\label{prop:wellposedness} The unbounded operator $\big(\mathcal A,D(\mathcal A)\big)$ generates a positive $C_{0}$-semigroup $(T_{t})_{t\geq0}$ in $E$, which yields the solutions of equation~\eqref{eq:evolution-A}. For any $u_{0}\in D(\mathcal A)$ the mapping $t\mapsto T_{t}u_{0} =: v(t)$ is the unique classical solution of~\eqref{eq:evolution-A} and for all $u_0\in E$ it is the unique mild solution. Moreover the Duhamel formulas \begin{equation}\label{eq:Duhamel} v(t) = T_{t}u_{0} = S_{t}u_{0} + \int_{0}^t S_{t-\tau}\big(K*T_{\tau}u_{0}\big)\,d\tau, \end{equation} and \begin{equation}\label{eq:Duhamel-bis} v(t) = T_{t}u_{0} = S_{t}u_{0} + \int_{0}^t T_{t-\tau}\big(K*S_{\tau}u_{0}\big)\,d\tau, \end{equation} hold for every $t \geq 0$ and $u_{0} \in E$. \end{proposition} It is noteworthy to observe that Proposition~\ref{prop:wellposedness} is still valid by replacing the choice of one of the above defined Banach spaces $E$ by the intersection $E_1\cap E_2$ of two such Banach spaces, endowed with the norm $\|\cdot\|_{E_1}+\|\cdot\|_{E_2}$. The uniqueness property in this intersection then guarantees that, if $u_0\in E_1\cap E_2$, then the solutions in $E_1$ and $E_2$ coincide for all time. \medskip Now we return to the study of the spectral gap for the operator $\mathcal A$ in $L^2(\mathbb R^N)$, and we use the Duhamel formula to prove that the semigroup is quasi-compact. \begin{lemma}\label{rouge} The semigroup $(T_{t})_{t\geq0}$ is quasi-compact in $L^p(\mathbb R ^N)$ for $1 \leq p < \infty$, and more precisely $\omega_\mathrm{ess}(A)\leq\lambda_1<0$. \end{lemma} \begin{proof} For a given $u_{0} \in L^p(\mathbb R^N)$, by the Duhamel formula \eqref{eq:Duhamel-bis} we have $$T_{t}u_{0} = S_{t} u_{0} + \int_{0}^t T_{t - \tau}\big(K*S_{\tau}u_{0}\big)\,d\tau .$$ We have $\|S_{t}\|_{L^p(\mathbb R^N)\to L^p(\mathbb R^N)}\leq {\rm e}^{\lambda_{1} t}<1$ for any $t>0$, hence setting $$R_{t}u_{0} := \int_{0}^t T_{t - \tau}\big(K*S_{\tau}u_{0}\big)\,d\tau $$ it suffices to prove that the operator $R_{t}$ is compact for all $t$ large enough. As a matter of fact, it turns out that $R_{t}$ is compact for any $t > 0$. To see this, we are going to use the Riesz-Fréchet-Kolmogorov theorem characterizing compact subsets of $L^p(\mathbb R^N)$ (see for instance K.~Yosida~\cite[Chapter X, section 1]{YosidaK}). First we check that, for any $\tau > 0$, the operator $$u_{0} \mapsto K*S_{\tau}u_{0}$$ is compact on $L^p(\mathbb R^N)$. Let $u_{0}\in L^p(\mathbb R^N)$ with $\|u_{0}\|_{L^p}\leq 1$. Observe first that $$\|K*S_{\tau}u_{0}\|_{L^p}\leq\|K\|_{L^1}\|S_{\tau}u_{0}\|_{L^p}\leq\sigma^2,$$ and thus the image of the unit ball of $L^p(\mathbb R^N)$ is bounded. Now, for $h \in \mathbb R^N$ define the translation operator $\tau_{h}$ by setting $\tau_{h}f = f(\cdot+h)$ for $f \in L^p(\mathbb R^N)$. We have $$\|\tau_{h}(K*S_{\tau}u_{0}) - K*S_{\tau}u_{0}\|_{L^p} = \|(\tau_{h} K - K)*S_{\tau}u_0\|_{L^p}\leq\|\tau_{h}K - K\|_{L^1}\xrightarrow[|h|\to0]{}0,$$ uniformly in $u_{0}$ in the unit ball of $L^p(\mathbb R^N)$. Next, by H\"older's inequality we have ($\frac 1p+\frac 1{p'}=1$) \begin{align*} |(K*S_{\tau}u_{0})(x)| & = \int_{\mathbb R^N} K(x-y)^{1/p'} K(x-y)^{1/p}{\rm e}^{(\lambda_{1} - W(y))\tau}u_{0}(y)\,dy\\ &\leq \left(\int_{\mathbb R^N} K(x-y) dy\right) ^{1/p'}\left(\int_{\mathbb R^N} K(x-y) {\rm e}^{p(\lambda_{1} - W(y))\tau}|u_{0}(y)|^p\,dy\right)^{1/p}\\ & \leq \sigma^{2/p'} \left(\int_{\mathbb R^N} K(x-y) \,{\rm e}^{p(\lambda_{1} - W(y))\tau}|u_{0}(y)|^p\,dy\right)^{1/p}. \end{align*} Hence by the Fubini-Tonelli theorem we may write (noting that $|x - y| \geq R/2$ when $|x| \geq R$ and $|y|\leq R/2$) \begin{align*} \int_{|x|\geq R}|(K*S_{\tau}u_{0})(x)|^pdx & \leq \sigma^{2p/p'} \int_{|x|\geq R}\int_{\mathbb R^N} K(x-y)\, {\rm e}^{p(\lambda_{1} - W(y))\tau}\vert u_{0}(y)\vert ^p\,dydx \\ &\leq \sigma^{2p/p'} \int_{|x|\geq R} \int_{|y| < R/2} K(x-y)\, \vert u_{0}(y)\vert ^p \,dydx\\ & \hskip8mm + \sigma^{2p/p'} \int_{|x|\geq R} \int_{|y|\geq R/2}K(x-y)\, {\rm e}^{-p\tau W(y)} \vert u_{0}(y)\vert ^p \,dydx\\ &\leq \sigma^{2p/p'} \int_{|z|\geq R/2} K(z)\,dz + \sigma^{2p}\sup_{|y|\geq R/2}{\rm e}^{-p\tau W(y)}\xrightarrow[R\to+\infty]{}0, \end{align*} uniformly in $u_{0}$ in the unit ball of $L^p(\mathbb R^N)$. Using the Riesz-Fréchet-Kolmogorov theorem we conclude that the mapping $u_0\mapsto K*S_{\tau}u_{0}$ is compact. Finally, since $T_{t-\tau}$ is a bounded operator, we infer that for any $0 < \epsilon \leq \tau \leq t$, the operators $$u_{0}\mapsto T_{t-\tau}(K*S_{\tau}u_{0}) \quad\text{and}\quad u_{0} \mapsto \int_{\epsilon}^t T_{t -\tau}(K*S_\tau u_{0})d\tau,$$ are compact operators on $L^p(\mathbb R^N)$. Since, as $\epsilon \to 0$ we have $$\int_{\epsilon}^t T_{t -\tau}(K*S_\tau u_{0})d\tau \to \int_{0}^t T_{t -\tau}(K*S_\tau u_{0})d\tau = R_{t}u_{0},$$ uniformly on the unit ball of $L^p(\mathbb R^N)$, we conclude that $R_{t}$ is compact. \end{proof} Now we can state our convergence result. \begin{corollary} \label{etleclochard} There exist $C,a>0$ such that, for all $u_{0} \in L^2(\mathbb R^N)$ and all $t \geq 0$, we have \begin{equation}\label{eq:conv-L2} \left\|T_{t}u_{0} - \langle u_{0},\phi\rangle \phi \right\|_{L^2}\leq C\, {\rm e}^{-at}\, \left\|u_{0}-\langle u_{0},\phi\rangle\phi\right\|_{L^2}. \end{equation} If additionally $b_\epsilon>\bar\Phi:=\sup_{[0,\infty)}\Phi$, where $b_\epsilon$ and $\Phi$ are defined in~\eqref{eq:b-eps} and~\eqref{def:Phi} respectively, then one can choose any $a<a_*=:b_\epsilon-\bar\Phi$. \end{corollary} \begin{proof} We proved that $\omega_\mathrm{ess}(A)\leq\lambda_1<0$, and we have the identity $\omega_\mathrm{ess}(A_{|(\mathbb R\varphi)^\perp})=\omega_\mathrm{ess}(A)$ (use for instance~\cite[Proposition IV.2.12]{EN}). We deduce that for any $w\in(\lambda_1,0)$ the set $\sigma(A_{|(\mathbb R\varphi)^\perp})\cap[w,0]$ is finite and made only of eigenvalues. Since the kernel of $A$ is the space generated by $\varphi$, zero is not an eigenvalue of $A_{|(\mathbb R\varphi)^\perp}$ and consequently there exists $a>0$ such that $\sigma(A_{|(\mathbb R\varphi)^\perp})\subset(-\infty,-a)$. This implies that \[\omega_0(A_{|(\mathbb R\varphi)^\perp})=\mathrm s(A_{|(\mathbb R\varphi)^\perp})<-a\] and accordingly the existence of $C>0$ such that for all $u_0\in(\mathbb R\varphi)^\perp$ \[\left\|T_{t}u_{0}\right\|_{L^2}\leq C\, {\rm e}^{-at}\, \left\|u_{0}\right\|_{L^2}.\] For $u_0\not\in(\mathbb R\varphi)^\perp$, applying this stability result to $u_{0}-\langle u_{0},\phi\rangle\phi\in(\mathbb R\varphi)^\perp$ gives~\eqref{eq:conv-L2} since $T_t\varphi=\varphi$ for all $t\geq0$. For proving the second part of Corollary~\ref{etleclochard}, it suffices to check that, if $b_\epsilon>\bar\Phi$, then there is no eigenvalue of $A$ in the interval $(\bar\Phi-b_\epsilon,0)$. Corollary~\ref{cor:spectral-gap} ensures that there is no non-zero eigenvalue above $\bar\Phi+\lambda_1$, and from Theorem~\ref{thm:eigen} we know that $\lambda_1\leq -b_\epsilon$. So the result is proved. \end{proof} \subsection{Analysis in $C_{0}(\mathbb R^N)$ and $\MM(\mathbb R^N)$} We start by checking that $(T_{t})_{t\geq0}$ is quasi-compact in $C_{0}(\mathbb R^N).$ \begin{lemma} The semigroup $(T_{t})_{t\geq0}$ is quasi-compact in $C_{0}(\mathbb R^N).$ \end{lemma} \begin{proof} Using the Duhamel formula~\eqref{eq:Duhamel-bis} in a similar way as we did in Lemma~\ref{rouge}, we only have to prove that $u_{0} \mapsto K*S_{\tau}u_{0}$ is compact for any $\tau > 0$. This property is a consequence of Ascoli's theorem: indeed we have, for any $u_{0} \in C_{0}(\mathbb R^N)$ with $\|u_0\|_{L^\infty} \leq 1$, $$\|\tau_{h}(K*S_{\tau}u_0)-K*S_{\tau}u_0\|_{L^\infty} \leq \|\tau_{h} K - K\|_{L^1}\xrightarrow[|h|\to0]{}0,$$ and $$|K*S_{\tau}u_0|\leq K*{\rm e}^{-\tau W}\in C_{0}(\mathbb R^N),$$ uniformly in $u_{0}$ with $\|u_{0}\|_{L^\infty} \leq 1$. The proof of the lemma is complete. \end{proof} Contrary to the $L^2$ case, we cannot argue through the orthogonal space of $\mathbb R\varphi$. Yet, the positivity of $\varphi$ combined to the quasi-compactness of $(T_t)_{t\geq0}$ is enough to prove the following results. \begin{corollary}\label{villoise} The eigenfunction $\phi$ belongs to $L^1(\mathbb R^N),$ and there exist $C,a>0$ such that, for all $u_{0}\in C_{0}(\mathbb R^N)$ and all $t \geq 0$, $$\left\|T_{t}u_{0}-\langle u_{0},\phi\rangle\phi\right\|_{L^\infty}\leq C\, {\rm e}^{-at}\left\|u_{0}-\langle u_{0},\phi\rangle\phi\right\|_{L^\infty}.$$ \end{corollary} \begin{proof} We have proved that $(T_{t})_{t\geq0}$ is quasi-compact and we know that $0$ is an eigenvalue of $\mathcal A$ associated to a strictly positive eigenfunction $\phi$. We deduce from Corollary B-IV-2.11 in~\cite{Nagel86} that there exists a positive projection $\mathbb P$ of finite rank and suitable constants $C,a>0$ such that, for all $t \geq 0$, $$\|T_{t} - \mathbb P\|_{C_{0}(\mathbb R^N) \to C_{0}(\mathbb R^N)}\leq C \, {\rm e}^{-at}.$$ Let us now identify this projection. From Corollary~\ref{etleclochard} we deduce that for all $u_0\in C_c(\mathbb R^N)$, $\mathbb P u_0=\langle u_0,\varphi\rangle\varphi$. Since $\mathbb P$ is a projection, this implies that for all $u_0\in C_c(\mathbb R^N)$, $|\langle u_0,\varphi\rangle|\leq{\|u_0\|}_{L^\infty}/{\|\varphi\|}_{L^\infty}$. Consequently $\varphi$ belongs to $L^1(\mathbb R^N)$, and the bounded operator $u_0\mapsto \langle u_0,\varphi\rangle\varphi$ on $C_0(\mathbb R^N)$ coincides with $\mathbb P$ on the dense subset $C_c(\mathbb R^N)$. So they are necessarily equal on $C_0(\mathbb R^N)$ and the proof is complete. \end{proof} \begin{corollary}\label{epoque} There exist $C,a>0$ such that, for all $\mu\in \MM(\mathbb R^N)$ and all $t\geq0$, $$\left\|T_{t}\mu - \langle\mu,\phi\rangle\phi\right\|_\mathrm{TV}\leq C{\rm e}^{-at}\left\|\mu-\langle \mu,\phi\rangle\phi\right\|_\mathrm{TV}.$$ \end{corollary} Notice that this implies also the exponential convergence in $L^1(\mathbb R^N)$, since for $u_{0} \in L^1(\mathbb R^N)$ and $\mu := u_{0}(x)dx$ we have ${\|\mu\|}_\mathrm{TV} = {\|u_0\|}_{L^1}$. \begin{proof} Due to the duality $\MM(\mathbb R^N) = (C_{0}(\mathbb R^N))'$ and the definition of the total variation norm as a duality norm $$\left\|\mu\right\|_\mathrm{TV}=\sup_{f\in C_{0},\,{\|f\|}_{L^\infty}\leq1}\langle \mu,f\rangle,$$ the result is a consequence of Corollary~\ref{villoise} applied to the dual semigroup ${(T_{t}^*)}_{t\geq0} = {(T_{t})}_{t \geq 0}$. \end{proof} \subsection{Study in $L^p(\mathbb R^N)$ with $1\leq p<\infty$} We have proved that $\phi\in L^1(\mathbb R^N)\cap C_{0}(\mathbb R^N),$ so that $\phi\in L^p(\mathbb R^N)$ for all $p\in[1,\infty].$ Also recall that in Lemma~\ref{rouge} we have shown that $(T_{t})_{t\geq0}$ is quasi-compact in $L^p(\mathbb R^N)$ for any $p\in[1,\infty)$. \begin{corollary} Let $p\in[1,\infty)$. There exist $C,a>0$ such that, for all $u_{0}\in L^p(\mathbb R^N)$ and all $t\geq0$, $$\left\|T_{t}u_{0} - \langle u_{0},\phi\rangle\phi\right\|_{L^p}\leq C{\rm e}^{-at}\left\|u_{0}-\langle u_{0},\phi\rangle\phi\right\|_{L^p}.$$ \end{corollary} \begin{proof} For $u_{0} \in L^p(\mathbb R^N)$ denote by $\mathbb P u_{0} := \langle u_{0},\phi \rangle\phi$. We know that there exist two constants $C > 0$ and $a > 0$ such that $$\|T_{t} - \mathbb P\|_{L^1(\mathbb R^N) \to L^1(\mathbb R^N)} \leq C {\rm e}^{-at} \quad\text{and} \quad \|T_{t} - \mathbb P\|_{L^2(\mathbb R^N) \to L^2(\mathbb R^N)} \leq C {\rm e}^{-at}. $$ Therefore, for $1 < p < 2$, by interpolation (see for instance L.~Tartar~\cite[Chapter 21, Theorem 21.2]{TartarL-Sobolev-Interpolation}) we have $$\|T_{t} - \mathbb P\|_{L^p(\mathbb R^N) \to L^p(\mathbb R^N)} \leq \|T_{t} - \mathbb P\|_{L^1(\mathbb R^N) \to L^1(\mathbb R^N)}^{1 - \theta} \|T_{t} - \mathbb P\|_{L^2(\mathbb R^N) \to L^2(\mathbb R^N)}^\theta,$$ where $\theta \in (0,1)$ is defined by $1/p = (1 - \theta) + (\theta/2) = 1 - (\theta/2)$. Thus we have $$\|T_{t} - \mathbb P\|_{L^p(\mathbb R^N) \to L^p(\mathbb R^N)} \leq C\, {\rm e}^{-at}.$$ When $2 < p < \infty$, then $p' := p/(p - 1) \in (1,2)$, and since $T_{t}^* = T_{t}$ and $\mathbb P^* = \mathbb P$ on the subspace $L^1(\mathbb R^N) \cap C_{0}(\mathbb R^N)$, which is dense both in $L^p(\mathbb R^N)$ and $L^{p'}(\mathbb R^N)$, the above inequality applied to $p'$ shows that $$\|T_{t} - \mathbb P\|_{L^p(\mathbb R^N) \to L^p(\mathbb R^N)} = \|T_{t}^* - \mathbb P^*\|_{L^{p'}(\mathbb R^N) \to L^{p'}(\mathbb R^N)}\leq C\, {\rm e}^{-at},$$ which concludes the proof. \end{proof} \section{Long time asymptotics of the replicator-mutator model}\label{sec:rep-mut} We begin by giving the definition of what we shall call classical and mild solutions for the nonlinear replicator-mutator equation \eqref{eq:rep-mut}. Let us denote by $E_{+}$ the positive cone of the Banach lattice $E$ and define $$E(W):=\{u\in E: Wu\in E\}$$ endowed with the norm $$\left\|u\right\|_{E(W)}:={\|u\|}_E+{\|Wu\|}_E.$$ Notice that in the framework of Section~\ref{sec:asympto}, this space is nothing but the domain of the operator $\mathcal A_{0}$ endowed with the graph norm. \begin{definition} A function $u:[0,+\infty)\longrightarrow E_{+}$ is called a \emph{classical solution} of Equation~\eqref{eq:rep-mut} if $u\in C^1([0,+\infty),E)\cap C([0,+\infty),E(W))$ and~\eqref{eq:rep-mut} holds. It is called a \emph{mild solution} of Equation~\eqref{eq:rep-mut} if $u\in C([0,\tau],E)\cap L^1([0,\tau],E(W))$ for all $\tau>0$, and $$u(t)={\rm e}^{-\lambda_{*} t}T_{t}u_{0} + \int_{0}^t\langle u(s),W\rangle\, {\rm e}^{-\lambda_{*} (t-s)}T_{t-s}u(s)\,ds,$$ for all $t\geq0$ (Recall that the eigenvalue $\lambda_{*}$ is defined in \eqref{eq:Def-lambda-*}). \end{definition} \begin{proposition}[The solution of \eqref{eq:rep-mut} in terms of that of \eqref{eq:parabolique}]\label{prop:wellposed_rep-mut} Suppose that Assumptions~\ref{ass:kernel} and~\ref{ass:potential} are verified, and let $u_{0}\in E_{+}$ with $\langle u_{0},\mathbf 1\rangle=1.$ There exists a unique mild solution to Equation~\eqref{eq:rep-mut} starting from $u_{0}$, and it is given by $$u(t)=\frac{T_{t}u_{0}}{\langle T_{t}u_{0},\mathbf 1\rangle}.$$ If additionally $u_{0}\in E(W)$ then it is a classical solution. \end{proposition} Before giving the proof of this result, note that the condition $\langle u_{0},\mathbf 1\rangle=1$ implies that when $E=L^p(\mathbb R^N)$, $u_0$ also belongs to $L^1(\mathbb R^N)$. By virtue of the comment after Proposition~\ref{prop:wellposedness}, this guarantees that $T_tu_0$ is also in $L^1(\mathbb R^N)$ for all $t\geq0$, and consequently $\langle T_{t}u_{0},\mathbf 1\rangle$ is finite. \begin{proof} We start with the case $u_{0}\in E(W)$ and show that the function $$u(t)=\frac{T_{t}u_{0}}{\langle T_{t}u_{0},\mathbf 1\rangle}$$ is a classical solution to Equation~\eqref{eq:rep-mut}. By Proposition~\ref{prop:wellposedness}, since $u_{0}\in E(W)=D(\mathcal A)$, the function $t\mapsto T_{t}u_{0}$ is continuously differentiable and we have $$\frac{d}{dt} u(t)=\frac{1}{\langle T_{t}u_{0},\mathbf 1\rangle}\bigg[\mathcal A T_{t}u_{0}-\frac{\langle\mathcal A T_{t}u_{0},\mathbf 1\rangle}{\langle T_{t}u_{0},\mathbf 1\rangle}T_{t}u_{0}\bigg].$$ Using the fact that $\mathcal A$ is self-adjoint and that $$\langle\mathcal A T_{t}u_{0},\mathbf 1\rangle=\langle T_{t}u_{0},\mathcal A\mathbf 1\rangle = \lambda_{*} \langle T_{t}u_{0},\mathbf 1\rangle-\langle T_{t}u_{0},W\rangle$$ we get $$\frac{d}{dt} u(t) = (\mathcal A - \lambda_{*} I) u(t)+\langle u(t),W\rangle u(t).$$ This proves the existence part. For the uniqueness of the solution, we use the uniqueness result for the linear equation. Let $u$ be a classical solution to Equation~\eqref{eq:rep-mut}. It is clear that $$v(t)=u(t)\, \exp\left(\lambda_{*} t - \int_{0}^t\langle u(s),W\rangle\,ds \right)$$ is continuously differentiable, and by differentiation we readily get that $v$ is the unique classical solution to Equation~\eqref{eq:evolution-A}. \medskip Let us now turn to the case where $u_{0}$ does not necessarily belong to $E(W)$. The approach is the same as before, but one has to deal with mild solutions. Defining ${\widetilde T}_{t}={\rm e}^{-\lambda_{*} t}T_{t}$ and setting again $$u(t)=\frac{T_{t}u_{0}}{\langle T_{t}u_{0},\mathbf 1\rangle}=\frac{{\widetilde T}_{t}u_{0}}{\langle {\widetilde T}_{t}u_{0},\mathbf 1\rangle},$$ we have for any $t \geq 0$ \begin{equation}\label{machin-chose} \int_{0}^t\langle u(s),W\rangle {\widetilde T}_{t-s}u(s)\,ds=\int_{0}^t\frac{\langle {\widetilde T}_{s}u_{0},W\rangle}{\langle {\widetilde T}_{s}u_{0},\mathbf 1\rangle}\frac{{\widetilde T}_{t}u_{0}}{\langle{\widetilde T}_{s}u_{0},\mathbf 1\rangle}ds=\bigg(\int_{0}^t\frac{\langle {\widetilde T}_{s}u_{0},W\rangle}{\langle {\widetilde T}_{s}u_{0},\mathbf 1\rangle^2}ds\bigg){\widetilde T}_{t}u_{0}. \end{equation} Since $t\mapsto T_{t}u_{0}$ is a mild solution of Equation~\eqref{eq:evolution-A} we get by integration, using Fubini-Tonelli's theorem (note that $T_{t}u_{0}$ and $W$ are nonnegative), $$\langle T_{t}u_{0},\mathbf 1\rangle=\langle u_{0},\mathbf 1\rangle+\left\langle\int_{0}^t T_{s}u_{0}\,ds,\mathcal A\mathbf 1\right\rangle=1+\int_{0}^t\langle T_{s}u_{0},\lambda_{*} -W\rangle\,ds.$$ Since $t\mapsto\langle T_{t}u_{0},\lambda_{*} - W\rangle$ is locally integrable, it ensures that $t\mapsto\langle T_{t}u_{0},\mathbf 1\rangle$ belongs to $W^{1,1}_{\rm loc}(0,+\infty)$ with, in the weak sense, $$\frac{d}{dt}\langle T_{t}u_{0},\mathbf 1\rangle=\langle T_{t}u_{0},\lambda_{*} -W\rangle,\quad\text{or equivalently}\quad \frac{d}{dt}\langle{\widetilde T}_{t}u_{0},\mathbf 1\rangle=-\langle{\widetilde T}_{t}u_{0},W\rangle.$$ As a result, since $\langle u_{0},\mathbf 1\rangle=1$, $$\int_{0}^t\frac{\langle{\widetilde T}_{s}u_{0},W\rangle}{\langle{\widetilde T}_{s}u_{0},\mathbf 1\rangle^2}ds=\frac{1}{\langle{\widetilde T}_{t}u_{0},\mathbf 1\rangle}-1$$ which, combined with \eqref{machin-chose}, yields $$\int_{0}^t\langle u(s),W\rangle{\widetilde T}_{t-s}u(s)\,ds=\frac{{\widetilde T}_{t}u_{0}}{\langle{\widetilde T}_{t}u_{0},\mathbf 1\rangle}-{\widetilde T}_{t}u_{0} = u(t)-{\widetilde T}_{t}u_{0}$$ and this exactly means that $u$ is a mild solution to Equation~\eqref{eq:rep-mut}. To prove the uniqueness of the solution, we consider a mild solution $u$ to Equation~\eqref{eq:rep-mut} and we define $$v(t)=u(t)\, \exp\left(\lambda_{*} t-\int_{0}^t\langle u(s),W\rangle\,ds\right).$$ Let us also take a function $f\in C_{c}(\mathbb R^N)\subset D(\mathcal A)$. Since $T_{t}$ is self-adjoint, by differentiation of the equality \begin{align*} \langle u(t),f\rangle&=\langle{\widetilde T}_{t}u_{0},f\rangle+\int_{0}^t\langle u(s),W\rangle\langle{\widetilde T}_{t-s} u(s),f\rangle\,ds\\ &=\langle u_{0},{\widetilde T}_{t} f\rangle+\int_{0}^t\langle u(s),W\rangle\langle u(s),{\widetilde T}_{t-s}f\rangle\,ds \end{align*} we get \begin{align*} \frac{d}{dt}\langle u(t),f\rangle & = \langle u_{0},{\widetilde T}_{t}(\mathcal A - \lambda_{*} ) f\rangle+\langle u(t),W\rangle\langle u(t),f\rangle+\int_{0}^t\!\langle u(s),W\rangle\langle u(s),{\widetilde T}_{t-s}^*(\mathcal A -\lambda_{*} ) f\rangle\,ds\\ &=\langle u(t),(\mathcal A-\lambda_{*} ) f\rangle+\langle u(t),W\rangle\langle u(t),f\rangle. \end{align*} Since $t\mapsto\int_{0}^t\langle u(s),W\rangle\,ds$ belongs to $W^{1,1}_{\rm loc}(0,+\infty)$, we obtain \begin{align*} \frac{d}{dt}\langle v(t),f\rangle&=\frac{d}{dt}\langle u(t),f\rangle {\rm e}^{\lambda_{*} t-\int_{0}^t\langle u(s),W\rangle\,ds} +\big(\lambda_{*} -\langle u(t),W\rangle \big)\langle u(t),f\rangle {\rm e}^{\lambda_{*} t-\int_{0}^t\langle u(s),W\rangle\,ds}\\ &=\langle u(t),\mathcal A f\rangle {\rm e}^{\lambda_{*} t-\int_{0}^t\langle u(s),W\rangle\,ds}=\langle v(t),\mathcal A f\rangle. \end{align*} Integrating between $0$ and $t$ we obtain $$\langle v(t),f\rangle-\langle u_{0},f\rangle=\int_{0}^t\langle v(s),\mathcal A f\rangle\,ds=\left\langle\int_{0}^t v(s)\,ds,\mathcal A f\right\rangle=\left\langle\mathcal A\int_{0}^t v(s)\,ds, f\right\rangle$$ for all $f\in C_{c}(\mathbb R^N)$. By density of $C_{c}(\mathbb R^N)$ in $E$ we get that $v$ is a mild solution to the linear equation, and so $v(t)=T_{t}u_{0}$. \end{proof} Due to the explicit expression of the solution to the replicator-mutator model in terms of the semigroup $(T_{t})_{t\geq0}$ obtained in Proposition~\ref{prop:wellposed_rep-mut}, the conclusion of Theorem~\ref{thm:rep-mut} follows from Theorem~\ref{thm:asympto}. \begin{proof}[Proof of Theorem~\ref{thm:rep-mut}] It suffices to write \begin{align*} \left\|\frac{T_{t}u_{0}}{\langle T_{t}u_{0},\mathbf 1\rangle}-\frac{\phi}{\langle\phi,\mathbf 1\rangle}\right\|_E&=\left\|\frac{\langle\langle u_{0},\phi\rangle\phi - T_{t}u_{0},\mathbf 1\rangle T_{t}u_{0} + \langle T_{t}u_{0},\mathbf 1\rangle(T_{t}u_{0} - \langle u_{0},\phi\rangle\phi)}{\langle T_{t}u_{0},\mathbf 1\rangle\langle\phi,\mathbf 1\rangle\langle u_{0},\phi\rangle}\right\|_E\\ &\leq\frac{\|T_{t}u_{0}\|_E}{\langle T_{t}u_{0},\mathbf 1\rangle}\frac{\|T_{t}u_{0} - \langle u_{0},\phi \rangle\phi\|_\mathrm{TV}}{\langle\phi,\mathbf 1\rangle\langle u_{0},\phi \rangle}+\frac{\|T_{t}u_{0} - \langle u_{0},\phi \rangle\phi\|_E}{\langle\phi,\mathbf 1\rangle\langle u_{0},\phi \rangle} \end{align*} and use the exponential convergence in Theorem~\ref{thm:asympto} to get the result which is valid for any~$E$. For $E=\MM(\mathbb R^N)$ or $E=L^1(\mathbb R^N)$ the stronger conclusion follows by noticing that in these two cases $\langle T_{t}u_{0},\mathbf 1\rangle=\|T_{t}u_{0}\|_E$. \end{proof} \section{About the non-symmetric case}\label{sec:non-sym} In this last section, we do not assume that $J$ is even, that is Assumption~\ref{ass:J-symmetric} is not supposed any more. In counterpart, we strengthen Assumption~\ref{ass:linkJW} by requiring that Assumption~\ref{ass:Coville} is verified. As in the symmetric case, we first consider the Banach space $E=L^2(\mathbb R^N)$. For $J$ non-even, the operator $-L$ is not self-adjoint, and we cannot use the variational approach to prove the existence of a first eigenfunction. We bypass this issue by taking advantage of the fact that the semigroup generated by $-L$ is irreducible. Proposition~\ref{prop:wellposedness} does not require $J$ to be even, and it ensures that $-L$ generates a positive semigroup $(U_t)_{t\geq0}$ in any of the considered Banach spaces $E$. The strong maximum principle in Lemma~\ref{lem:Max-principle} is also valid without evenness assumption on~$J$, and it precisely means that $(U_t)_{t\geq0}$ is irreducible in $L^p(\mathbb R^N)$ for $1\leq p<\infty$ (see~\cite[Definition C-III-3.1.(v)]{Nagel86}, recalling that the quasi-interior points in $L^p(\mathbb R^N)$ with $1\leq p<\infty$ are the functions strictly positive a.e.). We thus have the following result. \begin{theorem}[The non-symmetric case]\label{thm:L2-non-sym} Let Assumptions~\ref{ass:kernel}, \ref{ass:potential} and~\ref{ass:Coville} hold. Then there exist $\lambda_1<0$ and two positive functions $\varphi$ and $\varphi^*$ in $L^2(\mathbb R^N)$ with $\|\varphi\|_{L^2(\mathbb R^N)}=\langle\varphi^*,\varphi\rangle=1$ such that \[L\varphi=\lambda_1\varphi\qquad\text{and}\qquad L^*\varphi^*=\lambda_1\varphi^*.\] Moreover, there exist two constants $C,a>0$ such that, for any $u_0\in L^2(\mathbb R^N)$ and all $t\geq0$ \begin{equation}\label{eq:conv-L2-non-sym} \big\|{\rm e}^{\lambda_1t}\,U_tu_0-\langle\varphi^*,u_0\rangle\varphi\big\|_{L^2}\leq C{\rm e}^{-at}\left\|u_0-\langle\varphi^*,u_0\rangle\varphi\right\|_{L^2}. \end{equation} \end{theorem} Let us mention that the convergence~\eqref{eq:conv-L2-non-sym} ensures the uniqueness of the triplet $(\lambda_1,\varphi,\varphi^*)$. \begin{proof} First we prove that Assumption~\ref{ass:Coville} ensures that $\omega_0(-L)>0$. Take $\epsilon,\eta>0$ and $B\subset\mathbb R^N$ such that \[\sigma^2\essinf_{B_\epsilon}\int_{B_\epsilon}\frac{J(x-y)}{W(y)}dy\geq1+\eta\] and define the function \[\psi(x):=\frac{1}{W(x)}\mathbf 1_{B_\epsilon}(x).\] Clearly $B_\epsilon$ is necessarily essentially bounded, so that $\psi\in D(-L)=L^2(1+W)$, and for almost all $x\in B_\epsilon$ we have \[-L\psi(x)=\sigma^2\int_{B_\epsilon}\frac{J(x-y)}{W(y)}dy-1\geq\eta.\] Since $\psi\leq1/\epsilon$, we deduce that \[-L\psi\geq\epsilon\eta\,\psi,\quad\text{and consequently}\quad U_t\psi\geq{\rm e}^{\epsilon\eta t}\psi.\] This ensures that $\omega_0(-L)\geq\epsilon\eta>0$, and we set $\lambda_1:=-\omega_0(-L)$. Besides, the proof of Lemma~\ref{rouge} does not require the evenness of $J$, and it guarantees that $\omega_\mathrm{ess}(\lambda_1-L)\leq\lambda_1$. We thus have $\omega_\mathrm{ess}(\lambda_1-L)<\omega_0(\lambda_1-L)=0$ and this has two implications. First, the semigroup $({\rm e}^{\lambda_1 t}\,U_t)_{t\geq0}$ generated by $\lambda_1-L$ is quasi-compact and second, since $\omega_0(\lambda_1-L)=\max(\omega_\mathrm{ess}(\lambda_1-L),\mathrm s(\lambda_1-L))$ (see for instance K.J. Engel \& R. Nagel~\cite[Corollary IV.2.11]{EN}), the spectral bound $\mathrm s(\lambda_1-L)$ is zero. We are in position to apply the result in W. Arendt \& al.~\cite[Section~C-IV, Remark~2.2.(d)]{Nagel86} to the semigroup $(T_t)_{t\geq0}:=({\rm e}^{\lambda_1 t}\,U_t)_{t\geq0}$, and it guarantees the existence of $\varphi$ and $\varphi^*$ positive such that~\eqref{eq:conv-L2-non-sym} holds. Using this convergence to pass to the limit $s\to+\infty$ in $T_tT_su_0=T_{t+s}u_0$ we get that $T_t\varphi=\varphi$, which yields $L\varphi=\lambda_1\varphi$ and, choosing $u_0=\varphi$ in~\eqref{eq:conv-L2-non-sym}, $\langle\varphi^*,\varphi\rangle=1$. Finally, taking $Lu_0$ in place of $u_0$ in~\eqref{eq:conv-L2-non-sym} we get by passing to the limit $t\to+\infty$ that \[\langle\varphi^*,u_0\rangle L\varphi=\langle\varphi^*,Lu_0\rangle\varphi\] and so $\lambda_1\langle\varphi^*,u_0\rangle=\langle\varphi^*,Lu_0\rangle=\langle L^*\varphi^*,u_0\rangle$ for all $u_0\in D(L)=L^2(1+W)$. This obviously implies that $L^*\varphi^*=\lambda_1\varphi^*$, and the proof is complete. \end{proof} Now that we have Theorem~\ref{thm:L2-non-sym} at hand, we can use the same strategy as in the symmetric case to deduce the following counterpart of Theorem~\ref{thm:asympto} in the non-symmetric case. \begin{corollary} The functions $\varphi$ and $\varphi^*$ belong to $L^1(\mathbb R^N)\cap C_0(\mathbb R^N)$ and there exist two constants $C > 0$ and $a > 0$ such that, for any $u_{0}\in E$, the solution $u=u(t,x)$ of \eqref{eq:parabolique} starting from $u_{0}=u_0(x)$ satisfies, for all $t > 0$, $$ \big\| {\rm e}^{\lambda_{*} t}u(t,\cdot)-\langle \phi^*,u_{0} \rangle \phi\big\|_{E} \leq C\, {\rm e}^{-at}\left\|u_{0} - \langle \varphi^*, u_{0} \rangle\phi\right\|_{E} , $$ where $\lambda_*=\lambda_1+\sigma^2$. \end{corollary} \bigskip \noindent{\large{\bf Acknowledgements.}} MA is supported by the ANR project DEEV, ANR-20-CE40-0011-01, funded by the French Ministry of Research. PG is supported by the ANR project NOLO, ANR-20-CE40-0015, funded by the French Ministry of Research. \medskip

https://arxiv.org/abs/1704.02190

Symmetry and Nonexistence of Positive Solutions for Fractional Choquard Equations

This paper is devoted to study the following Choquard equation \begin{eqnarray*}\left\{ \begin{array}{lll}(-\triangle)^{\alpha/2}u=(|x|^{\beta-n}\ast u^p)u^{p-1},~~~&x\in R^n, u\geq0,\,\,&x\in R^n,\end{array}\right. \end{eqnarray*} where $0<\alpha,\beta<2$, $1\leq p<\infty$, and $n\geq2$. Using a direct method of moving planes, we prove the symmetry and nonexistence of positive solutions in the critical and subcritical case respectively.

\section{Introduction} We study the following Choquard equation involving the fractional Laplacian \begin{eqnarray}\label{1.1} \left\{ \begin{array}{ll} (-\triangle)^{\alpha/2}u=(|x|^{\beta-n}\ast u^p)u^{p-1},\,\,&x\in R^n,\\ u\geq0,\,\,&x\in R^n, \end{array} \right. \end{eqnarray} where $0<\alpha, \beta<2$, $1\preceq p<\infty$ and $n\geq2$. The fractional Laplacian in $R^n$ is a nonlocal pseudo-differential operator taking the form \begin{eqnarray}\label{1.2} (-\triangle)^{\alpha/2}u(x)=C_{n,\alpha}PV \int_{R^n}\frac{u(x)-u(y)}{|x-y|^{n+\alpha}}dy=C_{n,\alpha}\lim_{\varepsilon\rightarrow 0} \int_{R^n\setminus B_\varepsilon(x)}\frac{u(x)-u(y)}{|x-y|^{n+\alpha}}dy, \end{eqnarray} where $C_{n,\alpha}$ is a normalization constant. This operator is well defined in $\mathcal{S}$, the Schwartz space of rapidly decreasing $C^\infty$ functions in $R^n$. In this space, it can also be equivalently defined in terms of the Fourier transform \begin{eqnarray*} \mathcal{F}[(-\triangle)^{\alpha/ 2}u](\xi)=|\xi|^\alpha\mathcal{F}u(\xi), \end{eqnarray*} where $\mathcal{F}u$ is the Fourier transform of $u$. One can extend this operator to a wider space of distributions: \begin{eqnarray*} \mathcal{L}_\alpha= \{u:R^n\rightarrow R\mid\int_{R^n}\frac{|u(x)|}{1+|x|^{n+\alpha}}dx<\infty\}. \end{eqnarray*} Then in this space, we defined $(-\triangle)^{\alpha/2}u$ as a distribution by \begin{eqnarray*} \langle(-\triangle)^{\alpha/2}u(x), \phi\rangle=\int_{R^n} u(x)(-\triangle)^{\alpha/2}\phi(x)dx, \,\,\forall\phi\in C_0^\infty(R^n). \end{eqnarray*} In our paper, let \begin{equation}\label{011801} v(x)=|x|^{\beta-n}\ast u^p=\int_{R^n}\frac{u^p(y)}{|x-y|^{n-\beta}}dy, \end{equation} and act $(-\triangle)^{\beta/2}$ on both side of (\ref{011801}), we obtain $$(-\triangle)^{\beta/2}v(x)=u^p(x).$$ Then, (\ref{1.1}) is equivalent to \begin{eqnarray}\label{011802} \left\{ \begin{array}{ll} (-\triangle)^{\alpha/2}u=v(x)u^{p-1}(x),\,\,&x\in R^n,\\ (-\triangle)^{\beta/2}v=u^{p}(x),\,\,&x\in R^n,\\ u\geq0,v\geq 0,\,\,&x\in R^n. \end{array} \right. \end{eqnarray} Hence, to study (\ref{1.1}), it is sufficiently to investigate (\ref{011802}). In recent years, the fractional Laplacian has attracted much attention. It appears in diverse physical phenomena, such as anomalous diffusion and quasi-geostrophic flows. It also has various applications in probability and finance. In particular, the fractional Laplacian can be understood as the infinitesimal generator of a stable L\'{e}vy diffusion process and appear in anomalous diffusions in plasmas, flames propagation and chemical reactions in liquids, population dynamics, geographical fluid dynamics, and American options in finance. For readers who are interested in the application of the fractional Laplacian, please refer to \cite{A}, \cite{B} and the references therein. In \cite{BCDS}, the authors considered the following fractional Laplacian equation \begin{eqnarray}\label{021401} (-\triangle)^{\alpha/2}u=u^p,\,\, x\in R^n, \end{eqnarray} they used the extension method to deduce the nonlocal problem into a local one in a higher dimensional half space $R^n\times [0, \infty)$, then applied the method of moving planes to show the symmetry of $U(x,y)$ in $x$, then derived the nonexistence of positive solutions in the subcritical case. In \cite{CLL}, the authors developed a direct method of moving planes for the fractional Laplacian, and using this method, they derived the symmetry and nonexistence of positive solutions for \begin{eqnarray*} \left\{ \begin{array}{ll} &(-\triangle)^{\alpha/2}u=u^p,\\ &u\geq0, \end{array} \right. \end{eqnarray*} in $R^n$ and $R^n_+$. In \cite{LM}, the authors studied the system involving the fractional Laplacian \begin{eqnarray*} \left\{ \begin{array}{ll} (-\triangle)^{\alpha/2}u=f(v(x)),\,\,&x\in R^n,\\ (-\triangle)^{\beta/2}v(x)=g(u(x)),\,\,&x\in R^n,\\ u\geq0, v\geq0,\,\,&x\in R^n. \end{array} \right. \end{eqnarray*} First, they used the iteration method to establish the maximum principles for system, then derived the symmetry of non-negative solutions by the direct method of moving planes without any decay assumption at infinity. In our paper, we first establish the maximum principles by the iteration method introduced by \cite{LM}, and then used the direct method of moving planes introduced by \cite{CLL} to derive the symmetry of positive solutions and then deduce the nonexistence of positive solutions. The following is our main theorems. \begin{thm}(\textbf{Decay at Infinity})\label{t1} Let $\Omega$ be an unbounded region in $\Sigma_\lambda$. Assume $\varphi\in L_\alpha\cap C^{1,1}_{loc}(\Omega)$, $\phi\in L_\beta\cap C^{1,1}_{loc}(\Omega)$ and $\varphi(x), \phi(x)$ are lower semi-continuous. If \begin{eqnarray}\label{1.3} \left\{ \begin{array}{lll} (-\triangle)^{\alpha/2}\varphi(x)+C_2(x)\varphi(x)+C_3(x)\phi(x)\geq 0 ~~~ ~~~&in& \Omega,\\ (-\triangle)^{\beta/2}\phi(x)+C_1(x)\varphi(x)\geq 0 ~~~ ~~~&\mbox{in}& \Omega,\\ \varphi(x),\phi(x)\geq0 ~~~ ~~~&\mbox{in}& \Sigma_\lambda\backslash\Omega,\\ \varphi(x^\lambda)=-\varphi(x) ~~~ ~~~&\mbox{in}& \Sigma_\lambda,\\ \phi(x^\lambda)=-\phi(x) ~~~ ~~~&\mbox{in}& \Sigma_\lambda, \end{array} \right. \end{eqnarray} with \begin{eqnarray}\label{1.4} C_1(x),C_3(x)\sim\frac{1}{|x|^{\alpha+\beta}}, \,\,C_2(x)\sim\frac{1}{|x|^{2\alpha}},\,\,\mbox{for} \,\,|x|\,\,\,\mbox{large}, \end{eqnarray} and $$C_1(x), C_2(x), C_3(x)<0.$$ Then there exists a constant $R_0 > 0$ such that if \begin{eqnarray} \varphi(x_0) = \min_{\Omega}\varphi(x)<0,\,\,\phi(x_1)=\min_{\Omega}\phi(x)<0, \end{eqnarray} then at least one of $x_0$ and $x_1$ satisfies \begin{eqnarray}\label{1.12} |x|\preceq R_0. \end{eqnarray} \end{thm} \begin{thm}(\textbf{Narrow Region Principle})\label{t2} Let $\Omega$ be a bounded narrow region in $\Sigma_\lambda$, such that it is contained in $\{x|\lambda-\delta<x_1<\lambda\}$ with small $l$. Suppose that $\varphi(x)\in L_\alpha\cap C^{1,1}_{loc}(\Omega)$, $\phi(x)\in L_\beta\cap C^{1,1}_{loc}(\Omega)$ and $\varphi(x), \phi(x)$ are lower semi-continuous. If $C_1(x)$, $C_2(x)$ and $C_3(x)$ are bounded from below in $\Omega$, then \begin{eqnarray}\label{1.5} \left\{ \begin{array}{lll} (-\triangle)^{\alpha/2}\varphi(x)+C_2(x)\varphi(x)+C_3(x)\phi(x)\geq 0 ~~~ ~~~&in& \Omega,\\ (-\triangle)^{\beta/2}\phi(x)+C_1(x)\varphi(x)\geq 0 ~~~ ~~~&in& \Omega,\\ \varphi(x),\phi(x)\geq0 ~~~ ~~~&\mbox{in}& \Sigma_\lambda\backslash\Omega,\\ \varphi(x^\lambda)=-\varphi(x) ~~~ ~~~&\mbox{in}& \Sigma_\lambda,\\ \phi(x^\lambda)=-\phi(x) ~~~ ~~~&\mbox{in}& \Sigma_\lambda, \end{array} \right. \end{eqnarray} then for sufficiently small $\delta$, we have \begin{eqnarray}\label{2017010901} \varphi(x),\phi(x) \geq 0,\,x\in\Omega. \end{eqnarray} Furthermore, if $\varphi = 0$ or $\phi(x)=0$ at some point in $\Omega$, then \begin{eqnarray}\label{2017010902} \varphi(x)=\phi(x) \equiv 0~~~\text{almost everywhere in}~~~R^n. \end{eqnarray} These conclusions hold for unbounded region $\Omega$ if we further assume that \begin{eqnarray}\label{2017010903} \underline{\lim\limits}_{|x|\rightarrow\infty}\varphi(x),\phi(x)\geq0. \end{eqnarray} \end{thm} \begin{thm} Let $0<\alpha,\beta<2$, $\frac{n}{n-\alpha}\preceq p\preceq\frac{n+\beta}{n-\alpha}$. Assume $u\in L_\alpha\cap C^{1,1}_{loc}$ and $v\in L_\beta\cap C^{1,1}_{loc}$ satisfy (\ref{1.1}). Then, (i) in the subcritical case $\frac{n}{n-\alpha}\preceq p<\frac{n+\beta}{n-\alpha}$, (\ref{1.1}) has no positive solution; (ii) in the critical case $p=\frac{n+\beta}{n-\alpha}$, the positive solutions must be radially symmetric and monotone decreasing about some point in $R$. \end{thm} \section{Proof of Theorem \ref{t1} and \ref{t2}} We first give some basic notations. Let \begin{eqnarray*} T_\lambda=\{x\in R^n\mid x_1=\lambda,\lambda\in R\} \end{eqnarray*} be the moving plane, \begin{eqnarray*} \Sigma_\lambda=\{x\in R^n\mid x_1<\lambda\} \end{eqnarray*} be the region to the left of the plane, $\tilde{\Sigma}_{\lambda}=R^n\backslash\Sigma_{\lambda}$, and \begin{eqnarray*} x^\lambda=(2\lambda-x_1,x_2,...,x_n) \end{eqnarray*} be the reflection of the point $x = (x_1, x_2, \cdot\cdot\cdot , x_n)$ about the plane $\mathrm{T}_\lambda$. \subsection{Decay at Infinity} \textbf{Proof}. By the definition of the fractional Laplacian (\ref{1.2}), \begin{eqnarray}\label{1.6}\nonumber (-\triangle)^{\beta/2}\phi(x_1)&=&C_{n,\beta}PV\int_{R^n}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\tilde{\Sigma}_{\lambda}}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y^\lambda)}{|x-y^\lambda|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)+\phi(y)}{|x_1-y^\lambda|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)+\phi(y)}{|x_1-y^\lambda|^{n+\beta}}dy\\ \nonumber &\preceq&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y^\lambda|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)+\phi(y)}{|x_1-y^\lambda|^{n+\beta}}dy\\ \nonumber &\preceq&C_{n,\beta}\int_{\Sigma_\lambda}\frac{2\phi(x_1)}{|x_1-y^\lambda|^{n+\beta}}dy. \end{eqnarray} Fix $\lambda$, from the fact $x_1\in\Sigma_\lambda$ and $|x_1|$ sufficiently large, \begin{eqnarray}\label{1.7}\nonumber \int_{\sum_\lambda} \frac{1}{|x_1-y^\lambda|^{n+\beta}}dy&\geq&\int_{\{x_1\geq0\}} \frac{1}{|x_1-y^\lambda|^{n+\beta}}dy\\ \nonumber &=&\frac{1}{2}\int_{R^n}\frac{1}{(|x_1|+|y^\lambda|)^{n+\beta}}dy\\ \nonumber &=&\frac{1}{2}\int_{0}^{\infty}\int_{B_r^0}\frac{1}{(|x_1|+|r|)^{n+\beta}}d\sigma dr\\ \nonumber &=&\frac{1}{2}\int_0^{\infty}\frac{w_{n-1}r^{n-1}}{(|x_1|+|r|)^{n+\beta}}dr\\ \nonumber &=&\frac{w_{n-1}}{2|x_1|^\beta}\int_0^{\infty}\frac{t^{n-1}}{(1+t)^{n+\beta}}dr, r=t|x_2|\\ &\sim&\frac{C}{|x_1|^\beta}. \end{eqnarray} Then, \begin{eqnarray}\label{1.8} (-\triangle)^{\beta/2}\phi(x_1)\preceq\frac{C}{|x_1|^\beta}\phi(x_1)<0. \end{eqnarray} Combining this with (\ref{1.3}), we can show \begin{eqnarray}\label{1.9} \varphi(x_1)<0, \end{eqnarray} and \begin{eqnarray}\label{1.10} \phi(x_1)\geq -CC_1|x_1|^\beta\varphi(x_1). \end{eqnarray} We know there exists $x_0$ such that $$\varphi(x_0)=\min_\Omega\varphi(x)<0.$$ From a similar argument as in (\ref{1.6}), we can show \begin{eqnarray}\label{1.11} (-\triangle)^{\alpha/2}\varphi(x_0)\preceq\frac{C}{|x_0|^\alpha}\varphi(x_0). \end{eqnarray} From (\ref{1.3}) and (\ref{1.10}), we can deduce \begin{eqnarray*} 0&\preceq&(-\triangle)^{\alpha/2}\varphi(x_0)+C_2(x_0)\varphi(x_0)+C_3(x_0)\phi(x_0)\\ &\preceq&\frac{C}{|x_0|^\alpha}\varphi(x_0)+C_2(x_0)\varphi(x_0)+C_3(x_0)\phi(x_1)\\ &\preceq&\frac{C}{|x_0|^\alpha}\varphi(x_0)+C_2(x_0)\varphi(x_0)-CC_3(x_0)C_1(x_1)|x_1|^\beta\varphi(x_1)\\ &<&0. \end{eqnarray*} The last inequality follows from assumptions (\ref{1.4}). This is a contradiction. Then (\ref{1.12}) must be true for at least one of $x_0$ and $x_1$. \subsection{Narrow Region Principle} \textbf{Proof}. If (\ref{2017010901}) does not hold, because $\phi(x)$ is lower semi-continuous, there exists $x_1\in\bar{\Omega}$ such that \begin{eqnarray*} \phi(x_1)=\min_{\bar{\Omega}}\phi(x)<0. \end{eqnarray*} By the definition of $(-\triangle)^{\beta/2}$, we have \begin{eqnarray}\label{010904}\nonumber (-\triangle)^{\beta/2}\phi(x_1)&=&C_{n,\beta}PV\int_{R^n}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\tilde{\Sigma}_{\lambda}}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y^\lambda)}{|x-y^\lambda|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)+\phi(y)}{|x_1-y^\lambda|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)+\phi(y)}{|x_1-y^\lambda|^{n+\beta}}dy\\ \nonumber &\preceq&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)-\phi(y)}{|x_1-y^\lambda|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(x_1)+\phi(y)}{|x_1-y^\lambda|^{n+\beta}}dy\\ \nonumber &\preceq&C_{n,\beta}\int_{\Sigma_\lambda}\frac{2\phi(x_1)}{|x_1-y^\lambda|^{n+\beta}}dy. \end{eqnarray} Let $D=B_{2\delta}(x_1)\cap\tilde{\Sigma}_\lambda$, then \begin{eqnarray}\label{010902}\nonumber &&\int_{\Sigma_\lambda}\frac{1}{|x_1-y^\lambda|^{n+\beta}}dy\\ \nonumber &\geq&\int_{D}\frac{1)}{|x_1-y|^{n+\beta}}dy\\ \nonumber &\geq&\frac{1}{10}\int_{B_{2\delta}(x_1)}\int_{D}\frac{1)}{|x_1-y|^{n+\beta}}dy\\ &\sim&\frac{C}{\delta^\beta}. \end{eqnarray} Thus, \begin{eqnarray} (-\triangle)^{\beta/2}\phi(x_1)\preceq \frac{C\phi(x_1)}{\delta^\beta}<0. \end{eqnarray} Combining this with (\ref{1.5}), we have \begin{eqnarray} -C_1(x_1)\varphi(x_1)\preceq \frac{C\phi(x_1)}{\delta^\beta}. \end{eqnarray} We know there exists a $x_2$ such that $$\varphi(x_2)=\min_{\bar{\Omega}}\varphi(x)<0.$$ Similar to (\ref{010902}), we can derive that \begin{eqnarray*} (-\triangle)^{\alpha/2}\varphi(x_2)\preceq\frac{C\varphi(x_2)}{\delta^\alpha}<0. \end{eqnarray*} By (\ref{1.5}), for $\delta$ sufficiently small, we have \begin{eqnarray*} 0&\preceq&(-\triangle)^{\alpha/2}\varphi(x_2)+C_2(x_2)\varphi(x_2)+C_3(x_2)\phi(x_2)\\ &\preceq&\frac{C\varphi(x_2)}{\delta^\alpha}+C_2(x_2)\varphi(x_2)+C_3(x_2)\phi(x_1)\\ &\preceq&\frac{C\varphi(x_2)}{\delta^\alpha}+C_2(x_2)\varphi(x_2)-C_3(x_2)C_1(x_1)C\delta^\beta\phi(x_1)\\ &\preceq&\frac{C\varphi(x_2)}{\delta^\alpha}+C_2(x_2)\varphi(x_2)-C_3(x_2)C_1(x_1)\delta^\beta C\phi(x_2)\\ &=&\frac{C\varphi(x_2)}{\delta^\alpha}(1+\frac{C_2(x_2)\delta^\alpha}{C}-C_3(x_2)C_1(x_1)\delta^\alpha\delta^\beta)\\ &<&0, \end{eqnarray*} which is a contradiction. To prove (\ref{2017010902}), we suppose there exists $\bar{x}\in \Omega$ such that $$\phi(\bar{x})=0.$$ Then, \begin{eqnarray}\nonumber (-\triangle)^{\beta/2}\phi(\bar{x})&=&C_{n,\beta}PV\int_{R^n}\frac{\phi(\bar{x})-\phi(y)}{|\bar{x}-y|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{-\phi(y)}{|\bar{x}-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\tilde{\Sigma}_{\lambda}}\frac{-\phi(y)}{|\bar{x}-y|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{-\phi(y)}{|\bar{x}-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{-\phi(y^\lambda)}{|x-y^\lambda|^{n+\beta}}dy\\ \nonumber &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{-\phi(y)}{|\bar{x}-y|^{n+\beta}}dy+C_{n,\beta}PV\int_{\Sigma_\lambda}\frac{\phi(y)}{|\bar{x}-y^\lambda|^{n+\beta}}dy\\ &=&C_{n,\beta}PV\int_{\Sigma_\lambda}\big(\frac{1}{|\bar{x}-y^\lambda|^{n+\beta}}-\frac{1}{|\bar{x}-y|^{n+\beta}}\big)\phi(y)dy. \label{011701} \end{eqnarray} If $\phi(x)\neq 0$, (\ref{011701}) implies that $$(-\triangle)^{\beta/2}\phi(\bar{x})<0.$$ Combining this with (\ref{1.5}), we can derive $$\varphi(\bar{x})<0,$$ which is a contradiction with (\ref{2017010901}). Therefore $\phi(x)$ is identically $0$ in $\Sigma_\lambda$. Since $$\phi(x^\lambda)=-\phi(x),\,\,x\in\Sigma_\lambda,$$ it shows that $$\phi(x)=0,\,\,x\in R^n,$$ then $$(-\triangle)^{\beta/2}\phi(x)=0,\,\,x\in R^n.$$ From (\ref{1.5}), we know $$\varphi(x)\leq0,\,\,x\in \Sigma_\lambda.$$ We already proved $$\varphi(x)\geq0,\,\,x\in \Sigma_\lambda.$$ It must hold $$\varphi(x)=0,\,\,x\in \Sigma_\lambda.$$ Combining this with the fact $$\varphi(x^\lambda)=-\varphi(x),\,\,x\in\Sigma_\lambda,$$ we have $$\varphi(x)\equiv 0,\,\,x\in R^n.$$ From a similar argument, we can show if $\varphi(x)$ is $0$ at one point in $\Sigma_\lambda$, then $\phi(x)$ and $\varphi(x)$ are identically 0 in $R^n$. \section{The Symmetry of Positive Solutions} Without any decay conditions on $u$ and $v$, we are not able to carry the method of moving planes on $u$ and $v$ directly. To circumvent this difficulty, we make a Kelvin transform. For any $x_0\in R^n$, let \begin{eqnarray*} &\overline{u}(x)=\frac{1}{|x-x_0|^{n-\alpha}}u(\frac{x-x_0}{|x-x_0|^2}+x_0),\\ &\overline{v}(x)=\frac{1}{|x-x_0|^{n-\beta}}v(\frac{x-x_0}{|x-x_0|^2}+x_0). \end{eqnarray*} Without loss of generality, let $x_0=0$, then \begin{eqnarray*} &\overline{u}(x)=\frac{1}{|x|^{n-\alpha}}u(\frac{x}{|x|^2}),\\ &\overline{v}(x)=\frac{1}{|x|^{n-\beta}}v(\frac{x}{|x|^2}). \end{eqnarray*} Thus, \begin{eqnarray*} (-\triangle)^{\alpha/2}\overline{u}(x)&=&\frac{1}{|x|^{n+\alpha}}(-\triangle)^{\alpha/2}u(\frac{x}{|x|^2}),\\ &=&\frac{1}{|x|^{n+\alpha}}v(\frac{x}{|x|^2})u^{p-1}(\frac{x}{|x|^2}),\\ &=&\frac{1}{|x|^{\alpha+\beta-(p-1)(n-\alpha)}}\overline{v}(x)\overline{u}^{p-1}(x). \end{eqnarray*} In a similar way, we have \begin{eqnarray*} (-\triangle)^{\beta/2}\overline{v}(x)=\frac{1}{|x|^{\alpha+\beta-(p-1)(n-\alpha)}}\overline{u}^p(x). \end{eqnarray*} Let $\gamma=\alpha+\beta-(p-1)(n-\alpha)$, and $(\ref{011802})$ becomes \begin{eqnarray}\label{011803} \left\{ \begin{array}{ll} (-\triangle)^{\alpha/2}\overline{u}(x)=|x|^{-\gamma}\overline{v}(x)\overline{u}^{p-1}(x),\,\,&x\in R^n,\\ (-\triangle)^{\beta/2}\overline{v}=|x|^{-\gamma}\overline{u}^p(x),\,\,&x\in R^n,\\ \overline{u}\geq0, \overline{v}\geq 0,\,\,&x\in R^n. \end{array} \right. \end{eqnarray} We first give some basic notations before starting moving the planes. Then we start moving planes on system $(\ref{011802})$. Let \begin{eqnarray*} T_\lambda=\{x\in R^n\mid x_1=\lambda,\lambda\in R\} \end{eqnarray*} be the moving plane, \begin{eqnarray*} \Sigma_\lambda=\{x\in R^n\mid x_1<\lambda\} \end{eqnarray*} be the region to the left of the plane, and \begin{eqnarray*} x^\lambda=(2\lambda-x_1,x_2,...,x_n) \end{eqnarray*} be the reflection of the point $x = (x_1, x_2, \cdot\cdot\cdot , x_n)$ about the plane $\mathrm{T}_\lambda$. Assume that $(\overline{u},\overline{v})$ solves the fractional system $(\ref{011802})$. To compare the values of $\overline{u}(x)$ with $\overline{u}(x^\lambda)$ and $\overline{v}(x)$ with $\overline{v}(x^\lambda)$, we denote \begin{eqnarray*} \left\{ \begin{array}{ll} U_\lambda(x) = \overline{u}(x^\lambda)-\overline{u}(x),\\ V_\lambda(x) = \overline{v}(x^\lambda)-\overline{v}(x).\\ \end{array} \right. \end{eqnarray*} Then system $(\ref{011802})$ becomes \begin{eqnarray}\label{011804} \left\{ \begin{array}{ll} (-\triangle)^{\alpha/2} U_{\lambda}(x)=\frac{1}{|x^\lambda|^{\gamma}}\overline{v}(x^\lambda)\overline{u}^{p-1}(x^\lambda)-\frac{1}{|x|^{\gamma}}\overline{v}(x)\overline{u}^{p-1}(x),\\ (-\triangle)^{\beta/2} V_{\lambda}(x)=\frac{1}{|x^\lambda|^{\gamma}}\overline{u}^{p}(x^\lambda)-\frac{1}{|x|^{\gamma}}\overline{u}^{p}(x). \end{array} \right. \end{eqnarray} \subsection{Proof of Theorem 1.1} Now,we start moving planes. \subsubsection{Subcritical Case $\frac{n}{n-\alpha}\preceq p<\frac{n+\beta}{n-\alpha}$.} $\mathbf{Step.1}$: We show that when $\lambda$ sufficiently negative, \begin{eqnarray}\label{011805} U_\lambda(x), V_\lambda(x)\geq0, \,\forall x\in\Sigma_{\lambda}\setminus\{0^\lambda\}. \end{eqnarray} We claim that for $\lambda$ sufficiently negative, there exists a constant $C$ such that $$U_\lambda(x), V_\lambda(x)\geq C>0,\,\,x\in B_\varepsilon(0^\lambda)\backslash\{0^\lambda\},$$ we will prove it in Appendix. Hence, there must be a point $\bar{x}$ such that $$U_\lambda(\bar{x})=\min_{x\in\Sigma_\lambda }U_\lambda(x)<0.$$ Moreover, \begin{eqnarray*} &&(-\triangle)^{\alpha/2}U_\lambda(\bar{x})\\ &=&C_{n,\alpha} PV\int_{R^n}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy\\ &=&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy+C_{n,\alpha}PV\int_{\tilde{\Sigma}_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy\\ &=&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy+C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y^\lambda)}{|\bar{x}-y^\lambda|^{n+\alpha}}dy\\ &=&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy+C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})+U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy\\ &\preceq&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y^\lambda|^{n+\alpha}}dy+C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})+U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy\\ &=&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{2U_\lambda(\bar{x})}{|\bar{x}-y^\lambda|^{n+\alpha}}dy. \end{eqnarray*} From a similar argument in (\ref{1.8}), we have \begin{equation}\label{011902} (-\triangle)^{\alpha/2}U_\lambda(\bar{x})\preceq\frac{CU_\lambda(\bar{x})}{|\bar{x}|^\alpha}<0. \end{equation} We claim that \begin{equation}\label{011901} V_\lambda(\bar{x})<0. \end{equation} Indeed, if not, $V_\lambda(\bar{x})\geq 0$. From (\ref{011804}), we have \begin{eqnarray*} &&(-\triangle)^{\alpha/2}U_\lambda(\bar{x})\\ &=&\frac{1}{|\bar{x}^\lambda|^{\gamma}}\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x})\\ &\geq&\frac{1}{|\bar{x}^\lambda|^{\gamma}}\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)\\ &\geq&\frac{1}{|\bar{x}^\lambda|^{\gamma}}\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)\\ &\geq&\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x}^\lambda)V_\lambda(\bar{x})\\ &\geq&0. \end{eqnarray*} This is a contradiction with (\ref{011902}). Then (\ref{011901}) holds. And from (\ref{011901}), we know there exists $\tilde{x}$ such that $$V_\lambda({\tilde{x}})=\min_{\Sigma_\lambda}V_\lambda(x)<0.$$ Similar to (\ref{011901}), we can obtain $$U_\lambda(\tilde{x})<0.$$ Therefore, \begin{eqnarray}\nonumber &&(-\triangle)^{\alpha/2}U_\lambda(\bar{x})\\ \nonumber &=&\frac{1}{|\bar{x}^\lambda|^{\gamma}}\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x})\\ \nonumber &=&\frac{1}{|\bar{x}^\lambda|^{\gamma}}\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)+\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x})\\ \nonumber &\geq&\frac{1}{|\bar{x}|^{\gamma}}\big[\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)\big]+\frac{1}{|\bar{x}|^{\gamma}}\big[\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)-\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x})\big]\\ \nonumber &=&\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x}^\lambda)V_\lambda(\bar{x})+(p-1)\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\xi^{p-2}U_\lambda(\bar{x}),\,\,\,\,\xi\in[\overline{u}(\bar{x}^\lambda),\overline{u}(\bar{x})]\\ \label{012701} &\geq&\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x})V_\lambda(\bar{x})+(p-1)\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\bar{u}(\bar{x})^{p-2}U_\lambda(\bar{x}), \end{eqnarray} and \begin{eqnarray*} &&(-\triangle)^{\beta/2}V_\lambda(\tilde{x})\\ &=&\frac{1}{|\tilde{x}^\lambda|^{\gamma}}\overline{u}^{p}(\tilde{x}^\lambda)-\frac{1}{|\tilde{x}|^{\gamma}}\overline{u}^{p}(\tilde{x})\\ &=&\frac{1}{|\tilde{x}^\lambda|^{\gamma}}\overline{u}^{p}(\tilde{x}^\lambda)-\frac{1}{|\tilde{x}|^{\gamma}}\overline{u}^{p}(\tilde{x}^\lambda)+\frac{1}{|\tilde{x}|^{\gamma}}\overline{u}^{p}(\tilde{x}^\lambda)-\frac{1}{|\tilde{x}|^{\gamma}}\overline{u}^{p}(\tilde{x})\\ &=&(\frac{1}{|\tilde{x}^\lambda|^{\gamma}}-\frac{1}{|\tilde{x}|^{\gamma}})\overline{u}^{p}(\tilde{x}^\lambda)+p\frac{1}{|\tilde{x}|^{\gamma}}\eta^{p-1}U_\lambda(\tilde{x}),\,\,\,\eta\in[\overline{u}(\tilde{x}^\lambda),\overline{u}(\tilde{x})]\\ &\geq&p\frac{1}{|\tilde{x}|^{\gamma}}\overline{u}^{p-1}(\tilde{x})U_\lambda(\tilde{x}). \end{eqnarray*} Let \begin{eqnarray*} C_1(\tilde{x})&=&p\frac{1}{|\tilde{x}|^{\gamma}}\overline{u}^{p-1}(\tilde{x})\\ &\sim&\frac{1}{|\tilde{x}|^{\gamma}}\frac{1}{|\tilde{x}|^{(n-\alpha)(p-1)}}\\ &\sim&\frac{1}{|\tilde{x}|^{\alpha+\beta}},\,\,|\tilde{x}| \,\,\mbox{large \,\,enough}, \end{eqnarray*} \begin{eqnarray*} C_2(\bar{x})&=&(p-1)\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\bar{u}(\bar{x})^{p-2}\\ &\sim&\frac{1}{|\bar{x}|^{\gamma}}\frac{1}{|\bar{x}|^{n-\beta}}\frac{1}{|\bar{x}|^{(n-\alpha)(p-2)}}\\ &\sim&\frac{1}{|\bar{x}|^{2\alpha}},\,\,|\bar{x}| \,\,\mbox{large \,\,enough}, \end{eqnarray*} and \begin{eqnarray*} C_3(\bar{x})&=&\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x})\\ &\sim&\frac{1}{|\bar{x}|^{\gamma}}\frac{1}{|\bar{x}|^{(n-\alpha)(p-1)}}\\ &\sim&\frac{1}{|\bar{x}|^{\alpha+\beta}},\,\,|\bar{x}| \,\,\mbox{large \,\,enough}. \end{eqnarray*} Then by Theorem \ref{t1}, if $\lambda$ sufficiently negative, there must be one of $U_\lambda(x)$ and $V_\lambda(x)$ positive in $\Sigma_\lambda\backslash\{0^\lambda\}.$ Without loss of generality, we assume $$U_\lambda(x)\geq0,\,\,x\in\Sigma_\lambda\backslash\{0^\lambda\}.$$ And we claim that $$V_\lambda(x)\geq0,\,\,x\in\Sigma_\lambda\backslash\{0^\lambda\}.$$ If not, there exists $\tilde{x}$ such that $$V_\lambda(\tilde{x})=\min_{\Sigma_\lambda}V_\lambda(x)<0.$$ Then from a similar argument, we can show $$0>\frac{CV_\lambda(\tilde{x})}{|\tilde{x}|^\beta}\geq(-\triangle)^{\beta/2}V_\lambda(\tilde{x})\geq p\frac{1}{|\tilde{x}|^\gamma}\bar{x}^{p-1}(\tilde{x})U_\lambda(\tilde{x})>0.$$ This is a contradiction. This completes Step 1. $\mathbf{Step.2}$: Step 1 provides a starting point, from which we can now move the plane $T_\lambda$ to the right as long as $(\ref{011805})$ holds to its limiting position. Let$$\lambda_0=\{\lambda\preceq 0\mid U_\mu\geq0, V_\mu\geq 0, \forall x\in\Sigma_\mu\backslash\{0^\mu\},\mu\preceq\lambda\}.$$ By definition, $$U_{\lambda_0}(x), V_{\lambda_0}(x)\geq 0,\,\,\forall x\in \Sigma_{\lambda_0}\backslash\{0^{\lambda_0}\}.$$ (i): If $\lambda_0=0$, we can move $T_\lambda$ from $+\infty$ to the left and show that $$U_{\lambda_0}(x), V_{\lambda _0}(x)\preceq 0,\,\,x\in\Sigma_{\lambda_0}\backslash\{0^{\lambda_0}\},$$ with $\lambda_0=0$. We obtain $$U_0(x)=V_0(x)\equiv 0,\,\,x\in\Sigma_0.$$ For more general Kelvin transform, through a similar argument we can show that $$\lambda_0=x^0_1,$$ and $$U_{\lambda_0}(x), V_{\lambda_0}(x)\equiv 0,\,\,x\in\Sigma_{\lambda_0}.$$ Since the $x_1$ direction and $x^0$ can be chosen arbitrarily, we have actually shown that $\bar{u}$ and $\bar{v}$ is radially symmetric about any point in $R^n$. For any $x^1, x^2\in R^n$, let the midcenter be the point of Kelvin transform $$x^0=\frac{x^1+x^2}{2},$$ and $$y^1=\frac{x^1-x^0}{|x^1-x^0|^2}+x^0,\,\,y^2=\frac{x^2-x^0}{|x^2-x^0|^2}+x^0.$$ Then, $$\bar{u}(y^1)=\bar{u}(y^2),\,\,\bar{v}(y^1)=\bar{y^2}.$$ Thus, $$u(x^1)=u(x^2),\,\,v(x^1)=v(x^2).$$ Since $x^1, x^2$ is chosen arbitrarily, $u$ and $v$ must be constant. From (\ref{1.1}), we know $$(-\triangle)^{\alpha/2}u=0,$$ and $$(|x|^{\beta-n}\ast u^p)u^{p-1}>0,$$ a contradiction. Hence, $(u,v)=(0,0)$. (ii): If $\lambda_0<0$, there must be two cases. Case i: $$U_{\lambda_0}(x)=V_{\lambda_0}(x)\equiv 0,\,\,\forall x\in\Sigma_{\lambda_0}\backslash\{0^{\lambda_0}\}.$$ Indeed, we suppose there exists $\bar{x}$ such that $$U_{\lambda_0}(\bar{x})=\min_{\Sigma_{\lambda_0}}U_{\lambda_0}(x)=0.$$ Then it must be true that \begin{equation}\label{012601} U_{\lambda_0}(x)\equiv 0,\,\,\forall x\in\Sigma_{\lambda_0}. \end{equation} If not, \begin{eqnarray*} (-\triangle)^{\alpha/2}U_{\lambda_0}(\bar{x})=C_{n,\alpha}PV\int_{R^n}\frac{-U_{\lambda_0}(\bar{x})}{|\bar{x}-y|^{n+\alpha}}dy<0. \end{eqnarray*} On the other hand, \begin{eqnarray*} &&(-\triangle)^{\alpha/2}U_{\lambda_0}(\bar{x})\\ &=&\frac{1}{|\bar{x}^{\lambda_0}|^{\gamma}}\overline{v}(\bar{x}^{\lambda_0})\overline{u}^{p-1}(\bar{x}^{\lambda_0})-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x})\\ &\geq&\frac{1}{|\bar{x}^{\lambda_0}|^{\gamma}}\overline{v}(\bar{x}^{\lambda_0})\overline{u}^{p-1}(\bar{x}^{\lambda_0})-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^{\lambda_0})\\ &\geq&\frac{1}{|\bar{x}^{\lambda_0}|^{\gamma}}\overline{v}(\bar{x}^{\lambda_0})\overline{u}^{p-1}(\bar{x}^{\lambda_0})-\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^{\lambda_0})\\ &\geq&\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x}^{\lambda_0})V_{\lambda_0}(\bar{x})\\ &\geq&0, \end{eqnarray*} which is a contradiction. This proves (\ref{012601}). Since $$U_{\lambda_0}(x)=-U_{\lambda_0}(x^{\lambda_0}),$$ we have $$U_{\lambda_0}(x)\equiv 0,\,\,x\in R^n.$$ Then, \begin{eqnarray*} 0&=&(-\triangle)^{\alpha/2}U_{\lambda_0}(x)\\ &=&\frac{1}{|x^{\lambda_0}|^{\gamma}}\overline{v}(x^{\lambda_0})\overline{u}^{p-1}(x^{\lambda_0})-\frac{1}{|x|^{\gamma}}\overline{v}(x)\overline{u}^{p-1}(x)\\ &=&\frac{1}{|x^{\lambda_0}|^{\gamma}}\overline{v}(x^{\lambda_0})\overline{u}^{p-1}(x)-\frac{1}{|x|^{\gamma}}\overline{v}(x)\overline{u}^{p-1}(x), \end{eqnarray*} we can derive that $$\overline{v}(x^{\lambda_0})\preceq\overline{v}(x),\,\,x\in\Sigma_{\lambda_0}.$$ Combing this with the fact $$\overline{v}(x^{\lambda_0})\geq\overline{v}(x),\,\,x\in\Sigma_{\lambda_0}.$$ We can deduce $$V_{\lambda_0}(x)\equiv 0,\,\,x\in R^n.$$ From a similar argument, we can also have if $V_{\lambda_0}(x)=0$ somewhere, then $$U_{\lambda_0}(x)=V_{\lambda_0}(x)\equiv 0,\,\,x\in R^n.$$ Therefore, for all $x\in R^n$, \begin{eqnarray*} 0&=&(-\triangle)^{\alpha/2}U_{\lambda_0}(x)\\ &=&\frac{1}{|x^{\lambda_0}|^{\gamma}}\overline{v}(x^{\lambda_0})\overline{u}^{p-1}(x^{\lambda_0})-\frac{1}{|x|^{\gamma}}\overline{v}(x)\overline{u}^{p-1}(x), \end{eqnarray*} and \begin{eqnarray*} 0&=&(-\triangle)^{\beta/2}V_{\lambda_0}(x)\\ &=&\frac{1}{|x^{\lambda_0}|^{\gamma}}\overline{u}^{p}(x^{\lambda_0})-\frac{1}{|x|^{\gamma}}\overline{u}^{p}(x). \end{eqnarray*} We can easily deduce that $$\bar{u}(x)=\bar{v}(x)\equiv 0,\,\,x\in R^n.$$ Then $$u(x)=v(x)\equiv0,\,\,x\in R^n.$$ Case ii: $$U_{\lambda_0}(x), V_{\lambda_0}(x)> 0,\,\,\forall x\in\Sigma_{\lambda_0}\backslash\{0^{\lambda_0}\}.$$ We show that the plane $T_\lambda$ can be moved further right. To be more rigorous, there exists some $\varepsilon>0$, such that for any $\lambda\in(\lambda_0,\lambda_0+\varepsilon)$, we have $$U_\lambda(x), V_\lambda(x)\geq 0,\,\,x\in\Sigma_\lambda\backslash\{0^\lambda\},$$ This is a contradiction with the definition of $\lambda_0$, so this case will not happen. Indeed, first we claim that for $\lambda_0<0$ and $\varepsilon$ sufficiently small, $$U_{\lambda_0}(x), V_{\lambda_0}(x)\geq C>0,\,\,x\in B_\varepsilon(0^{\lambda_0})\backslash\{0^{\lambda_0}\},$$ this will be proved in Appendix. Then there exist $\delta>0$ small and a constant $C>0$ such that $$U_{\lambda_0}(x), V_{\lambda_0}(x)\geq C>0,\,\,x\in (\Sigma_{\lambda_0-\delta}\backslash\{0^{\lambda_0}\})\cap B_R(0).$$ Since $U_\lambda(x), V_\lambda(x)$ are continuous about $\lambda$, then $$U_{\lambda}(x), V_{\lambda}(x)\geq 0,\,\,x\in (\Sigma_{\lambda_0-\delta}\backslash\{0^{\lambda_0}\})\cap B_R(0).$$ Suppose $U_{\lambda}(x)<0,\,\,x\in \Sigma_{\lambda}\backslash\{0^{\lambda}\},$ then there must exist $\bar{x}$ such that $$U_{\lambda}(\bar{x})=\min_{\Sigma_\lambda}U_\lambda(x)<0.$$ From a similar argument in the proof of \emph{Decay at Infinity}, we have $$V_{\lambda}(\bar{x})<0,\,\,x\in\Sigma_{\lambda}\backslash\{0^{\lambda}\}.$$ Then there exists $\tilde{x}$ such that $$V_\lambda(\tilde{x})=\min_{\Sigma_\lambda}V_\lambda(x)<0.$$ By Theorem \ref{t1} (\emph{Decay at Infinity}), one of $\bar{x}$ and $\tilde{x}$ must be in $B_R(0)$. Without loss of generality, we assume $$|\bar{x}|<R.$$ Hence, $$\bar{x}\in(\Sigma_\lambda\backslash\Sigma_{\lambda_0-\delta})\cap B_R(0).$$ If $\tilde{x}\in(\Sigma_\lambda\backslash\Sigma_{\lambda_0-\delta})\cap B_R(0)$, then by (\ref{011902}), we have, $$(-\triangle)^{\alpha/2}U_\lambda(\bar{x})\preceq\frac{CU_\lambda(\bar{x})}{(\delta+\varepsilon)^\alpha}<0.$$ Similarly, $$(-\triangle)^{\beta/2}V_\lambda(\tilde{x})\preceq\frac{CV_\lambda(\tilde{x})}{(\delta+\varepsilon)^\beta}<0.$$ Then, \begin{eqnarray*} 0&\preceq& (-\triangle)^{\alpha/2}V_\lambda(\bar{x})+C_2(\bar{x})U_\lambda(\bar{x})+C_3(\bar{x})V_\lambda(\bar{x})\\ &\preceq&\frac{CU_\lambda(\bar{x})}{(\delta+\varepsilon)^\alpha}+C_2(\bar{x})U_\lambda(\bar{x})+C_3(\bar{x})V_\lambda(\tilde{x})\\ &\preceq&\frac{CU_\lambda(\bar{x})}{(\delta+\varepsilon)^\alpha}+C_2(\bar{x})U_\lambda(\bar{x})+C_3(\bar{x})V_\lambda(\tilde{x})\\ &\preceq&\frac{CU_\lambda(\bar{x})}{(\delta+\varepsilon)^\alpha}+C_2(\bar{x})U_\lambda(\bar{x})-C_3(\bar{x})C_1(\tilde{x})(\delta+\varepsilon)^\beta U_\lambda(\tilde{x})\\ &\preceq&\frac{CU_\lambda(\bar{x})}{(\delta+\varepsilon)^\alpha}+C_2(\bar{x})U_\lambda(\bar{x})-C_3(\bar{x})C_1(\tilde{x})(\delta+\varepsilon)^\beta U_\lambda(\bar{x})\\ &=&\frac{CU_\lambda(\bar{x})}{(\delta+\varepsilon)^\alpha}\{1-C_3(\bar{x})C_1(\tilde{x})(\delta+\varepsilon)^{\alpha+\beta}\}+C_2(\bar{x})U_\lambda(\bar{x}) \end{eqnarray*} Through an identical argument in Theorem \ref{t2} (\emph{Narrow Region Principle}), we have $$U_\lambda(x), V_\lambda(x)\geq 0,\,\,x\in (\Sigma_\lambda\backslash\Sigma_{\lambda_0-\delta})\cap B_R(0).$$ It implies that $\tilde{x}\in(\Sigma_\lambda\backslash\Sigma_{\lambda_0-\delta})\cap B_R(0)$ will not be happen. If $\tilde{x}\in B_R^c\cap\Sigma_\lambda$, then \begin{eqnarray*} 0&>&\frac{CU_\lambda(\bar{x})}{(\delta+\varepsilon)^\alpha}\\ &\geq&(-\triangle)^{\alpha/2}U_\lambda(\bar{x})\\ &\geq&\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x})V_\lambda(\bar{x})+(p-1)\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\bar{u}^{p-2}(\bar{x})U_\lambda(\bar{x})\\ &\geq&\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x})V_\lambda(\tilde{x})+(p-1)\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\bar{u}^{p-2}(\bar{x})U_\lambda(\bar{x}), \end{eqnarray*} and, \begin{eqnarray*} 0&>&\frac{CV_\lambda(\tilde{x})}{|\tilde{x}|^\beta}\\ &\geq&(-\triangle)^{\beta/2}U_\lambda(\tilde{x})\\ &\geq&p\frac{1}{|\tilde{x}|^{\gamma}}\overline{u}^{p-1}(\tilde{x})U_\lambda(\tilde{x})\\ &\geq&p\frac{1}{|\tilde{x}|^{\gamma}}\overline{u}^{p-1}(\tilde{x})U_\lambda(\bar{x}). \end{eqnarray*} Through a simple calculation, we have $$V_\lambda(\tilde{x})\geq|\tilde{x}|^{\beta-\gamma}\bar{u}^{p-1}(\tilde{x})U_\lambda(\bar{x}),$$ and, \begin{eqnarray*} U_\lambda(\bar{x})&\geq&(\delta+\varepsilon)^\alpha\big\{\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x})V_\lambda(\tilde{x})+(p-1)\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\bar{u}^{p-2}(\bar{x})U_\lambda(\bar{x})\big\}\\ &\geq&(\delta+\varepsilon)^\alpha\big\{\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x})|\tilde{x}|^{\beta-\gamma}\bar{u}^{p-1}(\tilde{x})U_\lambda(\bar{x})+(p-1)\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\bar{u}^{p-2}(\bar{x})U_\lambda(\bar{x})\big\}. \end{eqnarray*} Then, \begin{eqnarray}\label{012705}\nonumber 1&\preceq&(\delta+\varepsilon)^\alpha\big\{\frac{1}{|\bar{x}|^{\gamma}}\overline{u}^{p-1}(\bar{x})|\tilde{x}|^{\beta-\gamma}\bar{u}^{p-1}(\tilde{x})+(p-1)\frac{1}{|\bar{x}|^{\gamma}}\overline{v}(\bar{x})\bar{u}^{p-2}(\bar{x})\big\}\\ &=&(\delta+\varepsilon)^\alpha\big\{\frac{1}{|\bar{x}|^{\alpha+\beta}}u^{p-1}(\frac{\bar{x}}{|\bar{x}|^2})\frac{1}{|\tilde{x}|^\alpha}u^{p-1}(\frac{\tilde{x}}{|\tilde{x}|^2})+(p-1)\frac{1}{|\bar{x}|^{2\alpha}}v(\frac{\bar{x}}{|\bar{x}|^2})u^{p-2}(\frac{\bar{x}}{|\bar{x}|^2})\big\}. \end{eqnarray} For a fixed $\lambda_0<0$, when $\varepsilon$ is sufficiently small, we have $\lambda<\lambda+\varepsilon<\frac{\lambda_0}{2}$. Since $\lambda_0\in\Sigma_\lambda$, it deduces that $|\bar{x}|>-\frac{\lambda_0}{2}$. Notice that $|\tilde{x}|>R$, then $\frac{1}{|\bar{x}|^{\alpha+\beta}}u^{p-1}(\frac{\bar{x}}{|\bar{x}|^2})\frac{1}{|\tilde{x}|^\alpha}u^{p-1}(\frac{\tilde{x}}{|\tilde{x}|^2})+(p-1)\frac{1}{|\bar{x}|^{2\alpha}}v(\frac{\bar{x}}{|\bar{x}|^2})u^{p-2}(\frac{\bar{x}}{|\bar{x}|^2})$ is bounded. This shows that (\ref{012705}) must not be true for $\delta$ sufficiently small. This implies this case also will not happen. \subsubsection{Critical Case $p=\frac{n+\beta}{n-\alpha}$.} $\mathbf{Step.1}$: We show that when $\lambda$ sufficiently negative, \begin{eqnarray}\label{012801} U_\lambda(x), V_\lambda(x)\geq0, \,\forall x\in\Sigma_{\lambda}\setminus\{0^\lambda\}. \end{eqnarray} We claim that for $\lambda$ sufficiently negative, there exists a constant $C$ such that $$U_\lambda(x), V_\lambda(x)\geq C>0,\,\,x\in B_\varepsilon(0^\lambda)\backslash\{0^\lambda\},$$ we will prove it in Appendix. Hence, there must be a point $\bar{x}$ such that $$U_\lambda(\bar{x})=\min_{x\in\Sigma_\lambda }U_\lambda(x)<0.$$ Moreover, \begin{eqnarray*} &&(-\triangle)^{\alpha/2}U_\lambda(\bar{x})\\ &=&C_{n,\alpha} PV\int_{R^n}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy\\ &=&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy+C_{n,\alpha}PV\int_{\tilde{\Sigma}_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy\\ &=&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy+C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y^\lambda)}{|\bar{x}-y^\lambda|^{n+\alpha}}dy\\ &=&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy+C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})+U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy\\ &\preceq&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})-U_\lambda(y)}{|\bar{x}-y^\lambda|^{n+\alpha}}dy+C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{U_\lambda(\bar{x})+U_\lambda(y)}{|\bar{x}-y|^{n+\alpha}}dy\\ &=&C_{n,\alpha}PV\int_{\Sigma_\lambda}\frac{2U_\lambda(\bar{x})}{|\bar{x}-y^\lambda|^{n+\alpha}}dy. \end{eqnarray*} From a similar argument in (\ref{1.8}), we have \begin{equation}\label{012802} (-\triangle)^{\alpha/2}U_\lambda(\bar{x})\preceq\frac{CU_\lambda(\bar{x})}{|\bar{x}|^\alpha}<0. \end{equation} We claim that \begin{equation}\label{012803} V_\lambda(\bar{x})<0. \end{equation} Indeed, if not, $V_\lambda(\bar{x})\geq 0$. From (\ref{011804}), we have \begin{eqnarray*} &&(-\triangle)^{\alpha/2}U_\lambda(\bar{x})\\ &=&\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x})\\ &\geq&\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)\\ &\geq&\overline{u}^{p-1}(\bar{x}^\lambda)V_\lambda(\bar{x})\\ &\geq&0. \end{eqnarray*} This is a contradiction with (\ref{012802}). Then (\ref{012803}) holds. And from (\ref{012803}), we know there exists $\tilde{x}$ such that $$V_\lambda({\tilde{x}})=\min_{\Sigma_\lambda}V_\lambda(x)<0.$$ Similar to (\ref{012803}), we can obtain $$U_\lambda(\tilde{x})<0.$$ Therefore, \begin{eqnarray}\nonumber &&(-\triangle)^{\alpha/2}U_\lambda(\bar{x})\\ \nonumber &=&\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x})\\ \nonumber &=&\overline{v}(\bar{x}^\lambda)\overline{u}^{p-1}(\bar{x}^\lambda)-\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)+\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x}^\lambda)-\overline{v}(\bar{x})\overline{u}^{p-1}(\bar{x})\\ \nonumber &=&\overline{u}^{p-1}(\bar{x}^\lambda)V_\lambda(\bar{x})+(p-1)\overline{v}(\bar{x})\xi^{p-2}U_\lambda(\bar{x}),\,\,\,\,\xi\in[\overline{u}(\bar{x}^\lambda),\overline{u}(\bar{x})]\\ \label{012805} &\geq&\overline{u}^{p-1}(\bar{x})V_\lambda(\bar{x})+(p-1)\overline{v}(\bar{x})\bar{u}(\bar{x})^{p-2}U_\lambda(\bar{x}), \end{eqnarray} and \begin{eqnarray*} &&(-\triangle)^{\beta/2}V_\lambda(\tilde{x})\\ &=&\overline{u}^{p}(\tilde{x}^\lambda)-\overline{u}^{p}(\tilde{x})\\ &=&p\frac{1}{|\tilde{x}|^{\gamma}}\eta^{p-1}U_\lambda(\tilde{x}),\,\,\,\eta\in[\overline{u}(\tilde{x}^\lambda),\overline{u}(\tilde{x})]\\ &\geq&p\overline{u}^{p-1}(\tilde{x})U_\lambda(\tilde{x}). \end{eqnarray*} Let \begin{eqnarray*} C_1(\tilde{x})&=&p\overline{u}^{p-1}(\tilde{x})\\ &\sim&\frac{1}{|\tilde{x}|^{(n-\alpha)(p-1)}}\\ &\sim&\frac{1}{|\tilde{x}|^{\alpha+\beta}},\,\,|\tilde{x}| \,\,\mbox{large \,\,enough}, \end{eqnarray*} \begin{eqnarray*} C_2(\bar{x})&=&(p-1)\overline{v}(\bar{x})\bar{u}(\bar{x})^{p-2}\\ &\sim&\frac{1}{|\bar{x}|^{n-\beta}}\frac{1}{|\bar{x}|^{(n-\alpha)(p-2)}}\\ &\sim&\frac{1}{|\bar{x}|^{2\alpha}},\,\,|\bar{x}| \,\,\mbox{large \,\,enough}, \end{eqnarray*} and \begin{eqnarray*} C_3(\bar{x})&=&\overline{u}^{p-1}(\bar{x})\\ &\sim&\frac{1}{|\bar{x}|^{(n-\alpha)(p-1)}}\\ &\sim&\frac{1}{|\bar{x}|^{\alpha+\beta}},\,\,|\bar{x}| \,\,\mbox{large \,\,enough}. \end{eqnarray*} Then by Theorem \ref{t1}, if $\lambda$ sufficiently negative, there must be one of $U_\lambda(x)$ and $V_\lambda(x)$ positive in $\Sigma_\lambda\backslash\{0^\lambda\}.$ Without loss of generality, we assume $$U_\lambda(x)\geq0,\,\,x\in\Sigma_\lambda\backslash\{0^\lambda\}.$$ And we claim that $$V_\lambda(x)\geq0,\,\,x\in\Sigma_\lambda\backslash\{0^\lambda\}.$$ If not, there exists $\tilde{x}$ such that $$V_\lambda(\tilde{x})=\min_{\Sigma_\lambda}V_\lambda(x)<0.$$ Then from a similar argument, we can show $$0>\frac{CV_\lambda(\tilde{x})}{|\tilde{x}|^\beta}\geq(-\triangle)^{\beta/2}V_\lambda(\tilde{x})\geq p\bar{x}^{p-1}(\tilde{x})U_\lambda(\tilde{x})>0.$$ This is a contradiction. This completes Step 1. $\mathbf{Step.2}$: Step 1 provides a starting point, from which we can now move the plane $T_\lambda$ to the right as long as $(\ref{012801})$ holds to its limiting position. Let$$\lambda_0=\{\lambda\leq0\mid U_\mu\geq0, V_\mu\geq 0, \forall x\in\Sigma_\mu\backslash\{0^\mu\},\mu\preceq\lambda\}.$$ By definition, $$U_{\lambda_0}(x), V_{\lambda_0}(x)\geq 0,\,\,\forall x\in \Sigma_{\lambda_0}\backslash\{0^{\lambda_0}\}.$$ $\mathbf{Case. i}$: $\lambda_0<0$. Similar to the subcritical case, one can show that $$U_{\lambda_0}(x), V_{\lambda_0}(x)\geq 0, \,x\in \Sigma_{\lambda_0}\backslash\{0^{\lambda_0}\}.$$ It follows that $x^0$ is not a singular point of $\bar{u}$ and $\bar{v}$ and hence $$u(x)=O(\frac{1}{|x|^{n-\alpha}}), v(x)=O(\frac{1}{|x|^{n-\beta}}),\,\,|x|\rightarrow\infty$$ This enables us to apply the method of moving plane to $u$ and $v$ directly and show that $u$ and $v$ are symmetric about some point in $R^n$. $\mathbf{Case. ii}$: $\lambda_0=0$. Then by moving the planes from $+\infty$, we derive that $\bar{u}$ and $\bar{v}$ are symmetric about the origin, and so do $u$ and $v$. In any case, $u$ and $v$ are symmetric about some point in $R^n$. This completes the proof. \section{Appendices} \begin{lemma}\label{lem5.1} For $\lambda$ negative large, there exists a constant $C>0$, such that \begin{equation}\label{05.1} U_\lambda(x), V_\lambda(x)\geq C>0,\, x\in B_\varepsilon(0^\lambda)\backslash\{0^\lambda\}. \end{equation} \end{lemma} \begin{proof} For $x\in\Sigma_\lambda$, as $|x|\rightarrow -\infty$, it is easy to see that \begin{equation}\label{05.2} \overline{u}(x)\rightarrow 0. \end{equation} To prove (\ref{05.1}), it is sufficient to show $$\overline{u}_\lambda(x)\geq C>0, \,x\in B_\varepsilon(0^\lambda)\backslash\{0^\lambda\}.$$ Or equivalently, $$\overline{u}(x)\geq C>0, \,x\in B_\varepsilon(0)\backslash\{0\}.$$ Let $\eta$ be a smooth cut-off function such that $\eta\in[0,1]$ in $R^n$, $\mbox{supp}~ \eta\subset B_2$ and $\eta\equiv 1$ in $B_1$. Let $$(-\triangle)^{\alpha/2}\phi(x)=\eta(x)v(x)u^{p-1}(x).$$ Then, $$\phi(x)=C_{n,-\alpha}\int_{R^n}\frac{\eta(y)v(y)u^{p-1}(y)}{|x-y|^{n-\alpha}}dy =C_{n,-\alpha}\int_{B_2(0)}\frac{\eta(y)v(y)u^{p-1}(y)}{|x-y|^{n-\alpha}}dy.$$ It is trivial for $|x|$ sufficiently large, \begin{equation}\label{05.3} \phi(x)\sim\frac{1}{|x|^{n-\alpha}}. \end{equation} Since \begin{eqnarray}\label{05.4}\left\{ \begin{array}{lll} (-\triangle)^{\alpha/2}(u-\phi)\geq 0,~~~&x\in B_R,\\ (u-\phi)(x)\geq 0,~~~ &x\in B_R^c, \end{array} \right. \end{eqnarray} by the maximum principle, we have $$(u-\phi)(x)\geq 0,\,x\in B_R,$$ thus $$(u-\phi)(x)\geq0,\,x\in R^n.$$ For $|x|$ sufficiently large, from (\ref{05.3}), one can see that for some constant $C>0$, \begin{equation}\label{05.5} u(x)\geq\frac{C}{|x|^{n-\alpha}}. \end{equation} Hence for $|x|$ small $$u(\frac{x}{|x|^2})\geq C|x|^{n-\alpha},$$ and $$\overline{u}(x)=\frac{1}{|x|^{n-\alpha}}u(\frac{x}{|x|^2})\geq C.$$ Together with (\ref{05.2}), it yields that \begin{equation}\label{05.6} U_\lambda(x)\geq\frac{C}{2}>0, \,x\in B_\varepsilon(0^\lambda)\setminus\{0^\lambda\}. \end{equation} Through an identical argument, one can show that (\ref{05.6}) holds for $V_\lambda(x)$ as well. \end{proof} \begin{lemma}\label{5.3} Let $(u, v)$ be a pair of nonnegative solutions of (\ref{011802}), and $\bar{u}$, $\bar{v}$ be the Kelvin transform of $u$ and $v$, then $\bar{u}$ and $\bar{v}$ also satisfy \begin{eqnarray*}\left\{ \begin{array}{lll} \bar{u}(x)=\int_{R^n}\frac{|y|^{-\gamma}\bar{v}(y)\bar{u}^{p-1}(y)}{|x-y|^{n-\alpha}}dy\\ \bar{v}(x)=\int_{R^n}\frac{|y|^{-\gamma}\bar{u}^p(y)}{|x-y|^{n-\beta}}dy \end{array} \right. \end{eqnarray*} and vice versa. \end{lemma} \begin{proof} It is easy to see $(\bar{u}, \bar{v})$ is a pair of nonnegative solutions to (\ref{011803}). From (\ref{011801}), we have $$ \bar{v}(x)=\int_{R^n}\frac{|y|^{-\gamma}\bar{u}^p(y)}{|x-y|^{n-\beta}}dy.$$ Then, we only need to show $$ \bar{u}(x)=\int_{R^n}\frac{|y|^{-\gamma}\bar{v}(y)\bar{u}^{p-1}(y)}{|x-y|^{n-\alpha}}dy.$$ We first show that \begin{eqnarray}\label{02.1} \bar{u}(x)=c_1+\int_{R^n}\frac{|y|^{-\gamma}\bar{v}(y)\bar{u}^{p-1}(y)}{|x-y|^{n-\alpha}}dy. \end{eqnarray} Let \begin{eqnarray}\label{02.2} &\bar{u}_R(x)=\int_{B_R(0)}G_R(x,y)|y|^{-\gamma}\bar{v}(y)\bar{u}^{p-1}(y)dy, \end{eqnarray} where $G_R(x,y)$ is the Green's function of fractional Laplacian on $B_R(0)$. It is easy to see that \begin{eqnarray}\label{02.3} \left\{ \begin{array}{ll} (-\triangle)^{\alpha/2}\bar{u}_R(x)=|x|^{-\gamma}\bar{v}(x)\bar{u}^{p-1}(x),\,\,&\mbox{in}\,\,B_R(0)\backslash\{0\},\\ \bar{u}_R(x)=0,\,\,\,\,\,&\mbox{on}\,\,B_R^c(0). \end{array} \right. \end{eqnarray} Let $\varphi_R(x)=\bar{u}(x)-\bar{u}_R(x)$, from (\ref{011803}) and (\ref{02.3}), we have \begin{eqnarray*} \left\{ \begin{array}{ll} (-\triangle)^{\alpha/2}\varphi_R(x)=0,\,\,&\mbox{in}\,\,B_R(0),\\ \varphi_R(x)\geq 0,\,\,\,\,\,&\mbox{on}\,\,B_R^c(0). \end{array} \right. \end{eqnarray*} By the Maximum Principle, we derive \begin{eqnarray}\label{02.4} \varphi_R(x)\geq 0,\,\,\,x\in R^n. \end{eqnarray} Therefore, when $R\rightarrow \infty$, \begin{eqnarray}\label{02.5} \bar{u}_R(x)\rightarrow \tilde{u}(x)=\int_{R^n}\frac{|y|^{-\gamma}\bar{v}(y)\bar{u}^{p-1}(y)}{|x-y|^{n-\alpha}}dy, \end{eqnarray} Moreover, \begin{eqnarray}\label{02.6} (-\triangle)^{\alpha/2}\tilde{u}(x)=|x|^{-\gamma}\bar{v}(x)\bar{u}^{p-1}(x),\,\,\,&x\in R^n, \end{eqnarray} Now let $\Phi(x)=\bar{u}(x)-\tilde{u}(x)$. From (\ref{1.1}) and (\ref{02.6}), we have \begin{eqnarray*} \left\{ \begin{array}{ll} (-\triangle)^{\alpha/2}\Phi(x)=0,\,\,\,&x\in R^n,\\ \Phi(x)\geq 0,\,\,\,\,&x\in R^n. \end{array} \right. \end{eqnarray*} From Proposition 2 in \cite{ZCCY}, we have $$\Phi(x)=c_1.$$ Thus we proved (\ref{02.1}). Next, we will show that $c_1=0$. If $c_1>0$, then from (\ref{02.1}) and the fact $p\geq\frac{n}{n-\alpha}$, we have $$ v(x)=\int_{R^n}\frac{|y|^{-\gamma}u^{p}(y)}{|x-y|^{n-\beta}}dy\geq \int_{R^n}\frac{c_1^p|y|^{-\gamma}}{|x-y|^{n-\beta}}dy=\infty. $$ But it is impossible. Hence $c_1=0$. Therefore, \begin{eqnarray*} \left\{ \begin{array}{ll} &\bar{u}(x)=\int_{R^n}\frac{|y|^{-\gamma}\bar{v}(y)\bar{u}^{p-1}(y)}{|x-y|^{n-\alpha}}dy,\\ &\bar{v}(x)=\int_{R^n}\frac{|y|^{-\gamma}\bar{u}^p(y)}{|x-y|^{n-\beta}}dy. \end{array} \right. \end{eqnarray*} We complete our proof. \end{proof} \begin{lemma}\label{lem5.2} For $\lambda_0<0$, if either of $U_{\lambda_0}$, $V_{\lambda_0}$ is not identically 0, then there exist some constant $C$ and $\varepsilon>0$ small such that $$U_{\lambda_0}(x), V_{\lambda_0}(x)\geq C>0, \,\,x\in B_\varepsilon(0^{\lambda_0})\setminus\{0^{\lambda_0}\}.$$ \end{lemma} \begin{proof} From Lemma \ref{5.3}, we have the integral equation \begin{eqnarray*} V_{\lambda_0}(x)&=& \overline{v}_{\lambda_0}(x)-\overline{v}(x) \\ &=& C_{n,\beta}\int_{\Sigma_{\lambda_0}}(|y^{\lambda_0}|^{-\gamma}\bar{u}^{p}(y^{\lambda_0})-|y|^{-\gamma}\bar{u}^{p}(y)) (\frac{1}{|x-y|^{n+\beta}}-\frac{1}{|x-y^{\lambda_0}|^{n+\beta}})dy \\ &\geq& C_{n,\beta}\int_{\Sigma_{\lambda_0}}\frac{\overline{u}_{\lambda_0}^{p}(y)-\overline{u}^{p}(y)}{|y|^\gamma} \cdot(\frac{1}{|x-y|^{n+\beta}}-\frac{1}{|x-y^{\lambda_0}|^{n+\beta}})dy. \end{eqnarray*} Since $$U_{\lambda_0}(x)\not\equiv 0,\,x\in\Sigma_{\lambda_0},$$ there exists some $x_0$ such that $U_{\lambda_0}(x_0)>0.$ Thus, for some $\delta>0$ small, it holds that $$\overline{u}_{\lambda_0}^p(y)-\overline{u}^p(y)\geq C>0, \,y\in B_\delta(x_0).$$ Therefore, \begin{equation}\label{5.7} V_{\lambda_0}(x)\geq\int_{B_\delta(x_0)}Cdy\geq C>0. \end{equation} In a same way, one can show that $U_{\lambda_0}(x)$ also satisfies (\ref{5.7}). \end{proof} \noindent{\bf Acknowledgement} The research was supported by NSFC(NO.11571176). The authors would like to express sincere thanks to the anonymous referee for his/her carefully reading the manuscript and valuable comments and suggestions.

https://arxiv.org/abs/math/0606562

An Isomonodromy Cluster of Two Regular Singularities

We consider a linear $2\times2$ matrix ODE with two coalescing regular singularities. This coalescence is restricted with an isomonodromy condition with respect to the distance between the merging singularities in a way consistent with the ODE. In particular, a zero-distance limit for the ODE exists. The monodromy group of the limiting ODE is calculated in terms of the original one. This coalescing process generates a limit for the corresponding nonlinear systems of isomonodromy deformations. In our main example the latter limit reads as $P_6\to P_5$, where $P_n$ is the $n$-th Painlevé equation. We also discuss some general problems which arise while studying the above-mentioned limits for the Painlevé equations.

\section{Prelude} \label{sec:prelude} This work appeared first as Sfb 288 Preprint No.~149 (Teschnishe Universit\"at, Berlin) in December 1994. It was distributed to all leading mathematical institutions and many researchers. It was submitted to Comm. Math. Phys. (November 1994), but was not accepted as "not interesting for the readership" of that Journal. The copy of this preprint in various formats can be downloaded from CERN Document Server\footnote[1]{http://preprints.cern.ch/cgi-bin/ setlink?base=preprint\&categ=.\&id=SCAN-9501196}. During the past (more than) 11 years I presented this work in a number of talks on various conferences, colloquium talks, and many private discussions. During this period appeared also a number of works by different authors that discuss questions closely related with this work. A review of those works actually requires considerable space and time. So, I was forced to avoid this issue. The only changes I made in the bibliography is the inclusion of the brief announcement~\cite{IKF} of the works~\cite{FIK1, FIK2} and also an appropriate renumbering of the references. In spite of the development that have been achieved during the past decade, the main results obtained in my preprint were not reproduced. Moreover, the intensive development of the Random Matrix Theory and related topics in the theory of Orthogonal Polynomials reveals that various double scaling limits of the Painlev\'e equations, one of those we study in this work, play a significant role in various questions of these theories and applications. So, I believe that the publication of this work will be helpful as presenting an important property of the nonlinear special function - the sixth Painlev\'e equation ($P_6$) and for future studies of isomonodromy deformations. I decided not to make any changes into the mathematical content of this work, just correction of English and misprints in formulae. Since it past so much time from the date the work was written, it is important to mention that the discretization procedure of the transition limit that studied in this work was motivated by the author's joint work with Alexander Its and Athanasis Fokas on matrix models and orthogonal polynomials\cite{FIK1, FIK2}. I am very grateful to the Guest Editors of this volume, Nalini Joshi and Frank Nijhoff, for giving me an opportunity to publish this work. I am also indebt to Fedor Andreev who typed the original version of this manuscript\footnotemark[1] in \AmS-\TeX\; and Arthur Vartanian for helping me in correction of English. \section{Introduction} \label{sec:introduction} By the isomonodromy cluster of two regular singularities we mean two points in the complex plane ${\mathbb C}$ with respective positions $\lambda_i(t)$ $(i=0,1)$, considered as given functions of the parameter $t$ defined in a neighborhood ${\cal O}(t_0)\subset\mathbb C$ (or $\mathbb R$) of some fixed point $t_0\in\mathbb C$ (or $\mathbb R$), satisfying the following conditions: (1) $\lambda_0(t_0)=\lambda_1(t_0)$; (2) $\lambda_0(t)\neq\lambda_1(t)$ for $t\in{\cal O}(t_0)\setminus t_0$; and (3) they are the simple poles of the linear ODE \begin{equation} \label{eq:cluster-definition} \frac{d}{d\lambda}\Psi=\left(\frac{A_0(t)}{\lambda-\lambda_0(t)}+ \frac{A_1(t)}{\lambda-\lambda_1(t)}+ A(\lambda,t)\right)\Psi, \end{equation} where $A_0(t), A_1(t), A(\lambda, t)\in\text{Mat}(n,\mathbb C)$ are analytic functions of $t$, and $A(\lambda, t)$ is a rational function of $\lambda$. It is also assumed that there exists a fundamental solution of Eq.~(\ref{eq:cluster-definition}) with the manifold of monodromy data\footnote[2]{The manifold of monodromy data is defined in Section~\ref{sec:2}} independent of $t$ ({\it hard isomonodromy condition}).\\ Thus Eq.~(\ref{eq:cluster-definition}) is equipped with a system of nonlinear ODEs (with respect to $t$) governing isomonodromy deformations of its coefficients. For the simplest nontrivial backgrounds (the functions $A(\lambda,t)$), the systems of isomonodromy deformations can be reduced to classical Painlev\'e equations; in particular, in this work we consider the fifth Painlev\'e equation $P_5$ ($n=2$, $\lambda-0=0$, $\lambda_1=t$, $A(\lambda,t)={\rm const}\cdot\sigma_3\neq0$) and the sixth one $P_6$ ($n=2$, $\lambda_0=0$, $\lambda_1=t$, $A(\lambda, t)=A(t)/(\lambda-1)$, $A_0(t)+A_1(t)+A(t)={\rm const}\cdot\sigma_3\neq0$). The main idea in studying the cluster system~(\ref{eq:cluster-definition}), equipped with the isomonodromy condition, is to substitute into Eq.~(\ref{eq:cluster-definition}), instead of the cluster entries, some singularity of a regular or irregular type. Carrying out this procedure in accordance with the isomonodromy condition, we obtain, instead of (\ref{eq:cluster-definition}), a new equation, whose fundamental solution solution is denoted by $\Psi_{\rm new}$, and a novel system of isomonodromy deformations with respect to a parameter which is proportional to $t-t_0$. As a result we obtain a formal limit (or formal asymptotics) as $t\to t_0$ of the initial Eq.~(\ref{eq:cluster-definition}) to a new one, and of the initial system of isomonodromy deformations to a novel one. To translate this formal limit into the nonformal result, i.e., to understand the solutions of the new system of isomonodromy deformations as defining the master term of asymptotic expansions as $t\to t_0$ for the solutions of the initial system of isomonodromy deformations, one has to solve the following nontrivial problem: to calculate the manifold of monodromy data for $\Psi_{\rm new}$ in terms of the monodromy manifold for $\Psi$. This is the problem addressed here by taking the background in Eq.~(\ref{eq:cluster-definition}) corresponding to $P_6$. The latter result is easy to generalize to an arbitrary number of regular singularities. It was M. Jimbo~\cite{J} who considered isomonodromy cluster of two regular singularities for the linear ODEs associated with $P_5$ and $P_6$. In accordance with the aforementioned, Jimbo's work can be interpreted as follows: Jimbo considered Eq.~(\ref{eq:cluster-definition}) with $P_5$ and $P_6$ backgrounds and substituted for the cluster terms a regular singularity. The resulting ODE for $\Psi_{\rm new}$, by means of simple gauge and scaling transformations, can be reduced to matrix versions of the standard ODEs for the Gauss hypergeometric function in the case of $P_6$, and the Whittaker confluent hypergeometric function in the case of $P_5$. From this, one observes: \begin{enumerate} \item The coefficients of the standard ODEs, which are constant (with respect to $t$) parameters, define the manifolds of monodromy data for the Gauss hypergeometric and Whittaker functions and, thereby, define the manifold of monodromy data for $\Psi_{\rm new}$; \item The coefficients of the standard ODEs also define coefficients of the ODE for $\Psi_{\rm new}$, the latter coefficients, which define solutions of the new system of isomonodromy deformations, asymptotically, as $t\to0$ behave as linear combinations of power-like functions, $t^\alpha$. \end{enumerate} Because the deformations are isomonodromic, it is easy to relate the monodromy manifolds of the initial function $\Psi$ with the monodromy manifolds of the standard equations. Thus, having done items 1 and 2 for each Painlev\'e equation, $P_5$ and $P_6$, Jimbo was able to parameterize the small-$t$ asymptotics for $P_5$ and $P_6$ by points of the corresponding manifolds of monodromy data; for $P_6$, i.e., when Eq.~(\ref{eq:cluster-definition}) has, along with the cluster $\{0,t\}$, two additional singularities at $1$ and $\infty$, Jimbo considered the group of fractional-linear transformations of $\mathbb C$ which interchange the cluster with the singularities $1$ and $\infty$. Having found the action of this group on the manifold of monodromy data, he obtained connection formulae for asymptotic expansions of general solutions of $P_6$ as $t\to0$, $t\to1$, and $t\to\infty$. Subsequently, McCoy and Tang~\cite{MT1,MT2,MT3} used Jimbo's monodromy parametrization for small-$t$ asymptotics of $P_5$ to connect them with the large-$t$ asymptotic expansions for the same equation. Like Jimbo~\cite{J}, we consider here the cluster on the $P_6$ background; however, unlike the {\it hard} isomonodromy condition used by Jimbo, we introduce a {\it weak} one, i.e., the deformations which preserve the generators of the monodromy group, forming an essential part of the manifold of monodromy data, rather than the whole manifold as for the hard isomonodromy deformations. As we see later, the condition that isomonodromy deformations are weak but not hard implies their discretization. At the same time the weak isomonodromy condition gives us an opportunity to substitute for the cluster not only the regular singularity (as in the Jimbo case) but also an irregular one. In this work we consider two different clusters on the $P_6$ background: the first one described by an irregular singularity and the other by a regular one. Even the ``regular'' case studied here differs from that in \cite{J}; in particular, it is related with a formal limit $P_6\to P_5$, the``irregular'' case leads to another formal limit $P_6\to P_5$. The main goal of this paper is to give a proper asymptotic interpretation of the formal limits in the manner explained above. We will study the first limit passage of the following oriented graph, \begin{figure}[ht] \begin{center} \begin{picture}(260,65) \put(30,35){$P_6$} \put(45,38){\vector(1,0){30}} \put(80,35){$P_5$} \put(93,42){\vector(4,3){30}} \put(93,34){\vector(4,-3){30}} \put(125,62){$P_4$} \put(125,07){$P_3$} \put(138,64){\vector(4,-3){30}} \put(138,11){\vector(4,3){30}} \put(170,35){$P_2$} \put(185,38){\vector(1,0){30}} \put(220,35){$P_1$} \end{picture} \end{center} \caption{The Painlev\'e - Okamoto degeneration scheme} \label{fig:1} \end{figure} which represents the successive degeneration of the Painlev\'e equations ($P_n$). This degeneration scheme was obtained by Painlev\'e~\cite{P}, who showed that by substituting into the canonical Painlev\'e equations a parameter $\epsilon$ and transforming the subsequent equations via appropriate $\epsilon\to0$ limits, simpler Painlev\'e equations can be obtained\footnote[3]{The scheme was extensively studied by Okamoto, who also presented it as the graph on Figure 1.}. Okamoto~\cite{O} pointed out that these formal limits can be viewed as infinitesimal canonical transformations of the Hamiltonian systems associated with the Painlev\'e equations; thus he equipped the scheme of Figure 1 with the analogous scheme for Hamiltonians. Furthermore, Okamoto associated with this step-by-step degeneration process a step-by-step confluence scheme for certain scalar linear ODEs, whose isomonodromy deformations are governed by the Painlev\'e equations. These scalar equations are known to be intimately connected with matrix equations of the form~(\ref{eq:cluster-definition}). The degeneration scheme for the linear ODEs can be formulated in a manner analogous to that shown in Figure 1, in terms of a special symbol which represents the corresponding transformation of the Poincar\'e ranks of singularities of the linear ODEs. The authors works~\cite{K1,K2,KK} in conjunction with the present one can be understood as the isomonodromy regularization of the degeneration scheme of Figure~\ref{fig:1}. In addition to the scheme discussed above, our regularization suggests the following scheme of transformation for the manifolds of monodromy data of the associated linear matrix ODEs: \begin{figure}[ht] \begin{center} \begin{picture}(260,65) \put(19,35){${\cal M}_6$} \put(39,38){\vector(1,0){30}} \put(74,35){${\cal M}_5$} \put(93,42){\vector(4,3){30}} \put(93,34){\vector(4,-3){30}} \put(125,62){${\cal M}_4$} \put(125,07){${\cal M}_3$} \put(143,64){\vector(4,-3){30}} \put(143,11){\vector(4,3){30}} \put(174,35){${\cal M}_2$} \put(193,38){\vector(1,0){30}} \put(228,35){${\cal M}_1$} \put(110,-10){Figure 2} \end{picture} \end{center} \label{fig:2} \end{figure} Let us now discuss the schemes in Figures~\ref{fig:1} and \ref{fig:2} in more detail. Each oriented edge in the Figures means the existence of a corresponding limit passage; in fact, there could also exist some other limits; for example, the limit $P_6\to P_5$ ($t\to1$), found by Painlev\'e~\cite{P}, differs from both of ours (in which $t\to0$). We call the formal limits equivalent if the diagram in Figure~3 is commutative, where the $x$-arrows denote the action of the transformation groups on the solutions of the Painlev\'e equations (for $n=6$, $m=5$, see~\cite{O1,O2}) or the manifolds of monodromy data, the $y$-arrows denote the same formal limits as on Figures~\ref{fig:1} and \ref{fig:2}, respectively, and where the $z$-arrows ``enumerate'' the solutions by the points of the manifolds of monodromy data, i.e., they ``point out'' on solutions of the corresponding direct monodromy problems (see Section~\ref{sec:2}). \begin{figure}[ht] \begin{center} \begin{picture}(100,85) \put(28,65){$P_n$} \put(42,68){\vector(1,0){29}} \put(74,65){$P_n$} \put(24,25){${\cal M}_n$} \put(42,28){\vector(1,0){29}} \put(74,25){${\cal M}_n$} \put(30,63){\vector(0,-1){28}} \put(80,63){\vector(0,-1){28}} \put(08,53){$P_m$} \put(22,56){\vector(1,0){29}} \put(54,53){$P_m$} \put(03,13){${\cal M}_m$} \put(22,16){\vector(1,0){29}} \put(54,13){${\cal M}_m$} \put(10,51){\vector(0,-1){28}} \put(60,51){\vector(0,-1){28}} \put(26,64){\vector(-3,-2){10}} \put(28,24){\vector(-3,-2){10}} \put(78,24){\vector(-3,-2){10}} \put(75,64){\vector(-3,-2){10}} \put(10,-10){Figure 3} \end{picture} \hspace{4cm} \begin{picture}(100,85) \put(28,65){$P_n$} \put(42,68){\vector(1,0){29}} \put(74,65){$P_k$} \put(24,25){${\cal M}_n$} \put(42,28){\vector(1,0){29}} \put(74,25){${\cal M}_k$} \put(30,63){\vector(0,-1){28}} \put(80,63){\vector(0,-1){28}} \put(08,53){$P_m$} \put(22,56){\vector(1,0){29}} \put(54,53){$P_l$} \put(03,13){${\cal M}_m$} \put(22,16){\vector(1,0){29}} \put(54,13){${\cal M}_l$} \put(10,51){\vector(0,-1){28}} \put(60,51){\vector(0,-1){28}} \put(26,64){\vector(-3,-2){10}} \put(28,24){\vector(-3,-2){10}} \put(78,24){\vector(-3,-2){10}} \put(75,64){\vector(-3,-2){10}} \put(10,-10){Figure 4} \end{picture} \end{center} \end{figure} Note that if the transformation $P_n\to P_n$ acts not only between the initial Painlev\'e equations but also during the entire limiting process, i.e., for arbitrary $\epsilon\to0$\footnote[4]{More precisely, if we denote by $\varepsilon$ and $\varepsilon_1$ small parameters in two copies of $P_n$, then the transformation maps $\varepsilon\to\varepsilon_1$, such that the functions $\varepsilon(\varepsilon_1)\to0$ and $\varepsilon_1(\varepsilon)\to0$, when $\varepsilon_1$ or, respectively, $\varepsilon\to0$.}, then the entire ${\cal M}$-plane of Figure 3 is not needed , since this commutativity is valid automatically. In general, there are some transformations of the Painlev\'e equations which do not have this property, i.e., they cannot be extended for arbitrary $\varepsilon$, we present the diagram in extended form. Although, as explained above, the limits $P_6\to P_5$ that we study in this work have different asymptotic senses, they appear to be equivalent under the above definition; moreover, all limits $P_6\to P_5$ known as at the time of the presents are equivalent. Therefore, the following question arises naturally: {\it are there an $n$ and $m$ such that nonequivalent formal limits $P_n\to P_m$ exist?} There are some Painlev\'e equations which are known to be equivalent; for example, $P_5'$ ($P_5$ with $\delta_5=0$, see (\ref{eq:P5})) is equivalent to $P_3$ \cite{G} and $P_{34}$ is equivalent to $P_2$ \cite{I}. This equivalence of the equations means that there is a mapping which is invertible on transcendental solutions and birationally dependent of them and their derivatives. Along with the limit $P_4\to P_2$ (see Figure 1), the formal limit $P_4\to P_{34}$ was found in \cite{K1}. Are these limits ``substantially'' different? \\ To formulate the general notion of equivalence of the formal limits of different Painlev\'e equations, i.e., the equivalence of the limits $P_n\to P_m$ and $P_k\to P_l$, we consider Figure~4. In this figure the upper $x$-arrows denote the birational mappings (which are assumed to exist). The mappings ${\cal M}_n\to{\cal M}_k$ and ${\cal M}_m\to{\cal M}_l$ are uniquely defined by demanding commutativity for the $xz$-walls of the cube. The $y$-arrows denote the equivalence classes of the limits defined earlier via Figure 3. The $z$-arrows have the same meaning as that in Figure 3. {\it The limits $P_n\to P_m$ and $P_k\to P_l$ are said to be equivalent if the diagram in Figure~{\rm4} is commutative. Are there any nonequivalent limits for pairwise equivalent equations?} In particular, are the limits $P_4\to P_2$ and $P_4 \to P_{34}$ equivalent? Although according to \cite{JK} these limits show different actions on local expansions for solutions of $P_4$, they can appear to be equivalent under the above definition. The cluster point of view makes clear that there must exist limits other than those shown in Figure~\ref{fig:1}; e.g., along with the chain of successive limits $P_4\to P_2\to P_1$, there exist direct ones \cite{K1,K2}, which look, at first glance, to be different from the limit obtained via the successive procedure. Thus we can expect the existence of a more complicated limit structure (than that shown on Figure~\ref{fig:1}), which can be presented as the following oriented spatial graph: \begin{figure}[ht] \begin{center} \begin{picture}(270,140) \put(30,70){\circle*{3}} \put(15,70){$P_6$} \put(220,70){\circle*{3}} \put(225,70){$P_1$} \put(30,70){\vector(1,0){60}} \put(90,70){\line(1,0){130}} \put(30,70){\vector(2,1){45}} \put(75,92){\line(2,1){50}} \put(30,70){\vector(2,-1){50}} \put(77,47){\line(2,-1){50}} \put(30,70){\vector(3,-1){44}} \put(63,59){\line(3,-1){36}} \put(84,81){\vector(4,1){10}} \put(30,70){\line(5,1){54}} \put(92,83){\line(5,1){57}} \put(148,94){\vector(3,-1){36}} \put(184,82){\line(3,-1){38}} \put(125,23){\vector(1,3){13}} \put(148,93){\line(-1,-3){12}} \put(135,107){\line(1,-1){13}} \put(125,117){\vector(1,-1){12}} \put(125,22){\vector(2,1){45}} \put(159,39){\line(2,1){60}} \put(220,70){\line(-2,1){45}} \put(125,117){\vector(2,-1){50}} \put(101,47){\vector(1,3){12}} \put(105,59){\line(1,3){19}} \put(101,46){\line(5,1){60}} \put(151,56){\vector(4,1){10}} \put(159,58){\line(5,1){60}} \put(101,47){\vector(1,-1){12}} \put(112,36){\line(1,-1){11}} \put(101,47){\circle*{3}} \put(88,54){$P_5$} \put(125,117){\circle*{3}} \put(125,122){$P_4$} \put(125,23){\circle*{3}} \put(125,08){$P_3$} \put(101,47){\vector(1,1){18}} \put(113,60){\line(1,1){34}} \put(148,94){\circle*{3}} \put(149,80){$P_2$} \put(-30,-10){Figure 5: A spatial degeneration scheme for the Painlev\'e equations} \end{picture} \end{center} \label{fig:5} \end{figure} Furthermore, it seems that the vertices $P_2$ and $P_3$ could have a more complicated microlocal structure $P_2\sim P_2\rightleftarrows P_{34}$ and $P_3\sim P_3\rightleftarrows P_5'\rightleftarrows P_3'$ where $P_3'$ is a special case of $P_3$ defined by the following conditions: $\gamma\delta=0$ and $|\gamma|+|\delta|\neq0$, where $\gamma$ and $\delta$ are the coefficients of the canonical form of $P_3$ \cite{I}. If this is the case, then some of the edges which are incident to $P_3$ must be paired off in order to be incident to $P_3$ and $P_3'$. The answers to our ``equivalence'' questions (see above) could lead to a further complication of the diagram in Figure 5. The answer to the first question governs the number of its edges, and the answer to the second question could result in additional vertices and edges (say, $P_n\to P_{34}$ and $P_n\to P_5'$). In Figure 5 we observe that for most of the pairs of vertices there are several routes connecting them. The natural question therefore is: {\it are there any nonequivalent routes for some pair of vertices?} The notion of equivalent routes, $P_n\to P_k\to P_l$ and $P_n\to P_m\to P_l$, can also be formulated by means of the commutativity of the diagram in Figure~4. For this purpose we must identify the roles of the $x$- and $y$-arrows, understanding them as denoting the equivalence classes of the limits. For routes with three or four edges this definition must, naturally, be generalized. As mentioned, the original setting of the formal limits for ODEs of the Painlev\'e type includes the small parameter $\varepsilon$. The appearance of $\varepsilon$ endows the problem with an infinite-dimensional status, as the constants of integration can be considered as arbitrary functions of $\varepsilon$. Furthermore, it is possible that the latter functions do not have a limit as $\varepsilon\to0$, although the corresponding limiting equations exist and therefore possess no explicit $\varepsilon$ dependence. Thus, we see that for a proper understanding of such limits on the level of the solutions a {\it regularization procedure} may be required. The regularization procedure, which we adopt in this work, means that we will be making such choices of the functions mentioned above so that they would satisfy the weak isomonodromy condition. This requirement leads us (see \cite{K1,K2,KK} and Sections~\ref{sec:4} and \ref{sec:5}) to the {\it discretization} of $\varepsilon$. It is clear that one could propose different regularization procedures for the formal limits: different procedures depend predominantly on possible applications and/or the employed mathematical machinery. Of course, some regularizations could require no $\varepsilon$-discretization. In particular we mean the approach proposed by Joshi and Kruskal \cite{JK}. It is necessary to mention at once that not all the details are set out explicitly in \cite{JK}; therefore, \cite{JK} yields a variety of different possibilities: here, we consider one such possibility. Consider the Laurent expansion about a neighborhood of a pole for a solution of some Painlev\'e equation. It has the following parameters: (coefficients of the Painlev\'e equation; (2) one more constant of integration along with the position of the pole; and (3) dependent and independent variables, i.e., the Painlev\'e function itself and its argument. It is possible to define transformation of all the above parameters (at least, again, as formal Laurent-type $\varepsilon$-expansions near $\varepsilon=0$) which map the initial Laurent expansion for the solution of the Painlev\'e equation into a new one. The latter, in turn, can be considered as a pole expansion of the solution of some other (limiting) Painlev\'e equation. All the parameters of the new pole expansion are defined by the coefficients of the leading terms of the formal $\varepsilon$-expansions. These leading terms are precisely the formulae given by Painlev\'e in his original work \cite{P}. So in \cite{JK} no $\varepsilon$-discretization and any associated linear structures and thereby manifolds of monodromy data are involved. We can provide Joshi's and Kruskal's work \cite{JK} with, perhaps, a somewhat unexpected interpretation, which yields an opportunity to relate their method with isomonodromy deformations. This interpretation lead us, again, to a discretization of $\varepsilon$; but in this case, in a way different from that proposed in \cite{K1}, and below in Sections~\ref{sec:4} and \ref{sec:5}. Recall that all the Painlev\'e equations are known to possess transcendental solutions having an infinite number of poles, $\{t_n\}$, on the real axis accumulating at infinity. The above $\varepsilon$-transformation can be interpreted as a procedure for driving the poles of solutions of the Painlev\'e equation to $\infty$. We fix a particular solution, which has an infinite number of poles accumulating at infinity, and treat the ``pole-drive'' as a pole-to-pole jumping. This leads to a discretization of $\varepsilon$ ($\varepsilon\to\varepsilon_n$), because we are aloud now to do only discrete shifts of the poles $t_n\to t_{n+1}$. Simultaneously iterating the pole expansion via B\"acklund transformations by the number of times consistent with the $\varepsilon$-transformation, we arrive at the pole expansion of some particular solution of the limiting Painlev\'e equation. The consistency means that we take the number of B\"acklund transformations such that the coefficients of the corresponding Painlev\'e equation increase by a rate prescribed by the formulae for $\varepsilon$-transformation (with $\varepsilon\to\varepsilon_n$). In other words, on the level of the pole expansions, we can interpret the limiting passage between Painlev\'e equations as some special asymptotics of poles of (fixed) solutions. These asymptotics, in turn, are given in terms of a pole of some particular solution of the limiting Painlev\'e equation. Note that our interpretation of the limiting passage is consistent with isomonodromy deformations. We know that in the standard situation (without B\"acklund iterations) pole-asymptotics of solutions can be parameterized by the monodromy data (see \cite{IN}). {\it Is that possible to get an analogous parametrization of the poles in the new situation described above \footnote[5]{The latter parametrization, in fact, means that the monodromy data of the initial Painlev\'e equation are related with the monodromy data of the limiting one.}?} It is important to make a distinction with our previous works~\cite{K1,K2,KK} which were concerned with the WKB-method and where the clusters of turning points were considered, and this work, where no WKB-method is involved. Hence, now, we are considering a different asymptotic process. In general, for Painlev\'e equations all limiting procedures are related with various processes of merging turning points and singularities in the associated linear ODEs describing corresponding isomonodromy deformations. For example, the limit $P_4\to P_2$ could be organized by considering the following two possibilities for the asymptotic behavior of the $2\times2$ matrix linear ODE associated with $P_4$: (1) merging of four turning points; and (2) merging of the irregular and regular singular points, i.e., in our terminology, we can consider either an isomonodromy cluster of four turning points or an isomonodromy cluster of irregular and regular singularities. So it seems that, really, nonequivalent limits could exist, and the graph on Figure~5 could actually be incomplete. The discretization of $\varepsilon$ is not too important a feature of the technique we are developing. Actually, in the isomonodromy systems with more than one continuous variable, the so-called higher Painlev\'e equations or Garnier systems, we can observe the analogous asymptotic processes and apply the same technique without any $\varepsilon$-discretization~\cite{K3}. We mention also an interesting WKB-theory for special solutions of the Painlev\'e equations and its interrelation with the corresponding WKB objects for the associated scalar linear ODEs~\cite{KT}. The technique considered in \cite{K1,K2,KK} and here has an interesting scope of applications. One meets the clusters of turning points while studying via the isomonodromy approach double-scaling limits of partition functions for matrix models of $2D$ quantum gravity~\cite{FIK1,FIK2}. Clusters of stationary phase (saddle) points, which are similar to the clusters of turning points, appear in the description of caustics in $1+1$ systems integrable via the Inverse Scattering Transform~\cite{K3}. The clusters of regular singularities play an important role in the study of a zero-curvature limit for holonomic quantum field theory of bosons in the Poincare disc\cite{NT,PBT}\footnote[6]{Cited works are written from a different perspective}. The limit relates this theory with the original Euclidean Sato-Miwa-Jimbo theory~\cite{SMJ1,SMJ2,SMJ3,SMJ4,SMJ5,JMMS}. The correspondence of two-point correlators of these theories in our notation reads as $P_6\to P_3$ or $P_6\to P_5'$. Considering Figure~5 we see that if we add to the limit $P_6\to P_5$ the next one $P_5\to P_3$, then we obtain the desired limit $P_6\to P_3$. The limit $P_5\to P_3$ in our approach is also related with the isomonodromy cluster of two regular singularities but on the ``irregular background'' ($A(\lambda,t)=const\cdot\sigma_3\neq0$). Of course, there is a direct limit $P_6\to P_3$, and the question of equivalence discussed above should be studied. I hope to return to this question in a special publication. This paper is organized as follows: Section~\ref{sec:2} contains no new material. It is based mainly on the works~\cite{J,JMU,JM}. In this Section we set the notation related with the description of isomonodromy deformations for $P_6$ and $P_5$. Most of the formulae presented here are extensively used throughout Sections~\ref{sec:3}-\ref{sec:5}. Some of our definitions are slightly different from those of the cited works; in particular, I found it convenient to define two different manifolds of monodromy data for $P_5$ (${\cal M}_5$ and $\tilde{\cal M}_5$): they are related with the schemes of the calculations in Sections~\ref{sec:4} and \ref{sec:5}, namely, constructions of semi-infinite sequences of Schlesinger transformations, which are important issues in our approach. In Section~\ref{sec:3}, we follow the simple procedure explained at the beginning of the Introduction to derive two different and novel formal limits $P_6\to P_5$. The corresponding formulae for the canonical Painlev\'e functions are rather cumbersome, whilst those being rewritten for the corresponding $\tau$-functions become much simpler. In Section~\ref{sec:4} we derive a discrete regularized version of the first limit. We calculate the manifold of the monodromy data for the limiting $P_5$ transcendent in terms of the corresponding manifold for the initial $P_6$ transcendent. The result is stated in Theorem~\ref{th:1}, and the following part of the section presents details of the derivation. Theorem~\ref{th:2} of Section~\ref{sec:5} states results for a discrete regularization of the second limit analogous to those in Section~\ref{sec:4}; however, a derivation of these results is more complicated than those in Section~4, because, in Section~\ref{sec:5} we meet an additional problem with the normalization of {\it a priori} unknown function $\Psi_5$. It causes a number of additional technical detail one of them is considered in Appendix. As a by-product for these efforts, we get: (1) a clear picture of how the monodromy matrices of the merging regular singularities produce the Stokes multipliers of the resulting $\Psi_5$-function which has an irregular singularity describing the merging process; and (2) the latter derivation is easy to generalize for a background with an arbitrary number of regular singularities. In the case of $P_6$, when Equation~(\ref{eq:cluster-definition}) has four regular singularities, the cluster and exterior cluster domains can be mapped one into another via a fractional-linear transformation. The latter means a transformation which maps the cluster of Section~\ref{sec:4} into the one of Section~\ref{sec:5}. Thus we prove the equivalence of the limits for the Painlev\'e transcendents and the corresponding $\tau$-functions. In the Appendix we study the problem of the simultaneous reduction of a pair of ${\rm SL}(2,\mathbb C)$ matrices to the lower and upper triangular forms. The solution is stated in a theorem which refers to a number of special cases studied in corresponding propositions. These results have an important consequence on the solvability of the asymptotic problem studied in Section~(\ref{sec:5}). A generalization of the problem of the simultaneous transformation of $n\geq3$ ${\rm SL}(2,\mathbb C)$ matrices to triangular forms is important for the investigation of the isomonodromy clusters with $n$ regular singularities. \section{$P_6$ and $P_5$ as Isomonodromy Deformations} \label{sec:2} In this section we recall some facts from the isomonodromy theory of $P_6$ and $P_5$ following, mainly, the works~\cite{J,JMU,JM}. To make a distinction between analogous objects related to $P_n$ $(n=5,6)$ for different $n$, we supply all of them by the subscript $n$. The convenience of this agreement becomes evident in the following sections. Consider the $2\times2$ matrix linear ODE with four regular singularities at $\lambda=0,1,t,\infty$: \begin{equation} \label{eq:linear-4-regular} \frac{d}{d\lambda}\Psi_6=\left(\frac{A_{06}}\lambda+ \frac{A_{16}}{\lambda-1}+\frac{A_{t6}}{\lambda-t_6}\right)\Psi_6. \end{equation} It is assumed that \begin{equation} \label{eq:niormalization-for-P6} A_{06}+A_{16}+A_{t6}=-\frac{\Theta_{\infty6}}2\sigma_3,\qquad\sigma_3= \left(\begin{array}{cc}1&0\\0&-1\end{array}\right),\qquad \Theta_{\infty6}\in\mathbb{C}\setminus\mathbb{Z}, \end{equation} and there exist $R_{\nu6}\in{\rm SL}(2,\mathbb{C})$ such that \begin{equation} \label{eq:Rnu6} R_{\nu6}^{-1}A_{\nu6}R_{\nu6}=\frac{\Theta_{\nu6}}2\sigma_3,\qquad \Theta_{\nu6}\in\mathbb{C}\setminus\mathbb{Z}. \end{equation} In a neighborhood of the regular singularity $\lambda=\nu\;(\nu\neq\infty)$, the $\Psi_6$-function can be expanded as \begin{equation} \label{eq:psi-expansion-at-nu} \Psi_6\,\underset{\lambda\to\nu}=\,\sum\limits_{k=0}^\infty\Psi_{k\nu6} (\lambda-\nu)^{k+\frac{\Theta_{\nu6}}2\sigma_3}C_{\nu6}, \end{equation} where the matrices $\Psi_{k\nu6}$, $C_{\nu6}$ are independent of $\lambda$, because of the normalization conditions (\ref{eq:niormalization-for-P6}) and (\ref{eq:Rnu6}), we can assume that \begin{equation} \label{eq:condition-det-1} \Psi_6, R_{\nu6},\Psi_{0\nu6},C_{\nu6}\in{\rm SL}(2,\mathbb{C}),\qquad\nu=0,1,t. \end{equation} Because of Equation~(\ref{eq:niormalization-for-P6}), the expansion of the $\Psi_6$-function can be normalized at $\lambda=\infty$: \begin{equation} \label{eq:psi6-infty-normalization} \Psi_6\underset{\lambda\to\infty}=\left(I+ \sum\limits_{k=1}^\infty\Psi_{k\infty6}\lambda^{-k}\right)\lambda^{-\frac{\Theta_{\infty6}}2\sigma_3}. \end{equation} The single-valued $\Psi_6$-function with the stated properties can be defined on the $\lambda$-plane "cut" along the negative imaginary semi-axis $[-i\infty,0]$, the segment $[0,t_6]$, and the positive real semi-axis $[1,+\infty]$. Henceforth, we assume that $t_6\in(0,1)$. Following \cite{J} we choose the paths in $\bar{\mathbb{C}}\setminus\{0,t_6,1,\infty\}$ as is shown on Figure~6. \begin{figure}[ht] \begin{center} \begin{picture}(310,90) \put(70,30){\circle*{3}} \put(68,36){$0$} \put(120,30){\circle*{3}} \put(112,36){$t$} \put(190,30){\circle*{3}} \put(176,36){$1$} \put(300,30){\circle*{3}} \put(286,36){$\infty$} \put(60,90){\circle*{3}} \put(45,85){$\lambda_0$} \qbezier(60,90)(90,25)(70,20) \qbezier(60,90)(45,25)(70,20) \qbezier(60,90)(155,28)(130,20) \qbezier(60,90)(94,22)(130,20) \qbezier(60,90)(188,08)(214,20) \qbezier(60,90)(247,30)(214,20) \qbezier(60,90)(350,50)(320,25) \qbezier(60,90)(310,10)(320,25) \put(124,42){\vector(-1,1){10}} \put(189,43){\vector(-2,1){10}} \put(77,41){\vector(-1,4){3}} \put(250,58){\vector(-4,1){10}} \put(140,00){Figure 6} \end{picture} \end{center} \end{figure} Continuing the $\Psi_6$-function along these paths we find a multi-valued function with the monodromy matrices $M_{\nu6}$ $(\nu=0,1,t,\infty)$ which, in terms of the local expansions (\ref{eq:psi-expansion-at-nu}) and (\ref{eq:psi6-infty-normalization}), can be written as follows: \begin{equation} \label{eq:monodromy-local} M_{\nu6}=C_{\nu6}^{-1}e^{\pi i\Theta_{\nu6}\sigma_3}C_{\nu6},\qquad\nu=0,1,t,\infty,\quad C_{\infty6}=I. \end{equation} One proves the cyclic relation \begin{equation} \label{eq:cyclic6} M_{\infty6}M_{16}M_{t6}M_{06}=I. \end{equation} The monodromy group $({\cal MG}_6)$ is the subgroup of ${\rm SL}(2,\mathbb{C})$ generated by the matrices $M_{\nu6}$. The set of monodromy data $({\cal M}_6)$ is an ordered set of matrix elements of $M_{\nu6}$ completed with four complex parameters $\Theta_{\nu6}$ $(\nu=0,1,t,\infty)$, satisfying the equation \begin{equation} \label{eq: trace-M} {\rm tr}M_{\nu6}=2\cos(\pi\Theta_{\nu6}). \end{equation} The numbers $\Theta_{\nu6}$ are called the coefficients of formal monodromy. Thus we see that for a fixed parameter $t_6$ and matrices $A_{\nu6}$ the above procedure yields the unique set ${\cal M}_6$. One proves that a different set of $\{A_{\nu6}\}_{\nu=0,1,t}$, with the same $t_6$, defines a different set ${\cal M}_6$; therefore, considering the Riemann problem of reconstruction of all differential equations corresponding to the given set ${\cal M}_6$, we have to vary not only $A_{\nu6}$, but also $t_6$. Following the Schlesinger approach it is convenient to consider $t_6$ as an independent variable, then $A_{\nu6}=A_{\nu6}(t_6)$. To find these functions explicitly one has to notice that the {\it hard isomonodromy condition}, i.e., $\partial_{t_6}{\cal M}_6=0$, leads to the additional ODE with respect to $t_6$ for the function $\Psi_6$, which, under the assumptions on the formal monodromy (\ref{eq:niormalization-for-P6}) and (\ref{eq:Rnu6}), reads \begin{equation} \label{eq:t-6} \frac{\partial}{\partial t_6}\Psi_6=-\frac{A_{t6}}{\lambda-t_6}\Psi_6. \end{equation} The compatibility condition of Equations~(\ref{eq:linear-4-regular}) and (\ref{eq:t-6}) is \begin{equation} \label{eq:schlesinger-p6} \frac{d A_{\nu6}}{dt_6}=\frac{[A_{t6},A_{\nu6}]}{t_6-\nu},\qquad\nu=0,1, \end{equation} where $[A,B]\equiv AB-BA$. In the derivation of (\ref{eq:schlesinger-p6}) we took into account the normalization condition (\ref{eq:niormalization-for-P6}) together with $\partial_{t_6}\Theta_{\infty6}=0$. The system~(\ref{eq:schlesinger-p6}) is called the {\it system of iso\-mo\-no\-dro\-my deformations} or the Schle\-sin\-ger system. It consists of eight scalar nonlinear ODEs of the first order for the matrix elements of $A_{\nu6}$ $(\nu=0,1)$. By using Equations~(\ref{eq:niormalization-for-P6}) and (\ref{eq:Rnu6}), one finds the following five first integrals: $$ {\rm tr}\,A_{\nu6}=0,\quad\det\,A_{\nu6}=-\frac{\Theta_{\nu6}^2}4,\;\nu=0,1,\quad \det(\Theta_{\infty6}\sigma_3/2+A_{06}+A_{16})=-\Theta^2_{t6}/4, $$ reducing the number of independent scalar equations to three. This number coincides with with the complex dimension of the {\it manifold of monodromy data}, ${\cal M}_6(\Theta_{06},\Theta_{16},\Theta_{t6},\Theta_{\infty6})$. The points of this manifold are in one-to-one correspondence with the set ${\cal M}_6$ for the fixed parameters $\Theta_{\nu6}$. More precisely: identifying an ordered set of the matrix elements of ${\cal M}_6$ as a point in $\mathbb{C}^{14}$, one defines ${\cal M}_6(\Theta_{06},\Theta_{16},\Theta_{t6},\Theta_{\infty6})$ as an algebraic variety in $\mathbb{C}^{14}$ by Equations~(\ref{eq:cyclic6}), (\ref{eq: trace-M}), and $\det M_{\nu6}=1$, $(\nu=0,1,t,\infty)$. Following \cite{JM}, introduce the notation, \begin{equation} \label{eq:A-and-R} A_{\nu6}=\left(\begin{array}{cc} z_{\nu6}+\frac{\Theta_{\nu6}}2&-u_{\nu6}z_{\nu6}\\ \frac{z_{\nu6}+\Theta_{\nu6}}{u_{\nu6}}&-z_{\nu6}-\frac{\Theta_{\nu6}}2 \end{array}\right),\qquad R_{\nu6}=\left(\begin{array}{cc} \frac1{\Theta_{\nu6}s_{\nu6}}&u_{\nu6}z_{\nu6}s_{\nu6}\\ \frac1{\Theta_{\nu6}u_{\nu6}s_{\nu6}}&(z_{\nu6}+\Theta_{\nu6})s_{\nu6} \end{array}\right), \end{equation} where $\nu=0,1,t$, and $s_{\nu6}\in\mathbb{C}\setminus\{0\}$. Equation~(\ref{eq:niormalization-for-P6}) in the notation~(\ref{eq:A-and-R}) reads: \begin{gather} \label{eq:normalization-scalar-diag} z_{06}+z_{16}+z_{t6}=-\frac12(\Theta_{06}+\Theta_{16}+\Theta_{t6}+\Theta_{\infty6}),\\ u_{06}z_{06}+u_{16}z_{16}+u_{t6}z_{t6}=0,\qquad\frac{z_{06}+\Theta_{06}}{u_{06}}+\frac{z_{16}+ \Theta_{16}}{u_{16}}+\frac{z_{t6}+\Theta_{t6}}{u_{t6}}=0. \label{eq:normalization-scalar-off} \end{gather} Thus, only three parameters amongst those which define the coefficients of Equation~(\ref{eq:linear-4-regular}) are independent. In the hard isomonodromy case the parameters appear to be functions of $t_6$: $z_{\nu6}=z_{\nu6}(t_6)$ and $u_{\nu6}=u_{\nu6}(t_6)$ defined by System~(\ref{eq:schlesinger-p6}). As explained above, one can rewrite (\ref{eq:schlesinger-p6}) as a system of three scalar first-order ODEs. A convenient form of this system is given in \cite{JM}. Here we define two important functions closely related with System~(\ref{eq:schlesinger-p6}). The first one, \begin{equation} \label{eq:y6} y_6=\frac{t_6u_{06}z_{06}}{(t_6+1)u_{06}z_{06}+t_6u_{16}z_{16}+u_{t6}z_{t6}}= \frac1{1+\left(1-\frac1{t_6}\right)\frac{u_{16}z_{16}}{u_{06}z_{06}}}= \frac{t_6}{1+(1-t_6)\frac{u_{t6}z_{t6}}{u_{06}z_{06}}}, \end{equation} is a solution of the sixth Painlev\'e equation, \begin{align} \frac{d^2y_6}{dt_6}^2&=\frac12\left(\frac1{y_6}+\frac1{y_6-1}+\frac1{y_6-t_6}\right)\left (\frac{dy_6}{dt_6}\right)^2-\left(\frac1{t_6}+\frac1{t_6-1}+\frac1{y_6-t_6}\right)\frac{dy_6}{dt_6} \nonumber\\ &+\frac{y_6(y_6-1)(y_6-t_6)}{t_6^2(t_6-1)^2}\left(\alpha_6+\frac{\beta_6t_6}{y_6^2}+ \frac{\gamma_6(t_6-1)}{(y_6-1)^2}+\frac{\delta_6t_6(t_6-1)}{(y_6-t_6)^2}\right), \label{eq:p6}\\ &\alpha_6=\frac12(\Theta_{\infty6}-1)^2,\quad \beta_6=-\frac12\Theta_{06}^2,\quad \gamma_6=\frac12\Theta_{16}^2,\quad \delta_6=\frac12(1-\Theta_{t6}^2). \label{eq:p6coeff} \end{align} The second function, which is important for applications, is the $\tau$-function~\cite{JM}: \begin{equation} \label{eq:tau-6} \frac{d}{dt_6}\log\,\tau_6(t_6)={\rm tr}\,\left(\frac{A_{06}}{t_6}+\frac{A_{16}}{t_6-1}\right)A_{t6} \end{equation} The function \begin{equation} \label{eq:sigma-6} \hat\sigma_6(t_6)=t_6(t_6-1)\frac{d}{dt_6}\log\tau_6(t_6) \end{equation} satisfies an ODE of the second order, quadratic with respect to $\hat\sigma''_6(t_6)$~\cite{J,O,JM}\footnote[7]{In the cited papers, the authors use a definition of the $\sigma$-function shifted by a linear function of $t_6$. The latter function satisfies a differential equation symmetric with respect to the formal monodromies. We use a ``hat'' in our notation to make a difference between these functions.}. Conversely, starting from an arbitrary solution of Equation~(\ref{eq:p6}) and using the formulae given in \cite{JM} one can construct a solution of the Schlesinger System~(\ref{eq:schlesinger-p6}) satisfying Equation~(\ref{eq:y6}). Therefore, all solutions of $P_6$ (\ref{eq:p6}) can be obtained from some solution of the Schlesinger System (\ref{eq:schlesinger-p6}) via Equation~(\ref{eq:y6}). Let us consider the {\it weak isomonodromy deformations} of Equation~((\ref{eq:linear-4-regular}) as the deformations of $A_{\nu6}$ ($\nu,0,1,t$) preserving, up to the sign, the generators of ${\cal MG}_6$, i.e., isomonodromy deformations in the sense of the ${\rm PSL}_2(\mathbb{C})$ monodromy group. Using Equations~(\ref{eq:cyclic6}) and (\ref{eq: trace-M}) one proves that in the weak case the continuous Schlesinger deformations can be extended only by the group of discrete Schlesinger transformations acting on the formal monodromies as \begin{equation} \label{eq:theta-shift} \Theta_{\nu6}\to\Theta_{\nu6}+n_\nu,\quad\nu=0,1,t,\infty,\qquad n_0+n_1+n_t+n_\infty=0({\rm mod}\,2), \end{equation} and the reflections \begin{equation} \label{eq: theta-reflections} \Theta_{\nu6}\to-\Theta_{\nu6},\qquad\nu=0,1,t. \end{equation} Consider how these transformations act on the coefficients $A_{\nu6}$ of Equation~(\ref{eq:linear-4-regular}). First, notice that the reflections~(\ref{eq:theta-shift}) are actually related with a certain ambiguity in the parametrization of the matrices $A_{\nu6}$ rather than with any transformation of these matrices. The ambiguity in the parametrization is related with the ambiguity in writing the local expansion of the $\Psi_6$ function near the corresponding singular point: $(\Psi_{k\nu6},\Theta_{\nu6},C_{\nu6})$ $\mapsto$ $(\Psi_{k\nu6}\sigma_1,-\Theta_{\nu6},\sigma_1C_{\nu6})$, where $\sigma_1$ is the Pauli matrix (with the matrix elements $\sigma_1^{ij}=0$ if $i=j$, else $\sigma_1^{ij}=1$). This nonuniqueness leads, simply, to the reparametrization of the matrices $A_{\nu6}$: $(u_{\nu6},\Theta_{\nu6},z_{\nu6})$ $\mapsto$ $(\tilde u_{\nu6},\tilde\Theta_{\nu6},\tilde z_{\nu6})$, where $z_{\nu6}+\frac{\Theta_{\nu6}}2=\tilde z_{\nu6}+\frac{\tilde\Theta_{\nu6}}2,\tilde\Theta_{\nu6}=\Theta_{\nu6}$, and $u_{\nu6}z_{\nu6}=\tilde u_{\nu6}\tilde z_{\nu6}$, which doesn't affect the functions $y_6(t_6)$ and $\hat\sigma_6(t_6)$ (see Equations (\ref{eq:y6}), (\ref{eq:tau-6}), and (\ref{eq:sigma-6})). Note that the reflection \begin{equation} \label{eq:reflection-infinity} \tilde\Theta_{\infty6}=-\Theta_{\infty6} \end{equation} does not preserve the generators of ${\cal MG}_6$. This reflection is related with the change of the generators of ${\cal MG}_6$ to their inverse: $\widetilde{\cal MG}_6=\sigma_1{\cal MG}_6\sigma_1$. To see this one can define the following action of the reflection (\ref{eq:reflection-infinity}) on the $\Psi_6$ function, $\tilde\Psi_6=\sigma_1\Psi_6\sigma_1$; it yields: \begin{equation} \label{eq:action-reflect-infty} \tilde M_{\nu6}=\sigma_1M_{\nu6}\sigma_1,\qquad \tilde A_{\nu6}=\sigma_1A_{\nu6}\sigma_1. \end{equation} It is an immediate consequence of Equations (\ref{eq:action-reflect-infty}), (\ref{eq:tau-6}), and (\ref{eq:sigma-6}) that $\tilde{\hat{\sigma}}_6=\hat\sigma_6$. In terms of the matrix elements the last equation in (\ref{eq:action-reflect-infty}) reads: \begin{gather} \label{eq:trans-theta-reflect-infty} \tilde\Theta_{\nu6}=\pm\Theta_{\nu6},\quad\nu=0,1,t,\\ \tilde z_{\nu6}=-\left(z_{\nu6}+\frac{\tilde\Theta_{\nu6}+\Theta_{\nu6}}2\right),\quad \tilde u_{\nu6}=\frac1{u_{\nu6}}\cdot \frac{z_{\nu6}+\Theta_{\nu6}}{z_{\nu6}+\frac{\tilde\Theta_{\nu6}+\Theta_{\nu6}}2}, \label{eq:trans-coef-reflect-infty} \end{gather} where in Equation (\ref{eq:trans-theta-reflect-infty}) for every value of $\nu$ we can make arbitrary choices of the signs. One can use Equations (\ref{eq:trans-coef-reflect-infty}), (\ref{eq:y6}), and Equations (C.49), (C.52), and (C.55) of \cite{JM} to find $\tilde y_6$ as a rational function of $y_6$, $y_6'$, and $t_6$. Consider the group of the discrete Schlesinger Transformations (\ref{eq:theta-shift}). Evidently, it has four generators and it is isomorphic to the group of translations along special basis in $\mathbb{C}^4$. We denote by ${\cal L}_{\nu\nu'}^{\pm\pm}$, where $\nu\neq\nu'$, the elementary Schlesinger Transformations~\cite{JM}: $$ {\cal L}_{\nu\nu'}^{\pm\pm}:\quad \tilde\Theta_{\varkappa6}=\Theta_{\varkappa6}\pm1,\quad {\rm for}\quad\varkappa=\nu,\,\nu'; \qquad\tilde\Theta_{\varkappa6}=\Theta_{\varkappa6}\quad{\rm for}\quad\varkappa\in\{0,1,t,\infty\} \setminus\{\nu,\,\nu'\}. $$ In ${\cal L}_{\nu\nu'}^{\pm\pm}$ we choose the signs over $\kappa=\nu,\,\nu'$ in the same way as in the formula for $\tilde\Theta_{\varkappa6}$. Since $\tilde M_{\varkappa6}=\pm M_{\varkappa6}$ (the sign minus occurs only for $\varkappa=\nu,\,\nu'$), the action of ${\cal L}_{\nu\nu'}^{\pm\pm}$ on the $\Psi_6$-function can be defined as the left multiplication \begin{equation} \label{eq:L-dressing} \tilde\Psi_6={\cal L}_{\nu\nu'}^{\pm\pm}\Psi_6, \end{equation} where \begin{gather*} L_{\nu\nu'}^{\pm\pm}=\frac{\sqrt{\lambda-\nu'}}{\sqrt{\lambda-\nu}}J_{\nu\nu'}^{\pm\pm}+ \frac{\sqrt{\lambda-\nu}}{\sqrt{\lambda-\nu'}}J_{\nu'\nu}^{\pm\pm},\qquad\nu,\,\nu'=0,1,t,\\ L_{\nu\infty}^{\pm\pm}=\sqrt{\lambda-\nu}\sigma_\infty^\pm+\frac1{\sqrt{\lambda-\nu}}J_{\nu\infty}^{\pm\pm}, \quad\nu\neq\infty,\quad\sigma_\varkappa^+=\left(\begin{array}{cc}0&0\\0&1\end{array}\right),\;\; \sigma_\varkappa^-=\left(\begin{array}{cc}1&0\\0&0\end{array}\right). \end{gather*} Here, the branches of the roots are fixed as $\sqrt{\lambda-\nu}/\sqrt\lambda\to1$ and $\sqrt{\lambda-\nu}/\sqrt{\lambda-\nu'}\to1$ as $\lambda\to\infty$ and $\lambda$ belongs to the complex plane cut as explained above (see the paragraph right after Equation~(\ref{eq:psi6-infty-normalization})). The choice of the signs over all equal subscripts is the same (upper/lower). The matrices $J_{\nu\nu'}^{\pm\pm}$ for $\nu,\nu'\neq\infty$ are uniquely defined by the equations: $$ J_{\nu\nu'}^{\pm\pm}+J_{\nu'\nu}^{\pm\pm}=I,\qquad J_{\nu\nu'}^{\pm\pm}R_{\nu6}\sigma_\nu^{\mp}= \left(\begin{array}{c}0\\0\end{array}\right). $$ The result is as follows \begin{equation} \label{eq:J-and-Delta} J_{\nu'\nu}^{\pm\pm}=\frac1{\Delta_{\nu'\nu}^{\pm}} \left(\begin{array}{cc}b_\nu^{\pm}&0\\0&-a_\nu^{\pm}\end{array}\right) \left(\begin{array}{cc}a_{\nu'}^{\pm}&b_{\nu'}^{\pm}\\a_{\nu'}^{\pm}&b_{\nu'}^{\pm}\end{array}\right), \qquad\Delta_{\nu'\nu}^{\pm}=a_{\nu'}^{\pm}b_\nu^\pm-a_\nu^{\pm}b_{\nu'}^\pm, \end{equation} where $a_\nu^{\pm}$ and $b_\nu^{\pm}$ are different notations for the matrix elements of $R_{\nu6}$ (see Equation~(\ref{eq:Rnu6})) which are convenient here, namely, \begin{equation} \label{eq:Rnu6a-b} R_{\nu6}=\left(\begin{array}{cc}b_{\nu}^+&b_{\nu}^-\\-a_{\nu}^+&-a_{\nu}^-\end{array}\right). \end{equation} The matrix $J_{\nu\nu'}^{\pm\pm}$ can be found by the same Equations~(\ref{eq:J-and-Delta}) and (\ref{eq:Rnu6a-b}) by making the permutation of the subscripts $(\nu'\leftrightarrow\nu)$. Note also the following useful properties of $J_{\nu\nu'}^{\pm\pm}$: \begin{equation} \label{eq:J-projectors} \big(J_{\nu\nu'}^{\pm\pm}\big)^2=J_{\nu\nu'}^{\pm\pm},\qquad J_{\nu\nu'}^{\pm\pm}J_{\nu'\nu}^{\pm\pm}=0, \end{equation} and analogous equations with $\nu\leftrightarrow\nu'$. We see that Transformation~(\ref{eq:L-dressing}) exists iff $\Delta_{\nu\nu'}^\pm\neq0$. Using Equations (C.51), (C.52), and (C.55) of \cite{JM} one finds that the condition $\Delta_{\nu\nu'}^\pm=0$ is equivalent to the existence of a one-parameter solution $y_6$ of $P_6$, which solves an ODE of the first order, \begin{equation} \label{eq:first-order-y6} \frac{dy_6}{dt_6}=R(y_6,t_6), \end{equation} where $R$ is a rational function of its arguments with the coefficients defined by $\Theta_{\nu6}$ $(\nu=0,1,t,\infty)$. Thus, for general continuous Schlesinger deformations (\ref{eq:schlesinger-p6}) which correspond to transcendental (nonclassical) solutions of $P_6$, the condition \begin{equation} \label{eq:Delta-ne0} \Delta_{\nu\nu'}^\pm\neq0 \end{equation} is valid. Furthermore, we consider iterations of Transformations~(\ref{eq:L-dressing}). If in some step we find $\Delta_{\nu\nu'}^\pm=0$, then it means we start from a solution of $P_6$ which is the iteration of a special (classical) solution of (\ref{eq:p6}), i.e., it can be presented in the form $\tilde R(y_6,y_6',t_6)$, where $\tilde R$ is a rational function of its arguments with the coefficients defined by $\Theta_{\nu6}$ and where $y_6$ is a solution of an equation of the type (\ref{eq:first-order-y6}). Thus we can iterate general (transcendental) solutions of $P_6$ (\ref{eq:p6}) without any restrictions. The condition (\ref{eq:Delta-ne0}) is assumed throughout this paper. It would be interesting to perform the complete investigation of our problem including the special (classical) solutions of Equation~(\ref{eq:p6}). This investigation is in progress now \cite{U,W}. Let us also find the action of ${\cal L}_{\nu\nu'}^{\pm\pm}$ on the matrices $A_{\nu6}$: substituting transformation~(\ref{eq:L-dressing}) into Equation~(\ref{eq:p6}) written for $\tilde A_{\nu6}$ and $\tilde\Psi_6$, one obtains: \begin{gather} \label{eq:Atilde-nu} J_{\nu'\nu}^{\pm\pm}\left(A_{\nu'6}-\frac12\right)=\tilde A_{\nu'6}J_{\nu'\nu}^{\pm\pm},\qquad J_{\nu\nu'}^{\pm\pm}\left(A_{\nu6}-\frac12\right)=\tilde A_{\nu6}J_{\nu\nu'}^{\pm\pm},\\ \tilde A_{\mu6}=\left(J_{\nu'\nu}^{\pm\pm}+\frac{\nu-\mu}{\nu'-\mu}J_{\nu\nu'}^{\pm\pm}\right)A_{\mu6} \left(J_{\nu'\nu}^{\pm\pm}+\frac{\nu'-\mu}{\nu-\mu}J_{\nu\nu'}^{\pm\pm}\right),\qquad\mu\neq\nu,\nu'. \label{eq:Atilde-mu} \end{gather} Using these equations and noting that the set $\{\nu,\nu',\mu\}$ is the permutation of $\{0,1,t\}$ we find \begin{equation} \label{eq:Atilde-infty} \tilde A_{\nu6}+\tilde A_{\nu'6}+\tilde A_{\mu6}=-\frac{\Theta_{\infty6}}2\sigma_3. \end{equation} Multiplying Equation~(\ref{eq:Atilde-infty}) by $J_{\nu\nu'}^{\pm\pm}$ on the right and using Equations (\ref{eq:J-projectors}), (\ref{eq:Atilde-nu}), and (\ref{eq:Atilde-mu}), we find \begin{align} \label{eq:Atilde-nu-result} \tilde A_{\nu'6}=&-\frac{\Theta_{\infty6}}2\sigma_3J_{\nu\nu'}^{\pm\pm}+ J_{\nu'\nu}^{\pm\pm}\left(A_{\nu'6}-\frac12\right)-J_{\nu\nu'}^{\pm\pm}\left(A_{\nu6}-\frac12\right)-\\ &\left(\frac{\nu'-\mu}{\nu-\mu}J_{\nu'\nu}^{\pm\pm}+J_{\nu\nu'}^{\pm\pm}\right)A_{\mu6}J_{\nu\nu'}^{\pm\pm}, \nonumber \end{align} where $\tilde A_{\nu6}$ is given by the same Equation~(\ref{eq:Atilde-nu-result}) but with the permutation $\nu'\leftrightarrow\nu$; $\tilde A_{\mu6}$ can then be found from Equation~(\ref{eq:Atilde-infty}). Consider, now, Transformation~(\ref{eq:L-dressing}) for $\nu'=\infty$: we find that \begin{equation} \label{eq:J-nu-infty} J_{\nu\infty}^{\pm+}=\frac1{a_\nu^\pm}\left(\begin{array}{cc}1&0\\0&-\Psi_{1\infty6}^{21}\end{array}\right) \left(\begin{array}{cc}a_\nu^\pm&b_\nu^\pm\\a_\nu^\pm&b_\nu^\pm\end{array}\right),\quad J_{\nu\infty}^{\pm-}= \frac1{b_\nu^\pm}\left(\begin{array}{cc}-\Psi_{1\infty6}^{12}&0\\0&1\end{array}\right) \left(\begin{array}{cc}a_\nu^\pm&b_\nu^\pm\\a_\nu^\pm&b_\nu^\pm\end{array}\right), \end{equation} where $a_\nu^\pm$ and $b_\nu^\pm$ are defined by Equations~(\ref{eq:Rnu6a-b}) and (\ref{eq:A-and-R}), $\Psi_{1\infty6}^{ij}$ are the matrix elements of the first coefficient of the expansion (\ref{eq:psi6-infty-normalization}). These matrix elements can be calculated via the matrix elements of $A_{\nu6}$, since (recall $\Theta_{\infty6}\notin\mathbb{Z}$): $$ -\Psi_{1\infty6}+\frac{\Theta_{\infty6}}2\big[\sigma_3,\Psi_{1\infty6}\big]=A_{16}+t_6A_{t6}. $$ Thus, Transformation~(\ref{eq:L-dressing}) with $\nu'=\infty$ exists iff \begin{equation} \label{eq:L-exiatence} {\cal L}_{\nu\infty}^{\pm+}:\;\;a_\nu^\pm\neq0,\qquad {\cal L}_{\nu\infty}^{\pm-}:\;\;b_\nu^\pm\neq0. \end{equation} The violation of conditions~(\ref{eq:L-exiatence}) can be discussed in the same manner as the violation of Condition~(\ref{eq:Delta-ne0}) (see the paragraph between Equations~(\ref{eq:J-projectors}) and (\ref{eq:Atilde-nu})). The only difference is that Equation~(\ref{eq:first-order-y6}) now takes the form $$ \left(\frac{dy_6}{dt_6}\right)^2=R(y_6,t_6). $$ Hereafter we assume that Conditions (\ref{eq:L-exiatence}) are valid (as well as the previously assumed (\ref{eq:Delta-ne0})). Consider now the action of ${\cal L}_{\nu\infty}^{\pm\pm}$ on $A_{\mu6}$, namely: \begin{equation} \label{eq:A-tilde-mu-infty} \tilde A_{mu6}=\left(\sigma_\infty^\pm+\frac1{\mu-\nu}J_{\nu\infty}^{\pm\pm}\right)A_{\mu6} \left(\sigma_\infty^\pm+\frac1{\mu-\nu}J_{\nu\infty}^{\pm\pm}\right)^{-1},\qquad\mu\neq\nu. \end{equation} The action of ${\cal L}_{\nu\infty}^{\pm\pm}$ on $A_{\nu6}$ can be obtained by substituting Equation~(\ref{eq:A-tilde-mu-infty}) for $\tilde A_{mu6}$ and (with $\mu\leftrightarrow\mu'$) for $\tilde A_{mu'6}$, where $\mu'$ is defined from the condition that $\{\nu,\mu,\mu'\}$ is a permutation of $0,1,t$, into the equation \begin{equation} \label{eq:Atilde-infty2} \tilde A_{\nu6}+\tilde A_{\mu6}+\tilde A_{\mu'6}=-\frac{\tilde\Theta_{\infty6}^\pm}2\sigma_3= -\frac{\Theta_{\infty6}\pm1}2\sigma_3,\qquad \mu\neq\mu'\neq\nu\neq\mu. \end{equation} To summarize, let us fix $t_6^0\in\mathbb{C}\setminus\{0,1\}$ and $A_{\nu6}^0$, $\Theta_{\nu6}^0$ as demanded by the last conditions in (\ref{eq:niormalization-for-P6}) and (\ref{eq:Rnu6}). The general Schlesinger (isomonodromy) deformations (GSD) for Equation~(\ref{eq:linear-4-regular}) are the matrices $A_{\nu6}$ (or their matrix elements) which depend on the continuous variable $t_6$ and the discrete variables $\Theta_{\nu6}$: $\Theta_{\nu6}-\Theta_{\nu6}^0\in\mathbb{Z},\; \sum_\nu(\Theta_{\nu6}-\Theta_{\nu6}^0)=0({\rm mod}2)$. The continuous deformations of $A_{\nu6}$ are governed by (\ref{eq:schlesinger-p6}) and the discrete deformations by Equations~(\ref{eq:Atilde-mu})-- (\ref{eq:Atilde-nu-result}) and (\ref{eq:A-tilde-mu-infty}), (\ref{eq:Atilde-infty2}). The initial condition is stated as $A_{\nu6}(t_6,\Theta_{06}^0,\Theta_{16}^0,\Theta_{t6}^0,\Theta_{\infty6}^0)=A_{\nu6}^0$. The continuous and discrete deformations are commuting so that GSD are correctly defined. Any GSD can be uniquely characterized by a point on ${\cal M}_6(\Theta_{06}^0,\Theta_{16}^0,\Theta_{t6}^0,\Theta_{\infty6}^0)$. {\it The direct monodromy problem} is: construct ${\cal M}_6(\Theta_{06}^0,\Theta_{16}^0,\Theta_{t6}^0,\Theta_{\infty6}^0)$ for the given GSD. {\it The inverse monodromy problem} is: construct the GSD for given ${\cal M}_6(\Theta_{06}^0,\Theta_{16}^0,\Theta_{t6}^0,\Theta_{\infty6}^0)$ \cite{B}. Consider the $2\times2$ matrix linear ODE related with $P_5$: \begin{equation} \label{eq:linear-5} \frac{d\Psi_5}{d\lambda}=\left(\frac{t_5}2\sigma_3+\frac{A_{05}}\lambda+\frac{A_{15}}{\lambda-1}\right)\Psi_5. \end{equation} This equation possess two regular singular points, at $\lambda=0$ and $1$, and an irregular one at the point of $\infty$. We require the following conditions: \begin{equation} \label{eq:diag-A5} {\rm diag}(A_{05}+A_{15})=-\frac{\Theta_{\infty5}}2\sigma_3,\qquad\Theta_{\infty5}\in\mathbb{C}; \end{equation} and that there exists $R_{\nu5}\in{\rm SL}(2,\mathbb{C})$ such that \begin{equation} \label{eq:R5-def} R_{\nu5}^{-1}A_{\nu5}R_{\nu5}=\frac{\Theta_{\nu5}}2\sigma_3,\qquad \Theta_{\nu5}\in\mathbb{C}\setminus\mathbb{Z},\quad\nu=0,1. \end{equation} To define the monodromy data let us define the canonical solutions $\Psi_5^k$ of Equation~(\ref{eq:linear-5}) by setting their asymptotics at the infinity point as \begin{gather} \label{eq:Psi5-asymptotics-infty} \Psi_5^k=\left(I+\sum\limits_{m=1}^\infty\Psi_{m\infty5}\lambda^{-m}\right) \exp\!\left(\!\!\Big(\frac{\lambda t_5}2-\frac{\Theta_{\infty5}}2\ln\lambda\Big)\!\sigma_3\!\right),\\ \lambda\to\infty,\qquad -\frac32\pi+\pi k<\arg(\lambda t_5)<\frac{\pi}2+\pi k,\quad k\in\mathbb{Z}.\nonumber \end{gather} The single-valued function $\Psi_5^k$ can be defined in the domain $\mathbb{C}\setminus\big([0,1]\cup[0,\infty e^{\frac{i\pi}2+\pi ki}\big)$, where we choose the main branch of $\ln\lambda$: ${\rm Im}\ln\lambda={\rm arg}\lambda$. Using Asymptotics (\ref{eq:Psi5-asymptotics-infty}), one proves that \begin{equation} \label{eq:infinity-circle} \Psi_5^{k+2}\big(\lambda\, e^{2\pi i}\big)=\Psi_5^k(\lambda)e^{-\pi i\Theta_{\infty5}\sigma_3}. \end{equation} Now we define the Stokes matrices $S_k$ as \begin{equation} \label{eq:Stokes-definition} \Psi_5^{k+1}(\lambda)=\Psi_5^k(\lambda)S_k. \end{equation} Definitions (\ref{eq:Psi5-asymptotics-infty}) and (\ref{eq:Stokes-definition}) yield $$ S_{2l}=\left(\begin{array}{cc}1&0\\s_{2l}&1\end{array}\right),\qquad S_{2l+1}=\left(\begin{array}{cc}1&s_{2l+1}\\0&1\end{array}\right),\qquad l\in\mathbb{Z}, $$ where $s_k$ are called the Stokes multipliers. Using Equations~(\ref{eq:infinity-circle}) and (\ref{eq:Stokes-definition}) one finds \begin{equation} \label{eq:stokes-shift-2} S_{k+2}=e^{\pi i\Theta_{\infty5}\sigma_3}S_ke^{-\pi i\Theta_{\infty5}\sigma_3}. \end{equation} The monodromy matrix at the point of infinity, $M_{k\infty5}$, for the $\Psi_5^k$-function is given by the equation \begin{equation} \label{eq:monodromy-infty} \Psi_5^k\big(\lambda\,e^{-2\pi i}\big)=\Psi_5^k(\lambda)M_{k\infty5}. \end{equation} Comparing Equations~(\ref{eq:infinity-circle}) and (\ref{eq:Stokes-definition}), one arrives at \begin{equation} \label{eq:M-infty-stokes-k} M_{k\infty5}=S_kS_{k+1}e^{\pi i\Theta_{\infty5}\sigma_3}. \end{equation} In the following we set $$ M_{\infty5}\equiv M_{0\infty5}. $$ All the others $M_{k\infty5}$ can be expressed in terms of $M_{\infty5}$ via the recurrence formula $$ M_{k+1\infty5}=S_k^{-1}M_{k\infty5}S_k $$ and Equation~(\ref{eq:M-infty-stokes-k}). The monodromy matrices $M_{\nu5}$, $\nu=0,1$, at the regular singularities $\lambda=\nu$ of Equation~(\ref{eq:linear-5}) are defined with the help of the paths given in Figure~7 by the same formulae (\ref{eq:psi-expansion-at-nu}), (\ref{eq:condition-det-1}), (\ref{eq:monodromy-local}), and (\ref{eq: trace-M}) with the subscript $6$ changed to $5$. \begin{figure}[ht] \begin{center} \begin{picture}(160,100) \put(30,80){\circle*{3}} \put(28,66){$0$} \put(80,80){\circle*{3}} \put(72,66){$1$} \put(150,80){\circle*{3}} \put(136,70){$\infty$} \put(20,20){\circle*{3}} \put(5,15){$\lambda_0$} \qbezier(20,20)(60,100)(30,100) \qbezier(20,20)(10,100)(30,100) \qbezier(20,20)(118,80)(90,98) \qbezier(20,20)(72,106)(90,98) \qbezier(20,20)(188,68)(164,90) \qbezier(20,20)(160,110)(164,90) \put(80,62){\vector(1,1){10}} \put(142,62){\vector(2,1){10}} \put(37,58){\vector(1,3){5}} \put(60,00){Figure 7} \end{picture} \begin{picture}(40,90) \put(70,30){\circle*{3}} \put(68,36){$0$} \put(120,30){\circle*{3}} \put(112,36){$1$} \put(190,30){\circle*{3}} \put(176,36){$\infty$} \put(60,90){\circle*{3}} \put(45,85){$\tilde\lambda_0$} \qbezier(60,90)(90,25)(70,20) \qbezier(60,90)(45,25)(70,20) \qbezier(60,90)(155,28)(130,20) \qbezier(60,90)(94,22)(130,20) \qbezier(60,90)(188,08)(214,20) \qbezier(60,90)(247,30)(214,20) \put(124,42){\vector(-1,1){10}} \put(189,43){\vector(-2,1){10}} \put(77,41){\vector(-1,4){3}} \end{picture} \begin{picture}(160,100) \put(60,00){Figure 8} \end{picture} \end{center} \end{figure} The cyclic relation reads \begin{equation} \label{eq:P5-cyclic-M} M_{05}M_{15}M_{\infty5}=I. \end{equation} The monodromy group $MG_5$ is a subgroup of ${\rm SL}(2,\mathbb{C})$ generated by the matrices $M_{\nu5}$. So, under the monodromy group we mean a particular monodromy representation of the fundamental group $\pi(\lambda_0,\bar{\mathbb{C}}\setminus\{0,1,\infty\})$. Together with the group $MG_5$ defined above we will use another one, $\widetilde{MG}_5$, which is generated by the monodromy matrices $\widetilde M_{\nu6}\equiv\widetilde M_{0\nu6}$ obtained by the analytic continuation of the same canonical solution $\Psi_5^0(\lambda)$ but along the paths presented in Figure~8. These matrices obey the following cyclic relation: \begin{equation} \label{eq:P5-cyclic-M-tilde} \widetilde M_{\infty5}\widetilde M_{15}\widetilde M_{05}=I \end{equation} The relation between both monodromy groups can be obtained by comparing representations of the fundamental groups in Figures 7 and 8 together with our way of defining a singlevalued branch of the function $\Psi_5^0$ explained in the paragraph below the asymptotics at the point of infinity (\ref{eq:Psi5-asymptotics-infty}): $$ \widetilde M_{15}=M_{05}M_{15}M_{05}^{-1},\quad\widetilde M_{05}=M_{05},\quad \widetilde M_{\infty5}=M_{\infty5}. $$ {\it The sets of monodromy data} ${\cal M}_5$ (or $\widetilde{\cal M}_5$) are ordered sets of the matrix elements $\{M_{\nu5}\}_{\nu=0,1,\infty}$ (or $\{\widetilde{M}_{\nu5}\}_{\nu=0,1,\infty}$) completed with three complex parameters $\Theta_{\nu5}$ ($\nu=0,1,\infty$) satisfying the following equations \begin{equation} \label{eq:M5-traces} \begin{aligned} &\nu=0,1:\;\;{\rm tr}\,M_{\nu5}=2\cos(\pi\Theta_{\nu5}),\quad\Theta_{\nu5}\in\mathbb{C}\setminus\mathbb{Z},\\ &\nu=\infty:\;\;{\rm tr}\,M_{\infty5}=2\cos(\pi\Theta_{\infty5})+e^{-\pi i\Theta_{\infty5}}s_0s_1,\quad \Theta_{\infty5}\in\mathbb{C}. \end{aligned} \end{equation} {\it The manifold of monodromy data} ${\cal M}_5(\Theta_{05},\Theta_{15},\Theta_{\infty5})$ is an algebraic variety defined by Equations~(\ref{eq:P5-cyclic-M}), (\ref{eq:M5-traces}) and $$ \det M_{\nu5}=1,\qquad\nu=0,1,\infty $$ in $\mathbb{C}^{12}$ (we identify an ordered set of matrix elements $\{M_{\nu5}\}_{\nu=0,1,\infty}$ as a point in $\mathbb{C}^{12}$). In an analogous way we define $\widetilde{\cal M}_5(\Theta_{05},\Theta_{15},\Theta_{\infty5})$. It is easy to see that the complex dimension of ${\cal M}_5(\Theta_{05},\Theta_{15},\Theta_{\infty5})$ (respectively, $\widetilde{\cal M}_5(\Theta_{05},\Theta_{15},\Theta_{\infty5})$) equals $3$. Consider, following \cite{JM}, the parametrization of $A_{\nu5}$: \begin{gather} \label{eq:A05-parametrization} A_{05}=\left(\!\!\begin{array}{cc} z_5+\Theta_{05}/2&-u_5(z_5+\Theta_{05})\\ u_5^{-1}z_5&-z_5-\Theta_{05}/2 \end{array}\!\!\right),\\ \label{eq:A15-parametrization} A_{15}=\left(\!\!\begin{array}{cc} -z_5-(\Theta_{05}+\Theta_{\infty5})/2&u_5y_5(z_5+(\Theta_{05}-\Theta_{15}+\Theta_{\infty5})/2)\\ -\frac1{u_5y_5}(z_5+(\Theta_{05}+\Theta_{15}+\Theta_{\infty5})/2)&z_5+(\Theta_{05}+\Theta_{\infty5})/2 \end{array}\!\!\right). \end{gather} We see that for fixed $t_5$ and formal monodromies $\Theta_{\nu5}$ ($\nu=0,1,\infty$) the number of parameters ($u_5$, $z_5$, $y_5$) in Equation~(\ref{eq:linear-5}) is $3$: it exactly coincides with ${\rm dim}\,{\cal M}_5(\Theta_{05},\Theta_{15},\Theta_{\infty5})$. The hard isomonodromy condition: $\partial_{t_5}\Theta_{\nu5}=0$ and $\partial_{t_5}M_{\nu5}=0$, for $\nu=0,1,\infty$ implies an additional ODE for the function $\Psi_5$ with respect to $t_5$: \begin{equation} \label{eqP5-linear-t5} \frac{d\Psi_5}{dt_5}=\left(\frac\lambda2\sigma_3+ \frac1{t_5}\left(\frac{\Theta_{\infty5}}2\sigma_3+A_{05}+A_{15}\right)\!\!\right)\Psi_5. \end{equation} The compatibility condition of Equations~(\ref{eq:linear-5}) and (\ref{eqP5-linear-t5}) implies that the matrices $A_{\nu5}$ and hence the parameters: $u_5$, $z_5$, and $y_5$, are functions of $t_5$. These functions are governed by the system of isomonodromy deformations, \begin{equation} \label{eq:system-idm5} \frac{dA_{\nu5}}{dt_5}=\left[\frac1{t_5}\left(\frac{\Theta_{\infty5}}2\sigma_3+A_{05}+A_{15}\right)+ \frac\nu2\sigma_3,\,A_{\nu5}\right],\qquad\nu=0,1. \end{equation} In terms of the parameters $u_5$, $z_5$, and $y_5$, this system is given in \cite{JM}: eliminating the function $z_5$ from these equations one finds that $y_5(t_5)$ solves the fifth Painlev\'e equation: \begin{equation} \label{eq:P5} \begin{gathered} \frac{d^2y_5}{dt_5^2}=\!\left(\frac1{2y_5}+\frac1{y_5-1}\right)\!\!\left(\frac{dy_5}{dt_5}\right)^2- \frac1{t_5}\frac{dy_5}{dt_5}+\left(\frac{y_5-1}{t_5}\right)^2\!\! \left(\alpha_5y_5+\frac{\beta_5}{y_5}\right)\!+\\ \gamma_5\frac{y_5}{t_5}+\delta_5\frac{y_5(y_5+1)}{y_5-1}, \end{gathered} \end{equation} \begin{equation} \label{eq:P5coeff} \begin{gathered} \alpha_5=\frac12\left(\frac{\Theta_{05}-\Theta_{15}+\Theta_{\infty5}}2\right)^2,\quad \beta_5=-\frac12\left(\frac{\Theta_{05}-\Theta_{15}-\Theta_{\infty5}}2\right)^2,\\ \gamma_5=1-\Theta_{05}-\Theta_{15},\quad\delta_5=-\frac12. \end{gathered} \end{equation} There is an ambiguity in the definition of the formal monodromies $\Theta_{\nu5}\to-\Theta_{\nu5}$, $\nu=0,1$, as well as in the case for the function $\Psi_6$. The change $\Theta_{\nu5}\to-\Theta_{\nu5}$ leads to a repa\-ra\-met\-rization of the matrices $A_{\nu5}$. Contrary to the above case for $P_6$ this reparametrization yields a nontrivial transformation of the solution $y_5$. We won't discuss it here. It is important for us that we can choose the signs of $\theta_{\nu5}$, $\nu=0,1$, arbitrarily, and then use the corresponding parametrization of $A_{\nu5}$. The $\tau$-function for the isomonodromy deformations~(\ref{eq:system-idm5}) is defined in \cite{JM} as follows: \begin{equation} \label{eq:tau5-via-Psi} \frac{d}{dt_5}\ln\,\tau_5(t_5)=-\frac12{\rm tr}\,(\Psi_{1\infty5}\sigma_3). \end{equation} In terms of the matrix elements Equation~(\ref{eq:system-idm5}) reads: \begin{equation} \label{eq:tau5-yz} \begin{gathered} \frac{d}{dt_5}\ln\,\tau_5(t_5)=(A_{15})_{11}+\frac1{t_5}(A_{05}+A_{15})_{21}(A_{05}+A_{15})_{12}= -z_5-\frac{\Theta_{05}+\Theta_{\infty5}}2\\-\frac1{t_5}\left(z_5-\frac1{y_5} \left(\!\!z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right)\!\left(\!\!z_5+\Theta_{05}- y_5\!\left(\!z_5+\frac{\Theta_{05}-\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right). \end{gathered} \end{equation} We can also rewrite Equation~(\ref{eq:tau5-yz}) in the ``matrix'' form: \begin{equation} \label{eq:tau5-matrix-form} \frac{d}{dt_5}\ln\,\tau_5(t_5)=\frac1{t_5}\left(\left(\frac{\Theta_{05}}2\right)^2+ \left(\frac{\Theta_{15}}2\right)^2-\left(\frac{\Theta_{\infty5}}2\right)^2\right)+ {\rm tr}\,\Big(\frac1{t_5}A_{05}+\frac{\sigma_3}2\Big)A_{15}. \end{equation} In \cite{JM} the function $\sigma_5(t_5)$: \begin{equation} \label{eq:sigma5-def} \sigma_5(t_5)=\frac{\Theta_{05}+\Theta_{\infty5}}2t_5+t_5\frac{d}{dt_5}\ln\,\tau_5(t_5). \end{equation} This function satisfies a second-order ODE which is quadratic with respect to $\sigma_5^{''}$ (see \cite{JM}). Differentiating (\ref{eq:sigma5-def}) and using (\ref{eq:tau5-matrix-form}), (\ref{eq:system-idm5}), and parametrization~(\ref{eq:A05-parametrization}), (\ref{eq:A15-parametrization}), we find that \begin{equation} \label{eq:sigma'=-z5} \frac{d\sigma_5}{dt_5}=-z_5. \end{equation} In the corresponding formula (C.44) of \cite{JM} there is a misprint in the sign: this sign is important for us to establish the differentiable character of our asymptotic expansion for the function $\hat\sigma_6(t_6)$ (see Subsection~{\bf II.4} of Section~\ref{sec:3}). \section{Formal Limit Transitions $P_6\to P_5$} \label{sec:3} As mentioned in the Introduction we consider here two different limits. Our scheme for the derivation of these limits consists of the following steps: \begin{enumerate} \item A formal limit passage of Equation~(\ref{eq:linear-4-regular}) to Equation~(\ref{eq:linear-5}); \item Finding conditions which guarantee that simultaneously with the limit passage in item~1 we have the limit passage of Equation~(\ref{eq:t-6}) to Equation~(\ref{eqP5-linear-t5}). Actually these conditions are additional to those found in the first step. One proves that if the asymptotic expansions found in the first step are differentiable with respect to $t_5$, then it is possible to define asymptotic expansions for the functions $s_{\nu6}$, $\nu=1,t$, (see Equation~(\ref{eq:A-and-R})) to satisfy additional conditions appearing at this step. To prove that the asymptotics we found are really differentiable, one has to use the systems of isomonodromy deformations~(\ref{eq:schlesinger-p6}) (or, in terms of the matrix elements, Equations (C.51), (C.52), and (C.55) of \cite{JM}) and Equation~(\ref{eq:system-idm5}) (or (C.40) of \cite{JM}). We leave this proof to the reader and write down only the formulae for $s_{\nu6}$. \item Presentation of the formal limit passage in terms of the matrix elements of $A_{\nu6}$ and $A_{\nu5}$. \item Presentation of the limit as (formal) asymptotics for the $P_6$- and $\tau_6$-functions. \end{enumerate} To eliminate possible confusion, let us agree to supply the parameters $\lambda$ from Equations (\ref{eq:linear-4-regular}) and (\ref{eq:linear-5}) with the subscripts 6 or 5, respectively. {\bf I.1.} The first limit passage: \begin{gather} \label{eq:lim1-t-lambda} \varepsilon\to+0,\quad \Theta_{16}=-\frac1\varepsilon,\quad t_6=\varepsilon t_5={\cal O}(\varepsilon),\quad \lambda_6=\varepsilon t_5\lambda_5=o(\varepsilon),\\ \label{eq:lim1-Psi} \underset{\varepsilon\to+0}\lim\,R_{16}^{-1}\Psi_6(\lambda_6,t_6)=\Psi_5(\lambda_5,t_5),\\ \label{eq:lim1-A} R_{16}^{-1}A_{06}R_{16}=A_{05}+{\cal O}(\varepsilon),\qquad R_{16}^{-1}A_{t6}R_{16}=A_{15}+{\cal O}(\varepsilon). \end{gather} {\bf I.2.} \begin{equation} \label{eq:lim1-t-R} -R_{16}^{-1}\frac{d}{dt_5}R_{16}=\frac{\Theta_{\infty5}}{2t_5}\sigma_3+{\cal O}(\varepsilon), \end{equation} where $\Theta_{\infty5}$ appears as a parameter of the limit passage, i.e., an arbitrary complex number, its notation as one of the formal monodromies related with the fact that in derivation of Equation~(\ref{eq:lim1-t-R}) we took into account relation~(\ref{eq:diag-A5}). {\bf I.3.} In terms of the matrix elements $A_{\nu6}$ and $A_{\nu5}$ we can write the limit {\bf I} in two possible ways. So for the matrix elements we get the two different formal limits {\bf I.3.a.} and {\bf I.3.b.}. Both limits are related via the transformation generated by the Reflection~(\ref{eq:reflection-infinity}); however, the leading terms of one limit do not completely define the leading terms of the other: some further terms of the expansions are needed. So the formulae given below for the limit {\bf I.3.a.} do not completely define asymptotics {\bf I.3.b.} {\bf I.3.a.} \begin{equation} \label{eq:13a-theta} \Theta_{16}=-\frac1\varepsilon,\quad \Theta_{\infty6}+\Theta_{16}=\Theta_{\infty5},\quad \Theta_{06}=\Theta_{05},\quad \Theta_{t6}=-\Theta_{15}. \end{equation} In Section~\ref{sec:2}, in the paragraph following Equation~(\ref{eq:P5coeff}), we explained that the signs in the last two equations of (\ref{eq:13a-theta}) can be taken arbitrarily; however, from their choice, the following formulae are strongly depended. \begin{gather} \label{eq:13a-z06-zt6} z_{06}=z_5+{\cal O}(\varepsilon),\quad z_{t6}=-z_5-\frac{\Theta_{05}-\Theta_{15}+\Theta_{\infty5}}2+{\cal O}(\varepsilon),\\ z_{16}=-\varepsilon\left(z_5- \frac1{y_5}\left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right)\times \nonumber\\ \left(z_5+\Theta_{05}-y_5\left(z_5+\frac{\Theta_{05}-\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right)+ {\cal O}(\varepsilon^2), \label{eq:13a-z16}\\ u_{06}s_{16}^2=\varepsilon^2u_5\frac{z_5+\Theta_{05}}{z_5}+{\cal O}(\varepsilon^3),\qquad u_{t6}s_{16}^2=\varepsilon^2y_5u_5+{\cal O}(\varepsilon^3), \label{eq:13a-u06-ut6}\\ u_{16}s_{16}^2=\frac{\varepsilon u_5}{z_5-\frac1{y_5} \left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)}+{\cal O}(\varepsilon^2). \label{eq:13a-u16} \end{gather} As follows from Equation~(\ref{eq:lim1-t-R}) the function $s_{16}$ must satisfy the equation \begin{equation} \label{eq:13a-s16} \frac{d}{dt_5}\ln\,s_{16}=\frac{\Theta_{\infty5}}{2t_5}+{\cal O}(\varepsilon). \end{equation} {\bf I.3.b.} \begin{gather} \label{eq:13b-Theta} \Theta_{16}=-\frac1\varepsilon,\quad \Theta_{16}-\Theta_{\infty6}=\Theta_{\infty5},\quad \Theta_{06}=\Theta_{05},\quad \Theta_{t6}=-\Theta_{15},\\ \label{eq:13b-z06-zt6} z_{06}=-z_5-\Theta_{05}+{\cal O}(\varepsilon),\quad z_{t6}=z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2+{\cal O}(\varepsilon),\\ \label{eq:13b-z16} z_{16}=\frac1\varepsilon+\varepsilon\left(z_5- \frac1{y_5}\left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right)\times \nonumber\\ \left(z_5+\Theta_{05}-y_5\left(z_5+\frac{\Theta_{05}-\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right)+ {\cal O}(\varepsilon^2),\\ \label{eq:13b-u16/ut6} \frac{u_{16}}{u_{t6}}=\varepsilon y_5\left(z_5- \frac1{y_5}\left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right)+ {\cal O}(\varepsilon^2),\\ \label{eq:13b-u16/u06} \frac{u_{16}}{u_{06}}=\varepsilon\frac{z_5+\Theta_{05}}{z_5}\left(z_5- \frac1{y_5}\left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right)+ {\cal O}(\varepsilon^2),\\ \label{eq:13b-u16s16} u_{16}s_{16}^2=-\frac{\varepsilon u_5}{z_5-\frac1{y_5} \left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)}+{\cal O}(\varepsilon^2),\\ \label{eq:13b-s16} \frac{d}{dt_5}\ln\,s_{16}=\frac{d}{dt_5}\ln\frac{u_5}{z_5-\frac1{y_5} \left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)} -\frac{\Theta_{\infty5}}{2t_5}+{\cal O}(\varepsilon). \end{gather} {\bf I.4.} Substituting Asymptotics (\ref{eq:lim1-A}) and (\ref{eq:lim1-t-lambda}) into Equation~(\ref{eq:tau-6}) and using definition~(\ref{eq:tau5-matrix-form}) one finds that \begin{equation} \label{eq:tau-6-asympt} \frac{d}{dt_5}\ln\tau_6(t_6)= \frac1{t_5}\left(\left(\frac{\Theta_{\infty5}}2\right)^2-\left(\frac{\Theta_{05}}2\right)^2- \left(\frac{\Theta_{15}}2\right)^2\right)+\frac{d}{dt_5}\ln\tau_5(t_5)+{\cal O}(\varepsilon). \end{equation} Substituting Asymptotics (\ref{eq:13a-z06-zt6}) and (\ref{eq:13a-z16}) into Equation~(\ref{eq:y6}) we find for the limit {\bf I.3.a.} \begin{equation} \label{eq:13a-y6-asympt} y_6(t_6)=\frac{\varepsilon t_5}{1+y_5\left(1- \frac{\Theta_{05}+\Theta_{15}-\Theta_{\infty5}}{2(z_5+\Theta_{\infty5})}\right)}+ {\cal O}(\varepsilon^2). \end{equation} In case {\bf I.3.b.} Equations~(\ref{eq:13b-z06-zt6}), (\ref{eq:13b-u16/ut6}) and (\ref{eq:13b-u16/u06}) yield \begin{equation} \label{eq:13b-y6-asympt} y_6(t_6)=\frac{\varepsilon t_5}{1+\frac1{y_5}\left(1+ \frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}{2z_5}\right)}+{\cal O}(\varepsilon^2). \end{equation} Now we consider the second formal limit. {\bf II.1.} \begin{gather} \label{eq:lim2-t-lambda} \varepsilon\to+0,\quad \Theta_{t6}=-\frac1\varepsilon,\quad t_6=\varepsilon t_5={\cal O}(\varepsilon),\quad \lambda_6=\frac1{\lambda_5},\quad \frac\varepsilon{\lambda_6^3}=o(1),\\ \label{eq:lim2-Psi} \underset{\varepsilon\to+0}\lim R_{t6}^{-1}\Psi_6(\lambda_6,t_6)=\Psi_5(\lambda_5,t_5),\\ \label{eq:lim2-A05-A15} R_{t6}^{-1}\frac{\Theta_{\infty6}}2\sigma_3R_{t6}=A_{05}+{\cal O}(\varepsilon),\qquad R_{t6}^{-1}A_{16}R_{t6}=A_{15}+{\cal O}(\varepsilon) \end{gather} {\bf II.2.} \begin{equation} \label{eq:lim2-Rt6-derivative} -R_{t6}^{-1}\frac{d}{dt_5}R_{t6}=\frac1{t_5}\left(\frac{\Theta_{\infty5}}2\sigma_3+A_{05}+A_{15}\right)+ {\cal O}(\varepsilon), \end{equation} where $\Theta_{\infty5}$ is, as for the first limit, a parameter of the limit passage satisfying Equation~(\ref{eq:diag-A5}). {\bf II.3.} \begin{equation} \label{eq:lim2-theta} \Theta_{t6}=-\frac1\varepsilon,\quad \Theta_{t6}+\Theta_{06}=\Theta_{\infty5},\quad \Theta_{\infty6}=\Theta_{05},\quad \Theta_{16}=\Theta_{15}. \end{equation} Together with the case $\Theta_{t6}+\Theta_{06}=\Theta_{\infty5}$ we can consider another one: $\Theta_{t6}-\Theta_{06}=\Theta_{\infty5}$. As explained in Section~\ref{sec:2}, transformation $\Theta_{06}\to-\Theta_{06}$ simply means that the reparametrization of the matrix $A_{06}$, which does no effect on both the $\tau_6$- and the $P_6$-functions. Thus, contrary to the first limit, it is not worthwhile to consider separately these possibilities: \begin{gather} \label{eq:II-zt6-z06} z_{t6}=-\frac{z_5}{\varepsilon\Theta_{05}}+{\cal O}(1),\qquad z_{06}=\frac{z_5}{\varepsilon\Theta_{05}}+{\cal O}(1),\\ \label{eq:II-z16} z_{16}+\Theta_{16}=\left(\frac{z_5(1-y_5)}{\Theta_{05}}+1\right)\!\! \left(\Theta_{15}+\frac{1-y_5}{y_5}\left(z_5+ \frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right)+{\cal O}(\varepsilon),\\ \label{eq:II-ut6-u06} u_{t6}=-\varepsilon^2\frac{\Theta_{05}u_5}{z_5s_{t6}^2}+{\cal O}(\varepsilon^3),\qquad u_{06}=u_{t6}+{\cal O}(\varepsilon^3),\\ \label{eq:II-u16} u_{16}=u_{t6}\frac{\Theta_{15}+\left(\frac1{y_5}-1\right)\left(z_5+ \frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)} {\Theta_{15}+\left(\frac1{y_5}-1\right)\left(z_5+ \frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\left(1+\frac{\Theta_{05}}{z_5(1-y_5)}\right)}+ {\cal O}(\varepsilon^3),\\ \label{eq:II-st6} \frac{d}{dt_5}\ln\,s_{t6}=\frac{z_5}{\Theta_{05}}\frac{d}{dt_5}\ln\,u_{t6}+{\cal O}(\varepsilon),\\ \label{eq:II-ut6} t_5\frac{d}{dt_5}\ln\,u_{t6}=-\frac{\Theta_{05}}{z_5}\left(z_5- \frac1{y_5}\left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right) +{\cal O}(\varepsilon),\\ \label{eq:II-ut6zt6st6} t_5\frac{d}{dt_5}\ln(u_{t6}z_{t6}s_{t6})=-\left(z_5+\Theta_{05}- y_5\left(z_5+\frac{\Theta_{05}-\Theta_{15}+\Theta_{\infty5}}2\right)\!\!\right) +{\cal O}(\varepsilon). \end{gather} Equations (\ref{eq:II-ut6}) and (\ref{eq:II-ut6zt6st6}) play a more important role in the proof of the consistency of the second limit with the isomonodromy condition~(\ref{eq:lim2-Rt6-derivative}) than the analogous formulae for the first limit, that is why they are written here explicitly. Note that $\Theta_{05}\neq0$ due to the third equation in (\ref{eq:lim2-theta}) and the condition $\Theta_{\infty6}\neq0$ which is imposed in Section~\ref{sec:2}. {\bf II.4.} Substituting Equations~(\ref{eq:lim2-t-lambda}) and (\ref{eq:lim2-A05-A15}) into (\ref{eq:tau-6}) and using normalization~(\ref{eq:niormalization-for-P6}) we find the following asymptotics for the $\tau_6$-function: \begin{equation} \label{eq:II-tau6-asympt-rough} \frac{d}{dt_5}\ln\,\tau_6(t_6)=-\frac1{2\varepsilon^2t_5}-\frac{\Theta_{\infty5}}{2\varepsilon t_5}+ {\cal O}(1). \end{equation} Thus, the most interesting term of the asymptotic expansion (\ref{eq:II-tau6-asympt-rough}), that is ${\cal O}(1)$, cannot be obtained directly from our result: it requires more precise expansions in (\ref{eq:II-zt6-z06}) (up to ${\cal O}(\varepsilon)$). Nevertheless, there is the following trick to overcome this difficulty. Using Equations (\ref{eq:tau-6}), (\ref{eq:sigma-6}), (\ref{eq:schlesinger-p6}), and (\ref{eq:niormalization-for-P6}), we, following \cite{JM}, find \begin{equation} \label{eq:II-sigma-hat-derivative} \frac{d\hat\sigma_6(t_6)}{dt_6}=-{\rm tr}\left(\frac{\Theta_{\infty6}}2\sigma_3A_{t6}\right)- {\rm tr}\,A_{t6}^2. \end{equation} Now Equations~(\ref{eq:lim2-A05-A15}) yield \begin{equation} \label{eq:II-sigma-hat-deriv-A05-asympt} t_6\frac{d\hat\sigma_6(t_6)}{dt_6}=\frac{t_5}2{\rm tr}\,(A_{05}\sigma_3)-\frac{t_5}{2\varepsilon}+ {\cal O}(\varepsilon). \end{equation} Applying now (\ref{eq:diag-A5}) one obtains \begin{equation} \label{eq:II-sigma-hat-deriv-A15-asympt} t_6\frac{d\hat\sigma_6(t_6)}{dt_6}=-\frac{t_5}2\left(\frac1{\varepsilon}+\Theta_{\infty5}+ {\rm tr}(A_{15}\sigma_3)\right)+{\cal O}(\varepsilon). \end{equation} Now from (C.59) or directly from Equation~(\ref{eq:II-sigma-hat-derivative}) we find that \begin{gather*} t_6\frac{d\hat\sigma_6(t_6)}{dt_6}-\hat\sigma_6(t_6)={\rm tr}\,A_{06}A_{t6}= \frac12({\rm tr}((A_{06}+A_{t6})^2-{\rm tr}(A_{06})^2-{\rm tr}(A_{t6})^2)\\ =\frac12\big({\rm tr}\left(\frac{\Theta_{\infty6}}2\sigma_3+A_{16}\right)^2- {\rm tr}(A_{06})^2-{\rm tr}(A_{t6})^2\big)\\ =\frac{\Theta_{\infty6}}2{\rm tr}(A_{16}\sigma_3)+ \frac12\left({\rm tr}\left(\frac{\Theta_{\infty6}}2\sigma_3\right)^2+{\rm tr}\,(A_{16})^2- {\rm tr}\,(A_{06})^2-{\rm tr}\,(A_{t6})^2\right). \end{gather*} Substituting into the latter formula Equations~(\ref{eq:lim2-A05-A15}) and (\ref{eq:II-sigma-hat-deriv-A15-asympt}) and using definitions (\ref{eq:sigma-6}) and (\ref{eq:tau5-matrix-form}) we obtain \begin{gather} \frac{d}{dt_5}\ln\tau_6(t_6)=\left(\frac1{t_5}+\varepsilon+\varepsilon^2t_5\right)\!\! \left(-\frac1{2\varepsilon^2}-\frac{\Theta_{\infty5}-t_5}{2\varepsilon}+ t_5{\rm tr}\Big(A_{15}\left(\frac1{t_5}A_{05}+\frac{\sigma_3}2\right)\!\!\Big)\right. +\frac{\Theta_{\infty5}t_5}2\nonumber\\ \label{eq:II-tau6-asympt} \left.+\frac{\Theta_{05}^2+\Theta_{15}^2-\Theta_{\infty5}^2}2\right) +{\cal O}(\varepsilon)=-\frac1{2\varepsilon^2t_5}-\frac{\Theta_{\infty5}}{2\varepsilon t_5}+ \frac{d}{dt_5}\ln\,\tau_5(t_5)+{\cal O}(\varepsilon). \end{gather} The second bracket on the r.h.s. of Equation (\ref{eq:II-tau6-asympt}) is nothing but the asymptotic expansion (up to the order ${\cal O}(\varepsilon)$ for $-\hat\sigma_6(t_6)$. Differentiating it with $t_5$ and using Equation~(\ref{eq:sigma'=-z5}) we find exactly Equation ~(\ref{eq:II-sigma-hat-deriv-A15-asympt}). In an analogous manner the differentiable character of the asymptotic expansions for $u_{\nu6}$ and $z_{\nu6}$ can be proved. Inserting into the first Equation~(\ref{eq:y6}) asymptotics (\ref{eq:II-z16}) -- (\ref{eq:II-u16}) we find that \begin{equation} \label{eq:II-y6-asympt} y_6(t_6)=\frac{1+{\cal O}(\varepsilon)}{1+\frac{y_5-1}{t_5}\left(z_5+ \frac{\Theta_{05}-\Theta_{15}+\Theta_{\infty5}}2- \frac1{y_5}\left(z_5+\frac{\Theta_{05}+\Theta_{15}+\Theta_{\infty5}}2\right)\!\right)}. \end{equation} \section{The First Limit} \label{sec:4} In this section we study the situation when the $\Psi_6$ function in the neighborhood of the isomonodromic cluster of two regular singularities is described via the $\Psi_5$-function, i.e., the function with the irregular singular point. First of all let's define this cluster keeping in mind the formulae (\ref{eq:lim1-t-lambda}) -- (\ref{eq:13a-s16}) corresponding to the first limit {\bf I.3.a.} We won't consider the limit {\bf I.3.b.}, as it can be obtained via the transformation (\ref{eq:reflection-infinity}). Let us begin with the precise setting. We are considering the $\Psi_6(\lambda_6,t_6)$ function defined as in Section~\ref{sec:2}. We assume that its manifold of monodromy data $$ {\cal M}_6(\Theta_{06},\Theta_{16},\Theta_{t6},\Theta_{\infty6}) $$ is given. It is also assumed that \begin{gather} \label{eq:limit1-theta-nu-notZ} \Theta_{\nu6}\in\mathbb{C}\setminus\mathbb{Z},\qquad\nu=0,1,t,\infty,\\ \Theta_{\infty5}=\Theta_{06}+\Theta_{\infty6}\in\mathbb{C}. \label{eq:theta-infty5-definition} \end{gather} The function $\Psi_6(\lambda_6,t_6,n)$ is a GSD of $\Psi_6(\lambda_6,t_6)$ defined by the following recurrence procedure: \begin{gather*} \Psi_6(\lambda_6,t_6,0)=\Psi_6(\lambda_6,t_6),\\ \Psi_6(\lambda_6,t_6,n+1)={\cal L}_{1\infty}^{-+}\Psi_6(\lambda_6,t_6,n),\quad n\in\mathbb{Z}_+=\{0,1,\ldots\}. \end{gather*} The monodromy manifold ${\cal M}_6(\Theta_{06},\Theta_{16}^n,\Theta_{t6},\Theta_{\infty6}^n)$ for $\Psi_6(\lambda_6,t_6,n)$ coincides with that for $\Psi_6(\lambda_6,t_6)$ except for the values of the two parameters, $\Theta_{16}^n$ and $\Theta_{\infty6}^n$: \begin{gather} \label{eq:theta-16n-epsilon-n} \Theta_{16}^n=\Theta_{16}-2n\equiv-\frac1{\varepsilon_n},\qquad n\in\mathbb{Z}_+,\\ \label{eq:theta-infty6n} \Theta_{16}^n+\Theta_{\infty6}^n=\Theta_{\infty5} \end{gather} Equation~(\ref{eq:theta-16n-epsilon-n}) is the definition of the discrete small parameter $\varepsilon=\varepsilon_n\to+0$, while the parameter $\Theta_{\infty5}$ in Equation~(\ref{eq:theta-infty6n}) is defined by Equation~(\ref{eq:theta-infty5-definition}). The main object of our investigation are the GSD's $A_{\nu6}^n(t_6)$, $\nu=0,1,t$. They can be defined by means of Equation~(\ref{eq:linear-4-regular}): \begin{equation} \label{eq:1-A-GSD-Psi} \partial_{\lambda_6}\Psi_6(\lambda_6,t_6,n)\Psi_6^{-1}(\lambda_6,t_6,n)=\sum\limits_{\nu=0,1,t} \frac{A_{\nu6}^n(t_6)}{\lambda-\nu}. \end{equation} An alternative (equivalent) definition of $A_{\nu6}^n(t_6)$ (without the usage of the auxiliary object $\Psi_6(\lambda_6,t_6,n)$) can be given as the following recurrence procedure: \begin{equation} \label{eq:1-A-GSD-rec} A_{\nu6}^0(t_6)=A_{\nu6}(t_6),\qquad A_{\nu6}^{n+1}={\tilde A}_{\nu6}^n(t_6), \end{equation} where $A_{\nu6}(t_6)$ is some solution of (\ref{eq:schlesinger-p6}), and ${\tilde A}_{\nu6}^n(t_6)$ is obtained via Equations~(\ref{eq:A-tilde-mu-infty}) -- (\ref{eq:Atilde-infty2}) with $J_{1\infty}^{-+}$ by inserting $A_{\nu6}^n(t_6)$ into the r.-h.s.'s. In fact, in Theorem~\ref{th:1} we assume that the matrices $\{A_{\nu6}^0(t_6)\}_{\nu=0,1,t}$ correspond to the manifold of monodromy data ${\cal M}_6(\Theta_{06},\Theta_{16},\Theta_{t6},\Theta_{\infty6})$; thus the function $\Psi_6(\lambda_6,t_6,n)$ is also implicitly presented in the second definition of $A_{\nu6}^n(t_6)$. We interpret the formal limit transition {\bf I} as asymptotics (as $n\to+\infty$) of the even sequences $\{A_{\nu6}^{2n}(\epsilon_{n} t_5\}_{\nu=0,1,t}$, or as asymptotics of the corresponding sequences of their matrix elements. The asymptotic behavior of the odd sequences $\{A_{\nu6}^{2n+1}(\epsilon_{n}t_5\}_{\nu=0,1,t}$ is given by exactly the same formulae as for the even one (see formal limit {\bf I} in Section~\ref{sec:3} with $\varepsilon\to\varepsilon_{n}$. The monodromy data for ``odd'' limit can be obtained from the monodromy data for the ``even'' limit by simply changing in the latter formulae $\Theta_6\to\Theta_6-1$. Sometimes, when it does not cause any confusion, we omit subscripts/superscripts $n$. To make a difference between the initial values of the parameters $\Theta_{16}$ and $\Theta_{\infty6}$ in Equations~(\ref{eq:limit1-theta-nu-notZ}), (\ref{eq:theta-infty5-definition}) and the parameters $\Theta_{16}^n$ and $\Theta_{\infty6}^n$, we agree to denote the pair of initial values as \begin{equation} \label{eq:I-initial-theta} \Theta_{16}^0\equiv\Theta_6\in\mathbb{C}\setminus\mathbb{Z},\qquad \Theta_{\infty6}^0=\Theta_{\infty5}-\Theta_6\in\mathbb{C}\setminus\mathbb{Z}, \end{equation} while, instead of $\Theta_{16}^n$ and $\Theta_{\infty6}^n$, we write $\Theta_{16}$ and $\Theta_{\infty6}$, respectively. {\it Thus, hereafter in Section~{\rm\ref{sec:4}}, $\Theta_{6}$ and $\Theta_{\infty5}$ are fixed according to {\rm(\ref{eq:I-initial-theta})} and {\rm(\ref{eq:theta-infty5-definition})}, while $\Theta_{16}$ and $\Theta_{\infty6}$ are dependent on $n$ such that $\Theta_{16}\underset{n\to+\infty}\to-\infty$ and $\Theta_{\infty6}\underset{n\to+\infty}\to+\infty$ and the condition~{\rm(\ref{eq:limit1-theta-nu-notZ})} holds}. In fact the formal limit {\bf I} contains one more parameter $f_0\in\mathbb{C}\setminus\{0\}$. This parameter is hidden as the constant of integration in Equation~(\ref{eq:13a-s16}). The asymptotic expansions of the sequences under investigation \begin{equation} \label{eq:I-even-A} A_{\nu6}^{2n}(\epsilon_{n}t_5),\qquad n\in\mathbb{Z}_+,\quad\nu=0,1,t, \end{equation} don't depend on $f_0$. On the language of the formulae~(\ref{eq:13a-z06-zt6}) -- (\ref{eq:13a-u16}) this means that if $s_{16}\to s_{16}f_0$, then $u_5\to u_5f_0^2$. Nevertheless we include $f_0$ for completeness. Now, denoting the matrix elements of (\ref{eq:I-even-A}) exactly as that for $A_{\nu6}$ in (\ref{eq:A-and-R}), and the matrix elements of the monodromy matrices $M_{\nu6}$ as \begin{equation} \label{eq:I-Mnu6-elements-def} M_{\nu6}=\left(\begin{array}{cc} m_{11}^{\nu6}&m_{12}^{\nu6}\\ m_{21}^{\nu6}&m_{22}^{\nu6} \end{array}\right), \end{equation} we are ready to formulate our result. \begin{theorem} \label{th:1} Assume that the complex parameters $\Theta_06$, $\Theta_6$, $\Theta_{t6}$ satisfy Conditions {\rm(\ref{eq:limit1-theta-nu-notZ})} and {\rm(\ref{eq:I-initial-theta})}. Let a point $\mu_6\in{\cal M}_6(\Theta_{06},\Theta_6,\Theta_{t6},\Theta_{\infty5}-\Theta_6)$ and $\{m_{ik}^{\nu6}\}_{i,k=1,2}^{\nu=0,1,t}$ be its coordinates. Suppose that the following conditions are valid:\\ {\rm1}. $m_{21}^{16}\neq0$;\\ {\rm2}. \begin{equation} \label{eq:I-l-condition} \cos\,\pi l\equiv=im_{11}^{16}\sin\,\pi(\Theta_{\infty5}-\Theta_6)+ \exp(-\pi i(\Theta_{\infty5}-\Theta_6))\cos\,\pi\Theta_6\neq-1; \end{equation}\\ {\rm3}. The inverse monodromy problem defined by the pairs $(\mu_6,t_6)$ are solvable for all $t_6=\varepsilon_{n}t_5$ where $n\geq N\in\mathbb{Z}_+$ and $\varepsilon_{n}$ is defined in Equation~{\rm(\ref{eq:theta-16n-epsilon-n})}, and $t_5\in\mathbb{C}\setminus\{0\}$ with $|\arg t_5|<\pi-\delta$ for some $\delta>0$;\\ {\rm4}. The point $\mu_6$ corresponds to the general, i.e., transcendental solution of $P_6$\footnote[6]{We mean the $P_6$-transcendent, i.e., any solution which cannot be constructed in terms of the logarithmic derivatives of the hypergeometric functions.}. Let $l$ be the unique solution of Equation~(\ref{eq:I-l-condition}) under the conditions: \begin{equation} \label{eq:I-l-conditions} l\neq0,\qquad |{\rm Re}\,l|<1. \end{equation} Define the parameters $\alpha$, $\beta$, $d_0^2$: \begin{equation} \label{eq:alpha-beta-d0} \alpha=\frac{l-\Theta_{\infty5}}2,\qquad \beta=-\frac{l+\Theta_{\infty5}}2,\qquad d_0^2=\frac{m_{21}^{16}}{2\pi i}\Gamma(1-\alpha)\Gamma(1-\beta), \end{equation} where $\Gamma(\cdot)$ is the gamma function~{\rm\cite{BE}}. For arbitrary $f_0\in\mathbb{C}\setminus\{0\}$ define the matrix \begin{equation} \label{eq:I-K} K=-f_0^{\sigma_3}e^{-\frac{\pi i}2\Theta_6\sigma_3} \left(\begin{array}{cc} 1&\frac{\pi}{\Gamma(\alpha)\Gamma(\beta)\sin\,\pi(\Theta_{\infty5}-\Theta_6)}\\ 0&1 \end{array}\right) d_0^{\sigma_3}, \end{equation} and consider the following equations: \begin{gather} \label{eq:I-def-theta's} \Theta_{05}=\Theta_{06},\qquad \Theta_{15}=-\Theta_{t6},\qquad \Theta_{\infty5}=\Theta_{16}+\Theta_{\infty6}=-(\alpha+\beta),\\ \label{eq:I-def-monodromy5} \tilde M_{05}=KM_{06}K^{-1},\qquad \tilde M_{15}=KM_{t6}K^{-1},\qquad \tilde M_{\infty5}=KM_{\infty6}M_{16}K^{-1}, \end{gather} as defining the point $\tilde\mu_5\in\tilde{\cal M}_5(\Theta_{05},\Theta_{15},\Theta_{\infty5})$. Suppose that the inverse monodromy problem for $(\tilde\mu_5,t_5)$ is solvable and the functions $u_5=u_5(\tilde\mu_5,t_5)$, $y_5=y_5(\tilde\mu_5,t_5)$, and $z_5=z_5(\tilde\mu_5,t_5)$ represent this solution. Then it is possible to construct the sequences~{\rm(\ref{eq:I-even-A})}, corresponding to the given manifold ${\cal M}_6(\Theta_{06},\Theta_6,\Theta_{t6},\Theta_{\infty5}-\Theta_6)$. The formulae~{\rm(\ref{eq:13a-theta})--(\ref{eq:13a-s16})} are asymptotic expansions as $n\to+\infty$ of the matrix elements of {\rm(\ref{eq:I-even-A})} if:\\ {\rm1}. The functions $u_5$, $y_5$, and $z_5$ are identified with the solution of the inverse monodromy problem for $(\tilde\mu_5,t_5)$;\\ {\rm2}. Equation~{\rm(\ref{eq:13a-s16})} is supplemented with \begin{equation} \label{eq:I-asympt-s16} s_{16}=\varepsilon f_0d_0(\varepsilon t_5)^{\frac{\Theta_{\infty5}}2}(1+{\cal O}(\varepsilon)); \end{equation} {\rm3}. $\varepsilon=\varepsilon_{n}$ is substituted into {\rm(\ref{eq:13a-theta})--(\ref{eq:13a-s16})} and {\rm(\ref{eq:I-asympt-s16})}. \end{theorem} \begin{remark}{\rm The complex number $l$ is defined by Conditions (\ref{eq:I-l-condition}) and (\ref{eq:I-l-conditions}) up to a sign, which means simply the permutation $\alpha\leftrightarrow\beta$. This permutation doesn't influence the asymptotics that we study, as well as the indefiniteness of the sign $d_0$ in (\ref{eq:alpha-beta-d0}). }\end{remark} \begin{remark}{\rm Some of the conditions imposed on $\mu_6$ (see Conditions $1$, $2$, $4$) are not necessary for the possibility to interpret the formal limit as an asymptotic expansion; however, their violation requires special investigation. In particular, we can apply this theorem not only to the $P_6$ transcendent, but also for most of the solutions that can be constructed via the classical special functions. }\end{remark} \begin{remark}{\rm To find the Stokes multipliers for the limiting $P_5$ equation one can use Equation~(\ref{eq:M-infty-stokes-k}). }\end{remark} \begin{derivation} Here I outline only the calculational scheme for the solution of the direct and inverse monodromy problem, while an explanation of how such calculations work for the justification of the asymptotic expansions, the reader will find in Section~\ref{sec:5} in the derivation of Theorem~\ref{th:2}. Although the derivations are different, the scheme of justification is the same and can be based on the work~\cite{K4}. For some $\delta_{in}$, $\delta_{out}$ such that $\frac13<\delta_{in}<\delta_{out}<1$ and for all rather small $\varepsilon>0$ we can present the $\lambda_6$-complex plane as $\mathbb{C}=\Omega_{in}(\varepsilon)\cup\Omega_{out}(\varepsilon)$, where $$ \lambda_6\in\Omega_{in}(\varepsilon)\Leftrightarrow |\lambda_6|\leq{\cal O}(\varepsilon^{\frac12+\frac{\delta_{in}}2}),\qquad \lambda_6\in\Omega_{out}(\varepsilon)\Leftrightarrow |\lambda_6|\geq{\cal O}(\varepsilon^{\frac12+\frac{\delta_{out}}2}). $$ Thus the matching domain, $\Omega_{mat}(\varepsilon)=\Omega_{in}(\varepsilon)\cap\Omega_{out}(\varepsilon)$, is nonempty for all rather small $\varepsilon$. In the cluster domain we approximate the function $\Psi_6=\Psi_6(\lambda_6,t_6,n)$ as follows \begin{equation} \label{eq:I-psi-approx-in} \Psi_6\underset{\varepsilon\to0}=R_{16}(1+{\cal O}(\varepsilon^\delta))\Psi_5^0K,\qquad \lambda_6\in\Omega_{in}(\varepsilon),\quad\delta>0, \end{equation} where $\Psi_5^0=\Psi_5^0(\lambda_5,t_5)$ is the canonical solution of Equation~(\ref{eq:linear-5}) with monodromy data to be determined, as well as the constant matrix $K\in{\rm SL}(2,\mathbb{C})$. Using Approximation~(\ref{eq:I-psi-approx-in}) one finds Equations~(\ref{eq:13a-theta}) -- (\ref{eq:13a-s16}), (\ref{eq:I-def-monodromy5}), and the first two equations in (\ref{eq:I-def-theta's}). We have to prove (\ref{eq:I-K}), the last equation in (\ref{eq:I-def-theta's}) and (\ref{eq:I-asympt-s16}), as well as for the justification, it is important to establish the following \begin{equation} \label{eq:K-matching} K=\underset{\begin{array}{c}\varepsilon\to0\\\lambda_0\in\Omega_{mat}(\varepsilon)\end{array}}\lim (R_{16})(\Psi_5^0)^{-1}\Psi_{out}, \end{equation} where $\Psi_{out}$ is an approximation for $\Psi_6$ in the domain $\Omega_{out}(\varepsilon)$: \begin{equation} \label{eq:I-psi-out-approx} \Psi_6=(I+{\cal O}(\varepsilon^\delta))\Psi_{out},\qquad\lambda_6\in\Omega_{out}(\varepsilon), \;\;\delta>0. \end{equation} We construct $\Psi_{out}$ by making use of the function $Y(x;\alpha,\beta,\gamma)$ constructed by Jimbo~\cite{J}. While constructing $\Psi_{out}$ we find the conditions (\ref{eq:I-l-condition}) -- (\ref{eq:alpha-beta-d0}) and $m_{21}^{16}\neq0$. Since the function $Y(x;\alpha,\beta,\gamma)$ plays an important role not only in the present derivation but in the one in Section~\ref{sec:5}, we, following~\cite{J}, recall its basic properties: \begin{gather} Y(x)\equiv Y(x;\alpha,\beta,\gamma)x^{\frac{\gamma-1}2}(x-1)^{\frac{\alpha+\beta+1-\gamma}2}\nonumber\\ =\left(\begin{array}{c} F(\alpha,\alpha+1-\gamma,\alpha-\beta;\frac1x,\frac{\beta(\beta+1-\gamma)}{(\beta-\alpha)(\beta-\alpha+1)x} F(\beta,\beta+2-\gamma,\beta-\alpha+2;\frac1x)\\ \frac{\alpha(\alpha+1-\gamma)}{(\alpha-\beta)(\alpha-\beta+1)x} F(\alpha+1,\alpha+2-\gamma,\alpha-\beta+2;\frac1x),F(\beta,\beta+1-\gamma,\beta-\alpha;\frac1x) \end{array}\right) \label{eq:Y(x)-def} \\ \times x^{\frac{\beta-\alpha}2\sigma_3}x^{\frac{\gamma-1-\alpha-\beta}2} (x-1)^{\frac{\alpha+\beta+1-\gamma}2},\nonumber\\ \label{eq:alpha,beta,gamma-conditions-Y} \Theta_{\infty Y}\equiv\alpha-\beta,\;\Theta_{0Y}\equiv1-\gamma,\; \Theta_{1Y}\equiv-\alpha-\beta-1+\gamma\in\mathbb{C\setminus Z}, \end{gather} where $F(\cdot,\cdot,\cdot;\cdot)$ denotes the Gauss hypergeometric function~\cite{BE}. \begin{equation} \label{eq: Y(x)-asympt} Y(x)\underset{x\to\nu=0,1}=G_{\alpha\beta\gamma}^\nu(I+{\cal O}(x-\nu)) (x-\nu)^{\frac{\Theta_{\nu Y}}2\sigma_3}C_{\alpha\beta\gamma}^\nu\underset{x\to\infty} =\left(I+{\cal O}(x^{-1})\right)x^{\frac{\beta-\alpha}2\sigma_3}, \end{equation} where \begin{align} \label{eq:I-G0-G1-alpha-beta-gamma} G_{\alpha\beta\gamma}^0&=\frac1{\beta-\alpha}\left(\begin{array}{cc} \beta+1-\gamma&\beta\\ \alpha+1-\gamma&\alpha \end{array}\right),\quad G_{\alpha\beta\gamma}^1=\frac1{\beta-\alpha}\left(\begin{array}{cc} 1&\beta(\beta+1-\gamma)\\ 1&\alpha(\alpha+1-\gamma) \end{array}\right),\\ \label{eq:C0-alpha-beta-gamma} C_{\alpha\beta\gamma}^0&=\left(\begin{array}{cc} e^{-\pi i(\alpha+1-\gamma)} \frac{\Gamma(\gamma-1)\Gamma(\alpha-\beta+1)}{\Gamma(\gamma-\beta)\Gamma(\alpha)}& -e^{-\pi i(\beta+1-\gamma)} \frac{\Gamma(\gamma-1)\Gamma(\beta-\alpha+1)}{\Gamma(\gamma-\alpha)\Gamma(\beta)}\\ e^{-\pi i\alpha}\frac{\Gamma(1-\gamma)\Gamma(\alpha-\beta+1)}{\Gamma(1-\beta)\Gamma(\alpha+1-\gamma)}& -e^{-\pi i\beta}\frac{\Gamma(1-\gamma)\Gamma(\beta-\alpha+1)}{\Gamma(1-\alpha)\Gamma(\beta+1-\gamma)} \end{array}\right),\\ \label{eq:C1-alpha-beta-gamma} C_{\alpha\beta\gamma}^1&=\left(\begin{array}{cc} -\frac{\Gamma(\alpha+\beta+1-\gamma)\Gamma(\alpha-\beta+1)}{\Gamma(\alpha+1-\gamma)\Gamma(\alpha)}& \frac{\Gamma(\alpha+\beta+1-\gamma)\Gamma(\beta-\alpha+1)}{\Gamma(\beta+1-\gamma)\Gamma(\beta)}\\ -e^{-\pi i(\gamma-\alpha-\beta-1}\frac{\Gamma(\gamma-\alpha-\beta-1)\Gamma(\alpha-\beta+1)}{\Gamma(1-\beta)\Gamma(\gamma-\beta)}& e^{-\pi i(\gamma-\alpha-\beta-1)}\frac{\Gamma(\gamma-\alpha-\beta-1)\Gamma(\beta-\alpha+1)}{\Gamma(1-\alpha)\Gamma(\gamma-\alpha)} \end{array}\right), \end{align} \begin{align} \label{eq:hypergeo-Y(x)} \frac{dY(x)}{dx}&=\left(\frac{A_{0Y}}x+\frac{A_{1Y}}{x-1}\right)Y(x),\qquad A_{0Y}+A_{1Y}=\frac{\beta-\alpha}2\sigma_3,\\ \label{eq:A0Y} A_{0Y}&=\frac1{\beta-\alpha}\left(\begin{array}{cc} -\frac{(\alpha+\beta)(1-\gamma)+2\alpha\beta}2&\beta(\beta+1-\gamma)\\ -\alpha(\alpha+1-\gamma)&\frac{(\alpha+\beta)(1-\gamma)+2\alpha\beta}2 \end{array}\right),\\ \label{eq:A1Y} A_{1Y}&=\frac1{\beta-\alpha}\left(\begin{array}{cc} -\frac{(\alpha+\beta)(1-\gamma)+\alpha^2+\beta^2}2&-\beta(\beta+1-\gamma)\\ -\alpha(\alpha+1-\gamma)&-\frac{(\alpha+\beta)(1-\gamma)+\alpha^2+\beta^2}2 \end{array}\right). \end{align} To the above-mentioned properties of $Y(x)$ pointed out by Jimbo, we add up the following one, which is important in the derivation of our main results in this and the next section: \begin{gather} \label{eq:Y(x)-spec-asympt} Y(x)\underset{x\in(\ref{eq:x-spec-asympt-cond})}=\frac1{\sqrt{\alpha-\beta}} \left(\!\begin{array}{cc} 1&\beta\\ 1&\alpha \end{array}\!\right)\!\! \left(I+{\cal O}\Big(x^{\tilde\delta}\Big)\!\right) \exp\left(\!\!\Big(\frac{1-\gamma}{2x}+\frac{\alpha+\beta}2\ln\frac{1-\gamma}x\Big)\sigma_3\!\right)\! \hat K(\varkappa),\\ \label{eq:K(varkappa)} \hat K(\varkappa)=\left(I+{\cal O}\Big(x^{\tilde\delta}\Big)\!\right)\sqrt{\alpha-\beta}\sigma_3 \left(\!\begin{array}{cc} \frac{\Gamma(\alpha-\beta)}{\Gamma(\alpha)}&\frac{\Gamma(\beta-\alpha)}{\Gamma(\beta)}\\ -e^{i\pi\alpha\varkappa}\frac{\Gamma(1+\alpha-\beta)}{\Gamma(1-\beta)}&e^{i\pi\beta\varkappa} \frac{\Gamma(1+\beta-\alpha)}{\Gamma(1-\alpha)} \end{array}\!\right)(1-\gamma)^{\frac{\beta-\alpha}2\sigma_3}, \end{gather} \begin{equation} \label{eq:x-spec-asympt-cond} \left|\frac{1-\gamma}{x^2}\right|\underset{\varepsilon\to+0}={\cal O}\big(x^{\tilde\delta}\big),\quad \left|\frac{1-\gamma}x\right|\underset{\varepsilon\to+0}=+\infty,\quad -\pi+\frac{\varkappa\pi}2<\arg\frac{1-\gamma}x<\pi+\frac{\varkappa\pi}2,\;\;\varkappa=\pm1. \end{equation} In Equation~(\ref{eq:Y(x)-spec-asympt}), and thereafter, the notation ${\cal O}\big(x^{\tilde\delta}\big)$ means that the corresponding estimate holds for some $\tilde\delta>0$. The precise (the largest possible) value of $\tilde\delta$ in such estimates are not important in our scheme of derivation. For a proof of Asymptotics~(\ref{eq:Y(x)-spec-asympt}) we need a special asymptotic expansion for the Gauss hypergeometric function $F(a,b,c;x)$ which can be found in \cite{BE}: \begin{equation} \label{eq:hyper-asympt-spec} \begin{gathered} F(a,b,c;z)=e^{i\pi a\varkappa}\frac{\Gamma(c)}{\Gamma(c-a)}(bz)^{-a}(1+{\cal O}(|bz|^{-1})) +\frac{\Gamma(c)}{\Gamma(a)}e^{bz}(bz)^{a-c}(1+{\cal O}(|bz|^{-1})),\\ 0<|z|<1,\quad|bz|\to\infty,\quad-\pi+\frac{\varkappa\pi}2<\arg(bz)<\pi+\frac{\varkappa\pi}2,\;\; \varkappa=\pm1. \end{gathered} \end{equation} In the domain~(\ref{eq:x-spec-asympt-cond}) one finds: \begin{equation} \label{eq:x-power-aympt-simp} x^{\frac{\beta-\alpha}2\sigma_3}x^{\frac{\gamma-1-\alpha-\beta}2}(x-1)^{\frac{\alpha+\beta+1-\gamma}2}= x^{\frac{\beta-\alpha}2\sigma_3}e^{-\frac{\alpha+\beta+1-\gamma}{2x}} \left(1+{\cal O}\big(x^{\tilde\delta}\big)\!\right). \end{equation} Substituting Expansions (\ref{eq:hyper-asympt-spec}) and (\ref{eq:x-power-aympt-simp}) into Equation~(\ref{eq:Y(x)-def}) one arrives at Equations~(\ref{eq:Y(x)-spec-asympt})--(\ref{eq:x-spec-asympt-cond}). We construct the function $\Psi_{out}$ as follows: \begin{gather} \Psi_{out}(\lambda_6)=G^{-1}Y(x)C_0^{-1},\qquad x=\frac1{\lambda_6},\nonumber\\ G=G_{\alpha\beta\gamma}^0(\beta-\alpha)(1-\gamma)^{-\frac12}d_0^{\sigma_3} (1-\gamma)^{\frac{\alpha+\beta-1}2\sigma_3}, \label{eq:I-G-Psi-out}\\ C_0=e^{\frac{\pi i}2(1-\gamma+\alpha+\beta)}(1-\gamma)^{\frac12}(\beta-\alpha)^{-1} (1-\gamma)^{\frac{1-\alpha-\beta}2\sigma_3}d_0^{-\sigma_3}C_{\alpha\beta\gamma}^0, \label{eq:I-C-Psi-out} \end{gather} where $\alpha$ and $\beta$, satisfying the conditions~(\ref{eq:alpha,beta,gamma-conditions-Y}), and $d_0\in\mathbb{C}\setminus\{0\}$ are the parameters to be determined, and \begin{equation} \label{eq:I-gamma-cond} 1-\gamma=\Theta_{\infty5}+\frac1{\varepsilon_{n}}\to+\infty,\quad 1-\gamma=\Theta_{\infty5}-\Theta_6+2n,\quad \gamma-\alpha-\beta-1=\Theta_6-2n, \end{equation} as it follows from Equations~(\ref{eq:theta-16n-epsilon-n}), (\ref{eq:I-initial-theta}), (\ref{eq:I-def-theta's}), and (\ref{eq: Y(x)-asympt}). By using (\ref{eq: Y(x)-asympt}) and (\ref{eq:Y(x)-spec-asympt}) - (\ref{eq:x-spec-asympt-cond}) we find that the function $\Psi_{out}$ has the following asymptotic behavior \begin{align} \label{eq:Psi-out-asympt-infty} \lambda_6&\to\infty:&\Psi_{out}&\sim\left(\frac1{\lambda_6}\right)^{\frac{\Theta_{\infty6}}2\sigma_3},\\ \label{eq:Psi-out-asympt-1} \lambda_6&\to1:&\Psi_{out}&\sim G^{-1}G_{\alpha\beta\gamma}^1 \left(\frac1{\lambda_6}-1\right)^{\frac{\Theta_{16}}2\sigma_3}C_{\alpha\beta\gamma}^1C_0^{-1},\\ \lambda_6&\to0,&\lambda_6&\in\Omega_{mat}(\varepsilon),\qquad-\pi+\frac{\varkappa\pi}2<\arg\lambda_6< \pi+\frac{\kappa\pi}2,\nonumber \end{align} \begin{equation} \label{eq:Psi-out-asympt-0} \Psi_{out}\sim\frac{G^{-1}}{\sqrt{\alpha-\beta}} \left(\begin{array}{cc} 1&\beta\\ 1&\alpha \end{array}\right) \exp\left(\!\Big(\frac{(1-\gamma)\lambda_6}2- \frac{\alpha+\beta}2\ln\frac1{(1-\gamma)\lambda_6}\Big)\sigma_3\!\right)\hat K(\varkappa)C_0^{-1}. \end{equation} We see from (\ref{eq:Psi-out-asympt-infty}) that $\Psi_{out}$ satisfies the same normalization condition as the function $\Psi_6$. Thus, the monodromy matrices at $\lambda_6=\infty$ for both functions coincide, $M_{\infty\,out}=M_{\infty6}$. Our next step will be to set the parameters $\alpha-\beta$ and $d_0$ such that the following equation \begin{equation} \label{eq:I-mon-matrix-cond-1} M_{1\,out}=M_{16} \end{equation} holds. To calculate asymptotically $M_{1\,out}$ we use Expansion~(\ref{eq:Psi-out-asympt-1}), Equations (\ref{eq:I-C-Psi-out}), (\ref{eq:C0-alpha-beta-gamma}) and the following formulae for the $\Gamma$-function~\cite{BE}: \begin{gather} \label{eq:Gamma-ratio} \frac{\Gamma(z+\mu)}{\Gamma(z+\nu)}=z^{\mu-\nu}(1+{\cal O}(z^{-1})), \qquad |z|\to\infty,\qquad|\arg\,z|<\pi,\\ \label{eq:Gamma-product} \Gamma(z)\Gamma(1-z)=\frac\pi{\sin(\pi z)}. \end{gather} One notices that $$ \det C_{\alpha\beta\gamma}^0=\frac{\gamma-1}{\beta-\alpha}e^{-\pi i(\alpha+\beta+1-\gamma)}, $$ and finds \begin{equation} \label{eq:I-C0-C} \begin{gathered} C_0^{-1}=(1-\gamma)^{\frac{\alpha-\beta}2\sigma_3}e^{\frac{\pi i}2(\alpha+\beta+1-\gamma)} \left(C+{\cal O}\big(x^{\tilde\delta}\big)\!\right)d_0^{\sigma_3},\\ C=\left(\begin{array}{cc} \frac{\Gamma(\beta-\alpha+1)}{\Gamma(1-\alpha)}e^{\pi i(1-\beta)}& \frac{\Gamma(\beta-\alpha+1)}{\Gamma(\beta)}\frac{\sin\,\pi(\alpha+1-\gamma)}{\sin\,\pi(1-\gamma)} e^{-\pi i(\beta-\gamma)}\\ \frac{\Gamma(\alpha-\beta+1)}{\Gamma(1-\beta)}e^{\pi i(1-\alpha)}& \frac{\Gamma(\alpha-\beta+1)}{\Gamma(\alpha)}\frac{\sin\,\pi(\beta+1-\gamma)}{\sin\,\pi(1-\gamma)} e^{-\pi i(\alpha-\gamma)} \end{array}\right). \end{gathered} \end{equation} Note that according to the last two equations (\ref{eq:I-gamma-cond}) $C$ is the constant matrix. Using Equations~(\ref{eq:Gamma-ratio}) and (\ref{eq:Gamma-product}) we obtain the leading term of asymptotics for $C_{\alpha\beta\gamma}^1$ (\ref{eq:C1-alpha-beta-gamma}): \begin{equation} \label{eq:I-C1-alpha-beta-gamma-asympt} \begin{gathered} C_{\alpha\beta\gamma}^1\underset{\varepsilon\to0}\sim(1-\gamma)^{\frac{\alpha+\beta}2\sigma_3} \left(\begin{array}{cc} 1&0\\ 0&\frac{e^{\pi i(\alpha+\beta+1-\gamma)}}{1-\gamma} \end{array}\right)\\ \times\left(\begin{array}{cc} -\frac{\Gamma(\alpha-\beta+1)}{\Gamma(\alpha)}&\frac{\Gamma(\beta-\alpha+1)}{\Gamma(\beta)}\\ \frac{\sin\,\pi(\beta+1-\gamma)}{\sin\,\pi(\alpha+\beta+1-\gamma)} \frac{\Gamma(\alpha-\beta+1)}{\Gamma(1-\beta)}& -\frac{\sin\,\pi(\alpha+1-\gamma)}{\sin\,\pi(\alpha+\beta+1-\gamma)} \frac{\Gamma(\beta-\alpha+1)}{\Gamma(1-\alpha)} \end{array}\right) (1-\gamma)^{\frac{\beta-\alpha}2\sigma_3}. \end{gathered} \end{equation} Now Equations (\ref{eq:I-C0-C}) and (\ref{eq:I-C1-alpha-beta-gamma-asympt}) yield \begin{equation} \label{eq:M1out} M_{1\,out}=C_1^{-1}e^{\pi i(\gamma-1-\alpha-\beta)}C_1. \end{equation} Hence Equation~(\ref{eq:I-mon-matrix-cond-1}) in the notation~(\ref{eq:I-Mnu6-elements-def}) reads as \begin{align} \label{eq:I-m-11-16} m_{11}^{16}&=\frac{i}{\sin\,\pi(1-\gamma)}(\cos\,\pi(\alpha+\beta+1-\gamma)e^{-\pi i(1-\gamma)}- \cos\,\pi(\alpha-\beta)),\\ m_{22}^{16}&=\frac{i}{\sin\,\pi(1-\gamma)}(\cos\,\pi(\alpha-\beta)- \cos\,\pi(\alpha+\beta+1-\gamma)e^{\pi i(1-\gamma)}),\nonumber\\ \label{eq:I-m-21-16=m12-16} m_{21}^{16}&=\frac{2\pi id_0^2}{\Gamma(1-\alpha)\Gamma(1-\beta)},\quad m_{12}^{16}=-\frac{2\pi i\sin\,\pi(\alpha+1-\gamma)\sin\,\pi(\beta+1-\gamma)} {d_0^2\Gamma(\alpha)\Gamma(\beta)\sin^2\pi(1-\gamma)}. \end{align} Now define $l=\alpha-\beta$ and recall (\ref{eq:I-gamma-cond}) to derive from Equations~(\ref{eq:I-m-11-16}) and (\ref{eq:I-m-21-16=m12-16}) Formulae (\ref{eq:I-l-condition}) and (\ref{eq:alpha-beta-d0}), respectively. It was supposed that $l\neq0,\pm1$ (see Equation~(\ref{eq:alpha,beta,gamma-conditions-Y})). The natural requirement $|{\rm Re}|\,l<1$ means no additional restrictions, since, thanks to the equation $\alpha+\beta=-\Theta_{\infty5}$, the shift $l\to l+2$ leads, simply, to a redefinition of $d_0$; thus we get (\ref{eq:I-l-conditions}). Turning to the matching (\ref{eq:K-matching}), we notice that $(1-\gamma)/x=\lambda_5t_5$; and by recalling asymptotics of $\Psi_5^0$ (\ref{eq:Psi5-asymptotics-infty}) we confirm not only the last equation (\ref{eq:I-def-theta's}), but also obtain that \begin{equation} \label{eq:R16-asympt} R_{16}=G^{-1}\left(\begin{array}{cc} 1&\beta\\ 1&\alpha \end{array}\right) \left(1+{\cal O}\big(x^{\tilde\delta}\big)\!\right) t_5^{-\frac{\Theta_{\infty5}}2\sigma_3}f_0^{-\sigma_3},\qquad\in\mathbb{C}\setminus\{0\}, \end{equation} \begin{equation} \label{eq:K-limit} K=\underset{\varepsilon\to0}\lim\,f_0^{\sigma_3}\hat K(-1)C_0^{-1}/\sqrt{\alpha-\beta}. \end{equation} We can further simplify Equation~(\ref{eq:R16-asympt}) with the help of Equations~(\ref{eq:I-G-Psi-out}) and (\ref{eq:I-G0-G1-alpha-beta-gamma}): \begin{equation} \label{eq:r16-asympt-final} R_{16}=d_0^{-\sigma_3}(1-\gamma)^{\frac{\Theta_{\infty5}}2\sigma_3} \left(1+{\cal O}\big(x^{\tilde\delta}\big)\!\right) t_5^{-\frac{\Theta_{\infty5}}2\sigma_3}f_0^{-\sigma_3}. \end{equation} Equation~(\ref{eq:r16-asympt-final}) is equivalent, up to the leading term, to Equation~(\ref{eq:I-asympt-s16}). Finally, substituting into (\ref{eq:K-limit}) formulae (\ref{eq:K(varkappa)}), (\ref{eq:I-gamma-cond}), and (\ref{eq:I-C0-C}), we arrive at Equation~(\ref{eq:I-K}). \end{derivation} \section{The Second Limit} \label{sec:5} Comparing the formulae for $s_{16}$ (\ref{eq:13a-s16}) and (\ref{eq:13b-s16}) for the first limit passage with the analogous formula for $s_{t6}$ (\ref{eq:II-st6}) for the second limit, one finds the ``principle'' distinction between the limits: while Equations~(\ref{eq:13a-s16}) and (\ref{eq:13b-s16}) are ``integrable'' Equation \ref{eq:II-st6}) is not. Thus it is not clear how to set the constant of integration in the definition of $s_{t6}$. As the result asymptotics of $u_{\nu6}$, $\nu=0,1,t$ is found here up to the factor $c$: $s_{t6}\to cs_{t6}$ $\Leftarrow$ $u_{\nu6}\to c^{-2}u_{\nu6}$. This fact does not influence the functions $y_6$ and $\tau_6$, whose asymptotics are properly defined in the case of the second limit passage (see below Theorem~\ref{th:2}). The problem of how to cope with the ambiguity of $c$, i.e., to set $s_{t6}$ in terms of ${\cal M}_6(\Theta_{06},\Theta_{16},\Theta_{t6},\Theta_{\infty6})$, is left for further investigation. This explains some differences appearing in the formulations of Theorems~\ref{th:1} and ~\ref{th:2}; here we also omit the nonessential $f_0$-like parameter (see (\ref{eq:I-K})). As it is mentioned in Introduction on the level of the Painlev\'e and the corresponding $\tau$-functions both limits are equivalent. We show that at the end of this section in Proposition~\ref{prop:1}. Nevertheless we present also the direct derivation because:\\ 1. This result is easy to extend for the cluster on the arbitrary regular background, i.e., for the Garnier systems (see Corollary~\ref{cor:1}).\\ 2. It is interesting to see how the ``nonintegrability'' of Equation~(\ref{eq:II-st6}) manifests itself in our asymptotic calculations which do not contain any integration in the usual sense. Actually I started my studies with the second cluster as it naturally appeared when one inserts the expansion $(\lambda_6-t_6)^{-1}=\lambda_6^{-1}+t_6\lambda_6^{-2}+\ldots$ into Equation~(\ref{eq:linear-4-regular}). After I realized that the above ``$c$-problem'' is not the intrinsic ``cluster problem'', I found the first limit, where such a ``$c$-problem'' does not appear. Note that applying Proposition~\ref{prop:1} to Theorem~\ref{th:2} we cannot obtain the complete result for the first limit, as it stated in Theorem~\ref{th:1}. Let us begin with the regularization of the second limit. Suppose that\\ ${\cal M}_6(\Theta_{06},\Theta_{16},\Theta_{t6},\Theta_{\infty6})$ is given and define the parameters $\varepsilon_n$, $\Theta_{\infty5}$, $\Theta_6$ as follows: \begin{gather} \label{eq:II-theta-t6} \Theta_{t6}=\Theta_6-2n\equiv-\frac1{\varepsilon_n},\quad n\in\mathbb{Z}_+,\quad \Theta_6\in\mathbb{C}\setminus\mathbb{Z},\\ \label{eq:II-def-theta-infty5} \Theta_{t6}+\Theta_{06}=\Theta_{\infty5},\qquad\Theta_{\infty5}-\Theta_6\in\mathbb{C}\setminus\mathbb{Z}. \end{gather} Here again our parameter $\varepsilon=\varepsilon_n\to+0$ is discrete. The main object of our investigation is the GSD $A_{\nu6}^k(t_6)$ $\nu=0,1,t$ and $k\in\mathbb{Z}_+$, which is defined via the recurrence procedure: \begin{equation} \label{eq:II-A-nu6-k-rec} A_{\nu6}^0(t_6)=A_{\nu6}(t_6),\qquad A_{\nu6}^{k+1}(t_6)={\tilde A}_{\nu6}^k(t_6), \end{equation} where $A_{\nu6}(t_6)$ is the solution of System~(\ref{eq:schlesinger-p6}) corresponding to ${\cal M}_6(\Theta_{06},\Theta_{16},\Theta_{t6},\Theta_{\infty6})$, and ${\tilde A}_{\nu6}^k(t_6)$ is obtained via (\ref{eq:Atilde-nu})--(\ref{eq:Atilde-nu-result}) (with $J_{0t}^{+-}$) by inserting $A_{\nu6}^k(t_6)$ into these formulae instead of $A_{\nu6}$. We interpret the formal limit {\bf II} as the asymptotics as $n\to+\infty$ of the sequence $A_{\nu6}^{2n}(\varepsilon_nt_5)$, or, equivalently, as the asymptotics of the corresponding sequences of their matrix elements. Here $t_5$ is a real positive number, which is further assumed as a given parameter. Asymptotics of the odd sequence, $A_{\nu6}^{2n+1}(\varepsilon_nt_5)$, is given by exactly the same formulae as for the even one, $A_{\nu6}^{2n}(\varepsilon_nt_5)$, (see Equations~(\ref{eq:13b-Theta})-(\ref{eq:13b-s16}) with $\varepsilon\to\varepsilon_n$) but in the corresponding formulae for the monodromy data one has to change $\Theta_6\to\Theta_6-1$. In the derivation of the results stated in Theorem~\ref{th:2} below, we omit the super/subscript $n$, if it does not cause a confusion. To formulate our result we need to introduce some preliminary notation. It follows from Equations (\ref{eq:monodromy-local}), (\ref{eq:II-theta-t6}), and (\ref{eq:II-def-theta-infty5}) that $e^{\pm\pi i\Theta_6}$ are the eigenvalues of $M_{t6}$ and $e^{\pm\pi i(\Theta_{\infty5}-\Theta_6)}$ are the eigenvalues of $M_{06}$. Define \begin{gather} \label{eq:II-T} {\bf T}={\rm tr}(M_{t6}-e^{+\pi i\Theta_6})(M_{06}-e^{-\pi i(\Theta_{\infty5}-\Theta_6)}),\\ \label{eq:II-l} l=\frac1{\pi i}\ln\left(\cos(\pi\Theta_{\infty5})-\frac12{\bf T} \pm\sqrt{\Big(\cos(\pi\Theta_{\infty5})-\frac12{\bf T}\Big)^2-1}\right),\qquad-1<{\rm Re}\,l<1,\\ \label{eq:II-alpha-beta} \alpha=-\frac12(\Theta_{\infty5}-l),\qquad\beta=-\frac12(\Theta_{\infty5}+l). \end{gather} The sign before the square root in Equation~(\ref{eq:II-l}) can be chosen arbitrary: the change of the sign means simply the change $\alpha\leftrightarrow\beta$ in our construction, which is invariant under this transformation. \begin{theorem} \label{th:2} Let $\mu_6\in{\cal M}_6(\Theta_{\infty5}-\Theta_6,\Theta_{16},\Theta_6,\Theta_{\infty6})$. Suppose the following conditions are valid:\\ {\rm1. (\ref{eq:II-theta-t6}), (\ref{eq:II-def-theta-infty5})}, and $\Theta_{\nu6}\in\mathbb{C\setminus Z}$, $\nu=0,1,\infty$;\\ {\rm2}. $\alpha,\beta,l\in\mathbb{C\setminus Z}$;\\ {\rm3}. $(M_{t6}-e^{-\pi i\Theta_6})(M_{06}-e^{-\pi i(\Theta_{\infty5}-\Theta_6)})\neq0$.\\ {\rm4}. The inverse monodromy problem for Equation~{\rm(\ref{eq:linear-4-regular})} is solvable for all pairs $(\mu_6,t_6)$ such that $t_6=\varepsilon_nt_5>0$, where $\varepsilon_n$ is given by Equation~{\rm(\ref{eq:II-theta-t6})}.\\ {\rm5}. It is possible to define the sequence $A_{\nu6}^{2n}(\varepsilon_nt_5)$: it is true in particular, if $\mu_6$ corresponds to the non-classical solution of Equation~{\rm(\ref{eq:p6})} \footnote[8]{The last condition is not necessary: such sequence is possible to organize for the classical solutions too, just for some of them $n\in\mathbb{Z}_-$ instead of $\mathbb{Z}_+$; we do not discuss here the corresponding modifications.}. Then define $\mu_5\in{\cal M}_5(\Theta_{05},\Theta_{15},\Theta_{\infty5})$: \begin{gather} \label{eq:II-P5-theta} \Theta_{05}=\Theta_{\infty6},\qquad\Theta_{15}=\Theta_{16},\\ \label{eq:II-P5-Stokes} S_0=\left(\begin{array}{cc} 1&0\\ -\frac{2\pi i}{\Gamma(1-\alpha)\Gamma(1-\beta)}&1 \end{array}\right),\qquad S_1=\left(\begin{array}{cc} 1&-\frac{2\pi i}{\Gamma(\alpha)\Gamma(\beta)}e^{\pi i\Theta_{\infty5}}\\ 0&1 \end{array}\right),\\ \label{eq:P5-M05-M15} KM_{\infty6}K^{-1}=M_{05},\qquad KM_{16}K^{-1}=M_{15}, \end{gather} where $K$ is the unique (up to the sign) solution of the system: \begin{equation} \label{eq:II-P5-K-system} KM_{t6}K^{-1}=S_0e^{\pi i\Theta_6\sigma_3},\qquad KM_{06}K^{-1}=e^{-\pi i\Theta_6\sigma_3}S_1e^{\pi i\Theta_{\infty5}\sigma_3}. \end{equation} See Theorem~{\rm\ref{th:A}} in the Appendix and Equations~{\rm((\ref{eq:A.19})--(\ref{eq:A.21})}. If for a given pair $(\mu_5,t_5)$ the inverse monodromy problem is solvable, then an asymptotic expansion of the GSD $A_{\nu6}^{2n}(\varepsilon_nt_5)$ as $\varepsilon_n\to+0$ is given by Equations {\rm(\ref{eq:lim2-t-lambda})--(\ref{eq:lim2-Rt6-derivative})} with $\varepsilon=\varepsilon_n$. \end{theorem} \begin{derivation} Suppose that the matrices $A_{\nu6}$ in Equation~(\ref{eq:linear-4-regular}) satisfy Equations~(\ref{eq:lim2-t-lambda})--(\ref{eq:lim2-Rt6-derivative}) corresponding to the formal limit {\bf II}. Then one proves that in the domain $\lambda_6\in\Omega_{out}:=\{\lambda_6\in\mathbb{C}, \varepsilon/|\lambda_6|^3 \leq{\cal O}(\varepsilon^{3\delta_0-2}),\; |\arg\,\lambda_6+\frac{\pi}2|<\delta_1,\;\delta_0>\frac23,\;0\leq\delta_1<\pi\}$ the $\Psi_6$ function has the following asymptotics \begin{equation} \label{eq:psi6} \Psi_6(\lambda_6,t_6)K^{-1}=\varepsilon^{\sigma_3}R\Psi_5^0(\lambda_5,t_5)(I+o(1)), \end{equation} where $K,R\in{\rm SL}(2,\mathbb{C})$, $R=\varepsilon^{-\sigma_3}R_{t6}+o(1)$, $\varepsilon^{-\sigma_3}R_{t6}={\cal O}(1)$. The matrix $K$ in Equation~(\ref{eq:psi6}) is independent of $t_\nu$, $\lambda_\nu$, $\nu=5,6$ and $\varepsilon$. Equations~(\ref{eq:P5-M05-M15}) are an immediate consequence of (\ref{eq:psi6}) and (\ref{eq:lim2-t-lambda}). To this end our problem is to find $K$. To solve it we consider the function $\Phi$ which solves the hypergeometric equation: \begin{equation} \label{eq:II-Phi-hyper} \frac{d\Phi}{d\lambda_6}=\left(\frac{A_{06}}{\lambda_6}+\frac{A_{16}}{\lambda_6-t_6}\right)\Phi, \end{equation} and has the same monodromy matrices at regular singularities $\lambda_6=0$ and $t$ as the function $\Psi_6(\lambda_6,t_6)K^{-1}$. We call the latter property of $\Phi$ as the condition $M$. The fundamental solution of Equation ~(\ref{eq:II-Phi-hyper}) can be presented as \begin{equation} \label{eq:II-Phi-Y} \Phi=PY(x),\qquad x=\lambda_6/t_6, \end{equation} where $Y(x)$ is given by (\ref{eq:Y(x)-def}) with the parameters $\alpha$, $\beta$, $\gamma$ satisfying the equations \begin{equation} \label{eq:II-alpha-beta-gamma} 1-\gamma=\Theta_{06},\quad\gamma-1-\alpha-\beta=\Theta_{t6}, 0<|\beta-\alpha||<1, \end{equation} and $P^{-1}A_{\nu6}P\approx A_{\frac\nu{t}Y}$, where $A_{\frac\nu{t}Y}$ is defined by (\ref{eq:A0Y}) and (\ref{eq:A1Y}). The last condition in (\ref{eq:II-alpha-beta-gamma}) is assumed by taking into account the theorem formulated in \cite{J}[\S2, pp.1145-1146]. Now we have to determine $l=\alpha-\beta$. To do this we use condition $M$ in the following way. First, for $m=0,1$ we calculate the matrix $\hat K(2m-1)$ by help of the equation \begin{equation} \label{eq:hat-K-calc} \Phi\hat K^{-1}(2m-1)=\varepsilon^{\sigma_3}R\Psi_5^m, \qquad\lambda_6\sim{\cal O}(\varepsilon^{1-\delta_0}). \end{equation} Thus, using definitions of Section~\ref{sec:2} we find the Stokes multiplier \begin{equation} \label{eq:II-S0} S_0=\hat K(-1)\hat K^{-1}(1), \end{equation} and the monodromy matrix at the infinity point \begin{equation} \label{eq;II-M-infty} M_{\infty5}=\hat K(-1)M_{1Y}M_{0Y}\hat K^{-1}(-1). \end{equation} Now the direct calculation shows \begin{equation} \label{eq:hat-K-M1Y-S0} \hat K(-1)M_{1Y}\hat K^{-1}(-1)=S_0e^{\pi i\Theta_{t6}\sigma_3}. \end{equation} Comparing Equation~(\ref{eq:M-infty-stokes-k}) for $k=0$ with Equations~(\ref{eq;II-M-infty}) and (\ref{eq:hat-K-M1Y-S0}) we obtain the other Stokes multiplier: \begin{equation} \label{eq:II-S1-calc} \hat K(-1)M_{0Y}\hat K^{-1}(-1)=e^{\pi i\Theta_{06}\sigma_3}e^{-\pi i\Theta_{\infty5}\sigma_3} S_1e^{\pi i\Theta_{\infty5}\sigma_3}. \end{equation} Equations (\ref{eq:II-S0}) and (\ref{eq:II-S1-calc}) are equivalent to the ones presented in (\ref{eq:II-P5-Stokes}), but the values of the parameters $\alpha$ and $\beta$ still remain undetermined. Comparing Equations (\ref{eq:hat-K-calc}), (\ref{eq:II-Phi-Y}), and (\ref{eq:psi6}), one finds \begin{equation} \label{eq:II-M-nu6-M-nu-Y} KM_{\nu6}K^{-1}=\hat K(-1)M_{\frac\nu{t}Y}\hat K^{-1}(-1),\qquad\nu=0,t. \end{equation} Now Equations (\ref{eq:II-M-nu6-M-nu-Y}), (\ref{eq:II-S1-calc}), and (\ref{eq:hat-K-M1Y-S0}) yield (\ref{eq:II-P5-K-system}), which are the {\it exact} formulae, since both sides of these equations are independent of $\varepsilon$. Finally we use Theorem~\ref{th:A} of the Appendix to find formulae (\ref{eq:II-T}) -- (\ref{eq:II-alpha-beta}) and the matrix $K$. The rigorous aspect of the above calculation is based on the justification scheme suggested in \cite{K4}. The scheme can be explained as follows. Suppose that the inverse monodromy problem for Equation ~(\ref{eq:linear-5}) is solvable for a given value of the parameter $t_5$ in some neighborhood ${\cal O}(\mu_5)$ of the given point $\mu_5\in{\cal M}_5(\Theta_{05},\Theta_{15},\Theta_{\infty5})$. Thus for all $\tilde\mu_5\in{\cal O}(\mu_5)$ we have functions $\tilde u_5$, $\tilde z_5$, and $\tilde y_5$ whose monodromy data coincide with $\tilde\mu_5$. Then using Equations (\ref{eq:II-zt6-z06}) --(\ref{eq:II-ut6zt6st6}), and (\ref{eq:lim2-t-lambda}) with $\tilde u_5$, $\tilde z_5$, and $\tilde y_5$ instead of $u_5$, $z_5$, and $y_5$ and the discrete $\varepsilon=\varepsilon_n$, substitute them into the matrix elements of the residue matrices of Equation~(\ref{eq:linear-4-regular}). Then the derivation presented above can be interpreted as as the proof that the monodromy data $\hat\mu_5$, of thus obtained Equation~(\ref{eq:linear-4-regular}), differ from the data $\tilde\mu_5$, obtained by the inversion of Equations~(\ref{eq:P5-M05-M15}) and (\ref{eq:II-P5-K-system}), on quantities estimated as $\parallel\tilde\mu-\hat\mu\parallel<o(1)$. To be able to apply the scheme of justification suggested in \cite{K4} the last estimate must possess an important property of ``local uniformness'': this means that there exists a neighborhood of $\tilde\mu_5$ such that the above estimate is uniform for all points $\hat\mu_5$ in this neighborhood. In the terminology of the work \cite{K4}: the inverse monodromy problem for Equation~(\ref{eq:linear-4-regular}) for the monodromy data $\tilde\mu_5$ is asymptotically solvable. The main result of \cite{K4} says that if the monodromy problem is asymptotically solvable, then it is exactly solvable. Since the solution of the inverse monodromy problem is unique, this solution coincides with the GSD introduced in the beginning of this section. Now I am going to give some details to the calculation outlined above. Let us begin with the matching (\ref{eq:hat-K-calc}) as it is the most crucial point. Consider Equation~(\ref{eq:II-Phi-Y}). The matrix $P$ which maps the function $Y(x)$ into a fundamental solution of the hypergeometric equation (\ref{eq:II-Phi-hyper}) is defined as follows \begin{equation} \label{eq:II-P-details} P^{-1}A_{\nu6}P=A_{\frac\nu{t}Y}+{\cal O}\big(\varepsilon^{\delta_0-1}\big),\qquad\nu=0,1, \qquad P\in{\rm SL}(2,\mathbb{C}). \end{equation} Let's explain the estimate ${\cal O}\big(\varepsilon^{\delta_0-1}\big)$ in Equation (\ref{eq:II-P-details}): it is an important and rather subtle moment. First we show that the estimate cannot be less than ${\cal O}(1)$. Actually, summing up equations in (\ref{eq:II-P-details}) (with the error estimate ${\cal O}(1)$) and using the second equality in (\ref{eq:hypergeo-Y(x)}) and the first equation in (\ref{eq:niormalization-for-P6}), one arrives at \begin{equation} \label{eq:II-P-details-2} P^{-1}\left(\frac{\Theta_{\infty6}}2\sigma_3+A_{16}\right)P=-\frac{\beta-\alpha}2\sigma_3+{\cal O}(1). \end{equation} Suppose that in Equation~(\ref{eq:II-P-details-2}) one have a better estimate, $o(1)$, instead of ${\cal O}(1)$. Then it implies \begin{equation} \label{eq:II-det-details} \det\left(\frac{\Theta_{\infty6}}2\sigma_3+A_{16}\pm\frac{\beta-\alpha}2\right)=o(1). \end{equation} The last equation leads to the nontrivial dependence of $\alpha$ and $\beta$ from $t_5$ (see Equations~(\ref{eq:II-z16}) and the first equation (\ref{eq:A-and-R}) with $\nu=1$. This dependence contradicts the isomonodromy condition, since the monodromy data in particular Stokes multipliers $S_k$ must be independent of $t_5$ (see (\ref{eq:II-P5-Stokes})). Thus the minimal possible error estimate in Equation~(\ref{eq:II-P-details}) is ${\cal O}(1)$, i.e., $\delta_0\leq1$. In the last case Equation~(\ref{eq:II-P-details-2}) becomes uninformative. One of the most important moments of our derivation is the matching procedure of $\Phi$ with $\varepsilon^{\sigma_3}R_{t6}\Psi_5^0$. This matching occurs at the domain $\lambda_6\sim\varepsilon^{1-\delta_0}$, $\frac23<\delta_0<1$, where the lower limit for $\delta_0$ is defined via the matching of $\Psi_6$ and $R_{t6}\Psi_5^0$ (see Equation~(\ref{eq:psi6})): The term $t_6^2A_{t6}/\lambda_6^3\sim{\cal O}\big(\varepsilon^{3\delta_0-2}\big)$, which was neglected in Equation~(\ref{eq:linear-4-regular}) to obtain the equation for $R_{t6}\Psi_5^0$, should decrease. The upper limit for $\delta_0:\;\delta_0<1$ is originated from the matching of $\Psi_6$ and $\Phi$: the corresponding equations differ by the term ${\cal O}(t_6A_{t6}/\lambda_6^2)\sim{\cal O}(\varepsilon^{2(\delta_0-1)})$. Thus to guarantee the matching we must define $P$ in Equation~(\ref{eq:II-P-details}) with the error estimate not worse than $\lambda_6{\cal O}(\varepsilon^{2(\delta_0-1)})={\cal O}(\varepsilon^{\delta_0-1})$. To summarize: the primary qualities of $\Phi$ (according to \cite{K4}) are its monodromy data and the matching with $\varepsilon^{\sigma_3}R_{t6}\Psi_5^0$. It is because of these properties the function $\Phi$, in fact, is not the exact solution of Equation~(\ref{eq:II-Phi-hyper}), but solves it up to the leading order with the rather large error ${\cal O}(\varepsilon^{\delta_0-1})$. It seems that the function $\Phi$ with the required properties can be defined to satisfy Equation~(\ref{eq:II-Phi-hyper}) more precisely, with the error ${\cal O}(1)$, but for this goal we need to substantially increase complexity of our calculations. In any case we cannot define the function $\Phi$ to satisfy Equation~(\ref{eq:II-Phi-hyper}) with the error less than ${\cal O}(1)$, because it contradicts the matching procedure. Actually, the asymptotic expansion for $A_{16}$, which we substitute into Equation~(\ref{eq:II-det-details}) to get the contradiction, is obtained from the differential equation for $R_{t6}\Psi_5^0$ and, hence, it is inexplicit consequence of the matching. Taking into account the discussion above, we rewrite System~(\ref{eq:II-P-details}) with the error ${\cal O}(\varepsilon^{\delta_0-1})$ changed by ${\cal O}(1)$ as follows: \begin{equation} \label{eq:II-P-calc-elements} \begin{aligned} (-1)^{\frac{\nu}t}\frac{\beta(1-\gamma)}{\beta-\alpha}&= \left(d-\frac{b\varepsilon^2}{u_{\nu6}}\right) \left(b\Theta_{\nu6}-\frac{u_{\nu6}z_{\nu6}}{\varepsilon^2} \left(d-\frac{b\varepsilon^2}{u_{\nu6}}\right)\right)+{\cal O}(1),\\ (-1)^{1+\frac{\nu}t}\frac{\alpha(1-\gamma)}{\beta-\alpha}&= \left(c-\frac{a\varepsilon^2}{u_{\nu6}}\right) \left(\frac{u_{\nu6}z_{\nu6}}{\varepsilon^2} \left(c-\frac{a\varepsilon^2}{u_{\nu6}}\right) -a\Theta_{\nu6}\right)+{\cal O}(1),\\ (-1)^{\frac{\nu}t}\frac{(\alpha+\beta)(1-\gamma)}{2(\beta-\alpha)}&= \left(d-\frac{b\varepsilon^2}{u_{\nu6}}\right) \left(c-\frac{a\varepsilon^2}{u_{\nu6}}\right) \frac{u_{\nu6}z_{\nu6}}{\varepsilon^2}\\ &-\frac{\Theta_{\nu6}}2\left(a\Big(d-\frac{b\varepsilon^2}{u_{\nu6}}\Big) +b\Big(c-\frac{a\varepsilon^2}{u_{\nu6}}\Big)\right)+{\cal O}(1), \end{aligned} \end{equation} where $\nu=0,t$ and $a$, $b$, $c$, and $d$ are the matrix elements of $P$: \begin{gather} \label{eq:P-abcd} P=\varepsilon^{\sigma_3}\left(\begin{array}{cc} a&b\\ c&d \end{array}\right),\qquad\text{with}\;\;a,b,c,d={\cal O}(1),\\ \label{eq:ad-bc=1} ad-bc=1. \end{gather} Using now Formulae~(\ref{eq:II-zt6-z06}), (\ref{eq:II-ut6-u06}), (\ref{eq:II-alpha-beta-gamma}), and (\ref{eq:P-abcd}), one rewrites Equations~(\ref{eq:II-P-calc-elements}) as follows: \begin{equation} \label{eq:u-nu6-z-nu6} u_{\nu6}=\varepsilon^2\frac{a+b}{c+d}+{\cal O}(\varepsilon^3),\quad \frac{z_{\nu6}}{\Theta_{\nu6}}=\frac{(a+b)(c+d)}2 \left(\frac{b-a}{b+a}-\frac{\beta-\alpha}{\beta+\alpha}\right)\!+{\cal O}(\varepsilon),\;\;\nu=0,t. \end{equation} To determine the matrix $P$ uniquely we must add to (\ref{eq:ad-bc=1}) and (\ref{eq:u-nu6-z-nu6}) one more equation. We get it via the matching (\ref{eq:hat-K-calc}): the matching domain now is \begin{equation} \label{eq:II-matching-domain} x=\frac{\lambda_6}{t_6}=(\varepsilon\lambda_5t_5)^{-1}={\cal O}\big(\varepsilon^{-\delta_0}\big),\qquad \frac23<\delta_0<1, \end{equation} so that Asymptotics~(\ref{eq:Y(x)-spec-asympt}), (\ref{eq:x-spec-asympt-cond}) imply the following equations: \begin{gather} \label{eq:varkappa} \varkappa=2m-1,\qquad\text{for}\quad m=0,1\quad\alpha+\beta=-\Theta_{\infty5},\\ \label{eq:R-t6-abcd} \frac{\varepsilon^{\sigma_3}}{\sqrt{\alpha-\beta}}\left(\begin{array}{cc} a&b\\ c&d \end{array}\right) \left(\begin{array}{cc} 1&\beta\\ 1&\alpha \end{array}\right) t^{-\frac{\Theta_{\infty5}}2\sigma_3}=R_{t6}+\varepsilon^{\sigma_3} {\cal O}\big(\varepsilon^{\tilde\delta}\big). \end{gather} Recall the definition of ${\cal O}\big(\varepsilon^{\tilde\delta}\big)$ below Equation~(\ref{eq:x-spec-asympt-cond}). The first equation in (\ref{eq:varkappa}) and (\ref{eq:hat-K-calc}) show that (\ref{eq:II-S0}) and (\ref{eq;II-M-infty}) hold up to the error bound ${\cal O}\big(\varepsilon^{\tilde\delta}\big)$. From (\ref{eq:R-t6-abcd}) one finds that Equations~(\ref{eq:ad-bc=1}) and (\ref{eq:u-nu6-z-nu6}) are valid up to the estimate ${\cal O}\big(\varepsilon^{\tilde\delta}\big)$. Equation (\ref{eq:R-t6-abcd}) together with the first equation in (\ref{eq:II-ut6-u06}) and the equality $\varepsilon_n\Theta_{t6}=-1$ we arrive at: \begin{gather*} a=-\frac1{s_{t6}\sqrt{\alpha-\beta}}\Big(\alpha t_5^{\frac{\Theta_{\infty5}}2} +u_5t_5^{-\frac{\Theta_{\infty5}}2}\Big),\qquad b=\frac1{s_{t6}\sqrt{\alpha-\beta}}\Big(\beta t_5^{\frac{\Theta_{\infty5}}2} +u_5t_5^{-\frac{\Theta_{\infty5}}2}\Big),\\ c=\frac{s_{t6}}{\sqrt{\alpha-\beta}}\left(\frac\alpha{u_5} \frac{z_5}{\Theta_{05}}t_5^{\frac{\Theta_{\infty5}}2}+\Big(\frac{z_5}{\Theta_{05}}+1\Big) t_5^{-\frac{\Theta_{\infty5}}2}\right),\\ d=-\frac{s_{t6}}{\sqrt{\alpha-\beta}}\left(\frac\beta{u_5} \frac{z_5}{\Theta_{05}}t_5^{\frac{\Theta_{\infty5}}2}+\Big(\frac{z_5}{\Theta_{05}}+1\Big) t_5^{-\frac{\Theta_{\infty5}}2}\right). \end{gather*} Now one uses (\ref{eq:II-S0}) and (\ref{eq:varkappa}) to find, again up to ${\cal O}\big(\varepsilon^{\tilde\delta}\big)$, the first equation in (\ref{eq:II-P5-Stokes}). To finish the proof we have to confirm the second equation in (\ref{eq:II-P5-Stokes}). To do it, we have to prove (\ref{eq:hat-K-M1Y-S0}) and to find $S_1$ from (\ref{eq:II-S1-calc}). Details for these calculations are as follows: apply the formulae for the $\Gamma$-function (\ref{eq:Gamma-ratio}) and (\ref{eq:Gamma-product}) to Equations (\ref{eq:C0-alpha-beta-gamma}) and (\ref{eq:C1-alpha-beta-gamma}) to find: \begin{gather*} C_{\alpha\beta\gamma}^0{\hat K}^{-1}(-1)=D_0\left(1+{\cal O}\big(\varepsilon^{\tilde\delta}\big)\!\right)\! \left(\begin{array}{cc} 1&-\frac{\pi\exp(\pi(\alpha+\beta+1-\gamma))}{(\alpha-\beta)\sin(\pi\gamma)\Gamma(\alpha)\Gamma(\beta)}\\ 0&1 \end{array}\right)\! \frac1{\sqrt{\alpha-\beta}^{\sigma_3}}\left(1+{\cal O}\big(\varepsilon^{\tilde\delta}\big)\!\right),\\ C_{\alpha\beta\gamma}^1{\hat K}^{-1}(-1)=D_1\left(1+{\cal O}\big(\varepsilon^{\tilde\delta}\big)\!\right)\! \left(\begin{array}{cc} 1&0\\ \frac{\pi\exp(\pi(\gamma-\alpha-\beta))(\beta-\alpha)}{\sin(\pi(\alpha+\beta-\gamma))\Gamma(1-\alpha) \Gamma(1-\beta)}&1 \end{array}\right)\! \left(1+{\cal O}\big(\varepsilon^{\tilde\delta}\big)\!\right), \end{gather*} where $D_k$ are diagonal matrices. Finally, one uses the definition of $M_{\nu Y}$: $$ M_{\nu Y}=(C_{\alpha\beta\gamma}^\nu)^{-1}e^{\pi i\Theta_{t\cdot\nu6}\sigma_3}C_{\alpha\beta\gamma}^\nu, \qquad\nu=0,1 $$ to prove (\ref{eq:hat-K-M1Y-S0}) and (\ref{eq:II-S1-calc}). \end{derivation} The result of Theorem~\ref{th:2} can be immediately generalized as follows. For $k=1,\ldots, n$ define $a_k\in\mathbb{C}\setminus\{0\}$, $a_k\neq a_l$, for $k\neq l$, $\frac{\partial a_k}{\partial t_6}= \frac{\partial a_k}{\partial\lambda_6}=0$. Consider the functions $\Psi_\nu=\Psi_\nu(\lambda_\nu,t_\nu;a_1,\ldots,a_n)\in{\rm SL}(2,\mathbb{C})$ $\nu=5,6$, which generalize the functions $\Psi_\nu$ defined in Section~\ref{sec:2}, as the normalized (\ref{eq:psi6-infty-normalization}) and (\ref{eq:Psi5-asymptotics-infty}) fundamental solutions of the systems: \begin{align*} \frac{d}{d\lambda_6}\Psi_6&=\left(\frac{A_{06}}{\lambda_6}+\frac{A_{t6}}{\lambda_6-t_6}+ \sum\limits_{k=1}^{m}\frac{A_{k6}}{\lambda_6-a_k}\right)\Psi_6,\\ \frac{d}{dt_6}\Psi_6&=-\frac{A_{t6}}{\lambda_6-t_6}\Psi_6, \end{align*} \begin{align*} \frac{d}{d\lambda_5}\Psi_5&=\left(\frac{t_5}2\sigma_3+\frac{A_{05}}{\lambda_5}+ \sum\limits_{k=1}^{m}\frac{A_{k5}}{\lambda_5-1/a_k}\right)\Psi_5,\\ \frac{d}{dt_5}\Psi_5&=\left(\frac{\lambda_5}2\sigma_3+\frac1{t_5}\left(\frac{\Theta_{\infty5}}2\sigma_3+ \sum\limits_{k=0}^{m}A_{k5}\right)\right)\Psi_5, \end{align*} $$ A_{t6}+\sum\limits_{k=0}^{m}A_{k6}\equiv-\frac{\Theta_{\infty6}}2\sigma_3,\qquad {\rm diag}\sum\limits_{k=0}^{m}A_{k5}\equiv-\frac{\Theta_{\infty5}}2\sigma_3,\qquad A_{k\nu}\in sl_2(\mathbb{C}). $$ The local objects like $\Theta_{k\nu}$, $R_{k\nu}$, and $M_{k\nu}$, $k=1,\ldots,m$, are defined in the same manner as that in Section~\ref{sec:2} for $k=1$ and $a_1=1$. Systems of isomonodromy deformations are the compatibility conditions for the systems of the linear ODEs written above, and the GSDs $A_{k6}^{2n}(\varepsilon_nt_5)$ can be defined as in the beginning of this section. \begin{corollary} \label{cor:1} Let conditions of Theorem~{\rm\ref{th:2}} be valid for $$ {\cal M}_6={\cal M}_6(\Theta_{06},\Theta_{t6},\Theta_{16},\ldots,\Theta_{m6},\Theta_{\infty6}), $$ then asymptotics ($n\to+\infty$) of the GSDs $A_{k6}^{2n}(\varepsilon_nt_5)$, $(k=1,\ldots,m)$, are given by Equations~{\rm(\ref{eq:lim2-t-lambda}) -- (\ref{eq:lim2-A05-A15})} which are supplemented by \begin{equation} \label{eq:II-A-k6-Ak5} R_{t6}^{-1}A_{k6}R_{t6}=A_{k5}+{\cal O}(\varepsilon), \end{equation} Equation~{\rm(\ref{eq:lim2-Rt6-derivative})} should be generalized as follows \begin{equation} \label{eq:II-R-t6-derivative-generalized} -R_{t6}^{-1}\frac{d}{dt_5}R_{t6}=\frac1{t_5}\left(\frac{\Theta_{\infty5}}2\sigma_3+ \sum\limits_{k=0}^mA_{k5}\right)+{\cal O}(\varepsilon), \end{equation} and with the change $\varepsilon\to\varepsilon_n$ in all formulae. The monodromy data for the limiting $\Psi_5$ function is given by exactly the same formulae as that in Theorem~{\rm\ref{th:2}} supplemented with the equations: \begin{equation} \label{eq:II-theta-k6-M-k6} \Theta_{k6}=\Theta_{k5},\qquad KM_{k6}K^{-1}=M_{k5},\qquad k=1,\ldots,m. \end{equation} \end{corollary} \begin{proposition} \label{prop:1} For the sixth Painlev\'e function (and the corresponding $\tau$-function) both limits considered in Section~{\rm\ref{sec:3}} are equivalent under the definition given in the Introduction. \end{proposition} \begin{proof} Let us supply with the superscripts $I$ and $II$ the objects corresponding to the first/second limits. Then for $\Psi_6$ function one finds: \begin{equation} \label{eq:Psi-I-II-equivalence} \Psi_6^{II}(\lambda_6^{II},t_6)=t_6^{-\frac{\Theta_{05}}2\sigma_3}(R_{06}^I)^{-1} \Psi_6^I(\lambda_6^I,t_6)(C_{06}^I)^{-1},\qquad\lambda_6^{II}=t_6/\lambda_6^I. \end{equation} This transformation is controlling the limiting procedure. Thus it is not necessary to consider especially what is happening on the lower ${\cal M}$-plane of the diagram on Figure~3. Formula~(\ref{eq:Psi-I-II-equivalence}) generates the following transformation for the matrices $A_{\nu6}$: \begin{equation} \label{eq:A-I-II-equivalence} A_{\nu6}^{II}=t_6^{-\frac{\Theta_{06}^I}2\sigma_3}(R_{06}^I)^{-1}A_{\frac{t}\nu6}^IR_{06}^I t_6^{\frac{\Theta_{06}^I}2\sigma_3},\qquad\nu=1,t. \end{equation} For the formal monodromy one finds: \begin{gather*} \Theta_{05}^{II}=\Theta_{\infty6}^{II}=\Theta_{06}^I=\Theta_{05}^I,\\ \Theta_{15}^{II}=\Theta_{16}^{II}=\Theta_{t6}^I=-\Theta_{15}^I,\\ \Theta_{t6}^{II}=\Theta_{16}^I=-\frac1{\varepsilon},\\ \Theta_{06}^{II}=\Theta_{\infty6}^I\qquad\Theta_{\infty6}^I+\Theta_{16}^I= \Theta_{06}^{II}+\Theta_{t6}^{II}=\Theta_{\infty5}. \end{gather*} Substituting~(\ref{eq:lim1-Psi}) and (\ref{eq:lim2-Psi}) into Equation~(\ref{eq:Psi-I-II-equivalence}) one obtains $$ R_{t6}^{II}\Psi_5^I(\lambda_5,t_5)K^{II}=t_6^{-\frac{\Theta_{06}^I}2\sigma_3}(R_{06}^I)^{-1} R_{16}^I\Psi_5^I(\lambda_5,t_5)K^I(C_{06}^I)^{-1}. $$ Where $K^p$ is the matrix introduced in Theorem $p$ ($p=I,II$). Thus we have \begin{gather} \label{eq:R-relation} R_{t6}^{II}f_0^{-\sigma_3}=t^{-\frac{\Theta_{06}^I}2\sigma_3}(R_{06}^I)^{-1}R_{16}^I,\\ \label{eq:K-relation} K^{II}=f_0^{-\sigma_3}K^I(C_{06}^I)^{-1}. \end{gather} Using Relations (\ref{eq:R-relation}) and (\ref{eq:K-relation}) one proves that Equations~(\ref{eq:lim1-A}) and (\ref{eq:lim2-A05-A15}) are equivalent. For the matrix elements one has to take into account that $\Theta_{15}^{II}=-\Theta_{15}^I$ and, hence, Equations~(\ref{eq:A05-parametrization}) and (\ref{eq:A15-parametrization}) yield $$ u_5^{II}=u_5^I,\;\; z_5^{II}=z_5^I,\;\; y_5^{II}\!\left(z_5+\frac{\Theta_{05}-\Theta{15}^{II}+\Theta_{\infty5}}2\right)\!= y_5^I\!\left(z_5+\frac{\Theta_{05}-\Theta{15}^I+\Theta_{\infty5}}2\right)\!. $$ The so-called $c$-problem discussed in the beginning of this Section reveals itself in Equations~(\ref{eq:R-relation}) and (\ref{eq:K-relation}) as the ambiguity in the definitions of $R_{06}$, $C_{06}$: $$ R_{06}\to R_{06}c^{\sigma_3},\qquad C_{06}\to c^{-\sigma_3}C_{06}. $$ \end{proof}

https://arxiv.org/abs/1910.12348

Motivic Donaldson-Thomas Invariants of Parabolic Higgs Bundles and Parabolic Connections on a Curve

Let $X$ be a smooth projective curve over a field of characteristic zero and let $D$ be a non-empty set of rational points of $X$. We calculate the motivic classes of moduli stacks of semistable parabolic bundles with connections on $(X,D)$ and motivic classes of moduli stacks of semistable parabolic Higgs bundles on $(X,D)$. As a by-product we give a criteria for non-emptiness of these moduli stacks, which can be viewed as a version of the Deligne-Simpson problem.

\section{Introduction and main results} \subsection{Overview} Let $\mathbf{k}$ be a field of characteristic zero and $X$ be a smooth geometrically connected projective curve over~$\mathbf{k}$ (geometric connectedness means that $X$ remains connected after the base change to an algebraic closure of $\mathbf{k}$). In~\cite{FedorovSoibelmans} we calculated the motivic classes of moduli stacks of semistable Higgs bundles on $X$. These motivic classes are closely related to Donaldson--Thomas invariants, see~\cite{KontsevichSoibelman08,KontsevichSoibelman10}. In~\cite{FedorovSoibelmans} we also calculated the motivic classes of moduli stacks of vector bundles with connections on $X$ by relating them to the motivic classes of stacks of semistable Higgs bundles. In this paper, we extend these results to the parabolic case. Some of our results are parallel to the results of A.~Mellit in the case of finite fields (see~\cite{MellitPunctures}). One difference is that we fix the eigenvalues of the residues. Another difference is that by working over a field of characteristic zero, we are also able to treat bundles with connections. We also note that the calculation of the motivic class requires subtler techniques, than the calculation of the volume of the corresponding stack over a finite field. \subsection{Moduli stacks} Let us briefly describe the moduli stacks whose motivic classes we will be interested in. There will be three classes of stacks. \subsubsection{Parabolic bundles with connections}\label{sect:IntroConn} Let $D\subset X(\mathbf{k})$ be a non-empty set of rational points of $X$. A \emph{parabolic bundle} of type $(X,D)$ is a collection $\mathbf E=(E,E_{\bullet,\bullet})$, where $E$ is a vector bundle over $X$ and $E_{x,\bullet}$ is a flag in its fiber $E_x$ for $x\in D$: \[ E_x=E_{x,0}\supseteq E_{x,1}\supseteq\dots\supseteq E_{x,l}\supseteq\cdots,\qquad E_{x,l}=0\quad \text{for} \ l\gg0. \] A \emph{connection} on $E$ with \emph{poles bounded by $D$} is a morphism of sheaves of abelian groups $\nabla\colon E\to E\otimes\Omega_X(D)$ satisfying Leibniz rule. (Here $\Omega_X$ is the canonical line bundle on $X$.) In this case for $x\in D$ one defines the residue of the connection $\res_x\nabla\in\End(E_x)$. Let $\zeta=\zeta_{\bullet,\bullet}=(\zeta_{x,j})$ be a sequence of elements of $\mathbf{k}$ indexed by $D\times\mathbb{Z}_{>0}$ such that $\zeta_{x,j}=0$ for $j\gg0$. Let $\mathcal{C}{onn}(X,D,\zeta)$ denote the moduli stack parameterizing collections $(E,E_{\bullet,\bullet},\nabla)$, where $(E,E_{\bullet,\bullet})$ is a parabolic bundle of type $(X,D)$, $\nabla$~is a connection on $E$ with poles bounded by $D$ such that $(\res_x\nabla-\zeta_{x,j}1)(E_{x,{j-1}})\subset E_{x,j}$ for all $x\in D$ and $j>0$. We usually skip $X$ and $D$ from the notation as they are fixed, denoting $\mathcal{C}{onn}(X,D,\zeta)$ simply by $\mathcal{C}{onn}(\zeta)$. We call the points of $\mathcal{C}{onn}(\zeta)$ \emph{parabolic bundles with connections of type $(X,D)$ with eigenvalues $\zeta$.} For a parabolic bundle $\mathbf E=(E,E_{\bullet,\bullet})$ define the \emph{class} of $\mathbf E$ as the following collection of integers: \begin{equation}\label{eq:class} (\rk E,\dim{E_{x,j-1}}-\dim{E_{x,j}},\deg E)\in\mathbb{Z}_{\ge0}\times\mathbb{Z}_{\ge0}[D\times\mathbb{Z}_{>0}]\times\mathbb{Z}. \end{equation} We also set $\rk\mathbf E:=\rk E$. The stack $\mathcal{C}{onn}(\zeta)$ decomposes according to the classes of parabolic bundles; denote the component corresponding to parabolic bundles of class $\gamma$ by $\mathcal{C}{onn}_\gamma(\zeta)$. We will see that this stack is an Artin stack of finite type over $\mathbf{k}$. One of our main results (see Section~\ref{sect:ExplAnswers} and Theorem~\ref{th:ExplAnsw2}) is the calculation of the motivic class of this stack. \subsubsection{Parabolic Higgs bundles}\label{sect:IntroHiggs} Let $\zeta$ be as above. A \emph{parabolic Higgs bundle with eigenvalues $\zeta$} is a triple $(E,E_{\bullet,\bullet},\Phi)$, where $(E,E_{\bullet,\bullet})$ is a parabolic bundle of type $(X,D)$, $\Phi\colon E\to E\otimes\Omega_X(D)$ is a morphism of $\mathcal O_X$-modules (called a \emph{Higgs field on $(E,E_{\bullet,\bullet})$}) such that for all $x\in D$ and $j>0$ we have \[ (\Phi-\zeta_{x,j}1)(E_{x,{j-1}})\subset E_{x,j}\otimes\Omega_X(D)_x. \] Denote the category and the stack of such Higgs bundles by $\mathcal{H}{iggs}(\zeta)$. Unfortunately, this stack is not of finite type over $\mathbf{k}$, and in fact, has an infinite motivic volume. To resolve the problem we endow the category with a stability structure. Let $\sigma=\sigma_{\bullet,\bullet}$ be a sequence of real numbers indexed by $D\times\mathbb{Z}_{>0}$. Let $\kappa\in\mathbb{R}_{\ge0}$. We define \emph{the $(\kappa,\sigma)$-degree} of a parabolic bundle $\mathbf E=(E,E_{\bullet,\bullet})$ by \[ \deg_{\kappa,\sigma}\mathbf E:=\kappa\deg E+\sum_{x\in D}\sum_{j>0}\sigma_{x,j}(\dim E_{x,j-1}-\dim E_{x,j})\in\mathbb{R}. \] If $\mathbf E\ne0$, we define the \emph{$(\kappa,\sigma)$-slope} of $\mathbf E$ as $\deg_{\kappa,\sigma}\mathbf E/\rk\mathbf E$. We say that a sequence $\sigma=\sigma_{\bullet,\bullet}$ of real numbers indexed by $D\times\mathbb{Z}_{>0}$ is a \emph{sequence of parabolic weights} if for all $x\in D$ we have \begin{equation}\label{eq:StabCond} \sigma_{x,1}\le\sigma_{x,2}\le\cdots \end{equation} and for all $x$ and $j$ we have $\sigma_{x,j}\le\sigma_{x,1}+1$. Let $\sigma$ be a sequence of parabolic weights. Let $\mathbf E=(E,E_{\bullet,\bullet})$ be a parabolic bundle. Let $F\subset E$ be a saturated vector subbundle (that is, $E/F$ is torsion free). Set $F_{x,j}:=F_x\cap E_{x,j}$. Then $\mathbf F:=(F,F_{\bullet,\bullet})$ is a parabolic bundle. We say that a Higgs bundle~$(\mathbf E,\Phi)$ is \emph{$\sigma$-semistable}, if for all saturated subbundles $F$ of $E$ preserved by~$\Phi$ the $(1,\sigma)$-slope of the corresponding parabolic bundle $\mathbf F$ is less than or equal to that of $\mathbf E$. We have an open substack $\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)$ of $\mathcal{H}{iggs}_\gamma(\zeta)$ classifying $\sigma$-semistable parabolic Higgs bundles. This stack is of finite type over $\mathbf{k}$; we will calculate its motivic class (see Section~\ref{sect:ExplAnswers} and Theorem~\ref{th:ExplAnsw}). We note that condition~\eqref{eq:StabCond} is imposed on $\sigma$ to ensure that we have Harder--Narasimhan filtrations for parabolic Higgs bundles. We also note that, scaling $\kappa$ and $\sigma$ by the same positive real number scales all the slopes by the same number. This is why we restrict to the case $\kappa=1$ above (see Remark~\ref{rm:kappa} for more details). \subsubsection{Semistable parabolic bundles with connections}\label{sect:IntroConnSS} We can also impose stability conditions on parabolic bundles with connections. Moreover, for non-resonant connections we can work with more general stability conditions, than those for Higgs bundles defined in the previous paragraph. A sequence $\zeta$ as above is called \emph{non-resonant} if for all $x\in X$ and all $i,j>0$ we have $\zeta_{x,i}-\zeta_{x,j}\notin\mathbb{Z}_{\ne0}$. Take $\kappa\in\mathbb{R}_{\ge0}$ and a sequence $\sigma$ of real numbers indexed by $D\times\mathbb{Z}_{>0}$ and satisfying condition~\eqref{eq:StabCond}. \looseness=1 Assume that $\zeta$ is non-resonant. We define $(\kappa,\sigma)$-semistability of parabolic bundles with connections similarly to semistability of Higgs bundles but using the $(\kappa,\sigma)$-slope. Denote the corresponding moduli stack by $\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)$; this is an open substack of $\mathcal{C}{onn}_\gamma(\zeta)$. If $\zeta$ is resonant, then $\sigma$ has to satisfy some additional conditions (see Proposition~\ref{pr:HN2} and Re\-mark~\ref{rm:resonant}). \subsection{Motivic Donaldson--Thomas invariants}\label{sect:DT} Our formulas for motivic classes of the moduli stacks above are all given in terms of certain motivic classes $\overline B_\gamma$ called \emph{motivic Donaldson--Thomas invariants} (see Section~\ref{sect:Aftermath} for this terminology), which we are going to define. First of all, we recall that in~\cite[Section~2]{FedorovSoibelmans} we defined (following earlier works \cite[Section~1]{Ekedahl09}, \cite{Joyce07}, and~\cite{KontsevichSoibelman08}) the ring of motivic classes of Artin stacks denoted $\Mot(\mathbf{k})$. We also defined its dimensional completion $\cMot(\mathbf{k})$. For an Artin stack $\mathcal S$ of finite type over $\mathbf{k}$ we have its motivic class $[\mathcal S]\in\Mot(\mathbf{k})$. We denote its image in $\cMot(\mathbf{k})$ by the same symbol. For a curve $X$ and a partition $\lambda$ we defined the series $J_\lambda^{\rm mot}(z),H_\lambda^{\rm mot}(z)\in\cMot(\mathbf{k})[[z]]$ in~\cite[Section~1.3.2]{FedorovSoibelmans}. The definitions (especially of $H_\lambda^{\rm mot}(z)$) are somewhat long, so we will not recall them here inviting the reader to look into~\cite{FedorovSoibelmans}. We only note that $J_\lambda^{\rm mot}(z)$ and $H_\lambda^{\rm mot}(z)$ are defined in terms of the motivic zeta-function of~$X$ (cf.~\eqref{eq:MotZeta} below). In particular, they only depend on~$X$ but not on~$D$. In this paper, we will denote them by $J_{\lambda,X}^{\rm mot}(z)$ and $H_{\lambda,X}^{\rm mot}(z)$ respectively to emphasize that they depend on the curve $X$ and to ensure that they are not confused with motivic modified Macdonald polynomials $\tilde H_\lambda^{\rm mot}(w_\bullet;z)$ and with motivic Hall--Littlewood polynomials $H_\lambda^{\rm mot}(w_\bullet)$ defined below. The modified Macdonald polynomials $\tilde H_\lambda(w_\bullet;q,z)$ are symmetric functions in variables $w_\bullet=(w_1,w_2,\dots)$ with coefficients in $\mathbb{Z}[q,z]$. In~\cite[Definition~2.5]{MellitPunctures} the modified Macdonald polynomials are defined as symmetric functions with coefficients in $\mathbb{Q}[q,z]$ but it is well-known that the coefficients are integers (see, e.g., \cite{HaglundEtAlOnMacdonaldPoly} and references therein). Note that, formally speaking, symmetric functions are not polynomials (they become polynomials upon plugging in $w_{N+1}=w_{N+2}=\dots=0$). Let $\mathbb{L}=\big[\mathbb A_\mathbf{k}^1\big]$ be the motivic class of the affine line. We denote by $\tilde H_\lambda^{\rm mot}(w_\bullet;z)$ the symmetric function with coefficients in $\Mot(\mathbf{k})[z]$ obtained from $\tilde H_\lambda(w_\bullet;q,z)$ by substituting $\mathbb{L}$ for~$q$. We denote their images in the ring of symmetric functions with coefficients in $\cMot(\mathbf{k})[z]$ by $\tilde H_\lambda^{\rm mot}(w_\bullet;z)$ as well. Let $\Gamma_+$ denote the commutative monoid of sequences $(r,r_{\bullet,\bullet},d)$, where $r$ is a nonnegative integer, $r_{\bullet,\bullet}$ is a sequence of nonnegative integers indexed by $D\times\mathbb{Z}_{>0}$, $d$ is an integer, subject to the following conditions: \begin{enumerate}\itemsep=0pt \item[(i)] For all $x\in D$ we have $\sum\limits_{j=1}^{\infty}r_{x,j}=r$. In particular, $r_{x,j}=0$ for $j$ large enough. \item[(ii)] If $r=0$, then $d=0$ (and so $r_{x,j}=0$ for all $x$ and $j$). \end{enumerate} The operation on $\Gamma_+$ is the componentwise addition. For $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$ we set $\rk\gamma=r$. The significance of the monoid $\Gamma_+$ is that the class of a parabolic bundle $\mathbf E$ defined by~\eqref{eq:class} is an element of $\Gamma_+$. We also need a submonoid $\Gamma_+'\subset\Gamma_+$ given by $d\le 0$. Consider the completed monoid ring $\Mot(\mathbf{k})[[\Gamma_+']]$, we write its elements as $\sum\limits_{\gamma\in\Gamma_+'}A_\gamma e_\gamma$, where $A_\gamma\in\Mot(\mathbf{k})$, $e_\gamma$ are basis vectors. It is convenient to identify $e_\gamma$ with a monomial \[ w^r\prod_{x\in D}\prod_{j=1}^{\infty}w_{x,j}^{r_{x,j}}z^d, \] where $w_{\bullet,\bullet}=(w_{x,j})$ is a sequence of variables indexed by $D\times\mathbb{Z}_{>0}$. Then we identify $\Mot(\mathbf{k})[[\Gamma_+']]$ with a subring of $\Mot(\mathbf{k})\big[\big[w,w_{\bullet,\bullet},z^{-1}\big]\big]$. Similarly, we consider the completed monoid ring $\cMot(\mathbf{k})[[\Gamma_+']]$. We note that these rings are closely related to completed quantum tori considered in~\cite{KontsevichSoibelman08,KontsevichSoibelman10}. In our case, they are commutative, essentially because we are working with 2-dimensional Calabi--Yau categories; see Sections~\ref{sect:Aftermath} and~\ref{sect:CatsOverPar} for more details. Finally, we need the notion of \emph{plethystic exponent and logarithm}. Let $\Mot(\mathbf{k})[[\Gamma_+']]^0$ denote the subset of $\Mot(\mathbf{k})[[\Gamma_+']]$ consisting of elements with zero constant terms. Then we have a~bijection \[ \Exp:\Mot(\mathbf{k})[[\Gamma_+']]^0\to1+\Mot(\mathbf{k})[[\Gamma_+']]^0 \] called the plethystic exponent. We refer the reader to Section~\ref{sect:Plethystic} for the definition. Let the plethystic logarithm $\Log$ be the inverse bijection. Let us write \[ \mathbb{L}\cdot\Log\left(\sum_\lambda w^{|\lambda|} J_{\lambda,X}^{\rm mot}\big(z^{-1}\big)H_{\lambda,X}^{\rm mot}\big(z^{-1}\big)\prod_{x\in D}\tilde H_\lambda^{\rm mot}\big(w_{x,\bullet};z^{-1}\big)\right)= \sum_{\gamma\in\Gamma'_+}\overline B_\gamma e_\gamma, \] where the sum in the LHS is over all partitions. We call the elements $\overline B_\gamma$ the \emph{Donaldson--Thomas invariants}. Note that $\overline{B}_0=0$. When $X=\P^1_\mathbf{k}$, we can define motivic Donaldson--Thomas invariants $B_\gamma$ by a simpler formula valid in $\Mot(\mathbf{k})$: \begin{equation}\label{eq:DT_P1intro} \mathbb{L}\cdot\Log\left(\sum_\lambda\frac{w^{|\lambda|}\prod\limits_{x\in D}\tilde H_\lambda^{\rm mot}\big(w_{x,\bullet};z^{-1}\big)} {\prod\limits_{h\in\Hook(\lambda)} \big(\mathbb{L}^{a(h)}-z^{-l(h)-1}\big)\big(\mathbb{L}^{a(h)+1}-z^{-l(h)}\big)}\right)= \sum_{\gamma\in\Gamma'_+}B_\gamma e_\gamma, \end{equation} where $\Hook(\lambda)$ stands for the set of hooks of $\lambda$, $a(h)$ and $l(h)$ stand for the armlength and the leglength of the hook $h$ respectively. We show that for $X=\P^1$ the images of $B_\gamma$ in $\cMot(\mathbf{k})$ are equal to $\overline B_\gamma$. \subsection{Explicit formulas}\label{sect:ExplAnswers} The following explicit formulas for the motivic classes are parts of Theorem~\ref{th:ExplAnsw2}, Theorem~\ref{th:ExplAnsw}, and Theorem~\ref{th:ExplAnsw3} respectively. Let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$, $\gamma\ne0$. Let $\zeta$ be as in Section~\ref{sect:IntroConn}. For $\kappa\in\mathbf{k}$ we define the $(\kappa,\zeta)$-degree and the $(\kappa,\zeta)$-slope of parabolic bundles similarly to $(\kappa,\sigma)$-degree and $(\kappa,\sigma)$-slope defined in Section~\ref{sect:IntroConnSS}; the only difference is that the $(\kappa,\zeta)$-degree and $(\kappa,\zeta)$-slope take values in~$\mathbf{k}$. For each $\tau\in\mathbf{k}$, define the elements $C_\gamma(\zeta)\in\cMot(\mathbf{k})$, where $\gamma$ ranges over elements of $\Gamma_+'$ such that $\gamma=0$ or the $(1,\zeta)$-slope of $\gamma$ is $\tau$, by the following formula \begin{equation*} \sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{1,\zeta}\gamma=\tau\rk\gamma}}\mathbb{L}^{-\chi(\gamma)}C_\gamma(\zeta)e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{1,\zeta}\gamma=\tau\rk\gamma}} \overline B_\gamma e_\gamma \right), \end{equation*} where $\chi(\gamma):=(g-1)r^2+\sum\limits_{x\in D}\sum\limits_{j<j'}r_{x,j}r_{x,j'}$, $g$ is the genus of~$X$. Let $\gamma\in\Gamma_+$ be such that $\deg_{1,\zeta}\gamma=0$. Then, according to Theorem~\ref{th:ExplAnsw2}, the stack $\mathcal{C}{onn}_\gamma(\zeta)$ is of finite type over $\mathbf{k}$ and we have in $\cMot(\mathbf{k})$ \[ [\mathcal{C}{onn}_\gamma(\zeta)]=C_{(r,r_{\bullet,\bullet},d-Nr)}(\zeta), \] whenever $N$ is large enough. If $\deg_{1,\zeta}\gamma\ne0$, then the stack $\mathcal{C}{onn}_\gamma(\zeta)$ is empty. Next, assume that $\zeta$ and $\sigma$ are as in Section~\ref{sect:IntroHiggs}, for $\tau\in\mathbb{R}$ define the elements $H_\gamma(\zeta,\sigma)\in\cMot(\mathbf{k})$ by the following formula \begin{equation*} \sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{0,\zeta}\gamma=0\\ \deg_{1,\sigma}\gamma=\tau\rk\gamma}}\mathbb{L}^{-\chi(\gamma)}H_\gamma(\zeta,\sigma)e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{0,\zeta}\gamma=0\\ \deg_{1,\sigma}\gamma=\tau\rk\gamma}} \overline B_\gamma e_\gamma \right). \end{equation*} Assume that $\deg_{0,\zeta}\gamma\!=\!0$, where $\gamma\in\Gamma_+'$. Then, according to Theorem~\ref{th:ExplAnsw}, the stack $\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)$ is of finite type over $\mathbf{k}$ and we have \[ \big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]=H_{(r,r_{\bullet,\bullet},d-Nr)}(\zeta,\sigma), \] whenever $N$ is large enough. If $\deg_{0,\zeta}\gamma\ne0$, then the stack $\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)$ is empty. Finally, assume that $\zeta$ and $(\kappa,\sigma)$ are as in Section~\ref{sect:IntroConnSS}. For $\tau\in\mathbf{k}$, $\tau'\in\mathbb{R}$ define the elements $C_\gamma(\zeta,\kappa,\sigma)\in\cMot(\mathbf{k})$ by the following formula \begin{equation*} \sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{1,\zeta}\gamma=\tau\rk\gamma\\ \deg_{\kappa,\sigma}\gamma=\tau'\rk\gamma }}\mathbb{L}^{-\chi(\gamma)}C_\gamma(\zeta,\kappa,\sigma)e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{1,\zeta}\gamma=\tau\rk\gamma\\ \deg_{\kappa,\sigma}\gamma=\tau'\rk\gamma }} \overline B_\gamma e_\gamma \right). \end{equation*} Let $\gamma\in\Gamma_+$ be such that $\deg_{1,\zeta}\gamma=0$. Then, according to Theorem~\ref{th:ExplAnsw3}, we have \[ \big[\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]=C_{(r,r_{\bullet,\bullet},d-Nr)}(\zeta,\kappa,\sigma), \] whenever $N$ is large enough. If $\deg_{1,\zeta}\gamma\ne0$, then the stack $\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)$ is empty. If $X=\P^1$, we get similar results valid in $\Mot(\mathbf{k})$, by replacing $\overline B_\gamma$ with $B_\gamma$ defined by a~simpler formula~\eqref{eq:DT_P1intro}. \begin{Remark} We note that each of the above motivic classes depends only on finitely many DT-invariants. Indeed, $[\mathcal{C}{onn}_\gamma(\zeta)]$, $\big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]$, and $\big[\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]$ depend only on $\overline B_{\gamma'}$ with $\rk\gamma'\le\rk\gamma$ and for a given $\gamma$ there are only finitely many such $\gamma'\in\Gamma'_+$. \end{Remark} \begin{Remark} We note also that all the stacks whose motivic classes we are calculating are of finite type over $\mathbf{k}$, so their motivic classes are defined in $\Mot(\mathbf{k})$. However, we can only calculate their motivic classes in $\cMot(\mathbf{k})$ except when $X=\P^1$. The reason is that, our calculation is based on the calculation of motivic classes of stacks of vector bundles on $X$ (without parabolic structures) with nilpotent endomorphisms. This calculation is performed in~\cite{FedorovSoibelmans} and is, in turn, based on the motivic analogue of Harder's residue formula (see~\cite[Theorem~1.5.1 and Section~4]{FedorovSoibelmans} and~\cite[Theorem~2.2.3]{HarderAnnals}). This formula, which is essentially saying that ``all vector bundles have essentially the same motivic number of Borel reductions'' involves some limiting process and is, therefore, only valid in the completed ring~$\cMot(\mathbf{k})$. \end{Remark} \subsection{Aftermath}\label{sect:Aftermath} In Section~\ref{sect:DT} we defined the classes $\overline B_\gamma$. These classes should be thought of as the Donaldson--Thomas invariants of the stack $\mathcal{H}{iggs}(0)$ of parabolic Higgs bundles with nilpotent residues. Note that this stack is the cotangent bundle of $\mathcal{B}{un}^{\rm par}(X,D)$, while the stacks $\mathcal{H}{iggs}(\zeta)$ and $\mathcal{C}{onn}(\zeta)$ are \emph{twisted cotangent bundles}. We emphasize that $\overline B_\gamma$ do not depend on~$\zeta$,~$\kappa$, and~$\sigma$. The meaning of the formulas in Section~\ref{sect:ExplAnswers} is that the Donaldson--Thomas invariants of these twisted cotangent bundles are obtained by restricting the range of $\gamma$ to the submonoid $\deg_{0,\zeta}\gamma=0$ in the case of $\mathcal{H}{iggs}(\zeta)$ and to the submonoid $\deg_{1,\zeta}\gamma=0$ in the case of $\mathcal{C}{onn}(\zeta)$. Another feature of the formulas is that the motivic classes of the stacks depend on \emph{equations} satisfied by~$\kappa$ and~$\sigma$ rather than on inequalities. In other words, there is no \emph{wall-crossing} in our case. This is not very surprising, as the category $\mathcal{H}{iggs}(0)$, being a cotangent bundle of $\mathcal{B}{un}^{\rm par}(X,D)$, is a 2-dimensional Calabi--Yau category, cf.~\cite{RenSoibelman}. One can speculate that similar results should be valid for the twisted cotangent stacks to the moduli stack of objects of any reasonable 1-dimensional category. Note that such cotangent stacks were studied by G.~Dobrovolska, V.~Ginzburg, and R.~Travkin in~\cite{DobrovolskaGinzburgTravkin}. Another example of such a twisted cotangent stack is the category of vector bundles with irregular connections and appropriate level structures. This example is certainly more complicated as the corresponding abelian category has infinite homological dimension. We hope to return to this question in subsequent publications. The formulas in Section~\ref{sect:ExplAnswers} are explicit but complicated. However, one can see that the motivic classes under considerations belong to the sub-$\lambda$-ring of $\cMot(\mathbf{k})$ generated by $\mathbb{L}$, $X$, and the inverses of $\mathbb{L}^i-1$ for $i\ge1$. We note that this ring is probably \emph{strictly larger}, than the subring of $\cMot(\mathbf{k})$ generated by $\mathbb{L}$, the symmetric powers $X^{(i)}$, and the inverses of $\mathbb{L}^i-1$ for $i\ge1$. The reason is that $\cMot(\mathbf{k})$ is unlikely to be a special $\lambda$-ring, see Section~\ref{sect:OtherRel}. On the other hand, if $X=\P^1$, then all our motivic classes are rational functions in $\mathbb{L}$ with denominators being products of $\mathbb{L}^i-1$ for $i\ge1$. \subsection{Other results} It is clear from the above formulas that we have a lot of equalities between different motivic classes of Higgs bundles and bundles with connections. In particular, we show in Propositions~\ref{pr:ConnUniversal} and~\ref{pr:ConnUniversal2} that every motivic class of the form $\big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]$ or $\big[\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]$ is equal to some motivic class of the form $[\mathcal{C}{onn}_\gamma(\zeta)]$, provided that $\mathbf{k}$ is not a finite extension of $\mathbb{Q}$. As a~consequence, we derive from results of Crawley-Boevey~\cite{CrawleyBoeveyIndecompPar} a criterion of non-emptiness of our moduli stacks. It is not difficult to see that if $X\ne\P^1_\mathbf{k}$, then the stack $\big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]$ is non-empty if and only if $\deg_{0,\zeta}\gamma=0$, while the stack $\big[\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]$ is non-empty if and only if $\deg_{1,\zeta}\gamma=0$. For $X=\P^1_\mathbf{k}$ the question is much more subtle and is related to the so-called Deligne--Simpson problem. This problem was originally stated for $\mathbf{k} =\mathbb C$ in~\cite{SimsponProducts}. It may be reformulated for an arbitrary algebraically closed field $\mathbf{k}$ of characteristic~$0$ as follows: given a~sequence of $\mathfrak{gl}_r$-conjugacy classes $C_\bullet$ indexed by $D$, does there exist a pair $(E, \nabla)$ consisting of a~rank~$r$ vector bundle and a connection $\nabla$ on $E$ with poles bounded by $D$ such that $\Res_x \nabla\in C_x$ for all $x\in D$? Bundles with connections $(E,\nabla)$ parameterized by $\mathcal{C}{onn}_{\gamma}\big(\P^1_\mathbf{k},D,\zeta\big)$ are exactly bundles with connections such that each residue $\Res_x\nabla$ lies in the closure of a conjugacy class determined by $\gamma$ and $\zeta$ (see~\cite{Crawley-Boevey:Indecomposable}). If this conjugacy class is semisimple for each $x\in D$, then the elements of $\mathcal{C}{onn}_{\gamma}\big(\P^1_\mathbf{k},D,\zeta\big)$ are the solutions of the corresponding Deligne--Simpson problem. For a comprehensive survey of the Deligne--Simpson problem, see \cite{Crawley-Boevey:Indecomposable,Kostov2004Deligne,Simpson:MiddleConv}. We note that if $\mathbf{k}$ is not algebraically closed, one can ask a subtler question of whether there is a $\mathbf{k}$-rational point in $\big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]$ or $\big[\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]$. We do not know the answer to this question. A somewhat similar question for moduli spaces of quiver representations is considered by V.~Hoskins and F.~Schaffhauser in~\cite{hoskins2017rational}. At this point, we would like to emphasize that working with not necessarily algebraically closed fields is inevitable in the motivic setup: even if one is only interested in the case $\mathbf{k}=\mathbb C$, one still has to consider all fields of characteristic zero, see Remark~\ref{rm:ArbFields} below. One of the motivations for this work is the non-abelian Hodge theory of C.~Simpson (see~\cite{SimpsonHarmonicNoncompact}). In this paper, Simpson constructs an equivalence between a category of parabolic bundles with connections and a category of Higgs bundles. In Proposition~\ref{pr:Simpson}, we show that the corresponding stacks have equal motivic classes. We note that neither statement can be derived from the other (cf.~\cite[Remark~1.2.2]{FedorovSoibelmans}). We note also, that it is clear from our results that there are many more equalities of motivic classes, than those that one can guess from the non-abelian Hodge theory. We remark that V.~Hoskins and S.~Pepin Lehalleur have shown in~\cite[Theorem~4.2]{HoskinsLehalleurOnVoevodskyMotive} that the Voevodsky motives of the coarse moduli spaces of bundles with connections and of Higgs bundles are equal in the case when the rank and the degree are coprime. One can ask whether this can be upgraded to the parabolic situation. \subsection{Other relations with previous work}\label{sect:OtherRel} We have already noted that our results are closely related with the results of Mellit~\cite{MellitPunctures}. One difference is that Mellit counts the weighted number of points over a finite field, while we work over a field of characteristic zero and calculate motivic classes. Mellit counts the volumes of moduli stacks of Higgs bundles but is not considering bundles with connections. Another difference is that Mellit is not fixing the eigenvalues of Higgs fields. \looseness=-1 On the other hand, Mellit's answers are simpler as they do not involve Schiffmann's polynomials~$H_\lambda$. In fact, Mellit's simplification of Schiffmann's formula from~\cite{SchiffmannIndecomposable} has nothing to do with parabolic structures. This simplification is the content of Mellit's papers~\cite{MellitIntegrality,MellitPunctures,MR4105090}. We believe that this simplification \emph{does not go through} in the motivic case because Mellit is using the fact that the $\lambda$-ring structure on symmetric functions is special (which, roughly speaking, means that $\lambda_\bullet(xy)$ and $\lambda_\bullet(\lambda_\bullet(x))$ can be expressed in terms of $\lambda_\bullet(x)$ and $\lambda_\bullet(y)$). It is known (see~\cite{LarsenLuntsRational}) that the Grothendieck $\lambda$-ring of varieties is not a special $\lambda$-ring. We do not know whether the Grothendieck $\lambda$-ring of \emph{stacks} $\Mot(\mathbf{k})$ and its completion $\cMot(\mathbf{k})$ are special. In any case, if one replaces $\cMot(\mathbf{k})$ by some its quotient that is a special $\lambda$-ring, then one expects that the Mellit's simplifications are valid in this quotient. Examples of such quotients are the Grothendieck ring of the category of Chow motives and the maximal special quotient of $\cMot(\mathbf{k})$ (see~\cite{LarsenLuntsRational}). The paper~\cite{ChuangDiaconescuDonagiPantev}, although conjectural, contains an alternative approach to the problem via upgrading the computation of the motivic class of Higgs bundles to the problem about motivic Pandharipande--Thomas invariants on the non-compact Calabi--Yau 3-fold associated with the spectral curve. Note also that Mozgovoy and Schiffmann in~\cite{MozgovoySchiffmanOnHiggsBundles} consider Higgs bundles with a twist by an arbitrary line bundle of degree at least $2g-2$, where $g$ is the genus of $X$. However, they do not consider parabolic structures and do not fix eigenvalues. Finally we note that the general philosophy of Donaldson--Thomas invariants and the approach via motivic and cohomological Hall algebras (see~\cite{KontsevichSoibelman08,KontsevichSoibelman10}) are applicable to our situation. For more details about the approach that uses motivic Hall algebras we refer the reader to~\cite[Section~1.6, Remark~3.6.3]{FedorovSoibelmans}. \subsection{Organization of the article} In Section~\ref{sect:ParBundles} we define the category $\mathcal{B}{un}^{\rm par}(X,D)$ of parabolic bundles and its graded stack of objects denoted by the same letter. Most of our stacks below will be stacks over $\mathcal{B}{un}^{\rm par}(X,D)$. In Section~\ref{sect:ParPairs} we study the stack of bundles with endomorphisms. This stack is the main intermediate object in our calculations. First, we calculate the motivic classes of stacks of parabolic bundles with nilpotent endomorphisms with fixed generic type. The calculation is based on Theorem~\ref{th:Factorization}, saying that these motivic classes are products of motivic classes of similar stacks without parabolic structures and of ``local stacks'' independent of the curve. This is a motivic analogue of~\cite[Theorem~5.6]{MellitPunctures}. However, the proof in the motivic case is significantly more involved and, hopefully, more conceptual. The motivic classes of stacks of vector bundles (without parabolic structures) with nilpotent endomorphisms are calculated in the proof of~\cite[Theorem~1.4.1]{FedorovSoibelmans}. Since the motivic classes of ``local stacks'' are independent of the curve, it is enough to calculate them for $X=\P^1$; we do this using the ideas and results of Mellit~\cite{MellitPunctures}. Then we use the formalism of plethystic powers to calculate the motivic classes of stacks of parabolic bundles with arbitrary endomorphisms. In Section~\ref{sect:HiggsnEigenval} we study parabolic Higgs bundles with fixed eigenvalues. If the eigenvalues are equal to zero, then the endomorphisms of a given parabolic bundle and the Higgs fields on this bundle are parameterized by vector spaces whose dimensions differ by some Euler characteristic. Thus, it is easy to relate the motivic classes of the two stacks. If the eigenvalues are not zero, then not every parabolic bundle admits a Higgs fields with these eigenvalues. We give a criterion for existence of such a Higgs field in Lemma~\ref{lm:existence}. This allows us to express the motivic class of parabolic Higgs bundles with fixed eigenvalues in terms of the motivic class of the so-called isoslopy parabolic bundles with endomorphisms (see Proposition~\ref{pr:Sasha}). The motivic class of isoslopy parabolic bundles with endomorphisms is derived from the results of Section~\ref{sect:ParPairs} with the help of a factorization formula (see Proposition~\ref{pr:IsoslProd}). This is analogous to~\cite[Proposition~3.5.1]{FedorovSoibelmans}. In Section~\ref{sect:Stability} we use a version of Kontsevich--Soibelman factorization formula to calculate the motivic classes of stacks of semistable Higgs bundles. These depend on two sets of parameters: the eigenvalues and the stability condition. Somewhat surprisingly, these two sets come symmetrically in the answer. Up to Section~\ref{sect:Stabilization} we work with nonpositive vector bundles, that is, vector bundles having no subbundle of positive degree. Without stability this restriction is inevitable as otherwise the moduli stacks would have infinite motivic volume. With a stability condition we can drop this technical restriction; the motivic classes of semistable parabolic Higgs bundles whose underlying vector bundles are not necessarily nonpositive are calculated in Section~\ref{sect:Stabilization}. In Section~\ref{sect:Conn} we study the moduli stack of bundles with connections~-- with or without stability condition. The strategy for connections is similar to that for Higgs bundles except that the corresponding stacks are of finite type over $\mathbf{k}$ even without stability conditions. So we first calculate the motivic classes of bundles with connections with given eigenvalues without stability conditions and without nonpositivity assumptions and then use a version of Kontsevich--Soibelman factorization formula to calculate the motivic classes of stacks of semistable bundles with connections. {\samepage In Section~\ref{sect:NonEmpty} we give a precise criterion for non-emptiness of moduli stacks of Higgs bundles, bundles with connections, or semistable bundles with connections. The idea is that such a~stack is non-empty if and only if its motivic class is non-zero. Using our explicit formulas we re-write each of our motivic classes as the motivic class of a stack of bundles with connections (without stability conditions). The non-emptiness of such a stack is decided using Lemma~\ref{lm:existence2} and Crawley-Boevey's result~\cite[p.~1334, Corollary]{CrawleyBoeveyIndecompPar}. } \section{Preliminaries} \subsection{Conventions} We denote by $\mathbf{k}$ a field of characteristic zero. We denote by $X$ a smooth projective geometrically connected curve over $\mathbf{k}$ (recall that geometric connectedness means that $X$ remains connected after the base change to an algebraic closure of $\mathbf{k}$). We denote by $D$ a set of $\mathbf{k}$-rational points of $X$ and by $\deg D$ the number of elements of $D$. If $E$ is a vector space or a vector bundle, we denote by $E^\vee$ the dual vector space (resp.~vector bundle). We identify vector bundles with their sheaves of sections. If $F$ is a coherent sheaf, we denote by $\mathcal{E}{nd}(F)$ the sheaf of its endomorphisms, we have $\mathcal{E}{nd}(F)=F^\vee\otimes F$ if $F$ is a vector bundle. \subsubsection{Partitions and nilpotent matrices} By a partition we mean a non-increasing sequence of integers $\lambda=\lambda_1\ge\lambda_2\ge\cdots$, where $\lambda_l=0$ for $l\gg0$. Set $|\lambda|:=\sum_i\lambda_i$. For partitions $\lambda$ and $\mu$ we set $\langle\lambda,\mu\rangle=\sum_i\lambda'_i\mu'_i$, where $\lambda'$ and $\mu'$ are the conjugate partitions. We denote by $\mathfrak{gl}_{r,\mathbf{k}}$ or simply by $\mathfrak{gl}_r$ the set of $r\times r$ matrices with entries in $\mathbf{k}$. We say that a~nilpotent matrix $n\in\mathfrak{gl}_{|\lambda|}=\mathfrak{gl}_{|\lambda|,\mathbf{k}}$ is of type $\lambda$, if for all $i\ge1$ we have $\dim\Ker n^i-\dim\Ker n^{i-1}=\lambda_i$. For each partition $\lambda$ choose a nilpotent matrix $n_\lambda$ of type $\lambda$. For concreteness, we can take for $n_\lambda$ the direct sum of nilpotent Jordan blocks, where the number of blocks of size $i\times i$ is equal to $\lambda_i-\lambda_{i+1}$. A sequence $(w_1,\dots,w_l,\dots)$ we denote by $w_\bullet$. \subsubsection{Stacks} We will be working with stacks. All our stacks will have affine stabilizers. Our stacks will be Artin stacks locally of finite type over a field except in Section~\ref{sect:Factorization}, where we will have to work with stacks whose points have stabilizers of infinite type. For a stack $\mathcal S$ we often abuse notation by writing $s\in\mathcal S$ to mean that $s$ is an object of the groupoid $\mathcal S(\mathbf{k})$, or an object of the groupoid $\mathcal S(K)$, where $K$ is an extension of $\mathbf{k}$. Following~\cite[Chapter~5]{LaumonMoretBailly} we say that a~$K'$-point~$\xi'$ of~$\mathcal S$ is \emph{equivalent} to a $K''$-point $\xi''$ of $\mathcal S$ if there is an extension $K\supset\mathbf{k}$ and $\mathbf{k}$-embeddings $K'\hookrightarrow K$, $K''\hookrightarrow K$ such that $\xi'_K$ is isomorphic to $\xi''_K$ (as an object of $\mathcal S(K)$). The corresponding equivalence classes are called \emph{points of $\mathcal S$}; the set of points is denoted by $|\mathcal S|$. See~\cite[Section~2]{FedorovSoibelmans} for more details. We write ``morphism of stacks'' to mean ``1-morphism of stacks''. We write ``of finite type'' to mean ``of finite type over $\mathbf{k}$''. \subsection{Motivic functions and motivic classes} Recall that in~\cite[Section~2]{FedorovSoibelmans} we defined (following~\cite[Section~1]{Ekedahl09},~\cite{Joyce07}, and~\cite{KontsevichSoibelman08}) the ring of motivic classes of Artin stacks denoted $\Mot(\mathbf{k})$. More generally, for an Artin stack $\mathcal X$ locally of finite type over $\mathbf{k}$, we defined the $\Mot(\mathbf{k})$-module of motivic functions on $\mathcal X$ denoted $\Mot(\mathcal X)$. For a morphism $f\colon \mathcal X\to\mathcal Y$ we have the pullback homomorphism $f^*\colon \Mot(\mathcal Y)\to\Mot(\mathcal X)$. The pushforward homomorphism $f_!\colon \Mot(\mathcal X)\to\Mot(\mathcal Y)$ is defined when $f$ is of finite type. We also defined the ring of completed motivic classes, denoted $\cMot(\mathbf{k})$, and $\cMot(\mathbf{k})$-modules of completed motivic functions $\cMot(\mathcal X)$ with a~(probably non-injective) morphism $\Mot(\mathcal X)\to\cMot(\mathcal X)$. We also defined the pullbacks and the pushforwards of completed motivic functions. We usually work with $\Mot(\mathbf{k})$ but our final results are formulated in $\cMot(\mathbf{k})$. We defined the notion of a constructible subset of a stack. If $\mathcal X\to\mathcal Y$ is a morphism of finite type, and $\mathcal S\subset\mathcal X$ is a constructible subset, we defined the motivic function $[\mathcal S\to\mathcal Y]\in\Mot(\mathcal Y)$. Recall~\cite[Proposition~2.6.1]{FedorovSoibelmans}: \begin{Proposition}\label{pr:MotFunEqual} Assume that we are given $A,B\in\Mot(\mathcal X)$ are such that for all field extensions $K\supset\mathbf{k}$ and for all $\mathbf{k}$-morphisms $\xi\colon \Spec K\to\mathcal X$ we have $\xi^*A=\xi^*B$. Then $A=B$. \end{Proposition} \begin{Remark}\label{rm:ArbFields} The previous proposition is one of the reasons we have to work with arbitrary fields. Indeed, even if we start with $\mathbf{k}=\mathbb C$, to be able to apply the proposition we have to consider all finitely generated extensions of~$\mathbb C$; see, for example, Section~\ref{sect:ProofFact}. \end{Remark} In Section~\ref{sect:NonEmpty} we will need the following proposition. \begin{Proposition}\label{pr:NonEmpty} An Artin stack of finite type over $\mathbf{k}$ is non-empty if and only if its motivic class in $\cMot(\mathbf{k})$ is not equal to zero. \end{Proposition} \begin{proof} The `if' direction is obvious. For the other direction assume for a contradiction that $\mathcal S$ is a non-empty Artin stack of finite type over $\mathbf{k}$ such that $[\mathcal S]=0\in\cMot(\mathbf{k})$. This means that for all $m\in\mathbb{Z}$ we have $[\mathcal S]\in F^m\Mot(\mathbf{k})$, where $F^\bullet$ is the dimensional filtration on $\Mot(\mathbf{k})$. According to~\cite[Propositions~3.5.6 and~3.5.9]{KreschStacks} every Artin stack of finite type with affine stabilizers has a~stratification by global quotients of the form $T/{\rm GL}_n$, where $T$ is a scheme. Thus, replacing $\mathcal S$ with a stratification and clearing the denominators, we may assume that~$\mathcal S$ is a disjoint union of integral affine schemes. Recall from~\cite[Section~2.5]{FedorovSoibelmans} that $\Mot(\mathbf{k})$ is the localization of the K-ring of varieties $\Mot_{\rm var}(\mathbf{k})$ with respect to the multiplicative set generated by $\mathbb{L}$ and $\mathbb{L}^i-1$, where $i>0$. Thus, multiplying $\mathcal S$ by a certain product of these elements, we may assume that the class of $\mathcal S$ in $\Mot_{\rm var}(\mathbf{k})$ belongs to the subgroup $F^{m-1}\Mot_{\rm var}(\mathbf{k})$ generated by the classes of the varieties of dimension at most $m-1$, where $m=\dim\mathcal S$. Compactifying each top-dimensional connected component of $\mathcal S$ and taking the resolution of singularities, we may assume that $\mathcal S$ is the disjoint union of smooth projective $\mathbf{k}$-varieties. Recall that the Hodge--Deligne polynomial of a smooth projective variety $Y$ is \[ \sum_{p,q=0}^{\dim Y}(-1)^{p+q}h^{p,q}(Y)u^pv^q, \] where $h^{p,q}=\dim H^q(Y,\wedge^p\Omega_Y)$. This extends uniquely to a~homomorphism $E\colon \Mot_{\rm var}(\mathbf{k})\to\mathbb{Z}[u,v]$. Clearly, $E([Y])$ has degree $2m$, if $Y$ is a smooth projective variety of dimension~$m$. On the other hand, $E([Y])$ has degree at most $2m-2$, if $Y$ is any variety of dimension at most $m-1$. We see that, on the one hand $E(\mathcal S)$ has degree~$2m$, on the other hand it has degree $2m-2$. We come to contradiction. \end{proof} \subsection{Principal bundles and special groups} Let $H$ be an algebraic group of finite type over $\mathbf{k}$. Recall that a \emph{principal $H$-bundle} over a $\mathbf{k}$-stack $\mathcal B$ is a stack $E$ together with a schematic smooth surjective morphism of finite type $E\to\mathcal B$ and an action $a\colon H\times_\mathbf{k} E\to E$ such that $H$ acts simply transitively on the fibers of $E\to\mathcal B$. More precisely, the simple transitivity means that the morphism $(a,p_2)\colon H\times E\to E\times_\mathcal B E$ is an isomorphism, where $p_2\colon H\times E\to E$ is the projection. A principal bundle is \emph{trivial} if there is an isomorphism $E\approx H\times\mathcal B$ compatible with the action and the projection to~$\mathcal B$. If $E$ is a principal $H$-bundle over $\mathcal B$ and $\mathcal B'\to\mathcal B$ is a morphism, then one gets an induced principal $H$-bundle $E\times_\mathcal B\cB'$ over $\mathcal B'$. The group $H$ is called \emph{special} if every principal $H$-bundle~$E$ over a~scheme $B$ of finite type over~$\mathbf{k}$ is locally trivial over $B$ in the Zariski topology. The following lemma is standard, see, e.g., \cite[Section~2.3]{BehrendDhillon}. \begin{Lemma}\label{lm:SpecialMot} Let $H$ be a special group and $E\to\mathcal B$ be a principal $H$-bundle, where $\mathcal B$ is an Artin stack of finite type over $\mathbf{k}$. Then in $\Mot(\mathbf{k})$ we have $[E]=[H][\mathcal B]$. \end{Lemma} \begin{proof} The case when $\mathcal B$ is a scheme is easily proved by Noetherian induction. If $\mathcal B$ is a stack, then, using again~\cite[Propositions~3.5.6 and~3.5.9]{KreschStacks}, we may assume that $\mathcal B$ is a global quotient: $\mathcal B=S/{\rm GL}_n$, where $S$ is a scheme. Then we have the cartesian diagram \begin{equation*} \begin{CD} E' @>>> E\\ @VVV @VVV\\ S @>>> \mathcal B, \end{CD} \end{equation*} where $E'=E\times_\mathcal B S$ and $E=E'/{\rm GL}_n$. Applying~\cite[Corollary~2.2.2]{FedorovSoibelmans} with $\mathcal S=\Spec\mathbf{k}$, we get $[E']=[E][{\rm GL}_n]$ and $S=[\mathcal B][{\rm GL}_n]$. Next, $E'$ is a principal $H$-bundle over the scheme $S$ so $[E']=[H][S]$. Combining these equations we easily get the required statement. \end{proof} Recall that a $\mathbf{k}$-group $U$ is \emph{unipotent} if it can be embedded into a group of strictly upper triangular matrices. Every unipotent subgroup is obtained from the additive group of the 1-dimensional vector space by iterated extensions. \begin{Lemma}\label{lm:SpecialRad} Let $H$ be an algebraic group of finite type over $\mathbf{k}$ and let $U$ be a unipotent subgroup. Assume that $H/U$ is special. Then $H$ is special. In particular, every unipotent group is special. \end{Lemma} \begin{proof} Assume first that $H$ is a unipotent group. We claim that every principal $H$-bundle over an affine scheme is trivial. We prove this by induction on $\dim H$. If $\dim H=1$, then~$H$ is the additive group and the principal $H$-bundles over a scheme $B$ are classified by the coherent cohomology group $H^1(B,\mathcal O_B)$, which vanishes as soon as $B$ is affine. If $\dim H>1$, then there is a subgroup $H'\subset H$ such that $\dim H'<\dim H$ and $\dim(H/H')<\dim H$. The groups~$H'$ and~$H/H'$ are unipotent. Recall that the principal $H$-bundles over $B$ are classified by the first non-abelian \'etale cohomology group $H_{\text{\'et}}^1(B,H)$. Now the statement follows from the exact sequence $H_{\text{\'et}}^1(B,H')\to H_{\text{\'et}}^1(B,H)\to H_{\text{\'et}}^1(B,H/H')$. In particular, every unipotent group is special. Now assume that $H/U$ is special, where $U$ is unipotent, and consider, for an affine $B$, the exact sequence $1=H_{\text{\'et}}^1(B,U)\to H_{\text{\'et}}^1(B,H)\to H_{\text{\'et}}^1(B,H/U)$. We see that a principal $H$-bundle is trivial over an affine scheme if the induced principal $H/U$-bundle is trivial. The lemma follows. \end{proof} \begin{Lemma}\label{lm:Zspecial} Let $Z_\lambda$ be the centralizer of $n_\lambda$ in ${\rm GL}_{|\lambda|}$. Then $Z_\lambda$ is a special group. \end{Lemma} \begin{proof} It is well-known that the quotient of $Z_\lambda$ by its unipotent radical is the product of ${\rm GL}_{r_i}$ for some $r_i\in\mathbb{Z}_{>0}$. It remains to note that ${\rm GL}_{r_i}$ are special groups and the product of special groups is special. \end{proof} \section{Parabolic bundles}\label{sect:ParBundles} \subsection{Definitions and notations} Recall that $\mathbf{k}$ denotes a field of characteristic zero, $X$ stands for a smooth projective geometrically connected curve over $\mathbf{k}$, and $D$ is a set of rational points of $X$. We often assume that $D\ne\varnothing$; in this case $X$ has a divisor of degree one defined over $\mathbf{k}$. We will often have to consider sequences indexed by $D\times\mathbb{Z}_{>0}$ or by $D\times\mathbb{Z}_{\ge0}$. A typical notation will be $r_{\bullet,\bullet}$. If $x\in D$, then $r_{x,\bullet}$ stands for the sequence $r_{x,1}, r_{x,2}, \dots$ (or $r_{x,0}, r_{x,1}, \dots$). The monoid of all sequences $r_{\bullet,\bullet}$ indexed by $D\times\mathbb{Z}_{>0}$ with terms $r_{x,j}$ in a commutative monoid~$S$ and such that $r_{x,j}=0$ for $j\gg0$ (that is, functions on $D\times\mathbb{Z}_{>0}$ with finite support) will be denoted by $S[D\times\mathbb{Z}_{>0}]$. \begin{Definition} A \emph{parabolic bundle} of type $(X,D)$ is a collection $(E,E_{\bullet,\bullet})$, where $E$ is a vector bundle over~$X$ and $E_{x,\bullet}$ is a flag in $E_x$ for $x\in D$: \[ E_x=E_{x,0}\supseteq E_{x,1}\supseteq\dots\supseteq E_{x,l}\supseteq\cdots,\qquad E_{x,l}=0\quad \text{for} \ l\gg0. \] \end{Definition} We have the category $\mathcal{B}{un}^{\rm par}(X,D)$ of parabolic bundles. We sometimes denote a parabolic bundle by a single boldface letter: $\mathbf E=(E,E_{\bullet,\bullet})$. The morphism from $\mathbf E$ to $\mathbf E'$ is a morphism $\phi\colon E\to E'$ such that for all $x\in D$ and $j\ge0$ we have $\phi(E_{x,j})\subset E'_{x,j}$. This category is an additive $\mathbf{k}$-linear category. The direct sum of $(E,E_{\bullet,\bullet})$ and $(E',E'_{\bullet,\bullet})$ is $(E\oplus E',E_{\bullet,\bullet}\oplus E'_{\bullet,\bullet})$. We note that the decomposition of a parabolic bundle into a direct sum of indecomposable parabolic bundles is unique up to isomorphism, while the isotypic summands are unique; the proof is similar to~\cite[Theorem~3]{Atiyah-KrullSchmidt} (see also~\cite[Proposition~3.1.2]{FedorovSoibelmans}). We often skip $X$ and $D$ from the notation, writing $\mathcal{B}{un}^{\rm par}$ instead of $\mathcal{B}{un}^{\rm par}(X,D)$. Abusing notation, we denote by $\mathcal{B}{un}^{\rm par}$ the stack of objects of the category $\mathcal{B}{un}^{\rm par}$. Precisely, if $S$ is a $\mathbf{k}$-scheme, then $\mathcal{B}{un}^{\rm par}(S)$ is the groupoid of collections $(E,E_{\bullet,\bullet})$, where $E$ is a vector bundle over $S\times_\mathbf{k} X$, $E_{x,\bullet}$ is a filtration by vector subbundles of the restriction of $E$ to $S\times_\mathbf{k} x$ for $x\in D$. Here by a subbundle of $E|_{S\times_\mathbf{k} x}$ we mean a subsheaf that splits off as a direct summand Zariski locally over $S$. \subsection[Monoids $\Gamma_+$ and $\Gamma'_+$]{Monoids $\boldsymbol{\Gamma_+}$ and $\boldsymbol{\Gamma'_+}$}\label{sect:Gamma} Consider the free abelian group $\mathbb{Z}\times\mathbb{Z}[D\times\mathbb{Z}_{>0}]\times\mathbb{Z}$ and its subgroup $\Gamma$ consisting of $(r,r_{\bullet,\bullet},d)$ such that for all $x\in D$ we have $\sum_{j=1}^{\infty}r_{x,j}=r$. Let $\Gamma_+\subset\Gamma$ be the monoid of sequences $(r,r_{\bullet,\bullet},d)$ such that \begin{enumerate}\itemsep=0pt \item[(i)] $r\ge0$ and for all $x\in D$ and $j>0$ we have $r_{x,j}\ge0$; \item[(ii)] if $r=0$, then $d=0$. \end{enumerate} Note that it follows from these conditions that $r=0$ implies that $(r,r_{\bullet,\bullet},d)$ is the zero sequence; we denote it by 0. Define the \emph{class function}: \[ \cl\colon \ \mathcal{B}{un}^{\rm par}\to\Gamma_+,\qquad(E,E_{\bullet,\bullet})\mapsto(\rk E,\dim E_{x,j-1}-\dim E_{x,j},\deg E). \] For $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma$ we set $\rk\gamma:=r$. For a parabolic bundle $\mathbf E$ we set $\rk\mathbf E:=\rk\cl(\mathbf E)$. For $\gamma\in\Gamma_+$, we denote by $\mathcal{B}{un}^{\rm par}_\gamma$ the stack of objects of class $\gamma$; this is an open and closed substack of $\mathcal{B}{un}^{\rm par}$. It is often convenient to think of $\mathcal{B}{un}^{\rm par}=\bigsqcup\limits_{\gamma\in\Gamma_+}\mathcal{B}{un}_\gamma^{\rm par}$ as a $\Gamma_+$-graded stack. Note that $\mathcal{B}{un}^{\rm par}_0$ has a single object: the parabolic bundle of rank zero. Let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$. The projection $\mathcal{B}{un}_\gamma^{\rm par}(X,D)\to\mathcal{B}{un}_{r,d}(X)$ to the stack of rank~$r$ degree $d$ vector bundles on $X$ is schematic and of finite type (in fact, projective). Thus $\mathcal{B}{un}_\gamma^{\rm par}(X,D)$ and $\mathcal{B}{un}^{\rm par}(X,D)$ are Artin stacks locally of finite type. Let us call a vector bundle on $X$ \emph{nonpositive} if it does not have a subbundle of positive degree. Recall that in~\cite[Section~3.2]{FedorovSoibelmans}) we called a vector bundle on $X$ HN-nonnegative, if its Harder--Narasimhan spectrum is nonnegative. by~\cite[Lemma~3.2.1(i)]{FedorovSoibelmans} there is an open substack $\mathcal{B}{un}^+(X)\subset\mathcal{B}{un}(X)$ classifying HN-nonnegative vector bundles. Moreover, by~\cite[Lemma~3.2.1(iii)]{FedorovSoibelmans} the substack of $\mathcal{B}{un}^+(X)$ corresponding to vector bundles of rank~$r$ and degree~$d$ is of finite type (this substack was denoted by $\mathcal{B}{un}_{r,d}^{\ge0}(X)$ in loc.~cit.). By~\cite[Lemma~3.2.1(ii)]{FedorovSoibelmans} a bundle $E$ is HN-nonnegative if and only if it has no quotient bundles of negative degree. Thus a vector bundle $E$ is nonpositive if and only if its dual $E^\vee$ is HN-nonnegative. The following lemma is now clear. \begin{Lemma}\label{lm:PosNeg} There is an open substack $\mathcal{B}{un}^-(X)$ of $\mathcal{B}{un}(X)$ classifying nonpositive vector bundles. The assignment $E\mapsto E^\vee$ is an isomorphism between $\mathcal{B}{un}^-(X)$ and $\mathcal{B}{un}^+(X)$. The component of $\mathcal{B}{un}^-(X)$ corresponding to vector bundles of fixed degree and rank is of finite type. \end{Lemma} Set \[ \mathcal{B}{un}^{{\rm par},-}=\mathcal{B}{un}^{{\rm par},-}(X,D):=\mathcal{B}{un}^{\rm par}(X,D)\times_{\mathcal{B}{un}(X)}\mathcal{B}{un}^-(X). \] In other words, $\mathcal{B}{un}^{{\rm par},-}$ is the stack (and the category) of parabolic bundles on $X$ whose underlying vector bundle is nonpositive. Set also \[ \mathcal{B}{un}^{{\rm par},-}_\gamma=\mathcal{B}{un}^{{\rm par},-}_\gamma(X,D):=\mathcal{B}{un}^{\rm par}_\gamma(X,D)\times_{\mathcal{B}{un}(X)}\mathcal{B}{un}^-(X). \] \begin{Lemma}\label{lm:ParGamma} For all $\gamma\in\Gamma_+$ the stack $\mathcal{B}{un}^{{\rm par},-}_\gamma(X,D)$ is an Artin stack of finite type. \end{Lemma} \begin{proof} Follows from Lemma~\ref{lm:PosNeg}. \end{proof} Let $\Gamma'_+$ be the submonoid of $\Gamma_+$ consisting of sequences with $d\le0$. Clearly, $\mathcal{B}{un}_\gamma^{{\rm par},-}\ne\varnothing$ only if $\gamma\in\Gamma'_+$. \subsection[Categories over $\mathcal{B}{un}^{\rm par}$ and $\Gamma_+$-graded stacks]{Categories over $\boldsymbol{\mathcal{B}{un}^{\rm par}}$ and $\boldsymbol{\Gamma_+}$-graded stacks}\label{sect:CatsOverPar} We will consider below many categories with a forgetful functor to $\mathcal{B}{un}^{\rm par}$ (e.g., the category of parabolic Higgs bundles). Let $\mathcal C$ be such a category and denote by $\mathcal C$ its stack of objects as well. Assume that the morphism $\mathcal C\to\mathcal{B}{un}^{\rm par}$ is of finite type. Define the stacks \begin{gather}\label{eq:StacksOverPar} \mathcal C_\gamma:=\mathcal C\times_{\mathcal{B}{un}^{\rm par}}\mathcal{B}{un}^{\rm par}_\gamma,\qquad\!\! \mathcal C^-:=\mathcal C\times_{\mathcal{B}{un}^{\rm par}}\mathcal{B}{un}^{{\rm par},-},\qquad\!\! \mathcal C_\gamma^-:=\mathcal C\times_{\mathcal{B}{un}^{\rm par}}\mathcal{B}{un}^{{\rm par},-}_\gamma.\!\! \end{gather} By Lemma~\ref{lm:ParGamma}, the stack $\mathcal C_\gamma^-$ is of finite type. The stack $\mathcal C=\bigsqcup\limits_{\gamma\in\Gamma_+}\mathcal C_\gamma$ is $\Gamma_+$-graded, while the stack $\mathcal C^-=\bigsqcup\limits_{\gamma\in\Gamma_+}\mathcal C^-_\gamma$ is $\Gamma'_+$-graded. Moreover, the stack $\mathcal C^-$ is of finite type as a graded stack, that is, its graded components are stacks of finite type. The group ring $\Mot(\mathbf{k})[\Gamma_+]$ is closely related to \emph{quantum tori} (cf.~\cite{KontsevichSoibelman08}). This has a natural basis $e_\gamma$, where~$\gamma$ ranges over $\Gamma_+$; the multiplication is given by $e_\gamma e_{\gamma'}=e_{\gamma+\gamma'}$. The reason for the multiplication in the quantum torus to be commutative is that we are actually working with 2-dimensional Calabi--Yau categories, and hence the skewsymmetrization of the Euler form vanishes (cf.~\cite{RenSoibelman}). Let $\Mot(\mathbf{k})[[\Gamma_+]]$ be the completion of $\Mot(\mathbf{k})[\Gamma_+]$ (this can be viewed as the group of $\Mot(\mathbf{k})$-valued functions on $\Gamma_+$). If $\mathcal D$ is a $\Gamma_+$-graded stack of finite type, we consider the generating series \begin{equation}\label{eq:MotDT} [\mathcal D]:=\sum_{\gamma\in\Gamma_+}[\mathcal D_\gamma]e_\gamma\in\Mot(\mathbf{k})[[\Gamma_+]]. \end{equation} We call $[\mathcal D]$ the \emph{graded motivic class} of the stack $\mathcal D$. Recall that in~\cite{KontsevichSoibelman08} the motivic Donaldson--Thomas series was defined as an element of the completed motivic quantum torus. In our case, it associates to a $\Gamma_+$-graded stack $\mathcal D$ an infinite series in the {\it commutative} motivic quantum torus corresponding to the monoid $\Gamma_+$ endowed with the trivial bilinear form. Thus, it coincides with~$[\mathcal D]$. Sometimes it is convenient to write $e_\gamma$ explicitly as \[ e_\gamma=w^r\prod_{x\in D}\prod_{j=1}^{\infty}w_{x,j}^{r_{x,j}}z^d\in\mathbb{Z}\big[\big[w,w_{\bullet,\bullet},z,z^{-1}\big]\big]\subset\Mot(\mathbf{k}) \big[\big[w,w_{\bullet,\bullet},z,z^{-1}\big]\big], \] where $\gamma=(r,r_{\bullet,\bullet},d)$. Here $w=w_{\bullet,\bullet}$ stands for the collection of variables \[ w_{x,\bullet}:=(w_{x,1},w_{x,2},\dots,w_{x,j},\dots),\qquad x\in D. \] The variables $w$, $z$, and $w_{x,j}$ for $x\in D$, $j\ge1$ are commuting variables. \begin{Remark} Note that we do not fix the lengths of flags. Let us fix a function $l\colon D\to\mathbb{Z}_{>0}$ and consider only flags of length at most $l(x)$ at $x$. Let $\Gamma_{+,l}$ be the submonoid of $\Gamma_+$ consisting of sequences $(r,r_{\bullet,\bullet},d)$ such that $r_{x,j}=0$ whenever $j>l(x)$. We have the obvious projection $\Gamma_+\to\Gamma_{+,l}$, which induces a homomorphism $\Pi_l \colon \Mot(\mathbf{k})[[\Gamma_+]]\to\Mot(\mathbf{k})[[\Gamma_{+,l}]]$. Explicitly, this is just setting $w_{x,j}=0$ whenever $j>l(x)$. Let $\mathcal D$ be as in~\eqref{eq:MotDT}, then $\Pi_l[\mathcal D]=\sum\limits_{\gamma\in\Gamma_{+,l}}[\mathcal D_\gamma]e_\gamma\in\Mot(\mathbf{k})[[\Gamma_{+,l}]]$ is the graded motivic class of the substack of~$\mathcal D$ where the lengths of the flags are bounded by $l$. We see that the difference between fixing the length of flags and allowing flags of arbitrary lengths corresponds on the quantum torus side to the difference between polynomials in infinite number of variables and polynomials in finite number of variables. In our applications, $[\mathcal D]$ will be symmetric in each sequence of variables $w_{x,\bullet}$ (cf.~Remark~\ref{rm:Weyl}). In this case, the difference between fixing and not fixing the lengths corresponds on the side of motivic classes to the difference between symmetric polynomials and symmetric functions, cf.~\cite[Chapter~1, Section~2]{macdonald1998symmetric}. \end{Remark} We emphasize that $\Mot(\mathbf{k})[[\Gamma_+]]$ is not a ring. However, $\Mot(\mathbf{k})[[\Gamma'_+]]\subset\Mot(\mathbf{k})[[w,w_{\bullet,\bullet},z^{-1}]]$ is a ring and $[\mathcal C^-]\in\Mot(\mathbf{k})[[\Gamma'_+]]$ whenever $\mathcal C$ is a stack of finite type over $\mathcal{B}{un}^{\rm par}$. This is in accordance to the general theory in~\cite{KontsevichSoibelman08}, where one fixes a strict sector in $\mathbb{R}^2$ in order to have well-defined Donaldson--Thomas invariants. We can also replace $\Mot(\mathbf{k})$ with $\cMot(\mathbf{k})$ in all the above constructions. \subsection{Motivic zeta-functions and plethystic operations}\label{sect:Plethystic} Following~\cite{KapranovMotivic}, for a variety $Y$ define its \emph{motivic zeta-funcion} by \begin{equation}\label{eq:MotZeta} \zeta_Y(z):=\sum_{n=0}^\infty\big[Y^{(n)}\big]z^n\in\Mot(\mathbf{k})[[z]], \end{equation} where $Y^{(n)}=Y^n/\Sigma_n$ is the $n$-th symmetric power of $Y$ ($\Sigma_n$ denotes the group of permutations). Consider the group $(1+z\Mot(\mathbf{k})[[z]])^\times$, where the group operation is multiplication. According to~\cite[Theorem~2.3]{EkedahlLambdaStacks} $\zeta$ can be uniquely extended to a homomorphism \[ \zeta\colon \ \Mot(\mathbf{k})\to(1+z\Mot(\mathbf{k})[[z]])^\times\colon \ A\mapsto\zeta_A(z) \] such that we have $\zeta_{\mathbb{L}^n A}(z)=\zeta_A\big(\mathbb{L}^nz\big)$ for all $n\in\mathbb{Z}$ and $A\in\Mot(\mathbf{k})$. Clearly, \[ \zeta_A(z)\equiv1+Ax\pmod {z^2}.\] Thus we have equipped $\Mot(\mathbf{k})$ with a $\lambda$-ring structure. Note that $\Mot(\mathbf{k})$ is \emph{not} a~special $\lambda$-ring, in particular, $\zeta(AB)$ cannot be expressed in terms of $\zeta(A)$ and $\zeta(B)$ (so some authors would call this a pre-$\lambda$-ring structure). According to loc.~cit., this homomorphism $\zeta$ is continuous with respect to the dimensional filtration on $\Mot(\mathbf{k})$, so it extends to a homomorphism \begin{equation*} \zeta\colon \ \cMot(\mathbf{k})\to\big(1+z\cMot(\mathbf{k})[[z]]\big)^\times, \end{equation*} which coincides with the one constructed in~\cite[Section~1.3.1]{FedorovSoibelmans}. Let $\Mot(\mathbf{k})[[\Gamma_+']]^0$ stand for the series without constant term. We define the \emph{plethystic exponent} $\Exp\colon \Mot(\mathbf{k})[[\Gamma_+']]^0\to(1+\Mot(\mathbf{k})[[\Gamma_+']]^0)^\times$ by \[ \Exp\left(\sum_{\gamma\in\Gamma_+'} A_\gamma e_\gamma\right)=\prod_{\gamma\in\Gamma_+'}\Exp(A_\gamma e_\gamma)= \prod_{\gamma\in\Gamma_+'}\zeta_{A_\gamma}(e_\gamma). \] One shows easily that this is an isomorphism of abelian groups. Denote the inverse isomorphism by $\Log$ (the \emph{plethystic logarithm}). Finally, we define the \emph{plethystic power} by \[ \Pow\colon \ \big(1+\Mot(\mathbf{k})[[\Gamma_+']]^0\big)\times\Mot(\mathbf{k})\to1+\Mot(\mathbf{k})[[\Gamma_+']]^0\colon \ (f,A)\mapsto\Exp(A\Log(f)). \] We note that we can similarly define $\Exp$, $\Log$, and $\Pow$ for the completed ring $\cMot(\mathbf{k})$, which coincide with the operations defined in~\cite{FedorovSoibelmans} when $D=\varnothing$. \subsection{Parabolic subbundles and quotient bundles}\label{sect:Subobjects} Let $\mathbf E=(E,E_{\bullet,\bullet})$ be a parabolic bundle of type $(X,D)$. We say that $\mathbf E'=(E',E'_{\bullet,\bullet})$ is a \emph{strict} parabolic subbundle of $\mathbf E$ if $E'$ is a saturated subbundle of $E$ (that is, $E/E'$ is torsion free) and for all $x$ and $j$ we have $E'_{x,j}=E_{x,j}\cap E'_x$. Note that strict parabolic subbundles of $\mathbf E=(E,E_{\bullet,\bullet})$ are in bijective correspondence with saturated subbundles of $E$. Let $\mathbf E'=(E',E'_{\bullet,\bullet})$ be a strict parabolic subbundle of $\mathbf E=(E,E_{\bullet,\bullet})$; set $E'':=E/E'$. Then we have a parabolic structure on $E''$ given by $E''_{x,j}:=E_{x,j}/E'_{x,j}$. We call the parabolic bundle $\mathbf E/\mathbf E':=(E'',E''_{\bullet,\bullet})$ the \emph{quotient parabolic bundle} of $\mathbf E$. Thus, the quotient parabolic bundles of $\mathbf E$ are also in bijective correspondence with saturated subbundles of $E$. Finally, in the above situation, we say that \begin{equation}\label{eq:ExactSeq} 0\to\mathbf E'\to\mathbf E\to\mathbf E/\mathbf E'\to0 \end{equation} is a \emph{short exact sequence}. We also say that \emph{$\mathbf E$ is an extension of $\mathbf E/\mathbf E'$ by $\mathbf E'$.} It is clear that in this case we have $\cl(\mathbf E)=\cl(\mathbf E')+\cl(\mathbf E/\mathbf E')$. One can use the short exact sequences above to define the group $K_0(\mathcal{B}{un}^{\rm par})$; the class function $\cl$ extends to $\cl\colon K_0(\mathcal{B}{un}^{\rm par})\to\Gamma$. \begin{Remark}The category $\mathcal{B}{un}^{\rm par}$ is not abelian. It can be extended to an abelian category by viewing vector bundles with flags as coherent sheaves on orbifold curves. Then the abelian category is the category of coherent sheaves on this orbifold. This extension will not be used in the current paper. On the other hand, if we define short exact sequences in $\mathcal{B}{un}^{\rm par}$ as sequences isomorphic to some sequence of the form~\eqref{eq:ExactSeq}, then $\mathcal{B}{un}^{\rm par}$ becomes an exact category in the sense of Quillen. \end{Remark} Let $\phi\colon E\to F$ be a morphism of vector bundles on $X$. We say that $\phi$ is \emph{generically an isomorphism} if it is an isomorphism at the generic point of $X$. Equivalently, $\phi$ is an isomorphism over a non-empty Zariski open subset of $X$. Another reformulation is that $\phi$ is injective and~$F/\phi(E)$ is a torsion sheaf. Sometimes one says in this situation that $E$ is a lower modification of~$F$. \begin{Definition}We say that a morphism of parabolic bundles $\phi\colon (E,E_{\bullet,\bullet})\to(F,F_{\bullet,\bullet})$ is \emph{generically an isomorphism} if the underlying morphism $E\to F$ is generically an isomorphism. \end{Definition} \begin{Lemma}\label{lm:MorPar} Let $\phi\colon \mathbf E\to\mathbf F$ be a morphism of parabolic bundles. Then there are strict parabolic subbundles $\mathbf E'\subset\mathbf E$ and $\mathbf F'\subset\mathbf F$ such that $\phi$ can be decomposed as \[ \mathbf E\xrightarrow{\phi_1}\mathbf E/\mathbf E'\xrightarrow{\phi_2}\mathbf F'\xrightarrow{\phi_3}\mathbf F, \] where $\phi_1$ is the canonical projection, $\phi_2$ is generically an isomorphism, $\phi_3$ is the canonical embedding. \end{Lemma} \begin{proof} Write $\mathbf E=(E,E_{\bullet,\bullet})$, $\mathbf F=(F,F_{\bullet,\bullet})$, let $\phi'\colon E\to F$ be the underlying morphism of vector bundles. Note that $\Ker\phi'$ is a vector subbundle of $E$. Indeed, $E/\Ker\phi'$ is isomorphic to a subsheaf of $F$, so it is torsion free. Let $\mathbf E'$ be the strict subbundle of $\mathbf E$ whose underlying vector bundle is $\Ker(\phi')$. Let $F'$ be the saturation of the image of $\phi'$ (that is, $F'$ is the unique saturated vector subbundle of $F$ containing $\phi'(E)$ such that the quotient $F'/\phi'(E)$ is a torsion sheaf). Let $\mathbf F'$ be the strict subbundle of $\mathbf F$ whose underlying vector bundle is $F'$. Now the existence of the decomposition is clear. \end{proof} \subsection{Generalized degrees and slopes}\label{sect:DegreeSlope} Let $A$ be a $\mathbb{Q}$-vector space (in applications it will be $\mathbf{k}$ or $\mathbb{R}$). Let $\kappa\in A$, $\zeta=\zeta_{\bullet,\bullet}\in A[D\times\mathbb{Z}_{>0}]$. Then we define the homomorphism $\deg_{\kappa,\sigma}\colon \Gamma\to A$ by \[ \deg_{\kappa,\zeta}(r,r_{\bullet,\bullet},d)=\kappa d+\sum_{x\in D}\sum_{j>0}\zeta_{x,j}r_{x,j}. \] If $\rk\gamma\ne0$, we define the $(\kappa,\zeta)$-slope of $\gamma$ by $\deg_{\kappa,\zeta}\gamma/\rk\gamma$. We write $\deg_{\kappa,\zeta}\mathbf E$ for $\deg_{\kappa,\zeta}\cl(\mathbf E)$ and call $\deg_{\kappa,\zeta}\mathbf E/\rk\mathbf E$ \emph{the $(\kappa,\sigma)$-slope of $\mathbf E$}. We say that a parabolic bundle $\mathbf E$ is \emph{$(\kappa,\zeta)$-isoslopy} if the $(\kappa,\zeta)$-slope of any direct summand of $\mathbf E$ is equal to the $(\kappa,\zeta)$-slope of $\mathbf E$. We remark that it is common to write $\zeta\star\gamma$ for $\deg_{0,\zeta}\gamma$ and $\deg_\zeta\gamma$ for $\deg_{1,\zeta}\gamma$ but we prefer a uniform notation. We also remark that for an exact sequence~\eqref{eq:ExactSeq} we have $\deg_{\kappa,\zeta}\mathbf E=\deg_{\kappa,\zeta}\mathbf E'+\deg_{\kappa,\zeta}(\mathbf E/\mathbf E')$. \subsection{Parabolic weights and stability conditions}\label{sect:ParWeights} The following definition should be compared to~\cite[Definition~6.9]{MellitPunctures}. \begin{Definition}\label{def:StabilityCond} We say that a sequence $\sigma=\sigma_{\bullet,\bullet}$ of real numbers indexed by $D\times\mathbb{Z}_{>0}$ is a~\emph{sequence of parabolic weights} if for all $x\in D$ we have \begin{equation}\label{eq:StabCond2} \sigma_{x,1}\le\sigma_{x,2}\le\cdots \end{equation} and for all $x$ and $j$ we have $\sigma_{x,j}\le\sigma_{x,1}+1$. \end{Definition} To every sequence of parabolic weights we will associate a notion of stability on parabolic bundles in Definition~\ref{def:Semistable} below. Thus we denote the set of all sequences of parabolic weights by $\Stab=\Stab(X,D)$. Fix $\sigma\in\Stab$. \begin{Definition}\label{def:Semistable} A parabolic bundle $\mathbf E$ is \emph{$\sigma$-semistable} if for all strict parabolic subbundles $\mathbf E'\subset\mathbf E$ we have \begin{equation}\label{eq:ss} \frac{\deg_{1,\sigma}\mathbf E'}{\rk\mathbf E'}\le\frac{\deg_{1,\sigma}\mathbf E}{\rk\mathbf E}. \end{equation} \end{Definition} \begin{Remark}\label{rm:kappa} We can similarly define semistability for any $\kappa>0$ replacing the condition $\sigma_{x,j}\le\sigma_{x,1}+1$ with $\sigma_{x,j}\le\sigma_{x,1}+\kappa$. However, scaling $\kappa$ and $\sigma$ by the same positive real number scales all the slopes by the same number, so we would get the same notion of semistability. This is why we restrict to the case $\kappa=1$ above. The case $\kappa=0$ would not yield stacks of finite type; however, the possibility of taking $\kappa=0$ will be useful below, when we work with connections (see Section~\ref{sect:StabConn}). \end{Remark} \begin{Proposition}\label{pr:HN} Let $\mathbf E\in\mathcal{B}{un}^{\rm par}$ be a parabolic bundle. Then there is a unique filtration $0=\mathbf E_0\subset\mathbf E_1\subset\dots\subset\mathbf E_m=\mathbf E$ by strict parabolic subbundles such that all the quotients $\mathbf E_i/\mathbf E_{i-1}$ are $\sigma$-semistable and we have $\tau_1>\dots>\tau_m$, where $\tau_i$ is the $(1,\sigma)$-slope of $\mathbf E_i/\mathbf E_{i-1}$. \end{Proposition} \begin{proof} We start with a Lemma. \begin{Lemma}\label{lm:ModifDegree} Let the morphism $\mathbf E\to\mathbf F$ be generically an isomorphism. Then $\deg_{1,\sigma}\mathbf E\le\deg_{1,\sigma}\mathbf F$. \end{Lemma} \begin{proof} Write $\mathbf E=(E,E_{\bullet,\bullet})$ and $\mathbf F=(F,F_{\bullet,\bullet})$. Let $\phi\colon E\to F$ be the underlying morphism of vector bundles. For $x\in D$ let $d_x$ denote the dimension of the kernel of $\phi_x$. Then \[ \deg E=\deg F-\length(F/\phi(E))\le\deg F-\sum_{x\in D}d_x. \] On the other hand, for all $x\in D$ and $i>0$ we have $\dim E_{x,i}\le\dim F_{x,i}+d_x$. Hence \begin{align*} \deg_{1,\sigma}\mathbf E& =\deg E+\sum_{x,j>0}\sigma_{x,j}(\dim E_{x,j-1}-\dim E_{x,j})\\ & =\deg E+\sum_{x\in D}\left(\sigma_{x,1}\rk E+\sum_{i>0}(\sigma_{x,i+1}-\sigma_{x,i})\dim E_{x,i}\right)\\ & \le \deg F-\sum_{x\in D} d_x+\sum_{x\in D}\left(\sigma_{x,1}\rk F+\sum_{i>0}(\sigma_{x,i+1}-\sigma_{x,i})(\dim F_{x,i}+d_x)\right)\\ & = \deg_{1,\sigma}\mathbf F+\sum_{x\in D}d_x\left(-1+\sum_{i>0}(\sigma_{x,i+1}-\sigma_{x,i})\right)\le\deg_{1,\sigma}\mathbf F. \end{align*} Lemma~\ref{lm:ModifDegree} is proved. \end{proof} The rest of the proof of Proposition~\ref{pr:HN} is completely analogous to the proof of~\cite[Section~1.3]{HarderNarasimhan} in view of Lemma~\ref{lm:MorPar}. \end{proof} In the situation of Proposition~\ref{pr:HN}(i) we say that the filtration is the \emph{Harder--Narasimhan filtration} of $E$ (or \emph{HN-filtration} for short) and $\tau_1>\dots>\tau_m$ is the $\sigma$-HN spectrum of $\mathbf E$. We define $\mathcal{B}{un}^{{\rm par},\le\tau}$ and $\mathcal{B}{un}^{{\rm par},\ge\tau}$ as full subcategories of $\mathcal{B}{un}^{\rm par}$ whose objects are parabolic bundles with the $\sigma$-HN spectrum contained in $(-\infty,\tau]$ (resp.~$[\tau,\infty)$). We emphasize that the categories $\mathcal{B}{un}^{{\rm par},\le0}$ and $\mathcal{B}{un}^{{\rm par},-}$ should not be confused with each other: they coincide only if $\sigma=0$. The following lemma is standard. \begin{Lemma}\label{lm:NoMorphismSS} Let $\mathbf E$ be an object of $\mathcal{B}{un}^{{\rm par},\le\tau}$ and $\mathbf E'$ be an object of $\mathcal{B}{un}^{{\rm par},\ge\tau'}$, where $\tau<\tau'$. Then $\Hom_{\mathcal{B}{un}^{\rm par}}(\mathbf E',\mathbf E)=0$. \end{Lemma} \section{Parabolic pairs}\label{sect:ParPairs} \subsection{Parabolic pairs and their generic Jordan types} The notion of parabolic pair, interesting by itself, will be used as a technical tool for studying parabolic Higgs bundles in Section~\ref{sect:HiggsnEigenval} and parabolic bundles with connections in Section~\ref{sect:Conn}. Our main results in this section are Theorem~\ref{th:MotMellitPunctures} and Corollary~\ref{cor:Pairs}. They give explicit answers for graded motivic classes of stacks of nilpotent parabolic pairs and parabolic pairs respectively. We will also give a simplified answer in the case $X=\P^1$ in Section~\ref{sect:P1ManyPts}. \begin{Definition} A \emph{parabolic pair} $(\mathbf E,\Psi)$ consists of a parabolic bundle \[ \mathbf E=(E,E_{\bullet,\bullet})\in\mathcal{B}{un}^{\rm par}(X,D) \] and an endomorphism $\Psi$ of $\mathbf E$ (that is, an endomorphism of $E$ preserving each $E_{x,j}$). If $\Psi$ is nilpotent we will speak about \emph{nilpotent parabolic pairs}. \end{Definition} Parabolic pairs as well as nilpotent parabolic pairs form an additive $\mathbf{k}$-linear category denoted $\mathcal{P}{air}=\mathcal{P}{air}(X,D)$ (resp.~$\mathcal{P}{air}^{\rm nilp}=\mathcal{P}{air}^{\rm nilp}(X,D)$). Again, we abuse notation by denoting the stacks of objects by the same symbols. We define $\mathcal{P}{air}_\gamma$, $\mathcal{P}{air}_\gamma^{\rm nilp}$, $\mathcal{P}{air}_\gamma^-$, $\mathcal{P}{air}_\gamma^{{\rm nilp},-}$ etc.\ following the general construction~\eqref{eq:StacksOverPar} of Section~\ref{sect:CatsOverPar}. The forgetful morphisms $\mathcal{P}{air}\to\mathcal{B}{un}^{\rm par}$ and $\mathcal{P}{air}^{\rm nilp}\to\mathcal{B}{un}^{\rm par}$ are schematic and of finite type. In particular, $\mathcal{P}{air}^-$ and $\mathcal{P}{air}^{{\rm nilp},-}$ are $\Gamma'_+$-graded Artin stacks of finite type (in the graded sense). Let $K\supset\mathbf{k}$ be an extension and $(E,E_{\bullet,\bullet},\Psi)\in\mathcal{P}{air}^{\rm nilp}(K)$. If we trivialize $E$ at the generic point of $X_K=X\times_\mathbf{k}\Spec K$, $\Psi$ becomes a $\rk E\times\rk E$ nilpotent matrix. Its Jordan type is a partition $\lambda$ of $\rk E$. Thus we get a locally closed stratification of $\mathcal{P}{air}^{\rm nilp}$ according to the generic Jordan type of the nilpotent endomorphism \[ \mathcal{P}{air}^{\rm nilp}(X,D)=\bigsqcup_\lambda\mathcal{P}{air}^{\rm nilp}(X,D,\lambda), \] where the disjoint union is over all partitions. In other words, $\mathcal{P}{air}^{\rm nilp}(X,D,\lambda)$ classifies nilpotent parabolic pairs such that the endomorphism is generically conjugate to $n_\lambda$ (that is, conjugate to $n_\lambda$ at the generic point of $X$, or, equivalently, at each point of a non-empty Zariski open subset of $X$). We remark that any endomorphism generically conjugate to $n_\lambda$ is necessarily nilpotent. Again, we define the $\Gamma'_+$-graded stacks $\mathcal{P}{air}^{{\rm nilp},-}(X,D,\lambda)$ using the general formalism~\eqref{eq:StacksOverPar} of Section~\ref{sect:CatsOverPar}. \subsection{Motivic classes of parabolic bundles with nilpotent endomorphisms}\label{sect:NilpEnd} Our goal in this section is to calculate the graded motivic class (that is, the motivic Donaldson--Thomas series, cf.~\cite{KontsevichSoibelman08}) \begin{align} \big[\mathcal{P}{air}^{{\rm nilp},-}(X,D,\lambda)\big]& =\sum_{\gamma\in\Gamma'_+}\big[\mathcal{P}{air}_\gamma^{{\rm nilp},-}(X,D,\lambda)\big]e_\gamma\nonumber\\ & = w^{|\lambda|}\sum_{\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma'_+}\big[\mathcal{P}{air}_\gamma^{{\rm nilp},-}(X,D,\lambda)\big]\prod_{x,j}w_{x,j}^{r_{x,j}}z^d.\label{eq:Omega} \end{align} The partition $\lambda$ is fixed until the end of Section~\ref{sect:Factorization}. This graded motivic class is calculated as follows. First, in Section~\ref{sect:Factorization} we write this graded motivic class as the product of a term that is independent of the parabolic structures, and the ``local'' terms independent of the curve. The first term has been calculated in~\cite{FedorovSoibelmans}. Since the local terms are independent of the curve, it is enough to calculate them when $X=\P^1$. More precisely, we will work with $\P^1$ and two points with parabolic structures (that is, $D=\{0,\infty\}$) but we will calculate the sum over all partitions (Section~\ref{sect:P1}). This part is very similar to~\cite[Section~5.4]{MellitPunctures}. In Section~\ref{sect:MotEnd} we give the explicit answer for the graded motivic classes under consideration. Using the formalism of plethystic powers we then easily calculate the class of parabolic bundles with not necessarily nilpotent endomorphisms. \begin{Remark}\label{rm:Weyl} We will see in Theorem~\ref{th:MotMellitPunctures} that~\eqref{eq:Omega} is a symmetric function in $w_{x,\bullet}$ for each $x\in D$. This can be explained as follows. Note that the ``Weyl group'' $W:=\prod\limits_{x\in D}\Sigma_{\infty}$ acts on~$\Gamma_+$ and~$\Gamma'_+$ in the obvious way (here $\Sigma_\infty$ is the inductive limit of the permutation groups $\Sigma_l$). Using the commutativity of the motivic Hall algebra of the category of representations of the Jordan quiver (the quiver with one vertex and one loop), one can easily show that the motivic classes $\big[\mathcal{P}{air}_\gamma^{{\rm nilp},-}(X,D,\lambda)\big]$ are $W$-invariant. Thus, we can re-write~\eqref{eq:Omega} as \[ w^{|\lambda|}\mathop{\sum\nolimits'}\limits_{\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma'_+} \big[\mathcal{P}{air}_\gamma^{{\rm nilp},-}(X,D,\lambda)\big]\prod_{x\in D} m_{r_{x,\bullet}}(w_{x,\bullet})z^d, \] where the summation is only over $r_{\bullet,\bullet}$ such that we have $r_{x,j}\ge r_{x,j+1}$ for all $x$ and $j$ (that is, $r_{x,\bullet}$ is a partition of~$|\lambda|$). Here, for a partition $\mu$, $m_\mu$ is the symmetric function equal to the sum of all monomials whose ordered list of exponents is~$\mu$. \end{Remark} \subsection[Factorization of graded motivic classes of stacks of nilpotent parabolic pairs]{Factorization of graded motivic classes of stacks\\ of nilpotent parabolic pairs}\label{sect:Factorization} In this section we factorize~\eqref{eq:Omega} as the product of the global part (depending only on $X$ but not on $D$) and the local parts corresponding to points of $D$ (but independent of $X$). This is a~motivic version of~\cite[Theorem~5.6]{MellitPunctures}. We follow the same ideas, though some parts of Mellit's proof do not work in the motivic case and must be replaced by different arguments. On the other hand, we were able to simplify some parts of Mellit's proof, in particular, by working with stacks. Note that $\mathcal{P}{air}^{{\rm nilp},-}(X,\varnothing,\lambda)$ classifies pairs $(E,\Psi)$, where~$E$ is a nonpositive vector bundle on~$X$, $\Psi$ is an endomorphism of $E$ generically conjugate to $n_\lambda$ but there are no parabolic structures. \begin{Lemma}\label{lm:invertible} The motivic class $\big[\mathcal{P}{air}_\gamma^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]\in w^{|\lambda|}\Mot(\mathbf{k})\big[\big[z^{-1}\big]\big]$ is invertible. \end{Lemma} \begin{proof} It is enough to show that the degree 0 part $\big[\mathcal{P}{air}_{|\lambda|,0}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]$ is invertible in $\Mot(\mathbf{k})$. Note that a nonpositive vector bundle of degree 0 on $\P^1$ is necessarily trivial. Thus, this degree zero part classifies pairs $(E,\Psi)$, where $E$ is a trivial vector bundle and $\Psi$ is a constant endomorphism conjugate to $n_\lambda$. It is easy to see that this stack is isomorphic to the classifying stack of the centralizer $Z_\lambda$ of $n_\lambda$. By Lemma~\ref{lm:Zspecial} $Z_\lambda$ is special. Now it follows from Lemma~\ref{lm:SpecialMot} or~\cite[Lemma~2.2.3]{FedorovSoibelmans} that $\big[\mathcal{P}{air}_{|\lambda|,0}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]=1/[Z_\lambda]$. \end{proof} We need some notation. Note that $\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\infty,\lambda\big)\big]\in w^{|\lambda|}\Mot(\mathbf{k})\big[\big[w_{\infty,\bullet},z^{-1}\big]\big]$. For $x\in D$ let \[ \big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\infty,\lambda\big)\big]_x\in w^{|\lambda|}\Mot(\mathbf{k})\big[\big[w_{x,\bullet},z^{-1}\big]\big] \] denote the result of replacing $w_{\infty,\bullet}$ by $w_{x,\bullet}$ in this series. \begin{Theorem}\label{th:Factorization} We have in $\Mot(\mathbf{k})[[\Gamma'_+]]$. \begin{equation* \big[\mathcal{P}{air}^{{\rm nilp},-}(X,D,\lambda)\big]=\big[\mathcal{P}{air}^{{\rm nilp},-}(X,\varnothing,\lambda)\big] \prod_{x\in D} \frac{\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\infty,\lambda\big)\big]_x} {\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]}. \end{equation*} \end{Theorem} The proof of the theorem occupies the rest of Section~\ref{sect:Factorization}; it is based on the local study of stacks in the formal neighborhood of~$D$. The main idea of the proof is very simple. Let us assume that $D=\{x\}$ is a single rational point of $X$. In Section~\ref{sect:LocStacks} we will define the stack $\mathcal{P}{air}^{{\rm loc},{\rm fl}}$ classifying triples $(F,\Phi,F_\bullet)$, where $F$ is a rank $|\lambda|$ vector bundle over the formal completion of $X$ at $x$, $\Phi$ is a nilpotent endomorphism of $F$ generically conjugate to $n_\lambda$, $F_\bullet$ is a flag in $F_x$ preserved by $\Phi(x)$. We have an obvious restriction morphism $\mathcal{P}{air}^{{\rm nilp},-}(X,x,\lambda)\to\mathcal{P}{air}^{{\rm loc},{\rm fl}}$. We will see in Lemma~\ref{lm:fiber} that this restriction morphism has constant fiber. Thus, one is tempted to write the graded motivic class $\big[\mathcal{P}{air}^{{\rm nilp},-}(X,x,\lambda)\big]$ as the product of the graded motivic class of this fiber and of $\mathcal{P}{air}^{{\rm loc},{\rm fl}}$. This would quickly lead to the proof of the theorem. Unfortunately, $\mathcal{P}{air}^{{\rm loc},{\rm fl}}$ is not an Artin stack as its points have inertia groups of infinite type, so its motivic class does not make sense. The major part of the proof consists of going around this problem. Let us give the overview of the proof. In Section~\ref{sect:jets} we define and study the schemes of jets into $\mathfrak{gl}_{|\lambda|}$. In Section~\ref{sect:LocStacks} we study the local stacks; they are essentially the quotients of the schemes of jets by the group of jets of ${\rm GL}_{|\lambda|}$. In Section~\ref{sect:StratThm} we re-write the theorem as a~statement about motivic classes of graded components. In Section~\ref{sect:Res} we study the fibers of the localization map; this is the main part of the proof. We complete the proof in Section~\ref{sect:ProofFact}. \subsubsection{Jets}\label{sect:jets} We will denote the non-archimedean local field $\mathbf{k}((t))$ by $\mathbb K$ and its ring of integers $\mathbf{k}[[t]]$ by $\mathbb O$. The order of pole at $t=0$ gives rise to a valuation map $\val\colon \mathbb K\to\mathbb{Z}\cup\{-\infty\}$, where $\val(0)=-\infty$. Clearly, $\val$ extends to $\mathfrak{gl}_{r,\mathbb K}$ as the maximum of valuations of all matrix elements. Let $J(X)$ denote the jet scheme of a scheme $X$ (this is a scheme of infinite type), and let $J_N(X)$ denote the scheme of order $N-1$ jets. In particular, $J_1(X)=X$. For an algebraic group $G$ of finite type over $\mathbf{k}$ we have the jet group $G_\mathbb O:=J(G)$ and the jet group of finite type $J_N(G)$. The $N$-th congruence subgroup $G^{(N)}$ is the kernel of the projection $G_\mathbb O\to J_N(G)$. We also have the ind-group of loops $G_\mathbb K$ containing $G_\mathbb O$. Let $\Delta:=\Spec\mathbb O$ be the formal disc and $\mathring\Delta:=\Spec\mathbb K$ be its generic point (the punctured formal disc). Also set $\Delta_N:=\Spec k[[t]]/t^N$. The groups ${\rm GL}_{r,\mathbb O}$, ${\rm GL}_{r,\mathbb K}$, and $J_N({\rm GL}_r)$ are the groups of automorphisms of the trivial vector bundles on $\Delta$, $\mathring\Delta$, and $\Delta_N$ respectively. For more details on the jet and loop groups we refer the reader to~\cite{SorgerLecturesBundles}. Set $r=|\lambda|$. Consider the orbit stratification of $\mathfrak{gl}_r$ under the adjoint action of ${\rm GL}_r$: $\mathfrak{gl}_r=\bigsqcup\limits_{\mu\vdash r}\mathcal O_\mu$, where $\mathcal O_\mu$ is the adjoint orbit containing $n_\mu$. Let $\overline\mathcal O_\mu$ denote the Zariski closure of $\mathcal O_\mu$. Set \[ J(\lambda):=J\big(\overline\mathcal O_\lambda\big)-\bigcup_{\mathcal O_\mu\subset\overline\mathcal O_\lambda-\mathcal O_\lambda}J\big(\overline\mathcal O_\mu\big). \] Note that $J(\lambda)$ parameterizes morphisms $\Delta\to\mathfrak{gl}_r$ such that the image of the generic point of~$\Delta$ is in $\mathcal O_\lambda$, that is, jets that are generically conjugate to~$n_\lambda$. \begin{Definition} We say that a loop $g\in {\rm GL}_{r,\mathbb K}(\mathbf{k})={\rm GL}_r(\mathbb K)$ is \emph{kernel-strict} if $g^{-1}n_\lambda g\in\mathfrak{gl}_{r,\mathbb O}$ and $g^{-1}$ induces an isomorphism between the $\mathbb O$-modules $\Ker n_\lambda\otimes_\mathbf{k}\mathbb O$ and $\Ker\big(g^{-1}n_\lambda g\big)$. \end{Definition} \begin{Remark} Note that our definition is a little different from~\cite[Definition~3.8]{MellitPunctures}. Mellit's definition of kernel-strictness depends also on a choice of a matrix $\theta$. In terminology of Mellit our $g$ is kernel-strict for $\theta=g^{-1}n_\lambda g$. Note also that the results of Mellit we are using here and below are formulated over finite fields but are valid over any field, proofs being the same. \end{Remark} Let $\Phi$ be a $\mathbf{k}$-point of $J(\lambda)$, then there is a kernel-strict $g\in {\rm GL}_{r,\mathbb K}(\mathbf{k})={\rm GL}_r(\mathbb K)$ such that $\Phi=g^{-1}n_\lambda g$. Set $\deg\Phi:=\val(\det g)$. The existence of such $g$ and independence of the degree on the choice of $g$ is proved in~\cite[Lemma~3.7]{MellitPunctures}. It follows also from loc.~cit.~that the degree is nonnegative. If $K\supset\mathbf{k}$ is a field extension, we similarly define the degree of a $K$-point of $J(\lambda)$. The degree is compatible with field extensions. Thus we get a stratification of the set of points of $J(\lambda)$: $|J(\lambda)|=\bigsqcup\limits_{d\ge0}J_d(\lambda)$. Let $\pi_N\colon J(\lambda)\to J_N\big(\overline\mathcal O_\lambda\big)$ be the truncation map. Set $J_{d,N}(\lambda):=\pi_N(J_d(\lambda))$. \begin{Proposition}\label{pr:jets} For a nonnegative integer $d$ and a partition $\lambda$ there is a positive integer $N(d,\lambda)$ such that for $N>N(d,\lambda)$ we have \begin{enumerate}\itemsep=0pt \item[$(i)$] For all $\Phi$ in $J_d(\lambda)$ there is a kernel-strict $g$ with $\val(g)<N/2$, $\val\big(g^{-1}\big)<N/2$ such that $g\Phi g^{-1}=n_\lambda$. \item[$(ii)$] $J_d(\lambda)=\pi_N^{-1}(J_{d,N}(\lambda))$. \item[$(iii)$] Any two points in the same fiber of the projection $J_d(\lambda)\to J_{d,N}(\lambda)$ are conjugate by an element of ${\rm GL}_{r,\mathbb O}$. \end{enumerate} \end{Proposition} \begin{proof} By~\cite[Lemma~3.7]{MellitPunctures} there is $N_0\ge1$ (depending on $\lambda$ and $d$) such that for all $\Phi\in J_d(\lambda)$ there is a kernel-strict $g$ with $\val(g)<N_0$ such that $g\Phi g^{-1}=n_\lambda$. Then $\val(\det g)=d$. Set $N_1:=rN_0+d$. Then by Cramer's rule $\val\big(g^{-1}\big)<N_1$. We also have $\val(g)<N_1$. Take $N(d,\lambda):=4N_1$. With this choice of $N(d,\lambda)$ part~(i) of the proposition is clear. Note that (ii) is saying that the degree of an infinite jet depends only on its $N$-th truncation if $N>4N_1$. Thus to prove (ii) we need to show that if $\Phi\in J_d(\lambda)$ and $\Phi'\in J(\lambda)$ is such that $\Phi'\equiv\Phi\pmod{z^{4N_1}}$, then $\Phi'\in J_d(\lambda)$. Choose a kernel-strict $g$ with $\val(g)<N_1$, $\val\big(g^{-1}\big)<N_1$ such that $g\Phi g^{-1}=n_\lambda$. Then $g\Phi'g^{-1}\equiv n_\lambda\pmod{z^{2N_1}}$. We need a lemma. \begin{Lemma}\label{lm:conjugate} There is $g'\in {\rm GL}_r^{(2N_1)}$ such that $g'g\Phi'g^{-1}(g')^{-1}=n_\lambda$. \end{Lemma} \begin{proof} Set $\Phi'':=g\Phi'g^{-1}$ and $N_2:=2N_1$. Write $\Phi''\equiv n_\lambda+\Phi_{N_2}z^{N_2}\pmod{z^{N_2+1}}$. Since $\Phi''\in J\big(\overline\mathcal O_\lambda\big)$, $\Phi_{N_2}$ belongs to the tangent space to $\overline\mathcal O_\lambda$ at $n_\lambda$, which is naturally identified with $[n_\lambda,\mathfrak{gl}_r]$. Thus there is $g_{_{N_2}}\in\mathfrak{gl}_r$ such that $[n_\lambda,g_{_{N_2}}]=\Phi_{N_2}$. Then $\big(1+g_{_{N_2}}z^{N_2}\big)\Phi''\big(1+g_{_{N_2}}z^{N_2}\big)^{-1}\equiv n_\lambda\pmod{z^{N_2+1}}$. Repeating this process we find $g_{_{N_2+1}},g_{_{N_2+2}},\ldots\in\mathfrak{gl}_r$ such that for $j>0$ we have \begin{gather*} \big(1+g_{_{N_2+j}}z^{N_2+j}\big)\cdots\big(1+g_{_{N_2}}z^{N_2}\big)\Phi'' \big(1+g_{_{N_2}}z^{N_2}\big)^{-1}\cdots\big(1+g_{_{N_2+j}}z^{N_2+j}\big)^{-1}\\ \qquad{} \equiv n_\lambda\pmod{z^{N_2+j+1}}. \end{gather*} It remains to take $g'=\prod\limits_{j=0}^\infty\big(1+g_{_{N_2+j}}z^{N_2+j}\big)$. Lemma~\ref{lm:conjugate} is proved. \end{proof} We return to the proof Proposition~\ref{pr:jets}. Note that $g^{-1}g'g\in {\rm GL}_{r,\mathbb O}$. It is easy to see that the set of kernel-strict loops is invariant under the multiplication by points of ${\rm GL}_{r,\mathbb O}$ on the right, so $g'g=g\big(g^{-1}g'g\big)$ is kernel-strict. Clearly, $\val(\det(g'g))=\val(\det g)=d$, so (ii) follows. Further, $g^{-1}g'g$ conjugates $\Phi'$ to $\Phi$, so (iii) follows as well. Proposition~\ref{pr:jets} is proved. \end{proof} For every $d\ge0$ we fix $N(d,\lambda)$ satisfying the conditions of the above proposition. \begin{Definition}\label{def:stabilized} For $N>N(d,\lambda)$ we call any jet in $J_{d,N}(\lambda)$ \emph{stabilized} and call $d$ its \emph{degree}. \end{Definition} According to the proposition, the degree of a stabilized jet is well-defined and any two lifts of a stabilized jet to an infinite jet are conjugate by an element of ${\rm GL}_{r,\mathbb O}$. Note that for every jet $\Phi\in J_d(\lambda)$ its truncation $\pi_N(\Phi)$ is stabilized for $N$ large enough. \subsubsection{Local stacks}\label{sect:LocStacks} Consider the quotient stack $\mathcal{P}{air}^{\rm loc}=\mathcal{P}{air}^{\rm loc}(\lambda):=J(\lambda)/{\rm GL}_{r,\mathbb O}$, where ${\rm GL}_{r,\mathbb O}$ acts by conjugation. We skip $\lambda$ from the notation as it is fixed until the end of Section~\ref{sect:Factorization}. Since the degree function on $J(\lambda)$ is ${\rm GL}_{r,\mathbb O}$-invariant, we get the degree function on the points of $\mathcal{P}{air}^{\rm loc}$. Note that $\mathcal{P}{air}^{\rm loc}$ classifies pairs $(F,\Phi)$, where $F$ is a rank $r=|\lambda|$ vector bundle over~$\Delta$, $\Phi$~is a nilpotent endomorphism of $F$ generically conjugate to $n_\lambda$. This follows from the fact that every vector bundle on $\Delta$ is trivial. It also follows that every $K$-point of $\mathcal{P}{air}^{\rm loc}$ is isomorphic to a point of the form $(\mathbb O^r,\Phi)$, where $\Phi\in\mathfrak{gl}_{r,\mathbb O}$. We emphasize that $\mathcal{P}{air}^{\rm loc}$ is not an Artin stack (its isotropy groups are not of finite type). Define $\mathcal{P}{air}^{\rm loc}_N:=J_N\big(\overline\mathcal O_\lambda\big)/J_N({\rm GL}_r)$; this is an Artin stack of finite type. The points of~$\mathcal{P}{air}^{\rm loc}_N$ are the pairs $(F,\Phi)$ where $F$ is a vector bundle on~$\Delta_N$, $\Phi$~is an endomorphism of $F$ such that if we trivialize $F$, $\Phi$ becomes a jet with values in $\overline\mathcal O_\lambda$. We say that $(F,\Phi)$ is \emph{stabilized} if $\Phi$ is stabilized in the sense of Definition~\ref{def:stabilized}. Note that this does not depend on the trivialization of~$F$. If $(F,\Phi)$ is stabilized, then we have a well-defined notion of the degree of~$(F,\Phi)$. Explicitly, we can lift $(F,\Phi)$ to a point $(\mathbb O^r,\overline\Phi)$ of $\mathcal{P}{air}^{\rm loc}$, and the degree of $(F,\Phi)$ is equal to the degree of $\overline\Phi\in\mathfrak{gl}_{r,\mathbb O}$. We define the stack $\mathcal{P}{air}^{{\rm loc},{\rm fl}}$ as the stack classifying triples $(F,\Phi,F_\bullet)$, where $(F,\Phi)$ is a point of $\mathcal{P}{air}^{\rm loc}$, $F_\bullet$ is a flag in the fiber $F_0$ preserved by $\Phi(0)$. We define the stack $\mathcal{P}{air}^{{\rm loc},{\rm fl}}_N$ as the stack classifying triples $(F,\Phi,F_\bullet)$, where $(F,\Phi)$ is a point of $\mathcal{P}{air}^{\rm loc}_N$, $F_\bullet$ is a flag in $F_0$ preserved by $\Phi(0)$. \subsubsection{Preparation for the proof of Theorem~\ref{th:Factorization}}\label{sect:StratThm} We will assume that $D=x$ is a single rational point of $X$. This will unburden the notation; the general case is proved similarly. Thus we want to prove that \[ \big[\mathcal{P}{air}^{{\rm nilp},-}(X,x,\lambda)\big]\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]=\big[\mathcal{P}{air}^{{\rm nilp},-}(X,\varnothing,\lambda)\big]\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\infty,\lambda\big)\big]_x. \] Equating the graded components, we see that this reduces to the following proposition. \begin{Proposition}\label{pr:factorization} Let $d$ be a nonpositive integer, $r_\bullet=(r_1,r_2,\dots)$ be a sequence of nonnegative integers such that $\sum_ir_i=r=|\lambda|$. Then we have in $\Mot(\mathbf{k})$: \begin{gather* \sum_{d'+d''=d}\big[\mathcal{P}{air}^{{\rm nilp},-}_{r,r_\bullet,d'}(X,x,\lambda)\times\mathcal{P}{air}^{{\rm nilp},-}_{r,d''}\big(\P^1,\varnothing,\lambda\big)\big]\\ \qquad{} = \sum_{d'+d''=d}\big[\mathcal{P}{air}^{{\rm nilp},-}_{r,d'}(X,\varnothing,\lambda)\times\mathcal{P}{air}^{{\rm nilp},-}_{r,r_\bullet,d''}\big(\P^1,\infty,\lambda\big)\big]. \end{gather*} \end{Proposition} This proposition will be proved in Section~\ref{sect:ProofFact}. We emphasize that the sum is over all $d',d''\in\mathbb{Z}$ with $d'+d''=d$ but the terms are non-zero only if $d',d''\in[d,0]$. We note that the RHS is manifestly independent of $x$. Thus, the LHS is independent of $x$ as well. \subsubsection[The restriction to the formal neighborhood of $x$]{The restriction to the formal neighborhood of $\boldsymbol{x}$}\label{sect:Res} We keep the simplifying assumption that $D=\{x\}$ is a single point; we write $r_\bullet$ instead of~$r_{x,\bullet}$. Fix $\gamma=(r,r_{\bullet},d)\in\Gamma_+'$. For $x\in X$ let $\mathcal O_{X,x}$ be the local ring of $x$ and $\hat\mathcal O_{X,x}$ be its formal completion. Set $\Delta_x:=\Spec\hat\mathcal O_{X,x}$. Choose a formal coordinate at $x$, use it to identify $\Delta_x$ with $\Delta$ and the $N$-th infinitesimal neighborhood $\Delta_{x,N}$ of $x$ with $\Delta_N$. Consider the restriction morphism \begin{equation}\label{eq:res_x_fl} \loc_x^{\rm fl}\colon \ \mathcal{P}{air}_{r,r_\bullet,d}^{{\rm nilp},-}(X,x,\lambda)\to\mathcal{P}{air}^{{\rm loc},{\rm fl}}_{r_\bullet},\qquad (E,\Psi,E_{x,\bullet}) \mapsto(E|_{\Delta_x},\Psi|_{\Delta_x},E_{x,\bullet}). \end{equation} Similarly we have a morphism \begin{equation}\label{eq:res_x} \loc_x \colon \ \mathcal{P}{air}_{r,d}^{{\rm nilp},-}(X,\varnothing,\lambda)\to\mathcal{P}{air}^{\rm loc},\qquad (E,\Psi) \mapsto(E|_{\Delta_x},\Psi|_{\Delta_x}). \end{equation} Our nearest goal is to describe the fibers of these morphisms. For a nonpositive integer $e$, let ${\mathcal F}{ib}_e(X,x)$ denote the open substack of $\mathcal{P}{air}^{{\rm nilp},-}_{r,e}(X,\varnothing,\lambda)$ consisting of $(E,\Psi)$ such that $\Psi$ is conjugate to $n_\lambda$ at $x$. Let $\overline{{\mathcal F}{ib}}_e(X,x)$ denote the stack of triples $(E,\Psi,s)$, where $(E,\Psi)$ is a~point of ${\mathcal F}{ib}_e(X,x)$, $s$ is a trivialization of $E$ over $\Delta_x$ such that $\Psi=n_\lambda$ in this trivialization. Recall that in Section~\ref{sect:LocStacks} we defined the notion of degree for the points of $\mathcal{P}{air}^{\rm loc}$. \begin{Lemma}\label{lm:InfFiber}\quad \begin{enumerate}\itemsep=0pt \item[$(i)$] The fiber of $\loc_x^{\rm fl}$ over $(F,\Phi,F_\bullet)$ is isomorphic to $\overline{{\mathcal F}{ib}}_{d+e}(X,x)$, where $e$ is the degree of~$(F,\Phi)$. \item[$(ii)$] Similarly, the fiber of $\loc_x$ over $(F,\Phi)$ is isomorphic to $\overline{{\mathcal F}{ib}}_{d+e}(X,x)$, where $e$ is the degree of $(F,\Phi)$. \end{enumerate} \end{Lemma} \begin{proof} We prove (ii) first. Fix a trivialization of $F$ on the formal disc $\Delta$. Then $\Phi$ becomes an element of $\mathfrak{gl}_{r,\mathbb K}$ and we choose a kernel-strict $g$ such that $g\Phi g^{-1}=n_\lambda$. Then $\val(\det g)=e$. Denote the fiber under consideration by $\overline{{\mathcal F}{ib}}$. The fiber can be described as the stack of triples $(E,\Psi,s)$, where $E$ is a nonpositive vector bundle, $\Psi$ is an endomorphism, $s$ is the trivialization of~$E$ over $\Delta_x$ such that in this trivialization we have $\Psi|_{\Delta_x}=\Phi$. Note that such $\Psi$ is automatically conjugate to $n_\lambda$ at the generic point of $X$. If $(E,\Psi,s)$ is a point of $\overline{{\mathcal F}{ib}}$, then $E|_{X-x}$ is trivialized over the punctured disc $\mathring\Delta_x$, and we use the $g$ chosen above to glue $E|_{X-x}$ with the trivial bundle $\mathbf{k}^r\times\Delta_x$ on $\mathring\Delta_x$ (we recall that $g$ can be viewed as an automorphism of the trivial vector bundle on $\mathring\Delta_x$). We obtain a new vector bundle $E'$ on $X$ with an isomorphism $E'|_{X-x}\simeq E|_{X-x}$ and a trivialization over $\Delta_x$. Thus $\Psi$ gives rise to an endomorphism $\Psi'$ of $E'|_{X-x}$. It is easy to derive from the definition of $g$ that in the given trivialization we have $\Psi'|_{\mathring\Delta_x}=n_\lambda$. Thus $\Psi'$ extends to $x$ and, moreover, in the given trivialization of $E'$ over $\Delta_x$ we have $\Psi'|_{\Delta_x}=n_\lambda$. Note that $E'$ is nonpositive. Indeed, $\Ker\Psi$ is nonpositive as a subbundle of $E$. Since $g$ is kernel-strict, the isomorphism between $\Ker\Psi$ and $\Ker\Psi'$ extends from $X-x$ to $X$. Thus $\Ker\Psi'$ is also nonpositive. But by~\cite[Proposition~5.3]{MellitPunctures} this implies that $E'$ is nonpositive as well. Next, we have an isomorphism between $\wedge^rE$ and $\wedge^rE'$ over $X-x$, and it has a zero of order $\val(\det g)$ at $x$. Thus $\deg E'=\deg\wedge^rE'=\deg\wedge^rE+\val(\det g)=\deg E+e=d+e$. We have constructed a morphism $\overline{{\mathcal F}{ib}}\to\overline{{\mathcal F}{ib}}_{d+e}(X,x)$. Conversely, given a point $(E',\Psi',s')$ of $\overline{{\mathcal F}{ib}}_{d+e}(X,x)$, we use $g$ and $s'$ to construct a new bundle $E$ with an isomorphism to $E'$ over $X-x$ and a trivialization over $\Delta_x$. Then $\Psi'|_{X-x}$ give rise to an endomorphism of $E|_{X-x}$ and we check that it extends to $x$ and, moreover, in the trivialization of $E$ over $\Delta_x$ we have $\Psi|_{\Delta_x}=\Phi$. Now it is easy to see that the two constructions are inverse to each other. This proves (ii). Now (i) follows from the cartesian diagram \[ \begin{CD} \mathcal{P}{air}_{r,r_\bullet,d}^{{\rm nilp},-}(X,x,\lambda)@>\loc_x^{\rm fl}>>\mathcal{P}{air}^{{\rm loc},{\rm fl}}_{r_\bullet}\\ @VVV @VVV\\ \mathcal{P}{air}_{r,d}^{{\rm nilp},-}(X,\varnothing,\lambda)@>\loc_x>>\mathcal{P}{air}^{\rm loc}, \end{CD} \] where the vertical arrows correspond to forgetting the flags. \end{proof} Consider now the compositions of~\eqref{eq:res_x_fl} and~\eqref{eq:res_x} with restrictions to the $N$-th infinitesimal neighborhood of $x$. \begin{equation}\label{eq:res_x_flN} \loc_{x,N}^{\rm fl}\colon \ \mathcal{P}{air}_{r,r_\bullet,d}^{{\rm nilp},-}(X,x,\lambda)\to\mathcal{P}{air}^{{\rm loc},{\rm fl}}_{r_\bullet,N},\qquad (E,\Psi,E_{x,\bullet}) \mapsto(E|_{\Delta_{x,N}},\Psi|_{\Delta_{x,N}},E_{x,\bullet}). \end{equation} Similarly we have a morphism \begin{equation}\label{eq:res_xN} \loc_{x,N}\colon \ \mathcal{P}{air}_{r,d}^{{\rm nilp},-}(X,\varnothing,\lambda)\to\mathcal{P}{air}^{\rm loc}_N,\qquad (E,\Psi) \mapsto(E|_{\Delta_{x,N}},\Psi|_{\Delta_{x,N}}). \end{equation} Let $(F,\Phi,F_\bullet)$ be a point of $\mathcal{P}{air}^{{\rm loc},{\rm fl}}_{r_\bullet,N}$ and assume that $(F,\Phi)$ is stabilized. By Definition~\ref{def:stabilized} we can find $\Phi'\in\mathfrak{gl}_{r,\mathbb O}$ such that $(\mathbb O^r,\Phi')$ lifts $(F,\Phi)$ and $N>N(e,\lambda)$, where $e$ is the degree of $\Phi'$ and $N(e,\lambda)$ is the integer number from Proposition~\ref{pr:jets}, which was fixed just before Definition~\ref{def:stabilized}. Choose a kernel-strict $g\in\mathfrak{gl}_{r,\mathbb K}$ such that $g\Phi'g^{-1}=n_\lambda$. Denote by $Z^{(N)}_g$ the intersection $Z_\mathbb O\cap\big(g^{-1}{\rm GL}^{(N)}g\big)\subset {\rm GL}_{r,\mathbb K}$, where $Z=Z_\lambda$ is the centralizer of $n_\lambda$ in ${\rm GL}_{|\lambda|}$ (cf.~Lemma~\ref{lm:Zspecial}). Recall that by Proposition~\ref{pr:jets}(i) we may assume that the orders of the poles of $g$ and $g^{-1}$ are less than $N/2$ so we have $Z^{(N)}_g\subset Z^{(1)}$. This is a pro-unipotent group. Clearly, $Z_\mathbb O$ (and thus~$Z^{(N)}_g$ as well) acts on $\overline{{\mathcal F}{ib}}_{d+e}(X,x)$ by changing the trivialization of $E$ on $\Delta_x$. \begin{Lemma}\label{lm:fiber} \quad \begin{enumerate}\itemsep=0pt \item[$(i)$] Let $(F,\Phi,F_\bullet)\in\mathcal{P}{air}_{r_\bullet,N}^{{\rm loc},{\rm fl}}$ be such that $(F,\Phi)$ is stabilized, choose $g\in {\rm GL}_{r,\mathbb K}$ as in the previous paragraph. Then the fiber of $\loc_{x,N}^{\rm fl}$ over $(F,\Phi,F_\bullet)$ is isomorphic to $\overline{{\mathcal F}{ib}}_{d+e}(X,x)/Z^{(N)}_g$, where $e$ is the degree of $(F,\Phi)$. \item[$(ii)$] Similarly, the fiber of $\loc_{x,N}$ over $(F,\Phi)$ is isomorphic to $\overline{{\mathcal F}{ib}}_{d+e}(X,x)/Z^{(N)}_g$. \end{enumerate} \end{Lemma} \begin{proof} Let us prove (ii), the proof of (i) is completely analogous. Denote the fiber under consideration by ${\mathcal F}{ib}$ and let $\overline{{\mathcal F}{ib}}$ be the fiber of $\loc_x$ over $(\mathbb O^r,\Phi')$. Then we have a restriction morphism $\overline{{\mathcal F}{ib}}\to{\mathcal F}{ib}$. It follows from the stability of $(F,\Phi)$ and Proposition~\ref{pr:jets}(iii) that this morphism is surjective. On the other hand, it is easy to see that two points of $\overline{{\mathcal F}{ib}}$ map to the same point of ${\mathcal F}{ib}$ if and only if they differ by the action of an element of $gZ_\mathbb O g^{-1}\cap {\rm GL}^{(N)}$. On the other hand, according to Lemma~\ref{lm:InfFiber}, $\overline{{\mathcal F}{ib}}\simeq\overline{{\mathcal F}{ib}}_{d+e}(X,x)$. One checks that under this isomorphism, the action of $gZ_\mathbb O g^{-1}\cap {\rm GL}^{(N)}$ on $\overline{{\mathcal F}{ib}}$ corresponds to the action of $Z^{(N)}_g$ on $\overline{{\mathcal F}{ib}}_{d+e}(X,x)$. \end{proof} We need to calculate the motivic class of this fiber. Set $Z_g:=Z_\mathbb O/Z^{(N)}_g$; this is a group of finite type. \begin{Lemma}\label{lm:MotFiber} We have in $\Mot(\mathbf{k})$ \[ \big[\overline{{\mathcal F}{ib}}_{d+e}(X,x)/Z^{(N)}_g\big]=[{\mathcal F}{ib}_{d+e}(X,x)]/[Z_g]. \] \end{Lemma} \begin{proof}For large $M$ we have $Z^{(M)}\subset Z_g^{(N)}$ and this subgroup is normal. \emph{Claim.} The groups $Z^{(N)}_g/Z^{(M)}$, $Z_\mathbb O/Z^{(M)}$, and $Z_\mathbb O/Z^{(N)}_g$ are special. \emph{Proof of the claim.} Recall that $Z_g^{(N)}\subset Z^{(1)}$. The group $Z^{(N)}_g/Z^{(M)}$ is special, since every unipotent group is special by Lemma~\ref{lm:SpecialRad}. Next, the quotient of $Z_\mathbb O/Z^{(M)}$ by the unipotent subgroup $Z^{(1)}/Z^{(M)}$ is equal to $Z$ and the statement follows from Lemmas~\ref{lm:SpecialRad} and~\ref{lm:Zspecial}. A~similar argument shows that $Z_\mathbb O/Z^{(N)}_g$ is special.\qed We continue with the proof of Lemma~\ref{lm:MotFiber}. Next, $\overline{\mathcal F}{ib}_{d+e}(X,x)/Z^{(M)}$ is a $Z_\mathbb O/Z^{(M)}$-principal bundle over ${\mathcal F}{ib}_{d+e}(X,x)$. Since $Z_\mathbb O/Z^{(M)}$ is a special group, we get by Lemma~\ref{lm:SpecialMot} \[ \big[\overline{\mathcal F}{ib}_{d+e}(X,x)/Z^{(M)}\big]=\big[Z_\mathbb O/Z^{(M)}\big][{\mathcal F}{ib}_{d+e}(X,x)]. \] Similarly, $\overline{{\mathcal F}{ib}}_{d+e}(X,x)/Z^{(M)}$ is a $Z^{(N)}_g/Z^{(M)}$-principal bundle over $\overline{{\mathcal F}{ib}}_{d+e}(X,x)/Z^{(N)}_g$ and we get \[ \big[\overline{{\mathcal F}{ib}}_{d+e}(X,x)/Z^{(M)}\big]=\big[Z^{(N)}_g/Z^{(M)}\big]\big[\overline{{\mathcal F}{ib}}_{d+e}(X,x)/Z^{(N)}_g\big]. \] Finally, $Z_\mathbb O/Z^{(M)}$ is a $Z^{(N)}_g/Z^{(M)}$-principal bundle over $Z_g$ and we have \[ \big[Z_\mathbb O/Z^{(M)}\big]=\big[Z^{(N)}_g/Z^{(M)}\big][Z_g]. \] The lemma follows from these three equations. \end{proof} \begin{Lemma}\label{lm:unique} Let $N$ be an integer larger than $N(j,\lambda)$ for all $j=0,\dots,-d$. Assume that the fiber of~\eqref{eq:res_xN} over $(F,\Phi)$ is non-empty. Then $(F,\Phi)$ is stabilized. A similar statement holds for the fibers of~\eqref{eq:res_x_flN}. \end{Lemma} \begin{proof} We prove the statement about the fibers of~\eqref{eq:res_xN}, the other statement being analogous. Let $(E,\Psi)$ be a point of the fiber, then the fiber of $\loc_x$ over $(E|_{\Delta_x},\Psi|_{\Delta_x})$ is non-empty. Then by Lemma~\ref{lm:InfFiber} this fiber is isomorphic to $\overline{{\mathcal F}{ib}}_{d+e}(X,x)$, where $e$ is the degree of $(E|_{\Delta_x},\Psi|_{\Delta_x})$ (recall that $e\ge0$). Since $\overline{{\mathcal F}{ib}}_{d+e}(X,x)$ classifies nonpositive vector bundles, we get $d+e\le0$. Thus $N>N(e,\lambda)$ and we see that $(F,\Phi)$ is stabilized. \end{proof} \subsubsection{Proof of Proposition~\ref{pr:factorization}}\label{sect:ProofFact} Let us take an integer $N$ larger than $N(j,\lambda)$ for all $j=0,\dots,-d$. It is enough to show that the motivic functions \[ A:=\sum_{d'+d''=d}\big[\mathcal{P}{air}^{{\rm nilp},-}_{r,r_\bullet,d'}(X,x,\lambda)\times\mathcal{P}{air}^{{\rm nilp},-}_{r,d''}\big(\P^1,\varnothing,\lambda\big) \to\mathcal{P}{air}^{{\rm loc},{\rm fl}}_{r_\bullet,N}\times\mathcal{P}{air}^{\rm loc}_N\big] \] and \[ B:=\sum_{d'+d''=d}\big[\mathcal{P}{air}^{{\rm nilp},-}_{r,r_\bullet,d'}\big(\P^1,\infty,\lambda\big)\times\mathcal{P}{air}^{{\rm nilp},-}_{r,d''}(X,\varnothing,\lambda) \to\mathcal{P}{air}^{{\rm loc},{\rm fl}}_{r_\bullet,N}\times\mathcal{P}{air}^{\rm loc}_N\big] \] are equal. The morphisms are $\loc_{x,N}^{\rm fl}\times\loc_{\infty,N}$ and $\loc_{\infty,N}^{\rm fl}\times\loc_{x,N}$ respectively. Let $K\supset\mathbf{k}$ be an extension and let $\xi$ be a $K$-point of the stack $\mathcal{P}{air}^{{\rm loc},{\rm fl}}_{r_\bullet,N}\times\mathcal{P}{air}^{\rm loc}_{r,N}$ represented by $((F,\Phi,F_\bullet),(F',\Phi'))$. By Proposition~\ref{pr:MotFunEqual} it is enough to show that $\xi^*A=\xi^*B$. Using base change, we may assume that $K=\mathbf{k}$. According to Lemma~\ref{lm:unique}, these motivic functions are zero unless $(F,\Phi)$ and $(F',\Phi')$ are stabilized. Let us lift $(F,\Phi)$ and $(F',\Phi')$ to $(\mathbb O^r,\overline\Phi)$ and $(\mathbb O^r,\overline\Phi')$, where $\Phi',\overline\Phi'\in\mathfrak{gl}_{r,\mathbb O}$. Choose kernel-strict $g,g'\in {\rm GL}_{r,\mathbb K}$ such that $g$, $g^{-1}$, $g'$, and $(g')^{-1}$ have poles of order less than $N/2$ and such that $g\overline\Phi g^{-1}=g'\overline\Phi'(g')^{-1}=n_\lambda$. Then, according to Lemmas~\ref{lm:fiber} and~\ref{lm:MotFiber} (applied to $X$ and $\P^1$) we get \begin{align*} \xi^*A& =\frac{\sum\limits_{d'+d''=d}[{\mathcal F}{ib}_{d'+e}(X,x)][{\mathcal F}{ib}_{d''+e'}(\P^1,\infty)]}{[Z_g][Z_{g'}]}\\ & = \frac{\sum\limits_{d'+d''=d+e+e'}[{\mathcal F}{ib}_{d'}(X,x)][{\mathcal F}{ib}_{d''}(\P^1,\infty)]}{[Z_g][Z_{g'}]}. \end{align*} Similarly, \begin{align*} \xi^*B& =\frac{\sum\limits_{d'+d''=d}[{\mathcal F}{ib}_{d'+e}(\P^1,\infty)][{\mathcal F}{ib}_{d''+e'}(X,x)]}{[Z_g][Z_{g'}]}\\ & = \frac{\sum\limits_{d'+d''=d+e+e'}[{\mathcal F}{ib}_{d'}(\P^1,\infty)][{\mathcal F}{ib}_{d''}(X,x)]}{[Z_g][Z_{g'}]}. \end{align*} We see that $\xi^*A=\xi^*B$. This completes the proof of Proposition~\ref{pr:factorization} and thus the proof of Theorem~\ref{th:Factorization}.\qed \begin{Remark} We emphasize that we only worked with motivic classes of Artin stacks of finite type. It seems plausible that one can define motivic classes of stacks like $\big[\,\overline{{\mathcal F}{ib}}_d(X,x)\big]$ using some ideas of motivic integration. This would significantly simplify our argument. Unfortunately, we were not able to develop such a formalism. \end{Remark} \subsection[Case of $\P^1$ and two points]{Case of $\boldsymbol{\P^1}$ and two points}\label{sect:P1} Consider the case when $X=\P^1$, $D=\{0,\infty\}$. \begin{Proposition}\label{pr:P1} We have in $\Mot(\mathbf{k})[[\Gamma'_+]]\subset\Mot(\mathbf{k})\big[\big[w,w_{0,\bullet},w_{\infty,\bullet},z^{-1}\big]\big]$ \begin{equation}\label{eq:P10infty} \big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\{0,\infty\}\big)\big]=\Exp\left(w\frac{\sum_{j=1}^\infty\sum_{j'=1}^\infty w_{0,j}w_{\infty,j'}}{(\mathbb{L}-1)(1-z^{-1})}\right), \end{equation} where $\Exp$ is the plethystic exponent defined in Section~{\rm \ref{sect:Plethystic}}. \end{Proposition} \begin{proof} We note that the proof in~\cite[Section~5.4]{MellitPunctures} goes through in the motivic case as well. The only difference is that Mellit uses the Hall algebra of the Jordan quiver (that is, the Hall algebra of the category of vector spaces with nilpotent endomorphisms); this Hall algebra has to be replaced with the similar motivic Hall algebra in our case. Let us give more details. In~\cite[Section~5]{FedorovSoibelmans}, to any smooth projective geometrically connected curve over $\mathbf{k}$ we associated the Hall algebra of the category of coherent sheaves on this curve, denoted by~$\mathcal H$. Let us take the curve to be $\P^1_\mathbf{k}$ and let $\mathcal H_0$ be the subalgebra of torsion sheaves supported at $0\in\P^1_\mathbf{k}$. Since such sheaves are identified with finite dimensional representations of the Jordan quiver, we can view $\mathcal H_0$ as the Hall algebra of the category of such representations. Of course, we could have taken any other curve and any rational point. Following Mellit, we re-write the LHS of~\eqref{eq:P10infty} in terms of products of certain elements of this algebra. This part of Mellit's proof is geometric, so it is easily carried to the motivic case. The rest of the proof is a calculation in this Hall algebra; the necessary identities in the Hall algebra are easily derived from results of \cite[Section~5]{FedorovSoibelmans}. \end{proof} \begin{Remark} Note that the RHS of~\eqref{eq:P10infty} can be written without plethystic exponents as follows (cf.~\cite[Lemma~5.7.3]{FedorovSoibelmans}) \begin{equation* \prod_{d=0}^{-\infty}\prod_{j=1}^\infty\prod_{j'=1}^\infty\left( 1+\sum_{i\ge1}\frac{\mathbb{L}^{i(i-1)}}{\big(\mathbb{L}^i-1\big)\cdots\big(\mathbb{L}^i-\mathbb{L}^{i-1}\big)}z^{id}w^iw_{0,j}^iw_{\infty,j'}^i \right). \end{equation*} \end{Remark} \subsection{Motivic modified Macdonald polynomials} For a commutative unital ring $A$, let $\Sym_A[w_\bullet]$ be the ring of symmetric functions with coefficients in $A$ in variables $w_\bullet$. In this section we define axiomatically the images of modified Macdonald polynomials in $\Sym_{\Mot(\mathbf{k})[[z]]}[w_\bullet]$. Consider the modified Macdonald polynomials $\tilde H_\lambda(w_\bullet;q,z)\in\Sym_{\mathbb{Z}[q,z]}[w_\bullet]$. For a definition see, for example,~\cite[Definition~2.5]{MellitPunctures}. It is not clear from this definition that the coefficients of $\tilde H_\lambda(w_\bullet;q,z)$ are integers, but this is well-known (see, e.g.,~\cite{HaglundEtAlOnMacdonaldPoly} and references therein). Note that~$\tilde H_\lambda$ is a symmetric function so, formally speaking, it is not a polynomial. We denote by $\tilde H^{\rm mot}_\lambda(w_\bullet;z)\in\Sym_{\Mot(\mathbf{k})[z]}[w_\bullet]$ the image of the corresponding modified Macdonald polynomial under the homomorphism $\Sym_{\mathbb{Z}[q,z]}[w_\bullet]\to\Sym_{\Mot(\mathbf{k})[z]}[w_\bullet]$ sending $q$ to $\mathbb{L}$; we call these images \emph{motivic modified Macdonald polynomials}. Define the motivic Hall--Littlewood polynomials as the specialization \[ H_\lambda^{\rm mot}(w_\bullet):=\tilde H_\lambda^{\rm mot}(w_\bullet;0).\] Thus $H_\lambda^{\rm mot}$ is the image of the usual Hall--Littlewood polynomial under the homomorphism $\mathbb{Z}[q,w_\bullet]\to\Mot(\mathbf{k})[w_\bullet]$ sending $q$ to $\mathbb{L}$. The motivic Hall--Littlewood polynomials can be interpreted as follows: let $\Fl_\lambda$ stand for the scheme of all flags in $\mathbf{k}^{|\lambda|}$ preserved by $n_\lambda$. Then $\Fl_\lambda$ is graded by the type of the flag. It is not difficult to check that we have \[ H^{\rm mot}_\lambda(w_\bullet)=[\Fl_\lambda]\in\Mot(\mathbf{k})[w_\bullet] \] (cf.~\cite[Theorem~2.12, Corollary~2.13]{MellitPunctures}). It follows from~\cite[Chapter~3, equation~(2.7)]{macdonald1998symmetric} that $H_\lambda^{\rm mot}$ form a basis of the $\Mot(\mathbf{k})$-module $\Sym_{\Mot(\mathbf{k})}[w_\bullet]$. Thus $\tilde H_\lambda^{\rm mot}$ also form a basis of $\Sym_{\Mot(\mathbf{k})[[z]]}[w_\bullet]$. \begin{Proposition}\label{pr:axiomMacdonald}\quad \begin{enumerate}\itemsep=0pt \item[$(a)$] The motivic modified Macdonald polynomials $\tilde H_\lambda^{\rm mot}$ satisfy the following properties \begin{enumerate}\itemsep=0pt \item[$(i)$] \begin{equation}\label{eq:scalar} \Exp\left(\frac{\sum\limits_{j=1}^\infty\sum\limits_{j'=1}^\infty w_{0,j}w_{\infty,j'}}{(\mathbb{L}-1)(1-z)}\right)= \sum_\lambda a_\lambda(z)\tilde H_\lambda^{\rm mot}(w_{0,\bullet};z)\tilde H_\lambda^{\rm mot}(w_{\infty,\bullet};z), \end{equation} where $a_\lambda$ are invertible elements of $\Mot(\mathbf{k})[[z]]$. \item[$(ii)$] $\tilde H_\lambda^{\rm mot}=\sum\limits_{\mu\colon \mu'\prec\lambda'}b_{\lambda\mu}H_\mu^{\rm mot}$, where $b_{\lambda\mu}$ are some elements of $\Mot(\mathbf{k})[[z]]$ such that $b_{\lambda\lambda}$ are invertible. $($Here $\mu'$ and $\lambda'$ stand for the conjugate partitions, $\prec$ is the usual order on partitions.$)$ \item[$(iii)$] $\tilde H^{\rm mot}_\lambda(1_\bullet;z)=1$, where $1_\bullet$ stands for the sequence $(1,0,\dots,0,\dots)$. \item[$(iv)$] $\tilde H^{\rm mot}_\lambda$ is homogeneous in $w_\bullet$ of degree $|\lambda|$. \end{enumerate} \item[$(b)$] The motivic modified Macdonald polynomials are uniquely determined by properties $(i)$--$(iv)$. \item[$(c)$] Additionally we have \begin{equation}\label{eq:scalarLengh} a_\lambda(z)=\frac1{\prod\limits_{h\in\Hook(\lambda)}\big(\mathbb{L}^{a(h)}-z^{l(h)+1}\big)\big(\mathbb{L}^{a(h)+1}-z^{l(h)}\big)}, \end{equation} where $\Hook(\lambda)$ stands for the set of hooks of $\lambda$, $a(h)$ and $l(h)$ stand for the armlength and the leglength of the hook $h$ respectively. \end{enumerate} \end{Proposition} \begin{proof} (a) It is enough to check the corresponding properties for the usual modified Macdonald polynomials $\tilde H_\lambda\in\Sym_{\mathbb{Z}[q,z]}[w_\bullet]$ and Hall--Littlewood polynomials $H_\lambda\in\Sym_{\mathbb{Z}[q]}[w_\bullet]$. To prove property~(ii) we note first that according to~\cite[Definition~2.5]{MellitPunctures} we have \[ \tilde H_\lambda[(q-1)w_\bullet]=\sum_{\mu\colon \mu'\prec\lambda'}c_{\lambda\mu}(q,z)m_{\mu'}(w_\bullet), \] where $[(q-1)w_\bullet]$ stands for the plethystic action as in~\cite[Section~2.1]{MellitPunctures}. Recalling that the Hall--Littlewood polynomials are $z=0$ specializations of the modified Macdonald polynomials, we get \[ H_\lambda[(q-1)w_\bullet]=\sum_{\mu\colon \mu'\prec\lambda'}c_{\lambda\mu}(q,0)m_{\mu'}(w_\bullet). \] Now it is easy to see that an analogue of property~(ii) holds in $\Sym_{\mathbb{Q}[q,z]}(w_\bullet)$: we can write $\tilde H_\lambda=\sum\limits_{\mu\colon \mu'\prec\lambda'}b'_{\lambda\mu}H_\mu$, where $b'_{\lambda\mu}$ are some elements of $\mathbb{Q}[q,z]$. Next, $H_\lambda$ form a basis in $\Sym_{\mathbb{Z}[q]}(w_\bullet)$ (by~\cite[Chapter~3, equation~(2.7)]{macdonald1998symmetric}), so $\tilde H_\lambda$ form a basis in $\Sym_{\mathbb{Z}[q][[w]]}(w_\bullet)$. It follows that $b'_{\lambda\mu}\in\mathbb{Z}[w][[z]]$ and $b'_{\lambda\lambda}$ are invertible in this ring. Now property~(ii) follows. It is sufficient to prove properties (iii), and (iv) in $\Sym_{\mathbb{Q}(q,z)}[w_\bullet]$. Property~(iii) is clear from~\cite[Definition~2.5]{MellitPunctures}. Property~(iv) follows, for example, from the definition of $\tilde H$ given in~\cite{GarsiaHaiman1996remarkable}. We first prove an analogue of property~(i) in $\Sym_{\mathbb{Q}(q,z)}[w_\bullet]$. Recall from loc.~cit.~that $\Sym_{\mathbb{Q}(q,z)}[w_\bullet]$ carries the $q,z$-scalar product $(\cdot,\cdot)_{q,z}$. for which \[ \Exp\left(\frac{\sum\limits_{j=1}^\infty\sum\limits_{j'=1}^\infty w_{0,j}w_{\infty,j'}}{(q-1)(1-z)}\right) \] is the reproducing kernel. This means that if $f_\lambda(w_\bullet;q,z)$ is any graded $\mathbb{Q}(q,z)$-basis in \linebreak $\Sym_{\mathbb{Q}(q,z)}[w_\bullet]$ indexed by partitions and $f^\vee_\lambda(w_\bullet;q,z)$ is the dual basis with respect to $(\cdot,\cdot)_{q,z}$, then \begin{equation}\label{eq:RepKer} \Exp\left(\frac{\sum\limits_{j=1}^\infty\sum\limits_{j'=1}^\infty w_{0,j}w_{\infty,j'}}{(q-1)(1-z)}\right)=\sum_\lambda f_\lambda(w_{0,\bullet};q,z)f^\vee_\lambda(w_{\infty,\bullet};q,z). \end{equation} Next, by property~(iv) the basis $H_\lambda(w_\bullet;q,z)$ is a graded basis. Thus, by~\cite[Proposition~2.7]{MellitPunctures} the dual of $H_\lambda(w_\bullet;q,z)$ is equal to $a_\lambda(q,z)H_\lambda(w_\bullet;q,z)$ for some $a_\lambda(q,z)\in\mathbb{Q}(q,z)$. Further, in~\cite[Section~2.4]{MellitPunctures} it is shown that \begin{equation}\label{eq:alambda} a_\lambda(q,z)=\frac1{\prod\limits_{h\in\Hook(\lambda)}\big(q^{a(h)}-z^{l(h)+1}\big)\big(q^{a(h)+1}-z^{l(h)}\big)}. \end{equation} It is clear from this formula that $a_\lambda(q,z)\in\mathbb{Z}(q)[[z]]$. Now property~(i) follows from~\eqref{eq:RepKer}. The condition in part~(c) follows from~\eqref{eq:alambda}. Now we prove part~(b). Since $\tilde H_\lambda^{\rm mot}$ form a basis of $\Sym_{\Mot(\mathbf{k})[[z]]}[w_\bullet]$, there is a unique $\Mot(\mathbf{k})[[z]]$-linear scalar product on $\Sym_{\Mot(\mathbf{k})[[z]]}[w_\bullet]$ such that $\langle\tilde H_\lambda^{\rm mot},\tilde H_\mu^{\rm mot}\rangle=\delta_{\lambda\mu}\frac1{a_\lambda}$. (This is the scalar product such that the LHS of~\eqref{eq:scalar} is the reproducing kernel for this product). Let $H'_\lambda=H'_\lambda(w_\bullet;z)$ be symmetric functions satisfying conditions of part~(a), we need to show that $H'_\lambda=\tilde H_\lambda^{\rm mot}$. Applying condition (ii), we see that we can write \begin{equation}\label{eq:triangle} H'_\lambda=\sum_{\mu\colon \mu'\prec\lambda'}c_{\lambda\mu}\tilde H_\mu^{\rm mot}, \end{equation} where $c_{\lambda\lambda}$ is invertible. Condition~(i) shows that we have \[ \sum_\lambda a_\lambda(z)\tilde H_\lambda^{\rm mot}(w_{0,\bullet};z)\tilde H_\lambda^{\rm mot}(w_{\infty,\bullet};z)= \sum_\lambda a'_\lambda(z)H'_\lambda(w_{0,\bullet};z)H'_\lambda(w_{\infty,\bullet};z), \] with invertible $a'_\lambda(z)$. Recalling that $H'_\lambda$ form a graded basis in $\Sym_{\Mot(\mathbf{k})[[z]]}[w_\bullet]$, we see that $\langle H'_\lambda,H'_\mu\rangle=\delta_{\lambda\mu}\frac1{a'_\lambda}$. Indeed, $H_\lambda^{\rm mot}$ and $a_\lambda H_\lambda^{\rm mot}$ are dual basis for the scalar product, so the above equality shows that $H'_\lambda$ and $a'_\lambda H'_\lambda$ are dual basis as well. We prove that $H'_\lambda=\tilde H_\lambda^{\rm mot}$ by induction on the conjugate partition $\lambda'$ with respect to $\prec$. Thus we assume that $H'_\mu=\tilde H_\mu^{\rm mot}$ whenever $\mu'\prec\lambda'$. Taking the scalar product of~\eqref{eq:triangle} with $\tilde H_\mu^{\rm mot}=H'_\mu$, we see that $c_{\lambda\mu}\langle\tilde H_\mu^{\rm mot},\tilde H_\mu^{\rm mot}\rangle=0$ whenever $\mu\ne\lambda$. We see that $c_{\lambda\mu}=0$ so that $H'_\lambda=c_{\lambda\lambda}\tilde H_\lambda^{\rm mot}$. Now condition (iii) implies that $c_{\lambda\lambda}=1$. \end{proof} \subsection{Explicit formulas for the graded motivic classes of nilpotent pairs}\label{sect:MotEnd} Now we are ready to give the precise formula for $\big[\mathcal{P}{air}^{{\rm nilp},-}(X,D,\lambda)\big]$. Recall that for a partition~$\lambda$ we defined $J_\lambda^{\rm mot}(z),H_\lambda^{\rm mot}(z)\in\cMot(\mathbf{k})[[z]]$ in~\cite[Section~1.3.2]{FedorovSoibelmans}. In this paper, we will denote them by $J_{\lambda,X}^{\rm mot}(z)$ and $H_{\lambda,X}^{\rm mot}(z)$ respectively to emphasize that they depend on the curve $X$ and to ensure that they are not confused with motivic modified Macdonald polynomials $\tilde H_\lambda^{\rm mot}(w_\bullet;z)$ and with motivic Hall--Littlewood polynomials $H_\lambda^{\rm mot}(w_\bullet)$. Denote by $g$ the genus of~$X$. \begin{Theorem}\label{th:MotMellitPunctures} We have in $\cMot(\mathbf{k})[[\Gamma'_+]]$ \[ \big[\mathcal{P}{air}^{{\rm nilp},-}(X,D,\lambda)\big]=w^{|\lambda|}\mathbb{L}^{(g-1)\langle\lambda,\lambda\rangle}J_{\lambda,X}^{\rm mot}\big(z^{-1}\big) H_{\lambda,X}^{\rm mot}\big(z^{-1}\big)\prod_{x\in D}\tilde H_\lambda^{\rm mot}\big(w_{x,\bullet};z^{-1}\big). \] \end{Theorem} \begin{proof} According to~\cite[Theorem~1.4.1]{FedorovSoibelmans}, we have \[ \sum_\lambda\big[\mathcal{P}{air}^{{\rm nilp},+}(X,\varnothing,\lambda)\big]= w^{|\lambda|}\sum_\lambda\mathbb{L}^{(g-1)\langle\lambda,\lambda\rangle}J^{\rm mot}_{\lambda,X}(z)H^{\rm mot}_{\lambda,X}(z), \] where the superscript ``$+$'' stands for HN-nonnegative vector bundles (see Section~\ref{sect:Gamma} and~\cite[Section~3.2]{FedorovSoibelmans} for the definition). Inspecting the proof, we see that for each $\lambda$ the summands are equal: \begin{equation* \big[\mathcal{P}{air}^{{\rm nilp},+}(X,\varnothing,\lambda)\big]=w^{|\lambda|}\mathbb{L}^{(g-1)\langle\lambda,\lambda\rangle}J_{\lambda,X}^{\rm mot}(z)H_{\lambda,X}^{\rm mot}(z). \end{equation*} By Lemma~\ref{lm:PosNeg} we get an isomorphism of stacks \[ \mathcal{P}{air}^{{\rm nilp},+}_{r,d}(X,\varnothing,\lambda)\simeq\mathcal{P}{air}^{{\rm nilp},-}_{r,-d}(X,\varnothing,\lambda). \] Thus \[ \big[\mathcal{P}{air}^{{\rm nilp},-}(X,\varnothing,\lambda)\big]=w^{|\lambda|}\mathbb{L}^{(g-1)\langle\lambda,\lambda\rangle}J_{\lambda,X}^{\rm mot}\big(z^{-1}\big)H_{\lambda,X}^{\rm mot}\big(z^{-1}\big). \] To be able to apply Theorem~\ref{th:Factorization}, we need the following lemma. \begin{Lemma}\label{lm:MacDonald} We have in $\Mot(\mathbf{k})[[w_{\infty,\bullet},z^{-1}]]$ \[ \frac{\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\infty,\lambda\big)\big]}{\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]}= \tilde H_\lambda^{\rm mot}\big(w_{\infty,\bullet};z^{-1}\big). \] \end{Lemma} \begin{proof} Our proof is similar to that of~\cite[Theorem~5.5]{MellitPunctures}. Let $H'_\lambda\in\Mot(\mathbf{k})[[w_\bullet,z]]$ be the series such that \[ \frac{\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\infty,\lambda\big)\big]}{\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]}=H'_\lambda\big(w_{\infty,\bullet};z^{-1}\big). \] Denote by $\mathcal C_{\lambda\mu}$ the stack classifying pairs $(E,\Psi)$, where $E$ is a nonpositive vector bundle of rank~$|\lambda|$ on $\P^1$, $\Psi$ is an endomorphism of $E$ generically conjugate to $n_\lambda$ and conjugate to~$n_\mu$ at $x=\infty$. Then $\mathcal C_{\lambda\mu}$ is graded by the degree of $E$, and we have $[\mathcal C_{\lambda\mu}]\in\Mot(\mathbf{k})\big[\big[z^{-1}\big]\big]$. Clearly, we have \begin{equation}\label{eq:H'} H'_\lambda\big(w_\bullet;z^{-1}\big)=\sum_\mu\frac{[\mathcal C_{\lambda\mu}]}{\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]}H_\mu^{\rm mot}(w_\bullet), \end{equation} where $H_\mu^{\rm mot}$ are the motivic Hall--Littlewood polynomials. Note that $C_{\lambda\mu}=\varnothing$ unless $\mu'\prec\lambda'$ because for all $i$ the dimension of the fiber of $\Ker\Psi^i$ is semicontinuous on $\P^1$. Now it is easy to see that $H'_\lambda$ are symmetric functions with coefficients in $\Mot(\mathbf{k})[[z]]$. We will use Proposition~\ref{pr:axiomMacdonald}(b) to show that for all $\lambda$ we have $H'_\lambda=\tilde H_\lambda^{\rm mot}$. To show that $H'_\lambda$ satisfy property~(ii) of Proposition~\ref{pr:axiomMacdonald}(a) it remains to show that $[\mathcal C_{\lambda\lambda}]$ is invertible. This is completely similar to the proof of Lemma~\ref{lm:invertible}. To show that $H'_\lambda$ satisfy condition~(i) of Proposition~\ref{pr:axiomMacdonald}(a), we note that combining Theorem~\ref{th:Factorization} and Proposition~\ref{pr:P1}, we get \[ \Exp\left(\frac{\sum\limits_{j=1}^\infty w_{0,j}w_{\infty,j}}{(\mathbb{L}-1)(1-z)}\right)= \sum_\lambda a'_\lambda(z)H'_\lambda(w_{0,\bullet};z)H'_\lambda(w_{\infty,\bullet};z), \] where $a'_\lambda(z^{-1})=\big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]$ and the statement follows. Condition~(iii) of Proposition~\ref{pr:axiomMacdonald}(a) is obvious. Condition~(iv) follows from~\eqref{eq:H'}. We note also for the future use that it is clear from the argument that we have \begin{equation}\label{eq:P1emptyset} \big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,\varnothing,\lambda\big)\big]=a'_\lambda\big(z^{-1}\big)=a_\lambda\big(z^{-1}\big), \end{equation} where $a_\lambda(z)$ is given by~\eqref{eq:scalarLengh}. The proof of Lemma~\ref{lm:MacDonald} is complete. \end{proof} Now Theorem~\ref{th:Factorization} completes the proof of Theorem~\ref{th:MotMellitPunctures}. \end{proof} \begin{Corollary}\label{cor:Pairs} We have in $\cMot(\mathbf{k})[[\Gamma'_+]]$ \begin{gather*} \big[\mathcal{P}{air}^-(X,D)\big]= \Pow\left( \sum_\lambda w^{|\lambda|}\mathbb{L}^{(g-1)\langle\lambda,\lambda\rangle}J_{\lambda,X}^{\rm mot}\big(z^{-1}\big) H_{\lambda,X}^{\rm mot}\big(z^{-1}\big)\prod_{x\in D}\tilde H_\lambda^{\rm mot}\big(w_{x,\bullet};z^{-1}\big) ,\mathbb{L} \right). \end{gather*} \end{Corollary} \begin{proof} The argument is similar to the proof of~\cite[Proposition~3.8.1]{FedorovSoibelmans}. In more details, let $(E,\Psi,E_{\bullet,\bullet})$ be a $K$-point of $\mathcal{P}{air}^-(X,D)$. According to~\cite[Lemma~3.8.3]{FedorovSoibelmans}, we can uniquely decompose \[ (E,\Psi)\xrightarrow{\simeq}\bigoplus_i R_{\mathbf{k}(x_i)/K}(E_i,x_i\Id+\Psi_i), \] where $x_i$ are distinct closed points of $\mathbb A_K^1$ (the eigenvalues of $\Psi$), $(E_i,\Psi_i)$ are $\mathbf{k}(x_i)$-points of the stack $\mathcal{P}{air}^{{\rm nilp},-}(X,\varnothing)$, $\mathbf{k}(x_i)\supset K$ is the residue field of $x_i$, and $R_{\mathbf{k}(x_i)/K}$ is the pushforward functor. It follows easily from the proof of~\cite[Lemma~3.8.3]{FedorovSoibelmans}, that we can write uniquely \[ (E,\Psi,E_{\bullet,\bullet})\xrightarrow{\simeq}\bigoplus_i R_{\mathbf{k}(x_i)/K}(E_i,x_i\Id+\Psi_i,E_{i,\bullet,\bullet}), \] where $(E_i,\Psi_i,E_{i,\bullet,\bullet})$ are $\mathbf{k}(x_i)$-points of $\mathcal{P}{air}^{{\rm nilp},-}(X,D)$. It remains to use a version of~\cite[Lemma~3.8.2]{FedorovSoibelmans}. \end{proof} \subsection[Case of $\P^1$]{Case of $\boldsymbol{\P^1}$}\label{sect:P1ManyPts} In the case of $X=\P^1$ we can give a more explicit answer. Moreover, we get an answer valid in $\Mot(\mathbf{k})[[\Gamma'_+]]$ rather than in its completion $\cMot(\mathbf{k})[[\Gamma'_+]]$, which is desirable, since we do not know whether the natural homomorphism $\Mot(\mathbf{k})\to\cMot(\mathbf{k})$ is injective. We argue as in~\cite[Corollary~5.9]{MellitPunctures}. Combining Theorem~\ref{th:Factorization},~\eqref{eq:P1emptyset}, and~\eqref{eq:scalarLengh}, we get the following formula valid in $\Mot(\mathbf{k})[[\Gamma'_+]]$. \[ \big[\mathcal{P}{air}^{{\rm nilp},-}\big(\P^1,D,\lambda\big)\big]=w^{|\lambda|}\frac{\prod\limits_{x\in D}\tilde H_\lambda^{\rm mot}\big(w_{x,\bullet};z^{-1}\big)} {\prod\limits_{h\in\Hook(\lambda)}\big(\mathbb{L}^{a(h)}-z^{-l(h)-1}\big)\big(\mathbb{L}^{a(h)+1}-z^{-l(h)}\big)}, \] where $\Hook(\lambda)$ stands for the set of hooks of $\lambda$, $a(h)$ and $l(h)$ stand for the armlength and the leglength of the hook $h$ respectively. Arguing as in Corollary~\ref{cor:Pairs}, we get in $\Mot(\mathbf{k})[[\Gamma'_+]]$ \begin{equation}\label{eq:P1} \big[\mathcal{P}{air}^-\big(\P^1,D\big)\big]= \Pow\left( \sum_\lambda\frac{w^{|\lambda|}\prod\limits_{x\in D}\tilde H_\lambda^{\rm mot}(w_{x,\bullet};z)} {\prod\limits_{h\in\Hook(\lambda)} \big(\mathbb{L}^{a(h)}-z^{-l(h)-1}\big)\big(\mathbb{L}^{a(h)+1}-z^{-l(h)}\big)} ,\mathbb{L} \right). \end{equation} \section{Parabolic Higgs bundles with fixed eigenvalues}\label{sect:HiggsnEigenval} \subsection{Stacks of Higgs bundles} Let $X$ and $D$ be as above. From now on we denote by $g$ the genus of $X$. Our goal in this section is to calculate the motivic Donaldson--Thomas series of the category of parabolic Higgs bundles. More precisely, we calculate the motivic classes of the moduli stacks of Higgs bundles with fixed eigenvalues and with nonpositive underlying vector bundles. Our argument is similar to~\cite[Sections~3.4--3.5]{FedorovSoibelmans}. The main result is Corollary~\ref{cor:MAIN}. We denote by $\Omega_X$ the canonical line bundle on $X$. \begin{Definition} A \emph{parabolic Higgs bundle} of type $(X,D)$ is a triple $(E,E_{\bullet,\bullet},\Phi)$, where $(E,E_{\bullet,\bullet})$ is a point of $\mathcal{B}{un}^{\rm par}(X,D)$, $\Phi\colon E\to E\otimes\Omega_X(D)$ is an $\mathcal O_X$-linear morphism (called a \emph{Higgs field on $(E,E_{\bullet,\bullet})$}) such that for all $x\in D$ and $j\ge0$ we have $\Phi_x(E_{x,j})\subset E_{x,j}\otimes\Omega_X(D)_x$. \end{Definition} We denote the category (and the Artin stack) of parabolic Higgs bundles by $\mathcal{H}{iggs}=\mathcal{H}{iggs}(X,D)$. We define the $\Gamma'_+$-graded stack $\mathcal{H}{iggs}^-=\mathcal{H}{iggs}^-(X,D)$ following the general formalism of Section~\ref{sect:CatsOverPar}, that is, $\mathcal{H}{iggs}^-$ is the open substack of $\mathcal{H}{iggs}$ corresponding to Higgs bundles with nonpositive underlying vector bundle. Clearly, this stack is of finite type in the graded sense (that is, the graded components are of finite type). In this section $X$ and $D$ are fixed, so we skip them from the notation. \subsection{Existence of Higgs fields with prescribed residues}\label{Sect:ExistResidue} To determine a criterion for the existence of a Higgs bundle with prescribed residues, we use an approach similar to~\cite{AtiyahConnections,MihaiMonodromie,MihaiConnexions}. Let $E\to X$ be a vector bundle and let $\Phi\colon E\to E\otimes\Omega_X(D)$ be a morphism. In this case, for all $x\in D$ we have a residue $\Res_x\Phi\in\End(E_x)$. \begin{Proposition}\label{pr:exist} \label{existenceHiggs} Let $E$ be a vector bundle on $X$ and let for $x\in D$, $\rho_x\in\End(E_x)$ be an endomorphism of the fiber of $E$ at $x$. There exists a Higgs field $\Phi\colon E\to E\otimes\Omega_X(D)$ with $\Res_x\Phi=\rho_x$ for all $x\in D$ if and only if \[ \sum_{x\in D}\tr(\rho_x\phi_x)=0 \] for all $\phi\in\End(E)$, where $\tr$ stands for the trace. \end{Proposition} \begin{proof} Consider the short exact sequence of sheaves \[ 0\rightarrow\mathcal O_X\rightarrow\mathcal K_X\rightarrow\mathcal K_X/\mathcal O_X\rightarrow 0, \] where $\mathcal K_X$ is the constant sheaf corresponding to the function field of $X$. Let $\Omega_\mathcal K$ be the constant sheaf of meromorphic differential forms on $X$. We can obtain a new short exact sequence of sheaves: \[ 0\rightarrow\mathcal{E}{nd}(E)\otimes\Omega_X \rightarrow \mathcal{E}{nd}(E)\otimes \Omega_\mathcal K \rightarrow\mathcal{E}{nd}(E)\otimes(\Omega_\mathcal K/\Omega_X)\rightarrow0 \] by taking the tensor product of the first sequence with $\mathcal{E}{nd}(E)\otimes\Omega_X$. Note that the middle term in this sequence is a constant sheaf, while the last term is an (infinite) direct sum of skyscraper sheaves. That is, $\mathcal{E}{nd}(E)\otimes(\Omega_\mathcal K/\Omega_X)\cong\bigoplus_{x\in X} (i_x)_*(\End(E_x)\otimes(\Omega_\mathcal K/\Omega_X)_x)$, where $(\Omega_\mathcal K/\Omega_X)_x$ is the vector space of polar parts at $x$ of meromorphic 1-forms, the summation is taken over all closed points of $X$, and $i_x\colon x\rightarrow X$ is the inclusion. Passing to the long exact sequence for cohomology we obtain the following exact sequence of vector spaces: \[ \End(E)\otimes\Omega_\mathcal K\rightarrow \bigoplus_{x\in X}\End(E_x)\otimes(\Omega_\mathcal K/\Omega_X)_x\rightarrow H^1(X, \mathcal{E}{nd}(E)\otimes\Omega_X) \rightarrow 0. \] This implies that $H^1(X, \mathcal{E}{nd}(E)\otimes \Omega_X)$ may be presented as the quotient of \[ \bigoplus_{x\in X}\End(E_x)\otimes(\Omega_\mathcal K/\Omega_X)_x\] by the image of $\End(E)\otimes\Omega_\mathcal K$ (compare with the adelic description of cohomology given in~\cite[Chapter~2, Section~5]{SerreAlgGrClFields}). Further note that the required Higgs field $\Phi$ always exists locally, defined as $\Phi_x =\rho_x\frac{dz_x}{z_x}$, where $z_x$ is an \'etale coordinate near $x$ if $x\in D$ and $\Phi_x=0$ if $x\notin D$. Under the above presentation of $H^1(X, \mathcal{E}{nd}(E)\otimes \Omega_X)$, the local solutions $\Phi_x$ define a cohomology class $a(E,D,\rho_\bullet)\in H^1(X, \mathcal{E}{nd}(E)\otimes\Omega_X)$. Moreover, it follows from the exact sequence that $a(E,D,\rho_\bullet)=0$ if and only if $\Phi$ can be defined globally. Serre duality defines a bilinear pairing $H^1(X,\mathcal{E}{nd}(E)\otimes\Omega_X)\times\End(E)\rightarrow\mathbf{k}$. Using the above presentation for $H^1(X, \mathcal{E}{nd}(E)\otimes \Omega_X)$ this pairing may be evaluated on $\phi\in\End(E)$ as \[ \langle a(E,D,\rho_\bullet),\phi\rangle = \sum_{x\in X}\Res_x\tr(\Phi_x\phi_x) = \sum_{x\in D}\tr(\rho_x\phi_x). \] Since the pairing is perfect, $\sum\limits_{x\in D}\tr(\rho_x \phi_x) = 0$ for all $\phi\in\End(E)$ if and only if $a(E,D,\rho_\bullet)=0$. The proof is complete. \end{proof} \subsection{Parabolic Higgs bundles with fixed eigenvalues}\label{sect:ExistEigen} Recall that $\mathbf{k}[D\times\mathbb{Z}_{>0}]$ is the set of all $\mathbf{k}$-valued sequences $\zeta=\zeta_{\bullet,\bullet}=(\zeta_{x,j})$ indexed by $D\times\mathbb{Z}_{>0}$ such that $\zeta_{x,j}=0$ for $j\gg0$.\footnote{According to our convention we should denote $\zeta$ by $\zeta_{\bullet,\bullet}$ but it does not look nice in the formulas.} For $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ let $\mathcal{H}{iggs}(\zeta)=\mathcal{H}{iggs}(X,D,\zeta)$ denote the full subcategory of $\mathcal{H}{iggs}$ (and its stack of objects) corresponding to collections $(E,E_{\bullet,\bullet},\Phi)$ such that $(\Phi-\zeta_{x,j}1)(E_{x,{j-1}})\subset E_{x,j}\otimes\Omega_X(D)_x$ for all $x\in D$ and $j>0$. Again, the $\Gamma_+'$-graded stack $\mathcal{H}{iggs}^-(\zeta)$ is defined following the formalism of Section~\ref{sect:CatsOverPar}. Let $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ and let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$. Recall from Section~\ref{sect:DegreeSlope} that we have set \[ \deg_{0,\zeta}\gamma:=\sum_{x\in D}\sum_{j=1}^{\infty}\zeta_{x,j}r_{x,j}\in\mathbf{k}. \] \begin{Lemma}\label{lm:existence} Let $\mathbf E\in\mathcal{B}{un}^{\rm par}(\mathbf{k})$ and $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$. There exists an object $(\mathbf E,\Phi)\in\mathcal{H}{iggs}(\zeta)(\mathbf{k})$ if and only if $\deg_{0,\zeta}\mathbf E'=0$ for any direct summand $\mathbf E'$ of $\mathbf E$. \end{Lemma} Note that, in particular, this condition implies that $\deg_{0,\zeta}\mathbf E=0$. \begin{proof} The proof is the same as the proof of~\cite[Theorem~7.1]{Crawley-Boevey:Indecomposable} after replacing $b(E)$ with 0 and replacing~\cite[Theorem~7.2]{Crawley-Boevey:Indecomposable} with Proposition~\ref{pr:exist}. (The Atiyah class $b(E)$ represents the obstruction to existence of a connection (without poles) on $E$. Thus, it is absent for Higgs fields essentially because every vector bundle possesses the zero Higgs field.) \end{proof} \subsection{Parabolic pairs with isoslopy underlying parabolic bundles}\label{sect:Isoslopy} Recall that for $\kappa\in\mathbf{k}$ and $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$, a parabolic bundle $\mathbf E\in\mathcal{B}{un}^{\rm par}$ is $(\kappa,\zeta)$-isoslopy if \[ \frac{\deg_{\kappa,\zeta}\mathbf E'}{\rk\mathbf E'}=\frac{\deg_{\kappa,\zeta}\mathbf E}{\rk\mathbf E} \] whenever $\mathbf E'$ is a direct summand of $\mathbf E$. Similarly to~\cite[Lemma~3.2.2]{FedorovSoibelmans}, one checks that the notion of $(\kappa,\zeta)$- isoslopy parabolic bundle is invariant with respect to field extensions. Thus for each $\gamma\in\Gamma_+'$ we have a well-defined subset $\mathcal{B}{un}^{{\rm par},(\kappa,\zeta)-{\rm iso},-}_\gamma\subset|\mathcal{B}{un}^{{\rm par},-}_\gamma|$. As in~\cite[Lemma~3.2.3]{FedorovSoibelmans} we show that this subset is constructible. Let $\mathcal{P}{air}^{(\kappa,\zeta)-{\rm iso},-}_\gamma\!$ be the preimage of $\mathcal{B}{un}^{{\rm par},(\kappa,\zeta)-{\rm iso},-}_\gamma\!$ under the projection $|\mathcal{P}{air}|\!\to\!|\mathcal{B}{un}^{\rm par}|$. \begin{Proposition}\label{pr:Sasha} For $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$, set $\chi(\gamma):=(g-1)r^2+\sum\limits_{x\in D \sum\limits_{j<j'}r_{x,j}r_{x,j'}$. Then \[ [\mathcal{H}{iggs}_\gamma^-(\zeta)]= \begin{cases} \mathbb{L}^{\chi(\gamma)}\big[\mathcal{P}{air}^{(0,\zeta)-{\rm iso},-}_\gamma\big] & \text{if }\deg_{0,\zeta}\gamma=0,\\ 0 &\text{otherwise.} \end{cases} \] \end{Proposition} \begin{proof} The case of $\deg_{0,\zeta}\gamma\ne0$ is obvious in view of Lemma~\ref{lm:existence}. Assume that $\deg_{0,\zeta}\gamma=0$. It is enough to show the equality of motivic functions in $\Mot\big(\mathcal{B}{un}^{{\rm par},-}_\gamma\big)$: \begin{equation}\label{eq:IsoslopyHiggs} \big[\mathcal{H}{iggs}_\gamma^-(\zeta)\to\mathcal{B}{un}^{{\rm par},-}_\gamma\big]= \mathbb{L}^{\chi(\gamma)}\big[\mathcal{P}{air}^{(0,\zeta)-{\rm iso},-}_\gamma\to\mathcal{B}{un}^{{\rm par},-}_\gamma\big]. \end{equation} Let $K\supset\mathbf{k}$ be a field extension. Let $\xi\colon \Spec K\to\mathcal{B}{un}^{{\rm par},-}_\gamma$ be a point represented by a parabolic bundle $\mathbf E=(E,E_{\bullet,\bullet})$. In view of Proposition~\ref{pr:MotFunEqual}, we only need to check that the $\xi$-pullbacks of~\eqref{eq:IsoslopyHiggs} are equal. If $\mathbf E$ is not $(0,\zeta)$-isoslopy, then, by Lemma~\ref{lm:existence}, the pullbacks are equal to zero, so we assume that $\mathbf E$ is $(0,\zeta)$-isoslopy. Let $\higgs(\mathbf E,\zeta)$ denote the space of Higgs fields on $\mathbf E$ with eigenvalues $\zeta$ (that is, the $\mathbf E$-fiber of the projection $\mathcal{H}{iggs}(\zeta)\to\mathcal{B}{un}^{\rm par}$). By Lemma~\ref{lm:existence}, $\higgs(\mathbf E,\zeta)$ is non-empty, so it is a torsor over the vector space $\higgs(\mathbf E,0)$. Thus, \begin{equation* \xi^*\big[\mathcal{H}{iggs}_\gamma^-(\zeta)\to\mathcal{B}{un}^{{\rm par},-}_\gamma\big]=\mathbb{L}^{\dim \higgs(\mathbf E,0)}. \end{equation*} On the other hand, we have \[ \xi^*\big[\mathcal{P}{air}^{(0,\zeta)-{\rm iso},-}_\gamma\to\mathcal{B}{un}^{{\rm par},-}_\gamma\big]=\mathbb{L}^{\dim\End(\mathbf E)}. \] It remains to prove the following lemma. \begin{Lemma}\label{lm:PairHiggs} Let $\mathbf E\in\mathcal{B}{un}^{\rm par}_\gamma$ be a parabolic bundle. Then \[ \dim\End(\mathbf E)-\dim\higgs(\mathbf E,0)=-\chi(\gamma). \] \end{Lemma} \begin{proof} Write $\mathbf E=(E,E_{\bullet,\bullet})$. Let $\mathcal{E}{nd}(\mathbf E)\subset\mathcal{E}{nd}(E)$ be the subsheaf of endomorphisms preserving flags. One checks easily that the trace pairing gives an isomorphism between the dual sheaf $\mathcal{E}{nd}(\mathbf E)^\vee\otimes\Omega_X$ and $\mathcal{H}{iggs}(\mathbf E,0)$, where $\mathcal{H}{iggs}(\mathbf E,0)$ stands for the sheaf of Higgs fields on~$\mathbf E$ with zero eigenvalues. Thus by Riemann--Roch theorem we have \begin{align*} \dim\End(\mathbf E)-\dim\higgs(\mathbf E,0)& =h^0(X,\mathcal{E}{nd}(\mathbf E))-h^0\big(X,\mathcal{E}{nd}(\mathbf E)^\vee\otimes\Omega_X\big)\\ & =(1-g)\rk\mathcal{E}{nd}(\mathbf E)+\deg\mathcal{E}{nd}(\mathbf E). \end{align*} It remains to calculate $\deg\mathcal{E}{nd}(\mathbf E)$. For $x\in D$ consider the fiber $E_x$, its ring of endomorphisms $\End(E_x)$, its subspace $V_x$ of endomorphisms preserving the flag $E_{x,\bullet}$, and the quotient of vector spaces $W_x:=\End(E_x)/V_x$. Further, consider the torsion sheaf $W:=\oplus_{x\in D}(i_x)_*W_x$, where, as before, $i_x\colon x\to X$ is the inclusion. We have an exact sequence \[ 0\to\mathcal{E}{nd}(\mathbf E)\to\mathcal{E}{nd}(E)\to W\to0, \] so \[ \deg\mathcal{E}{nd}(\mathbf E)=\deg\mathcal{E}{nd}(E)-\length(W)=0-\sum_{x\in D \sum_{j<j'}r_{x,j}r_{x,j'} \] and the lemma follows. \end{proof} The lemma completes the proof of Proposition~\ref{pr:Sasha}. \end{proof} \begin{Proposition}\label{pr:IsoslProd} We have in $\Mot(\mathbf{k})[[\Gamma'_+]]$ \[ [\mathcal{P}{air}^-]= \prod_{\tau\in\mathbf{k}}\left( \sum_{\substack{\gamma\in\Gamma'_+\\ \deg_{\kappa,\zeta}\gamma=\tau\rk\gamma}}\big[\mathcal{P}{air}^{(\kappa,\zeta)-{\rm iso},-}_\gamma\big]e_\gamma \right). \] \end{Proposition} We note that the product makes sense because for a $\gamma\in\Gamma_+'$ there are only finitely many ways to write $\gamma$ as the sum of elements of $\Gamma_+'$. Also, the order of the multiples is irrelevant, since we are working with a commutative quantum torus. \begin{proof} The proof is almost the same as the proof of~\cite[Lemma~3.5.3 ]{FedorovSoibelmans} (see also~\cite[Proposition~3.5.1]{FedorovSoibelmans}). \end{proof} We need some notation. Let us write \[ \mathbb{L}\cdot\Log\left(\sum_\lambda w^{|\lambda|} J_{\lambda,X}^{\rm mot}\big(z^{-1}\big)H_{\lambda,X}^{\rm mot}\big(z^{-1}\big)\prod_{x\in D}\tilde H_\lambda^{\rm mot}\big(w_{x,\bullet};z^{-1}\big)\right)= \sum_{\gamma\in\Gamma'_+}\overline B_\gamma e_\gamma, \] where $\Log$ is the plethystic logarithm defined in Section~\ref{sect:Plethystic}, the summation is over all partitions. We note that $\overline B_\gamma$ are $W$-invariant, where $W=\prod\limits_{x\in D}\Sigma_\infty$ (cf.~Remark~\ref{rm:Weyl}). Note also that $\overline B_0=0$ by the definition of plethystic logarithm. \begin{Definition}\label{def:DT} The motivic classes $\overline B_\gamma\in\cMot(\mathbf{k})$ are called \emph{motivic Donaldson--Thomas invariants} of the pair $(X,D)$. \end{Definition} \begin{Corollary}\label{cor:isoslopy} For each $\tau\in\mathbf{k}$ we have in $\cMot(\mathbf{k})[[\Gamma'_+]]$ \[ \sum_{\substack{\gamma\in\Gamma'_+\\ \deg_{\kappa,\zeta}\gamma=\tau\rk\gamma}}\big[\mathcal{P}{air}^{(\kappa,\zeta)-{\rm iso},-}_\gamma\big]e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma'_+\\ \deg_{\kappa,\zeta}\gamma=\tau\rk\gamma}}\overline B_\gamma e_\gamma\right), \] where $\Exp$ is the plethystic exponent defined in Section~{\rm \ref{sect:Plethystic}}. \end{Corollary} \begin{proof} First of all, using Corollary~\ref{cor:Pairs} and properties of plethystic operations, we get \[ [\mathcal{P}{air}^-]=\Exp\left(\sum_{\gamma\in\Gamma'_+}\overline B_\gamma e_\gamma\right)= \prod_{\tau\in\mathbf{k}}\Exp\left(\sum_{\substack{\gamma\in\Gamma'_+\\ \deg_{\kappa,\zeta}\gamma=\tau\rk\gamma}}\overline B_\gamma e_\gamma\right). \] Now, it remains to use Proposition~\ref{pr:IsoslProd} and equate the slopes (cf.~\cite[Lemma~3.7.1]{FedorovSoibelmans}). \end{proof} \begin{Corollary}\label{cor:MAIN} We have in $\cMot(\mathbf{k})[[\Gamma'_+]]$ \[ \sum_{\gamma\in\Gamma'_+}\mathbb{L}^{-\chi(\gamma)}[\mathcal{H}{iggs}^-_\gamma(\zeta)]e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma'_+\\ \deg_{0,\zeta}\gamma=0}}\overline B_\gamma e_\gamma\right), \] where $\Exp$ is the plethystic exponent defined in Section~{\rm \ref{sect:Plethystic}}. \end{Corollary} \begin{proof} Use Proposition~\ref{pr:Sasha} and Corollary~\ref{cor:isoslopy}. \end{proof} \subsection[Case of $\P^1$]{Case of $\boldsymbol{\P^1}$} Assume now that $X=\P^1$. Then we have a simpler result. Moreover, it is more precise in the sense that we get an answer in $\Mot(\mathbf{k})$ rather than in $\cMot(\mathbf{k})$. Define the Donaldson--Thomas invariants $B_\gamma\in\Mot(\mathbf{k})$ by \begin{equation}\label{eq:DT_P1} \mathbb{L}\cdot\Log\left(\sum_\lambda\frac{w^{|\lambda|}\prod\limits_{x\in D}\tilde H_\lambda^{\rm mot}\big(w_{x,\bullet};z^{-1}\big)} {\prod\limits_{h\in\Hook(\lambda)} \big(\mathbb{L}^{a(h)}-z^{-l(h)-1}\big)\big(\mathbb{L}^{a(h)+1}-z^{-l(h)}\big)}\right)= \sum_{\gamma\in\Gamma'_+}B_\gamma e_\gamma. \end{equation} We have precisely the same formula as in Corollary~\ref{cor:MAIN}, where $\overline B_\gamma$ is replaced with $B_\gamma$ (thus, the formula is valid in $\Mot(\mathbf{k})$). The proof is the same as of Corollary~\ref{cor:MAIN} except that one uses~\eqref{eq:P1} instead of Corollary~\ref{cor:Pairs}. Comparing Corollary~\ref{cor:Pairs} with~\eqref{eq:P1}, we see that the images of $B_\gamma$ in $\cMot(\mathbf{k})$ are equal to $\overline B_\gamma$. \section{Stability conditions for Higgs bundles}\label{sect:Stability} \subsection{Harder--Narasimhan filtration} Recall that in Section~\ref{sect:ParWeights} we defined the set $\Stab$ of sequences of parabolic weights. To every sequence of parabolic weights we associated a stability condition on parabolic bundles in Definition~\ref{def:StabilityCond}. We want to extend this to Higgs bundles and to calculate the motivic classes of stacks of semistable parabolic Higgs bundles with nonpositive underlying vector bundles. Let~$X$ and~$D$ be as before and let $\sigma\in\Stab$. \begin{Definition}\quad \begin{enumerate}\itemsep=0pt \item[(i)] A parabolic Higgs bundle $(\mathbf E,\Phi)$ is \emph{$\sigma$-semistable} if~\eqref{eq:ss} is satisfied for all strict subbundles preserved by $\Phi$. \item[(ii)] A parabolic Higgs bundle $(\mathbf E,\Phi)=(E,E_{\bullet,\bullet},\Phi)$ is \emph{$\sigma$-nonpositive-semistable}, if the underlying vector bundle of $\mathbf E$ is nonpositive and~\eqref{eq:ss} is satisfied for all strict subbundles $\mathbf E'=(E',E'_{\bullet,\bullet})$ such that $\Phi$ preserves $\mathbf E'$ and $E/E'$ is a nonpositive vector bundle. \end{enumerate} \end{Definition} The notion of $\sigma$-nonpositive-semistable Higgs bundle is similar to that of nonnegative-semi\-stable Higgs bundle (see~\cite[Section~3.3]{FedorovSoibelmans} and~\cite{MozgovoySchiffmanOnHiggsBundles}). We emphasize that a $\sigma$-nonpositive-semistable parabolic Higgs bundle is not necessarily $\sigma$-semistable; cf.~\cite[Remark~3.3.1]{FedorovSoibelmans}. Denote the substack of $\mathcal{H}{iggs}(\zeta)$ corresponding to $\sigma$-semistable (resp.~$\sigma$-nonpositive-semi\-stable) parabolic Higgs bundles by $\mathcal{H}{iggs}^{\sigma-{\rm ss}}(\zeta)$ (resp.~$\mathcal{H}{iggs}^{\sigma-{\rm ss},-}(\zeta)$). An argument similar to~\cite[Lemma~3.7]{Simpson1} shows that these are open substacks of $\mathcal{H}{iggs}(\zeta)$ and $\mathcal{H}{iggs}^-(\zeta)$ respectively. Note that if $(\mathbf E,\Phi)$ is a parabolic Higgs bundle and $\mathbf E'\subset\mathbf E$ is a strict parabolic subbundle preserved by $\Phi$, then we get an induced Higgs field on $\mathbf E/\mathbf E'$; denote it $\Phi'$. Then $(\mathbf E/\mathbf E',\Phi')\in\mathcal{H}{iggs}^-(\zeta)$. One can use this construction to give $\mathcal{H}{iggs}^-(\zeta)$ the structure of an exact category. The proof of the following proposition is completely similar to the proof of Proposition~\ref{pr:HN}. \begin{Proposition}\label{pr:HN3}\quad \begin{enumerate}\itemsep=0pt \item[$(i)$] If $(\mathbf E,\Phi)\in\mathcal{H}{iggs}(\zeta)$ is a parabolic Higgs bundle with eigenvalues $\zeta$, then there is a unique filtration $0=\mathbf E_0\subset\mathbf E_1\subset\dots\subset\mathbf E_m=\mathbf E$ by strict parabolic subbundles preserved by $\Phi$ such that all the quotients $\mathbf E_i/\mathbf E_{i-1}$ with induced Higgs fields are $\sigma$-semistable parabolic Higgs bundles and we have $\tau_1>\dots>\tau_m$, where $\tau_i$ is the $(1,\sigma)$-slope of $\mathbf E_i/\mathbf E_{i-1}$. \item[$(ii)$] If $(\mathbf E,\Phi)\in\mathcal{H}{iggs}^-(\zeta)$, then there is a unique filtration of $\mathbf E$ as in~(i) by strict parabolic subbundles preserved by $\Phi$ with quotients being $\sigma$-nonpositive-semistable parabolic Higgs bundles. \end{enumerate} \end{Proposition} \subsection{Kontsevich--Soibelman factorization formula}\label{sect:KS} The general formalism of~\cite{KontsevichSoibelman08} implies the following factorization formula valid in $\Mot(\mathbf{k})[[\Gamma'_+]]$. One can also give a direct proof along the lines of the proof of~\cite[Proposition~3.6.1]{FedorovSoibelmans}\footnote{Note that all but countably many multiples are equal to one. We can understand the countable product as a~clockwise product as in~\cite{KontsevichSoibelman08,KontsevichSoibelman10}. Note, however, that this is a product in a commutative ring.} \begin{equation}\label{eq:KS} \sum_{\gamma\in\Gamma'_+}\mathbb{L}^{-\chi(\gamma)}[\mathcal{H}{iggs}^-_\gamma(\zeta)]e_\gamma= \prod_{\tau\in\mathbb{R}}\left( \sum_{\substack{\gamma\in\Gamma'_+\\ \deg_{1,\sigma}\gamma=\tau\rk\gamma}}\mathbb{L}^{-\chi(\gamma)}\big[\mathcal{H}{iggs}^{\sigma-{\rm ss},-}_\gamma(\zeta)\big]e_\gamma \right). \end{equation} Now, taking the plethystic logarithms of both sides and using Corollary~\ref{cor:MAIN}, we get the following statement. \begin{Proposition}\label{pr:expl-ss>=0} We have in $\cMot(\mathbf{k})[[\Gamma'_+]]$ \[ \sum_{\substack{\gamma\in\Gamma'_+\\ \deg_{1,\sigma}\gamma=\tau\rk\gamma}}\mathbb{L}^{-\chi(\gamma)}\big[\mathcal{H}{iggs}^{\sigma-{\rm ss},-}_\gamma(\zeta)\big]e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma'_+\\ \deg_{0,\zeta}\gamma=0\\ \deg_{1,\sigma}\gamma=\tau\rk\gamma}}\overline B_\gamma e_\gamma\right). \] If $X=\P^1$, then the same formula holds in $\Mot(\mathbf{k})[[\Gamma'_+]]$ with $\overline B_\gamma$ replaced by $B_\gamma$, where $B_\gamma$ are defined by~\eqref{eq:DT_P1}. \end{Proposition} \section{Stabilization}\label{sect:Stabilization} \subsection{Stabilization of semistable Higgs bundles} Let $X$ and $D$ be as before. We will be assuming that $D\ne\varnothing$. Note that this implies that $X$ has a $\mathbf{k}$-rational divisor of degree one. Set $\delta:=\max(2g-2+\deg D,0)$. Fix a stability condition $\sigma\in\Stab$. Our goal in this section is to calculate the motivic class of the moduli stack of $\sigma$-semistable parabolic Higgs bundles without nonnegativity assumption. The main result in this section is Theorem~\ref{th:ExplAnsw}. Recall that in the end of Section~\ref{sect:ParWeights} we defined the categories $\mathcal{B}{un}^{{\rm par},\le\tau}$ and $\mathcal{B}{un}^{{\rm par},\ge\tau}$. These are the full subcategories of $\mathcal{B}{un}^{\rm par}$ whose objects are parabolic bundles with the $\sigma$-HN spectrum contained in $(-\infty,\tau]$ and~$[\tau,\infty)$ respectively. We start with the following analogue of~\cite[Lemma~3.1]{MozgovoySchiffmanOnHiggsBundles}. \begin{Lemma}\label{lm:gap} Let $\mathbf E\in\mathcal{B}{un}^{\rm par}$ be a parabolic bundle with the $\sigma$-HN-spectrum $\tau_1>\tau_2>\dots>\tau_m$. Assume that for some $i\in\{1,\dots,m-1\}$ we have $\tau_i-\tau_{i+1}>\delta$. Then there are no $\sigma$-semistable Higgs bundles of the form $(\mathbf E,\Phi)$. \end{Lemma} \begin{proof} Assume the contrary. We have an exact sequence \[ 0\to\mathbf E^{\ge}\to\mathbf E\to\mathbf E^{\le}\to0, \] where $\mathbf E^{\ge}\in\mathcal{B}{un}^{{\rm par},\ge\tau_i}$, $\mathbf E^{\le}\in\mathcal{B}{un}^{{\rm par},\le\tau_{i+1}}$. Then $\Phi$ induces a morphism $\Phi':\mathbf E^{\ge}\to\mathbf E^{\le}\otimes\Omega_X(D)$. Note that $\mathbf E^{\le}\otimes\Omega_X(D)\in\mathcal{B}{un}^{{\rm par},\le\tau_{i+1}+\delta}$. By Lemma~\ref{lm:NoMorphismSS}, $\Phi'=0$ and we see that $\mathbf E^{\ge}$ is preserved by $\Phi$ contradicting $\sigma$-semistability of $(\mathbf E,\Phi)$. \end{proof} Next, we have an analogue of~\cite[Lemma~3.2]{MozgovoySchiffmanOnHiggsBundles}. \begin{Lemma}\label{lm:MozSchif} Let $(\mathbf E,\Phi)$ be a $\sigma$-semistable Higgs bundle. Assume that $\deg_{1,\sigma}\mathbf E<-\frac{r(r-1)}2\delta$, where $r=\rk\mathbf E$. Then $\mathbf E\in\mathcal{B}{un}^{{\rm par},\le0}$. \end{Lemma} \begin{proof} Let $\tau_1>\tau_2>\dots>\tau_m$ be the $\sigma$-HN-spectrum of $\mathbf E$. Denote by $r_i$ the jumps of the ranks of $\sigma$-HN-filtration. By Lemma~\ref{lm:gap} we have $\tau_i\ge\tau_1-(i-1)\delta$. We have \[ -\frac{r-1}2\delta>\frac{\deg_{1,\sigma}\mathbf E}r=\frac{\sum\limits_{i=1}^m\tau_ir_i}r\ge\frac{\sum\limits_{i=1}^m(\tau_1-(i-1)\delta)r_i}r\ge \frac mr\tau_1-\frac{r-1}2\delta \] and the statement follows. \end{proof} Set $|\sigma|:=\sum\limits_{x\in D}(\sup_i\sigma_{x,i}-\sigma_{x,1})$. We remark that we always have $|\sigma|\le\deg D$. We have an analogue of~\cite[Corollary~3.3]{MozgovoySchiffmanOnHiggsBundles}. \begin{Lemma}\label{lm:ss=ss>=0} Let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$ be such that $d<-r|\sigma|-\frac{r(r-1)}2\delta$. Then \[ \mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)=\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss},-}(\zeta). \] \end{Lemma} \begin{proof} For all $x$ and $i$ replace $\sigma_{x,i}$ by $\sigma_{x,i}-\sigma_{x,1}$. This does not change $|\sigma|$ and the notion of semistability but we now have $\sigma_{x,i}\ge0$ for all $x$ and $i$. Next \begin{equation}\label{eq:|sigma|} \deg_{1,\sigma}\gamma=d+\sum_x\sum_{i=1}^\infty\sigma_{x,i}r_{x,i}\le d+ \sum_x\Big(\sup_i\sigma_{x,i}\Big)\left(\sum_{i=1}^\infty r_{x,i}\right)=d+r|\sigma|. \end{equation} Let $(\mathbf E,\Phi)\in\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)$. By~\eqref{eq:|sigma|} we have{\samepage \[ \deg_{1,\sigma}\gamma<-\frac{r(r-1)}2\delta. \] By Lemma~\ref{lm:MozSchif} we have $\mathbf E\in\mathcal{B}{un}^{{\rm par},\le0}\subset\mathcal{B}{un}^{{\rm par},-}$ (the last inclusion follows from $\sigma_{x,i}\ge0$).} Conversely, assume that $(\mathbf E,\Phi)\in\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss},-}(\zeta)$. Assume for contradiction that $(\mathbf E,\Phi)$ is not $\sigma$-semistable. Then by Proposition~\ref{pr:HN3}(i) we have an exact sequence $0\to\mathbf E'\to\mathbf E\to\mathbf E''\to0$ in~$\mathcal{B}{un}^{\rm par}$ such that~$\Phi$ preserves $\mathbf E'$, and $(\mathbf E'',\Phi'')$ is $\sigma$-semistable, where $\Phi''$ is the induced Higgs field. Using~\eqref{eq:|sigma|} we get \[ \frac{\deg_{1,\sigma}\mathbf E''}{\rk\mathbf E''}<\frac{\deg_{1,\sigma}\mathbf E}{\rk\mathbf E}\le\frac{d+r|\sigma|}r-\frac{r-1}2\delta\le -\frac{\rk\mathbf E''-1}2\delta. \] Now it follows from Lemma~\ref{lm:MozSchif} that $\mathbf E''\in\mathcal{B}{un}^{{\rm par},\le0}$. \looseness=1 Write $\mathbf E''=(E'',E''_{\bullet,\bullet})$. Since $(\mathbf E,\Phi)$ is $\sigma$-nonpositive-semistable, $E''$ cannot be nonpositive. Thus there is $E'''\subset E''$ with $\deg E'''>0$. Let $\mathbf E'''=(E''',E'''_{\bullet,\bullet})$ be the corresponding parabolic subbundle of $\mathbf E''$. Then $\deg_{1,\sigma}\mathbf E'''\ge\deg E'''>0$, which gives contradiction with $\mathbf E''\in\mathcal{B}{un}^{{\rm par},\le0}$. \end{proof} Set $\mathbf1=(0,0_{\bullet,\bullet},1)\in\Gamma$ (here $0_{\bullet,\bullet}$ is the sequence of zeroes indexed by $D\times\mathbb{Z}_{>0}$). If $\gamma\in\Gamma_+$ and $\gamma\ne0$, then $\gamma+N\mathbf1\in\Gamma_+$ for all $N\in\mathbb{Z}$. \begin{Corollary}\label{cor:Stabilization} Let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$, $\gamma\ne0$, and $N>|\sigma|+\frac{r-1}2\delta+d/r$. Then \[ \mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\simeq\mathcal{H}{iggs}_{\gamma-Nr\mathbf1}^{\sigma-{\rm ss},-}(\zeta). \] \end{Corollary} \begin{proof} Since $X$ has a divisor of degree one, it has a line bundle of degree $N$. Tensorisation with this line bundle gives $\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\simeq\mathcal{H}{iggs}_{\gamma-Nr\mathbf1}^{\sigma-{\rm ss}}(\zeta)$. Now Lemma~\ref{lm:ss=ss>=0} completes the proof. \end{proof} Recall from Definition~\ref{def:DT} the Donaldson--Thomas invariants $\overline B_\gamma\in\cMot(\mathbf{k})$. For each $\tau\in\mathbb{R}$ define the elements $H_\gamma(\zeta,\sigma)\in\cMot(\mathbf{k})$, where $\gamma\in\Gamma_+$ is such that the $(1,\sigma)$-slope of $\gamma$ is $\tau$ (or $\gamma=0$), by the following formula. \begin{equation}\label{eq:ExplAnswer} \sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{0,\zeta}\gamma=0\\ \deg_{1,\sigma}\gamma=\tau\rk\gamma}}\mathbb{L}^{-\chi(\gamma)}H_\gamma(\zeta,\sigma)e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{0,\zeta}\gamma=0\\ \deg_{1,\sigma}\gamma=\tau\rk\gamma}} \overline B_\gamma e_\gamma \right). \end{equation} Thus $H_\gamma(\zeta,\sigma)$ is defined for all $\gamma$ such that $\deg_{0,\zeta}\gamma=0$. Note that $H_0(\zeta,\sigma)=1$. Now we can formulate our first main result. \begin{Theorem}\label{th:ExplAnsw} Let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$, $\gamma\ne0$. \begin{enumerate}\itemsep=0pt \item[$(i)$] The elements $H_\gamma(\zeta,\sigma)$ are periodic in the following sense: for $d<-|\sigma|-\frac{r-1}2\delta$ we have $H_\gamma(\zeta,\sigma)=H_{\gamma-r\mathbf1}(\zeta,\sigma)$. \item[$(ii)$] The stack $\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)$ is of finite type and we have in $\cMot(\mathbf{k})$ \begin{equation}\label{eq:ThExpl1} \big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]=H_{\gamma-Nr\mathbf1}(\zeta,\sigma) \end{equation} whenever $N$ is large enough, provided that $\deg_{0,\zeta}\gamma=0$ {\rm(}it suffices to take $N>|\sigma|+\frac{r-1}2\delta+d/r${\rm)}. If $\deg_{0,\zeta}\gamma\ne0$, then the stack is empty. \end{enumerate} \end{Theorem} \begin{proof} For part~(ii) combine Corollary~\ref{cor:Stabilization} with Proposition~\ref{pr:expl-ss>=0}. Part~(i) is clear from \linebreak part~(ii). \end{proof} An immediate corollary of the above theorem and formula~\eqref{eq:ExplAnswer} is the following curious observation. \begin{Corollary}\label{cor:EqualMot} Assume that we are given $\gamma\in\Gamma_+$, sets of eigenvalues $\zeta$ and $\zeta'$, and sequences of parabolic weights $\sigma$, $\sigma'$. Let $\tau$ and $\tau'$ be $(1,\sigma)$ and $(1,\sigma')$-slopes of $\gamma$ respectively. Assume also that \begin{gather*} \{\gamma'\in\Gamma_+\colon \deg_{0,\zeta}\gamma'=0, \deg_{1,\sigma}\gamma'=\tau\rk\gamma\}= \{\gamma'\in\Gamma_+\colon \deg_{0,\zeta'}\gamma'=0, \deg_{1,\sigma'}\gamma'=\tau'\rk\gamma\}.\! \end{gather*} Then we have an equality of motivic classes \[ \big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]=\big[\mathcal{H}{iggs}_\gamma^{\sigma'-{\rm ss}}(\zeta')\big]. \] \end{Corollary} \subsection[Case of $\P^1$]{Case of $\boldsymbol{\P^1}$}\label{sect:StabP1} If $X=\P^1$, we obtain simpler and more precise results. Namely, if we define elements $H_\gamma(\zeta,\sigma)$ by the same formula~\eqref{eq:ExplAnswer} but with $B_\gamma$ instead of $\overline B_\gamma$, then~\eqref{eq:ThExpl1} holds in $\Mot(\mathbf{k})$. \section{Motivic classes of parabolic connections}\label{sect:Conn} \subsection{Stacks of parabolic connections} Let $X$ and $D$ be as above. Our goal in this section is to calculate the motivic classes of the moduli stacks of parabolic bundles with connections with prescribed eigenvalues of residues. In Section~\ref{sect:StabConn} we put stability conditions on these moduli stacks and calculate the motivic classes of substacks of semistable parabolic bundles with connections. Our argument is similar to the argument for Higgs bundles. Let $E$ be a vector bundle on $X$. A \emph{connection} on $E$ with \emph{poles bounded by $D$} is a morphism of sheaves of abelian groups $\nabla\colon E\to E\otimes\Omega_X(D)$ satisfying Leibniz rule. In this case for $x\in D$ one defines the residue of the connection $\res_x\nabla\in\End(E_x)$. \begin{Definition} A \emph{parabolic connection} of type $(X,D)$ is a triple $(E,E_{\bullet,\bullet},\nabla)$, where $(E,E_{\bullet,\bullet})$ is a point of $\mathcal{B}{un}^{\rm par}(X,D)$, $\nabla\colon E\to E\otimes\Omega_X(D)$ is a connection on $E$ such that for all $x\in D$ and $j\ge0$ we have $(\res_x\nabla)(E_{x,j})\subset E_{x,j}$. \end{Definition} We denote the category (and the Artin stack) of parabolic connections by $\mathcal{C}{onn}\!=\!\mathcal{C}{onn}(X,D)$. In this section $X$ and $D$ are fixed, so we skip them from the notation. Recall that $\mathbf{k}[D\times\mathbb{Z}_{>0}]$ is the set of all $\mathbf{k}$-valued sequences $\zeta=\zeta_{\bullet,\bullet}=(\zeta_{x,j})$ indexed by $D\times\mathbb{Z}_{>0}$ such that $\zeta_{x,j}=0$ for $j\gg0$. For $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ let $\mathcal{C}{onn}(\zeta)=\mathcal{C}{onn}(X,D,\zeta)$ denote the full subcategory of $\mathcal{C}{onn}$ (and its stack of objects) corresponding to collections $(E,E_{\bullet,\bullet},\nabla)$ such that $(\res_x\nabla-\zeta_{x,j}1)(E_{x,{j-1}})\subset E_{x,j}$ for all $x\in D$ and $j>0$. We call the points of $\mathcal{C}{onn}(\zeta)$ the parabolic bundles with connections \emph{with eigenvalues $\zeta$.} Let $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ and let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$. Recall from Section~\ref{sect:DegreeSlope} that \[ \deg_{1,\zeta}\gamma=d+\sum_{x\in D}\sum_{j=1}^{\infty}\zeta_{x,j}r_{x,j}\in\mathbf{k}. \] The following lemma is~\cite[Theorem~7.1]{Crawley-Boevey:Indecomposable} if $\mathbf{k}=\mathbb C$. The proof in the general case is completely similar. \begin{Lemma}\label{lm:existence2} Let $\mathbf E\in\mathcal{B}{un}^{\rm par}(\mathbf{k})$ and $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$. There exists an object $(\mathbf E,\nabla)\in\mathcal{C}{onn}(\zeta)(\mathbf{k})$ if and only if $\deg_{1,\zeta}\mathbf E'=0$ for any direct summand $\mathbf E'$ of $\mathbf E$. \end{Lemma} Note that, in particular, for every $(\mathbf E,\nabla)\in\mathcal{C}{onn}(\zeta)(\mathbf{k})$ we have $\deg_{1,\zeta}\mathbf E=0$. Recall from Section~\ref{sect:Isoslopy} the notion of $(\kappa,\zeta)$-isoslopy parabolic bundle and the stacks $\mathcal{P}{air}^{(\kappa,\zeta)-{\rm iso}}_\gamma$. Recall also that for $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$, we have set $\chi(\gamma):=(g-1)r^2+\sum\limits_{x\in D}\sum\limits_{j<j'}r_{x,j}r_{x,j'}$. \begin{Proposition}\label{pr:Sasha2} We have \[ [\mathcal{C}{onn}_\gamma(\zeta)]= \begin{cases} \mathbb{L}^{\chi(\gamma)}\big[\mathcal{P}{air}^{(1,\zeta)-{\rm iso}}_\gamma\big]& \text{if }\deg_{1,\zeta}\gamma=0,\\ 0 & \text{otherwise}. \end{cases} \] \end{Proposition} \begin{proof} The proof is completely analogous to the proof of Proposition~\ref{pr:Sasha} with Lemma~\ref{lm:existence} replaced by Lemma~\ref{lm:existence2}. \end{proof} \subsection{Stabilization of isoslopy parabolic bundles}\label{sect:Stabilization2} As in Section~\ref{sect:Stabilization} we will be assuming that $D\ne\varnothing$. Recall that this implies that $X$ has a $\mathbf{k}$-rational divisor of degree one. As before, set $\delta:=\max(2g-2+\deg D,0)$. Recall that every vector bundle $E$ on $X$ has a unique HN-filtration and the slopes of the quotients form a sequence called the HN-spectrum of $E$. We start with the following analogue of~\cite[Lemma~4.1]{MozgovoySchiffmanOnHiggsBundles}. \begin{Lemma}\label{lm:gap2} Let $\mathbf E=(E,E_{\bullet,\bullet})\in\mathcal{B}{un}^{\rm par}$ be a parabolic bundle such that $E$ has HN-spectrum $\tau_1>\tau_2>\dots>\tau_m$. Assume that for some $i\in\{1,\dots,m-1\}$ we have $\tau_i-\tau_{i+1}>\delta$. Then $\mathbf E$ is decomposable. \end{Lemma} \begin{proof}One shows that the extensions of a parabolic bundle $\mathbf E''$ by a parabolic bundle $\mathbf E'$ (in the sense of Section~\ref{sect:Subobjects}) are classified by a vector space $\Ext^1(\mathbf E'',\mathbf E')$ dual to $\Hom(\mathbf E',\mathbf E''(\Omega_X(D))$. Let $0=E_0\subset E_1\subset\dots\subset E_m=E$ be the Harder--Narasimhan filtration of $E$. Let $\mathbf E_i$ be the strict parabolic subbundle with the underlying vector bundle $E_i$. We have an exact sequence $0\to\mathbf E_i\to\mathbf E\to\mathbf E/\mathbf E_i\to0$. Note that by the assumption the Harder--Narasimhan spectrum of $E_i$ is contained in $[\tau_i,\infty)$, while the Harder--Narasimhan spectrum of $(E/E_i)(\Omega_X(D))$ is contained in $(-\infty,\tau_i)$. It follows that $\Hom\bigl(E_i,(E/E_i)(\Omega_X(D))\bigr)=0$. Thus \[ \Ext^1(\mathbf E/\mathbf E_i,\mathbf E_i)=\Hom\bigl(\mathbf E_i,(\mathbf E/\mathbf E_i)(\Omega_X(D))\bigr)^\vee=0. \] Thus $\mathbf E\simeq\mathbf E_i\oplus(\mathbf E/\mathbf E_i)$ is decomposable. \end{proof} Next, we have an analogue of~\cite[Corollary~4.2]{MozgovoySchiffmanOnHiggsBundles} whose proof is similar to loc.~cit.~and to that of Lemma~\ref{lm:MozSchif}. \begin{Lemma}\label{lm:MozSchif2} Let $\mathbf E\in\mathcal{B}{un}^{\rm par}$ be indecomposable and $\cl(\mathbf E)=(r,r_{\bullet,\bullet},d)$. Assume that $d<-\frac{r(r-1)}2\delta$. Then $\mathbf E\in\mathcal{B}{un}^{{\rm par},-}$. \end{Lemma} \begin{proof} Write $\mathbf E=(E,E_{\bullet,\bullet})$. Let $\tau_1>\tau_2>\dots>\tau_m$ be the HN-spectrum of $E$. Denote by $r_i$ the jumps of the ranks of HN-filtration. By Lemma~\ref{lm:gap} we have $\tau_i\ge\tau_1-(i-1)\delta$. We have \[ -\frac{r-1}2\delta>\frac dr=\frac{\sum\limits_{i=1}^m\tau_ir_i}r\ge\frac{\sum\limits_{i=1}^m(\tau_1-(i-1)\delta)r_i}r\ge \frac mr\tau_1-\frac{r-1}2\delta \] and the statement follows. \end{proof} Fix $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$. Let $|\bullet|$ be any norm on the $\mathbb{Q}$-vector subspace of $\mathbf{k}$ generated by the components of~$\zeta$. If~$\mathbf{k}$ is embedded into $\mathbb C$, we can take the usual absolute value for $|\bullet|$. We set $|\zeta|:=\sum\limits_{x\in D}(\max_i|\zeta_{x,i}|)$. We have an analogue of~\cite[Lemma~3.2.3(i)]{FedorovSoibelmans} (cf.~also Lemma~\ref{lm:ss=ss>=0}). \begin{Lemma}\label{lm:ss=ss>=02} Let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$ be such that $d<-2r|\zeta|-\frac{r(r-1)}2\delta$. Then \[ \mathcal{B}{un}_\gamma^{{\rm par},(1,\zeta)-{\rm iso}}=\mathcal{B}{un}_\gamma^{{\rm par},(1,\zeta)-{\rm iso},-}. \] \end{Lemma} \begin{proof} Take $\mathbf E=(E,E_{\bullet,\bullet})\in\mathcal{B}{un}_\gamma^{{\rm par},(1,\zeta)-{\rm iso}}$. Assume for a contradiction that $\mathbf E\notin\mathcal{B}{un}^{{\rm par},-}$. By Lemma~\ref{lm:MozSchif2}, $\mathbf E$ is decomposable. Let $\mathbf E'$ be an indecomposable summand of $\mathbf E$ such that $\mathbf E'\notin\mathcal{B}{un}^{{\rm par},-}$. By the definition of isoslopy bundles, we have $\frac{\deg_{1,\zeta}\mathbf E'}{\rk\mathbf E'}=\frac{\deg_{1,\zeta}\mathbf E}r$. Write $\cl(\mathbf E')=(r',r'_{\bullet,\bullet},d')$. We have \[ \frac{d'}{r'}\le\frac{\deg_{1,\zeta}\mathbf E'}{\rk\mathbf E'}+|\zeta|=\frac{\deg_{1,\zeta}\mathbf E}r+|\zeta|\le\frac dr+2|\zeta|< -\frac{(r-1)}2\delta\le-\frac{(r'-1)}2\delta \] and Lemma~\ref{lm:MozSchif2} gives a contradiction. \end{proof} Recall that we have $\mathbf1=(0,0_{\bullet,\bullet},1)\in\Gamma$. \begin{Corollary}\label{cor:ExplAnswer2} Let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$, $\gamma\ne0$, and $N>2|\zeta|+\frac{r-1}2\delta+d/r$. Then $\mathcal{P}{air}_\gamma^{(1,\zeta)-{\rm iso}}\simeq\mathcal{P}{air}_{\gamma-Nr\mathbf1}^{(1,\zeta)-{\rm iso},-}$. In particular, $\mathcal{P}{air}_\gamma^{(1,\zeta)-{\rm iso}}$ is a constructible subset of $\mathcal{P}{air}_\gamma$ of finite type. \end{Corollary} \begin{proof} Since $X$ has a divisor of degree one, it has a line bundle of degree $N$. Tensorisation with this line bundle gives $\mathcal{P}{air}_\gamma^{(1,\zeta)-{\rm iso}}\simeq\mathcal{P}{air}_{\gamma-Nr\mathbf1}^{(1,\zeta)-{\rm iso}}$. Now Lemma~\ref{lm:ss=ss>=02} completes the proof. \end{proof} Define the elements $C_\gamma(\zeta)\in\cMot(\mathbf{k})$, where $\gamma$ ranges over $\Gamma_+$ by the following formula (cf.~\eqref{eq:ExplAnswer}) \begin{equation}\label{eq:ExplAnswer2} \sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{1,\zeta}\gamma=\tau\rk\gamma}}\mathbb{L}^{-\chi(\gamma)}C_\gamma(\zeta)e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma_+'\\ \deg_{1,\zeta}\gamma=\tau\rk\gamma}} \overline B_\gamma e_\gamma \right). \end{equation} Now we can formulate our second main result. Recall that $\mathcal{C}{onn}(\zeta)=\varnothing$ unless $\deg_{1,\zeta}\gamma=0$. \begin{Theorem}\label{th:ExplAnsw2} Let $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma_+$, $\gamma\ne0$ and $\deg_{1,\zeta}\gamma=0$. Let $\zeta$ be an element of $\mathbf{k}[D\times\mathbb{Z}_{>0}]$ and let $|\bullet|$ be a norm on the $\mathbb{Q}$-vector subspace of $\mathbf{k}$ generated by the components of~$\zeta$. Then \begin{enumerate}\itemsep=0pt \item[$(i)$] The elements $C_\gamma(\zeta)$ are periodic in the following sense: for $d<-2|\zeta|-\frac{r-1}2\delta$ we have $C_\gamma(\zeta)=C_{\gamma-r\mathbf1}(\zeta)$. \item[$(ii)$] The stack $\mathcal{C}{onn}_\gamma(\zeta)$ is of finite type and we have \[ [\mathcal{C}{onn}_\gamma(\zeta)]=C_{\gamma-Nr\mathbf1}(\zeta), \] whenever $N$ is large enough {\rm(}it suffices to take $N>2|\zeta|+\frac{r-1}2\delta+d/r${\rm)}. \end{enumerate} \end{Theorem} \begin{proof} Combine Proposition~\ref{pr:Sasha2}, Corollary~\ref{cor:ExplAnswer2}, and Corollary~\ref{cor:isoslopy}. \end{proof} \subsubsection[Case of $\P^1$]{Case of $\boldsymbol{\P^1}$}\label{sect:StabP12} If $X=\P^1$, we obtain simpler and more precise results. Define $B_\gamma\in\Mot(\mathbf{k})$ by~\eqref{eq:DT_P1}. Define $C_\gamma(\zeta)\in\Mot(\mathbf{k})$ by the same formula~\eqref{eq:ExplAnswer2} but with $\overline B_\gamma$ replaced by $B_\gamma$. Then Theorem~\ref{th:ExplAnsw2} holds in $\Mot(\mathbf{k})$. \subsection{Stability conditions for bundles with connections}\label{sect:StabConn} Recall that in Definition~\ref{def:StabilityCond} we defined the notion of a sequence of parabolic weights. For non-resonant connections one can work with more general sequences of parabolic weights. Let us give the definitions. \begin{Definition} We say that $\zeta=\zeta_{\bullet,\bullet}\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ is \emph{non-resonant} if for all $x\in X$ and all $i,j>0$ we have $\zeta_{x,i}-\zeta_{x,j}\notin\mathbb{Z}_{\ne0}$. \end{Definition} The importance of this definition is in the following lemma. \begin{Lemma}\label{lm:non-resonant} Let $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ be non-resonant and let $\phi$ be a morphism in $\mathcal{C}{onn}(\zeta)$ such that the underlying morphism of vector bundles is generically an isomorphism. Then the underlying morphism of vector bundles is an isomorphism. \end{Lemma} \begin{proof} One easily reduces to the case $\mathbf{k}=\mathbb C$. Take $x\in D$. Since $\zeta$ is non-resonant, one can find a subset $\Omega$ of $\mathbb C$ containing $\{\zeta_{x,j}|j>0\}$ and such that the exponential function induces a bijection between $\Omega$ and $\mathbb C-0$. Then it is well-known that every regular connection on the punctured formal disc has a unique extension to the puncture such that the eigenvalues of the residues are in $\Omega$. The statement follows. \end{proof} Define the space of extended sequences of parabolic weights $\Stab'$ as the set of pairs $(\kappa,\sigma)$, where $\kappa\in\mathbb{R}_{\ge0}$, $\sigma=\sigma_{\bullet,\bullet}$ is a sequence of real numbers, indexed by $D\times\mathbb{Z}_{>0}$, such that for all $x\in D$ we have~\eqref{eq:StabCond2}. \begin{Definition} Let $(\kappa,\sigma)\in\Stab'$. A parabolic connection $(\mathbf E,\nabla)$ is \emph{$(\kappa,\sigma)$-semistable} if for all strict parabolic subbundles $\mathbf E'\subset\mathbf E$ preserved by $\nabla$ we have \begin{equation* \frac{\deg_{\kappa,\sigma}\mathbf E'}{\rk\mathbf E'}\le\frac{\deg_{\kappa,\sigma}\mathbf E}{\rk\mathbf E}. \end{equation*} \end{Definition} We denote by $\mathcal{C}{onn}^{(\kappa,\sigma)-{\rm ss}}(\zeta)$ the substack of $\mathcal{C}{onn}(\zeta)$ corresponding to $(\kappa,\sigma)$-semistable connections. An argument similar to~\cite[Lemma~3.7]{Simpson1} shows that this is an open substacks of $\mathcal{C}{onn}(\zeta)$. \begin{Proposition}\label{pr:HN2} Assume that $(\kappa,\sigma)\in\Stab'$. Assume also that either $\zeta$ is non-resonant, or $\kappa=1$, $\sigma\in\Stab$. If $(\mathbf E,\nabla)\in\mathcal{C}{onn}(\zeta)$, then there is a unique filtration $0=\mathbf E_0\subset\mathbf E_1\subset\dots\subset\mathbf E_m=\mathbf E$ by strict parabolic subbundles preserved by $\nabla$ such that all the quotients $\mathbf E_i/\mathbf E_{i-1}$ with induced connections are $(\kappa,\sigma)$-semistable parabolic bundles with connections and we have $\tau_1>\dots>\tau_m$, where $\tau_i$ is the $(\kappa,\sigma)$-slope of $\mathbf E_i/\mathbf E_{i-1}$. \end{Proposition} \begin{proof} In the non-resonant case the proof is completely analogous to the proof of~\cite[Section~1.3]{HarderNarasimhan} in view of Lemma~\ref{lm:MorPar} and the following lemma. \begin{Lemma} Let $\zeta$ be non-resonant and let $\mathbf E\to\mathbf F$ be a morphism in $\mathcal{C}{onn}(\zeta)$, which is generically an isomorphism. Then $\deg_{\kappa,\sigma}\mathbf E\le\deg_{\kappa,\sigma}\mathbf F$. \end{Lemma} \begin{proof} Write $\mathbf E=(E,E_{\bullet,\bullet})$ and $\mathbf F=(F,F_{\bullet,\bullet})$. Let $\phi\colon E\to F$ be the underlying morphism of vector bundles. By Lemma~\ref{lm:non-resonant}, $\phi$ is an isomorphism. Thus $\dim E_{x,j}\le\dim F_{x,j}$ for all~$x$ and~$j$. Therefore \begin{align*} \deg_{\kappa,\sigma}\mathbf E& =\kappa\deg E+\sum_{x,j>0}\sigma_{x,j}(\dim E_{x,j-1}-\dim E_{x,j}) \\ & = \kappa\deg E+\sum_{x\in D}\left(\sigma_{x,1}\rk E+\sum_{i>0}(\sigma_{x,i+1}-\sigma_{x,i})\dim E_{x,i}\right)\\ & \le \kappa\deg F+\sum_{x\in D}\left(\sigma_{x,1}\rk F+\sum_{i>0}(\sigma_{x,i+1}-\sigma_{x,i})\dim F_{x,i}\right)=\deg_{\kappa,\sigma}\mathbf F.\tag*{\qed} \end{align*}\renewcommand{\qed}{} \end{proof} In the resonant case, the proof is completely analogous to the proof of~\cite[Section~1.3]{HarderNarasimhan} in view of Lemma~\ref{lm:ModifDegree} (cf.\ Propositions~\ref{pr:HN} and~\ref{pr:HN3}). \end{proof} \begin{Remark}\label{rm:resonant} More generally, If $\zeta$ is resonant, one can work with any $(\kappa,\sigma)\in\Stab'$ such that $\sigma_{x,j}-\sigma_{x,1}\le\kappa$ for all $x$ and $j$. However, the notion of stability does not change if we scale~$(\kappa,\sigma)$. Thus, we can always assume that $\kappa=1$, in which case $\sigma\in\Stab$, or $\kappa=0$, in which case $\sigma=0$. The latter case corresponds to the trivial stability condition; the corresponding motivic class has been calculated in Theorem~\ref{th:ExplAnsw2}. \end{Remark} Similarly to Proposition~\ref{pr:expl-ss>=0} the Kontsevich--Soibelman factorization formula implies the following proposition. \begin{Proposition}\label{pr:expl-ss>=02} Let $(\kappa,\sigma)\in\Stab'$. Assume that either $\zeta$ is non-resonant or $\kappa=1$ and $\sigma\in\Stab$. Then we have in $\Mot(\mathbf{k})[[\Gamma_+]]$ \begin{equation}\label{eq:KS2} \sum_{\gamma\in\Gamma_+}\mathbb{L}^{-\chi(\gamma)}[\mathcal{C}{onn}_\gamma(\zeta)]e_\gamma= \prod_{\tau\in\mathbb{R}}\left( \sum_{\substack{\gamma\in\Gamma_+\\ \deg_{\kappa,\sigma}\gamma=\tau\rk\gamma}}\mathbb{L}^{-\chi(\gamma)}\big[\mathcal{C}{onn}^{(\kappa,\sigma)-{\rm ss}}_\gamma(\zeta)\big]e_\gamma \right). \end{equation} \end{Proposition} Define the elements $C_\gamma(\zeta,\kappa,\sigma)\in\cMot(\mathbf{k})$, where $\gamma$ ranges over $\Gamma_+$ by the following formula (cf.~\eqref{eq:ExplAnswer}) valid for any $\tau\in\mathbf{k}$, $\tau'\in\mathbb{R}$ \begin{equation}\label{eq:ExplAnswer3} \sum_{\substack{\gamma\in\Gamma_+\\ \deg_{1,\zeta}\gamma=\tau\rk\gamma\\ \deg_{\kappa,\sigma}\gamma=\tau'\rk\gamma }}\mathbb{L}^{-\chi(\gamma)}C_\gamma(\zeta,\kappa,\sigma)e_\gamma= \Exp\left(\sum_{\substack{\gamma\in\Gamma_+\\ \deg_{1,\zeta}\gamma=\tau\rk\gamma\\ \deg_{\kappa,\sigma}\gamma=\tau'\rk\gamma }} \overline B_\gamma e_\gamma \right). \end{equation} Note that we have $C_\gamma(\zeta,0,0)=C_\gamma(\zeta)$, where $C_\gamma(\zeta)$ are defined in~\eqref{eq:ExplAnswer2}. Now we can formulate our third main result. \begin{Theorem}\label{th:ExplAnsw3} Assume that $\gamma\in\Gamma_+$, $\gamma\ne0$, and $\deg_{1,\zeta}\gamma=0$. Then for $N>3|\zeta|+(\rk\gamma-1)\delta/2$ we have \[ \big[\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]=C_{\gamma-Nr\mathbf1}(\zeta,\kappa,\sigma). \] \end{Theorem} We note that if $\kappa=0$, and $\sigma=0$, then this theorem is essentially Theorem~\ref{th:ExplAnsw2}. \begin{proof} We defined $C_\gamma(\zeta,\kappa,\sigma)$ when $\gamma\in\Gamma_+'$. It will be convenient for us to extend this notation to the case when $\gamma$ is a multiple of $\mathbf1$ by setting $C_{n\mathbf1}(\zeta,\kappa,\sigma)=1$. We will prove that $\big[\mathcal{C}{onn}_\beta^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]=C_{\beta-Nr\mathbf1}(\zeta,\kappa,\sigma)$ whenever $\beta\in\Gamma_+$, $\deg_{1,\zeta}\beta=0$, and $N>3|\zeta|+(\rk\beta-1)\delta/2$ by induction on $\rk\beta$. The base case of $\rk\beta=0$ is obvious. Note that replacing $\gamma$ with $\gamma-N(\rk\gamma)\mathbf1$ shifts the $(\kappa,\sigma)$-slopes by $\kappa N$. Consider the product in the RHS of~\eqref{eq:KS2} and \begin{equation}\label{eq:2} \prod_{\tau\in\mathbb{R}}\left( \sum_{\substack{\gamma\in\Gamma_+\\ \deg_{\kappa,\sigma}\gamma=\tau\rk\gamma\\ \deg_{1,\zeta}\gamma=N\rk\gamma}}\mathbb{L}^{-\chi(\gamma)}[C_{\gamma-N\rk\gamma\mathbf1}(\zeta,\kappa,\sigma)]e_\gamma \right). \end{equation} We note that this product makes sense because we set $C_{-N\rk\gamma\mathbf1}(\zeta,\kappa,\sigma)=1$ in the beginning of the proof. Note also that $\deg_{1,\zeta}\beta=0$ implies that $d/r\le|\zeta|$, where $\beta=(r,r_{\bullet,\bullet},d)$, so $N>2|\zeta|+(r-1)\delta/2+d/r$. Using~\eqref{eq:KS2} and Theorem~\ref{th:ExplAnsw2} we easily see that the coefficients of these products at $e_\beta$ are both equal to $\mathbb{L}^{-\chi(\beta)}[\mathcal{C}{onn}_\beta(\zeta)]$. Expanding the product in the RHS of~\eqref{eq:KS2}, we see that the coefficient at $e_\beta$ is equal to \[ \mathbb{L}^{-\chi(\beta)}\big[\mathcal{C}{onn}_\beta^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]+ \sum_{\beta_1,\dots,\beta_n}\prod_{i=1}^n\mathbb{L}^{-\chi(\beta_i)}\big[\mathcal{C}{onn}_{\beta_i}^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big], \] where the summation is over all decompositions of $\beta$ into the sum of $n\ge2$ non-zero elements of~$\Gamma_+$ such that $\deg_{1,\zeta}\beta_i=0$ and $\deg_{\kappa,\sigma}\beta_i=\tau\rk\beta_i$. Similarly, expanding the product~\eqref{eq:2}, we see that the coefficient at $e_\beta$ is equal to \[ \mathbb{L}^{-\chi(\beta)}[C_{\beta-N\rk\beta\mathbf1}(\zeta,\kappa,\sigma)]+ \sum_{\beta_1,\dots,\beta_n}\prod_{i=1}^n\mathbb{L}^{-\chi(\beta_i)}[C_{\beta_i-N\rk\beta_i\mathbf1}(\zeta,\kappa,\sigma)]. \] It remains to show that the respective terms in the sums are equal. Clearly, we have $N>3|\zeta|+(\rk\beta_i-1)\delta/2$ and the statement follows from the induction hypothesis. \end{proof} As usual, if $X=\P^1$ we obtain similar formulas valid in $\Mot(\mathbf{k})$ by replacing $\overline B_\gamma$ with $B_\gamma$ in~\eqref{eq:ExplAnswer3}. \section[Equalities of motivic classes and non-emptiness of moduli stacks]{Equalities of motivic classes and non-emptiness\\ of moduli stacks}\label{sect:NonEmpty} \subsection{Equalities between motivic classes of stacks} In this section we will give a criterion for non-emptiness of stacks $\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)$ or $\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta).\!$ For the stacks $\mathcal{C}{onn}_\gamma(\zeta)$ such a criterion follows easily from~\cite{Crawley-Boevey:Indecomposable,CrawleyBoeveyIndecompPar}. We reduce the case, when stability condition is present, to the case of $\mathcal{C}{onn}_\gamma(\zeta)$ using some equalities between motivic classes and Proposition~\ref{pr:NonEmpty}. We will also discuss a relation with Simpson's non-abelian Hodge theory. It follows from Theorem~\ref{th:ExplAnsw3} that the motivic class of $\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)$ depends only on the submonoid of $\Gamma_+$ given by the equations $\deg_{1,\zeta}\gamma'=0$, $\deg_{\kappa,\sigma}\gamma'=\tau\rk\gamma'$, where $\tau=\deg_{\kappa,\sigma}\gamma/\rk\gamma$. Using this fact, one can give a lot of examples of seemingly unrelated moduli stacks $\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)$ having the same motivic class. An analogous statement for moduli spaces of parabolic Higgs bundles is the content of Corollary~\ref{cor:EqualMot} above. Finally, we can get a lot of equalities between motivic classes of parabolic Higgs bundles and motivic classes of connections. In the following proposition, we show that motivic classes of the form $[\mathcal{C}{onn}_\gamma(\zeta)]$ are universal. This proposition should be compared to Proposition~\ref{cor:EqualMot}. \begin{Proposition}\label{pr:ConnUniversal} Assume that $\mathbf{k}$ is not a finite extension of $\mathbb{Q}$. Let $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$. Let $\gamma\in\Gamma_+$. Let $(\kappa,\sigma)\in\Stab'$. Assume that $\zeta$ is non-resonant or $\kappa=1$ and $\sigma\in\Stab$. Set $\tau:=\deg_{\kappa,\sigma}\gamma/\rk\gamma$. Then there is $\zeta'\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ such that \begin{equation}\label{eq:EqualMot2} \{\gamma'\in\Gamma_+\colon \deg_{1,\zeta}\gamma'=0, \deg_{\kappa,\sigma}\gamma'=\tau\rk\gamma'\}= \{\gamma'\in\Gamma_+\colon \deg_{1,\zeta'}\gamma'=0\}. \end{equation} Moreover, we have $\big[\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]=[\mathcal{C}{onn}_\gamma(\zeta')]$. \end{Proposition} \begin{proof} Choose $x\in D$ and set $\sigma'_{y,j}=\sigma_{y,j}$ if $y\ne x$, $\sigma'_{x,j}=\sigma_{x,j}-\tau$. Since for all $(r',r'_{\bullet,\bullet},d)\in\Gamma'$ we have $\sum_jr'_{x,j}=r'$, we see that $\deg_{\kappa,\sigma}\gamma'=\tau\rk\gamma$ if and only if $\deg_{\kappa,\sigma'}\gamma'=0$. Now, let $U$ be the $\mathbb{Q}$-vector subspace of $\mathbb{R}$ generated by $1$ and all the numbers $\sigma'_{y,j}$, where~$y$ ranges over~$D$ and $j$ ranges over positive integers. Similarly, let $V$ be the $\mathbb{Q}$-vector subspace of~$\mathbf{k}$ generated by all the components $\zeta_{y,j}$ of $\zeta$. Since $\mathbf{k}$ is not a~finite extension of $\mathbb{Q}$, there is a~$\mathbb{Q}$-linear embedding $U\oplus V\to\mathbf{k}$. Moreover, we may assume that $(\kappa,1)$ maps to $1$. This follows from a general fact: if $L_1$ is a finite dimensional $\mathbb{Q}$-vector space, $L_2$ is an infinite dimensional $\mathbb{Q}$-vector space, $v_1$ and $v_2$ are non-zero vectors of $L_1$ and $L_2$ respectively, then there is a $\mathbb{Q}$-linear embedding $L_1\hookrightarrow L_2$ such that $v_1$ maps to $v_2$. Next, $(\sigma'_{\bullet,\bullet},\zeta)$ is an element of $(U\oplus V)[D\times\mathbb{Z}_{>0}]$ and we define $\zeta'$ to be its image in $\mathbf{k}[D\times\mathbb{Z}_{>0}]$ under the embedding. It is clear that we have~\eqref{eq:EqualMot2}. Indeed, the equations $\deg_{1,\zeta}\gamma'=0$, $\deg_{\kappa,\sigma'}\gamma'=0 $ can be written as the following equation in $U\oplus V$: \[ d'(1,\kappa)+\sum_{y\in D}\sum_{j\ge0}(r'_{y,j-1}-r'_{y,j})(\zeta_{y,,j},\sigma'_{y,j})=0, \] where $\gamma'=(r',r'_{\bullet,\bullet},d')$. But the image of this equation in $\mathbf{k}$ is exactly $\deg_{1,\zeta'}\gamma'=0$. Now the equality of motivic classes follows from Theorems~\ref{th:ExplAnsw2} and~\ref{th:ExplAnsw3} (see formulas~\eqref{eq:ExplAnswer2} and~\eqref{eq:ExplAnswer3}). \end{proof} Similarly, we have the following proposition. \begin{Proposition}\label{pr:ConnUniversal2} Assume that $\mathbf{k}$ is not a finite extension of $\mathbb{Q}$. Let $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$. Let $\sigma\in\Stab$. Let $\gamma\in\Gamma_+$. Set $\tau:=\deg_{1,\sigma}\gamma/\rk\gamma$. Then there is $\zeta'\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ such that \begin{equation}\label{eq:EqualMot3} \{\gamma'\in\Gamma_+\colon \deg_{0,\zeta}\gamma'=0, \, \deg_{1,\sigma}\gamma'=\tau\rk\gamma'\}= \{\gamma'\in\Gamma_+\colon \deg_{1,\zeta'}\gamma'=0\}. \end{equation} Moreover, we have $\big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]=[\mathcal{C}{onn}_\gamma(\zeta')]$. \end{Proposition} \begin{proof} The proof is analogous to that of Proposition~\ref{pr:ConnUniversal} except that one finds an embedding $U\oplus V\hookrightarrow\mathbf{k}$ taking $(0,\kappa)$ to $1$ and uses Theorem~\ref{th:ExplAnsw} and~\eqref{eq:ExplAnswer} instead of Theorem~\ref{th:ExplAnsw3} and~\eqref{eq:ExplAnswer3}. \end{proof} \begin{Remark} Many motivic classes of parabolic Higgs bundles and parabolic connections are equal to classes of the form $\big[\mathcal{H}{iggs}^{\sigma-{\rm ss}}(0)\big]$, cf.~\cite[Theorem~1.2.1]{FedorovSoibelmans}. However, whether these classes are universal (that is, whether one can write every motivic class $\big[\mathcal{C}{onn}_\gamma^{(\kappa,\sigma)-{\rm ss}}(\zeta)\big]$ and $\big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]$ in this form) is not clear; the reason is that there are restrictions on $\sigma\in\Stab$. \end{Remark} \subsection{Simpson's non-abelian Hodge theory} Assume now that $\mathbf{k}=\mathbb C$. Let $\sigma\in\Stab$ and $\zeta\in\mathbb C[D\times\mathbb{Z}_{>0}]$. Let $\Re\zeta$ and $\Im\zeta$ denote the real and imaginary parts of $\zeta$. Assume that either $\sigma-2\Re\zeta\in\Stab$ or $\sigma+2\sqrt{-1}\Im\zeta$ is non-resonant. In this case, for $\gamma\in\Gamma_+$ we have moduli stacks \[ \mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\qquad \text{and}\qquad \mathcal{C}{onn}_\gamma^{(1,\sigma-2\Re\zeta)-{\rm ss}}\big(\sigma+2\sqrt{-1}\Im\zeta\big). \] Assume that $\deg_{1,\sigma}\gamma=0$. Then, according to the results of~\cite{SimpsonHarmonicNoncompact}, the corresponding categories are equivalent (see especially that table on p.~720 in loc.~cit.). Note also that this equivalence of categories can be upgraded to a diffeomorphism of coarse moduli spaces (cf.~\cite{Biquard1997Higgs,BiquardBoalch2004, Nakajima1996Hyperkaehler}). \begin{Proposition}\label{pr:Simpson} We have in $\cMot(\mathbb C)$: $\big[\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}}(\zeta)\big]=\big[\mathcal{C}{onn}_\gamma^{(1,\sigma-2\Re\zeta)-{\rm ss}}\big(\sigma+2\sqrt{-1}\Im\zeta\big)\big]$. \end{Proposition} \begin{proof} Note that the system of equations $\deg_{0,\zeta}\gamma'=\deg_{1,\sigma}\gamma'=0$ is equivalent to the system of three real equations $\deg_{0,\Re\zeta}\gamma'=\deg_{0,\Im\zeta}\gamma'=\deg_{1,\sigma}\gamma'=0$. This system is, in turn, equivalent to the system $\deg_{1,\sigma-2\Re\zeta}\gamma'=\deg_{1,\sigma+2\sqrt{-1}\Im\zeta}\gamma'=0$. It remains to use Theorems~\ref{th:ExplAnsw} and~\ref{th:ExplAnsw3}. \end{proof} We emphasize that this result does not follow from the diffeomorphism of coarse moduli spaces or from the equivalence of categories. Nor the diffeomorphism of coarse moduli spaces or equivalence of categories can be derived from our result. We would also like to mention~\cite[Theorem~4.2]{HoskinsLehalleurOnVoevodskyMotive}, where the equality of Voevodsky motives is proved in the case when parabolic structures are absent and the rank and degree are coprime. \subsection{Indecomposable parabolic bundles and non-emptiness of moduli stacks} \subsubsection{Indecomposable parabolic bundles}\label{sect:Indecomp} Here we recall some results of~\cite{CrawleyBoeveyIndecompPar}. Recall that $X$ is a smooth projective curve of genus $g$, $D\subset X(\mathbf{k})$ is a non-empty set. Let $\gamma\in\Gamma_+$. We would like to know whether there exists an indecomposable parabolic bundle of class $\gamma$. The following simple statement is noted in~\cite[Introduction]{Crawley-Boevey:Indecomposable}. \begin{Lemma}\label{lm:g>0} Assume that $\mathbf{k}$ is algebraically closed. If $g>0$, then for all $\gamma\in\Gamma_+$ there is an indecomposable parabolic bundle of class $\gamma$. \end{Lemma} \begin{proof} It is well-known that there is an indecomposable vector bundle on $X$ of rank $\rk\gamma$. Now one extends it arbitrarily to a parabolic bundle of class $\gamma$. \end{proof} Next, let $X=\P^1$. Fix $\gamma=(r,r_{\bullet,\bullet},d)\in\Gamma'$ and choose a sequence of positive integers $w_\bullet$ indexed by $D$ such that $r_{x,j}=0$ for $j\ge w_x$. Consider the star-shaped graph $G_{w_\bullet}$ with vertices~$v_*$ and~$v_{x,j}$ where $x\in D$, $j$ is between~1 and~$w_x-1$. The vertex $v_*$ is connected to all the vertices of the form $v_{x,1}$, the vertex $v_{x,i}$ is connected to $v_{x,i\pm1}$ (see picture). $$ \begin{tikzpicture} \draw[fill](0,5) circle [radius=0.1]; \node [below] (02) at (0,5) {}; \node [above] (01) at (0,5) {}; \node [right] (0) at (0,5) {}; \node [left] (*) at (-0.1,5) {$v_*$}; \draw[fill](2,8) circle [radius=0.1]; \node [left] (11l) at (2,8) {}; \node [right] (11r) at (2,8) {}; \draw [thick, -] (01) -- (11l); \draw[fill](4,8) circle [radius=0.1]; \node [left] (12l) at (4,8) {}; \node [right] (12r) at (4,8) {}; \draw [thick, -] (12l) -- (11r); \draw[opacity=0, fill](6.25,8) circle [radius=0.1]; \node [left] (13l) at (6.25,8) {}; \draw [thick, -] (13l) -- (12r); \draw[fill](6.5,8) circle [radius=0.05]; \draw[fill](7,8) circle [radius=0.05]; \draw[fill](7.5,8) circle [radius=0.05]; \draw[opacity=0, fill](7.75,8) circle [radius=0.1]; \node [right] (1w12r) at (7.75,8) {}; \draw[fill](10,8) circle [radius=0.1]; \node [left] (1w11l) at (10,8) {}; \node [right] (1w11r) at (10,8) {}; \draw [thick, -] (1w11l) -- (1w12r); \draw[fill](2,6) circle [radius=0.1]; \node [above] (vx1) at (2,6.1) {$v_{x,1}$}; \node[left] (21l) at (2,6) {}; \node[right] (21r) at (2,6) {}; \draw [thick, -] (1.8,6) -- (0.3,5.1); \draw[fill](4,6) circle [radius=0.1]; \node [above] (vx2) at (4,6.1) {$v_{x,2}$}; \node [left] (22l) at (4,6) {}; \node [right] (22r) at (4,6) {}; \draw [thick, -] (22l) -- (21r); \draw[opacity=0, fill](6.25,6) circle [radius=0.1]; \node [left] (23l) at (6.25,6) {}; \draw [thick, -] (23l) -- (22r); \draw[fill](6.5,6) circle [radius=0.05]; \draw[fill](7,6) circle [radius=0.05]; \draw[fill](7.5,6) circle [radius=0.05]; \draw[opacity=0, fill](7.75, 6) circle [radius=0.1]; \node [right] (2w22r) at (7.75, 6) {}; \draw[fill](10,6) circle [radius=0.1]; \node [above] (vxw) at (10,6.1) {$v_{x,w_x-1}$}; \node [left] (2w21l) at (10,6) {}; \node [right] (2w21r) at (10,6) {}; \draw [thick, -] (2w21l) -- (2w22r); \draw[fill](2,5) circle [radius=0.05]; \draw[fill](2,4) circle [radius=0.05]; \draw[fill](2,3) circle [radius=0.05]; \draw[fill](4,5) circle [radius=0.05]; \draw[fill](4,4) circle [radius=0.05]; \draw[fill](4,3) circle [radius=0.05]; \draw[fill](10,5) circle [radius=0.05]; \draw[fill](10,4) circle [radius=0.05]; \draw[fill](10,3) circle [radius=0.05]; \draw[fill](2,2) circle [radius=0.1]; \node[left] (k1l) at (2,2) {}; \node[right] (k1r) at (2,2) {}; \draw [thick, -] (0.15,4.7) -- (k1l); \draw[fill](4,2) circle [radius=0.1]; \node [left] (k2l) at (4,2) {}; \node [right] (k2r) at (4,2) {}; \draw [thick, -] (k2l) -- (k1r); \draw[opacity=0, fill](6.25,2) circle [radius=0.1]; \node [left] (k3l) at (6.25,2) {}; \draw [thick, -] (k3l) -- (k2r); \draw[fill](6.5,2) circle [radius=0.05]; \draw[fill](7,2) circle [radius=0.05]; \draw[fill](7.5,2) circle [radius=0.05]; \draw[opacity=0, fill](7.75,2) circle [radius=0.1]; \node [right] (kwk2r) at (7.75,2) {}; \draw[fill](10,2) circle [radius=0.1]; \node [left] (kwk1l) at (10,2) {}; \node [right] (kwk1r) at (10,2) {}; \draw [thick, -] (kwk1l) -- (kwk2r); \node[] at (6,0.9) {\emph{Star-shaped graph}}; \end{tikzpicture} $$ Consider the Kac--Moody Lie algebra $\mathfrak g_{w_\bullet}$ associated to the generalized Cartan matrix defined by this graph (see, e.g.,~\cite[Section~1]{Kac82}). Let $\Lambda_{w_\bullet}$ be the root lattice of $\mathfrak g_{w_\bullet}$; we identify it with the free abelian group generated by the set of vertices. Then $\gamma$ gives rise to an element of $\Lambda_{w_\bullet}$ given by \[ \rho_{\gamma,w_\bullet}:=rv_*+\sum_{x\in D}\sum_{j=1}^{w_x-1}\left(\sum_{i=1}^jr_{x,i}\right)v_{x,j}. \] Now~\cite[p.~1334, Corollary]{CrawleyBoeveyIndecompPar} can be re-formulated as follows. \begin{Proposition}\label{pr:ExistIndecomp} In the above notation, there is a non-zero indecomposable parabolic bundle $\mathbf E\in\mathcal{B}{un}^{\rm par}_\gamma$ if and only if $\rho_{\gamma,w_\bullet}$ is a root of $\mathfrak g_{w_\bullet}$. \end{Proposition} We see that $\rho_{\gamma,w_\bullet}$ does not depend on $d$. Thus if there is an indecomposable parabolic bundle $\mathbf E$ with $\cl(\mathbf E)=(r,r_{\bullet,\bullet},d)$, then for any $d'$ there is an indecomposable parabolic bundle of class $(r,r_{\bullet,\bullet},d')$. Secondly, we see that the property of $\rho_{\gamma,w_\bullet}$ being a root does not depend on the choice of $w_\bullet$ as long as the components of $w_\bullet$ are large enough. By a slight abuse of terminology we say that $\gamma$ is a root in this case. \begin{Remark} In fact, one can consider an infinite star-shaped graph $G_D$ with $\deg D$ infinite rays, and the corresponding Kac--Moody Lie algebra $\mathfrak g_D$, which is the inductive limit of $\mathfrak g_{w_\bullet}$. Then we have a homomorphism $\rho$ from $\Gamma_+$ to the root lattice of $\mathfrak g_D$ and the classes of indecomposable parabolic bundles are exactly the $\rho$-preimages of roots. \end{Remark} \subsubsection{Non-emptiness of moduli stacks} Now we can give a full answer to the question of when $\mathcal{H}{iggs}_\gamma^{\sigma-{\rm ss}} (\zeta)$, $\mathcal{C}{onn}_\gamma(\zeta)$ and $\mathcal{C}{onn}^{(\kappa,\sigma)-{\rm ss}}_\gamma (\zeta)\!\!$ are non-empty. The first statement follows immediately from results of Crawley--Boevey. \begin{Theorem}\label{th:NonEmpty} Assume that $\gamma\in\Gamma_+$ and $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$. \begin{enumerate}\itemsep=0pt \item[$(i)$] If $g=g(X)>0$, then $\mathcal{C}{onn}_\gamma(\zeta)$ is non-empty if and only if $\deg_{1,\zeta}\gamma=0$. \item[$(ii)$] If $g=0$, then $\mathcal{C}{onn}_\gamma(\zeta)$ is non-empty if and only if $\gamma$ can be written as $\sum\limits_{i=1}^n\gamma_i$, where $\gamma_i\in\Gamma'$ are roots and $\deg_{1,\zeta}\gamma_i=0$. \end{enumerate} \end{Theorem} \begin{proof} Note that a $\mathbf{k}$-stack $\mathcal X$ is non-empty if and only if $\mathcal X\times_\mathbf{k}\overline\mathbf{k}$ is non-empty, where $\overline\mathbf{k}$ is the algebraic closure of $\mathbf{k}$. Thus, we can assume that $\mathbf{k}$ is algebraically closed from the very beginning. By Lemma~\ref{lm:existence2}, $\mathcal{C}{onn}_\gamma(\zeta)$ is non-empty if and only if there is a $(1,\zeta)$-isoslopy parabolic bundle of class $\gamma$ such that $\deg_{1,\zeta}\gamma=0$. Now (i) follows from Lemma~\ref{lm:g>0}, while (ii) follows from Proposition~\ref{pr:ExistIndecomp}. \end{proof} \begin{Theorem}\label{th:NonEmpty2} Assume that $\gamma\in\Gamma_+$, $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$, and $\sigma\in\Stab$. \begin{enumerate}\itemsep=0pt \item[$(i)$] If $g=g(X)>0$, then $\mathcal{H}{iggs}^{\sigma-{\rm ss}}_\gamma(\zeta)$ is non-empty if and only if $\deg_{0,\zeta}\gamma=0$. \item[$(ii)$] If $g=0$, then $\mathcal{H}{iggs}^{\sigma-{\rm ss}}_\gamma(\zeta)$ is non-empty if and only if $\gamma$ can be written as $\sum\limits_{i=1}^n\gamma_i$, where $\gamma_i\in\Gamma'$ are roots, $\deg_{0,\zeta}\gamma_i=0$, and the $(1,\sigma)$-slope of each $\gamma_i$ is equal to the $(1,\sigma)$-slope of $\gamma$. \end{enumerate} \end{Theorem} \begin{proof} By Proposition~\ref{pr:NonEmpty}, $\mathcal{H}{iggs}^{\sigma-{\rm ss}}_\gamma(\zeta)$ is non-empty if and only if $\big[\mathcal{H}{iggs}^{\sigma-{\rm ss}}_\gamma(\zeta)\big]\ne0$. Let $\zeta'$ be as in Proposition~\ref{pr:ConnUniversal2}. Applying Proposition~\ref{pr:NonEmpty} again, we see that $\mathcal{H}{iggs}^{\sigma-{\rm ss}}_\gamma(\zeta)$ is non-empty if and only if $\mathcal{C}{onn}_\gamma(\zeta')$ is non-empty. It remains to use Theorem~\ref{th:NonEmpty} and~\eqref{eq:EqualMot3}. \end{proof} \begin{Theorem}\label{th:NonEmpty3} Assume that $\gamma\in\Gamma_+$, $\zeta\in\mathbf{k}[D\times\mathbb{Z}_{>0}]$ and $(\kappa,\sigma)\in\Stab'$. Assume that either~$\zeta$ is non-resonant, or $\kappa=1$ and $\sigma\in\Stab$. \begin{enumerate}\itemsep=0pt \item[$(i)$] If $g=g(X)>0$, then $\mathcal{C}{onn}^{(\kappa,\sigma)-{\rm ss}}_\gamma(\zeta)$ is non-empty if and only if $\deg_{1,\zeta}\gamma=0$. \item[$(ii)$] If $g=0$, then $\mathcal{C}{onn}^{(\kappa,\sigma)-{\rm ss}}_\gamma(\zeta)$ is non-empty if and only if $\gamma$ can be written as $\sum\limits_{i=1}^n\gamma_i$, where $\gamma_i\in\Gamma'$ are roots, $\deg_{1,\zeta}\gamma_i=0$, and the $(\kappa,\sigma)$-slope of each $\gamma_i$ is equal to the $(\kappa,\sigma)$-slope of $\gamma$. \end{enumerate} \end{Theorem} \begin{proof} Same as of Theorem~\ref{th:NonEmpty2} except that one uses Proposition~\ref{pr:ConnUniversal} instead of Proposition~\ref{pr:ConnUniversal2} and~\eqref{eq:EqualMot2} instead of~\eqref{eq:EqualMot3}. \end{proof} \subsection*{Acknowledgements} We thank E.~Diaconescu, J.~Heinloth, O.~Schiffmann, and especially A.~Mellit for useful discussions and correspondence. We thank P.~Boalch for a useful comment on an earlier version. A part of this work was done while R.F.~was visiting Max Planck Institute of Mathematics in Bonn, and a part when he was visiting A.~Mellit at the University of Vienna. The work of R.F.~was partially supported by NSF grant DMS--1406532. A.S.~and Y.S.~thank IHES for excellent research conditions and hospitality. The work of Y.S.~was partially supported by NSF grants and Munson--Simu Faculty Award at Kansas State University. The authors would like to thank the anonymous referees for carefully reading the paper and for useful comments. \pdfbookmark[1]{References}{ref}

https://arxiv.org/abs/math/0512554

On weakly bounded empirical processes

Let $F$ be a class of functions on a probability space $(\Omega,\mu)$ and let $X_1,...,X_k$ be independent random variables distributed according to $\mu$. We establish high probability tail estimates of the form $\sup_{f \in F} |\{i : |f(X_i)| \geq t \}$ using a natural parameter associated with $F$. We use this result to analyze weakly bounded empirical processes indexed by $F$ and processes of the form $Z_f=|k^{-1}\sum_{i=1}^k |f|^p(X_i)-\E|f|^p|$ for $p>1$. We also present some geometric applications of this approach, based on properties of the random operator $\Gamma=k^{-1/2}\sum_{i=1}^k \inr{X_i,\cdot}e_i$, where the $(X_i)_{i=1}^k$ are sampled according to an isotropic, log-concave measure on $\R^n$.

\section{Introduction} Empirical Processes theory focuses on understanding the behavior of the supremum of the process $$ f \to Z_f = \left|\frac{1}{k} \sum_{i=1}^k f(X_i) -\mathbb{E} f \right| $$ where $F$ is a class of functions on a probability space $(\Omega,\mu)$, $f \in F$ and $(X_i)_{i=1}^k$ are independent random variables distributed according to $\mu$. Let $\mu_k$ denote the random empirical measure $k^{-1}\sum_{i=1}^k \delta_{X_i}$, and for a class $F$ we denote the supremum of the empirical process indexed by $F$ by $\|\mu_k-\mu\|_F$. Often, one would like to bound this supremum using geometric properties of the set $F$, but the question we tackle here is slightly different; our aim is to bound the supremum of the empirical process indexed by powers of the class $F$, that is, the supremum of the process indexed by the set $F^p \equiv \{ |f|^p : f \in F\}$ for $p>1$ using the geometry of the set $F$ rather than the geometry of $F^p$. The difficulty arises when elements in $F$ are not necessarily bounded functions, or in cases where the $L_\infty$ bound is weak - while the situation is considerably simpler in the bounded case. For example, if $F$ consists of functions bounded by $1$ then the empirical process indexed by $F^p$ can be bounded using a combination of symmetrization and contraction arguments. Indeed, by the Gin\'{e}-Zinn symmetrization method (see, for example, \cite{GZ84,VW}), \begin{align*} \mathbb{E} \|\mu_k-\mu\|_{F^p} \leq & 2\mathbb{E} \sup_{f \in F} \left| \frac{1}{k} \sum_{i=1}^k \varepsilon_i |f|^p(X_i) \right| \\ \leq & 2p \mathbb{E} \sup_{f \in F} \left| \frac{1}{k} \sum_{i=1}^k \varepsilon_i f(X_i) \right|, \end{align*} where $(\varepsilon_i)_{i=1}^k$ are independent, symmetric $\{-1,1\}$-valued random variables. The last inequality is evident from a contraction principle \cite{LT,VW} and the fact that $|x|^p$ is a Lipschitz function on $[-1,1]$ with constant $p$. Moreover, for a class of uniformly bounded functions, the supremum of the empirical process $\|\mu_k-\mu\|_F$ is highly concentrated around its mean, as the following theorem, due to Talagrand, shows. \begin{Theorem} \label{thm:Talagrand} \cite{Tal94,Led} Let $F$ be a class of mean zero functions defined on $(\Omega,\mu)$ such that for every $f \in F$, $\|f\|_\infty \leq b$. Let $X_1,...,X_k$ be independent random variables distributed according to $\mu$ and set $\sigma^2=k \sup_{f \in F} {\rm var}(f)$. Define \begin{equation*} Z =\sup_{f \in F} \sum_{i=1}^k f(X_i), \ \ \ \bar{Z} = \sup_{f \in F} \left| \sum_{i=1}^k f(X_i) \right|. \end{equation*} Then, for every $x>0$, \begin{equation} \label{eq:Talagrand} Pr \left(\left\{\left|Z-\mathbb{E} Z\right| \geq x \right\} \right) \leq c_1\exp\left(-\frac{x}{c_2b}\log\left(1+\frac{bx}{\sigma^2+b\mathbb{E} \bar{Z}}\right)\right), \end{equation} where $c_1$ and $c_2$ are absolute constants. The same inequality is also true when $\bar{Z}$ replaces $Z$ in \eqref{eq:Talagrand}. \end{Theorem} Unfortunately, in many applications the function class at hand does not consist of uniformly bounded functions, or even if the functions are, the uniform bound is very bad. One such example which motivated this study is the class of linear functionals of Euclidean norm 1 on $\mathbb{R}^n$, and the variables $X_i$ are distributed according to a Borel measure on $\mathbb{R}^n$ which is natural from the geometric viewpoint, namely, a measure which is isotropic and log-concave. \begin{Definition} A probability measure $\mu$ on $\mathbb{R}^n$ is called isotropic if for every $y \in \mathbb{R}^n$, $\int |\inr{x,y}|^2 d\mu(x) = \|y\|^2$. The measure $\mu$ is log-concave if for every $0<\lambda<1$ and every Borel measurable $A,B \subset \mathbb{R}^n$, $\mu(\lambda A+(1-\lambda)B) \geq \mu(A)^\lambda\mu(B)^{1-\lambda}$, where $A+B$ is the Minkowski sum of $A$ and $B$. \end{Definition} A question of particular interest in this case can be formulated as follows: \begin{Question} \label{qu:sphere} Let $\mu$ be an isotropic measure on $\mathbb{R}^n$ and let $X_1,...,X_k$ be independent, distributed according to $\mu$. Given $T \subset \mathbb{R}^n$, for every $0<\varepsilon,\delta<1$ and $p>1$, what is the smallest integer $k_0$ such that for every $k \geq k_0$, with probability at least $1-\delta$, $$ \sup_{t \in T} \left|\frac{1}{k}\sum_{i=1}^k |\inr{X_i,t}|^p -\mathbb{E} |\inr{X,t}|^p \right| < \varepsilon? $$ \end{Question} Two simple examples which come to mind are when $p=2$, $T=S^{n-1}$ and $\mu$ is the Gaussian measure on $\mathbb{R}^n$ or the uniform measure on the vertices of the unit cube. \begin{Example} \label{ex:gauss} {\rm For every $t \in \mathbb{R}^n$ define the linear functional $f_t=\inr{t,\cdot}$ and set $F=\{f_t:t \in S^{n-1}\}$. Let $\mu_G$ be the Gaussian measure on $\mathbb{R}^n$ and note that for every $t \in S^{n-1}$, $\mathbb{E} f_t^2 =1$. Then, $$ \|\mu_k-\mu\|_{F^2}=\sup_{t\in S^{n-1}} \left| \frac{1}{k} \sum_{i=1}^k \inr{t,X_i}^2 -1 \right| = \sup_{t \in S^{n-1}} \left| \frac{1}{k} \|\Gamma t\|^2 - 1 \right|, $$ where $\Gamma$ is a random $k \times n$ matrix with independent, standard Gaussian random variables as entries. Hence, if $\|\mu_k-\mu\|_{F^2}<\varepsilon$, the gaussian matrix is an almost isometric embedding of $\ell_2^n$ in $\ell_2^k$ which is a well known and useful fact and occurs as long as $k \geq c(\varepsilon,\delta)n$ (see \cite{Pis}). Another example is when $\mu=\mu_R$ is the uniform probability measure on $\{-1,1\}^n$. Thus, if $\|\mu_k-\mu\|_{F^2} \leq \varepsilon$ then a random $k \times n$ matrix with independent, symmetric, $\{-1,1\}$-valued entries is an almost isometric embedding of $\ell_2^n$ in $\ell_2^k$. Unfortunately, functions in $F$ on the probability space $(\mathbb{R}^n,\mu_G)$ are not bounded, while on $(\mathbb{R}^n, \mu_R)$ the best uniform $L_\infty$ bound is $\sup_{t \in S^{n-1}} \|f_t\|_\infty \leq \sqrt{n}$ which is too weak to be useful. Therefore, symmetrization and concentration methods which are so helpful in the bounded case can not assist in resolving Question \ref{qu:sphere} here, as well as in other, more general examples we will explore. The useful property of linear functionals (with respect to both $\mu_G$ and $\mu_R$) is that for every $f_t \in F$, $$ Pr \left(|f_t| \geq u\right) \leq 2\exp\left(-cu^2\right) $$ for a suitable absolute constant $c$, implying that functions in $F$ exhibit a subgaussian behavior. Moreover, using Borell's inequality \cite{Bor,MS}, one can show that if $\mu$ is an arbitrary isotropic log-concave measures, linear functionals exhibit a subexponential decay. To formulate these decay properties in a more accurate way, we require the definition of Orlicz norms \cite{LT,VW}}. \end{Example} \begin{Definition} \ For $\alpha \geq 1$ the $\psi_\alpha$ norm of a random variable $Y$ is defined by $$ \|Y\|_{\psi_\alpha} = \inf \left\{ u>0 : \mathbb{E} \exp(|Y|^\alpha/u^\alpha) \leq 2 \right\}. $$ \end{Definition} It is standard to verify that if $Y$ has a bounded $\psi_\alpha$ norm then $Pr \left(|Y| \geq t \right) \leq 2\exp(-c t^\alpha/\|Y\|^\alpha_{\psi_\alpha})$ where $c$ is an absolute constant. The reverse direction is also true, and if $Y$ has a tail bounded by $\exp(-t^\alpha/K^\alpha)$ then $\|Y\|_{\psi_\alpha} \leq c_1 K$. Out main goal is to show how decay properties of individual class members can be combined to control $\|\mu_k-\mu\|_{F^p}$. As a starting point, let us consider the linear case where is addition, functionals are subgaussian with respect to the $\ell_2^n$ norm, i.e. for every $y \in \mathbb{R}^n$, $\|\inr{y,X}\|_{\psi_2} \leq c\|y\|_2$. In particular, the diameter of $F=S^{n-1}$ is bounded with respect to the $\psi_2$ norm. This fact by itself is not enough to bound $\|\mu_k - \mu \|_{F^p}$, and to that end we require the following notion of complexity of the class $F$. \begin{Definition} \label{def:gamma-2} \cite{Tal:book} For a metric space $(T,d)$, an {\it admissible sequence} of $T$ is a collection of subsets of $T$, $\{T_s : s \geq 0\}$, such that for every $s \geq 1$, $|T_s|=2^{2^s}$ and $|T_0|=1$. For $\beta \geq 1$, define the $\gamma_\beta$ functional by $$ \gamma_\beta(T,d) =\inf \sup_{t \in T} \sum_{s=0}^\infty 2^{s/\beta}d(t,T_s), $$ where the infimum is taken with respect to all admissible sequences of $T$. \end{Definition} In \cite{MPT} the question of estimating $\|\mu_k-\mu\|_{F^2}$ has been studied for sets of functions which have a bounded diameter with respect to the $\psi_2$ metric and a finite $\gamma_2(F,\psi_2)$, under the additional assumption that for every $f \in F$, $\mathbb{E} f^2 =1$. \begin{Theorem} \label{thm:MPT} \cite{MPT} There exist absolute constants $c_1, c_2, c_3$ and for which the following holds. Let $(\Omega,\mu)$ be a probability space, set $F$ to be a subset of the unit sphere of $L_2(\mu)$ and assume that ${\rm diam}(F, \psi_2) = \alpha$. Then, for any $\theta>0$ and $k\ge 1$ satisfying $$ c_1\alpha \gamma_2(F, \psi_2)\leq \theta\sqrt{k}, $$ with probability at least $1-\exp(-c_2\theta^2k/\alpha^4)$, $ \|\mu_k-\mu\|_{F^2} \leq \theta$. Moreover, if $F$ is symmetric, then $\mathbb{E} \|\mu_k-\mu\|_{F^2} \leq c_3\alpha \gamma_2(F,\psi_2)/\sqrt{k}$. \end{Theorem} Theorem \ref{thm:MPT} gives an answer to Question \ref{qu:sphere} for $p=2$ under a $\psi_2$ assumption in a very general situation. It is particularly helpful when the $\psi_2$ metric endowed on $F$ is equivalent to the $L_2$ metric, that is, if for every $f,g \in F$, $\|f-g\|_{\psi_2} \leq K \|f-g\|_{L_2}$. In such a case, ${\rm diam}(F,\psi_2) \sim {\rm diam}(F,L_2)$ and $\gamma_2(F,\psi_2) \sim \gamma_2(F,L_2)$, where by $A \sim B$ we mean that there are absolute constants $c$ and $C$ such that $cA \leq B \leq CA$. By the majorizing measures Theorem (see \cite{Tal:book} for the most recent survey on the subject), $\gamma_2(F,L_2)$ is equivalent to the expectation of the supremum of the Gaussian processes indexed by $F$, denoted by $\mathbb{E}\|G\|_F$. Therefore, under a $\psi_2$ assumption, Theorem \ref{thm:MPT} implies that if $F \subset S(L_2)$ then for every $0<\delta<1$, with probability at least $1-\delta$, $$ \|\mu_k-\mu\|_{F^2} \leq c \frac{\mathbb{E}\|G\|_F}{\sqrt{k}}, $$ where $c$ depends on $\delta$ and on the equivalence constant between the $\psi_2$ and $L_2$ metrics. In the geometric context of Example \ref{ex:gauss}, Theorem \ref{thm:MPT} is helpful when the indexing set in an arbitrary subset of $S^{n-1}$. Moreover, if the measure $\mu$ happens to be isotropic, then the Gaussian process indexed by $F$ is the isonormal one and thus $\gamma_2(F,L_2) \sim \mathbb{E}\sup_{t \in T} \left|\sum_{i=1}^n g_it_i \right|$, where $g_1,..,g_n$ are independent, standard Gaussian variables. Unfortunately, the assumption that the $\psi_2$ metric is equivalent to the $L_2$ metric is overly optimistic. In particular, the class may not have a well bounded diameter in $\psi_2$, or the diameter could be of the same order of magnitude as $\gamma_2(F,\psi_2)$. For example, if $\mu$ is log-concave and isotropic, then for every $y \in \mathbb{R}^n$, the function $f_y=\inr{y,\cdot}$ satisfies $\|f_y\|_{\psi_1(\mu)} \leq K \|f_y\|_{L_2(\mu)}$ and the $\psi_1$ and $L_2$ norms are equivalent on $\mathbb{R}^n$, but in contrast, the $\psi_2$ diameter of $S^{n-1}$ might be polynomial in the dimension (e.g. $\sqrt{n}$ when $\mu$ is the normalized volume measure on the isotropic position of the unit ball of $\ell_1^n$). Hence, the bound one can establish from Theorem \ref{thm:MPT} is useless in such cases because of the way it depends on the $\psi_2$ diameter of the set. It would be desirable to prove a result of a similar flavor to Theorem \ref{thm:MPT}, with the $\psi_2$ diameter of $F$ replaced by the $\psi_1$ diameter and also removes the restrictions that $p=2$ and that $T \subset S(L_2)$. Our main result implies just that. To see why the $\psi_1$ case is considerably more difficult than the $\psi_2$ one, consider a single function $h \in L_{\psi_1}$. By Bernstein's inequality (Lemma \ref{lemma:Bern} below), empirical means of $h$ are highly concentrated around its expectation, with a tail which decays exponentially in sample size. Clearly, if a function $f \in L_{\psi_2}$ then $f^2 \in L_{\psi_1}$ and hence exhibits the degree of concentration needed in the proof of Theorem \ref{thm:MPT}. On the other hand, if $f \in L_{\psi_1}$ the degree of concentration of empirical means of $f^2$ around $\mathbb{E} f^2$ is not strong enough for that approach. To overcome this obstacle, the method we suggest here is to decompose $F$ to two subsets $F_1$ and $F_2$ which satisfy that $F \subset F_1+F_2$. Fix $\theta(k)>0$ and consider the sets $F_1=\{f1_{\{|f| \leq \theta\}} : f \in F\}$ and $F_2=\{f1_{\{|f|> \theta\}} : f \in F\}$. Since all the functions in $F_1$ are bounded by $\theta$, the empirical mean $\mu_k(f)$ is highly concentrated around the true mean for any $f \in F_1$ and $\|\mu_k-\mu\|_{F_1}$ (or $\|\mu_k-\mu\|_{F_1^p}$, using a contraction argument) is well behaved. The key point in this approach is to control the ``large part" of the process, namely, $$ \sup_{f \in F} k^{-1}\sum_{i=1}^k |f|^p1_{\{|f| > \theta \}}, $$ and to show that the supremum is small even for a relatively low level of truncation $\theta$. The reason this supremum is small has nothing to do with the concentration of each individual class member around its mean, but rather with the fact that with high probability, all the functions in $F$ have an empirical distribution which decays quickly. And indeed, the main Theorem we present is an ``empirical processes" version of result due to Bourgain on the distribution of functions in $F$ with respect to the (random) empirical measure $\mu_k$. \noindent{\bf Theorem A.} {\it There exist absolute constants $c_1$, $c_2$ and $c_3$ for which the following holds. Let $F$ be a class of mean zero functions on $(\Omega,\mu)$. For every $v_1,v_2 \geq c_1$, with probability at least $1-\exp(-c_2 \min\{v_1,v_2 \} )$, for any $f \in F$ and $t>0$, \begin{equation*} \left| \left\{ i : \left|f(X_i)\right| \geq t \right\} \right| \leq \max \left\{ \frac{c_3v_1 \gamma_2^2(F,\psi_2)}{t^2}, ek\exp \left(-\frac{\theta}{c_3\alpha v_2}\right) \right\}, \end{equation*} where $\alpha = {\rm diam}(F,\psi_1)$. } Bourgain's argument \cite{Bour} is very different from ours and is tailored to the specific case $F=\{\inr{y,\cdot} : y \in S^{n-1}\}$, where $X_1,...,X_k$ are selected according to a log-concave measure on $\mathbb{R}^n$ (see Section \ref{sec:main} for a more detailed discussion). The proof of Theorem A is based on the following estimate (which will be shown to be optimal) on the $\ell_1$ structure of a random coordinate projection of $F$. \noindent{\bf Theorem B.} {\it For every $0<\delta<1$ there is a constant $c(\delta)$ for which the following holds. For every integer $k$, with probability at least $1-\delta$, for every $f \in F$ and $I \subset \{1,...,k\}$, \begin{align*} \sum_{i \in I} |f(X_i)| & \leq c(\delta) \left(\sqrt{|I|}\gamma_2(F,\psi_2) + {\rm diam}(F,\psi_1)|I|\log\left(\frac{ek}{|I|}\right)\right), \\ \sum_{i \in I} |f(X_i)| & \leq c(\delta) \left(\sqrt{|I|}\gamma_2(F,\psi_2) + {\rm diam}(F,\psi_2)|I|\sqrt{\log\left(\frac{ek}{|I|}\right)}\right). \end{align*} } We present several geometric applications of Theorem A. The first of which is a ``log-concave" version of the celebrated result of Pajor and Tomczak-Jaegermann \cite{PT} on sections of small diameter of a convex, symmetric body $K$ (see also \cite{MiPi,Mi,MiPa} for results along the same lines). We show that if $X_1,...,X_k$ are selected according to an isotropic log-concave measure on $\mathbb{R}^n$, then with high probability, the intersection of the kernel of the operator $\Gamma=\sum_{i=1}^k \inr{X_i,\cdot}e_i$ with $K$ will have a small diameter. \noindent {\bf Theorem C.} {\it For every $0<\delta<1$ there exists a constant $c(\delta)$ for which the following holds. Let $\mu$ be an isotropic, log-concave measure on $\mathbb{R}^n$ and let $K \subset \mathbb{R}^n$ be a convex symmetric body. If $X_1,...,X_k$ are independent, distributed according to $\mu$, then with probability at least $1-\delta$, $$ {\rm diam}({\rm ker} \Gamma \cap K) \leq q_k^*(K), $$ where $$ q_k^*(K) = \inf\left\{\rho>0 : \rho \geq c(\delta)\frac{V_\rho\sqrt{\log{V_\rho}}}{\sqrt{k}}\right\}, $$ and $V_\rho=\gamma_2(K\cap \rho S^{n-1} ,\psi_2)$. } \vskip 0.4cm If $\mu$ is a subgaussian measure, Theorem C gives a weaker result (by up to a factor of $\sqrt{\log{n}}$) and with a weaker probability estimate than Theorem \ref{thm:MPT}. On the other hand, it is applicable for a wider set of measures. The downside of our approach is that it depends on the parameter $\gamma_2(F,\psi_2)$ which is often hard to bound. However, as we show, a completely $\psi_1$ version of Theorem B is not true and one might have to use the additional structural assumptions on the indexing set to improve our estimate. Luckily, in the case $F=\{f_t : t \in S^{n-1}\}$, it is possible to bound $\|\mu_k-\mu\|_{F^p}$ in a rather strong sense (though probably suboptimal by a logarithmic factor) using a truncation of the measure $\mu$. Let $\mu$ be a probability measure on $\mathbb{R}^n$, for every integer $k$, let $X_1,...,X_k$ be independent, distributed according to $\mu$ and set $H_k=\mathbb{E} \max_{1 \leq i \leq k} \|X_i\|$. Observe that if $Y_i=X_i 1_{\{\|X\| \leq c_1(\delta)H_k\}}$, then with probability at least $1-\delta$, $X_i=Y_i$ for $1 \leq i \leq k$. Thus, one can consider the process $\|\nu_k-\nu\|_{F^p}$ instead of the original process $\|\mu_k-\mu\|_{F^p}$. Moreover, one can show \noindent{\bf Theorem D.} {\it There exist absolute constants $c_1$, $c_2$ and $c_3$ for which the following holds. If $F=\{f_t : t \in S^{n-1}\}$ then $$ \gamma_2(F,\psi_2(\nu)) \leq c_1 H_k\sqrt{\log{n}}. $$ } \vskip0.5cm Note that if $\mu$ is an isotropic log-concave measure on $\mathbb{R}^n$ and if $n \leq k \leq \exp(c_2\sqrt{n})$ then $H_k \leq c_3 \sqrt{n}$, which is a fact recently proved by Paouris \cite{Pao}. As we demonstrate in Section \ref{sec:applications}, the combination of Theorem B and Theorem D allows us to bound $$ \sup_{t \in S^{n-1}} \left| \frac{1}{k}\sum_{i=1}^k |\inr{t,X_i}|^p - \mathbb{E}|\inr{t,X}|^p \right| $$ for any log-concave measure. \section{Preliminary Results} In this section we present basic results which are used throughout this article. First, a notational convention. All absolute constants are positive numbers, denoted by $c,c_1,c_2,..$ etc. Their value may change from line to line. We denote the Euclidean norm by $\| \ \|$, while all other norms will be clearly specified. There are several useful results regarding the concentration and tail behavior of sums of independent random variables. The first one we present here deals with subgaussian random variables and can be easily seen using the moment generating function. \begin{Lemma} \label{lemma:sum-subgauss} \cite{VW} There exists an absolute constant $c$ for which the following holds. Let $X$ be a subgaussian random variable and let $X_1,...,X_k$ be independent, distributed as $X$. Then, for every $a=(a_1,...,a_k) \in \mathbb{R}^k$ $$ \|\sum_{i=1}^k a_i X_i \|_{\psi_2} \leq c\|X\|_{\psi_2} \|a\|. $$ \end{Lemma} If $X$ is not a $\psi_2$ random variable and only exhibits a subexponential tail then Bernstein's inequality describes the way the average of independent copies of $X$ concentrate around their mean - with a tail which is a mixture of subgaussian and subexponential. \begin{Lemma} \label{lemma:Bern} \cite{VW} There exists an absolute constant $c$ for which the following holds. Let $X_1,...,X_k$ be independent copies of a mean zero random variable. Then, for any $t>0$, \begin{equation*} \Pr \left(\left|\frac{1}{k}\sum_{i=i}^k X_i \right| >t \right) \leq 2\exp \left(-c\, k\min \left( \frac{t}{\|X\|_{\psi_1}},\, \frac{t^2}{\|X\|^2_{\psi_1}}\right)\right). \end{equation*} \end{Lemma} It turns out that using the generic chaining method \cite{Tal:book} combined with Lemma \ref{lemma:sum-subgauss} or Lemma \ref{lemma:Bern}, one can bound the supremum of the empirical process indexed by $F$. \begin{Theorem} \cite{Tal:book} \label{thm:generic-chaining} There exists an absolute constant $c$ for which the following holds. If $F$ is a class of functions on $(\Omega,\mu)$, then for every integer $k$, \begin{align*} \mathbb{E} \|\mu_k-\mu\|_F & \leq c \frac{\gamma_2(F,\psi_2)}{\sqrt{k}}, \\ \mathbb{E} \|\mu_k-\mu\|_F & \leq c\left(\frac{\gamma_2(F,\psi_1)}{\sqrt{k}}+ \frac{\gamma_1(F,\psi_1)}{k} \right), \end{align*} and similar bounds hold with high probability. \end{Theorem} In many cases, computing the $\gamma$ functionals is a difficult task. It is possible to upper bound them using a metric entropy integral, similar to Dudley's integral in the context of Gaussian process. \begin{Definition} Let $(T,d)$ be a metric space. The covering number of $T$ at scale $\varepsilon$ is the minimal number of open balls (with respect to the metric $d$) of radius $\varepsilon$ needed to cover $T$. The covering numbers of $(T,d)$ are denoted by $N(\varepsilon,T,d)$. \end{Definition} Since one way of forming an admissible sequence for $(T,d)$ is to use an almost optimal cover (the set $T_s$ is a cover at the scale at which one needs $2^{2^s}$ balls to cover $T$), the following is evident: \begin{Lemma} There exists an absolute constant $c$ for which the following holds. Let $(T,d)$ be a metric space. Then, $$ \gamma_2(T,d) \leq c\int_0^\infty \sqrt{\log{N(\varepsilon,T,d)}}d\varepsilon. $$ \end{Lemma} A much more difficult result, due to Talagrand \cite{Tal94,Tal:book}, is that if $T$ is a unit ball of a $2$-convex normed space, $\gamma_2$ could be bounded from above by a sharper version of the entropy integral. \begin{Definition} A Banach space is called 2-convex if there is $\rho>0$ such that for $\|x\|,\|y\| \leq 1$, $\|x+y\| \leq 2-2\rho\|x-y\|^2$. \end{Definition} \begin{Theorem} \label{thm:talagrand} \cite{Tal1} For every $\rho>0$ there exists a constant $c(\rho)$ for which the following holds. If $Y$ is a $2$-convex Banach space with parameter $\rho$ and if the metric $d$ is given by some other norm $| \ |$, then $$ \gamma_2(B_Y, d) \leq c(\rho) \left(\int_0^\infty \varepsilon \log N \left(B_Y, B_{| \ |},\varepsilon \right) d\varepsilon \right)^{\frac{1}{2}}. $$ \end{Theorem} Theorem \ref{thm:talagrand} is used in the case $Y=\ell_2^n$, the $n$-dimensional Euclidean space, where $d$ is the metric endowed on $\mathbb{R}^n$ by the $\psi_2$ norm (see Section \ref{sec:applications}). \section{Decomposing classes of functions} \label{sec:main} Let $F$ be a class of functions on the probability space $(\Omega,\mu)$ and assume that for every $f \in F$, $\mathbb{E} f =0$. Let us formulate the main technical tool we require. \begin{Theorem} \label{thm:sets} There exists absolute constants $c_1$ and $c_2$ for which the following holds. Let $F$ be a class of mean zero functions on $(\Omega,\mu)$ and set $X_1,...,X_k$ to be independent random variables distributed according to $\mu$. Then, for every $v_1, v_2 \geq c_1$, with probability at least $1-\exp(-c_2 \min\{v_1^2,v_2\})$, for every $I \subset \{1,...,k\}$, $$ \sup_{f \in F} \left| \sum_{i \in I} f(X_i) \right| \leq v_1 \sqrt{|I|} \gamma_2(F,\psi_2) + v_2 {\rm diam}(F,\psi_1) |I| \log\left(\frac{ek}{|I|}\right). $$ \end{Theorem} Theorem \ref{thm:sets} has a similar version in which one assumes that the set of functions is well bounded in $\psi_2$. \begin{Theorem} \label{thm:sets-psi_2} There exist absolute constants $c_1$ and $c_2$ for which the following holds. Let $F$ and $X_1,...,X_k$ be as in Theorem \ref{thm:sets}. Then, for every $v \geq c_1$, with probability at least $1-\exp(-c_2 v^2)$, for every $I \subset \{1,...,k\}$, $$ \sup_{f \in F} \left| \sum_{i \in I} f(X_i) \right| \leq v \left(\sqrt{|I|} \gamma_2(F,\psi_2) + {\rm diam}(F,\psi_2)|I| \sqrt{ \log\left(\frac{ek}{|I|}\right)}\right). $$ \end{Theorem} Theorem \ref{thm:sets} is an empirical processes version of a lemma due to Bourgain (\cite{Bour}, see also \cite{GiaMil}) which deals with the case when $F$ is $S^{n-1}$, considered as a class of linear functionals on $\mathbb{R}^n$ and $\mu$ is an isotropic log-concave measure. Unlike Bourgain's argument, which relies heavily on the fact that the functions in the class are linear functionals and on that the indexing set is the whole sphere, Theorem \ref{thm:sets} is very general. Observe that if the $L_2$ and $\psi_2$ metrics are equivalent on $F$ with a constant $\beta$ and if $\mathbb{E} \|G\|_F$ denotes the expectation of the supremum of the Gaussian process indexed by $F$, then by the majorizing measures Theorem there are absolute constants $c$ and $C$ and a constant $c_1(\beta)$ depending only on $\beta$ such that $$ c_1(\beta) \gamma_2(F,\psi_2) \leq c\gamma_2(F,L_2) \leq \mathbb{E}\|G\|_F \leq C\gamma_2(F,L_2) \leq C \gamma_2(F,\psi_2). $$ Therefore, by Theorem \ref{thm:sets}, with probability at least $1-\delta$, for every $I \subset \{1,...,k\}$, \begin{align*} \sup_{f \in F} \left| \sum_{i \in I} f(X_i) \right| & \leq c(\delta,\beta) \left( \sqrt{|I|} \mathbb{E}\|G\|_F + {\rm diam}(F,\psi_1) |I| \log\left(\frac{ek}{|I|}\right)\right), \\ \sup_{f \in F} \left| \sum_{i \in I} f(X_i) \right| & \leq c(\delta,\beta) \left(\sqrt{|I|} \mathbb{E}\|G\|_F + {\rm diam}(F,\psi_2)|I| \sqrt{\log\left(\frac{ek}{|I|}\right)}\right). \end{align*} Let us point out that it is impossible to obtain a fully $\psi_1$ version of Theorem \ref{thm:sets}. Indeed, suppose the converse was true, and that for every set $F$ and integer $k$, with probability at least $1-\delta$, for every $I \subset \{1,...,k\}$, \begin{equation} \label{eq:wrong} \sup_{f \in F} \left| \sum_{i \in I} f(X_i) \right| \leq c(\delta) \left(\sqrt{|I|} \gamma_2(F,\psi_1) + {\rm diam}(F,\psi_1) |I| \log\left(\frac{ek}{|I|}\right)\right). \end{equation} Let $Y$ be an exponential random variable and let $X^{(n)}=\sum_{i=1}^n \frac{Y_i}{\sqrt{\log{(i+1)}}}e_i \in \mathbb{R}^n$ where $(e_i)_{i=1}^n$ is the standard basis in $\mathbb{R}^n$ and $(Y_i)_{i=1}^n$ are independent copies of $Y$. Setting $\mu^{(n)}$ to be the measure on $\mathbb{R}^n$ which endows $X^{(n)}$, \eqref{eq:wrong} can not be true for $\mu^{(n)}$ and $F_n=B_1^n$, the unit ball in $\ell_1^n$ even when $k=1$. Indeed, using Borell's inequality (see, e.g. \cite{MS}) or by a direct computation as in \cite{BGMN}, it is evident the for every $t \in \mathbb{R}^n$, $$\left\|\inr{t,X^{(n)}}\right\|_{\psi_1} \leq c\left\|\inr{t,X^{(n)}}\right\|_{L_2}=c\left(\sum_{i=1}^n \frac{t_i^2}{\log{(i+1)}}\right)^{1/2} \equiv c|t|^{(n)}, $$ where $| \ |^{(n)}$ is the weighted Euclidean norm with weights $(\sqrt{\log(i+1)})_{i=1}^n$ and $c$ is an absolute constant. Hence, by the majorizing measures Theorem and a standard computation, there are absolute constants $c$, $c_1$ and $c_2$ such that for every $n$, $$ \gamma_2\left(F_n,\psi_1(\mu^{(n)})\right) \leq c\gamma_2(F_n,| \ |^{(n)}) \leq c_1\mathbb{E} \sup_{t \in B_1^n} \left|\sum_{i=1}^n g_i \frac{t_i}{\sqrt{\log(i+1)}} \right| \leq c_2. $$ Therefore, if \eqref{eq:wrong} were correct for $k=1$, it would follow that with probability of at least $1/2$, $\sup_{t \in B_1^n} \inr{t,X^{(n)}} \leq c_3$, for a suitable $c_3$ which is independent of $n$. On the other hand, an easy computation shows that with probability larger than some constant $c_4$, $$ \sup_{t \in B_1^n} \inr{t,X^{(n)}} \geq \sqrt{\log{(n+1)}}, $$ and thus it is impossible to get a completely $\psi_1$ version of Theorem \ref{thm:sets}. Next, observe that Theorem \ref{thm:sets-psi_2} is optimal, in the sense that both the $\gamma_2$ term and the term that depends on the $\psi_2$-diameter are required. To see this, fix an integer $k$ and let $1 \leq \ell <k$. Set $F=\{a,-a\} \subset S^{n-1}$, acting as linear functional of $\mathbb{R}^n$ and let $X=(g_1,...,g_n)$ be a Gaussian vector in $\mathbb{R}^n$. With this choice of $X$, the $\psi_2$ and $\ell_2$ metrics on $\mathbb{R}^n$ are equivalent with an absolute constant, and thus $\gamma_2(F,\psi_2) \leq c \gamma_2(F,\ell_2) =c_1$. On the other hand, writing $a=(a_1,...,a_n)$, for every $1 \leq \ell \leq k$, \begin{equation*} \sup_{f \in F} \sup_{I \in E_\ell} \left|\sum_{i \in I} f(X_i)\right|=\sup_{I \in E_\ell} \left|\sum_{i \in I} \sum_{j=1}^n a_jg_{i,j} \right|, \end{equation*} which is the supremum of the Gaussian process indexed by $$ E_\ell=\left\{I\ : \ I \subset \left\{1,...,k\right\}, \ |I|=\ell \right\} $$ with the covariance structure endowed by the Hamming metric on $E_\ell$, given by $d_H(I,J)=|I \vartriangle J|^{1/2}$. Recall the well known entropy estimate for $E_\ell$ with respect to this metric (see, for example, \cite{MPR}): \begin{Lemma} \label{Lemma:MPR} For $0<\lambda\le 1/2$ there exists a constant $c_\lambda$ for which the following holds. For every integers $k$ and $1 \leq \ell \leq k$, there is a subset $P \subset E_\ell$ which satisfies that $\log |P| \geq (1-\lambda) \ell \log\left(c_\lambda \frac{k}{\ell}\right)$ and if $I, J \in P$ and $I \not=J$ then $d_H(I,J) \geq \sqrt{\lambda \ell}$. In other words, $$ \log N \left(E_\ell,\sqrt{\lambda \ell}, d_H\right) \geq (1-\lambda) \ell \log \left(c_\lambda \frac{k}{\ell}\right). $$ \end{Lemma} Combining Lemma \ref{Lemma:MPR} for $\lambda =1/4$ with Sudakov's minoration (see, e.g. \cite{LT}), it is evident that $$ \mathbb{E} \sup_{I \in E_\ell} \left|\sum_{i \in I} \sum_{j=1}^n a_jg_{i,j} \right| \geq c\ell \sqrt{\log\left(\frac{ck}{\ell}\right)}, $$ proving that the second term in Theorem \ref{thm:sets-psi_2} is indeed necessary. To show that the $\gamma_2$ term is necessary, let $F=\{-1,1\}^n$ acting as linear functionals, and again set $X$ to be the Gaussian vector on $\mathbb{R}^n$. Then, for every $1 \leq \ell \leq k$, \begin{equation*} \sup_{a \in \{-1,1\}^n} \sup_{I \in E_\ell} \left|\sum_{i \in I} \sum_{j=1}^n g_{i,j}a_j \right| \geq \sup_{a \in \{-1,1\}^n} \left|\sum_{i = 1}^\ell \sum_{j=1}^n g_{i,j}a_j \right|. \end{equation*} The latter is the supremum of the Gaussian process indexed by $\{-1,1\}^n$ with the covariance structure given by the metric $d(u,v)=\sqrt{\ell}\|u-v\|_{\ell_2^n}$. Thus, it is standard to verify that $$ \mathbb{E}\sup_{a \in \{-1,1\}^n} \left|\sum_{i \in 1}^\ell \sum_{j=1}^n g_{i,j}a_j \right| \geq c\sqrt{\ell}n. $$ On the other hand, ${\rm diam}(\{-1,1\}^n,\psi_2) \leq c{\rm diam}(\{-1,1\}^n,\ell_2^n) \leq c\sqrt{n}$. Thus, the upper bound from Theorem \ref{thm:sets-psi_2} is of the order of $\sqrt{\ell}n+\sqrt{n}\sqrt{\ell \log(ek/\ell)}$, showing that the $\gamma_2$ term can not be removed from the bound. \noindent {\bf proof of Theorem \ref{thm:sets}.} To control $\sup_{f \in F} \left|\sum_{i \in I} f(X_i) \right|$, consider the following $k$ processes. Recall that for every $1 \leq \ell \leq k$, $E_\ell = \{ I : I \subset \{1,...,k\}, \ |I| =\ell \}$ and define the random process $$ f \to Z_f^\ell = \sup_{I \in E_\ell} \left| \sum_{i \in I} f(X_i) \right|, $$ where $X_1,...,X_k$ are independent random variables distributed according to $\mu$. Fix $1 \leq \ell <k$ (the result for $\ell=k$ requires minor changes and is omitted) and consider the process $Z_f^\ell$. Observe that for every $f,g \in F$, \begin{align*} Pr \left( \left| Z_f^\ell - Z_g^\ell \right| \geq t \right) & \leq Pr \left( \sup_{I \in E_\ell} \left| \sum_{i \in I} (f-g)(X_i) \right| \geq t \right) \\ & \leq |E_\ell| Pr \left( \left| \sum_{i=1}^\ell (f-g)(X_i) \right| \geq t \right) \\ & \leq 2|E_\ell| \exp \left(-\frac{ct^2}{\|f-g\|^2_{\psi_2} \ell} \right), \end{align*} where $c$ is an absolute constant. Without loss of generality, assume that $\gamma_2(F,\psi_2) < \infty$, let $(F_s)_{s \geq 0}$ be an almost optimal admissible sequence for the metric space $(F,\psi_2)$ and set $\pi_s(f)$ to be a nearest element to $f$ in $F_s$ with respect to the $\psi_2$ metric. Thus, $|F_s| \leq 2^{2^s}$. Define $s_0$ as the first index such that $2^{s_0-1} < \log |E_\ell| \leq 2^{s_0}$, and note that $$ Z_f^\ell = Z^\ell_{\pif{s_0}} + \sum_{i=s_0}^\infty Z^\ell_{\pif{i+1}} - Z^\ell_{\pif{i}}. $$ Fix $u>0$ to be specified later and $s \geq s_0$, and consider $t_s = u\sqrt{\ell} \| \pif{s+1}-\pif{s}\|_{\psi_2} 2^{s/2} \sqrt{\log |E_\ell|}$. Then, \begin{align*} Pr \left( \left|Z^\ell_{\pif{s+1}} - Z^\ell_{\pif{s}} \right| \geq t_s \right) & \leq 2 |E_\ell| \exp(-cu^2 2^{s-1} \log |E_\ell|) \\ & \leq 2 \exp \left(-c\log |E_\ell| \left(u^22^{s-1} -1 \right) \right). \end{align*} Take $u=v_1 /\sqrt{\log |E_\ell|}$ for $v_1 \geq c_1$ and note that $2^s > \log |E_\ell|$, implying that the tail is upper bounded by $2\exp(-c_2v_1^2 2^s)$. Summing over $s_0 \leq s < \infty$ it is evident that with probability at least $$ 1-2\sum_{s_0}^\infty \exp(-c_2v_1^2 2^s) \geq 1 - 2\exp(-c_32^{s_0}v_1^2) \geq 1 -2\exp(-c_3v_1^2 \log |E_\ell|), $$ for every $f \in F$ \begin{align*} \sum_{i=s_0+1}^\infty \left| Z^\ell_{\pif{s+1}} - Z^\ell_{\pif{s}} \right| & \leq v_1 \sqrt{\ell} \sum_{i=s_0+1}^\infty 2^{s/2} \| \pif{s+1} -\pif{s} \|_{\psi_2} \\ & \leq c_4 v_1 \sqrt{\ell} \gamma_2(F,\psi_2). \end{align*} To handle $F_{s_0}=\{\pif{s_0} : f \in F \}$, note that the cardinality of this set is at most $2^{2^{s_0}} \leq 2^{2 \log |E_\ell|}$. Applying Bernstein's inequality (Lemma \ref{lemma:Bern}), for every $t>0$ and every $f \in F$, \begin{align*} Pr \left( \left| Z_f^\ell \right| \geq t\ell \right) & \leq |E_\ell| Pr \left( \left| \sum_{i=1}^\ell f(X_i) \right| \geq t \ell \right) \\ & \leq 2 |E_\ell| \exp \left(-c \ell \min \left\{ \frac{t^2}{\|f\|_{\psi_1}^2}, \frac{t}{\|f\|_{\psi_1}} \right\} \right). \end{align*} Let $t=\frac{\log |E_\ell|}{\ell} \|f\|_{\psi_1} v_2$ for $v_2 \geq 1$. Since $1 \leq \ell < k$ then $\ell^{-1} \log |E_\ell| \geq 1$ and $t \geq \|f\|_{\psi_1}$. Therefore, with probability at least $$ 1-2|E_\ell|^2 \exp \left(-c_5v_2 \log |E_\ell|\right) \geq 1-2\exp(-\log |E_\ell| (c_5v_2 - c_6) ), $$ for every $ f \in F_{s_0}$, $$ Z_f^\ell \leq v_2 \|f\|_{\psi_1} \log |E_\ell| \leq v_2 {\rm diam}(F,\psi_1) \log |E_\ell|. $$ To conclude, there are absolute constants $c_7$, $c_8$ and $c_9$ such that for every $1 \leq \ell \leq k$, if $v_1,v_2 \geq c_7$, with probability at least $1-2\exp(-c_8 \log |E_\ell| \min\{v_1^2,v_2\})$, $$ \sup_{f \in F} |Z_f^\ell| \leq c_9 \left(v_1 \sqrt{\ell} \gamma_2(F,\psi_2) + v_2 {\rm diam}(F,\psi_1) \log |E_\ell|\right). $$ Summing the probabilities, the latter holds for every $1 \leq \ell \leq k$ with probability at least $1-\exp(-c_{10} \min\{v_1^2,v_2\})$, completing the proof. {\mbox{}\nolinebreak\hfill\rule{2mm}{2mm}\par\medbreak} The proof of Theorem \ref{thm:sets-psi_2} is similar and is omitted. \vskip 0.2cm \noindent{\bf Proof of Theorem B.} In the $\psi_1$ case, take $v_1=\sqrt{\log(1/\delta)}$ and $v_2=\log(1/\delta)$ for $\delta$ small enough. Fix any $f \in F$ and for $I \in E_\ell$ let $I^+(f)=\{i : f(X_i)>0\} \cap I$ and $I^-(f)=\{i : f(X_i)<0\} \cap I$. Then, by Theorem \ref{thm:sets}, with probability at least $1-\delta$, \begin{align*} \sum_{i \in I} |f(X_i)| & = \big| \sum_{i \in I^+(f)} f(X_i) \big| + \big| \sum_{i \in I^-(f)} f(X_i)\big| \\ & \leq c(\delta) \left(\sqrt{\ell}\gamma_2(F,\psi_2)+{\rm diam}(F,\psi_1) \ell \log\left(\frac{ek}{\ell}\right)\right), \end{align*} as claimed. The $\psi_2$ case is equally easy. {\mbox{}\nolinebreak\hfill\rule{2mm}{2mm}\par\medbreak} For Theorem \ref{thm:sets} one can derive the following uniform empirical tail estimate for functions in $F$, which was formulated as Theorem A in the introduction. \begin{Corollary} \label{cor:tail-of-sets} There exist absolute constants $c_1$, $c_2$ and $c_3$ for which the following holds. Let $F$ be as in Theorem \ref{thm:sets}. For every $v_1,v_2 \geq c_1$, with probability at least $1-\exp(-c_2 \min\{v_1^2,v_2 \} )$, for any $f \in F$ and $t >0$, \begin{equation} \label{eq:theta-est} \left| \left\{ i : \left|f(X_i)\right| \geq t \right\} \right| \leq \max \left\{ \frac{c_3v_1^2 \gamma_2^2(F,\psi_2)}{t^2}, ek\exp \left(-\frac{t}{c_3\alpha v_2}\right) \right\}, \end{equation} where $\alpha = {\rm diam}(F,\psi_1)$. \end{Corollary} \noindent {\bf Proof.}\ \ Fix $v_1,v_2$ as in Theorem \ref{thm:sets} and consider the set for which the assertion of Theorem \ref{thm:sets} holds. Let $(X_1,...,X_k)$ be in that set and for $f \in F$ and $t>0$ put $$ I_t(f) =\{i : |f(X_i)| \geq t \}. $$ Setting $\alpha= {\rm diam}(F,\psi_1)$ there are two possibilities. First, if $$ 2\alpha v_2 | I_t(f)| \log \left(\frac{ek}{|I_t(f)|}\right) \leq \frac{t}{2} |I_t(f)|, $$ then by Theorem \ref{thm:sets}, \begin{align*} t |I_t(f)| & \leq 2v_1 \sqrt{|I_t(f)|} \gamma_2(F,\psi_2) + 2v_2 \alpha |I_t(f)| \log \left(\frac{ek}{|I_t(f)|}\right) \\ & \leq 2v_1\sqrt{|I_t(f)|} \gamma_2(F,\psi_2) + \frac{t}{2} |I_t(f)|. \end{align*} Thus, $t |I_t(f)|/2 \leq 2v_1 \sqrt{|I_t(f)|} \gamma_2(F,\psi_2)$, implying that $$ |I_t(f)| \leq 16v_1^2 \frac{\gamma_2^2(F,\psi_2)}{t^2}. $$ Otherwise, $2\alpha v_2 |I_t(f)| \log(ek/|I_t(f)|) \geq t |I_t(f)|/2$, or in other words, $$ |I_t(f)| \leq ek \exp\left(-\frac{t}{4v_2 {\rm diam}(F,\psi_1)}\right). $$ {\mbox{}\nolinebreak\hfill\rule{2mm}{2mm}\par\medbreak} Now we are ready to formulate and prove the main theorem of this section, which is a decomposition result for the class $F$. \begin{Theorem} \label{thm:main} There exist absolute constants $c_1$, $c_2$ and $c_3$, and for every $1 \leq p<\infty$ there exists a constants $c_4(p)$ for which the following holds. Let $F$ be a class of mean zero functions. For $v \geq c_1$, $A \geq \gamma_2(F,\psi_2)$, $B \geq {\rm diam}(F,\psi_1)$ and an integer $k$, set $$ \theta \geq \max\left\{c_2vB \log \left(c_2\frac{B^2k}{A^2}v+1\right), c_2pB\log(c_2pB+1) \right\}. $$ Then, there are Lipschitz functions $\phi:\mathbb{R} \to \mathbb{R}$ and $\psi:\mathbb{R} \to \mathbb{R}$ which depend on $\theta$, such that $\|\phi\|_{{\rm lip}}, \ \|\psi\|_{{\rm lip}} \leq 1$ and setting $F_{1}=\{\phi(f) : f \in F\}$ and $F_{2}=\{\psi(f) : f \in F\}$, \begin{description} \item{1.} $F \subset F_{1} + F_{2}$. \item{2.} For every $h \in F_{1}$, $\|h\|_{\infty} \leq \theta$ and for every $h \in F_2$, $\mathbb{E}|h|^p \leq A^2/k$ . \item{3.} With probability at least $1-\exp(-c_3v)$, $$ \sup_{h \in F_{2}} \sum_{i=1}^k |h(X_i)|^p \leq c_2vA^2\left(\theta^{p-2}+\kappa_p\right), $$ where $\kappa_p =c_4(p)\theta^{p-2}$ for $p<2$, $\kappa_2=c_4(2)\log A$, while for $p>2$, $\kappa_p=c_4(p)A^{p-2}$. \end{description} \end{Theorem} Theorem \ref{thm:main} implies that $F$ can be decomposed into two simple sets $F_{1}$ and $F_{2}$ (which depend on $k,p$ and $v$). The fact that these sets are as simple as $F$ is evident because they are images of $F$ via Lipschitz functions with constant $1$. In particular, $\gamma_\beta(F_{i},d) \leq \gamma_\beta(F,d)$ with respect to any reasonable metric $d$. The sets $F_{1}$ and $F_{2}$ have additional properties. $F_{1}$ has a bounded diameter in $L_\infty$ - up to a logarithmic term, its diameter in $L_\infty$ is proportional to the $\psi_1$ diameter of $F$. Thus, if $F$ has a well bounded diameter with respect to the $\psi_1$ metric then functions in $F_{1}$ are highly concentrated around their means, and one can safely use a contraction argument when bounding the empirical process indexed by a power of $F_{1}$. The main difficulty is in controlling the ``large part" of $F$ - i.e. $F_{2}$. The empirical process indexed by $F_{2}^p$ is small not because of concentration, but because the $\ell_p^k$ diameter of a random coordinate projection of $F_{2}$ and its $L_p$ diameter are small. \vskip0.2cm \noindent {\bf Proof of Theorem \ref{thm:main}.} Fix an integer $k$ and $v$ for which Corollary \ref{cor:tail-of-sets} holds. The first step is to select the Lipschitz functions $\phi$ and $\psi$; those are simply truncation functions at the level $\theta$. For $f \in F$, set $\phi(f)={\rm sgn}(f)\min\{|f|,\theta\}$ and $\psi(f)=f-\phi(f)$. Clearly, both functions have Lipschitz constant $1$, $F \subset \phi(F)+\psi(F)$, and for $p>1$, because $\phi(f)$ and $\psi(f)$ are supported on disjoint sets, \begin{align*} |f|^p & = \min\{|f|^p,\theta^p\}+\left(|f|^p - \theta^p\right) 1_{\left\{|f| \geq \theta \right\}} \\ & \leq |\phi(f)|^p + |f|^p1_{\{|f| \geq \theta\}}. \end{align*} Let $A \geq \gamma_2(F,\psi_2)$ and $B \geq {\rm diam}(F,\psi_1)$. It is evident that \begin{equation} \label{eq:p-moment} \mathbb{E} |f|^p 1_{\{|f| \geq \theta\}} \leq c_1(2pB)^p \exp\left(-\frac{\theta}{c_2B}\right) \leq \frac{A^2}{k}, \end{equation} as long as \begin{equation} \label{eq:theta0a} \theta \geq (c_3pB)\log (c_3pB) + c_3B \log \left(\frac{B^2k}{A^2}\right), \end{equation} which is satisfied by our choice of $\theta$. Thus (2) is established. Turning to (3), recall that for every $t>0$ and $f \in F$, $I_t(f)=\{i: |f(X_i)| \geq t \}$. By our choice of $v$, with probability at least $1-\exp(-c_4v)$, for every $f \in F$, for every $t>0$ \begin{equation} \label{eq:I(f)} |I_t(f)| \leq \max \left\{ \frac{c_5v A^2}{t^2}, k\exp \left(-\frac{t}{c_5 B v}\right) \right\}. \end{equation} Therefore, if \begin{equation} \label{eq:theta0b} t \geq t_0 = c_6 B v \log(c_6B v \sqrt{k}/A) \end{equation} for a suitable absolute constant $c_6$, the first term in \eqref{eq:I(f)} is dominant. Note that if $t \geq \max\{t_0,\sqrt{c_5v} A\}$ then $|I_t(f)|=0$, and since $\theta \geq t_0$ then by a standard integration argument with respect to the random empirical measure $\mu_k$, with probability at least $1-\exp(-c_4v)$, for every $f \in F$ \begin{align*} \mathbb{E}_{\mu_k} |f|^p 1_{\{|f| \geq \theta\}} & \leq \theta^p Pr_{\mu_k} \left(|f| \geq \theta \right) + \int_{\theta}^{\sqrt{c_5v} A} pt^{p-1} Pr_{\mu_k} \left(|f| > t\right) dt \\ & \leq c_6 \frac{v A^2}{k}\left( \theta^{p-2} + \int_{\theta}^{\sqrt{c_5v}A} pt^{p-3} dt \right), \end{align*} for which the claim follows. {\mbox{}\nolinebreak\hfill\rule{2mm}{2mm}\par\medbreak} \begin{Remark} \label{rem:F_{1,k}} Note that Theorem \ref{thm:main} enables one to bound $\|\mu_k-\mu\|_{F_{1}}$ and thus $\|\mu_k-\mu\|_{F}$. Indeed, pointwise, $\left|\left(\phi(f)\right)^p - \left(\phi(g)\right)^p \right| \leq 2 p \theta^{p-1} |f-g|$, implying that $$ \gamma_2\left(\left(\phi(F)\right)^p,\psi_2\right) \leq cp \theta^{p-1} \gamma_2(F,\psi_2). $$ By a standard generic chaining argument (see Theorem \ref{thm:generic-chaining} and \cite{Tal:book}), for every $v>0$, with probability at least $1-\exp(-c_1v^2)$, \begin{align*} \sup_{f \in F} \left| \frac{1}{k} \sum_{i=1}^k \left(\phi(f)\right)^p(X_i) -\mathbb{E} \left(\phi(f)\right)^p \right| \leq & c_2 v \frac{\gamma_2((\phi(F))^p,\psi_2)}{\sqrt{k}} \\ \leq & c_3p\theta^{p-1}v \frac{\gamma_2(F,\psi_2)}{\sqrt{k}}. \end{align*} \end{Remark} Combining this with Theorem \ref{thm:main}, it follows that with probability at least $1-2\exp(-c_1v)$, $$ \|\mu_k-\mu\|_F \leq c_2 v\left(p \theta^{p-1} \frac{\gamma_2(F,\psi_2)}{\sqrt{k}}+ \frac{\gamma_2^2(F,\psi_2)}{k} \left(\theta^{p-2}+\kappa_p+1\right)\right). $$ \begin{Remark} \label{rem:p-momnet} Observe that by \eqref{eq:p-moment}, $\sup_{h \in F_2} \mathbb{E} |h|^p \leq (cpB)^p \exp\left(-\frac{\theta}{cB}\right)$, a fact we shall use below. \end{Remark} We end this section with another observation which follows easily from the proof of Theorem \ref{thm:main}. To avoid complications, we will formulate it only is the case we need it, which is when $F$ is a class of linear functionals on $\mathbb{R}^n$ and $\mu$ is a measure on $\mathbb{R}^n$. Consider the random variable $U=\sup_{f \in F} |\inr{f,X}|$, and for every integer $k$ set $H_k = \mathbb{E} \max_{1 \leq i \leq k} U_i$, where $(U_i)_{i=1}^k$ are independent copies of $U$. \begin{Theorem} \label{thm:main-trunk} For every $p \geq 1$ and $0<\delta,\varepsilon<1$ there are constants $c_1(\delta,\varepsilon,p)$, $c_2(\delta,p)$ and $c_3(p)$ for which the following holds. Let $F$ and $\mu$ be as above, consider the random variable $Y^k=X1_{\{U \leq c_1(\delta,\varepsilon,p)H_k\}}$ and let $\nu$ be the probability measure on $\mathbb{R}^n$ corresponding to $Y^k$. If $A \geq \gamma_2(F,\psi_2(\nu))$ and $B \geq {\rm diam}(F,\psi_1(\nu))$, then with probability at least $1-\delta$, \begin{equation} \label{eq:main-trunk} \|\mu_k-\mu\|_{F^p} \leq c_2 \left(\theta^{p-1}\frac{\gamma_2(F,\nu)}{\sqrt{k}}+\frac{\gamma_2^2(F,\nu)}{k}\left(\theta^{p-2} + \tilde{\kappa}_p \right) \right)+c_3B^{1/2}\varepsilon, \end{equation} where $$ \theta=\max\left\{c_2B \log\left(c_2\frac{k B^2}{A^2}+1\right),c_2pB\log(c_2pB+1)\right\}, $$ and $\tilde{\kappa}_p=1$ for $1 \leq p < 2$, $\tilde{\kappa}_2=\log{H_k}$ and $\tilde{\kappa}_p=H_k^{p-2}$ for $p>2$. \end{Theorem} Because the proof is based on the same arguments used in Theorem \ref{thm:main} and Remark \ref{rem:F_{1,k}}, we will only give a brief sketch of the required modifications which are that with high probability, $X_i=Y_i$ for $1 \leq i \leq k$ and that by the Cauchy-Schwarz inequality, $$ \sup_{f \in F} \left| \mathbb{E}_\mu |f|^p -\mathbb{E}_{\nu}|f|^p \right| =\sup_{f \in F} \mathbb{E} |f|^p 1_{\{U \geq c_1(\delta,\varepsilon,p)H_k\}} \leq c_3(p)B^{1/2}\varepsilon, $$ for the right choice of constants. Thus, one can replace the measure $\mu$ with the measure $\nu$ and consider the empirical process $\|\nu-\nu_k\|_{F^p}$ instead of $\|\mu-\mu_k\|_{F^p}$. The advantage of using the measure $\nu$ is that it is a truncated version of $\mu$ at the ``correct" level for $F$ and the sample size $k$. This truncation enables us to bound $\gamma_2(S^{n-1},\nu)$, where $\nu$ is a truncation of an isotropic, log-concave measure on $\mathbb{R}^n$. \section{Applications} \label{sec:applications} The first geometric application we present deals with sections of small diameter of a convex, symmetric body. Let $\Gamma=\sum_{i=1}^k \inr{X_i,\cdot} e_i$, where $X_1,...,X_k$ are selected according an isotropic log concave measure. As we show below, if $K$ is a convex, symmetric body in $\mathbb{R}^n$, then with high probability, the diameter of $K \cap {\rm ker}(\Gamma)$ is small. This extends a celebrated result of Pajor and Tomczak-Jaegermann \cite{PT} which was proved in the case where the random subspace was selected according to the Haar measure on the Grassmann manifold ${\cal G}(n,k)$, but the same proof works in the Gaussian case. Various versions and extensions of this result may be found, for example, \cite{MiPi,Mi,MiPa}. The following theorem is a formulation of version of this result for a general $\psi_2$ operator (see \cite{MPT}). Let us introduce the following notation: for a set $T \subset \mathbb{R}^n$ we denote by $\ell_*(T)=\mathbb{E} \sup_{t \in T} \left|\sum_{i=1}^n g_i t_i \right|$, the expectation of the supremum of the gaussian process indexed by $T$. Recall that by the majorizing measures Theorem \cite{Tal:book}, there are absolute constants $c_1$ and $c_2$ such that for every $T \subset \mathbb{R}^n$, \begin{equation} \label{eq:MM} c_1 \gamma_2(T,\| \ \|) \leq \ell_*(T) \leq c_2 \gamma_2(T,\| \ \|). \end{equation} \begin{Theorem} \label{thm:MPT1} \cite{MPT} There exists a absolute constant $c$ and $c_1$ for which the following holds. Let $X_1,...,X_k$ be distributed according to an isotropic measure $\mu$ on $\mathbb{R}^n$ and assume that for every $t \in \mathbb{R}^n$, $\|\inr{t,\cdot}\|_{\psi_2} \leq \alpha \|t\|$ for some $\alpha \geq 1$. If $K \subset \mathbb{R}^n$ is a convex symmetric body then with probability at least $1-\exp(-c_1k/\alpha^4)$, $$ {\rm diam}({\rm ker} \Gamma \cap K) \leq r_k^*(K), $$ where $$ r_k^*(K) = \inf\left\{\rho>0 : \rho \geq c_1\alpha^2\ell_*(K\cap \rho S^{n-1})/\sqrt{k} \right\}. $$ \end{Theorem} Our result is similar (though with a weaker estimate) to Theorem \ref{thm:MPT1}. Other than the different ways of estimating the empirical process $\|\mu_k-\mu\|_{F^2}$, the two proofs are identical, and thus the proof of Theorem \ref{thm:PT-log-concave} is omitted. \begin{Theorem} \label{thm:PT-log-concave} For every $0<\delta<1$ there exist constant $c(\delta)$ for which the following holds. Let $\mu$ be an isotropic, log-concave measure on $\mathbb{R}^n$ and let $K \subset \mathbb{R}^n$ be a convex symmetric body. If $X_1,...,X_k$ are independent, distributed according to $\mu$, then with probability at least $1-\delta$, $$ {\rm diam}({\rm ker} \Gamma \cap K) \leq q_k^*(K), $$ where $$ q_k^*(K) = \inf\left\{\rho>0 : \rho \geq c(\delta)\gamma_2(K\cap \rho S^{n-1} ,\psi_2)\sqrt{\frac{\gamma_2(K\cap \rho S^{n-1} ,\psi_2)}{k}} \right\}. $$ \end{Theorem} If $\mu$ is a subgaussian measure then for every $A \subset \mathbb{R}^n$, $\gamma_2(A,\psi_2) \leq c\ell_*(A)$, and thus $\gamma_2(K\cap \rho S^{n-1} ,\psi_2) \leq \sqrt{n}$. Therefore, Theorem \ref{thm:PT-log-concave} recovers Theorem \ref{thm:MPT1} up to a $\sqrt{\log{n}}$ factor. Of course, the bound on the probability is considerably weaker. On the other hand, Theorem \ref{thm:PT-log-concave} holds for a much wider family of measures because the bound given in Theorem \ref{thm:MPT1} depends on the equivalence constant between the $\psi_2$ and $\ell_2$ metrics endowed on $\mathbb{R}^n$. \subsection{Sampling from an isotropic, log-concave measure} A question which was originally studied in \cite{KLS,Bour,Rud,GiaMil,GueRud} is the following: how many points sampled from an isotropic, convex, symmetric body are needed to ensure that the random operator $k^{-1}\sum_{i=1}^k \inr{X_i,\cdot}e_i$ is an almost isometric embedding of $\ell_2^n$ in $\ell_2^k$? In other words, that with probability at least $1-\delta$, for every $\theta \in S^{n-1}$, $$ 1-\varepsilon \leq \frac{1}{k} \sum_{i=1}^k \inr{X_i,\theta}^2 \leq 1+\varepsilon. $$ \begin{Theorem} \label{thm:best-prev-est} \cite{Rud,GiaMil,GueRud} For every $0<\varepsilon,\delta<1$ there is a constant $c(\varepsilon,\delta)$ for which the following holds. Let $X_1,...,X_k$ be independent random variables, distributed according to the volume measure of a convex, symmetric body in isotropic position. If $k \geq c(\varepsilon,\delta)n\log^2n$, then with probability at least $1-\delta$, for every $\theta \in S^{n-1}$, $$ 1-\varepsilon \leq \frac{1}{k} \sum_{i=1}^k \inr{\theta,X_i}^2 \leq 1+\varepsilon. $$ \end{Theorem} The estimate of $k \sim n \log^2n $ was first proved by Rudelson \cite{Rud}. Previously, Bourgain showed \cite{Bour} how to obtain this result with a slightly weaker estimate of $k \sim n \log^3n$, but then Giannopoulos and Milman \cite{GiaMil} demonstrated that Bourgain's method can actually give the same estimate as Rudelson's. The proofs of Bourgain and Rudelson use very different arguments. Rudelson's proof is based on a noncommutative Khintchine inequality, due to Lust-Piquard and Pisier \cite{LuPi}, namely, a bound on Rademacher averages of the form $\mathbb{E} \|\sum_{i=1}^k \varepsilon_i x_i \otimes x_i\|^p_{\ell_2 \to \ell_2}$ for $p \geq 1$, where $(x_i)_{i=1}^k \in \ell_2$ and $(\varepsilon_i)_{i=1}^k$ are independent, symmetric, $\{-1,1\}$-valued random variables. The fact that the set indexing the empirical process is exactly the Euclidean sphere is essential in the proof and the argument can not be modified to handle any other indexing sets - not even other subsets of the sphere. Bourgain's proof uses a similar technique to the one we used here, which relies on the following version of Theorem \ref{thm:sets}. The formulation we present here is from \cite{GiaMil}. \begin{Lemma} \label{lemma:Bour} Let $\delta \in (0,1)$ and let $X_1,...,X_k$ be points in $\mathbb{R}^n$ sampled according to an isotropic log-concave measure. If $k \leq c\delta \exp(\sqrt{n})$ then with probability at least $1-\delta$, for every $I \subset \{1,...,k\}$, $$ \left\|\sum_{i \in I} X_i \right\| \leq c_1(\delta) \left(\sqrt{\log{k}}\sqrt{|I|}\sqrt{n}+|I| \log{k} \right). $$ In particular, with probability at least $1-\delta$, for every $t \geq c(\delta)\log{k}$ and every $x \in S^{n-1}$, \begin{equation} \label{eq:tail1} \left|\left\{ i : \inr{x,X_i} \geq t\right\} \right| \leq c_2(\delta)\frac{n\log{k}}{t^2}. \end{equation} \end{Lemma} Bourgain's method was generalized in \cite{GiaMil} in which the following theorem was established: \begin{Theorem} \label{thm:GiaMil} Let $p>0$ and $0 < \delta < 1$. There exists $n_0(\delta)$ such that for every $n \geq n_0(\delta)$, every log-concave measure on $\mathbb{R}^n$, every $k \geq k_0(\delta,p)$ and every $\theta \in S^{n-1}$, $$ c_p \leq \frac{1}{k}\sum_{i=1}^k |\inr{X_i,\theta}|^p \leq C_p $$ where $c_p$ and $C_p$ depend only on $p$ and \begin{equation*} k_0(\delta,p) =c(\delta,p) \begin{cases} n & \text{if $0<p<1$}, \\ n\log^p{n} & \text{if $1 \leq p \leq 2$}, \\ \min\{(p-2)^{-1},\log{n}\} (n\log{n})^{p/2} & \text{if $p>2$}. \end{cases} \end{equation*} \end{Theorem} Note that this bound is isomorphic in nature rather than almost isometric, though in the case $p=2$ the proof of Theorem \ref{thm:GiaMil} can be modified to give an almost isometric estimate. Recently, Gu\'{e}don and Rudelson \cite{GueRud} were able to bound $$ \mathbb{E}_\varepsilon \sup_{y \in K} \sum_{i=1}^k \varepsilon_i |\inr{x_i,y}|^p, $$ for any $x_1,...,x_k \in \mathbb{R}^n$, where $K \subset B_2^n$ is a convex, symmetric body which has a $q$-power type modulus of convexity. The method of proof is based on majorizing measures, and can be used to bound $\mathbb{E}\|\mu_k-\mu\|_{F^p}$ for $F=\{\inr{x,-}:x \in K\}$ as long as $p \geq q \geq 2$. It turns out that the dominant factor in the bound is $$ \left(\mathbb{E} \max_{1 \leq i \leq k}\|X_i\|^2 \cdot \mathbb{E} \max_{1 \leq i \leq k}\|X_i\|^{p-2}_{K^\circ}\right)^{\frac{1}{p}}. $$ For $K=B_2^n$ this approach yields the best known estimates for $\mathbb{E}\|\mu_k-\mu\|_{F^p}$ for $p \geq 2$, and the resulting estimate on the required size of the sample is $k \sim c(\varepsilon,\delta,p)n^{p/2}\log{n}$, and in particular, for $p=2$ gives the best known estimate of $k \sim c(\varepsilon,\delta)n\log{n}$. Let us mention that for $p=2$ this result is not helpful for ``small" subsets of the sphere, and the best bound that one can establish for such subsets coincides with the one obtained for the whole sphere. All the known bounds, including \cite{GueRud} and ours, are based on the behavior of the random variable $\|X\|$. The best estimates on $\|X\|$ are due to Paouris \cite{Pao}: \begin{Theorem} \label{thm:Paouris} There are absolute constants $c_1$ and $c_2$ for which the following holds. Let $X$ be distributed according to an isotropic log-concave measure on $\mathbb{R}^n$. Then, for every $p \leq c_1\sqrt{n}$, $(\mathbb{E}\|X\|^p)^{1/p} \leq c_2 \sqrt{n}$. \end{Theorem} Theorem \ref{thm:Paouris} immediately leads to a removal of a logarithmic factor in \eqref{eq:tail1}, though not to an improved level of truncation; thus, the estimate of Theorem \ref{thm:GiaMil} in the case $p=2$ remains unchanged despite the improved tail estimate. The properties of an isotropic log-concave measure which will be used below are that for suitable absolute constants $C$ and $C_1$, \begin{description} \item{1.} linear functionals have a subexponential tail - that is, for every $x \in \mathbb{R}^n$, $$ \|\inr{X,\cdot}\|_{\psi_1} \leq C\|x\|, $$ and \item{2.} By Theorem \ref{thm:Paouris}, for $n \leq k \leq \exp(c\sqrt{n})$, $$ \mathbb{E}\max_{1 \leq i \leq k} \|X_i\| = \mathbb{E}\max_{1 \leq i \leq k} \sup_{x \in S^{n-1}} |\inr{x,X_i}| \leq c_1\sqrt{n}. $$ \end{description} Therefore, in light of Theorem \ref{thm:main-trunk}, it is enough to consider the truncated measure $\nu$ on $\mathbb{R}^n$ which is supported on a ball of radius $c_2(\delta)\sqrt{n}$ to bound $\|\mu_k-\mu\|_{F^p}$. The main ingredient in our method is to bound $\gamma_2(S^{n-1},\nu)$ and to that end we shall estimate $$ \ell_E = \mathbb{E}\|\sum_{i=1}^n g_ie_i\|_E, $$ where $g_1,...,g_n$ are standard, independent Gaussian variables. The particular norm $\| \ \|_E$ we consider is the one endowed on $\mathbb{R}^n$ by the $\psi_2(\nu)$ structure, formally defined for every $t \in \mathbb{R}^n$ by $\|t\|_E = \|\inr{t,Y}\|_{\psi_2}$. \begin{Lemma} \label{lemma:M-estimate} There exists an absolute constant $c$ for which the following holds. Let $\nu$ be a probability measure on $\mathbb{R}^n$ and set $Y$ to be distributed according to $\nu$. If $Z=\|Y\|$ and $E=(\mathbb{R}^n, \| \ \|_{\psi_2})$, then $\ell_{E} \leq c\|Z\|_{\infty}$. \end{Lemma} \noindent {\bf Proof.}\ \ Fix $\rho$ to be named later and consider the gaussian vector $G=(g_1,...,g_n)$, where $(g_i)_{i=1}^n$ are independent, standard gaussian random variables. Let $\|Z\|_\infty =D $ and since $\|f\|_{\psi_2} \leq \mathbb{E} \exp(f^2)$ then \begin{align*} \frac{\ell_E}{\rho} \leq & \mathbb{E}_Y \mathbb{E}_g \exp \left(\frac{\sum_{i=1}^n g_i \inr{e_i,Y}}{\rho} \right)^2. \end{align*} Recall that $\sum_{i=1}^n g_i \inr{Y,e_i}$ is distributed as $g\|Y\|$, and thus, \begin{align*} \frac{\ell_E}{\rho} \leq & \mathbb{E}_Y \mathbb{E}_g \left(1+\sum_{m=1}^\infty \frac{\left(\sum_{j=1}^n g_i \inr{Y,e_i} \right)^{2m}}{m! \rho^{2m}} \right) \\ \leq & \left(1+\sum_{i=1}^\infty \frac{1}{m!} \mathbb{E}|g|^{2m}\left( \frac{D}{\rho}\right)^{2m}\right) \\ = & \mathbb{E} \exp \left(\frac{Dg}{\rho } \right)^2 \leq 2, \end{align*} if one selects $\rho=cD$. Therefore, $\ell_E \leq cD$, as claimed. {\mbox{}\nolinebreak\hfill\rule{2mm}{2mm}\par\medbreak} \begin{Definition} For two sets $A,B \subset \mathbb{R}^n$, let $N(A,B)$ be the minimal number of translates of $B$ needed to cover $A$, that is, the minimal cardinality of a set $\{x_1,...,x_m\}$ such that $A \subset \bigcup_{i=1}^m (B+x_i)$. \end{Definition} Note that if $B$ is a unit ball of a norm on $\mathbb{R}^n$ then $N(A,\varepsilon B)$ are the covering numbers of $A$ with respect to the metric endowed by $B$. \begin{Corollary} \label{cor:entropy-est} There exists an absolute constant $c$ such that for every $\varepsilon \geq 1/2$, $$ \log N(B_2^n, \varepsilon B_E) \leq c\frac{n}{\varepsilon^2}, $$ and for $0<\varepsilon<1/2$, $$ \log N(B_2^n, \varepsilon B_E) \leq cn\log \left(\frac{1}{\varepsilon}\right), $$ where $B_E$ is the unit ball of $(\mathbb{R}^n,\psi_2(\nu))$ and $B_2^n$ is the Euclidean unit ball. \end{Corollary} \noindent {\bf Proof.}\ \ Recall that if $Z=\|Y\|$ then $\|Z\|_{\infty} \leq c_1 \sqrt{n}$ for a suitable absolute constant. By the dual Sudakov Theorem \cite{PajTom2}, $\log N(B_2^n, \varepsilon B_E) \leq c_2\ell_E/\varepsilon^2$, and applying Lemma \ref{lemma:M-estimate}, $\ell_E \leq c_3\sqrt{n}$, from which the first part of the claim follows. Turning to the second part, by a standard volumetric estimate (see, e.g. \cite{Pis}), and since $B_E$ is a unit ball of a norm on $\mathbb{R}^n$, $N(\frac{1}{2}B_E,\varepsilon B_E) \leq (1/2\varepsilon)^n$. Therefore, $$ N(B_2^n,\varepsilon B_E) \leq N\left(B_2^n, \frac{1}{2}B_E\right) \cdot N\left(\frac{1}{2}B_E, \varepsilon B_E \right) \leq \exp(c_4n) \left(\frac{1}{2\varepsilon}\right)^n. $$ {\mbox{}\nolinebreak\hfill\rule{2mm}{2mm}\par\medbreak} Using Corollary \ref{cor:entropy-est} one can bound $\gamma_2(F,\psi_2(\nu))$ by applying Theorem \ref{thm:talagrand} for the space $\ell_2^n$ which is $2$-convex, and with $d$ being the $\psi_2(\nu)$ metric endowed on $\mathbb{R}^n$. \begin{Corollary} \label{cor:gamma-est} Let $\mu$ be an isotropic log-concave measure on $\mathbb{R}^n$ and set $\nu$ to be its truncation as above. Then for $n \leq k \leq \exp(c_1\sqrt{n})$, $$ \gamma_2(S^{n-1}, \psi_2(\nu) ) \leq c_2\sqrt{n\log{n}}, $$ where $c_1$ and $c_2$ are absolute constants. \end{Corollary} \noindent {\bf Proof.}\ \ The proof is immediate from Theorem \ref{thm:talagrand}, the entropy estimate in Corollary \ref{cor:entropy-est}, combined with the fact that for every $\theta \in S^{n-1}$, $|\inr{Y,\theta}| \leq \|Y\| \leq c\sqrt{n}$, and thus ${\rm diam}(S^{n-1},\psi_2(\nu)) \leq c\sqrt{n\log{k}} \leq c\sqrt{n\log{n}}$. {\mbox{}\nolinebreak\hfill\rule{2mm}{2mm}\par\medbreak} Let us remark that we believe this estimate is suboptimal by a factor of $\sqrt{\log{n}}$. Combining Corollary \ref{cor:gamma-est} with Theorem \ref{thm:main-trunk}, we obtain the following (most likely, suboptimal) estimate of $\|\mu_k-\mu\|_{F^p}$, which we only state for $p > 2$. This estimate recovers the best known result for $p>2$, and was originally established in \cite{GueRud}. \begin{Theorem} \label{thm:p-sphere} For every $0<\varepsilon$, $0<\delta<1$ and $p >2$ there exists a constant $c(\varepsilon,\delta,p)$ for which the following holds. With probability at least $1-\delta$, if $k \geq k_0$, $$ \sup_{\theta \in S^{n-1}} \left| \frac{1}{k} \sum_{i=1}^k |\inr{X_i,\theta} |^p - \mathbb{E}|\inr{X,\theta}|^p \right| < \varepsilon, $$ provided that $k_0 \geq c(\varepsilon,\delta,p)n^{p/2}\log{n}$ for $p>2$. \end{Theorem} \noindent {\bf Proof.}\ \ Let $n \leq k \leq \exp(c_1\sqrt{n})$. Using the notation of Theorem \ref{thm:main-trunk}, observe that $\gamma_2(S^{n-1},\psi_2(\nu)) \leq c_2\sqrt{n\log{n}}$, ${\rm diam}(S^{n-1},\psi_1) \leq c_2$, $H_k \leq c_2\sqrt{n}$. Also, if $k \geq c_3n\log{n}$, then for $p > 2$, $\theta$ can be taken as $\theta \leq c_4 \log \log{n}$, from which the claim is evident. {\mbox{}\nolinebreak\hfill\rule{2mm}{2mm}\par\medbreak} Let us remark that if one could select $k \leq cn\log{n}$ it would be possible to take $\theta$ at the level of an absolute constant. This would be the case if the logarithmic term in the estimate on $\gamma_2(S^{n-1},\psi_2(\nu))$ were to be removed and would lead to the optimal bound for any $p>1$. \footnotesize {

https://arxiv.org/abs/1608.08975

Improved accuracy for time-splitting methods for the numerical solution of parabolic equations

In this work, we study time-splitting strategies for the numerical approximation of evolutionary reaction-diffusion problems. In particular, we formulate a family of domain decomposition splitting methods that overcomes some typical limitations of classical alternating direction implicit (ADI) schemes. The splitting error associated with such methods is observed to be $\mathcal{O}(\tau^2)$ in the time step $\tau$. In order to decrease the size of this splitting error to $\mathcal{O}(\tau^3)$, we add a correction term to the right-hand side of the original formulation. This procedure is based on the improved initialization technique proposed by Douglas and Kim in the framework of ADI methods. The resulting non-iterative schemes reduce the global system to a collection of uncoupled subdomain problems that can be solved in parallel. Computational results comparing the newly derived algorithms with the Crank-Nicolson scheme and certain ADI methods are presented.

\section{Introduction}\label{sec:introduction} The numerical approximation of parabolic problems using time-splitting procedures has been a wide field of research since the pioneering works of Douglas, Peaceman and Rachford in the decade of the 1950s (cf. \cite{dou:55,dou:pea:55,dou:rac:56,pea:rac:55}). In such works, they introduced the so-called alternating direction implicit (ADI) methods by noting that, in a $d$-dimensional spatial domain $\Omega$, the diffusion operator $-\nabla\cdot(a\nabla)$ can be expressed as the sum of $d$ one-dimensional operators $\{\partial_k(a\partial_k)\}_{k=1,2,\ldots,d}$, $a$ being a uniformly positive function on $\overline{\Omega}$ and $\partial_k=\partial/\partial x_k$. Using this idea, multidimensional parabolic problems can be solved as a sequence of one-dimensional problems, each formulated for one of the spatial variables under consideration. Any time-stepping procedure based on a component-wise splitting of this kind is called a locally one-dimensional method (see, e.g., \cite{hun:ver:03}). For an extensive study of ADI and related time-splitting methods, we refer to the monographs \cite{far:hav:09,hun:ver:03,mar:90}. Significantly, an ADI method can be viewed as a perturbation of some underlying implicit scheme, such as the backward Euler or the Crank--Nicolson method. In this setting, the ADI method may be formulated as the corresponding implicit scheme plus a perturbation term called the splitting error. In general, this splitting error is of the same --or higher-- order in the time step $\tau$ as the truncation error associated with the underlying unsplit method. As a result, the asymptotic rate of convergence for both the ADI and its underlying method should be the same; in practice, however, the actual errors associated with the former are much larger than those associated with the latter. This fact is due to the presence of the splitting error and, typically, it is considered to be the main drawback of time-splitting methods. In order to reduce the size of such an error, Douglas and Kim introduced in \cite{dou:kim:01} the so-called alternating direction method with improved initialization (AD-II) (sometimes referred to as the modified alternating direction (AD-M) method, see \cite{arb:hua:yan:07}). Essentially, they proposed to add a correction term to the right-hand side of the original ADI scheme with the aim of reducing the splitting error from $\mathcal{O}(\tau^2)$ to $\mathcal{O}(\tau^3)$. This idea was used in \cite{arb:hua:yan:07} to formulate improved ADI methods for regular and mixed finite element discretizations, and further studied in \cite{kar:09} to derive linear multistep methods by the approximate factorization technique. In this paper, we extend the improved initialization procedure of Douglas and Kim to the case of domain decomposition-based splittings. This kind of splittings was first introduced in \cite{vab:89,vab:94} to obtain the so-called regionally-additive schemes, and has been subsequently studied in \cite{mat:pol:rus:wan:98,por:buj:jor:04,vab:08} for the solution of evolutionary problems. In the context of linear and semilinear parabolic equations, it has been successfully used in combination with various spatial discretization techniques, such as mimetic finite differences (cf. \cite{arr:por:jor:09,por:arr:jor:11}), mixed finite elements (cf. \cite{arr:por:14}) or multipoint flux approximation methods (cf. \cite{arr:por:yot:14}). Some additional results regarding nonlinear parabolic equations can be found in \cite{arr:por:jor:10,por:arr:jor:10}. The monographs \cite{mat:08,sam:mat:vab:02} show an overview of some recent contributions to the topic. The key to the efficiency of domain decomposition splitting methods lies in reducing the system matrix to a collection of uncoupled submatrices of lower dimension. As compared to classical overlapping domain decomposition algorithms (cf. \cite{qua:val:99}), this approach does not involve any Schwarz iteration procedure, thus reducing the computational cost of the overall solution process. In addition, it overcomes two typical limitations of alternating direction splittings, namely: (a) their need to deal with rectangular or hexahedral spatial grids (in two- or three-dimensional problems, respectively); and (b) their difficulty to handle mixed derivative terms. In this respect, although several ADI methods have been specifically designed along the years to overcome this latter constraint (see, e.g., \cite{dou:gun:64,mck:mit:70,mck:wal:wil:96} or, more recently, \cite{int:mis:13,int:wel:07,int:wel:09}), no AD-II scheme with this property has been developed so far. In order to introduce the improved time-splitting procedures, we consider a parabolic initial-boundary value problem of the form \begin{subequations}\label{continuous:problem} \begin{align} u_t-\nabla\cdot(a\nabla u)+cu&=f,&&\hspace*{-1.5cm}\mathbf{x}\in\Omega, &&\hspace*{-1.5cm}0<t\le T,\label{continuous:problem:a}\\ u&=0,&&\hspace*{-1.5cm}\mathbf{x}\in\partial\Omega, &&\hspace*{-1.5cm}0<t\le T,\label{continuous:problem:b}\\ u&=u_0,&&\hspace*{-1.5cm}\mathbf{x}\in\Omega, &&\hspace*{-1.5cm}t=0,\label{continuous:problem:c} \end{align} \end{subequations} where $\Omega\subset\mathds{R}^2$ is a bounded open domain with boundary $\partial\Omega$, $a(\mathbf{x})\in\mathds{R}^{2\times2}$ is a symmetric positive definite matrix function, with elements $\{a_{ij}(\mathbf{x})\}_{i,j=1,2}$, $c(\mathbf{x})$ is a uniformly positive function on $\overline{\Omega}$, and the subscript $t$ denotes partial differentiation with respect to time. The entries of $a(\mathbf{x})$ and the terms $c(\mathbf{x})$, $f(\mathbf{x},t)$ and $u_0(\mathbf{x})$ are assumed to be sufficiently smooth. In the sequel, we denote by $Au=-\nabla\cdot(a\nabla u)+cu$ the elliptic operator applied to the exact solution $u(\mathbf{x},t)$. For simplicity in the exposition, we consider homogeneous Dirichlet boundary conditions, although more general conditions can also be handled. For the sake of completeness, it must be mentioned that the solution of problem (\ref{continuous:problem}) can also be approximated with a high order of accuracy by means of unsplit methods. Among the time-stepping schemes that have been recently designed for this purpose, it is worth referring to the so-called ADER (Arbitrary DERivative in space and time) approach. This method was initially developed to provide high-order approximations to the solution of hyperbolic conservation laws (see, e.g., \cite{dum:ena:tor:08,tit:tor:02}), and was subsequently formulated for nonlinear reaction--diffusion problems (cf. \cite{tor:hid:09}). More recently, it has been extended in \cite{hid:dum:11} to nonlinear systems of advection--diffusion--reaction equations involving stiff source terms. Some novel contributions to the topic can be found in \cite{mon:tor:14,tor:mon:14}. The rest of the paper is organized as follows. In \S\ref{sec:unsplit}, we formulate some classical unsplit implicit methods for the numerical solution of problem (\ref{continuous:problem}). Two time-splitting strategies are described in \S\ref{sec:split}: on one hand, the well-known Douglas (cf. \cite{dou:62}) and Douglas--Rachford (cf. \cite{dou:rac:56}) alternating direction methods; on the other, a non-iterative domain decomposition splitting technique based on a family of partition of unity functions. The splitting error associated with the preceding schemes is studied in \S\ref{sec:splitting:error}. In this section, we also introduce a correction procedure to reduce the size of such an error. Numerical experiments comparing the previous time-splitting schemes with the Crank--Nicolson method are reported in \S\ref{sec:experiments}. The paper ends with some concluding remarks summarized in \S\ref{sec:conclusions}. \section{Classical implicit methods}\label{sec:unsplit} Let us consider a suitable partition $\mathcal{T}_h$ of the spatial domain $\Omega$, where $h$ denotes the maximal grid spacing. In principle, $\mathcal{T}_h$ is an unstructured grid composed of either triangular or quadrilateral elements. We will see in the next section that, if an ADI method is used for solving (\ref{continuous:problem}), $\mathcal{T}_h$ must be assumed to be a rectangular grid. Domain decomposition splittings, by contrast, will remain valid for unstructured partitions. On the other hand, let us consider a constant time step $\tau>0$ and define the discrete times $t_n = n\tau$, for $n=0,1,\ldots,N_T+1$, with $N_T=[T/\tau]-1$. In this framework, let $A_h$ be a symmetric positive definite matrix obtained from finite difference or finite element discretization of the elliptic operator $A$ with order of accuracy $\mathcal{O}(h^s)$. Then, if $U^n_h$ denotes the fully discrete solution at time $t_n$, the classical implicit time-stepping schemes can be written together, for $n=0,1,\ldots,N_T$, in the form \begin{equation}\label{be:cn} \frac{U_h^{n+1}-U_h^n}{\tau}+A_h(\theta U_h^{n+1}+(1-\theta)U_h^n)=F_h^{n+\theta}, \end{equation} where $\theta=1$ for backward Euler and $\theta=\frac{1}{2}$ for Crank--Nicolson. Note that $F_h^{n+\theta}=\theta F_h(t_{n+1})+(1-\theta)F_h(t_{n})$, $F_h(t_n)$ being the discrete forcing term at time $t_n$. The initial condition is given by $U_h^0=\mathcal{P}_hu_0$, where $\mathcal{P}_h$ is a suitable restriction or projection operator. It is well known that, for a spatial discretization of order $\mathcal{O}(h^s)$, the backward Euler solution converges to the true solution with order $\mathcal{O}(\tau+h^s)$; in turn, the Crank--Nicolson solution converges with order $\mathcal{O}(\tau^2+h^s)$. In compact form, both results can be written together as $\mathcal{O}(\tau^{3-2\theta}+h^s)$. \section{Time-splitting methods}\label{sec:split} In this section, we introduce two classes of time-splitting procedures for the solution of problem (\ref{continuous:problem}), namely: (a) the classical alternating direction methods based on a component-wise splitting; and (b) a family of non-iterative schemes based on an overlapping domain decomposition splitting. \subsection{Alternating direction splitting} Let us assume that $a(\mathbf{x})$ in (\ref{continuous:problem:a}) is a diagonal matrix with elements $a_{11}(\mathbf{x})$ and $a_{22}(\mathbf{x})$. This assumption avoids the existence of mixed derivative terms in the parabolic problem. In this way, the elliptic operator $A$ can be expressed as the sum $A=A_1+A_2$, where \begin{equation}\label{adi:splitting} A_{1}u=-(a_{11}u_x)_x+\dfrac{1}{2}\,cu,\qquad A_{2}u=-(a_{22}u_y)_y+\dfrac{1}{2}\,cu. \end{equation} If the spatial domain $\Omega$ admits a rectangular grid $\mathcal{T}_h$, problem (\ref{continuous:problem}) can be approximated, for $n=0,1,\ldots,N_T$, by an ADI method of the form \begin{equation}\label{adi:scheme} \begin{aligned} \frac{V_h^{n,1}-V_h^n}{\tau}+A_{1h}(\theta V_h^{n,1}+(1-\theta)V_h^n)+A_{2h}V_h^n&=F_h^{n+\theta},\\ \frac{V_h^{n+1}-V_h^n}{\tau}+A_{1h}(\theta V_h^{n,1}+(1-\theta)V_h^n)+A_{2h}(\theta V_h^{n+1}+(1-\theta)V_h^n)&=F_h^{n+\theta}. \end{aligned} \end{equation} Here, $A_{1h}$ and $A_{2h}$ are suitable discretizations of $A_{1}$ and $A_{2}$, respectively, such that $A_h=A_{1h}+A_{2h}$. Note that we introduce the notation $V_h^n$ to distinguish the ADI discrete solution from the unsplit discrete solution $U_h^n$ of the preceding section. In this case, the initial condition is also given by $V_h^0=\mathcal{P}_hu_0$. The previous discretization scheme yields some classical ADI methods for certain values of the parameter $\theta$. More precisely, if $\theta=1$, we recover the Douglas--Rachford method, first proposed in \cite{dou:rac:56}; i.e., for $n=0,1,\ldots,N_T$, \begin{equation}\label{douglas:rachford} \begin{aligned} (I+\tau A_{1h})\,V_h^{n,1}&=(I-\tau A_{2h})\,V_h^n+\tau F_h^{n+1},\\ (I+\tau A_{2h})\,V_h^{n+1}&=V_h^{n,1}+\tau A_{2h}V_h^n, \end{aligned} \end{equation} for a given $V_h^0$, where $I$ denotes the identity matrix. In turn, if $\theta=\frac{1}{2}$, the Douglas method (cf. \cite{dou:62}) is obtained; i.e., for $n=0,1,\ldots,N_T$, \begin{equation}\label{douglas:gunn} \begin{aligned} \left(I+\dfrac{\tau}{2} A_{1h}\right)V_h^{n,1}&=\left(I-\dfrac{\tau}{2} A_{1h}-\tau A_{2h}\right)V_h^n+\tau F_h^{n+\frac{1}{2}},\\ \left(I+\dfrac{\tau}{2} A_{2h}\right)V_h^{n+1}&=\textstyle V_h^{n,1}+\dfrac{\tau}{2} A_{2h}V_h^n, \end{aligned} \end{equation} for a given $V_h^0$. \subsection{Domain decomposition splitting} As an alternative to component-wise splitting methods, we present a time-stepping procedure based on a suitable decomposition of the spatial domain. Unlike the ADI methods introduced above, this technique permits to handle mixed derivative terms (i.e., full tensor coefficients $a(\mathbf{x})$ in (\ref{continuous:problem:a})), and is also valid for unstructured partitions $\mathcal{T}_h$ of the spatial domain $\Omega$. In this case, we consider the splitting $A=A_{1}+A_{2}+\ldots+A_{m}$ into an arbitrary number $m$ of split terms, where \begin{equation}\label{dd:splitting} A_{k}u=\textstyle -\nabla\cdot(\rho_k a\nabla u)+\rho_k cu,\quad\mbox{for}\,\,k=1,2,\ldots,m. \end{equation} The family of functions $\{\rho_k(\mathbf{x})\}_{k=1,2,\ldots,m}$ is subordinate to an overlapping decomposition of $\Omega$ and conforms a smooth partition of unity. The construction of such a partition is discussed in the sequel. Let $\{\Omega_k^{\ast}\}_{k=1,2,\ldots,m}$ form a non-overlapping decomposition of $\Omega$ into $m$ subdomains. Such a decomposition fulfills the conditions $\overline{\Omega}=\bigcup_{k=1}^m\overline{\Omega}^{\ast}_k$ and $\Omega_{k}^{\ast}\cap\Omega_{l}^{\ast}=\emptyset$, for $k\neq l$. In turn, each $\Omega_{k}^{\ast}\subset\Omega$ is considered to be an open disconnected set involving $m_k$ connected components, i.e., \begin{equation*} \textstyle\Omega_k^{\ast}=\bigcup_{l=1}^{m_{k}}\Omega^{\ast}_{kl},\quad\mathrm{for}\,\,k=1,2,\ldots,m. \end{equation*} Such components are pairwise disjoint (that is, $\Omega_{ki}^{\ast}\cap\Omega_{kj}^{\ast}=\emptyset$, for $i\neq j$) and typically chosen to be shape regular of diameter $h_0$. For instance, the components $\Omega_{kl}^{\ast}$ may correspond to the elements in a coarse partition $\mathcal{T}_{h_0}$ of $\Omega$ with mesh size $h_0$. Let $\Omega_{kl}$ be the extension of $\Omega_{kl}^{\ast}$ obtained by suitably translating its internal boundaries, $\partial\Omega_{kl}^{\ast}\cap\Omega$, within a distance $\beta h_0$ in $\Omega$. The parameter $\beta>0$ is usually referred to as the overlapping factor and its value must be chosen in such a way that the extended components are also pairwise disjoint (i.e., $\Omega_{ki}\cap\Omega_{kj}=\emptyset$, for $i\neq j$). The distance $\xi=2\beta h_0$ is called the overlapping size. If we denote by $\Omega_k\subset\Omega$ the open disconnected set defined as \begin{equation}\label{disjoint:components} \textstyle\Omega_k=\bigcup_{l=1}^{m_{k}}\Omega_{kl},\quad\mathrm{for}\,\,k=1,2,\ldots,m, \end{equation} then the collection $\{\Omega_k\}_{k=1,2,\ldots,m}$ forms an overlapping decomposition of $\Omega$ into $m$ subdomains, that is, $\Omega=\bigcup_{k=1}^m\Omega_k$. Subordinate to this overlapping covering of $\Omega$, we construct a smooth partition of unity consisting of a family of $m$ non-negative and $\mathcal{C}^{\infty}(\Omega)$ functions $\{\rho_k(\mathbf{x})\}_{k=1,2,\ldots,m}$. Each function $\rho_{k}:\overline{\Omega}\rightarrow[0,1]$ is chosen to be \begin{equation*} \rho_{k}(\mathbf{x})= \begin{cases} \,0,&\mathrm{if}\,\,\mathbf{x}\in\overline{\Omega}\setminus\overline{\Omega}_{k},\\ \,h_{k}(\mathbf{x}),&\mathrm{if}\,\,\mathbf{x}\in\bigcup_{l=1;\,l\neq k}^{m}\,(\overline{\Omega}_{k}\cap\overline{\Omega}_l),\\ \,1,&\mathrm{if}\,\,\mathbf{x}\in\overline{\Omega}_{k}\setminus\bigcup_{l=1;\,l\neq k}^{m}\,(\overline{\Omega}_{k}\cap\overline{\Omega}_l), \end{cases} \end{equation*} where $h_k(\mathbf{x})$ is $\mathcal{C}^{\infty}(\Omega)$ and satisfies the conditions \begin{equation*} 0\leq h_k(\mathbf{x})\leq1,\qquad\sum_{k=1}^mh_k(\mathbf{x})=1, \end{equation*} for any $\mathbf{x}\in\bigcup_{l=1;\,l\neq k}^{m}\,(\overline{\Omega}_{k}\cap\overline{\Omega}_l)$. The family of functions $\{\rho_k(\mathbf{x})\}_{k=1,2,\ldots,m}$ is such that \begin{equation}\label{partition:of:unity:conditions} \mathrm{supp}(\rho_k(\mathbf{x}))\subset\overline{\Omega}_k,\qquad 0\leq\rho_k(\mathbf{x})\leq1,\qquad \sum_{k=1}^m\rho_k(\mathbf{x})=1, \end{equation} for any $\mathbf{x}\in\overline{\Omega}$. For numerical purposes, $h_k(\mathbf{x})$ may not necessarily be $\mathcal{C}^{\infty}(\Omega)$, but only a continuous and piecewise smooth function (cf. \cite{mat:pol:rus:wan:98}); this fact will be illustrated in the numerical experiments provided below. In component-wise splittings, the number $m$ of split terms is equal to the dimension $d$ of the spatial domain $\Omega$. In particular, the alternating direction splitting corresponding to the two-dimensional problem (\ref{continuous:problem}) consists of two split terms given by (\ref{adi:splitting}). However, for domain decomposition splittings as that considered in (\ref{dd:splitting}), $m$ is not necessarily equal to $d$. As a result, such splittings may require a generalization of the two-stage procedure (\ref{adi:scheme}) to an arbitrary number $m$ of stages. This general formulation was first proposed in \cite{dou:gun:64} in the context of ADI methods. In our case, it is given, for $n=0,1,\ldots,N_T$ and $k=1,2,\ldots,m$, by \begin{equation}\label{dd:scheme} \begin{aligned} \frac{W_h^{n,k}-W_h^n}{\tau}+\sum_{i=1}^{k}A_{ih}(\theta W_h^{n,i}+(1-\theta)W_h^n)+\sum_{i=k+1}^{m}A_{ih}W_h^n&=F_h^{n+\theta},\\ W_h^{n+1}&=W_h^{n,m}, \end{aligned} \end{equation} where $W_h^0=\mathcal{P}_hu_0$. Note that the domain decomposition discrete solution is denoted by $W_h^n$. Here, $\{A_{kh}\}_{k=1,2,\ldots,m}$ are suitable discretizations of the split operators $\{A_{k}\}_{k=1,2,\ldots,m}$ introduced in (\ref{dd:splitting}); such discretizations are constructed to satisfy $A_h=A_{1h}+A_{2h}+\ldots+A_{mh}$. For convenience, the algorithm (\ref{dd:scheme}) is rewritten in the equivalent form \begin{subequations}\label{generalized:splitting} \begin{align} (I+\theta\tau A_{1h})\,W_h^{n,1}&=\left(I-(1-\theta)\tau A_{1h}-\tau\sum_{i=2}^mA_{ih}\right)W_h^n+\tau F_h^{n+\theta},\label{generalized:splitting:a}\\[0.5ex] (I+\theta\tau A_{kh})\,W_h^{n,k}&=W_h^{n,k-1}+\theta\tau A_{kh}W_h^n,\quad\mbox{for\,\,}k=2,3,\ldots,m,\label{generalized:splitting:b}\\[2.2ex] W_h^{n+1}&=W_h^{n,m}.\label{generalized:splitting:c} \end{align} \end{subequations} Thus, the linear system to be solved at the $k$-th internal stage is of the form \begin{equation}\label{global:system} (I+\theta\tau{A}_{kh})W_h^{n,k}=Q_h^{n,k},\quad\mbox{for\,\,}k=1,2,\ldots,m, \end{equation} where $Q_h^{n,k}$ stands for the corresponding right-hand side. As stated in (\ref{partition:of:unity:conditions}), the function $\rho_{k}(\mathbf{x})$ has compact support on $\overline{\Omega}_{k}$ and, by construction, the entries of $A_{kh}$ corresponding to the nodes that lie outside of this subdomain are zero. In addition, since $\Omega_k$ involves $m_k$ disjoint connected components $\Omega_{kl}$ due to (\ref{disjoint:components}), the linear system (\ref{global:system}) is indeed a collection of $m_k$ uncoupled subsystems of the form \begin{equation}\label{local:system} (I_{\Omega_{kl}}+\theta\tau{A}_{\Omega_{kl}h})\mathcal{R}_{\Omega_{kl}}W_h^{n,k}=\mathcal{R}_{\Omega_{kl}}Q_h^{n,k},\quad\mbox{for\,\,}k=1,2,\ldots,m, \end{equation} where $I_{\Omega_{kl}}=\mathcal{R}_{\Omega_{kl}}I\mathcal{R}_{\Omega_{kl}}^T$ and ${A}_{\Omega_{kl}h}=\mathcal{R}_{\Omega_{kl}}A_{kh}\mathcal{R}_{\Omega_{kl}}^T$. The rectangular matrices $\mathcal{R}_{\Omega_{kl}}$ and $\mathcal{R}_{\Omega_{kl}}^T$ are usually called restriction and extension matrices, respectively, and represent a type of domain decomposition preconditioners (cf. \cite{mat:08}). To obtain such matrices, we first order the nodes of $\mathcal{T}_h$ in $\Omega_{kl}$ in some local ordering. Let us denote by $N_{\mathcal{S}}$ and $N$, respectively, the number of nodes of $\mathcal{T}_h$ in $\Omega_{kl}$ and the total number of nodes of $\mathcal{T}_h$. Then, $\mathcal{R}_{\Omega_{kl}}$ is an $N_{\mathcal{S}}\times N$ matrix defined as \begin{equation*} (\mathcal{R}_{\Omega_{kl}})_{ij}=\begin{cases} \,1,&\mathrm{if}\,\,{index}(\Omega_{kl},i)=j,\\ \,0,&\mathrm{if}\,\,{index}(\Omega_{kl},i)\neq j, \end{cases} \end{equation*} where the index function \emph{index}$(\Omega_{kl},i)$ obtains the global index of the $i$-th local node of $\mathcal{T}_h$ in $\Omega_{kl}$, for $i=1,2,\ldots,N_{\mathcal{S}}$. In other words, this matrix maps a vector in $\mathds{R}^{N}$ of nodal values in $\mathcal{T}_h$ into a vector in $\mathds{R}^{N_{\mathcal{S}}}$ of nodal values corresponding to the nodes of $\mathcal{T}_h$ in $\Omega_{kl}$ (in the local ordering). Likewise, the $N\times N_{\mathcal{S}}$ transpose matrix $\mathcal{R}_{\Omega_{kl}}^{T}$ maps a vector in $\mathds{R}^{N_{\mathcal{S}}}$ into a vector in $\mathds{R}^{N}$, inserting zero entries for the global indices which do not belong to $\Omega_{kl}$. Finally, given a global matrix $M_h\in\mathds{R}^{N\times N}$, the submatrix $M_{\Omega_{kl}h}\in\mathds{R}^{N_{\mathcal{S}}\times N_{\mathcal{S}}}$ associated with the nodes in $\Omega_{kl}$ may be obtained from the restriction and extension mappings as $M_{\Omega_{kl}h}=\mathcal{R}_{\Omega_{kl}}M_h\mathcal{R}_{\Omega_{kl}}^{T}$. From a computational viewpoint, this domain decomposition splitting procedure has three main parameters to adjust its efficiency, namely: the number $m$ of subdomains, the number $m_k$ of disjoint connected components inside a subdomain, and the overlapping size $\xi$. In order to achieve an efficient performance of the algorithm, $m$ should be chosen as small as possible to minimize the number of sequential steps (i.e., internal stages) in (\ref{local:system}), $m_k$ should be chosen as large as possible to maximize the number of parallel components, and $\xi$ should be chosen as small as possible to minimize the number of unknowns within the overlapping regions. In addition, to ensure that the loads assigned to each processor are balanced, there should be approximately the same number of disjoint connected components in each subdomain and, moreover, each such component should have approximately the same diameter. For instance, suppose that we have $q$ available processors for parallel computing. Then, if $m_k$ is approximately the same in each subdomain $\Omega_k$ and also a multiple of $q$, each subdomain can be partitioned into $q$ groups of $\frac{m_k}{q}$ components and each group assigned to one of the processors. Remarkably, unlike most classical overlapping domain decomposition algorithms (cf. \cite{qua:val:99}), the solution to (\ref{local:system}) does not require any Schwarz iteration procedure, since the internal stages in the algorithm are solved sequentially (i.e., interface conditions need not be imposed on subdomains during the solution process). \section{The splitting error}\label{sec:splitting:error} As noted in the introduction, the time-splitting methods presented above can be interpreted as perturbations of some unsplit implicit schemes, such as the backward Euler and Crank--Nicolson methods. Recall that the scheme (\ref{adi:scheme}) considers the alternating direction splitting (\ref{adi:splitting}), while the algorithm (\ref{dd:scheme}) makes use of the domain decomposition splitting (\ref{dd:splitting}). In essence, the former method is a particular instance of the latter for $m=2$. In this section, we reformulate (\ref{dd:scheme}) as a classical implicit scheme (backward Euler, for $\theta=1$, and Crank--Nicolson, for $\theta=\frac{1}{2}$) plus a perturbation term, which will be referred to as the splitting error. As we will see, this splitting error is $\mathcal{O}(\tau^2)$, i.e., one order higher than the truncation error associated with the backward Euler method and of the same order as that associated with the Crank--Nicolson method. As a result, one would expect to obtain the same asymptotic rate of convergence for both the time-splitting and its underlying implicit method. In practice, however, the size of the splitting error can be much larger than that of the truncation error corresponding to the underlying scheme, thus degrading the efficiency of the resulting algorithm. In the sequel, we derive a general expression for the splitting error and present a strategy to reduce its actual size. Let us consider the formulation given by (\ref{generalized:splitting}). In order to eliminate the intermediate values $W_h^{n,1},W_h^{n,2},\ldots,W_h^{n,m-1}$, we multiply the equations (\ref{generalized:splitting:b}) by $(I+\theta\tau A_{1h})(I+\theta\tau A_{2h})\cdots(I+\theta\tau A_{(k-1)h})$, sum on $k$ and add the equation (\ref{generalized:splitting:a}) to obtain, for $n=0,1,\ldots,N_T$, \begin{equation}\label{perturbed:implicit} \frac{W_h^{n+1}-W_h^n}{\tau}+A_h(\theta W_h^{n+1}+(1-\theta)W_h^n)+B_h(W_h^{n+1}-W_h^n)=F_h^{n+\theta}, \end{equation} where \begin{align*} B_h=&\,\,\theta^2\tau\sum_{1\leq i_1<i_2\leq m}A_{i_1h}A_{i_2h}+\theta^3\tau^2\sum_{1\leq i_1<i_2<i_3\leq m}A_{i_1h}A_{i_2h}A_{i_3h}\\ &\,\,+\ldots+\theta^m\tau^{m-1}A_{1h}A_{2h}\cdots A_{mh}. \end{align*} Note that, depending on the value of $\theta$, (\ref{perturbed:implicit}) is, in fact, a perturbed backward Euler or Crank--Nicolson method (as compared with (\ref{be:cn})). The perturbation term $B_h(W_h^{n+1}-W_h^n)$ is the aforementioned splitting error and satisfies \begin{equation*} B_h(W_h^{n+1}-W_h^n)=\tau B_h\left(\frac{W_h^{n+1}-W_h^n}{\tau}\right)=\mathcal{O}\left(\tau^2\sum_{i,j}\left|\frac{\partial}{\partial t}A_iA_ju\right|\right)=\mathcal{O}(\tau^2), \end{equation*} where the split operators $A_k$ are given by either (\ref{adi:splitting}) or (\ref{dd:splitting}), provided that $u$, $c$ and the entries of $a$ are sufficiently smooth. In \cite{dou:kim:01}, Douglas and Kim noted that, if we could add $B_h(W_h^{n+1}-W_h^n)$ to the right-hand side of (\ref{perturbed:implicit}), then the perturbation term would be cancelled. However, since $W_h^{n+1}$ is not known at time $t_n$, this modification cannot be done. Taking into account that the best available approximation to the difference $W_h^{n+1}-W_h^n$ at time $t_n$ is $W_h^{n}-W_h^{n-1}$, the algorithm (\ref{generalized:splitting}) can be modified, for $n=1,2,\ldots,N_T$, to obtain \begin{equation}\label{improved:splitting} \begin{aligned} \widehat{F}_h^{n+\theta}&=F_h^{n+\theta}+B_h(Z_h^{n}-Z_h^{n-1}),\\[0.5ex] (I+\theta\tau A_{1h})Z_h^{n,1}&=\left(I-(1-\theta)\tau A_{1h}-\tau\sum_{i=2}^mA_{ih}\right)Z_h^n+\tau \widehat{F}_h^{n+\theta}\\[0.5ex] (I+\theta\tau A_{kh})Z_h^{n,k}&=Z_h^{n,k-1}+\theta\tau A_{kh}Z_h^n,\qquad\mbox{for\,\,}k=2,3,\ldots,m,\\[2.2ex] Z_h^{n+1}&=Z_h^{n,m}, \end{aligned} \end{equation} where $Z_h^0=\mathcal{P}_hu_0$ and $Z_h^1$ must be suitably approximated. In this case, eliminating the intermediate values $Z_h^{n,1},Z_h^{n,2},\ldots,Z_h^{n,m-1}$, or referring to (\ref{perturbed:implicit}), we get, for $n=1,2,\ldots,N_T$, \begin{equation*} \frac{Z_h^{n+1}-Z_h^n}{\tau}+A_h(\theta Z_h^{n+1}+(1-\theta)Z_h^n)+B_h(Z_h^{n+1}-2Z_h^n+Z_h^{n-1})=F_h^{n+\theta}. \end{equation*} Now, the perturbation term $B_h(Z_h^{n+1}-2Z_h^n+Z_h^{n-1})$ satisfies \begin{align*} B_h(Z_h^{n+1}-2Z_h^n+Z_h^{n-1})&=\tau^2 B_h\left(\frac{Z_h^{n+1}-2Z_h^n+Z_h^{n-1}}{\tau^2}\right)\\ &=\mathcal{O}\left(\tau^3\sum_{i,j}\left|\frac{\partial^2}{\partial t^2}A_iA_ju\right|\right)=\mathcal{O}(\tau^3), \end{align*} provided that $u$, $c$ and the entries of $a$ are sufficiently smooth. In other words, the corrected right-hand side $\widehat{F}_h^{n+\theta}$ introduced in (\ref{improved:splitting}) yields a reduction in the order of the splitting error from $\mathcal{O}(\tau^2)$ to $\mathcal{O}(\tau^3)$. As a result, the actual discretization error of the modified algorithm (\ref{improved:splitting}) is of the same size as that associated with its underlying unsplit method. Remarkably, if the domain decomposition splitting (\ref{dd:splitting}) is used to obtain $A_k$, the operator $B_h$ will be non-zero only within the overlapping regions. Thus, the newly added term $B_h(Z_h^{n}-Z_h^{n-1})$ will have a very low extra computational cost. The convergence analysis for the improved accuracy method (\ref{improved:splitting}) must take into account the splitting strategy under consideration. The case in which the split operators $A_k$ are obtained via the alternating direction splitting (\ref{adi:splitting}) is discussed in \cite{arb:hua:yan:07} and \cite{dou:kim:01}, and error estimates of order $\mathcal{O}(\tau^{3-2\theta}+h^s)$ are derived. In the latter work, the discrete split operators are assumed to commute pairwise, while the former analyzes the non-commuting case. As for the domain decomposition splitting (\ref{dd:splitting}), a convergence analysis of this kind is lacking so far. A similar analysis for the unimproved algorithm (\ref{generalized:splitting}), considering a finite difference spatial discretization, is carried out in \cite{mat:pol:rus:wan:98}. The derivation of error estimates for the newly proposed method (\ref{improved:splitting}) is a topic of current research. \section{Numerical experiments}\label{sec:experiments} The numerical experiments performed in this section are partly inspired by those contained in \cite{arb:hua:yan:07,dou:kim:01}. Let us consider the parabolic initial-boundary value problem (\ref{continuous:problem}), where $\Omega=(0,1)\times(0,1),$ $T=1$ and $c(\mathbf{x})\equiv 0$. Given a certain diffusion term $a(\mathbf{x})$ (to be specified below), the functions $f(\mathbf{x},t)$ and $u_0(\mathbf{x})$ are defined in such a way that the exact solution is \begin{equation*} u(x,y,t)=\sin(2\pi t)\sin(2\pi x)\sin(2\pi y). \end{equation*} Throughout this section, a finite difference spatial discretization is combined with several second-order time integrators (i.e., we fix $\theta=\frac{1}{2}$), considering the discretization parameters $h=\tau=\frac{1}{M}$. In particular, we compare the accuracy of the Crank--Nicolson (CN) method (\ref{be:cn}) to that of the Douglas--Gunn scheme (\ref{generalized:splitting}) and the Douglas--Kim method (\ref{improved:splitting}), combined with either the ADI splitting (\ref{adi:splitting}) (DG$_{\mathrm{ADI}}$ and DK$_{\mathrm{ADI}}$, respectively) or the domain decomposition splitting (\ref{dd:splitting}) (DG$_{\mathrm{DD}}$ and DK$_{\mathrm{DD}}$, respectively). For the Douglas--Kim procedures, DK$_{\mathrm{ADI}}$ and DK$_{\mathrm{DD}}$, the approximation of the exact solution at time $t_1=\tau$ is obtained by running a single step of the Crank--Nicolson scheme. The computed errors are measured in the discrete $\ell^{\infty}(\ell^2)$-norm, that is, \begin{equation*} \|e\|_{\ell^{\infty}(\ell^2)}=\max_{1\leq n\leq N_T}\left(\sum_{i=1}^{M-1}\sum_{j=1}^{M-1}(e_{i,j}^n)^2h^2\right)^{1/2}, \end{equation*} where $e_{i,j}^n$ denotes the error at the $(i,j)$-th node and time $t_n$, obtained as the difference between the exact solution at $(ih,jh,t_n)$ and the corresponding numerical solution. Unless otherwise stated, both the DG$_{\mathrm{DD}}$ and the DK$_{\mathrm{DD}}$ methods consider two overlapping subdomains, each consisting of four non-overlapping connected components. More precisely, let $I=(0,1)$ and define the intervals \begin{align*} I_1&=\textstyle\left(0,\frac{1}{8}+\frac{\xi}{2}\right)\cup\left(\frac{1}{4}-\frac{\xi}{2},\frac{3}{8}+\frac{\xi}{2}\right)\cup\left(\frac{1}{2}-\frac{\xi}{2},\frac{5}{8}+\frac{\xi}{2}\right) \cup\left(\frac{3}{4}-\frac{\xi}{2},\frac{7}{8}+\frac{\xi}{2}\right),\\[0.5ex] I_2&=\textstyle\left(\frac{1}{8}-\frac{\xi}{2},\frac{1}{4}+\frac{\xi}{2}\right)\cup\left(\frac{3}{8}-\frac{\xi}{2},\frac{1}{2}+\frac{\xi}{2}\right)\cup\left(\frac{5}{8}-\frac{\xi}{2},\frac{3}{4}+\frac{\xi}{2}\right) \cup\left(\frac{7}{8}-\frac{\xi}{2},1\right), \end{align*} where $\xi$ denotes the overlapping size (to be specified below). Then, we can define the subdomains $\Omega_1=I_1\times I$ and $\Omega_2=I_2\times I$. Figure \ref{fig:1D:DD:splitting} shows a domain decomposition of this kind on the unit square, for an arbitrary number $q$ of connected components per subdomain. According to \cite{mat:pol:rus:wan:98}, we construct a piecewise smooth partition of unity as follows. The closure of subdomain $\Omega_k$ can be expressed as the union of four closed disjoint rectangles given by $R_k^i=[a_k^i,b_k^i]\times[0,1]$, for $i=1,2,3,4$ and $k=1,2$. Then, for each rectangle $R_k^i$, we define a function of the form \begin{equation*} w_{k}^i(x,y)= \begin{cases} \,\sin\left(\displaystyle\frac{\pi(x-a_k^i)}{b_k^i-a_k^i}\right),&\mbox{if }(x,y)\in R_k^i,\\ \,0,&\mbox{if }(x,y)\in\overline{\Omega}\setminus R_k^i. \end{cases} \end{equation*} The partition of unity functions are thus obtained as \begin{equation}\label{piecewise:smooth:partition:unity} \rho_k(x,y)=\frac{\displaystyle\sum_{i=1}^{4}w_k^i(x,y)}{\displaystyle\sum_{l=1}^{2}\displaystyle\sum_{i=1}^{4}w_l^i(x,y)},\quad\mbox{for } k=1,2. \end{equation} Table \ref{table:N:varies} shows the global errors and mean rates of convergence obtained for different values of the discretization parameter $M$, with the aforementioned methods. In this experiment, the diffusion coefficient is $a(x,y)=1$, and the overlapping size is $\xi=\frac{1}{8}$. We observe that the global errors corresponding to the DG$_{\mathrm{ADI}}$ and the DG$_{\mathrm{DD}}$ methods are about 10 and 14 times larger, respectively, than the CN errors. The DK$_{\mathrm{ADI}}$ errors are larger but comparable to the CN errors, since the splitting error for the former method is much smaller than that for the DG$_{\mathrm{ADI}}$ scheme. In this case, the mean rate of convergence is greater than 2, because the splitting error is being removed at a rate of $\mathcal{O}(\tau^3)$. Accordingly, the DK$_{\mathrm{DD}}$ errors are much smaller than the DG$_{\mathrm{DD}}$ errors, and very similar to those obtained for the CN scheme. In fact, due to the observed higher rate of convergence, the DK$_{\mathrm{DD}}$ errors are indeed smaller than the CN errors, for $M=160$ and $M=320$. In Table \ref{table:a:varies}, we fix the values $M=160$ and $\xi=\frac{1}{8}$, and compute the global errors corresponding to the previous schemes, for five different choices of the diffusion coefficient $a(x,y)$. Besides the one tested in the preceding table, $a_1(x,y)=1$, we also consider \begin{align*} a_2(x,y)&=\displaystyle\frac{1}{2+\cos(3\pi x)\cos(2\pi y)},\\[1.75ex] a_3(x,y)&= \begin{cases} \,1+0.5 \sin(5\pi x)+y^3,&\mbox{if }x\leq 0.5,\\ \,\dfrac{1.5}{1+(x-0.5)^2}+y^3,&\mbox{otherwise}, \end{cases}\\[1.75ex] a_4(x,y)&=\left(\begin{array}{cc}a_2(x,y) &0\\0&a_3(x,y)\end{array}\right),\\[1.75ex] a_5(x,y)&=\left(\begin{array}{cc}a_2(x,y) &1/4\\1/4&a_2(x,y)\end{array}\right). \end{align*} If we take $a(x,y)=a_i(x,y)$, for $i=1,2,3,4$, the computed DG$_{\mathrm{ADI}}$ errors are much larger than the corresponding CN errors. In turn, the errors associated with the DK$_{\mathrm{ADI}}$ method are comparable to, and only slightly larger than, those obtained with the CN scheme. On the other hand, if the diffusion coefficient is chosen to be $a_5(x,y)$, none of these two ADI methods can be applied, due to the presence of mixed derivative terms. As for the domain decomposition splitting techniques, in all five cases, the DK$_{\mathrm{DD}}$ scheme is shown to be comparable to the CN scheme and superior to the DG$_{\mathrm{DD}}$ method. Table \ref{table:d:varies} shows the effect of the overlapping size $\xi$ on the accuracy of the implemented methods. In particular, we consider $a(x,y)=a_2(x,y)$ and $M=160$, and compute the global errors corresponding to the CN, the DG$_{\mathrm{DD}}$ and the DK$_{\mathrm{DD}}$ schemes, for different values of $\xi$. The observed DG$_{\mathrm{DD}}$ errors increase as $\xi$ becomes smaller, due to the presence of negative powers of $\xi$ in the splitting error (cf. \cite{mat:pol:rus:wan:98}). This undesirable behaviour is avoided by the DK$_{\mathrm{DD}}$ scheme, whose global errors get slightly smaller as $\xi$ decreases. Finally, we study the influence of the number of disjoint connected components per subdomain on the accuracy of the domain decomposition procedures. For that purpose, we consider the same number $q$ of components in $\Omega_1$ and $\Omega_2$; i.e., recalling the notation introduced in \S\ref{sec:split}, $m_1=m_2=q$. Table \ref{table:q:varies} shows the global errors corresponding to the CN, the DG$_{\mathrm{DD}}$ and the DK$_{\mathrm{DD}}$ schemes, for different values of $q$, when $a(x,y)=a_2(x,y)$, $M=160$ and $\xi=\frac{1}{16}$. Note that, if $q=2$ or $q=8$, the partition of unity functions (\ref{piecewise:smooth:partition:unity}) must be modified accordingly. As in the preceding experiments, the DG$_{\mathrm{DD}}$ errors are larger than the CN errors; moreover, they even grow as $q$ increases, which represents a serious drawback from a computational viewpoint. By contrast, the errors for the DK$_{\mathrm{DD}}$ scheme are comparable to, and even smaller than, those associated with the CN method. In this case, such errors decrease as $q$ increases, thus permitting to exploit the parallel capabilities of the new algorithms. \section{Conclusions}\label{sec:conclusions} A family of improved domain decomposition splitting methods for solving parabolic equations has been presented. These methods can be formulated as unsplit implicit schemes (such as backward Euler or Crank--Nicolson) perturbed by a term that is $\mathcal{O}(\tau^3)$ in the time step $\tau$. This perturbation term, also known as splitting error, is one order higher than its $\mathcal{O}(\tau^2)$ analogue stemming from the unimproved formulation. Such an increase in the order of the splitting error is achieved by adding a correction term to the right-hand side of the original scheme. This technique was proposed in \cite{dou:kim:01} in the context of ADI methods, and has been extended here to domain decomposition-based splittings for the first time. The newly derived methods can be implemented with a very low additional cost, since the correction term only requires certain computations within the overlapping regions. In addition, the resulting subdomain problems can be solved in parallel with no need for Schwarz-type iteration procedures. Numerical results show that the proposed modification leads to a significant reduction in the splitting error. Moreover, the global discretization error is comparable to that associated with the Crank--Nicolson method, independently of the number of connected components per subdomain and the overlapping size.

https://arxiv.org/abs/1912.05366

$L^\infty$ bounds for numerical solutions of noncoercive convection-diffusion equations

In this work, we apply an iterative energy method à la de Giorgi in order to establish $L^{\infty}$ bounds for numerical solutions of noncoercive convection-diffusion equations with mixed Dirichlet-Neumann boundary conditions.

\section{Introduction} {\bf The continuous problem.} Let $\Omega$ be an open bounded polygonal domain of $\mathbb R^p$ with $p=2$ or $3$. We denote by ${\rm m(\cdot)}$ both the Lebesgue and $p-1$ dimensional Hausdorff measure. We assume that $\partial \Omega= \Gamma^D\cup\Gamma^N$ with $\Gamma^D\cap\Gamma^N=\emptyset$ and ${\rm m}(\Gamma^D)>0$ and we denote by ${\mathbf n}$ the exterior normal to $\partial \Omega$. Let ${\mathbf U}\in C({\bar \Omega})^2$ be a velocity field, $b\in L^{\infty}(\Omega)$ assumed to be nonnegative, $f\in L^\infty(\Omega)$ a source term and $v^D\in L^{\infty} (\Gamma^D)$ a boundary condition. We consider the following convection-diffusion equation with mixed boundary conditions: \begin{subequations}\label{pb_depart} \begin{align} &\mathrm{div} (-\nabla v + {\mathbf U} v)+ bv=f&&\qquad \mbox{in }\Omega, \label{eq_v}\\ & (-\nabla v + {\mathbf U} v)\cdot {\mathbf n}=0 &&\qquad \mbox{on } \Gamma^N, \label{Neum_bc}\\ &v=v^D&&\qquad \mbox{on } \Gamma^D \label{Dir_bc}. \end{align} \end{subequations} This noncoercive elliptic linear problem has been widely studied by Droniou and coauthors, even with less regularity on the {data}, see for instance \cite{droniou_potan_2002, DG_M2AN_2002, Droniou_jnm_2003,DGH_sinum_2003}. Nevertheless, up to our knowledge, the derivation of explicit $L^\infty$ bounds on numerical solutions has not been done in the literature. \bigskip \noindent {\bf The numerical scheme.} The mesh of the domain $\Omega$ is denoted by $\cal M=(\mathcal T,\mathcal E,\cal P)$ and classically given by: $\mathcal T$, a set of open polygonal { or polyhedral} control volumes; $\mathcal E$, a set of edges {or faces}; ${\mathcal P}=(x_K)_{K\in\mathcal T}$ a set of points. {In the following, we also use the denomination ``edge'' for a face in dimension $3$}. As we deal with a Two-Point Flux Approximation (TPFA) of convection-diffusion equations, we assume that the mesh is admissible in the sense of~\cite{Eymard2000} (Definition 9.1). We distinguish in $\mathcal E$ the interior edges, $\sigma =K|L$, from the exterior edges: $\mathcal E=\mathcal E_{int}\cup {\mathcal E}_{ext}$. Among the exterior edges, we distinguish the edges included in $\Gamma^D$ from the edges included in $\Gamma^N$: ${\mathcal E}_{ext}={\mathcal E}^D\cup {\mathcal E}^N$. For a given control volume $K\in{\mathcal T}$, we define ${\mathcal E}_K$ the set of its edges, which is also split into ${\mathcal E}_K={\mathcal E}_{K,int}\cup{\mathcal E}_{K}^D\cup{\mathcal E}_{K}^N$. For each edge $\sigma\in\mathcal E$, we pick one cell in the non empty set $\{K:\sigma\in\mathcal E_K\}$ and denote it by $K_\sigma$. In the case of an interior edge $\sigma=K|L$, $K_{\sigma}$ is either $K$ or $L$. { Let ${\rm d}(\cdot,\cdot)$ denote the Euclidean distance.} For all edges $\sigma\in{\mathcal E}$, we set ${\rm d}_{\sigma}={\rm d}(x_K,x_L)$ if $\sigma=K|L\in{\mathcal E}_{int}$ and ${\rm d}_{\sigma}={\rm d}(x_K,\sigma)$ if $\sigma\in{\mathcal E}_{ext}$ with $\sigma\in \mathcal E_K$ and the transmissibility coefficient is defined by $\tau_{\sigma}={\rm m}(\sigma)/{\rm d}_{\sigma}$, for all $\sigma\in{\mathcal E}$. We also denote by ${\mathbf n}_{K,\sigma}$ the normal to $\sigma\in{\mathcal E}_K$ outward $K$. We assume that the mesh satisfies the regularity constraint: \begin{equation}\label{reg-mesh} \exists \xi >0 \mbox{ such that } {\rm d}(x_K,\sigma)\geq \xi \, {\rm d}_{\sigma},\quad \forall K\in\mathcal T, \forall \sigma\in\mathcal E_K. \end{equation} As a consequence, we obtain that \begin{equation}\label{inegvol} \sum_{\sigma\in\mathcal E_K} {\rm m}(\sigma){\rm d}_{\sigma}\leq \displaystyle\frac{p}{\xi} {\rm m} (K)\quad \forall K\in\mathcal T. \end{equation} The size of the mesh is defined by $h=\max\{\mbox{diam }(K)\,:\,K\in\mathcal T\}$. Let us define $$ \begin{aligned} & f_K=\displaystyle\frac{1}{{\rm m}(K)}\int_{K} f, \quad b_K=\displaystyle\frac{1}{{\rm m}(K)}\int_{K} b \quad \forall K\in\mathcal T,\\ &U_{K,\sigma}=\displaystyle\frac{1}{{\rm m}(\sigma)}\int_{\sigma} {\mathbf U}\cdot {\mathbf n}_{K,\sigma},\quad \forall K\in\mathcal T,\ \forall\sigma \in {\mathcal E}_K,\\ & v_\sigma^D=\displaystyle\frac{1}{{\rm m}(\sigma)}\int_\sigma v^D,\quad \forall \sigma\in {\mathcal E}^D. \end{aligned} $$ Given a Lipschitz-continuous function on $\mathbb R$ which satisfies \begin{equation}\label{hyp_B} B(0)=1,\quad\ B(s)>0\quad \mbox{ and }\quad B(s)-B(-s)=-s\quad\forall s\in\mathbb R, \end{equation} we consider the B-scheme defined by \begin{equation}\label{scheme} \sum_{\sigma\in \mathcal E_K} {\mathcal F}_{K,\sigma}+ {\rm m}(K) b_K v_K= {\rm m}(K)f_K, \quad \forall K\in{\mathcal T}, \end{equation} where the numerical fluxes are defined by \begin{equation}\label{numflux} {\mathcal F}_{K,\sigma}=\left\{ \begin{aligned} &0,\quad \forall K\in\mathcal T,\forall \sigma \in \mathcal E_K^N,\\ &\tau_{\sigma} \Bigl(B(-U_{K,\sigma}{\rm d}_{\sigma})v_K-B(U_{K,\sigma}{\rm d}_{\sigma})v_{K,\sigma}\Bigl),\quad\forall K\in\mathcal T, \forall \sigma\in\mathcal E_{K}\setminus \mathcal E_K^N, \end{aligned} \right. \end{equation} with the convention $v_{K,\sigma}=v_L$ if $\sigma =K|L$ and $v_{K,\sigma}=v_\sigma^D$ if $\sigma \in\mathcal E_K^D$. Let us recall that the upwind scheme corresponds to the case $B(s)=1+s^-$ ($s^-$ is the negative part of $s$, while $s^+$ is its positive part) and the Scharfetter-Gummel scheme to the case $B(s)=s/(e^s-1)$. They both satisfy \eqref{hyp_B}. The centered scheme which corresponds to $B(s)=1-s/2$ does not satisfy the positivity assumption. It can however be used if $|U_{K,\sigma}|{\rm d}_{\sigma}\leq 2$ for all $K\in\mathcal T$ and $\sigma \in\mathcal E_K$. Thanks to the hypotheses \eqref{hyp_B}, we notice that the numerical fluxes through the interior and Dirichlet boundary edges rewrite \begin{equation}\label{numflux2} {\mathcal F}_{K,\sigma}= \tau_\sigma B(|U_{K,\sigma}|{\rm d}_{\sigma})(v_K-v_{K,\sigma})+ {\rm m}(\sigma) \left(U_{K,\sigma}^+ v_K - U_{K,\sigma}^-v_{K,\sigma}\right) . \end{equation} \noindent {\bf Main result.} The scheme \eqref{scheme}-\eqref{numflux} defines a linear system of equations ${\mathbb M} {\mathbf v}={\mathbf S}$ whose unknown is ${\mathbf v}=(v_K)_{K\in\mathcal T}$; It is well-known that ${\mathbb M}$ is an M-matrix, which ensures existence and uniqueness of a solution to the scheme. Moreover, we may notice that, if $v^D$ and $f$ are nonnegative functions, then ${\mathbf S}$ has nonnegative values and therefore $v_K\geq 0$ for all $K\in\mathcal T$. Our purpose is now to establish $L^{\infty}$ bounds on ${\mathbf v}$ as stated in Theorem \ref{mainthm}. \begin{theorem}\label{mainthm} Assume that ${\mathbf U}\in C({\bar \Omega})^2$, $b\in L^{\infty}(\Omega)$ with $b\geq 0$ {\em a.e.}, ${f\in L^\infty(\Omega)}$ and $v^D\in L^{\infty} (\Gamma^D)$. There exists non-negative constants $\overline{M}$ (\emph{resp.} $\underline{M}$) depending only on $\Omega$, $\xi$, the function $B$, $\|{\bf U}\|_{L^\infty}$, $\|f^+\|_{L^\infty}$ and $\|(v^D)^+\|_{L^\infty}$ (\emph{resp.} $\|f^-\|_{L^\infty}$ and $\|(v^D)^-\|_{L^\infty}$) such that the solution ${\mathbf v}$ to the scheme \eqref{scheme}-\eqref{numflux} verifies \[ -\underline{M}\ \leq\ v_K\ \leq\ \overline{M}, \quad \forall K\in\mathcal T. \] \end{theorem} The rest of this paper is dedicated to the proof of Theorem~\ref{mainthm}. It relies on a De Giorgi iteration method (see \cite{Vasseur_lectnotes} and references therein). In Section~\ref{sec:particular}, we start by studying a particular case where the data is normalized. Then, we give the proof of the theorem in Section~\ref{sec:proof}. Let us mention that from the bounds of Theorem~\ref{mainthm}, it is possible to establish global-in-time $L^\infty$ bounds for the corresponding evolution equation by using an entropy method (see \cite[Theorem 2.7]{chainais_2019_large}). \section{Study of a particular case}\label{sec:particular} In this section, we consider the particular case where the source $f$ is non-negative and the boundary condition $v^D$ is non-negative and bounded by $1$. Let us start with some notations. Given $m\geq 1$, we denote the $m$-th truncation threshold by \begin{equation}\label{eq:trunc} C_m=2(1-2^{-m})\,, \end{equation} Then, we introduce the $m$-th energy \begin{equation}\label{eq:energy} E_m({\bf v})=\displaystyle\sum_{\sigma\in\mathcal E_{int}\cup\mathcal E^D} \tau_\sigma \left[\log (1 +(v_{K,\sigma}-C_m)^+)-\log (1 +(v_{K}-C_m)^+)\right]^2. \end{equation} When there is no ambiguity we write $E_m = E_m({\bf v})$. The first proposition is a fundamental estimate of the energy. \begin{proposition}\label{prop:fund_ineq} Assume that $f_K\geq 0$ for all $K\in\mathcal T$ and $v_\sigma^D \in [0,1]$ for all $\sigma \in \mathcal E^D$, so that the solution ${\mathbf v}$ to \eqref{scheme}-\eqref{numflux} satisfies $v_K\geq 0$ for all $K\in\mathcal T$. Then one has for all $m\geq1$ that \begin{equation}\label{majEm} E_m\ \leq\ \frac{4p}{\beta_{\mathbf U}^2}\left(\Vert {\mathbf U}\Vert_{L^\infty}^2+{\Vert f\Vert_{L^\infty}}\right) \sum_{\substack{K\in\mathcal T\\v_{K}>C_m}}{{\rm m}(K)}\,. \end{equation} where $\beta_{\mathbf U} := \inf_{x\in[-\|\mathbf{U}\|_{L^\infty},\|\mathbf{U}\|_{L^\infty}]}B({\rm diam}(\Omega)\,x)$ (because of \eqref{hyp_B}, $\beta_{\mathbf U}\in(0,1]$). \end{proposition} \begin{proof} In order to shorten some expressions hereafter, let us introduce $w_K^m = v_K-C_m$ for all $K\in \mathcal T$ and $w_\sigma^{m,D}=v_{\sigma}^D-C_m$ for all $\sigma \in\mathcal E^D$. Let us note that we identify ${\mathbf w}^m=(w_K^m)_{K\in\mathcal T}$ and the associate piecewise constant function. Therefore, we can write $$ {\rm m} (\{{\mathbf w}^m>0\}) = \sum_{w_{K}^m>0}{{\rm m}(K)}. $$ First, observe that $E_m$ is the discrete counterpart of $$ \int_\Omega \left\vert \nabla \log (1+w^m)\right\vert^2 {\mathbf 1}_{\{w^m>0\}}=\int_\Omega \nabla w^m\cdot\frac{\nabla w^m}{(1+w^m)^2}{\mathbf 1}_{\{w^m>0\}},\ \mbox{ with }w^m=v-C_m\,, $$ {where ${\mathbf 1}_{A}$ is the indicator function of $A$}. Let us define $\varphi :s\mapsto s/(1+s){\mathbf 1}_{\{s\geq 0\}}$, which satisfies $\varphi'(s)=1/(1+s)^2{\mathbf 1}_{\{s\geq 0\}}$ and let us introduce $F_m$ another discrete counterpart of the preceding quantity $$ F_m=\displaystyle\sum_{\sigma\in\mathcal E_{int}\cup\mathcal E^D} \tau_\sigma \left((w_{K,\sigma}^m)^+-(w_{K}^m)^+\right)\left(\varphi(w_{K,\sigma}^m)-\varphi(w_{K}^m) \right). $$ It is clear that $E_m\leq F_m$ for all $m\geq 1$, as for all $x,y\in\mathbb R$ we have $$ \left(\log(1+x^+)-\log(1+y^+)\right)^2\leq (x^+-y^+)\left(\varphi(x)-\varphi(y)\right). $$ Let us now multiply the scheme \eqref{scheme} by $\varphi(w_K^m)$ and sum over $K\in\mathcal T$. Due to the non-negativity of $b$ and ${\mathbf v}$, we obtain, after a discrete integration by parts, $$ \displaystyle\sum_{\sigma\in\mathcal E_{int}\cup\mathcal E^D}{\mathcal F}_{K,\sigma}(\varphi(w_K^m)-\varphi(w_{K,\sigma}^m))\leq \sum_{K\in\mathcal T} {\rm m}(K) f_K \varphi(w_K^m). $$ Using that $\varphi$ is bounded by 1 and vanishes on $\mathbb R_-$, we deduce that \begin{equation}\label{inegdep} \displaystyle\sum_{\sigma\in\mathcal E_{int}\cup\mathcal E^D}{\mathcal F}_{K,\sigma}(\varphi(w_K^m)-\varphi(w_{K,\sigma}^m))\leq { \Vert f\Vert_{L^\infty}\, {\rm m}(\{{\mathbf w}^m>0\})}. \end{equation} We focus now on the left-hand-side of \eqref{inegdep}. Due to \eqref{numflux2} and the definition of $w_K^m$, we can rewrite $ {\mathcal F}_{K,\sigma}$ as $$ {\mathcal F}_{K,\sigma}= \tau_\sigma B(|U_{K,\sigma}|{\rm d}_{\sigma})(w_K^m-w^m_{K,\sigma})+ {\rm m}(\sigma) \left(U_{K,\sigma}^+ (w_K^m+C_m) - U_{K,\sigma}^-(w_{K,\sigma}^m+C_m)\right) . $$ Observe that since $\varphi$ is a non-decreasing function, one has $$ (x-y)\left(\varphi(x)-\varphi(y)\right)\geq (x^+-y^+)(\varphi(x)-\varphi(y)),\quad \forall x,y\in\mathbb R. $$ Therefore, using the definition of $\beta_{\mathbf U}$ we obtain that \begin{equation}\label{ineg1} \displaystyle\sum_{\sigma\in\mathcal E_{int}\cup\mathcal E^D}{\mathcal F}_{K,\sigma}(\varphi(w_K^m)-\varphi(w_{K,\sigma}^m))\geq \beta_{\mathbf U} F_m -G_m, \end{equation} with $$ G_m=-\sum_{\sigma\in\mathcal E_{int}\cup\mathcal E^D}{\rm m}(\sigma) \left(U_{K,\sigma}^+ (w_K^m+C_m) - U_{K,\sigma}^-(w_{K,\sigma}^m+C_m)\right) (\varphi(w_K^m)-\varphi(w_{K,\sigma}^m)). $$ For an interior edge, $w_K^m$ and $w_{K,\sigma}^m$ play a symmetric role in the preceding sum. As $w_{\sigma}^{m,D}\leq 0$ for all $\sigma\in\mathcal E^D$ and $\varphi$ vanishes on $\mathbb R_-$, we can always assume that $w_K^m\geq w_{K,\sigma}^m$ and an edge has a contribution in the sum if at least $w_K^m> 0$. Then, under these assumptions one has \begin{multline*} -{\rm m}(\sigma) \left(U_{K,\sigma}^+ (w_K^m+C_m) - U_{K,\sigma}^-(w_{K,\sigma}^m+C_m)\right) (\varphi(w_K^m)-\varphi(w_{K,\sigma}^m))\\ \leq \Vert {\mathbf U}\Vert_{L^\infty} {\rm m}(\sigma) (w_{K,\sigma}^m+C_m)(\varphi(w_K^m)-\varphi(w_{K,\sigma}^m)). \end{multline*} But, $w_{K,\sigma}^m+C_m\leq 2(1+(w_{K,\sigma}^m)^+)$ and applying the definition of $\varphi$, we get $$ \begin{array}{rcl} (w_{K,\sigma}^m+C_m)(\varphi(w_K^m)-\varphi(w_{K,\sigma}^m))&\leq& 2\displaystyle\frac{(w_K^m)^+-(w_{K,\sigma}^m)^+}{1+(w_{K}^m)^+}\\[1em] &\leq&2\displaystyle\frac{(w_K^m)^+-(w_{K,\sigma}^m)^+}{\sqrt{1+ (w_K^m)^+}\sqrt{1+ (w_{K,\sigma}^m)^+}}. \end{array} $$ Therefore, $$ G_m\leq 2\Vert {\mathbf U}\Vert_{L^\infty} \sum_{\sigma\in\mathcal E_{int}\cup\mathcal E^D}{\rm m}(\sigma)\displaystyle\frac{\vert(w_K^m)^+-(w_{K,\sigma}^m)^+\vert}{\sqrt{1+ (w_K^m)^+}\sqrt{1+ (w_{K,\sigma}^m)^+}}. $$ We apply now Cauchy-Schwarz inequality in order to get \begin{equation}\label{ineg2} G_m\leq 2\Vert {\mathbf U}\Vert_{L^\infty}(F_m)^{1/2} \left(\sum_{\sigma\in\mathcal E^{sp}}{\rm m}(\sigma){\rm d}_{\sigma}\right)^{1/2}, \end{equation} where $\mathcal E^{sp}$ is the set of interior and Dirichlet boundary edges on which $(w_K^m)^+-(w_{K,\sigma}^m)^+\neq 0$. It appears that, due to \eqref{inegvol}, \begin{equation}\label{ineg3} \sum_{\sigma\in\mathcal E^{sp}}{\rm m}(\sigma){\rm d}_{\sigma}\leq \displaystyle\sum_{K\in\mathcal T; w_K^m>0}\left(\sum_{\sigma\in \mathcal E_{K,int}\cup \mathcal E_K^D} {\rm m}(\sigma){\rm d}_{\sigma}\right)\leq \displaystyle\frac{p}{\xi} {\rm m} (\{{\mathbf w}^m>0\}). \end{equation} We deduce from \eqref{inegdep}, \eqref{ineg1}, \eqref{ineg2} and \eqref{ineg3} that $$ \beta_{\mathbf U} F_m\leq {2}\Vert {\mathbf U}\Vert_{L^\infty} (F_m)^{1/2}(\frac{p}{\xi}{\rm m} (\{{\mathbf w}^m>0\}))^{1/2}+{\Vert f\Vert_{L^\infty}{\rm m} (\{{\mathbf w}^m>0\})}, $$ which yields \eqref{majEm} using Young's inequality and the bounds $E_m\leq F_m$ and $\beta_{\mathbf U}\leq1$. \end{proof} {Before stating the main result of the section, we need a technical lemma. \begin{lemma}\label{lem:sequence} Let $(u_n)_{n\in\mathbb N}$ be a sequence of non-negative real numbers and let $K, \rho>0$ and $\alpha >1$. Then if for all $n\in\mathbb N$ \[ u_{n+1}\,\leq\, K\,\rho^{n}\,u_{n}^\alpha\,, \] one has \[ 0\leq u_n\,\leq\, \left(u_0\,\rho^{\frac{1}{(\alpha-1)^2}}\,K^{\frac{1}{\alpha-1}}\right)^{\alpha^n}\,\rho^{-\frac{n(\alpha-1)+1}{(\alpha-1)^2}}\,K^{-\frac{1}{\alpha-1}} \] for all $n\in\mathbb N$ and the bound is optimal. In particular, if $u_0\leq \rho^{-\frac{1}{(\alpha-1)^2}}\,K^{-\frac{1}{\alpha-1}}$, then $\lim u_n=0$. \end{lemma} \begin{proof} Just observe that the sequence $v_n = u_n\,\rho^{\frac{n(\alpha-1) + 1}{(\alpha-1)^2}}\,K^{\frac{1}{\alpha-1}}$ satisfies $0\leq v_{n+1}\leq v_{n}^\alpha$ for all $n\geq0$ which directly yields the result. \end{proof}} { \begin{proposition}\label{mainprop} Assume that $f_K\geq 0$ for all $K\in\mathcal T$ and $v_\sigma^D \in [0,1]$ for all $\sigma \in \mathcal E^D$, so that $v_K\geq 0$ for all $K\in\mathcal T$. Then, there exists $\eta>0$ depending only on $\Omega$, $p$ and $\xi$ such that one has the implication \begin{equation}\label{resprop} \displaystyle E_1\leq\ \eta\ \frac{\beta_{\mathbf U}^4}{(\|\mathbf{U}\|_{L^\infty}^2+\|f\|_{L^\infty})^2}\quad\Rightarrow\quad (v_K\leq 2,\ \forall K\in\mathcal T)\,. \end{equation} \end{proposition} } \begin{proof} The proof consists in establishing an induction property on $E_m$ which guarantees that if $E_1$ is small enough then $\lim E_m =0$. Then, as $\lim C_m=2$ and thanks to the discrete Poincar\'e inequality, we deduce that $$ \displaystyle\sum_{K\in\mathcal T} {\rm m}(K) \left( \log (1 +(v_K-2)^+)\right)^2=0, $$ which implies $v_K\leq 2$ for all $K\in\mathcal T$. For establishing the induction, first observe that as $C_m=C_{m-1} +2^{-m+1}$, for any $q>0$ we have: \begin{equation}\label{eq:nonlinbound} {\mathbf 1}_{\{{\mathbf w}^m>0\}}\leq \frac{\left(\log (1+({\mathbf w}^{m-1})^+)\right)^q}{(\log (1+2^{-m+1}))^q}{\mathbf 1}_{\{{\mathbf w}^{m-1}>0\}}, \end{equation} and thus $$ {\rm m} (\{{\mathbf w}^m>0\})\leq \frac{1}{(\log (1+2^{-m+1}))^q}\displaystyle\sum_{K\in\mathcal T} {\rm m}(K) \left( \log (1 +(w_K^{m-1})^+)\right)^q. $$ We may choose for instance $q=3$ and apply a discrete Poincar\'e-Sobolev inequality (whose constant $C_{\Omega,p}$ depends only on $\Omega$ and $p$), which leads to \begin{equation}\label{majmes} {\rm m} (\{{\mathbf w}^m>0\})\leq\frac{1}{(\log (1+2^{-m+1}))^3}\frac{C(\Omega)}{\xi^{3/2}} E_{m-1}^{3/2}. \end{equation} Noticing that for $x\in[0,1]$, $(\log(1+x))^3\geq (\log 2)^3 x^3$, we deduce from \eqref{majEm} and \eqref{majmes} that $$ E_m\leq \frac{4}{\beta_{\mathbf U}^2}\left(\Vert {\mathbf U}\Vert_{L^\infty}^2+\Vert f\Vert_{L^\infty}\right)\frac{{\tilde C}_{\Omega,p}}{\xi^{3/2}}8^{m-1}E_{m-1}^{3/2}. $$ Thus the sequence $(E_m)_{m\geq 0}$ satisfies the hypothesis of Lemma~\ref{lem:sequence} with $\alpha=3/2$ and $K$ proportional to $(\Vert {\mathbf U}\Vert_{L^\infty}^2+\Vert f\Vert_{L^\infty})/\beta_{\mathbf U}^2$. We deduce the upper bound for $E_1$ under which $\lim E_m =0$. \end{proof} {\em Remark:} The arguments developed in this section still hold, up to minor adaptation, for $f\in L^r(\Omega)$ with $r>p/2$. \section{Proof of Theorem \ref{mainthm}}\label{sec:proof} First observe that if one replaces the data $f$ and $v^D$ by either $f^+$ and $(v^D)^+$, or $f^-$ and $(v^D)^-$, in the scheme \eqref{scheme}-\eqref{numflux}, then the corresponding solutions, say respectively $\mathbf{P}=(P_K)_{K\in\mathcal T}$ and $\mathbf{N}=(N_K)_{K\in\mathcal T}$, are non-negative and such that ${\mathbf v}=\mathbf{P}-\mathbf{N}$ is the solution to \eqref{scheme}-\eqref{numflux} in the original framework. From there let us show that there is ${\overline M}>V^D_+:=\max(\|(v^D)^+\|_{L^\infty},1)$ such that for all $K\in\mathcal T$ one has $0\leq P_K\leq {\overline M}$. The bound for $\mathbf{N}$, which is denoted by ${\underline M}$, can be obtained in the same way. Let $M>V^D_+$. First observe that $\mathbf{P}^M:=\mathbf{P}/M$ satisfies the scheme \eqref{scheme}-\eqref{numflux} where the source term and boundary data have been replaced by $f^+/M$ and $(v^D)^+/M$ respectively. Moreover, one can apply Proposition~\ref{prop:fund_ineq}, which yields \begin{equation}\label{eq:ineqPMagain} E_1(\mathbf{P}^M)\leq \frac{4p}{\beta_{\mathbf U}^2}\left(\Vert {\mathbf U}\Vert_{L^\infty}^2 + \frac{\Vert f^+\Vert_{L^\infty}}{M}\right) {\rm m} (\{{\mathbf P}^M>1\})\,. \end{equation} Now observe that $\mathbf{P}\,=\,M\,\mathbf{P}^M\,=\,V^D_+\,\mathbf{P}^{V^D_+}$. Therefore, $$ \begin{array}{rcl} \displaystyle E_1(\mathbf{P}^M)&\leq& \displaystyle \frac{4p}{\beta_{\mathbf U}^2}\Big{(}\Vert {\mathbf U}\Vert_{L^\infty}^2\,{\rm m} (\{{\mathbf P}^{ V^D_+}>M / V^D_+\}) + \frac{\Vert f^+\Vert_{L^\infty}}{M}{\rm m}(\Omega)\Big{)} \\[1em] &\leq&\displaystyle\frac{4p}{\beta_{\mathbf U}^2}\Big{(}\Vert {\mathbf U}\Vert_{L^\infty}^2 \sum_{K\in\mathcal T}{\rm m}(K)\,\frac{\log(1+(P_K^{ V^D_+}-1)^+)^2}{\log(M / V^D_+)^2} + \frac{\Vert f^+\Vert_{L^\infty}}{M}{\rm m}(\Omega)\Big{)}\\[1em] &\leq&\displaystyle\frac{C_{\Omega,p}}{\xi\beta_{\mathbf U}^2}\Vert {\mathbf U}\Vert_{L^\infty}^2\, \frac{E_1(\mathbf{P}^{V^D_+})}{\log(M / V^D_+)^2} + \frac{4p\,{\rm m}(\Omega)}{\beta_{\mathbf U}^2}\frac{\Vert f^+\Vert_{L^\infty}}{M}\,\,, \end{array} $$ where we used an argument similar to \eqref{eq:nonlinbound} in the second inequality and a discrete Poincar\'e inequality in the third one. Then, by using \eqref{eq:ineqPMagain} again we get $$ E_1(\mathbf{P}^{V^D_+})\ \leq\ \frac{4\,p\,{\rm m}(\Omega)}{\beta_{\mathbf U}^2}\left(\Vert {\mathbf U}\Vert_{L^\infty}^2 + \frac{\Vert f^+\Vert_{L^\infty}}{V^D_+}\right) $$ Therefore, the smallness condition of Proposition~\ref{mainprop} is satisfied by $E_1(\mathbf{P}^M)$ if \begin{multline}\label{boundM} \left[\Vert{\mathbf U}\Vert_{L^\infty}^2\left(\Vert {\mathbf U}\Vert_{L^\infty}^2 + \frac{\Vert f^+\Vert_{L^\infty}}{V^D_+}\right)+ \frac{\Vert f^+\Vert_{L^\infty}}{M}\log\left(\frac{M}{V^D_+}\right)^2\right]\,\left(\Vert {\mathbf U}\Vert_{L^\infty}^2 + \frac{\Vert f^+\Vert_{L^\infty}}{M}\right)^2\\\ \leq\ \,C_{\Omega,\xi,p}\,\beta_{\mathbf U}^4\,\log\left(\frac{M}{V^D_+}\right)^2\,. \end{multline} It is clear that \eqref{boundM} is satisfied for $M$ large enough, which permits to define ${\overline M}$. Observe that if $v^D_+=0$ ($V_+^D=1$) and $\mathbf U=0$, ${\overline M} = \widetilde{C}_{\Omega,\xi,p}\|f^+\|_{L^\infty}$ works as expected. \smallskip \textbf{Acknowledgements.} The authors thank the Labex CEMPI (ANR-11-LABX-0007-01) and the ANR MOHYCON (ANR-17-CE40-0027-01) for their support. They also want to thank Alexis F. Vasseur for fruitful exchanges on the subject. \bibliographystyle{spmpsci}

https://arxiv.org/abs/2204.12214

Derivations of a family of quantum second Weyl algebras

In view of a well-known theorem of Dixmier, its is natural to consider primitive quotients of $U_q^+(\mathfrak{g})$ as quantum analogues of Weyl algebras. In this work, we study these primitive quotients in the $G_2$ case and compute their Lie algebra of derivations.

\section{Introduction} Weyl algebras have been extensively studied in the last 60 years due to their link to Lie theory, differential operators, quantum mechanics, etc. One of the main questions remaining is the famous Dixmier Conjecture that asserts that every endomorphism of a complex Weyl algebra is an automorphism. Let ${\mathbb K}$ be a field and $q$ be a non-zero element of ${\mathbb K}$ that is not a root of unity. The aim of this article is to produce quantum analogues of the second Weyl algebra and to compare their properties to those of the second Weyl algebra. There exist in the literature various families of ``quantum Weyl algebras'', e.g. the so-called quantum Weyl algebras and generalised Weyl algebras (GWA for short). Most of the time, they are obtained by generators and relations through a deformation of the classical defining relation of the first Weyl algebra: $xy-yx=1$. To produce potential quantisations, we take a different approach in this article. Our inspiration comes from a Theorem of Dixmier (see, for instance, \cite[Th\'eor\`eme 4.7.9]{dix}) that asserts that primitive quotients of enveloping algebras of complex nilpotent Lie algebras are isomorphic to Weyl algebras. We have at hand a quantum analogue of at least some enveloping algebras of complex nilpotent Lie algebras, namely the positive part $U_q^+(\mathfrak{g})$ of a quantised enveloping algebra $U_q^+(\mathfrak{g})$ of a complex simple Lie algebra $\mathfrak{g}$. As a consequence, it is natural to consider primitive quotients of $U_q^+(\mathfrak{g})$ as quantum analogues of Weyl algebras. In the $A_2$ and $B_2$ cases, primitive ideals of $U_q^+(\mathfrak{g})$ have been classified and it turns out that in the $B_2$ case, some of the resulting primitive quotients provide `nice' quantum analogues of the first Weyl algebra. For instance, they are simple---this is not the case of quantum Weyl algebras---and do not possess non-trivial units---this is not the case of a quantum GWA over a Laurent polynomial ring. (See \cite{sl} for details.) The present article is concerned with the $G_2$ case. More precisely, we identify a family of primitive ideals of $U_q^+(G_2)$ and then proceed in proving that the corresponding primitive quotients have (at least for some choice of the parameters) properties similar to those of the second Weyl algebra. More precisely, the center of $U_q^+(G_2)$ is a polynomial algebra in two variables ${\mathbb K}[\Omega_1,\Omega_2]$ and we prove that the quotient algebra $$A_{\alpha,\beta}:= U_q^+(G_2) / \langle \Omega_1 - \alpha, \Omega_2 - \beta\rangle$$ is simple for all $(\alpha , \beta)\neq (0,0)$. We then proceed and study these quotient algebras. In particular, we show that $A_{\alpha,\beta}$ has the same (Gelfand-Kirillov) dimension as the second Weyl algebra $A_2({\mathbb K})$. We also establish that for certain choice of the parameters $\alpha$ and $\beta$, the algebra $A_{\alpha,\beta}$ is a deformation of a quadratic extension of $A_2({\mathbb K})$ at $q=1$. In the final section, we compute the derivations of $A_{\alpha,\beta}$. Our results show that when $\alpha$ and $\beta$ are both non-zero, all derivations of $A_{\alpha,\beta}$ are inner, a property that is well known to hold in $A_2({\mathbb K})$. In view of the celebrated Dixmier Conjecture, it would be interesting to describe automorphisms and endomorphisms of $A_{\alpha,\beta}$ when $\alpha$ and $\beta$ are both non-zero. We intend to come back to these questions in the future. This article is organized as follows. In Section 2, we recall the presentation of $U_q^+(G_2)$ as a so-called quantum nilpotent algebra (QNA for short). This allows the use of two different tools to study the prime and primitive spectra of $U_q^+(G_2)$: the ${\mathcal H}$-stratification theory of Goodearl and Letzter, and the Deleting Derivation Theory of Cauchon. We recall both theories in the context of $U_q^+(G_2)$ in Section 2. In Section 3, we use these two theories to establish that $\langle \Omega_1 - \alpha, \Omega_2 - \beta\rangle$ is a maximal ideal of $U_q^+(G_2)$ when $(\alpha, \beta) \neq (0,0)$. In Section 4, we focus on comparing $A_{\alpha,\beta}$ with the second Weyl algebra $A_2({\mathbb K})$. In particular, we show that both have Gelfand-Kirillov dimension equal to 4. Through a direct computation, we also establish that $A_{1,\frac{1}{9(q^6-1)}}$ is a quadratic extension of $A_2({\mathbb K})$ at $q=1$. In this section we also compute a linear basis for $A_{\alpha,\beta}$. In the final section, we compute the derivations of $A_{\alpha,\beta}$. Our strategy here is to make use of the following tower of algebras arising from the Deleting Derivation Algorithm: $$A_{\alpha,\beta}\subset {\mathcal R}_6={\mathcal R}_7\Sigma_6^{-1}\subset {\mathcal R}_5={\mathcal R}_6\Sigma_5^{-1}\subset {\mathcal R}_4={\mathcal R}_5\Sigma_4^{-1} \subset {\mathcal R}_3.$$ The later algebra ${\mathcal R}_3$ is a simple quantum torus whose derivations have been described by Osborn and Passman in \cite{op}. We pull back their description to obtain a description of the derivations of $A_{\alpha,\beta}$ through a step-by-step process consisting in ``reverting'' the Deleting Derivation Algorithm. Our results show that when $\alpha$ or $\beta$ is equal to zero, then the first Hochschild cohomology group of $A_{\alpha,\beta}$ is a 1-dimension vector space, whereas when both $\alpha$ and $\beta$ are non-zero, all derivations are inner. \section{The quantum nilpotent algebra $U_q^+(G_2)$ and its primitive ideals} \subsection{The quantum nilpotent algebra $U_q^+(G_2)$} Let ${\mathbb K}$ be a field and $q$ be a non-zero element of ${\mathbb K}$ that is not a root of unity. The algebra of $U_q^+(G_2)$ is the so-called positive part of the quantum enveloping algebra $U_q(\mathfrak{g})$ of a Lie algebra $\mathfrak{g}$ of type $G_2$. It is well known, see for instance \cite{bg}, that this algebra is generated over ${\mathbb K}$ by two indeterminates $E_{\alpha}$ and $E_{\beta}$ subject to the following quantum Serre relations: \begin{align*} (S1) \ \ E_{\alpha}^4E_{\beta}&- \left[\begin{smallmatrix} 4 \\ \\ 1 \end{smallmatrix}\right]_q E_{\alpha}^3E_{\beta}E_{\alpha}+\left[\begin{smallmatrix} 4 \\ \\ 2 \end{smallmatrix}\right]_q E_{\alpha}^2E_{\beta}E_{\alpha}^2-\left[\begin{smallmatrix} 4 \\ \\ 1 \end{smallmatrix}\right]_q E_{\alpha}E_{\beta}E_{\alpha}^3+E_{\beta}E_{\alpha}^4=0,\\ (S2) \ \ E_{\beta}^2E_{\alpha}&-\left[\begin{smallmatrix} 2 \\ \\ 1 \end{smallmatrix}\right]_{q^3} E_{\beta}E_{\alpha}E_{\beta}+E_{\alpha}E_{\beta}^2=0,\\ \end{align*} where $\left[\begin{smallmatrix} n \\ \\ i \end{smallmatrix}\right]_{z}$ denotes the quantum binomial coefficients (see \cite[I.6.1]{bg}). One can construct a PBW-basis of $U_q^+(G_2)$ using the so-called Lusztig automorphisms of $U_q(\mathfrak{g})$, see for instance \cite[I.6.8]{bg}. In the present case, such a basis was computed by De Graaf in \cite{degraaf1}. We will use the convention of that paper, but with $E_1:=E_{\alpha}, \ E_2:=E_{3\alpha+\beta}, \ E_3:=E_{2\alpha+\beta}, \ E_4:=E_{3\alpha+2\beta}, \ E_5:=E_{\alpha+\beta}$ and $E_6:=E_{\beta}$. With these notations, the defining relations of $U_q^+(G_2)$ are as follows: \begin{align*} E_2E_1&=q^{-3}E_1E_2& E_3E_1&=q^{-1}E_1E_3-(q+q^{-1}+q^{-3})E_2\\ E_3E_2&=q^{-3}E_2E_3&E_4E_1&=E_1E_4+(1-q^2)E_3^2\\ E_4E_2&=q^{-3}E_2E_4-\frac{q^4-2q^2+1}{q^4+q^2+1}E_3^3&E_4E_3&=q^{-3}E_3E_4 \\ E_5E_1&=qE_1E_5-(1+q^2)E_3&E_5E_2&=E_2E_5+(1-q^2)E_3^2\\ E_5E_3&=q^{-1}E_3E_5-(q+q^{-1}+q^{-3})E_4&E_5E_4&=q^{-3}E_4E_5\\ E_6E_1&=q^3E_1E_6-q^3E_5&E_6E_2&=q^3E_2E_6+(q^4+q^2-1)E_4+(q^2-q^4)E_3E_5\\ E_6E_3&=E_3E_6+(1-q^2)E_5^2 &E_6E_5&=q^{-3}E_5E_6\\ E_6E_4&=q^{-3}E_4E_6-\frac{q^4-2q^2+1}{q^4+q^2+1}E_5^3, \end{align*} and the monomials $E_1^{k_1} \dots E_6^{k_6}$ ($k_1, \dots ,k_6 \in \mathbb{N} $) form a basis of $U_q^+(G_2)$ over ${\mathbb K}$. Even better, one may write $U_q^+(G_2)$ as a Quantum Nilpotent Algebra (QNA for short) or Cauchon-Goodearl-Letzter extension in the sense of \cite[Definition 3.1]{llr-cgl}, by adjoining the generators $E_i$ in lexicographic order. This means in particular that $U_q^+(G_2)$ can be presented as an iterated Ore extension: \[ U_q^+(G_2)={\mathbb K}[E_{1}][E_{2};\sigma_{2},\delta_{2}]\cdots [E_{6};\sigma_{6},\delta_{6}], \] where the $\sigma_{i}$ are automorphisms and the $\delta_{i}$ are left $\sigma_{i}$-derivations of the appropriate subalgebras. We would not need the precise definition of a QNA for what follows, but it is worth reminding the reader of the algebraic torus action involved in writing $U_q^+(G_2)$ as a QNA. The algebraic torus ${\mathcal H}=({\mathbb K}^\times)^{2}$ acts by automorphisms on $U_q^+(G_2)$ as follows: \[ h\cdot E_i=h_i E_{i} \mbox{ for all } i \in \{1,6\} \mbox{ and } h=(h_1,h_6) \in {\mathcal H} . \] Note that the action of the automorphism $h$ on the generators $E_2, \dots , E_5$ follows from the above defining relations. By \cite[Theorem II.2.7]{bg}, the action of ${\mathcal H}$ on $U_q^+(G_2)$ is rational in the sense of \cite[Definition II.2.6]{bg}. A consequence of the QNA condition is that important tools such as Cauchon's deleting derivations procedure and the Goodearl-Letzter stratification theory (this is the origin of the CGL terminology, see \cite{llr-cgl}) are available to study prime and primitive ideals. These ideas will be introduced in following sections. At the moment, we merely note that it is immediate that $U_q^+(G_2)$ is a noetherian domain and that all prime ideals are completely prime (in the case of $U_q^+(G_2)$, it was proved in \cite[Section 5]{picp}). We denote by $F_q$ its skew-field of fractions, i.e. $F_q:= \mathrm{Frac}(U_q^+(G_2))$. \subsection{Prime ideals in $U_q^+(G_2)$ and ${\mathcal H}$-stratification} A two-sided ideal $I$ of $U_q^+(G_2)$ is said to be {\em ${\mathcal H}$-invariant} if $h\cdot I=I$ for all $h \in {\mathcal H}$. An {\em ${\mathcal H}$-prime ideal} of $U_q^+(G_2)$ is a proper ${\mathcal H}$-invariant ideal $J$ of $U_q^+(G_2)$ such that if $J$ contains the product of two ${\mathcal H}$-invariant ideals of $U_q^+(G_2)$ then $J$ contains at least one of them. We denote by ${\mathcal H}$-${\rm Spec}(U_q^+(G_2))$ the set of all ${\mathcal H}$-prime ideals of $U_q^+(G_2)$. Observe that if $P$ is a prime ideal of $U_q^+(G_2)$ then \begin{equation} (P:{\mathcal H})\ := \ \bigcap_{h\in {\mathcal H}} h\cdot P \end{equation} is an ${\mathcal H}$-prime ideal of $U_q^+(G_2)$. Indeed, let $J$ be an ${\mathcal H}$-prime ideal of $U_q^+(G_2)$. We denote by ${\rm Spec}_J (U_q^+(G_2))$ the {\em ${\mathcal H}$-stratum} associated to $J$; that is, \begin{equation} {\rm Spec}_J (U_q^+(G_2))=\{ P \in {\rm Spec}(U_q^+(G_2)) \mbox{ $\mid$ } (P:{\mathcal H})=J \}. \end{equation} Then the ${\mathcal H}$-strata of ${\rm Spec}(U_q^+(G_2))$ form a partition of ${\rm Spec}(U_q^+(G_2))$ \cite[Chapter II.2]{bg}; that is, \begin{equation} \label{eq:Hstratification} {\rm Spec}(U_q^+(G_2))= \bigsqcup_{J \in {\mathcal H}\mbox{-}{\rm Spec}(U_q^+(G_2))}{\rm Spec}_J(U_q^+(G_2)). \end{equation} This partition is the so-called {\em ${\mathcal H}$-stratification} of ${\rm Spec}(U_q^+(G_2))$. It follows from the work of Goodearl and Letzter \cite{gz2} that every ${\mathcal H}$-prime ideal of $U_q^+(G_2)$ is completely prime, so ${\mathcal H}$-${\rm Spec}(U_q^+(G_2))$ coincides with the set of ${\mathcal H}$-invariant completely prime ideals of $U_q^+(G_2)$. Moreover there are precisely $|W|$ ${\mathcal H}$-prime ideals in $U_q^+(G_2)$, where $W$ denotes the Weyl group of type $G_2$ (see \cite[Remark 6.2.2]{meriaux}). As a consequence, the ${\mathcal H}$-stratification of ${\rm Spec} ( U_q^+(G_2) ) $ is finite and so the full strength of the ${\mathcal H}$-stratification theory of Goodearl and Letzter is available to study ${\rm Spec} (U_q^+(G_2))$. For each ${\mathcal H}$-prime ideal $J$ of $U_q^+(G_2)$, the space ${\rm Spec}_J(U_q^+(G_2))$ is homeomorphic to the prime spectrum ${\rm Spec} ({\mathbb K}[z_1^{\pm 1},\ldots ,z_d^{\pm 1}])$ of a commutative Laurent polynomial ring whose dimension depends on $J$, \cite[Theorems II.2.13 and II.6.4]{bg}. These dimensions were computed in \cite{bcl,yak}. Finally, let us mention that the primitive ideals of $A$ are precisely the prime ideals that are maximal in their ${\mathcal H}$-strata \cite[Theorem II.8.4]{bg}. In this article, we will mainly focus on one specific ${\mathcal H}$-stratum. Since $U_q^+(G_2)$ is a domain, $0$ (technically, $\langle0\rangle$) is clearly an ${\mathcal H}$-invariant completely prime ideal of $U_q^+(G_2)$, and so an ${\mathcal H}$-prime, and we will focus on computing its stratum, the so-called $0$-stratum. The motivation here is twofold: first, in the $B_2$ case, we obtain ``new'' quantum deformation of the first Weyl algebra as $U_q^+(B_2) / P$, where $P$ is a primitive ideal from the $0$-stratum of ${\rm Spec} (U_q^+(B_2))$ \cite{sl}. Next, in the present case, we would like to construct algebras of GK dimension 4 as explained in the introduction. Since Tauvel's height formula holds in $U_q^+(G_2)$ \cite{gll}, we need to quotient $U_q^+(G_2)$ by a primitive ideals of height 2. Given that the ${\mathcal H}$-spectrum of $U_q^+(G_2)$ is homeomorphic to the Weyl group of type $G_2$, such primitive ideals can only be found in the $0$-stratum and the strata associated to one of the two height 1 ${\mathcal H}$-primes. In this article, we mainly present results for the $0$-stratum, but we will also indicate results obtained for the primitive quotients coming from the height 1 ${\mathcal H}$-prime strata. \subsection{Deleting derivations algorithms in $U_q^+(G_2)$} As $U_q^+(G_2)$ is a QNA, we can apply Cauchon's Deleting Derivation Algorithm to study its prime spectrum. Recall first that $U_q^+(G_2)$ is an itereated Ore extension of the form: $$U_q^+(G_2)={\mathbb K}[E_1][E_2;\sigma_2][E_3;\sigma_3,\delta_3][E_4;\sigma_4,\delta_4][E_5;\sigma_5,\delta_5][E_6;\sigma_6,\delta_6];$$ where, $\sigma_2$ denotes the automorphism of ${\mathbb K}[E_1]$ defined by: $$\sigma_2(E_1)=q^{-3}E_1,$$ $\sigma_3$ denotes the automorphism of ${\mathbb K}[E_1][E_2;\sigma_2]$ defined by: $$\sigma_3(E_1)=q^{-1}E_1 \ \ \ \ \ \sigma_3(E_2)=q^{-3}E_2,$$ $\delta_3$ denotes the $\sigma_3$-derivation of ${\mathbb K}[E_1][E_2;\sigma_2]$ defined by: $$\delta_3(E_1)=-(q+q^{-1}+q^{-3})E_2 \ \ \ \ \ \delta_3(E_2)=0,$$ $\sigma_4$ denotes the automorphism of ${\mathbb K}[E_1]\cdots[E_3;\sigma_3,\delta_3]$ defined by: $$\sigma_4(E_1)=E_1 \ \ \ \ \ \sigma_4(E_2)=q^{-3}E_2 \ \ \ \ \sigma_4(E_3)=q^{-3}E_3,$$ $\delta_4$ denotes the $\sigma_4$-derivation of ${\mathbb K}[E_1]\cdots[E_3;\sigma_3,\delta_3]$ defined by: $$\delta_4(E_1)=(1-q^2)E_3^2 \ \ \ \ \ \delta_4(E_2)=\frac{-q^4+2q^2-1}{q^4+q^2+1}E_3^3 \ \ \ \ \ \delta_4(E_3)=0,$$ $\sigma_5$ denotes the automorphism of ${\mathbb K}[E_1]\cdots[E_4;\sigma_4,\delta_4]$ defined by: $$\sigma_5(E_1)=qE_1 \ \ \ \ \ \sigma_5(E_2)=E_2 \ \ \ \ \sigma_5(E_3)=q^{-1}E_3 \ \ \ \ \ \sigma_5(E_4)=q^{-3}E_4,$$ $\delta_5$ denotes the $\sigma_5$-derivation of ${\mathbb K}[E_1]\cdots[E_4;\sigma_4,\delta_4]$ defined by: $$\delta_5(E_1)=-(1+q^2)E_3 \ \ \ \ \ \delta_5(E_2)=(1-q^2)E_3^2 \ \ \ \ \ \delta_5(E_3)=-(q+q^{-1}+q^{-3})E_4 \ \ \ \ \delta_5(E_4)=0,$$ $\sigma_6$ denotes the automorphism of ${\mathbb K}[E_1]\cdots[E_5;\sigma_5,\delta_5]$ defined by: $$\sigma_6(E_1)=q^3E_1 \ \ \ \ \ \sigma_6(E_2)=q^3E_2 \ \ \ \ \sigma_6(E_3)=E_3 \ \ \ \ \ \sigma_6(E_4)=q^{-3}E_4 \ \ \ \ \sigma_6(E_5)=q^{-3}E_5,$$ and $\delta_6$ denotes the $\sigma_6$-derivation of ${\mathbb K}[E_1]\cdots[E_5;\sigma_5,\delta_5]$ defined by: $$\delta_6(E_1)=-q^3E_5 \ \ \ \ \ \delta_6(E_2)=(q^2-q^4)E_3E_5+(q^4+q^2-1)E_4 \ \ \ \ \ \delta_6(E_3)=(1-q^2)E_5^2 $$ $$\delta_6(E_4)=\dfrac{-q^4+2q^2-1}{q^4+q^2+1}E_5^3 \ \ \ \ \ \delta_6(E_5)=0.$$ The deleting derivations algorithm (DDA for short) constructs by a decreasing induction a family $\{E_{1,j},\dots,E_{6,j}\}$ of elements of the division ring of fractions $F_q={\rm Fract}(U_q^+(G_2))$ of $U_q^+(G_2)$ for each $2\leq j \leq 7$. The precise definition of these elements in the general context of QNAs can be found in \cite{ca}. In the present case, direct computation leads to: \begin{align*} E_{1,6}&=E_1+rE_5E_6^{-1}\\ E_{2,6}&=E_2+tE_3E_5E_6^{-1}+uE_4E_6^{-1}+nE_5^3E_6^{-2}\\ E_{3,6}&=E_3+sE_5^2E_6^{-1}\\ E_{4,6}&=E_4+bE_5^3E_6^{-1}\\ E_{1,5}&=E_{1,6}+hE_{3,6}E_{5,6}^{-1}+gE_{4,6}E_{5,6}^{-2}\\ E_{2,5}&=E_{2,6}+fE_{3,6}^2E_{5,6}^{-1}+pE_{3,6}E_{4,6}E_{5,6}^{-2}+eE_{4,6}^2E_{5,6}^{-3 }\\ E_{3,5}&=E_{3,6}+aE_{4,6}E_{5,6}^{-1}\\ E_{1,4}&=E_{1,5}+sE_{3,5}^2E_{4,5}^{-1}\\ E_{2,4}&=E_{2,5}+bE_{3,5}^3E_{4,5}^{-1}\\ E_{1,3}&=E_{1,4}+aE_{2,4}E_{3,4}^{-1}\\ T_1&:=E_{1,2}=E_{1,3}\\ T_2&:= E_{2,2}=E_{2,3}=E_{2,4}\\ T_3&:=E_{3,2}=E_{3,3}=E_{3,4}=E_{3,5}\\ T_4&:=E_{4,2}=E_{4,3}=E_{4,4}=E_{4,5}=E_{4,6}\\ T_5&:=E_{5,2}=E_{5,3}=E_{5,4}=E_{5,5}=E_{5,6}=E_5\\ T_6&:=E_{6,2}=E_{6,3}=E_{6,4}=E_{6,5}=E_{6,6}=E_6, \end{align*} where the parameters $a,b,e,f,g,h,n,p,r,s,t,u$ are all defined in Appendix \ref{appenc}. In the following, we set $A:=U_q^+(G_2)$ and we denote by $A^{(j)}$ the subalgebra of $F_q$ generated by $E_{1,j}$, ..., $E_{6,j}$. The following results were proved by Cauchon \cite[Th\'eor\`eme 3.2.1 and Lemme 4.2.1]{ca}. For $2\leq j \leq 7$, we have: \begin{enumerate} \item when $j=7$, $(E_{1,7},\dots,E_{6,7}) =(E_1, \dots , E_6)$, so that $A^{(7)}=A=U_q^+(G_2)$; \item $A^{(j)}$ is isomorphic to an iterated Ore extension of the form $${\mathbb K}[y_1]\dots[y_{j-1};\sigma_{j-1},\delta_{j-1}][y_j;\tau_j]\cdots[y_6;\tau_6]$$ by an isomorphism that sends $E_{i,j}$ to $y_i$ ($1 \leq i \leq 6$), where $\tau_j,\dots,\tau_6$ denote the ${\mathbb K} $-linear automorphisms such that $\tau_{\ell}(y_i)=\lambda_{\ell,i} y_i$ ($1 \leq i \leq \ell$). \item Assume that $j \neq 7$ and set $\Sigma_j:=\{E_{j,j+1}^n ~|~ n\in \mathbb{N} \}= \{E_{j,j}^n ~|~ n\in \mathbb{N} \}$. \\This is a multiplicative system of regular elements of $A^{(j)}$ and $A^{(j+1)}$, that satisfies the Ore condition in $A^{(j)}$ and $A^{(j+1)}$. Moreover we have $$A^{(j)}\Sigma_j^{-1}=A^{(j+1)}\Sigma_j^{-1}.$$ \end{enumerate} It follows from these results that $A^{(j)}$ is a noetherian domain, for all $2\leq j \leq 7$. As in \cite{ca}, we use the following notation. \begin{notation} We set $\overline{A}:=A^{(2)}$ and $T_i:=E_{i,2}$ for all $1\leq i \leq 6$. \end{notation} It follows from \cite[Proposition 3.2.1]{ca} that $\overline{A}$ is a quantum affine space in the indeterminates $T_1,\dots,T_6$ and so can be presented as an iterated Ore extension in the $T_i$s with no skew-derivations: it is for this reason that Cauchon used the expression ``effacement des d\'erivations". More precisely, let $M=\left( \mu_{i,j} \right) \in M_6({\mathbb K} ^*)$ be the multiplicatively antisymmetric matrix defined as follows: $$M = \begin{bmatrix} 0&3&1&0&-1&-3\\ -3&0&3&3&0&-3\\ -1&-3&0&3&1&0\\ 0&-3&-3&0&3&3\\ 1&0&-1&-3&0&3\\ 3&3&0&-3&-3&0 \end{bmatrix}.$$ Then we have \begin{equation}\overline{A}={\mathbb K}_{q^M} [T_1,\dots,T_6]=\mathcal{O}_{q^M}({\mathbb K} ^6), \end{equation} where ${\mathbb K}_{q^M} [T_1,\dots,T_6]=\mathcal{O}_{q^M}({\mathbb K} ^6)$ denotes the ${\mathbb K}$-algebra generated by $T_1, \dots , T_6$ with relations $T_iT_j = q^{\mu_{i,j}}T_jT_i$ for all $i,j$. \subsection{Canonical embedding} \label{c1s1} Since $A=U_q^+(G_2)$ is a QNA, one can use Cauchon's DDA in order to relate the prime spectrum of $A$ to the prime spectrum of the associated quantum affine space $\overline{A}$. More precisely, the DDA allows the construction of embeddings \begin{equation} \psi_j:{\rm Spec}(A^{(j+1)}) \longrightarrow {\rm Spec}(A^{(j)}) \qquad (2\leq j \leq 6). \end{equation} Recall from \cite[Section 4.3]{ca} that these embeddings are defined as follows. Let $P \in {\rm Spec}(A^{(j+1)})$. Then $$ \psi_j (P) = \left\{\begin{array}{ll} P\Sigma_j^{-1} \cap A^{(j)} & \mbox{ if } E_{j,j+1}=T_j \notin P \\ g_j^{-1} \left( P/\langle E_{j,j+1}\rangle \right) & \mbox{ if } E_{j,j+1} \in P \\ \end{array}\right. $$ where $g_j$ denotes the surjective homomorphism $$ g_j:A^{(j)}\rightarrow A^{(j+1)}/\langle E_{j,j+1} \rangle $$ defined by $$ g_j(E_{i,j}):=E_{i,j+1} + \langle E_{j,j+1} \rangle $$ (for more details, see \cite[Lemme 4.3.2]{ca}). It was proved by Cauchon \cite[Proposition 4.3.1]{ca} that $\psi_j$ induces an increasing homeomorphism from the topological space $$\{P \in {\rm Spec}(A^{(j+1)}) \mid E_{j,j+1} \notin P \}$$ onto $$\{Q \in {\rm Spec}(A^{(j)}) \mid E_{j,j} \notin Q \}$$ whose inverse is also an increasing homeomorphism. Also, $\psi_j$ induces an increasing homeomorphism from $$\{P \in {\rm Spec}(A^{(j+1)}) \mid E_{j,j+1} \in P \}$$ onto its image by $\psi_j$ whose inverse similarly is an increasing homeomorphism. Note however that, in general, $\psi_j$ is not an homeomorphism from ${\rm Spec}(A^{(j+1)})$ onto its image. Composing these embeddings, we get an embedding \begin{equation} \psi:=\psi_2 \circ \dots \circ \psi_6 : {\rm Spec}(A) \longrightarrow {\rm Spec}(\overline{A}), \end{equation} which is called the \emph{canonical embedding} from ${\rm Spec}(A)$ into ${\rm Spec}(\overline{A})$. The canonical embedding $\psi$ is ${\mathcal H}$-equivariant so that $\varphi (\hspec(A)) \subseteq \hspec(\overline{A})$. Interestingly, the set $\hspec(\overline{A})$ has been described by Cauchon as follows. For any subset $C$ of $\{1,\dots,6\}$, let $K_C$ denote the ${\mathcal H}$-prime ideal of $\overline{A}$ generated by the $T_i$ with $i\in C$, that is \[ K_C= \langle T_i\mid i\in C\rangle . \] It follows from \cite[Proposition 5.5.1]{ca} that $$\hspec(\overline{A}) = \{K_C ~|~ C \subseteq \{1,\dots,6\} \},$$ so that $$\psi (\hspec(A)) \subseteq \{K_C ~|~ C \subseteq \{1,\dots,6\} \}.$$ \section{Primitive ideals of $U_q^+(G_2)$ in the $0$-stratum} The aim of this section is to give explicit generating sets for the primitive ideals of $U_q^+(G_2)$ that belong to the $0$-stratum. They are intimately related to the centre of $U_q^+(G_2)$ and so we start this section by making explicit the centre of $U_q^+(G_2)$ and related algebras. \subsection{Centre of $U_q^+(G_2)$} \label{c2s1} Recall that $\overline{A}:=A^{(2)}={\mathbb K}_{\Lambda}[T_1,\cdots, T_6]$ is a quantum affine space. Set $\Omega_1:=T_1T_3T_5$ and $\Omega_2:=T_2T_4T_6.$ One can easily verify that $\Omega_1$ and $\Omega_2$ are central elements of $\overline{A}$ by checking they commute with all $T_i$s. We now want to successively pull $\Omega_1$ and $\Omega_2$ from the quantum affine space $\overline{A}$ into the algebra $A$ using the data of DDA of $A$ discussed above. Direct computation shows that \begin{align*} \Omega_1&:=T_1T_3T_5\\&=E_{1,4}E_{3,4}E_{5,4}+aE_{2,4}E_{5,4}\\ &=E_{1,5}E_{3,5}E_{5,5} +aE_{2,5}E_{5,5}\\ &=E_{1,6}E_{3,6}E_{5,6} +aE_{1,6}E_{4,6} +aE_{2,6} E_{5,6}+a'E_{3,6}^2\\ &=E_1E_3E_5+aE_1E_4+aE_2E_5+a'E_3^2, \end{align*} and \begin{align*} \Omega_2&:=T_2T_4T_6\\ &=E_{2,5}E_{4,5}E_{6,5}+bE_{3,5}^3E_{6,5}\\ &=E_{2,6}E_{4,6}E_{6,6}+bE_{3,6}^3E_{6,6}\\ &=E_2E_4E_6+bE_2E_5^3+bE_3^3E_6 +b'E_3^2E_5^2+c'E_3E_4E_5+d'E_4^2, \end{align*} where the parameters $a,b,a',b',c',d'$ can be found in Appendix \ref{appenc}. Note, $\Omega_1$ and $\Omega_2$ are central elements of $A^{(j)}$ for each $2\leq j\leq 7$ since Fract$(A^{(j)})=$ Fract$(\overline{A}).$ We now want to show that the centre of $A$ and other related algebras is a polynomial ring generated by $\Omega_1$ and $\Omega_2$ over ${\mathbb K}.$ The following discussions will lead us to the proof. Set $S_j:=\{\lambda T_j^{i_j}T_{j+1}^{i_{j+1}}\cdots T_6^{i_6}\mid i_j,\cdots,i_6\in \mathbb{N} \ \text{and} \ \lambda\in {\mathbb K}^* \}$ for each $2\leq j\leq 6.$ One can observe that $S_j$ is a multiplicative system of non-zero divisors of $A^{(j)}={\mathbb K}\langle E_{i,j}\mid \ \text{for all} \ i=1,\cdots, 6\rangle.$ Furthermore, the elements $T_j, \cdots, T_6$ are all normal in $A^{(j)}.$ Hence, $S_j$ is an Ore set in $A^{(j)}.$ We can therefore localize $A^{(j)}$ at $S_j$ as follows: $$R_j:=A^{(j)}S_j^{-1}.$$ Recall that $\Sigma_j:=\{T_j^n\mid n\in \mathbb{N}\}$ is an Ore set in both $A^{(j)}$ and $A^{(j+1)}$ for each $2\leq j\leq 6,$ and that $$A^{(j)}\Sigma_j^{-1}=A^{(j+1)}\Sigma_j^{-1}.$$ For all $2\leq j\leq 6,$ we have that: \begin{align} \label{emb0} R_j= A^{(j)}S_j^{-1}=(A^{(j)}\Sigma_j^{-1})S_{ 1+j}^{-1}=(A^{(j+1)}\Sigma_j^{-1})S_{ 1+j}^{-1}=(A^{(j+1)}S_{j+1}^{-1})\Sigma_{ j}^{-1}=R_{j+1}\Sigma_j^{-1}. \end{align} Note, $R_7:=A.$ Again, one can also observe that $T_1$ is normal in $R_2.$ As a result, we can form the localization $R_1:=R_2[T_1^{-1}]$. The algebra $R_1$ is the quantum torus associated to the quantum affine space $\overline{A}.$ As a result, $R_1=\mathbb{{\mathbb K}}_{q^{M}}[T_1^{\pm 1},\cdots, T_6^{\pm 1}],$ where $T_iT_j=q^{\mu_{ij}}T_jT_i$ for all $1\leq i,j\leq 6$ and $\mu_{ij}\in M.$ Similar to \cite[\S 31]{ss}, we construct the following tower of algebras: \begin{align} \label{emb} A=R_7&\subset R_6=R_7\Sigma_6^{-1}\subset R_5=R_6\Sigma_5^{-1}\subset R_4=R_5\Sigma_4^{-1}\nonumber\\ &\subset R_3=R_4\Sigma_3^{-1}\subset R_2=R_3\Sigma_2^{-1}\subset R_1. \end{align} Note, the family $(E_{1,j}^{k_1} \cdots E_{6,j}^{k_6}),$ where $k_i\in \mathbb{N}$ if $i<j$ and $k_i\in \mathbb{Z}$ otherwise is a PBW-basis of $R_j$ for all $1\leq i,j\leq 7.$ In particular, the family $( T_1^{k_1}T_2^{k_2}T_3^{k_3}T_4^{k_4}T_5^{k_5}T_6^{k_6})_{k_1,\cdots, k_6\in \mathbb{Z}}$ is a basis of $R_1.$ \begin{lemma} \label{rlr} \begin{enumerate} \item $Z(R_1)={\mathbb K}[\Omega_1^{\pm 1},\Omega_2^{\pm 1}].$ \item $Z(R_3)={\mathbb K}[\Omega_1,\Omega_2].$ \item $Z(\overline{A})={\mathbb K}[\Omega_1,\Omega_2].$ \end{enumerate} \end{lemma} \begin{proof} \begin{enumerate} \item It follows from \cite[1.3]{gz2} that $Z(R_1)$ is a commutative Laurent polynomial ring generated by certain monomials in the $T_i$s. A direct computation proves the result. \item Clearly, ${\mathbb K}[\Omega_1,\Omega_2]\subseteq Z(R_3).$ For the reverse inclusion, let $y\in Z(R_3).$ Then, $y$ can be written in terms of the basis of $R_3$ (recall, $T_i=E_{i,3}$) as: $$y=\displaystyle \sum_{(i,\cdots,n)\in \mathbb{N}^2\times \mathbb{Z}^4}a_{(i,\cdots,n)} T_1^iT_2^{j}T_3^{k}T_4^{l}T_5^{m}T_6^{n}.$$ Using the fact that $T_1, \cdots, T_6$ are all normal elements in $R_3$ and $yT_i=T_i y$ for all $i$, one easily conclude that $i=k=m$ and $j=l=n$ for all monomials appearing in $y$. Since $i,j\geq 0$, we have that $y=\sum_{(i,j)\in \mathbb{N}^2}q^{\bullet}a_{(i,j)} T_1^iT_3^{i}T_5^{i}T_2^{j}T_4^{j}T_6^{j}=\sum_{(i,j)\in \mathbb{N}^2}q^{\bullet}a_{(i,j)}\Omega_1^i\Omega_2^j.$ This implies that $y\in {\mathbb K}[\Omega_1,\Omega_2]$ as expected. \item Similar to the previous case, and so left to the reader. \end{enumerate} \end{proof} \begin{lemma} \label{c3l1} $Z(A)={\mathbb K}[\Omega_1,\Omega_2].$ \end{lemma} \begin{proof} Since $R_i$ is a localization of $R_{i+1},$ it follows that $Z(R_{i+1})\subseteq Z(R_i).$ From \eqref{emb}, we have that $Z(A)\subseteq Z(R_3).$ Observe that ${\mathbb K}[\Omega_1,\Omega_2]\subseteq Z(A)\subseteq Z(R_3)={\mathbb K}[\Omega_1,\Omega_2].$ Hence, $Z(A)={\mathbb K}[\Omega_1,\Omega_2].$ \end{proof} \begin{remark} \label{remark-centre-Ri} Since $Z(A)=Z(R_3)={\mathbb K}[\Omega_1,\Omega_2]$ and $Z(R_{i+1})\subseteq Z(R_i),$ it follows from \eqref{emb} that $Z(A)=Z(R_6)=Z(R_5)=Z(R_4)=Z(R_3)={\mathbb K}[\Omega_1,\Omega_2].$ One can also deduce from Lemma \ref{rlr} that $Z(R_2)={\mathbb K}[\Omega_1, \Omega_2^{\pm 1}].$ \end{remark} \begin{remark} The centre of the positive part of a quantised enveloping algebra of a simple Lie algebra has been described by Caldero in \cite{caldero} but we will need Remark \ref{remark-centre-Ri} later on. \end{remark} \subsection{$\Omega_1$ and $\Omega_2$ generate completely prime ideals of $U_q^+(G_2)$} The aim of this paragraph is to show that $\langle\Omega_1\rangle$ and $\langle\Omega_2\rangle$ are (completely) prime. We will make use of DDA to establish these facts. Note that we could also have used the results of \cite{goodearl2016quantum} to obtain these results. However, we will need some of the intermediate steps obtained here to compute the derivations of certain primitive quotients of $U_q^+(G_2)$ in the final section. From Section \ref{c1s1} we know that there is a bijection between $\{P\in \text{Spec}(A^{(j+1)})\mid P\cap \Sigma_j=\emptyset\}$ and $\{Q\in \text{Spec}(A^{(j)})\mid Q\cap \Sigma_j=\emptyset\}$ via $P=Q\Sigma_j^{-1}\cap A^{(j+1)}.$ Note, $\langle T_1\rangle$ and $\langle T_2\rangle$ are prime ideals of the quantum affine space $\overline{A},$ since each of the factor algebras $\overline{A}/\langle T_1\rangle$ and $\overline{A}/\langle T_2\rangle$ is isomorphic to a quantum affine space of rank 5 which is well known to be a domain. The following result and its proof show that $\langle T_1\rangle$ belongs to the image $\mathrm{Im}(\psi)$ of the canonical embedding $\psi$ and that $\langle\Omega_1\rangle$ is the completely prime ideal of $A$ such that $\psi (\langle\Omega_1\rangle)=\langle T_1\rangle.$ \begin{lemma} \label{lem1} $\langle \Omega_1\rangle\in {\rm Spec}(A)$. \end{lemma} \begin{proof} We will prove this result in several steps by showing that: \begin{itemize} \item[1.] $\langle T_1\rangle_{A^{(3)}} \in$ Spec$(A^{(3)})$. \item[2.] $\langle E_{1,4}T_3+a T_2\rangle =\langle T_1\rangle_{A^{(3)}}[T_3^{-1}]\cap A^{(4)},$ hence $Q_1:=\langle E_{1,4}T_3+a T_2\rangle \in $ Spec$(A^{(4)})$. \item[3.] $\langle E_{1,5}T_3+a E_{2,5}\rangle =Q_1[T_4^{-1}]\cap A^{(5)},$ hence $Q_2:=\langle E_{1,5}T_3+aE_{2,5} \rangle \in $ Spec$(A^{(5)})$. \item[4.] $\langle \Omega_{1}\rangle_{A^{(6)}} = Q_2[T_5^{-1}]\cap A^{(6)},$ hence $\langle \Omega_{1}\rangle_{A^{(6)}} \in $ Spec$(A^{(6)})$. \item[5.] $\langle \Omega_1\rangle_{A}=\langle \Omega_{1}\rangle_{A^{(6)}}[T_6^{-1}]\cap A,$ hence $\langle \Omega_1\rangle_{A}\in$ Spec$(A)$. \end{itemize} We now proceed to prove the above claims. 1. One can easily verify that $A^{(3)}/\langle T_1\rangle$ is isomorphic to a quantum affine space of rank 5, which is a domain, hence $\langle T_1\rangle$ is a prime ideal in $A^{(3)}.$ 2. Note, $T_1=E_{1,4}+a T_2T_3^{-1}.$ We want to show that $\langle E_{1,4}T_3+a T_2\rangle =\langle T_1\rangle_{A^{(3)}}[T_3^{-1}]\cap A^{(4)}.$ Observe that $\langle E_{1,4}T_3+a T_2\rangle\subseteq \langle T_1\rangle_{A^{(3)}}[T_3^{-1}]\cap A^{(4)}.$ We established the reverse inclusion. Let $y\in\langle T_1\rangle_{A^{(3)}}[T_3^{-1}]\cap A^{(4)}.$ Then, $y\in \langle T_1\rangle_{A^{(3)}}[T_3^{-1}].$ Therefore, there exists $i\in\mathbb{N}$ such that $y T_3^i\in \langle T_1\rangle_{A^{(3)}}.$ This implies that $y T_3^i=T_1v ,$ for some $v\in A^{(3)}.$ Since $A^{(3)}[T_3^{-1}]=A^{(4)}[T_3^{-1}],$ there exists $j\in\mathbb{N}$ such that $v T_3^j=v',$ for some $v'\in A^{(4)}.$ It follows that $y T_3^{i+j}=T_1v T_3^j=T_1v'=(E_{1,4}+aT_2T_3^{-1})v'=(E_{1,4}T_3+aT_2)T_3^{-1}v'.$ The multiplicative system generated by $T_3$ satisfies the Ore condition in $A^{(4)}$, hence, there exists $k\in \mathbb{N}$ and $v''\in A^{(4)}$ such that $T_3^{-1}v'=v''T_3^{-k}.$ One can therefore write $y T_3^{i+j}=(E_{1,4}T_3+aT_2)v''T_3^{-k}.$ This implies that $yT_3^{\delta}=\Omega_{1}'v'',$ where $\Omega_{1}':=E_{1,4}T_3+aT_2$ and $\delta=i+j+k.$ Set $S:=\{s\in\mathbb{N}\mid \exists v''\in A^{(4)}: y T_3^{s}=\Omega_{1}'v''\}.$ Note, $S\neq \emptyset,$ since $\delta\in S.$ Let $s=s_0$ be the minimum element of $S$ such that $y T_3^{s_0}=\Omega_{1}'v''.$ We want to show that $s_0=0.$ Remember, $\Omega_1'T_5=\Omega_1$ in $A^{(4)}.$ Since $\Omega_1$ is central in $A^{(4)},$ and $T_5$ is normal in $A^{(4)},$ we must have $\Omega_1'$ to be a normal element in $A^{(4)},$ otherwise, there will be a contradiction. Therefore, there exists $w\in A^{(4)}$ such that $y T_3^{s_0}=\Omega_{1}'v''=w\Omega_1'.$ Now, $A^{(4)}$ can be viewed as a free left ${\mathbb K}\langle E_{1,4},T_2,T_4,T_5,T_6\rangle-$module with basis $\left( T_3^{\xi}\right) _{\xi\in\mathbb{N}}.$ One can therefore write $y=\sum_{\xi=0}^n\alpha_\xi T_3^{\xi}$ and $w=\sum_{\xi=0}^n\beta_\xi T_3^{\xi},$ where $\alpha_\xi,\beta_\xi\in {\mathbb K}\langle E_{1,4},T_2,T_4,T_5,T_6\rangle.$ This implies that $\sum_{\xi=0}^n\alpha_\xi T_3^{\xi+s_0}=\sum_{\xi=0}^n\beta_\xi T_3^{\xi}\Omega_1'=\sum_{\xi=0}^n q^{\bullet} \beta_\xi \Omega_1'T_3^{\xi}$ (note, $T_3\Omega_1'=q^{-1}\Omega_1'T_3$). Given that $\Omega_{1}'=E_{1,4}T_3+aT_2,$ we have that $\sum_{\xi=0}^n\alpha_\xi T_3^{\xi+s_0}= \sum_{\xi=0}^nq^{\bullet} \beta_\xi E_{1,4}T_3^{1+\xi}+\sum_{\xi=0}^nq^{\bullet} a\beta_\xi T_2T_3^{\xi}.$ Suppose that $s_0>0.$ Then, identifying the constant coefficients, we have $q^{\bullet} a\beta_0T_2=0.$ As a result, $\beta_0=0,$ since $q^{\bullet} aT_2\neq 0.$ Hence, $w$ can be written as $w=\sum_{\xi=1}^n\beta_\xi T_3^{\xi}.$ Returning to $yT_3^{s_0}=w\Omega_{1}',$ we have that $y T_3^{s_0}=\sum_{\xi=1}^n\beta_\xi T_3^{\xi}\Omega_{1}'=\sum_{\xi=1}^nq^{\bullet}\beta_\xi \Omega_{1}'T_3^{\xi}=\Omega_{1}'\sum_{\xi=1}^nq^{\bullet}\beta_\xi' T_3^{\xi}.$ This implies that $y T_3^{s_0-1}=\Omega_{1}'w',$ where $w'=\sum_{\xi=1}^nq^{\bullet}\beta_\xi' T_3^{\xi-1}\in A^{(4)},$ with $\beta_{\xi}'\in {\mathbb K}\langle E_{1,4},T_2,T_4,T_5,T_6\rangle.$ Consequently, $s_0-1\in S,$ a contradiction! Therefore, $s_0=0$ and $y=\Omega_{1}'v''\in\langle\Omega_{1}'\rangle=\langle E_{1,4}T_3+aT_2\rangle.$ Hence, $\langle T_1\rangle_{A^{(3)}}[T_3^{-1}]\cap A^{(4)}\subseteq\langle E_{1,4}T_3+aT_2\rangle$ as desired. The following steps are proved in a similar manner to Step 2. They are left to the reader who might want to check details in \cite{io-thesis}. \comment{3. We want to show that $\langle E_{1,5}T_3+a E_{2,5}\rangle =\langle \Omega_1'\rangle_{A^{(4)}}[T_4^{-1}]\cap A^{(5)}.$ Note, $\Omega_1'=E_{1,4}T_3+a T_2=E_{1,5}T_3+a E_{2,5}.$ Observe that $\langle E_{1,5}T_3+a E_{2,5}\rangle\subseteq \langle \Omega_1'\rangle_{A^{(4)}}[T_4^{-1}]\cap A^{(5)}.$ We establish the reverse inclusion. Let $y\in\langle \Omega_1'\rangle_{A^{(4)}}[T_4^{-1}]\cap A^{(5)}.$ Then, $y\in \langle \Omega_1'\rangle_{A^{(4)}}[T_4^{-1}].$ Therefore, there exists $i\in\mathbb{N}$ such that $y T_4^i\in \langle \Omega_1'\rangle_{A^{(4)}}.$ This implies that $y T_4^i=\Omega_1'v,$ for some $v\in A^{(4)}.$ Since $A^{(4)}[T_4^{-1}]=A^{(5)}[T_4^{-1}],$ there exists $j\in\mathbb{N}$ such that $v T_4^j=v',$ for some $v'\in A^{(5)}.$ It follows that $y T_4^{i+j}=\Omega_1'v T_4^j=\Omega_1'v'.$ This implies that $yT_4^{\delta}=\Omega_{1}'v',$ where $\delta=i+j.$ Set $S:=\{s\in\mathbb{N}\mid \exists v'\in A^{(5)} : y T_4^{s}=\Omega_{1}'v'\}.$ Since $\delta\in S,$ we have that $S\neq \emptyset.$ Let $s=s_0$ be the minimum element of $S$ such that $y T_4^{s_0}=\Omega_{1}'v'.$ We want to show that $s_0=0.$ Note, $\Omega_1'T_5=\Omega_1$ in $A^{(5)}.$ Since $\Omega_1$ is central in $A^{(5)},$ and $T_5$ is normal in $A^{(5)},$ we must have $\Omega_1'$ as a normal element in $A^{(5)}.$ Therefore, there exists some $v''\in A^{(5)}$ such that $y T_4^{s_0}=\Omega_{1}'v'=v''\Omega_1'.$ Now, $A^{(5)}$ can be viewed as a free left ${\mathbb K}\langle E_{1,5},E_{2,5},T_3,T_5,T_6\rangle-$module with basis $\left( T_4^{\xi}\right) _{\xi\in\mathbb{N}}.$ One can write $y=\sum_{\xi=0}^n\alpha_\xi T_4^{\xi}$ and $v''=\sum_{\xi=0}^n\beta_\xi T_4^{\xi},$ where $\alpha_\xi,\beta_\xi\in {\mathbb K}\langle E_{1,5},E_{2,5},T_3,T_5,T_6\rangle.$ This implies that $\sum_{\xi=0}^n\alpha_\xi T_4^{\xi+s_0}=\sum_{\xi=0}^n \beta_\xi T_4^{\xi}\Omega_1'=\sum_{\xi=0}^nq^\bullet \beta_\xi \Omega_1'T_4^{\xi}$ (note, $T_4\Omega_1'=q^{-3}\Omega_1'T_4$). Suppose that $s_0>0.$ Then, identifying the constant coefficients, we have that $q^\bullet \beta_0 \Omega_1'=0.$ Hence, $\beta_0=0,$ since $q^{\bullet}\Omega_1'\neq 0.$ One can therefore write $v''$ as $v''=\sum_{\xi=1}^n\beta_\xi T_4^{\xi}.$ Returning to $yT_4^{s_0}=v''\Omega_{1}',$ we have that $y T_4^{s_0}=\sum_{\xi=1}^n\beta_\xi T_4^{\xi}\Omega_{1}'=\sum_{\xi=1}^nq^{\bullet}\beta_\xi \Omega_{1}' T_4^{\xi}=\Omega_{1}'\sum_{\xi=1}^nq^{\bullet}\beta_\xi' T_4^{\xi},$ where $\beta_\xi'\in {\mathbb K}\langle E_{1,5},E_{2,5},T_3,T_5,T_6\rangle.$ This implies that $y T_4^{s_0-1}=\Omega_{1}'w,$ where $w=\sum_{\xi=1}^nq^{\bullet}\beta_\xi' T_4^{\xi-1}\in A^{(5)}.$ Consequently, $s_0-1\in S,$ a contradiction! Therefore, $s_0=0$ and $y=\Omega_{1}'v'\in\langle\Omega_{1}'\rangle=\langle E_{1,5}T_3+aE_{2,5}\rangle.$ As a result, $\langle \Omega_1'\rangle_{A^{(4)}}[T_4^{-1}]\cap A^{(5)}\subseteq\langle E_{1,5}T_3+aE_{2,5}\rangle$ as desired. 4. Observe that $\Omega_1'=E_{1,5}T_3+aE_{2,5}=\Omega_1T_5^{-1}$ in $A^{(6)}[T_5^{-1}].$ We want to show that $\langle\Omega_1'\rangle_{A^{(5)}}[T_5^{-1}]\cap A^{(6)}=\langle\Omega_1\rangle_{A^{(6)}}.$ Obviously, $\langle\Omega_1\rangle_{A^{(6)}}\subseteq \langle\Omega_1'\rangle_{A^{(5)}}[T_5^{-1}]\cap A^{(6)}.$ We establish the reverse inclusion. Let $y\in \langle\Omega_1'\rangle_{A^{(5)}}[T_5^{-1}]\cap A^{(6)}.$ This implies that $y\in \langle\Omega_1'\rangle_{A^{(5)}}[T_5^{-1}].$ There exists $i\in\mathbb{N}$ such that $yT_5^i\in \langle\Omega_1\rangle_{A^{(5)}}.$ Hence, $yT_5^i=\Omega_1'v,$ for some $v\in A^{(5)}.$ Furthermore, since $A^{(5)}[T_5^{-1}]=A^{(6)}[T_5^{-1}],$ there exists $j\in\mathbb{N}$ such that $vT_5^j=v',$ for some $v'\in A^{(6)}.$ It follows from $yT_5^i=\Omega_1'v$ that $yT_5^{i+j}=\Omega_1'vT_5^j=\Omega_1'v'=\Omega_1T_5^{-1}v'$ (note, $\Omega_1'T_5=\Omega_1$ in $A^{(6)}$). The multiplicative system generated by $T_5$ satisfies the Ore condition in $A^{(6)}$, hence, there exists $k\in \mathbb{N}$ and $v''\in A^{(6)}$ such that $T_5^{-1}v'=v''T_5^{-k}.$ One can therefore write $yT_5^{i+j} =\Omega_1v''T_5^{k}.$ Hence, $yT_5^\delta=\Omega_1v'',$ where $\delta=i+j+k.$ Set $S:=\{s\in\mathbb{N}\mid \exists v''\in A^{(6)} : y T_5^{s}=\Omega_{1}v''\}.$ Since $\delta\in S,$ we have that $S\neq \emptyset.$ Let $s=s_0$ be the minimum element of $S$ such that $ y T_5^{s_0}=\Omega_{1}v''.$ We want to show that $s_0=0.$ Now, $A^{(6)}$ can be viewed as a free ${\mathbb K}\langle E_{1,6},E_{2,6},E_{3,6},T_4,T_6\rangle-$module with basis $\left( T_5^{\xi}\right) _{\xi\in\mathbb{N}}.$ One can write $y=\sum_{\xi=0}^n\alpha_\xi T_5^{\xi}$ and $v''=\sum_{\xi=0}^n\beta_\xi T_5^{\xi},$ where $\alpha_\xi,\beta_\xi\in {\mathbb K}\langle E_{1,6},E_{2,6},E_{3,6},T_4,T_6\rangle.$ With this, $y T_5^{s_0}=\Omega_{1}v''$ implies that $\sum_{\xi=0}^n\alpha_\xi T_5^{\xi+s_0}= \sum_{\xi=0}^n\beta_\xi\Omega_1T_5^{\xi}.$ Write $\Omega_1=\gamma_1T_5+\gamma_2,$ where $\gamma_1=E_{1,6}E_{3,6}+aE_{2,6}$ and $\gamma_2=a'E_{3,6}^2+aE_{1,6}E_{4,6}.$ It follows that $\sum_{\xi=0}^n\alpha_\xi T_5^{\xi+s_0}= \sum_{\xi=0}^n \beta_\xi \gamma_1T_5^{\xi+1}+\sum_{\xi=0}^n \beta_\xi\gamma_2 T_5^{\xi}.$ Suppose that $s_0>0.$ Then, identifying the constant coefficients, we have that $\beta_0 \gamma_2=0.$ Hence, $\beta_0=0,$ since $\gamma_2\neq 0.$ One can therefore write $v''=\sum_{\xi=1}^n\beta_\xi T_5^{\xi}.$ Returning to $yT_5^{s_0}=\Omega_{1}v'',$ we have that $y T_5^{s_0}=\Omega_{1}\sum_{\xi=1}^n\beta_\xi T_5^{\xi}.$ This implies that $y T_5^{s_0-1}=\Omega_{1}w,$ where $w=\sum_{\xi=1}^n\beta_\xi T_5^{\xi-1}\in A^{(6)}.$ As a result, $s_0-1\in S,$ a contradiction! Therefore, $s_0=0$ and $y=\Omega_{1}v''\in\langle\Omega_{1}\rangle_{A^{(6)}}.$ Consequently, $\langle\Omega_1'\rangle_{A^{(5)}}[T_5^{-1}]\cap A^{(6)}\subseteq \langle\Omega_1\rangle_{A^{(6)}}$ as desired. 5. We want to show that $\langle\Omega_1\rangle_{A^{(6)}}[T_6^{-1}]\cap A=\langle\Omega_1\rangle_{A}.$\newline Obviously, $\langle\Omega_1\rangle_{A}\subseteq \langle\Omega_1\rangle_{A^{(6)}}[T_6^{-1}]\cap A.$ We establish the reverse inclusion. Let $y\in \langle\Omega_1\rangle_{A^{(6)}}[T_6^{-1}]\cap A.$ This implies that $y\in \langle\Omega_1\rangle_{A^{(6)}}[T_6^{-1}].$ There exists $i\in\mathbb{N}$ such that $yT_6^i\in \langle\Omega_1\rangle_{A^{(6)}}.$ Hence, $yT_6^i=\Omega_1v,$ for some $v\in A^{(6)}.$ Furthermore, since $A^{(6)}[T_6^{-1}]=A[T_6^{-1}],$ there exists $j\in\mathbb{N}$ such that $vT_6^j=v',$ for some $v'\in A.$ It follows from $yT_6^i=\Omega_1v$ that $yT_6^{i+j}=\Omega_1vT_6^j=\Omega_1v'.$ Hence, $yT_6^\delta=\Omega_1v',$ where $\delta=i+j.$ Set $S:=\{s\in\mathbb{N}\mid \exists v'\in A : y T_6^{s}=\Omega_{1}v'\}.$ Note, $\delta\in S,$ hence $S$ is non-empty. Let $s=s_0$ be the minimum element of $S$ such that $ y T_6^{s_0}=\Omega_{1}v'.$ We want to show that $s_0=0.$ Now, $A$ can be viewed as a free ${\mathbb K}\langle E_{1},E_{2},E_{3},E_{4},T_5\rangle-$module with basis $\left(T_6^\xi\right) _{\xi\in\mathbb{N}}.$ One can write $y=\sum_{\xi=0}^n\alpha_\xi T_6^{\xi}$ and $v'=\sum_{\xi=0}^n\beta_\xi T_6^{\xi},$ where $\alpha_\xi,\beta_\xi\in {\mathbb K}\langle E_{1},E_{2},E_{3},E_4,T_5\rangle.$ With this, $y T_6^{s_0}=\Omega_{1}v'$ implies that $\sum_{\xi=0}^n\alpha_\xi T_6^{\xi+s_0}= \sum_{\xi=0}^n\beta_\xi\Omega_1T_6^{\xi}.$ Suppose that $s_0>0.$ Then, identifying constant coefficients, we have that $\beta_0 \Omega_1=0.$ As a result, $\beta_0=0,$ since $\Omega_1\neq 0.$ One can therefore write $v'=\sum_{\xi=1}^n\beta_\xi T_6^{\xi}.$ Returning to $yT_6^{s_0}=\Omega_{1}v',$ we have that $y T_6^{s_0}=\Omega_{1}\sum_{\xi=1}^n\beta_\xi T_6^{\xi}.$ This implies that $y T_6^{s_0-1}=\Omega_{1}v^{\prime\prime},$ where $v^{\prime\prime}=\sum_{\xi=1}^n\beta_\xi T_6^{\xi-1}\in A.$ Hence, $s_0-1\in S,$ a contradiction! Therefore, $s_0=0$ and $y=\Omega_{1}v'\in\langle\Omega_{1}\rangle_{A}.$ Consequently, $\langle\Omega_1\rangle_{A^{(6)}}[T_6^{-1}]\cap A\subseteq \langle\Omega_1\rangle_{A}$ as desired.} \end{proof} Using similar techniques, one can prove that $\langle T_2\rangle\in$ Im$(\psi)$ and that $\langle\Omega_2\rangle$ is the completely prime ideal of $A$ such that $\psi (\langle\Omega_2\rangle)=\langle T_2\rangle$. Again, we refer the interested reader to \cite{io-thesis} for details. We record these facts in the following lemma. \begin{lemma} \label{lem2} $\langle T_2\rangle\in \mathrm{Im}(\psi)$ and $\langle\Omega_2\rangle$ is the completely prime ideal of $A$ such that $\psi (\langle\Omega_2\rangle)=\langle T_2\rangle.$ \end{lemma} Since ${\mathcal H}$-${\rm Spec} (U_q^+(G_2))$ is homeomorphic to the Weyl group $W$ of type $G_2$ by \cite{my2}, there are only 2 ${\mathcal H}$-primes in $U_q^+(G_2)$ of height 1. Since $\Omega_1$ and $\Omega_2$ are central, the prime ideals that they generate have height less than or equal to 1, and so equal to 1. As an immediate consequence, we get the following result. \begin{lemma} \label{rem1} \begin{enumerate} \item $\langle\Omega_1\rangle$ and $\langle\Omega_2\rangle$ are the only height one ${\mathcal H}$-invariant prime ideals of $A$. \item Every non-zero ${\mathcal H}$-invariant prime ideal of $A$ contains either $\langle\Omega_1\rangle$ or $\langle\Omega_2\rangle$. \end{enumerate} \end{lemma} \subsection{Description of the $0$-stratum and beyond} In this section, we will often assume that our base field ${\mathbb K}$ is algebraically closed. This assumption is actually not necessary for the main result of this section, Theorem \ref{c3p29}, but makes the description of the $0$-stratum easier to present. This section focuses on finding the height two maximal ideals of $A=U_q^+(G_2)$. Note first that such ideals can only belong to the ${\mathcal H}$-stratum of an ${\mathcal H}$-prime of height less than or equal to 1 (since ${\mathcal H}$-${\rm Spec}(A)$ is isomorphic to $W$). It follows from the previous sections that we need to compute the ${\mathcal H}$-strata of 3 ${\mathcal H}$-primes: $0$, $\langle\Omega_1\rangle$ and $\langle\Omega_2\rangle$. We start with the $0$-stratum. The strategy is similar to \cite[Propositions 2.3 and 2.4]{sl}. Note, in this subsection, all ideals in $A$ will simply be written as $\langle \Theta\rangle,$ where $\Theta\in A.$ However, if we want to refer to an ideal in any other algebra, say $R$, then that ideal will be written as $\langle \Theta\rangle_R,$ where in this case, $\Theta\in R.$ \begin{proposition} \label{p0} Assume ${\mathbb K}$ is algebraically closed. Let $\mathcal{P}$ be the set of those unitary irreducible polynomials $P(\Omega_1,\Omega_2)\in {\mathbb K}[\Omega_1,\Omega_2]$ with $P(\Omega_1,\Omega_2)\neq \Omega_1$ and $P(\Omega_1,\Omega_2)\neq \Omega_2$. Then, ${\rm Spec}_{\langle 0\rangle}(A)= \{\langle 0\rangle\}\cup \{\langle P(\Omega_1,\Omega_2)\rangle\mid P(\Omega_1,\Omega_2)\in \mathcal{P}\}\cup \{\langle \Omega_1-\alpha,\Omega_2-\beta\rangle\mid\alpha,\beta\in {\mathbb K}^*\}.$ \end{proposition} \begin{proof} We claim that Spec$_{\langle 0\rangle}(A)=\{Q \in \text{Spec}(A)\mid\Omega_1,\Omega_2\not\in Q\}.$ To establish this claim, let us assume that this is not the case. That is, suppose there exists $ Q\in$ Spec$_{\langle 0\rangle}(A)$ such that $\Omega_1,\Omega_2\in Q;$ then the product $\Omega_1\Omega_2$ which is an $\mathcal{H}-$eigenvector belongs to $Q.$ Consequently, $\Omega_1\Omega_2\in \bigcap_{h\in \mathcal{H}}h\cdot Q=\langle 0\rangle,$ a contradiction. Hence, we have shown that Spec$_{\langle 0\rangle}(A)\subseteq \{Q \in \text{Spec}(A)\mid\Omega_1,\Omega_2\not\in Q\}.$ Conversely, suppose that $Q\in$ Spec$(A)$ such that $\Omega_1,\Omega_2\not\in Q,$ then $\bigcap_{h\in \mathcal{H}}h\cdot Q$ is an $\mathcal{H}-$invariant prime ideal of $A,$ which contains neither $\Omega_1$ nor $\Omega_2.$ Obviously, the only possibility for $\bigcap_{h\in \mathcal{H}}h\cdot Q$ is $\langle0\rangle$ since every non-zero $\mathcal{H}-$invariant prime ideal contains at least $\Omega_1$ or $\Omega_2$. Thus, $\bigcap_{h\in \mathcal{H}}h\cdot Q=\langle 0\rangle.$ Hence, $Q\in$ Spec$_{\langle 0\rangle}(A).$ Therefore, $\{Q \in \text{Spec}(A)\mid\Omega_1,\Omega_2\not\in Q\}\subseteq$ Spec$_{\langle 0\rangle}(A).$ This confirms our claim. Since $\Omega_1,\Omega_2\in Z(A),$ we have that the set $\{\Omega_1^i\Omega_2^j\mid i,j\in\mathbb{N}\}$ is a right denominator set in the noetherian domain $A.$ One can now localize $A$ as $R:=A[\Omega_1^{-1},\Omega_2^{-1}].$ Let $Q\in$ Spec$_{\langle 0\rangle}(A),$ the map $\phi:Q\longrightarrow Q[\Omega_1^{-1},\Omega_2^{-1}]$ is an increasing bijection from Spec$_{\langle 0\rangle}(A)$ onto Spec$(R).$ Since $\Omega_1$ and $\Omega_2$ are $\mathcal{H}-$eigenvectors, and $\mathcal{H}$ acts on $A$, we have that $\mathcal{H}$ also acts on $R.$ Since every ${\mathcal H}$-prime ideal of $A$ contains $\Omega_1$ or $\Omega_2$, one can easily check that $R$ is $\mathcal{H}-$simple (in the sense that the only ${\mathcal H}$-invariant ideal of $R$ is the $0$ ideal). We proceed to describe Spec$(R)$ and Spec$_{\langle 0\rangle}(A).$ We deduce from \cite[Exercise II.3.A]{bg} that the action of $\mathcal{H}$ on $R$ is rational. This rational action coupled with $R$ being $\mathcal{H}-$simple implies that the extension and contraction maps provide a mutually inverse bijection between Spec$(R)$ and Spec$(Z(R))$ \cite[Corollary II.3.9]{bg}. From Lemma \ref{c3l1}, $Z(A)={\mathbb K}[\Omega_1,\Omega_2],$ and so $Z(R)={\mathbb K}[\Omega_1^{\pm 1},\Omega_2^{\pm 1}].$ Since ${\mathbb K}$ is algebraically closed, we have that Spec$(Z(R))=\{\langle 0\rangle_{Z(R)}\}\cup \{\langle P(\Omega_1,\Omega_2)\rangle_{Z(R)}\mid P(\Omega_1,\Omega_2)\in \mathcal{P}\}\cup \{\langle \Omega_1-\alpha,\Omega_2-\beta\rangle_{Z(R)}\mid\alpha,\beta\in {\mathbb K}^*\}.$ Since there is an inverse bijection between Spec$(R)$ and Spec$(Z(R)),$ and also $R$ is $\mathcal{H}-$simple, one can recover Spec$(R)$ from Spec$(Z(R))$ as follows: Spec$(R)=\{\langle 0\rangle_{R}\}\cup \{\langle P(\Omega_1,\Omega_2)\rangle_{R}\mid P(\Omega_1,\Omega_2)\in \mathcal{P}\}\cup \{\langle \Omega_1-\alpha,\Omega_2-\beta\rangle_R\mid\alpha,\beta\in {\mathbb K}^*\}.$ It follows that Spec$_{\langle 0\rangle}(A)=\{\langle 0\rangle_{R}\cap A\}\cup \{\langle P(\Omega_1,\Omega_2)\rangle_{R}\cap A\mid P(\Omega_1,\Omega_2)\in \mathcal{P}\}\cup \{\langle \Omega_1-\alpha,\Omega_2-\beta\rangle_R\cap A\mid \alpha,\beta\in {\mathbb K}^*\}.$ Undoubtedly, $\langle 0\rangle_{R}\cap A=\langle 0 \rangle.$ We now have to show that $\langle P(\Omega_1,\Omega_2)\rangle_{R}\cap A=\langle P(\Omega_1,\Omega_2)\rangle,$ $\forall P(\Omega_1,\Omega_2)\in \mathcal{P},$ and $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle_R\cap A=\langle \Omega_1-\alpha,\Omega_2-\beta\rangle,$ $\forall \alpha,\beta\in {\mathbb K}^*$ to complete the proof. Fix $P(\Omega_1,\Omega_2)\in \mathcal{P}.$ Observe that $\langle P(\Omega_1,\Omega_2) \rangle \subseteq \langle P(\Omega_1,\Omega_2)\rangle_{R}\cap A.$ To show the reverse inclusion, let $y\in \langle P(\Omega_1,\Omega_2)\rangle_{R}\cap A$. This implies that $y=dP(\Omega_1,\Omega_2), $ where $d\in R,$ since $y\in \langle P(\Omega_1,\Omega_2) \rangle_R.$ Also, $d\in R$ implies that there exist $ i,j\in \mathbb{N}$ such that $d=a\Omega_1^{-i}\Omega_2^{-j},$ where $a\in A.$ Therefore, $y=a\Omega_1^{-i}\Omega_2^{-j}P(\Omega_1,\Omega_2), $ which implies that $y\Omega_1^i\Omega_2^j=aP(\Omega_1,\Omega_2).$ Choose $(i,j)\in \mathbb{N}^2$ minimal (in the lexicographic order on $\mathbb{N}^2$) such that the equality holds. Without loss of generality, suppose that $i>0,$ then $aP(\Omega_1,\Omega_2)\in \langle\Omega_1\rangle.$ Given that $\langle\Omega_1\rangle$ is a completely prime ideal, it implies that $a\in \langle\Omega_1\rangle$ or $P(\Omega_1,\Omega_2)\in \langle\Omega_1\rangle.$ Since $P(\Omega_1,\Omega_2)\in \mathcal{P},$ it implies that $P(\Omega_1,\Omega_2)\not\in \langle\Omega_1\rangle,$ hence $a\in \langle \Omega_1\rangle.$ This further implies that $a=t\Omega_1,$ where $ t\in A.$ Returning to $y\Omega_1^i\Omega_2^j=aP(\Omega_1,\Omega_2),$ we have that $y\Omega_1^i\Omega_2^j=t\Omega_1 P(\Omega_1,\Omega_2).$ Therefore, $y\Omega_1^{i-1}\Omega_2^j=tP(\Omega_1,\Omega_2).$ This clearly contradicts the minimality of $(i,j),$ hence $(i,j)=(0,0),$ and $y=aP(\Omega_1,\Omega_2)\in\langle P(\Omega_1,\Omega_2)\rangle.$ Consequently, $\langle P(\Omega_1,\Omega_2)\rangle_{R}\cap A=\langle P(\Omega_1,\Omega_2)\rangle$ for all $P(\Omega_1,\Omega_2)\in \mathcal{P}$ as desired. Similarly, we show that $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle_R\cap A=\langle \Omega_1-\alpha,\Omega_2-\beta\rangle;$ $\forall \alpha,\beta\in {\mathbb K}^*.$ Fix $ \alpha,\beta\in {\mathbb K}^*.$ Observe that $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle\subseteq\langle \Omega_1-\alpha,\Omega_2-\beta\rangle_R\cap A.$ We establish the reverse inclusion. Let $y\in \langle \Omega_1-\alpha,\Omega_2-\beta\rangle_R \cap A.$ Since $y\in \langle \Omega_1-\alpha,\Omega_2-\beta\rangle_R,$ there exist $i,j\in\mathbb{N}$ such that $ y\Omega_1^{i}\Omega_2^{j}=m(\Omega_1-\alpha)+n(\Omega_2-\beta),$ where $m,n\in A.$ Choose $(i,j)\in \mathbb{N}^2$ minimal (in the lexicographic order on $\mathbb{N}^2$) such that the equality holds. Without loss of generality, suppose that $i>0$ and let $f:A\longrightarrow {A}/{\langle \Omega_2-\beta\rangle}$ be a canonical surjection. We have that $f(y)f(\Omega_1)^if(\Omega_2)^j=f(m)f(\Omega_1-\alpha).$ It follows that $f(m)f(\Omega_1-\alpha)\in \langle f(\Omega_1)\rangle$ and so $\alpha f(m) \in \langle f(\Omega_1)\rangle$. Observe that $f(\Omega_1-\alpha)\not\in \langle f(\Omega_1)\rangle,$ hence $f(m)\in \langle f(\Omega_1)\rangle.$ As $\alpha \neq 0$, we obtain the existence of $\lambda\in A$ such that $f(m)=f(\lambda)f(\Omega_1).$ Consequently, $f(y)f(\Omega_1)^if(\Omega_2)^j=f(\lambda)f(\Omega_1)f(\Omega_1-\alpha).$ Since $f(\Omega_1)\neq 0,$ it implies that $f(y)f(\Omega_1)^{i-1}f(\Omega_2)^j=f(\lambda)f(\Omega_1-\alpha).$ Therefore, $y\Omega_1^{i-1}\Omega_2^j=\lambda(\Omega_1-\alpha)+\lambda^\prime(\Omega_2-\beta)$ for some $\lambda^\prime\in A.$ This contradicts the minimality of $(i,j).$ Hence, $(i,j)=(0,0)$ and so $y=m(\Omega_1-\alpha)+n(\Omega_2-\beta)\in \langle \Omega_1-\alpha,\Omega_2-\beta\rangle.$ In conclusion, $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle_R\cap A=\langle \Omega_1-\alpha,\Omega_2-\beta\rangle,$ $\forall \alpha,\beta\in {\mathbb K}^*.$ \end{proof} Using similar techniques, we obtain the following description for the ${\mathcal H}$-strata of $\langle\Omega_1\rangle$ and $\langle\Omega_2\rangle$. \begin{proposition} \label{p1} Assume ${\mathbb K}$ is algebraically closed. \begin{enumerate} \item ${\rm Spec}_{\langle\Omega_1\rangle}(A)= \{\langle\Omega_1\rangle\}\cup \{\langle\Omega_1,\Omega_2-\beta\rangle\mid \beta\in {\mathbb K}^*\}.$ \item ${\rm Spec}_{\langle\Omega_2\rangle}(A)= \{\langle\Omega_2\rangle\}\cup \{\langle\Omega_1-\alpha,\Omega_2\rangle\mid\alpha\in {\mathbb K}^*\}.$ \end{enumerate} \end{proposition} Since maximal ideals in their stratum are primitive for a QNA, we obtain the following result. \begin{corollary} Assume ${\mathbb K}$ is algebraically closed and let $(\alpha, \beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ The ideal $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle$ of $A$ is primitive. \end{corollary} \begin{remark} The statement of the above corollary is still valid without the assumption that ${\mathbb K}$ is algebraically closed. The proof is actually similar as we only use this assumption to get a full description of the strata we were interested in. \end{remark} We can actually prove a stronger result. \begin{theorem} \label{c3p29} Let $(\alpha, \beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ The prime ideal $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle$ of $A$ is maximal. \end{theorem} \begin{proof} Let $(\alpha, \beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ Suppose that there exists a maximal ideal $I$ of $A$ such that $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle\varsubsetneq I\varsubsetneq A.$ Let $J$ be the $\mathcal{H}-$invariant prime ideal in $A$ such that $I\in {\rm Spec}_J(A).$ We claim that $J$ cannot be $\langle 0\rangle,$ $\langle\Omega_1\rangle$ or $\langle\Omega_2\rangle$. For instance, if $\alpha, \beta \neq 0$, then $J$ cannot be equal to $\langle 0\rangle$ since in this case $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle $ is maximal in the $0$-stratum. Moreover, $J \neq \langle\Omega_1\rangle$ as otherwise $I$ would contain $\alpha=\Omega_1-(\Omega_1-\alpha)$, a contradiction. The other cases are similar and left to the reader. This means that $J$ is an ${\mathcal H}$-prime of height at least equal to 2. As the poset of ${\mathcal H}$-primes is isomorphic to $W$, this forces $J$ to contain both $\Omega_1$ and $\Omega_2.$ Moreover, since $ J\subseteq I,$ it implies that $\Omega_1, \Omega_2\in I.$ Given that $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle\subset I$, we have that $\Omega_1-\alpha,\Omega_2-\beta\in I.$ It follows that $\alpha, \beta\in I,$ hence $I=A,$ a contradiction! This confirms that $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle$ is a maximal ideal in $A.$ \end{proof} \section{Simple quotients of $U_q^+(G_2)$ and their relation to the second Weyl algebra} \label{c3s1} Now that we have found primitive ideals of $A=U_q^+(G_2),$ we are going to study the corresponding simple quotient algebras. In view of Dixmier's theorem, we consider these simple quotients as deformations of a Weyl algebra (of appropriate dimension), and so we compare their properties with some known properties of the Weyl algebras. In this section, we prove that the Gelfand-Kirillov dimension of $A_{\alpha,\beta}$ is 4 and consequently prove that the height of the maximal ideal $\langle \Omega_1-\alpha,\Omega_2-\beta\rangle$ is $2$ as expected. Then we focus on describing a linear basis of $A_{\alpha,\beta}$; we use this basis in the following section to study the derivations of $A_{\alpha,\beta}$. Finally, we show that with appropriate choices of $\alpha$ and $\beta,$ the algebra $A_{\alpha,\beta}$ is a quadratic extension of the second Weyl algebra $A_2({\mathbb K})$ at $q=1.$ Recall from Theorem \ref{c3p29} that $\Omega_1-\alpha$ and $\Omega_2-\beta,$ where $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\},$ generate a maximal ideal of $A.$ As a result, the corresponding quotient $$A_{\alpha,\beta}:=\frac{A}{\langle \Omega_1-\alpha,\Omega_2-\beta\rangle}$$ is a simple noetherian domain. Denote the canonical images of $E_i$ in $A_{\alpha,\beta}$ by $e_i:=E_i+\langle\Omega_1-\alpha, \Omega_2-\beta\rangle$ for all $1\leq i\leq 6$. The algebra $A_{\alpha,\beta}$ satisfies the following relations: \begin{align*} e_2e_1&=q^{-3}e_1e_2& e_3e_1&=q^{-1}e_1e_3-(q+q^{-1}+q^{-3})e_2\\ e_3e_2&=q^{-3}e_2e_3&e_4e_1&=e_1e_4+(1-q^2)e_3^2\\ e_4e_2&=q^{-3}e_2e_4-\frac{q^4-2q^2+1}{q^4+q^2+1}e_3^3&e_4e_3&=q^{-3}e_3e_4\\ e_5e_1&=qe_1e_5-(1+q^2)e_3&e_5e_2&=e_2e_5+(1-q^2)e_3^2\\ e_5e_3&=q^{-1}e_3e_5-(q+q^{-1}+q^{-3})e_4&e_5e_4&=q^{-3}e_4e_5\\ e_6e_1&=q^3e_1e_6-q^3e_5&e_6e_2&=q^3e_2e_6+(q^4+q^2-1)e_4+(q^2-q^4)e_3e_5\\ e_6e_3&=e_3e_6+(1-q^2)e_5^2&e_6e_4&=q^{-3}e_4e_6-\frac{q^4-2q^2+1}{q^4+q^2+1}e_5^3\\ e_6e_5&=q^{-3}e_5e_6, \end{align*} and \begin{align} e_1e_3e_5+ae_1e_4+ae_2e_5+a'e_3^2&=\alpha,\label{e1e}\\ e_2e_4e_6+be_2e_5^3+be_3^3e_6+b'e_3^2e_5^2+c'e_3e_4e_5+d'e_4^2&=\beta.\label{e2e} \end{align} The following additional relations of $A_{\alpha,\beta}$ in the lemma below will be helpful when computing linear basis for $A_{\alpha,\beta}$. Note, we put constant coefficients of monomials in a square bracket $[ \ ]$ in order to distinguish them from monomials easily. These constants are defined in Appendix \ref{appenc}. \begin{lemma} \label{c3l2} \begin{align*} (1)\ e_3^2=&c_1\alpha+[c_2]e_2e_5+[c_2]e_1e_4+ [c_3]e_1e_3e_5.\\ \\ (2)\ e_4^2=&b_1\beta+[b_2]e_2e_4e_6 +[b_3]e_2e_5^3+[b_4\alpha]e_3e_6+[b_5]e_2e_3 e_5e_6+[b_6]e_1e_3e_4e_6 \\& +[b_7\alpha]e_1e_5e_6+[b_8]e_1e_2e_5^2 e_6+[b_9]e_1^2e_4e_5e_6+[b_{10}] e_1^2e_3e_5^2e_6+[b_{11}\alpha]e_5^2\\&+ [b_{12}]e_1e_3e_5^3+ [b_{13}]e_3e_4e_5+[b_{14}]e_1e_4e_5 ^2.\\ \\ (3)\ e_3^2e_4=& [c_1\alpha] e_4+[q^{-3}c_2]e_2e_4e_5 +[c_2b_4\alpha]e_1e_3e_6 +[b_{15}]e_1e_3e_4e_5 +[\beta b_1c_2]e_1\\&+[c_2b_3]e_1e_2e_5^3+[c_2b_5]e_1e_2e_3 e_5e_6+[c_2b_6]e_1^2e_3e_4e_6+ [c_2b_7\alpha]e_1^2e_5e_6\\&+[c_2b_{11}\alpha]e_1e_5^2+[c_2b_{12}]e_1^2e_3 e_5^3+[c_2b_8]e_1^2e_2e_5 ^2e_6+[c_2b_9]e_1^3e_4e_5e_6\\ &+[c_2b_{10}]e_1^3 e_3e_5^2e_6+[c_2b_{14}]e_1^2 e_4e_5^2 +[c_2b_2]e_1e_2e_4e_6.\\ \\ (4) \ e_3e_4^2=& [\beta b_1]e_3+[k_1]e_2e_3 e_4e_6+ [k_2]e_2e_3e_5^3 +[k_3\alpha^2] e_6+[k_4\alpha] e_2e_5e_6+[k_5\alpha]e_1e_4 e_6 \\&+[k_6\alpha]e_1e_3e_5 e_6+[k_7]e_2 ^2e_5^2e_6+[k_8 \beta] e_1e_5 +[k_9] e_1e_2e_4e_5e_6+ [k_{10}]e_1^3e_2e_5^2e_6^2 \\& + [k_{11}\alpha] e_4 e_5+[k_{12}\alpha]e_1^3e_5e_6^2+[k_{13}] e_1^2e_3 e_4e_5e_6+[k_{14}\beta]e_1^2e_6 +[b_{11}\alpha]e_3e_5^2 \\&+[k_{15}]e_1^2 e_2e_4e_6^2+[k_{16}]e_1^2e_2 e_5^3e_6+[k_{17}\alpha]e_1^2e_3e_6^2+ [k_{18}]e_1^2e_2e_3e_5e_6^2 +[k_{19}]e_1^3e_3e_4e_6^2 \\&+[k_{20}]e_1^4e_4e_5e_6^2+ [k_{21}\alpha]e_1^2 e_5^2e_6+[k_{22}]e_1^4e_3 e_5^2e_6^2+[k_{23}] e_1^3e_3e_5^3e_6 +[k_{24}]e_1^3 e_4e_5^2e_6 \\& +[k_{25}]e_1e_5^3+[k_{26}]e_1^2e_4 e_5^3+[k_{27}]e_1^2e_3e_5^4+[k_{28}]e_2e_4e_5^2+[k_{29}]e_1e_3e_4 e_5^2 \\&+[k_{30}]e_1e_2 e_5^4+[k_{31}] e_1e_2e_3e_5^2e_6. \end{align*} \end{lemma} \begin{proof} This is proved by brute-force computation, left to the reader. \end{proof} \subsection{Gelfand-Kirillov dimension of $A_{\alpha,\beta}$} \label{c3p3} We refer the reader to \cite{gt} for background on the Gelfand-Kirillov dimension (GKdim for short). Assume first that $\alpha,\beta \neq 0.$ Recall from Section \ref{c2s1} that $R_1={\mathbb K}_{q^M}[T_1^{\pm1},\cdots,T_6^{\pm1}]$ is the quantum torus associated to the quantum affine space $\overline{A}=A^{(2)}$. Also, $\Omega_1=T_1T_3T_5$ and $\Omega_2=T_2T_4T_6$ in $\overline{A}.$ It follows from \cite[Theorem 5.4.1]{ca} that there exists an Ore set $S_{\alpha,\beta}$ in $A_{\alpha,\beta}$ such that $A_{\alpha,\beta}S^{-1}_{\alpha,\beta}\cong {R_1}/{\langle T_1T_3T_5-\alpha,T_2T_4T_6-\beta\rangle}.$ Now, set $$\mathscr{A}_{\alpha,\beta}:=\frac{R_1}{\langle T_1T_3T_5-\alpha,T_2T_4T_6-\beta\rangle}.$$ Let $t_i:=T_i+\langle T_1T_3T_5-\alpha, T_2T_4T_6-\beta\rangle$ denote the canonical images of the generators $T_i$ of $R_1$ in $\mathscr{A}_{\alpha,\beta}.$ The algebra $\mathscr{A}_{\alpha,\beta}$ is generated by $t_1^{\pm 1}, \cdots, t_6^{\pm 1} $ subject to the following relations: \begin{align*} t_it_j&=q^{\mu_{ij}}t_jt_i&t_1&=\alpha t_5^{-1}t_3^{-1}& t_2&=\beta t_6^{-1}t_4^{-1}, \end{align*} for all $1\leq i,j\leq 6$ and $ M=(\mu_{ij})$ (the skew-symmetric matrix in Section \ref{c2s1}). Observe that $\mathscr{A}_{\alpha,\beta}\cong {\mathbb K}_{q^N}[t_3^{\pm1},t_4^{\pm1},t_5^{\pm1},t_6^{\pm 1}],$ where the skew-symmetric matrix $N$ can easily be deduced from $M$ (by deleting the first two rows and columns) as follows: \begin{align*} N:=\begin{bmatrix} 0&3&1&0\\ -3&0&3&3\\ -1&-3&0&3\\ 0&-3&-3&0 \end{bmatrix}. \end{align*} Secondly, suppose that $\alpha=0$ and $\beta\neq 0.$\newline Then, $A_{0,\beta}S^{-1}_{0,\beta}\cong \mathscr{A}_{0,\beta}= {R_1}/{\langle T_1,T_2T_4T_6-\beta\rangle}.$ The algebra $\mathscr{A}_{0,\beta}$ is generated by $t_2^{\pm 1}, \cdots, t_6^{\pm 1}$ subject to the relations \begin{align*} t_it_j&=q^{\mu_{ij}}t_jt_i&t_1&=0& t_2&=\beta t_6^{-1}t_4^{-1}, \end{align*} for all $1\leq i,j\leq 6$ and $ \mu_{ij}\in M.$ We also have that $\mathscr{A}_{0,\beta}\cong {\mathbb K}_{q^N}[t_3^{\pm1},t_4^{\pm1},t_5^{\pm1},t_6^{\pm 1}].$ Finally, when $\alpha\neq 0$ and $\beta=0,$ then one can also verify that $\mathscr{A}_{\alpha,0}\cong {\mathbb K}_{q^N}[t_3^{\pm1},t_4^{\pm1},t_5^{\pm1},t_6^{\pm 1}].$ As a result of the above discussion, in all cases, we have that $A_{\alpha,\beta}S^{-1}_{\alpha,\beta}\cong \mathscr{A}_{\alpha,\beta}\cong {\mathbb K}_{q^N}[t_3^{\pm1},t_4^{\pm1},t_5^{\pm1},t_6^{\pm 1}].$ With a slight abuse of notation, we write $A_{\alpha,\beta}S^{-1}_{\alpha,\beta}= \mathscr{A}_{\alpha,\beta}= {\mathbb K}_{q^N}[t_3^{\pm1},t_4^{\pm1},t_5^{\pm1},t_6^{\pm 1}]$ for all $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ It follows from \cite[Theorem 6.3]{gll} that ${\rm GKdim} (A_{\alpha,\beta})={\rm GKdim} (A_{\alpha,\beta}S^{-1}_{\alpha,\beta})={\rm GKdim} (\mathscr{A}_{\alpha,\beta})=4.$ Since Tauvel's height formula holds in $A=U_q^+(G_2)$ \cite{gll}, we have that ${\rm GKdim} (A)= {\rm ht} (\langle \Omega_1-\alpha,\Omega_2-\beta \rangle) \ + {\rm GKdim} (A_{\alpha,\beta}).$ Since ${\rm GKdim} (A)=6$, we conclude that $ {\rm ht} (\langle\Omega_1-\alpha,\Omega_2-\beta\rangle) =2$ for all $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ \begin{proposition} ${\rm GKdim} (A_{\alpha,\beta})=4$ for all $(\alpha, \beta) \neq (0,0)$. \end{proposition} \subsection{Linear basis for $A_{\alpha,\beta}$} \label{c3s2} Set $A_{\beta}:={A}/{\langle \Omega_2-\beta\rangle},$ where $\beta\in{\mathbb K}.$ Now, denote the canonical images of $E_i$ by $\widehat{e}_i:=E_i+\langle \Omega_2-\beta\rangle$ in $A_{\beta}.$ Clearly, $A_{\alpha,\beta} \cong {A_{\beta}}/{\langle \widehat{\Omega}_1-\alpha\rangle}.$ As a result, one can identify $A_{\alpha,\beta}$ with ${A_{\beta}}/{\langle \widehat{\Omega}_1-\alpha\rangle}.$ Moreover, the algebra $A_{\beta}$ satisfies the relations of $A=U_q^+(G_2)$ and \begin{align} \widehat{e_2}\widehat{e_4}\widehat{e_6}+b\widehat{e_2}\widehat{e_5}^3+ b\widehat{e_3}^3\widehat{e_6}+b'\widehat{e_3}^2\widehat{e_5}^2+c' \widehat{e_3}\widehat{e_4}\widehat{e_5}+d'\widehat{e_4}^2=\beta.\label{c3eo} \end{align} From Propositions \ref{p0} and \ref{p1}, one can conclude that $\langle \Omega_2-\beta\rangle$ is a completely prime ideal (since it is a prime ideal) of $A$ for all $\beta\in {\mathbb K}.$ Hence, the algebra $A_{\beta}$ is a noetherian domain. We are now going to find a linear basis for $A_{\alpha,\beta},$ where $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ Since $A_{\alpha,\beta}$ is identified with ${A_{\beta}}/{\langle \widehat{\Omega}_1-\alpha\rangle},$ we will first and foremost find a basis for ${A_{\beta}},$ and then proceed to find a basis for $A_{\alpha,\beta}.$ Note, the relations in Lemma \ref{c3l3} are also valid in $A_{\beta}$ and $A_{\alpha,\beta},$ and are going to be very useful in this section. \begin{proposition} The set $\mathfrak{S}=\{\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k\widehat{e_4}^{\xi}\widehat{e_5}^l\widehat{e_6}^m\mid i,j,k,l,m\in \mathbb{N} \ \text{and} \ \xi=0,1\}$ is a ${\mathbb K}-$basis of $A_\beta.$ \label{c3p4} \end{proposition} \begin{proof} Since the family $(\Pi_{s=1}^6E_s^{i_s})_{i_s\in\mathbb{N}}$ is a PBW-basis of $A$ over ${\mathbb K},$ it follows that the family $(\Pi_{s=1}^6\widehat{e_s}^{i_s})_{i_s\in\mathbb{N}}$ is a spanning set of $A_{\beta}$ over ${\mathbb K}.$ We want to show that $\mathfrak{S}$ spans $A_{\beta}.$ We do this by showing that $\Pi_{s=1}^6\widehat{e_s}^{i_s}$ can be written as a finite linear combination of the elements of $\mathfrak{S}$ for all $i_1,\cdots,i_6\in \mathbb{N}$ by an induction on $i_4.$ The result is obvious when $i_4=0$ or $1.$ For $i_4\geq 1,$ assume that $$\Pi_{s=1}^6\widehat{e_s}^{i_s}= \sum_{(\xi,\underline{v})\in I}a_{(\xi,\underline{v})}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k\widehat{e_4}^\xi \widehat{e_5}^l \widehat{e_6} ^m,$$ where $\underline{v}:=(i,j,k,l,m)\in \mathbb{N}^5$ and $a_{(\xi,\underline{v})}$ are all scalars. Note, $I$ is a finite subset of $\{0,1\}\times \mathbb{N}^5.$ It follows from the commutation relations of $A_{\beta}$ (see Lemma \ref{c3l3}) that $$\widehat{e_1}^{i_1}\widehat{e_2}^{i_2}\widehat{e_3}^{i_3}\widehat{e_4}^{i_4+1} \widehat{e_5}^{i_5}\widehat{e_6}^{i_6}=q^{\bullet}\Pi_{s=1}^6\widehat{e_s}^{i_s}\widehat{e_4}-q^{\bullet}d_1[i_6]\widehat{e_1}^{i_1}\widehat{e_2}^{i_2}\widehat{e_3}^{i_3} \widehat{e_4}^{i_4}\widehat{e_5}^{i_5+3}\widehat{e_6}^{i_6-1}.$$ From the inductive hypothesis, $\widehat{e_1}^{i_1}\widehat{e_2}^{i_2}\widehat{e_3}^{i_3} \widehat{e_4}^{i_4}\widehat{e_5}^{i_5+3}\widehat{e_6}^{i_6-1}\in$ Span($\mathfrak{S}$). Hence, we proceed to show that $\Pi_{s=1}^6\widehat{e_s}^{i_s}\widehat{e_4}$ is also in the span of $\mathfrak{S}.$ From the inductive hypothesis, we have \begin{align*} \Pi_{s=1}^6\widehat{e_s}^{i_s}\widehat{e_4}=\sum_{(\xi,\underline{v})\in I}a_{(\xi,\underline{v})}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k\widehat{e_4}^\xi \widehat{e_5}^l \widehat{e_6}^m \widehat{e_4}. \end{align*} Using the commutation relations in Lemma \ref{c3l3}, we have that \begin{align*} \Pi_{s=1}^6\widehat{e_s}^{i_s}\widehat{e_4} =&\sum_{(\xi,\underline{v})\in I}q^\bullet a_{(\xi,\underline{v})}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k\widehat{e_4}^{\xi+1} \widehat{e_5}^l \widehat{e_6} ^m \\&+\sum_{(\xi,\underline{v})\in I} q^\bullet d_1[m]a_{(\xi,\underline{v})}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k \widehat{e_4}^{\xi}\widehat{e_5}^{l+3} \widehat{e_6}^{m-1}. \end{align*} All the terms in the above expression belong to the span of $\mathfrak{S}$ except $\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k\widehat{e_4}^2\widehat{e_5}^l \widehat{e_6}^m.$ From \eqref{c3eo}, we have that \begin{align} \widehat{e_4}^2=\beta_0\widehat{e_2}\widehat{e_4}\widehat{e_6}+b\beta_0\widehat{e_2} \widehat{e_5}^3+ b\beta_0\widehat{e_3}^3\widehat{e_6} +b'\beta_0\widehat{e_3}^2\widehat{e_5}^2+c'\beta_0\widehat{e_3}\widehat{e_4}\widehat{e_5} -\beta\beta_0,\label{c3e3} \end{align} where $\beta_0=-1/d'.$ Substituting \eqref{c3e3} into $\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k\widehat{e_4}^2\widehat{e_5}^l \widehat{e_6}^m,$ one can easily verify that $$\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k\widehat{e_4}^2\widehat{e_5}^l \widehat{e_6}^m\in \mathrm{Span}(\mathfrak{S}).$$ Therefore, $\widehat{e_1}^{i_1}\widehat{e_2}^{i_2}\widehat{e_3}^{i_3}\widehat{e_4}^{i_4+1} \widehat{e_5}^{i_5}\widehat{e_6}^{i_6}$ can be written as a finite linear combination of the elements of $\mathfrak{S}$ over ${\mathbb K}$ for all $i_1,\cdots,i_6\in \mathbb{N}.$ By the principle of mathematical induction, $\mathfrak{S}$ is a spanning set of $A_\beta$ over ${\mathbb K}.$\\ Next we show that $\mathfrak{S}$ is a linearly independent set. Suppose that $$\sum_{(\xi,\underline{v})\in I}a_{(\xi,\underline{v})}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k\widehat{e_4}^\xi \widehat{e_5}^l \widehat{e_6} ^m=0.$$ Since $A_{\beta}=A/{\langle\Omega_2-\beta\rangle}$, it implies that $$\sum_{(\xi,\underline{v})\in I}a_{(\xi,\underline{v})} E_1^iE_2^jE_3^kE_4^\xi E_5^l E_6^m=(\Omega_2-\beta)\nu,$$ with $\nu\in A.$ Write $\nu=\displaystyle\sum_{(i,\cdots,n)\in J}b_{(i,\cdots,n)}E_1^iE_2^jE_3^kE_4^lE_5^mE_6^n,$ where $J$ is a finite subset of $\mathbb{N}^6$ and $b_{(i,\cdots,n)}$ are all scalars. It follows that \begin{align} \label{xx} \sum_{(\xi,\underline{v})\in I}a_{(\xi,\underline{v})} E_1^iE_2^jE_3^kE_4^\xi E_5^l E_6^m&=\displaystyle\sum_{(i,\cdots,n)\in J}b_{(i,\cdots,n)}E_1^iE_2^jE_3^k(\Omega_2-\beta)E_4^lE_5^mE_6^n. \end{align} Before we continue the proof, the following needs to be noted. \begin{itemize} \item[$\clubsuit$ ] Let $(i',j',k',l',m',n'),$ $(i,j,k,l,m,n)\in$ $ \mathbb{N}^6.$ We say that $(i,j,k,l,m,n)<_4 (i',j',k',l',m',n')$ if $[l<l']$ or $[l=l'$ and $i<i']$ or $[l=l', \ i=i'$ and $j<j']$ or $[l=l', \ i=i', \ j=j'$ and $k<k']$ or $[l=l', \ i=i', \ j=j', \ k=k'$ and $m<m']$ or $[l=l', \ i=i', \ j=j', \ k=k', \ m=m'$ and $n\leq n'].$ Note, the purpose of the square bracket $[ \ ]$ is to differentiate the options. \end{itemize} From Section \ref{c2s1}, we have that $\Omega_2=E_2E_4E_6+bE_2E_5^3+bE_3^3E_6 +b'E_3^2E_5^2+c'E_3E_4E_5+d'E_4^2$ in $A=U_q^+(G_2).$ Now, \begin{align*} \sum_{(\xi,\underline{v})\in I}a_{(\xi,\underline{v})} E_1^iE_2^jE_3^kE_4^\xi E_5^l E_6^m&=\displaystyle\sum_{(i,\cdots,n)\in J}b_{(i,\cdots,n)}E_1^iE_2^jE_3^k(\Omega_2-\beta)E_4^lE_5^mE_6^n \\&=\sum_{(i,\cdots,n)\in J}d'b_{(i,\cdots,m)}E_1^iE_2^jE_3^kE_4^{l+2}E_5^mE_6^n + \text{LT}_{<_4}, \end{align*} where $\text{LT}_{<_4}$ contains lower order terms with respect to $<_4$ (as in $\clubsuit$). Moreover, $\text{LT}_{<_4}$ vanishes when $b_{(i,\cdots,n)}=0$ for all $(i,\cdots,n)\in J$ (one can easily confirm this by fully expanding the right hand side of \eqref{xx}). Now, suppose that there exists $(i,j,k,l,m,n)\in J$ such that $b_{(i,j,k,l,m,n)}\neq 0.$ \newline Let $(i',j',k',l',m',n')$ be the greatest element of $J$ with respect to $<_4$ (defined in $\clubsuit$ above) such that $b_{(i',j',k',l',m',n')}\neq 0.$ Note, the family $(E_1^iE_2^jE_3^kE_4^lE_5^mE_6^n)_{(i,\cdots,n)\in \mathbb{N}^6}$ is a basis of $A.$ Since $\text{LT}_{<_4}$ contains lower order terms, identifying the coefficients of $E_1^{i'}E_2^{j'}E_3^{k'}E_4^{l'+2}E_5^{m'}E_6^{n'}$ in the above equality, we have that $d'b_{(i',\cdots,n')}=0.$ Since $b_{(i',j',k',l',m',n')}\neq 0,$ it follows that $d'={q^{12}}/(q^6-1)=0,$ a contradiction (see Appendix \ref{appenc} for the expression of $d'$). As a result, $b_{(i,j,k,l,m,n)}=0$ for all $(i,j,k,l,m,n)\in J.$ Therefore, $\sum_{(\xi,\underline{v})\in I}a_{(\xi,\underline{v})} E_1^iE_2^jE_3^kE_4^\xi E_5^l E_6^m=0.$ Since $(E_1^iE_2^jE_3^kE_4^lE_5^mE_6^n)_{(i,\cdots,n)\in\mathbb{N}^6}$ is a basis of $A,$ it follows that $a_{(\xi,\underline{v})}=0$ for all $(\xi,\underline{v})\in I.$ In conclusion, $\mathfrak{S}$ is a linearly independent set and hence forms a basis of $A_\beta$ as desired. \end{proof} \begin{proposition} \label{c3p5} Let $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ The set $\mathcal{B}=\{e_1^ie_2^je_3^{\epsilon_1}e_4^{\epsilon_2}e_5^ke_6^l\mid i,j,k,l\in \mathbb{N} \ \text{and} \ \epsilon_1,\epsilon_2\in \{0,1\} \}$ is a ${\mathbb K}-$basis of $A_{\alpha,\beta}.$ \end{proposition} \begin{proof} Since the set $\mathfrak{S}=\{\widehat{e_1}^{i_1}\widehat{e_2}^{i_2}\widehat{e_3}^{i_3}\widehat{e_4}^{\xi} \widehat{e_5}^{i_5}\widehat{e_6}^{i_6}\mid i_1,i_2,i_3,i_5,i_6\in\mathbb{N} \ \text{and} \ \xi=0,1\} $ is a ${\mathbb K}-$basis of $A_\beta$ (Proposition \ref{c3p4}), and $A_{\alpha,\beta}$ is identified with ${A_\beta}/{\langle \widehat{\Omega}_1-\alpha\rangle},$ it follows that $$(e_1^{i_1}e_2^{i_2}e_3^{i_3}e_4^\xi e_5^{i_5}e_6^{i_6})_{i_1,i_2,i_3,i_5,i_6\in \mathbb{N}, \ \xi\in \{0,1\}}$$ is a spanning set of $A_{\alpha,\beta}$ over ${\mathbb K}.$ We want to show that $\mathcal{B}$ spans $A_{\alpha,\beta}$ by showing that $e_1^{i_1}e_2^{i_2}e_3^{i_3}e_4^\xi e_5^{i_5}e_6^{i_6}$ can be written as a finite linear combination of the elements of $\mathcal{B}$ over ${\mathbb K}$ for all $i_1,i_2,i_3,i_5,i_6\in\mathbb{N}$ and $\xi=0,1.$ By Proposition \ref{c3p4}, it is sufficient to do this by an induction on $i_3.$ The result is obvious when $i_3=0$ or $1.$ For $ i_3\geq 1,$ suppose that \begin{align*} e_1^{i_1}e_2^{i_2}e_3^{i_3}e_4^\xi e_5^{i_5}e_6^{i_6} &=\sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})} e_1^ie_2^je_3^{\epsilon_1}e_4^{\epsilon_2}e_5^ke_6^l, \end{align*} where $\underline{\upsilon}=(i,j,k,l)\in \mathbb{N}^4,$ \ $I$ is a finite subset of $\{0,1\}^2\times \mathbb{N}^4$, and $(a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})})_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}$ is a family of scalars. Using the commutation relations in $A_{\alpha,\beta}$ (Lemma \ref{c3l3}), we have that: $$e_1^{i_1}e_2^{i_2}e_3^{i_3+1}e_4^\xi e_5^{i_5}e_6^{i_6} =q^{\bullet}e_3e_1^{i_1}e_2^{i_2}e_3^{i_3}e_4^\xi e_5^{i_5}e_6^{i_6}-q^{\bullet}d_2[i_1]e_1^{i_1-1}e_2^{1+i_2}e_3^{i_3} e_4^\xi e_5^{i_5}e_6^{i_6}.$$ From the inductive hypothesis, $e_1^{i_1-1}e_2^{1+i_2}e_3^{i_3} e_4^\xi e_5^{i_5}e_6^{i_6}\in$ Span($\mathcal{B}$) for all $i_1>0$ (note: $d_2[0]=0$). As a result, we proceed to show that $e_3e_1^{i_1}e_2^{i_2}e_3^{i_3}e_4^\xi e_5^{i_5}e_6^{i_6}$ is also in the span of $\mathcal{B}.$ It follows from the inductive hypothesis that \begin{align*} e_3e_1^{i_1}e_2^{i_2}e_3^{i_3}e_4^\xi e_5^{i_5}e_6^{i_6}=& \sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})} e_3e_1^ie_2^je_3^{\epsilon_1}e_4^{\epsilon_2}e_5^ie_6^l\\ =&\sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})} e_1^ie_2^je_3^{\epsilon_1+1}e_4^{\epsilon_2}e_5^ke_6^l\\ &+\sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}d_2[i]a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})} e_1^{i-1}e_2^{1+j}e_3^{\epsilon_1}e_4^{\epsilon_2}e_5^ke_6^l. \end{align*} Clearly, the monomial $e_1^{i-1}e_2^{1+j}e_3^{\epsilon_1}e_4^{\epsilon_2}e_5^ke_6^l$ belongs to the span of $\mathcal{B}$ for all $(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I$ (with $i>0$). Again, the monomial $e_1^ie_2^je_3^{\epsilon_1+1}e_4^{\epsilon_2}e_5^ke_6^l$ belongs to the span of $\mathcal{B}$ for all $(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I;$ with $(\epsilon_1,\epsilon_2)=(0,0), (0,1).$ For $(\epsilon_1, \epsilon_2) =(1,0), (1,1);$ we must show that $e_1^ie_2^je_3^2e_5^ke_6^l$ and $e_1^ie_2^je_3^{2}e_4e_5^ke_6^l$ belong to the span of $\mathcal{B}.$ From Lemma \ref{c3l2}, one can write $e_1^ie_2^je_3^2e_5^ke_6^l$ and $ e_1^ie_2^je_3^2e_4e_5^ke_6^l$ as finite linear combinations of the elements of $\mathcal{B}$ over ${\mathbb K}.$ Hence, $e_1^ie_2^je_3^2e_5^ke_6^l$ and $ e_1^ie_2^je_3^2e_4e_5^ke_6^l$ belong to the span of $\mathcal{B}$ for all $(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I;$ with $(\epsilon_1,\epsilon_2) =(1,0), (1,1).$ We have therefore established that $e_3e_1^{i_1}e_2^{i_2}e_3^{i_3}e_4^\xi e_5^{i_5}e_6^{i_6}\in$ Span($\mathcal{B}$). Consequently, each $e_1^{i_1}e_2^{i_2}e_3^{i_3+1}e_4^\xi e_5^{i_5}e_6^{i_6}$ belongs to the span of $\mathcal{B}.$ By the principle of mathematical induction, $\mathcal{B}$ is a spanning set of $A_{\alpha,\beta}$ over ${\mathbb K}$ as expected.\\ Finally, we show that $\mathcal{B}$ is a linearly independent set. Suppose that \begin{align*} \sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})} e_1^ie_2^je_3^{\epsilon_1}e_4^{\epsilon_2}e_5^ke_6^l=0. \end{align*} In $A_\beta,$ we have that \begin{align*} \sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^{\epsilon_1}\widehat{e_4}^{\epsilon_2} \widehat{e_5}^k\widehat{e_6}^l &= (\widehat{\Omega}_1-\alpha)\nu, \end{align*} with $\nu\in A_{\beta}.$ One can write $\nu$ in terms of the basis $\mathfrak{S}$ of $A_{\beta}$ (Proposition \ref{c3p4}) as: $$\nu=\displaystyle\sum_{\underline{w}\in J_1}b_{\underline{w}}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k \widehat{e_4}\widehat{e_5}^m \widehat{e_6}^n+\displaystyle\sum_{\underline{w}\in J_2}c_{\underline{w}}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k \widehat{e_5}^l \widehat{e_6}^m,$$ where $b_{\underline{w}}$ and $c_{\underline{w}}$ are all scalars and $w:=(i,j,k,l,m)$. Hence, \begin{align*} \sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})} \widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^{\epsilon_1}\widehat{e_4}^{\epsilon_2} \widehat{e_5}^k\widehat{e_6}^l=&\sum_{\underline{w}\in J_1}b_{\underline{w}}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k(\widehat{\Omega}_1-\alpha) \widehat{e_4}\widehat{e_5}^l \widehat{e_6}^m\\&+\displaystyle\sum_{\underline{w}\in J_2}c_{\underline{w}}\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^k (\widehat{\Omega}_1-\alpha)\widehat{e_5}^l \widehat{e_6}^m. \end{align*} Note, $\widehat{\Omega}_1=\widehat{e_1}\widehat{e_3}\widehat{e_5}+a\widehat{e_1}\widehat{e_4}+ a\widehat{e_2}\widehat{e_5}+a'\widehat{e_3}^2.$ Using \eqref{c3e3} and the relation $e_3^ke_1=q^{-k}e_1e_3^k+d_2[k]e_2e_3^{k-1}$ (see Lemma \ref{c3l3}), one can verify that the above equality can be written in terms of the basis of $A_{\beta}$ (Propositions \ref{c3p4}) as: \begin{align} \label{aa} \sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})} \widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^{\epsilon_1}\widehat{e_4}^{\epsilon_2} \widehat{e_5}^k\widehat{e_6}^l =&\sum_{\underline{w}\in J_1} b_{\underline{w}} a'\widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^{k+2}\widehat{e_4} \widehat{e_5}^l\widehat{e_6}^m\nonumber\\ &+\sum_{\underline{w}\in J_1}q^\bullet b_{\underline{w}} ab\beta_0\widehat{e_1}^{i+1}\widehat{e_2}^j\widehat{e_3}^{k+3} \widehat{e_5}^l\widehat{e_6}^{m+1}\nonumber\\ &+\sum_{\underline{w}\in J_2}q^{\bullet} c_{\underline{w}}a\widehat{e_1}^{i+1}\widehat{e_2}^{j}\widehat{e_3}^k\widehat{e_4} \widehat{e_5}^{l}\widehat{e_6}^m\nonumber\\ &+\sum_{\underline{w}\in J_2}c_{\underline{w}} a'\widehat{e_1}^i\widehat{e_2}^{j}\widehat{e_3}^{k+2} \widehat{e_5}^l\widehat{e_6}^m+\Upsilon, \end{align} where $\Upsilon$ is defined as follows: \begin{align*} \Upsilon =&\sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}} \widehat{e_1}^{i+1}\widehat{e_2}^j\widehat{e_3}^{k+1}\widehat{e_4} \widehat{e_5}^{l+1}\widehat{e_6}^m + \sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}d_2[k] \widehat{e_1}^{i}\widehat{e_2}^{1+j}\widehat{e_3}^{k}\widehat{e_4} \widehat{e_5}^{l+1}\widehat{e_6}^m\\ &-\sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}a\beta\beta_0 \widehat{e_1}^{i+1}\widehat{e_2}^{j}\widehat{e_3}^{k} \widehat{e_5}^{l}\widehat{e_6}^m+ \sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}a\beta_0 \widehat{e_1}^{i+1}\widehat{e_2}^{1+j}\widehat{e_3}^{k}\widehat{e_4} \widehat{e_5}^{l}\widehat{e_6}^{m+1}\\ &+ \sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}ab\beta_0 \widehat{e_1}^{i+1}\widehat{e_2}^{j+1}\widehat{e_3}^{k} \widehat{e_5}^{l+3}\widehat{e_6}^m+ \sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}ab'\beta_0 \widehat{e_1}^{i+1}\widehat{e_2}^{j}\widehat{e_3}^{k+2} \widehat{e_5}^{l+2}\widehat{e_6}^m\\ &+\sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}ac'\beta_0 \widehat{e_1}^{i+1}\widehat{e_2}^{j}\widehat{e_3}^{k+1} \widehat{e_4}\widehat{e_5}^{l+1}\widehat{e_6}^m -\sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}a\beta\beta_0 d_2[k] \widehat{e_1}^{i}\widehat{e_2}^{1+j}\widehat{e_3}^{k-1} \widehat{e_5}^{l}\widehat{e_6}^m\\&+ \sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}a\beta_0d_2[k] \widehat{e_1}^{i}\widehat{e_2}^{j+2}\widehat{e_3}^{k-1} \widehat{e_4} \widehat{e_5}^{l}\widehat{e_6}^{m+1}+ \sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}ab\beta_0d_2[k] \widehat{e_1}^{i}\widehat{e_2}^{j+2}\widehat{e_3}^{k-1} \widehat{e_5}^{l+3}\widehat{e_6}^m\\ & +\sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}ab\beta_0d_2[k] \widehat{e_1}^{i+1}\widehat{e_2}^{j}\widehat{e_3}^{k+2} \widehat{e_5}^{l}\widehat{e_6}^{m+1} +\sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}ab'\beta_0d_2[k] \widehat{e_1}^{i}\widehat{e_2}^{j+1}\widehat{e_3}^{k+1} \widehat{e_5}^{l+2}\widehat{e_6}^m\\ &+\sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}ac'\beta_0d_2[k] \widehat{e_1}^{i}\widehat{e_2}^{j+1}\widehat{e_3}^{k} \widehat{e_4}\widehat{e_5}^{l+1}\widehat{e_6}^m +\sum_{\underline{w}\in J_1}q^{\bullet} b_{\underline{w}}a \widehat{e_1}^{i}\widehat{e_2}^{j+1}\widehat{e_3}^{k} \widehat{e_4}\widehat{e_5}^{l+1}\widehat{e_6}^m\\ &-\sum_{\underline{w}\in J_1}b_{\underline{w}}\beta \widehat{e_1}^{i}\widehat{e_2}^{j}\widehat{e_3}^{k}\widehat{e_4} \widehat{e_5}^{l}\widehat{e_6}^m+\sum_{\underline{w}\in J_2}q^{\bullet} c_{\underline{w}}\widehat{e_1}^{i+1}\widehat{e_2}^{j}\widehat{e_3}^{k+1} \widehat{e_5}^{l+1}\widehat{e_6}^m\\ & +\sum_{\underline{w}\in J_2}q^{\bullet} c_{\underline{w}}d_2[k]\widehat{e_1}^{i}\widehat{e_2}^{1+j}\widehat{e_3}^{k} \widehat{e_5}^{l+1}\widehat{e_6}^m+\sum_{\underline{w}\in J_2}q^{\bullet}c_{\underline{w}}ad_2[k]\widehat{e_1}^{i}\widehat{e_2}^{j+1}\widehat{e_3}^{k-1}\widehat{e_4} \widehat{e_5}^{l}\widehat{e_6}^m\\ & +\sum_{\underline{w}\in J_2}q^{\bullet}c_{\underline{w}}a\widehat{e_1}^{i}\widehat{e_2}^{j+1}\widehat{e_3}^{k} \widehat{e_5}^{l+1}\widehat{e_6}^m-\sum_{\underline{w}\in J_2}c_{\underline{w}}\alpha\widehat{e_1}^{i}\widehat{e_2}^{j}\widehat{e_3}^{k} \widehat{e_5}^{l}\widehat{e_6}^m. \end{align*} Before we continue the proof, the following point needs to be noted. \begin{itemize} \item[$\spadesuit$] Let $(\vartheta_1,\vartheta_2,\vartheta_3,\vartheta_5,\vartheta_6)$, $(\varsigma_1,\varsigma_2,\varsigma_3,\varsigma_5,\varsigma_6)\in\mathbb{N}^5.$ We say that $(\varsigma_1,\varsigma_2,\varsigma_3,\varsigma_5,\varsigma_6)<_3 (\vartheta_1,\vartheta_2,\vartheta_3,\vartheta_5,\vartheta_6)$ if $[\vartheta_3>\varsigma_3]$ or $[\vartheta_3=\varsigma_3$ and $\vartheta_1>\varsigma_1]$ or $[\vartheta_3=\varsigma_3, \ \vartheta_1=\varsigma_1$ and $\vartheta_2>\varsigma_2]$ or $[\vartheta_3=\varsigma_3,\ \vartheta_1=\varsigma_1, \ \vartheta_2=\varsigma_2$ and $\vartheta_5>\varsigma_5]$ or $[\vartheta_3=\varsigma_3,\ \vartheta_1=\varsigma_1, \ \vartheta_2=\varsigma_2,\ \vartheta_5=\varsigma_5$ and $\vartheta_6\geq \varsigma_6].$ \end{itemize} Observe that $\Upsilon$ contains lower order terms with respect to $<_3$ (defined in $\spadesuit$ above) in each monomial type (note, there are two different types of monomials in the basis of $A_{\beta}$: one with $\widehat{e_4}$ and the other without $\widehat{e_4}$). Now, suppose that there exists $(i,j,k,l,m)\in J_1$ and $(i,j,k,l,m)\in J_2$ such that $b_{(i,j,k,l,m)}\neq 0$ and $c_{(i,j,k,l,m)}\neq 0.$ Let $(v_1,v_2,v_3,v_5,v_6)$ and $(w_1,w_2,w_3,w_5,w_6)$ be the greatest elements of $J_1$ and $J_2$ respectively with respect to $<_{3}$ such that $b_{(v_1,v_2,v_3, v_5,v_6)}$ and $ c_{(w_1,w_2,w_3,w_5,w_6)}$ are non-zero. Since $\mathfrak{S}$ is a linear basis for $A_\beta,$ and $\Upsilon$ contains lower order terms with respect to $<_3$, we have the following: if $w_3-v_3<2,$ then identifying the coefficients of $\widehat{e_1}^{v_1}\widehat{e_2}^{v_2}\widehat{e_3}^{v_3+2}\widehat{e_4} \widehat{e_5}^{v_5}\widehat{e_6}^{v_6}$ in \eqref{aa}, we have that $a'b_{(v_1,v_2,v_3, v_5,v_6)}=0$. But $b_{(v_1,v_2,v_3, v_5,v_6)}\neq 0,$ hence $ a'=q^6/(q^2-1)=0,$ a contradiction (see Appendix \ref{appenc} for the expression of $a'$). Finally, if $w_3-v_3\geq 2,$ then identifying the coefficient of $\widehat{e_1}^{w_1}\widehat{e_2}^{w_2}\widehat{e_3}^{w_3+2}\widehat{e_5}^{w_5}\widehat{e _6}^{w_6},$ we have that $a'c_{(w_1,w_2,w_3,w_5,w_6)}=0$. But $c_{(w_1,w_2,w_3,w_5,w_6)}\neq 0,$ hence $a'=0,$ a contradiction! This implies that either all $b_{(i,j,k,l,m)}$ or all $ c_{(i,j,k,l,m)}$ are zero. Without loss of generality, suppose that there exists $(i,j,k,l,m)\in J_2$ such that $c_{(i,j,k,l,m)}$ is not zero. Then, $ b_{(i,j,k,l,m)}$ are all zero. Let $(w_1,w_2,w_3,w_5,w_6)$ be the greatest element of $J_2$ such that $ c_{(w_1,w_2,w_3,w_5,w_6)}\neq 0.$ Identifying the coefficients of $\widehat{e_1}^{w_1}\widehat{e_2}^{w_2} \widehat{e_3}^{w_3+2} \widehat{e_5}^{w_5}\widehat{e_6}^{w_6}$ in the above equality, we have that $a'c_{(w_1,w_2,w_3,w_5,w_6)}=0.$ Since $c_{(w_1,w_2,w_3,w_5,w_6)}\neq 0,$ it follows that $a'=0,$ a contradiction! We can therefore conclude that $b_{(i,j,k,l,m)}$ and $ c_{(i,j,k,l,m)}$ are all zero. Consequently, $ \sum_{(\epsilon_1,\epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})} \widehat{e_1}^i\widehat{e_2}^j\widehat{e_3}^{\epsilon_1}\widehat{e_4}^{\epsilon_2} \widehat{e_5}^k\widehat{e_6}^l = 0. $ Since $(\widehat{e_1}^{i_1}\widehat{e_2}^{i_2}\widehat{e_3}^{i_3}\widehat{e_4}^{\xi} \widehat{e_5}^{i_5}\widehat{e_6}^{i_6})_{(\xi,i_1,\cdots,i_6)\in \{0,1\}\times \mathbb{N}^5}$ is a basis of $A_{\beta}$, it implies that $a_{(\epsilon_1,\epsilon_2,\underline{v})}=0$ for all $(\epsilon_1,\epsilon_2,\underline{v})\in I.$ Therefore, $\mathcal{B}$ is a linearly independent set. \end{proof} We note for future use the following immediate consequence of Proposition \ref{c3p5}. \begin{corollary} \label{c4c2} Let $\underline{v}=(i,j,k,l)\in \mathbb{N}^2\times \mathbb{Z}^2, \ I$ represent a finite subset of $\{0,1\}\times \mathbb{N}^2\times \mathbb{Z}^2$ and $(a_{(\epsilon_1,\epsilon_2,\underline{v})})_{(\epsilon_1,\epsilon_2, \underline{v})\in I}$ be a family of scalars. If $$\sum_{(\epsilon_1,\epsilon_2, \underline{v})\in I}a_{(\epsilon_1,\epsilon_2, \underline{v})}e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k}t_6^{l}=0,$$ then $a_{(\epsilon_1,\epsilon_2, \underline{v})}=0$ for all $(\epsilon_1,\epsilon_2, \underline{v})\in I.$ \end{corollary} \begin{remark} Given the basis of $A_{\alpha,\beta},$ we have computed the group of units of $A_{\alpha,\beta},$ however, we do not include the details in this manuscript due to the voluminous computations involved. We only summarise our findings below. Set $$h_1:=e_3e_5+ae_4 \ \ \text{and} \ \ h_2:=(q^{-3}-q^{-9})e_2e_4-(q^4-2q^2+1)/(q^4+q^2+1)e_3^3.$$ \begin{theorem} Let $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\}$ and $\mathcal{U}(A_{\alpha,\beta})$ denote the group of units of $A_{\alpha,\beta}.$ We have that: $$\mathcal{U}(A_{\alpha,\beta})=\begin{cases} \{\lambda h_1^i\mid \lambda\in{\mathbb K}^*, \ i\in\mathbb{Z}\}& \text{if $\alpha=0$}\\ \{\lambda h_2^i\mid \lambda\in{\mathbb K}^*, \ i\in\mathbb{Z}\}& \text{if $\beta=0$}\\ {\mathbb K}^*& \text{$otherwise$.} \end{cases} $$ \end{theorem} \end{remark} \subsection{$A_{\alpha,\beta}$ as a $q$-deformation of a quadratic extension of $A_2({\mathbb K})$} \label{rec} Recall that ${\rm GKdim} A_{\alpha,\beta} =4$ and so we should compare $A_{\alpha,\beta}$ to the second Weyl algebra. In this section, we prove that, for a suitable choice of $\alpha$ and $\beta$, the simple algebra $A_{\alpha,\beta}$ is a $q$-deformation of (a quadratic extension of) $A_2({\mathbb K}).$ Recall that $A_2({\mathbb K})$ is generated by $x_1,x_2,y_1$ and $y_2$ subject to the relations: \begin{align*} y_1y_2&=y_2y_1 & x_2y_1&=y_1x_2 & x_1x_2&=x_2x_1 & x_1y_1&-y_1x_1=1\\ y_1y_2&=y_2y_1 & x_1y_2&=y_2x_1& x_2y_1&=y_1x_2 & x_ 2y_2&-y_2x_2=1. \end{align*} Given the relations of $A_{\alpha,\beta}$ at the onset of this section, we have that $A_{1,\frac{1}{9(q^6-1)}}$ satisfies the following relations: \begin{align*} e_2e_1&=q^{-3}e_1e_2& e_3e_1&=q^{-1}e_1e_3-(q+q^{-1}+q^{-3})e_2\\ e_3e_2&=q^{-3}e_2e_3&e_4e_1&=e_1e_4+(1-q^2)e_3^2\\ e_4e_2&=q^{-3}e_2e_4-\frac{q^4-2q^2+1}{q^4+q^2+1}e_3^3&e_4e_3&=q^{-3}e_3e_4\\ e_5e_1&=qe_1e_5-(1+q^2)e_3&e_5e_2&=e_2e_5+(1-q^2)e_3^2\\ e_5e_3&=q^{-1}e_3e_5-(q+q^{-1}+q^{-3})e_4&e_5e_4&=q^{-3}e_4e_5\\ e_6e_1&=q^3e_1e_6-q^3e_5&e_6e_2&=q^3e_2e_6+(q^4+q^2-1)e_4+(q^2-q^4)e_3e_5\\ e_6e_3&=e_3e_6+(1-q^2)e_5^2&e_6e_4&=q^{-3}e_4e_6-\frac{q^4-2q^2+1}{q^4+q^2+1}e_5^3\\ e_6e_5&=q^{-3}e_5e_6, \end{align*} and \begin{align*} (q^{-2}-1)e_1e_3e_5+(q^2+1+q^{-2}) e_1e_4+(q^2+1+q^{-2})e_2e_5-{q^4}e_3^2&=q^{-2}-1,\\ \\ (q^6-1)e_2e_4e_6+\frac{2q^{-1}-q^{-3}-q}{q^4+q^2+1} e_2e_5^3+\frac{2q^{-1}-q^{-3}-q}{q^4+q^2+1}e_3^3e_6&\\+\frac{(q^6-1)(q^{13}-q^{11})}{(q^4+q^2+1)^2}e_3^2e_5^2-\frac{q^9(q^6-1)}{q^4+q^2+1}e_3e_4e_5+q^{12}e_4^2&=\frac{1}{9}. \end{align*} Note, we have made the necessary substitutions for $a,\ a', \ b, \ b', \ c'$ and $d'$ from Appendix \ref{appenc}. Set $F:={\mathbb K}[z^{\pm 1}].$ One can define a $F[(z^4+z^2+1)^{-1}]-$algebra $A_{z}$ generated by $e_1, \cdots, e_6$ subject to the following relations: \begin{align*} e_2e_1&=z^{-3}e_1e_2& e_3e_1&=z^{-1}e_1e_3-(z+z^{-1}+z^{-3})e_2\\ e_3e_2&=z^{-3}e_2e_3&e_4e_1&=e_1e_4+(1-z^2)e_3^2\\ e_4e_2&=z^{-3}e_2e_4-\frac{z^4-2z^2+1}{z^4+z^2+1}e_3^3&e_4e_3&=z^{-3}e_3e_4\\ e_5e_1&=z e_1e_5-(1+z^2)e_3&e_5e_2&=e_2e_5+(1-z^2)e_3^2\\ e_5e_3&=z^{-1}e_3e_5-(z+z^{-1}+z^{-3})e_4&e_5e_4&=z^{-3}e_4e_5\\ e_6e_1&=z^3e_1e_6-z^3e_5&e_6e_2&=z^3e_2e_6+(z^4+z^2-1)e_4+(z^2-z^4)e_3e_5\\ e_6e_3&=e_3e_6+(1-z^2)e_5^2&e_6e_4&=z^{-3}e_4e_6-\frac{z^4-2z^2+1}{z^4+z^2+1}e_5^3\\ e_6e_5&=z^{-3}e_5e_6, \end{align*} \begin{align*} (z^{-2}-1)e_1e_3e_5+(z^2+1+z^{-2}) e_1e_4+(z^2+1+z^{-2})e_2e_5-{z^4}e_3^2&=z^{-2}-1, \ \ \text{and} \end{align*} \begin{align*} (z^6-1)e_2e_4e_6+\frac{2z^{-1}-z^{-3}-z}{z^4+z^2+1} e_2e_5^3+\frac{2z^{-1}-z^{-3}-z}{z^4+z^2+1}e_3^3e_6&\\+\frac{(z^6-1)(z^{13}-z^{11})}{(z^4+z^2+1)^2}e_3^2e_5^2-\frac{z^9(z^6-1)}{z^4+z^2+1}e_3e_4e_5+z^{12}e_4^2&=\frac{1}{9}. \end{align*} Set $A_1:= A_z / \langle z-1 \rangle$ and observe that $A_1$ satisfies the following relations: \begin{align*} e_2e_1&=e_1e_2& e_3e_1&=e_1e_3-3 e_2& e_3e_2&=e_2e_3\\ e_4 e_1&=e_1e_4& e_4e_2&=e_2e_4&e_4e_3&= e_3 e_4\\ e_5e_1&=e_1e_5- 2e_3&e_5e_2&=e_2e_5 &e_5e_3&=e_3e_5 -3e_4\\ e_5e_4&=e_4 e_5& e_6e_1&=e_1e_6-e_5& e_6e_2&=e_2e_6 +e_4\\ e_6e_3&=e_3e_6&e_6e_4&= e_4e_6 & e_6e_5&=e_5e_6\\ e_4^2-&1/9=0&&& e_3^2-&3e_1e_4-3e_2e_5=0. \end{align*} \begin{lemma} $e_4\in Z(A_1)$ and it is also invertible. \end{lemma} \begin{proof} Since $e_4e_i=e_ie_4$ for all $1\leq i\leq 6,$ we have that $e_4\in Z(A_1).$ Again, from $e_4^2-1/9=0,$ we have that $e_4(9e_4)=(9e_4)e_4=1.$ Hence $e_4$ is invertible with $e_4^{-1}=9e_4.$ \end{proof} Given that $e_4^{-1}=9e_4$ and $e_4\in Z(A_1),$ it follows from the relation $e_3^2-3e_1e_4-3e_2e_5=0$ that $e_1=3e_3^2e_4-9e_2e_4e_5.$ Therefore, $A_1$ can be generated by only $e_2, \cdots, e_6.$ All these generators commute except that $$e_6e_2=e_2e_6+e_4 \ \ \text{and} \ \ e_5e_3=e_3e_5 -3e_4.$$ Since $e_4$ is invertible, one can also verify that $9e_2e_4, 3e_3e_4, e_4, e_5$ and $e_6$ generate $A_1.$ Let $R$ be an algebra generated by $f_2, f_3, f_4, f_5, f_6$ subject to the following defining relations: \begin{align*} f_3f_2&=f_2f_3 & f_4f_2&=f_2f_4& f_4f_3&=f_3f_4\\ f_5f_2&=f_2f_5 & f_5f_4&=f_4f_5 & f_6f_3&=f_3f_6\\ f_6f_4&=f_4f_6& f_6f_5&=f_5f_6& f_4^2&=1/9\\ f_6f_2&=f_2f_6+1&f_5f_3&=f_3f_5-1. \end{align*} \begin{proposition} $R\cong A_1.$ \end{proposition} \begin{proof} One can easily check that we define a homomorphism $\phi: R\longrightarrow A_1$ via \begin{align*} \phi(f_2)&=9e_2e_4& \phi(f_3)&=3e_3e_4& \phi(f_4)&=e_4& \phi(f_5)&=e_5& \phi(f_6)&=e_6. \end{align*} Recall, $e_4^2=1/9.$ To check that $\phi$ is indeed a homomorphism, we just need to check its compatibility with the defining relations of $R.$ We check this on the relation $f_6f_2-f_2f_6=1$ and $f_3f_5-f_5f_3=1,$ and leave the remaining ones for the reader to verify. We do that as follows: $\phi(f_6)\phi(f_2)-\phi(f_2)\phi(f_6)=9e_6e_2e_4-9e_2e_4e_6=9e_4(e_6e_2-e_2e_6)=9e_4^2=9(1/9)=1$ as needed. Also, $\phi(f_3)\phi(f_5)-\phi(f_5)\phi(f_3)=3e_3e_4e_5-3e_5e_3e_4=3e_4(e_3e_5-e_5e_3)=3e_4(3e_4)=9e_4^2=1.$ Conversely, one can check that we define a homomorphism $\varphi: A_1\longrightarrow R$ via \begin{align*} \varphi(e_1)&=3f_3^2f_4-f_2f_5& \varphi(e_2)&=f_2f_4& \varphi(e_3)&=3f_3f_4\\ \varphi(e_4)&=f_4& \varphi(e_5)&=f_5 & \varphi(e_6)&= f_6. \end{align*} We check this on the relation $e_3^2-3e_1e_4-3e_2e_5=0,$ and leave the remaining ones for the reader to verify. We do that as follows: $\varphi(e_3)^2-3\varphi(e_1)\varphi(e_4)-3\varphi(e_2)\varphi(e_5)= (3f_3f_4)^2-3(3f_3^2f_4-f_2f_5)f_4-3f_2f_4f_5=9f_3^2f_4^2-9f_3^2f_4^2+ 3f_2f_4f_5-3f_2f_4f_5=0$ as expected. To conclude we just observe that $\phi$ and $\varphi$ are inverse of each other. \end{proof} The corollary below can easily be deduced from the above proposition. \begin{corollary} Set $\mathbb{F}:={\mathbb K}[f_4]/\langle f_4^2-1/9\rangle,$ we have that $R\cong A_2(\mathbb{F}),$ where $A_2(\mathbb{F})$ is the second Weyl algebra over the ring $\mathbb{F}.$ \end{corollary} \begin{remark} Observe that the subalgebra $B$ of $R$ generated by $f_2, f_3,f_5, f_6$ is isomorphic to $A_2({\mathbb K})$ and $R\cong B[f_4]\cong A_2({\mathbb K})[f_4].$ Therefore, $R$ is a \textit{quadratic extension} of $A_2({\mathbb K}).$ Note, $A_{1,\frac{1}{9(q^6-1)}}$ is a $q$-deformation of $A_1\cong R\cong A_2(\mathbb{F})\cong A_2({\mathbb K})[f_4].$ \end{remark} \section{Derivations of the simple quotients of $U_q^+(G_2)$} \label{chapd} In this section, we compute the derivations of the algebra $A_{\alpha,\beta}$ using DDA that allows to embed $A_{\alpha,\beta}$ into a suitable quantum torus. Derivations of quantum tori are known, thanks to the work of Osborn and Passman \cite{op}. In our cases, such derivations are always the sum of an inner derivation and a scalar derivation (of the quantum torus). Since $A_{\alpha,\beta}$ can be embedded into a quantum torus, we first extend every derivation of $A_{\alpha,\beta}$ to a derivation of such quantum torus, and then pull back their description as a derivation of the quantum torus to a description of their action on the generators of $A_{\alpha,\beta}$ by ``reverting'' DDA process. We conclude that every derivation of $A_{\alpha,\beta}$ is inner when $\alpha$ and $\beta$ are both non-zero. However, when either $\alpha$ or $\beta$ is zero, we conclude that every derivation of $A_{\alpha,\beta}$ is the sum of an inner and a scalar derivation. In fact, the first Hochschild cohomology group of $A_{\alpha,\beta}$ is of dimension $0$ when $\alpha$ and $\beta$ are both non-zero and $1$ when either $\alpha$ or $\beta$ is zero. \subsection{Preliminaries and strategy} \label{c4ss1} Let $2\leq j\leq 7$ and $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ Set $$A_{\alpha,\beta}^{(j)}:=\frac{A^{(j)}}{\langle \Omega_1-\alpha,\Omega_2-\beta\rangle},$$ where $A^{(j)}$ is defined in Section \ref{c2s1} and, $\Omega_1$ and $\Omega_2$ are the generators of the center of $A^{(j)}$, see Remark \ref{remark-centre-Ri}. Note in particular that $A_{\alpha,\beta}^{(7)}=A_{\alpha,\beta}.$ For each $2\leq j\leq 7,$ denote the canonical images of the generators $E_{i,j}$ of $A^{(j)}$ in $A_{\alpha,\beta}^{(j)}$ by $e_{i,j}$ for all $1\leq i\leq 6.$ As usual we denote by $t_i$ the canonical image of $T_i$ in $A_{\alpha,\beta}^{(2)}$ for each $1\leq i\leq 6.$ For each $3\leq j\leq 6,$ define $S_j:=\left\lbrace \lambda t_j^{i_j}t_{j+1}^{i_{j+1}}\cdots t_6^{i_6}\mid i_j,\cdots,i_6\in\mathbb{N} \ \text{and} \ \lambda\in {\mathbb K}^* \right\rbrace.$ One can observe that $S_j$ is a multiplicative system of non-zero divisors (or regular elements) of $A_{\alpha,\beta}^{(j)}.$ Furthermore; $t_j, \cdots, t_6$ are all normal elements of $A_{\alpha,\beta}^{(j)}$ and so $S_j$ is an Ore set in $A_{\alpha,\beta}^{(j)}$. One can localize $A_{\alpha,\beta}^{(j)}$ at $S_j$ as follows: $${\mathcal R}_j:=A_{\alpha,\beta}^{(j)}S_j^{-1}.$$ Let $3\leq j\leq 6$, and set $\Sigma_j:=\{t_j^k\mid k\in \mathbb{N}\}.$ By \cite[Lemme 5.3.2]{ca}, $\Sigma_j$ is an Ore set in both $A_{\alpha,\beta}^{(j)}$ and $A_{\alpha,\beta}^{(j+1)},$ and $$A_{\alpha,\beta}^{(j)}\Sigma_j^{-1}=A_{\alpha,\beta}^{(j+1)}\Sigma_j^{-1}.$$ As a consequence, similar to \eqref{emb0}, we have that \begin{align} \label{der} {\mathcal R}_j={\mathcal R}_{j+1}\Sigma_j^{-1}, \end{align} for all $2\leq j\leq 6.$ By convention, ${\mathcal R}_7:=A_{\alpha,\beta}.$ We also construct the following tower of algebras in a manner similar to \eqref{emb}: \begin{align} \label{c4ee1} {\mathcal R}_7=A_{\alpha,\beta}\subset {\mathcal R}_6={\mathcal R}_7\Sigma_6^{-1}\subset {\mathcal R}_5={\mathcal R}_6\Sigma_5^{-1}\subset {\mathcal R}_4={\mathcal R}_5\Sigma_4^{-1} \subset {\mathcal R}_3. \end{align} Note, ${\mathcal R}_3=A_{\alpha,\beta}^{(3)}S_3^{-1}={\mathcal R}_4\Sigma_3^{-1}$ is the quantum torus $\mathscr{A}_{\alpha,\beta}={\mathbb K}_{q^N}[t_3^{\pm1},t_4^{\pm1},t_5^{\pm1},t_6^{\pm 1}]$ studied in Section \ref{c3p3}. Our strategy to compute the derivations of ${\mathcal R}_7$ is to extend these derivations to derivations of the quantum torus ${\mathcal R}_3$. Then we can use the description of the derivations of a quantum torus obtained by Osborn and Passman in \cite{op}. Once this is done, we will have a ``nice'' description but involving elements of ${\mathcal R}_3$ and we will then use the fact that these derivations fix (globally) all ${\mathcal R}_i$ to obtain a description only involving elements of ${\mathcal R}_7$. This is a step by step process requiring knowing linear bases for ${\mathcal R}_i$. We find such bases in the next section. Before doing so, we note from \cite[Lemme 5.3.2]{ca} that DDA theory predicts the following relations between the elements $e_{i,j}$: \begin{align*} e_{1,6}&=e_1+re_5e_6^{-1}\\ e_{2,6}&=e_2+te_3e_5e_6^{-1}+ue_4e_6^{-1}+ne_5^3e_6^{-2}\\ e_{3,6}&=e_3+se_5^2e_6^{-1}\\ e_{4,6}&=e_4+be_5^3e_6^{-1}\\ e_{1,5}&=e_{1,6}+he_{3,6}e_{5,6}^{-1}+ge_{4,6}e_{5,6}^{-2}\\ e_{2,5}&=e_{2,6}+fe_{3,6}^2e_{5,6}^{-1}+pe_{3,6}e_{4,6}e_{5,6}^{-2}+ee_{4,6}^2e_{5,6}^{-3 }\\ e_{3,5}&=e_{3,6}+ae_{4,6}e_{5,6}^{-1}\\ e_{1,4}&=e_{1,5}+se_{3,5}^2e_{4,5}^{-1}\\ e_{2,4}&=e_{2,5}+be_{3,5}^3e_{4,5}^{-1}\\ e_{1,3}&=e_{1,4}+ae_{2,4}e_{3,4}^{-1}\\ t_1&:=e_{1,2}=e_{1,3}\\ t_2&:=e_{2,2}=e_{2,3}=e_{2,4}\\ t_3&:=e_{3,2}=e_{3,3}=e_{3,4}=e_{3,5}\\ t_4&:=e_{4,2}=e_{4,3}=e_{4,4}=e_{4,5}=e_{4,6}\\ t_5&:=e_{5,2}=e_{5,3}=e_{5,4}=e_{5,5}=e_{5,6}=e_5\\ t_6&:=e_{6,2}=e_{6,3}=e_{6,4}=e_{6,5}=e_{6,6}=e_6, \end{align*} where, as usual, the necessary parameters can be found in Appendix \ref{appenc}. We also note that we have complete control over the centers of the algebras ${\mathcal R}_i$. \begin{lemma} \label{ev25} Let $Z({\mathcal R}_i)$ denote the center of ${\mathcal R}_i,$ then $Z({\mathcal R}_i)={\mathbb K}$ for each $3\leq i\leq 7.$ \end{lemma} \begin{proof} One can easily verify that $Z({\mathcal R}_3)={\mathbb K}.$ Note, ${\mathcal R}_7=A_{\alpha,\beta}.$ Since ${\mathcal R}_i$ is a localization of ${\mathcal R}_{i+1}$ (see \eqref{der}), we have that ${\mathbb K}\subseteq Z({\mathcal R}_7)\subseteq Z({\mathcal R}_6)\subseteq Z({\mathcal R}_5)\subseteq Z({\mathcal R}_4)\subseteq Z({\mathcal R}_3)={\mathbb K}.$ Therefore, $Z({\mathcal R}_7)=Z({\mathcal R}_6)=Z({\mathcal R}_5)=Z({\mathcal R}_4)=Z({\mathcal R}_3)={\mathbb K}.$ \end{proof} \subsection{Linear bases for ${\mathcal R}_3$, ${\mathcal R}_4$ and ${\mathcal R}_5$} Let $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\}.$ We aim to find a basis of ${\mathcal R}_j$ for each $j=3,4,5.$ Since ${\mathcal R}_3=\mathscr{A}_{\alpha,\beta},$ the set $\{t_3^{i}t_4^{j}t_5^{k}t_6^{l}\mid i,j,k,l \in \mathbb{Z}\}$ is a ${\mathbb K}-$basis of $R_3.$ {For simplicity,} we set \begin{align*} f_1:&=e_{1,4} & F_1:&=E_{1,4}\\ z_1:&=e_{1,5} & Z_1:&=E_{1,5} \\ z_2:&=e_{2,5} & Z_2:&=E_{2,5}. \end{align*} \textbf{Basis for ${\mathcal R}_4$.} Observe that $$A_{\alpha,\beta}^{(4)}=\frac{A^{(4)}}{\langle \Omega_{1}-\alpha, \Omega_{2}-\beta\rangle},$$ where $\Omega_{1}=F_1T_3T_5+aT_2T_5$ and $\Omega_{2}=T_2T_4T_6$ in $A^{(4)}$. Recall from Section \ref{c3s2} that finding a basis for the algebra $A_\beta$ served as a good ground for finding a basis for $A_{\alpha,\beta}.$ In a similar manner, to find a basis for ${\mathcal R}_4,$ we will first and foremost find a basis for the algebra $$A_\beta^{(4)}S_4^{-1}=\frac{A^{(4)}S_4^{-1}}{\langle \Omega_{2}-\beta \rangle}=\frac{A^{(4)}S_4^{-1}}{\langle T_2T_4T_6-\beta \rangle},$$ where $\beta\in {\mathbb K}^*$. We will denote the canonical images of $E_{i,4}$ (resp. $T_i$) in ${A}_{\beta}^{(4)}$ by $\widehat{e_{i,4}}$ (resp. $\widehat{t_i}$) for all $1\leq i\leq 6.$ Observe that $\widehat{t_2}=\beta \widehat{t_6}^{-1}\widehat{t_4}^{-1}$ in $A_\beta^{(4)}S_4^{-1}.$ Note, when $\beta=0,$ then one can easily deduce that $A_\beta^{(4)}S_4^{-1}={A^{(4)}S_4^{-1}}/{\langle T_2 \rangle}$, hence, $\widehat{t_2}=0.$ \begin{proposition} \label{c4p2} The set $\mathfrak{S}_4=\left\lbrace \widehat{f_1}^{i_1}\widehat{t_3}^{i_3} \widehat{t_4}^{i_4}\widehat{t_5}^{i_5} \widehat{t_6}^{i_6}\mid (i_1,i_3,i_4,i_5,i_6)\in \mathbb{N}^2\times \mathbb{Z}^3 \right\rbrace $ is a ${\mathbb K}-$basis of $A_\beta^{(4)}S_4^{-1},$ where $\beta\in {\mathbb K}.$ \end{proposition} \begin{proof} We begin by showing that $\mathfrak{S}_4$ is a spanning set for $A_\beta^{(4)}S_4^{-1}.$ It is sufficient to do this by showing that $\widehat{f_1}^{k_1}\widehat{t_2}^{k_2}\widehat{t_3}^{k_3} \widehat{t_4}^{k_4}\widehat{t_5}^{k_5} \widehat{t_6}^{k_6}$ can be written as a finite linear combination of the elements of $\mathfrak{S}_4$ for all $(k_1,\cdots,k_6)\in \mathbb{N}^3\times \mathbb{Z}^3.$ This can easily be done through an induction on $k_2$ using the fact that $\widehat{t_2}=\beta \widehat{t_6}^{-1}\widehat{t_4}^{-1}$ (note that, if $\beta=0$, then $\widehat{t_2}=0$). We now prove that $\mathfrak{S}_4$ is a linearly independent set. Suppose that $$\sum_{\underline{i}\in I} a_{\underline{i}}\widehat{f_1}^{i_1}\widehat{t_3}^{i_3} \widehat{t_4}^{i_4}\widehat{t_5}^{i_5} \widehat{t_6}^{i_6}=0.$$ This implies that $$\sum_{\underline{i}\in I} a_{\underline{i}}F_{1}^{i_1}T_3^{i_3}T_4^{i_4}T_5^{i_5}T_6^{i_6}=(\Omega_{2}-\beta) \nu,$$ for some $\nu \in A^{(4)}S_4^{-1}.$ Write $\nu=\displaystyle\sum_{\underline{j}\in J}b_{\underline{j}}F_{1}^{i_1}T_2^{i_2}T_3^{i_3}T_4^{i_4}T_5^{i_5}T_6^{i_6},$ where $\underline{j}=(i_1,i_2,i_3,i_4,i_5,i_6) \in J\subset \mathbb{N}^3\times\mathbb{Z}^3$ and $b_{\underline{j}}$ is a family of scalars. Given that $\Omega_2=T_2T_4T_6,$ it follows from the above equality that \begin{align*} \sum_{\underline{i}\in I} a_{\underline{i}}F_{1}^{i_1}T_3^{i_3}T_4^{i_4}T_5^{i_5}T_6^{i_6}=& \sum_{\underline{j}\in J}q^\bullet b_{\underline{j}} F_{1}^{i_1}T_2^{i_2+1}T_3^{i_3}T_4^{i_4+1}T_5^{i_5}T_6^{i_6+1}-\sum_{\underline{j}\in J}\beta b_{\underline{j}} F_{1}^{i_1}T_2^{i_2}T_3^{i_3}T_4^{i_4}T_5^{i_5}T_6^{i_6}. \end{align*} We denote by $<_2$ the total order on $\mathbb{Z}^6$ defined by $(i_1,i_2,i_3,i_4,i_5,i_6)<_2(w_1,w_2,w_3,w_4,w_5,w_6)$ if $[w_2>i_2]$ or $[w_2=i_2, \ w_1>i_1]$ or $[w_2=i_2, \ w_1=i_1, \ w_3>i_3]$ or $\cdots$ or [$w_l=i_l, \ w_6\geq t_6, \ l=2,1,3,4,5]$. Suppose that there exists $(i_1,\cdots,i_6)\in J$ such that $b_{(i_1,\cdots,i_6)}\neq 0.$ Let $(w_1,\cdots,w_6)\in J$ be the greatest element of $J$ with respect to $<_2$ such that $b_{(w_1,\cdots,w_6)}\neq 0.$ Note, $\left( {F}_{1}^{i_1}{T}_2^{i_2}{T}_3^{i_3}T_4^{i_4}T_5^{i_5}T_6^{i_6}\right) _{(i_1,\cdots,i_6)\in J}$ is a basis of $A^{(4)}S_4^{-1}.$ Identifying the coefficients of ${F}_{1}^{w_1}{T}_2^{w_2+1}{T}_3^{w_3}{T}_4^{w_4+1}{T}_5^{w_5}{T_6^{w_6+1}},$ we have that $b_{(w_1,\cdots,w_6)}=0.$ This is a contradiction to our assumption, hence $b_{(i_1,\cdots,i_6)}= 0$ for all $(i_1,\cdots, i_6)\in J.$ This implies that $$\sum_{\underline{i}\in I} a_{\underline{i}}F_{1}^{i_1}T_3^{i_3}T_4^{i_4}T_5^{i_5}T_6^{i_6}=0.$$ Consequently, $a_{\underline{i}}=0$ for all $\underline{i}\in I.$ Therefore, $\mathfrak{S}_4$ is a linearly independent set. \end{proof} In ${\mathcal R}_4=A_{\alpha,\beta}^{(4)}S_4^{-1},$ we have the following two relations: $f_1t_3t_5+at_2t_5=\alpha$ and $t_2t_4t_6=\beta.$ This implies that $f_{1}t_3=\alpha t_5^{-1}-at_2$ and $t_2=\beta t_6^{-1}t_4^{-1}.$ Putting these two relations together, we have that \begin{align} f_{1}t_3= \alpha t_5^{-1}-a\beta t_6^{-1}t_4^{-1}\label{de1}. \end{align} Note, we will usually identify ${\mathcal R}_4$ with ${A_{\beta}^{(4)}S_4^{-1}}/{\langle \widehat{\Omega}_{1}-\alpha\rangle}.$ \begin{proposition} \label{c4p1} The set $\mathcal{B}_4=\left\lbrace f_{1}^{i_1}t_4^{i_4}t_5^{i_5} t_6^{i_6}, t_3^{i_3}t_4^{i_4}t_5^{i_5} t_6^{i_6}\mid i_1,i_3\in \mathbb{N} \ \text{and} \ i_4, i_5, i_6\in \mathbb{Z}\right\rbrace$ is a ${\mathbb K}-$basis of ${\mathcal R}_4.$ \end{proposition} \begin{proof} Since $\left( \widehat{f_{1}}^{k_1}\widehat{t_3}^{k_3} \widehat{t_4}^{k_4}\widehat{t_5}^{k_5}\widehat{t_6}^{k_6}\right)_{(k_1,k_3,\cdots,k_6)\in \mathbb{N}^2\times \mathbb{Z}^3}$ is a basis of ${A_{\beta}^{(4)}S_4^{-1}}$ (Proposition \ref{c4p2}), the set $\left( f_{1}^{k_1}t_3^{k_3} t_4^{k_4}t_5^{k_5}t_6^{k_6}\right)_{(k_1,k_3,\cdots,k_6)\in \mathbb{N}^2\times \mathbb{Z}^3}$ spans ${\mathcal R}_4.$ We show that $\mathcal{B}_4$ is a spanning set of ${\mathcal R}_4$ by showing that $f_{1}^{k_1}t_3^{k_3} t_4^{k_4}t_5^{k_5}t_6^{k_6}$ can be written as a finite linear combination of the elements of $\mathcal{B}_4$ for all $(k_1,k_3,\cdots,k_6)\in \mathbb{N}^2\times \mathbb{Z}^3.$ By Proposition \ref{c4p2}, it is sufficient to do this by induction on $k_1.$ The result is clear when $k_1=0.$ Assume that the statement is true for $k_1\geq 0.$ That is, $$f_{1}^{k_1}t_3^{k_3}t_4^{k_4}t_5^{k_5}t_6^{k_6} =\sum_{\underline{i}\in I_1}{a}_{\underline{i}}f_{1}^{i_1}t_4^{i_4}t_5^{i_5}t_6^{i_6}+\sum_{\underline{j}\in I_2}b_{\underline{j}} t_3^{i_3}t_4^{i_4}t_5^{i_5}t_6^{i_6},$$ where $\underline{i}=(i_{1},i_{4},i_{5},i_{6})\in I_1\subset \mathbb{N}\times\mathbb{Z}^3$ and $\underline{j}=(i_{3},i_{4},i_{5},i_{6})\in I_2\subset \mathbb{N}\times\mathbb{Z}^3.$ Note, $a_{\underline{i}}$ and $b_{\underline{j}}$ are all scalars. It follows that \begin{align*} f_{1}^{k_1+1}t_3^{k_3}t_4^{k_4}t_5^{k_5}t_6^{k_6}&= f_{1}\left( f_{1}^{k_1}t_3^{k_3}t_4^{k_4}t_5^{k_5}t_6^{k_6}\right) =\sum_{\underline{i}\in I_1}{a}_{\underline{i}}f_{1}^{i_1+1}t_4^{i_4}t_5^{i_5}t_6^{i_6}+\sum_{\underline{j}\in I_2}b_{\underline{j}} f_{1}t_3^{i_3}t_4^{i_4}t_5^{i_5}t_6^{i_6}. \end{align*} Clearly, the monomial $f_{1}^{i_1+1}t_4^{i_4}t_5^{i_5}t_6^{i_6}\in$ Span($\mathcal{B}_4$). We have to also show that $f_{1}t_3^{i_3}t_4^{i_4}t_5^{i_5}t_6^{i_6}\in$ Span($\mathcal{B}_4$) for all $i_3\in \mathbb{N}$ and $i_4,i_5,i_6\in\mathbb{Z}.$ This can easily be achieved by an induction on $i_3,$ and the use of the relation $f_1t_3=\alpha t_5^{-1}-a\beta t_6^{-1}t_4^{-1}.$ Therefore, by the principle of mathematical induction, $\mathcal{B}_4$ is a spanning set of ${\mathcal R}_4$ over ${\mathbb K}.$\\ We prove that $\mathcal{B}_4$ is a linearly independent set. Suppose that $$\sum_{\underline{i}\in I_1}a_{\underline{i}}f_{1}^{i_1}t_4^{i_4}t_5^{i_5}t_6^{i_6} +\sum_{\underline{j}\in I_2}b_{\underline{j}}t_3^{i_3}t_4^{i_4}t_5^{i_5}t_6^{i_6}=0.$$ It follows that there exists $\nu\in A_\beta^{(4)}S_4^{-1}$ such that $$\displaystyle\sum_{\underline{i}\in I_1}a_{\underline{i}}\widehat{f_{1}}^{i_1}\widehat{t_4}^{i_4} \widehat{t_5}^{i_5}\widehat{t_6}^{i_6} +\displaystyle\sum_{\underline{j}\in I_2}b_{\underline{j}}\widehat{t_3}^{i_3}\widehat{t_4}^{i_4} \widehat{t_5}^{i_5}\widehat{t_6}^{i_6} =\left( \widehat{\Omega}_{1}-\alpha\right) \nu.$$ Write $\nu=\displaystyle\sum_{\underline{l}\in J}c_{\underline{l}}\widehat{f_{1}}^{i_1}\widehat{t_3}^{i_3}\widehat{t_4}^{i_4} \widehat{t_5}^{i_5}\widehat{t_6}^{i_6},$ where $\underline{l}=(i_1,i_3,i_4,i_5,i_6)\in J\subset\mathbb{N}^2\times\mathbb{Z}^3$ and $c_{\underline{l}}\in {\mathbb K}.$ Note, $\widehat{t_2}=\beta\widehat{t_6}^{-1}\widehat{t_4}^{-1}.$ We have that $\widehat{\Omega}_1= \widehat{f_{1}}\widehat{t_3}\widehat{t_5}+a\widehat{t_2}\widehat{t_5}=\widehat{f_{1}}\widehat{t_3}\widehat{t_5}+a\beta\widehat{t_6}^{-1}\widehat{t_4}^{-1}\widehat{t_5}.$ Therefore, \begin{align*} \displaystyle\sum_{\underline{i}\in I_1}a_{\underline{i}}\widehat{f_{1}}^{i_1}\widehat{t_4}^{i_4} \widehat{t_5}^{i_5}\widehat{t_6}^{i_6} +\displaystyle\sum_{\underline{j}\in I_2}b_{\underline{j}}\widehat{t_3}^{i_3}\widehat{t_4}^{i_4} \widehat{t_5}^{i_5}\widehat{t_6}^{i_6}=&\displaystyle\sum_{\underline{l}\in J}q^\bullet c_{\underline{l}} \widehat{f_{1}}^{i_1+1}\widehat{t_3}^{i_3+1} \widehat{t_4}^{i_4} \widehat{t_5}^{i_5+1}\widehat{t_6}^{i_6}\\ &+\sum_{\underline{l}\in J}q^\bullet\beta ac_{\underline{l}}\widehat{f_{1}}^{i_1} \widehat{t_3}^{i_3}\widehat{t_4}^{i_4-1} \widehat{t_5}^{i_5+1}\widehat{t_6}^{i_6-1}\\ &-\displaystyle\sum_{\underline{l}\in J}\alpha c_{\underline{l}}\widehat{f_{1}}^{i_1}\widehat{t_3}^{i_3} \widehat{t_4}^{i_4} \widehat{t_5}^{i_5}\widehat{t_6}^{i_6}. \end{align*} Suppose that there exists $(i_1,i_3,i_4,i_5,i_6)\in J$ such that $c_{(i_1,i_3,i_4,i_5,i_6)}\neq 0.$\newline Let $(w_1,w_3,w_4,w_5,w_6)\in J$ be the greatest element (in the lexicographic order on $\mathbb{N}^2\times \mathbb{Z}^3$) of $J$ such that $c_{(w_1,w_3,w_4,w_5,w_6)}\neq 0.$ Since $\left( \widehat{f_{1}}^{k_1}\widehat{t_3}^{k_3} \widehat{t_4}^{k_4}\widehat{t_5}^{k_5}\widehat{t_6}^{k_6}\right) _{(k_1,k_3,\cdots,k_6)\in \mathbb{N}^2\times \mathbb{Z}^3}$ is a basis of $A^{(4)}S_4^{-1},$ it implies that the coefficients of $\widehat{f_{1}}^{w_1+1}\widehat{t_3}^{w_3+1}\widehat{t_4}^{w_4}\widehat{t_5}^{w_5+1} \widehat{t_6}^{w_6}$ in the above equality can be identified as: $q^\bullet c_{(w_1,w_3,w_4,w_5,w_6)}=0$. Hence, $c_{(w_1,w_3,w_4,w_5,w_6)}=0,$ a contradiction! Therefore, $c_{(i_1,i_3,i_4,i_5,i_6)}= 0$ for all $(i_1,i_3,i_4,i_5,i_6)\in J.$ This further implies that $$\displaystyle\sum_{\underline{i}\in I_1}a_{ \underline{i}}\widehat{f_{1}}^{i_1}\widehat{t_4}^{i_4} \widehat{t_5}^{i_5}\widehat{t_6}^{i_6} +\displaystyle\sum_{\underline{j}\in I_2}b_{\underline{j}}\widehat{t_3}^{i_3}\widehat{t_4}^{i_4} \widehat{t_5}^{i_5}\widehat{t_6}^{i_6} =0.$$ It follows from the previous proposition that $a_{\underline{i}}$ and $ b_{\underline{j}}$ are all zero. In conclusion, $\mathcal{B}_4$ is a linearly independent set. \end{proof} \textbf{Basis for ${\mathcal R}_5.$} We will identify ${\mathcal R}_5$ with ${A_{\alpha}^{(5)}S_5^{-1}}/{\langle \widehat{\Omega}_{2}-\beta\rangle},$ where $A_\alpha^{(5)}S_5^{-1}$ $=\dfrac{{A^{(5)}S_5^{-1}}}{{\langle \Omega_{1}-\alpha \rangle}}.$ Note, the canonical images of $E_{i,j}$ (resp. $T_i$) in $A_{\alpha}^{(5)}$ will be denoted by $\widehat{e_{i,j}}$ (resp. $\widehat{t_i}$). We now find a basis for $A_\alpha^{(5)}S_5^{-1}.$ Recall that $\Omega_{1}=Z_1T_3T_5+a Z_2T_5$ and $\Omega_2=Z_2T_4T_6+bT_3^3T_6$ in $A^{(5)}$ (remember, $Z_1:=E_{1,5}$ and $Z_2:=E_{2,5}$). Since $z_2t_4t_6+bt_3^3t_6=\beta$ and $\widehat{z_1}\widehat{t_3}\widehat{t_5}+a\widehat{z_2}\widehat{t_5}=\alpha$ in ${\mathcal R}_5$ and $A_\alpha^{(5)}S_5^{-1}$ respectively, we have the relation $\widehat{z_{2}}= \dfrac{1}{a}\left( \alpha\widehat{t_5}^{-1}-\widehat{z_{1}}\widehat{t_3}\right) $ in $A_\alpha^{(5)}S_5^{-1}$ and, in ${\mathcal R}_5,$ we have the following two relations: \begin{align} z_{2}&=\dfrac{1}{a}\left( \alpha t_5^{-1}-z_{1}t_3\right). \label{de2} \\ t_3^3&=\dfrac{1}{b}\left( \beta t_6^{-1}-z_{2}t_4\right) =\dfrac{\beta}{b} t_6^{-1}-\dfrac{q^3\alpha}{ab}t_4t_5^{-1}+\dfrac{1}{ab} z_{1}t_3t_4.\label{de3} \end{align} \begin{proposition} \label{c4p3} The set $\mathfrak{S}_5=\left\lbrace \widehat{z_{1}}^{i_1}\widehat{t_3}^{i_3}\widehat{t_4}^{i_4}\widehat{t_5 }^{i_5}\widehat{t_6}^{i_6}\mid {(i_1,i_3,\cdots,i_6)\in \mathbb{N}^3\times \mathbb{Z}^2}\right\rbrace $ is a ${\mathbb K}-$basis of $A_\alpha^{(5)}S_5^{-1},$ where $\alpha\in {\mathbb K}.$ \end{proposition} \begin{proof} The proof is similar to that of Proposition \ref{c4p2} and so is left to the reader. Details can be found in \cite{io-thesis}. \end{proof} \begin{proposition} \label{c4p20} The set $\mathcal{B}_5=\left\lbrace z_{1}^{i_1}t_3^{\xi}t_4^{i_4}t_5^{i_5}t_6^{i_6} \mid (\xi, i_1,i_4,i_5,i_6)\in \{0,1,2\}\times \mathbb{N}^2\times\mathbb{Z}^2\right\rbrace $ is a ${\mathbb K}-$basis of ${\mathcal R}_5$. \end{proposition} \begin{proof} The proof is similar to that of Proposition \ref{c4p1} and so is left to the reader. Details can be found in \cite{io-thesis}. \end{proof} We note for future reference the following immediate corollary. \begin{corollary} \label{c4c1} Let $I$ be a finite subset of $\{0,1,2\}\times \mathbb{N}\times\mathbb{Z}^3$ and $(a_{(\xi,\underline{i})})_{\underline{i}\in I}$ be a family of scalars. If $$\sum_{(\xi,\underline{i})\in I}a_{(\xi.\underline{i})} z_{1}^{i_1}t_3^\xi t_4^{i_4} t_5^{i_5}t_6^{i_6}=0, $$ then $a_{(\xi,\underline{i})}=0$ for all $(\xi, \underline{i})\in I.$ \end{corollary} \begin{remark} We were not successful in finding a basis for ${\mathcal R}_6.$ However, this has no effect on our main results in this section. Since ${\mathcal R}_7=A_{\alpha,\beta},$ we already have a basis for ${\mathcal R}_7$ (Proposition \ref{c3p5}). \end{remark} \subsection{Derivations of $A_{\alpha,\beta}$} We are now going to study the derivations of $A_{\alpha,\beta}.$ We will only treat the case when both $\alpha$ and $\beta$ are non-zero, and mention results when either $\alpha$ or $\beta$ is zero without details. Throughout this subsection, we assume that $\alpha$ and $\beta$ are non-zero. Let ${\rm Der}(A_{\alpha,\beta})$ denote the ${\mathbb K}-$derivations of $A_{\alpha,\beta}$ and $D\in {\rm Der}(A_{\alpha,\beta}).$ Via localization, $D$ extends uniquely to a derivation of each of the series of algebras in \eqref{c4ee1}. Therefore, $D$ extends to a derivation of the quantum torus ${\mathcal R}_3={\mathbb K}_{q^N}[t_3^{\pm1},t_4^{\pm1},t_5^{\pm1},t_6^{\pm 1}].$ It follows from \cite[Corollary 2.3]{op} that $D$ can uniquely be written as: $$D=\text{ad}_x+\delta,$$ where $x\in {\mathcal R}_3,$ and $\delta$ is a scalar derivation of ${\mathcal R}_3$ defined as $\delta(t_i)=\lambda_it_i$ for each $i=3,4,5,6.$ Note, $\lambda_i\in Z({\mathcal R}_3)={\mathbb K}.$ Also, $\text{ad}_x$ is an inner derivation of ${\mathcal R}_3$ defined as $\text{ad}_x(L)=xL-Lx$ for all $L\in {\mathcal R}_3.$ We aim to describe $D$ as a derivation of $A_{\alpha,\beta}={\mathcal R}_7.$ We do this in several steps. Before starting the process we note the following relations that will be used in this section. They all follow from \cite[Lemme 5.3.2]{ca}. \begin{remark} \label{rrr} Recall the notations: \begin{align*} f_1:&=e_{1,4} & F_1:&=E_{1,4}\\ z_1:&=e_{1,5} & Z_1:&=E_{1,5} \\ z_2:&=e_{2,5} & Z_2:&=E_{2,5}. \end{align*} Then \begin{align*} f_1&=t_1-at_2t_3^{-1} & e_{3,6}&=t_3-at_4t_5^{-1}\\ z_1&=f_1-st_3^2t_4^{-1} &e_1&=e_{1,6}-rt_5t_6^{-1}\\ z_2&=t_2-bt_3^3t_4^{-1}& e_3&=e_{3,6}-st_5^2t_6^{-1}\\ e_{1,6}&=z_1-he_{3,6}t_5^{-1}-gt_4t_5^{-2}& e_4&=t_4-bt_5^3t_6^{-1}. \end{align*} \end{remark} We first describe $D$ as a derivation of ${\mathcal R}_4.$ \begin{lemma} \begin{itemize}\label{c4l1} \item[1.] $x\in {\mathcal R}_4.$ \item[2.] $\lambda_5=\lambda_4+\lambda_6,$ $\delta(f_{1})=-(\lambda_3+\lambda_5)f_{1}$ and $\delta(t_2)=-\lambda_5t_2.$ \item[3.] Set $\lambda_1:=-(\lambda_3+\lambda_5)$ and $\lambda_2:=-\lambda_5.$ Then, $D(e_{\kappa,4})=\text{ad}_x(e_{\kappa,4})+\lambda_\kappa e_{\kappa,4}$ for all $\kappa\in \{1,\cdots, 6\}.$ \end{itemize} \end{lemma} \begin{proof} 1. Set $\mathcal{Q}_q:={\mathbb K}_{q^\mu}[t_4^{\pm 1},t_5^{\pm 1},t_6^{\pm 1}],$ where $\mu$ is the skew-symmetric sub-matrix of $N$ (see Section \ref{c3p3}) obtained by deleting the first row and first column of $N$. Observe that $\mathcal{Q}_q$ is a subalgebra of both ${\mathcal R}_3$ and ${\mathcal R}_4$ with central element $$z:=t_4t_5^{-1}t_6.$$ Furthermore, since ${\mathcal R}_3$ is a quantum torus, we can present it as a free left $\mathcal{Q}_q-$module with basis $(t_3^s)_{s\in\mathbb{Z}}.$ With this presentation, $x\in {\mathcal R}_3$ can be written as $$x=\sum_{s\in \mathbb{Z}}y_st_3^s,$$ where $y_s\in \mathcal{Q}_q.$ Set $$x_{+}:=\sum_{s\geq 0}y_st_3^s \ \ \text{and} \ \ x_{-}:=\sum_{s<0}y_st_3^s.$$ Clearly, $x=x_{+}+x_{-}.$ Obviously, $x_+\in {\mathcal R}_4,$ hence we aim to also show that $x_-$ belongs to ${\mathcal R}_4$ by following a pattern similar to \cite[Proposition 7.1.2]{ak}. As $D$ is a derivation of ${\mathcal R}_4$, we have that $D(z^j)\in {\mathcal R}_4$ for all $j\in \mathbb{N}_{\geq 1}.$ Now $D(z^j)=\text{ad}_{x}(z^j)+\delta(z^j)=\text{ad}_{x_+}(z^j)+\text{ad}_{x_-}(z^j)+\delta(z^j).$ Observe that $\text{ad}_{x_+}(z^j)\in {\mathcal R}_4$; since $x_+, z^j \in {\mathcal R}_4.$ Also, $\delta(z)=\delta(t_4t_5^{-1}t_6)=(\lambda_4-\lambda_5+\lambda_6)t_4t_5^{-1}t_6=(\lambda_4-\lambda_5+\lambda_6)z,$ where $\lambda_4,\lambda_5,\lambda_6\in {\mathbb K}.$ It follows that $\delta(z^j)=j(\lambda_4-\lambda_5+\lambda_6)z^j\in {\mathcal R}_4.$ We can therefore conclude that each $\text{ad}_{x_-}(z^j)$ belongs to ${\mathcal R}_4$ since $D(z^j), \text{ad}_{x_+}(z^j), \delta(z^j)\in {\mathcal R}_4.$ We have: $$\text{ad}_{x_-}(z^j)=x_-z^j-z^jx_-=\sum_{s=-1}^{-n}y_st_3^sz^j-\sum_{s=-1}^{-n}y_sz^jt_3^s.$$ One can verify that $zt_3=q^{-2}t_3z.$ Therefore, $$\text{ad}_{x_-}(z^j)=\sum_{s=-1}^{-n}(1-q^{-2js})y_st_3^sz^j, \ \text{hence,} \ \ \text{ad}_{x_-}(z^j)z^{-j}=\sum_{s=-1}^{-n}(1-q^{-2js})y_st_3^s.$$ Set $\nu_j:=\text{ad}_{x_-}(z^j)z^{-j}\in {\mathcal R}_4.$ It follows that $$\nu_j=\sum_{s=-1}^{-n}(1-q^{-2js})y_st_3^s,$$ for each $j\in \{1,\cdots,n\}.$ One can therefore write the above equality as a matrix equation as follows: $$ \begin{bmatrix} (1-q^2)&(1-q^4)&(1-q^6)&\cdots &(1-q^{2n})\\ (1-q^4)&(1-q^8)&(1-q^{12})&\cdots &(1-q^{4n})\\ (1-q^6)&(1-q^{12})&(1-q^{18})&\cdots &(1-q^{6n})\\ \vdots&\vdots&\vdots&\ddots&\vdots\\ (1-q^{2n})&(1-q^{4n})&(1-q^{6n})&\cdots &(1-q^{2n^2})\\ \end{bmatrix} \begin{bmatrix} y_{-1}t_3^{-1}\\ y_{-2}t_3^{-2}\\ y_{-3}t_3^{-3}\\ \vdots\\ y_{-n}t_3^{-n}\\ \end{bmatrix}= \begin{bmatrix} \nu_1\\ \nu_2\\ \nu_3\\ \vdots\\ \nu_n\\ \end{bmatrix}. $$ We already know that each $\nu_j$ belongs to ${\mathcal R}_4.$ We want to show that $y_st_3^s$ also belongs to ${\mathcal R}_4$ for each $s\in\{-1,\cdots, -n\}.$ To establish this, it is sufficient to show that the coefficient matrix of the above matrix equation is invertible. Let $U$ represent this matrix. Thus, $$ U=\begin{bmatrix} (1-q^2)&(1-q^4)&(1-q^6)&\cdots &(1-q^{2n})\\ (1-q^4)&(1-q^8)&(1-q^{12})&\cdots &(1-q^{4n})\\ (1-q^6)&(1-q^{12})&(1-q^{18})&\cdots &(1-q^{6n})\\ \vdots&\vdots&\vdots&\ddots&\vdots\\ (1-q^{2n})&(1-q^{4n})&(1-q^{6n})&\cdots &(1-q^{2n^2})\\ \end{bmatrix}. $$ Apply row operations: $ -r_{n-1}+r_n\rightarrow r_n,\cdots, -r_2+r_3\rightarrow r_3, -r_1+r_2\rightarrow r_2$ to $U$ to obtain: $$ U'=\begin{bmatrix} l_1&l_2&l_3&\cdots &l_n\\ q^2l_1&q^4l_2&q^6l_3&\cdots &q^{2n}l_n\\ q^4l_1&q^8l_2&q^{12}l_3&\cdots &q^{4n}l_n\\ \vdots&\vdots&\vdots&\ddots&\vdots\\ q^{2(n-1)}l_1&q^{4(n-1)}l_2&q^{6(n-1)}l_3&\cdots &q^{2n(n-1)}l_n\\ \end{bmatrix}, $$ where $l_i:=1-q^{2i};$ $i\in \{1,2,\cdots,n\}.$ Clearly, $U'$ is similar to a Vandermonde matrix (since the terms in each column form a geometric sequence) which is well known to be invertible when all parameters are pairwise distinct (this is the case here as $q$ is not a root of unity). This further implies that $U$ is invertible. So each $y_{s}t_3^{s}$ is a linear combination of the $\nu_j\in {\mathcal R}_4.$ As a result, $y_{s}t_3^{s}\in {\mathcal R}_4$ for all $s\in \{-1,\cdots, -n\}.$ Consequently, $x_{-}=\sum_{s=-1}^{-n}y_st_3^s\in {\mathcal R}_4$ as desired. 2. Recall that $\delta (t_\kappa)=\lambda_\kappa t_\kappa$ for all $\kappa\in \{3,4,5,6\}$ and $\lambda_\kappa\in {\mathbb K}.$ From Remark \ref{rrr}, we have that $f_{1}=t_1-at_2t_3^{-1}.$ Recall from Section \ref{c3p3} that $t_1=\alpha t_5^{-1}t_3^{-1}$ and $t_2^{-1}=\beta t_6^{-1}t_4^{-1}$ in ${\mathcal R}_3=\mathscr{A}_{\alpha,\beta}.$ As a result, $f_1=\alpha t_5^{-1}t_3^{-1}-a\beta t_6^{-1}t_4^{-1}t_3^{-1}.$ Hence, \begin{align} \delta(f_{1}) =&-(\lambda_5+\lambda_3)\alpha t_5^{-1}t_3^{-1}+(\lambda_6+\lambda_4+\lambda_3)a\beta t_6^{-1}t_4^{-1}t_3^{-1}. \label{c4e2} \end{align} From Proposition \ref{c4p1}, the set $\mathcal{B}_4=\left\lbrace f_{1}^{i_1}t_4^{i_4}t_5^{i_5} t_6^{i_6}, t_3^{i_3}t_4^{i_4}t_5^{i_5} t_6^{i_6}\mid i_1,i_3\in \mathbb{N} \ \text{and} \ i_4, i_5, i_6\in \mathbb{Z}\right\rbrace$ is a ${\mathbb K}-$basis of ${\mathcal R}_4.$ Since $t_4, t_5$ and $t_6$ $q$-commute with $f_1$ and $t_3,$ one can also write $\delta(f_{1})\in {\mathcal R}_4$ in terms of $\mathcal{B}_4$ as follows: \begin{align} \label{de4} \delta(f_{1})=&\displaystyle\sum_{r> 0}a_rf_{1}^r+\displaystyle\sum_{s\geq 0}b_st_3^s, \end{align} where $a_r$ and $b_s$ belong to $\mathcal{Q}_q={\mathbb K}_{q^\mu}[t_4^{\pm 1},t_5^{\pm 1},t_6^{\pm 1}].$ \begin{align} f_{1}^r&=(\alpha t_5^{-1}t_3^{-1}-a\beta t_6^{-1}t_4^{-1}t_3^{-1})^r=\sum_{i=0}^r {r\choose i}_{q^\bullet}(\alpha t_5^{-1}t_3^{-1})^i(-a\beta t_6^{-1}t_4^{-1}t_3^{-1})^{r-i} \nonumber\\ &=\sum_{i=0}^r {r\choose i}_{q^\bullet}\alpha^i(-a\beta)^{r-i}q^{\frac{1}{2}i(i-1)+\frac{3}{2}(r-i)(r-i-1)+3i(i-r)}t_5^{-i}(t_6^{-1}t_4^{-1})^{r-i}t_3^{-r}\nonumber\\ &=c_rt_3^{-r},\label{c4e23} \end{align} where \begin{align} c_r=\displaystyle\sum_{i=0}^r {r\choose i}_{q^\bullet}q^{\frac{1}{2}i(i-1)+\frac{3}{2}(r-i)(r-i-1)+3i(i-r)}\alpha^i(-a\beta)^{r-i} t_5^{-i}(t_6^{-1}t_4^{-1})^{r-i} \in \mathcal{Q}_q\setminus \{0\}.\label{c4e24} \end{align} Substitute \eqref{c4e23} into \eqref{de4} to obtain; \begin{align} \delta(f_{1})=&\displaystyle\sum_{r> 0}a_rc_rt_3^{-r}+\displaystyle\sum_{s\geq 0}b_st_3^s.\label{c4e3} \end{align} One can rewrite \eqref{c4e2} as \begin{equation} \delta (f_{1})=dt_3^{-1}, \label{c4e4} \end{equation} where $d=-(\lambda_5+\lambda_3)\alpha t_5^{-1}+(\lambda_6+\lambda_4+\lambda_3)a\beta t_6^{-1}t_4^{-1}\in\mathcal{Q}_q.$ Comparing \eqref{c4e3} to \eqref{c4e4} shows that $b_s=0$ for all $s\geq 0,$ and $a_rc_r=0$ for all $r\neq 1$. Therefore $\delta(f_{1})=a_1c_1t_3^{-1}.$ Moreover, from \eqref{c4e24}, $c_1=-a\beta t_6^{-1}t_4^{-1}+\alpha t_5^{-1}.$ Hence, \begin{align} \delta(f_{1})=a_1c_1t_3^{-1}&=a_1(-a\beta t_6^{-1}t_4^{-1}+\alpha t_5^{-1})t_3^{-1} =a_1\alpha t_5^{-1}t_3^{-1}-a_1a\beta t_6^{-1}t_4^{-1}t_3^{-1}.\label{c4e25} \end{align} Comparing \eqref{c4e25} to \eqref{c4e2} reveals that $a_1=-(\lambda_5+\lambda_3)=-(\lambda_6+\lambda_4+\lambda_3).$ Consequently, $\lambda_5=\lambda_6+\lambda_4.$ Hence, $\delta(f_{1})=-(\lambda_5+\lambda_3)\alpha t_5^{-1} t_3^{-1}+(\lambda_5+\lambda_3)a\beta t_6^{-1} t_4^{-1}t_3^{-1}=-(\lambda_5+\lambda_3)f_{1}.$ Finally, since $t_2=\beta t_6^{-1}t_4^{-1}$ in ${\mathcal R}_4,$ it follows that $\delta(t_2)= -(\lambda_6+\lambda_4)\beta t_6^{-1}t_4^{-1}=-(\lambda_6+\lambda_4) t_2=-\lambda_5t_2.$ 3. Set $\lambda_1:=-(\lambda_3+\lambda_5)$ and $\lambda_2:=-\lambda_5.$ it follows from points (1) and (2) that $D(e_{\kappa,4})=\text{ad}_x(e_{\kappa,4})+\delta(e_{\kappa,4})=\text{ad}_x(e_{\kappa,4}) +\lambda_\kappa e_{\kappa,4}$ for all $\kappa\in \{1,\cdots, 6\}.$ In conclusion, $D=\text{ad}_x+\delta,$ with $x\in {\mathcal R}_4.$ \end{proof} We proceed to describe $D$ as a derivation of ${\mathcal R}_5.$ \begin{lemma} \label{ev5} \begin{itemize} \item[1.] $x\in {\mathcal R}_5.$ \item[2.] $\lambda_4=3\lambda_3+\lambda_5, \ \lambda_6=-3\lambda_3, \ \delta(z_{1})=-(\lambda_3+\lambda_5)z_{1}$ and $ \delta(z_{2})=-\lambda_5z_{2}.$ \item[3.] Set $\lambda_1:=-(\lambda_3+\lambda_5), \ \lambda_2:=-\lambda_5$ and $\lambda_6:=-3\lambda_3.$ Then, $D(e_{\kappa,5})=\text{ad}_x(e_{\kappa,5})+\lambda_\kappa e_{\kappa,5}$ for all $\kappa\in \{1,\cdots, 6\}.$ \end{itemize} \end{lemma} \begin{proof} In this proof, we denote $\underline{\upsilon}:=(i,j,k,l)\in \mathbb{N}\times \mathbb{Z}^3.$ 1. We already know that $x\in {\mathcal R}_4={\mathcal R}_5[t_4^{-1}].$ Given the basis $\mathcal{B}_5$ of ${\mathcal R}_5$ (Proposition \ref{c4p20}), $x$ can be written as $x=\displaystyle\sum_{(\xi, \underline{\upsilon})\in I}a_{(\xi,\underline{\upsilon})}z_{1}^{i}t_3^{\xi} t_4^{j}t_5^{k}t_6^{l},$ where $I$ is a finite subset of $\{0,1,2\}\times \mathbb{N}\times \mathbb{Z}^3$ and $a_{(\xi, \underline{\upsilon})}$ are scalars. Write $x=x_-+x_+,$ where $$x_+=\sum_{\substack{(\xi, \underline{\upsilon})\in I\\ j\geq 0}} a_{(\xi, {\underline{\upsilon})}}z_{1}^{i}t_3^{\xi} t_4^{j}t_5^{k}t_6^{l}\ \ \text{and} \ \ x_-=\displaystyle \sum_{\substack{(\xi, \underline{\upsilon})\in I \\ j<0}}a_{(\xi,\underline{\upsilon})}z_{1}^{i}t_3^{\xi} t_4^{j}t_5^{k}t_6^{l}.$$ Suppose that there exists a minimum $j_0<0$ such that $a_{(\xi, i,j_0,k,l)}\neq 0$ for some $(\xi,i,j_0,k,l)\in I$ and $a_{(\xi,i,j,k,l)}=0$ for all $(\xi,i,j_0,k,l)\in I$ with $j<j_0.$ Given this assumption, write \begin{align*} x_-=\displaystyle\sum_{ \substack{(\xi,\underline{\upsilon})\in I\\ j_0 \leq j\leq -1}}a_{(\xi,\underline{\upsilon})}z_{1}^{i}t_3^{\xi} t_4^{j}t_5^{k}t_6^{l}. \end{align*} Now, $D(t_6)=\text{ad}_{x_+}(t_6)+\text{ad}_{x_-}(t_6)+ \delta(t_6)\in {\mathcal R}_5.$ This implies that $\text{ad}_{x_-}(t_6)\in {\mathcal R}_5,$ since $\text{ad}_{x_+}(t_6)+\delta(t_6)=\text{ad}_{x_+}(t_6)+ \lambda_6t_6\in {\mathcal R}_5.$ We aim to show that $x_-=0.$ Since $t_6$ is normal in ${\mathcal R}_5,$ one can easily verify that \begin{align*} \text{ad}_{x_-}(t_6)&= \sum_{\substack{(\xi,\underline{\upsilon})\in I\\ j_0 \leq j\leq -1}}\left( q^{3(i-j-k)}-1\right) a_{(\xi, \underline{\upsilon})} z_{1}^{i}t_3^{\xi} t_4^{j}t_5^{k}t_6^{l+1}. \end{align*} Set $\underline{w}:=(i,j,k,l)\in \mathbb{N}^2\times \mathbb{Z}^2.$ One can equally write $\text{ad}_{x_-}(t_6)\in {\mathcal R}_5$ in terms of the basis $\mathcal{B}_5$ of ${\mathcal R}_5$ (Proposition \ref{c4p20}) as: \begin{align*} \text{ad}_{x_-}(t_6)&= \sum_{(\xi,\underline{w})\in J} b_{(\xi, {\underline{w}})}z_{1}^{i} t_3^{\xi}t_4^{j}t_5^{k}t_6^{l} \end{align*} where $ J$ is a finite subset of $\{0,1,2\}\times \mathbb{N}^2\times \mathbb{Z}^2$ and $b_{(\xi, \underline{w})}$ are all scalars. It follows that $$ \sum_{\substack{(\xi,\underline{\upsilon})\in I\\ j_0 \leq j\leq -1}}\left( q^{3(i-j-k)}-1\right) a_{(\xi, \underline{\upsilon})} z_{1}^{i}t_3^{\xi} t_4^{j}t_5^{k}t_6^{l+1}= \sum_{(\xi,\underline{w})\in J} b_{(\xi, {\underline{w}})}z_{1}^{i} t_3^{\xi}t_4^{j}t_5^{k}t_6^{l}. $$ As $\mathcal{B}_5$ is a basis for ${\mathcal R}_5,$ we deduce from Corollary \ref{c4c1} that $\left(z_{1}^{i} t_3^{\xi}t_4^{j}t_5^{k}t_6^{l}\right)_{(i\in \mathbb{N}; j,k,l\in \mathbb{Z}; \xi\in \{0,1,2\})}$ is a basis for ${\mathcal R}_5[t_4^{-1}].$ Now, at $j=j_0,$ denote $\underline{\upsilon}=(i,j,k,l)$ by $\underline{\upsilon}_0:= (i,j_0,k,l).$ Since $\underline{v}_0\in \mathbb{N}\times \mathbb{Z}^3$ (with $j_0<0$) and $\underline{w}=(i,j,k,l)\in \mathbb{N}^2\times \mathbb{Z}^2$ (with $j\geq 0$), it follows from the above equality that, at $\underline{\upsilon}_0,$ we must have $$\left( q^{3(i-j_0-k)}-1\right) a_{(\xi, \underline{\upsilon}_0)}=0.$$ From our initial assumption, the coefficients $a_{(\xi, \underline{\upsilon}_0)}$ are all not zero, therefore $q^{3(i-j_0-k)}-1=0. $ This implies that \begin{align} k=i-j_0,\label{2e} \end{align} for some $(\xi,\underline{\upsilon}_0)\in I.$ In a similar manner, $D(t_3)=\text{ad}_{x_+}(t_3)+\text{ad}_{x_-}(t_3)+ \delta(t_3)\in {\mathcal R}_5.$ This implies that $\text{ad}_{x_-}(t_3)\in {\mathcal R}_5,$ since $\text{ad}_{x_+}(t_3)+\delta(t_3)=\text{ad}_{x_+}(t_3)+ \lambda_3t_3\in {\mathcal R}_5.$ We have that \begin{align*} \text{ad}_{x_-}(t_3)= \sum_{\substack{(\xi,\underline{\upsilon})\in I\\ j_0\leq j\leq -1}}a_{(\xi,\underline{\upsilon})} z_{1}^{i}t_3^{\xi} t_4^{j}t_5^{k}t_6^{l}t_3- \sum_{\substack{(\xi,\underline{\upsilon})\in I\\ j_0 \leq j\leq -1}} a_{(\xi, \underline{\upsilon})} t_3z_{1}^{i}t_3^{\xi} t_4^{j}t_5^{k}t_6^{l}. \end{align*} One can deduce from Lemma \ref{c3l3}(3a) that $$t_3z_1^{i}=q^{-i}z_1^it_3+d_2[i]z_1^{i-1}z_2,$$ where $d_2[i]=q^{1-i}d_2[1]\left(\dfrac{1-q^{-2i}}{1-q^{-2}} \right),$ $d_2[1]=-(q+q^{-1}+q^{-3})$ and $d_2[0]=0.$ Therefore, the above expression for $ \text{ad}_{x_-}(t_3)$ can be expressed as: \begin{align*} \text{ad}_{x_-}(t_3)=&\sum_{\substack{(0,\underline{\upsilon})\in I\\ j_0 \leq j\leq -1}}f[i,j,k] a_{(0,\underline{\upsilon})}z_{1}^{i}t_3t_4^{j}t_5^{k} t_6^{l}+ \sum_{\substack{(1,\underline{\upsilon})\in I\\ j_0 \leq j\leq -1}}f[i,j,k] a_{(1,\underline{\upsilon})}z_{1}^{i}t_3^{2}t_4^{j}t_5^{k} t_6^{l} + \\ &+\sum_{\substack{(2,\underline{\upsilon})\in I\\ j_0 \leq j\leq -1}}f[i,j,k] a_{(2,\underline{\upsilon})}z_{1}^{i}t_3^{3}t_4^{j}t_5^{k} t_6^{l}-\sum_{\substack{(\xi,\underline{\upsilon})\in I\\ j_0 \leq j\leq -1}}a_{(\xi,\underline{\upsilon})}d_2[i]z_{1}^{i-1}z_{2}t_3^{\xi}t_4^{j}t_5^{k} t_6^{l}, \end{align*} where $f[i,j,k]:= q^{-(k+3j)}-q^{-i}.$ Recall from \eqref{de2} and \eqref{de3} that $$z_{2}=\dfrac{1}{a}\left( \alpha t_5^{-1}-z_{1}t_3\right) \ \ \text{and} \ \ t_3^3=\dfrac{\beta}{b} t_6^{-1}-\dfrac{q^3\alpha}{ab}t_4t_5^{-1}+\dfrac{1}{ab} z_{1}t_3t_4,$$ where $a$ and $b$ are non-zero scalars (Appendix \ref{appenc}). Using these two expressions, one can write $\text{ad}_{x_-}(t_3)$ in terms of the basis of ${\mathcal R}_5$ as: \begin{align} \text{ad}_{x_-}(t_3)=&\mathcal{K}+\sum_{\substack{(0,\underline{\upsilon}_0)\in I}} g[i,j_0,k]a_{(0,{\underline{\upsilon}_0})}z_{1}^{i}t_3 t_4^{j_0}t_5^{k}t_6^{l}+\sum_{\substack{(1,\underline{\upsilon}_0)\in I}} g[i,j_0,k]a_{(1,\underline{\upsilon}_0)}z_{1}^{i}t_3^{2} t_4^{j_0}t_5^{k}t_6^{l}\nonumber \\ &+ \sum_{\substack{(2,\underline{\upsilon}_0)\in I}}\frac{q^{\bullet}\beta}{b} a_{(2,{\underline{\upsilon}_0})}g[i,j_0,k] z_{1}^{i} t_4^{j_0}t_5^{k}t_6^{l-1} - \sum_{\substack{(\xi,\underline{\upsilon}_0)\in I}}\frac{q^{\bullet}\alpha}{a} d_2[i] a_{(\xi,\underline{\upsilon}_0)}z_{1}^{i-1}t_3^{\xi} t_4^{j_0}t_5^{k-1}t_6^{l}\nonumber\\ =&\sum 1/b\left( {q^{\bullet}\beta} g[i,j_0,k]a_{(2,i,j_0,k,l+1)}+(q^{\bullet}\alpha b{d_2[i+1]}/{a})a_{(0, i+1,j_0,k+1,l)}\right) z_{1}^it_4^{j_0}t_5^kt_6^l\nonumber\\ &+\sum \left( g[i,j_0,k]a_{(0,i,j_0,k,l)}+(q^{\bullet}\alpha {d_2[i+1]}/{a}) a_{(1,i+1,j_0,k+1,l)}\right) z_{1}^it_3t_4^{j_0}t_5^kt_6^l\nonumber\\ &+\sum \left( g[i,j_0,k]a_{(1,i,j_0,k,l)}+(q^{\bullet}\alpha {d_2[i+1]}/{a})a_{(2,i+1,j_0,k+1,l)}\right) z_{1}^it_3^2t_4^{j_0}t_5^kt_6^l+\mathcal{K} ,\label{3e} \end{align} where $g[i,j_0,k]:=q^{-(k+3j_0)}-q^{-i}+{d_2[i]}/{a}$ and $$\mathcal{K}\in \text{Span}\left( \mathcal{B}_5\setminus \{z_{1}^{i}t_3^{\xi} t_4^{j_0}t_5^{k}t_6^{l}\mid (\xi, i,j_0,k,l)\in \{0,1,2\}\times \mathbb{N}\times \mathbb{Z}^3 \}\right).$$ One can also write $\text{ad}_{x_-}(t_3)\in {\mathcal R}_5$ in terms of the basis $\mathcal{B}_5$ of ${\mathcal R}_5$ (Proposition \ref{c4p20}) as: \begin{align} \text{ad}_{x_-}(t_3)=& \sum_{(\xi,\underline{w})\in J}b_{(\xi, \underline{w})} z_{1}^{i} t_3^{\xi}t_4^{j}t_5^{k}t_6^{l},\label{1e} \end{align} where $J$ is a finite subset of $\{0,1,2\}\times \mathbb{N}^2\times \mathbb{Z}^2$, and $b_{(\xi, {\underline{w})}}\in {\mathbb K}.$ Recall: $\underline{w}=(i,j,k,l)\in \mathbb{N}^2\times \mathbb{Z}^2.$ Now, \eqref{3e} and \eqref{1e} imply that \begin{align*} \sum_{(\xi,\underline{w})\in J}b_{(\xi, \underline{w})} &z_{1}^{i} t_3^{\xi}t_4^{j}t_5^{k}t_6^{l}\\ =&\sum 1/b\left( {q^{\bullet}\beta} g[i,j_0,k]a_{(2,i,j_0,k,l+1)}+(q^{\bullet}\alpha b{d_2[i+1]}/{a})a_{(0, i+1,j_0,k+1,l)}\right) z_{1}^it_4^{j_0}t_5^kt_6^l\nonumber\\ &+\sum \left( g[i,j_0,k]a_{(0,i,j_0,k,l)}+(q^{\bullet}\alpha {d_2[i+1]}/{a}) a_{(1,i+1,j_0,k+1,l)}\right) z_{1}^it_3t_4^{j_0}t_5^kt_6^l\nonumber\\ &+\sum \left( g[i,j_0,k]a_{(1,i,j_0,k,l)}+(q^{\bullet}\alpha {d_2[i+1]}/{a})a_{(2,i+1,j_0,k+1,l)}\right) z_{1}^it_3^2t_4^{j_0}t_5^kt_6^l+\mathcal{K}. \end{align*} We have already established that $\left(z_{1}^{i} t_3^{\xi}t_4^{j}t_5^{k}t_6^{l}\right)_{(i\in \mathbb{N}; j,k,l\in \mathbb{Z}; \xi\in \{0,1,2\})}$ is a basis for ${\mathcal R}_5[t_4^{-1}].$ Given that $\underline{v}_0=(i,j_0,k,l)\in \mathbb{N}\times \mathbb{Z}^3$ (with $j_0<0$) and $\underline{w}=(i,j,k,l)\in \mathbb{N}^2\times \mathbb{Z}^2$ (with $j\geq 0$), it follows that \begin{align} q^{\bullet}\beta g[i,j_0,k]a_{(2,i,j_0,k,l+1)}+(q^{\bullet}\alpha b {d_2[i+1]}/{a})a_{(0,i+1,j_0,k+1,l)}&=0.\label{2ee}\\ g[i,j_0,k]a_{(0,i,j_0,k,l)}+(q^{\bullet}\alpha {d_2[i+1]}/{a})a_{(1,i+1,j_0,k+1,l)}&=0.\label{3ee}\\ g[i,j_0,k]a_{(1,i,j_0,k,l)}+(q^{\bullet}\alpha {d_2[i+1]}/{a})a_{(2,i+1,j_0,k+1,l)}&=0.\label{4e} \end{align} Suppose that there exists $(\xi, i,j_0,k,l)\in I$ such that $g[i,j_0,k]=0.$ Then, $$g[i,j_0,k]=q^{-(k+3j_0)}-q^{-i}+d_2[i]/a=0.$$ Note, $d_2[i]=d_2[1]q^{1-i}\left( \dfrac{1-q^{-2i}}{1-q^{-2}}\right),$ where $d_2[1]=-(q+q^{-1}+q^{-3})$ and $d_2[0]=0.$ Again, recall from Appendix \ref{appenc} that $a=(q^2+1+q^{-2})/(q^{-2}-1)=\dfrac{qd_2[1]}{1-q^{-2}}.$ Given these expressions for $d_2[i]$ and $a$, we have that $$g[i,j_0,k]=q^{-(k+3j_0)}-q^{-i}+d_2[i]/a=q^{-3j_0-k}-q^{-3i}=0.$$ Since $q$ is not a root of unity, we get \begin{align} k=3(i-j_0).\label{4ee} \end{align} Comparing \eqref{4ee} to \eqref{2e} shows that $i-j_0=0$ which implies that $ i=j_0<0,$ a contradiction (note, $i\geq 0$). Therefore, $g[i,j_0,k]\neq 0$ for all $(\xi, i,j,k,l)\in I.$ Now, observe that if there exists $\xi\in \{0,1,2\}$ such that $a_{(\xi, i,j_0,k,l)}=0$ for all $(i,j_0,k,l)\in \mathbb{N}\times \mathbb{Z}^3$, then one can easily deduce from equations \eqref{2ee}, \eqref{3ee} and \eqref{4e} that $a_{(\xi, i,j_0,k,l)}=0$ for all $(\xi, i,j_0,k,l)\in I.$ This will contradict our initial assumption. Therefore, there exists some $(i,j_0,k,l)\in \mathbb{N}\times \mathbb{Z}^3$ such that $a_{(\xi, i,j_0,k,l)}\neq 0$ for each $\xi\in \{0,1,2\}.$ Without loss of generality, let $(u,j_0,v,w)$ be the greatest element in the lexicographic order on $\mathbb{N}\times \mathbb{Z}^3$ such that $a_{(0,u,j_0,v,w)}\neq 0$ and $a_{(0,i,j_0,k,l)}=0$ for all $i>u.$ From \eqref{3ee}, at $(i,j_0,k,l)=(u,j_0,v,w)$, we have: \begin{align*} g[u,j_0,v]a_{(0,u,j_0,v,w)}+(q^{\bullet}\alpha {d_2[u+1]}/{a})a_{(1,u+1,j_0,v+1,w)}&=0. \end{align*} From \eqref{4e}, at $(i,j_0,k,l)=(u+1,j_0,v+1,w)$, we have: \begin{align*} g[u+1,j_0,v+1]a_{(1, u+1,j_0,v+1,w)}+(q^{\bullet}\alpha {d_2[u+2]}/{a})a_{(2, u+2,j_0,v+2,w)}&=0. \end{align*} Finally, from \eqref{2ee}, at $(i,j_0,k,l)=(u+2,j_0,v+2,w-1),$ we have: \begin{align*} q^{\bullet}\beta g[u+2,j_0,v+2]a_{(2,u+2,j_0,v+2,w)}+(q^{\bullet}\alpha b {d_2[u+3]}/{a})a_{(0,u+3,j_0,v+3,w-1)}&=0. \end{align*} Note: $ a, b,$ $\alpha,$ $\beta,$ $q^\bullet\neq 0;$ $g[i,j_0,k]\neq 0$ for all $(\xi, i,j_0,k,l)\in I;$ and $d_2[i]\neq 0$ for $i>0.$ Since $u+3>u,$ it follows from the above list of equations (starting from the last one) that $$a_{(0,u+3,j_0,v+3,w-1)}=0\Rightarrow a_{(2,u+2,j_0,v+2,w)}=0\Rightarrow a_{(1,u+1,j_0,v+1,w)}=0\Rightarrow a_{(0,u,j_0,v,w)}=0,$$ a contradiction! Hence, $a_{(0,i,j_0,k,l)}=0$ for all $(i,j_0,k,l)\in \mathbb{N}\times \mathbb{Z}^3.$ From \eqref{2ee}, \eqref{3ee} and \eqref{4e}, one can easily conclude that $a_{(\xi, i,j_0,k,l)}=0$ for all $(\xi, i,j_0,k,l)\in I.$ This contradicts our initial assumption, hence $x_-=0.$ Consequently, $x=x_+\in {\mathcal R}_5$ as desired. 2. From Remark \ref{rrr}, we have $z_{2}=t_{2}-bt_3^3t_4^{-1}$. Since $\delta(t_\kappa)=\lambda_\kappa t_\kappa, \ \kappa\in \{2,\cdots,6\}, $ with $\lambda_2:=-\lambda_5$ (see Lemma \ref{c4l1}), it follows that \begin{align*} \delta(z_{2})=&-\lambda_5t_{2}-b(3\lambda_3-\lambda_4) t_3^3t_4^{-1} =-\lambda_5z_{2}-b(3\lambda_3-\lambda_4+\lambda_5) t_3^3t_4^{-1}. \end{align*} Furthermore, \begin{align*} D(z_{2})&=\text{ad}_x(z_{2})+\delta(z_{2}) =\text{ad}_x(z_{2}) -\lambda_5z_{2}-b(3\lambda_3-\lambda_4+\lambda_5) t_3^3t_4^{-1}\in {\mathcal R}_5. \end{align*} Hence $b(3\lambda_3-\lambda_4+\lambda_5) t_3^3t_4^{-1}\in {\mathcal R}_5,$ since $\text{ad}_x(z_{2}) -\lambda_5z_{2}\in {\mathcal R}_5.$ This implies that $b(3\lambda_3-\lambda_4+\lambda_5) t_3^3\in {\mathcal R}_5{t}_4$ (note, from Appendix \ref{appenc}, $b\neq 0$). Set $w:=3\lambda_3-\lambda_4+\lambda_5.$ Suppose that $w\neq 0.$ From \eqref{de3}, we have: $$t_3^3=\dfrac{\beta}{b} t_6^{-1}-\dfrac{q^3\alpha}{ab}t_4t_5^{-1}+\dfrac{1}{ab} z_{1}t_3t_4.$$ It follows that $$wbt_3^3= {w\beta}t_6^{-1}-\frac{q^3w\alpha}{a} t_4t_5^{-1}+\frac{w}{a}z_{1}t_3t_4\in {\mathcal R}_5t_4.$$ Since $t_3^3, \ t_4t_5^{-1}$ and $z_1t_3t_4$ are all elements of ${\mathcal R}_5t_4,$ it implies that $t_6^{-1}\in {\mathcal R}_5t_4.$ Hence, $1\in {\mathcal R}_5t_4t_6.$ Using the basis $\mathcal{B}_5$ of ${\mathcal R}_5$ (Proposition \ref{c4p20}), this leads to a contradiction. Therefore, $w=0$. That is, $3\lambda_3-\lambda_4+\lambda_5=0,$ and so $\lambda_4= 3\lambda_3+\lambda_5.$ This further implies that $\delta(z_{2})=-\lambda_5 z_{2}$ as desired. Again, from Lemma \ref{c4l1}, we have that $\delta(f_1)=-(\lambda_3+\lambda_5)f_1.$ Recall from Remark \ref{rrr} that $z_{1}=f_{1}-st_3^2t_4^{-1}.$ It follows that \begin{align*} \delta(z_{1})=&-(\lambda_3+\lambda_5)f_{1}-s(2\lambda_3-\lambda_4)t_3^2t_4^{-1} =-(\lambda_3+\lambda_5)z_1-s(3\lambda_3-\lambda_4+\lambda_5)t_3^2t_4^{-1}\\ =&-(\lambda_3+\lambda_5)z_1-s(3\lambda_3-(3\lambda_3+\lambda_5)+\lambda_5)t_3^2t_4^{-1} =-(\lambda_3+\lambda_5)z_{1}. \end{align*} Finally, we know that $\delta(t_6)=\lambda_6t_6.$ This implies that $ \delta(t_6^{-1})=-\lambda_6t_6^{-1}.$ From \eqref{de3}, we have that $$t_3^3=\dfrac{\beta}{b} t_6^{-1}-\dfrac{q^3\alpha}{ab}t_4t_5^{-1}+\dfrac{1}{ab} z_{1}t_3t_4,$$ where $a$ and $b$ are non-zero scalars (Appendix \ref{appenc}). This implies that $$t_6^{-1}=\dfrac{b}{\beta}t_3^3+\dfrac{q^3\alpha }{a\beta} t_4t_5^{-1}-\dfrac{1}{a\beta}z_1t_3t_4.$$ Given that $\delta(z_1)=-(\lambda_3+\lambda_5)z_1, \ \delta(t_3)=\lambda_3t_3, \ \delta(t_4)=(3\lambda_3+\lambda_5)t_4$ and $\delta(t_5)=\lambda_5t_5,$ applying $\delta$ to the above relation gives $$-\lambda_6t_6^{-1}=3\lambda_3\left( \dfrac{b}{\beta}t_3^3+\dfrac{q^3\alpha }{a\beta} t_4t_5^{-1}-\dfrac{1}{a\beta}z_1t_3t_4\right).$$ It follows that $\lambda_6=-3\lambda_3$ as desired. 3. Set $\lambda_1:=-(\lambda_3+\lambda_5)$ and $\lambda_2:=-\lambda_5.$ It follows from points (1) and (2) that $D(e_{\kappa,5})=\text{ad}_x(e_{\kappa,5})+\delta (e_{\kappa,5})=\text{ad}_x(e_{\kappa,5})+\lambda_\kappa e_{\kappa,5}$ for all $\kappa\in \{1,\cdots,6\}.$ In conclusion, $D=\text{ad}_x+\delta$ with $x\in {\mathcal R}_5.$ \end{proof} We are now ready to describe $D$ as a derivation of $A_{\alpha,\beta}.$ \begin{lemma} \label{ev19} \begin{itemize} \item[1.] $x\in A_{\alpha,\beta}.$ \item[2.] $\delta(e_\kappa)=0$ for all $\kappa\in \{1,\cdots, 6\}.$ \item[3.] $D=\text{ad}_x.$ \end{itemize} \end{lemma} \begin{proof} In this proof, we denote $\underline{\upsilon}:=(i,j,k,l)\in \mathbb{N}^2\times\mathbb{Z}^2.$ Also, recall from DDA of $A_{\alpha,\beta}$ at the beginning of this section that $t_5=e_5$ and $t_6=e_6.$ 1. Given the basis $\mathcal{B}$ of $A_{\alpha,\beta}$ (Proposition \ref{c3p5}), one can write $x\in {\mathcal R}_5=A_{\alpha,\beta}[t_5^{-1},t_6^{-1}]$ as: $$x=\displaystyle\sum_{(\epsilon_1, \epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2, \underline{\upsilon})}e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k}t_6^{l},$$ where $I$ is a finite subset of $\{0,1\}^2\times \mathbb{N}^2\times\mathbb{Z}^2,$ and $a_{(\epsilon_1,\epsilon_2,\underline{\upsilon})}$ are scalars. Write $x=x_-+x_+,$ where $$x_+=\sum_{\substack{(\epsilon_1, \epsilon_2,\underline{\upsilon})\in I\\ k, \ l\geq 0 }}a_{(\epsilon_1,\epsilon_2, \underline{\upsilon})}e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k}t_6^{l},$$ and $$ x_-=\displaystyle\sum_{\substack{(\epsilon_1, \epsilon_2,\underline{\upsilon})\in I\\ k<0 \ \text{or} \ l<0}}a_{(\epsilon_1,\epsilon_2, \underline{\upsilon})}e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k}t_6^{l}.$$ Suppose that there exists a minimum negative integer $k_0$ or $l_0$ such that $a_{(\epsilon_1,\epsilon_2, i,j,k_0,l)}\neq 0$ or $a_{(\epsilon_1,\epsilon_2, i,j,k,l_0)}\neq 0$ for some $(\epsilon_1,\epsilon_2, i,j,k_0,l), (\epsilon_1,\epsilon_2, i,j,k,l_0)\in I,$ and $a_{(\epsilon_1,\epsilon_2, i,j,k,l)}=0$ whenever $k<k_0$ or $l<l_0.$ Write $$x_-=\displaystyle\sum_{\substack{(\epsilon_1, \epsilon_2,\underline{\upsilon})\in I\\ k_0\leq k \leq -1 \ \text{or} \ l_0\leq l\leq -1}}a_{(\epsilon_1,\epsilon_2, \underline{\upsilon})}e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k}t_6^{l}.$$ Now, $D(e_3)=\text{ad}_{x_+}(e_3)+\text{ad}_{x_-}(e_3)+ \delta(e_3)\in A_{\alpha,\beta}.$ From Remark \ref{rrr}, we have that $e_3=e_{3,6}-st_5^2t_6^{-1}$ and $e_{3,6}=t_3-at_4t_5^{-1}.$ Putting these two together gives \begin{align*} e_3 =t_3-at_4t_5^{-1}-st_5^2t_6^{-1}. \end{align*} Again, from Remark \ref{rrr}, we also have that $t_4=e_4+bt_5^3t_6^{-1}.$ Note, $\delta(t_\kappa) =\lambda_\kappa t_\kappa, \ \kappa\in \{3,4,5,6\}.$ Now, \begin{align} \delta(e_3)&=\lambda_3t_3-a(\lambda_4-\lambda_5)t_4t_5^{-1}-s(2\lambda_5-\lambda_6)t_5^2t_6^{-1}\nonumber\\ &=\lambda_3(e_{3,6}+at_4t_5^{-1})+ a(\lambda_5-\lambda_4)t_4t_5^{-1}+s(\lambda_6-2\lambda_5) t_5^2t_6^{-1}\nonumber\\ &=\lambda_3e_{3,6}+a(\lambda_3-\lambda_4+\lambda_5)t_4 t_5^{-1}+s(\lambda_6-2\lambda_5)t_5^2t_6^{-1}\nonumber\\ &=\lambda_3(e_3+st_5^2t_6^{-1})+ a(\lambda_3-\lambda_4+\lambda_5)(e_{4}+bt_5^3t_6^{-1}) t_5^{-1}+s(\lambda_6-2\lambda_5)t_5^2t_6^{-1}\nonumber\\ &=\lambda_3e_3+\alpha_1e_4t_5^{-1} +\alpha_2t_5^2t_6^{-1},\label{fe1} \end{align} where $\alpha_1=a(\lambda_3-\lambda_4+\lambda_5)$ and $\alpha_2=s(\lambda_3-2\lambda_5+\lambda_6)+q^{-3}ab(\lambda_3-\lambda_4+ \lambda_5).$ Therefore, $D(e_3)=\text{ad}_{x_+}(e_3)+\text{ad}_{x_-}(e_3)+ \lambda_3e_3+\alpha_1e_4t_5^{-1} +\alpha_2t_5^2t_6^{-1}\in A_{\alpha,\beta}.$ It follows that $D(e_3)t_5t_6=\text{ad}_{x_+}(e_3)t_5t_6+\text{ad}_{x_-}(e_3)t_5t_6+ \lambda_3e_3t_5t_6+ \alpha_1e_4t_6 +q^3\alpha_2t_5^3\in A_{\alpha,\beta}.$ Hence, $\text{ad}_{x_-}(e_3)t_5t_6\in A_{\alpha,\beta},$ since $\text{ad}_{x_+}(e_3)t_5t_6+ \lambda_3e_3t_5t_6+ \alpha_1e_4t_6 +q^3\alpha_2t_5^3\in A_{\alpha,\beta}.$ Now, \begin{align} \text{ad}_{x_-}(e_3)&=\displaystyle\sum_{(\epsilon_1, \epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2, \underline{\upsilon})}e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k}t_6^{l} e_3- \displaystyle\sum_{(\epsilon_1, \epsilon_2,\underline{\upsilon})\in I}a_{(\epsilon_1,\epsilon_2, \underline{\upsilon})}e_3e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k}t_6^{l}.\label{c4e5} \end{align} Using Lemma \ref{c3l3}, we have the following: \begin{align} t_5^{k}t_6^{l}e_3&=q^{-k}e_3t_5^{k}t_6^{l}+d_2[k]e_4 t_5^{k-1}t_6^{l}+d_3[l]t_5^{k+2} t_6^{l-1},\label{c4e6}\\ e_3e_1^{i}e_2^{j}&=q^{-i-3j}e_1^{i}e_2^{j}e_3+d_2[i]e_1^{i-1} e_2^{j+1},\label{c4e7} \end{align} Substitute \eqref{c4e6} and \eqref{c4e7} into \eqref{c4e5}, simplify and multiply (on the right) by $t_5t_6$ to obtain \begin{align} \text{ad}_{x_-}(e_3)t_5t_6=\nonumber\\ \sum_{(\epsilon_1, \epsilon_2,\underline{\upsilon})\in I}&a_{(\epsilon_1,\epsilon_2, \underline{\upsilon})}\left(g[i,j,\epsilon_2,l] e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1+1} e_4^{\epsilon_2}t_5^{k+1}t_6^{l+1}+q^{-3l}d_2[k]e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2+1}t_5^{k}t_6^{l+1}\right. \nonumber\\ +&\left.q^{-3(l-1)} d_3[l] e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k+3}t_6^{l}-q^{-3l} d_2[i]e_{1}^{i-1}e_{2}^{j+1}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k+1}t_6^{l+1}\right), \label{c4e8} \end{align} where $g[i,j,\epsilon_2,l]:=q^{-k-3\epsilon_2-3l}-q^{-i-3j-3l}.$ Assume that there exists $l<0$ such that $a_{(\epsilon_1,\epsilon_2,\underline{v})}\neq 0.$ It follows from our initial assumption that $a_{(\epsilon_1,\epsilon_2,i,j,k,l_0)}\neq 0.$ Now, at $l=l_0,$ denote $\underline{\upsilon}=(i,j,k,l)$ by $\underline{\upsilon}_0:=(i,j,k,l_0).$ From \eqref{c4e8}, we have that \begin{align*} \text{ad}_{x_-}(e_3)t_5t_6=&\displaystyle\sum_{(\epsilon_1, \epsilon_2,\underline{\upsilon}_0)\in I} q^{-3(l_0-1)} a_{(\epsilon_1,\epsilon_2, \underline{\upsilon}_0)}d_3[l_0] e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k+3}t_6^{l_0} +\mathcal{J}_1 \end{align*} where $\mathcal{J}_1 \in \text{Span}\left( \mathcal{B}\setminus \{e_1^{i}e_2^{j}e_3^{\epsilon_1}e_4^{\epsilon_2} t_5^{k}t_6^{l_0}\mid \epsilon_1,\epsilon_2\in \{0,1\}, \ k\in\mathbb{Z} \ \text{and} \ i,j\in \mathbb{N} \}\right) .$ Set $\underline{w}:=(i,j,k,l)\in \mathbb{N}^4.$ One can also write $\text{ad}_{x_-}(e_3)t_5t_6\in A_{\alpha,\beta}$ in terms of the basis $\mathcal{B}$ of $A_{\alpha,\beta}$ (Proposition \ref{c3p5}) as: \begin{align} \text{ad}_{x_-}(e_3)t_5t_6=&\sum_{(\epsilon_1,\epsilon_2, \underline{w})\in J}b_{(\epsilon_1,\epsilon_2,\underline{w})}e_1^{i} e_2^{j}e_3^{\epsilon_1}e_4^{\epsilon_2} t_5^{k}t_6^{l},\label{c4e10} \end{align} where $J$ is a finite subset of $\{0,1\}^2\times \mathbb{N}^4$, and $b_{(\epsilon_1,\epsilon_2,\underline{w})}\in {\mathbb K}.$ It follows that $$\sum_{(\epsilon_1,\epsilon_2, \underline{w})\in J}b_{(\epsilon_1,\epsilon_2,\underline{w})}e_1^{i} e_2^{j}e_3^{\epsilon_1}e_4^{\epsilon_2} t_5^{k}t_6^{l}= \displaystyle\sum_{(\epsilon_1, \epsilon_2,\underline{\upsilon}_0)\in I} q^{-3(l_0-1)} a_{(\epsilon_1,\epsilon_2, \underline{\upsilon}_0)}d_3[l_0] e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2}t_5^{k+3}t_6^{l_0} +\mathcal{J}_1.$$ Since $\mathcal{B}$ is a basis for $A_{\alpha,\beta},$ we deduce from Corollary \ref{c4c2} that $\left(e_1^{i} e_2^{j}e_3^{\epsilon_1}e_4^{\epsilon_2} t_5^{k}t_6^{l}\right)_{((\epsilon_1,\epsilon_2, \underline{v})\in \{0,1\}^2\times \mathbb{N}^2 \times \mathbb{Z}^2)}$ is also a basis for $A_{\alpha,\beta}[t_5^{-1}, t_6^{-1}].$ Since $\underline{v}_0=(i,j,k,l_0)\in \mathbb{N}^2\times \mathbb{Z}^2$ (with $l_0<0$) and $\underline{w}=(i,j,k,l)\in \mathbb{N}^4$ (with $l\geq 0$) in the above equality, we must have $$q^{-3(l_0-1)}a_{(\epsilon_1,\epsilon_2,\underline{\upsilon}_0)}d_3[l_0]=0.$$ Given that $q^{-3(l_0-1)}d_3[l_0]\neq 0,$ it follows that $a_{(\epsilon_1,\epsilon_2,\underline{\upsilon}_0)}=a_{(\epsilon_1,\epsilon_2,i,j,k,l_0)}$ are all zero. This is a contradiction. Therefore, $l\geq 0$ (i.e. there is no negative exponent for $t_6$). Since $l\geq 0,$ it follows from our initial assumption that there exists $k=k_0<0$ such that $a_{(\epsilon_1,\epsilon_2,i,j,k_0,l)}\neq 0.$ The rest of the proof will show that this assumption cannot also hold. Set $\underline{\upsilon}_0:=(i,j,k_0,l)\in \mathbb{N}^2\times \mathbb{Z}\times \mathbb{N}.$ From \eqref{c4e8}, we have that \begin{align*} \text{ad}_{x_-}(e_3)t_5t_6=& \displaystyle\sum_{(\epsilon_1,\epsilon_2, \underline{\upsilon}_0)\in I}q^{-3l} a_{(\epsilon_1,\epsilon_2, \underline{\upsilon}_0)}d_2[k_0]e_{1}^{i}e_{2}^{j}e_3^{\epsilon_1} e_4^{\epsilon_2+1}t_5^{k_0}t_6^{l+1}+V, \end{align*} where $V\in \mathcal{J}_2 := \text{Span}\left( \mathcal{B}\setminus \{e_1^{i}e_2^{j}e_3^{\epsilon_1}e_4^{\epsilon_2} t_5^{k_0}t_6^{l}\mid \epsilon_1,\epsilon_2\in \{0,1\} \ \text{and} \ i,j,l\in \mathbb{N} \}\right).$ It follows that: \begin{align} \text{ad}_{x_-}&(e_3)t_5t_6=\nonumber\\ &\sum_{(0,0,\underline{\upsilon}_0)\in I}q^{-3l} a_{(0,0,\underline{\upsilon}_0)}d_2[k_0]e_{1}^{i}e_{2}^{j} e_4t_5^{k_0}t_6^{l+1} +\sum_{(1,0, \underline{\upsilon}_0)\in I} a_{(1,0,\underline{\upsilon}_0)}d_2[k_0]e_{1}^{i}e_{2}^{j}e_3 e_4t_5^{k_0}t_6^{l+1}\nonumber\\ & +\sum_{(0,1, \underline{\upsilon}_0)\in I}q^{-3l} a_{(0,1,\underline{\upsilon}_0)}d_2[k_0]e_{1}^{i}e_{2}^{j} e_4^2t_5^{k_0}t_6^{l+1}\sum_{(1,1, \underline{\upsilon}_0)\in I} a_{(1,1,\underline{v}_0)}d_2[k_0]e_{1}^{i}e_{2}^{j}e_3 e_4^2t_5^{k_0}t_6^{l+1}+V.\label{c4e11} \end{align} Write the relations in Lemma \ref{c3l2}(2),(4) as: \begin{align} e_4^2=&b_1\beta+b_2e_2e_4e_6 +b_4\alpha e_3e_6+b_6e_1e_3e_4e_6 +L_1,\label{c4e12}\\ e_3e_4^2=& \beta b_1e_3+k_1e_2e_3 e_4e_6 +k_3\alpha^2 e_6+k_5\alpha e_1e_4 e_6 +k_{14}\beta e_1^2e_6 \nonumber\\&+k_{15}e_1^2 e_2e_4e_6^2 +k_{17}\alpha e_1^2e_3e_6^2+ k_{19} e_1^3e_3e_4e_6^2+ L_2,\label{c4e13} \end{align} where $L_1$ and $L_2$ are some elements of the left ideal $A_{\alpha,\beta}t_5\subseteq \mathcal{J}_2.$ Substitute \eqref{c4e12} and \eqref{c4e13} into \eqref{c4e11}, and simplify to obtain: \begin{align} \text{ad}_{x_-}(e_3)t_5t_6= \sum&[\lambda_{1,1}\beta a_{(0,1,i,j,k_0,l-1)}+\lambda_{1,2}\alpha^2a_{(1,1, i,j,k_0,l-2)}\nonumber\\ &+\lambda_{1,3}\beta a_{(1,1, i-2,j,k_0,l-2)}]e_1^ie_2^jt_5^{k_0}t_6^l\nonumber \\ +&\sum[\lambda_{2,1}\alpha a_{(0,1, i,j,k_0,l-2)} +\lambda_{2,2}\beta a_{(1,1, i,j,k_0,l-1)}\nonumber\\&+\lambda_{2,3}\alpha a_{(1,1,i-2,j,k_0,l-3)}]e_1^ie_2^je_3t_5^{k_0}t_6^l\nonumber\\ +&\sum[\lambda_{3,1}a_{(0,1, i,j-1,k_0,l-2)} +\lambda_{3,2}\alpha a_{(1,1, i-1,j,k_0,l-2)}\nonumber\\ &+\lambda_{3,3}a_{(1,1, i-2,j-1,k_0,l-3)}+\lambda_{3,4}a_{(0,0,i,j,k_0,l-1)}]e_1^ie_2^je_4t_5^{k_0}t_6^l\nonumber\\ +&\sum[\lambda_{4,1}a_{(0,1,i-1,j,k_0,l-2)} +\lambda_{4,2}a_{(1,1,i,j-1,k_0,l-2)}\nonumber\\ &+\lambda_{4,3}a_{(1,1,i-3,j,k_0,l-3)} +\lambda_{4,4}a_{(1,0,i,j,k_0,l-1)}]e_1^ie_2^je_3e_4t_5^{k_0}t_6^l+V', \label{c4e14} \end{align} where $V'\in \mathcal{J}_2$. Also, $\lambda_{s,t}:=\lambda_{s,t}(j,k_0,l)$ are some families of scalars which are non-zero for all $s,t\in\{1,2,3,4\}$ and $j,l\in\mathbb{N}$, except $\lambda_{1,4}$ and $\lambda_{2,4}$ which are assumed to be zero since they do not exist in the above expression. Note, although each $\lambda_{s,t}$ depends on $j,k_0,l,$ we have not made this dependency explicit in the above expression since the minimum requirement we need to complete the proof is for all the $\lambda_{s,t}$ existing in the above expression to be non-zero, which we already have. Observe that \eqref{c4e14} and \eqref{c4e10} are equal, hence, \begin{align*} \sum_{(\epsilon_1,\epsilon_2, \underline{w})\in J}b_{(\epsilon_1,\epsilon_2,\underline{w})}e_1^{i} e_2^{j}e_3^{\epsilon_1}e_4^{\epsilon_2} t_5^{k}t_6^{l} =& \sum[\lambda_{1,1}\beta a_{(0,1,i,j,k_0,l-1)}+\lambda_{1,2}\alpha^2a_{(1,1, i,j,k_0,l-2)}\nonumber\\ &+\lambda_{1,3}\beta a_{(1,1, i-2,j,k_0,l-2)}]e_1^ie_2^jt_5^{k_0}t_6^l\nonumber \\ +&\sum[\lambda_{2,1}\alpha a_{(0,1, i,j,k_0,l-2)} +\lambda_{2,2}\beta a_{(1,1, i,j,k_0,l-1)}\nonumber\\&+\lambda_{2,3}\alpha a_{(1,1,i-2,j,k_0,l-3)}]e_1^ie_2^je_3t_5^{k_0}t_6^l\nonumber\\ +&\sum[\lambda_{3,1}a_{(0,1, i,j-1,k_0,l-2)} +\lambda_{3,2}\alpha a_{(1,1, i-1,j,k_0,l-2)}\nonumber\\ &+\lambda_{3,3}a_{(1,1, i-2,j-1,k_0,l-3)}+\lambda_{3,4}a_{(0,0,i,j,k_0,l-1)}]e_1^ie_2^je_4t_5^{k_0}t_6^l\nonumber\\ +\sum&[\lambda_{4,1}a_{(0,1,i-1,j,k_0,l-2)} +\lambda_{4,2}a_{(1,1,i,j-1,k_0,l-2)}\nonumber\\ +\lambda_{4,3}&a_{(1,1,i-3,j,k_0,l-3)} +\lambda_{4,4}a_{(1,0,i,j,k_0,l-1)}]e_1^ie_2^je_3e_4t_5^{k_0}t_6^l+V'. \end{align*} We have previously established that $\left(e_1^{i} e_2^{j}e_3^{\epsilon_1}e_4^{\epsilon_2} t_5^{k}t_6^{l}\right)_{((\epsilon_1,\epsilon_2, \underline{v})\in \{0,1\}^2\times \mathbb{N}^2 \times \mathbb{Z}^2)}$ is a basis for $A_{\alpha,\beta}[t_5^{-1}, t_6^{-1}]$ (note, in this part of the proof $l\geq 0$). Since $\underline{v}_0=(i,j,k_0,l)\in \mathbb{N}^2\times \mathbb{Z}\times \mathbb{N}$ (with $k_0<0$) and $\underline{w}=(i,j,k,l)\in \mathbb{N}^4$ (with $k\geq 0$) in the above equality, it follows that \begin{align} \lambda_{1,1}&\beta a_{(0,1, i,j,k_0,l-1)}+\lambda_{1,2}\alpha^2 a_{(1,1, i,j,k_0,l-2)} +\lambda_{1,3}\beta a_{(1,1, i-2,j,k_0,l-2)}=0,\label{c4e15}\\ \lambda_{2,1}&\alpha a_{(0,1, i,j,k_0,l-2)} +\lambda_{2,2}\beta a_{(1,1, i,j,k_0,l-1)}+\lambda_{2,3}\alpha a_{(1,1, i-2,j,k_0,l-3)} =0,\label{c4e16}\\ \lambda_{3,1}&a_{(0,1, i,j-1,k_0,l-2)} +\lambda_{3,2}\alpha a_{(1,1, i-1,j,k_0,l-2)}+\lambda_{3,3}a_{(1,1, i-2,j-1,k_0,l-3)}\nonumber\\&+\lambda_{3,4}a_{(0,0, i,j,k_0,l-1)}=0,\label{c4e17}\\ \lambda_{4,1}&a_{(0,1, i-1,j,k_0,l-2)} +\lambda_{4,2}a_{(1,1, i,j-1,k_0,l-2)}+\lambda_{4,3}a_{(1,1,i-3,j,k_0,l-3)}\nonumber\\ &+\lambda_{4,4}a_{(1,0, i,j,k_0,l-1)}=0.\label{c4e18} \end{align} From \eqref{c4e15} and \eqref{c4e16}, one can easily deduce that \begin{align} &a_{(0,1, i,j,k_0,l)}=-\frac{\alpha^2\lambda_{1,2}}{\beta\lambda_{1,1}}a_{(1,1, i,j,k_0,l-1)} -\frac{\lambda_{1,3}}{\lambda_{1,1}}a_{(1,1, i-2,j,k_0,l-1)},\label{c4e19}\\ & a_{(1,1, i,j,k_0,l)}=-\frac{\alpha\lambda_{2,1}}{\beta\lambda_{2,2}}a_{(0,1, i,j,k_0,l-1)}-\frac{\alpha\lambda_{2,3}}{\beta\lambda_{2,2}}a_{(1,1, i-2,j,k_0,l-2)}.\label{c4e20} \end{align} Note, $a_{(\epsilon_1,\epsilon_2, i,j,k_0,l)}:=0$ whenever $i<0$ or $j<0$ or $l<0$ for all $\epsilon_1,\epsilon_2\in \{0,1\}.$ \textbf{Claim.} The coefficients $a_{(0,1, i,j,k_0,l)}$ and $a_{(1,1, i,j,k_0,l)}$ are all zero for all $l\geq 0$. We now justify the claim by an induction on $l.$ From \eqref{c4e19} and \eqref{c4e20}, the result is obviously true when $l=0.$ For $l\geq 0,$ assume that $a_{(0,1, i,j,k_0,l)}=a_{(1,1, i,j,k_0,l)}=0.$ Then, it follows from \eqref{c4e19} and \eqref{c4e20} that $$a_{(0,1, i,j,k_0,l+1)}=-\frac{\alpha^2\lambda_{1,2}}{\beta\lambda_{1,1}}a_{(1,1, i,j,k_0,l)} -\frac{\lambda_{1,3}}{\lambda_{1,1}}a_{(1,1, i-2,j,k_0,l)},$$ $$a_{(1,1, i,j,k_0,l+1)}=-\frac{\alpha\lambda_{2,1}}{\beta\lambda_{2,2}}a_{(0,1,i,j,k_0,l)}-\frac{\alpha\lambda_{2,3}}{\beta\lambda_{2,2}}a_{(1,1, i-2,j,k_0,l-1)}.$$ From the inductive hypothesis, $a_{(1,1,i,j,k_0,l)}=a_{(1,1,i-2,j,k_0,l)}=a_{(0,1,i,j,k_0,l)}=a_{(1,1, i-2,j,k_0,l-1)}=0.$ Hence, $a_{(1,1, i,j,k_0,l+1)}=a_{(0,1,i,j,k_0,l+1)}=0.$ By the principle of mathematical induction, $a_{(0,1, i,j,k_0,l)}=a_{(1,1,i,j,k_0,l)}=0$ for all $l\geq 0$ as desired. Given that the families $a_{(0,1, i,j,k_0,l)}$ and $a_{(1,1, i,j,k_0,l)}$ are all zero, it follows from \eqref{c4e17} and \eqref{c4e18} that $a_{(0,0, i,j,k_0,l)}$ and $a_{(1,0, i,j,k_0,l)}$ are also zero for all $(i,j,k_0,l)\in \mathbb{N}^2\times\mathbb{Z}\times \mathbb{N}.$ Since $a_{(\epsilon_1,\epsilon_2, i,j,k_0,l)}$ are all zero, it contradicts our assumption. Hence, $x_-=0.$ Consequently, $x=x_+\in A_{\alpha,\beta}$ as desired. 2. From Remark \ref{rrr}, we have $e_4 =t_4-bt_5^3t_6^{-1}.$ Again, from Lemma \ref{ev5}, we have that $\lambda_4=3\lambda_3+\lambda_5$ and $\lambda_6=-3\lambda_3.$ Therefore, \begin{align*} \delta(e_4) &=\lambda_4t_{4}-b(3\lambda_5-\lambda_6)t_5^3 t_6^{-1}\\& =(3\lambda_3+\lambda_5)e_{4,6}-3b(\lambda_3+\lambda_5)t_5^3 t_6^{-1}\\& =(3\lambda_3+\lambda_5)(e_4+bt_5^3t_6^{-1})-3b(\lambda_3+\lambda_5)t_5^3 t_6^{-1}\\ &=(3\lambda_3+\lambda_5)e_4-2b\lambda_5t_5^3t_6^{-1}. \end{align*} Moreover, $D(e_4)=\text{ad}_x(e_4)+\delta(e_4) =\text{ad}_x(e_4)+(3\lambda_3+\lambda_5)e_4-2b\lambda_5t_5^3t_6^{-1}\in A_{\alpha,\beta}.$ It follows that $b\lambda_5t_5^3t_6^{-1}\in A_{\alpha,\beta},$ since $\text{ad}_x(e_4)+(3\lambda_3+\lambda_5)e_4\in A_{\alpha,\beta}.$ Consequently, $ b\lambda_5t_5^3\in A_{\alpha,\beta}t_6.$ Since $b\neq 0$ (Appendix \ref{appenc}), we must have $\lambda_5 =0,$ otherwise, there will be a contradiction using the basis of $A_{\alpha,\beta}$ (Proposition \ref{c3p5}). Therefore, $\delta(e_4) =3\lambda_3e_4$ and $\delta(t_5) =0.$ We already know from Lemma \ref{ev5} that $\delta(t_6) =-3\lambda_3t_6$. From \eqref{fe1}, we have that $\delta(e_3)=\lambda_3e_3+a(\lambda_3-\lambda_4+\lambda_5)e_4t_5^{-1} +[s(\lambda_3-2\lambda_5+\lambda_6)+q^{-3}ab(\lambda_3-\lambda_4+\lambda_5)]t_5^2t_6^{-1}.$ Given that $\lambda_4=3\lambda_3, \ \lambda_5=0$ and $\lambda_6=-3\lambda_3,$ we have that $\delta(e_3)=\lambda_3e_3-2a\lambda_3e_4t_5^{-1}$ (note, from Appendix \ref{appenc}, one can confirm that $q^{-3}ab+s=0$). Now, $D(e_3)=\text{ad}_x(e_3)+\delta(e_3)=\text{ad}_x(e_3)+\lambda_3e_3-2a\lambda_3e_4t_5^{-1} \in A_{\alpha,\beta}.$ Observe that $\text{ad}_x(e_3)+\lambda_3e_3\in A_{\alpha,\beta}.$ Hence, $2a\lambda_3e_4t_5^{-1} \in A_{\alpha,\beta}, \ \text{and so} \ 2a\lambda_3e_4 \in A_{\alpha,\beta}t_5.$ Since $a\neq 0,$ it implies that $\lambda_3=0,$ otherwise, there will be a contradiction using the basis of $A_{\alpha,\beta}.$ We now have that $\delta(e_3)=\delta(e_4)=\delta(e_5)=\delta(e_6)=0.$ We finish the proof by showing that $\delta(e_1)=\delta(e_2)=0.$ Recall from \eqref{e2e} that $$e_2e_4e_6+be_2e_5^3+be_3^3e_6+b'e_3^2e_5^2+c'e_3e_4e_5+d'e_4^2=\beta.$$ Apply $\delta$ to this relation to obtain $\delta(e_2)e_4e_6+b\delta(e_2)e_5^3=0.$ This implies that $\delta(e_2)(e_4e_6+be_5^3)=0.$ Since $e_4e_6+be_5^3\neq 0,$ it follows that $\delta(e_2)=0.$ Similarly, from \eqref{e1e}, we have that $$e_1e_3e_5+ae_1e_4+ae_2e_5+a'e_3^2=\alpha.$$ Apply $\delta$ to this relation to obtain $\delta(e_1)(e_3e_5+ae_4)=0.$ Since $e_3e_5+ae_4\neq 0,$ we have that $\delta(e_1)=0.$ In conclusion, $\delta(e_\kappa)=0$ for all $\kappa\in \{1,\cdots, 6\}.$ 3. As a result of (1) and (2), we have that $D(e_\kappa)=\text{ad}_x(e_\kappa).$ Therefore, $D=\text{ad}_x$ as desired. \end{proof} Using similar techniques, one can describe the derivations of $A_{\alpha,0}$ and $A_{0,\beta}$. Details can be found in \cite{io-thesis}. There are fundamental differences in these two cases. Indeed, there exist in both cases derivations which are not inner. More precisely, one can check that the linear map $\theta$ of $A_{\alpha,0}$ defined by $$\theta(e_1)=-e_1, \ \ \theta(e_2)=-e_2, \ \ \theta(e_3)=0, \ \ \theta(e_4)=e_4, \ \ \theta(e_5)=e_5, \ \ \ \theta(e_6)=2e_6$$ is a ${\mathbb K}-$derivation of $A_{\alpha,0}$. Similarly, the linear map $\tilde{\theta}$ of $A_{0,\beta}$ by $$\tilde{\theta}(e_1)=-2e_1, \ \ \tilde{\theta}(e_2)=-3e_2, \ \ \tilde{\theta}(e_3)=-e_3, \ \ \tilde{\theta}(e_4)=0, \ \ \tilde{\theta}(e_5)=e_5, \ \ \ \tilde{\theta}(e_6)=3e_6$$ is a ${\mathbb K}-$derivation of $A_{0,\beta}$. We summarize our main results in the theorem below. \begin{theorem} Given that $A_{\alpha,\beta}=U_q^+(G_2)/\langle \Omega_1-\alpha,\Omega_2-\beta\rangle,$ with $(\alpha,\beta)\in {\mathbb K}^2\setminus \{(0,0)\},$ we have the following results: \begin{itemize} \item[1.] if $\alpha, \beta\neq 0;$ then every derivation $D$ of $A_{\alpha,\beta}$ can uniquely be written as $D=\text{ad}_x,$ where $x\in A_{\alpha,\beta}.$ \item[2.]if $\alpha\neq 0$ and $\beta=0,$ then every derivation $D$ of $A_{\alpha,0}$ can uniquely be written as $D=\text{ad}_x+\lambda\theta,$ where $\lambda\in {\mathbb K}$ and $x\in A_{\alpha,0}.$ \item[3.] if $\alpha=0$ and $\beta\neq 0,$ then every derivation $D$ of $A_{0,\beta}$ can uniquely be written as $D=\text{ad}_x+\lambda\tilde{\theta},$ where $\lambda\in {\mathbb K}$ and $x\in A_{0,\beta}.$ \item[4.] $HH^1(A_{\alpha,0})={\mathbb K}[\theta]$ and $HH^1(A_{0,\beta})={\mathbb K}[\tilde{\theta}],$ where $[\theta]$ and $[\tilde{\theta}]$ respectively denote the classes of $\theta$ and $\tilde{\theta}$ modulo the space of inner derivations. \item[5.] if $\alpha,\beta\neq 0;$ then $HH^1(A_{\alpha,\beta})=\{[0]\},$ where $[0]$ denotes the class of $0$ modulo the space of inner derivations. \end{itemize} \end{theorem} The above theorem shows that $A_{\alpha,\beta}$ when both $\alpha$ and $\beta$ are nonzero shares a number of properties with the second Weyl algebra over ${\mathbb K}$: it is simple, units are reduced to scalars, and all derivations are inner. It would be interesting to compute the automorphism group of these algebras and verify if all endomorphisms are automorphisms, i.e. an analogue of the celebrated Dixmier Conjecture \cite{dj2}. In general, the present work and \cite{sl} suggest that the primitive quotients of $U_q^+(\mathfrak{g})$ by primitive ideals from the 0-stratum provide algebras that could (should?) be regarded (and studied) as quantum analogue of Weyl algebras.

https://arxiv.org/abs/1711.00908

Feynman-Kac formula for the stochastic Bessel operator

We introduce a stochastic process and functional that should describe the semigroup generated by the stochastic Bessel operator. Recently Gorin and Shkolnikov showed that the largest eigenvalues for certain random matrix ensembles with soft edge behavior can be understood by analyzing large powers of tridiagonal matrices, which converge to operators in the stochastic Airy semigroup. In this article we make some progress towards realizing Gorin and Shkolnikov's program at the random matrix hard edge. We analyze large powers of a suitable tridiagonal matrix model (a slight modification of the $\beta$-Laguerre ensemble). For finite $n$ we represent the matrix powers using Feynman-Kac type formulas, which identifies a sequence of stochastic processes $X^n$ and functionals $\Phi_n$. We show that $\Phi_n(X^n)$ converges in probability to the limiting functional $\Phi(X)$ for our proposed stochastic Bessel semigroup. We also discuss how the semigroup method may be used to understand transitions from a hard edge to a soft edge in the $\beta$-Laguerre models.

\section{Introduction} \subsection{$\beta$-ensembles and their differential operator limits} Perhaps the best known random matrix models are the classical ``invariant ensembles". In these models the random matrices have real, complex or quaternion entries and thus it makes sense to talk about the number of degrees of freedom $\beta=1,2,4$ for the matrix entries. However the joint densities of eigenvalues depend on $\beta$ in such a way that one is led to consider values of $\beta$ other than $1,2,4$. For example the $\beta$-Laguerre eigenvalue density is \begin{align} \label{Lag1} dP(\lambda_1,\ldots , \lambda_n) = \frac{1}{Z_{\beta,a}} |\Delta( \lambda)|^{\beta} \prod_{k=1}^n \lambda_{k}^{\frac{\beta}{2}(a+1)-1} e^{-\frac{\beta}{2} \lambda_k} \, d\lambda_k, \end{align} where $\Delta(\lambda)$ is the determinant of the Vandermonde matrix $(\lambda_{i}^{j-1})_{i,j=1}^n$. There are also $\beta$-Hermite and $\beta$-Jacobi ensembles. In these models the eigenvalues can be thought of as an interacting particle system called a ``log gas" \cite{ForresterBook}, and in this interpretation $\beta$ is the reciprocal of temperature. In the celebrated paper \cite{DE02}, Dumitriu and Edelman used Householder conjugations to put the classical invariant ensembles in tridiagonal form, and showed that these tridiagonal random matrices can be constructed from independent samples from Gaussian or Chi distributions. For example the tridiagonal $\beta$-Laguerre ensembles are $n\times n$ matrices given by \begin{align} \label{eq1} L=& n^{-1} BB^t ,\qquad B_{kk}\eqd \chi [(k+a)\beta]/\sqrt{\beta}, \qquad B_{k+1,k}\eqd-\chi [k\beta]/\sqrt{\beta} . \end{align} All other entries of $B$ are $0$. Here $\chi[k]$ indicates the Chi distribution with parameter $k$. The tridiagonal models naturally make sense for all $\beta>0$ and their eigenvalues agree with the general $\beta$ extensions of the classical eigenvalue densities, for example (\ref{Lag1}). As is usual in random matrix theory, we are interested in limits taken as the matrix size $n$ approaches infinity. Edelman and Sutton \cite{ES07} showed that in this limit matrices from the tridiagonal $\beta$-ensembles act approximately like differential operators on appropriate $L^2$ spaces, and introduced the stochastic Airy and Bessel operators. This observation sparked a series of papers \cite{RRV,RR09,RR17,KRV13,VV16,RW16} exploring differential operator limits of the $\beta$-ensembles. In this paper we are interested in limits associated with the spectral edges of the $\beta$-ensembles. This excludes the sine-$\beta$ operator \cite{VV16} which has interesting connections to the Riemann hypothesis. In the context of the Laguerre $\beta$-ensembles, the spectral edges can be understood as follows. As $n\rightarrow \infty$ the mean density of eigenvalues for $L$ converges weakly in distribution to a point mass on the following Marchenko-Pastur law: \begin{align} d\mu(\lambda) =&\frac{1}{2\pi} \sqrt{(4-\lambda)/\lambda}\,\one_{[0,4]}(\lambda)\, d\lambda. \end{align} The matrices $L$ in display (\ref{eq1}) are positive definite, so the Laguerre ensembles are said to have a ``hard edge" at $\lambda=0$. In contrast, for finite values of $n$ the largest eigenvalue of $L$ may be greater than $4$, so the ensemble is said to have a soft edge at $\lambda=4$. Ramirez, Rider and Virag \cite{RRV} showed that the largest eigenvalues converge in finite dimensional distributions to those of the stochastic Airy operator (SAO) which we denote $\mathcal{A}$. For some constant $c$ they showed that \begin{align} cn^{2/3}\left(4I -L\right) \rightarrow \mathcal{A}=- \partial_{x}^2 +x +\frac{2}{\sqrt{\beta}}W'(x) . \label{SAO1} \end{align} Here $W'(x)$ is Gaussian white noise. One of the achievements of \cite{RRV} was to rigorously define the eigenvalues of $\mathcal{A}$. In the same paper they also characterized a $\beta>0$ generalization of the Airy point process in terms of a diffusion related to the eigenfunction equation for the SAO and used this to calculate tail asymptotics of the Tracy Widom-$\beta$ distributions. Later Krishnapur, Rider and Virag \cite{KRV13} showed that this limit is universal in the sense that in describes the largest eigenvalues for a family of models indexed by convex polynomial potential functions. The limiting operator for the hard edge is the stochastic Bessel operator $\mathcal{B}$. Ramirez and Rider \cite{RR09} showed that the smallest eigenvalues of $n^2 L$ converge\footnote{ The $n^2$ scaling is suggested by the following formal calculation: if $\lambda_k$ is the $k^{th}$ smallest eigenvalue then one would expect $k/n \approx \int_{0}^{\lambda_k} d\mu \approx C \sqrt{\lambda_k}$. A similar calculation explains the $n^{2/3}$ scaling in (\ref{SAO1}). } in finite dimensional distributions to those of the operator \begin{align} \mathcal{B}=&-x\partial_{x}^2 -\partial_x + \frac{a^2}{ 4x} -2\sqrt{\frac{x}{\beta}}W'(x)\partial_x + \frac{a}{\sqrt{\beta x}}W'(x) .\label{SBO1} \end{align} Later it was shown \cite{RW16} that the SBO limit is observed universally at the hard edge for a family of models indexed by polynomial potential functions. \subsection{Semigroup formula for the hard edge} We now briefly discuss the results of Gorin and Shkolnikov \cite{GS16}. They defined a family of tridiagonal random matrix models, which includes the $\beta$-Hermite ensembles, and showed that as $n\rightarrow \infty$ the semigroups generated by the tridiagonal matrices converge in distribution to a stochastic Airy semigroup. As a consequence they concluded the following: \begin{theorem}[Gorin-Shkolnikov \cite{GS16}] \label{GSTheorem} For almost all $W$ the following holds. If $f\in L^2[0,\infty)$ then \begin{align} \exp(-t \mathcal{A}/2) f(x) =& \mathbb{E}_x\left[ \one_E \exp\left( -\frac{1}{2}\int_{0}^t B_s\, ds + \frac{1}{\sqrt{\beta}}\int_{0}^{\infty}L_B(x,t)\,dW(x)\right)f(B_t)\right], \label{GS1} \end{align} where $B$ is a standard Brownian motion. The subscript $\mathbb{E}_x$ indicates the starting position of $B$, and $E$ is the event that $\inf_{0\leq u\leq t}B_u\geq 0$. \end{theorem} In a typical statement of the Feynman-Kac formula the hypotheses would include a smoothness condition for the potential function, for example a uniform Lipschitz condition. With the white noise term $\beta^{-1/2} W'(x)$, the operator $\mathcal{A}$ does not meet the hypotheses of any such theorem. Furthermore the authors of \cite{GS16} showed that the traces of both sides in equation (\ref{GS1}) agree. This extends to general $\beta$ the following formula of Okounkov \cite{OkRMRP}: \begin{align} \mathbb{E}\left[\exp \left\{ tn^{1/6} \tr ( H -2\sqrt{n}I)\right\}\right] \rightarrow \frac{\exp(t^3/12)}{2\sqrt{\pi}t^{3/2}}, \end{align} where $H$ is an $n\times n$ Gaussian unitary ensemble matrix. By comparing the SAO semigroup functional to Okounkov's formula, Gorin and Shkolnikov were able to deduce a new identity for the local time of a Brownian excursion. In a subsequent paper Lamarre and Shkolnikov \cite{LS17} extended this technique to study the largest eigenvalues of spiked Laguerre ensembles. In this article we make some progress towards realizing Gorin and Shkolnikov's program at the random matrix hard edge. We will analyze large powers of a suitable tridiagonal matrix model, a slight modification of the Laguerre $\beta$-ensemble. For finite $n$ we represent the matrix powers using Feynman-Kac type formulas, which identifies a sequence of stochastic processes $X^n$ and functionals $\Phi_n$. We show that $\Phi_n(X^n)$ converges in probability to the $\Phi(X)$ where \begin{align} dX=&\sqrt{2X}\,dB +\left(1+2\sqrt{X/\beta} W'(X)\right)\, dt, \label{SDE1} \\ \Phi(X)=&-\frac{a^2}{4}\int_{0}^t \frac{du}{X_u} - \frac{a}{\sqrt{\beta}} \int_{0}^1 \frac{L_X(x,t)}{\sqrt{x}} \circ dW(x). \label{PhiDef1111} \end{align} Typical existence and uniqueness theorems for SDE's require the drift coefficient to be uniformly Lipschitz, so at first it is not clear that (\ref{SDE1}) defines a stochastic process. In section \ref{SectionInterpretations} we use the Ito-McKean speed and scale construction to define a process $X$ that formally satisfies (\ref{SDE1}). It was recently explained in \cite{HLM16} how to rigorously interpret the speed and scale construction as a strong solution to (\ref{SDE1}). Even after we explain the precise definition of the process $X$, it will not be obvious that the Stratonovich integral on line (\ref{PhiDef1111}) is well defined, but (as we will show in section \ref{SectionInterpretations}) this point can also reconciled using the methods of \cite{HLM16}. Formally applying the Feynman-Kac formula to the SBO leads to the following conjecture: \begin{conjecture} The stochastic Bessel operator satisfies the following Feynman-Kac formula. If $f\in L^2[0,1]$ then \begin{align} \exp(-t\mathcal{B}) f(x) \label{SBO_semigroup} =& \mathbb{E}_x \left[ \one_A \exp \Phi(X) f(X_t) \right], \end{align} where $\Phi$ and $X$ are as in displays (\ref{SDE1},\ref{PhiDef1111}), and $A$ is the event $\sup_{0\leq u\leq t}X_u\leq 1$. Furthermore if $f\in L^2[0,1]$ is fixed and $f_n$ is the orthogonal projection of $f$ onto the span of $\{ \one_{[(k-1)/n,k/n)}:\; k=1,\ldots ,n\}$, then $(I-L)^{tn^2} f_n $ should converge in distribution to the right hand side of (\ref{SBO_semigroup}), with $L$ being a matrix from the $n\times n$ tridiagonal $\beta$-Laguerre ensemble. \end{conjecture} In this paper we make some progress towards the above conjecture. We have been somewhat conservative in only stating the above conjecture for the $\beta$-Laguerre ensembles. It seems very plausible that it would also hold for the models considered in \cite{RW16}, and for perturbations of the $\beta$-Laguerre ensembles where the matrix entries of $B$ are independent and satisfy some moment conditions as in \cite{GS16}. The model we consider will actually be a slight perturbation of the $\beta$-Laguerre ensemble. As we discuss in section \ref{DefSection}, we have chosen the model deliberately to simplify our analysis. We let $B_n$ be a bidiagonal matrix with diagonal entries $\mathbf{x}_1,\ldots ,\mathbf{x_n}$ and subdiagonal entries $-\mathbf{y}_1,\ldots ,-\mathbf{y}_{n-1}$. We will define the entries $\mathbf{x}_k$ and $\mathbf{y}_k$ in section \ref{matrixModelSection}, but for now we just comment that they are approximately independent and approximately distributed like the $\chi$ random variables in display (\ref{eq1}). We then define \begin{align} A=\frac{1}{n} P B B^t P^{-1}, \end{align} where $P$ is an $n\times n$ diagonal matrix with $P_{k,k}=(\mathbf{y}_k/\mathbf{x}_k)\sqrt{1+a/j}$. This conjugation is necessary because formal calculations suggest that the differential operator limit for $n^{-1}BB^t$ would have a white noise squared term, and therefore not even make sense as a mapping of functions to distributions. In the interest of stating our main theorem in this section, we now introduce the stochastic processes $X^n$ and functionals $\Phi_n$ that appear in our discrete Feynman-Kac formula. However, to be brief we will point the reader to later sections for a couple of the details. The processes $X^n$ are left continuous, piecewise constant and have the form $X^{n}(i \Delta t)=X(\t_i )$ where $\Delta t=1/(4tn^2)$ and $\t_i$ are some stopping times defined in section \ref{DiscreteProcessSection}. As we show in section \ref{ApproxJumpTimesSection}, if $i$ and $n$ are large then $\t_i$ and $i\Delta t$ are approximately equal. The discrete functionals are \begin{align} \label{phiFormula00021} \Phi_n(t)=& \frac{-1}{4n^2} \sum_{i=1}^{\lfloor 4tn^2 \rfloor} \frac{a^2}{4 x(i)}+\frac{a\sqrt{n}G_{n x(i)}}{\sqrt{\beta x(i)}} +\frac{a G^{(2)}_{nx(i)} }{\beta x(i)} ,\qquad \text{where }x(i)=X^{n}(i\Delta t), \end{align} and the quantities $G_k$ and $G_{k}^{(2)}$ are random variables constructed from the Wiener process $W$ in display (\ref{GDEF}) of section \ref{phiNSection}. We will show in section \ref{discFKSection2} that the random matrices $A$ satisfy the following discrete Feynman-Kac formula. If $v \in \mathbb{R}^n$ then \begin{align}\label{DFK2} ((I-A/4)^{4tn^2} v)_k =\mathbb{E}_{k/n}\left[ \one_{E} \exp \left(\Phi_n +\mathcal{E}_n\right)v_{nX^n(t)} \right], \end{align} where $E$ is the event $\sup_{u\leq t}X^{n}_u\leq 1$ and $\mathcal{E}_n$ is an error term which, on the event $\inf_{u\leq t}X^{n}_u \geq 1/\log n$, satisfies the estimate $\mathcal{E}_n =O(\sqrt{\log (n)/n})$. The subscript $\mathbb{E}_{k/n}$ indicates the starting position for $X^n$. In section \ref{HardEdgeSection} we show that the smallest eigenvalues of $n^2A$ converge to those of $\mathcal{B}$. Assuming this, it is easy to see that the eigenvalues of the left hand side $(I-A/4)^{4tn^2}$ in (\ref{DFK2}) should converge to those of $\exp(-t\mathcal{B})$. The left hand side of \ref{DFK2} is the same kind of matrix power considered in \cite{GS16}. At this point it seems appropriate to comment that throughout the paper, the quality of our approximations will degrade whenever $X(t)$ approaches $0$. That is one of the main technical difficulties that must be overcome in order to prove conjecture 1. Our choice of the cutoff $X>1/\log n$ is somewhat arbitrary; we could for example use a cutoff of $X>n^{-p}$ for some $p>0$, but this would change the error estimates in all our propositions. Our main result is the following. \begin{theorem} \label{mainThm} On the event $\inf_{u\leq 1.01t} X(u)\geq 1/\log n$, we have \begin{align} \left|\Phi -\Phi_n\right| \leq Ct n^{-0.249}. \end{align} This holds for almost all $B,W$, and $C$ depends only on $B,W,\beta$. \end{theorem} Below in lemma \ref{pathLowerBoundLemma} we show that given any fixed starting position, the probability of $X$ hitting $1/\log n$ before $1$ converges to zero. Also by lemma \ref{TimeApproxProp} we have that if $X^n(t)$ converges to $X(t)$. It follows that if $f$ is continuous on $[0,1]$ and we let \begin{align} v_k =n\int_{(k-1)/n}^{k/n} f(x)\, dx, \end{align} then the quantity inside the expected value of our discrete Feynman-Kac formula (\ref{DFK2}) converges in probability to the integrand of our proposed limiting Feynman-Kac formula (\ref{SBO_semigroup}). \subsection{Semigroup in terms of Brox's diffusion} Before moving on with the main thrust of the paper, we would like to comment briefly on one more point. In just this section we will relax the mathematical rigor of our exposition and present some formal calculations. A connection to Brox's diffusion at the random matrix hard edge has been observed before in \cite{RR09}. The SDE for our process $X$ is not quite Brox's diffusion, but formally can be transformed to it as follows. Let \begin{align} Y_t =& -\frac{1}{\beta}\log X_{\theta(t)} ,\qquad \text{where}\qquad\theta(t)= \beta^2 \int_{0}^t \exp\left ( -\beta Y_u \right)\, du. \end{align} Formally applying Ito's lemma gives the SDE \begin{align} dY=dB+2 W'(Y)\, dt,\label{BroxSDE} \end{align} but to see this, one must be careful to use the special chain rule for Brownian motion\footnote{ This chain rule can be seen from $\text{Var}\int f(x)\,dW(x) =\int f(x)^2 dx = \text{Var}\int f(h(y))\sqrt{dx/dy}\,d\wt{W}(y)$. }: if $x=h(y)$ then $W'(x)=|h'(y)|^{-1/2}\wt{W}'(y)$ for a new Brownian motion $\wt{W}$. In terms of the transformed process $Y$, the Feynman-Kac formula becomes \begin{align} \exp(-t\mathcal{B})f(x) =& \mathbb{E}_x \left[\one_E \exp\left( \frac{(a\beta)^2}{4} \tau+ a\beta \int_{0}^\infty L_Y(y,\tau)\circ dW(y) \right) f\left(e^{-\beta Y_\tau}\right)\right] , \label{BroxSemigroup} \end{align} where $\tau=\tau(t)=\theta^{-1}(t)$. Some interesting simplifications happen when calculating the functional in (\ref{BroxSemigroup}), for example \begin{align} &\int_{0}^t \frac{W'(X_u)}{\sqrt{X_u}}\, du = \int_{0}^{\theta^{-1}t} \frac{W'(X_{\theta(u)})}{\sqrt{X_{\theta(u)}}} \beta^2 e^{-\beta Y_u}\, du \\ &\qquad = \int_{0}^{\theta^{-1}t} \frac{dW}{dX} ( e^{-\beta Y_u}) \beta^2 e^{-\beta Y_{u}/2}\, du =\beta^2 \int_{0}^{\theta^{-1}t} \wt{W}'(Y_u)\, du. \end{align} One could now attempt to remove the white noise drift term from the SDE (\ref{BroxSDE}) using Girsonov's theorem. But this would transform the limiting functional $\Phi$ to a formal expression with a term $-\int_{0}^t W'(Y_u)^2\, du$, which must be interpreted as $-\infty$ for almost all sample paths $Y$. This reflects the fact that the sample paths for Brox' diffusion are concentrated on a set (depending on $W$) which for almost all $W$ will have Wiener measure $0$ with respect to the ``quenched" measure on $B$ given $W$. Thus it does not appear feasible to modify the hard edge semigroup formula so that the driving process becomes standard Brownian motion. \subsection{Rigorous interpretation of $X$ and $\Phi$} \label{SectionInterpretations} We now rigorously define a stochastic process $X$ which formally satisfies the SDE (\ref{SDE1}). Start with the following diffusion coefficient and potential function: \begin{align} \sigma(x) =&\sqrt{2x} ,\qquad V(x)= - \log(x)+\frac{2}{\sqrt{\beta}} \int_{x}^1 \frac{dW(y)}{\sqrt{y}} . \end{align} Let $B$ be a standard Brownian motion started at $s(X_0)$, and define $X$ by \begin{alignat}{2} \label{speedScale1} X_t=&s^{-1}(B_{T^{-1}t}),\qquad & s(x)=&-\int_{x}^1 e^{V(x)}\, dx,\qquad T^{-1}(t) = \int_{0}^{t} \sigma(X_u)^2 e^{2V(X_u)}\, du. \end{alignat} Here $s$ and $T$ are respectively called the scale and time functions. Just before stating theorem \ref{mainThm}, we explained that throughout the paper our approximations will require a lower bound on the process $X$. We now prove that $X$ satisfies such a lower bound with high probability. \begin{lemma} \label{pathLowerBoundLemma} For almost all $W$, all $x\in (0,1)$ and all $t>0$ we have \begin{align} P_{x}\left(\inf_{\tau\leq t} X(\tau) \leq \frac{1}{ \log n } \;\; \mathrm{or}\;\; \sup_{\tau \leq t}X(\tau)\geq 1 \middle| W\right) \rightarrow 0\text{ as }n\rightarrow \infty. \end{align} \end{lemma} \begin{proof} It suffices to estimate the probability of hitting $1/\log n$ before $1$, so call this event $E$. Since $ P(E|W) = (s(1)-s(x))/(s(1)-s(1/\log n)), $ and the numerator does not depend on $n$, it suffices to estimate the denominator. \begin{align} s(1)-s(1/\log n)=& -\int_{1/\log n}^1 y^{-1}e^{\frac{2}{\sqrt{\beta}}\wt{W}_{-\log y}}\, dy= \int_{0}^{\log \log n} e^{\frac{2}{\sqrt{\beta}}\wt{W}_z}\, dz. \end{align} It is easy to see that the above integral will increase to $+\infty$ as $n\rightarrow \infty$ for almost all $W$. \end{proof} We now address the interpretation of the Stratonovich integral on line (\ref{SBO_semigroup}). Because $X$ depends on the random environment $W$, the local time $L_{X}(x,t)$ will not be adapted to the filtration generated by $W$. Therefore the existence of our integral is not clear from the standard theory of stochastic calculus. We assign a meaning to the integral following the approach of \cite{HLM16}. Local time for $X$ can be expressed in terms of local time for $B$ by\footnote{ Derivation of the local time relation: \begin{align} \int_{0}^1 f(x)L_{X}(x,t)\, dx =& \int_{0}^t f(X_t)\, dt =\int_{0}^{T^{-1}(t)} f(s^{-1}(B_u)T'(u)\, du =\int_{0}^{T^{-1}(t)} \frac{f(s^{-1}(B_u)}{(\sigma \cdot s')^2\circ s^{-1}(B_u)}\, du\\ =& \int_{0}^\infty \frac{f(b)L_B(b,T^{-1}(t))}{(\sigma \cdot s')^2\circ s^{-1}(b)}\, db =\int_{0}^\infty \frac{f(b)L_B(s(x),T^{-1}(t))}{(\sigma(x)^2 s'(x)}\, dx \end{align} } \begin{align} L_X(x,t)=\frac{L_B(s(x),T^{-1}(t))}{\sigma(x)^2 s'(x)}.\label{changeLocTime} \end{align} The time $T^{-1}(t)$ is random and depends on $W$ in a complicated way, but we can replace it by a deterministic time, evaluate the integral and substitute the random time afterwards. Note that on the right hand side of (\ref{changeLocTime}) the $x$ dependence is through $s(x)$ which only depends on the process $(W_y)_{y\geq x}$. Therefore with $T^{-1}(t)$ replaced by a deterministic time $u$, the local time is adapted to the filtration generated by $W$ \emph{in the $-x$ direction}. For this reason we will work with ``anti-Ito" integrals\footnote{ Let $\Delta W_i=W_{(i+1)/n}-W_{i/n}$. The different stochastic integrals can be summarized as follows: \begin{align} \int_{0}^1 Z_x \, dW_x \approx \sum_{i=0}^{n-1}Z_{i/n}\Delta W_i,\qquad \int_{0}^1 Z_x \circ dW_x \approx \sum_{i=0}^{n-1}Z_{\frac{i+0.5}{n}}\Delta W_i,\qquad \int_{0}^1 Z_x \, d\ola{W}_x \approx \sum_{i=0}^{n-1}Z_{\frac{i+1}{n}}\Delta W_i . \end{align} } throughout the paper, which we write using the notation $d\ola{W}$. Finally, we can explain the meaning of the Stratonovich integral in equation (\ref{SBO_semigroup}): \begin{align} \int_{0}^\infty \frac{L_X(x,t)}{\sqrt{x}} \circ dW(x) := \left.\int_{-\infty}^\infty \frac{L_B(s(x),u)}{\sqrt{x}\sigma(x)^2 s'(x)} d\ola{W}(x) \right|_{u=T^{-1}(t)} - \int_{-\infty}^\infty \frac{L_B(s(x),T^{-1}(t))}{\sqrt{\beta}x\sigma(x)^2 s'(x)} \, dx \label{integral_interp} \end{align} The second integral in (\ref{integral_interp}) is the Ito-Stratonovich correction term. It arises since $L_B(s(x),u)$ has no quadratic variation with respect to $W$, and because by Ito's lemma we have \begin{align} d s'(x) =& \left(2/\beta -1\right) \frac{s'(x)}{x} \, dx - \frac{2s'(x)}{\sqrt{\beta x}} \,dW(x). \end{align} \subsection{Organization of the paper} Our paper is organized as follows. In section \ref{DefSection} we define a matrix model such that large powers of the matrix should converge to the stochastic Bessel semigroup operators. For finite $n$ a discrete Feynman-Kac type formula holds exactly, and this leads us to define the processes $X^n$ and functionals $\Phi_n$. Most of the paper is devoted to proving theorem \ref{mainThm}, but before undertaking this we discuss in section \ref{EdgeTransitionSection} the hard edge to soft edge transition from the semigroup point of view. The functional $\Phi$ has two integrals, a fairly tame additive functional of the process $X$, and the much more ornery white noise local time integral. The discrete functionals $\Phi_n$ we define in section \ref{phiNSection} decompose in the same way as $\Phi$, and so we divide our proof of theorem \ref{mainThm} into two parts. First in section \ref{NoNoiseSection} we show convergence of the better behaved parts of the functionals which we call $\Phi^{A}_n$ and $\Phi^A$. Our method is essentially to construct some approximate martingales and use them to bound a moment generating function. In the remaining sections of the paper we carry out this method for the more difficult noisy parts of the functional. This relies on a branching structure for the numbers of visits of the processes $X^n$ to different lattice points. The crux of the proof is essentially to show that network of potential energy traps due to the random environment $W$ do not cause the discrete processes $X^n$ to get stuck in one place for too long. Our argument for analyzing the noisy parts of the functionals roughly follows the template in section \ref{NoNoiseSection}, but is more complicated so we separate it into two parts, sections \ref{section5} and \ref{section6}. \subsection{Acknowledgements} It is a pleasure to thank Brian Rider for many useful discussions. \section{Random environment, driving process and matrix model} \label{DefSection} In this section we define the random processes, functionals and matrix model that we will study in this paper. We do this in such a way that they are all coupled, and are only random through their dependence on two independent Brownian motions: the ``random environment" $W$, and another Brownian motion $B$ which we call the ``driving process". At first it might appear simpler to work directly with the $\beta$-Laguerre ensembles. Our approach has two advantages: first that all of the $n\times n$ random matrices are naturally coupled to the limiting objects. This makes it easy to work in terms of almost sure convergence in many places. The second advantage is that the matrix model we choose simplifies some parts of our analysis. If one instead uses the $\beta$-Laguerre ensembles, then $\Phi_n$ must be modified, and will depend explicitly on the step types of $X^n$ (i.e.\ whether it jumps up, down or does not jump at each time point). This makes the formulas more complicated, but in the end the same methods still work; so to streamline our exposition we prefer to avoid these complications. Also if one works directly with the $\beta$-Laguerre ensembles, then the most convenient way to compare the processes $X^n$ and $X$ is to use the construction of Seignourel \cite{Seignourel00}. There the scale function $s(x)$ is approximated using discrete versions $s_n(x)$, and estimating the associated error terms would add an additional section to the present paper. Our construction of $X^n$ from $X$ is in fact a modified version of Seignourel's. \subsection{Discrete process} \label{DiscreteProcessSection} First we define lattice path processes $X^n$ that will approximate $X$. Discretize space and time by \begin{align} x_k=k/n,\qquad t_i =i \Delta t,\qquad \Delta t =1/(4n^2),\qquad i,k\in \mathbb{N}\cup\{0\}. \end{align} We will use the notation $x_k$ very often. By ``lattice path" we mean that $X^n$ is piecewise constant with jumps of only $\pm 1/n$ allowed only at the times $t_i$. \begin{definition} $X^n$ is stochastic process defined by: $X^{n}$ is constant on any time interval not containing a time lattice point $t_i$, $i\in \mathbb{N}$; and $X^n(t_i)=X(\t_i)$, where $\t_i$ are stopping times defined below in display (\ref{gothicTDef}). \end{definition} Let $\t_1$ be the first time $X$ hits one of the lattice points $x_k$. Let \begin{align} r_k=1-\frac{k}{2n}. \end{align} It will turn out that $r_k$ is the probability that the next step of $X^n$ is $\rightarrow$ given that the present location is $x_k$. Suppose that the stopping times $\t_i$ have been determined for $i=1,\ldots ,\ell-1$ and let $x_k =X^{n}(\frac{ \ell-1}{n^2})$ be the present location of $X^n$. Using the abbreviations $s_{k}=s(x_k)$ for values of the scale function, choose $\alpha_+,\alpha_-$ so that \begin{align} r_k =&\frac{s(\alpha_+)-s_k}{s_{k+1}-s_{k}} =\frac{s_k- s(\alpha_-) }{s_{k }-s_{k-1}} . \end{align} The next stopping time is then defined by \begin{align} \label{gothicTDef} \t^{-}_{\ell} =&\inf\left\{ t \geq \t_{\ell-1}:\; nX_{t} \in \{ \alpha_-,\alpha_+ \} \right\ \\ \t_{\ell} =&\inf\left\{ t \geq \t^{-}_{\ell}:\; nX_{t} \in \{ k(\ell-1)-1 , k(\ell-1) , k(\ell-1)+1\} \right\},\qquad k(\ell)=nX_{t_\ell}.\nonumber \end{align} The above construction can be summarized as follows: the next jump of $X^n$ will be $\uparrow$ or $\downarrow$ depending on which of $x_{k-1}$ or $x_{k+1}$ the process $X$ hits next; however, before this happens $X^n$ may take several $\rightarrow$ steps if $X$ ventures out of $(\alpha_-,\alpha_+)$ and then returns to $x_k$. As a general convention, we will not refer to $\rightarrow$ steps as jumps. \subsection{Discrete functional} \label{phiNSection} Formally applying the Feynman-Kac formula to the stochastic Bessel operator gives the limiting functional \begin{align} \Phi(X)= -\int_{0}^t \frac{a^2}{4X_s} + \frac{a W'(X_s)}{\sqrt{ \beta X_s}}\, ds. \end{align} To discretize this functional correctly we need to make some definitions. Let $ x(i):=X^n(t_i)=X(\t_i), $ and let $G_k$ and $G^{(2)}_k$ be random variables defined by \begin{align} \label{GDEF} G_k =& -\sqrt{n}\int_{x_{k-1}}^{x_{k+1}} \left(\begin{cases}n( s-x_{k-1}) &\text{if }s<x_k \\ 1-n(x_{k+1}-s) &\text{if } x_k <s \end{cases} \right)\, dW(s) ,\\ G^{(2)}_k =& n \int_{x_{k-1}}^{x_{k+1}} \int_{x_{k-1}}^{s_2}\left(\begin{cases} n(s_1 -x_{k-1})(1- n(s_2-x_{k-1})) & \text{if }s_1<s_2<x_k \\ 0 & \text{if } s_1<x_k <s_2 \\ n(s_1 -x_k)(1 -n(s_2 -x_k)) & \text{if }x_k <s_1<s_2 \end{cases}\right)\, dW(s_1)\, dW(s_2).\nonumber \end{align} Note that $G_k$ and $G^{(2)}_k$ depend on $n$, and are scaled so that they will typically be of order $1$ as $n\rightarrow\infty$. $G_k$ is Gaussian but $G^{(2)}_k$ is not. The discretization of $\Phi$ that we will consider in this paper is \begin{align} \label{phiFormula} \Phi_n(t)=& \frac{-1}{4n^2} \sum_{i=1}^{\lfloor 4tn^2 \rfloor} \frac{a^2}{4x(i)}+\frac{a\sqrt{n}G_{n x(i)}}{\sqrt{\beta x(i)}} +\frac{a G^{(2)}_{nx(i)} }{\beta x(i)} . \end{align} At this point it is obscure why the third ``$G^{(2)}$ term" inside the sum should be included; the reason this term is included will be made clear in section \ref{discFKSection2}. Let $p_k$ be the probability that the next jump of $X^n$ is $\uparrow$ given that the present location is $x_k$, and $q_k=1-p_k$. The following lemma explains the origins of the definitions $G_k,G^{(2)}_k$. Its proof is a straitforward but tedious calculation which we defer to appendix A. \begin{lemma} \label{pkApproxLemma} For almost all $W$ the approximation \begin{align} \label{pkExpansion} \log \frac{p_k}{q_k} =& \frac{2G_k}{\sqrt{\beta k} }+\frac{1}{k} +\frac{2G_{k}^{(2)}}{\beta k} + O\left( \frac{\sqrt{\log k}}{k^{3/2}}\right) \end{align} holds for all $n\in \mathbb{N}$ and $k=1 ,\ldots , n$. The implied constant depends only on $W,\beta$. \end{lemma} \subsection{Matrix model} \label{matrixModelSection} In this section we construct random matrix ensembles $A=A^{(n)}$. Start by naming the entries of a bidiagonal matrix as follows: for $k=1,\ldots ,n$ let $B_{k,k}= \mathbf{x}_k $ and $B_{k+1,k} =-\mathbf{y}_k$. Now let $A=\frac{1}{n}P B B^t P^{-1}$, where $P$ is the $n\times n$ diagonal matrix given by \begin{align} P_{k,k}= \prod_{j=1}^{k-1} \frac{\mathbf{y}_j}{\mathbf{x}_{j}}\sqrt{1+\frac{a}{j} } . \label{basis3892} \end{align} Some motivation for this change of basis is explained in the footnote\footnote{ The definitions in this section imply the approximations \begin{align} \mathbf{x}_{k}\approx &\sqrt{k} +\frac{G_k}{2\sqrt{\beta}}+\frac{a+1}{4\sqrt{k}} +\frac{G^{(2)}_k}{2\beta\sqrt{k}}-\frac{G_{k}^2}{8\beta\sqrt{k}},\qquad \mathbf{y}_{k-1}\approx \sqrt{k} -\frac{G_k}{2\sqrt{\beta}}-\frac{ a+1}{ 4\sqrt{k}}-\frac{G^{(2)}_k}{2\beta\sqrt{k}}-\frac{G_{k}^2}{8\beta\sqrt{k}} . \end{align} By a calculation as in \cite{ES07}, this suggests the limit $\sqrt{n} B_n \rightarrow \mathcal{L},$ where $\mathcal{L}= \sqrt{x}\frac{d}{dx} +\frac{W'(x)}{\sqrt{\beta}} + \frac{a+1}{2\sqrt{x}}$. The expression $\mathcal{L}\mathcal{L}^t$ does not even make sense because it has a term $W'(x)^2/\beta$. However, $\mathcal{L}= \sqrt{x}e^{I(x)} (\frac{d}{dx}+\frac{a+1}{2x})\circ e^{-I(x)}$ where $I(x)=\int_{x}^1 \frac{dW(y)}{\sqrt{\beta y}}$ and $\circ$ indicates composition of differential operators. This leads us to the change of basis $e^{I(x)}\mathcal{L}\mathcal{L}^t e^{-I(x)}=\mathcal{B}$, and $\mathcal{B}$ turns out to be the operator in equation (\ref{SBO1}). Using the approximations for $x_k$ and $y_{k-1}$, a formal calculation suggests $ \prod_{k=\lfloor x n\rfloor}^n \frac{y_k}{x_k}\sqrt{1+a/k} \rightarrow e^{I(x)}, $ and thus the change of basis in equation (\ref{basis3892}) is a natural discretization of the conjugation leading to the operator $\mathcal{B}$. }. We have not yet defined $\mathbf{x}_k$ and $\mathbf{y}_k$, but clearly they will be related to $A$ by \begin{align} \label{wtaEntries} nA_{k,k-1}=- \mathbf{y}_{k-1}^2 \sqrt{1+\frac{a}{k-1}},\qquad nA_{k,k}= \mathbf{y}_{k-1}^2 + \mathbf{x}_{k}^2,\qquad n A_{k,k+1}=-\frac{\mathbf{x}_{k}^2 }{\sqrt{1+\frac{a}{k}}}. \end{align} Define $\mathbf{x}_k$ and $\mathbf{y}_k$ by \begin{align} \label{HDef} H_k =& \frac{-1}{4n^2} \left(\frac{a^2}{4x_k}+ \frac{a\sqrt{n} G_{k}}{ \sqrt{\beta x_k} }+ \frac{ aG^{(2)}_k}{\beta x_{k}} \right)\\ A_{k,k+1}=& -4 (1-r_k) p_k \exp(H_k), \qquad A_{k,k-1}= -4 (1-r_k) q_k \exp(H_k). \label{q0091} \end{align} To be clear: substituting (\ref{wtaEntries}) in (\ref{q0091}) give formulas which we take as the definitions of $\mathbf{x}_k,\mathbf{y}_{k-1}$. Recall that the $r_k$ and $p_k$ were defined in section \ref{DiscreteProcessSection} and that $q_k=1-p_k$. This defines finishes the construction of our matrix model. In the next two subsections we show that the matrices $A^n$ have the following properties: \begin{proposition} \label{HardEdgeProp} For each $k$, the first $k$ eigenvalues of $n^2 A^{(n)}$ converge in distribution to those of the stochastic Bessel operator. \end{proposition} \begin{proposition} \label{DiscFKProp} There exists a functional $R(t)$ such that on the event $\ul{X}(t)\geq 1/\log n$ we have $|R(t)| \leq C(\log n)^3/\sqrt{n}$ such that the following discrete Feynman-Kac formula holds: \begin{align} ((I-A^{(n)}/4)^m v)_k =\mathbb{E}_{x_k}\left[\one_E \exp( \Phi_n(t_m)+R(t_m))v(n X^n(t_m))\right]. \label{sg4} \end{align} where $v$ is any vector in $\mathbb{R}^n$ and $\Phi_n$ is the functional in equation (\ref{phiFormula}). \end{proposition} \textbf{Comment.} Since the eigenvalues of $A^{(n)}$ are of order $n^{-2}$ we will take $m=4tn^2$ in equation (\ref{sg4}), so that for any $k$ the first $k$ eigenvalues of the operator on the left hand side converge to $(\exp (-t \Lambda_i))_{i=1}^k$ where $\Lambda_i$ are the eigenvalues of the SBO. Because the matrices $A^{(n)}$ are not symmetric, it is not obvious that the matrices on the left hand side of (\ref{sg4}) converge to $\exp (-t\mathcal{B})$; but it is at least clear that these are the right matrix powers to consider, if one hopes to extend the semigroup method to the random matrix hard edge. \subsection{Discrete Feynman-Kac formula for $A^{(n)}$} \label{discFKSection2} \begin{proof}[Proof of proposition \ref{DiscFKProp}] Let $M=I-A^{(n)}/4$. It is well known that the entries of a power of a tridiagonal matrix can be expressed using a sum over lattice paths. Let $\mathcal{P}^{m}_k$ be the set of functions $p:\;\{0,\ldots ,m\}\rightarrow \mathbb{Z}$ such that $p(0)=k$ and satisfying the conditions $p(i)-p(i-1) \in \{ -1,0,1\}$ for all $i=1,\ldots ,m$. Let $E$ be the event that the range of $p$ is a subset of $\{1,\ldots ,n\}$. Then \begin{align} (M^m v)_k =\sum_{p \in \mathcal{P}^{m}_k} \one_E M_{p(0),p(1)} M_{p(1),p(2)}\ldots M_{p(m-1),p(m)}v_{p(m)} . \end{align} Letting $p(k) =n X^n(t_k)$, we can recognize the right hand side above as an expectation with respect to the process $X^n$ by factoring $M_{k,k+\Delta}= P(k,\Delta) e^{H(k,\Delta)}$ where $P(k,\Delta)$ is the probability that the next step of $X^n$ is an increment $\Delta/n$ given that the present location is $x_k$. This defines the quantities $H(k,\Delta)$. We now have \begin{align} (M^mv)_k =\mathbb{E}_{x_k}\left[ \one_{E} v_{nX(t_m)}\exp \sum_{i=1}^m H(n X(t_i),\Delta X_i)\right] . \end{align} Therefore it suffices to show that $H(k,\Delta)=H(k)+O((\log(k) k^{-5/2})$ for all $\Delta$, where $H(k)$ is the function defined in equation (\ref{HDef}). Actually, definitions (\ref{wtaEntries}-\ref{q0091}) are cooked up precisely so that \begin{align} e^{H(k,1)} = \frac{-A_{k,k+1}/4}{(1-r_k)p_k}=e^{H_k},\qquad e^{H(k,-1)} = \frac{-A_{k,k-1}/4}{(1-r_k)q_k}=e^{H_k}, \end{align} where $H_k$ is defined in equation (\ref{HDef}). Thus it remains only to calculate $H(k,0)$. Equations (\ref{wtaEntries}-\ref{q0091}) give the relation \begin{align} r_k e^{H(k,0)} =1- \left( q_k/\sqrt{1+a/(k-1)} +p_k \sqrt{1+a/k}\right) (1-r_k)e^{H_k}. \label{aeorgibae} \end{align} A short calculation using formula (\ref{pkExpansion}) and Newton's binomial theorem gives \begin{align} q_k/\sqrt{1+a/(k-1)} +p_k \sqrt{1+a/k} =1- \frac{2n}{k} H_k +O\left(\frac{\sqrt{\log k}}{k^{5/2}}\right). \end{align} Substituting this in (\ref{aeorgibae}) yields $H(k,0)=H_k+O(\sqrt{\log k}/ k^{5/2})$. \end{proof} \subsection{Hard edge for the matrix model} \label{HardEdgeSection} We now show that our matrix model has a stochastic Bessel operator limit. By this we mean that for any $k$, the smallest $k$ eigenvalues of $A^{(n)}$ converge in joint distribution to those of the SBO. In the case that the matrix entries are independent, the technique established in \cite{RR09} for proving convergence to the SBO is quite robust. Therefore we will just give the essential calculations here, and point the reader to \cite{RR09} for the technical details. It is equivalent to show that the smallest $k$ singular values of $B=B^{(n)}$ converge in distribution to those of the operator \begin{align} \mathcal{L}=\sqrt{x} \frac{d}{dx} +\frac{a+1}{2\sqrt{x}} +\frac{1}{\sqrt{\beta}}W'(x). \end{align} The inverse of $\mathcal{L}$ is the following integral operator on $L^2[0,1]$: \begin{align} (\mathcal{L}^{-1}f)(x) =& \int_{0}^1 K(x,y)f(y)\, dy,\qquad K(x,y) =\one_{y\leq x} \frac{1}{\sqrt{x}} (x/y)^{a/2} \exp \int_{y}^x \frac{dW(z)}{\sqrt{\beta z}} \end{align} We check this by computing an integral kernel for $B^{-1}$, starting with the formula \begin{align} (B^{-1})_{i,j} =& \frac{1}{\mathbf{x}_j}\prod_{k=j+1}^{i} \frac{\mathbf{y}_{k-1}}{\mathbf{x}_k}. \end{align} Recall that $\mathbf{x}_k$ and $\mathbf{y}_k$ were defined in formulas (\ref{wtaEntries}-\ref{q0091}). Unpacking these definitions, we get \begin{align} \frac{\mathbf{y}_{k-1}^2}{\mathbf{x}_{k}^2} =& \frac{A_{k,k-1}(1+a/(k-1))^{-1/2}}{A_{k,k+1}(1+a/k)^{1/2}} = \frac{1}{ 1+a/k } \frac{q_k }{p_k }(1+O(k^{-2})) \nonumber \\ =& \frac{1}{1+a/k} \frac{s_{k+1}-s_k}{s_{k}-s_{k-1}}(1+O(k^{-2})) . \end{align} Recall that $s_k=s(x_k)$ are values of the scale function for $X$. For the scale ratio we make the following approximation: \begin{align} \prod_{k=j+1}^i \frac{s_{k+1}-s_k}{s_{k}-s_{k-1}} = &\frac{s_{i+1}-s_i}{s_{j+1}-s_j} =\exp\bigg(\int_{x_{j+1}}^{x_{i+1}} dV\bigg) \frac{ \int_{x_{i}}^{x_{i+1}} \exp(-\int_{y}^{x_{i+1}} dV)\, dy} { \int_{x_{j}}^{x_{j+1}} \exp(-\int_{y}^{x_{j+1}} dV)\, dy} \nonumber \\ =& \exp(V(x_i)-V(x_j)) (1+O(j^{-1/2})) \nonumber \\ =& \frac{x_j}{x_i } \exp\left( \int_{x_j}^{x_i} \frac{2\,dW(y)}{\sqrt{\beta y}}\right)(1+O(j^{-1/2})). \end{align} Also we have \begin{align} \prod_{k=j+1}^i \frac{1}{1+a/k} =\exp \sum_{k=j+1}^i \left(\frac{-a}{k} +O(k^{-2})\right) =& (x_{j}/x_{i})^{a}(1+O(j^{-1})). \end{align} Finally \begin{align} \mathbf{x}_{k}^2 =4(1-r_k )p_k \exp(H(k))\sqrt{1+a/k} = x_k (1+O(k^{-1/2})). \end{align} Our discussion so far shows that if $j\leq i$ then \begin{align} (B^{-1})_{ij} =& \frac{1}{\sqrt{x_i}} \left( x_{j}/x_{i} \right)^{a/2} \exp \left( \int_{x_j}^{x_i} \frac{dW(y)}{\sqrt{\beta y}} \right)(1+O(k^{-1/2})).\label{kernel7} \end{align} Define a linear mapping of $\mathbb{R}^n$ into $L^2[0,1]$ sending the $k^{th}$ basis vector to $\one_{[x_{k-1},x_k]}/\sqrt{n}$. This identifies the matrices $B^{-1}$ with integral operators on $L^2[0,1]$. The integral operators have piecewise constant kernel functions $K(x,y)$ given by equation (\ref{kernel7}) on the rectangle $[x_{i-1},x_i]\times [x_{j-1},x_j]$. Thus the discrete integral kernels converge pointwise to the limiting kernel from \cite{RR09}. \section{Hard edge to soft edge transition} \label{EdgeTransitionSection} In this section we let \begin{align} \alpha=a/2. \end{align} Letting $\mathcal{B}_{a,\beta}(x)$ and $\mathcal{A}_\beta(x)$ respectively be the operators in equation (\ref{SBO1}) and (\ref{SAO1}), a formal calculation suggests that if one lets $x=1-\alpha^{-2/3}z$ then \begin{align} \alpha^{-4/3}\left( \mathcal{B}_{a,\beta}(x) -\alpha^2 \right)\rightarrow \mathcal{A}_{\beta}(z)\qquad \text{as}\qquad a\rightarrow \infty. \end{align} In particular one expects that under this scaling the smallest eigenvalue of the SBO should converge to that of the SAO. This was in fact shown by Ramirez and Rider \cite{RR09}, up to a slightly different scaling which should not affect the limit. Using a Ricatti transformation, they showed that the eigenfunction equation for the SBO is equivalent to a diffusion process; they were able to show that this diffusion converges to one characterizing SAO eigenfunctions as $a\rightarrow \infty$. It was shown in \cite{DTM16} that if $a=C\sqrt{n}$ then the smallest eigenvalue of the Laguerre unitary ensemble converges, properly scaled, to a Tracy-Widom distribution, i.e.\ the smallest eigenvalue of the SAO. The authors of \cite{DTM16} rely on Riemann-Hilbert methods, and therefore their technique is specific to the classical parameter values $\beta=1,2,4$. First in section \ref{LimitTransitionSection} we will formally show how the SBO semigroup degenerates to the SAO semigroup. Within the method of \cite{RR09} it is not convenient to let $a$ depend on $n$, so it would be interesting to obtain a result describing how $a$ can depend on $n$ if one is to observe an SBO to SAO transition. In section \ref{FiniteNTransitionSection} we explore this point from the semigroup perspective. \subsection{Transition of limiting Feynman-Kac formulas} \label{LimitTransitionSection} We now give some partial results which suggest how the edge transition problem can be approached from the semigroup point of view. First decompose the SBO as a generator plus source term \begin{align} -\mathcal{B} =& \frac{1}{2}\sigma^2 \partial_{x}^2 +\mu \partial_x - K;\\ \sigma=&\sqrt{2x}, \qquad \mu= 1+2\sqrt{ \frac{x}{\beta}}W'(x) ,\qquad K= -\frac{\alpha^2}{ x}- \frac{2\alpha W'(x)}{\sqrt{\beta x}}. \end{align} Assuming that the Feynman-Kac formula holds for the SBO, we have \begin{align} &\exp\left( -t \alpha^{-4/3} (B-\alpha^2 I) \right)f(\alpha^{2/3}(x-1)) \label{transition1} \\ &\hspace{70pt}=\mathbb{E}_x \exp\bigg(-\int_{0}^{\alpha^{-4/3}t} K(X_u)+\alpha^2\, du \bigg) f(\alpha^{2/3}(X_{\alpha^{-4/3}t}-1)) \nonumber. \end{align} For each $a>0$, define new stochastic processes $\wt{B}$ and $w$ by \begin{align} \wt{B}(t) =- \alpha^{2/3} B(2\alpha^{-4/3} t), \qquad w(z) =\alpha^{1/3}W(1-\alpha^{-2/3}z). \label{ScaledProcessDef} \end{align} Note that $\wt{B}$ is a Brownian motion with a diffusion coefficient of $\sqrt{2}$. We now state a result which suggests that the right hand side of (\ref{transition1}) should converge to $\exp(-t\mathcal{A} )f(x)$ as $a\rightarrow \infty$, where $\mathcal{A}$ is the SAO. Note that the authors of \cite{GS16} consider $\exp(-t\mathcal{A}/2)$, so our formulas will differ by some powers of $2$. \begin{proposition}\label{SemigroupConvProp} Suppose $X_0=1-\alpha^{-2/3}z$ where $|z|\leq \log \alpha $, then \begin{align} \left| \alpha^{2/3}(1-X_{\alpha^{-4/3}t}) - \wt{B}_t \right| \leq & C a^{-1/3}(\log a)^2 \\ \int_{0}^{\alpha^{-4/3}t} K(X_u)+\alpha^2\, du \rightarrow & \int_{0}^t \wt{B}_u \, du+\frac{2}{\sqrt{\beta}}\int_{0}^\infty L_{\wt{B}}(x,t)\, dw(x) \end{align} \end{proposition} To prove the above proposition we will use the following two lemmas. \begin{lemma} \label{ShortTimeLemma} For almost all $W$ there exists a finite constant $C=C(W,\beta)$ such that if $|z|\leq \log \alpha$ then \begin{align} \left| \alpha^{2/3}s(1-\alpha^{-2/3}z) - z \right| \leq C \alpha^{-1/3} (\log \alpha)^2 . \end{align} For almost all $W,B$ there exists a finite constant $C=C(W,B,\beta)$ such that if $|Z_0 |\leq \log \alpha$ and $T\leq \log \alpha$, then \begin{align} \left| \alpha^{4/3}T^{-1}(\alpha^{-4/3} t) -2 t\right| \leq & C \alpha^{-1/3} (\log a)^2. \end{align} \end{lemma} The proof of lemma \ref{ShortTimeLemma} is a direct but tedious calculation, so we will postpone it until the end of the section. We will need some continuity estimates for Brownian local time. It is easy to find a variety of very sharp estimates in the literature, for example \cite{BorodinSurvey}, but not exactly in forms that are convenient for us. However for our purposes it is not important to have an optimal modulus of continuity, and therefore the lemma we need is accessible using the method developed in \cite{McKeanLocalTime}. Since the method is relatively old and well understood, and since the following is not a very ambitious statement, we will not give a proof. \begin{lemma} \label{LocalTimeLemma} For almost all $B$ the following quantities are finite: \begin{align} \sup_{t \geq 0} \sup_{\substack{x,y\in \mathbb{R}\\ |x-y|<1}} \frac{|L_B(x,t)-L_B(y,t)|}{1 \vee t^{0.251} |x-y|^{0.499}},\qquad \sup_{\substack{t,t'\geq 0\\ |t-t'|<1}} \sup_{x\in \mathbb{R}} \frac{|L_B(x,t)-L_B(x,t')|}{1 \vee t^{0.01} |t-t'|^{0.499}}. \end{align} \end{lemma} \begin{proof}[Proof of proposition \ref{SemigroupConvProp}] Throughout the proof we will see a proliferation of error terms that are bounded in absolute value by expressions of the form $C(B,W,\beta) \alpha^{p}(\log \alpha)^{q}$, which we will abbreviate using the notation $O(\alpha^{p})^-$. First we prove that the rescaled process is approximately a Brownian motion. Our convention is that $B$ starts at $X_0$. But to understand how the starting position affects our calculation, just for now we let $X_t=s^{-1}(X_0 +B_{T^{-1}t})$ where $B$ starts at $0$. \begin{align} \alpha^{2/3}\left(1-X_{\alpha^{-4/3}t} \right) =& \alpha^{2/3} \left(1- s^{-1}(s(X_0)+ B_{T^{-1}(\alpha^{-4/3}t)}\right) \\ = & \alpha^{2/3}\left(1-s^{-1} \left( \alpha^{-2/3}z +O(a^{-1})^- +B_{2 \alpha^{-4/3}t +O(\alpha^{-5/3})^-}\right)\right)\nonumber \\ =& \alpha^{2/3}\left(1-s^{-1} \left( \alpha^{-2/3} z+ B_{2 \alpha^{-4/3}t } + O(\alpha^{-5/6} )^-\right) \right)\label{lil111} \\ =& \alpha^{2/3} \left(1 -\left(1-\alpha^{-2/3}z -\alpha^{-2/3} \wt{B}_{t } + O(\alpha^{-1})^-\right)\right) .\nonumber \end{align} For line (\ref{lil111}) we used the law of the iterated logarithm. For the other steps we used lemma \ref{ShortTimeLemma}. Now we turn our attention to the convergence of the functionals. The first term is easy: \begin{align} &\int_{0}^{\alpha^{-4/3}t} \frac{\alpha^2 }{X_u}-\alpha^2 \, du = \alpha^2\int_{0}^{\alpha^{-4/3}t} \frac{1}{1-\alpha^{-2/3}( \wt{B}_{\alpha^{4/3}u}+O(\alpha^{-1/3})^- )}-1 \, du \\ &\qquad = \alpha^{2} \int_{0}^{\alpha^{-4/3}t} \alpha^{-2/3}\wt{B}_{\alpha^{4/3}u} \, du +O(\alpha^{-1/3})^- = \int_{0}^t \wt{B}_u \, du +O(\alpha^{-1/3})^- . \end{align} Now for white noise integral term. Using the interpretation defined in the introduction, we have \begin{align} \int_{0}^{\alpha^{-4/3}t} \frac{W'(X_u)}{\sqrt{X_u}}\, du :=&\int_{0}^1 \frac{L_B(s(x),T^{-1}(\alpha^{-4/3}t))}{2\sqrt{x}e^{I(x)}} d\ola{W}(x) +\int_{0}^1 \frac{L_B(s(x),T^{-1}(\alpha^{-4/3}t))}{2\sqrt{\beta }x e^{I(x)}} dx. \end{align} We now change variables by $x=1-\alpha^{-2/3}z$. Since $W(x) =\alpha^{-1/3} w(\alpha^{2/3}(1-x))$, the Jacobian factor is $\alpha^{-1/3}$ in the stochastic integral, but $\alpha^{-2/3}$ in the ordinary integral. Therefore we can ignore the ordinary integral. \begin{align} \alpha\int_{0}^1 \frac{L_B(s(x),T^{-1}(\alpha^{-4/3}t))}{2\sqrt{x}e^{I(x)}} d\ola{W}(x) = \alpha^{2/3}\int_{0}^{\alpha^{2/3}} \frac{L_B(s(1-\alpha^{-2/3} z),T^{-1}(\alpha^{-4/3}t))}{2\sqrt{1-\alpha^{-2/3}z}e^{I(1-\alpha^{2/3}z)}} d\ola{w}(z) \label{aeorug} \end{align} We now approximate the local time using lemmas \ref{ShortTimeLemma} and \ref{LocalTimeLemma}: \begin{align} L_B(s(1-\alpha^{-2/3} z),T^{-1}(\alpha^{-4/3}t)) =& L_B( \alpha^{-2/3} z +O(\alpha^{-1})^-,2\alpha^{-4/3}t+O(\alpha^{-5/3})^-)\\ =&2\alpha^{-2/3}L_{\wt{B}}( z +O(\alpha^{-1/3})^-, t+O(\alpha^{-1/3})^- ) \\ =& 2\alpha^{-2/3}L_{\wt{B}}( z , t )+O(\alpha^{-1/6})^-. \end{align} Applying the above result in the integral (\ref{aeorug}) we see that it equals \begin{align} \alpha^{2/3}\int_{0}^{\alpha^{2/3}} \frac{ 2\alpha^{-2/3} L_{\wt{B}}( z , t)}{2 +O(\alpha^{-1/3} )^-} d\ola{w}(z) +O(\alpha^{-1/6})^- = \int_{0}^{\infty} L_{\wt{B}}( z , t )\, dw(z) +O(\alpha^{-1/6})^- . \end{align} Since $L_{\wt{B}}(z,t)$ does not depend on $w$, it does not matter whether we write the last integral as Ito or anti-Ito. Clearly we incur a negligible error by extending the domain of integration to infinity. \end{proof} \subsection{Edge transition with $a$ depending on $n$}\label{FiniteNTransitionSection} We begin by decomposing our functionals as follows: \begin{align} \label{ABDecomposition} \Phi_n =& a^2\Phi^{A}_n + a \Phi^{B}_n, \qquad \Phi^{A}_n = \frac{-1}{4n^2} \sum_{i=1}^{\lfloor 4tn^2 \rfloor} \frac{1}{4x(i)},\qquad \Phi^{B}_n= \frac{-1}{4n^2} \sum_{i=1}^{\lfloor 4tn^2 \rfloor} \frac{ \sqrt{n}G_{n x(i)}}{\sqrt{\beta x(i)}} +\frac{ G^{(2)}_{nx(i)} }{\beta x(i)} \\ \Phi =& a^2\Phi^{A} + a \Phi^{B} , \qquad \Phi^{A} = -\int_{0}^t \frac{1}{4X_s} ,\qquad \Phi^B = -\int_{0}^t \frac{ W'(X_s)}{\sqrt{ \beta X_s}}\, ds. \end{align} Also let $\Psi$ be the limiting functional for the SAO semigroup: \begin{align} \Psi =-\int_{0}^t \wt{B}_u \, du - \frac{2}{\sqrt{\beta}}\int_{0}^{\infty } L_{\wt{B}}(x,t) \, dw(x), \end{align} where $\wt{B}$ and $w$ were defined in display (\ref{ScaledProcessDef}). In the following proposition we write our functionals with a second argument to indicate the starting position of the process. So $\Phi(t,x)$ means the functional $\Phi$ integrating to time $t$, with the process $X$ started at $x$. Also we use the notation $\ul{X}(t)=\inf_{u\leq t}X_u$. \begin{proposition} Allow $a$ to depend on $n$ in such a way that $a(n)=o(n^{0.249})$. For almost all $B,W$ and any $z$, on the event $\ul{X}(\alpha^{-4/3}t)\geq 1/\log n$ we then have \begin{align} \left| \Phi_n(\alpha^{-4/3}t, 1-\alpha^{-2/3}z) -\Psi(t,z) \right| \rightarrow 0 \qquad \text{as}\qquad n \rightarrow \infty. \end{align} \end{proposition} Note that the hypothesis on $\ul{X}$ will be satisfied almost surely if $a(n)\geq n^p$ for some $p>0$. \begin{proof} Our assertion is basically a corollary of the main results of later sections of the paper. Suppressing the arguments of the functionals we estimate \begin{align} \left|\Phi_n -\Psi \right| \leq & a(n)^2 \left| \Phi^{A}_n -\Phi^A \right| +a(n)\left| \Phi^{B}_n -\Phi^{B}\right| +\left| \Phi -\Psi \right|. \end{align} In proposition \ref{SemigroupConvProp} we showed that the third term converges to zero. Proposition \ref{ATermsProposition} asserts that if $\ul{X}(\alpha^{-4/3}t)\geq 1/\log n$, then $\Phi^{A}_n -\Phi^A =O(n^{-0.499})$. Thus $a(n)^2 \left|\Phi^{A}_n -\Phi^A\right|$ converges to zero. Applying propositions \ref{PhiNPhiTildeThm} and \ref{WtPhiConverges}, we have $\Phi^{B}_n -\Phi^B =O(n^{-0.249})$ and thus $a(n)\left|\Phi^{B}_n -\Phi^B\right|$ converges to zero. \end{proof} \subsection{Proof of lemma \ref{ShortTimeLemma}} \begin{proof}[Proof of lemma \ref{ShortTimeLemma}] We use the representation \begin{align} X_t=& s^{-1}(B_{T^{-1}(t)}), \qquad s(x)=-\int_{x}^1 e^{V(y)}\, dy ,\qquad T(t)=\int_{0}^t \frac{du}{(s'\cdot \sigma)^2\circ s^{-1}(B_u)}, \end{align} where $B$ is a Wiener process started from $s(X_0)$. By the reflection principle $\sup_{t\leq \alpha^{-4/3}\log \alpha} B_t -B_0$ is normal with mean zero and variance $\alpha^{-4/3}\log \alpha$, and so by a standard tail bound and the Borel-Cantelli lemma \begin{align} \sup_{t\leq \alpha^{-4/3}\log \alpha} |B_t -B_0|\leq 2\alpha^{-2/3}\log \alpha \end{align} for all sufficiently large $\alpha$. Recall that \begin{align} s'(x) =x^{-1} e^{I(x)},\qquad \text{where}\qquad I(x)=\int_{x}^1 \frac{2 dW(y)}{\sqrt{\beta y}} =\frac{2}{\sqrt{\beta}} \wt{W}_{-\log x} ,\label{sPrimeFormula} \end{align} and $\wt{W}$ is another Wiener process. Using once again that the running maximum of $\wt{W}_\xi$ is a normal random variable, for sufficiently large $\alpha$ we have \begin{align} \sup_{ x\in [1-\alpha^{-2/3}\log \alpha,1]} \wt{W}_{-\log x} \leq \sup_{ \xi\leq 2 \alpha^{-2/3}\log \alpha} \wt{W}_{\xi } \leq 2\alpha^{-1/3}\log \alpha\qquad \text{a.s.} \end{align} Taylor expanding the exponential in (\ref{sPrimeFormula}) and using the above estimate, we get the bound \begin{align} |s'(x) -1|\leq \alpha^{-1/3}\log \alpha \qquad \text{for all }x\in [1-\alpha^{-2/3}\log \alpha,1]\text{ if $\alpha$ is large enough}. \end{align} The bounds we asserted on the scale function and its inverse then follow easily. We now approximate the derivative of the time change function. For $u\leq \alpha^{-4/3} \log \alpha$ we have so far shown that \begin{align} |B_u -B_0|\leq \alpha^{-2/3}\log \alpha ,\qquad s^{-1}(B_u)-1 =B_u-B_0 +O(\alpha^{-1}(\log \alpha)^2) =O(\alpha^{-2/3}\log \alpha). \end{align} Using $\sigma(x)=\sqrt{2x}$ and the above approximation for $s'$, it follows that if $a$ is sufficiently large then for all $t\leq \alpha^{-4/3}\log \alpha$ we have \begin{align} T'(t)= \frac{1}{2} +O(\alpha^{-1/3}\log \alpha). \end{align} We conclude that \begin{align} T^{-1}(t) -2t =O(a^{-5/3}(\log \alpha)^2)\qquad \text{for }t\leq \alpha^{-4/3}\log \alpha. \end{align} \end{proof} \section{Convergence of the noiseless terms} \label{NoNoiseSection} Recall that in display (\ref{ABDecomposition}) we decomposed our functionals as \begin{align} \Phi_n =& a^2\Phi^{A}_n + a \Phi^{B}_n, \qquad \Phi = a^2\Phi^{A} + a \Phi^{B} , \end{align} where the ``$A$" and ``$B$" terms do not depend on $a$. We now further divide the task of comparing $\Phi_n$ and $\Phi$ into two steps as follows: \begin{align} \label{MakePhiTilde} |\Phi -\Phi_n| \leq & |\Phi -\wt{\Phi}_n| + | \wt{\Phi}_n-\Phi_n|, \qquad \text{where } \wt{\Phi}_n = a^2 \wt{\Phi}^{A}_n +a \wt{\Phi}^{B}_n \\ \wt{\Phi}^{A}_n (t) =& -\sum_{k=1}^n \frac{1}{4 x_k}L_{X}([x_k,x_{k+1}) ,t),\\ \wt{\Phi}^{B}_n (t) =& -\sum_{k=1}^n \left(\frac{ \sqrt{n} (G_{k+1}+G_k)}{2\sqrt{\beta x_k}} +\frac{ G^{(2)}_k}{ \beta x_k} \right) L_{X}([x_k,x_{k+1}) ,t). \end{align} It is much more difficult to analyze the ``$B$" terms (the white noise/ local time integral and its discretization). In this section we carry out the easier task of analyzing the noiseless ``$A$" terms; we will prove the following. \begin{proposition} \label{ATermsProposition} On the event $\ul{X}(t)\geq 1/\log n$ we have \begin{align} |\Phi^{A}_n -\Phi^A | \leq C t n^{-0.499}. \end{align} This holds for almost all $B,W$, and $C=C(B,W,\beta)$. \end{proposition} \subsection{Estimating $|\wt{\Phi}^{A}_n -\Phi^{A}_n|$} \label{SectionAAN} Recall that the steps of $X^n$ are equal to increments of $X$ between the stopping times $\t_i$ constructed in section \ref{DiscreteProcessSection}. However, it will sometimes be convenient to only consider times at which $X^n$ jumps (i.e.\ not a $\rightarrow$ step). Therefore define a sequence of times $\tau_i$ by letting $\tau_0=0$ and \begin{align} \tau_{i}=\inf\{ t_j\geq \tau_{i-1}:\; X^{n}(t_{j})\neq X^{n}(t_{j-1})\}. \end{align} Also, we define another sequence of times $\wt{\tau}_i$ by \begin{align} \wt{\tau}_i=\t_j \qquad \text{if and only if}\qquad \tau_i=t_j. \end{align} So $\tau_i$ are the jump times for $X^n$, and $\wt{\tau}_i$ are the jump times for $X$. We also introduce a notation for the sequence of lattice points visited after each jump: \begin{align} x(i) =X^n(\tau_i) =X(\wt{\tau}_i). \end{align} Notice that increments of $\Phi_n$ depend on the current position of $X^n$, which is the most recent lattice point visited by $X$; whereas $\wt{\Phi}_n$ depends on which interval $[x_k,x_{k+1})$ the process $X^n$ lies in. Therefore during each jump it is important to distinguish between time spent by $X$ above and below the lattice point $x_k$, so we define \begin{align} \Delta\tau_i=&\tau_{i+1}-\tau_{i},\qquad \Delta \wt{\tau}_i = \wt{\tau}_{i+1}-\wt{\tau}_i, \\ \Delta \wt{\tau}^{\uparrow}_i =&\text{meas}\,\{t\in [\wt{\tau}_{i},\wt{\tau}_{i+1}):\; X\geq x(i),\\ \Delta \wt{\tau}^{\downarrow}_i =&\text{meas}\,\{t\in [\wt{\tau}_{i},\wt{\tau}_{i+1}):\; X< x(i). \end{align} Here ``meas" is Lebesgue measure. With all this notation in hand, we can state the following formula which is clear from the definitions: \begin{align} \Phi^{A}_n(\tau_j) - \wt{\Phi}^{A}_n(\wt{\tau}_j) =& \sum_{i=1}^{j}\xi_i ,\qquad \xi_i= \frac{\Delta\tau_i -\Delta\wt{\tau}_i}{x(i)} + \Delta\wt{\tau}^{\downarrow}_i \left(\frac{1}{x(i)} -\frac{1}{x(i)-1/n}\right) . \end{align} Since we have evaluated the functionals at different times, we take on the burden of also bounding $|\Phi^{A}_n(\tau_j) - \Phi^{A}_n(\wt{\tau}_j)|$, but this will be relatively easy. The first step is to compute the means of the time increments. The approximation we prove here is stronger than we need for this section but will be necessary later. Define the following variables: \begin{align} \gamma_k=& 4n^{5/2} \int_{x_k}^{x_{k+1}} (x-x_k)(x_{k+1}-x)\, dW(x). \end{align} Note that with the explicit factor of $n^{5/2}$ the $\gamma_k$'s are typically order 1. \begin{lemma} \label{exitTimeLemma01} For almost all $W$, all $n\in \mathbb{N}$ and all $k=\frac{n}{\log n},\ldots ,n$ we have \begin{align} \mathbb{E} \left[ \Delta \wt{\tau}^{\uparrow}_i \middle| W,x(i)=x_k \right] =&\frac{1}{x_k}\left( 1 +\frac{G_k+\gamma_k}{\sqrt{\beta k}}\right)\Delta t+O(n^{-3}\log (n)^3), \\ \mathbb{E} \left[ \Delta \wt{\tau}^{\downarrow}_i \middle| W,x(i)=x_k \right] =&\frac{1}{x_k}\left( 1 -\frac{G_k+\gamma_{k-1}}{\sqrt{\beta k}}\right)\Delta t+O(n^{-3}\log (n)^3). \end{align} The implied constants depend only on $W$ and $\beta$. \end{lemma} \begin{proof} By the general theory of one dimensional diffusion process, if the process starts at $x_k$, then the expected time the process spends in any subinterval $A\subset [x_{k-1} ,x_{k+1}]$ before the first exit is \begin{align} \mathbb{E}[T(A)] =& \int_A \frac{ 2(s(y)\wedge s_k -s_{k-1})(s_{k+1} - s (y) \vee s_k)}{(s_{k+1}-s_{k-1})s'(y) \sigma(y)^2} \, dy. \end{align} Here $\wedge,\vee$ precede $+,-$ in order of operations. Letting $A=[x_k,x_{k+1}]$ and substituting some of the definitions given in displays (\ref{speedScale1}) and (\ref{pkExpansion}), we get \begin{align} \mathbb{E}[\Delta\wt{\tau}^{\uparrow}_i|W,x(i)=x_k] =& p_k \int_{x_k}^{x_{k+1}} \int_{y}^{x_{k+1}} \exp\left( \int_{y}^z dV\right) y^{-1} \,dy. \end{align} For the quality of approximation claimed in the lemma statement, it suffices to substitute a linear approximation for the exponential function, and $x_{k}^{-1}$ for the explicit factor of $y^{-1}$. After a straitforward calculation one obtains the first formula in the lemma statement, and the second formula is proven similarly. \end{proof} \begin{lemma} \label{mgfLemma01} Let $M(\lambda) = \mathbb{E}[ \exp(\lambda \xi_i ) |x(i)=x_k ]$. For all $n\in \mathbb{N}$ and $k=\frac{n}{\log n},\ldots n$ we have \begin{align} \left| M(\lambda)-1-\lambda M'(0) \right| \leq \frac{C\lambda^2 (\log n)^2}{k^{4}} \qquad \text{for all }\lambda \leq \frac{k^{2}}{C\log n} . \label{taylorError1} \end{align} \end{lemma} \begin{proof} Suppose a random variable $Q$ satisfies a tail bound \begin{align} P(Q\geq r)\leq C\exp(-c r) \label{tb1}. \end{align} Then for $\lambda< c/2$ the second derivative of the moment generating function for $Q$ can be bounded by \begin{align} M_{Q}''(\lambda) =&\int_{0}^\infty (2q+\lambda q^2)e^{\lambda q} P(Q>q)\, dq \leq\int_{0}^\infty (2q+\frac{c}{2} q^2)Ce^{-c q/2}\, dq = 24C /c^2. \label{TailToMGF} \end{align} We will apply the above bound to $Q=k^{2}\xi_i /\log n$. Since we assumed that $k\geq n/\log n$ we have $1/x(i)\leq \log n$. After conditioning on $x(i)$, $\xi_i$ is random only through $\Delta \tau_i , \Delta \wt{\tau}_{i}^{\uparrow},\Delta \tau_{i}^{\downarrow}$. Since the diffusion coefficient for $X$ is bounded below by $\sqrt{(k-1)/n}$ until it leaves $(x_{k-1},x_{k+1})$, it is clear that the probability of the process exiting in the next $1/(kn)$ time units is bounded below by a constant regardless of the position of $X$ within $[x_{k-1},x_{k+1}]$. At this point it is clear that $k^{2}\xi_i /\log n$ satisfies (\ref{tb1}) with constants $C,c$ depending only on $\beta$, so the lemma statement follows from Taylor's theorem with the remainder in Lagrange form. \end{proof} \begin{lemma} \label{HMartProp01} On the event $\ul{x}(4tn^2)\geq 1/\log n$ we have, for all $n\in \mathbb{N}$ and all positive integers $m\leq 4tn^2$ that \begin{align} \left|\Phi^{A}_n(\tau_m) -\wt{\Phi}^{A}_n(\wt{\tau}_m) \right| \leq C t n^{-0.499} . \end{align} The above holds for almost all $B,W$, and $C=C(B,W,\beta )$. \end{lemma} \begin{proof} Given $x(i)$, each $\Delta \tau_i /\Delta t$ is a geometric random variable with parameter $x(i)/2$. From this and lemma \ref{exitTimeLemma01}, it is easy to see that \begin{align} \left|\mathbb{E}\left[ \xi_i \middle| x(i)\right] \right| \leq C \log( n) (nx(i))^{-5/2}, \end{align} where $C$ depends on $W,\beta$. The factor of $\log n$ came from using elementary techniques to get an almost sure bound on the gaussian random variables $G_k +\gamma_k$ in lemma \ref{exitTimeLemma01}. For $x(i)\geq 1/\log n$ using this mean estimate in lemma \ref{mgfLemma01} gives \begin{align} \mathbb{E}[\exp(\lambda \xi_i)|F_i] =& 1 + \lambda O( (\log n)^p n^{-5/2} ) +\lambda^2 O((\log n)^p n^{-4}) \end{align} for all $\lambda $ of order $n^2/ (\log n)^p $. Here $p$ is a universal constant which in this case can be taken to be 3.5, but throughout the paper we will use this notation to avoid keeping track of irrelevant powers of $\log n$. Define the following notations: \begin{align} \ul{x}(i) =\min_{j=1}^i x(j), \qquad Z_i(\lambda) =\exp\bigg( \lambda \sum_{j=1}^i \xi_j \bigg), \end{align} and let $F_i$ be the $\sigma$-field generated by $\{W,x(1),\ldots ,x(i) \}$. Let $E(i)$ be the event where $ \ul{x}(i)\geq 1/\log n$, and calculate \begin{align} \label{MartCalc1} \mathbb{E}[\one_{E(j)}Z_j|W ] \leq & \mathbb{E}[\one_{E(j)}Z_{j-1} \mathbb{E}[ \exp (\lambda \xi_j )|F_j ]|W ] \\ \leq & \mathbb{E}\left[ \one_{E (j)}Z_{j-1} \middle|W \right]\left(1+C (\log n)^p \left(\lambda n^{-5/2} +\lambda^2 n^{-4}\right) \right) \label{MartCalc2a} \end{align} By induction on $j$, the inequality $e^x\geq 1+x$, and $m\leq 4tn^2$ it follows that \begin{align}\label{mgfBound1} \E[\one_{ E(i)}Z_i|W] \leq \exp\left( C t (\log n)^p \left(\lambda n^{-1/2} +\lambda^2 n^{-2}\right)\right). \end{align} By Chebychev's inequality we get \begin{align} &\log P\left(\Phi^{A}_n(\tau_i) -\wt{\Phi}^{A}_n(\wt{\tau}_i) \geq tn^{-0.499} ,\ul{x}(i)\geq \log(n)^{-1}\middle| W \right)\\ & \leq -\lambda t n^{-0.499} + \log \E[\one_{ E(i)}Z_i|W]. \end{align} Now choose $\lambda =\sqrt{n}/t$ and bound the expectation using (\ref{mgfBound1}). By making a similar argument with $Z_i(\lambda)$ replaced by $Z_i(\lambda)^{-1}$ we get the reverse inequality, thus the absolute value in the inequality \begin{align} \log P\left(\Phi^{A}_n(\tau_i) -\wt{\Phi}^{A}_n(\wt{\tau}_i) \geq t n^{-0.499} ,\ul{x}(i)\geq \log(n)^{-1}\middle| W \right)\leq -c n^{0.001}. \end{align} Applying the Borel-Cantelli lemma finishes the proof. \end{proof} \begin{lemma} On the event $\ul{X}(1.01t)\geq 1/\log n$ we have, for all $n\in \mathbb{N}$ that \begin{align} \left|\Phi^{A}_n(t) -\wt{\Phi}^{A}_n(t) \right| \leq C t n^{-0.499} . \end{align} The above holds for almost all $B,W$, and $C=C(B,W,\beta )$. \end{lemma} \begin{proof} First we claim for almost all $B,W$ and all $i\leq 4t n^2$ that $|\tau_i -\wt{\tau}_i|\leq C tn^{-0.499}$, with the constant depending only on $B,W,\beta$. This can be proven in exactly the same way as lemma \ref{HMartProp01}, so we will not give the details. Let $I$ be the smallest integer such that $\tau_I\geq t$. Since the times $\tau_i$ are a subset of the time lattice points $\mathbb{Z}/(4n^2)$, it follows that $I \leq 4tn^2$. We then estimate as follows: \begin{align} \label{pfEq938} \left|\Phi^{A}_n(t) -\wt{\Phi}^{A}_n(t) \right| \leq & \left|\Phi^{A}_n(t) - \Phi^{A}_n(\tau_I) \right| + \left|\Phi^{A}_n(\tau_I) -\wt{\Phi}^{A}_n(\wt{\tau}_I) \right| + \left|\wt{\Phi}^{A}_n(\wt{\tau}_I) -\wt{\Phi}^{A}_n(t) \right|. \end{align} Recall that given $x(i)$ the jump wait time $\Delta \tau_i$ has a $\text{Geom}[x(i)/2]/(4n^{2})$ distribution. Using this it is elementary to show that $\tau_I-t\leq Cn^{-1.99}$ almost surely, with the constant depending only on $B,W,\beta$. Since $X_n(\tau_I)=X(\wt{\tau}_I)$ and $|\tau_I-\wt{\tau}_I|\leq Ctn^{-0.499}$ we have $\wt{\tau}_I\leq 1.01t$ for all sufficiently large $n$. Therefore the hypothesis $\ul{X}(1.01t)\geq 1/\log n$ implies $\ul{x}_I\geq 1/\log n$. Using this we can bound the first and third terms on the right hand side of (\ref{pfEq938}) within the tolerance of the lemma statement. Using lemma \ref{HMartProp01}, we bound the second term. \end{proof} \subsection{Estimating $|\Phi^{A}-\wt{\Phi}^{A}_n |$} Throughout this section we use the notation \begin{align} I(x)=\int_{x}^1 \frac{2\,dW(y)}{\sqrt{\beta y}}. \end{align} \begin{lemma} \label{TBound} On the event $\ul{X}(t)\geq 1/\log n$ we have the following bounds for almost all $W$: \begin{align} \sup_{1/\log n\leq x\leq 1}| I(x) | \leq & C (\log \log n)^{0.51}, \label{IBound}\\ \sup_{1/\log n\leq x\leq 1}|s'(x)| \leq & C \beta^{-1/2} (\log n)^{1.01}, \\ T^{-1}(t) \leq & C\beta^{-1/2} t (\log n)^{1.02}. \end{align} The constant $C$ depends on $W,\beta$. \end{lemma} \begin{proof} For another Wiener process $\wt{W}$ we have \begin{align} I(x) =\frac{2}{\sqrt{\beta}} \wt{W}_{-\log x}. \end{align} The bound (\ref{IBound}) now follows from the law of the iterated logarithm. Recall that \begin{align} s'(x) = &x^{-1} e^{I(x)} ,\qquad T^{-1}(t) = \int_{0}^t 2 X_{u}^{-1}e^{2 I(X_u)} \, du. \end{align} The other two bounds now follow from the above display and $\ul{X}(t)\geq 1/\log n$. \end{proof} We will use the following bound on the supremum of Brownian local time. \begin{lemma} \label{LocTimeBound} For almost all $B$ we have \begin{align} \sup_{x\in \mathbb{R},t>0} \frac{L_B(x,t)}{\sqrt{t \log t}\vee 1} <\infty . \end{align} \end{lemma} \begin{proof} The key ingredient is inequality (4.6) of A.\ N.\ Borodin's survey \cite{BorodinSurvey}, which is \begin{align} P\left(\sup_{x\in \mathbb{R}} L_B(x,t)>r\right) \leq \frac{A r^2}{t} \exp\left(\frac{-r^2}{2t}\right) \label{BorodinIneq} \end{align} for some constant $A$. We start with a union bound \begin{align} P\left(\sup_{x\in \mathbb{R},t>0} \frac{L_B(x,t)}{\sqrt{t\log t}\vee 1} > R \right) \leq &\sum_{k=0}^{\infty} P\left(\sup_{x\in \mathbb{R}} \frac{L_B(x,k +1)}{\sqrt{\lfloor k \log k\rfloor }\vee 1} > R \right). \end{align} Applying (\ref{BorodinIneq}), we see that the sum on the right hand side converges, and the result is a rapidly decaying function of $R$. Considering only positive integer values of $R$, our assertion follows from the Borel-Cantelli lemma. \end{proof} \begin{lemma} On the event $\ul{X}(t)\geq 1/\log n$, for all $B,W$ we have \begin{align} |\Phi^{A}-\wt{\Phi}^{A}_n |\leq C tn^{-0.99}. \end{align} The constant depends on $B,W,\beta$. \end{lemma} \begin{proof} From the definitions it is easy to see that \begin{align} \Phi^{A}-\wt{\Phi}^{A}_n = \sum_{k=1}^n A_k ,\qquad A_k=\int_{x_{k-1}}^{x_{k}} \left(\frac{1}{x_k} -\frac{1}{x} \right) L_X(x,t) \, dx. \end{align} Using formula (\ref{changeLocTime}) to express local time for $X$ in terms of local time for $B$ gives \begin{align} A_k = \int_{x_{k-1}}^{x_k} \left(\frac{1}{x_{k-1}}-\frac{1}{x}\right) \frac{L_B(s(x),T^{-1}(t))}{2\exp(I(x))}\, dx \label{fpmdt} \end{align} For $k\geq \frac{n}{\log n}$ we have $|x_{k-1}^{-1}-x^{-1}|\leq C\log (n)/n$. The lemma statement now follows immediately by using lemmas \ref{TBound} and \ref{LocTimeBound} to estimate the right hand side of (\ref{fpmdt}). \end{proof} \section{Convergence of noisy terms, part I} \label{section5} Recall that in display (\ref{MakePhiTilde}) we introduced an intermediate functional $\wt{\Phi}_n$, and decomposed the functionals $\Phi,\wt{\Phi}_n,\Phi_n$ into ``noiseless" parts $\Phi^A$, etc., and ``noisy" parts $\Phi^B$, etc. It this section we estimate the difference between the terms \begin{align} \Phi^{B}_n(t)=& \frac{-1}{4n^2} \sum_{i=1}^{\lfloor 4tn^2 \rfloor} \frac{ \sqrt{n}G_{n x(i)}}{\sqrt{\beta x(i)}} +\frac{ G^{(2)}_{nx(i)} }{\beta x(i)}\\ \wt{\Phi}^{B}_n (t) =& -\sum_{k=1}^n \left(\frac{ \sqrt{n} (G_{k+1}+G_k)}{2\sqrt{\beta x_k}} +\frac{ G^{(2)}_k}{ \beta x_k} \right) L_{X}([x_k,x_{k+1}) ,t). \end{align} The results of this section build up to a proof of the following. \begin{proposition} \label{PhiNPhiTildeThm} On the event $\ul{X}(t)\geq 1/\log n$, for almost all $B,W$ we have \begin{align} \left| \Phi^{B}_n(t)-\wt{\Phi}^{B}_n(t)\right| \leq C tn^{-0.249} . \end{align} The constant depends on $B,W,\beta$. \end{proposition} \subsection{Mean increments for $\Phi^{B}_n - \wt{\Phi}^{B}_n$} \label{hMartSection} In this section we will again use the notations $\tau_i,\wt{\tau}_i,\Delta \wt{\tau}^{\uparrow}_i$, etc.\ defined in section \ref{SectionAAN}. We will however reassign the notation $\xi_i$ to the increments that we study here, that is \begin{align} \Phi^{B}_n(\tau_{m}) - \wt{\Phi}^{B}_n( \tau _{m}) =& \sum_{i=1}^m \xi_i \\ \xi_ =& \frac{ G^{(2)}_{k}}{ \beta x_k} \left( \Delta \tau_i - \Delta \wt{\tau}_i \right) + \frac{ G^{(2)}_k}{\beta} \left(\frac{1}{x_{k-1}}-\frac{1}{x_k}\right)\Delta \wt{\tau}^{\downarrow} \label{xiDef2} \\ &\qquad + \frac{ \sqrt{n}(G_k+G_{k-1})}{2\sqrt{\beta}} \left( \frac{1}{\sqrt{x_{k-1}}}-\frac{1}{\sqrt{x_k}}\right) \Delta \wt{\tau}^{\downarrow} \nonumber \\ &\qquad +\frac{ \sqrt{n}}{2\sqrt{\beta x_k}}\left(2G_{k }\Delta \tau_i -(G_{k +1}+G_{k })\Delta \wt{\tau}^{\uparrow}_i -(G_{k }+ G_{k -1} ) \Delta \wt{\tau}^{\downarrow}_i \right).\nonumber \end{align} Let us recall the argument we used in section \ref{NoNoiseSection} to analyze the noiseless ``$A$" terms. In that section the increments we considered had conditional means of order $O(n^{-5/2})$ (forgetting the powers of $\log n$) and variances of order $O(n^{-4})$. Since we had a sum of $O(n^2)$ terms, a central limit theorem heuristic suggests a result with mean of order $O(n^{-1/2})$ and fluctuations of $O(n^{-1})$. Here this heuristic no longer works because, as the next lemma shows, the conditional means of the $\xi_i$'s are on about order $O(n^{-3/2})$: \begin{lemma} \label{meanCor} For almost all $W$, there exists a constant $C=C(W ,\beta )$ such that for all $n\in \mathbb{N}$ and all $k=\frac{n}{\log n},\ldots ,n$, the approximation \begin{align} |\E\left[ \xi_i \middle| W,x(i)=x_k\right] -\mathcal{E}_k |\leq C n^{-5/2}\log( n)^{9/2} \end{align} holds, where \begin{align} \label{EDef} \mathcal{E}_k &= \frac{ \sqrt{n} \Delta t}{2\sqrt{\beta }x_{k}^{3/2}}\bigg\{ G_{k+1}-2G_k +G_{k-1} + \frac{G_{k+1}G_k -G_k G_{k-1}}{ \sqrt{\beta k}} \\ & \hspace{80pt}+ \frac{1}{ \sqrt{\beta k}}\left((G_{k+1}+G_k)\gamma_k -(G_k +G_{k-1})\gamma_{k-1}\right) \bigg\} . \nonumber \end{align} \end{lemma} \begin{proof} Substitute the formulas from lemma \ref{exitTimeLemma01} in formula (\ref{xiDef2}). \end{proof} Therefore at first glance the situation appears desperate-- there is an entire extra power of $n$ we must overcome! However, the argument from section \ref{SectionAAN} at least reduces our task from analyzing a sum of $\xi_i$'s to analyzing the following sum: \begin{align} \mathcal{S}_i =\sum_{j=1}^i \mathcal{E}_{n x(j)} \label{SDef}. \end{align} Indeed we find the following analogue of lemma \ref{HMartProp01}. \begin{lemma} \label{HMartProp} On the event $\ul{X}(1.01t)\geq 1/\log n$ we have, for all $n\in \mathbb{N}$ that \begin{align} \left|\Phi_n(\tau_i) -\wt{\Phi}_n(\wt{\tau}_i)- \mathcal{S}_i\right| \geq C t n^{-0.499}. \end{align} The above holds for almost all $B,W$ and $C=C(B,W,\beta)$. \end{lemma} \begin{proof} The proof is the same as that of lemma \ref{HMartProp01}, so we omit the details. \end{proof} \subsection{Analysis of $\mathcal{S}_i$} \label{BranchingSection} In the previous section we reduced the comparison of $\Phi^{B}_n$ and $\wt{\Phi}^{B}_n$ to the estimation of the sum $\mathcal{S}_i$ defined in formulas (\ref{EDef},\ref{SDef}). Let $D_k(t)$ be the number of downward jumps by $X^n$ from height $x_k$ up to time $t$ and $\ol{D}(t)=\max_{k=1}^n D_k(t)$. In this section we prove the following proposition. \begin{proposition} \label{VMartProp} For almost all $W$, all $n\in\mathbb{N}$, and all positive integers $i$, we have \begin{align} P\left(|\mathcal{S}_i | \geq n^{-0.249}, \ul{x}(i)\geq \frac{1}{\log n},\ol{D}(\tau_i)\leq n (\log n)^3\right) \leq & \exp\left(-cn^{0.5}\right), \end{align} where $c=c(W,\beta)>0$. \end{proposition} Clearly this proposition is only useful together with a bound on the probability that $\ol{D}(\tau_i)$ is large; that will be the subject of section \ref{NStarSection}. \subsubsection{Branching structure for numbers of jumps} Let $J_k(t)$ be the total number of jumps by $X^n$ from height $x_k$ up to time $t$. Then letting $U_k(t)$ and $D_k(t)$ respectively be the numbers of upward and downward jumps from height $x_k$, we clearly have $J_k(t)=U_k(t)+D_k(t)$ and \begin{align} \label{EFormula2} \mathcal{S}_i=\sum_{j=1}^i \mathcal{E}_{nx(j)} =\sum_{k=1}^n J_k\left(\tau_i\right) \mathcal{E}_k.\qquad \end{align} Among the first $i$ jumps, by definition $J_k(\tau_i)$ of them are from $x_k$. It is somewhat redundant to keep track of both upward and downward jump counts since it is easy to see that \begin{align} \label{UFormula1} U_k(t) =&D_{k+1}(t)+\begin{cases} 1 &\text{ if } X^n(0) \leq x_k< X^n(t) \\ -1 & \text{ if } X^n(0)>x_k\geq X^n(t) \\ 0 &\text{ else.} \end{cases} \end{align} A key ingredient in our analysis will be a branching structure for the numbers of upward and downward jumps. However the branching structure can only be applied at appropriate stopping times. Let $T_{i,k'}$ be the time of the $i^{th}$ visit by $X^n$ to height $x_{k'}$. To state the branching structure, we define the following notation: \begin{align} d_k = \begin{cases} 1 & \text{if }X^n(0) \leq x_k \leq x_{k'} \text{ or } X^{n}(0)\geq x_k \geq x_{k'} \\ 0 & \text{else.}\end{cases} \end{align} The branching structure is the following lemma: \begin{lemma} \label{BranchingLemma} Fix positive integers $i,k'$ and let $D_k=D_k(T_{i,k'})$, $U_k=U_k(T_{i,k'})$. There exist independent geometric random variables $\{\gamma_{\ell,k}:\; 1\leq k\leq n,\ell\geq 1\}$ such that: for $k=1,\ldots ,k'-1$ we have \begin{align} \label{branching1} D_{k} =& \sum_{\ell=1}^{D_{k+1}+d_k} \gamma_{k,\ell} ,\qquad \gamma_{k,\ell} +1 \eqd \mathrm{Geom}[p_{k}], \end{align} and for $k=k'+1,\ldots ,n-1$ we have \begin{align} \label{branching2} U_{k} =& \sum_{\ell=1}^{U_{k-1}+d_k} \gamma_{k,\ell} ,\qquad \gamma_{k,\ell}+1 \eqd \mathrm{Geom}[q_{k}]. \end{align} \end{lemma} \begin{proof} First assume that $x_k<x_k'$ and $X^n(0)<x_{k}$. Let $\gamma_{k,1}$ be the number of downward steps from $x_k$ before the first upward step from height $x_{k}$, and for $\ell\geq 1$ let $\gamma_{k,\ell}$ be the number of downward steps from $x_k$ between the $(\ell-1)^{th}$ and $\ell^{th}$ upward steps from $x_k$. Clearly the variables $\{\gamma_{k,\ell}:\;\ell\geq 1\}$ are independent. Also clearly each $\gamma_{k,\ell}+1$ has a geometric distribution with parameter equal to the probability that the next jump from $x_k$ is upwards, which is $p_k$. Since $X^n(0)\leq x_k$, the number of downward steps from $x_k$ before the first downward step from $x_{k+1}$ is $\gamma_{k,1}$ (in the other case this would be zero). Furthermore, the number of downward steps from $x_k$ before the $\ell^{th}$ downward step from $x_{k+1}$ is \begin{align} \sum_{i=1}^{ \ell} \gamma_{k,i}. \end{align} Since $D_{k+1}=D_{k+1}(T_{i,k'})$ and since $k'>k$, there cannot be any downward steps from $x_k$ between $T_{i,k'}$ and the $(D_{k+1}+1)^{th}$ downward step from $x_{k+1}$. Thus the number of downward steps from $x_k$ before $T_{i,k'}$ equals the number before the $(D_{k+1}+1)^{th}$ downward step from $x_{k+1}$. This establishes the formula for $D_{k}$ in the lemma statement for the case we are considering. The proofs of the other possible cases of the lemma are similar. We only comment that if $k>k'$, then $\gamma_{k,\ell}$ is defined instead to be the number of upward steps from $x_k$ between the $(\ell-1)^{th}$ and $\ell^{th}$ downward steps from $x_k$. \end{proof} \subsubsection{Apply summation by parts to $\mathcal{S}_i$} By lemma \ref{BranchingLemma} and since a geometric random variable with parameter $p$ has mean $1/p$, we have \begin{align} \E \left[ D_{k}-D_{k+1} \middle| D_k,W\right] =& \left( \frac{q_{k}}{p_{k}}-1\right)D_k =\left(\frac{2G_k}{\sqrt{\beta k}} +O( \log(n)^3 n^{-1}) \right)D_k. \end{align} For the above approximation we used lemma \ref{pkApproxLemma}. The above calculation suggests that differences of the $D_k$ variables are relatively small. Therefore it is useful to write $\mathcal{S}_i$ in terms of the differences $D_k -D_{k+1}$. In the next lemma we achieve this using summation by parts. There is one subtle detail though: we do not use summation by parts on the term $D_{k+1}f_{2,k}$ in the lemma statement; this is because we need this term ``left behind" so that it can be part of a cancellation in the proof of lemma \ref{EQMGFLemma}. In the following lemma we use the notation $D_\ast$ for a hypothetical bound on the maximum number of downward steps from any height; a reasonable value to use would be $D_\ast =n\log n$. \begin{lemma} \label{sumByPartsLemma} Let $D_k=D_k(\tau_i)$. For almost all $W$, on the event $\{\ol{D}(\tau_i)\leq D_\ast , \ul{x}(i)\geq \frac{1}{\log n}\}$ we have $|\mathcal{S}_i-S_i|\leq C (\log n)^4D_\ast n^{-3/2}$ where \begin{align} \label{fDef} S_i =& \frac{1}{n^2}\sum_{k=1}^n \left\{ (D_{k}-D_{k+1 })(\sqrt{n}f_{1,k} +f_{3,k})+ D_{k+1}f_{2,k} \right\} \\ f_{1,k} =&\frac{ G_{k+1}-G_{k-1} }{8\sqrt{\beta} x_{k}^{3/2}},\qquad f_{2,k}=\frac{1}{4\beta x_{k}^2}G_{k}(G_{k+1}-G_{k-1}) \\ f_{3,k}=&\frac{1}{8\beta x_{k}^2} (G_{k+1}G_k -G_{k}G_{k-1}+\gamma_k(G_{k+1}-G_{k})+\gamma_{k-1}(G_{k}-G_{k-1})). \end{align} \end{lemma} \begin{proof} By formulas (\ref{EFormula2},\ref{UFormula1}) there exist numbers $\wt{d}_k \in\{-1,0,1\}$ such that \begin{align} \mathcal{S}_i =& \sum_{k=1}^n (D_k +D_{k+1}+\wt{d}_k ) \mathcal{E}_k. \label{aaeoringa} \end{align} By the bound $|G_k|\leq \log n$ and the condition $\ul{x}(i) \geq 1/\log n$, we have $|\mathcal{E}_k| \leq C n^{-3/2} (\log n)^4$ almost surely. Therefore the term $\wt{d}_k \mathcal{E}_k$ in equation (\ref{aaeoringa}) can be ignored. To explain precisely how we apply summation by parts, decompose $\mathcal{E}_k$ as follows: \begin{align} \mathcal{E}_k =&c_k (H_{k+1}-H_{k})+c'_k ( H'_{k+1}-H'_k) +c''_k (H''_{k+1} - H''_{k}), \qquad\text{where} \label{kajbr}\\ c_{k} =&\frac{1 }{8\sqrt{\beta}k^{3/2}},\qquad c'_{k}=c''_k =\frac{1 }{8\beta k^2}, \\ H_{k}=& G_{k}-G_{k-1},\qquad H'_k = G_{k}G_{k-1},\qquad H''_{k}=(G_k+G_{k-1})\gamma_{k-1}. \end{align} We will then use \begin{align}\label{wrugb} &\sum_{k=1}^n (D_k +D_{k+1})c_k (H_{k+1}-H_k)\\ &= 2D_{n+1}c_{n+1}H_{n+1}-2D_{1}c_1 H_1 -\sum_{k=1}^n (D_{k+1}-D_k)c_k (H_{k+1}+H_k)+2D_{k+1}(c_{k+1}-c_k)H_{k+1}\\ &= \bigg(-\sum_{k=1}^n (D_{k+1}-D_k)c_k (H_{k+1}+H_k) \bigg) +O(D_\ast \log(n)^{7/2} n^{-3/2}). \end{align} To ignore the second term inside the sum we used that $(k+1)^{-3/2}-k^{-3/2} =O(k^{ -5/2})$ and that since $\ul{x}(i)\geq 1/\log n$ we have $D_k=0$ for all $k\leq n/\log(n)$. We treat the $c''_k(H''_{k+1}-H''_k)$ term on line (\ref{kajbr}) in the same way as calculation (\ref{wrugb}). For the $c'_k (H'_{k+1}-H'_{k})$ term on line (\ref{kajbr}), we do not sum by parts. Instead we make the following algebraic manipulation: \begin{align} &\sum_{k=1}^{n} (D_{k} + D_{k+1})c'_k (H'_{k+1}-H'_{k}) \\ &=\sum_{k=1}^{n}2 D_{k+1} c'_k (H'_{k+1}-H'_{k}) +(D_{k}-D_{k+1})c'_k (H'_{k+1}-H'_k). \end{align} It follows that the term $f_2$ in the lemma statement is given by $f_2=2n^2 c'_k (H'_{k+1}-H'_k)$. Collecting all the remaining terms gives the formula asserted in the lemma statement. \end{proof} \subsubsection{MGF bound for the remainder term} In the previous section we showed that instead of $\mathcal{S}_i$, it suffices to analyze terms $S_i$ that were obtained after applying summation by parts. We now write the $S_i$ terms as follows: \begin{align} S_i=\sum_{k=1}^n E_{k,i},\qquad E_{k,i}=\frac{1}{n^2}(D_k(\tau_i) -D_{k+1}(\tau_i))(\sqrt{n} f_{1,k}+f_{3,k}) +\frac{1}{n^2}D_{k+1}(\tau_i)f_{2,k}, \end{align} Recall that the terms $f_{i,k}$ were defined on line (\ref{fDef}) as functionals of $W$; they are local to the region around $x_k$ and typically of order $O(1)$. Throughout this section we will abbreviate $D_k=D_k(\tau_i)$ when it is clear at what time we are counting the downward steps. As we mentioned above, the purpose of the summation by parts was to take advantage of the differences $D_{k}-D_{k+1}$, which we now do: \begin{lemma}\label{EQMGFLemma} Suppose $j$ is defined by $\tau_j =T_{i,k'}$ for some $i,k'$. For almost all $W$, all $n\in \mathbb{N}$, all $k=\frac{n}{\log n},\ldots ,n$ and $|\lambda| \leq n$ the following approximation holds: \begin{align} \log \mathbb{E}\left[ \exp\left(\lambda E_{k,j}\right)\middle| D_{k+1},W\right] =D_{k+1}\left(\frac{f_{1,k}^2}{n^3}\lambda^2 +O\left((\log n)^p n^{-3/2}\right)\right), \end{align} where the implied constant depends only on $W,\beta$. Here $p$ is a positive universal constant. \end{lemma} \begin{proof} We will assume that $k\leq k'$. A similar proof based on formula (\ref{branching2}) can be given for the case $k\geq k'$. Given $D_{k+1}$ and $W$, the variable $D_{k}$ is (by lemma \ref{BranchingLemma}) a sum of $N_{k+1}$ independent $\text{Geom}[p_k]-1$ random variables. It follows that \begin{align} \mathbb{E}\left[\exp(\lambda D_k\middle| D_{k+1}=D,W\right] =& \left(1 -(q_{k}/p_{k})(e^\lambda -1)\right)^{-D}. \end{align} The lemma now follows by substituting the definition of $E_{k,j}$ in this explicit formula for the moment generating function, and making a Taylor approximation: \begin{align} \log \mathbb{E}\left[ \exp\left(\lambda E_{k,j}\right)\middle| D_{k+1},W\right] =& \frac{D_{k+1}\lambda}{n^2}(\sqrt{n}f_{1,k}+f_{2,k}+f_{3,k})\\ &-D_{k+1}\log\left(1-\frac{q_k}{p_k}\left(\exp \left(-(\sqrt{n}f_{1,k}+f_{3,k}) \lambda/n^2\right)-1\right)\right). \end{align} Lemma \ref{pkApproxLemma} gives the series expansion $q_{k}/p_k = 1-2G_k/\sqrt{\beta k}+O(\log(n)^3/n)$. The lemma statement now follows by using Taylor series for the logarithm and exponential. \end{proof} \subsubsection{Proof of proposition \ref{VMartProp}} \begin{proof}[Proof of proposition \ref{VMartProp}] By lemma \ref{sumByPartsLemma} it suffices to replace $\mathcal{S}_i$ with $S_i$ in the proposition statement. For some $i',k'$ we have either $\tau_i=T_{i',k'}$. Although $k',i'$ are random, by conditioning on their values it suffices to obtain a bound that holds uniformly in $k',i'$. Within this proof we abreviate $D_k=D_k(T_{i',k'})$. We will also use the notation \begin{align} \ol{D}_k =\max\{ D_j :\; j=k,\ldots n\}. \end{align} Suppose $k\leq k'$ and let $Z_k(\lambda)=\exp (\lambda \sum_{j=k}^{k'}E_{j,i})$. Also for each $j$ let $F_j$ be the $\sigma$-field generated by $D_{j},\ldots ,D_n$ and $W$. In the following calculation we let $D_\ast=n(\log n)^3$. For the second inequality we use lemma \ref{EQMGFLemma}. Also we assume $k\geq n/\log n$. \begin{align} \E\left[ Z_k(\lambda) \one_{\ol{D}_k\leq D_\ast}\middle|W \right] \leq & \E\left[ Z_{k+1}(\lambda) \one_{\ol{D}_{k+1}\leq D_\ast} \E \left[ \exp\left(\lambda E_{k,i} \right) \middle| F_{k+1}\right]\middle|W \right] \nonumber \\ \leq & \E\left[ Z_{k+1}(\lambda) \one_{\ol{D}_{k+1}\leq D_\ast} \exp\left(D_{k+1}f_{1,k}^2 n^{-3}\lambda^2 +CD_{k+1}\log(n)^p n^{-3/2} \right)\middle|W \right]\nonumber \\ \leq & \E\left[ Z_{k+1}(\lambda) \one_{\ol{D}_{k+1}\leq D_\ast}\middle|W \right] \exp\left(C D_{\ast}\log(n)^p \left(n^{-3}\lambda^2 + n^{-3/2}\right)\right). \end{align} It follows by induction that \begin{align} \label{389429} \E\left[ Z_{\lfloor n/\log n \rfloor}(\lambda) \one_{\ol{D} \leq D_\ast}\middle| W\right] \leq & \exp\left(C D_{\ast}\log(n)^p \left(n^{-2}\lambda^2 + n^{-1/2}\right)\right). \end{align} Since we are working on the event $\ul{x}(i)\geq 1/\log n$, we have that $Z_1 =Z_{\lfloor n/\log n\rfloor }$. By Chebychev's inequality and line (\ref{389429}) we get \begin{align} &\log P\bigg(\sum_{j=1}^{k'} E_{j,i} \geq r,\ol{D}(\tau_i) \leq D_\ast,\ul{x}(i)\geq \frac{1}{\log(n)}\bigg|W\bigg)\\ &\qquad \leq \log P\bigg( \sum_{j=n/\log n}^{k'} E_{j,i}\geq r,\ol{D}(\tau_i)\leq D_\ast \bigg|W\bigg)\\ &\qquad \leq -\lambda r +C D_{\ast}\log(n)^p \left(n^{-2}\lambda^2 + n^{-1/2}\right). \end{align} Now obtain a bound by letting $r=(D_\ast/n)n^{-0.2499}$ and $\lambda=n^{0.75}$. The same bound can be obtained for the probability that $\sum_{j=k'}^n E_{j,i} \geq n^{-0.249}$ by a similar argument. \end{proof} \subsection{Bound on numbers of jumps} \label{NStarSection} In this section we will use the notations \begin{align} \ol{D}_k(t)=\max_{i=k}^n D_i(t),\qquad \ul{X}(t)=\inf \{X(s):\; 0\leq s\leq t\}. \end{align} After several lemmas we will prove the following \begin{lemma} \label{NStarLemma} Let $D_\ast \geq t n (\log n)^3$. For almost all $W$ there exists $c=c(W,\beta)>0$ such that \begin{align} P\left(\ol{D}(t) \geq D_\ast ,\ul{X}(t)\geq \log(n)^{-1} \middle| W \right)\leq \exp\left( -\frac{c D_{\ast}^2}{ t n^2} \right). \end{align} \end{lemma} We will divide the proof of lemma \ref{NStarLemma} into three pieces, two of them being lemmas \ref{mgflemma4} and \ref{DeltaNLemma}. Before getting into the details, we would like to explain roughly the idea of the proof. By a union bound it suffices to estimate the probability of an unusually large number of downward steps at some height $x_k\geq 1/\log n$. Because of the branching structure from lemma \ref{BranchingLemma}, if there are many downward steps from $x_k$ there will also probably be many downward steps from nearby heights. To make this precise use the branching structure to compute a conditional moment generating function (lemma \ref{mgflemma4}) and then apply Chebychev's inequality (lemma \ref{DeltaNLemma}). There can be at most $4tn^2$ steps by time $t$, so if there are sufficiently many downward steps from nearby heights that the total number of steps exceeds $4tn^2$, then we have a contradiction. In some sense we are reducing the problem of bounding $\ol{D}(t)$ to the task of proving a high probability modulus of continuity estimate on $D_k(t)$ considered as a function of $k$. In this section we choose and fix a starting height $k_0/n$ for the process $X^n$, and also integers $i$ and $k'$ which index a stopping time $T_{i,k'}$. In this section the time variable will always be this $T_{i,k'}$, so for example $D_{k}=D_{k}(T_{i,k'})$. \begin{lemma} \label{mgflemma4} If $k<k'$ then the conditional moment generating function for $D_k$ is \begin{align} \mathbb{E}[e^{\lambda D_k} | W, D_{k'}=D] =& \left( 1+ \frac{ (e^\lambda-1)\Delta S}{1- (e^\lambda -1)S} \right)^{D} \end{align} where $\Delta S=\prod_{j=k}^{k'-1} q_j/p_j $ and $S=\sum_{i=k}^{k'-1} \prod_{j=k}^{i} q_j/p_j $. \end{lemma} \begin{proof} Let $F_j$ and $f_j$ be the following probability generating functions: \begin{align} F_j(z)=\sum_{j=0}^\infty P(D_j=\ell | W,D_{k'})z^\ell,\qquad f_j(z)=\sum_{j=0}^\infty P(D_j=\ell | W,D_{j+1})z^\ell. \end{align} If $j<k'$ then lemma \ref{BranchingLemma} asserts that $D_j$ is a sum of independent variables $\gamma_{j,\ell}$ such that $\gamma_{j,\ell}+1$ has a geometric distribution. A short calculation using the formula for the probability generating function for a geometric random variable shows that \begin{align} f_j(z) =& \left(1 -\rho_j (z-1)\right)^{-1},\qquad \rho_j =\frac{q_j}{p_j}. \end{align} By the general theory of braching processes, $F_j (z) =F_{j+1}\circ f_j(z)$. Suppose that $D_{k'}=1$, so that $F_{k'}(z)=z$. It is easy then to check by induction for $k<k'$ we have \begin{align} F_k(z) =& 1+\frac{(z-1)\prod_{j=k}^{k '-1}\rho_j}{1-(z-1)\sum_{i=k}^{k'-1}\prod_{j=k}^{i} \rho_j}. \end{align} The formula in the lemma statement now follows since the moment generating function and probability generating function are related by $M(\lambda)=F(e^\lambda)$. The general case $D_{k'}\geq 1$ follows from the composition formula for probability generating functions. \end{proof} \begin{lemma} \label{DeltaNLemma} Suppose that $k <k'$, and suppose the random environment $W$ is such that \begin{align} \Delta S=& \prod_{j=k}^{k'-1}\frac{p_j}{q_j} \in [0.99,1.01],\qquad \text{and}\qquad S=\sum_{i=k}^{k'-1}\prod_{j=k}^i \frac{p_j}{q_j}\geq 100. \end{align} Then the following bound holds: \begin{align} P\left(D_{k}< D_{k'}/2 \middle|W, D_{k'} \right)\leq & \exp\left(-0.05 D_{k'}/S\right). \end{align} \end{lemma} \begin{proof} If $X$ is a random variable with a finite moment generating function and $\lambda>0$, then Chebychev's inequality gives the following bound on lower tail probabilities: \begin{align} \log P(X<r)\leq & \lambda r +\log \mathbb{E}[\exp(-\lambda X)]. \end{align} Starting with the result of the previous lemma, we let \begin{align} \lambda=-\log(1+\alpha),\qquad \text{where }\alpha=\frac{-1}{4S}. \end{align} We then apply Taylor's theorem to the logarithm function to get \begin{align} &\log P(D_{k } <D/2 |W,D_{k'}=D ) \leq \lambda D/2 + D \log \left( 1+ \frac{ (e^{-\lambda}-1)\Delta S}{1- (e^{-\lambda} -1)S} \right) \\ &= - (D/2) \log (1+\alpha) +D \log \left(1-\alpha(S- \Delta S)\right)-D\log\left(1-\alpha S\right) \label{eq114} \\ &=(D \Delta S -D/2 ) \alpha +\left(D/2 + 2 D S \Delta S - D (\Delta S)^2\right)\frac{\alpha^2}{2} +\frac{\alpha^3}{6} f'''(\xi) , \label{eq115} \end{align} where $\xi=\xi(\alpha) \in [\alpha,0]$. Here $f$ is the right hand side of line (\ref{eq114}) considered as a function of $\alpha$. Since $\alpha^3 \leq 0$ we now want to get a lower bound for $f'''(\xi)$. Assume that $r>0$, and note that $-1<\xi<0$ and $S>\Delta S$. It follows that \begin{align} f'''(\xi)=& -\frac{D}{(1+\xi)^3} - \frac{2D(S-\Delta S)^3}{(1-\xi (S-\Delta S))^3} +\frac{2DS^3}{(1-\xi S)^3} \geq -\frac{D}{(1+\xi)^3}. \end{align} The above inequality is valid because $x\mapsto x/(1-\xi x)$ is an increasing function. Now we use the hypotheses that $\Delta S\in [0.99,1.01]$ and $S\geq 100$. These conditions imply the following bounds for the three terms on line (\ref{eq115}): \begin{align} (D\Delta S -D/2)\alpha \leq & -0.1225 D/S,\\ \left(D/2 + 2 D S \Delta S - D (\Delta S)^2\right)\frac{\alpha^2}{2} \leq & 0.0630 D/S, \\ \frac{\alpha^3}{6} f'''(\xi) \leq & 3\times 10^{-7} D/S. \end{align} Our assertion follows by adding these inequalities. \end{proof} \begin{lemma} \label{DeltaSLemma} Almost all realizations of the random environment $W$ satisfy the following conditions. There exists $c=c(\beta,W)>0$ such that for sufficiently large $n$ (depending on $\beta,W$), and all $k,k'$ such that $n/\log n\leq k<k' \leq n$ and $k'-k\leq c n/( \log n)^2$, we have \begin{align} \prod_{j=k}^{k'-1} \frac{p_j}{q_j} \in \left[ 0.99,1.01\right]. \end{align} \end{lemma} \begin{proof} By lemma \ref{pkApproxLemma} we have (for almost all $W$) \begin{align} \log \prod_{j=k}^{k'-1} \frac{p_j}{q_j} =\sum_{j=k}^{k'-1} \left\{ \frac{-2G_j}{\sqrt{\beta j}}+\frac{1}{j} +\frac{2G_{j}^{(2)}}{\beta j} +O( ( \log n)^3 n^{-3/2})\right\}. \label{eq937} \end{align} We will show that the right hand side is $O(1/\log n)$ for almost all $W$. Since $k'-k\leq n/ \log n $ we see that ``$O$'' terms can be ignored. Recall that the variables $G_k$ are defined in equation (\ref{GDef}) as certain integrals of the random environment. It follows that \begin{align} \sum_{j=k}^{k'-1} \frac{G_j}{\sqrt{j}} =& \sum_{j=k}^{k'-1} \int_{x_{j-1}}^{x_{j+1}} \frac{1}{\sqrt{x_j}} (1-n|x-x_j|)\, dW(x). \label{eq938} \end{align} The right hand side is a Gaussian random variable and it is straitforward to bound its variance by $C \log (n)(k'-k)/n \leq Cc/ \log n $. By a standard tail bound for the normal distribution, the probability that the left hand side of (\ref{eq938}) exceeds $0.01$ is bounded by $ \exp(- (0.01)^2 \log (n)/(Cc))$. Taking $c$ sufficiently small and making a union bound over $k,k'$, the Borel-Cantelli lemma now shows that the left hand side of (\ref{eq937}) is less than $0.01$ for almost all $W$. The $G^{(2)}_j/j$ terms in equation (\ref{eq937}) can be analyzed in a similar way but are smaller by a factor of $\sqrt{j}$. The $1/j$ terms make a contribution of $O(1/\log n)$ and therefore can be ignored. The lemma statement then follows by taking the exponential of both sides and using a Taylor approximation. \end{proof} \begin{proof}[Proof of lemma \ref{NStarLemma}] In this proof we fix a realization of $W$ such that the conclusion of lemma \ref{DeltaSLemma} holds, and all probabilities are conditional on this $W$. In order to use lemma \ref{DeltaNLemma} we must first we reduce to the case that $t$ is one of the stopping times $T_{i,k'}$. The number of downward steps from any height can only change at one of the stopping times $T_{i,k'}$. Thus by a union bound we have \begin{align} & P\left(\ol{D}(t) \geq D_\ast,\ul{X}(t)\geq 1/\log n\right) \\ & \leq \sum_{i=1}^{\lfloor 4tn^2\rfloor} \sum_{k'= \lfloor n/\log n \rfloor}^{n} P\left(\ol{D}(T_{i,k'})\geq D_\ast, \ul{X}(t)\geq 1/\log n,T_{i,k'}\leq t\right). \end{align} The condition $X(t)\geq 1/\log n$ let us assume $k'\geq n/\log n$. By another union bound we get \begin{align} P\left(\ol{D}(T_{i,k'})\geq D_\ast ,\ul{X}(t)\geq 1/\log n,T_{i,k'}\leq t\right)\leq \sum_{k=\lfloor n/\log n \rfloor}^n P(D_k(T_{i,k'}) \geq D_\ast,T_{i,k'}\leq t). \end{align} We will make yet one more union bound. Fix $n$ and suppose $x_k$ is a height from which there are more than $D_\ast$ downward steps by time $t$. If $D_{k-i}>D_\ast/2$ for all $i=\lfloor 8tn^2/D_\ast \rfloor,\ldots ,\lceil 16tn^2/D_\ast \rceil$, then this accounts for more than $4tn^2$ steps altogether. Since $\Delta t=1/(4n^2)$ it is impossible for there to be so many steps by time $t$. It follows that \begin{align} \label{nStarEq1} P\left(D_k(T_{i,k'} )\geq D_\ast,T_{i,k'}\leq t\right) \leq \sum_{j=\lfloor 8tn^2/D_\ast \rfloor}^{\lceil 16tn^2/D_\ast\rceil} P\left(D_k(T_{i,k'}) \geq D_\ast ,D_{k-j}(T_{i,k'})\leq \frac{D_\ast}{2} \right). \end{align} If $k$ is so small so that $k-j <n/\log n$ for one of the terms in this sum, then we can change $D_{k-j}$ to $D_{k+j}$ in the above formula. Since we assumed in the lemma statement that $D_\ast \geq t n (\log n)^3 $, for sufficiently large $n$ the indices of the sum must be between $n/\log n$ and $n$ for at least one of the options $k\pm j$. In both cases the proof is similar, so we will assume equation (\ref{nStarEq1}). Since $D_\ast \geq t n (\log n)^3$, the indices $k$ and $k-i$ in display (\ref{nStarEq1}) differ by at most $ n/ ( \log n)^{ 3} $, so by lemma \ref{DeltaSLemma} we have for sufficiently large $n$ that \begin{align} \Delta S(k-i,k) \in[0.99,1.01],\qquad \text{where} \qquad \Delta S(k,k')= \prod_{j=k}^{k'-1} \frac{p_j}{q_j} . \end{align} Also $k$ and $k-i$ differ by at least $ n/(2 ( \log n)^{ 3})$, so \begin{align} S(k-i,k)\geq \frac{n}{2C(\log n)^2}\times 0.99 \gg 100, \qquad \text{where}\qquad S(k,k')=\sum_{i=k}^{k'}\Delta S(i,k'). \end{align} Therefore for sufficiently large $n$ the hypotheses of lemma (\ref{DeltaNLemma}) are satisfied so we get \begin{align} &P\left(D_k(T_{i,k'}) \geq D_\ast ,D_{k-i}(T_{i,k'})\leq \frac{D_\ast}{2} \right) \\ &\leq P\left(D_{k-i}(T_{i,k'})\leq \frac{D_\ast}{2} \middle| D_k(T_{i,k'}) = D_\ast \right) \leq \exp\left(\frac{-0.05 D_\ast}{ S(k-i,k)}\right),\\ & \text{and }S(k-i,k)\leq 1.01 \times 16 tn^2/D_\ast. \end{align} This gives the lemma statement, except an extra factor of $n^4$ from union bounds. For sufficiently large $n$ this can be absorbed by slightly decreasing the constant in the exponent. \end{proof} \subsection{Approximate equality of jump times for $X$ and $X^{(n)}$} \label{ApproxJumpTimesSection} \begin{proposition} \label{TimeApproxProp} For almost all $W$, all $n\in \mathbb{N}$, and for all $i\in \mathbb{N}$ we have \begin{align} P(|\tau_{i} -\wt{\tau}_{i}|\geq t_i n^{-0.99} , \ul{x}(i) \geq 1/\log n, \ol{D}_{i}\leq n (\log n)^3|W) \leq \exp(-cn^{0.01}) \end{align} where $c=c(W, \beta)$. \end{proposition} \begin{proof} By lemma \ref{exitTimeLemma01} we have \begin{align} \E\left[\Delta \tau_i -\Delta \wt{\tau}_i\middle| x(i)=x_k ,W\right] =& \theta_k +O( (\log n)^3/n^3),\qquad \text{where }\theta_k=\frac{\gamma_k -\gamma_{k-1}}{4\sqrt{\beta }x_{k}^{3/2}n^{5/2}}. \end{align} Suppose $x(i)=x_k\geq 1/\log n$. Recall that $4 n^2 \Delta \tau_i$ is a geometric random variable with parameter $x_k/2$. Using lemma \ref{exitTimeLemma01} and reasoning as in lemma \ref{mgfLemma01}, it follows that for all $k\geq n/\log n$ we have \begin{align} &|M(\lambda)-1-M'(0)\lambda|\leq C\log(n)^{2}n^{-4} \text{ for all }\lambda \leq \frac{Cn^2}{\log n},\\ &\text{ where } M(\lambda)=\E[\exp(\lambda(\Delta \tau_i -\Delta \wt{\tau}_i - \theta_{k(i)}))|W,x(i)=x_k]. \end{align} The above formula is valid as long as $x_k\geq 1/\log n$. Now let \begin{align} Z_j(\lambda)=\exp(\lambda(\tau_j -\wt{\tau}_j-\theta_{k(j)})). \end{align} For some universal constant $p$ we have \begin{align} \E\left[\one_{\ul{x}(j)\geq \log(n)^{-1}} Z_j(\lambda)\middle| W\right] \leq & \E\left[\one_{\ul{x}(j-1)\geq \log(n)^{-1}} Z_{j-1}(\lambda) \E[\exp(\lambda (\Delta \tau_j -\Delta \wt{\tau}_j -\theta_{k(i)})) | W,x(j)]\middle| W\right] \nonumber \\ \leq & \E\left[\one_{\ul{x}(j-1)\geq \log(n)^{-1}} Z_{j-1}(\lambda)\middle| W\right](1+C\log(n)^p (\lambda n^{-3}+\lambda^2 n^{-4})). \end{align} It follows by induction that \begin{align} \E\left[\one_{\ul{x}(j)\geq \log(n)^{-1}} Z_j(\lambda)\middle| W\right] \leq \exp\left(C t (\log n)^p (\lambda n^{-1} +\lambda^2 n^{-2})\right), \end{align} with $C$ depending only on $W,\beta$. Applying Chebychev's inequality as in proposition \ref{HMartProp}, and using $\lambda=n/(t\vee 1)$ we obtain \begin{align} \label{p62} P\bigg( \tau_i -\wt{\tau}_i -\sum_{j=1}^{i}\theta_{k(j)}\geq t n^{-0.99},\ul{x}(i)\geq 1/\log n \bigg| W\bigg) \leq \exp(-c n^{0.01}), \end{align} for some constant $c=c(W,\beta )$. By the same argument but using $1/Z_i(\lambda)$, the same bound holds for $-\tau_i +\wt{\tau}_i+\sum_{j=1}^i \theta_{k(j)}$. To finish the proof we must show that $|\sum_{j=1}^i \theta_{k(j)}|\leq n^{-0.99}$ with high probability. The sum can be separated as $\sum_{j=1}^i \theta_{k(j)} \one_{k(j)\leq k'} +\sum_{j=1}^i \theta_{k(j)} \one_{k(j)\geq k'}$, and we will only consider the ``$\leq$" term since the same bound can be obtained for the ``$\geq$" term similarly using the branching formula (\ref{branching2}) for numbers of upward steps. For some $k',i'$ we have $\tau_i=T_{i',k'}$. Let $D_k=D_k(\tau_i)$ and $d_k=d_{k}(\tau_i)$. Summing by parts as in lemma \ref{sumByPartsLemma} we get \begin{align} \sum_{j=1}^i \theta_{k(j)} \one_{k(j)<k'} =&\sum_{k=1}^{k'} (D_k+D_{k+1}+b_k) \theta_k\\ =&\bigg( \sum_{k=1}^{k'} (D_{k+1}-D_k) \frac{\gamma_k +\gamma_{k-1}}{4\sqrt{\beta}x_{k}^{3/2}n^{-5/2}}\bigg)+O(D_\ast (\log n)^p /n^{ 2}). \end{align} As in lemma \ref{sumByPartsLemma}, the $b_k$ term became part of the $O(D_\ast (\log n)^p /n^{ 2})$; on the event $\ol{D}_i\leq n \log n$ this error term is within the tolerance of the lemma statement. Now let \begin{align} Z_k (\lambda)=& \exp \bigg(\lambda \sum_{j=k}^{k'} (D_{k+1}-D_k)\frac{f_k}{n^{5/2}}\bigg),\qquad f_k =\frac{\gamma_k +\gamma_{k-1}}{4\sqrt{\beta}x_{k}^{3/2}}. \end{align} For $k\geq n/\log n$ we have the following estimate: \begin{align} \E\left[Z_k \one_{\ol{D}_k\leq D_\ast} \middle|W\right] \leq & \E\left[Z_{k+1} \one_{\ol{D}_{k+1}\leq D_\ast} \E\left[ \exp(\lambda(D_{k+1}-D_k)f_k n^{-5/2})) \middle|D_{k+1},W\right] \middle|W\right] \\ =& \E\left[Z_{k+1} \one_{\ol{D}_{k+1}\leq D_\ast} \left(1+\frac{q_k}{p_k}\left(\exp(-\lambda f_k n^{-5/2})-1\right) \right)^{-D_{k+1}} \middle|W\right] \\ \leq & \E\left[Z_{k+1} \one_{\ol{D}_{k+1}\leq D_\ast} \middle|W\right] \exp\left(D_\ast C\log(n)^p (\lambda n^{-3} +\lambda^2 n^{-5})\right). \end{align} By induction we get (again for $k\geq n/\log n$) \begin{align} \E\left[Z_k \one_{\ol{D}_k\leq D_\ast} \middle|W\right] \leq & \exp\left(D_\ast C\log(n)^p (\lambda n^{-2} +\lambda^2 n^{-4})\right). \end{align} Applying Chebychev's inequality and using $\lambda =n$ we get \begin{align} &\log P\bigg( \sum_{j=1}^i \theta_{k(i)}\one_{k(i)\leq k'} \geq n^{-0.99},\ol{D}_i \leq n (\log n)^3,\ul{x}(i)\geq \log(n)^{-1} \bigg|W\bigg)\\ &\leq \log P\bigg( \sum_{j=1}^i \theta_{k(i)}\one_{\frac{n}{\log n}\leq k(i)\leq k'} \geq n^{-0.99},\ol{D}_i \leq n \log n \bigg|W\bigg)\\ & \leq -\lambda n^{-0.99}+\E\left[Z_k \one_{\ol{D}_k\leq D_\ast} \middle|W\right] \leq -n^{0.01}+ C\log(n)^p \left( 1+n^{-1}\right) . \label{p63} \end{align} One can obtain the same bound for $-\sum_{j=1}^i \theta_{k(i)}\one_{k(i)\leq k'}$ using the same argument but with $1/Z_k(\lambda)$. The lemma statement now follows from equations (\ref{p62}) and (\ref{p63}). \end{proof} \subsection{Combining the pieces to bound $|\Phi^{B}_n -\wt{\Phi}^{B}_n|$.} \label{PhiNPhiTildeProofSection} \begin{proof}[Proof of theorem \ref{PhiNPhiTildeThm}] By basic properties of sets and probability we have \begin{align} &P\left( \left| \Phi^{B}_n(t)-\wt{\Phi}^{B}_n(t)\right| \geq n^{-.499} , \ul{X}(1.01 t)\geq 1/\log n\middle| W\right) \\ &\leq P\left( \left| \Phi^{B}_n(t)-\wt{\Phi}^{B}_n(t)\right| \geq n^{-.499} , \ul{X}(t)\geq 1/\log n,\ol{D}(t)\leq t n (\log n)^3 \middle| W\right)\\ &\qquad+P\left(\ul{X}(t)\geq 1/\log n,\ol{D}(t)\geq tn (\log n)^3 \middle| W\right) \label{a918} \end{align} The term on line (\ref{a918}) converges to zero by proposition \ref{NStarLemma}. Let $i=i(t,W,B)$ be the greatest integer $i$ such that $\tau_i \leq t$. Define terms $\RN{1},\ldots ,\RN{4}$ by \begin{align} \left| \Phi_n(t) -\wt{\Phi}_n(t)\right| \leq & \left| \Phi_n(t) -\Phi_n(\tau_i)\right| +\left| \Phi_n(\tau_i) -\wt{\Phi}_n(\wt{\tau}_i)-\mathcal{E}_i\right| +\left| \mathcal{E}_i \right| +\left| \wt{\Phi}_n(\wt{\tau}_i) -\wt{\Phi}_n(t)\right| \\ &= \RN{1}+\RN{2}+\RN{3}+\RN{4}. \end{align} By a union bound and proposition \ref{HMartProp} we have \begin{align} & P\left(\RN{2} \geq n^{-.499} , \ul{X}(t)\geq \log(n)^{-1},\ol{D}(t)\leq n (\log n)^3 \middle| W\right)\\ &\leq \sum_{j=1}^{4tn^2} P\left(\left| \Phi_n(\tau_j) -\wt{\Phi}_n(\wt{\tau}_j)-\mathcal{E}_j\right| \geq \medmath{\frac{1}{4}}n^{-.499}, \ul{X}(t)\geq \log(n)^{-1},\ol{D}(t)\leq n \log(n) \middle| W\right) \\ & \leq n^2 \exp\left(-c n^{0.001}\right). \label{a913} \end{align} By construction of $i$ and the definitions of $\Phi_n$ and $\wt{\Phi}_n$ the same estimate (\ref{a913}) holds for term $\RN{1}$, and by proposition \ref{VMartProp} we get \begin{align} P\left(\RN{3} \geq n^{-.249} , \ul{X}(t)\geq \log(n)^{-1},\ol{D}(t)\leq n (\log n)^3 \middle| W\right) & \leq n^2 \exp\left(-c n^{0.001}\right). \end{align} By definition of $\wt{\Phi}_n$ it is clear that \begin{align} \left| \wt{\Phi}_n(\wt{\tau}_i) -\wt{\Phi}_n(t)\right| \leq C\left| \wt{\tau}_i-t\right| \sqrt{n}, \end{align} where $C=C(W,\beta,a,t)$. From this and proposition \ref{TimeApproxProp}, it follows that the estimate (\ref{a913}) holds yet again for term $\RN{4}$. \end{proof} \section{Comparison of $\wt{\Phi}^{B}_n$ and $\Phi ^{B}$} \label{section6} In this section we prove the following \begin{proposition} \label{WtPhiConverges} On the event $\ul{X}(t)\geq 1/\log n$ we have \begin{align} \left|\wt{\Phi}^{B}_n(t) -\Phi^{B}(t)\right|\leq & C t n^{-0.249}. \end{align} This estimate holds for almost all $B,W$. The constant $C$ depends only on $B,W,\beta$. \end{proposition} From the definitions of the functionals we have \begin{align} \wt{\Phi}^{B}_n(t) -\Phi^{B}(t) =& \sum_{k=0}^{n-1} \frac{B_k}{\sqrt{\beta}} +\frac{ G^{(2)}_k}{\beta x_k} L_X([x_{k},x_{k+1}),t), \qquad \text{where} \label{ABdef}\\ B_k=& \frac{\sqrt{n}(G_k+G_{k+1})}{2\sqrt{x_{k}}}L_X([x_{k},x_{k+1}),t) -\int_{x_{k}}^{x_{k+1}} \frac{L_X(x,t)}{\sqrt{x}}\circ dW(x).\nonumber \end{align} To prove proposition \ref{WtPhiConverges}, we must estimate the right hand side of (\ref{ABdef}). \subsection{Convergence of $\sum_k B_k$} For the $B_k$ terms we must estimate some stochastic integrals, and to do this we will rely on the following lemma. This kind of result is very well known, with the ideas going back at least to \cite{McKeanBook}. \begin{lemma} \label{ItoIntEst1} Let $W_t$ be the Wiener process and $Y^{(n)}_t$ be a sequence of stochastic processes adapted to the filtration generated by $W$. If for each $n$ we have $|Y^{(n)}_t| \leq f(n)$ almost surely, then the bound \begin{align} \left| \int_{0}^t Y^{(n)}_s \, dW_s \right| \leq 4 f(n) \sqrt{\log n} \end{align} holds almost surely as well. \end{lemma} \begin{proof} Consider the martingales \begin{align} Z^{(n)}_t =& \exp\left( \lambda \int_{0}^t Y^{(n)}_s dW_s - \frac{\lambda^2}{2}\int_{0}^t (Y^{(n)}_s)^2\, ds \right). \end{align} Applying Doob's martingale inequality we get that \begin{align} P\left( \int_{0}^t Y^{(n)}_s dW_s - \frac{\lambda}{2}\int_{0}^t (Y^{(n)}_s)^2\, ds > r\right) \leq & e^{-\lambda r}, \end{align} and it follows that \begin{align} P\left( \int_{0}^t Y^{(n)}_s dW_s > 2r\right) \leq e^{-\lambda r} + P\left( \int_{0}^t (Y^{(n)}_s)^2\, ds > \frac{2r}{\lambda} \right). \label{eq183561} \end{align} Let $r =2f(n)\sqrt{\log n}$ and $\lambda =\sqrt{\log n}/f(n)$. The right hand side of (\ref{eq183561}) is then bounded by $1/n^2$, so our assertion follows from the Borel-Cantelli lemma. \end{proof} Recall that \begin{align} L_X(x,t) =&\frac{L_B(s(x),T^{-1}(t))}{2\exp(I(x))},\qquad I(x)= \int_{x}^1 \frac{2\, dW(y)}{\sqrt{\beta y}}. \end{align} The variables $G_k$ are defined (eq.\ \ref{GDEF}) in terms of $W$, but are not increments of $W$. Therefore it is useful to decompose $G_k$ as follows \begin{align} G_{k}=&g_k +\wt{g}_k,\qquad g_k=n^{3/2}\int_{x_{k-1}}^{x_k} (x-x_{k-1})\,dW(x),\qquad \wt{g}_k =n^{3/2}\int_{x_k}^{x_{k+1}} (x_{k+1}-x)\, dW(x). \end{align} The point is that increments of $W$ can be expressed as $W(x_{k})-W(x_{k-1})=n^{-1/2}(\wt{g}_{k-1}+g_k)$. For the calculations below to fit reasonably on the page, we will need the following abbreviations: \begin{align} W_k=&W(x_k),\qquad \Delta W_k =W_{k+1}-W_{k},\qquad I_k =I(x_k),\\ \qquad L_B s(x)=&L_B(s(x),T^{-1}(t)),\qquad L_B s_k =L_B(s_k,T^{-1}(t)). \nonumber \end{align} It is straitforward to check that \begin{align} \label{q1001} &\sum_{k=1}^n \frac{\sqrt{n}(G_k +G_{k-1})}{2\sqrt{x_{k-1}} } L_X([x_{k-1},x_{k}),t)\\ &\qquad = \sum_{k=0}^{n-1} \bigg\{ \frac{\Delta W_{k-1} L_Bs_k}{2\sqrt{x_k}e^{ I_k}} +\frac{\sqrt{n}g_{k}}{4\sqrt{x_k}} \left( \int_{x_{k-1}}^{x_{k+1}} \frac{L_Bs(x)}{e^{I(x)}}\, dx -\frac{2L_B s_k}{ne^{I_k} }\right) \nonumber \\ &\hspace{100pt} + \frac{\sqrt{n}\wt{g}_{k-1}}{4\sqrt{x_k}} \left( \int_{x_{k-2}}^{x_{k}} \frac{L_Bs(x)}{e^{ I(x)}}\, dx -\frac{2L_B s_k}{ne^{I_k} }\right) +R_k\hspace{1pt} \bigg\},\nonumber \end{align} where for almost all $W$ the remainder term satisfies \begin{align} \label{IgnoreR} R_k \leq C \log(n)^p n^{-3/2}\sup_{x\in \mathbb{R}}L_B(x,T^{-1}(t)), \end{align} where $p$ is a universal constant and $C=C(W,\beta)$. Thus we may ignore the remainders $R_k$. The terms $B_k$ are the left hand side of (\ref{q1001}) minus the following Stratonovich integral, which we rewrite in terms of an anti-Ito integral: \begin{align} \label{q1002} \int_{x_k}^{x_{k+1}} \frac{L_B s(x)}{2\sqrt{x}e^{ I(x)}} \circ dW(x) =& \int_{x_k}^{x_{k+1}} \frac{L_B s(x)}{2\sqrt{x}e^{ I(x)}} d\overleftarrow{W}(x) - \int_{x_k}^{x_{k+1}} \frac{L_B s(x)}{2\sqrt{\beta}x e^{ I(x)}} dx. \end{align} Recall that $s(x)$ and $I(x)$ are functionals of $\{ W(y):\; y\geq x\}$, so that the integrand above is nonanticipating for an anti-Ito integral but not an Ito integral. We now show that the first term on the right hand side of (\ref{q1001}) converges to the anti-Ito integral in (\ref{q1002}): \begin{lemma} \label{AntiItoConv} On the event $\ul{X}(t)< 1/\log (n)$, the bound \begin{align} \left| \sum_{k=0}^{n-1} \frac{\Delta W_{k-1} L_Bs_k}{2\sqrt{x_k}e^{ I_k}} - \int_{0}^{1} \frac{L_B s(x)}{2\sqrt{x}e^{ I(x)}} d\overleftarrow{W}(x) \right| \leq & C t n^{-0.49} \label{conv2048} \end{align} holds for almost all $B,W$. The constant $C$ depends only on $B,W,\beta$. \end{lemma} \begin{proof} We may assume $\ul{X}(t) \geq 1/\log n$, and thus that \begin{align} \sum_{k=0}^{\lfloor n/\log n\rfloor} \frac{\Delta W_{k-1} L_Bs_k}{2\sqrt{x_k}e^{ I_k}} =0= \int_{0}^{1/\log n} \frac{L_B s(x)}{2\sqrt{x}e^{ I(x)}} d\overleftarrow{W}(x) . \label{ignoreSmallK83} \end{align} We may therefore assume throughout the proof that $x>1/\log n$. Let $\lceil x\rceil_n$ indicate the smallest $x_k$ not less than $x$. The sum we are considering can be represented as a stochastic integral as follows: \begin{align} \sum_{k=0}^{n-1} \frac{\Delta W_{k-1} L_Bs_k}{2\sqrt{x_k}e^{ I_k}} = \int_{0}^1 \frac{L_B s(\ck{x})}{2\sqrt{\ck{x}} e^{I(\ck{x})}}\, d\ola{W}(x). \end{align} Thus the difference we are to estimate is \begin{align} \left| \int_{0}^1 \frac{L_B s(\ck{x})}{2\sqrt{\ck{x}} e^{I(\ck{x})}} - \frac{L_B s(x)}{2\sqrt{x}e^{ I(x)}} d\overleftarrow{W}(x) \right| =\left| \int_{0}^1 Y^{(n)}_x \, d\ola{W}(x)\right| , \end{align} and we claim that $ |Y^{(n)}_x|\leq Ct n^{-0.49} . $ If we can verify this claim, then the present lemma follows from lemma \ref{ItoIntEst1}. We emphasize that each $Y^{(n)}_x$ is nonanticipating in the $-x$ direction, so making a change of variable $x\mapsto 1-x$ transforms the expression to an ordinary Ito integral of a nonanticipating process. By (\ref{ignoreSmallK83}) we may assume that $\sqrt{x}- \sqrt{\ck{x}}=O(\sqrt{\log n}/n)$. By lemma \ref{TBound} we have that the local times are up to time $T_{n}^{-1}(t)=O( (\log n)^{1.02})$, and thus by lemma \ref{LocalTimeLemma} we have $L_B s(\ck{x}) -L_B s(x) =O(n^{-0.499})$. Lemma \ref{TBound} also gives the bound $e^{I(x)}-e^{I(\ck{x})}=O(n^{-0.499})$. Combining these bounds finishes the proof. \end{proof} With lemma \ref{AntiItoConv} in hand, there remain two terms on the right hand side of (\ref{q1001}), and we now show that their sum will converge to the last integral on line (\ref{q1002}). \begin{lemma} \label{BkLemma1111} On the event $\ul{X}(t)\geq 1/\log n$ we have \begin{align} \sum_{k=0}^{n-1}\frac{\sqrt{n}g_{k}}{4\sqrt{x_k}} \left( \int_{x_{k-1}}^{x_{k+1}} \frac{L_Bs(x)}{e^{I(x)}}\, dx -\frac{2L_B s_k}{ne^{I_k}}\right) = & -\frac{1}{3} \int_{0}^{1} \frac{L_B s(x)}{2\sqrt{\beta}x e^{ I(x)}} dx +O( tn^{-.249}),\label{eq857} \\ \sum_{k=0}^{n-1}\frac{\sqrt{n}\wt{g}_{k-1}}{4\sqrt{x_k}} \left( \int_{x_{k-2}}^{x_{k}} \frac{L_Bs(x)}{e^{I(x)}}\, dx -\frac{2L_B s_k}{ne^{I_k}}\right) = & -\frac{2}{3} \int_{0}^{1} \frac{L_B s(x)}{2\sqrt{\beta}x e^{ I(x)}} dx+O( tn^{-.249}). \end{align} These approximations are valid for almost all $B,W$ and the implied constants depend on $B,W,\beta$. \end{lemma} \begin{proof} The proofs for the two assertions are the same so we will only prove the first statement. In the same way as lemma \ref{AntiItoConv}, by our assumption that $\ul{X}(t)\geq 1/\log n$, we may restrict the domain of integration to $x>1/\log n$. Decompose terms of the sum as follows: \begin{align} &\frac{\sqrt{n}g_{k}}{4\sqrt{x_k}} \left( \int_{x_{k-1}}^{x_{k+1}} \frac{L_Bs(x)}{e^{I(x)}}\, dx -\frac{2L_B s_k}{ne^{I_k}}\right) \\ &\qquad = \frac{\sqrt{n}g_{k}}{4\sqrt{x_k}} \int_{x_{k-1}}^{x_{k+1}}e^{-I(x)} ( L_Bs(x) -L_B s_k )\,dx + \frac{\sqrt{n}g_{k}}{4\sqrt{x_k}} L_Bs_k\int_{x_{k-1}}^{x_{k+1}} \left( e^{- I(x)} - e^{- I_k}\right)\, dx \nonumber \\ &\qquad =A'_k +B'_k. \end{align} First we consider the $A'$ terms. The idea is to express the $g_k$'s as stochastic integrals and then apply lemma \ref{ItoIntEst1}. However, the resulting integrand has a small anticipating component which we must remove and analyze separately. Define the following modifications of the scale function: \begin{align} \wt{s}(x)=-\int_{x}^1 y^{-1}\exp\left(\begin{cases} I(\ck{y}) & \text{ if }y < \ck{x} \\ I(y) & \text{ else}\end{cases} \right)\, dy \end{align} Note that since $I(y)$ is in the $\sigma$-field generated by $\{W_z:\; y\leq z\leq 1\}$, each $g_k$ is independent of all $\wt{s}(x)$ for $x>x_{k-1}$. Furthermore, assuming that $x>1/\log n$ we have by lemma \ref{TBound} that \begin{align} |s(x)-\wt{s}(x)|\leq C (\log n)^p n^{-3/2} \end{align} for some universal constant $p$. We decompose $A'_k$ into nonanticipating and small parts as follows: \begin{align} A'_k =& \frac{\sqrt{n}g_{k}}{4\sqrt{x_k}} \int_{x_{k-1}}^{x_{k+1}}\frac{ L_B \wt{s}(x) -L_B s_k }{e^{I(\ck{x})} }\,dx + \frac{\sqrt{n}g_{k}}{4\sqrt{x_k}} \int_{x_{k-1}}^{x_{k+1}}\left(\frac{ L_B s(x) }{e^{I(x)} } -\frac{ L_B \wt{s}(x) }{e^{I(\ck{x})} }\right)\,dx \nonumber \\ =&A'_{k,1}+A'_{k,2}. \end{align} Using the condition $x>1/\log n$, lemmas \ref{TBound} and \ref{LocalTimeLemma}, and also our bound on $|s(x)-\wt{s}|$; it is easy to see that $A'_{k,2} \leq C (\log n)^p n^{-5/4}$ for some constant $C$ and universal constant $p$. Thus the sum of $A'_{k,2}$ terms is within the tolerance of the lemma statement. We rewrite $A'_{k,1}$ as a stochastic integral: \begin{align} \sum_{k=1}^n A'_{k,1} =\int_{0}^1 Y_x \, d\ola{W}(x),\qquad Y_x = \frac{n^2 (x-\ck{x})}{4\sqrt{\ck{x}}} \int_{\fk{x}}^{\ck{x}+1/n} \frac{ L_B \wt{s}(y) -L_B s(\ck{x}) }{e^{I(\ck{y})} }\,dy \end{align} Applying lemmas \ref{LocalTimeLemma} and \ref{ItoIntEst1}, one can see that \begin{align} \sum_{k=1}^n A'_{k,1} \leq C t(\log n)^p n^{-1/2} \end{align} for some constant $C$ and universal constant $p$. We now turn our attention to the $B'_k$ terms. Substituting a linear Taylor approximation with a quadratic error term for the exponential function in $\exp (-I(x))$ it is easy to see that \begin{align} B'_k =& \frac{\sqrt{n} L_Bs_k}{2\sqrt{\beta}x_k e^{I_k}} g_{k}\int_{x_{k-1}}^{x_{k+1}}\int_{x_k}^x dW(y)\, dx +R'_k, \end{align} where $R'_k$ satisfies the estimate (\ref{IgnoreR}) and thus can be ignored. Observe that \begin{align} &\E\left[g_{k}\int_{x_{k-1}}^{x_{k+1}}\int_{x_k}^x dW(y)\, dx\right]\\ &= \E\left[\int_{x_{k-1}}^{x_{k}} n^{3/2}(x-x_{k-1})dW(x) \left( \int_{x_{k-1}}^{x_k}(x_{k-1}-x)\, dW(x) + \int_{x_{k}}^{x_{k+1}} (x_{k+1}-x) \, dW(x) \right) \right]\\ & = -n^{3/2} \int_{x_{k-1}}^{x_{k}} (x-x_{k-1})^2 \, dx = -\frac{1}{3n^{3/2}}. \end{align} In fact, a short calculation gives the following decomposition: \begin{align} n^{3/2} g_{k}\int_{x_{k-1}}^{x_{k+1}} \int_{x}^{x_k}dW(y)\, dx =& \frac{-1}{3} + n^{1/2} \int_{x_{k-1}}^{x_{k}} Y_x \, d\ola{W}(x) \\ Y_x=&-2n^{5/2}\int_{x }^{\ck{x}} (x-\fk{x})(y-\fk{x})\, d\ola{W}(y)\\ &\qquad + n^{5/2}\int_{\ck{x}}^{\ck{x}+1/n}(x-\fk{x})(\ck{x}+1/n-y)\, d\ola{W}(y) \end{align} The important property of this decomposition is that for each $x$ we have $Y_x$ measurable with respect to $\{W_y:\; y\in (x,1)\}$. Notice also that we split up the power of $n$ so that $Y_x$ is typically order 1. We can then write the sum of $B'_k$ terms as follows \begin{align} \sum_{k=1}^n B'_k = \frac{-1}{3}\int_{0}^1 \frac{L_B s(\ck{x}) \, dx}{2\sqrt{\beta}\ck{x} e^{I(\ck{x})}} + \frac{1}{\sqrt{n}} \int_{0}^1 \frac{ L_B s(\ck{x}) Y_x }{2\sqrt{\beta}\ck{x} e^{I(\ck{x})}} \, d\ola{W}(x) + \sum_k R'_k \label{aoeurab} \end{align} It is easy to see that first term on the right hand side is within $O(t(\log n)^p n^{-1/2})$ of the right hand side of (\ref{eq857}). Due to the explicit factor of $1/\sqrt{n}$ and lemma \ref{ItoIntEst1}, the stochastic integral term on line (\ref{aoeurab}) is small within the tolerance of the lemma statement. As we have already said, the terms $R'_k$ are also sufficiently small. \end{proof} \subsection{Convergence of the $G^{(2)}$ terms} To complete the proof of proposition \ref{WtPhiConverges} it suffices to show the following \begin{lemma} On the event $\ul{X}(t)\geq 1/\log n$ we have \begin{align} \sum_{k=1}^n \frac{ G^{(2)}_k}{ x_k} L_X([x_{k},x_{k+1}),t) \leq C tn^{-0.499}. \end{align} The above estimate holds for almost all $B,W$ and the implied constant depends on $B,W,\beta $. \end{lemma} \begin{proof} As observed several times before, on the assumption $\ul{X}(t)\geq \log n$ we have $L_X(x,t)=0$ for all $x<1/\log n$ and therefore it suffices to assume that $x\geq 1/\log n$ throughout this proof. Recall the definition \begin{align} G^{(2)}_k =n\int_{x_{k-1}}^{x_{k+1}} \int_{x}^{x_{k+1}} f_2(y,x) \, d\ola{W}(y)\, d\ola{W}(x), \end{align} where $f_2$ is the function defined in display (\ref{GDEF}). By lemma \ref{TBound} we have that $|f_2|\leq C(\log n)^p$ for some universal constant $p$. A short calculation reveals that \begin{align} \sum_{k=1}^n \frac{G^{(2)}_k}{x_k} L_X([x_{k},x_{k+1}),t) =& \int_{0}^1 (Y_x+\wt{Y}_x)\, d\ola{W}(x) +\sum_{k=1}^n \frac{G^{(2)}_k}{x_k} \int_{x_{k}}^{x_{k+1}} \frac{L_B s(x)-L_B\wt{s}(x)}{2e^{I(x)}}\, dx ,\label{qoiajf} \end{align} where \begin{align} Y_x =& \frac{n}{\ck{x} } \int_{x}^{\ck{x}+1/n} f_2(y,x)\, d\ola{W}(y) \int_{\ck{x}}^{\ck{x}+1/n} \frac{L_B s(y)}{2e^{I(y)}}, \\ \wt{Y}_x =& \frac{n}{\fk{x} } \int_{x}^{\ck{x}} f_2(y,x)\, d\ola{W}(y) \int_{\fk{x}}^{\ck{x}} \frac{L_B \wt{s}(y)}{2e^{I(y)}}. \end{align} On the right hand side of (\ref{qoiajf}), the sum comes from removing a small anticipating component from the stochastic integral just like we did before in lemma \ref{BkLemma1111}. Using the bound $|s(x)-\wt{s}(x)|\leq C (\log n)^pn^{-3/2}$ and lemma \ref{LocalTimeLemma}, and also bounding $|G^{(2)}_k| \leq C(\log n)^p$ almost surely, we get that \begin{align} \sum_{k=1}^n \frac{G^{(2)}_k}{x_k} \int_{x_{k}}^{x_{k+1}} \frac{L_B s(x)-L_B\wt{s}(x)}{2e^{I(x)}}\, dx =O(t(\log n)^p n^{-3/4}). \end{align} Using lemmas \ref{TBound}, \ref{LocTimeBound} and \ref{ItoIntEst1} we see that the stochastic integral on the right hand side of (\ref{qoiajf}) is of order $t(\log n)^p n^{-1/2}$. \end{proof} \section*{Appendix A} \begin{proof}[Proof of lemma \ref{pkApproxLemma}] By the definition of $p_k$ we have \begin{align} \label{pk1} \frac{p_k}{q_k}= \frac{s(x_k)-s(x_{k-1})}{s(x_{k+1})-s(x_k)} = \frac{n\int_{x_{k-1}}^{x_{k}} \exp(-\int_{y}^{x_{k+1}}dV)\, dy} {n\int_{x_{k}}^{x_{k+1}} \exp(-\int_{y}^{x_{k+1}}dV)\, dy} \end{align} By definition, the upper limits of integration for the $dV$ integrals would be $1$, but the present expression is equivalent by cancelling a common factor. The stochastic integral in (\ref{pk1}) can be represented as \begin{align} -\int_{y}^{x_{k+1}} dV = \log \frac{k+1}{yn } -\frac{2}{\sqrt{\beta}} \int_{y}^{x_{k+1}} \frac{dW(x)}{\sqrt{x}}=\log \frac{k+1}{yn } +\frac{2}{\sqrt{\beta}} \wt{W}_{\log \frac{k+1}{yn}}, \end{align} where $\wt{W}$ is another standard Brownian motion. Since $ x_k\leq y \leq x_{k+1}$ we have that $\log \frac{k+1}{yn}=O(1/k)$ and so, by the law of the iterated logarithm, we have for almost all $W$ that \begin{align} |\wt{W}_{\log \frac{k+1}{yn}} | \leq \sqrt{(2/k) \log \log k}. \end{align} Applying Taylor's theorem with the remainder in Lagrange form, the denominator of (\ref{pk1}) can be expressed as \begin{align} \label{pk2} &n\int_{x_k}^{x_{k+1}} \frac{k+1}{yn}\bigg\{ 1-\frac{2}{\sqrt{\beta}}\int_{y}^{x_{k+1}} \frac{dW(x)}{\sqrt{x}}+\frac{2}{\beta} \left(\int_{y}^{x_{k+1}} \frac{dW(x)}{\sqrt{x}}\right)^2 - \frac{ 4 e^{-\xi(y)}}{3 \beta^{3/2}} \wt{W}_{\log \frac{k+1}{yn}}^3 \bigg\}\, dy, \end{align} where $|\xi(y)|\leq |\int_{y}^{x_{k+1}}\frac{dW(x)}{\sqrt{x}}|$. Using Taylor's theorem again, line (\ref{pk2}) can be expressed as \begin{align} 1 -\frac{2n}{\sqrt{\beta}}\int_{x_k}^{x_{k+1}} \int_{y}^{x_{k+1}}\frac{dW(x)}{\sqrt{x}}\,dy +\frac{1}{2k} +\frac{2n}{\beta} \int_{y}^{x_{k+1}}\left(\int_{x_k}^{x_{k+1}} \frac{dW(x)}{\sqrt{x}}\right)^2 \, dy+O\left( \frac{ (\log \log k)^{3/2}}{k^{3/2}}\right). \nonumber \end{align} We also have for almost all $W$ that for all $n,k$ the estimate \begin{align} \int_{y}^{x_{k+1}} \frac{dW(x)}{\sqrt{x}} = \frac{1}{\sqrt{x_k}}\int_{y}^{x_{k+1}}dW +O(\sqrt{\log k}/k^{3/2}) \end{align} holds. This can be checked by observing that the error term is a Gaussian random variable with variance of order $k^{-3/2}$, then applying a tail bound and Borel-Cantelli. Making the same manipulations, the numerator of (\ref{pk1}) can be expressed as \begin{align} 1+\frac{3}{2k} +\int_{x_{k-1}}^y \bigg\{- \frac{2}{\sqrt{\beta x_k}}\int_{y}^{x_{k+1}} dW +\frac{2}{\beta x_k} \left(\int_{y}^{x_{k+1} }dW \right)^2 \bigg\}\, dy +O\left(\frac{\sqrt{\log k}}{k^{3/2}}\right). \end{align} Using our approximations for the numerator and denominator in (\ref{pk1}), we get \begin{align} \label{pk3} \log \frac{p_k}{q_k} \sim & -\frac{2n}{\sqrt{\beta x_k}} (I_1 -I'_{1}) +\frac{1}{x_k n} +\frac{2n}{\beta x_k} (I_2 -I'_2) - \frac{2n^2}{\beta x_k} (I^{2}_1 -(I'_1)^2) +O\left(\frac{\sqrt{\log k}}{k^{3/2}}\right), \end{align} where \begin{align} I_1 =& \int_{x_{k-1}}^{x_{k+1}}(s\wedge x_k -x_{k-1})\, dW(s) ,\qquad I'_1 =\int_{x_{k-1}}^{x_{k+1}} (s\vee x_k -x_k)\, dW(s)\\ I_2 =&\int_{x_{k-1}}^{x_{k+1}} \int_{x_{k-1}}^{x_{k+1}} (s_1 \wedge s_2 \wedge x_k -x_{k-1})\, dW(s_1)\, dW(s_2) , \qquad \mathbb{E} I_2 =\frac{3}{2n^2}\\ I'_{2} =& \int_{x_k}^{x_{k+1}} \int_{x_k}^{x_{k+1}} (s_1 \wedge s_2 -x_{k})\, dW(s_1)\, dW(s_1), \qquad \mathbb{E} I'_2 =\frac{1}{2n^2} \end{align} Double Ito integrals of a symmetric function can be decomposed by \begin{align} \int_{0}^1 \int_{0}^1 f(s,t)\, dW(s)\, dW(t) =2 \int_{0}^1 \int_{0}^t f(s,t)\, dW(s)\, dW(t) + \int_{0}^1 f(s,s)\, ds.\label{aeorbga} \end{align} The point of doing this is that the stochastic integral on the right hand side of (\ref{aeorbga}) has mean zero. Using (\ref{aeorbga}) in (\ref{pk3}) some straitforward calculations give the formulas in our lemma statement. \end{proof}

https://arxiv.org/abs/1802.07015

Generalized nil-Coxeter algebras

Motivated by work of Coxeter (1957), we study a class of algebras associated to Coxeter groups, which we term 'generalized nil-Coxeter algebras'. We construct the first finite-dimensional examples other than usual nil-Coxeter algebras; these form a $2$-parameter type $A$ family that we term $NC_A(n,d)$. We explore the combinatorial properties of these algebras, including the Coxeter word basis, length function, maximal words, and their connection to Khovanov's categorification of the Weyl algebra.Our broader motivation arises from complex reflection groups and the Broue-Malle-Rouquier freeness conjecture (1998). With generic Hecke algebras over real and complex groups in mind, we show that the 'first' finite-dimensional examples $NC_A(n,d)$ are in fact the only ones, outside of the usual nil-Coxeter algebras. The proofs use a diagrammatic calculus akin to crystal theory.

\section{Introduction and main results} We study a new class of finite-dimensional algebras arising out of Coxeter theory, with connections to old work by Coxeter and new work on generic Hecke algebras, combinatorics, and categorification. We work throughout over a ground field $\bk$ for ease of exposition, although our results hold over any commutative unital ground ring. We begin with background and notation. Real reflection groups $W$ and their Iwahori--Hecke algebras $\mathcal{H}_W(q)$ are classical objects that have long been studied in algebraic combinatorics, representation theory, and mathematical physics. Recall that every Coxeter group is specified by a Coxeter matrix $M \in \Z^{I \times I}$, with finite index set $I$ and entries $m_{ii} = 2 \leqslant m_{ij} \leqslant \infty\ \forall i \neq j$. The corresponding Artin monoid $\B_M^{\geqslant 0}$ has generators $\{ T_i : i \in I \}$, and braid relations $T_i T_j T_i \cdots = T_j T_i T_j \cdots$ with $m_{ij}$ factors on each side whenever $m_{ij} < \infty$. We will denote the corresponding Coxeter group by $W(M)$. Three prominent algebras associated to $W(M)$ are its group algebra $\bk W(M)$, its $0$-Hecke algebra, and its nil-Coxeter algebra $NC(M)$ (also known in the literature as the nil Hecke ring, nil Coxeter algebra, and nilCoxeter algebra). All three algebras are quotients of the monoid algebra $\bk \B_M^{\geqslant 0}$ by quadratic relations for the $T_i$, and are `generic Hecke algebras' \cite[Chapter 7]{Hum}. Among such algebras, the $T_i$ satisfy homogeneous relations only in the case of the nil-Coxeter algebra $NC(M)$. Using this, one shows that \[ NC(M) := \bk \B_M^{\geqslant 0} / (T_i^2 : i \in I) \] is the monoid algebra of a monoid with $|W(M)|+1$ elements, say $\{ T_w : w \in W(M) \} \sqcup \{ O_{W(M)} \}$ quotiented by the `absorbing' central ideal $\bk O_{W(M)}$. Nil-Coxeter algebras were introduced by Fomin and Stanley \cite{FS}, and are related to flag varieties \cite{KK}, symmetric function theory \cite{BSS}, and categorification \cite{Kho}. We now present a larger family of algebras, which constitute the main object of study. Note that the algebras $NC(M)$ are the associated graded versions of the group algebra $\bk W(M)$; indeed, taking the top-degree component of the non-homogeneous relations $T_i^2 = 1$ yields the nil-Coxeter relations $T_i^2 = 0$. The following construction is motivated by both real and complex reflection groups, and allows the nilpotence degree to vary. \begin{defn}\label{Dnilcox} Define a {\em generalized Coxeter matrix} to be a symmetric matrix $M := (m_{ij})_{i,j \in I}$ with $I$ finite, $m_{ii} < \infty\ \forall i \in I$, and $2 \leqslant m_{ij} \leqslant \infty\ \forall i \neq j$. Now fix such a matrix $M$. \begin{enumerate} \item Given an integer tuple ${\bf d} = (d_i)_{i \in I}$ with all $d_i \geqslant 2$, let $M({\bf d})$ denote the matrix where the diagonal in $M$ is replaced by the coordinates of ${\bf d}$. Let $M_2 := M((2,\dots,2))$. \item The {\em generalized Coxeter group} $W(M)$ is the group generated by $\{ s_i : i \in I \}$ modulo the braid relations $s_i s_j s_i \cdots = s_j s_i s_j \cdots$ whenever $m_{ij} < \infty$, and the relations $s_i^{m_{ii}} = 1\ \forall i$. \item Define the corresponding {\em generalized nil-Coxeter algebra} to be: \begin{equation}\label{Egen} NC(M) := \frac{\bk \tangle{T_i, i \in I}} {(\underbrace{T_i T_j T_i \cdots}_{m_{ij}\ times} = \underbrace{T_j T_i T_j \cdots}_{m_{ij}\ times}, \ T_i^{m_{ii}} = 0, \ \forall i \neq j \in I)} = \frac{\bk \B_{M_2}^{\geqslant 0}}{(T_i^{m_{ii}} = 0\ \forall i)}, \end{equation} where we omit the braid relation $T_i T_j T_i \cdots = T_j T_i T_j \cdots$ if $m_{ij} = \infty$. \item As an important special case, we denote by $M_{A_n}$ the usual type $A$ Coxeter matrix with $|I| = n$, given by: $m_{ij} = 3$ if $|i-j|=1$, and $2$ otherwise. \end{enumerate} \end{defn} Working with generalized nil-Coxeter algebras $NC(M)$ yields a larger class of objects than the corresponding groups $W(M)$. For example, Marin \cite{Ma} has shown that in rank $2$ in type $A$, the algebra $NC(M_{A_2}((3,n)))$ is not finite-dimensional for $n \geqslant 3$, in particular for even $n$. However, the corresponding generalized Coxeter group $W(M_{A_2}((3,n)))$ is trivial for $3 \nmid n$, since in it the generators $s_1, s_2$ are conjugate, hence have equal orders. In fact, this reasoning shows that for for all integers $d_1, \dots, d_n \geqslant 2$, we have \[ W(M_{A_n}({\bf d})) = W(M_{A_n}((d,\dots,d))), \quad \text{where} \quad d = \gcd(d_1, \dots, d_n). \] Now it is natural to ask for which integers $n,d \geqslant 2$ is the group $W(M_{A_n}((d,\dots,d)))$ finite -- and what is its order. These questions were considered by Coxeter \cite{Co}, and he proved that $W = W(M_{A_n}((d, \dots, d)))$ is finite if and only if $\frac{1}{n} + \frac{1}{d} > \frac{1}{2}$; moreover, in this case $W$ has size $\left( \frac{1}{n} + \frac{1}{d} - \frac{1}{2} \right)^{1-n} \cdot n! / n^{n-1}$. In his thesis \cite{Ko}, Koster extended Coxeter's results to classify all finite generalized Coxeter groups; apart from the finite `usual' Coxeter groups, one obtains precisely the Shephard groups. In a parallel vein to these works, we explore for which matrices is the algebra $NC(M)$ finite-dimensional. In this we are also strongly motivated by the larger picture, which involves \textit{complex} reflection groups and the \textit{BMR freeness conjecture} \cite{BMR2}. We elaborate on these motivations presently; for now we remark that since complex reflections can have order $\geqslant 3$, working with them provides a natural reason to define and study generalized nil-Coxeter algebras. Returning to real groups: recall that `usual' nil-Coxeter algebras $NC(M((2,\dots,2)))$ are finite-dimensional precisely for finite Coxeter groups, since for `usual' Coxeter matrices $M_2 := M((2,\dots,2))$ one has $\dim NC(M_2) = |W(M_2)|$. To our knowledge there are no other finite-dimensional examples $NC(M)$ known to date. Our first main result parallels Coxeter's construction, and exhibits the first such `non-usual' family of finite-dimensional algebras $NC(M)$ in type $A$: \begin{theorem}\label{ThmA} For integers $n \geqslant 1$ and $d \geqslant 2$, define the $\bk$-algebra \begin{equation} NC_A(n,d) := NC(M_{A_n}((2,\dots,2,d))). \end{equation} Thus, $NC_A(n,d)$ has generators $T_1, \dots, T_n$, with relations: \begin{alignat}{5} T_i T_{i+1} T_i & = T_{i+1} T_i T_{i+1}, & & \forall\ 0 < i < n;\\ T_i T_j & = T_j T_i, & & \forall\ |i-j| > 1;\\ T_1^2 & = \cdots = T_{n-1}^2 & = T_n^d = &\ 0. \end{alignat} \noindent Then $NC_A(n,d)$ has a Coxeter word basis of $n! (1 + n(d-1))$ generators \[ \{ T_w : \ w \in S_n \} \ \sqcup \ \{ T_w T_n^k T_{n-1} T_{n-2} \cdots T_{m+1} T_m :\ w \in S_n,\ k \in [1,d-1],\ m \in [1,n] \}. \] In particular, the subalgebra $R_l$ generated by $T_1, \dots, T_l$ is isomorphic to the type $A$ nil-Coxeter algebra $NC(M_{A_l}((2,\dots,2)))$, for all $0 < l < n$. \end{theorem} \begin{remark} We adopt the following notation in the sequel without further reference: let \begin{equation} w_\circ \in S_{n+1}, \quad w'_\circ \in S_n \quad \text{denote the respective longest elements}, \end{equation} \noindent where the symmetric group $S_{l+1}$ corresponds to the basis of the algebra $R_l$ for $l = n-1, n$. \end{remark} In a later section, we will discuss additional properties of the algebras $NC_A(n,d)$, including identifying the `maximal' words, and exploring the Frobenius property. \subsection*{Classification of finite-dimensional nil-Coxeter algebras} Our next main result classifies the matrices $M$ for which the generalized nil-Coxeter algebra $NC(M)$ is finite-dimensional. In combinatorics and in algebra, classifying Coxeter-type objects of finite size, dimension, or type is a problem of significant classical as well as modern interest. Such settings include real and complex reflection groups \cite{Cox,Cox2,ST} and associated Hecke algebras; finite type quivers, simple Lie algebras, the McKay--Slodowy correspondence, and Kleinian singularities (as well as the above results by Coxeter and Koster). The recent classification of finite-dimensional pointed Hopf algebras \cite{AnSc} reveals connections to small quantum groups. Even more recently, the classification of finite-dimensional Nichols algebras has been well-received (see \cite{HV2} and the references therein); some ingredients used in proving those results show up in the present work as well. We now classify the generalized Coxeter matrices $M$ for which $NC(M)$ is finite-dimensional. Remarkably, outside of the usual nil-Coxeter algebras, our first family of examples $NC_A(n,d)$ turns out to be the only one: \begin{theorem}\label{ThmC} Suppose $W$ is a Coxeter group with connected Dynkin diagram. Fix an integer vector ${\bf d}$ with $d_i \geqslant 2\ \forall i$, i.e., a generalized Coxeter matrix $M({\bf d})$. The following are equivalent: \begin{enumerate} \item The generalized nil-Coxeter algebra $NC(M({\bf d}))$ is finite-dimensional. \item Either $W$ is a finite Coxeter group and $d_i = 2 \ \forall i$, or $W$ is of type $A_n$ and ${\bf d} = (2, \dots, 2, d)$ or $(d, 2, \dots, 2)$ for some $d>2$. \end{enumerate} \end{theorem} \begin{remark} The above results are characteristic-free; in fact they hold over arbitrary ground rings $\bk$, in which case Theorem \ref{ThmA} yields a $\bk$-basis of the free $\bk$-module $NC_A(n,d)$; and Theorem \ref{ThmC} classifies the finitely generated $\bk$-algebras $NC(M)$. In sketching the proofs of these results below, we will continue to assume $\bk$ is a field; for the general case over a ring $\bk$, for full details, and for further ramifications, we refer the reader to \cite{Kh}, of which this note is an extended abstract. \end{remark} Before proceeding further, we mention another strong motivation for Theorem \ref{ThmC}, arising from generic Hecke algebras over \textit{complex} reflection groups. As mentioned above, the varying nilpotence degree of the $T_i$ is natural in the setting of complex reflection groups $W$. A prominent area of research has been the study of the associated generic Hecke algebras $\mathcal{H}_W$ and the Brou\'e--Malle--Rouquier freeness conjecture \cite{BMR2}. The conjecture says that $\mathcal{H}_W$ is a free $R$-module of rank $|W|$, where $R$ is the ground ring. See also its recent resolution in characteristic zero \cite{Et}, and the references therein. In studying these topics, Marin \cite{Ma} remarks that the lack of nil-Coxeter algebras of dimension $|W|$ is a striking difference between complex and real reflection groups $W$. This was verified in some cases in \textit{loc.~cit.}; and it motivated us to define generalized nil-Coxeter algebras over all complex reflection groups. We do so in \cite{Kh}, and then completely classify the finite-dimensional algebras over all such groups. Remarkably, Theorem \ref{ThmC} extends to all complex $W$ as well, and the only finite-dimensional families are real (usual) nil-Coxeter algebras, and the family $NC_A(n,d)$. In particular, this shows the above statement of Marin. \begin{remark} Our result holds even more generally: following the classification of finite complex reflection groups in the celebrated work \cite{ST}, Popov classified in \cite{Po1} the \textit{infinite} discrete groups generated by unitary reflections. In \cite{Kh} we extend Theorem \ref{ThmC} to also cover all of these groups; once again, we show there are no finite-dimensional nil-Coxeter analogues. \end{remark} The equidimensionality (or not) of $\mathcal{H}_W$ and its nil-Coxeter analogue amounts to whether the former -- a filtered algebra -- is a \textit{flat} deformation of the latter, which is $\Z^{\geqslant 0}$-graded. The study of flat deformations goes back to classical work of Gerstenhaber, and also by Braverman--Gaitsgory, Drinfeld, Etingof--Ginzburg, and the recent program by Shepler and Witherspoon; see \cite{SW2,Kh} for more on this. In this formalism, Theorem \ref{ThmC} -- or its extension to complex groups -- says that over complex reflection groups, generic Hecke algebras are not flat deformations of their nil-Coxeter analogues. This is in stark contrast to the real case, where $\dim NC(M) = |W(M)|$. \section{A finite-dimensional generalized nil-Coxeter algebra} We now outline the proof of Theorem \ref{ThmA}, using a diagrammatic calculus as well as braid monoid computations. Note that $NC_A(1,d) = \bk[T_1] / (T_1^d)$, while $NC_A(n,2)$ is the usual type $A$ nil-Coxeter algebra, for which the theorem is well-known (see e.g. \cite{Hum}). Thus, in this section we will assume $d \geqslant 3$ and $n \geqslant 2$. We begin by showing that the set from Theorem \ref{ThmA} spans $NC_A(n,d)$. As a first step: \begin{lemma}\label{L1} A word in the generators $T_i$ either vanishes in $NC_A(n,d)$, or can be equated with a word in which all occurrences of $T_n$ are successive. \end{lemma} \begin{proof} Suppose a word $\mathcal{T}$ has a sub-word of the form $T_n^a T_{i_1} \cdots T_{i_k} T_n^b$ for some $a,b > 0$, with $0 < i_j < n\ \forall j$. Using the relations, we may assume the above representation of $\mathcal{T}$ is such that $k$ is minimal. Thus $i_1 = i_k = n-1$, $i_2 = i_{k-1} = n-2$, and so on (else we may push some $T_{i_j}$ outside of the sub-string). Hence the sub-string is of the form \[ T_{n-1} T_{n-2} \cdots T_{m+1} T_m T_{m+1} \cdots T_{n-2} T_{n-1}, \quad \text{for some } 1 \leqslant m \leqslant n-1. \] Now one shows by descending induction on $m \leqslant n-1$ that in the Artin monoid $\B_{M_{A_n}}^{\geqslant 0}$, \[ T_{n-1} \cdots T_m \cdots T_{n-1} = T_m T_{m+1} \cdots T_{n-2} T_{n-1} T_{n-2} \cdots T_{m+1} T_m. \] Hence, \begin{align}\label{Ereduce} T_n^a \cdot (T_{n-1} \cdots T_m \cdots T_{n-1}) \cdot T_n^b = &\ T_n^a \cdot (T_m \cdots T_{n-2} T_{n-1} T_{n-2} \cdots T_m) \cdot T_n^b\\ = &\ (T_m \cdots T_{n-2}) T_n^{a-1} (T_n T_{n-1} T_n) T_n^{b-1} (T_{n-2} \cdots T_m)\notag\\ = &\ (T_m \cdots T_{n-2}) T_n^{a-1} (T_{n-1} T_n T_{n-1}) T_n^{b-1} (T_{n-2} \cdots T_m).\notag \end{align} \noindent If $a,b \leqslant 1$ then the lemma follows. If instead $b>1$ then this expression contains as a substring $T_{n-1} (T_n T_{n-1} T_n) = T_{n-1}^2 T_n T_{n-1} = 0$, so we are done. Similarly if $a>1$. \end{proof} Now the subalgebra $R_{n-1}$ generated by $T_1, \dots, T_{n-1}$ satisfies the relations of the usual nil-Coxeter algebra $NC_A(n-1,2)$, so the words $\{ T_w : w \in S_n \}$ span it. By \eqref{Ereduce}, every nonzero word not in $R_{n-1}$ is of the form $T_w T_n^k T_{w'}$; writing $T_{w'}$ as a sub-string of minimal length, by above we may rewrite the word such that $T_{w'} = T_{n-1} \cdots T_m$. Hence, \[ NC_A(n,d) = R_{n-1} + \sum_{k=1}^{d-1} \sum_{m=1}^n R_{n-1} \cdot T_n^k \cdot (T_{n-1} \cdots T_m). \] Since $\dim R_{n-1} \leqslant n!$, the upper bound on $\dim NC_A(n,d)$ follows. The proof of the converse -- i.e., linear independence of the claimed word basis -- repeatedly uses some results about the permutation group $S_n$ and its nil-Coxeter algebra: \begin{lemma}\label{Lsymm} Suppose $W = S_n$, with simple reflections $s_1, \dots, s_{n-1}$ labelled as usual. Let $S_{n-1}$ be generated by $s_1, \dots, s_{n-2}$; then for all $w \in S_n \setminus S_{n-1}$, $w$ has a reduced expression as $w = w' s_{n-1} \cdots s_{m'}$, where $w' \in S_{n-1}$ and $m' \in [1,n-1]$ are unique. Given such an element $w \in S_n$, we have in the usual nil-Coxeter algebra $NC(M_{A_n}((2,\dots,2)))$: \begin{equation}\label{EnilcoxA} T_n \cdot T_w \cdot T_n \cdots T_m = \begin{cases} T_{w'} T_{n-1} \cdots T_{m-1} \cdot T_n \cdots T_{m'}, & \qquad \text{if } m' < m,\\ 0 & \qquad \text{otherwise}. \end{cases} \end{equation} \end{lemma} Now we introduce a diagrammatic calculus reminiscent of crystal theory from combinatorics and quantum groups. For simplicity, we begin by presenting the $n=2$ case. Let $\scrm$ be a $\bk$-vector space, with basis given by the nodes in Figure \ref{Fig1}. \begin{figure}[ht] \hspace*{7mm}\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm] \draw(1,2) circle (0.5cm); \draw(1,4) circle (0.5cm); \draw(5,2) circle (0.5cm); \draw(5,4) circle (0.5cm); \draw(7,2) circle (0.5cm); \draw(7,4) circle (0.5cm); \draw(9,2) circle (0.5cm); \draw(9,4) circle (0.5cm); \draw(11,2) circle (0.5cm); \draw(11,4) circle (0.5cm); \draw(15,2) circle (0.5cm); \draw(15,4) circle (0.5cm); \draw(8,5.5) circle (0.5cm); \draw(8,0.5) circle (0.5cm); \draw [->] (1,2.5) -- (1,3.5); \draw [->] (2.5,2) -- (1.5,2); \draw [->] (4.5,2) -- (3.5,2); \draw [->] (5,2.5) -- (5,3.5); \draw [->] (6.5,2) -- (5.5,2); \draw [->] (7,3.5) -- (7,2.5); \draw [->] (9,3.5) -- (9,2.5); \draw [->] (9.5,4) -- (10.5,4); \draw [->] (13.5,4) -- (14.5,4); \draw [->] (11,3.5) -- (11,2.5); \draw [->] (11.5,4) -- (12.5,4); \draw [->] (15,3.5) -- (15,2.5); \draw [->] (7.5,5.2) -- (7.1,4.55); \draw [->] (8.5,5.2) -- (8.9,4.55); \draw [->] (7.1,1.45) -- (7.5,0.8); \draw [->] (8.9,1.45) -- (8.5,0.8); \draw (0.52,2.4) node[anchor=north west] {$ 2^{d'} 1 $}; \draw (0.46,4.4) node[anchor=north west] {$ 1 2^{d'} 1 $}; \draw (4.58,2.35) node[anchor=north west] {$ 2^2 1 $}; \draw (4.46,4.35) node[anchor=north west] {$ 1 2^2 1 $}; \draw (6.65,2.27) node[anchor=north west] {$ 21 $}; \draw (6.75,4.3) node[anchor=north west] {$ 1 $}; \draw (8.65,2.27) node[anchor=north west] {$ 12 $}; \draw (8.75,4.3) node[anchor=north west] {$ 2 $}; \draw (10.58,2.35) node[anchor=north west] {$ 1 2^2 $}; \draw (10.7,4.35) node[anchor=north west] {$ 2^2 $}; \draw (14.5,2.35) node[anchor=north west] {$ 1 2^{d'} $}; \draw (14.7,4.35) node[anchor=north west] {$ 2^{d'} $}; \draw (2.7,2.22) node[anchor=north west] {$ \cdots $}; \draw (2.7,4.22) node[anchor=north west] {$ \cdots $}; \draw (12.7,2.22) node[anchor=north west] {$ \cdots $}; \draw (12.7,4.22) node[anchor=north west] {$ \cdots $}; \draw (7.76,5.8) node[anchor=north west] {$ \emptyset $}; \draw (7.65,0.7) node[anchor=north west] {$ w_\circ $}; \draw (0.6,3.2) node[anchor=north west] {$1$}; \draw (1.8,2) node[anchor=north west] {$2$}; \draw (3.8,2) node[anchor=north west] {$2$}; \draw (4.6,3.2) node[anchor=north west] {$1$}; \draw (5.8,2) node[anchor=north west] {$2$}; \draw (6.6,3.4) node[anchor=north west] {$2$}; \draw (6.8,1.4) node[anchor=north west] {$1$}; \draw (6.9,5.35) node[anchor=north west] {$1$}; \draw (8.75,1.4) node[anchor=north west] {$2$}; \draw (8.7,5.35) node[anchor=north west] {$2$}; \draw (8.95,3.4) node[anchor=north west] {$1$}; \draw (9.7,4.5) node[anchor=north west] {$2$}; \draw (10.95,3.4) node[anchor=north west] {$1$}; \draw (11.7,4.5) node[anchor=north west] {$2$}; \draw (13.7,4.5) node[anchor=north west] {$2$}; \draw (14.95,3.4) node[anchor=north west] {$1$}; \end{tikzpicture} \caption{Regular representation for $NC_A(2,d)$, with $d' = d-1$} \label{Fig1} \end{figure} In this figure, the node $2^2 1$ should be thought of as applying $T_2^2 T_1$ to the generating basis vector/node $\emptyset$; similarly for all other nodes. The arrows show the action of $T_1, T_2$ on the basis vectors (i.e., nodes), and the lack of an arrow labeled $i$ with source $v \in \scrm$ means $T_i v = 0$. Now verify by inspection that the relations in $NC_A(2,d)$ are satisfied in ${\rm End}_\bk(\scrm)$, whence $\scrm$ is a cyclic $NC_A(2,d)$-module generated by the vector $\emptyset$ -- in fact, the regular representation. This gives the desired result for $NC_A(2,d)$. For general $n \geqslant 2$, the strategy is similar but with more involved notation. For $w \in S_n$, let $T_w$ denote the (well-defined) word in $T_1, \dots, T_{n-1} \in NC_A(n,d)$. Now define a vector space $\scrm$ with basis given by \eqref{Ebasis} and $NC_A(n,d)$-action as in Figure \ref{Fig2} below: \begin{equation}\label{Ebasis} \B := \{ B(w,k,m) : w \in S_n,\ k \in [1, d-1],\ m \in [1,n] \} \sqcup \{ B(w) : w \in S_n \}. \end{equation} \begin{figure}[ht] \hspace*{2cm}\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm] \draw(1.5,1.5) circle (0.5cm); \draw(1.5,3.5) circle (0.5cm); \draw(3.5,2.5) circle (0.5cm); \draw(3.5,0.5) circle (0.5cm); \draw(5,3.5) circle (0.5cm); \draw(6,0.5) circle (0.5cm); \draw(7,2.5) circle (0.5cm); \draw(11,3.5) circle (0.5cm); \draw(12,0.5) circle (0.5cm); \draw(13,2.5) circle (0.5cm); \draw [->] (1.93,3.2) -- (3.05,2.75); \draw [->] (5.43,3.2) -- (6.55,2.75); \draw [->] (11.43,3.2) -- (12.55,2.75); \draw [->] (2,3.5) -- (4.46,3.5); \draw [->] (4,2.5) -- (6.46,2.5); \draw [->] (5.5,3.5) -- (7.2,3.5); \draw [->] (7.5,2.5) -- (9.2,2.5); \draw [->] (8.5,3.5) -- (10.46,3.5); \draw [->] (10.5,2.5) -- (12.46,2.5); \draw (1.5,3)-- (1.5,2); \draw (1.93,1.2)-- (3.05,0.75); \draw (3.5,2)-- (3.5,1); \draw (5.17,3)-- (5.85,1); \draw (7,2)-- (6.35,0.85); \draw (11.17,3)-- (11.85,1); \draw (13,2)-- (12.35,0.85); \draw (1,3.78) node[anchor=north west] {$ 11m $}; \draw (2.9,2.8) node[anchor=north west] {$ w'1m $}; \draw (4.5,3.78) node[anchor=north west] {$ 12m $}; \draw (6.4,2.8) node[anchor=north west] {$ w'2m $}; \draw (10.45,3.83) node[anchor=north west] {$ 1d'm $}; \draw (12.37,2.84) node[anchor=north west] {$ w'd'm $}; \draw (7.5,3.71) node[anchor=north west] {$ \cdots $}; \draw (9.5,2.71) node[anchor=north west] {$ \cdots $}; \draw (2.2,2) node[anchor=north west] {$ V_1 $}; \draw (5.6,2) node[anchor=north west] {$ V_{2,m} $}; \draw (11.6,2) node[anchor=north west] {$ V_{d',m} $}; \draw (4.5,2.5) node[anchor=north west] {$n$}; \draw (8,2.5) node[anchor=north west] {$n$}; \draw (10.7,2.5) node[anchor=north west] {$n$}; \draw (3,3.9) node[anchor=north west] {$n$}; \draw (6,3.9) node[anchor=north west] {$n$}; \draw (9,3.9) node[anchor=north west] {$n$}; \end{tikzpicture} \caption{Regular representation for $NC_A(n,d)$, with $d' = d-1$} \label{Fig2} \end{figure} Note that $\dim_\bk \scrm = n! (1 + n(d-1))$; that the basis vectors in \eqref{Ebasis} are to be thought of as akin to $T_w T_n^k T_{n-1} \cdots T_m$ and $T_w$ respectively; and the nodes $(wkm), (w)$ precisely denote the basis vectors $B(w,k,m), B(w)$ respectively. Now let $V_1$ denote the span of the vectors $\{ B(w) : w \in S_{n-1} \} \sqcup \{ B(w,1,m) : w \in S_{n-1}, m \in [1,n] \}$. These vectors are in bijection with the word basis of the usual nil-Coxeter algebra $NC_A(n,2)$. Similarly for $k \in [1,d-1]$ and $m \in [1,n]$, define $V_{k,m}$ to be the span of the vectors $B(w,k,m), w \in S_n$. Now we define the $NC_A(n,d)$-action on $\scrm$. First for the action of $T_1, \dots, T_{n-1}$, write $V_{1,n+1} := {\rm span} \{ B(w) : w \in S_n \}$; and equip each space $V(k,m)$ for $k \in [1,d-1], m \in [1,n]$ and also $V(1,n+1)$ with the structure of the regular representation of $R_{n-1}$. Next, if $w \in S_{n-1}$, then we define \[ T_n \cdot B(w,k,m) := {\bf 1}(k \leqslant d-2) B(w,k+1,m), \qquad T_n \cdot B(w) := B(w,1,n). \] Now suppose $w \in S_n \setminus S_{n-1}$. Using Lemma \ref{Lsymm}, write $w = w' s_{n-1} \cdots s_m$; then $m \leqslant n-1$. Define $T_n \cdot B(w,k,m) := 0$ if $k>1$; also set $T_n \cdot B(w) := B(w',1,m')$; finally, \begin{equation}\label{Ereln} T_n \cdot B(w,1,m) := \begin{cases} B(w' s_{n-1} \cdots s_{m-1},1,m'), & \qquad \text{if } m' < m,\\ 0 & \qquad \text{otherwise}. \end{cases} \end{equation} Then some involved computations using the identities mentioned above show that the proposed action indeed equips $\scrm$ with the structure of a cyclic $NC_A(n,d)$-module, generated by $B(1)$. This allows us to complete the proof of Theorem \ref{ThmA}. \qed \section{Further properties} We now discuss several additional properties of the algebras $NC_A(n,d)$. For proofs of results in this section, we refer the reader to \cite{Kh}. The first set of properties shows how these algebras resemble usual nil-Coxeter algebras. \begin{theorem}[see \cite{Kh}]\label{ThmB} Fix integers $n \geqslant 1$ and $d \geqslant 2$. \begin{enumerate} \item The algebra $NC_A(n,d)$ has a length function that restricts to the usual length function $\ell_{A_{n-1}}$ on $R_{n-1} \simeq NC_{A_{n-1}}((2,\dots,2))$ (from Theorem \ref{ThmA}), and \begin{equation}\label{Elength} \ell(T_w T_n^k T_{n-1} \cdots T_m) = \ell_{A_{n-1}}(w) + k + n-m, \end{equation} for all $w \in S_n$, $k \in [1,d-1]$, and $m \in [1,n]$. \item There is a unique longest word $T_{w'_\circ} T_n^{d-1} T_{n-1} \cdots T_1$ of length \[ l_{n,d} := \ell_{A_{n-1}}(w'_\circ) + d+n-2. \] \item The algebra $NC_A(n,d)$ is local, with unique maximal (augmentation) ideal $\m$ generated by $T_1, \dots, T_n$. The ideal $\m$ is nilpotent with $\m^{1 + l_{n,d}} = 0$. \end{enumerate} \end{theorem} Thus there is a variant of the Coxeter word length, as well as a unique longest word and nilpotent augmentation ideal. As an immediate consequence, one can compute the Hilbert polynomial of the graded algebra $NC_A(n,d)$: \begin{cor}\label{Chilb} If $T_1, \dots, T_n$ all have degree $1$, then $NC_A(n,d)$ has Hilbert--Poincar\'e series \[ [n]_q! \; (1 + [n]_q \; [d-1]_q), \qquad \text{where } [n]_q := \frac{q^n-1}{q-1}, \ [n]_q! := \prod_{j=1}^n [j]_q. \] \end{cor} Here, we also use the standard result that the usual nil-Coxeter algebra $NC_A(n,2)$ has Hilbert--Poincar\'e series $[n]_q!$ (see e.g.~\cite[\S 3.12, 3.15]{Hum}). Having discussed similarities with usual nil-Coxeter algebras, we next present certain differences in structure. For any generalized Coxeter matrix $M$, define $x \in NC(M)$ to be \textit{left-primitive} if $T_i x = 0\ \forall i \in I$. Similarly define \textit{right-primitive} elements; and an element that is both is said to be \textit{primitive}. We denote these subspaces of $NC(M)$ by \[ \Prim_L(NC(M)), \quad \Prim_R(NC(M)), \quad \Prim(NC(M)). \] \begin{proposition}[see \cite{Kh}]\label{Pprim} Every generalized nil-Coxeter algebra $NC(M)$ is equipped with an anti-involution $\theta$ that fixes each generator $T_i$. Now $\theta$ is an isomorphism $: \Prim_L(NC(M)) \longleftrightarrow \Prim_R(NC(M))$. Moreover, the following hold. \begin{enumerate} \item If $NC(M) = NC_A(1,d)$, then \[ \Prim_L(NC(M)) = \Prim_R(NC(M)) = \Prim(NC(M)) = \bk \cdot T_1^{d-1}. \] \item If $NC(M) = NC_A(n,d)$ with $n \geqslant 2$ and $d \geqslant 2$, then: \begin{enumerate} \item $\Prim_L(NC(M))$ is spanned by $T_{w_\circ} := T_{w'_\circ} T_n T_{n-1} \cdots T_1$ and the $n(d-2)$ words \[ \{ T_{w'_\circ} T_n^k T_{n-1} \cdots T_m : \ k \in [2, d-1], \ m \in [1,n] \}. \] \item $\Prim(NC(M))$ is spanned by the words $T_{w'_\circ} T_n^k T_{n-1} \cdots T_1$, where $1 \leqslant k \leqslant d-1$. \end{enumerate} \end{enumerate} \noindent In all cases, the map $\theta$ fixes both $\Prim(NC(M))$ as well as the lengths of all nonzero words. \end{proposition} \noindent (Thus there are multiple primitive words for $d>2$.) Using Proposition \ref{Pprim}, we address another difference with usual nil-Coxeter algebras: the latter are always Frobenius \cite{Kho}. It is natural to ask when the finite-dimensional algebras $NC_A(n,d)$ share this property. \begin{proposition} The algebra $NC_A(n,d)$ is Frobenius if and only if $n=1$ or $d=2$. \end{proposition} In fact this happens if and only if the group algebra $\bk W(M_{A_n}({\bf d}))$ is a flat deformation of $NC_A(n,d)$. Flat deformations will be further discussed in the final section. \begin{proof} If $W(M)$ is a finite Coxeter group, \cite[\S 2.2]{Kho} shows that $NC(M)$ is Frobenius. Next, one easily verifies $NC_A(1,d) = \bk[T_1] / (T_1^d)$ is Frobenius, via the symmetric bilinear form given by: $\sigma(T_1^i, T_1^j) = {\bf 1}(i+j=d-1)$. Now suppose for some $n,d$ that $NC_A(n,d)$ is Frobenius, with nondegenerate invariant bilinear form $\sigma$. For each primitive $p \neq 0$, there exists $a_p$ such that $0 \neq \sigma(p,a_p) = \sigma(p a_p, 1)$. Thus, we can take $a_p = 1, \ \forall p$. Since the functional $\sigma(-,1) : \Prim(NC_A(n,d)) \to \bk$ is nonsingular, we obtain $\dim_\bk \Prim(NC_A(n,d)) = 1$. Applying Proposition \ref{Pprim}, we get $n=1$ or $d=2$. \end{proof} Finally, recall the famous result by Khovanov \cite{Kho} that the Weyl algebra $W_n := \Z \langle x, \partial \rangle / (\partial x = 1 + x \partial)$ can be represented by functors on bimodule categories over usual nil-Coxeter algebras. (Here we use that the nil-Coxeter algebra $\mathcal{A}_n := NC_A(n,2)$ is a bimodule over $\mathcal{A}_{n-1}$.) We now explain how $NC_A(n,d)$ fits into Khovanov's framework for $d \geqslant 2$, noting that for $d=2$ it was proved in \cite{Kho}: \begin{proposition}[see \cite{Kh}]\label{Pkhovanov} For $n \geqslant 1$ and $d \geqslant 2$, there is an isomorphism of $\mathcal{A}_{n-1}$-bimodules: \[ NC_A(n,d) \simeq \mathcal{A}_{n-1} \oplus \bigoplus_{k=1}^{d-1} \left( \mathcal{A}_{n-1} \otimes_{\mathcal{A}_{n-2}} \mathcal{A}_{n-1} \right). \] \end{proposition} For $d \geqslant 2$, in the notation of \cite{Kho} this result implies that over $\mathcal{A}_{n-1}$-bimodules, the algebra $NC_A(n,d)$ corresponds to $1 + (d-1) x \partial$. Thus, Proposition \ref{Pkhovanov} strengthens Theorems \ref{ThmA} and \ref{ThmB}, which discussed a left $\mathcal{A}_{n-1}$-module structure on $NC_A(n,d)$ (namely, $NC_A(n,d)$ is free of rank $1 + n(d-1)$). \section{All finite-dimensional generalized nil-Coxeter algebras} We conclude by proving Theorem \ref{ThmC}. Clearly $(2) \implies (1)$ by Theorem \ref{ThmA} and \cite[Chapter 7]{Hum}. Now suppose $(1)$ holds and ${\bf d} \neq (2,\dots,2)$. We again use the diagrammatic calculus above, now for the diagrams in Figure \ref{Fig3}. \begin{figure}[ht] \hspace*{14mm}\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm] \draw(5.8,13) circle (0.25cm); \draw(7.5,14.1) circle (0.4cm); \draw(9.5,14.1) circle (0.4cm); \draw(11.5,14.1) circle (0.4cm); \draw(13.5,14.1) circle (0.4cm); \draw(13.5,11.9) circle (0.4cm); \draw(11.5,11.9) circle (0.4cm); \draw(9.5,11.9) circle (0.4cm); \draw(7.5,11.9) circle (0.4cm); \draw(14.9,13) circle (0.25cm); \draw (5.5,13.3) node[anchor=north west] {$A$}; \draw (7.1,14.4) node[anchor=north west] {$B_1$}; \draw (9.1,14.4) node[anchor=north west] {$B_2$}; \draw (11,14.4) node[anchor=north west] {$B_{m'}$}; \draw (13.05,14.4) node[anchor=north west] {$B_m$}; \draw (13,12.2) node[anchor=north west] {$B'_m$}; \draw (11,12.2) node[anchor=north west] {$B'_{m'}$}; \draw (9.1,12.2) node[anchor=north west] {$B'_2$}; \draw (7.1,12.2) node[anchor=north west] {$B'_1$}; \draw (14.6,13.3) node[anchor=north west] {$C$}; \draw [->] (7,12.2) -- (6.1,12.8); \draw [->] (6.1,13.2) -- (7,14); \draw [->] (8,14.1) -- (9,14.1); \draw [->] (12,14.1) -- (13,14.1); \draw [->] (13,11.9) -- (12,11.9); \draw [->] (9,11.9) -- (8,11.9); \draw [->] (13.9,13.8) -- (14.6,13.2); \draw [->] (14.6,12.8) -- (13.9,12.1); \draw (6.2,12.5) node[anchor=north west] {$ \alpha $}; \draw (6.2,14) node[anchor=north west] {$ \alpha $}; \draw (8.1,14.7) node[anchor=north west] {$ \beta_1 $}; \draw (10.1,14.3) node[anchor=north west] {$ \cdots $}; \draw (12,14.7) node[anchor=north west] {$ \beta_{m'} $}; \draw (14.2,14.1) node[anchor=north west] {$ \gamma $}; \draw (14.2,12.5) node[anchor=north west] {$ \gamma $}; \draw (12.3,11.9) node[anchor=north west] {$ \beta_{m'} $}; \draw (10.1,12.1) node[anchor=north west] {$ \cdots $}; \draw (8.3,11.9) node[anchor=north west] {$ \beta_1 $}; \draw (6,13.3) node[anchor=north west] {+}; \draw (8.7,11.1) node[anchor=north west] {Fig. 3.1 ($m' = m-1$)}; \draw(3,5.7) circle (0.25cm); \draw(5,5.7) circle (0.25cm); \draw(3,3.7) circle (0.25cm); \draw (2.7,5.95) node[anchor=north west] {\textit{A}}; \draw (4.7,5.95) node[anchor=north west] {\textit{B}}; \draw (2.7,3.95) node[anchor=north west] {\textit{C}}; \draw [->] (3,4.1) -- (3,5.3); \draw [->] (3.4,5.7) -- (4.6,5.7); \draw [->] (4.7,5.4) -- (3.3,4); \draw (2.5,4.9) node[anchor=north west] {$ t $}; \draw (3.7,6.2) node[anchor=north west] {$ s $}; \draw (4,4.8) node[anchor=north west] {$ u $}; \draw (3.1,5.6) node[anchor=north west] {+}; \draw (3.4,3.6) node[anchor=north west] {Fig. 3.2}; \draw(5.8,7.7) circle (0.25cm); \draw(7.5,8.8) circle (0.4cm); \draw(9.5,8.8) circle (0.4cm); \draw(11.5,8.8) circle (0.4cm); \draw(13.5,8.8) circle (0.4cm); \draw(13.5,6.6) circle (0.4cm); \draw(11.5,6.6) circle (0.4cm); \draw(9.5,6.6) circle (0.4cm); \draw(7.5,6.6) circle (0.4cm); \draw(12.1,7.7) circle (0.25cm); \draw(14.9,7.7) circle (0.25cm); \draw (5.5,8) node[anchor=north west] {$A$}; \draw (7.1,9.1) node[anchor=north west] {$B_1$}; \draw (9.1,9.1) node[anchor=north west] {$B_2$}; \draw (11,9.1) node[anchor=north west] {$B_{m'}$}; \draw (13.05,9.1) node[anchor=north west] {$B_m$}; \draw (13,6.9) node[anchor=north west] {$B'_m$}; \draw (11,6.9) node[anchor=north west] {$B'_{m'}$}; \draw (9.1,6.9) node[anchor=north west] {$B'_2$}; \draw (7.1,6.9) node[anchor=north west] {$B'_1$}; \draw (11.8,8) node[anchor=north west] {$D$}; \draw (14.6,8) node[anchor=north west] {$C$}; \draw [->] (7,6.9) -- (6.1,7.5); \draw [->] (6.1,7.9) -- (7,8.7); \draw [->] (8,8.8) -- (9,8.8); \draw [->] (12,8.8) -- (13,8.8); \draw [->] (13,6.6) -- (12,6.6); \draw [->] (9,6.6) -- (8,6.6); \draw [->] (13.1,8.6) -- (12.3,7.9); \draw [->] (12.3,7.5) -- (13.1,6.9); \draw [->] (13.9,8.5) -- (14.6,7.9); \draw [->] (14.6,7.5) -- (13.9,6.8); \draw (6.2,7.2) node[anchor=north west] {$ \alpha $}; \draw (6.2,8.7) node[anchor=north west] {$ \alpha $}; \draw (8.1,9.4) node[anchor=north west] {$ \beta_1 $}; \draw (10.1,9) node[anchor=north west] {$ \cdots $}; \draw (12,9.4) node[anchor=north west] {$ \beta_{m'} $}; \draw (12.7,8.4) node[anchor=north west] {$ \gamma $}; \draw (12.7,7.6) node[anchor=north west] {$ \delta $}; \draw (14.2,8.8) node[anchor=north west] {$ \delta $}; \draw (14.2,7.2) node[anchor=north west] {$ \gamma $}; \draw (12.3,6.6) node[anchor=north west] {$ \beta_{m'} $}; \draw (10.1,6.8) node[anchor=north west] {$ \cdots $}; \draw (8.3,6.6) node[anchor=north west] {$ \beta_1 $}; \draw (6,8) node[anchor=north west] {+}; \draw (8.7,5.8) node[anchor=north west] {Fig. 3.3 ($m' = m-1$)}; \draw(5.8,2.1) circle (0.25cm); \draw(7.5,3.2) circle (0.4cm); \draw(9.5,3.2) circle (0.4cm); \draw(11.5,3.2) circle (0.4cm); \draw(13.5,3.2) circle (0.4cm); \draw(13.5,1) circle (0.4cm); \draw(11.5,1) circle (0.4cm); \draw(9.5,1) circle (0.4cm); \draw(7.5,1) circle (0.4cm); \draw (5.5,2.4) node[anchor=north west] {$A$}; \draw (7.1,3.5) node[anchor=north west] {$B_1$}; \draw (9.1,3.5) node[anchor=north west] {$B_2$}; \draw (11,3.5) node[anchor=north west] {$B_{m'}$}; \draw (13.05,3.5) node[anchor=north west] {$B_m$}; \draw (13,1.3) node[anchor=north west] {$B'_m$}; \draw (11,1.3) node[anchor=north west] {$B'_{m'}$}; \draw (9.1,1.3) node[anchor=north west] {$B'_2$}; \draw (7.1,1.3) node[anchor=north west] {$B'_1$}; \draw [->] (7,1.3) -- (6.1,1.9); \draw [->] (6.1,2.3) -- (7,3.1); \draw [->] (8,3.2) -- (9,3.2); \draw [->] (12,3.2) -- (13,3.2); \draw [->] (13.5,2.7) -- (13.5,1.5); \draw [->] (13,1) -- (12,1); \draw [->] (9,1) -- (8,1); \draw (6.2,1.6) node[anchor=north west] {$ \alpha $}; \draw (6.2,3.1) node[anchor=north west] {$ \alpha $}; \draw (8.1,3.8) node[anchor=north west] {$ \beta_1 $}; \draw (10.1,3.4) node[anchor=north west] {$ \cdots $}; \draw (12,3.8) node[anchor=north west] {$ \beta_{m'} $}; \draw (13.6,2.4) node[anchor=north west] {$ \gamma $}; \draw (12.3,1) node[anchor=north west] {$ \beta_{m'} $}; \draw (10.1,1.2) node[anchor=north west] {$ \cdots $}; \draw (8.3,1) node[anchor=north west] {$ \beta_1 $}; \draw (6,2.4) node[anchor=north west] {+}; \draw (8.7,0) node[anchor=north west] {Fig. 3.4 ($m' = m-1$)}; \end{tikzpicture} \caption{Modules for the infinite-dimensional generalized nil-Coxeter algebras} \label{Fig3} \end{figure} We consider the possible cases, showing in each case that the algebra $NC(M)$ is infinite-dimensional, until we are left with only $NC_A(n,d)$. First suppose there exist two nodes $\alpha, \gamma \in I$ with $m_{\alpha \alpha}, m_{\gamma \gamma} \geqslant 3$. Since the Dynkin diagram of $I$ is connected by assumption, there exist $\beta_1, \dots, \beta_{m-1} \in I$ such that \[ \alpha \quad \longleftrightarrow \quad \beta_1 \quad \longleftrightarrow \quad \cdots \quad \longleftrightarrow \quad \beta_{m-1} \quad \longleftrightarrow \quad \gamma \] are all connected in $I$, i.e., a path. Now define an $NC(M)$-module $\scrm$ with basis \[ A_r, B_{1r}, \dots, B_{mr}, C_r, B'_{1r}, \dots, B'_{mr}, \quad r \geqslant 1, \] and where every $T_i$ kills all basis vectors, \textit{except} for the actions described in Figure 3.1, namely $T_\alpha(A_r) := B_{1r}, T_{\beta_1}(B'_{2r}) := B'_{1r}$, and so on for all $r \geqslant 1$. The `$+$' indicates that $T_\alpha(B'_{1r}) := A_{r+1}\ \forall r \geqslant 1$. One verifies that the $T_i$ satisfy the $NC(M)$-relations on every basis vector, whence on $\scrm$. Now as $\scrm$ is cyclic and infinite-dimensional, so is $NC(M)$. The strategy is similar for the remainder of the proof. Henceforth we fix the unique node $\alpha \in I$ such that $m_{\alpha \alpha} \geqslant 3$. If $\alpha$ is connected in $I$ to $\gamma$ with $m_{\alpha \gamma} \geqslant 4$, then we work with Figure 3.2, setting $(s,t,u) \leadsto (\alpha, \alpha, \gamma)$, and define $\scrm := {\rm span}_\bk \{ A_r, B_r, C_r : r \geqslant 1 \}$. Now check that $\scrm$ is an infinite-dimensional cyclic $NC(M)$-module. Next, if $\alpha$ is adjacent to two nodes $\gamma, \delta \in I$, work with Figure 3.3 for $m=1$. This shows $\alpha$ must be extremal. Note that if $NC(M)$ is finite-dimensional then so is its quotient $NC(M((2,\dots,2)))$, which is a nil-Coxeter algebra. Hence $W(M((2,\dots,2)))$ is a finite Coxeter group, and these are known \cite{Cox,Cox2}. We now sketch how to eliminate all cases not of type $A$, whence from above, $NC(M) \cong NC_A(n,d)$, where we set $n := |I|$. The dihedral types $G_2, H_2, I$ do not hold from above. Suppose $I$ is of type $B,C,H$: \[ \alpha \quad \longleftrightarrow \quad \beta_1 \quad \longleftrightarrow \quad \cdots \quad \longleftrightarrow \quad \beta_{m-1} \quad \longleftrightarrow \quad \gamma, \] \noindent with $m_{\alpha \alpha} \geqslant 3, m_{\gamma \gamma} = 2, m_{\beta_{m-1} \gamma} \geqslant 4$. Then work with Figure 3.4. Note this also rules out the $F_4$ case, since $NC(M_{F_4}) \twoheadrightarrow NC(M_{B_3})$ or $NC(M_{C_3})$ by killing the extremal generator $T_\delta$ with $m_{\delta \delta} = 2$, and we showed that the latter two algebras are infinite-dimensional. Next suppose $I$ is of type $D$. If $\alpha$ is a (extremal) node on the `long arm', work with Figure 3.3 with $m=n-2$, with the other extremal nodes $\gamma, \delta$. Else if $\alpha$ is extremal on a short arm, we work as above with the quotient algebra $NC(M_{D_4})$ of Dynkin type $D_4$, by killing all $T_j$ with node $j$ on the long arm having degree $\leqslant 2$. Working with Figure 3.3 for $m=2$, it follows that $NC(M_{D_4})$, and hence $NC(M)$, is infinite-dimensional. Finally, in all remaining cases, the Coxeter graph of $I$ is of type $E$. Akin to above, these cases are ruled out by quotienting to reduce to type $D$. This concludes the proof.\qed

https://arxiv.org/abs/1010.0722

A Littlewood-Richardson rule for Macdonald polynomials

Macdonald polynomials are orthogonal polynomials associated to root systems, and in the type A case, the symmetric kind is a common generalization of Schur functions, Macdonald spherical functions, and Jack polynomials. We use the combinatorics of alcove walks to calculate products of monomials and intertwining operators of the double affine Hecke algebra. From this, we obtain a product formula for Macdonald polynomials of general type.

\section{Introduction} In~\cite{M88}, Macdonald introduced a remarkable family of orthogonal polynomials $P_\lambda(q,t)$ associated with root systems. For special values of $q$ and $t$, they specialize to various well-known functions, including Weyl characters and spherical functions for $p$-adic groups. These polynomials are a basis for symmetric functions, and are a common generalization of Schur functions $s_\lambda$, monomial symmetric functions, Hall-Littlewood polynomials, and the symmetric Jack polynomials. The symmetric Macdonald polynomials are indexed by dominant weights of the weight lattice $P$. Classically, the Littlewood-Richardson coefficients $c_{\lambda\mu}^\nu$ are the structure constants of the ring of symmetric functions with respect to the Schur basis: $$s_\lambda s_\mu = \sum_{\nu} c_{\lambda\mu}^\nu s_\nu.$$ In the representation theory of the general linear group $\GL_n(\bbC)$, the Littlewood-Richardson coefficients also give the multiplicity of the irreducible highest weight module $V(\nu)$ in $V(\lambda)\otimes V(\mu)$. The coefficient $c_{\lambda\mu}^\nu$ is given combinatorially as the number of Young tableaux of shape $\nu\backslash \lambda$ admitting a Littlewood-Richardson filling of type $\mu$. Littelmann introduced the path model in~\cite{Li94} as a tool for calculating formulas for characters of complex symmetrizable Kac-Moody algebras, and showed that it can also be used to compute Littlewood-Richardson coefficients. Instead of a sum over tableaux, his formula for $c_{\lambda\mu}^\nu$ is a sum over certain paths in the vector space $P\otimes_\bbZ \bbR$, where the endpoint (weight) of a path takes the place of the filling of a tableau. Several variations of the Littelmann path model were introduced to obtain character formulas, including the gallery model of Gaussent-Littelmann~\cite{GL02}, and the model of Lenart-Postnikov~\cite{LP08} based on $\lambda$-chains. In~\cite{R06}, Ram developed the alcove walk model for working in the affine Hecke algebra, and the paper~\cite{RY} showed that alcove walks are a useful tool for expanding products of intertwining operators of the double affine Hecke algebra. Cherednik developed the theory of double affine Hecke algebras, using it to solve Macdonald's constant term conjectures~\cite{C95a}, and in~\cite{C95b}, he showed that products of intertwining operators of the double affine Hecke algebra generate the nonsymmetric Macdonald polynomials $E_\lambda(q,t)$, which are a family of orthogonal polynomials indexed by points of the weight lattice. These polynomials were first introduced by Opdam~\cite{O95} in the case $q\rightarrow 1$ (see~\cite[p.147]{M03}). By applying a symmetrizing operator $\mathbf{1}_0$ to $E_\lambda$, one can obtain the symmetric polynomials $P_\lambda$. In this paper, we use the alcove walk model to calculate products of monomials and intertwining operators of the double affine Hecke algebra (Theorem~\ref{thm.X-tau}), and give a product formula for two symmetric Macdonald polynomials (Theorem~\ref{thm.PP-P}). This is a generalization of the classical formula for products of Weyl characters, where the generalized Littlewood-Richardson coefficents are rational functions in $q$ and $t$. In particular at $q=0$, Theorem~\ref{thm.PP-P} reduces to Schwer's product formula~\cite[Theorem 1.3]{Sc06} for Hall-Littlewood polynomials in terms of positively folded galleries, and at $q=t$, the formula reduces to the product formula of Littelmann~\cite{Li94} for Weyl characters phrased in terms of Littelmann paths. Section~\ref{sec.daha} of this paper introduces the basic definitions and properties of double affine Hecke algebras for reduced root systems. The alcove walk statistics needed in the later sections are addressed here. Section~\ref{sec.poly} discusses how alcove walks can be used to calculate the coefficients of products of monomials and intertwining operators in the double affine Hecke algebra. In Section~\ref{sec.lrrule}, we state and prove the main results: \noindent{\bf Theorem~\ref{thm.EP-E}} Let $E_\mu$ be the nonsymmetric Macdonald polynomial indexed by the weight $\mu$, and let $P_\lambda$ be the symmetric Macdonald polynomial indexed by the dominant weight $\lambda$. Let $m_{\varpi(h)}^{-1}$ be the alcove where the walk $h$ ends. Then $$E_\mu P_\lambda = \sum_{h} a_h(q,t) E_{\varpi(h)},$$ where the sum is over alcove walks of type determined by $\mu$ and contained in the dominant chamber, and the coefficients $a_h(q,t)$ are certain rational functions in $q$ and $t$. \noindent{\bf Theorem~\ref{thm.PP-P}} Let $P_\lambda$ be the symmetric Macdonald polynomial indexed by the dominant weight $\lambda$. Let $\ttwt(h)$ be the weight of the path $h$, and let $w_0$ be the longest element of the Weyl group. Then $$P_\mu P_\lambda = \sum_{h} c_h(q,t) P_{-w_0\ttwt(h)},$$ where the sum is over alcove walks of type determined by $\mu$ and contained in the dominant chamber, and the coefficients $c_h(q,t)$ are rational functions in $q$ and $t$. This section concludes by explaining how special cases of Theorem~\ref{thm.PP-P} relate to Macdonald's Pieri formula for symmetric Macdonald polynomials in terms of tableaux, and also explains the connection to Schwer's formula for Hall-Littlewood polynomials in terms of positively folded galleries. The final Section~\ref{sec.eg} contains many examples and illustrations. A number of calculations can be made completely explicit in the case of the reduced rank-one root system of type $A_1$. {\bf Acknowledgements.} This research was supported by the National Sciences and Engineering Research Council of Canada. The author would like to thank A. Ram for his guidance and insight, and also J. Haglund and C. Lenart for helpful conversations. \section{Alcoves, Weyl groups, and double affine Hecke algebras}\label{sec.daha} \subsection{Root systems and Weyl groups} Let $(\fh_\bbZ^*, R, \fh_\bbZ, R^\vee)$ be a reduced root datum with a pairing $$\langle \cdot , \cdot \rangle : \fh_\bbZ^* \times \fh_\bbZ \rightarrow \bbZ.$$ That is, $\fh_\bbZ^*$ and $\fh_\bbZ$ are lattices of finite rank, containing the finite subsets $R$ and $R^\vee$ respectively, and there is a bijection $R \rightarrow R^\vee: \alpha \mapsto \alpha^\vee$ such that $\langle \alpha, \alpha^\vee \rangle =2$. Let $\alpha_1, \ldots, \alpha_n \in R$ be the simple roots, and $\alpha_1^\vee, \ldots, \alpha_n^\vee \in R^\vee$ be the simple coroots. The fundamental weights $\{\omega_1,\ldots, \omega_n\}$ are defined by $\langle \omega_i, \alpha_j^\vee \rangle = \delta_{ij}$, and the fundamental coweights $\{\omega_1^\vee, \ldots, \omega_n^\vee \}$ are defined by $\langle \alpha_i, \omega_j^\vee \rangle = \delta_{ij}$. Let $$Q = \sum_{i=1}^n \bbZ \alpha_i,\quad Q^\vee = \sum_{i=1}^n \bbZ \alpha_i^\vee,\quad P = \sum_{i=1}^n \bbZ \omega_i,\quad P^\vee = \sum_{i=1}^n \bbZ \omega_i^\vee$$ be the root, coroot, weight, and coweight lattice respectively. Then $Q \subseteq \fh_\bbZ^* \subseteq P$ and $Q^\vee \subseteq \fh_\bbZ \subseteq P^\vee$ as lattices. Let $\fh_\bbR^* = \fh_\bbZ^* \otimes \bbR$ and $\fh_\bbR = \fh_\bbZ \otimes \bbR$. For $\alpha \in R$, the map $$s_\alpha : \fh_\bbR^* \rightarrow \fh_\bbR^*:x\mapsto x-\langle x, \alpha^\vee\rangle \alpha$$ acts on the lattice $\fh_\bbZ^*$ and is a reflection in the hyperplane $$\ttH_{\alpha^\vee} = \{ x\in \fh_\bbR^* \mid \langle x, \alpha^\vee \rangle =0 \},$$ and sends $\alpha$ to $-\alpha$. For $1\leq i \leq n$, let $$s_i = s_{\alpha_i}.$$ The {\em Weyl group} $W_0$ is generated by $s_1, \ldots, s_n$ subject to the relations $$s_i^2 =1, \qquad \textrm{and}\qquad s_is_js_i \cdots = s_js_is_j \cdots \quad \textrm{$(m_{ij}$ factors each side),}$$ where $\pi/m_{ij}$ is the angle between $\ttH_{\alpha_i^\vee}$ and $\ttH_{\alpha_j^\vee}$. See~\cite[p.69]{K01}. \subsubsection{Double affine Weyl groups} Let $e$ be the smallest positive integer which satisfies $\langle \fh_\bbZ^*, \fh_\bbZ \rangle \subseteq \hbox{$\frac{1}{e}$}\bbZ$. Let $X=\{x^\mu \mid \mu \in \fh_\bbZ^*\}$ and $Y=\{y^{\lambda^\vee} \mid \lambda^\vee \in \fh_\bbZ\}$ be abelian groups isomorphic to $\fh_\bbZ^*$ and $\fh_\bbZ$ respectively, with multiplication \begin{equation}\label{eqn.xyabeliangroups} x^\mu x^\lambda = x^{\mu+\lambda}, \quad \textrm{and} \quad y^{\lambda^\vee}y^{\mu^\vee} = y^{\lambda^\vee+\mu^\vee}. \end{equation} The {\em double affine Weyl group} $\widetilde W$ is $$\left\{q^k x^\mu w y^{\lambda^\vee} \mid k \in \hbox{$\frac{1}{e}$}\bbZ, \mu \in \fh_\bbZ^*, w\in W_0, \lambda^\vee \in \fh_\bbZ\right\},$$ subject to the relations \eqref{eqn.xyabeliangroups} and $$wx^\mu = x^{w\mu} w, \quad wy^{\lambda^\vee} = y^{w\lambda^\vee}w,\quad x^\mu y^{\lambda^\vee} = q^{\langle \mu, \lambda^\vee \rangle} y^{\lambda^\vee}x^\mu, \quad q^{1/e}\in Z(\widetilde W).$$ See~\cite[Corollary 4.6]{H06}. The {\em extended affine Weyl groups} \begin{equation}\label{eqn.awg1} W = \{w y^{\lambda^\vee}\mid w\in W_0, \lambda^\vee\in \fh_\bbZ \} = W_0 \ltimes Y, \end{equation} \begin{equation}\label{eqn.awg2} W^\vee = \{x^\mu w \mid \mu \in \fh_\bbZ^*, w\in W_0\} = X \rtimes W_0 \end{equation} are subgroups of $\widetilde W$, and $W$ acts by conjugation on $\{q^kx^\mu \mid k\in\hbox{$\frac{1}{e}$}\bbZ, \mu \in \fh_\bbZ^* \}$. Define \begin{equation} x^{\mu + k\delta} = q^kx^\mu \quad\textrm{and}\quad y^{\lambda^\vee + kd} = q^{-k}y^{\lambda^\vee}. \end{equation} Then $W$ acts on the lattice $\fh_\bbZ^*\oplus \bbZ\delta$, where for $w\in W$ and $\nu = \mu+k\delta\in \fh_\bbZ^*\oplus \bbZ\delta$, $w\nu$ is defined by \begin{equation}\label{eqn.conjaction} x^{w\nu} = wx^\nu w^{-1} \textrm{ in }\widetilde W. \end{equation} For $\alpha \in R$ and $j\in \bbN$, the map $$x^{j\alpha}s_\alpha:\fh_\bbR^* \rightarrow \fh_\bbR^* : x \mapsto s_\alpha x + j\alpha$$ acts on the lattice $\fh_\bbZ^*$ and is a reflection in the affine hyperplane $$\ttH_{-\alpha^\vee+jd} = \{x\in \fh_\bbR^* \mid \langle x, \alpha^\vee \rangle = j\}.$$ Let $\varphi^\vee \in R^\vee$ be the maximal coroot, and $\varphi\in R$ be the maximal (short) root. Define $$\alpha_0 = -\varphi + \delta,\quad \alpha_0^\vee= -\varphi^\vee + d,\quad s_0 = y^{\varphi^\vee}s_\varphi \in W, \quad \textrm{and}\quad s_0^\vee = x^{\varphi}s_\varphi \in W^\vee.$$ Then $s_0^\vee$ is a reflection in the hyperplane $\ttH_{\alpha_0^\vee} = \ttH_{-\varphi^\vee+d}$. The affine Weyl group $W_a = W_0 \ltimes Q^\vee$ is generated by $s_0, s_1,\ldots, s_n$ subject to the relations $$s_i^2 =1, \qquad \textrm{and}\qquad s_is_js_i \cdots = s_js_is_j \cdots \quad \textrm{$(m_{ij}$ factors each side),}$$ where $\pi/m_{ij}$ is the angle between $\ttH_{\alpha_i^\vee}$ and $\ttH_{\alpha_j^\vee}$. See~\cite[p.123]{K01}. The extended affine Weyl group $W$ has an alternate presentation~\cite[p.132]{K01} $$W = W_a \rtimes \Pi,$$ where $\Pi \cong \fh_\bbZ/Q^\vee$. The dual version of the above statements for $W$ holds for $W^\vee$ as well. That is, $$W^\vee = Q \rtimes W_0 = \Pi^\vee \ltimes W_a^\vee,$$ $\Pi^\vee \cong \fh_\bbZ^*/Q$, and $W_a^\vee$ is the group generated by $s_0^\vee, s_1, \ldots, s_n$. For notational convenience, we sometimes write $s_i^\vee = s_i$ for $i=1,\ldots, n$. \subsection{The alcove picture}\label{sec.alcovepicture} See for example, a picture for the $\fsl_2$ root system in Section~\ref{sec.eg}. Denote the positive roots and coroots by $R_+$ and $R_+^\vee$. The {\em positive affine coroots} are $$S_+^\vee = \left\{\alpha^\vee+jd\mid \alpha^\vee\in R_+^\vee, j \in \bbZ_{\geq0} \right\} \cup \left\{-\alpha^\vee+jd\mid \alpha^\vee\in R_+^\vee, j \in \bbZ_{\geq1} \right\}.$$ The {\em chambers} of $W_0$ are the connected components of $\fh_\bbR^* \backslash \cup_{\alpha \in R_+}\ttH_{\alpha^\vee}$, and the {\em alcoves} of $W_a^\vee$ are the connected components of $\fh_\bbR^* \backslash \cup_{a\in S_+} \ttH_{a^\vee}$. The {\em fundamental chamber} or {\em dominant chamber} is the region $$\ttC =\left \{x\in \fh_\bbR^* \mid 0 < \langle x, \alpha^\vee \rangle \textrm{ for } \alpha \in R_+\right\} = \bigcap_{i=1}^n \{x\in \fh_\bbR^* \mid 0<\langle x, \alpha_i^\vee \rangle\},$$ whose {\em walls} (the hyperplanes which have nonempty intersection with the closure of $\ttC$) are~ $\ttH_{\alpha_1^\vee}, \ldots, \ttH_{\alpha_n^\vee}$. The {\em fundamental alcove} is the region $$\ttA = \{x\in \fh_\bbR^* \mid 0 < \langle x, \alpha^\vee \rangle <1 \textrm{ for } \alpha \in R^+\} = \ttC \cap \{x\in \fh_\bbR^* \mid \langle x, \varphi^\vee \rangle <1\},$$ and its walls are the hyperplanes $\ttH_{\alpha_0^\vee}, \ldots, \ttH_{\alpha_n^\vee}$. By Proposition 4-6 and Proposition 11-5~\cite{K01}, $W_0$ acts freely transitively on the chambers, and $W_a^\vee$ acts freely transitively on the alcoves so that there is a bijection \begin{eqnarray*} W_a^\vee &\longleftrightarrow & \{\textrm{alcoves}\} \\ w & \leftrightarrow & w\ttA. \end{eqnarray*} In the above correspondence, the elements of $\Pi^\vee \subseteq W^\vee = \Pi^\vee \rtimes W_a^\vee$ fix the fundamental alcove $\ttA$. Since $|P/Q| = \det[\langle \alpha_i, \alpha_j^\vee\rangle]_{1\leq i, j \leq n}$ is finite, then $\Pi^\vee\cong \fh_\bbZ^*/Q \subseteq P/Q$ is a finite abelian group. The extended affine Weyl group $W^\vee$ acts freely transitively on $|\Pi^\vee|$ copies (sheets) of alcoves so that there is a bijection \begin{eqnarray*} W^\vee &\longleftrightarrow & \{\textrm{alcoves}\} \times |\Pi^\vee|\\ w & \leftrightarrow & w\ttA, \end{eqnarray*} where elements in $W_a^\vee$ permute alcoves in the base sheet, and elements $\pi_j^\vee \in \Pi^\vee$ send the fundamental alcove to the copy of the fundamental alcove on the $j$th sheet. This correspondence will be used frequently, and we will often use the shorthand $w=w\ttA$. The {\em periodic orientation} is the orientation of the hyperplanes $\left\{\ttH_{a^\vee} \mid a^\vee \in S_+^\vee\right\}$ such that \begin{enumerate} \item $\ttA$ is on the positive side of $\ttH_{\alpha^\vee}$ for $\alpha^\vee \in R_+^\vee$, \item $\ttH_{\alpha^\vee+jd}$ and $\ttH_{\alpha^\vee}$ have parallel orientations. \end{enumerate} The figure in Section~\ref{sec.eg} illustrates the alcove picture of the extended affine Weyl group $W^\vee$ for the $\fsl_2$ root system, showing the periodic orientation of the hyperplanes. \subsubsection{The length function} Given $w\in W^\vee$ with a reduced expression $w = \pi_j^\vee s_{i_1}^\vee \cdots s_{i_r}^\vee$, the set of positive coroots \begin{equation}\label{calLw} \calL(w) = \left\{\pi_j^\vee\alpha_{i_1}^\vee,\quad \pi_j^\vee s_{i_1}^\vee \alpha_{i_2}^\vee,\quad \pi_j^\vee s_{i_1}^\vee s_{i_2}^\vee \alpha_{i_3}^\vee, \quad \ldots,\quad \pi_j^\vee s_{i_1}^\vee \cdots s_{i_{r-1}}^\vee \alpha_{i_r}^\vee \right\} \end{equation} indexes the hyperplanes that separate the fundamental alcove $\ttA$ and the alcove $w\ttA$. In Macdonald's notation~\cite[(2.2.1), (2.2.9)]{M03}, $$\calL(w) = \{b^\vee \in S_+^\vee \mid w^{-1}b^\vee \in S_-^\vee\} =S(w^{-1}).$$ The {\em length} of $w$ is $$\ell(w) = |\calL(w)|,$$ the number of hyperplanes that separate $\ttA$ and $w\ttA$. In particular, $\ell(\pi^\vee) =0 $ for all $\pi^\vee \in \Pi^\vee$. More generally, for $v, w\in W^\vee$, the set of positive coroots \begin{equation}\label{calLvw} \calL(v,w) = \left(\calL(v) \cup \calL(w)\right) \backslash (\calL(v) \cap \calL(w)) \end{equation} indexes the set of hyperplanes that separate the alcoves $v\ttA$ and $w\ttA$. If $v\leq w$, we may write $\calL(v,w) = v\calL(1,v^{-1}w)$, where $\calL(1,w) = \calL(w)$. \subsubsection{The group $\Pi^\vee$} Let $w_0\in W_0$ be the unique longest element in the finite Weyl group $W_0$. For $\mu \in \fh_\bbZ^*$, let $\mu_+$ denote the unique dominant weight in the orbit, so that $\mu_- = w_0\mu_+$ is the unique antidominant weight. Let $v_\mu$ be the shortest element of $W_0$ such that $v_\mu\mu = \mu_-$. Define \begin{equation}\label{eqn.mlcr} m_\mu = x^\mu v_\mu^{-1} \in W^\vee. \end{equation} By~\cite[(2.4.5)]{M03}, $m_\mu$ is the unique shortest element in the coset $x^\mu W_0$. The {\em minuscule weights} are the fundamental weights $\omega_j$ that satisfy $\langle \omega_j, \alpha^\vee \rangle \leq 1$ for $\alpha^\vee \in R_+^\vee$. In other words, these are the fundamental weights which are contained in the closure of the fundamental alcove. Let $$J = \{j \mid \omega_j \in \fh_\bbZ^* \textrm{ is a minuscule weight} \} \cup \{0\}.$$ Let $\pi_0^\vee = 1$, and for $j \in J\backslash \{0\}$, let \begin{equation} \pi_j^\vee = m_{\omega_j} = x^{\omega_j}v_{\omega_j}^{-1}. \end{equation} By~\cite[(2.5.4)]{M03}, the subgroup of length zero elements in $W^\vee$ is $$\Pi^\vee = \left\{\pi_j^\vee \mid j\in J\right\}.$$ \subsection{Braid groups and Hecke algebras} \subsubsection{Double affine braid groups} The relevant facts about braid groups from~\cite{M03} are stated here in our notation. Let $\left\{q^k X^\mu \mid k\in\hbox{$\frac{1}{e}$}\bbZ, \mu \in \fh_\bbZ^*\right\}$ be the multiplicative group isomorphic to $\fh_\bbZ^* \oplus \bbZ\delta$, and write $X^{\mu+k\delta} = q^kX^\mu$. By equation~\eqref{eqn.conjaction}, the conjugation action of $W = \langle s_0, \ldots, s_n \rangle \rtimes \Pi$ on $\left\{q^kX^\mu \mid k\in \hbox{$\frac{1}{e}$}\bbZ, \mu\in\fh_\bbZ^*\right\}$ is given by $$w(\mu+k\delta)= w\mu + k\delta, \quad y^{\lambda^\vee}(\mu+k\delta)=\mu -\langle \mu, \lambda^\vee \rangle \delta +k\delta,\quad \textrm{ for } w\in W_0, \lambda^\vee \in \fh_\bbZ.$$ The {\em double affine braid group} $\widetilde B$ is generated by $\left\{q^k X^\mu \mid k\in\bbZ, \mu \in \fh_\bbZ^*\right\}$, $T_0, T_1, \ldots, T_n$, and $\Pi$, subject to the relations \begin{enumerate} \item $T_iT_jT_i \cdots = T_jT_iT_j \cdots$ \quad ($m_{ij}$ factors each side), \item $T_i X^\mu = X^{\mu}T_i$ \quad if $\langle \mu,\alpha_i^\vee\rangle=0$ for $0\leq i \leq n$, \item $T_i X^\mu T_i = X^{s_i\mu}$ \quad if $\langle \mu,\alpha_i^\vee\rangle=1$ for $0\leq i \leq n$, \item $\pi T_i \pi^{-1} = T_j$ \quad if $\pi\alpha_i = \alpha_j$ for $\pi \in \Pi$, \item $\pi X^\mu \pi^{-1} = X^{\pi\mu}$ \quad for $\pi \in \Pi$, \end{enumerate} where $\langle \mu, \alpha_0^\vee\rangle = \langle \mu, -\varphi^\vee\rangle$. Since $W= W_0 \rtimes Y$ fixes $\delta$, then $q^{1/e}$ is a central element of $\widetilde B$. See~\cite[Sec 3.4]{M03}. For $w\in W$ with a reduced expression $w=\pi_j s_{i_1}\cdots s_{i_r}$ where $\pi\in \Pi$ and $s_{i_j}\in \langle s_0,\ldots, s_n\rangle$, define \begin{equation} T_w = \pi_j T_{i_1}\cdots T_{i_r}. \end{equation} By~\cite[(3.1.1)]{M03}, the element $T_w$ is independent of the choice of a reduced word for $w$. Identify the reduced expression $w=\pi_j s_{i_1}\cdots s_{i_r}$ with the minimal path $p$ from $\ttA$ to $w\ttA$ via the sequence of alcoves $\pi_j\ttA, \ \pi_js_{i_1}\ttA, \ \ldots,\ \pi_js_{i_1}\cdots s_{i_r}\ttA$, and define \begin{equation}\label{eqn.Y} Y^w = \pi_j T_{i_1}^{\epsilon_1}\cdots T_{i_r}^{\epsilon_r}, \quad \textrm{ where } \epsilon_k = \begin{cases} +1, & \textrm{if the $k$th step of $p$ is \beginpicture \setcoordinatesystem units <1cm,1cm> \setplotarea x from -0.8 to 0.8, y from -0.5 to 0.5 \put{$\scriptstyle{-}$}[b] at -0.3 0.25 \put{$\scriptstyle{+}$}[b] at 0.3 0.25 \plot 0 -0.4 0 0.5 / \arrow <5pt> [.2,.67] from -0.5 0 to 0.5 0 % \endpicture ,}\\ -1, & \textrm{if the $k$th step of $p$ is \beginpicture \setcoordinatesystem units <1cm,1cm> \setplotarea x from -0.8 to 0.8, y from -0.5 to 0.5 \put{$\scriptstyle{-}$}[b] at -0.3 0.25 \put{$\scriptstyle{+}$}[b] at 0.3 0.25 \plot 0 -0.4 0 0.5 / \arrow <5pt> [.2,.67] from 0.5 0 to -0.5 0 % \endpicture ,} \end{cases} \end{equation} with respect to the periodic orientation (section~\ref{sec.alcovepicture}) of the hyperplanes. For simplicity, write $Y^{y^{\lambda^\vee}} = Y^{\lambda^\vee}$. Then $Y^{\lambda^\vee}Y^{\mu^\vee} = Y^{\lambda^\vee+\mu^\vee}$ for $\lambda^\vee, \mu^\vee \in \fh_\bbZ,$ and $$Y=\{Y^{\lambda^\vee} \mid \lambda^\vee \in \fh_\bbZ\}$$ is a multiplicative group isomorphic to $\fh_\bbZ$. The elements $Y^{\lambda^\vee}$ satisfy the relations $$\begin{array}{ll} T_i^{-1} Y^{\lambda^\vee} = Y^{\lambda^\vee}T_i^{-1}, & \hbox{if $\langle \alpha_i, \lambda^\vee\rangle=0$ for $1\leq i \leq n$,}\\ T_i^{-1} Y^{\lambda^\vee} T_i^{-1} = Y^{s_i\lambda^\vee}, & \hbox{if $\langle \alpha_i, \lambda^\vee\rangle=1$ for $1\leq i \leq n$.} \end{array} $$ See~\cite[(3.2.4)]{M03} for details. Define \begin{equation} T_0^\vee = (X^\varphi T_{s_\varphi})^{-1} \quad \textrm{and}\quad Y^{-\alpha_0^\vee} = qY^{\varphi^\vee}. \end{equation} For notational convenience, we sometimes write $T_i^\vee=T_i$ for $i=1,\ldots, n$. Then, identifying the reduced expression $z = s_{i_r}^\vee \cdots s_{i_1}^\vee (\pi_j^\vee)^{-1} \in W^\vee$ with the minimal path $b$ from $z\ttA$ to $\ttA$ via the sequence of alcoves $$z\ttA,\quad z\pi_j^\vee s_{i_1}^\vee \ttA,\quad z\pi_j^\vee s_{i_1}^\vee s_{i_2}^\vee \ttA,\quad \ldots,\quad (z\pi_j^\vee s_{i_1}^\vee \cdots s_{i_r}^\vee) \ttA= \ttA,$$ let \begin{equation}\label{eqn.epsilonk} X^z = (T_{i_r}^\vee)^{\epsilon_r} \cdots (T_{i_1}^\vee)^{\epsilon_1} (\pi_j^\vee)^{-1}, \quad \hbox{where } \epsilon_k = \begin{cases} +1, &\hbox{if the $k$th step of $b$ is \beginpicture \setcoordinatesystem units <1cm,1cm> \setplotarea x from -0.7 to 0.7, y from -0.5 to 0.5 \put{$\scriptstyle{-}$}[b] at -0.3 0.25 \put{$\scriptstyle{+}$}[b] at 0.3 0.25 \plot 0 -0.4 0 0.5 / \arrow <5pt> [.2,.67] from -0.5 0 to 0.5 0 % \endpicture ,}\\ -1, &\hbox{if the $k$th step of $b$ is \beginpicture \setcoordinatesystem units <1cm,1cm> \setplotarea x from -0.7 to 0.7, y from -0.5 to 0.5 \put{$\scriptstyle{-}$}[b] at -0.3 0.25 \put{$\scriptstyle{+}$}[b] at 0.3 0.25 \plot 0 -0.4 0 0.5 / \arrow <5pt> [.2,.67] from 0.5 0 to -0.5 0 % \endpicture ,} \end{cases} \end{equation} with respect to the periodic orientation of the hyperplanes (section~\ref{sec.alcovepicture}). Note that for $\mu\in\fh_\bbZ^*\subseteq W^\vee$, $X^\mu=X^{x^{\mu}}$. \subsubsection{Double affine Hecke algebras} Let $\bbK$ be a field. Fix $t_0, t_1, \ldots, t_n \in \bbK$ such that $t_i =t_j$ if $s_i$ and $s_j$ are conjugate in $W$. For $\alpha\in R$ and $k\in \frac{1}{e}\bbZ$, define $t_{\alpha+k\delta} = t_i$ if $\alpha = w\alpha_i$ for some $w\in W$. The {\em double affine Hecke algebra} $\widetilde H$ is the quotient of the group algebra $\bbK\widetilde B$ of the double affine braid group by the relations $$T_i^2 = (t_i^{1/2}-t_i^{-1/2})T_i +1, \quad \textrm{for } 0 \leq i \leq n.$$ The {\em intertwining operators} (also called {\em creation operators} in~\cite[Sec 5.10]{M03}) are \begin{align*} \tau_i^\vee &= T_i^\vee + \frac{t_i^{-1/2}-t_i^{1/2}}{1-Y^{-\alpha_i^\vee}} = (T_i^\vee)^{-1} + \frac{(t_i^{-1/2}-t_i^{1/2})Y^{-\alpha_i^\vee}}{1-Y^{-\alpha_i^\vee}}, \qquad\textrm{for } 0\leq i \leq n,\\ \pi_j^\vee &= X^{\omega_j}T_{v_{\omega_j}^{-1}}, \qquad\textrm{for } j\in J. \end{align*} For $w\in W^\vee$ with a reduced expression $w = \pi_j^\vee s_{i_1}^\vee \cdots s_{i_r}^\vee$, define \begin{equation} \tau_w^\vee = \pi_j^\vee \tau_{i_1}^\vee \cdots \tau_{i_r}^\vee. \end{equation} By~\cite[(5.10.13)]{M03}, $\tau_w^\vee$ is independent of the choice of a reduced word for $w$. Moreover, the intertwiners satisfy \begin{equation}\label{eqn.intertwiningrelns} \tau_w^\vee Y^{\lambda^\vee} = Y^{w\lambda^\vee}\tau_w^\vee, \qquad \textrm{for } w\in W^\vee, \lambda^\vee \in \fh_\bbZ. \end{equation} For $0\leq i \leq n$ and $w\in W^\vee$, \begin{equation} \label{eqn.tauisquared} \tau_i^\vee \tau_w^\vee = \begin{cases} \tau_{s_i^\vee w}^\vee, & \textrm{if } s_i^\vee w > w,\\ (\tau_i^\vee)^2 \tau_{s_i^\vee w}^\vee, & \textrm{if } s_i^\vee w < w, \end{cases} \hbox{ where } (\tau_i^\vee)^2 = \left(\frac{1-t_i^{-1}Y^{-\alpha_i^\vee}}{1-Y^{-\alpha_i^\vee}}\right) \!\!\!\left(\frac{1-t_iY^{-\alpha_i^\vee}}{1-Y^{-\alpha_i^\vee}}\right). \end{equation} \subsubsection{Polynomial representation} The {\em affine Hecke algebra} $H$ is the subalgebra of the double affine Hecke algebra $\widetilde H$ generated by $T_0, \ldots, T_n$ and $\Pi$. A basis for $H$ is $\{T_w Y^{\lambda^\vee} \mid w\in W_0, \lambda^\vee \in \fh_\bbZ\}$. Let $\bbK\mathbf{1}$ be the $H$-module given by $$\pi \mathbf{1} = \mathbf{1}, \quad T_i\mathbf{1} = t_i^{1/2}\mathbf{1},\quad \textrm{for $\pi\in \Pi$ and $0\leq i \leq n$.}$$ The {\em polynomial representation} of $\widetilde H$ is $$\bbK[X]\mathbf{1} = \mathrm{Ind}_H^{\widetilde H}\mathbf{1},$$ with basis $\{X^\mu\mathbf{1} \mid \mu\in \fh_\bbZ^*\}$. For $0\leq i \leq n$, the operators $T_i$ act on $\bbK[X]\mathbf{1}$ by \begin{equation} T_i X^\mu \mathbf{1} = t_i^{1/2} X^{s_i\mu}\mathbf{1} + \left(t_i^{1/2}-t_i^{-1/2}\right) \frac{X^\mu - X^{s_i\mu}}{1-X^{\alpha_i}}\mathbf{1}. \end{equation} For an affine coroot $\beta^\vee + jd$, define {\em shift} and {\em height} \begin{equation}\label{eqn.shiftheight} q^{\shf(\beta^\vee + jd)} = q^{-j}, \qquad \textrm{and} \qquad t^{\hgt(\beta^\vee + jd)} = \prod_{\alpha \in R_+} t_\alpha^{\frac12\langle \alpha, \beta^\vee \rangle}, \end{equation} so that \begin{equation} Y^{\beta^\vee+jd}\mathbf{1} = q^{-j}Y^{\beta^\vee}\mathbf{1} = q^{-j} \prod_{\alpha \in R_+} t_{\alpha}^{\frac12 \langle \alpha, \beta^\vee \rangle}\mathbf{1} = q^{\shf(\beta^\vee + jd)}t^{\hgt(\beta^\vee + jd)}\mathbf{1}. \end{equation} If $t_\alpha=t$ for all $\alpha \in R_+$, then $t^{\hgt(\beta^\vee + jd)} = t^{\langle \rho, \beta^\vee \rangle}$ for $\rho = \frac12\sum_{\alpha\in R_+} \alpha^\vee$. \subsection{Alcove walks}\label{sec.alcovewalks} Fix a reduced factorization of $w=\pi_j^\vee s_{i_1}^\vee\cdots s_{i_r}^\vee \in W^\vee$. An {\em alcove walk} of type $\vec{w} = (i_1, \ldots, i_r)$ beginning at $z$ is a sequence of steps in the alcove picture, where for $i=0,\ldots, n$, a step of type $i$ is one of the following: \begin{equation} \beginpicture \setcoordinatesystem units <0.75cm,0.75cm> \setplotarea x from -1 to 1, y from -1 to 1 \plot 0 -0.9 0 1 / \setplotsymbol(.) \arrow <5pt> [.2,.67] from -0.7 0.3 to 0.7 0.3 \put{\footnotesize $v$} at -0.7 -0.5 \put{\footnotesize $vs_i^\vee$} at 0.7 -0.5 \put{$i$-crossing,} at 0 -1.4 \endpicture \qquad\qquad\qquad \beginpicture \setcoordinatesystem units <0.75cm,0.75cm> \setplotarea x from -1 to 1, y from -1 to 1 \plot 0 -0.9 0 1 / \setplotsymbol(.) \plot -0.75 0.15 0 0.15 / \plot -0.02 0.15 -0.02 0.35 / \arrow <5pt> [.2,.67] from 0 0.35 to -0.75 0.35 \put{\footnotesize $v$} at -0.7 -0.5 \put{\footnotesize $vs_i^\vee$} at 0.7 -0.5 \put{$i$-folding.} at 0 -1.4 \endpicture \end{equation} In addition, a `step' of type $\vec{\pi}^\vee$ for $\pi^\vee\in \Pi^\vee$ can be thought of as a change of sheets from the alcove $v$ to the alcove $v\pi_j^\vee$. See the figure in section~\ref{sec.eg}. Let $\Gamma(\vec{w},z)$ be the set of alcove walks of type $\vec{w}$ beginning in $z$. There are $2^r$ walks in $\Gamma(\vec{w},z)$, since each step can be either a crossing or a folding. For a walk $h\in\Gamma(\vec{w},z)$, let \begin{equation}\label{eqn.pkvee} h_k^\vee \hbox{ be the positive coroot such that $\ttH_{h_k^\vee}$ separates the alcoves $v\ttA$ and $vs_{i_k}^\vee\ttA$,} \end{equation} in which $v\ttA$ is the alcove where $k$th step of $h$ begins. We call $\ttH_{h_k^\vee}$ the $k$th {\em active hyperplane} of the walk $h$. For $k=1,\ldots, r$, let \begin{equation}\label{eqn.bkvee} b_k^\vee = s_{i_r}^\vee s_{i_{r-1}}^\vee \cdots s_{i_{k+1}}^\vee\alpha_{i_k}^\vee, \quad\hbox{and}\quad t_{b_k^\vee} = t_{i_k}, \end{equation} so that $b_r^\vee, \ldots, b_1^\vee$ is the sequence of labels of hyperplanes crossed by the walk of type $\vec{w}^{-1}=(i_r,\ldots, i_1)$ beginning in $1$. See example~\ref{eg.1} for an illustration. {\em Positive} and {\em negative} steps are defined with respect to the periodic orientation of the hyperplanes as follows: \begin{equation}\label{eqn.posnegsteps} \beginpicture \setcoordinatesystem units <0.75cm,0.75cm> \setplotarea x from -1 to 1, y from -1 to 1.2 \plot 0 -0.7 0 1 / \setplotsymbol(.) \arrow <5pt> [.2,.67] from -0.7 0.3 to 0.7 0.3 \put{$\scriptstyle{+}$} at 0.4 0.9 \put{$\scriptstyle{-}$} at -0.4 0.9 \put{\scriptsize $v$}[t] at -0.7 -0.5 \put{\scriptsize $vs_i$}[t] at 0.7 -0.5 \put{positive crossing,} at 0 -1.5 \endpicture \qquad \beginpicture \setcoordinatesystem units <0.75cm,0.75cm> \setplotarea x from -1 to 1, y from -1 to 1.2 \plot 0 -0.7 0 1 / \setplotsymbol(.) \arrow <5pt> [.2,.67] from 0.7 0.3 to -0.7 0.3 \put{$\scriptstyle{+}$} at 0.4 0.9 \put{$\scriptstyle{-}$} at -0.4 0.9 \put{\scriptsize $vs_i$}[t] at -0.7 -0.5 \put{\scriptsize $v$}[t] at 0.7 -0.5 \put{negative crossing,} at 0 -1.5 \endpicture \qquad \beginpicture \setcoordinatesystem units <0.75cm,0.75cm> \setplotarea x from -1 to 1, y from -1 to 1.2 \plot 0 -0.7 0 1 / \setplotsymbol(.) \plot 0.75 0.15 0.05 0.15 / \plot 0.02 0.15 0.02 0.35 / \arrow <5pt> [.2,.67] from 0.05 0.35 to 0.75 0.35 \put{$\scriptstyle{+}$} at 0.4 0.9 \put{$\scriptstyle{-}$} at -0.4 0.9 \put{\scriptsize $vs_i$}[t] at -0.7 -0.5 \put{\scriptsize $v$}[t] at 0.7 -0.5 \put{positive folding,} at 0 -1.5 \endpicture \qquad \beginpicture \setcoordinatesystem units <0.75cm,0.75cm> \setplotarea x from -1 to 1, y from -1 to 1.2 \plot 0 -0.7 0 1 / \setplotsymbol(.) \plot -0.75 0.15 -0.05 0.15 / \plot -0.02 0.15 -0.02 0.35 / \arrow <5pt> [.2,.67] from -0.05 0.35 to -0.75 0.35 \put{$\scriptstyle{+}$} at 0.4 0.9 \put{$\scriptstyle{-}$} at -0.4 0.9 \put{\scriptsize $v$}[t] at -0.7 -0.5 \put{\scriptsize $vs_i$}[t] at 0.7 -0.5 \put{negative folding.} at 0 -1.5 \endpicture \end{equation} Moreover, \begin{equation}\label{eqn.longshortsteps} \hbox{a step is an {\em ascent} if $vs_i > v$, and it is a {\em descent} if $vs_i < v$ in the Bruhat order.} \end{equation} \begin{example}\label{eg.1}{\em See section~\ref{sec.eg} for more details on the alcove picture of type $\fsl_2$. Let $w= (s_1 s_0^\vee)^4 = x^{-8\omega}$. The following is an alcove walk $h \in \Gamma(\vec{w}, 1)$. $$\beginpicture \setcoordinatesystem units <1.5cm,1cm> \setplotarea x from -3 to 6, y from -1.5 to 1.5 \put{$\bullet$} at 0 0 {\small \put{$\ttH_{-\alpha^\vee+2d}$}[t] at 2 -0.6 \put{$\ttH_{\alpha^\vee}$}[t] at 0 -0.6 \put{$\ttH_{\alpha^\vee+2d}$}[t] at -2 -0.6 \put{$\ttH_{-\alpha^\vee+d}$}[t] at 1 -0.6 \put{$\ttH_{-\alpha^\vee+3d}$}[t] at 3 -0.6 \put{$\ttH_{\alpha^\vee+d}$}[t] at -1 -0.6 \put{$\ttH_{-\alpha^\vee+4d}$}[t] at 4 -0.6 \put{$\ttH_{-\alpha^\vee+5d}$}[t] at 5 -0.6 }\plot -2.5 0 5.5 0 / \plot -2 1.2 -2 -0.3 / \plot -1 1.2 -1 -0.3 / \plot 0 1.2 0 -0.3 / \plot 1 1.2 1 -0.3 / \plot 2 1.2 2 -0.3 / \plot 3 1.2 3 -0.3 / \plot 4 1.2 4 -0.3 / \plot 5 1.2 5 -0.3 / \put{$\scriptstyle{+}$} at -1.8 1.1 \put{$\scriptstyle{-}$} at -2.2 1.1 \put{$\scriptstyle{+}$} at -0.8 1.1 \put{$\scriptstyle{-}$} at -1.2 1.1 \put{$\scriptstyle{+}$} at 0.2 1.1 \put{$\scriptstyle{-}$} at -0.2 1.1 \put{$\scriptstyle{+}$} at 1.2 1.1 \put{$\scriptstyle{-}$} at 0.8 1.1 \put{$\scriptstyle{+}$} at 2.2 1.1 \put{$\scriptstyle{-}$} at 1.8 1.1 \put{$\scriptstyle{+}$} at 3.2 1.1 \put{$\scriptstyle{-}$} at 2.8 1.1 \put{$\scriptstyle{+}$} at 4.2 1.1 \put{$\scriptstyle{-}$} at 3.8 1.1 \put{$\scriptstyle{+}$} at 5.2 1.1 \put{$\scriptstyle{-}$} at 4.8 1.1 \put{$1$} at 0.5 -0.3 \put{$x^{4\omega}s_1$} at 3.5 -0.3 \setplotsymbol(.) \arrow <6pt> [.2,.67] from 0.5 0.2 to -0.5 0.2 \plot -0.5 0.2 -1 0.2 / \plot -0.99 0.2 -0.99 0.4 / \arrow <6pt> [.2,.67] from -1 0.4 to -0.5 0.4 \arrow <6pt> [.2,.67] from -0.5 0.4 to 0.5 0.4 \arrow <6pt> [.2,.67] from 0.5 0.4 to 1.5 0.4 \arrow <6pt> [.2,.67] from 1.5 0.4 to 2.5 0.4 \plot 2.5 0.4 3 0.4 / \plot 2.99 0.4 2.99 0.6 / \arrow <6pt> [.2,.67] from 3 0.6 to 2.5 0.6 \plot 2.5 0.6 2 0.6 / \plot 2.01 0.6 2.01 0.8 / \arrow <6pt> [.2,.67] from 2 0.8 to 2.5 0.8 \arrow <6pt> [.2,.67] from 2.5 0.8 to 3.5 0.8 \endpicture $$ This walk of length eight has type $\vec{w}=(1,0,1,0,1,0,1,0)$. The coroots $h_k^\vee$ are \begin{align*} h_1^\vee, \ldots, h_8^\vee &= \alpha^\vee,\ \alpha^\vee+d,\ \alpha^\vee,\ -\alpha^\vee+d,\ -\alpha^\vee+2d,\ -\alpha^\vee+3d,\ -\alpha^\vee+2d,\ -\alpha^\vee +3d. \end{align*} For $k=1,\ldots, 8$, the coroots $b_k^\vee= s_{i_8}^\vee \cdots s_{i_{k+1}}^\vee \alpha_{i_k}^\vee = -\alpha^\vee+(9-k)d$ are the labels of hyperplanes crossed by the walk of type $\vec{w}^{-1}$ starting in $1$. $$\beginpicture \setcoordinatesystem units <1.2cm,1.2cm> \setplotarea x from -2 to 11, y from -1 to 0.8 \put{$\bullet$} at 0 0 \plot -1.5 0 10.5 0 / \plot -1 0.6 -1 -0.3 / \plot 0 0.6 0 -0.3 / \plot 1 0.6 1 -0.3 / \plot 2 0.6 2 -0.3 / \plot 3 0.6 3 -0.3 / \plot 4 0.6 4 -0.3 / \plot 5 0.6 5 -0.3 / \plot 6 0.6 6 -0.3 / \plot 7 0.6 7 -0.3 / \plot 8 0.6 8 -0.3 / \plot 9 0.6 9 -0.3 / \plot 10 0.6 10 -0.3 / \put{\tiny$\scriptstyle{+}$} at 0.2 0.6 \put{\tiny$\scriptstyle{-}$} at -0.2 0.6 \put{$w^{-1}$} at 8.5 0.75 \put{$1$} at 0.65 0.75 {\small \put{$\ttH_{b_8^\vee}$}[t] at 1 -0.4 \put{$\ttH_{b_7^\vee}$}[t] at 2 -0.4 \put{$\ttH_{b_6^\vee}$}[t] at 3 -0.4 \put{$\ttH_{b_5^\vee}$}[t] at 4 -0.4 \put{$\ttH_{b_4^\vee}$}[t] at 5 -0.4 \put{$\ttH_{b_3^\vee}$}[t] at 6 -0.4 \put{$\ttH_{b_2^\vee}$}[t] at 7 -0.4 \put{$\ttH_{b_1^\vee}$}[t] at 8 -0.4 } \setplotsymbol(.) \arrow <6pt> [.2,.67] from 0.5 0.2 to 1.5 0.2 \arrow <6pt> [.2,.67] from 1.5 0.2 to 2.5 0.2 \arrow <6pt> [.2,.67] from 2.5 0.2 to 3.5 0.2 \arrow <6pt> [.2,.67] from 3.5 0.2 to 4.5 0.2 \arrow <6pt> [.2,.67] from 4.5 0.2 to 5.5 0.2 \arrow <6pt> [.2,.67] from 5.5 0.2 to 6.5 0.2 \arrow <6pt> [.2,.67] from 6.5 0.2 to 7.5 0.2 \arrow <6pt> [.2,.67] from 7.5 0.2 to 8.5 0.2 \endpicture $$ \hfill$\diamond$ }\end{example} \section{Macdonald polynomials} \label{sec.poly} \subsection{Nonsymmetric Macdonald polynomials}\label{sec.nonsymmMac} Given a weight $\mu \in \fh_\bbZ^*$, let $m_\mu \in W^\vee$ be the shortest element in the coset $x^\mu W_0$. See~\eqref{eqn.mlcr}. For $w\in W^\vee$, define $t_w = t_{i_1}\cdots t_{i_r}$, if $w=s_{i_1}^\vee \cdots s_{i_r}^\vee \in W^\vee$ is a reduced expression. \begin{defn}{\em Let $m_\mu = \pi_j^\vee s_{i_1}^\vee\cdots s_{i_r}^\vee$ be a reduced expression. The {\em nonsymmetric Macdonald polynomial} $E_\mu\in \bbK[X]$ indexed by $\mu \in \fh_\bbZ^*$ is defined by \begin{equation}\label{eqn.E} E_\mu \mathbf{1} = \tau_{m_\mu}^\vee \mathbf{1} = \pi_j^\vee \tau_{i_1}^\vee \cdots \tau_{i_r}^\vee\mathbf{1}. \end{equation} }\end{defn} \begin{remark}{\em In equation~\eqref{eqn.E}, the coefficient of $X^\mu$ in $E_\mu$ is $t_{v_\mu^{-1}}^{1/2}$, where $x^\mu = m_\mu v_\mu$~\eqref{eqn.mlcr}. In the literature, $E_\mu$ is often normalized so that the coefficient of $X^\mu$ in $E_\mu$ is $1$. By~\eqref{eqn.intertwiningrelns}, the nonsymmetric Macdonald polynomials are eigenfunctions for the operators $\{Y^{\lambda^\vee} \mid\lambda^\vee \in \fh_\bbZ\}$, and the set $\{E_\mu\mathbf{1} \mid \mu\in \fh_\bbZ^*\}$ is a basis for the polynomial representation $\bbK[X]\mathbf{1}$. }\end{remark} The following result gives an expansion of a product of monomials $X^\nu$ and intertwining operators in terms of monomials in $\widetilde H$. This leads to an expression for $E_\mu$ in the monomial basis. For a walk $p\in \Gamma(\vec{w},z)$ of type $\vec{w}$ beginning in $z$, let \begin{equation} \begin{array}{rl} \phi(p)\!\!\!&= \{k\ |\ \hbox{the $k$th step of $p$ is a fold}\},\\ \phi_-(p)\!\!\!&= \{k\ |\ \hbox{the $k$th step of $p$ is a negative fold}\}, \end{array} \end{equation} \begin{equation}\label{endpoint} \begin{array}{l} \ttb(p) \in W^\vee \hbox{ be the alcove where $p$ begins,}\\ \tte(p) \in W^\vee \hbox{ be the alcove where $p$ ends.} \end{array} \end{equation} The {\em weight} $\ttwt(p) \in \fh_\bbZ^*$ and {\em final direction} $\ttd(p)\in W_0$ of a walk $p$ is defined by \begin{equation} X^{\tte(p)} = X^{\ttwt(p)}T_{\ttd(p)}. \end{equation} \begin{theorem}\label{thm.Xtau-X}{\ } \begin{enumerate} \item[(a)] \cite[Theorem 2.2]{RY10} Let $z, w\in W^\vee$, and fix a reduced expression $w = \pi_j^\vee s_{i_1}^\vee\cdots s_{i_r}^\vee$. Then $$ X^z \tau_w^\vee = \sum_{p \in \Gamma(\vec{w},z)} X^{\tte(p)} \left(\prod_{k\in \phi(p)}\frac{t_{b_k^\vee}^{-1/2}-t_{b_k^\vee}^{1/2}}{1-Y^{-b_k^\vee}}\right) \left( \prod_{k\in \phi_-(p)} Y^{-b_k^\vee} \right) ,$$ where $b_k^\vee = s_{i_r}^\vee\cdots s_{i_{k+1}}^\vee \alpha_{i_k}^\vee$ and $t_{b_k^\vee} = t_{i_k}$, see~\eqref{eqn.bkvee}.\\ \item[(b)] \cite[Theorem 3.1]{RY10} Let $\mu \in \fh_\bbZ^*$ and fix a reduced expression $m_\mu = \pi_j^\vee s_{i_1}^\vee \cdots s_{i_r}^\vee\in W^\vee$ for the minimal length representative of the coset $x^\mu W_0$. The nonsymmetric Macdonald polynomial $$E_\mu = \sum_{p\in \Gamma(\vec{m}_\mu)} \left(\prod_{k\in \phi(p)}\frac{t_{b_k^\vee}^{-1/2}-t_{b_k^\vee}^{1/2}}{1-q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)}}\right) \left( \prod_{k\in \phi_-(p)} q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)}\right) t_{\ttd(p)}^{1/2} X^{\ttwt(p)},$$ where the sum is over the set $\Gamma(\vec{m}_\mu) = \Gamma(\vec{m}_\mu, 1)$ of alcove walks of type $\vec{m}_\mu= (\pi_j^\vee, i_1,\ldots , i_r)$ beginning in the fundamental alcove. \end{enumerate} \end{theorem} \proof {\ } \begin{enumerate} \item[(a)] This is proved by using the definition of $\tau_i^\vee$ and the commutation relations~\eqref{eqn.intertwiningrelns}.\\ \item[(b)] Since $E_\mu\mathbf{1} = \tau_{m_\mu}^\vee\mathbf{1}$, then the result follows from (a) by setting $w=m_\mu$, $z=1$, using $$Y^{-b_k^\vee}\mathbf{1} = q^{\shf(-b_k^\vee)}t^{-\hgt(-b_k^\vee)}\mathbf{1}, \qquad X^{\tte(p)}\mathbf{1} = X^{\ttwt(p)}T_{\ttd(p)} \mathbf{1} = t_{\ttd(p)}^{1/2} X^{\ttwt(p)}\mathbf{1}.$$ \end{enumerate} \qed The next result gives an expansion of a product of monomials and intertwining operators in terms of intertwining operators in $\tilde H$. Recall the definition of $X^z$ from Equation~\eqref{eqn.epsilonk}: identify a reduced expression $z = s_{i_r}^\vee \cdots s_{i_1}^\vee (\pi_j^\vee)^{-1} \in W^\vee$ with the minimal walk $b$ from $z\ttA$ to $\ttA$ of type $\vec{z}^{-1}$ via the sequence of alcoves $$z\ttA,\quad z\pi_j^\vee s_{i_1}^\vee \ttA,\quad z\pi_j^\vee s_{i_1}^\vee s_{i_2}^\vee \ttA,\quad \ldots,\quad (z\pi_j^\vee s_{i_1}^\vee \cdots s_{i_r}^\vee) \ttA= \ttA, \qquad \hbox{so}$$ $$X^z = (T_{i_r}^\vee)^{\epsilon_r} \cdots (T_{i_1}^\vee)^{\epsilon_1} (\pi_j^\vee)^{-1}, \quad \hbox{where } \epsilon_k = \begin{cases} +1, &\hbox{if the $k$th step of $b$ is \beginpicture \setcoordinatesystem units <1cm,1cm> \setplotarea x from -0.7 to 0.7, y from -0.5 to 0.5 \put{$\scriptstyle{-}$}[b] at -0.3 0.25 \put{$\scriptstyle{+}$}[b] at 0.3 0.25 \plot 0 -0.4 0 0.5 / \arrow <5pt> [.2,.67] from -0.5 0 to 0.5 0 % \endpicture ,}\\ -1, &\hbox{if the $k$th step of $b$ is \beginpicture \setcoordinatesystem units <1cm,1cm> \setplotarea x from -0.7 to 0.7, y from -0.5 to 0.5 \put{$\scriptstyle{-}$}[b] at -0.3 0.25 \put{$\scriptstyle{+}$}[b] at 0.3 0.25 \plot 0 -0.4 0 0.5 / \arrow <5pt> [.2,.67] from 0.5 0 to -0.5 0 % \endpicture .} \end{cases} $$ Given a walk $h\in \Gamma(\vec{z}^{-1},w^{-1})$, let \begin{equation} \begin{array}{rl} \xi_{\rmdes}(h)\!\!\!&= \{k\mid \hbox{the $k$th step of $h$ is a descending crossing}\},\\ \phi_{\rmasc}(h)\!\!\!&= \{k\mid \hbox{the $k$th step of $h$ is an ascending fold}\}, \end{array} \end{equation} \begin{equation}\label{eqn.psip} \psi(h) = \left\{k\ \bigg| \begin{array}{c} \hbox{the $k$th step of $h$ is an ascending fold and } \epsilon_k = -1 \\ \hbox{or the $k$th step of $h$ is a descending fold and } \epsilon_k = +1 \end{array} \right\}. \end{equation} \begin{theorem}\label{thm.X-tau} Let $z, w\in W^\vee$, and fix a reduced expression $z = s_{i_r}^\vee \cdots s_{i_1}^\vee (\pi_j^\vee)^{-1}$. Then $$ X^z \tau_w^\vee =\sum_{h\in \Gamma(\vec{z}^{-1},w^{-1})} (-1)^{|\phi_{\rmasc}(h)|}\tau_{\tte(h)^{-1}}^\vee g_h(Y) n_h(Y),$$ where the sum is over all alcove walks of type $\vec{z}^{-1} = (i_1,\ldots, i_r)$ beginning in $w^{-1}(\pi_j^\vee)^{-1}$, $$g_h(Y) = \left(\prod_{k\in\phi(h)} \frac{t_{i_k}^{-1/2}-t_{i_k}^{1/2}}{1-Y^{-h_k^\vee}}\right)\!\!\! \left(\prod_{k\in\psi(h)} Y^{-h_k^\vee}\right),\quad n_h(Y) = \prod_{k\in \xi_\rmdes(h)} \frac{1-t_{i_k}^{-1} Y^{-h_k^\vee}}{1-Y^{-h_k^\vee}} \frac{1-t_{i_k} Y^{-h_k^\vee}}{1-Y^{-h_k^\vee}}, $$ and $h_k^\vee$ indexes the $k$th active hyperplane of $h$, see~\eqref{eqn.pkvee}. \end{theorem} \proof The proof is by induction on the length of $z$. The base case is the formulas $$(T_{i}^\vee)^{\epsilon_i} = \tau_i^\vee - \frac{(t_i^{-1/2}-t_i^{1/2})(Y^{-\alpha_i^\vee})^{\frac12(1-\epsilon_i)}}{1-Y^{-\alpha_i^\vee}} = \tau_i^\vee + \frac{(t_i^{-1/2}-t_i^{1/2})(Y^{\alpha_i^\vee})^{\frac12(1+\epsilon_i)}}{1-Y^{\alpha_i^\vee}},$$ where $\epsilon_i \in \{\pm1\}$. For the induction step, let $h \in \Gamma(\vec{z}^{-1}, w^{-1})$, $$g_h(Y) = \prod_{k\in\phi(h)} \frac{t_{i_k}^{-1/2}-t_{i_k}^{1/2}}{1-Y^{-h_k^\vee}} \prod_{k\in\psi(h)} Y^{-h_k^\vee},\qquad n_h(Y) = \prod_{k\in \xi_\rmdes(h)} \frac{1-t_{i_k}^{-1} Y^{-h_k^\vee}}{1-Y^{-h_k^\vee}} \frac{1-t_{i_k} Y^{-h_k^\vee}}{1-Y^{-h_k^\vee}}, $$ and let $h_1, h_2 \in \Gamma(\vec{v}^{-1}s_{i}^\vee, w^{-1})$ be the two extensions of $h$ by a crossing and a folding of type $i$, respectively. By induction, a term in $(T_i^\vee)^{\epsilon_i}X^z\tau_w^\vee$ is \begin{align*} (&-1)^{|\phi_a(h)|} \left(T_i^\vee\right)^{\epsilon_i} \tau_{\tte(h)^{-1}}^\vee g_h(Y)n_h(Y)\\ &= \begin{cases} (-1)^{|\phi_a(h)|} \displaystyle\left(\tau_{i}^\vee - \frac{(t_{i}^{-1/2}-t_{i}^{1/2}) \left(Y^{-\alpha_{i}^\vee} \right)^{\frac12(1-\epsilon_i)}} {1-Y^{-\alpha_{i}^\vee}} \right) \tau_{\tte(h)^{-1}}^\vee g_h(Y) n_h(Y), &\hbox{if } \tte(h)s_i^\vee > \tte(h),\\ (-1)^{|\phi_a(h)|} \displaystyle\left(\tau_{i}^\vee + \frac{(t_{i}^{-1/2}-t_{i}^{1/2}) \left(Y^{\alpha_{i}^\vee} \right)^{\frac12(1+\epsilon_i)}} {1-Y^{\alpha_{i}^\vee}} \right) \tau_{\tte(h)^{-1}}^\vee g_h(Y) n_h(Y), &\hbox{if } \tte(h)s_i^\vee < \tte(h), \end{cases}\\ &= (-1)^{|\phi_a(h_1)|} \tau_{\tte(h_1)^{-1}}^\vee g_{h_1}(Y) n_{h_1}(Y) + (-1)^{|\phi_a(h_2)|} \tau_{\tte(h_2)^{-1}}^\vee g_{h_2}(Y) n_{h_2}(Y), \end{align*} since if $\tte(h)s_i^\vee < \tte(h)$ (the last step is descending), then $$\tau_i^\vee \tau_{\tte(h)^{-1}}^\vee = (\tau_i^\vee)^2 \tau_{(s_i^\vee\tte(h_1))^{-1}}^\vee = \tau_{\tte(h_1)^{-1}}^\vee \left(\frac{1-t_{i}^{-1}Y^{-\tte(h_1)\alpha_{i}^\vee}} {1-Y^{-\tte(h_1)\alpha_{i}^\vee}} \right) \left(\frac{1-t_{i}Y^{-\tte(h_1)\alpha_{i}^\vee}} {1-Y^{-\tte(h_1)\alpha_{i}^\vee}} \right) ,$$ and the hyperplane crossed by the last step of $h_1$ is indexed by $$h_{r+1}^\vee = -\tte(h)\alpha_i^\vee = -\tte(h_1)s_i^\vee \alpha_i^\vee = \tte(h_1)\alpha_i^\vee.$$ \qed The bijection between left and right $W_0$-cosets of $W^\vee$ gives a bijection between minimal length left coset representatives and alcoves in the dominant chamber (minimal length right coset representatives) via taking inverses: \begin{eqnarray*} W^\vee/W_0 & \longleftrightarrow & W_0\backslash W^\vee\\ m_\mu & \leftrightarrow & m_\mu^{-1}. \end{eqnarray*} If $h$ is a walk whose endpoint $\tte(h)$ is in the dominant chamber, define $\varpi(h)\in\fh_\bbZ^*$ by \begin{equation}\label{eqn.varweight} \tte(h)^{-1}=m_{\varpi(h)}. \end{equation} Let \begin{equation} \begin{array}{rl} \phi_{\mathrm{aff}}(h)\!\!\! &= \{ k \mid \hbox{the $k$th step of $h$ is a fold touching an affine hyperplane}\},\\ \phi_{\mathrm{o}}(h)\!\!\! &= \{ k \mid \hbox{the $k$th step of $h$ is a fold touching a hyperplane containing $0$}\}, \end{array} \end{equation} \begin{equation}\label{eqn.psiCaff} \psi_{\mathrm{aff}}(h) = \left\{k\in \phi_\mathrm{aff}(h) \bigg| \begin{array}{c} \hbox{the $k$th step of $h$ is a negative fold and } \epsilon_k = -1 \\ \hbox{or the $k$th step of $h$ is a positive fold and } \epsilon_k = +1 \end{array} \right\}. \end{equation} The closure of the dominant chamber is $\overline{\ttC}= \{x\in \fh_\bbR^* \mid 0 \leq \langle x, \alpha^\vee \rangle \hbox{ for } \alpha \in R_+\}$. \begin{corollary}\label{cor.X-tau} Let $\mu\in \fh_\bbZ^*$, and fix a reduced expression $m_\mu = s_{i_r}^\vee \cdots s_{i_1}^\vee (\pi_j^\vee)^{-1}$ for the minimal length element $m_\mu$ in the coset $x^\mu W_0$. The monomial $X^\mu$ as a linear combination of nonsymmetric Macdonald polynomials is $$ X^\mu = t_{v_\mu^{-1}}^{-1/2}\sum_{h\in \Gamma^{\overline\ttC}(\vec{m}_\mu^{-1})} (-1)^{\phi_-(h)}g_h n_h E_{\varpi(h)},$$ where the sum is over all alcove walks of type $\vec{m_\mu}^{-1}=(\pi_j^\vee, i_1,\ldots, i_r)$ beginning in the fundamental alcove and contained in $\overline{\ttC}$, \begin{align*} g_h &=\left(\prod_{k\in\phi_{\mathrm{o}}(h)} t_{i_k}^{\epsilon_k/2}\right) \left(\prod_{k\in\phi_\mathrm{aff}(h)} \frac{t_{i_k}^{-1/2}-t_{i_k}^{1/2}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}} \right) \left(\prod_{k\in \psi_\mathrm{aff}(h)} q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}\right),\\ n_h &= \prod_{k\in \xi_-(h)} \frac{1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}t_{i_k}^{-1}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}} \frac{1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}t_{i_k}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}}, \end{align*} $h_k^\vee$ indexes the $k$th active hyperplane of $h$, and $\epsilon_1,\ldots, \epsilon_r$ are as defined in~\eqref{eqn.epsilonk}. \end{corollary} \proof Since $X^\mu\mathbf{1} = X^{m_\mu} T_{v_\mu^{-1}}^{-1}\mathbf{1} = t_{v_\mu^{-1}}^{-1/2} X^{m_\mu}\mathbf{1}$, then set $z = m_\mu$, $w=1$ in Theorem~\ref{thm.X-tau} and use $$Y^{-h_k^\vee}\mathbf{1}= q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}\mathbf{1}, \quad \hbox{and} \quad \tau_{\tte(h)^{-1}}^\vee\mathbf{1} = \tau_{m_{\varpi(h)}}^\vee\mathbf{1} = E_{\varpi(h)}\mathbf{1}. $$ The main difference in applying the formula for $X^\mu$ in Theorem~\ref{thm.X-tau} to $\mathbf{1}$ is that if a walk $h\in \Gamma(\vec{m}_\mu^{-1})$ is not contained in the dominant chamber, then it has zero contribution to the sum. Identify the expression $\tau_{\tte(h)^{-1}}^\vee n_h(Y)$ with the product of interwining operators which correspond to the crossing steps of $h$. That is, $$\tau_{\tte(h)^{-1}}^\vee n_h(Y) = \tau_{i_{c_l}}^\vee \cdots \tau_{i_{c_1}}^\vee(\pi_j^\vee)^{-1},$$ where the $c_d$th step of $h$ is a crossing for $d=1,\ldots, l$. If $h$ leaves the dominant chamber at the $k$th crossing, then $\left(s_{i_{c_k}}^\vee \cdots s_{i_{c_1}}^\vee (\pi_j^\vee)^{-1}\right)^{-1}$ is not a minimal length coset representative, and \begin{align*} \tau_{\tte(h)^{-1}}^\vee n_h(Y)g_h(Y)\mathbf{1} &= g_h \ \tau_{i_{c_l}}^\vee \cdots \tau_{i_{c_{k+1}}}^\vee \left(\tau_{i_{c_k}}^\vee \cdots \tau_{i_{c_1}}^\vee (\pi_j^\vee)^{-1}\mathbf{1}\right) = 0. \end{align*} Since all walks which are now under consideration are contained in $\overline\ttC$, a few simplifications can be made. The steps of a walk $h$ contained in $\overline\ttC$ may be analyzed according to whether the active hyperplane is a wall of the dominant chamber (ie.\! is one of $\ttH_{\alpha_1^\vee},\ldots, \ttH_{\alpha_n^\vee}$), or is an affine hyperplane. There are three possibilities: \begin{enumerate} \item[(a)] $k\in \xi_\mathrm{aff}(h)$: the $k$th step of $h$ crosses an affine hyperplane. In this case, ascending crossings are equivalent to positive crossings, and descending crossings are equivalent to negative crossings. So $\xi_\rmdes(h) = \xi_-(h)$. \item[(b)] $k\in \phi_\mathrm{o}(h)$: the $k$th step of $h$ is a fold touching a wall of $\ttC$. In this case, $Y^{-h_k^\vee}\mathbf{1} = q^0t^{-\langle \alpha_{i_k}^\vee,\rho \rangle} = t_{i_k}^{-1}$, and the fold is necessarily ascending and positive. In particular, $$ -\frac{(t_{i_k}^{-1/2}-t_{i_k}^{1/2})(Y^{-h_k^\vee})^{\frac12(1-\epsilon_k)}} {1-Y^{-h_k^\vee}}\mathbf{1} = t_{i_k}^{\epsilon_k/2}\mathbf{1}. $$ \item[(c)] $k\in \phi_\mathrm{aff}(h)$: the $k$th step of $h$ is a fold touching an affine hyperplane. In this case, ascending folds are equivalent to negative folds, and descending folds are equivalent to positive folds. \end{enumerate} Therefore, (b) and (c) give \begin{align*} (-1)&^{|\phi_\rmasc(h)|} \left(\prod_{k\in\phi(h)} \frac{t_{i_k}^{-1/2}-t_{i_k}^{1/2}}{1-Y^{-h_k^\vee}} \right) \left(\prod_{k\in\psi(h)} Y^{-h_k^\vee}\right) \mathbf{1} \\ &= \left(\prod_{k\in\phi_{\mathrm{o}}(h)} t_{i_k}^{\epsilon_k/2}\right) \left((-1)^{\phi_-(h)} \prod_{k\in\phi_{\mathrm{aff}}(h)} \frac{t_{i_k}^{-1/2}-t_{i_k}^{1/2}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}} \prod_{k\in \psi_{\mathrm{aff}}(h)} q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}\right)\mathbf{1}. \end{align*} \qed \begin{remark}{\label{rem.b}\ } {\em \begin{enumerate} \item[(a)] If the walk $h$ is contained in the dominant chamber and the $k$th step of $h$ is a fold against an affine hyperplane indexed by $h_k^\vee = -\beta^\vee + jd$, then $q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}\mathbf{1} = q^j t^{\langle \beta^\vee, \rho \rangle}\mathbf{1},$ where $j \in \bbN$ and $\langle \beta^\vee, \rho\rangle \in \bbN$, since $\beta^\vee \in R_+^\vee$. \item[(b)] Moreover, if $\mu$ is a dominant weight, then $\epsilon_k = -1$ for all $k=1,\ldots, r$ in equation~\eqref{eqn.epsilonk}, and so $$\psi_\mathrm{aff}(h) = \phi_-(h).$$ \end{enumerate} }\end{remark} \subsection{Symmetric Macdonald polynomials}\label{sec.symmMac} The {\em dominant weights} are $$(\fh_\bbZ^*)^+ = \{\mu \in \fh_\bbZ^* \mid \langle \mu, \alpha_i^\vee \rangle \geq 0 \hbox{ for } i=1,\ldots, n \}.$$ Given a dominant weight $\mu \in (\fh_\bbZ^*)^+$, let $m_\mu \in W^\vee$ be the shortest element in the coset $x^\mu W_0$ (see equation~\eqref{eqn.mlcr}). Let $W_\mu \subseteq W_0$ be the stabilizer of $\mu$, $W^\mu$ be the minimal length representatives of $W_0/W_\mu$, and $w_\mu\in W_\mu$, $v_\mu \in W^\mu$ be the longest element in their respective set (see~\cite{BB05}). Also let $W_\mu(t)= \sum_{u\in W_\mu} t_u$ be the {\em Poincar\'{e} polynomial} of $W_\mu$. \begin{lemma}\label{lem.Poincare} Let $\mu \in (\fh_\bbZ^*)^+$, and let $W_\mu(t)$ be the Poincar\'e polynomial of the stabilizer of $\mu$. Then $$\sum_{u\in W_\mu} \left( \prod_{\alpha^\vee \in \calL(1,u)} t_{\alpha^\vee}^{\frac12} \frac{1-t^{\hgt(-\alpha^\vee)}t_{\alpha^\vee}^{-1}} {1-t^{\hgt(-\alpha^\vee)}} \right) \left( \prod_{\alpha^\vee \in \calL(u,w_\mu)} t_{\alpha^\vee}^{-\frac12} \frac{1-t^{\hgt(-\alpha^\vee)}t_{\alpha^\vee}} {1-t^{\hgt(-\alpha^\vee)}} \right) = t_{w_\mu}^{-\frac12} W_\mu(t).$$ \end{lemma} \proof For $u\in W_\mu$, since $$\calL(1,u) = \left\{u\beta^\vee \in R_+^\vee: \beta^\vee \in R_-^\vee\right\}, \quad\hbox{ and }\quad \calL(u,w_\mu) = \left\{u\beta^\vee \in R_-^\vee: \beta^\vee \in R_-^\vee\right\},$$ then \begin{align*} \sum_{u\in W_\mu} & \left( \prod_{\alpha^\vee \in \calL(1,u)} t_{\alpha^\vee}^{\frac12} \frac{1-t^{\hgt(-\alpha^\vee)}t_{\alpha^\vee}^{-1}} {1-t^{\hgt(-\alpha^\vee)}} \right) \left( \prod_{\alpha^\vee \in \calL(u,w_\mu)} t_{\alpha^\vee}^{-\frac12} \frac{1-t^{\hgt(-\alpha^\vee)}t_{\alpha^\vee}} {1-t^{\hgt(-\alpha^\vee)}} \right)\mathbf{1}\\ &= \sum_{u\in W_\mu} \left( \prod_{\beta^\vee \in R_-^\vee: \atop u\beta^\vee \in R_+^\vee} t_{-u\beta^\vee}^{\frac12} \frac{1-Y^{-u\beta^\vee}t_{-u\beta^\vee}^{-1}} {1-Y^{-u\beta^\vee}} \right) \left( \prod_{\beta^\vee \in R_-^\vee: \atop u\beta^\vee \in R_-^\vee} t_{u\beta^\vee}^{-\frac12} \frac{1-Y^{u\beta^\vee}t_{u\beta^\vee}} {1-Y^{u\beta^\vee}} \right)\mathbf{1}\\ &= t_{w_\mu}^{-\frac12} \sum_{u\in W_\mu} u \left( \prod_{\beta^\vee \in R_-^\vee} \frac{1-Y^{\beta^\vee}t_{\beta^\vee}} {1-Y^{\beta^\vee}} \right)\mathbf{1}\\ &= t_{w_\mu}^{-\frac12} W_\mu(t)\mathbf{1}, \end{align*} where the last equality is~\cite[Corollary 2.6(a)]{NR03}. \qed \begin{defn}{\em The {\em symmetric Macdonald polynomial} $P_\mu\in \bbK[X]^{W_0}$ indexed by $\mu \in (\fh_\bbZ^*)^+$ is defined by \begin{equation}\label{eqn.defnP} P_\mu \mathbf{1} = \frac{1}{t_{w_\mu}^{-1/2}W_\mu(t)}\ \mathbf{1}_0\tau_{m_\mu}^\vee \mathbf{1},\qquad \hbox{where } \mathbf{1}_0 = \sum_{w\in W_0} t_{w_0w}^{-1/2}T_w. \end{equation} See~\cite[(5.5.7) and (5.7.10)]{M03}. }\end{defn} The symmetric Macdonald polynomials $\{P_\mu\mathbf{1} \mid \mu\in (\fh_\bbZ^*)^+\}$ is a basis for the $W_0$-invariant polynomials $$\bbK[X]^{W_0}\mathbf{1}= \{f\mathbf{1} \mid wf\mathbf{1} = f\mathbf{1}\hbox{ for all } w\in W_0\}.$$ \begin{remark}{\em The polynomial $P_\mu$ is normalized so that the coefficient of $X^\mu$ in the monomial expansion of $P_\mu$ is $1$. The normalization chosen here is different from that in~\cite{RY10}. }\end{remark} The following result provides an expression for $\mathbf{1}_0$ in terms of intertwining operators. From this, $P_\mu$ can be expressed as a linear combination of $E_\nu$ for $\nu \in W_0\mu$. \begin{prop}\label{prop.10-tau}\label{cor.P-E} \cite[p.\! 203]{M94}. Also see~\cite[(5.7.8)]{M03} and~\cite[(4.13)]{C95b}. Recall $\calL(v,w) = \{a^\vee \in S_+^\vee \mid \ttH_{a^\vee} \hbox{ separates } v\ttA \hbox{ and } w\ttA\}$. \begin{enumerate} \item[(a)] The symmetrizing operator is $$\mathbf{1}_0 = \sum_{w\in W_0} \tau_w^\vee \left(\prod_{a^\vee \in \calL(w^{-1},w_0)} t_{a^\vee}^{1/2}\left( \frac{1-t_{a^\vee}^{-1}Y^{-a^\vee}}{1-Y^{-a^\vee}} \right)\right),$$ \item[(b)] For $\mu\in (\fh_\bbZ^*)^+$, $$P_\mu = \sum_{v\in W^\mu} \prod_{a^\vee \in m_\mu^{-1}\calL(v^{-1},v_\mu^{-1})} t_{a^\vee}^{1/2} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)} t_{a^\vee}^{-1}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}} \ E_{v\mu}. $$ \end{enumerate} \end{prop} \proof \ (a) Since $\mathbf{1}_0= \sum_{w\in W_0} t_{w_0w}^{-1/2} T_w$, then $\mathbf{1}_0$ can be written in the form $\sum_{w\in W_0} \tau_w^\vee b_w$, where each $b_w$ is a rational function in $Y$. The coefficient of $T_{w_0}$ in $\mathbf{1}_0$ is $1$, hence $b_{w_0}=1$. The other $b_w$ can be computed by induction on the length of $w$, comparing coefficients in $$\sum_{w\in W_0}\tau_w^\vee b_w t_i^{1/2} = \mathbf{1}_0 t_i^{1/2} = \mathbf{1}_0 T_i = \sum_{w\in W_0}\tau_w^\vee b_w T_i = \sum_{w\in W_0}\tau_w^\vee b_w\left(\tau_i^\vee -\frac{t_i^{-1/2}-t_i^{1/2}}{1-Y^{-\alpha_i^\vee}}\right), $$ using $\mathbf{1}_0T_i = \mathbf{1}_0 t_i^{1/2}$ for $i=1,\ldots, n$ from~\cite[(5.5.9)]{M03}. (b) Each element $w$ in $W_0$ has a unique factorization of the form $w=vu$ where $v\in W^\mu$ and $u\in W_\mu$, so following (a), \begin{align*} \mathbf{1}_0 &= \sum_{w\in W_0} \tau_w^\vee b_{(w^{-1}, w_0^{-1})} = \left(\sum_{v\in W^\mu} \tau_v^\vee b_{(v^{-1},v_\mu^{-1})}\right) \left(\sum_{u\in W_\mu}\tau_u^\vee b_{(u^{-1},w_\mu^{-1})}\right), \end{align*} where \begin{equation}\label{eqn.bvw} b_{(v,w)} = \prod_{a^\vee \in \calL(v,w)} t_{a^\vee}^{1/2} \frac{1-t_{a^\vee}^{-1}Y^{-a^\vee}}{1-Y^{-a^\vee}}. \end{equation} Applying $\mathbf{1}_0$ to $\tau_{m_\mu}^\vee\mathbf{1}$, and using the fact that $T_u f\mathbf{1} = t_u^{1/2} f\mathbf{1}$ if $u (f\mathbf{1}) = f\mathbf{1}$, then \begin{align*} \mathbf{1}_0 \tau_{m_\mu}^\vee\mathbf{1} &= \left(\sum_{v\in W^\mu} \tau_v^\vee b_{(v^{-1},v_\mu^{-1})}\right) \left(\sum_{u\in W_\mu} t_{w_\mu u}^{-1/2}T_u \right) \tau_{m_\mu}^\vee \mathbf{1} = t_{w_\mu}^{-1/2} W_\mu(t) \left(\sum_{v\in W^\mu} \tau_v^\vee b_{(v^{-1},v_\mu^{-1})} \tau_{m_\mu}^\vee \right)\mathbf{1}\\ &= t_{w_\mu}^{-1/2} W_\mu(t) \sum_{v\in W^\mu} \left(\prod_{a^\vee \in m_\mu^{-1}\calL(v^{-1},v_\mu^{-1})} t_{a^\vee}^{1/2} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}t_{a^\vee}^{-1}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}} \right) E_{v\mu}\mathbf{1}. \end{align*} Observe that $m_\mu^{-1} v_\mu^{-1} = (v_\mu m_\mu)^{-1} = x^{-w_0\mu}$. \qed The following calculation shows that if the weights $\mu$ and $\nu$ lie in the same $W_0$-orbit, then $\mathbf{1}_0 \tau_{m_\mu}^\vee\mathbf{1}$ and $\mathbf{1}_0 \tau_{m_\nu}^\vee\mathbf{1}$ differ by a scalar multiple. \begin{prop}\label{prop.domP} Let $\mu\in (\fh_\bbZ^*)^+$ and $v\in W^\mu$ with $vm_\mu > m_\mu$. Then $$\mathbf{1}_0\tau_v^\vee \tau_{m_\mu}^\vee\mathbf{1} = \left(\prod_{a^\vee \in \calL(m_\mu^{-1}, m_\mu^{-1}v^{-1})} t_{a^\vee}^{-\frac12} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}t_{a^\vee}}{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}}\right) \mathbf{1}_0 \tau_{m_\mu}^\vee\mathbf{1}, $$ where $\calL(v,w) = \{a^\vee\in S_+^\vee \mid \ttH_{a^\vee} \hbox{ separates $v$ and $w$} \}$. \end{prop} \proof By~\cite[(5.5.9)]{M03}, $\mathbf{1}_0T_i = \mathbf{1}_0 t_i^{\frac12}$ for $i=1,\ldots, n$, so $$\mathbf{1}_0 \tau_i^\vee \tau_w^\vee \mathbf{1} = \mathbf{1}_0 \left(t_i^{\frac12} + \frac{t_i^{-\frac12}-t_i^{\frac12}}{1-Y^{-\alpha_i^\vee}}\right) \tau_w^\vee \mathbf{1} = \mathbf{1}_0 \tau_w^\vee \left(t_i^{\frac12} + \frac{t_i^{-\frac12}-t_i^{\frac12}}{1-q^{\shf(-w^{-1}\alpha_i^\vee)} t^{\hgt(-w^{-1}\alpha_i^\vee)}} \right)\mathbf{1}.$$ Let $v=s_{i_1} \cdots s_{i_r}\in W^\mu$ be a reduced expression, so that $\mathbf{1}_0 \tau_v^\vee \tau_{m_\mu}^\vee\mathbf{1} = \mathbf{1}_0 \tau_{i_1}^\vee \cdots \tau_{i_r}^\vee \tau_{m_\mu}^\vee \mathbf{1}.$ Starting with $\tau_{i_1}^\vee$, commuting each $\tau_{i_j}^\vee$ from left to right gives \begin{align*} \mathbf{1}_0\tau_v^\vee \tau_{m_\mu}^\vee \mathbf{1} &= \mathbf{1}_0 \tau_{m_\mu}^\vee \prod_{j=1}^r \left(t_{i_j}^{\frac12}+\frac{t_{i_j}^{-\frac12}-t_{i_j}^{\frac12}} {1-Y^{-m_\mu^{-1}s_{i_r} \cdots s_{i_{j+1}} \alpha_{i_j}^\vee}} \right)\mathbf{1} \\ &= \mathbf{1}_0 \tau_{m_\mu}^\vee \prod_{a^\vee\in \calL(m_\mu^{-1},m_\mu^{-1}v^{-1})} \left(t_{a^\vee}^{\frac12} +\frac{t_{a^\vee}^{-\frac12}-t_{a^\vee}^{\frac12}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}} \right)\mathbf{1}. \end{align*} \qed See Remark~\ref{rem.goodeh} for an alternate formulation of this statement. \section{Multiplication formulas}\label{sec.lrrule} In this section, all alcove walks under consideration are contained in the closure $\overline{\ttC}$ of the dominant chamber. \subsection{Littlewood-Richardson formulas} Various multiplication formulas for Macdonald polynomials can be obtained by applying Theorem~\ref{thm.X-tau} to $\mathbf{1}$. \begin{corollary}\label{cor.XE-E} Let $\mu,\lambda \in \fh_\bbZ^*$, and fix a reduced expression $x^\mu = s_{i_r}^\vee \cdots s_{i_1}^\vee (\pi_j^\vee)^{-1}$. Then $$ X^\mu E_\lambda = \sum_{h\in \Gamma^{\overline\ttC}(\vec{x^{-\mu}}, m_\lambda^{-1})} (-1)^{\phi_-(h)}g_h n_h E_{\varpi(h)},$$ where the sum is over the set of alcove walks of type $\vec{x^{-\mu}}=(\pi_j^\vee, i_1,\ldots, i_r)$ beginning in $m_\lambda^{-1}$ and contained in $\overline\ttC$, \begin{align*} g_h &=\left(\prod_{k\in\phi_{\mathrm{o}}(h)} t_{i_k}^{\epsilon_k/2}\right) \left(\prod_{k\in\phi_\mathrm{aff}(h)} \frac{t_{i_k}^{-1/2}-t_{i_k}^{1/2}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}} \right) \left(\prod_{k\in \psi_\mathrm{aff}(h)} q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}\right),\\ n_h &= \prod_{k\in \xi_-(h)} \frac{1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}t_{i_k}^{-1}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}} \frac{1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}t_{i_k}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}}, \end{align*} $h_k^\vee$ indexes the $k$th active hyperplane of $h$, and $\epsilon_1,\ldots, \epsilon_r$ are as defined in~\eqref{eqn.epsilonk}. \end{corollary} \proof Apply Theorem~\ref{thm.X-tau} to $\mathbf{1}$ with $z=x^\mu$, $w=m_\lambda$ , and the proof is the same as for Corollary~\ref{cor.X-tau}. \qed Next, we examine multiplication involving symmetric Macdonald polynomials. The following notation will be used throughout the rest of this section. For $\mu \in \fh_\bbZ^*$, $\lambda \in (\fh_\bbZ^*)^+$, fix a reduced expression for the minimal length representative $m_\mu = s_{i_r}^\vee \cdots s_{i_1}^\vee (\pi_j^\vee)^{-1}$ of the coset $x^\mu W_0$. For $v\in W^\lambda$, let $\Gamma_2^{\overline\ttC}\left(\vec{m}_\mu^{-1}, (vm_\lambda)^{-1}\right)$ denote the set of alcove walks satisfying the following properties: \begin{enumerate} \item has type $\vec{m}_\mu^{-1} = (\pi_j^\vee, i_1,\ldots, i_r)$, \item begins in $(vm_\lambda)^{-1}$, \item is contained in $\overline\ttC$, \item each fold is coloured either black or grey. \end{enumerate} Given $h\in \Gamma_2^{\overline\ttC}$, let \begin{eqnarray} &&\label{eqn.ph} \begin{array}{l} \hbox{$p(h)$ be the walk obtained by straightening all grey folds of $h$,}\\ \qquad\hbox{and translated so that it ends in $1$,} \end{array}\\ &&\label{eqn.epsilonkh} \epsilon_k(h) = \begin{cases} +1 & \hbox{if the $k$th step of $p(h)$ is positive,}\\ -1 & \hbox{if the $k$th step of $p(h)$ is negative.} \end{cases} \end{eqnarray} See Example~\ref{eg.2} for an illustration. \begin{theorem}\label{thm.EP-E} Let $\mu \in \fh_\bbZ^*$, $\lambda \in (\fh_\bbZ^*)^+$, and fix a reduced expression for the minimal length representative $m_\mu = s_{i_r}^\vee \cdots s_{i_1}^\vee (\pi_j^\vee)^{-1}$ of the coset $x^\mu W_0$. Then $$E_\mu P_\lambda = \sum_{v\in W^\lambda} \sum_{h \in \Gamma_2^{\overline\ttC}\left(\vec{m}_\mu^{-1}, (vm_\lambda)^{-1}\right) } \left( (-1)^{|\phi_{\mathrm{grey}}^-(h)|} b_h\prod_{k=1}^r c_{k(h)}\right) E_{\varpi(h)},$$ where $|\phi_{\mathrm{grey}}^-(h)|$ is the number negative grey folds of $h$, \begin{equation} b_h = \prod_{a^\vee \in \calL(\ttb(h),x^{-w_0\lambda})} t_{a^\vee}^{1/2} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}t_{a^\vee}^{-1}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}}, \qquad \ttb(h) \hbox{ is where $h$ begins,} \end{equation} and the coefficient $c_{k(h)}$ arising from the $k$th step of $h$ is \begin{equation}\label{eqn.verylonglist} \begin{array}{ll} 1, & \begin{array}{l} \hbox{for a positive crossing,} \end{array}\\ \displaystyle \frac{1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}t_{i_k}^{-1}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}} \frac{1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}t_{i_k}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}}, & \begin{array}{l} \hbox{for a negative crossing,} \end{array}\\ \\ \displaystyle t_{i_k}^{\epsilon_k/2}, & \begin{array}{l} \hbox{for a grey fold}\\ \quad\hbox{touching a wall of }\ttC,\\ \end{array}\\ \\ \displaystyle \frac{(t_{i_k}^{-1/2}-t_{i_k}^{1/2}) (q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)})^{\frac12(1-\epsilon_k(h))}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}}, & \begin{array}{l} \hbox{for a negative grey fold}\\ \quad\hbox{touching an affine hyperplane,}\\ \end{array}\\ \displaystyle \frac{(t_{i_k}^{-1/2}-t_{i_k}^{1/2}) (q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)})^{\frac12(1+\epsilon_k(h))}} {1-q^{\shf(-h_k^\vee)}t^{\hgt(-h_k^\vee)}}, & \begin{array}{l} \hbox{for a positive grey fold}\\ \quad\hbox{touching an affine hyperplane,}\\ \end{array}\\ \\ \displaystyle \frac{t_{b_k^\vee}^{-1/2}-t_{b_k^\vee}^{1/2}} {1-q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)}}, & \begin{array}{l} \hbox{for a black fold such that the}\\ \quad\hbox{$k$th step of $p(h)$ is positive,}\end{array}\\ \displaystyle \frac{(t_{b_k^\vee}^{-1/2}-t_{b_k^\vee}^{1/2}) (q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)})} {1-q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)}}, & \begin{array}{l} \hbox{for a black fold such that the}\\ \quad\hbox{$k$th step of $p(h)$ is negative,}\end{array} \end{array} \end{equation} $\varpi(h)$, $h_k^\vee$, $\epsilon_k(h)$ are respectively defined in~\eqref{eqn.varweight}, \eqref{eqn.pkvee}, \eqref{eqn.epsilonkh}, and $b_k^\vee = s_{i_1}^\vee s_{i_2}^\vee\cdots s_{i_{k-1}}^\vee \alpha_{i_k}^\vee$. \end{theorem} \proof The idea is to expand $E_\mu$ in terms of monomials $X^\nu$ (Theorem~\ref{thm.Xtau-X}), and then expand $X^\nu P_\lambda$ in terms of nonsymmetric Macdonald polynomials (Proposition~\ref{prop.10-tau} and Corollary~\ref{cor.XE-E}). Since $$E_\mu = \sum_{p\in \Gamma(\vec{m}_\mu)} f_p t_{\ttd(p)}^{1/2} X^{\ttwt(p)}, \quad\hbox{where} \quad f_p= \prod_{k\in\phi(p)} \frac{t_{b_k^\vee}^{-1/2}-t_{b_k^\vee}^{1/2}}{1-q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)}} \prod_{k\in \phi_-(p)} q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)},$$ with notation from Theorem~\ref{thm.Xtau-X}. Then \begin{align*} E_\mu \mathbf{1}_0 \tau_{m_\lambda}^\vee\mathbf{1} &= \left(\sum_{p\in \Gamma(\vec{m}_\mu)} f_p X^{\ttwt(p)} T_{\ttd(p)}^{1/2}\right) \mathbf{1}_0 \tau_{m_\lambda}^\vee\mathbf{1} = \sum_{p\in \Gamma(\vec{m}_\mu)} f_p X^{\tte(p)} \mathbf{1}_0 \tau_{m_\lambda}^\vee\mathbf{1}\\ &= t_{w_\lambda}^{-1/2}W_{\lambda}(t) \sum_{v\in W^\lambda} b_{((vm_\lambda)^{-1},x^{-w_0\lambda})} \sum_{p\in \Gamma(\vec{m}_\mu)} f_p X^{\tte(p)} \tau_{vm_\lambda}^\vee\mathbf{1}, \end{align*} where $\displaystyle b_{((vm_\lambda)^{-1}, x^{-w_0\lambda})} = \prod_{a^\vee \in m_\lambda^{-1}\calL(v^{-1},v_\lambda^{-1})} t_{a^\vee}^{1/2} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}t_{a^\vee}^{-1}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}}, $ as in Proposition~\ref{prop.10-tau}. The next step is to express $X^{\tte(p)}\tau_{vm_\lambda}^\vee\mathbf{1}$ in terms of nonsymmetric Macdonald polynomials. The way this is achieved in Corollary~\ref{cor.XE-E} is to interpret $X^{\tte(p)}$ as a minimal length walk (without any folds) of type $\tte(p)^{-1}$ and beginning in the alcove $(vm_\lambda)^{-1}$, then the terms in the expansion of $X^{\tte(p)}\tau_{vm_\lambda}^\vee\mathbf{1}$ are generated by folding this walk in all possible ways (and only keeping those which are contained in the dominant chamber). Instead, we interpret $f_p X^{\tte(p)}$ as a walk (possibly having folds) of type $m_\mu^{-1}$ and beginning in $(vm_\lambda)^{-1}$. Then the terms in the expansion of $f_pX^{\tte(p)}\tau_{vm_\lambda}^\vee\mathbf{1}$ in the nonsymmetric Macdonald polynomial basis are generated by folding this walk in all possible ways (and keeping those contained in the dominant chamber). These newly introduced folds contribute coefficients which are different from those coming from the folds of the seminal walk, so we keep track of the new folds by colouring them grey. What remains to be done is to unify the previous concepts of various kinds of crossings and foldings from Theorem~\ref{thm.Xtau-X} and Corollary~\ref{cor.XE-E}. That is, \begin{align*} f_pX^{\tte(p)}\tau_{vm_\lambda}^\vee\mathbf{1} &= f_{p(h)} \sum_{h} (-1)^{|\phi_{\mathrm{grey}}^-(h)|} g_h n_h E_{\varpi(h)}\mathbf{1} \end{align*} where the sum is over all alcove walks of type $\vec{m}_\mu^{-1}$ beginning in $(vm_\lambda)^{-1}$, which are contained in $\overline\ttC$, such that the $k$th step is a black fold if and only if the $k$th step of $p(h)$ is a fold (see~\eqref{eqn.ph}), and all other folds are grey. Here, the coefficient $g_h$ involves grey folds only (see Corollary~\ref{cor.XE-E}), and $$f_{p(h)} = \prod_{k\in\phi(p(h))} \frac{t_{b_k^\vee}^{-1/2}-t_{b_k^\vee}^{1/2}}{1-q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)}} \prod_{k\in \phi_-(p(h))} q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)}, \qquad\hbox{for } b_k^\vee = s_{i_1}^\vee \cdots s_{i_{k-1}}^\vee \alpha_{i_k}^\vee. $$ Thus, \begin{align*} E_\mu P_\lambda\mathbf{1} &= \sum_{v\in W^\lambda} \sum_{p\in \Gamma(\vec{m}_\mu)} b_{((vm_\lambda)^{-1},x^{-w_0\lambda})} f_p X^{\tte(p)} \tau_{vm_\lambda}^\vee\mathbf{1},\\ &= \sum_{v\in W^\lambda} \sum_{h\in \Gamma_2^{\overline\ttC}(\vec{m}_\mu^{-1}, (vm_\lambda)^{-1})} \left((-1)^{|\phi_\mathrm{grey}^-(h)|} b_h \sum_{k=1}^r c_{k(h)}\right) E_{\varpi(h)}\mathbf{1}. \end{align*} The steps of $h$ may be a crossing, a grey fold, or a black fold, and depending on the step, the coefficient $c_k(h)$ contributed by the $k$th step of $h$ is summarized in table~\eqref{eqn.verylonglist}. \qed The next result expresses the product of two symmetric Macdonald polynomials in terms of symmetric Macdonald polynomials. For the purposes of simplifying the notation, we assume for the remainder of this section that the walls of $\ttC$ are labeled by the {\em negative} coroots $\{-\alpha_1^\vee, \ldots, -\alpha_n^\vee\}$ instead of the usual positive coroots. The labeling of the affine hyperplanes remain unchanged from before; they are labeled by positive affine coroots $\{ -\alpha^\vee + jd \mid \alpha^\vee \in R_+^\vee, \ j\in \bbZ_{\geq1}\}$. \begin{remark}\label{rem.goodeh}{\em It is now useful to phrase Proposition~\ref{prop.domP} another way. Given $\mu \in (\fh_\bbZ^*)^+$, \begin{align*} \mathbf{1}_0 \tau_{m_\mu}^\vee\mathbf{1} &= \left(t_{w_\mu}^{-\frac12}W_\mu(t)\right) P_\mu\mathbf{1} = \left(\prod_{\alpha^\vee\in \calL(m_\mu^{-1}, m_\mu^{-1}w_\mu^{-1})} t^{-\frac12}_{\alpha^\vee} \frac{1-q^{\shf(-\alpha^\vee)}t^{\hgt(-\alpha^\vee)}t_{\alpha^\vee}} {1-q^{\shf(-\alpha^\vee)}t^{\hgt(-\alpha^\vee)}}\right) P_\mu \mathbf{1}, \end{align*} (this is where we use the convention that the walls of $\ttC$ are labeled by negative coroots rather than positive coroots), then for $z\in W^\mu$, \begin{equation}\label{eqn.goodeh} \mathbf{1}_0 \tau_z^\vee \tau_{m_\mu}^\vee\mathbf{1} = \left(\prod_{a^\vee \in\calL(m_\mu^{-1}w_\mu^{-1}, m_\mu^{-1}z^{-1})} t^{-\frac12}_{a^\vee} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}t_{a^\vee}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}} \right) P_\mu\mathbf{1}. \end{equation} Observe that $m_\mu^{-1} w_\mu^{-1} = \left(v_\mu x^{-\mu}\right) w_\mu^{-1} = x^{-w_0\mu} w_0.$ }\end{remark} \begin{theorem}\label{thm.PP-P} Let $\mu, \lambda \in (\fh_\bbZ^*)^+$, and fix a reduced expression for the minimal length representative $m_\mu = s_{i_r}^\vee \cdots s_{i_1}^\vee (\pi_j^\vee)^{-1}$ of the coset $x^\mu W_0$. With the same notation from Theorem~\ref{thm.EP-E}, then $$P_\mu P_\lambda = \frac{1}{t_{w_\mu}^{-1/2}W_\mu(t)} \sum_{v\in W^\lambda} \sum_{h \in \Gamma_2^{\overline\ttC}\left( \vec{m}_\mu^{-1}, (vm_\lambda)^{-1}\right)} \left((-1)^{\phi_{grey}^-(h)} b_h e_h \prod_{k=1}^r c_{k(h)} \right) P_{-w_0\ttwt(h)},$$ where \begin{equation} e_h = \prod_{a^\vee \in \calL(x^{\ttwt(h)}w_0,\tte(h))} t_{a^\vee}^{-\frac12} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}t_{a^\vee}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}}, \qquad \tte(h) \hbox{ is where $h$ ends.} \end{equation} \end{theorem} \proof Since $t_{w_\mu}^{-\frac12}W_\mu(t) P_\mu\mathbf{1}_0 = \mathbf{1}_0 E_\mu\mathbf{1}_0$, then the result can obtained by applying the operator $\mathbf{1}_0$ to both sides of the equation in Theorem~\ref{thm.EP-E}. The weight $\varpi(h)$ defined by $\tte(h)^{-1} = m_{\varpi(h)}$ is not necessarily dominant, so $\mathbf{1}_0 E_{\varpi(h)}\mathbf{1}$ is a certain scalar multiple of a symmetric Macdonald polynomial. To find this scalar, let $z\in W_0$ be of minimal length such that $z\varpi(h)_+ = \varpi(h)$. Note that $m_{\varpi(h)_+}^{-1} = m_{-w_0\varpi(h)_+}$, as $\varpi(h)_+$ is a dominant weight. Moreover, $m_{-w_0\varpi(h)_+}$ is the minimal length representative of the coset $x^{\ttwt(h)}W_0$, because the endpoint $\tte(h)$ is in that coset by definition, therefore $$\ttwt(h) = -w_0\varpi(h)_+. $$ Hence remark~\ref{rem.goodeh} gives \begin{align*} \mathbf{1}_0 E_{\varpi(h)} \mathbf{1} &=\mathbf{1}_0\tau_z^\vee \tau_{m_{\varpi(h)_+}}^\vee \mathbf{1} = \left(\prod_{a^\vee \in \calL(x^{\ttwt(h)}w_0, \tte(h))} t_{a^\vee}^{-\frac12} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}t_{a^\vee}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}}\right) P_{-w_0\ttwt(h)} \mathbf{1}, \end{align*} since $\tte(h) = m_{\varpi(h)}^{-1} = m_{\varpi(h)_+}^{-1}z^{-1}$. and \begin{align*} P_\mu P_\lambda \mathbf{1} &= \frac{1}{t_{w_\mu}^{-\frac12}W_\mu(t)} \mathbf{1}_0 E_\mu P_\lambda \mathbf{1} = \frac{1}{t_{w_\mu}^{-\frac12}W_\mu(t)} \sum_{v\in W^\lambda} \sum_{h \in \Gamma_2^{\overline\ttC}\left( \vec{m}_\mu^{-1}, (vm_\lambda)^{-1}\right)} a_h(q,t)e_h P_{-w_0\ttwt(h)}\mathbf{1}, \end{align*} where $a_h(q,t)= \left((-1)^{\phi_{grey}^-(h)} b_h \prod_{k=1}^r c_{k(h)} \right)$. \qed \begin{remark}{\em Theorem~\ref{thm.Xtau-X} gives the expansion of $E_\mu$ in terms of monomials, so in theory, the product of nonsymmetric Macdonald polynomials may be computed as a sum over pairs of paths. That is, since $$E_\mu = \sum_{p\in \Gamma(\vec{m}_\mu)} f_p t_{\ttd(p)}^{\frac12} X^{\ttwt(p)}, \quad\hbox{where} \quad f_p= \prod_{k\in\phi(p)} \frac{t_{b_k^\vee}^{-\frac12}-t_{b_k^\vee}^{\frac12}}{1-q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)}} \prod_{k\in \phi_-(p)} q^{\shf(-b_k^\vee)}t^{\hgt(-b_k^\vee)},$$ then \begin{align*} E_\mu E_\lambda\mathbf{1} = E_\mu \tau_{m_\lambda}^\vee \mathbf{1} &= \sum_{p\in \Gamma(\vec{m}_\mu)} f_p t_{\ttd(p)}^{\frac12} X^{\ttwt(p)} \tau_{m_\lambda}^\vee \mathbf{1}\\ &=\sum_{p\in \Gamma(\vec{m}_\mu)} f_p t_{\ttd(p)}^{\frac12} \sum_{h\in \Gamma(\vec{x}^{-\ttwt(p)}, m_\lambda^{-1})} (-1)^{\phi_-(h)}g_h n_h E_{\varpi(h)} \mathbf{1}, \end{align*} with the same notations from Theorem~\ref{thm.Xtau-X} and Corollary~\ref{cor.XE-E}. }\end{remark} \begin{remark}{\em \ \begin{enumerate} \item Using interpolation Macdonald polynomials, Baratta obtained type A Pieri-type formulas~\cite[Proposition 8, 10]{B09} for the expansion of $E_\mu(q,t) P_{\omega_1}(0,0)$ and $E_\mu(q,t) P_{\omega_n}(0,0)$ in terms of nonsymmetric Macdonald polynomials. Also see~\cite{La08}. \item Haglund, Luoto, Mason, and van Willigenburg considered the type A case at $q=t=0$, when $P_\lambda(0,0)$ is a Schur polynomial and $E_\lambda(0,0)$ is a Demazure character, and obtained a formula for the expansion of $E_\mu(0,0) P_\lambda(0,0)$ in terms of $E_\mu(0,0)$ with positive coefficients~\cite[Theorem 6.1]{HLMvW09}. The coefficients $\sum_{h: \varpi(h) = \gamma} a_h(0,0)$ count certain fillings of skew tableau-like diagrams called skyline diagrams. \item Recently, a generalized Macdonald operator was introduced by van Diejen and Emsiz~\cite{vDE10}, who used it to obtain Pieri formulas for Macondald polynomials of arbitrary type, and also a Littlewood-Richardson formula for `small weights'. \end{enumerate} }\end{remark} \begin{example}\label{eg.2}{\em See section \ref{sec.eg} for more details on the alcove picture of type $\fsl_2$. The following walks $h_1$ and $h_2 \in \Gamma_2^{\overline\ttC}(\vec{m}_{-8\omega}^{-1} , (s_1m_{2\omega})^{-1})$ are used in the expansion of $E_{-8\omega}P_{2\omega}$ in terms of nonsymmetric Macdonald polynomials. Each is of type $\vec{m}_{-8\omega}^{-1} = (0,1,0,1,0,1,0,1)$ and begins in $(s_1 m_{2\omega})^{-1} = x^{2\omega}$. $$\beginpicture \setcoordinatesystem units <1.5cm,1cm> \setplotarea x from -2 to 9, y from -1.2 to 1.5 \put{$\bullet$} at 0 0 {\small \put{$\ttH_{\alpha^\vee}$}[t] at 0 -0.6 \put{$\ttH_{-\alpha^\vee+2d}$}[t] at 2 -0.6 \put{$\ttH_{-\alpha^\vee+4d}$}[t] at 4 -0.6 \put{$\ttH_{-\alpha^\vee+6d}$}[t] at 6 -0.6 \put{$\ttH_{-\alpha^\vee+8d}$}[t] at 8 -0.6 }\plot -1.5 0 8.5 0 / \plot -1 1.2 -1 -0.3 / \plot 0 1.2 0 -0.3 / \plot 1 1.2 1 -0.3 / \plot 2 1.2 2 -0.3 / \plot 3 1.2 3 -0.3 / \plot 4 1.2 4 -0.3 / \plot 5 1.2 5 -0.3 / \plot 6 1.2 6 -0.3 / \plot 7 1.2 7 -0.3 / \plot 8 1.2 8 -0.3 / \put{$\scriptstyle{+}$} at 0.2 1.1 \put{$\scriptstyle{-}$} at -0.2 1.1 \put{$1$} at 0.5 -0.3 \put{$x^{2\omega}$} at 2.5 -0.3 \put{$h_1:$} at -1.5 0.7 \setplotsymbol(.) \arrow <6pt> [.2,.67] from 2.5 0.2 to 3.5 0.2 \plot 3.5 0.2 4 0.2 / \plot 3.99 0.2 3.99 0.4 / \arrow <6pt> [.2,.67] from 4 0.4 to 3.5 0.4 \plot 3.5 0.4 3 0.4 / \plot 3.01 0.4 3.01 0.6 / \arrow <6pt> [.2,.67] from 3 0.6 to 3.5 0.6 \arrow <6pt> [.2,.67] from 3.5 0.6 to 4.5 0.6 \arrow <6pt> [.2,.67] from 4.5 0.6 to 5.5 0.6 \arrow <6pt> [.2,.67] from 5.5 0.6 to 6.5 0.6 \plot 6.5 0.6 7 0.6 / \plot 6.99 0.6 6.99 0.8 / \arrow <6pt> [.2,.67] from 7 0.8 to 6.5 0.8 \arrow <6pt> [.2,.67] from 6.5 0.8 to 5.5 0.8 \endpicture $$ The endpoint is in $\tte(h_1) = s_0^\vee s_1 s_0^\vee s_1 s_0^\vee = m_{6\omega}^{-1}$, so $\varpi(h_1)= 6\omega$. $$\beginpicture \setcoordinatesystem units <1.5cm,1cm> \setplotarea x from -2 to 9, y from -1 to 1.5 \put{$\bullet$} at 0 0 {\small \put{$\ttH_{\alpha^\vee}$}[t] at 0 -0.6 \put{$\ttH_{-\alpha^\vee+2d}$}[t] at 2 -0.6 \put{$\ttH_{-\alpha^\vee+4d}$}[t] at 4 -0.6 \put{$\ttH_{-\alpha^\vee+6d}$}[t] at 6 -0.6 \put{$\ttH_{-\alpha^\vee+8d}$}[t] at 8 -0.6 }\plot -1.5 0 8.5 0 / \plot -1 1.2 -1 -0.3 / \plot 0 1.3 0 -0.3 / \plot 1 1.3 1 -0.3 / \plot 2 1.3 2 -0.3 / \plot 3 1.3 3 -0.3 / \plot 4 1.3 4 -0.3 / \plot 5 1.3 5 -0.3 / \plot 6 1.3 6 -0.3 / \plot 7 1.3 7 -0.3 / \plot 8 1.3 8 -0.3 / \put{$\scriptstyle{+}$} at 0.2 1.1 \put{$\scriptstyle{-}$} at -0.2 1.1 \put{$1$} at 0.5 -0.3 \put{$x^{2\omega}$} at 2.5 -0.3 \put{$h_2$:} at -1.5 0.7 \setplotsymbol(.) \arrow <6pt> [.2,.67] from 2.5 0.2 to 3.5 0.2 \plot 3.5 0.2 4 0.2 / \plot 3.99 0.2 3.99 0.4 / \arrow <6pt> [.2,.67] from 4 0.4 to 3.5 0.4 \plot 3.5 0.4 3 0.4 / \plot 3.01 0.4 3.01 0.6 / \arrow <6pt> [.2,.67] from 3 0.6 to 3.5 0.6 \textcolor{Gray}{\plot 3.5 0.6 4 0.6 / \plot 3.99 0.6 3.99 0.8 / \arrow <6pt> [.2,.67] from 4 0.8 to 3.5 0.8 }\arrow <6pt> [.2,.67] from 3.5 0.8 to 2.5 0.8 \arrow <6pt> [.2,.67] from 2.5 0.8 to 1.5 0.8 \plot 1.5 0.8 1 0.8 / \plot 1.01 0.8 1.01 1 / \arrow <6pt> [.2,.67] from 1 1 to 1.5 1 \textcolor{Gray}{\plot 1.5 1 2 1 / \plot 1.99 1 1.99 1.2 / \arrow <6pt> [.2,.67] from 2 1.2 to 1.5 1.2 }\endpicture $$ The endpoint is in $\tte(h_2) = s_0^\vee = m_{2\omega}^{-1}$, so $\varpi(h_2) = 2\omega$. Notice that $h_2$ is obtained from $h_1$ by folding the fourth and eighth steps; these are indicated in grey. Since $h_2$ is generated from $h_1$, then $p=p(h_1)=p(h_2)$ defined in~\eqref{eqn.ph}, is $$\beginpicture \setcoordinatesystem units <1.5cm,1cm> \setplotarea x from -2 to 9, y from -1 to 1.5 \put{$\bullet$} at 0 0 {\small \put{$\ttH_{\alpha^\vee}$}[t] at 0 -0.6 \put{$\ttH_{-\alpha^\vee+2d}$}[t] at 2 -0.6 \put{$\ttH_{-\alpha^\vee+4d}$}[t] at 4 -0.6 \put{$\ttH_{-\alpha^\vee+6d}$}[t] at 6 -0.6 \put{$\ttH_{-\alpha^\vee+8d}$}[t] at 8 -0.6 }\plot -1.5 0 8.5 0 / \plot -1 1.2 -1 -0.3 / \plot 0 1.2 0 -0.3 / \plot 1 1.2 1 -0.3 / \plot 2 1.2 2 -0.3 / \plot 3 1.2 3 -0.3 / \plot 4 1.2 4 -0.3 / \plot 5 1.2 5 -0.3 / \plot 6 1.2 6 -0.3 / \plot 7 1.2 7 -0.3 / \plot 8 1.2 8 -0.3 / \put{$\scriptstyle{+}$} at 0.2 1.1 \put{$\scriptstyle{-}$} at -0.2 1.1 \put{$1$} at 0.5 -0.3 \put{$x^{4\omega}s_1$} at 3.5 -0.3 \put{$p:$} at -1.5 0.7 \setplotsymbol(.) \arrow <6pt> [.2,.67] from 3.5 0.8 to 2.5 0.8 \plot 2.5 0.8 2 0.8 / \plot 2.01 0.6 2.01 0.8 / \arrow <6pt> [.2,.67] from 2 0.6 to 2.5 0.6 \plot 2.5 0.6 3 0.6 / \plot 2.99 0.4 2.99 0.6 / \arrow <6pt> [.2,.67] from 3 0.4 to 2.5 0.4 \arrow <6pt> [.2,.67] from 2.5 0.4 to 1.5 0.4 \arrow <6pt> [.2,.67] from 1.5 0.4 to 0.5 0.4 \arrow <6pt> [.2,.67] from 0.5 0.4 to -0.5 0.4 \plot -0.5 0.4 -1 0.4 / \plot -0.99 0.2 -0.99 0.4 / \arrow <6pt> [.2,.67] from -1 0.2 to -0.5 0.2 \arrow <6pt> [.2,.67] from -0.5 0.2 to 0.5 0.2 \endpicture $$ and the crossing steps give $(\epsilon_1, \epsilon_4, \epsilon_5, \epsilon_6, \epsilon_8) = (-1 ,-1, -1, -1, +1)$. For $k=1,\ldots, 8$, the coroots $b_k^\vee = s_{i_1}^\vee \cdots s_{i_{k-1}}^\vee \alpha_{i_k}^\vee = -\alpha^\vee + kd$ is the sequence of labels of hyperplanes crossed by the walk of type $\vec{m}_{-8\omega}^{-1}$ beginning in $1$: $$\beginpicture \setcoordinatesystem units <1.5cm,1cm> \setplotarea x from -2 to 9, y from -1.2 to 1.5 \put{$\bullet$} at 0 0 {\small \put{$\ttH_{\alpha^\vee}$}[t] at 0 -0.6 \put{$\ttH_{-\alpha^\vee+2d}$}[t] at 2 -0.6 \put{$\ttH_{-\alpha^\vee+4d}$}[t] at 4 -0.6 \put{$\ttH_{-\alpha^\vee+6d}$}[t] at 6 -0.6 \put{$\ttH_{-\alpha^\vee+8d}$}[t] at 8 -0.6 }\plot -1.5 0 8.5 0 / \plot -1 1.2 -1 -0.3 / \plot 0 1.2 0 -0.3 / \plot 1 1.2 1 -0.3 / \plot 2 1.2 2 -0.3 / \plot 3 1.2 3 -0.3 / \plot 4 1.2 4 -0.3 / \plot 5 1.2 5 -0.3 / \plot 6 1.2 6 -0.3 / \plot 7 1.2 7 -0.3 / \plot 8 1.2 8 -0.3 / \put{$\scriptstyle{+}$} at 0.2 1.1 \put{$\scriptstyle{-}$} at -0.2 1.1 \put{$1$} at 0.5 -0.3 \put{$x^{8\omega}$} at 8.5 -0.3 \setplotsymbol(.) \arrow <6pt> [.2,.67] from 0.5 0.4 to 1.5 0.4 \arrow <6pt> [.2,.67] from 1.5 0.4 to 2.5 0.4 \arrow <6pt> [.2,.67] from 2.5 0.4 to 3.5 0.4 \arrow <6pt> [.2,.67] from 3.5 0.4 to 4.5 0.4 \arrow <6pt> [.2,.67] from 4.5 0.4 to 5.5 0.4 \arrow <6pt> [.2,.67] from 5.5 0.4 to 6.5 0.4 \arrow <6pt> [.2,.67] from 6.5 0.4 to 7.5 0.4 \arrow <6pt> [.2,.67] from 7.5 0.4 to 8.5 0.4 \endpicture $$ Using Theorem~\ref{thm.EP-E}, we compute the terms in the expansion of $E_{-8\omega}P_{2\omega}$ arising from $h_1$ and $h_2$. Both walks begin in $x^{2\omega}$, so $b_{(\ttb(h_j), x^{2\omega})} =1 $ for $j=1,2$. From $h_1$, $$\begin{array}{cccc} c_1 = 1, & c_2 = \displaystyle \frac{t^{-\frac12}(1-t)}{1-q^2t}, & c_3 = \displaystyle \frac{t^{-\frac12}(1-t)q^3t}{1-q^3t}, & c_4 = 1,\\ c_5 = 1, & c_6 = 1, & c_7 = \displaystyle \frac{t^{-\frac12}(1-t)}{1-q^7t}, & c_8 = \displaystyle \frac{1-q^6}{1-q^6t} \frac{1-q^6t^2}{1-q^6t}, \end{array}$$ so $h_1$ gives rise to the term $$\frac{1-t}{1-q^2t} \frac{1-t}{1-q^3t} \frac{1-t}{1-q^7t} \frac{1-q^6}{1-q^6t} \frac{1-q^6t^2}{1-q^6t} q^3t^{-\frac12}E_{6\omega}.$$ From $h_2$, $$\begin{array}{cccc} c_1 = 1, & c_2 = \displaystyle \frac{t^{-\frac12}(1-t)}{1-q^2t}, & c_3 = \displaystyle \frac{t^{-\frac12}(1-t)q^3t}{1-q^3t}, & c_4 = \displaystyle \frac{t^{-\frac12}(1-t)q^4t}{1-q^4t},\\ c_5 = \displaystyle \frac{1-q^3}{1-q^3t} \frac{1-q^3t^2}{1-q^3t}, & c_6 = \displaystyle \frac{1-q^2}{1-q^2t} \frac{1-q^2t^2}{1-q^2t}, & c_7 = \displaystyle \frac{t^{-\frac12}(1-t)}{1-q^7t}, & c_8 = \displaystyle \frac{t^{\frac12}(1-t)}{1-q^2t}, \end{array}$$ and $h_2$ has two negative grey folds, so it gives rise to the term $$(-1)^2 \frac{1-t}{1-q^2t} \frac{1-t}{1-q^3t} \frac{1-t}{1-q^4t} \frac{1-q^3}{1-q^3t} \frac{1-q^3t^2}{1-q^3t} \frac{1-q^2}{1-q^2t} \frac{1-q^2t^2}{1-q^2t} \frac{1-t}{1-q^7t} \frac{1-t}{1-q^2t} q^7t^{-\frac12}E_{2\omega}.$$ \hfill$\diamond$ }\end{example} \subsection{Pieri formulas} \label{sec.pieri} This section concerns the special cases of Theorems~\ref{thm.EP-E} and~\ref{thm.PP-P} when $\mu = \omega_j\in (\fh_\bbZ^*)^+$ is a minuscule weight. Recall that $x^{\omega_j} = m_{\omega_j} v_{\omega_j}$. In this section, we will use the notation $v_j = v_{\omega_j}$. Also, in this case, the minimal length coset representative $m_{\omega_j} = \pi_j^\vee \in \Pi^\vee$. So $$E_{\omega_j}\mathbf{1} = \pi_j^\vee \mathbf{1} = X^{\omega_j}T_{v_j^{-1}}\mathbf{1} = t_{v_j^{-1}}^{\frac12}X^{\omega_j}\mathbf{1}. $$ Moreover, the walks appearing in Theorems~\ref{thm.EP-E} and~\ref{thm.PP-P} have type $(\pi_j^\vee)^{-1}$, which is a ``change in sheets''. Such walks do not have crossings or foldings, so the product formulas simplify significantly. For $\lambda \in (\fh_\bbZ^*)^+$, \begin{align} \label{eqn.PieriEP} E_{\omega_j} P_\lambda &= \sum_{v\in W^\lambda} \sum_{h \in \Gamma\left((\pi_j^\vee)^{-1}, (vm_\lambda)^{-1}\right)} \!\!\!\! b_{(\ttb(h),x^{-w_0\lambda})} E_{\varpi(h)},\\ \label{eqn.PieriPP} P_{\omega_j} P_\lambda &= \frac{1}{t_{w_{\omega_j}}^{-\frac12} W_{\omega_j}(t)} \sum_{v\in W^\lambda} \sum_{h \in \Gamma\left((\pi_j^\vee)^{-1}, (vm_\lambda)^{-1}\right)} \!\!\!\! b_{(\ttb(h),x^{-w_0\lambda})} e_{(x^{\ttwt(h)}w_0, \tte(h))} P_{-w_0\ttwt(h)}, \end{align} where each sum is over the set of alcove walks of type $(\pi_j^\vee)^{-1}$ beginning in $(vm_\lambda)^{-1}$ for $v\in W^\lambda$. Recall that the weight $\varpi(h)$ is defined by $\tte(h)^{-1} = m_{\varpi(h)}$, and \begin{align}\label{eqn.bhqt} b_{\left(\ttb(h),x^{-w_0\lambda}\right)} &= \prod_{a^\vee\in\calL\left(\ttb(h), x^{-w_0\lambda}\right)} t_{a^\vee}^{\frac12} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)} t_{a^\vee}^{-1}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}},\\ \label{eqn.ehqt} e_{\left(x^{\ttwt(h)}w_0, \tte(h)\right)} &= \prod_{a^\vee\in\calL\left(x^{\ttwt(h)}w_0, \tte(h)\right)} t_{a^\vee}^{-\frac12} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)} t_{a^\vee}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}}. \end{align} Given a walk $h$ beginning in $\ttb(h) = m_\lambda^{-1}v^{-1}$, its endpoint is in $\tte(h) = m_\lambda^{-1} v^{-1} (\pi_j^\vee)^{-1}$, so $m_{\varpi(h)} = \tte(h)^{-1} = \pi_j^\vee v m_\lambda$ means that $\varpi(h)=\pi_j^\vee v\lambda = x^{\omega_j}v_j^{-1}v\lambda = v_j^{-1}v\lambda + \omega_j$. Thus we may also write \begin{align*} E_{\omega_j}P_\lambda &= \sum_{v\in W^\lambda} \left( \prod_{a^\vee\in m_\lambda^{-1} \calL\left(v^{-1}, v_\lambda^{-1}\right)} t_{a^\vee}^{\frac12} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)} t_{a^\vee}^{-1}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}}\right) E_{v_j^{-1}v\lambda+\omega_j}. \end{align*} \subsection{Compression of the Pieri formula} Formula~\eqref{eqn.PieriPP} is a sum over $|W^\lambda|$ walks, and many of the walks have the same weight. By imposing a condition on the final direction of the walks and modifying the coefficients appropriately, the formula can be compressed to contain the minimal number of terms. \begin{corollary}\label{cor.PPshort} Let $\omega_j$ be a minuscule weight, and $\lambda$ be a dominant weight. Then with the same notation as~\eqref{eqn.bhqt} and~\eqref{eqn.ehqt}, $$P_{\omega_j}P_\lambda = \sum_{{h\in \Gamma((\pi_j^\vee)^{-1},(zm_\lambda)^{-1})} \atop {z\in W^\lambda,\ \ttd(h) \in W^{\omega_j}}} b_{\left(\ttb(h),x^{-w_0\lambda}\right)} e_{\left(x^{\ttwt(h)}w_0,\tte(h)w_{\omega_j}\right)} P_{-w_0\ttwt(h)}, $$ where the sum is over the set of alcove walks of type $(\pi_j^\vee)^{-1}$, beginning in $m_\lambda^{-1} z^{-1}$ for $z\in W^\lambda$, and has final direction $\ttd(h) \in W^{\omega_j}$. \end{corollary} \begin{remark}{\em $\ $ \begin{enumerate} \item The {\em initial direction} of $h$ is defined by $X^{-w_0\lambda}T_{\tti(h)}$. It follows from equation~\eqref{eqn.tteh} that the final direction condition $\ttd(h) \in W^{\omega_j}$ is equivalent to the initial direction condition $\tti(h) \in W^{-w_0\omega_j}$. \item Since the walks which are under consideration do not have folds, then the condition that the walks begin in the alcoves $m_\lambda^{-1}z^{-1}$ for $z \in W^\lambda$ is enough to guarantee that the walks are contained in the closure of the dominant chamber. \end{enumerate} }\end{remark} \proof If a walk $h$ begins in the alcove \begin{equation}\label{eqn.ttbh} \ttb(h) = m_\lambda^{-1}z^{-1} = m_{-w_0\lambda} z^{-1} = x^{-w_0\lambda} \left(v_\lambda z^{-1}\right), \end{equation} for some $z\in W^\lambda$, then $h$ has initial direction $\tti(h) = v_\lambda z^{-1}$. And since $h$ is a walk of type $(\pi_j^\vee)^{-1}$, then $h$ ends in the alcove \begin{equation}\label{eqn.tteh} \tte(h) = \ttb(h)(\pi_j^\vee)^{-1} = x^{-w_0\lambda}\tti(h) (x^{\omega_j}v_j^{-1})^{-1} = x^{-w_0\lambda} x^{-\tti(h)v_j\omega_j} \left(\tti(h) v_j\right), \end{equation} so $h$ has final direction $\ttd(h) = \tti(h) v_j$, and weight \begin{equation}\label{eqn.ttwth} \ttwt(h) = -w_0\left(\lambda + w_0^{-1}\tti(h) v_j\omega_j\right) = -w_0\left(\lambda + w_0^{-1}\ttd(h)\omega_j\right). \end{equation} The weights of the walks $h\in \Gamma((\pi_j^\vee)^{-1}, (vm_\lambda)^{-1})$ are not distinct, since the stabilizer of $\omega_j$ is not trivial. The idea is to group together the walks with the same weight by factoring $\tti(h) = vw$ for $v\in W^{-w_0\omega_j}$ and $w\in W_{-w_0\omega_j}$. We make one more observation before proceeding with the calculation. Suppose $h$ is a walk beginning in an alcove $x^{-w_0\lambda} u$, for some $u\in W_0$, which is not contained in the dominant chamber, and that the hyperplane $\ttH_{-\alpha_k^\vee}$ separates the alcove from the dominant chamber. Notice that one of the factors of the coefficient \begin{equation}\label{eqn.startoutsideC} b_{\left(\ttb(h),x^{-w_0\lambda}\right)} = \prod_{a^\vee\in\calL\left(\ttb(h), x^{-w_0\lambda}\right)} t_{a^\vee}^{\frac12} \frac{1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)} t_{a^\vee}^{-1}} {1-q^{\shf(-a^\vee)}t^{\hgt(-a^\vee)}}, \end{equation} is $\displaystyle t_k^{\frac12} \frac{1-t_k t_k^{-1}}{1-t_k} =0$, so that $b_{\left(\ttb(h),x^{-w_0\lambda}\right)} =0$ in this case. Therefore, although such walks are not counted in equation~\eqref{eqn.PieriPP}, it will still make sense to include them in the following calculation. Consider the bijections $$\bar{\quad} : W_{-w_0{\omega_j}} \rightarrow W_{\omega_j}: w \mapsto \bar{w}= v_j^{-1} w v_j, \quad\hbox{and}\quad \bar{\quad} :W^{-w_0{\omega_j}} \rightarrow W^{\omega_j} : v \mapsto \bar{v}=vv_j.$$ With this notation, then $\tti(h) = vw$ if and only if $\ttd(h) = \bar{v}\bar{w}$. By equations~\eqref{eqn.ttbh} and~\eqref{eqn.tteh}, the walks $h$ have the same weight only if $$\ttb(h) \in \calB_v = \left\{x^{-w_0\lambda}vw \mid w\in W_{-w_0\omega_j}\right\}, $$ for a fixed $v\in W^{-w_0\omega_j}$, or equivalently, if $$\tte(h) \in \calE_{\bar{v}} = \left\{x^{\ttwt(h)}\bar{v}\bar{w} \mid \bar{w}\in W_{\omega_j}\right\}, $$ for a fixed $\bar{v}\in W^{\omega_j}$. Since $b_{(x^{-w_0\lambda}vw, x^{-w_0\lambda}v)} = b_{(x^{\ttwt(h)}\bar{v}\bar{w}, x^{\ttwt(h)}\bar{v})}$, then \begin{align*} \sum_{h: \tte(h) \in \calE_{\bar{v}}} & b_{\left(\ttb(h),x^{-w_0\lambda}\right)} e_{\left(x^{\ttwt(h)}w_0, \tte(h)\right)} \mathbf{1} \\ &= b_{\left(x^{-w_0\lambda}v,x^{-w_0\lambda}\right)} e_{\left(x^{\ttwt(h)}w_0, x^{\ttwt(h)}\bar{v} w_{\omega_j}\right)} \sum_{\bar{w}\in W_{\omega_j}} b_{\left(x^{\ttwt(h)}\bar{v}\bar{w}, x^{\ttwt(h)}\bar{v}\right)} e_{\left(x^{\ttwt(h)}\bar{v} w_{\omega_j}, x^{\ttwt(h)}\bar{v}\bar{w} \right)} \mathbf{1}\\ &= b_{\left(x^{-w_0\lambda}v,x^{-w_0\lambda}\right)} e_{\left(x^{\ttwt(h)}w_0, x^{\ttwt(h)}\bar{v} w_{\omega_j}\right)} \left( x^{\ttwt(h)}\bar{v} \cdot \sum_{\bar{w}\in W_{\omega_j}} b_{(\bar{w}, 1)} e_{(w_{\omega_j}, \bar{w})} \right)\mathbf{1}\\ &= b_{\left(x^{-w_0\lambda}v,x^{-w_0\lambda}\right)} e_{\left(x^{\ttwt(h)}w_0, x^{\ttwt(h)}\bar{v} w_{\omega_j}\right)} t_{w_{\omega_j}}^{-\frac12} W_{\omega_j}(t)\mathbf{1}, \end{align*} by Lemma~\ref{lem.Poincare}. Therefore, we can restrict to walks which have final direction $\ttd(h) = \bar{v} \in W^{\omega_j}$, $$ P_{\omega_j}P_\lambda = \sum_{h: \ttd(h) \in W^{\omega_j}} b_{\left(\ttb(h),x^{-w_0\lambda}\right)} e_{\left(x^{\ttwt(h)}w_0, \tte(h) w_{\omega_j}\right)} P_{-w_0\ttwt(h)}.$$ \qed \subsection{Type $A$ Pieri formulas and partitions} In the case of the type $A_n$ root systems, Macdonald gave Pieri formulas for symmetric Macdonald polynomials in terms of partitions. This section is a brief sketch of the relation between the alcove walk combinatorics in Corollary~\ref{cor.PPshort} and the partition combinatorics in Macdonald's formula~\cite[(6.24)(iv)]{M88}, reproduced below. In the type $A_n$ setting, the parameters $t=t_0=\cdots = t_n$ are necessarily all equal. The correspondence between partitions with at most $n+1$ parts and type $A_n$ dominant weights is as follows: \begin{eqnarray*} \hbox{column partitions $(1^j)$} &\leftrightarrow& \hbox{minuscule weights $\omega_j$ for $j=1,\ldots,n$,}\\ \hbox{column partition $(1^{n+1})$} &\leftrightarrow& \hbox{weight $0$,} \end{eqnarray*} so $$e_j = s_{(1^j)} = P_{\omega_j}(q,t) \hbox{ is the $j$th elementary symmetric polynomial.}$$ Viewing a partition $\kappa$ as a collection of boxes justified to the top and the left, the {\em arm-length} and {\em leg-length} of a box $s$ in the $i$th row and $j$th column of $\kappa$ are \begin{align*} a_\kappa(s) &= \kappa_i-j = \textrm{ number of boxes to the east of } s,\\ l_\kappa(s) &= \kappa_j'-i = \textrm{ number of boxes to the south of } s. \end{align*} If $\kappa \supseteq \lambda$, let $C_{\kappa-\lambda}$ (respectively $R_{\kappa-\lambda}$) be the set of columns (respectively rows) of $\kappa$ that intersect the skew partition $\kappa-\lambda$. \begin{theorem} \cite[(6.24)(iv)]{M88} \label{eqn.MacPP} $$P_{\omega_j}P_\lambda = \sum_\kappa \left( \prod_{s\in C_{\kappa-\lambda} - R_{\kappa-\lambda}} \frac{1-q^{a_\lambda(s)+1}t^{l_\lambda(s)}} {1-q^{a_\kappa(s)+1}t^{l_\kappa(s)}} \frac{1-q^{a_\kappa(s)}t^{l_\kappa(s)+1}} {1-q^{a_\lambda(s)}t^{l_\lambda(s)+1}} \right) P_\kappa, $$ where the sum is over the set of partitions $\kappa$ with at most $n+1$ parts, obtained from $\lambda$ by adding $j$ boxes, with no two in the same row. \qed \end{theorem} To draw the connection between walks and partitions, we begin by considering Corollary~\ref{cor.PPshort} with parameters $t_i=t$: {\em \noindent{\bf Corollary 4.8.} \begin{align*} &P_{\omega_j}P_\lambda\\ &=\sum_h \left( \prod_{\alpha^\vee \in \calL(\tti(h))} \left(x^\lambda\!\cdot\! \frac{1-q^{\shf(-\alpha^\vee)}t^{\hgt(-\alpha^\vee)-1} } {1-q^{\shf(-\alpha^\vee)}t^{\hgt(-\alpha^\vee)}}\right) \!\! \left(x^{\ttwt(h)}\!\cdot\! \frac{1-q^{\shf(-\alpha^\vee)}t^{\hgt(-\alpha^\vee)+1} } {1-q^{\shf(-\alpha^\vee)}t^{\hgt(-\alpha^\vee)}}\right) \right) P_{\ttwt(h)}, \end{align*} where the sum is over the set of alcove walks of type $\pi_j^\vee$, beginning in the alcoves $x^{\lambda}W^{\omega_j}$. } \proof First, notice that $\calL\left(\ttb(h),x^{-w_0\lambda}\right) = x^{-w_0\lambda}\calL\left(1,\tti(h)\right)$, while \begin{align*} \calL\left(x^{\ttwt(h)}w_0, \tte(h)w_{\omega_j}\right) &=\calL\left(x^{\ttwt(h)}, \tte(h) w_{\omega_j}w_0^{-1}\right) =x^{\ttwt(h)} \calL\left(1, \ttd(h)v_j^{-1}\right) =x^{\ttwt(h)} \calL\left(1, \tti(h)\right), \end{align*} so the coefficients in Corollary~\ref{cor.PPshort} can be rewritten as \begin{align*} b_{\left(\ttb(h),x^{-w_0\lambda}\right)} &e_{\left(x^{\ttwt(h)}w_0, \tte(h) w_{\omega_j}\right)}\mathbf{1} = \prod_{\alpha^\vee\in \calL(1,\tti(h))} \left(x^{-w_0\lambda}\cdot \frac{1-Y^{-\alpha^\vee}t^{-1} } {1-Y^{-\alpha^\vee}} \right) \left(x^{\ttwt(h)}\cdot \frac{1-Y^{-\alpha^\vee}t } {1-Y^{-\alpha^\vee}}\right)\mathbf{1}, \end{align*} where $x^{-w_0\lambda} = x^{\ttwt(h) + \ttd(h)\omega_j}$ by~\eqref{eqn.ttwth}. Next, consider the action of the automorphism $-w_0:\fh_\bbR^* \rightarrow \fh_\bbR^*$, which acts on the (type $A_n$) lattice $\fh_\bbZ^*$ by sending $\omega_j \mapsto \omega_{n+1-j}$. Hence, this automorphism also acts on the set of alcoves, with action given by $$-w_0(x^\beta w\ttA) = x^{-w_0\beta} w_0^{-1}ww_0\ttA, \qquad \hbox{for } \beta\in \fh_\bbZ^*, w\in W_0.$$ Since $\lambda$ is a dominant weight, then $$(zm_\lambda)^{-1} = m_\lambda^{-1}z^{-1} = m_{-w_0\lambda}z^{-1} = x^{-w_0\lambda}v_{-w_0\lambda}^{-1} z^{-1} = x^{-w_0\lambda} (zv_{-w_0\lambda})^{-1},$$ where $zv_{-w_0\lambda}\in W^{-w_0\omega_j}$. The Pieri formula stated above is the result of applying $-w_0$ to the set of alcoves. The last observation is, while it seems that the condition that walks must be contained in the dominant chamber is dropped in this version of Corollary~\ref{cor.PPshort}, in fact if the walk $h$ begins outside the the dominant chamber, then the coefficient $b_h =0$ for the same reason as explained in~\eqref{eqn.startoutsideC}. So the Pieri formula remains unchanged even if the set of walks on which the sum runs over is expanded to include those which begin outside the dominant chamber. \qed In what follows, we fix $\lambda\in (\fh_\bbZ^*)^+$ to be a regular dominant weight, so that when viewed as a partition, $\lambda$ does not have equal parts. Also fix a minuscule weight $\omega_j$. Let $\calK$ be the lattice of partitions which are obtained from $\lambda$ by adding $j$ boxes with no two in the same row. Given $\kappa\in \calK$, any box not contained in $\lambda$ will be called an {\em added box}. Since we assume that $\lambda$ has no equal parts, $\calK$ may be viewed as the set of $j$-subsets of $[n+1]=\{1,\ldots, n+1\}$, identifying $\kappa$ with the subset $\{r_1,\ldots, r_j\}$ if the $j$ boxes were added in rows $r_1,\ldots, r_j$ of $\lambda$ to obtain $\kappa$. The order in $\calK$ is defined by the covering relation: $\kappa$ covers $\theta$ if and only if $\theta$ can be obtained from $\kappa$ by moving one of the added boxes down one row. Thus, the unique maximal element $\hat{1}$ in $\calK$ is $\lambda$ with boxes added in rows $1$ through $j$, and the unique minimal element $\hat{0}$ in $\calK$ is $\lambda$ with boxes added in rows $(n+1)-(j-1)$ through $n+1$. Further, in the Hasse diagram representing the lattice $\calK$, if $\kappa$ covers $\theta$ and a box was moved from row $i$ to row $i+1$, then the edge between $\kappa$ and $\theta$ is labeled $s_i$. Recall $\Gamma(\pi_j^\vee, x^\lambda W^{\omega_j})$ is the set of alcove walks of type $\pi_j^\vee$ which begin in the alcoves $x^\lambda W^{\omega_j}$. \begin{lemma} Assume $\lambda$ is a regular dominant weight, or equivalently, a partition with no equal parts. There is a bijection \begin{eqnarray*} \calK &\longleftrightarrow & \Gamma\left(\pi_j^\vee, x^\lambda W^{\omega_j}\right) \\ \kappa &\leftrightarrow& h_\kappa \end{eqnarray*} such that \begin{enumerate} \item $h_\kappa$ is the walk with $\ttwt(h_\kappa) = \kappa$, \item if a shortest path from $\kappa$ to $\hat{1}$ in $\calK$ traverses the sequence of edges $s_{i_1},\ldots, s_{i_\ell}$, then $\tti(h) = s_{i_1}\cdots s_{i_\ell}$. \item Moreover, for each pair $\kappa \leftrightarrow h_\kappa$, there is a bijection \begin{eqnarray*} C_{\kappa-\lambda}-R_{\kappa-\lambda} & \longleftrightarrow & \calL(\tti(h_\kappa))\\ \Box_{\alpha^\vee} & \leftrightarrow & \alpha^\vee \end{eqnarray*} satisfying $$ l_\kappa(\Box_{\alpha^\vee}) = \hgt(-x^{\kappa}\alpha^\vee), \qquad\hbox{and}\qquad a_\kappa(\Box_{\alpha^\vee}) = \shf(-x^{\kappa}\alpha^\vee). $$ \end{enumerate} \end{lemma} \proof \noindent (1) Since $\lambda$ is a regular weight, then no matter which $j$ rows we choose to add boxes, the resulting shape corresponds to a partition, hence the lattice $\calK$ has $\binom{n+1}{j}$ elements. In the type $A_n$ root system, $W_0 = \fS_{n+1}$ and $W_{\omega_j} \cong \fS_j \times \fS_{n+1-j}$, thus $|W^{\omega_j}| = (n+1)!/j!(n+1-j)! = \binom{n+1}{j}.$ So the sets in question have the same cardinality. To each partition $\kappa\in \calK$, we assign the walk $h_\kappa$ such that $\ttwt(h_\kappa) = \kappa$. \noindent (2) For $i=1,\ldots, n+1$, adding a box in row $i$ of the partition $\lambda$ yields the composition corresponding to the weight $\lambda + \omega_i-\omega_{i-1}$, where we use the notation $\omega_0 = \omega_{n+1}=0$. Hence, adding a box in rows $1$ through $i$ of $\lambda$ yields the partition $\lambda + \omega_i$. If $\kappa$ is the partition obtained from $\lambda$ by adding $j$ boxes in rows $r_1,\ldots, r_j$ (assuming $r_1< r_2< \cdots < r_j$), then $\kappa = \lambda + \sum_{i=1}^j (\omega_{r_i}-\omega_{r_i-1})$. Observe that $\hat{1} = \lambda+\omega_j$, while $\hat{0}=\lambda+v_j\omega_j$. A minimal path in $\calK$ from $\hat{1}$ to $\kappa$ can be constructed as follows: begin by traveling along the edges $s_j, s_{j+1}, \ldots, s_{r_j-2}, s_{r_j-1}$ to reach the partition in which the last box in row $j$ of $\hat{1}$ has been moved to row $r_j$, then travel along the edges $s_{j-1}, s_{j}, \ldots, s_{r_{j-1}-1}$ to reach the partition in which the last box in row $j-1$ of the previous partition has been moved to row $r_{j-1}$, and so on, and end by traveling along the edges $s_1, s_2, \ldots, s_{r_1-1}$ to reach the partition $\kappa$. Inverting these $j$ sequences of edges gives a path from $\kappa$ to $\hat{1}$. Let $$\sigma_k = s_{r_k-1} s_{r_k-2} \cdots s_{k+1}s_k, \hbox{ for } k=1,\ldots, j, \qquad \hbox{ and let } \sigma = \sigma_1 \sigma_2 \cdots \sigma_j.$$ Then $\lambda+ \tti(h)\omega_j = \ttwt(h_\kappa) = \kappa = \lambda + \sigma\omega_j.$ Since this is a path of minimal length in $\calK$, then the above factorization is a reduced expression for $\tti(h) = \sigma \in W^{\omega_j}$. \noindent (3) By equation~\eqref{calLw}, $$ \begin{array}{l} \calL(\tti(h_\kappa)) = \Big\{ \alpha_{r_1-1}^\vee,\ (s_{r_1-1}\alpha_{r_1-2}^\vee),\ \ldots,\ (s_{r_1-1} s_{r_1-2} \cdots s_2\alpha_1^\vee),\\ \\ \qquad\qquad\qquad\qquad (\sigma_1\alpha_{r_2-1}^\vee),\ (\sigma_1 s_{r_2-1}\alpha_{r_2-2}^\vee),\ \ldots,\ (\sigma_1 s_{r_2-1} s_{r_2-2} \cdots s_3\alpha_2^\vee),\\ \\ \qquad\qquad\qquad\qquad \qquad \qquad\qquad\qquad\qquad \vdots \\ \\ \qquad\qquad\qquad\qquad (\sigma_1\cdots \sigma_{j-1}\alpha_{r_j-1}^\vee),\ \ldots,\ (\sigma_1\cdots \sigma_{j-1}s_{r_j-1} s_{r_j-2} \cdots s_{j+1}\alpha_j^\vee)\Big\}.\end{array} $$ In contrast, $C_{\kappa-\lambda}- R_{\kappa-\lambda}$ is the set of boxes in $\kappa$ which are in columns containing an added box, and in rows not containing an added box. For example, when $\kappa=\hat{1}$, $C_{\kappa-\lambda}- R_{\kappa-\lambda} = \emptyset$. Consider the box that was added to $\lambda$ in row $r_k$ (and column $c_k$). Column $c_k$ contains $(r_k-1) - (k-1)$ boxes which are in $C_{\kappa-\lambda}- R_{\kappa-\lambda}$. Altogether, $$|C_{\kappa-\lambda}- R_{\kappa-\lambda}| = \sum_{k=1}^j (r_k-k) = \sum_{k=1}^j\ell(\sigma_k) = \ell(\sigma) = \ell(\tti(h_\kappa)) = |\calL(\tti(h_\kappa))|,$$ noting that column $c_k$ of $\kappa$ has $\ell(\sigma_k)$ boxes. Thus, define the correspondence $C_{\kappa-\lambda}- R_{\kappa-\lambda} \leftrightarrow \calL(\tti(h_\kappa))$ as follows: to the boxes in column $c_k$ which belong to $C_{\kappa-\lambda}- R_{\kappa-\lambda}$, assign in order from bottom to top, the coroots $$(\sigma_1 \cdots \sigma_{k-1} \alpha_{r_k-1}^\vee),\ (\sigma_1 \cdots \sigma_{k-1}s_{r_k-1}\alpha_{r_k-2}^\vee),\ \ldots,\ (\sigma_1 \cdots \sigma_{k-1} s_{r_k-1} s_{r_k-2} \cdots s_{k+1}\alpha_k^\vee).$$ Using this particular choice of factorization for $\tti(h_\kappa)$, and noting that in the type $A_n$ root system $$s_i\alpha_k^\vee =\begin{cases} \alpha_k^\vee + \alpha_i^\vee, &\hbox{if } i=k\pm1,\\ -\alpha_k^\vee, &\hbox{if } i=k,\\ \alpha_k^\vee, &\hbox{otherwise,} \end{cases}$$ we see that if the coroot $$\gamma^\vee = \left(\sigma_1\cdots \sigma_{k-1}\right)\left( s_{r_k-1}s_{r_k-2}\cdots s_{r_k-m+1}\right)\alpha_{r_k-m}^\vee $$ is the one which corresponds to the box $\Box_{\gamma^\vee}$ in column $c_k$ and row $t$ of $\kappa$ under the above bijection, then $$\gamma^\vee = \alpha_{r_k-1}^\vee + \alpha_{r_k-2}^\vee + \cdots + \alpha_{r_k-m}^\vee + \cdots + \alpha_{t+1}^\vee +\alpha_{t}^\vee.$$ Hence it follows that \begin{align*} \hgt(-x^{\kappa}\gamma^\vee) &= \hgt(\gamma^\vee) = r_k-t = l_\kappa(\Box_{\gamma^\vee}),\\ \shf(-x^{\kappa}\gamma^\vee) &= \left\langle \kappa, \gamma^\vee \right\rangle = \sum_{i=1}^{n} d_i\left\langle \omega_i, \gamma^\vee \right\rangle = a_\kappa(\Box_{\gamma^\vee}), \end{align*} where $d_i$ in $\kappa = \sum_{i=1}^n d_i\omega_i$ is the number of columns in $\kappa$ with height $i$. \qed Under the assumption that $\lambda$ has no equal parts, then $a_\kappa(s) = a_\lambda(s)$ and $l_\kappa(s) = l_\lambda(s)+1$, for all $\kappa \in\calK$. On the other hand, $\hgt(-x^\lambda\alpha^\vee) = \hgt(\alpha^\vee) = \hgt(-x^{\ttwt(h)}\alpha^\vee)$. Also, $\shf(-x^{\ttwt(h)}\alpha^\vee) =\langle \ttwt(h), \alpha^\vee \rangle = \langle \lambda + \tti(h)^{-1}\omega_j, \alpha^\vee \rangle = \shf(-x^\lambda\alpha^\vee) + \langle \omega_j, \tti(h)\alpha^\vee \rangle$ for all $\alpha^\vee \in \calL(\tti(h))$, and since $\tti(h)^{-1}\alpha^\vee$ is of the form $s_j s_{i_\ell-1}\cdots s_{i_\ell-m+1}\alpha_{i_\ell-m}^\vee$, then $\langle \omega_j, \tti(h)\alpha^\vee\rangle = -1$, and so $\shf(-x^{\ttwt(h)}\alpha^\vee) = \shf(-x^\lambda\alpha^\vee)-1$. It is now clear that Theorem~\ref{eqn.MacPP} and Corollary~\ref{cor.PPshort} are identical, in the case that $\lambda$ is a regular dominant weight (a partition with no equal parts). \begin{remark} {\em \ \begin{enumerate} \item In the case when $\lambda$ is not a regular dominant weight (a partition having equal parts), then $|\calL(\tti(h))| \geq C_{\kappa-\lambda}-R_{\kappa-\lambda}$, and there are many canceling pairs of factors in the alcove walk formula which do not appear in the tableau formula. Note that $a_\kappa(s) = a_\lambda(s)$, but $l_\kappa(s) \geq l_\lambda(s)+1$ in this case. \item The Pieri case involving multiplication with $P_{j\omega_1}$, where $P_{j\omega_1}(q,q) = s_{(j)} = h_j$ is the $j$th complete symmetric polynomial, is more involved, and is likely related to the compression phenomenon discussed in~\cite{Le}. \end{enumerate}}\end{remark} \subsection{Hall-Littlewood polynomials} \label{sec.specialization} The Hall-Littlewood polynomials $P_\mu(t)=P_\mu(0,t)$ are the symmetric Macdonald polynomials under the specialization of the parameter $q=0$. In this section, we assume all parameters $t_i = t$ for $i=0,\ldots, n$. Combinatorial formulas for Hall-Littlewood polynomials were given in ~\cite[Theorem 1.1]{Sc06}, and a Pieri formula was given in~\cite[p.\!\! 217]{M88}. We restate the alcove walk formula \cite[Theorem 4.9]{R06} for the Littlewood-Richardson rule for $P_\lambda(t)$ in the notation of this paper. Recall that the final direction $\ttd(h)$ is defined by $X^{\tte(h)} = X^{\ttwt(h)}T_{\ttd(h)}$, and the initial direction $\tti(h)$ is defined by $X^{\ttb(h)} = X^{\lambda} T_{\tti(h)}$. Let $\ttf(h)$ be the number of folds in $h$, and $\ttf_0(h)$ be the number of folds in $h$ touching a wall of $\ttC$. A {\em positively folded walk} is a walk with no negative folds. \begin{theorem}\cite[Theorem 4.9]{R06}. Also see~\cite[Theorem 1.3]{Sc06}. \label{thm.HL1} Let $\mu,\lambda \in (\fh_\bbZ^*)^+$, and fix a reduced expression for $x^\mu$. Then $$P_\mu(t) P_\lambda(t) = \sum_{v\in W^\mu} \sum_{h\in \Gamma^{\ttC-\rho}_+ \left(\vec{x}^{\mu}, x^\lambda v\right)} t^{\frac12(\ell(\tti(h)) + \ell(\ttd(h)) - \ttf(h))} (1-t)^{\ttf(h)-\ttf_0(h)} P_{\ttwt(h)}(t), $$ where the sum is over all positively folded alcove walks of type $\vec{x}^\mu$ beginning in $x^\lambda v$ for $v\in W^\mu$, contained in the shifted dominant chamber $\ttC-\rho$. \qed \end{theorem} As a corollary to Theorem~\ref{thm.PP-P}, we derive a different version of the above result. The main differences are that the walks under consideration are of type $\vec{m}_\mu$ instead of $\vec{x}^\mu$, and the walks are contained in $\overline{\ttC}$ instead of $\ttC$ shifted by $-\rho$. \begin{corollary}\label{cor.HL2} Let $\mu,\lambda \in (\fh_\bbZ^*)^+$, and fix a reduced expression for $m_\mu$. Then $$P_\mu(t) P_\lambda(t) = t^{-\frac12\ell(v_\mu)} \sum_{v\in W^\lambda} \sum_{h \in \Gamma^{\overline{\ttC}}_+ \left(\vec{m}_\mu, x^\lambda v^{-1}\right)} t^{\frac12( \ell(\tti(h))+\ell(\ttd(h))-\ttf(h))} (1-t)^{\ttf(h)-\ttf_0(h)} \frac{W_{\ttwt(h)}(t)}{W_\mu(t)} P_{\ttwt(h)}(t),$$ where the sum is over all positively folded alcove walks of type $\vec{m}_\mu$ beginning in $x^\lambda v^{-1}$ for $v\in W^\lambda$, contained in $\overline\ttC$. \end{corollary} \proof Keeping the notation from Theorem~\ref{thm.PP-P}, let $$a_h(q,t) = (-1)^{\phi_{grey}^-(h)} b_h e_h \prod_{k=1}^r c_{k(h)}. $$ The first step is to show that the walks $h\in \Gamma_2^\ttC(\vec{m}_\mu^{-1}, (vm_\lambda)^{-1})$ for which $a_h(0,t) \neq 0$ are precisely those whose folds must be positive and grey. In this section, these are referred to as the `grey positively folded walks'. The monomial expansion of $E_\mu$ is a sum over the set of walks $\Gamma(\vec{m}_\mu)$. There is a unique walk $l\in\Gamma(\vec{m}_\mu)$ without folds, and since $\mu$ is a dominant weight, then every step of $l$ is a positive crossing because $l$ is contained in the dominant chamber. The walks in $\Gamma(\vec{m}_\mu)$ are generated by folding $l$ in all possible ways, so if a walk $p\in \Gamma(\vec{m}_\mu)$ has a fold, then it has at least one negative fold. At $q=0$, Theorem~\ref{thm.Xtau-X} gives $$E_\mu(0,t) = t_{v_\mu^{-1}}^{\frac12} X^\mu, \quad\hbox{if $\mu$ is dominant}.$$ In other words, if $h\in \Gamma_2^\ttC(\vec{m}_\mu^{-1}, (vm_\lambda)^{-1})$ has a black fold, then $p(h)$ has at least one negative black fold, so the last row of~\eqref{eqn.verylonglist} gives $a_h(0,t)=0$ in this case. Thus, any walk with a black fold does not contribute to the expansion of $P_\mu(t)P_\lambda(t)$ in the basis of Hall-Littlewood polynomials. Another way to put this is if $h$ survives setting $q=0$, then $p(h)$ is the unfolded walk of type $\vec{m}_\mu^{-1}$ ending in $1$. Accordingly, $\epsilon_k(h) = -1$ for $k=1,\ldots, r$, since $\mu$ is a dominant weight. Using this observation, if $h$ has a negative grey fold (so the fold must be against an affine wall), then the fourth row of~\eqref{eqn.verylonglist} gives $a_h(0,t)=0$ in this case. Since the walks now under consideration do not have negative grey folds, then the expansion of the product of Hall-Littlewood polynomials does not involve negative signs. Moreover, every fold is grey, so we may forget the colour of the folds for the remainder of this section. The remaining coefficients in Theorem~\ref{thm.PP-P} simplify: \begin{align*} b_h &= t^{\frac12(\ell(x^{-w_0\lambda}) - \ell(\ttb(h)) )} = t^{\frac12 \ell(\tti(h))},\\ e_h &= t^{-\frac12(\ell(\tte(h)) - \ell(m_{\ttwt(h)}) )} = t^{\frac12(\ell(\ttd(h)) - \ell(v_{\ttwt(h)}) )},\\ c_{k(h)} &= \begin{cases} t^{-1/2}, & \hbox{for (grey) folds touching a wall of }\ttC,\\ t^{-1/2}(1-t), & \hbox{for (grey) folds touching an affine hyperplane.} \end{cases} \\ &= t^{-\frac12 \ttf(h)} (1-t)^{\ttf(h)-\ttf_0(h)}. \end{align*} One last observation to make is that the action of the automorphism $-w_0: \fh_\bbR^* \rightarrow \fh_\bbR^*$ acts on the lattice $\fh_\bbZ^*$ by permuting $\omega_1,\ldots, \omega_n$, so $-w_0$ acts on the set of alcoves, given by $$-w_0(x^\beta w \ttA) = x^{-w_0\beta} w_0^{-1} w w_0\ttA, \qquad\hbox{for } \beta\in \fh_\bbZ^*, w\in W_0.$$ Since $\lambda$ is a dominant weight, then $$(vm_\lambda)^{-1} = m_\lambda^{-1}v^{-1} = m_{-w_0\lambda} v^{-1} = x^{-w_0\lambda} v_{-w_0\lambda}^{-1} v^{-1} = x^{-w_0\lambda} v_\lambda v^{-1}. $$ Therefore, if $h$ is a walk of type $\vec{m}_\mu^{-1} = \vec{m}_{-w_0\mu}$ beginning in $(vm_\lambda)^{-1}$, then $-w_0(h)$ is a walk of type $\vec{m}_\mu$ beginning in $x^\lambda w_0^{-1} v_\lambda v^{-1}w_0.$ Moreover, $\{v_\lambda v^{-1} \mid v\in W^\lambda\} = \{y^{-1} \mid y\in W^{-w_0\lambda} \}$, so $\{ w_0 v_\lambda v^{-1} w_0^{-1} \mid v\in W^\lambda \} = \{ v^{-1} \mid v\in W^\lambda\}$. Putting all the observations together, $$P_\mu(t) P_\lambda(t) = \sum_{v\in W^\lambda} \sum_{h \in \Gamma_+^{\overline{\ttC}} \left(\vec{m}_\mu, x^\lambda v^{-1}\right)} t^{\frac12(\tti(h)+\ttd(h)-\ttf(h))} (1-t)^{\ttf(h)-\ttf_0(h)} \frac{ t^{-\frac12}_{v_{\ttwt(h)}} t_{w_{\ttwt(h)}}^{-\frac12}W_{\ttwt(h)}(t)} {t_{w_\mu}^{-\frac12}W_\mu(t)} P_{\ttwt(h)}(t),$$ where the sum is over the set of positively folded walks of type $\vec{m}_\mu$, beginning in $x^\lambda v^{-1}$ for $v\in W^\lambda$, which are contained in $\overline{\ttC}$. To finish, note that $t^{-\frac12}_{v_{\ttwt(h)}}t_{w_{\ttwt(h)}}^{-\frac12}t_{w_\mu}^{\frac12} = t_{w_0}^{-\frac12} t_{w_\mu}^{\frac12} = t_{v_\mu}^{-\frac12}.$ \qed \begin{example}{\em See section~\ref{sec.eg} for more details on the alcove picture of type $\fsl_3$. Let $\varphi =\omega_1+\omega_2$. Using either Theorem~\ref{thm.HL1} or Corollary~\ref{cor.HL2}, \begin{align*} P_\varphi(t)^2 = P_{2\varphi}(t) &+ (1+t) P_{3\omega_1}(t) + (1+t)P_{3\omega_2}(t)\\ &+ (2 + t(1-t)) P_\varphi(t) + (1+2t+2t^2+t^3)P_0(t). \end{align*} We remark that 16 walks were used in the first formula, and seven walks were used in the second formula. $\hfill\diamond$ }\end{example} \section{Mainly Lots of Examples}\label{sec.eg} \subsection{Type $\fsl_2$} The root datum is $$\fh_\bbZ^* =\bbZ\omega,\quad R = \{\pm\alpha\},\quad \fh_\bbZ = \bbZ\omega^\vee,\quad R^\vee = \{\pm\alpha^\vee\},$$ with $\alpha = 2\omega$ and $\alpha^\vee = 2\omega^\vee$. Let $s_0^\vee = x^\alpha s_1^\vee$ and $\pi^\vee = x^\omega s_1^\vee$. The extended affine Weyl group $W^\vee$ is generated by $s_1^\vee, s_0^\vee, \pi^\vee$, subject to the relations \begin{equation}\label{eqn.presn1} (\pi^\vee)^2 =1,\qquad \pi^\vee s_0^\vee = s_1^\vee \pi^\vee,\qquad (s_i^\vee)^2 =1,\hbox{ for } i=0,1. \end{equation} Alternatively, \begin{equation}\label{eqn.presn2} W^\vee = \{x^{k\omega} w \mid k\in\bbZ, w\in \{1, s_1^\vee\} \}. \end{equation} The following is the alcove picture for the extended affine Weyl group $W^\vee$, showing the correspondence between the alcoves and the elements of $W^\vee$. The periodic orientation is indicated by $\scriptstyle{+}$ and $\scriptstyle{-}$ on either side of the hyperplanes. The two ways of indexing the alcoves correspond to the two presentations~\eqref{eqn.presn1} and~\eqref{eqn.presn2} for $W^\vee$. $$\beginpicture \setcoordinatesystem units <1.8cm,1cm> \setplotarea x from -4 to 4.5, y from -1.5 to 2.8 \put{$\bullet$} at 0 0 {\small \put{$\ttH_{-\alpha^\vee+2d}$}[t] at 2 -0.9 \put{$\ttH_{\alpha^\vee}$}[t] at 0 -0.9 \put{$\ttH_{\alpha^\vee+2d}$}[t] at -2 -0.9 \put{$\ttH_{-\alpha^\vee+d}$}[t] at 1 -0.9 \put{$\ttH_{-\alpha^\vee+3d}$}[t] at 3 -0.9 \put{$\ttH_{\alpha^\vee+d}$}[t] at -1 -0.9 \put{Sheet $1$} at -4.2 0 \plot -3.5 0 4.5 0 / \plot -3 0.7 -3 -0.4 / \plot -2 0.7 -2 -0.4 / \plot -1 0.7 -1 -0.4 / \plot 0 0.7 0 -0.4 / \plot 1 0.7 1 -0.4 / \plot 2 0.7 2 -0.4 / \plot 3 0.7 3 -0.4 / \plot 4 0.7 4 -0.4 / \put{$\scriptstyle{+}$} at -1.9 0.7 \put{$\scriptstyle{-}$} at -2.1 0.7 \put{$\scriptstyle{+}$} at -0.9 0.7 \put{$\scriptstyle{-}$} at -1.1 0.7 \put{$\scriptstyle{+}$} at 0.1 0.7 \put{$\scriptstyle{-}$} at -0.1 0.7 \put{$\scriptstyle{+}$} at 1.1 0.7 \put{$\scriptstyle{-}$} at 0.9 0.7 \put{$\scriptstyle{+}$} at 2.1 0.7 \put{$\scriptstyle{-}$} at 1.9 0.7 \put{$\scriptstyle{+}$} at 3.1 0.7 \put{$\scriptstyle{-}$} at 2.9 0.7 \put{$1$}[b] at 0.5 0.15 \put{$s_0^\vee$}[b] at 1.5 0.15 \put{$s_0^\vee s_1^\vee$}[b] at 2.5 0.15 \put{$s_0^\vee s_1^\vee s_0^\vee$}[b] at 3.5 0.15 \put{$s_1^\vee$}[b] at -0.5 0.15 \put{$s_1^\vee s_0^\vee$}[b] at -1.5 0.15 \put{$s_1^\vee s_0^\vee s_1^\vee$}[b] at -2.5 0.15 \put{$1$}[t] at 0.5 -0.15 \put{$x^{\alpha}s_1^\vee$}[t] at 1.5 -0.15 \put{$x^{\alpha}$}[t] at 2.5 -0.15 \put{$x^{2\alpha}s_1^\vee$}[t] at 3.5 -0.15 \put{$s_1^\vee$}[t] at -0.5 -0.15 \put{$x^{-\alpha}$}[t] at -1.5 -0.15 \put{$x^{-\alpha}s_1^\vee$}[t] at -2.5 -0.15 \put{Sheet $\pi^\vee$} at -4.2 2 \plot -3.5 2 4.5 2 / \plot -3 2.7 -3 1.5 / \plot -2 2.7 -2 1.5 / \plot -1 2.7 -1 1.5 / \plot 0 2.7 0 1.5 / \plot 1 2.7 1 1.5 / \plot 2 2.7 2 1.5 / \plot 3 2.7 3 1.5 / \plot 4 2.7 4 1.5 / \put{$\scriptstyle{+}$} at -1.9 2.7 \put{$\scriptstyle{-}$} at -2.1 2.7 \put{$\scriptstyle{+}$} at -0.9 2.7 \put{$\scriptstyle{-}$} at -1.1 2.7 \put{$\scriptstyle{+}$} at 0.1 2.7 \put{$\scriptstyle{-}$} at -0.1 2.7 \put{$\scriptstyle{+}$} at 1.1 2.7 \put{$\scriptstyle{-}$} at 0.9 2.7 \put{$\scriptstyle{+}$} at 2.1 2.7 \put{$\scriptstyle{-}$} at 1.9 2.7 \put{$\scriptstyle{+}$} at 3.1 2.7 \put{$\scriptstyle{-}$} at 2.9 2.7 \put{$\pi^\vee$}[b] at 0.5 2.25 \put{$\pi^\vee s_1^\vee$}[b] at 1.5 2.15 \put{$\pi^\vee s_1^\vee s_0^\vee$}[b] at 2.5 2.15 \put{$\pi^\vee s_1^\vee s_0^\vee s_1^\vee$}[b] at 3.5 2.15 \put{$\pi^\vee s_0^\vee$}[b] at -0.5 2.15 \put{$\pi^\vee s_0^\vee s_1^\vee$}[b] at -1.5 2.15 \put{$\pi^\vee s_0^\vee s_1^\vee s_0^\vee$}[b] at -2.5 2.15 \put{$x^{\omega}s_1^\vee$}[t] at 0.5 1.85 \put{$x^{\omega}$}[t] at 1.5 1.85 \put{$x^{3\omega}s_1^\vee$}[t] at 2.5 1.85 \put{$x^{3\omega}$}[t] at 3.5 1.85 \put{$x^{-\omega}$}[t] at -0.5 1.85 \put{$x^{-\omega}s_1^\vee$}[t] at -1.5 1.85 \put{$x^{-3\omega}$}[t] at -2.5 1.85 \endpicture $$ The double affine Hecke algebra over the field $\bbK = \bbQ(q^{1/2},t^{1/2})$ is the algebra generated by $\pi, T_0, T_1$, and the group $X=\{q^kX^{j\omega} \mid k\in\hbox{$\frac12$}\bbZ, j\in \bbZ\}$ subject to the relations $$ \pi^2=1,\qquad T_i^2 = (t^{1/2}-t^{-1/2}) T_i+1, \textrm{ for } i=0,1, \qquad X^{j\omega}X^{k\omega} = X^{(j+k)\omega} \textrm{ for } j,k\in \bbZ, $$ $$\pi T_0 \pi^{-1} = T_1, \qquad T_1 X^\omega T_1 = X^{-\omega},\qquad \pi X^\omega \pi^{-1} = q^{1/2} X^{-\omega}. $$ Let $(T_0^\vee)^{-1} = X^\alpha T_1$, $Y^{-\alpha^\vee} = T_1^{-1}T_0^{-1}$, $Y^{-\alpha_0^\vee} = qY^{\alpha^\vee} = qT_0T_1$. The intertwiners are $\pi^\vee = X^\omega T_1$, and $$\tau_i^\vee = T_i^\vee + \frac{t^{-1/2}-t^{1/2}}{1-Y^{-\alpha_i^\vee}} = (T_i^\vee)^{-1} + \frac{(t^{-1/2}-t^{1/2})Y^{-\alpha_i^\vee}}{1-Y^{-\alpha_i^\vee}}, \quad\textrm{ for } i=0,1.$$ For $\mu \in \fh_\bbZ^*$, the minimal coset representatives are given by $$m_{k\omega} =\begin{cases} x^{k\omega}s_1 = \pi^\vee(s_1\pi^\vee)^{k-1}, & k \geq 1,\\ x^{k\omega} = (s_1\pi^\vee)^k, & k \leq 0. \end{cases}$$ \begin{example} {\em {\bf Littlewood-Richardson formulas.} In this example, we calculate $E_{3\omega}P_{k\omega}$ and $P_{3\omega}P_{k\omega}$ for $k\geq 3$. Let $\mu=3\omega$ and $\lambda = k\omega$. The minimal length representative of the coset $x^{3\omega}W_0$ is $m_{3\omega} = s_0^\vee s_1 \pi^\vee$, so Theorem~\ref{thm.EP-E} and Theorem~\ref{thm.PP-P} give \begin{align*} E_{3\omega}P_{k\omega} &= \sum_{h\in \Gamma_2^{\overline{\ttC}}} \left( (-1)^{|\phi_{\mathrm{grey}}^-(h)|} b_h c_{1(h)} c_{2(h)} \right) E_{\varpi(h)},\\ P_{3\omega}P_{k\omega} &= \sum_{h\in \Gamma_2^{\overline{\ttC}}} \left( (-1)^{|\phi_{\mathrm{grey}}^-(h)|} b_h e_h c_{1(h)} c_{2(h)} \right) P_{-w_0\ttwt(h)},\\ \end{align*} where the walks are of type $\vec{m}_{3\omega}^{-1} = (\pi^\vee, 1, 0)$, and begin in $x^{k\omega}$ or $x^{k\omega}s_1$. There are 18 walks in all. According to where each walk begins and ends (the alcoves $\ttb(h)$ and $\tte(h)$), we have $$b_h = \begin{cases} 1, & \hbox{if } \ttb(h) = x^{k\omega},\\ \displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}, & \hbox{if } \ttb(h) = x^{k\omega}s_1, \end{cases} \qquad\hbox{and}\qquad e_h = \begin{cases} 1, & \hbox{if } \tte(h) = x^{j\omega}s_1,\\ \displaystyle t^{-\frac12}\frac{1-q^jt^2}{1-q^jt} & \hbox{if } \tte(h) = x^{j\omega}. \end{cases} $$ Also note that for these calculations, we have $$Y^{-b_1^\vee}\mathbf{1} = Y^{\alpha^\vee-d}\mathbf{1} = qt\mathbf{1}, \qquad\hbox{and}\qquad Y^{-b_2^\vee}\mathbf{1} = Y^{\alpha^\vee-2d}\mathbf{1} = q^2 t\mathbf{1}. $$ In the following figures, they are organized into four groups so that walks in a group are generated by the same walk $p(h)$ (see~\eqref{eqn.ph}). In this root system, $-w_0 \ttwt(h) = \varpi(h)_+$. \newpage \noindent{\bf Group I.} The walks in this group are generated by $$\beginpicture \setcoordinatesystem units <0.75cm, 0.75cm> \setplotarea x from -4 to 10, y from -0.5 to 1.5 \plot -1.5 1 3 1 / \plot -1 0.8 -1 1.4 / \plot 0 0.8 0 1.4 / \plot 1 0.8 1 1.4 / \plot 2 0.8 2 1.4 / \plot -1.5 0 3 0 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \put{$\bullet$} at 0 0 \put{$p(h)=$} at -3 0.5 \setplotsymbol(.) \arrow<6pt> [.2,.67] from 1.5 0.2 to 0.5 0.2 \arrow<6pt> [.2,.67] from 2.5 0.2 to 1.5 0.2 \setdots<3pt> \plot 2.5 0.3 2.5 1.3 / \put{which gives $\epsilon_1=-1$, $\epsilon_2=-1$.}[l] at 4 0.5 \endpicture $$ \hspace{-0.4cm} \begin{tabular}{ccccc} $h$ & $b_h$ & $e_h$ & $(-1)^{|\phi_{\mathrm{grey}}^-(h)|}c_{1(h)}c_{2(h)}$ & $\varpi(h)$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -2.5 to 3, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k+3)\omega$}[r] at 3 1 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0.5 0.3 to 1.5 0.3 \arrow <4pt> [.2,.67] from 1.5 0.3 to 2.5 0.3 \setdots<3pt> \plot 0.5 -0.8 0.5 0.3 / \endpicture &$1$ &$1$ &$1$ &\scriptsize$(k+3)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -2.5 to 3, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k+1)\omega$} at 1.5 1 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0.5 0.3 to 1.5 0.3 \textcolor{Gray} { \plot 1.5 0.3 2 0.3 / \plot 1.97 0.3 1.97 0.5 / \arrow <4pt> [.2,.67] from 2 0.5 to 1.5 0.5 }\setdots<3pt> \plot 0.5 -0.8 0.5 0.3 / \endpicture &$1$ &\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k+1}t^2}{1-q^{k+1}t}$ &\footnotesize$\displaystyle -\frac{(t^{-\frac12}-t^{\frac12})q^{k+2}t}{1-q^{k+2}t}$ &\scriptsize$-(k+1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -2.5 to 3, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k-1)\omega$} at -0.5 1 \setplotsymbol(.) \textcolor{Gray} { \plot 0.5 0.3 1 0.3 / \plot 0.97 0.3 0.97 0.5 / \arrow <4pt> [.2,.67] from 1 0.5 to 0.5 0.5 }\arrow <4pt> [.2,.67] from 0.5 0.5 to -0.5 0.5 \setdots<3pt> \plot 0.5 -0.8 0.5 0.3 / \endpicture &$1$ &\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}$ &\footnotesize$\displaystyle -\left(\frac{(t^{-\frac12}-t^{\frac12})q^{k+1}t}{1-q^{k+1}t}\right) \left(\frac{1-q^k}{1-q^kt}\frac{1-q^kt^2}{1-q^kt}\right)$ &\scriptsize$-(k-3)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -2.5 to 3, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k+1)\omega$} at 0.5 1 \setplotsymbol(.) \textcolor{Gray} {\plot 0.5 0.2 1 0.2 / \plot 0.97 0.2 0.97 0.35 / \arrow <4pt> [.2,.67] from 1 0.35 to 0.5 0.35 \plot 0.5 0.35 0 0.35 / \plot 0.03 0.35 0.03 0.5 / \arrow <4pt> [.2,.67] from 0 0.5 to 0.5 0.5 }\setdots<3pt> \plot 0.5 -0.8 0.5 0.2 / \endpicture &$1$ &$1$ &\footnotesize$\displaystyle - \left(\frac{(t^{-\frac12}-t^{\frac12})q^{k+1}t}{1-q^{k+1}t}\right) \left(\frac{t^{-\frac12}-t^{\frac12}}{1-q^kt}\right)$ &\scriptsize$(k+1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -2.5 to 3, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k-3)\omega$}[l] at -2.5 1 \setplotsymbol(.) \arrow <4pt> [.2,.67] from -0.5 0.3 to -1.5 0.3 \arrow <4pt> [.2,.67] from -1.5 0.3 to -2.5 0.3 \setdots<3pt> \plot -0.5 -0.8 -0.5 0.3 / \endpicture &\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k-3}t^2}{1-q^{k-3}t}$ &\footnotesize$\displaystyle \left(\!\frac{1-q^{k-1}}{1-q^{k-1}t}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}\!\right) \!\!\! \left(\!\frac{1-q^{k-2}}{1-q^{k-2}t}\frac{1-q^{k-2}t^2}{1-q^{k-2}t}\!\right)$ &\scriptsize$-(k-3)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -2.5 to 3, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k-1)\omega$} at -1.5 1 \setplotsymbol(.) \arrow <4pt> [.2,.67] from -0.5 0.3 to -1.5 0.3 \textcolor{Gray} { \plot -1.5 0.3 -2 0.3 / \plot -1.97 0.3 -1.97 0.5 / \arrow <4pt> [.2,.67] from -2 0.5 to -1.5 0.5 }\setdots<3pt> \plot -0.5 -0.8 -0.5 0.3 / \endpicture &\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &$1$ &\footnotesize$\displaystyle \left(\frac{1-q^{k-1}}{1-q^{k-1}t}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}\right) \left(\frac{t^{-\frac12}-t^{\frac12}}{1-q^{k-2}t}\right)$ &\scriptsize$(k-1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -2.5 to 3, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k+1)\omega$} at 0.5 1 \setplotsymbol(.) \textcolor{Gray} { \plot -0.5 0.3 -1 0.3 / \plot -0.97 0.3 -0.97 0.5 / \arrow <4pt> [.2,.67] from -1 0.5 to -0.5 0.5 }\arrow <4pt> [.2,.67] from -0.5 0.5 to 0.5 0.5 \setdots<3pt> \plot -0.5 -0.8 -0.5 0.3 / \endpicture &\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &$1$ &\footnotesize$\displaystyle \frac{t^{-\frac12}-t^{\frac12}}{1-q^{k-1}t}$ &\scriptsize$(k+1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -2.5 to 3, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k-3)\omega$} at -0.5 1 \setplotsymbol(.) \textcolor{Gray} {\plot -0.5 0.2 -1 0.2 / \plot -0.97 0.2 -0.97 0.35 / \arrow <4pt> [.2,.67] from -1 0.35 to -0.5 0.35 \plot -0.5 0.35 0 0.35 / \plot -0.03 0.35 -0.03 0.5 / \arrow <4pt> [.2,.67] from 0 0.5 to -0.5 0.5 }\setdots<3pt> \plot -0.5 -0.8 -0.5 0.2 / \endpicture &\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}$ &\footnotesize$\displaystyle -\left(\frac{t^{-\frac12}-t^{\frac12}}{1-q^{k-1}t}\right) \left(\frac{(t^{-\frac12}-t^{\frac12})q^kt}{1-q^kt}\right)$ &\scriptsize$-(k-3)\omega$\\ \\ \end{tabular} \noindent{\bf Group II.} The walks in this group are generated by $$\hspace{-0.4cm} \beginpicture \setcoordinatesystem units <0.75cm, 0.75cm> \setplotarea x from -4 to 10, y from -0.35 to 2 \plot -1.5 1 3 1 / \plot -1 0.8 -1 1.4 / \plot 0 0.8 0 1.4 / \plot 1 0.8 1 1.4 / \plot 2 0.8 2 1.4 / \plot -1.5 0 3 0 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \put{$\bullet$} at 0 0 \put{$p(h)=$} at -3 0.5 \setplotsymbol(.) \arrow <6pt> [.2,.67] from 1.5 0.2 to 0.5 0.2 \plot 1.5 0.4 2 0.4 / \plot 1.97 0.2 1.97 0.4 / \arrow <6pt> [.2,.67] from 2 0.2 to 1.5 0.2 \setdots<3pt> \plot 1.5 0.4 1.5 1.3 / \put{which gives $\epsilon_2=-1$.}[l] at 4 0.5 \endpicture $$ \begin{center} \begin{tabular}{ccccc} $h$ & $b_h$ & $e_h$ & $(-1)^{\phi_{\mathrm{grey}}^-(h)}c_{1(h)}c_{2(h)}$ & $\varpi(h)$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -4 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k-1)\omega$} at -0.5 1 \setplotsymbol(.) \textcolor{Black} { \plot 0.5 0.3 1 0.3 / \plot 0.97 0.3 0.97 0.5 / \arrow <4pt> [.2,.67] from 1 0.5 to 0.5 0.5 }\arrow <4pt> [.2,.67] from 0.5 0.5 to -0.5 0.5 \setdots<3pt> \plot 0.5 -0.8 0.5 0.3 / \endpicture &$1$ &\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}$ &\footnotesize$\displaystyle \left(\frac{(t^{-\frac12}-t^{\frac12})qt}{1-qt}\right) \left(\frac{1-q^k}{1-q^kt}\frac{1-q^kt^2}{1-q^kt}\right)$ &\footnotesize$-(k-1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -4 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k+1)\omega$} at 0.5 1 \setplotsymbol(.) \textcolor{Black} {\plot 0.5 0.2 1 0.2 / \plot 0.97 0.2 0.97 0.35 / \arrow <4pt> [.2,.67] from 1 0.35 to 0.5 0.35 }\textcolor{Gray} {\plot 0.5 0.35 0 0.35 / \plot 0.03 0.35 0.03 0.5 / \arrow <4pt> [.2,.67] from 0 0.5 to 0.5 0.5 }\setdots<3pt> \plot 0.5 -0.8 0.5 0.2 / \endpicture &$1$ &$1$ &\footnotesize$\displaystyle \left(\frac{(t^{-\frac12}-t^{\frac12})qt}{1-qt}\right) \left(\frac{t^{-\frac12}-t^{\frac12}}{1-q^kt}\right)$ &\footnotesize$(k+1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -4 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k+1)\omega$} at 0.5 1 \setplotsymbol(.) \textcolor{Black} { \plot -0.5 0.3 -1 0.3 / \plot -0.97 0.3 -0.97 0.5 / \arrow <4pt> [.2,.67] from -1 0.5 to -0.5 0.5 }\arrow <4pt> [.2,.67] from -0.5 0.5 to 0.5 0.5 \setdots<3pt> \plot -0.5 -0.8 -0.5 0.3 / \endpicture &\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &$1$ &\footnotesize$\displaystyle\frac{(t^{-\frac12}-t^{\frac12})qt}{1-qt}$ &\footnotesize$(k+1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -4 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k-1)\omega$} at -0.5 1 \setplotsymbol(.) \textcolor{Black} {\plot -0.5 0.2 -1 0.2 / \plot -0.97 0.2 -0.97 0.35 / \arrow <4pt> [.2,.67] from -1 0.35 to -0.5 0.35 }\textcolor{Gray} {\plot -0.5 0.35 0 0.35 / \plot -0.03 0.35 -0.03 0.5 / \arrow <4pt> [.2,.67] from 0 0.5 to -0.5 0.5 }\setdots<3pt> \plot -0.5 -0.8 -0.5 0.2 / \endpicture &\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}$ &\footnotesize$\displaystyle -\left(\frac{(t^{-\frac12}-t^{\frac12})qt}{1-qt}\right) \left(\frac{(t^{-\frac12}-t^{\frac12})q^kt}{1-q^kt}\right)$ &\footnotesize$-(k-1)\omega$\\ \\ \end{tabular} \end{center} \newpage \noindent{\bf Group III.} The walks in this group are generated by $$\hspace{-0.4cm} \beginpicture \setcoordinatesystem units <0.75cm, 0.75cm> \setplotarea x from -4 to 10, y from -0.35 to 2 \plot -1.5 1 3 1 / \plot -1 0.8 -1 1.4 / \plot 0 0.8 0 1.4 / \plot 1 0.8 1 1.4 / \plot 2 0.8 2 1.4 / \plot -1.5 0 3 0 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \put{$\bullet$} at 0 0 \put{$p(h)=$} at -3 0.5 \setplotsymbol(.) \plot 0.5 0.4 1 0.4 / \plot 0.97 0.2 0.97 0.4 / \arrow <6pt> [.2,.67] from 1 0.2 to 0.5 0.2 \arrow <6pt> [.2,.67] from -0.5 0.4 to 0.5 0.4 \setdots<3pt> \plot -0.5 0.4 -0.5 1.3 / \put{which gives $\epsilon_1=+1$.}[l] at 4 0.5 \endpicture $$ \begin{center} \begin{tabular}{cccccc} $h$ & $b_h$ & $e_h$ & $(-1)^{\phi_{\mathrm{grey}}^-(h)}c_{1(h)}c_{2(h)}$ &$\varpi(h)$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -3 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k+1)\omega$} at 1.5 1 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0.5 0.3 to 1.5 0.3 \textcolor{Black} { \plot 1.5 0.3 2 0.3 / \plot 1.97 0.3 1.97 0.5 / \arrow <4pt> [.2,.67] from 2 0.5 to 1.5 0.5 }\setdots<3pt> \plot 0.5 -0.8 0.5 0.3 / \endpicture &$1$ &\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k+1}t^2}{1-q^{k+1}t}$ &\footnotesize$\displaystyle \frac{(t^{-\frac12}-t^{\frac12})q^2t}{1-q^2t}$ &\footnotesize$-(k+1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -3 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k+1)\omega$} at 0.5 1 \setplotsymbol(.) \textcolor{Gray} {\plot 0.5 0.2 1 0.2 / \plot 0.97 0.2 0.97 0.35 / \arrow <4pt> [.2,.67] from 1 0.35 to 0.5 0.35 }\textcolor{Black} {\plot 0.5 0.35 0 0.35 / \plot 0.03 0.35 0.03 0.5 / \arrow <4pt> [.2,.67] from 0 0.5 to 0.5 0.5 }\setdots<3pt> \plot 0.5 -0.8 0.5 0.2 / \endpicture &$1$ &$1$ &\footnotesize$\displaystyle -\left(\frac{t^{-\frac12}-t^{\frac12}}{1-q^{k+1}t}\right) \left(\frac{(t^{-\frac12}-t^{\frac12})q^2t}{1-q^2t}\right)$ &\footnotesize$(k+1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -3 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k-1)\omega$} at -1.5 1 \setplotsymbol(.) \arrow <4pt> [.2,.67] from -0.5 0.3 to -1.5 0.3 \textcolor{Black} { \plot -1.5 0.3 -2 0.3 / \plot -1.97 0.3 -1.97 0.5 / \arrow <4pt> [.2,.67] from -2 0.5 to -1.5 0.5 }\setdots<3pt> \plot -0.5 -0.8 -0.5 0.3 / \endpicture &\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &$1$ &\footnotesize$\displaystyle \left(\frac{1-q^{k-1}}{1-q^{k-1}t}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}\right) \left(\frac{(t^{-\frac12}-t^{\frac12})q^2t}{1-q^2t}\right)$ &\footnotesize$(k-1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -3 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k-1)\omega$} at -0.5 1 \setplotsymbol(.) \textcolor{Gray} {\plot -0.5 0.2 -1 0.2 / \plot -0.97 0.2 -0.97 0.35 / \arrow <4pt> [.2,.67] from -1 0.35 to -0.5 0.35 }\textcolor{Black} {\plot -0.5 0.35 0 0.35 / \plot -0.03 0.35 -0.03 0.5 / \arrow <4pt> [.2,.67] from 0 0.5 to -0.5 0.5 }\setdots<3pt> \plot -0.5 -0.8 -0.5 0.2 / \endpicture &\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}$ &\footnotesize$\displaystyle \left(\frac{(t^{-\frac12}-t^{\frac12})q^{k-1}t}{1-q^{k-1}t}\right) \left(\frac{(t^{-\frac12}-t^{\frac12})q^2t}{1-q^2t}\right)$ &\footnotesize$-(k-1)\omega$\\ \\ \end{tabular} \end{center} \noindent{\bf Group IV.} The walks in this group are generated by $$\hspace{-0.4cm} \beginpicture \setcoordinatesystem units <0.75cm, 0.75cm> \setplotarea x from -4 to 4, y from -0.35 to 2 \plot -1.5 1 3 1 / \plot -1 0.8 -1 1.4 / \plot 0 0.8 0 1.4 / \plot 1 0.8 1 1.4 / \plot 2 0.8 2 1.4 / \plot -1.5 0 3 0 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \put{$\bullet$} at 0 0 \put{$p(h)=$} at -3 0.5 \setplotsymbol(.) \plot 0.5 0.35 1 0.35 / \plot 0.97 0.15 0.97 0.35 / \arrow <6pt> [.2,.67] from 1 0.15 to 0.5 0.15 \plot 0.5 0.55 0 0.55 / \plot 0.03 0.35 0.03 0.55 / \arrow <6pt> [.2,.67] from 0 0.35 to 0.5 0.35 \setdots<3pt> \plot 0.5 0.55 0.5 1.3 / \endpicture $$ \begin{center} \begin{tabular}{cccccc} $h$ & $b_h$ & $e_h$ & $(-1)^{\phi_{\mathrm{grey}}^-(h)}c_{1(h)}c_{2(h)}$ &$\varpi(h)$ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -4 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$(k+1)\omega$} at 0.5 1 \setplotsymbol(.) \textcolor{Black} {\plot 0.5 0.2 1 0.2 / \plot 0.97 0.2 0.97 0.35 / \arrow <4pt> [.2,.67] from 1 0.35 to 0.5 0.35 \plot 0.5 0.35 0 0.35 / \plot 0.03 0.35 0.03 0.5 / \arrow <4pt> [.2,.67] from 0 0.5 to 0.5 0.5 }\setdots<3pt> \plot 0.5 -0.8 0.5 0.2 / \endpicture &$1$ &$1$ &\footnotesize$\displaystyle \left(\frac{(t^{-\frac12}-t^{\frac12})q^2t}{1-q^2t}\right) \left(\frac{(t^{-\frac12}-t^{\frac12})}{1-qt}\right)$ &\footnotesize$(k+1)\omega$ \\ \\ \beginpicture \setcoordinatesystem units <0.5cm, 0.6cm> \setplotarea x from -4 to 4, y from -1 to 1 \plot -3 0 3 0 / \plot -3 -1 3 -1 / \plot -2 -0.2 -2 0.6 / \plot -1 -0.2 -1 0.6 / \plot 0 -0.2 0 0.6 / \plot 1 -0.2 1 0.6 / \plot 2 -0.2 2 0.6 / \plot -2 -1.2 -2 -0.8 / \plot -1 -1.2 -1 -0.8 / \plot 0 -1.2 0 -0.8 / \plot 1 -1.2 1 -0.8 / \plot 2 -1.2 2 -0.8 / \put{\tiny$k\omega$} at 0 -1.4 \put{\tiny$-(k-1)\omega$} at -0.5 1 \setplotsymbol(.) \textcolor{Black} {\plot -0.5 0.2 -1 0.2 / \plot -0.97 0.2 -0.97 0.35 / \arrow <4pt> [.2,.67] from -1 0.35 to -0.5 0.35 \plot -0.5 0.35 0 0.35 / \plot -0.03 0.35 -0.03 0.5 / \arrow <4pt> [.2,.67] from 0 0.5 to -0.5 0.5 }\setdots<3pt> \plot -0.5 -0.8 -0.5 0.2 / \endpicture &\quad\footnotesize$\displaystyle t^{\frac12}\frac{1-q^k}{1-q^kt}$ &\quad\footnotesize$\displaystyle t^{-\frac12}\frac{1-q^{k-1}t^2}{1-q^{k-1}t}$ &\footnotesize$\displaystyle \left(\frac{(t^{-\frac12}-t^{\frac12})q^2t}{1-q^2t}\right) \left(\frac{(t^{-\frac12}-t^{\frac12})}{1-qt}\right)$ &\footnotesize$(k+1)\omega$ \\ \\ \end{tabular} \end{center} Thus $E_{3\omega}P_{k\omega}$ is a linear combination of $E_{(k+3)\omega}$, $E_{-(k+1)\omega}$, $E_{(k+1)\omega}$, $E_{-(k-1)\omega}$, $E_{(k-1)\omega}$, $E_{-(k-3)\omega}$, and after simplification, \begin{align*} E_{3\omega}P_{k\omega} = E_{(k+3)\omega} &+ q^2\frac{1-t}{1-q^2t} \frac{1-q^k}{1-q^{k+2}t} t^{1/2} E_{-(k+1)\omega}\\ &+ \frac{1-t}{1-q}\frac{1-q^2}{1-q^2t}\frac{1-q^k}{1-q^{k-1}t} \frac{1-q^{k+1}t^2}{1-q^{k+1}t} E_{(k+1)\omega}\\ &+ q\frac{1-t}{1-q} \frac{1-q^2}{1-q^2t} \frac{1-q^{k-1}}{1-q^{k-1}t} \frac{1-q^k}{1-q^kt} \frac{1-q^kt^2}{1-q^{k+1}t} t^{1/2}E_{-(k-1)\omega}\\ &+ \frac{1-t}{1-q^2t}\frac{1-q^{k-1}}{1-q^{k-2}t} \frac{1-q^k}{1-q^{k-1}t}\frac{1-q^{k-1}t^2}{1-q^{k-1}t} \frac{1-q^kt^2}{1-q^kt} E_{(k-1)\omega}\\ &+ \frac{1-q^{k-2}}{1-q^{k-2}t}\frac{1-q^{k-2}t^2}{1-q^{k-2}t} \frac{1-q^{k-1}}{1-q^{k-1}t}\frac{1-q^{k-1}t^2}{1-q^{k-1}t} \frac{1-q^k}{1-q^kt} t^{1/2}E_{-(k-3)\omega}. \end{align*} Proposition~\ref{prop.domP} gives $$P_{-j\omega} = t^{-\frac12}\frac{1-q^jt^2}{1-q^jt} P_{j\omega} = e_{h} P_{j\omega}$$ if $\tte(h) = x^{j\omega}$ is a dominant weight. Thus, after simplification, \begin{align*} P_{3\omega}P_{k\omega} = P_{(k+3)\omega} &+ \frac{1-t}{1-q} \frac{1-q^3}{1-q^2t} \frac{1-q^k}{1-q^{k-1}t} \frac{1-q^{k+1}t^2}{1-q^{k+2}t} P_{(k+1)\omega} \\ &+ \frac{1-t}{1-q} \frac{1-q^3}{1-q^2t} \frac{1-q^{k-1}}{1-q^{k-2}t} \frac{1-q^k}{1-q^{k-1}t} \frac{1-q^{k-1}t^2}{1-q^kt} \frac{1-q^kt^2}{1-q^{k+1}t} P_{(k-1)\omega}\\ &+ \frac{1-q^{k-2}}{1-q^{k-2}t}\frac{1-q^{k-2}t^2}{1-q^{k-2}t} \frac{1-q^{k-1}}{1-q^{k-1}t}\frac{1-q^{k-1}t^2}{1-q^{k-1}t} \frac{1-q^k}{1-q^kt} \frac{1-q^{k-3}t^2}{1-q^{k-3}t} P_{(k-3)\omega}. \end{align*} Consider the case $q=0$ where the symmetric Macdonald polynomials become Hall-Littlewood polynomials. Following the discussion in Section~\ref{sec.specialization}, the walks giving a nonzero contribution to the sum are those whose only folds are positive and grey. The expression $$P_{3\omega}(t)P_{k\omega}(t) = P_{(k+3)\omega}(t) +(1-t)P_{(k+1)\omega}(t) +(1-t)P_{(k-1)\omega}(t) +P_{(k-3)\omega}(t),$$ is given by four positively folded walks, and coincides with the Littlewood-Richardson formulas~\cite[Theorem 1.3]{Sc06} and~\cite[Theorem 4.9]{R06} for Hall-Littlewood polynomials. We also mention that $$E_{3\omega}(t)P_{k\omega}(t) = E_{(k+3)\omega}(t) +(1-t)E_{(k+1)\omega}(t) +(1-t)E_{(k-1)\omega}(t) +t^{1/2}E_{-(k-3)\omega}(t).$$ In the case $q=t=0$, the symmetric Macdonald polynomials are Schur polynomials $P_\mu(0,0)= s_\mu$, and the nonsymmetric Macdonald polynomials are Demazure characters $E_\mu(0,0) = \mathcal{A}_\mu$. The dominant weights $k\omega$ correspond to the partitions $(k,0)$, and the weights $-k\omega$ correspond to the compositions $(0,k)$ for $k\geq 0$. The four positively folded walks correspond to the four Littlewood-Richardson tableaux that give the classical Littlewood-Richardson formula for Schur functions: $$s_{3\omega}s_{k\omega}= s_{(k+3)\omega} +s_{(k+1)\omega} +s_{(k-1)\omega} +s_{(k-3)\omega}.$$ By normalizing the nonsymmetric polynomials $E_\mu$ so that the coefficient of the monomial $X^\mu$ in $E_\mu$ is $1$, (see the paragraph following~\eqref{eqn.E}), then each term in $$\calA_{3\omega}s_{k\omega} = \calA_{(k+3)\omega} +\calA_{(k+1)\omega} +\calA_{(k-1)\omega} +\calA_{-(k-3)\omega},$$ corresponds to a skyline filling in the formula~\cite[Theorem 6.1]{HLMvW09}. \hfill$\diamond$ }\end{example} \begin{example}{\em {\bf Pieri formulas.} The weight $\omega$ is minuscule in the $\fsl_2$ root system. The two walks which give the expressions \begin{align*} E_{\omega}P_{k\omega} &= E_{(k+1)\omega} + \frac{1-q^k}{1-q^kt}t^{1/2} E_{-(k-1)\omega},\\ P_{\omega}P_{k\omega} &= P_{(k+1)\omega} + \frac{1-q^k}{1-q^kt} \frac{1-q^{k-1}t^2}{1-q^{k-1}t} P_{(k-1)\omega}, \end{align*} are walks of type $\vec{m}_{\omega}^{-1} = \pi^\vee$ corresponding to `changes in sheets', and begin in the alcoves $m_{k\omega}^{-1}$ or $m_{k\omega}^{-1}s_1$. \hfill$\diamond$ }\end{example} \subsection{Type $\fsl_3$}\label{sec.eg2} Let $\{\varepsilon_1, \varepsilon_2, \varepsilon_3\}$ be an orthonormal basis for $\bbR^3$, and let $\{\varepsilon_1^\vee, \varepsilon_2^\vee, \varepsilon_3^\vee\}$ be its dual basis, where $\langle \varepsilon_i, \varepsilon_j^\vee \rangle = \delta_{ij}$. The simple roots and simple coroots are \begin{eqnarray*} &\alpha_1 = \varepsilon_1-\varepsilon_2,\quad \alpha_2 = \varepsilon_2-\varepsilon_3,\quad \varphi = \varepsilon_1 - \varepsilon_3,\\ &\alpha_1^\vee = \varepsilon_1^\vee-\varepsilon_2^\vee,\quad \alpha_2^\vee = \varepsilon_2^\vee-\varepsilon_3^\vee,\quad \varphi^\vee = \varepsilon_1^\vee - \varepsilon_3^\vee, \end{eqnarray*} so that the root and coroot lattices are $Q = \bbZ\alpha_1+\bbZ\alpha_2$ and $Q = \bbZ\alpha_1^\vee+\bbZ\alpha_2^\vee$. The fundamental weights and fundamental coweights are \begin{eqnarray*} &\omega_1 = \varepsilon_1 - \hbox{$\frac13$}(\varepsilon_1+\varepsilon_2+\varepsilon_3), \quad \omega_2 = \varepsilon_1+\varepsilon_2 - \hbox{$\frac23$}(\varepsilon_1+\varepsilon_2+\varepsilon_3),\\ &\omega_1^\vee = \varepsilon_1^\vee - \hbox{$\frac13$}(\varepsilon_1^\vee+\varepsilon_2^\vee+\varepsilon_3^\vee), \quad \omega_2^\vee = \varepsilon_1^\vee+\varepsilon_2^\vee - \hbox{$\frac23$}(\varepsilon_1^\vee+\varepsilon_2^\vee+\varepsilon_3^\vee), \end{eqnarray*} and the weight and coweight lattices are $\fh_\bbZ^* = \bbZ \omega_1\oplus \bbZ \omega_2$ and $\fh_\bbZ = \bbZ \omega_1^\vee \oplus \bbZ \omega_2^\vee$. The group $\Pi^\vee \cong \fh_\bbZ^*/Q$ is the cyclic group of order 3. Note that $\langle \fh_\bbZ^*, \fh_\bbZ \rangle \subseteq \hbox{$\frac13$}\bbZ$. The Weyl group of this root system is the symmetric group on three symbols $$W_0=\fS_3 = \left\langle s_1, s_2, \mid s_1s_2s_1=s_2s_1s_2 = s_\varphi, s_i^2=1 \hbox{ for } i =1,2 \right\rangle,$$ and the extended affine Weyl group $W^\vee$ is generated by the group $W_0$ and the element $\pi^\vee= x^{\omega_1}s_1s_2$, subject to the relations \begin{equation} (\pi^\vee)^3=1, \quad \pi^\vee s_0^\vee = s_1\pi^\vee,\quad \pi^\vee s_1 = s_2\pi^\vee,\quad s_0s_1s_0 = s_1s_0s_1, \quad s_0s_2s_0 = s_2s_0s_2, \end{equation} where $s_0^\vee= x^\varphi s_\varphi$ and $(\pi^\vee)^2= x^{\omega_2}s_2s_1$. Alternatively, \begin{equation} W^\vee = \{x^{k_1\omega_1+k_2\omega_2} w \mid k_1,k_2\in\bbZ, w\in \fS_3\}. \end{equation} What follows is the alcove picture for the extended affine Weyl group $W^\vee$, showing the correspondence between the alcoves and the elements of $W^\vee$. The periodic orientation is indicated by $\scriptstyle{+}$ and $\scriptstyle{-}$ on either side of the hyperplanes. Since the $\fsl_n$ root system is self-dual, the dual alcove picture is identical. \newpage \begin{figure}\label{fig.sl3alcoves} $$\beginpicture \setcoordinatesystem units <0.65cm,0.65cm> \setplotarea x from -11 to 6, y from -5 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \put{$\ttH_{\alpha_2^\vee}$} at 2.8 4.6 \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \put{$\ttH_{\alpha_1^\vee}$} at -2.9 4.7 \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \put{$\ttH_{\varphi^\vee}$} at 5.7 0 \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\scriptstyle +$} at 2.2 4.2 \put{$\scriptstyle -$} at 2.7 4 \put{$\scriptstyle +$} at -2.2 4.2 \put{$\scriptstyle -$} at -2.7 4 \put{$\scriptstyle +$} at 4.9 0.25 \put{$\scriptstyle -$} at 4.9 -0.2 \put{Sheet $(\pi^\vee)^2$}[l] at -11 0 \put{$\bullet$} at 0 0 \put{\small ${\pi^\vee}^2$} at 0 1.3 \put{\small${\pi^\vee}^2s_2$} at -1 0.4 \put{\small${\pi^\vee}^2s_1$} at 0 2.1 \put{\small${\pi^\vee}^2s_0^\vee$} at 1 0.4 \put{\small$x^{\omega_2}$} at -1 3 \put{\small $x^{\omega_1-\omega_2}$} at 2 1.4 \put{\small $x^{-\omega_1}$} at -1 -0.4 \endpicture $$ $$ \beginpicture \setcoordinatesystem units <0.65cm,0.65cm> \setplotarea x from -11 to 6, y from -5 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \put{$\ttH_{\alpha_2^\vee}$} at 2.8 4.6 \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \put{$\ttH_{\alpha_1^\vee}$} at -2.9 4.7 \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \put{$\ttH_{\varphi^\vee}$} at 5.7 0 \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\scriptstyle +$} at 2.2 4.2 \put{$\scriptstyle -$} at 2.7 4 \put{$\scriptstyle +$} at -2.2 4.2 \put{$\scriptstyle -$} at -2.7 4 \put{$\scriptstyle +$} at 4.9 0.25 \put{$\scriptstyle -$} at 4.9 -0.2 \put{Sheet $\pi^\vee$}[l] at -11 0 \put{$\bullet$} at 0 0 \put{\small $\pi^\vee$} at 0 1.2 \put{\small $\pi^\vee s_1$} at 1 0.4 \put{\small $\pi^\vee s_2$} at 0 2.1 \put{\small $\pi^\vee s_0^\vee$} at -1 0.4 \put{\small $x^{\omega_1}$} at 1 3 \put{\small $x^{-\omega_1+\omega_2}$} at -2 1.4 \put{\small $x^{-\omega_2}$} at 1 -0.4 \endpicture $$ $$ \beginpicture \setcoordinatesystem units <0.65cm,0.65cm> \setplotarea x from -11 to 6, y from -5 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \put{$\ttH_{\alpha_2^\vee}$} at 2.8 4.6 \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \put{$\ttH_{\alpha_1^\vee}$} at -2.9 4.7 \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \put{$\ttH_{\varphi^\vee}$} at 5.7 0 \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\scriptstyle +$} at 2.2 4.2 \put{$\scriptstyle -$} at 2.7 4 \put{$\scriptstyle +$} at -2.2 4.2 \put{$\scriptstyle -$} at -2.7 4 \put{$\scriptstyle +$} at 4.9 0.25 \put{$\scriptstyle -$} at 4.9 -0.2 \put{\hbox{Sheet} $1$}[l] at -11 0 \put{$\bullet$} at 0 0 \put{\small $1$} at 0 1.1 \put{\small $s_1$} at -1 0.5 \put{\small $s_2$} at 1 0.5 \put{\small $s_1s_2$} at -1 -0.5 \put{\small $s_2s_1$} at 1 -0.5 \put{\small $s_\varphi$} at 0 -1.2 \put{\small $s_0^\vee$} at 0 2.3 \put{\small $x^{\alpha_1}$} at 3 2.82 \put{\small $x^{\alpha_2}$} at -3 2.82 \put{\small $x^{-\alpha_1}$} at 3 -0.5 \put{\small $x^{-\alpha_2}$} at -3 -0.5 \put{\small $x^{-\varphi}$} at 0 -2.3 \endpicture $$ \end{figure} \begin{example}{\em Using Theorem~\ref{thm.Xtau-X}, we calculate the expansion of $E_{-\alpha_2}$ in the basis of monomials. The minimal length representative is $m_{-\alpha_2} = s_2s_1s_0^\vee$, and the eight walks of type $\vec{m}_{-\alpha_2}=(2,1,0)$ beginning in the fundamental alcove are: \begin{center}\begin{tabular}{ccc} \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -8 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle X^{-\alpha_2}T_1$} at 0 -6 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0 1.15 to 1 0.57 \arrow <4pt> [.2,.67] from 1 0.57 to 1 -0.57 \arrow <4pt> [.2,.67] from 1 -0.57 to 2 -1.15 \endpicture & \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -8 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle X^0T_2T_1 \frac{t^{-1/2}-t^{1/2}}{1-Y^{\varphi^\vee-d}}$} at 0 -6 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0 1.15 to 1 0.57 \arrow <4pt> [.2,.67] from 1 0.57 to 1 -0.57 \plot 1 -0.58 1.5 -0.87 / \plot 1.5 -0.87 1.4 -1.05 / \arrow <4pt> [.2,.67] from 1.4 -1.05 to 0.9 -0.76 \endpicture & \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -8 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle X^{\alpha_1}T_1T_2 \frac{t^{-1/2}-t^{1/2}}{1-Y^{\alpha_2^\vee-d}}$} at 0 -6 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0 1.15 to 0.9 0.57 \plot 0.9 0.57 0.9 0 / \plot 1.1 0 0.9 0 / \arrow <4pt> [.2,.67] from 1.1 0 to 1.1 0.57 \arrow <4pt> [.2,.67] from 1.1 0.57 to 2 1.15 \endpicture \\ \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -8 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\scriptsize$\displaystyle X^0T_2 \frac{(t^{-1/2}-t^{1/2})Y^{\varphi^\vee-d}} {1-Y^{\varphi^\vee-d}} \frac{t^{-1/2}-t^{1/2}}{1-Y^{\alpha_2^\vee-d}} $} at 0 -6 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0 1.15 to 0.9 0.57 \plot 0.9 0.57 0.9 0 / \plot 1.1 0 0.9 0 / \arrow <4pt> [.2,.67] from 1.1 0 to 1.1 0.57 \plot 1.1 0.57 1.5 0.86 / \plot 1.4 1.04 1.5 0.86 / \arrow <4pt> [.2,.67] from 1.4 1.04 to 0.95 0.75 \endpicture \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -8 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle X^{\alpha_2}T_2T_1 \frac{t^{-1/2}-t^{1/2}}{1-Y^{\varphi^\vee-2d}}$} at 0 -6 \setplotsymbol(.) \plot 0 1.15 0.5 0.86 / \plot 0.5 0.86 0.4 0.68 / \arrow <4pt> [.2,.67] from 0.4 0.68 to -0.1 0.97 \arrow <4pt> [.2,.67] from -0.1 0.97 to -1 0.57 \arrow <4pt> [.2,.67] from -1 0.57 to -2 1.15 \endpicture \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -8 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\!\!\!\!\!\scriptsize$\displaystyle X^0T_1 \frac{t^{-1/2}-t^{1/2}}{1-Y^{\varphi^\vee-2d}} \frac{(t^{-1/2}-t^{1/2})Y^{\varphi^\vee-d}} {1-Y^{\varphi^\vee-d}}$} at 0 -6 \setplotsymbol(.) \plot 0 1.15 0.5 0.86 / \plot 0.5 0.86 0.4 0.68 / \arrow <4pt> [.2,.67] from 0.4 0.68 to -0.1 0.97 \arrow <4pt> [.2,.67] from -0.1 0.97 to -0.9 0.57 \plot -0.9 0.57 -1.5 0.86 / \plot -1.5 0.86 -1.6 0.68 / \arrow <4pt> [.2,.67] from -1.6 0.68 to -1 0.38 \endpicture \end{tabular}\end{center} $$ \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -8 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle X^\varphi T_1T_2T_1 \frac{t^{-1/2}-t^{1/2}}{1-Y^{\varphi^\vee-2d}} \frac{t^{-1/2}-t^{1/2}}{1-Y^{\alpha_2^\vee-d}}$} at 0 -6 \setplotsymbol(.) \plot 0 0.8 0.4 0.6 / \plot 0.4 0.6 0.5 0.78 / \arrow <4pt> [.2,.67] from 0.5 0.78 to 0 1.05 \plot 0 1.05 -0.4 0.83 / \plot -0.4 0.83 -0.5 1.01 / \arrow <4pt> [.2,.67] from -0.5 1.01 to 0 1.3 \arrow <4pt> [.2,.67] from 0 1.3 to 0 2.4 \endpicture \qquad \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -8 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle X^0 \frac{t^{-1/2}-t^{1/2}}{1-Y^{\varphi^\vee-2d}} \frac{t^{-1/2}-t^{1/2}}{1-Y^{\alpha_2^\vee-d}} \frac{(t^{-1/2}-t^{1/2})Y^{\varphi^\vee-d}} {1-Y^{\varphi^\vee-d}}$} at 0 -6 \setplotsymbol(.) \plot 0 0.8 0.4 0.6 / \plot 0.4 0.6 0.5 0.78 / \arrow <4pt> [.2,.67] from 0.5 0.78 to 0 1 \plot 0 1 -0.4 0.7 / \plot -0.4 0.7 -0.5 0.88 / \arrow <4pt> [.2,.67] from -0.5 0.88 to -0.1 1.15 \plot -0.1 1.15 -0.1 1.72 / \plot -0.1 1.72 0.1 1.72 / \arrow <4pt> [.2,.67] from 0.1 1.72 to 0.1 1.15 \endpicture $$ Since $Y^{\varphi^\vee-d}\mathbf{1}= qt^2\mathbf{1}$, $Y^{\alpha_2^\vee-d}\mathbf{1}= qt\mathbf{1}$, and $Y^{\varphi^\vee-2d}\mathbf{1} = q^2t^2\mathbf{1}$, then in the polynomial representation, \begin{align*} t^{-\frac12}&E_{-\alpha_2} \mathbf{1} = X^{-\alpha_2}\mathbf{1} + X^{\alpha_1}\frac{1-t}{1-qt}\mathbf{1} + X^{\alpha_2}\frac{1-t}{1-q^2t^2}\mathbf{1} + X^\varphi \frac{1-t}{1-q^2t^2}\frac{1-t}{1-qt}\mathbf{1} \\ &+\! X^0\! \left(\frac{1-t}{1-qt^2} +\! \frac{1-t}{1-qt^2}\frac{1-t}{1-qt}qt +\! \frac{1-t}{1-q^2t^2}\frac{1-t}{1-qt^2}qt +\! \frac{1-t}{1-q^2t^2}\frac{1-t}{1-qt}\frac{1-t}{1-qt^2}q\right) \!\mathbf{1}. \end{align*} \hfill$\diamond$ }\end{example} \begin{example}\label{eg.XtoE}{\em Using Corollary~\ref{cor.X-tau}, we calculate the expansion of $X^{-\alpha_2}$ in the basis of nonsymmetric Macdonald polynomials. We consider the eight walks of type $\vec{m}_{-\alpha_2}^{-1}=(0,1,2)$ beginning in the fundamental alcove. Only five of these are contained in the dominant chamber. \begin{center}\begin{tabular}{ccc} \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -7 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle t^{-1/2}\tau_{m_{-\alpha_2}}^\vee \mathbf{1} $} at 0 -6 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0 1.15 to 0 2.3 \arrow <4pt> [.2,.67] from 0 2.3 to -1 2.88 \arrow <4pt> [.2,.67] from -1 2.88 to -1 4.03 \endpicture & \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -7 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle -t^{-1/2}\tau_{m_{\alpha_2}}^\vee \frac{t^{-1/2}-t^{1/2}}{1-Y^{\varphi^\vee-2d}} \mathbf{1}$} at 0 -6 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0 1.15 to 0 2.3 \arrow <4pt> [.2,.67] from 0 2.3 to -0.9 2.88 \plot -0.9 2.88 -0.9 3.46 / \plot -1.1 3.45 -0.9 3.45 / \arrow <4pt> [.2,.67] from -1.1 3.46 to -1.1 2.88 \endpicture & \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -7 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle -t^{-1/2}\tau_{m_{\alpha_1}}^\vee \frac{t^{-1/2}-t^{1/2}}{1-Y^{\alpha_2^\vee-d}} \mathbf{1}$} at 0 -6 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0 1.15 to 0 2.2 \plot -0.6 2.42 0 2.12 / \plot -0.5 2.6 -0.6 2.42 / \arrow <4pt> [.2,.67] from -0.5 2.6 to 0.1 2.31 \arrow <4pt> [.2,.67] from 0.1 2.31 to 1 2.88 \endpicture \end{tabular} \end{center} $$ \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -7 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle t^{-1/2}\tau_{m_\varphi}^\vee \frac{t^{-1/2}-t^{1/2}}{1-Y^{\alpha_2^\vee-d}} \frac{t^{-1/2}-t^{1/2}}{1-Y^{\alpha_1^\vee-d}} \mathbf{1}$} at 0 -6 \setplotsymbol(.) \arrow <4pt> [.2,.67] from 0 1.15 to 0 2.1 \plot 0 2.1 -0.6 2.4 / \plot -0.6 2.4 -0.5 2.58 / \arrow <4pt> [.2,.67] from -0.5 2.58 to 0 2.3 \plot 0 2.3 0.5 2.59 / \plot 0.4 2.77 0.5 2.59 / \arrow <4pt> [.2,.67] from 0.4 2.77 to 0 2.55 \endpicture \qquad \beginpicture \setcoordinatesystem units <0.45cm,0.45cm> \setplotarea x from -5 to 5, y from -7 to 5 \plot 1.5 -4.34 4.5 0.87 / \plot -0.5 -4.34 3.5 2.59 / \plot -2.5 -4.34 2.5 4.34 / \plot -3.5 -2.6 0.5 4.34 / \plot -4.5 -0.87 -1.5 4.34 / \plot -1.5 -4.34 -4.5 0.87 / \plot 0.5 -4.34 -3.5 2.59 / \plot 2.5 -4.34 -2.5 4.34 / \plot 3.5 -2.6 -0.5 4.34 / \plot 4.5 -0.87 1.5 4.34 / \plot -3 3.46 3 3.46 / \plot -4 1.73 4 1.73 / \plot -5 0 5 0 / \plot -4 -1.73 4 -1.73 / \plot -3 -3.46 3 -3.46 / \put{$\bullet$} at 0 0 \put{\footnotesize$\displaystyle -t^{-1/2} \frac{t^{-1/2}-t^{1/2}}{1-Y^{\varphi^\vee-d}}\ t^{1/2} \ t^{1/2} \mathbf{1}$} at 0 -6 \setplotsymbol(.) \plot -0.3 1.1 -0.3 1.72 / \plot -0.3 1.72 -0.05 1.72 / \arrow <4pt> [.2,.67] from -0.05 1.72 to -0.05 1 % \plot -0.05 1 -0.4 0.75 / \plot -0.4 0.75 -0.27 0.57 / \arrow <4pt> [.2,.67] from -0.27 0.57 to 0.02 0.82 % \plot 0.02 0.82 0.33 0.62 / \plot 0.33 0.62 0.4 0.78 / \arrow <4pt> [.2,.67] from 0.4 0.78 to 0.15 1.05 \endpicture $$ Thus, \begin{align*} X^{-\alpha_2} &= t^{-\frac12} E_{-\alpha_2} - \frac{1-t}{1-q^2t^2} t^{-\frac22} E_{\alpha_2} - \frac{1-t}{1-qt}t^{-\frac22} E_{\alpha_1} + \frac{1-t}{1-qt}\frac{1-t}{1-qt} t^{-\frac32} E_\varphi - \frac{1-t}{1-qt^2} E_0. \end{align*} }\end{example}

https://arxiv.org/abs/1805.11928

On 2-absorbing ideals of commutative semirings

In this paper, we investigate 2-absorbing ideals of commutative semirings and prove that if $\mathfrak{a}$ is a nonzero proper ideal of a subtractive valuation semiring $S$ then $\mathfrak{a}$ is a 2-absorbing ideal of $S$ if and only if $\mathfrak{a}=\mathfrak{p}$ or $\mathfrak{a}=\mathfrak{p}^2$ where $\mathfrak{p}=\sqrt\mathfrak{a}$ is a prime ideal of $S$. We also show that each 2-absorbing ideal of a subtractive semiring $S$ is prime if and only if the prime ideals of $S$ are comparable and if $\mathfrak{p}$ is a minimal prime over a 2-absorbing ideal $\mathfrak{a}$, then $\mathfrak{am} = \mathfrak{p}$, where $\mathfrak{m}$ is the unique maximal ideal of $S$.

\section{Introduction} Though in different references, the term ``semiring'' is applied for different meanings, in many references (see the explanations on page 3 of the book \cite{Glazek2002}), a semiring is an algebraic structure $(S,+,\cdot,0,1)$ with the following properties: \begin{enumerate} \item $(S,+,0)$ is a commutative monoid, \item $(S,\cdot,1)$ is a monoid with $1\neq 0$, \item $a(b+c) = ab+ac$ and $(b+c)a = ba+ca$ for all $a,b,c\in S$, \item $a\cdot 0 = 0\cdot a = 0$ for all $a\in S$. \end{enumerate} A semiring $S$ is commutative if $ab = ba$ for all $a,b\in S$. In this paper, all semirings are commutative. Semirings are important ring-like structures with many applications in science and engineering \cite[p. 225]{Golan2003} and considered to be interesting generalizations of bounded distributive lattices and commutative rings with nonzero identities \cite[Example 1.5]{Golan1999(b)}. For more on ring-like algebraic structures and their applications, one may refer to \cite{AhsanMordesonShabir2012,Bistarelli2004,Glazek2002,Golan1999(a),Golan1999(b),Golan2003,GondranMinoux2008,HebischWeinert1998,KuichSalomaa1986,Pilz1983}. A nonempty subset $\mathfrak{a}$ of a semiring $S$ is said to be an ideal of $S$, if $x+y \in \mathfrak{a}$ for all $x,y \in \mathfrak{a}$ and $sx \in \mathfrak{a}$ for all $s \in S$ and $x \in \mathfrak{a}$ \cite{Bourne1951}. We denote the set of all ideals of $S$ by $\Id(S)$. It is easy to verify that $\Id(S)$ with the addition and multiplication of ideals, defined as follows, configures a semiring (check Example 1.4 and Proposition 6.29 in \cite{Golan1999(a)}): \begin{itemize} \item $I+J := \{a+b : a\in I, b\in J\},$ \item $I\cdot J := \{\Sigma^n_{i=1} a_i b_i: a_i \in I, b_i \in J, n\in \mathbb N\}.$ \end{itemize} An ideal $\mathfrak{a}$ of a semiring $S$ is called a proper ideal of the semiring $S$ if $\mathfrak{a} \neq S$. An ideal $\mathfrak{a}$ of a semiring $S$ is said to be subtractive if $x+y\in \mathfrak{a}$ and $x\in \mathfrak{a}$ imply $y\in \mathfrak{a}$ for all $x,y\in S$. A semiring $S$ is subtractive if each ideal of $S$ is subtractive \cite[\S 6]{Golan1999(b)}. It is clear that any (commutative) ring is subtractive. Let us recall that by definition a proper ideal $\mathfrak{p}$ of a semiring $S$ is prime if $\mathfrak{ab} \subseteq \mathfrak{p}$ implies either $\mathfrak{a} \subseteq \mathfrak{p}$ or $\mathfrak{b} \subseteq \mathfrak{p}$. It is, then, easy to observe that a proper ideal $\mathfrak{p}$ of a semiring $S$ is prime if and only if $ab\in \mathfrak{p}$ implies either $a\in \mathfrak{p}$ or $b\in \mathfrak{p}$ \cite[Corollary 7.6]{Golan1999(b)}. A proper ideal $\mathfrak{m}$ is maximal if $\mathfrak{m} \subseteq \mathfrak{a} \subseteq S$ implies that either $\mathfrak{m} = \mathfrak{a}$ or $\mathfrak{a} = S$ for any ideal $\mathfrak{a}$ of $S$. It is easy to prove that any proper ideal of a semiring is contained in a maximal ideal and any maximal ideal of a semiring is prime. A semiring $S$ is called local if it has only one maximal ideal. It is easy to show that a semiring $S$ is local if and only if $S-U(S)$ is an ideal of $S$, where $U(S)$ is the set of all unit elements of $S$ (see Example 6.1 and Proposition 6.61 in \cite{Golan1999(b)}). We also recall that an ideal $\mathfrak{q}$ of a semiring $S$ is called a primary ideal if $\mathfrak{q}$ is a proper ideal of $S$ and $xy\in \mathfrak{q}$ implies either $x\in \mathfrak{q}$ or $y^n \in \mathfrak{q}$ for some $n\in \mathbb N$ \cite[p. 92]{Golan1999(b)}. It is straightforward to see that if $\mathfrak{q}$ is a primary ideal of a semiring $S$, then $\sqrt\mathfrak{q}$ is the smallest prime ideal containing $\mathfrak{q}$. In this case, if we set $\mathfrak{p} = \sqrt\mathfrak{q}$, then we say that $\mathfrak{q}$ is $\mathfrak{p}$-primary \cite{Nasehpour2018(a)}. For more on primary ideals of semirings, refer to \cite{Nasehpour2018(a)}. Note that a semiring $S$ is a semidomain if it is multiplicatively cancellative, i.e. if $xy = xz$, and $x\neq 0$, then $y=z$ for all $x,y,z \in S$. A semiring $S$ is a valuation semiring if it is a semidomain and the set of its ideals $\Id(S)$ is totally ordered by inclusion. For more on valuation semirings, refer to \cite{Nasehpour2018(b)}. In Section \ref{sec:val}, we investigate ideals of valuation semirings. Let us recall that, in commutative algebra, a prime ideal $\mathfrak{p}$ of a domain $R$ is called divided if in $R_\mathfrak{p}$, we have the following equality: \[\set{z/s \;\delimsize\vert\;}#1 z\in \mathfrak{p}, s\notin \mathfrak{p}} = \set{z/1 \;\delimsize\vert\;}#1 z\in \mathfrak{p}}.\] It is easy to verify that a prime ideal $\mathfrak{p}$ of a ring $R$ is a divided prime ideal if and only if $\mathfrak{p} \subset (x)$ for every $x \in R-\mathfrak{p}$. We also note that a domain $D$ is called divided if each prime ideal of $D$ is divided \cite{Dobbs1976}. Divided domains are called AV-domains by Akiba in \cite{Akiba1967}. Inspired by this, we define a prime ideal $\mathfrak{p}$ of a semiring $S$ to be a divided prime ideal of $S$ if $\mathfrak{p} \subset (x)$ for every $x \in S-\mathfrak{p}$ (check Definition \ref{dividedprimeDef}). In Proposition \ref{valuationisdivided}, we prove that every valuation semiring is divided. We also recall that a proper ideal $\mathfrak{a}$ of a semiring $S$ is called 2-absorbing ideal of $S$ if $xyz\in \mathfrak{a}$ implies either $xy\in \mathfrak{a}$, or $yz \in \mathfrak{a}$, or $xz\in \mathfrak{a}$ \cite[Definition 2.1]{Darani2012}. Section \ref{sec:2-Abs} is devoted to studying 2-absorbing ideals of semirings. For example in Theorem \ref{p2subsetofideal}, we prove that if $\mathfrak{a}$ is a 2-absorbing ideal of a subtractive semiring $S$, then one of the following statements must hold: \begin{enumerate} \item $\sqrt{\mathfrak{a}}=\mathfrak{p}$ is a prime ideal of $S$ such that $\mathfrak{p}^2 \subseteq \mathfrak{a}$. \item $\sqrt{\mathfrak{a}} = \mathfrak{p}_1 \cap \mathfrak{p}_2 $, $\mathfrak{p}_1 \mathfrak{p}_2 \subseteq \mathfrak{a}$ and $(\sqrt{\mathfrak{a}})^2 \subseteq \mathfrak{a}$ where $\mathfrak{p}_1$, $\mathfrak{p}_2$ are the only distinct prime ideals of $S$ that are minimal over $\mathfrak{a}$. \end{enumerate} We also show in Theorem \ref{dividedprimeThm1} that if $\mathfrak{p}$ is a nonzero divided prime ideal of a subtractive semiring $S$ and $\mathfrak{a}$ is an ideal of $S$ such that $\sqrt{\mathfrak{a}}=\mathfrak{p}$, then the following statements are equivalent: \begin{enumerate} \item $\mathfrak{a}$ is a 2-absorbing ideal of $S$; \item $\mathfrak{a}$ is a $\mathfrak{p}$-primary ideal of $S$ such that $\mathfrak{p}^2 \subseteq \mathfrak{a}$. \end{enumerate} We also prove that if $\mathfrak{p}$ is a nonzero divided prime ideal of a subtractive semidomain $S$, then $\mathfrak{p}^2$ is a 2-absorbing ideal of $S$ (see Theorem \ref{dividedprimeThm2}). Finally, we add that with the help of the results in Section \ref{sec:val}, we show in Theorem \ref{2-absorbing} that if $S$ is a subtractive valuation semiring and $\mathfrak{a}$ is a nonzero proper ideal of $S$, then the following statements are equivalent: \begin{enumerate} \item $\mathfrak{a}$ is a 2-absorbing ideal of $S$; \item $\mathfrak{a}$ is a $\mathfrak{p}$-primary ideal of $S$ such that $\mathfrak{p}^2\subseteq\mathfrak{a}$; \item $\mathfrak{a}=\mathfrak{p}$ or $\mathfrak{a}=\mathfrak{p}^2$ where $\mathfrak{p}=\sqrt\mathfrak{a}$ is a prime ideal of $S$. \end{enumerate} Note that examples for subtractive valuation semirings include the semiring $\Id(D)$, where $D$ is a Dedekind domain (see Theorem 3.8 and Proposition 3.10 in \cite{Nasehpour2018(b)}).\\ It is obvious that each prime ideal of a semiring is 2-absorbing, though the converse of this statement is not correct. Therefore, it is natural to investigate those semirings in which every 2-absorbing ideal is prime. We define a semiring $S$ to be a 2-AB semiring if every 2-absorbing ideal of $S$ is prime (see Definition \ref{2-ABdef}). We devote Section \ref{sec:2-AB} to 2-AB semirings and in Theorem \ref{2-ABThm2}, we show that if $S$ is a subtractive semiring, then the following statements are equivalent: \begin{enumerate} \item The semiring $S$ is 2-AB. \item (a) The prime ideals of $S$ are comparable and (b) if $\mathfrak{p}$ is a minimal prime over a 2-absorbing ideal $\mathfrak{a}$, then $\mathfrak{am} = \mathfrak{p}$, where $\mathfrak{m}$ is the unique maximal ideal of $S$. \item (a) The prime ideals of $S$ are comparable and (b) for every prime ideal $\mathfrak{p}$ of $S$, $2-\operatorname{Min}_S(\mathfrak{p}^2)=\set{\mathfrak{p}}$. \end{enumerate} Assume that $\mathfrak{a}$ is an ideal of a semiring $S$. We define a 2-absorbing ideal $\mathfrak{p}$ of $S$ to be a minimal 2-absorbing ideal over $\mathfrak{a}$ if there is not a 2-absorbing ideal $\mathfrak{q}$ of $S$ such that $\mathfrak{a}\subseteq\mathfrak{q}\subset\mathfrak{p}$. We denote the set of minimal 2-absorbing ideals over $\mathfrak{a}$ by $2-\operatorname{Min}_S(\mathfrak{a})$ (See Definition \ref{2-MinDef}). It is worth mentioning that in Section \ref{sec:2-AB}, we also prove that if $S$ is a subtractive valuation semiring, then, $S$ is 2-AB if and only if $\mathfrak{p}^2 = \mathfrak{p}$ for every prime ideal $\mathfrak{p}$ of $S$ (check Theorem \ref{2-ABThm3}). Finally, the reader is warned that we use ``$\subseteq$'' for inclusion and ``$\subset$'' for strict inclusion \cite[p. 17]{Monk1969}. \section{Ideals of valuation semirings}\label{sec:val} Let us recall that a semidomain (i.e. multiplicatively cancellative semiring) is a valuation semiring if and only if the ideals of $S$ are totally ordered by inclusion \cite[Theorem 2.4]{Nasehpour2018(b)}. We also recall that by a fractional ideal of a semidomain $S$, we mean a subset $I$ of the semifield of fractions $F(S)$ such that the following conditions are satisfied: \begin{enumerate} \item I is an $S$-subsemimodule of $F(S)$, that is, if $a, b \in I$ and $ s \in S$, then $a + b \in I $ and $sa \in I$. \item There exists a nonzero element $d\in S$ such that $dI \subseteq S$. \end{enumerate} Also, a fractional ideal $I$ of a semidomain $S$ is invertible if there exists a fractional ideal $J$ of $S$ such that $IJ=S$ \cite{GhalandarzadehNasehpourRazavi2017}. It is clear that any invertible ideal $I$ is cancellation, i.e. $IJ =IK$ implies $J=K$, for all ideals $J$ and $K$ of $S$ \cite{LaGrassa1995}. First, we prove the following lemma: \begin{lemma} \label{idealsofvaluation1} Let $S$ be a valuation semiring and $\mathfrak{a}$ a proper ideal of $S$. Then, the following statements hold: \begin{enumerate} \item $\bigcap^{\infty}_{n=1} \mathfrak{a}^n = \mathfrak{c}$ is a prime ideal of $S$. \item If $\mathfrak{b}$ is an ideal of $S$ such that $\mathfrak{a} \subset \sqrt\mathfrak{b}$, then $\mathfrak{b}$ contains a power of $\mathfrak{a}$. \end{enumerate} \begin{proof} (1): Suppose $s,t\in S$ are not elements of $\bigcap^{\infty}_{n=1} \mathfrak{a}^n = \mathfrak{c}$. Therefore, there are natural numbers $m,n \in \mathbb N$ such that $s \notin \mathfrak{a}^m$ and $t \notin \mathfrak{a}^n$. Since $S$ is a valuation semiring, $\mathfrak{a}^m \subset (s)$ and $\mathfrak{a}^n \subset (t)$. Since $(t)$ is an invertible ideal of $S$ \cite[Proposition 1.4]{GhalandarzadehNasehpourRazavi2017}, $\mathfrak{a}^m (t) \subset (st)$. On the other hand, since $\mathfrak{a}^m \subset (s)$, $\mathfrak{a}^{m+n} \subseteq \mathfrak{a}^n (s)$. So, $\mathfrak{a}^{m+n} \subset (st)$, which means that $st \notin \mathfrak{c}$. (2): If $\mathfrak{b}$ contains no power of $\mathfrak{a}$, then $\mathfrak{b} \subseteq \mathfrak{a}^n$, for each $n\in \mathbb N$. Therefore, $\mathfrak{b} \subseteq \mathfrak{c}$. Since $\mathfrak{c}$ is prime, $\sqrt\mathfrak{b} \subseteq \mathfrak{c} \subseteq \mathfrak{a}$. This finishes the proof. \end{proof} \end{lemma} \begin{theorem} \label{idealsofvaluation2} Let $\mathfrak{p}$ be a proper prime ideal of a valuation semiring $S$. \begin{enumerate} \item If $\mathfrak{q}$ is $\mathfrak{p}$-primary and $x \in S-\mathfrak{p}$, then $\mathfrak{q}=\mathfrak{q}(x)$. \item A finite nonempty product of $\mathfrak{p}$-primary ideals of $S$ is a $\mathfrak{p}$-primary ideal. If $\mathfrak{p} \neq \mathfrak{p}^2$, then the only $\mathfrak{p}$-primary ideals of $S$ are powers of $\mathfrak{p}$. \end{enumerate} \begin{proof} (1): As $x \notin \mathfrak{p}$, $\mathfrak{q} \subset (x)$. Suppose that $K$ is the quotient semifield of $S$ and $A=\set{ y \colon y \in K, xy \in \mathfrak{q} }$. As $\mathfrak{q} \subset (x)$, $A$ is a subset of $S$. Now, it is easy to verify that $A$ is, in fact, an ideal of $S$. Our claim is that $\mathfrak{q}=A(x)$. First we prove that $A(x) \subseteq \mathfrak{q}$. Let $s\in A(x)$. Since $A$ is an ideal of $S$, we have that $s =ax$, where $a\in A$. But $s=ax$ is an element of $\mathfrak{q}$ by the definition of $A$. The proof of the other containment is very easy if one considers the definition of $A$ and this point that $\mathfrak{q} \subset (x)$. Now, we prove that $A \subseteq \mathfrak{q}$. Let $a\in A$. So, $ax\in \mathfrak{q}$, by definition of $A$. Since $\mathfrak{q}$ is $\mathfrak{p}$-primary, we have that either $a\in \mathfrak{q}$ or $x\in \mathfrak{p}=\sqrt{\mathfrak{q}}$. But by assumption, $x \notin \mathfrak{p}$. So, $a\in \mathfrak{q}$ and this shows that $A \subseteq \mathfrak{q}$. Thus $\mathfrak{q}=A$ and $\mathfrak{q}=\mathfrak{q}(x)$. \newline (2): Suppose that $\mathfrak{q}_1,\mathfrak{q}_2$ are $\mathfrak{p}$-primary ideals of $S$. It is straightforward to see that $\sqrt{\mathfrak{q}_1\mathfrak{q}_2}=\mathfrak{p}$. Suppose that $x,y \in S$ and $xy \in \mathfrak{q}_1 \mathfrak{q}_2$ and $x \notin \mathfrak{p}$. By the first part of this theorem, we have $\mathfrak{q}_1=\mathfrak{q}_1(x)$. Therefore, $xy \in (x)\mathfrak{q}_1\mathfrak{q}_2$. As $S$ is a semidomain, $y\in \mathfrak{q}_1 \mathfrak{q}_2$. Thus $\mathfrak{q}_1\mathfrak{q}_2$ is $\mathfrak{p}$-primary. \newline Now let $\mathfrak{p} \neq \mathfrak{p}^2$ and $\mathfrak{q}$ be a $\mathfrak{p}$-primary ideal of $S$. By Lemma \ref{idealsofvaluation1}, $\mathfrak{q}$ contains a power of $\mathfrak{p}^2$ and so contains a power of $\mathfrak{p}$. Thus, there is a positive integer $m$ such that $\mathfrak{p}^m\subseteq\mathfrak{q}$ and $\mathfrak{p}^{m-1} \nsubseteq \mathfrak{q}$. Suppose that $x \in \mathfrak{p}^{m-1}, x \notin \mathfrak{q}$. Clearly, $\mathfrak{q}\subset(x)$. We define $A=\set{ y \;\delimsize\vert\;}#1 y \in K, xy \in \mathfrak{q} }$, so $\mathfrak{q}=A(x)$. As $\mathfrak{q}$ is $\mathfrak{p}$-primary and $x \notin \mathfrak{q}$, we have $A \subseteq \mathfrak{p}$. Thus, $\mathfrak{q}=A(x)\subseteq\mathfrak{p}(x)\subseteq\mathfrak{p}^m$, and so we conclude that $\mathfrak{q}=\mathfrak{p}^m$. \end{proof} \end{theorem} Now, we bring the definition of divided primes and semirings: \begin{definition} \label{dividedprimeDef} We define a prime ideal $\mathfrak{p}$ of a semiring $S$ to be a divided prime ideal of $S$ if $\mathfrak{p} \subset (x)$ for every $x \in S-\mathfrak{p}$. We call a semiring $S$ divided if each prime ideal of $S$ is divided. \end{definition} \begin{proposition} \label{valuationisdivided} Any valuation semiring is divided. \begin{proof} Let $S$ be a valuation semiring and $\mathfrak{p}$ be a prime ideal of $S$. Also, let $x\notin \mathfrak{p}$. This means that the principal ideal $(x)$ is not a subset of $\mathfrak{p}$. Since in valuation semirings, ideals are totally ordered by inclusion \cite[Theorem 2.4]{Nasehpour2018(b)}, $\mathfrak{p} \subset (x)$ and the proof is complete. \end{proof} \end{proposition} One may define the localization of semirings similar to ring theory \cite{Kim1985,Nasehpour2018(a)}. It is, then, straightforward to see that a prime ideal $\mathfrak{p}$ of a semiring $S$ is divided if and only if in $S_\mathfrak{p}$, we have the following equality: \[\set{z/s \;\delimsize\vert\;}#1 z\in \mathfrak{p}, s\notin \mathfrak{p}} = \set{z/1 \;\delimsize\vert\;}#1 z\in \mathfrak{p}}.\] By using this point, one can easily prove the following: \begin{proposition} Let $S$ be a divided semiring. Then, each arbitrary ideal of $S$ is comparable with each prime ideal of $S$. In particular, prime ideals of $S$ are comparable and $S$ is local. \end{proposition} \section{On 2-Absorbing Ideals of Semirings}\label{sec:2-Abs} The concept of 2-absorbing ideals of semirings was introduced by Darani \cite[Definition 2.1]{Darani2012}. His definition is the semiring version of the 2-absorbing ideals of rings introduced by Badawi \cite{Badawi2007}. We recall this concept in the following: \begin{definition} \label{2-absorbingdef} A proper ideal $\mathfrak{a}$ of a semiring $S$ is called 2-absorbing ideal of $S$ if $xyz\in \mathfrak{a}$ implies either $xy\in \mathfrak{a}$, or $yz \in \mathfrak{a}$, or $xz\in \mathfrak{a}$. \end{definition} Here is a good place to mention that prime ideals of a semiring are 2-absorbing. Now, we give some other methods to make 2-absorbing ideals: \begin{proposition} \label{2-absorbingintersection} If $\mathfrak{p}_1$ and $\mathfrak{p}_2$ are both prime ideals of a semiring $S$, then $\mathfrak{p}_1 \cap \mathfrak{p}_2$ is a 2-absorbing ideal of $S$. \begin{proof} Let $x_1 x_2 x_3 \in \mathfrak{p}_1 \cap \mathfrak{p}_2$ for the elements $x_1, x_2, x_3 \in S$. Then, $x_1 x_2 x_3 \in \mathfrak{p}_1$ and $x_1 x_2 x_3 \in \mathfrak{p}_2$. Since $\mathfrak{p}_1$ is prime, either $x_1 \in \mathfrak{p}_1$, or $x_2 \in \mathfrak{p}_1$, or $x_3 \in \mathfrak{p}_1$. Similarly, one of these three elements is in $\mathfrak{p}_2$. Suppose that $x_1 \in \mathfrak{p}_1$. Now, if $x_1\in \mathfrak{p}_2$, then $x_1 x_2$ is in the both $\mathfrak{p}_1$ and $\mathfrak{p}_2$, so $x_1x_2 \in \mathfrak{p}_1 \cap \mathfrak{p}_2$. If $x_2\in \mathfrak{p}_2$ then $x_1 x_2\in \mathfrak{p}_1 \cap \mathfrak{p}_2$. And if $x_3\in \mathfrak{p}_2$, then $x_1 x_3\in \mathfrak{p}_1 \cap \mathfrak{p}_2$. The same proof is valid for the cases $x_2 \in \mathfrak{p}_1$ and $x_3 \in \mathfrak{p}_1$. Therefore, there are two of these three elements whose product is always in $\mathfrak{p}_1 \cap \mathfrak{p}_2$. Hence, $\mathfrak{p}_1 \cap \mathfrak{p}_2$ is a 2-absorbing ideal of $S$ and the proof is complete. \end{proof} \end{proposition} \begin{proposition} \label{2-absorbingradical} Let $\mathfrak{a}$ be a 2-absorbing ideal of a semiring $S$. Then, $\sqrt{\mathfrak{a}}$ is a 2-absorbing ideal of $S$ and $s^2 \in \mathfrak{a}$ for each $s\in \sqrt{\mathfrak{a}}$. \begin{proof} The proof is just a mimicking of the proof of Theorem 2.1 in \cite{Badawi2007} and therefore, omitted. \end{proof} \end{proposition} \begin{proposition} \label{2-absorbingmulti} Let $(S, \mathfrak{m})$ be a local semiring. For each prime ideal $\mathfrak{p}$ of $S$, $\mathfrak{pm}$ is a 2-absorbing ideal of $S$. Furthermore, $\mathfrak{pm}$ is a prime ideal of $S$ if and only if $\mathfrak{pm} = \mathfrak{p}$. \begin{proof} Let $s_1 s_2 s_3 \in \mathfrak{pm}$. Since $\mathfrak{pm} \subseteq \mathfrak{p}$ and $\mathfrak{p}$ is a prime ideal of $S$, we have either $s_1 \in \mathfrak{p}$, or $s_2 \in \mathfrak{p}$, or $s_3 \in \mathfrak{p}$. Now let for the moment $s_1 \in \mathfrak{p}$. If $s_2$ is a unit, then it is obvious that $s^{-1}_2 (s_1 s_2 s_3) = s_1 s_3$ is an element of $\mathfrak{pm}$. If not, then since $S$ is local, $s_2$ is an element of $\mathfrak{m}$. So, $s_1 s_2 \in \mathfrak{pm}$. Now let $\mathfrak{pm}$ be a prime ideal of $S$. Clearly, $\mathfrak{pm} \subseteq \mathfrak{p}$. Assume $s\in \mathfrak{p}$. It is obvious that $s\in \mathfrak{m}$ and so, $s^2 \in \mathfrak{pm}$. But $\mathfrak{pm}$ is prime, so $s\in \mathfrak{pm}$ and the proof is complete. \end{proof} \end{proposition} Let us recall that a prime ideal $\mathfrak{p}$ of a semiring $S$ is said to be a minimal prime ideal over an ideal $\mathfrak{a}$ of $S$ if it is minimal among all prime ideals containing $\mathfrak{a}$. We collect all minimal prime ideals over an ideal $\mathfrak{a}$ in $\operatorname{Min}(\mathfrak{a})$. A subset $W$ of a semiring is called a multiplicatively closed set (for short MC-set) if $1_S \in W$ and if $w_1$ and $w_2$ are elements of $S$, then $w_1 w_2 \in W$. We also recall the following: \begin{theorem}[Theorem 3.5 in \cite{Nasehpour2018}] \label{minimalprimehuckaba} Let $\mathfrak{p} \supseteq \mathfrak{a}$ be ideals of a semiring $S$, where $\mathfrak{p}$ is prime. Then, the following statements are equivalent: \begin{enumerate} \item $\mathfrak{p}$ is a minimal prime ideal of $\mathfrak{a}$. \item $S-\mathfrak{p}$ is an MC-set maximal with respect to being disjoint to $\mathfrak{a}$. \item For each $x\in \mathfrak{p}$, there is some $y\in S-\mathfrak{p}$ and a nonnegative integer $i$ such that $yx^i \in \mathfrak{a}$. \end{enumerate} \end{theorem} \begin{theorem} \label{minimalprimeover2-ab} Let $\mathfrak{a}$ be a 2-absorbing ideal of a semiring $S$ such that each minimal prime ideal over $\mathfrak{a}$ is subtractive. Then, $|\operatorname{Min}(\mathfrak{a})| \leq 2$. \begin{proof} By considering Theorem \ref{minimalprimehuckaba} and this point that each minimal prime ideal over the ideal $\mathfrak{a}$ is subtractive, the proof is nothing but a mimicking of the proof of Theorem 2.3 in \cite{Badawi2007}. \end{proof} \end{theorem} Let us recall that, by definition, a semiring $S$ is weak Gaussian if $c(f)c(g) \subseteq \sqrt {c(fg)} $ for all $f,g \in S[X]$, where the content of a polynomial $f\in S[X]$, denoted by $c(f)$, is an ideal of $S$ generated by the coefficients of $f$. Note that a semiring $S$ is weak Gaussian if and only if each prime ideal of $S$ is subtractive \cite[Definition 18, Theorem 19]{Nasehpour2016}. \begin{corollary} Let $S$ be a weak Gaussian semiring and $\mathfrak{a}$ a 2-absorbing ideal of $S$. Then, $|\operatorname{Min}(\mathfrak{a})| \leq 2$. \end{corollary} \begin{theorem} \label{p2subsetofideal} Let $\mathfrak{a}$ be a 2-absorbing ideal of a subtractive semiring $S$. Then, one of the following statements must hold: \begin{enumerate} \item $\sqrt{\mathfrak{a}}=\mathfrak{p}$ is a prime ideal of $S$ such that $\mathfrak{p}^2 \subseteq \mathfrak{a}$. \item $\sqrt{\mathfrak{a}} = \mathfrak{p}_1 \cap \mathfrak{p}_2 $, $\mathfrak{p}_1 \mathfrak{p}_2 \subseteq \mathfrak{a}$ and $(\sqrt{\mathfrak{a}})^2 \subseteq \mathfrak{a}$ where $\mathfrak{p}_1$, $\mathfrak{p}_2$ are the only distinct prime ideals of $S$ that are minimal over $\mathfrak{a}$. \end{enumerate} \end{theorem} \begin{proof} By considering Theorem \ref{minimalprimeover2-ab} and this point that $S$ is a subtractive semiring, the proof is nothing but a mimicking of the proof of Theorem 2.4 in \cite{Badawi2007}. \end{proof} \begin{theorem} \label{m2twoabsorbing} Let $\mathfrak{a}$ be a $\mathfrak{p}$-primary ideal of a subtractive semiring $S$. Then, $\mathfrak{a}$ is a 2-absorbing ideal of $S$ if and only if $\mathfrak{p}^2 \subseteq \mathfrak{a}$. In particular, for each maximal ideal $\mathfrak{m}$ of $S$, $\mathfrak{m}^2$ is a 2-absorbing ideal of $S$. \begin{proof} Let $\mathfrak{a}$ be a 2-absorbing ideal of a subtractive semiring $S$. Then, by Theorem \ref{p2subsetofideal}, we have $\mathfrak{p}^2 \subseteq \mathfrak{a}$. Conversely, let $xyz \in \mathfrak{a}$ and $\mathfrak{p}^2 \subseteq \mathfrak{a}$. If $x \in \mathfrak{a}$ or $yz \in \mathfrak{a}$, then there is nothing to prove. If neither $x \in \mathfrak{a}$ nor $yz \in \mathfrak{a}$, as $\mathfrak{a}$ is a $\mathfrak{p}$-primary ideal of $S$, then $x \in \mathfrak{p}$ and $yz \in \mathfrak{p}$. Therefore, either $x,y \in \mathfrak{p}$ or $x,z \in \mathfrak{p}$. Now from the assumption $\mathfrak{p}^2 \subseteq \mathfrak{a}$, we get that either $xy \in \mathfrak{a}$ or $xz \in \mathfrak{a}$ and this completes the proof. \end{proof} \end{theorem} \begin{theorem} \label{dividedprimeThm1} Let $\mathfrak{p}$ be a nonzero divided prime ideal of a subtractive semiring $S$ and $\mathfrak{a}$ be an ideal of $S$ such that $\sqrt{\mathfrak{a}}=\mathfrak{p}$. Then, the following statements are equivalent: \begin{enumerate} \item $\mathfrak{a}$ is a 2-absorbing ideal of $S$; \item $\mathfrak{a}$ is a $\mathfrak{p}$-primary ideal of $S$ such that $\mathfrak{p}^2 \subseteq \mathfrak{a}$. \end{enumerate} \end{theorem} \begin{proof} (1) $\Rightarrow$ (2): Let $\mathfrak{a}$ be a 2-absorbing ideal of $S$. By using Theorem \ref{p2subsetofideal} and considering this point that $\sqrt{\mathfrak{a}}=\mathfrak{p}$ is a nonzero prime ideal of $S$, we have $\mathfrak{p}^2 \subseteq \mathfrak{a}$. Suppose that for some $x,y \in S$, we have $xy \in \mathfrak{a}$ and $y \notin \mathfrak{p}$. As $\mathfrak{p}$ is a divided prime ideal of $S$ and $x \in \mathfrak{p}$, we conclude that $x=my$ for some $m \in S$. Therefore, $xy=my^2 \in \mathfrak{a}$. Since $y^2 \notin \mathfrak{a}$ and $\mathfrak{a}$ is a 2-absorbing ideal of $S$, we have $my=x \in \mathfrak{a}$. Thus, $\mathfrak{a}$ is a $\mathfrak{p}$-primary ideal of $S$. \newline (2) $\Rightarrow$ (1): Using Theorem \ref{m2twoabsorbing}, the proof of this implication is obvious. \end{proof} \begin{theorem} \label{dividedprimeThm2} Let $\mathfrak{p}$ be a nonzero divided prime ideal of a subtractive semidomain $S$. Then, $\mathfrak{p}^2$ is a 2-absorbing ideal of $S$. \begin{proof} By Theorem \ref{dividedprimeThm1}, we only need to show that $\mathfrak{p}^2$ is a $\mathfrak{p}$-primary ideal of $S$. Let $st\in \mathfrak{p}^2$, while $s\notin \mathfrak{p}$. Suppose that $st = \Sigma^n_{i=1} x_i y_i$, where $x_i,y_i \in \mathfrak{p}$. Since $\mathfrak{p}$ is a divided prime ideal of $S$ and $s\notin \mathfrak{p}$, we have $\mathfrak{p} \subset (s)$. Now since $x_i \in \mathfrak{p}$, we have $x_i = s z_i$, for each $1\leq i \leq n$. On the other hand, since $s\notin \mathfrak{p}$, we have $z_i \in \mathfrak{p}$. Now $st = s(\Sigma^n_{i=1} z_i y_i)$. By assumption $S$ is multiplicatively cancellative. So, $t = \Sigma^n_{i=1} z_i y_i$, which means that $t\in \mathfrak{p}^2$. Hence, $\mathfrak{p}^2$ is $\mathfrak{p}$-primary and the proof is complete. \end{proof} \end{theorem} \begin{question} Is there any example of a semiring $S$ that contains a prime ideal $\mathfrak{p}$ for which $\mathfrak{p}^2$ is not 2-absorbing? \end{question} Related to the above question, we also invite the reader to check Corollary \ref{not-2-absorbing}. Now we proceed to investigate 2-absorbing ideals of valuation semirings. For doing this, we need to prove some statements for the ideals of valuation semirings. \begin{theorem} \label{2-absorbing} Let $S$ be a subtractive valuation semiring and $\mathfrak{a}$ be a nonzero proper ideal of $S$. Then, the following statements are equivalent: \begin{enumerate} \item $\mathfrak{a}$ is a 2-absorbing ideal of $S$; \item $\mathfrak{a}$ is a $\mathfrak{p}$-primary ideal of $S$ such that $\mathfrak{p}^2\subseteq\mathfrak{a}$ where $\mathfrak{p} = \sqrt{\mathfrak{a}}$ is a prime ideal of $S$; \item $\mathfrak{a}=\mathfrak{p}$ or $\mathfrak{a}=\mathfrak{p}^2$ where $\mathfrak{p}=\sqrt\mathfrak{a}$ is a prime ideal of $S$. \end{enumerate} \begin{proof} (1) $\Rightarrow$ (2): Let $\mathfrak{a}$ be a 2-absorbing ideal of $S$. It is easy to verify that $\sqrt\mathfrak{a}=\mathfrak{p}$ is a prime ideal of $S$. Since $S$ is a valuation semiring, by Proposition \ref{valuationisdivided}, $S$ is a divided semidomain. Now by using Theorem \ref{dividedprimeThm1}, $\mathfrak{a}$ is a $\mathfrak{p}$-primary ideal of $S$ such that $\mathfrak{p}^2\subseteq\mathfrak{a}$. \newline (2) $\Rightarrow$ (3): Let $\mathfrak{a}$ be a $\mathfrak{p}$-primary ideal of $S$ such that $\mathfrak{p}^2 \subseteq \mathfrak{a}$. As $S$ is a valuation semiring, by using Theorem \ref{idealsofvaluation2} we can conclude that either $\mathfrak{a}=\mathfrak{p}$ or $\mathfrak{a}=\mathfrak{p}^2$. \newline (3) $\Rightarrow$ (1): Suppose that either $\mathfrak{a}=\mathfrak{p}$ or $\mathfrak{a}=\mathfrak{p}^2$ where $\mathfrak{p}=\sqrt{\mathfrak{a}}$ is a prime ideal of $S$. If $\mathfrak{a}=\mathfrak{p}$, then $\mathfrak{a}$ is a 2-absorbing ideal of $S$. If $\mathfrak{a}=\mathfrak{p}^2$, then by using Theorem \ref{dividedprimeThm2}, $\mathfrak{a}$ is a 2-absorbing ideal of $S$. \end{proof} \end{theorem} \section{2-AB semirings}\label{sec:2-AB} Let us recall that a ring $R$ is called 2-AB if every 2-absorbing ideal of $R$ is prime \cite[Definition 2.1]{BennisFahid2017}. Inspired by this, we give the following definition: \begin{definition} \label{2-ABdef} We define a semiring $S$ to be a 2-AB semiring if every 2-absorbing ideal of $S$ is prime. \end{definition} Let us recall that a 2-absorbing ideal $\mathfrak{p}$ of a ring $R$ is said to be a minimal 2-absorbing ideal over an ideal $\mathfrak{a}$ of $R$ if it is minimal among all 2-absorbing ideals containing $\mathfrak{a}$ \cite{MoghimiNaghani2016}. The following is the semiring version of the definition of minimal 2-absorbing ideals in commutative rings: \begin{definition} \label{2-MinDef} Let $\mathfrak{a}$ be an ideal of a semiring $S$. We define a 2-absorbing ideal $\mathfrak{p}$ of $S$ to be a minimal 2-absorbing ideal over $\mathfrak{a}$ if there is not a 2-absorbing ideal $\mathfrak{q}$ of $S$ such that $\mathfrak{a}\subseteq\mathfrak{q}\subset\mathfrak{p}$. We denote the set of minimal 2-absorbing ideals over $\mathfrak{a}$ by $2-\operatorname{Min}_S(\mathfrak{a})$. \end{definition} \begin{lemma} \label{2minsemiring} Let $S$ be a subtractive semiring such that the prime ideals of $S$ are comparable and $\mathfrak{m}$ be the unique maximal ideal of $S$. Then, the following statements are equivalent: \begin{enumerate} \item For every minimal prime ideal $\mathfrak{p}$ over a 2-absorbing ideal $\mathfrak{a}$, $\mathfrak{am}=\mathfrak{p}$; \item For every prime ideal $\mathfrak{p}$ of $S$, $2-\operatorname{Min}_S(\mathfrak{p}^2)=\set{\mathfrak{p}}$. \end{enumerate} Moreover, if one of the above equivalent statements holds, then $S$ is 2-AB. \begin{proof} $(1) \Rightarrow (2)$: Let $\mathfrak{p}$ be a prime ideal of $S$ and $\mathfrak{b}$ be a 2-absorbing ideal of $S$ such that $\mathfrak{b}\in 2-\operatorname{Min}_S(\mathfrak{p}^2)$. First, we prove that $\mathfrak{p} \in \operatorname{Min}_S(\mathfrak{b})$. Suppose there exists a prime ideal $\mathfrak{q}$ such that $\mathfrak{b} \subseteq \mathfrak{q} \subseteq \mathfrak{p}$. Clearly, $\mathfrak{p}^2 \subseteq \mathfrak{b} \subseteq \mathfrak{q} \subseteq \mathfrak{p}$. Let $x\in \mathfrak{p}$. So, we have $x^2 \in \mathfrak{p}^2$ and therefore, $x^2 \in \mathfrak{q}$. Since $\mathfrak{q}$ is prime, $x \in \mathfrak{q}$. Thus, $\mathfrak{p}=\mathfrak{q}$. But $\mathfrak{b}\mathfrak{m} \subseteq \mathfrak{b} \subseteq \mathfrak{p}$. Now since by hypothesis $\mathfrak{b}\mathfrak{m}=\mathfrak{p}$, we have $\mathfrak{b}=\mathfrak{p}$. \newline $(2) \Rightarrow (1)$: Suppose that $\mathfrak{p}$ is a minimal prime ideal over a 2-absorbing ideal $\mathfrak{a}$. Since $\sqrt\mathfrak{a}$ is a prime ideal of $S$ and $\sqrt\mathfrak{a} \subseteq \mathfrak{p}$, we have $\sqrt\mathfrak{a} = \mathfrak{p}$. So, by Theorem \ref{p2subsetofideal}, $\mathfrak{p}^2 \subseteq \mathfrak{a} \subseteq \mathfrak{p}$. Now by hypothesis $2-\operatorname{Min}_S(\mathfrak{p}^2)=\set{\mathfrak{p}}$. Therefore, $\mathfrak{a}=\mathfrak{p}$. By Proposition \ref{2-absorbingmulti}, $\mathfrak{pm}$ is 2-absorbing. Now, $\mathfrak{p}^2 \subseteq \mathfrak{pm} \subseteq \mathfrak{p}$. So, $\mathfrak{am}=\mathfrak{pm}=\mathfrak{p}$. Now, let the statement (1) hold. If $\mathfrak{a}$ is a 2-absorbing ideal of $S$, then there is a minimal prime ideal $\mathfrak{p}$ over $\mathfrak{a}$ and clearly, we have $\mathfrak{am} \subseteq \mathfrak{a} \subseteq \mathfrak{p}$, which implies that $\mathfrak{a} = \mathfrak{p}$. \end{proof} \end{lemma} \begin{corollary} Let $S$ be a subtractive divided semidomain with the unique maximal ideal $\mathfrak{m}$. Then, the following statements are equivalent: \begin{enumerate} \item For every minimal prime ideal $\mathfrak{p}$ over a 2-absorbing ideal $\mathfrak{a}$, $\mathfrak{am}=\mathfrak{p}$; \item For every prime ideal $\mathfrak{p}$ of $S$, $2-\operatorname{Min}_S(\mathfrak{p}^2)=\set{\mathfrak{p}}$. \end{enumerate} Moreover, if one of the above equivalent statements holds, then $S$ is 2-AB. \end{corollary} Now, we prove the following lemma: \begin{lemma} \label{2-ABThm1} Let $S$ be a 2-AB semiring. Then, the following statements hold: \begin{enumerate} \item Prime ideals of $S$ are comparable; in particular, $S$ is a local semiring. \item Let $\mathfrak{m}$ be the unique maximal ideal of $S$. If $\mathfrak{p}$ is a minimal prime over a 2-absorbing ideal $\mathfrak{a}$, then $\mathfrak{am} = \mathfrak{p}$. In particular, $\mathfrak{m}$ is an idempotent ideal of $S$. \end{enumerate} \begin{proof} (1): Let $\mathfrak{p}_1$ and $\mathfrak{p}_2$ be prime ideals of $S$. By Proposition \ref{2-absorbingintersection}, $\mathfrak{p}_1 \cap \mathfrak{p}_2$ is a 2-absorbing ideal of $S$ and since $S$ is 2-AB, $\mathfrak{p} = \mathfrak{p}_1 \cap \mathfrak{p}_2$ is prime. Therefore, either $\mathfrak{p}_1 = \mathfrak{p}_1 \cap \mathfrak{p}_2$ or $\mathfrak{p}_2 = \mathfrak{p}_1 \cap \mathfrak{p}_2$. This means that prime ideals of $S$ are comparable and so, $S$ is local. (2): Let $\mathfrak{p}$ be a minimal prime over a 2-absorbing ideal $\mathfrak{a}$. Since $S$ is 2-AB, $\mathfrak{a}$ is prime and so, $\mathfrak{p} = \mathfrak{a}$. Now, by Proposition \ref{2-absorbingmulti}, $\mathfrak{am} = \mathfrak{p}$ and this finishes the proof. \end{proof} \end{lemma} \begin{theorem} \label{2-ABThm2} Let $S$ be a subtractive semiring. Then the following statements are equivalent: \begin{enumerate} \item \label{2-ABThm2-1} The semiring $S$ is 2-AB. \item \label{2-ABThm2-2} (a) The prime ideals of $S$ are comparable and (b) if $\mathfrak{p}$ is a minimal prime over a 2-absorbing ideal $\mathfrak{a}$, then $\mathfrak{am} = \mathfrak{p}$, where $\mathfrak{m}$ is the unique maximal ideal of $S$. \item \label{2-ABThm2-3} (a) The prime ideals of $S$ are comparable and (b) for every prime ideal $\mathfrak{p}$ of $S$, $2-\operatorname{Min}_S(\mathfrak{p}^2)=\set{\mathfrak{p}}$. \end{enumerate} \begin{proof} According to Lemma \ref{2-ABThm1}, $(1) \Rightarrow (2)$. For the proof of $(2) \Rightarrow (1)$, let $\mathfrak{a}$ be a 2-absorbing ideal of $S$. By assumption, prime ideals of $S$ are comparable. Therefore, by Theorem \ref{p2subsetofideal}, we have $\mathfrak{p}^2 \subseteq \mathfrak{a} \subseteq \mathfrak{p}$, where $\mathfrak{p}$ is a minimal prime over $\mathfrak{a}$. Now by assumption, $\mathfrak{am} = \mathfrak{p}$. So, $\mathfrak{p} \subseteq \mathfrak{a} \cap \mathfrak{m} = \mathfrak{a}$ and the proof is complete. By Lemma \ref{2minsemiring} and this point that the statements (1) and (2) are equivalent, clearly, the statements (1) and (3) are equivalent and this finishes the proof.\end{proof} \end{theorem} \begin{corollary} \label{not-2-absorbing} Let $S$ be a subtractive 2-AB semiring. For each prime ideal $\mathfrak{p}$ of $S$, either $\mathfrak{p} = \mathfrak{p}^2$ or $\mathfrak{p}^2$ is not 2-absorbing. \end{corollary} \begin{question} Is it true that a subtractive semiring $S$ is 2-AB if primes are comparable and if $\mathfrak{pm} = \mathfrak{p}$ for all prime ideals $\mathfrak{p}$? \end{question} \begin{remark} Let $R$ be a commutative ring with a nonzero identity. By Nakayama's lemma in commutative algebra and Proposition \ref{2-absorbingmulti}, it is straightforward to see that if $R$ is a Noetherian 2-AB ring, then $R$ is a field. Now let $S$ be a Noetherian 2-AB semiring. One may ask if $S$ is a semifield. In fact, this is not the case. Consider the semiring $S=\{0,u,1\}$, where $1+u = u+1 = u$, $1+1=1$, and $u+u = u\cdot u =u$ \cite{LaGrassa1995}. The only ideals of $S$ are $(0)$, $S$, and $\{0,u\}$. Note that the ideal $\{0,u\}$ is maximal (and prime). Also, it is clear that $(0)$ is prime. Therefore, $S$ is 2-AB. Clearly, $S$ is Noetherian, but it is not a semifield, because the element $u$ is not a unit. \end{remark} \begin{theorem} \label{2-ABThm3} Let $S$ be a subtractive valuation semiring. Then, $S$ is 2-AB if and only if $\mathfrak{p}^2 = \mathfrak{p}$ for every prime ideal $\mathfrak{p}$ of $S$. \begin{proof} $(\Rightarrow)$: Suppose that $S$ is 2-AB. Since $S$ is a subtractive valuation semiring, $\mathfrak{p}^2$ is a $\mathfrak{p}$-primary ideal of $S$. Therefore, by Theorem \ref{2-absorbing}, $\mathfrak{p}^2$ is 2-absorbing and so by hypothesis, a prime ideal of $S$. This implies that $\mathfrak{p}^2 = \mathfrak{p}$. $(\Leftarrow)$: This is obvious by Theorem \ref{2-absorbing}. \end{proof} \end{theorem} \begin{remark} In some parts of our paper, we suppose a semiring to be a (subtractive) valuation semiring. On the other hand, in some theorems, we suppose a semiring to be divided. Note that any valuation semiring is divided. Now, one may ask if there are some subtractive valuation semirings which are not rings and so we have really generalized some results in commutative algebra. We point out that examples for subtractive valuation semirings include the semiring $\Id(D)$, where $D$ is a Dedekind domain (refer to Theorem 3.8 and Proposition 3.10 in \cite{Nasehpour2018(b)}). \end{remark} \subsection*{Acknowledgments} The second named author is supported by the Department of Engineering Science at the Golpayegan University of Technology and his special thanks go to the Department for providing all necessary facilities available to him for successfully conducting this research. The authors are also grateful to the esteemed referee for her/his through reading and review which improved the paper.

https://arxiv.org/abs/1505.02545

An extension of the LMO functor

Cheptea, Habiro and Massuyeau constructed the LMO functor, which is defined on a certain category of cobordisms between two surfaces with at most one boundary component. In this paper, we extend the LMO functor to the case of any number of boundary components, and our functor reflects relations among the parts corresponding to the genera and boundary components of surfaces. We also discuss a relationship with finite-type invariants and Milnor invariants.

\section{Introduction}\label{sec:Intro} In the early 1990s, Kontsevich \cite{Kon93} defined the Kontsevich invariant (the universal finite-type invariant) of knots by the integral on the configuration space of finite distinct points in $\mathbb{C}$. All rational-valued Vassiliev invariants are recovered from the Kontsevich invariant through weight systems. In the late 1990s, Ohtsuki \cite{Oht96b} showed that one can consider an arithmetic expansion of the quantum $SO(3)$-invariant of rational homology spheres. The result of this expansion is called the perturbative $SO(3)$-invariant. Ohtsuki \cite{Oht96a} also introduced integer-valued finite-type invariants of integral homology spheres. Kricker and Spence \cite{KrSp97} proved that the coefficients of the perturbative $SO(3)$-invariant are of finite-type. On the other hand, the perturbative $SO(3)$-invariant was extended to the perturbative $PG$-invariant for any simply connected simple Lie group $G$, where $PG$ is the quotient Lie group of $G$ by its center. Moreover, using the Kontsevich invariant, Le, Murakami and Ohtsuki \cite{LMO98} introduced the LMO invariant of connected oriented closed 3-manifolds. It is known that the LMO invariant is universal among perturbative invariants of rational homology spheres. Bar-Natan, Garoufalidis, Rozansky and Thurston \cite{BGRT02a,BGRT02b,BGRT04} gave an alternative construction of the LMO invariant of rational homology spheres by introducing the \r{A}rhus integral that is also called the formal Gaussian integral. In these papers, it is suggested that the \r{A}rhus integral can be extended to an invariant of tangles in a rational homology sphere, which is called the Kontsevich-LMO invariant in \cite{HabN00} and \cite{CHM08}. Using the Kontsevich-LMO invariant, Cheptea, Habiro and Massuyeau \cite{CHM08} defined the LMO functor as a functorial extension of the LMO invariant. In fact, the value for a rational homology cube $M$ (in which case the boundary of $M$ is $S^2$) coincides with the LMO invariant of the closed 3-manifold obtained from $S^3\setminus\operatorname{Int}[-1,1]^3$ and $M$ by gluing their boundaries, see \cite[Section~3.5]{CHM08}. One of the advantage of the LMO functor is that we can use its functoriality to calculate its values and to prove its properties. \begin{figure}[h] \centering \begin{minipage}{0.45\columnwidth} \centering \includegraphics[height=6em]{OldCob.pdf} \end{minipage} \begin{minipage}{0.45\columnwidth} \centering \includegraphics[height=6em]{NewCob.pdf} \end{minipage} \caption{A cobordism introduced in \cite{CHM08} (left) and a new one (right) both with parametrizations on their boundaries} \label{fig:OldNewCob} \end{figure}% The LMO functor is defined for a connected, oriented, compact 3-manifold regarded as a certain cobordism between two surfaces. Here, these surfaces are assumed to be with at most one boundary component. The purpose of this paper is to construct an extension of the LMO functor to the case of \emph{any} number of boundary components (compare two figures in Figure~\ref{fig:OldNewCob}). It is expected that this extension enables us to introduce many \emph{categorical} operations on cobordisms, for instance, which corresponds to the pairing or shelling product defined in \cite{ABMP10}. We define the braided non-strict monoidal category $\mathcal{LC}\mathit{ob}_q$ to be Lagrangian $q$-cobordisms extended as above (see Definition~\ref{subsec:qCob} and Remark~\ref{rem:braided}). The main result is the following theorem. \begin{introtheorem}[Theorem~\ref{thm:Zt}] There is a tensor-preserving functor $\widetilde{Z}\colon \mathcal{LC}\mathit{ob}_q \to {}^{\mathit{ts}}\!\mathcal{A}$ between two monoidal categories, which is an extension of the LMO functor. \end{introtheorem} A generating set of $\mathcal{LC}\mathit{ob}_q$ as a monoidal category is determined in Proposition~\ref{prop:Gen} and the values on them are listed in Table~\ref{tab:Value}. Therefore, the functoriality and tensor-preservingness of $\widetilde{Z}$ enable us to compute the value on a Lagrangian $q$-cobordism by decomposing it into the generators. It should be emphasized that there are diagrams colored by both $+$ (or $-$) and $0$ in Table~\ref{tab:Value}, which implies that our extension is non-trivial. Habiro \cite{HabK00} and Goussarov \cite{GGP01} introduced claspers and clovers respectively that play a crucial role in a theory of finite-type invariants of arbitrary 3-manifolds. In \cite{CHM08}, using claspers, it was shown that the LMO functor is universal among rational-valued finite-type invariants. We prove that our LMO functor $\widetilde{Z}$ has the same property. \begin{introtheorem}[Theorem~\ref{thm:universal}] $\widetilde{Z}$ is universal among rational-valued finite-type invariants. \end{introtheorem} In \cite{Mil54}, Milnor defined an invariant of links, which is called the Milnor $\overline{\mu}$-invariant. Habegger and Lin \cite{HaLi90} showed that this invariant is essentially an invariant of string links. In \cite{CHM08}, it was proven that the tree reduction of the LMO functor is related to Milnor invariants of string links. By extending the Milnor-Johnson correspondence (see, for example, \cite{HabN00}) suitably, we show that the same is true for $\widetilde{Z}$. \begin{introtheorem}[Theorem~\ref{thm:MilnorInv}] Let $(B,\sigma)$ be a string link in an integral homology sphere $B$. Then the first non-trivial term of $\widetilde{Z}^{Y,t}(\operatorname{MJ}_w^{-1}(B,\sigma))$ is determined by the first non-vanishing Milnor invariant of $(B,\sigma)$, and vice versa. \end{introtheorem} Finally, in \cite{ChLe06}, \cite{ABMP10} and \cite{Kat14}, one can find related researches from different points of view from this paper. \subsection*{Organization of this paper} In Section~\ref{sec:CobBTtangle}, we define cobordisms and bottom-top tangles that are main object in this paper. Section~\ref{sec:JacobiDiagOp} is devoted to reviewing Jacobi diagrams and the formal Gaussian integral. In Section~\ref{sec:KontsevichLMOinv}, the Kontsevich-LMO invariant of tangles in a homology cube is explained, which plays a key role in the subsequent sections. The main part of this paper is Section~\ref{sec:ExtLMOfunc}, where we construct an extension of the LMO functor. In Section~\ref{sec:GenValue}, we shall give a generating set of the category $\mathcal{LC}\mathit{ob}$ and calculate the values on them. These values will be used later. Section~\ref{sec:Universal} is devoted to reviewing clasper calculus and proving the universality among finite-type invariants. Finally, in Section~\ref{sec:Knot}, we apply our LMO functor $\widetilde{Z}$ to some cobordisms arising from knots or string links. In particular, the relationship between $\widetilde{Z}$ and Milnor invariants of string links will be discussed. \subsection*{Notation} Throughout this paper, we denote by $H_\ast(X)$ the homology groups of a topological space $X$ with coefficients in $\mathbb{Z}$. (However, arguments in Sections~\ref{sec:CobBTtangle}--\ref{sec:ExtLMOfunc} are valid for the case of coefficients in $\mathbb{Q}$, see Remark~\ref{rem:Rational}.) Let $\fc{n}^\ast$ denote the set $\{1^\ast,\dots,n^\ast\}$, where $\ast$ is $+,\ -,\ 0$ or we omit $\ast$. Let $X$ and $Y$ be geometric objects equipped with the ``top'' and ``bottom'', for example, cobordisms or tangles. Then, the composition of $X$ and $Y$ is always defined by stacking the bottom of $Y$ on the top of $X$. We use almost the same notation and terminology as in \cite{CHM08}. However, their definitions are suitably extended. \subsection*{Acknowledgements} The author would like to thank Takuya Sakasai, who provided him with helpful comments and suggestions. He also feels thankful to Kazuo Habiro, who helped him to prove Proposition~\ref{prop:Gen}. He would also like to thank Gw\'ena\"el Massuyeau for pointing out many crucial mistakes in his first draft and recommending him to introduce the functor $\mathrm{K}$ in Remark~\ref{rem:Kill}. Finally, this work was supported by the Program for Leading Graduate Schools, MEXT, Japan. \section{Cobordisms and bottom-top tangles}\label{sec:CobBTtangle} In this section, we give the definition of cobordisms and a way to express them as certain tangles. \subsection{Cobordisms}\label{subsec:Cob} We first introduce some notation. Let $\mathrm{Mon}(x_1,\dots,x_k)$ denote the free monoid generated by letters $x_1,\dots,x_k$. Similarly, we denote by $\mathrm{Mag}(x_1,\dots,x_k)$ the free magma generated by $x_1,\dots,x_k$. We call an element of $\mathrm{Mon}(x_1,\dots,x_k)$ or $\mathrm{Mag}(x_1,\dots,x_k)$ a \emph{word}. Let $w$ be a word. We denote by $w^{x_i}$ the element obtained from $w$ by replacing all letters except for $x_i$ with the empty word $e$. Let $|w|$ denote the word length of $w$. For example, if $w=\bullet((\circ\bullet)\bullet) \in \mathrm{Mag}(\bullet,\circ)$, then $|w^\bullet|=|\bullet(\bullet\bullet)|=3$. Next, we prepare two kinds of surfaces. Let $w \in \mathrm{Mon}(\bullet,\circ)$ and let $g=|w^\bullet|$, $n=|w^\circ|$. We define the oriented compact surface $F_w$ in $\mathbb{R}^3$ as illustrated in Figure~\ref{fig:F}, which has handles and inner boundary components corresponding to letters $\bullet$ and $\circ$ in $w$, respectively. Moreover, let $\alpha_1,\dots,\alpha_g$, $\beta_1,\dots,\beta_g$ and $\delta_1,\dots,\delta_n$ denote the oriented simple closed curves based at $\ast$ as drawn in Figure~\ref{fig:F}. We often regard the above closed curves as free loops. \begin{figure}[h] \centering \includegraphics[width=0.6\columnwidth]{F.pdf} \caption{An example of $F_w$ with $w=\bullet\circ\bullet\:\circ$} \label{fig:F} \end{figure} Let $w_{+},w_{-} \in \mathrm{Mag}(\bullet,\circ)$ such that $|w_{+}^\circ|=|w_{-}^\circ|=:n$, and let $\sigma$ be an element of the $n$th symmetric group $\SS_n$. We define the \emph{reference surface} $R^{w_+}_{w_-,\sigma}$ as illustrated in Figure~\ref{fig:R}, which is an oriented closed surface consisting of four kinds of surfaces: the \emph{top surface} $F_{w_+}$, the \emph{bottom surface} $-F_{w_-}$, $n$ tubes and four sides, where $-F_{w_-}$ is the surface $F_{w_-}$ with the opposite orientation. The boundary of the tubes are attached to the inner boundary of the top surface and bottom surface according to the permutation $\sigma$. \begin{figure}[h] \centering \includegraphics[width=0.6\columnwidth]{R-crop.pdf} \caption{An example of $R^{w_+}_{w_-,\sigma}$ with $w_{+}=\bullet\circ\circ\:\bullet$, $w_{-}=\circ\bullet\circ$} \label{fig:R} \end{figure} \begin{definition} Let $w_+, w_- \in \mathrm{Mon}(\bullet,\circ)$ such that $|w_+^\circ|=|w_-^\circ|=:n$. A \emph{cobordism} from $F_{w_+}$ to $F_{w_-}$ is an equivalence class of triples $(M,\sigma,m)$, where \begin{itemize} \item $M$ is a connected, oriented, compact 3-manifold, \item $\sigma$ is an element of the $n$th symmetric group $\SS_{n}$, \item $m\colon R^{w_+}_{w_-,\sigma} \to \partial M$ is an orientation-preserving homeomorphism. \end{itemize} Here, $(M,\sigma,m)$ is equivalent to $(N,\tau,n)$ if $\sigma=\tau$ and there is an orientation-preserving homeomorphism $f\colon M \to N$ satisfying $f|_{\partial M}\circ m=n$. Let $m_\pm$ denote the restriction of $m$ to $\pm F_{w_\pm}$. \end{definition} We define the category $\mathcal{C}\mathit{ob}$ of cobordisms as follows: The set of objects is the monoid $\mathrm{Mon}(\bullet,\circ)$. A morphism from $w_+$ to $w_-$ is a cobordism from $F_{w_+}$ to $F_{w_-}$. The composition of $(M,\sigma,m) \in\mathcal{C}\mathit{ob}(v,u)$ and $(N,\tau,n) \in \mathcal{C}\mathit{ob}(w,v)$ is defined by \[(M,\sigma,m)\circ(N,\tau,n) := (M \cup_{m_+\circ n_-^{-1}} N, \sigma\tau, m_{-} \cup n_{+}) \in \mathcal{C}\mathit{ob}(w,u),\] namely, the composition is obtained by stacking the bottom of $N$ on the top of $M$. The identity of $w$ is the cobordism $(F_w\times[-1,1],\mathrm{Id}_{\SS_{|w^\circ|}},\mathrm{Id})$, where the last $\mathrm{Id}$ means the obvious parametrization. Moreover, $\mathcal{C}\mathit{ob}$ is a strict monoidal category, where the tensor product of two cobordisms is their horizontal juxtaposition and the unit is the empty word $e$. \begin{remark}\label{rem:Kill} We denote by $\mathcal{C}\mathit{ob}^\bullet$ the full subcategory whose objects belong to the monoid $\mathrm{Mon}(\bullet)$, namely, $\mathcal{C}\mathit{ob}^\bullet$ is the category $\mathcal{C}\mathit{ob}$ defined in \cite{CHM08}. The full functor $\mathrm{K}\colon \mathcal{C}\mathit{ob} \to \mathcal{C}\mathit{ob}^\bullet$ is defined to be killing the extra boundary, that is, for $M=(M,\sigma,m)\in \mathcal{C}\mathit{ob}(w_+,w_-)$, $\mathrm{K}(M)$ is obtained by attaching $|w_{+}^\circ|$ tubes to $M$. \end{remark} \begin{definition} Let $w \in \mathrm{Mon}(\bullet,\circ)$ such that $|w^\circ|=n$, and let $\ast_1,\dots,\ast_n$ be points on each inner boundary component. We recall that $\ast$ is the point on the outer boundary as drawn in Figure~\ref{fig:F}. The \emph{mapping class group} of $F_w$ is defined by \begin{align*} \mathcal{M}(F_w) &:= \left\{h \in \operatorname{Homeo}_{+}(F_w) \mathrel{}\middle|\mathrel{} \parbox{13em}{$h(\ast)=\ast, \\ h(\{\ast_1,\dots,\ast_b\})=\{\ast_1,\dots,\ast_b\}$} \right\} \bigg/ \text{isotopy}, \end{align*} where an isotopy fixes the set $\{\ast,\ast_1,\dots,\ast_n\}$ pointwise. For any $h \in \mathcal{M}(F_w)$, we denote by $\mathrm{C}(h)$ the \emph{mapping cylinder} of $h$, that is, the cobordism \[(F_w\times[-1,1], \pi(h), \mathrm{Id}\times(-1)\cup h\times 1),\] where $\pi(h) \in \SS_n$ is defined by $h(\ast_i)=\ast_{\pi(h)(i)}$. Note that an isotopy does not necessarily fix $\partial F_w$, however, by definition, $\mathrm{C}(h)$ is well-defined. One can show that the map \[\mathrm{C}\colon \mathcal{M}(F_w) \to \mathcal{C}\mathit{ob}(w,w)\] is an injective monoid homomorphism. \end{definition} It is obvious that the image of $\mathrm{C}$ is contained in the group consisting of invertible elements of $\mathcal{C}\mathit{ob}(w,w)$. In fact, the following proposition holds. \begin{proposition}\label{prop:InvertElem} Let $w \in \mathrm{Mon}(\bullet,\circ)$ and $M \in \mathcal{C}\mathit{ob}(w,w)$. Then $M$ is left-invertible if and only if it is right-invertible. Moreover, the group $\mathcal{C}\mathit{ob}(w,w)^\times$ consisting of invertible elements of $\mathcal{C}\mathit{ob}(w,w)$ is equal to $\Im \mathrm{C}$. \end{proposition} The definition of the mapping class group is different from \cite{HaMa12}, although the same proof as \cite[Proposition~2.4]{HaMa12} is valid for our monoid homomorphism $\mathrm{C}$. \begin{remark} The mapping class group of $F_w$ is usually defined by \[\mathcal{M}_0(F_w) := \{f\in\operatorname{Homeo}(F_w) \mid f|_{\partial F_w}=\mathrm{Id}_{\partial F_w}\} \mathrel{/} \text{isotopy rel $\partial F_w$},\] which is isomorphic to the kernel of the homomorphism $\pi\colon \mathcal{M}(F_w) \twoheadrightarrow \SS_n$. If $w \in \mathrm{Mon}(\circ)$, then $\mathcal{M}(F_w)$ and $\mathcal{M}_0(F_w)$ are called the \emph{framed braid group} and the \emph{framed pure braid group} on $n$ strands, respectively. \end{remark} We should restrict cobordisms to the good ones in order to define the Kontsevich-LMO invariant (see Section~\ref{subsec:DefKontsevichLMOinv}) and to let it take value in the space of top-substantial Jacobi diagrams (see Section~\ref{subsec:TopSubstantial}). Let $w \in \mathrm{Mon}(\bullet,\circ)$. We define the subgroups of $H_1(F_w)$ by \[A_w:=\ang{\alpha_1,\dots,\alpha_{|w^\bullet|}},\quad B_w:=\ang{\beta_1,\dots,\beta_{|w^\bullet|}},\quad D_w:=\ang{\delta_1,\dots,\delta_{|w^\circ|}}.\] \begin{definition} A cobordism $(M,\sigma,m) \in \mathcal{C}\mathit{ob}(w_+,w_-)$ is \emph{Lagrangian} if the following two conditions are satisfied: \begin{enumerate} \item $H_1(M) = m_{-,\ast}(A_{w_-}) + m_{+,\ast}(H_1(F_{w_+}))$, \item $m_{+,\ast}(A_{w_+}) \subset m_{-,\ast}(A_{w_-}) + m_{+,\ast}(D_{w_+})$. \end{enumerate} \end{definition} It follows from a Mayer-Vietoris argument that the composition of two Lagrangian cobordisms is again Lagrangian. (One can get an alternative proof by Lemma \ref{lem:Lagrangian} and \cite[Claim 3.18]{CHM08} or the fact that $M$ is Lagrangian if and only if $\mathrm{K}(M)$ is Lagrangian in the sense of \cite{CHM08}.) Thus, we obtain the wide subcategory $\mathcal{LC}\mathit{ob}$ whose morphisms are Lagrangian cobordisms. Actually, we need a non-strict monoidal category $\mathcal{LC}\mathit{ob}_q$ defined in Section~\ref{subsec:qCob}. \begin{lemma}\label{1'} Under the condition \textup{(2)}, \textup{(1)} is equivalent to the following condition: \textup{(1')} the map \[m_{-,\ast}\oplus m_{+,\ast}\colon A_{w_-}\oplus(B_{w_+}\oplus D_{w_+}) \to H_1(M)\] is an isomorphism. \end{lemma} \begin{proof} It is easy to see that (1') implies (1). Let us prove the converse. It follows from (1) and (2) that $b_1(M) \le g_{-}+g_{+}+n$, where $b_1(M)$ is the first Betti number of $M$. Since $M$ is a compact odd-dimensional manifold, we have $\chi(\partial M)=2\chi(M)$, and thus $b_1(M)=g_{-}+g_{+}+n+b_2(M)$. Hence, the above inequality is an equality, that is, $m_{-,\ast}\oplus m_{+,\ast}$ is an isomorphism. \end{proof} \begin{remark}\label{rem:HomCube} By the proof of the above lemma, we have $b_2(M)=0$. Since $H_2(M)$ is free, we conclude $H_2(M)=0$, namely $M$ is a \emph{homology handlebody}. In particular, a homology handlebody of genus zero is called a \emph{homology cube}. \end{remark} \begin{example} The mapping cylinder of $h \in \mathcal{M}(F_w)$ is Lagrangian if and only if $h_\ast(A_w \oplus D_w) = A_w \oplus D_w$. Indeed, if $\mathrm{C}(h)$ is Lagrangian, then we have $h_\ast(A_w \oplus D_w) \subset A_w \oplus D_w$. Since $h_\ast$ is an isomorphism, so is $h_\ast|_{A_w \oplus D_w}$. The converse is obvious. \end{example} \subsection{Bottom-top tangles}\label{subsec:BTtangle} In this subsection, we translate cobordisms to certain tangles in a homology cube. Let $w \in \mathrm{Mon}(\bullet,\circ)$. We prepare the reference surface $F_e$ with the pairs of points $\{p_1,q_1\},\dots,\{p_{|w^\bullet|},q_{|w^\bullet|}\}$ and the points $r_1\dots,r_{|w^\circ|}$ corresponding to the letters $\bullet$'s and $\circ$'s in $w$, respectively. These points are on the $x$-axis according to $w$ as drawn in Figure~\ref{fig:PonitedReferenceSurface}. \begin{figure}[h] \centering \includegraphics[width=0.5\columnwidth]{PointedReferenceSurface.pdf} \caption{The reference surface with the points corresponding to $w$ such that $|w^\bullet|=g$.} \label{fig:PonitedReferenceSurface} \end{figure} \begin{definition} Let $w_+,w_- \in \mathrm{Mon}(\bullet,\circ)$ such that $|w_+^\circ|=|w_-^\circ|$ and let $\sigma \in \SS_{|w_+^\circ|}$. A \emph{bottom-top tangle} of \emph{type} $(w_+,w_-,\sigma)$ is an equivalence class of pairs $(B,\gamma)$, where \begin{itemize} \item $B=(B,\emptyset,b)$ is a cobordism from $F_e$ to $F_e$, \item $\gamma$ is a framed oriented tangle in $B$, which consists of the top components $\gamma_1^+,\dots,\gamma_{|w_+^\bullet|}^+$, the bottom components $\gamma_1^-,\dots,\gamma_{|w_-^\bullet|}^-$ and the vertical components $\gamma_1^0,\dots,\gamma_{|w_+^\circ|}^0$, where \begin{itemize} \item each $\gamma^{+}_i$ runs from $p_i\times1$ to $q_i\times1$, \item each $\gamma^{-}_i$ runs from $q_i\times(-1)$ to $p_i\times(-1)$, \item each $\gamma^0_i$ runs from $r_i\times1$ to $r_{\sigma(i)}\times(-1)$. \end{itemize} \end{itemize} Here, $(B,\gamma)$ and $(B',\gamma')$ are equivalent if they are of the same type and $B$ is equivalent to $B'$ as cobordisms by a homeomorphism that sends $\gamma$ to $\gamma'$ and respects their framings. \end{definition} We define the category ${}_{\mathit{b}}^{\mathit{t}}\mathcal{T}$ as follows: An object of ${}_{\mathit{b}}^{\mathit{t}}\mathcal{T}$ is an element of $\mathrm{Mon}(\bullet,\circ)$. A morphism from $w_+$ to $w_-$ is a bottom-top tangle of type $(w_+,w_-,\sigma)$ for some $\sigma \in \SS_{|w_+^\circ|}$. The composition of $(B,\gamma)\in{}_{\mathit{b}}^{\mathit{t}}\mathcal{T}(v,u)$ and $(C,\omega)\in{}_{\mathit{b}}^{\mathit{t}}\mathcal{T}(w,v)$ is the bottom-top tangle obtained by inserting $([-1,1]^3, T_v)$ between them and performing the surgery along $(2|v^\bullet|)$-components link $\gamma^{+}\cup T_{w^\bullet}\cup\omega^{-}$ as illustrated in Figure~\ref{fig:TvIdw}. The identity of $w$ is the bottom-top tangle as depicted in Figure~\ref{fig:TvIdw}. Indeed, one can check that \[\mathrm{Id}_{w_-}\circ(B,\gamma) = (B,\gamma) = (B,\gamma)\circ\mathrm{Id}_{w_+},\] using the Kirby move II and the useful fact that if a disk $D$ in a 3-manifold $M$ intersects a framed knot $K$ once, then the result of surgery along $K$ and $\partial D$ with 0-framing is again $M$ (see Figure~\ref{fig:composition}). \begin{figure}[h] \centering \includegraphics[width=0.8\columnwidth]{TvIdw.pdf} \caption{The bottom-top tangle $([-1,1]^3, T_v)$ and the identity $\mathrm{Id}_w$} \label{fig:TvIdw} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.8\columnwidth]{composition-crop.pdf} \caption{An example of the equality $\mathrm{Id}_{\bullet\circ}\circ(B,\gamma)=(B,\gamma)$} \label{fig:composition} \end{figure} Moreover, ${}_{\mathit{b}}^{\mathit{t}}\mathcal{T}$ is a strict monoidal category, where the tensor product of two bottom-top tangles is their horizontal juxtaposition and the unit is the empty word $e$. The next lemma is used in the following theorem and so on. \begin{lemma}[see {\cite[Definition 3.15]{CHM08}}] Let $(B,\gamma)$ be a bottom-top tangle. Then there exist a framed link $L$ and a tangle $\gamma'$ in $[-1,1]^3$ such that $(B,\gamma)=([-1,1]^3_L,\gamma')$, where $[-1,1]^3_L$ is the \textup{3}-manifold obtained by surgery along $L$. The pair $(L,\gamma')$ is called a \emph{surgery presentation} of $(B,\gamma)$. \end{lemma} \begin{theorem}[{\cite[Theorem 2.10]{CHM08}}] The operation as drawn in Figure~\textup{\ref{fig:dig}} gives an isomorphism $\mathrm{D}\colon {}_{\mathit{b}}^{\mathit{t}}\mathcal{T} \to \mathcal{C}\mathit{ob}$ between two strict monoidal categories. \end{theorem} \begin{figure}[h] \centering \includegraphics[width=0.9\columnwidth]{dig-crop.pdf} \caption{Correspondence between a bottom-top tangle and a cobordism} \label{fig:dig} \end{figure} \begin{proof} Precisely, $\mathrm{D}$ is defined by digging a bottom-top tangle $((B,b),\gamma)$ and parametrizing the boundary of the resulting manifold in accordance with the homeomorphism $b\colon R^e_{e,\mathrm{Id}} \to \partial B$ and the framings of $\gamma$. It is easy to see that $\mathrm{D}$ satisfies the definition of a monoidal functor except \[\mathrm{D}((B,\gamma)\circ(C,\omega)) = \mathrm{D}(B,\gamma)\circ\mathrm{D}(C,\omega).\] To prove this equality, we take a surgery presentation $(L,\gamma')$ of $(B,\gamma)$. Moreover, using an additional surgery link $L'$, $(B,\gamma)$ is expressed as the bottom-top tangle obtained from a bottom-top tangle $([-1,1]^3,\widetilde{\gamma})$ such that $\widetilde{\gamma}^+$ is trivial by surgery along $L \cup L'$. Each $\widetilde{\gamma}^+_i$ bounds a half-disk $D_i$ that could intersect $\widetilde{\gamma}^- \cup \widetilde{\gamma}^0$ or $L \cup L'$. We now gather the intersections on $D_i$, and we regard the components intersecting $D_i$ as a bundle in a neighborhood of $D_i$. Since $D_i$ intersects the bundle once, we can apply the argument in the proof of $\mathrm{Id}_{w_-} \circ (B,\gamma) = (B,\gamma)$, and thus the desired equality is obtained. Next, let $(M,\sigma,m)$ be a cobordism. The inverse of $\mathrm{D}$ can be constructed by attaching 2-handles $[0,1]^2\times[0,1]$ to the boundary of $M$. (We recall that the way of attaching a 3-dimensional 2-handle is uniquely determined by the way one glues the attaching sphere of the 2-handle with $\partial M$.) We obtain the 3-manifold $B$ by attaching 2-handles along $m_{+}(\beta_i)$'s, $m_{-}(\alpha_i)$'s and $m_{+}(\delta_i)$'s and define $\gamma'$ by the co-cores of the 2-handles. By construction, this correspondence gives the inverse of $\mathrm{D}$. \end{proof} \begin{example} Let $c$ be a simple closed curve on the surface $\operatorname{Int} F_w$ that is the interior of $F_w$. We denote by $\tau_c$ the Dehn twist along $c$. By the definition of a Dehn twist and surgery, the cobordisms $\mathrm{C}(\tau_{\alpha_i})$, $\mathrm{C}(\tau_{\beta_i})$ and $\mathrm{C}(\tau_{\delta_i})$ are sent by $\mathrm{D}^{-1}$ to \begin{center} \raisebox{-2.2em}{ \begin{overpic}[height=7em]{DehnTwist.pdf} \put (35,12){,} \put (75,12){and} \end{overpic}} \end{center} respectively. $\mathrm{C}(\tau_{\alpha_i})$, $\mathrm{C}(\tau_{\delta_i})$ are Lagrangian while $\mathrm{C}(\tau_{\beta_i})$ is not. \end{example} There is a natural question when a bottom-top tangle is sent to a Lagrangian cobordism by $\mathrm{D}$. We introduce an extension of the linking matrix defined in \cite[Definition 2.11]{CHM08} to answer this question. \begin{definition} Let $((B,b),\gamma)$ be a bottom-top tangle of type $(w_+,w_-,\sigma)$ such that $B$ is a homology cube, and let $g_\pm:=|w_\pm^\bullet|$, $n:=|w_+^\circ|$. Prepare the manifold $S^3\setminus\operatorname{Int}[-1,1]^3$ containing four kinds of arcs: $g_+$ copies of $\cap$-like arcs, $g_-$ $\cup$-like arcs, $n$ parallel arcs and a braid $\beta \in B_n$ such that $\pi(\beta)=\sigma^{-1}$, as illustrated in Figure~\ref{fig:Lk}. Then the \emph{linking matrix} of $\gamma$ in $B$ is defined by \[\operatorname{Lk}_B(\gamma) := \operatorname{Lk}_{\widehat{B}}(\widehat{\gamma})-O_{g_{+}+g_{-}}\oplus\sigma^{-1}\cdot\operatorname{Cr}(\beta) \in \tfrac{1}{2}\operatorname{Sym}_{\pi_0 \gamma}(\mathbb{Z}),\] where $\operatorname{Lk}_{\widehat{B}}(\widehat{\gamma})$ is the usual linking matrix of the link \[\widehat{\gamma} := \gamma \cup (\text{the arcs and braid in $S^3\setminus\operatorname{Int}[-1,1]^3$})\] in the homology sphere $\widehat{B} := B \cup_b (S^3 \setminus\operatorname{Int}[-1,1]^3)$. Here, $\operatorname{Cr}(\beta)$ is the matrix whose $(i,j)$-entry is half the number of positive crossings of the $i$th strand and the $j$th strand minus the number of negative ones, and $\sigma^{-1}\cdot\operatorname{Cr}(\beta)$ is the matrix obtained from $\operatorname{Cr}(\beta)$ by permuting its rows and columns by $\sigma^{-1}$. \end{definition} \begin{figure}[h] \centering \includegraphics[width=0.8\columnwidth]{Lk.pdf} \caption{The union of $(B,\gamma)$ and additions, that is, $(\widehat{B},\widehat{\gamma})$} \label{fig:Lk} \end{figure} Moreover, we set $\operatorname{Lk}(M):=\operatorname{Lk}_B(\gamma)$ for a cobordism $(M,\sigma,m)=\mathrm{D}(B,\gamma)$, where $B$ is a homology cube. \begin{remark} The above map $\operatorname{Cr}$ represents a non-trivial class of the first cohomology group $H^1(B_n^\mathrm{op};\tfrac{1}{2}\operatorname{Sym}_n(\mathbb{Z}))$, where $B_n^\mathrm{op}$ is the group obtained from $B_n$ by reversing the order of multiplication. Indeed, $B_n^\mathrm{op}$ acts on $\tfrac{1}{2}\operatorname{Sym}_n(\mathbb{Z})$, via the anti-homomorphism $\pi\colon B_n^\mathrm{op} \to \SS_n$, such that $\beta \cdot X:=\pi(\beta)^{-1} \cdot X$. Then $\operatorname{Cr} \colon B_n^\mathrm{op} \to \tfrac{1}{2}\operatorname{Sym}_n(\mathbb{Z})$ is a crossed homomorphism, that is, it satisfies \[\operatorname{Cr}(\beta\beta') = \operatorname{Cr}(\beta)+\beta\cdot\operatorname{Cr}(\beta').\] Furthermore, $\left|\Im\left(d^0(X)\colon B_n^\mathrm{op}\to\tfrac{1}{2}\operatorname{Sym}_n(\mathbb{Z})\right)\right| \le n!$ for any $X \in \tfrac{1}{2}\operatorname{Sym}_n(\mathbb{Z})$. While, $|\Im(\operatorname{Cr})| = \infty$. \end{remark} The next lemma is proved in much the same way as \cite[Lemma 2.12]{CHM08}. \begin{lemma}\label{lem:Lagrangian} Let $(B,\gamma)$ be a bottom-top tangle. Then $\mathrm{D}(B,\gamma)$ is Lagrangian if and only if $B$ is a homology cube and $\operatorname{Lk}_B(\gamma^+)=O$ holds. \end{lemma} \section{Jacobi diagrams and some operations}\label{sec:JacobiDiagOp} \subsection{Jacobi diagrams}\label{subsec:JacobiDiag} Let $D$ be a uni-trivalent graph, which is depicted by dashed lines. A univalent (resp.\ trivalent) vertex of $D$ is called an \emph{internal} (resp.\ \emph{external}) \emph{vertex}. We denote by $\operatorname{i-deg} D$ (resp.\ $\operatorname{e-deg} D$) the number of internal (resp.\ external) vertices of $D$ and we set $\deg D := (\operatorname{i-deg} D +\operatorname{e-deg} D)/2$. Let $X$ be a compact oriented 1-manifold and let $C$ be a finite set. A \emph{Jacobi diagram} based on $(X,C)$ is a vertex-oriented uni-trivalent graph whose vertices are either embedded into $X$ or are colored by elements of $C$. When identifications of $\pi_0 X$ with $\pi_0 X'$ and $C$ with $C'$ are given, two Jacobi diagrams $D$ based on $(X,C)$ and $D'$ based on $(X',C')$ are identified if there exists a homeomorphism $f\colon X\cup D \to X'\cup D'$ respecting the identification, orientations of $X$ and $X'$ and the vertex-orientations of $D$ and $D'$. A Jacobi diagram is drawn in the plane whose vertex-orientation agrees with the counter-clockwise order. We denote by $\mathcal{A}(X,C)$ the quotient $\mathbb{Q}$-vector space spanned by Jacobi diagrams based on $(X,C)$ subject to the AS, IHX and STU relations depicted in Figure~\ref{fig:relations}. \begin{figure}[h] \centering \begin{minipage}{0.36\columnwidth} \centering \includegraphics[width=0.8\columnwidth]{AS-crop.pdf} \end{minipage} \hspace{12pt} \begin{minipage}{0.52\columnwidth} \centering \includegraphics[width=0.8\columnwidth]{IHX-crop.pdf} \end{minipage} \begin{minipage}{0.52\columnwidth} \vspace{0.5em}% \centering \includegraphics[width=0.8\columnwidth]{STU-crop.pdf} \end{minipage} \caption{The AS, IHX and STU relations. These figures represent local relations, namely, the lest of the diagrams are identical.} \label{fig:relations} \end{figure} Since these relations preserve the degree of a Jacobi diagram, $\mathcal{A}(X,C)$ is a graded $\mathbb{Q}$-vector space. Thus, we can take the degree completion of $\mathcal{A}(X,C)$ and denote it by $\mathcal{A}(X,C)$ again. Note that the internal degree is not well-defined for Jacobi diagrams of $\mathcal{A}(X,C)$. A Jacobi diagram $D \in \mathcal{A}(X,C)$ is said to have \emph{\textrm{i}-filter at least} $n$ if $D$ is written as a series of Jacobi diagrams whose internal degrees are at least $n$, and then we use the notation $\operatorname{i-filter} D \ge n$. If $X=\emptyset$ (resp.\ $C=\emptyset$), then we simply write $\mathcal{A}(C)$ (resp.\ $\mathcal{A}(X)$) for $\mathcal{A}(X,C)$ when no confusion can arise. It is well known that $\mathcal{A}(C)$ is naturally equipped with a connected, commutative, cocommutative, graded Hopf algebra structure. There are two important subsets of $\mathcal{A}(C)$. We first set \[\mathcal{G}(\mathcal{A}(C)) := \{x\in\mathcal{A}(C) \mid \Delta(x)=x\otimes x,\ \varepsilon(x)=\varnothing\},\] where $\varnothing$ denotes the empty diagram. $\mathcal{G}(\mathcal{A}(C))$ is a group whose multiplication is disjoint union $\sqcup$, and its element is said to be \emph{group-like}. Let $\mathcal{A}^c(C)$ denote the subspace of $\mathcal{A}(C)$ spanned by non-empty connected Jacobi diagrams, which is identified with the quotient by the subspace generated by $\varnothing$ and disconnected Jacobi diagrams. The image of $x \in \mathcal{A}(C)$ by the quotient map is denoted by $x^c$. Moreover, it is well known that the map $\exp_\sqcup\colon \mathcal{A}^c(C) \to \mathcal{G}(\mathcal{A}(C))$ defined by $\exp_\sqcup(x) := \sum_{n\ge0}x^{\sqcup n}/n!$ is a group isomorphism. We further use two subalgebras of $\mathcal{A}(C)$ which play an important role in this paper. The subalgebra generated by Jacobi diagrams only with struts (resp.\ without struts) is denoted by $\mathcal{A}^s(C)$ (resp.\ $\mathcal{A}^Y(C)$) that is identified with the quotient by the ideal generated by diagrams with $\operatorname{i-deg}>0$ (resp.\ the ideal generated by struts). The quotient maps are called the \emph{$s$-reduction} (resp.\ \emph{$Y$-reduction}), the images of $x\in\mathcal{A}(C)$ is denoted by $x^s$ (resp.\ $x^Y$). \begin{lemma}[{\cite[Lemma~3.5]{CHM08}}] \label{lem:group-like} The group $\mathcal{G}(\mathcal{A}(C))$ coincides with the set \[\{x\in\mathcal{A}(C) \mid \text{$x^s, x^Y \in \mathcal{G}(\mathcal{A}(C))$ and $x=x^s \sqcup x^Y$}\}.\] \end{lemma} \begin{remark} The subspace $\mathcal{A}^Y(C)$ can also be regarded as the result of the internal degree completion, since $(\deg D)/2 \le \operatorname{i-deg} D \le 2\deg D$ holds for all $D \in \mathcal{A}^Y(C)$. We denote by $\mathcal{A}_i(C)$ the subspace of all homogeneous elements of internal degree $i$. \end{remark} Next, we review an analogue of the Poincar\'e-Birkhoff-Witt isomorphism. Let $\downarrow^S$ (resp.\ $\circlearrowleft^S$) denote $|S|$ intervals (resp.\ circles) indexed by elements of a finite set $S$. \begin{definition} The graded linear map $\chi_S \colon \mathcal{A}(X, C\cup S) \to \mathcal{A}(X\downarrow^S, C)$ is defined for a Jacobi diagram $D \in \mathcal{A}(X, C\cup S)$ to be the average of all possible ways of attaching the $s$-colored vertices in $D$ to the $s$-indexed interval, for all $s \in S$, as depicted in Figure~\ref{fig:PBW}. \end{definition} \begin{figure}[h] \centering \includegraphics[width=0.6\columnwidth]{PBW-crop.pdf} \caption{An example of $\chi_S$ with $X=\ \circlearrowleft$, $S=\{1,2,3\}$ and $c \in C$} \label{fig:PBW} \end{figure} It is well known that $\chi_S$ is a graded coalgebra isomorphism. Similarly, an isomorphism $\chi_S \colon \mathcal{A}(X, C\cup S)/{\sim} \to \mathcal{A}(X\!\circlearrowleft^S, C)$ is defined, where $\sim$ denotes the $S$-link relation defined in \cite[Section~5.2]{BGRT02b} (see also \cite[Remark~3.7]{CHM08}). We need an extension of $\chi_S$ to construct the composition law of the category ${}^{\mathit{ts}}\!\mathcal{A}$ defined later. \begin{definition} Let $S'$ be a copy of the above set $S$. The graded linear map \[\chi_{S,S'} \colon \mathcal{A}(X,C\cup S\cup S')\to\mathcal{A}(X\downarrow^S,C)\] is defined for a Jacobi diagram $D \in \mathcal{A}(X, C\cup S\cup S')$ to be stacking of the $s'$-indexed interval on the $s$-indexed interval in $\chi_S \circ \chi_{S'}(D)$, for all $s \in S$, as illustrated in Figure~\ref{fig:PBW2}. \end{definition} \begin{figure}[h] \centering \includegraphics[width=0.8\columnwidth]{PBW2-crop.pdf} \caption{An example of $\chi_{S,S'}$ with $X=C=\emptyset$, $S=\{1,2\}$ and $S'=\{1',2'\}$} \label{fig:PBW2} \end{figure} Similar ideas appear, for example, in \cite[Claim~5.6]{CHM08}, \cite[Proposition~5.4]{BGRT02b} and \cite[(3.3)]{HaMa09}. Before finishing this subsection, we recall the following lemma that is used several times. \begin{lemma}[{\cite[Lemma 8.19]{CHM08}}]\label{lem:filter} Let $D$ be a Jacobi diagram based on $(X\!\downarrow^S,C)$. Then we have \[\chi_S^{-1}(D) = \widetilde{D}+(\operatorname{i-filter}>\operatorname{i-deg} D),\] where $\widetilde{D}$ is obtained from $D$ by deleting the $S$-indexed intervals. \end{lemma} \subsection{The formal Gaussian integral}\label{subsec:GaussianInt} We first review $S$-substantial Jacobi diagrams according to \cite{CHM08}. \begin{definition} A Jacobi diagram $D \in \mathcal{A}(X,C\cup S)$ is \emph{$S$-substantial} if $D$ contains no strut both of whose vertices are colored by the elements of $S$, and an element of $\mathcal{A}(X,C\cup S)$ which can be written as a series of such diagrams is also called $S$-substantial. Let $D$ and $E$ be Jacobi diagrams based on $(X,C\cup S)$ such that at least one of them is $S$-substantial. If $D$ and $E$ have the same number of $s$-colored vertices, for all $s \in S$, then we define $\ang{D,E}_S \in \mathcal{A}(X,C)$ by \[\ang{D,E}_S:= \left(\parbox{0.6\textwidth}{sum of all ways of connecting the $s$-colored vertices in $D$ with those of $E$, for all $s \in S$}\right).\] Otherwise, we set $\ang{D,E}_S :=0$. If $D$ and $E$ have no $S$-colored vertex, we naturally interpret $\ang{D,E}_S$ as their disjoint union. \end{definition} From now on, we identify a symmetric matrix $(a_{ij})_{i,j\in S}$ with a linear combination of struts $\sum\limits_{i,j\in S} a_{ij} \strutgraph{j}{i}$, and let $[x]:=\exp_\sqcup(x)$ for any $x \in \mathcal{A}(C)$. \begin{definition}\label{def:GaussianInt} An element $G \in \mathcal{A}(X,C\cup S)$ is \emph{Gaussian} in the \emph{variable $S$} if there are a symmetric matrix $L \in \operatorname{Sym}_S(\mathbb{Q})$ and an $S$-substantial element $P \in \mathcal{A}(X,C\cup S)$ such that $G = [L/2]\sqcup P$. Here, if $\det L\ne 0$, then $G$ is said to be \emph{non-degenerate}, and we set \[\int_S G := \ang{[-L^{-1}/2],P}_S\ \in \mathcal{A}(X,C)\] that is called the \emph{formal Gaussian integral} of $G$ along $S$. \end{definition} \begin{remark} If $L$ and $P$ exist, they are unique. $L$ is called the \emph{covariance matrix} of $G$ in \cite{BGRT02a}. \end{remark} \section{The Kontsevich-LMO invariant}\label{sec:KontsevichLMOinv} In this section, we review the domains and codomains of the Kontsevich invariant and Kontsevich-LMO invariant in accordance with \cite{CHM08}. \subsection{Domain of the Kontsevich-LMO invariant}\label{subsec:Domain} In order to define the Kontsevich invariant and the Kontsevich-LMO invariant as functors, we use the categories $\mathcal{T}_q$, $\mathcal{T}_q\mathcal{C}\mathit{ub}$ and $\mathcal{A}$. We define the non-strict monoidal category $\mathcal{T}_q\mathcal{C}\mathit{ub}$ as follows: The set of objects is the magma $\mathrm{Mag}(+,-)$. A morphism from $w_+$ to $w_-$ is a $q$-tangle in a homology cube of type $(w_+,w_-)$, that is, a properly embedded compact 1-manifold whose boundary corresponds to letters in $w_\pm$. The composition of $(B,\gamma)\in\mathcal{T}_q\mathcal{C}\mathit{ub}(v,u)$ and $(C,\omega)\in\mathcal{T}_q\mathcal{C}\mathit{ub}(w,v)$ is stacking $C$ on the top of $B$. The identity of $w$ is the $q$-tangle $([-1,1]^3,\downarrow^w)$, where $\downarrow^w$ is the union of $\downarrow$'s and $\uparrow$'s and each $\downarrow$ (resp.\ $\uparrow$) corresponds to a letter $+$ (resp.\ $-$) in $w$. The tensor product of $q$-tangles is their juxtaposition. The unit is the empty word $e$. Moreover, we denote by $\mathcal{T}_q$ the wide subcategory whose morphisms are $q$-tangles in $[-1,1]^3$. \subsection{Codomain of the Kontsevich-LMO invariant}\label{subsec:Codomain} Next, we define the strict monoidal category $\mathcal{A}$ as follows: The set of objects is $\mathrm{Mon}(+,-)$. A morphism from $w$ to $v$ is an element of $\coprod_X \mathcal{A}(X,\emptyset)$, the union is taken over all oriented compact 1-manifolds $X$ such that each start (resp.\ end) point corresponds to a letter $+$ in $w$ or $-$ in $v$ (resp.\ $-$ in $v$ or $+$ in $w$). Roughly speaking, we assign $+$ to the boundary of an interval which faces downward. The composition of two elements $D\in\mathcal{A}(v,u)$ and $E\in\mathcal{A}(w,v)$ is defined to be stacking $E$ on $D$ according to the letters in $v$. We often denote by $D\natural E$ the composite $D\circ_{\mathcal{A}} E$. The identity of $w$ is the intervals $\downarrow^w$. The tensor product of two morphisms is defined to be their disjoint union. The unit is the empty word $e$. Here, we prepare three operations in accordance with \cite{CHM08}. Let $\Delta$ and $S$ denote the doubling map and orientation reversal map defined as usual: \[\Delta\colon \mathcal{A}(X \downarrow,C) \to \mathcal{A}(X \downarrow \downarrow,C),\quad S\colon \mathcal{A}(X \downarrow,C) \to \mathcal{A}(X \uparrow,C).\] Let $w,w_1,\dots,w_n \in \mathrm{Mon}(+,-)$ such that $|w|=n$. We define the linear map \[\Delta^{w}_{w_1,\dots,w_n}\colon \mathcal{A}(X \downarrow^w,C) \to \mathcal{A}(X \downarrow^{w_1 \dotsm w_n},C)\] by applying $\Delta$ as much as $(|w_i|-1)$ times to the interval indexed by the $i$th letter in $w$ for $i=1,\dots,n$, and applying $S$ to the intervals where the two letters differ. \subsection{Definition of the Kontsevich-LMO invariant}\label{subsec:DefKontsevichLMOinv} We first fix an (rational) associator $\Phi \in \mathcal{A}(\downarrow \downarrow \downarrow)$ to define the Kontsevich invariant of a $q$-tangle in $[-1,1]^3$. The \emph{Kontsevich invariant} $Z^{\mathrm{K}}$ is defined by $Z^{\mathrm{K}}(\gamma):=Z(\gamma)$ for a $q$-tangle $\gamma$, where $Z$ is the Kontsevich integral in \cite{CHM08}. By definition, $Z^{\mathrm{K}}$ defines a tensor-preserving functor $\mathcal{T}_q \to \mathcal{A}$, that is, it satisfies the following conditions: \begin{enumerate} \item $Z^{\mathrm{K}}(\mathrm{Id}_w) = \ \downarrow^w$, \item $Z^{\mathrm{K}}(\gamma\circ\omega) = Z^{\mathrm{K}}(\gamma)\circZ^{\mathrm{K}}(\omega)$, \item $Z^{\mathrm{K}}(e) = e$, \item $Z^{\mathrm{K}}(\gamma\otimes\omega) = Z^{\mathrm{K}}(\gamma)\otimesZ^{\mathrm{K}}(\omega)$. \end{enumerate} Next, we review the Kontsevich-LMO invariant of a $q$-tangle in a homology cube $B$. We define $U_\pm$ by \[U_\pm := \int_{\{\ast\}} \chi_{\{\ast\}}^{-1}Z^{\mathrm{K}}(\circlearrowleft_{\pm1}) \sharp \nu \in \mathcal{A}(\emptyset,\emptyset),\] where $\ast$ is the label of the $(\pm1)$-framed unknot, and $\nu$ denotes the Kontsevich invariant of the 0-framed unknot, and $\sharp$ means the multiplication induced by the connected sum of two circles. The \emph{Kontsevich-LMO invariant} $Z^{\mathrm{K\text{-}LMO}}$ is defined by \[Z^{\mathrm{K\text{-}LMO}}(B,\gamma) := U_{+}^{-\sigma_{+}(L)} \sqcup U_{-}^{-\sigma_{-}(L)} \sqcup \int_{\pi_0 L} \chi_{\pi_0 L}^{-1} Z^{\mathrm{K}}(L^\nu \cup \gamma) \in \mathcal{A}(\gamma,\emptyset)\] for a $q$-tangle $(B,\gamma)$, where $(L,\gamma)$ is a surgery presentation of $(B,\gamma)$, $\sigma_\pm(L)$ is the number of positive/negative eigenvalues of the linking matrix of $L$, and let $Z^{\mathrm{K}}(L^\nu \cup \gamma)$ denote $Z^{\mathrm{K}}(L \cup \gamma)\sharp\nu^{\otimes L}$. One can check that $Z^{\mathrm{K\text{-}LMO}}$ defines a tensor-preserving functor $\mathcal{T}_q\mathcal{C}\mathit{ub} \to \mathcal{A}$. Here, a bottom-top $q$-tangle in a homology cube $B$ is regarded as a $q$-tangle in $B$ via the embedding ${}_{\mathit{b}}^{\mathit{t}}\mathcal{T}_q(w,v) \to \mathcal{T}_q\mathcal{C}\mathit{ub}(w',v')$, where words $v',w' \in \mathrm{Mag}(\bullet,\circ)$ are obtained from words $v,w \in \mathrm{Mag}(+,-)$ by the rule \[\bullet\mapsto(+-),\quad \circ\mapsto+.\] Then we denote by $Z(B,\gamma)$ the Kontsevich-LMO invariant of a bottom-top $q$-tangle $(B,\gamma)$. The next lemma is used in Section~\ref{subsec:ConstExtLMOfunc} and proved similarly to \cite[Lemma 3.17]{CHM08}. The only difference is that the definition of $\operatorname{Lk}_B(\gamma)$ is extended. \begin{lemma}\label{lem:bottom-top tangle} Let $(B,\gamma)$ be a bottom-top $q$-tangle. Then $\chi_{\pi_0 \gamma}^{-1}Z(B,\gamma) \in \mathcal{A}(\pi_0 \gamma)$ is group-like and its $s$-reduction is equal to $[\operatorname{Lk}_B(\gamma)/2]$. \end{lemma} \section{Extension of the LMO functor}\label{sec:ExtLMOfunc} In this section, we construct categories $\mathcal{LC}\mathit{ob}_q$, ${}^{\mathit{ts}}\!\mathcal{A}$ and an extension of the LMO functor $\widetilde{Z}\colon \mathcal{LC}\mathit{ob}_q \to {}^{\mathit{ts}}\!\mathcal{A}$. \subsection{$q$-cobordisms}\label{subsec:qCob} We refine cobordisms by replacing the monoid $\mathrm{Mon}(\bullet,\circ)$ with the magma $\mathrm{Mag}(\bullet,\circ)$. Let $w \in \mathrm{Mag}(\bullet,\circ)$. We denote by $F_w$ the compact surface $F_{w'}$ defined in Section~\ref{subsec:Cob}, where $w'$ is the word in $\mathrm{Mon}(\bullet,\circ)$ obtained from $w$ by forgetting its parenthesization. In the same manner, we use the notation $R^{w_+}_{w_-,\sigma}$. \begin{definition}\label{def:qCob} Let $w_+, w_- \in \mathrm{Mag}(\bullet,\circ)$ with $|w_+^\circ|=|w_-^\circ|$. A \emph{$q$-cobordism} from $F_{w_+}$ to $F_{w_-}$ is a cobordism from $F_{w_+}$ to $F_{w_-}$ equipped with the parenthesizations of $w_+$ and $w_-$. \end{definition} We define the category $\mathcal{C}\mathit{ob}_q$ of $q$-cobordisms by the same way as the category $\mathcal{C}\mathit{ob}$ in Section~\ref{subsec:Cob}. Similarly, the wide subcategory $\mathcal{LC}\mathit{ob}_q$ of \emph{Lagrangian $q$-cobordisms} is obtained, which is the domain of an extension of the LMO functor. Furthermore, we set $Z(M):=Z(\mathrm{D}^{-1}(M))$ for a $q$-cobordism $M=(M,\sigma,m)$, where the functor $\mathrm{D}$ has been defined in Section~\ref{subsec:BTtangle}, and $Z(M)$ is called the \emph{unnormalized LMO invariant} of $M$ in \cite{CHM08}. \subsection{Top-substantial Jacobi diagrams}\label{subsec:TopSubstantial} An element $x \in \mathcal{A}(\fc{g}^{+}\cup\fc{f}^{-}\cup\fc{n}^0)$ is said to be \emph{top-substantial} if $x$ is $\fc{g}^+$-substantial. Recall that the term $S$-substantial is defined in Section~\ref{subsec:GaussianInt}. We define the strict monoidal category ${}^{\mathit{ts}}\!\mathcal{A}$ as follows: The set of objects is the set $\mathbb{Z}_{\ge0}^2$ of pairs of non-negative integers. A morphism from $(g,n)$ to $(f,n)$ is a pair $(x,\sigma)$, where $x \in \mathcal{A}(\fc{g}^{+}\cup\fc{f}^{-}\cup\fc{n}^0)$ is top-substantial and $\sigma \in \SS_n$. There is no morphism from $(g,n)$ to $(f,m)$ if $n \neq m$, and we simply denote ${}^{\mathit{ts}}\!\mathcal{A}((g,n),(f,n))$ by ${}^{\mathit{ts}}\!\mathcal{A}(g,f,n)$. The composition of $(x,\sigma) \in {}^{\mathit{ts}}\!\mathcal{A}(g,f,n)$ and $(y,\tau) \in {}^{\mathit{ts}}\!\mathcal{A}(h,g,n)$ is defined to be \[\left( \chi_{\fc{n}^0}^{-1} \circ \chi_{\fc{n}^0,\fc{n}^{0'}} \ang{ \left(x \middle/ \parbox{5.6em}{$i^{+}\mapsto i^\ast,\\ i^0\mapsto \tau^{-1}(i)^0$}\right), \left(y \middle/ \parbox{3.8em}{$i^{-}\mapsto i^\ast,\\ i^0\mapsto i^{0'}$}\right) }_{\fc{g}^\ast}, \sigma\tau \right),\] where two kinds of $\chi$'s have been defined in Section~\ref{subsec:JacobiDiag}. The identity of $(g,n)$ is $\left[\sum\limits_{i=1}^g \strutgraph{i^+}{i^-}\right]$. The tensor product of $(x,\sigma) \in {}^{\mathit{ts}}\!\mathcal{A}(g,f,n)$ and $(y,\tau) \in {}^{\mathit{ts}}\!\mathcal{A}(g',f',n')$ is defined to be \[\left( x\sqcup(y/j^{\pm}\mapsto(n+j)^{\pm},j^0\mapsto(n+j)^0), \sigma\oplus\tau \right),\] where $\sigma\oplus\tau \in \SS_{n+n'}$ maps $i$ to $\sigma(i)$ (resp.\ $n+\tau(i-n)$) if $i\leq n$ (resp.\ $i>n$). Finally, $(g,n)\otimes(g',n'):=(g+g',n+n')$ and the unit is $(0,0)$. Note that we omit $\sigma$ from $(x,\sigma)$ when no confusion can arise. Let us describe the above composition law concretely. \begin{lemma}[see {\cite[Lemma 4.4]{CHM08}}]\label{lem:PreComposition} Let $x=(x,\sigma)\in{}^{\mathit{ts}}\!\mathcal{A}(g,f,n)$, $y=(y,\tau)\in{}^{\mathit{ts}}\!\mathcal{A}(h,g,n)$ and let $C=(c_{i^-,j^+})$ be a $\fc{f}^{-}\times\fc{g}^{+}$ matrix that is regarded as a linear map $C\colon \mathbb{Q}\fc{g}^{+} \to \mathbb{Q}\fc{f}^{-}$. Then $([C]\sqcup x)\circ y$ is equal to \[\chi_{\fc{n}^0}^{-1}\chi_{\fc{n}^0,\fc{n}^{0'}}\ang{(x/j^{+}\mapsto j^\ast, j^0\mapsto \tau^{-1}(j)^{0}), (y/j^{-}\mapsto j^\ast +Cj^{+},j^0\mapsto j^{0'})}_{\fc{g}^\ast},\] where $(y/j^{-}\mapsto Cj^+) := \sum_{i=1}^f (c_{i^-,j^+}y / j^-\mapsto i^-)$. Similarly, let $D$ be a $\fc{g}^-\times\fc{h}^+$ matrix. Then $x\circ([D]\sqcup y)$ is equal to \[\chi_{\fc{n}^0}^{-1}\chi_{\fc{n}^0,\fc{n}^{0'}}\ang{(x/i^{+}\mapsto i^\ast +Di^{-}, i^0\mapsto \tau^{-1}(i)^{0}), (y/i^-\mapsto i^\ast,i^0\mapsto i^{0'})}_{\fc{g}^\ast}.\] \end{lemma} The next lemma is proven by applying the previous lemma repeatedly. \begin{lemma}[see {\cite[Lemma 4.5]{CHM08}}] \label{lem:Composition} Let $x,y$ be as above and suppose that they are decomposed as $x=[X/2]\sqcup x^Y$ and $y=[Y/2]\sqcup y^Y$, where \[X=\begin{pmatrix} O & X^{+-} & X^{+0} \\ X^{-+} & X^{--} & X^{-0} \\ X^{0+} & X^{0-} & X^{00} \end{pmatrix},\quad Y=\begin{pmatrix} O & Y^{+-} & Y^{+0} \\ Y^{-+} & Y^{--} & Y^{-0} \\ Y^{0+} & Y^{0-} & Y^{00} \end{pmatrix}.\] Then $x\circ y$ is equal to \begin{align*} & [X^{--}/2+Y^{+-}X^{+-}+X^{-+}Y^{--}X^{+-}/2] \\ &\quad \sqcup\chi_{\fc{n}^0}^{-1}\chi_{\fc{n}^0,\fc{n}^{0'}}\left([\tau^{-1}(X^{00}/2+X^{-0})+Y^{00}/2+Y^{+0}]\sqcup\ang{L,R}_{\fc{g}^\ast}\right), \end{align*} where \begin{align*} L &:= \left([\tau^{-1}X^{+0}]\sqcup x^Y/i^{+} \mapsto i^\ast +X^{-+}Y^{--}i^{-} +Y^{+-}i^{-}, i^0 \mapsto \tau^{-1}(i)^0 \right), \\ R &:= \left([Y^{--}/2]/i^{-} \mapsto i^\ast\right)\sqcup\left([Y^{-0}]\sqcup y^Y/i^{-} \mapsto i^\ast +X^{-+}i^{+}, j^0 \mapsto j^{0'}\right). \end{align*} \end{lemma} \begin{corollary}\label{cor:Composition} In the above lemma, if $Y= \begin{pmatrix} O & I_g & O \\ I_g & O & O \\ O & O & O \end{pmatrix}$ and $\varepsilon(x)=\varepsilon(y)=\varnothing$, then $(x\circ y)^s = [X/2]$. \end{corollary} \begin{proof} By the previous Lemma, we have \begin{align*} x\circ y &= [X^{--}/2+X^{+-}]\sqcup\chi_{\fc{n}^0}^{-1}\chi_{\fc{n}^0,\fc{n}^{0'}}\bigl([X^{00}/2+X^{-0}/2] \\ &\quad \sqcup\ang{([X^{+0}]\sqcup x^Y/i^{+}\mapsto i^\ast+i^{+}),(y^Y/i^{-}\mapsto i^\ast +X^{-+}i^{+})}_{\fc{g}^\ast}\bigr). \end{align*} Hence, using Lemma~\ref{lem:filter}, $(x\circ y)^s$ is equal to \[[X^{--}/2+X^{+-}]\sqcup\chi_{\fc{n}^0}^{-1}\chi_{\fc{n}^0,\fc{n}^{0'}}\left([X^{00}/2+X^{-0}/2]\sqcup[X^{+0}]\right) = [X/2].\] \end{proof} \subsection{Construction of an extension of the LMO functor}\label{subsec:ConstExtLMOfunc} In \cite{CHM08}, a certain element $\mathsf{T}_g \in \mathcal{A}(\fc{g}^+\cup\fc{g}^-)$ is introduced, which roughly corresponds to the bottom-top tangle $T_w$ defined in Section \ref{subsec:BTtangle}. We first set \[\lambda(x,y;r):=\chi_{\{r\}}^{-1}\left(\BCHgraph{x}{y}{r}\right) \in \mathcal{A}(\emptyset,\{x,y,r\}),\] where \rotatebox{90}{$[\ ]$} is the map $\exp_\natural\colon \mathcal{A}_{>0}(\downarrow) \to \mathcal{A}(\downarrow)$. Next, $\mathsf{T}(x_+,x_-) \in \mathcal{A}(\{x_+,x_-\})$ is defined to be \[U_+^{-1} \sqcup U_-^{-1} \sqcup \int_{\{r_+,r_-\}} \ang{\lambda(y_-,x_-;r_-)\sqcup\lambda(x_+,y_+;r_+),\chi^{-1}Z(T_\bullet^\nu)}_{\{y_+,y_-\}}.\] Finally, $\mathsf{T}_g$ is defined by \[\mathsf{T}_g := \mathsf{T}(1^+,1^-) \sqcup \dots \sqcup \mathsf{T}(g^+,g^-) \in \mathcal{A}(\fc{g}^+\cup\fc{g}^-).\] Since $\mathsf{T}(x_+,x_-)$ is group-like and its $s$-reduction is equal to $\left[\strutgraph{x_+}{x_-}\right]$ (\cite[Lemma~4.9]{CHM08}), $\mathsf{T}_g$ is also group-like and its $s$-reduction is equal to $[I_g]$. The proof of the following key lemma is similar to \cite[Lemma~4.10]{CHM08}, however, we must pay attention to the additional map $\chi_{\fc{n}^0}^{-1} \circ \chi_{\fc{n}^0,\fc{n}^{0'}}$ in the definition of the composition law in the category ${}^{\mathit{ts}}\!\mathcal{A}$. \begin{lemma} \label{lem:key} Let $B=(B,\gamma)$, $C=(C,\omega)$ be bottom-top $q$-tangles of type $(v,u,\sigma)$, $(w,v,\tau)$ respectively and suppose that $\mathrm{D}(B)$, $\mathrm{D}(C)$ are Lagrangian. Then we have \[\left(\chi_{\pi_0(\gamma^{-}\cup\omega^{+}\cup\gamma^0\omega^0)}^{-1} Z(B\circ C), \sigma\tau\right) = (\chi_{\pi_0 \gamma}^{-1}Z(B), \sigma) \circ \mathsf{T}_{|v^\bullet|} \circ (\chi_{\pi_0 \omega}^{-1}Z(C), \tau),\] where $\mathsf{T}_{|v^\bullet|}=(\mathsf{T}_{|v^\bullet|},\mathrm{Id}_{\SS_{|v^\circ|}})$ is regarded as an element of ${}^{\mathit{ts}}\!\mathcal{A}(|v^\bullet|,|v^\bullet|,|v^\circ|)$. \end{lemma} \begin{proof} Let $(K,\gamma)$, $(L,\omega)$ be surgery presentations of $B$, $C$ respectively. Let $T^-$ (resp.\ $T^+$) be the framed link in $B\circ C$ obtained from $\gamma^+$ and the bottom components of $T_v$ (resp.\ $\omega^-$ and the top components of $T_v$) by gluing their boundaries. Finally, let $T:=T^-\cup T^+$ and $S:=K\cup T\cup L$. Then $(S,\gamma^{-}\cup\omega^{+}\cup\gamma^0\omega^0)$ is a surgery presentation of $B\circ C$, and thus \[Z(B\circ C) = U_+^{-\sigma_+(S)} \sqcup U_-^{-\sigma_-(S)} \sqcup \int_{\pi_0 S} \chi_{\pi_0 S}^{-1}Z^{\mathrm{K}}\left( S^\nu \cup (\gamma^{-}\cup\omega^{+}\cup\gamma^0\omega^0) \right).\] Here, the above integration is computed as follows: \begin{align*} & \int_{\pi_0 S}\chi_{\pi_0 S}^{-1} \left(Z^{\mathrm{K}}(K^\nu\cup\gamma)\natural (Z(T_w)\sharp\nu^{\otimes 2g}) \naturalZ^{\mathrm{K}}(L^\nu\cup\omega)\right) \\ &= \int_{\pi_0 T}\chi_{\pi_0 T}^{-1}\int_{\pi_0 K}\chi_{\pi_0 K}^{-1}Z^{\mathrm{K}}(K^\nu\cup\gamma)\natural (Z(T_w)\sharp\nu^{\otimes 2g}) \natural\int_{\pi_0 L}\chi_{\pi_0 L}^{-1}Z^{\mathrm{K}}(L^\nu\cup\omega), \end{align*} where $g:=|v^\bullet|$, and $\natural$ means the composition in the category $\mathcal{A}$. It follows from the identity $\sigma_\pm(S) =\sigma_\pm(K) +g +\sigma_\pm(L)$ (\cite[(4-2)]{CHM08}) that \[Z(B\circ C) = U_{+}^{-g}\sqcup U_{-}^{-g}\sqcup\int_{\pi_0 T}\chi_{\pi_0 T}^{-1} \left( Z(K^\nu\cup\gamma)\natural (Z(T_w)\sharp\nu^{\otimes 2g}) \natural Z(L^\nu\cup\omega) \right).\] Therefore, $\chi_{\pi_0(\gamma^{-}\cup\omega^{+}\cup\gamma^0\omega^0)}^{-1} Z(B\circ C)$ is equal to \begin{align*} & U_{+}^{-g}\sqcup U_{-}^{-g}\sqcup \int_{\pi_0 T} \chi_{\pi_0(T\cup\gamma^0\omega^0)}^{-1} \left( \chi_{\pi_0 \gamma^{-}}^{-1}Z(K^\nu\cup\gamma)\natural (Z(T_w)\sharp\nu^{\otimes 2g}) \natural\chi_{\pi_0 \omega^{+}}^{-1}Z(L^\nu\cup\omega) \right). \end{align*} Let us introduce notation to calculate the above integration: \begin{align*} \pi_0(T^{-}) &= \fc{g}^\bot, & \pi_0(T_{w^\bullet}^{+}) &= \fc{g}^\triangledown, & \pi_0(\omega^{+}) &= \fc{h}^+, \\ \pi_0(\gamma^{+}) &= \fc{f}^\cup, & \pi_0(T_{w^\bullet}^{-}) &= \fc{g}^\vartriangle, & \pi_0(\omega^{-}) &= \fc{h}^\cap, \\ \pi_0(\gamma^{-}) &= \fc{f}^{-}, & & & \pi_0(T^{+}) &= \fc{g}^\top, \end{align*} where $f:=|u^\bullet|$, $h:=|w^\bullet|$. Then, the above integrand is computed as follows: \begin{align*} & \chi_{\pi_0(T\cup\gamma^0\omega^0)}^{-1} \left( \chi_{\pi_0 \gamma^+}\chi_{\pi_0 \gamma^\pm}^{-1}Z(B) \natural \chi_{\pi_0 T_{w^\bullet}}\chi_{\pi_0 T_{w^\bullet}}^{-1}(Z(T_{w^\bullet}^\nu)\downarrow^{\fc{n}^0}) \natural \chi_{\pi_0 \omega^-}\chi_{\pi_0 \omega^\pm}^{-1}Z(C) \right) \\ &= \chi_{\pi_0(T\cup\gamma^0\omega^0)}^{-1} \biggl\langle \bigsqcup_{i=1}^g \BCHgraph{i^\vartriangle}{i^\cup}{i^\bot} \sqcup \bigsqcup_{i=1}^g \BCHgraph{i^\cap}{i^\triangledown}{i^\top}\ , \chi_{\pi_0 T_{w^\bullet}}^{-1}Z(T_{w^\bullet}^\nu) \sqcup \chi_{\pi_0 \gamma^\pm}^{-1}Z(B) \natural \chi_{\pi_0 \omega^\pm}^{-1}Z(C) \biggr\rangle_{\parbox{2em}{\scriptsize $\mathord{\vartriangle}\mathord{\cap} \\[-4pt] \mathord{\cup}\mathord{\triangledown}$}} \\ &= \chi_{\pi_0(\gamma^0\omega^0)}^{-1} \ang{ \ang{ \Lambda, \chi_{\pi_0 T_{w^\bullet}}^{-1}Z(T_{w^\bullet}^\nu) }_{\vartriangle \triangledown}, \chi_{\pi_0 \gamma^\pm}^{-1}Z(B) \natural \chi_{\pi_0 \omega^\pm}^{-1}Z(C) }_{\cup \cap} \\ &= \chi_{\pi_0(\gamma^0\omega^0)}^{-1} \ang{ \ang{ \Lambda, \chi_{\pi_0 T_{w^\bullet}}^{-1}Z(T_{w^\bullet}^\nu) }_{\vartriangle \triangledown}, \chi_{\pi_0 \gamma^0, \pi_0 \omega^0} \left(\chi_{\pi_0 \gamma}^{-1}Z(B) \sqcup \chi_{\pi_0 \omega}^{-1}Z(C)\right) }_{\cup \cap}, \end{align*} where the $i$th component of $\gamma^0$ corresponds to the $\tau^{-1}(i)$th component of $\omega^0$, and $\Lambda \in \mathcal{A}(\fc{g}^\vartriangle \cup \fc{f}^\cup \cup \fc{g}^\bot \cup \fc{h}^\cap \cup \fc{g}^\triangledown \cup \fc{g}^\top)$ is defined by \[\Lambda := \bigsqcup_{i=1}^g \lambda(i^\vartriangle,i^\cup;i^\bot) \sqcup \bigsqcup_{i=1}^g \lambda(i^\cap,i^\triangledown;i^\top).\] Therefore, the above integration is written as follows: \[\chi_{\pi_0(\gamma^0\omega^0)}^{-1} \chi_{\pi_0 \gamma^0, \pi_0 \omega^0} \ang{ \int_{\pi_0 T} \ang{\Lambda, \chi_{\pi_0 T_{w^\bullet}}^{-1}Z(T_{w^\bullet}^\nu) }_{\vartriangle \triangledown}, \chi_{\pi_0 \gamma}^{-1}Z(B) \sqcup \chi_{\pi_0 \omega}^{-1}Z(C) }_{\cup \cap}.\] Here, \begin{align*} & U_{+}^{-g} \sqcup U_{-}^{-g} \sqcup \int_{\pi_0 T} \ang{\Lambda, \chi_{\pi_0 T_{w^\bullet}}^{-1}Z(T_{w^\bullet}^\nu) }_{\vartriangle \triangledown} \\ &= \bigsqcup_{i=1}^g U_{+}^{-1} \sqcup U_{-}^{-1} \sqcup \int_{\{i^\bot,i^\top\}} \ang{ \lambda(i^\vartriangle,i^\cup;i^\bot) \sqcup \lambda(i^\cap,i^\triangledown;i^\top)} \\ &= \bigsqcup_{i=1}^g \mathsf{T}(i^\cap,i^\cup) = \left(\mathsf{T}_g \middle/ \parbox{40pt}{$i^+ \mapsto i^\cap \\ i^- \mapsto i^\cup$} \right). \end{align*} Therefore, the left-hand side of this lemma is equal to \begin{align*} \chi_{\pi_0(\gamma^0\omega^0)}^{-1} \chi_{\pi_0 \gamma^0, \pi_0 \omega^0} \ang{ \left(\mathsf{T}_g \middle/ \parbox{40pt}{$i^+ \mapsto i^\cap \\ i^- \mapsto i^\cup$} \right), \chi_{\pi_0 \gamma}^{-1}Z(B) \sqcup \chi_{\pi_0 \omega}^{-1}Z(C) }_{\cup \cap}. \end{align*} On the other hand, the right-hand side of this lemma is obtained as follows: Connect the $i^-$-colored vertices and $i^+$-colored vertices in $\mathsf{T}_g$ with the $i^\cup$-colored vertices in $\chi_{\pi_0 \gamma}^{-1}Z(B)$ and $i^\cap$-colored vertices in $\chi_{\pi_0 \omega}^{-1}Z(C)$ respectively and apply the composite map $\chi_{\pi_0(\gamma^0\omega^0)}^{-1} \chi_{\pi_0 \gamma^0, \pi_0 \omega^0}$. \end{proof} We can now define $\widetilde{Z}$, which is the main purpose of this paper. \begin{definition} Let $v, w \in \mathrm{Mag}(\bullet,\circ)$ with $|v^\circ|=|w^\circ|=n$ and let $f:=|v^\bullet|$, $g:=|w^\bullet|$. The \emph{normalized LMO invariant} $\widetilde{Z}$ of $(M,\sigma,m) \in \mathcal{LC}\mathit{ob}_q(w,v)$ is defined by \[\widetilde{Z}(M,\sigma,m) := \left(\chi_{\pi_0 \gamma}^{-1} Z(B,\gamma), \sigma\right) \circ_{{}^{\mathit{ts}}\!\mathcal{A}} \left(\mathsf{T}_g, \mathrm{Id}_{\SS_n}\right) \in {}^{\mathit{ts}}\!\mathcal{A}(g,f,n).\] \end{definition} Let $\widetilde{Z}^s(M)$ (resp.\ $\widetilde{Z}^Y(M)$) denote the $s$- (resp.\ $Y$-) reduction of $\widetilde{Z}(M)=\widetilde{Z}(M,\sigma,m)$. The next lemma follows from Lemma~\ref{lem:bottom-top tangle} and Corollary~\ref{cor:Composition}~(1). \begin{lemma} $\widetilde{Z}(M,\sigma,m)$ is group-like and $\widetilde{Z}^s(M,\sigma,m) = [\operatorname{Lk}(M)/2]$. \end{lemma} \begin{theorem}\label{thm:Zt} The normalized LMO invariant defines a tensor-preserving functor \[\widetilde{Z} \colon \mathcal{LC}\mathit{ob}_q\to{}^{\mathit{ts}}\!\mathcal{A},\] which is called \emph{(}an extension of\/\textup{)} the \emph{LMO functor}, that is, $\widetilde{Z}$ satisfies the following conditions: \begin{enumerate \item $\widetilde{Z}(\mathrm{Id}_w)=\mathrm{Id}_{(|w^\bullet|,|w^\circ|)}$, \item $\widetilde{Z}(M\circ N)=\widetilde{Z}(M)\circ\widetilde{Z}(N)$, \item $\widetilde{Z}(e)=(0,0)$, \item $\widetilde{Z}(M\otimes N)=\widetilde{Z}(M)\otimes\widetilde{Z}(N)$. \end{enumerate} \end{theorem} \begin{proof} (3) and (4) follows from these of $Z^{\mathrm{K}}$ (see Section~\ref{subsec:DefKontsevichLMOinv}). (2) is due to Lemma~\ref{lem:key}. Let us prove (1). By the previous lemma, $\widetilde{Z}(\mathrm{Id}_w)$ is group-like, thus it is written as $\varnothing+(\operatorname{i-deg}>0)$. We assume that the higher-degree terms are not zero, and write \[\widetilde{Z}(\mathrm{Id}_w) = \varnothing +x +(\operatorname{i-deg}>k) \quad(\deg x=k).\] It follows from (2) that $\widetilde{Z}(\mathrm{Id}_w) = \widetilde{Z}(\mathrm{Id}_w)\circ\widetilde{Z}(\mathrm{Id}_w)$. By comparing both side of this equality, \begin{align*} & \chi_{\fc{n}^0}^{-1} \chi_{\fc{n}^0,\fc{n}^{0'}} \ang{ \left( \widetilde{Z}^Y(\mathrm{Id}_w) \middle/ i^{+}\mapsto i^\ast \right), \left( \widetilde{Z}^Y(\mathrm{Id}_w) \middle/ \parbox{3.7em}{$i^{-}\mapsto i^\ast\\ i^0\mapsto i^{0'}$} \right) }_{\fc{g}^\ast} \\ &= \chi_{\fc{n}^0}^{-1} \chi_{\fc{n}^0,\fc{n}^{0'}}\left( \varnothing +x +(x/i^0\mapsto i^{0'}) +(\operatorname{i-deg}>k) \right) \\ &= \varnothing +2x +(\operatorname{i-deg}>k). \end{align*} Thus we have $x=2x$, and it is a contradiction. Therefore, $\widetilde{Z}^Y(\mathrm{Id}_w)=\varnothing$, and we conclude (4). \end{proof} \begin{remark} By definition, if $w \in \mathrm{Mag}(\bullet)$, then $\widetilde{Z}$ coincides with the original LMO functor defined in \cite{CHM08}. Moreover, the diagram \[\xymatrix{\mathcal{LC}\mathit{ob}_q \ar[r]^-{\mathrm{K}} \ar[d]_-{\widetilde{Z}} & \mathcal{LC}\mathit{ob}^\bullet_q \ar[d]^-{\text{original $\widetilde{Z}$}} \\ {}^{\mathit{ts}}\!\mathcal{A} \ar[r]^-{(-/i^0\mapsto0)} & {}^{\mathit{ts}}\!\mathcal{A}}\] commutes, where $\mathrm{K}$ is the functor defined in the same manner as $\mathrm{K}$ in Remark~\ref{rem:Kill}. In contrast, if $w \in \mathrm{Mag}(\circ)$, then we have \[\widetilde{Z}(\mathrm{D}([-1,1]^3,-\sigma)) = \left(\chi_{\pi_0 \sigma}^{-1} Z^{\mathrm{K}}(-\sigma), \mathrm{Id}_{\SS_n}\right)\] for any (not necessarily pure) string link $\sigma$ in $[-1,1]^3$ on $n$ strands. \end{remark} \begin{remark}\label{rem:Rational} By the same manner, one can define the category $\mathbb{Q}\mathcal{LC}\mathit{ob}$ of \emph{rational Lagrangian $q$-cobordisms} and the functor $\widetilde{Z}\colon \mathbb{Q}\mathcal{LC}\mathit{ob}_q \to {}^{\mathit{ts}}\!\mathcal{A}$, see \cite[Remark~2.8, 3.19, 4.14]{CHM08}. Indeed, the formal Gaussian integral is originally defined under rational condition and linking matrices in this paper are defined homologically. \end{remark} \section{Generators and the values on them}\label{sec:GenValue} In \cite{CHM08}, it was proven that the category $\mathcal{C}\mathit{ob}^\bullet$, see Remark~\ref{rem:Kill}, is generated as a monoidal category by the morphisms $\mu$, $\eta$, $\Delta$, $\varepsilon$, $S^{\pm 1}$, $\psi_{\bullet,\bullet}^{\pm 1}$, $v_\pm$ and $Y$ listed in Table~\ref{tab:Gen}, and the values on them was calculated up to internal degree 2. We shall add some morphisms and calculate the values on them. \subsection{Generators of the category $\mathcal{LC}\mathit{ob}$}\label{subsec:Gen} We first define some bottom-top tangles as Table~\ref{tab:Gen}. By Lemma~\ref{lem:Lagrangian}, the corresponding cobordisms are Lagrangian, and we use the same notation to represent the cobordisms. The bottom-top tangles except $\psi_{\bullet,\circ}$, $\psi_{\circ,\bullet}$, $\psi_{\circ,\circ}$, $\tau$ and $\beta$ are same as in \cite{CHM08}. \begin{table} \centering $ \begin{array}{cccc} \eta:=\raisegraph{-1em}{2.5em}{eta.pdf} & \varepsilon:=\raisegraph{-0.8em}{2.5em}{epsilon.pdf} & v_{+}:=\raisegraph{-1em}{2.5em}{v+.pdf} & v_{-}:=\raisegraph{-1em}{2.5em}{v-.pdf} \\ \mu:=\raisegraph{-2em}{5em}{mu.pdf} & \Delta:=\raisegraph{-2em}{5em}{Delta.pdf} & \psi_{\bullet,\bullet}:=\raisegraph{-2em}{5em}{psibb.pdf} & \psi_{\circ,\circ}:=\raisegraph{-2em}{5em}{psiww.pdf} \\ \tau:=\raisegraph{-2em}{5em}{tau.pdf} & \beta:=\raisegraph{-2em}{5em}{beta.pdf} & \psi_{\bullet,\circ}:=\raisegraph{-2em}{5em}{psibw.pdf} & \psi_{\circ,\bullet}:=\raisegraph{-2em}{5em}{psiwb.pdf} \\ S:=\raisegraph{-2em}{5em}{S.pdf} & \multicolumn{3}{l}{Y:=\raisegraph{-2em}{5em}{Ycob.pdf}} \end{array} $ \vspace{1em} \caption{Some morphisms of the category $\mathcal{LC}\mathit{ob}$} \label{tab:Gen} \end{table}% \begin{remark} When we write $Y$ as a morphism of $\mathcal{LC}\mathit{ob}_q$, we always regard it as a morphism from $(\bullet\bullet)\bullet$ to $e$. In general, when we regard a morphism of $\mathcal{C}\mathit{ob}$ as one of $\mathcal{C}\mathit{ob}_q$, we use the left-handed parenthesization $(\cdots((\ast\ast)\ast)\dots\ast)$ unless otherwise stated. \end{remark} \begin{remark}\label{rem:braided} Four kinds of $\psi$'s and their inverses induce braidings $\psi_{v,w}\colon v\otimes w \to w\otimes v$, which make the category $\mathcal{LC}\mathit{ob}$ into a braided strict monoidal category. \end{remark} It is easy to see that the cobordisms \begin{center} \raisegraph{-2em}{5em}{psibb-1.pdf}\ ,\quad \raisegraph{-2em}{5em}{psibw-1.pdf}\ ,\quad \raisegraph{-2em}{5em}{beta-1.pdf} and \raisegraph{-2em}{5em}{S-1.pdf} \end{center} are the (two-sided) inverses of $\psi_{\bullet,\bullet}$, $\psi_{\bullet,\circ}$, $\beta$ and $S$ in the category $\mathcal{LC}\mathit{ob}_q$. In the same manner, one finds the inverses of $\psi_{\circ,\circ}$, $\psi_{\circ,\bullet}$ and $\tau$. (By Proposition~\ref{prop:InvertElem}, it suffices to check that they are left or right inverses.) \begin{proposition}\label{prop:Gen} The category $\mathcal{LC}\mathit{ob}$ is generated as monoidal category by the morphisms $\mu$, $\eta$, $\Delta$, $\varepsilon$, $S^{\pm1}$, $\psi_{\bullet,\bullet}^{\pm 1}$, $v_\pm$, $Y$, $\psi_{\bullet,\circ}^{\pm1}$, $\psi_{\circ,\bullet}^{\pm1}$, $\psi_{\circ,\circ}^{\pm1}$ and $\tau$. Moreover, $\mathcal{LC}\mathit{ob}_q$ is generated by the morphism \[P_{u,v,w}:=\raisegraph{-2em}{5em}{Puvw.pdf}\] for $u,v,w \in \mathrm{Mag}(\bullet,\circ)$ and their inverses together with the above morphisms. \end{proposition} \begin{proof} The latter follows from the former. We prove that every $M \in \mathcal{LC}\mathit{ob}(w,v)$ belongs to the category $\mathcal{C}$ generated by the above morphisms. Using the braidings, it suffices to consider the case of $M=(M,\mathrm{Id}_{\SS_n},m) \in \mathcal{LC}\mathit{ob}(\bullet^{\otimes g}\circ^{\otimes n},\bullet^{\otimes f}\circ^{\otimes n})$. Namely, the bottom-top tangle $(B,\gamma):=\mathrm{D}^{-1}(M)$ is represented as follows: \begin{center} \includegraphics[width=0.99\textwidth]{GenTrans.pdf}. \end{center} (This correspondence is essentially same as the preferred bijection $\tau_n$ in \cite[Section~13]{Hab06}.) Since $\operatorname{Lk}_B(\gamma^+)=O$, it suffice to show that the bottom-top tangle $\alpha_n$ at the lower right in the right-hand side of the above equality belongs to $\mathcal{C}$. Here, one can show the following equalities \[\alpha_1 = (\varepsilon \otimes \mathrm{Id}_\circ) \circ \psi_{\bullet,\circ}^{-1} \circ \tau \circ \psi_{\circ,\bullet}^{-1},\quad \alpha_n = (\text{braidings})\circ\alpha_1^{\otimes n}.\] Therefore, the proof is completed. \end{proof} \begin{example} One can check the following decomposition \[\raisegraph{-2.2em}{5em}{StringLink.pdf} = ((\varepsilon\otimes\mathrm{Id}_\circ)\check{\tau}\otimes\mathrm{Id}_\circ)(\mathrm{Id}_\bullet\otimes\psi_{\circ,\circ}^2)(\psi_{\bullet,\circ}^{-1}\otimes\mathrm{Id}_\circ)(\tau\otimes\mathrm{Id}_\circ)(\mathrm{Id}_\circ \otimes v_{-}\otimes\mathrm{Id}_\circ),\] where the composition $\circ$ is omitted and the morphism $\check{\tau}$ is defined by \[\check{\tau} := \raisegraph{-2em}{5em}{ttau.pdf} = \psi_{\bullet,\circ}^{-1} \circ \tau \circ \psi_{\bullet,\circ}\ .\] \end{example} \subsection{The values on the generators}\label{subsec:Value} From now on, we assume an associator $\Phi \in \mathcal{A}(\downarrow\downarrow\downarrow)$ is the one derived from a rational even Drinfel'd series $\varphi(A,B) \in \mathbb{Q}\angg{A,B}$, which was introduced in \cite[Section~3]{LeMu97}. By the pentagon and hexagon relations and the evenness, $\varphi(A,B)$ must be of the form \[1 +\frac{1}{24}[A,B] +(\deg>3).\] Thus, under the above assumption, we have \[\Phi = \varphi\left(\raisegraph{-1em}{2.5em}{ass23.pdf},\raisegraph{-1em}{2.5em}{ass12.pdf}\right) = \raisegraph{-1em}{2.5em}{ass0.pdf} +\frac{1}{24}\raisegraph{-1em}{2.5em}{ass123.pdf} +(\deg>3).\] We can calculate the values on $\psi_{\bullet,\circ}^{\pm1}$, $\psi_{\circ,\bullet}^{\pm1}$, $\psi_{\circ,\circ}^{\pm1}$ and $\beta$ using \cite[Lemma~5.7]{CHM08}, and the results are written in Table~\ref{tab:Value}. \begin{table} \centering $ \begin{array}{c|c|c} M & \log_\sqcup\widetilde{Z}^s(M) & \log_\sqcup\widetilde{Z}^Y(M) \bmod (\operatorname{i-deg}>2) \\ \hline \mu & \strutgraph{1^+}{1^-} +\strutgraph{2^+}{1^-} & -\frac{1}{2}\Ygraph{1^+}{2^+}{1^-} +\frac{1}{12}\yengraph{1^+}{1^+}{2^+}{1^-} +\frac{1}{12}\myengraph{1^+}{2^+}{2^+}{1^-} \rule[-1.5em]{0em}{3.7em} \\ \hline \eta & 0 & 0 \rule[-0.7em]{0em}{2em} \\ \hline \Delta & \strutgraph{1^+}{1^-} +\strutgraph{1^+}{1^-} & \frac{1}{2}\dYgraph{1^+}{1^-}{2^-} +\frac{1}{12}\mlambdagraph{1^+}{1^-}{2^-}{2^-} +\frac{1}{12}\lambdagraph{1^+}{1^-}{1^-}{2^-} - \frac{1}{4}\Hgraph{1^+}{1^+}{1^-}{2^-} \rule[-1.5em]{0em}{3.7em} \\ \hline \varepsilon & 0 & 0 \rule[-0.7em]{0em}{2em} \\ \hline S^{\pm1} & -\strutgraph{1^+}{1^-} & \mp\frac{1}{4} \Phigraph{1^+}{1^-} \mp\frac{1}{4}\Hgraph{1^+}{1^+}{1^-}{1^-} \rule[-1.5em]{0em}{3.7em} \\ \hline \psi_{\bullet,\bullet}^{\pm1} & \strutgraph{1^+}{2^-} +\strutgraph{2^+}{1^-} & \mp\frac{1}{2}\Hgraph{1^+}{2^+}{2^-}{1^-} \rule[-1.2em]{0em}{3em} \\ \hline v_\pm & \mp\frac{1}{2}\dstrutgraph{1^-}{1^-} & \frac{1}{48} \dPhigraph{1^-}{1^-} \rule[-1em]{0em}{3em} \\ \hline Y & 0 & -\uYgraph{1^+}{2^+}{3^+} +\frac{1}{2}\uHgraph{1^+}{1^+}{2^+}{3^+} +\frac{1}{2}\uHgraph{2^+}{2^+}{3^+}{1^+} +\frac{1}{2}\uHgraph{3^+}{3^+}{1^+}{2^+} \rule[-1.5em]{0em}{3.7em} \\ \hline P_{u,v,w} & \sum\limits_{i=1}^{|u^\bullet|+|v^\bullet|+|w^\bullet|} \!\strutgraph{i^+}{i^-} & 0 \rule[-1.1em]{0em}{3em} \\ \hline \psi_{\bullet,\circ}^{\pm1} & \strutgraph{1^+}{1^-} & \pm\frac{1}{2}\rTgraph{1^+}{1^-}{1^0} +\frac{1}{8}\rPigraph{1^+}{1^-}{1^0}{1^0} \rule[-1.5em]{0em}{3.7em} \\ \hline \psi_{\circ,\bullet}^{\pm1} & \strutgraph{1^+}{1^-} & \pm\frac{1}{2}\rTgraph{1^+}{1^-}{1^0} +\frac{1}{8}\rPigraph{1^+}{1^-}{1^0}{1^0} \rule[-1.5em]{0em}{3.7em} \\ \hline \psi_{\circ,\circ}^{\pm1} & \pm\frac{1}{2}\dstrutgraph{1^0}{2^0} & -\frac{1}{32}\lrPhigraph{1^0}{2^0} \pm\frac{1}{24}\lrHgraph{1^0}{2^0}{1^0}{2^0} \rule[-1.2em]{0em}{3em}\\ \hline \tau & \strutgraph{1^+}{1^-} +\lustrutgraph{1^+}{1^0} & -\frac{1}{2}\lTgraph{1^+}{1^-}{1^0} +\frac{1}{4}\uchairgraph{1^+}{1^+}{1^-}{1^0} +\frac{1}{12}\chairgraph{1^+}{1^-}{1^-}{1^0} +\frac{1}{12}\lPigraph{1^+}{1^-}{1^0}{1^0} -\frac{1}{4}\luPhigraph{1^+}{1^0} +\frac{1}{4}\luHgraph{1^+}{1^+}{1^0}{1^0} \rule[-1.5em]{0em}{3.7em}\\ \hline \tau^{-1} & \strutgraph{1^+}{1^-} -\lustrutgraph{1^+}{1^0} & \frac{1}{2}\lTgraph{1^+}{1^-}{1^0} -\frac{1}{4}\uchairgraph{1^+}{1^+}{1^-}{1^0} -\frac{1}{12}\chairgraph{1^+}{1^-}{1^-}{1^0} +\frac{1}{12}\lPigraph{1^+}{1^-}{1^0}{1^0} \rule[-1.5em]{0em}{3.7em}\\ \hline \beta^{\pm1} & \strutgraph{1^+}{1^-} \pm\ldstrutgraph{1^-}{1^0} & \mp\frac{1}{2}\lTgraph{1^+}{1^-}{1^0} -\frac{1}{8}\ldPhigraph{1^-}{1^0} \pm\frac{1}{12}\uchairgraph{1^+}{1^+}{1^-}{1^0} -\frac{1}{12}\lPigraph{1^+}{1^-}{1^0}{1^0} \pm\frac{1}{8}\ldHgraph{1^-}{1^-}{1^0}{1^0} \rule[-1.5em]{0em}{3.7em}\\ \end{array} $ \vspace{1em} \caption{The values on the generators, $\tau^{-1}$ and $\beta^{\pm1}$.} \label{tab:Value} \end{table}% On the other hand, to calculate the value on $\tau^{\pm1}$, we need Lemma~\ref{lem:updown} that is an upside-down version of \cite[Lemma~5.5]{CHM08}. One can prove Lemma~\ref{lem:updown} in a similar way. \begin{lemma}\label{lem:updown} Let $M \in \mathcal{LC}\mathit{ob}_q(w,v)$. Suppose $\mathrm{D}^{-1}(M)$ is the composition of the $q$-tangle \[U:=\raisegraph{-4em}{8em}{U.pdf}\] and some $q$-tangle $L$ in $[-1,1]^3$. Then $\mathsf{T}_f\circ\widetilde{Z}(M)\circ\mathsf{T}_g^{-1} \in {}^{\mathit{ts}}\!\mathcal{A}(g,f,n)$ is equal to the image by $\chi_{\pi_0(U^{+}\cup L)}^{-1}$ of the composition of the series of Jacobi diagrams \[\raisegraph{-5em}{10em}{V.pdf}\] and $Z^{\mathrm{K}}(L) \in \mathcal{A}(L,\emptyset)$ in the category $\mathcal{A}$. Here, $C_{w_i}$ and a directed rectangle is same as in \cite[Lemma~5.5]{CHM08}. \end{lemma} \begin{remark} The lower degree terms of $\mathsf{T}_1$ and $\mathsf{T}_1^{-1}$ are given in \cite[Lemma~5.7]{CHM08} and its proof. \end{remark} \section{Universality among finite-type invariants}\label{sec:Universal} Le~\cite{Le97} proved that the LMO invariant is universal among rational-valued finite-type invariants of rational homology spheres (see \cite[p.~89, Remark~1]{Le97}). As mentioned before, Cheptea, Habiro and Massuyeau showed that the LMO functor is universal among rational-valued finite-type invariants of 3-manifolds. We prove that the extension $\widetilde{Z}$ has the same property. \subsection{Clasper calculus}\label{subsec:Clasper} In this subsection, we prepare terminology of clasper calculus based on \cite{HabK00}. Consider pairs $(M,\gamma)$ where $M$ is a compact oriented 3-manifold and $\gamma$ is a framed oriented tangle in $M$. If $M$ has a boundary, we take into account a parametrization of $\partial M$. Moreover, if $\gamma$ has a boundary, it must be attached to assigned points in $\partial M$. \begin{definition} A \emph{graph clasper} is an oriented compact surface with a certain decomposition into three kinds of pieces: disks, bands and annuli, which are called \emph{nodes}, \emph{edges} and \emph{leaves} respectively. Each leaf should be connected to a node and no leaf by a band. Each node should be connected to nodes or leaves by exactly three bands. \end{definition} The surface on the left of Figure~\ref{fig:Ygraph} is one of the simplest example of a graph clasper. For simplicity, it is represented as the graph on the right of Figure~\ref{fig:Ygraph}, that is, leaves, nodes and bands of a graph clasper $G$ are replaced with circles, points and arcs respectively, here we should remember information required for recovering $G$. \begin{figure}[h] \centering \includegraphics[width=0.7\columnwidth]{Ygraph.pdf} \caption{An example of a graph clasper and its alternative expression} \label{fig:Ygraph} \end{figure}% Next, consider surgery along a graph clasper $G$ in a 3-manifold $M$. Let $Y(G)$ denotes the graph clasper obtained from $G$ by applying the \emph{fission rule} illustrated in Figure~\ref{fig:FissionRule} to each edge connecting two nodes. The resulting graph is the disjoint union of $\operatorname{i-deg} G$ copies of $Y$-graphs, where $\operatorname{i-deg} G$ is the \emph{internal degree} of $G$, that is, the number of nodes of $G$. \begin{figure}[h] \centering \includegraphics[width=0.6\columnwidth]{FissionRule.pdf} \caption{The fission rule} \label{fig:FissionRule} \end{figure}% The pair $(M_G,\gamma_G)$ obtained from $(M,\gamma)$ by surgery along $G$ is defined as follows: Consider the case in which $G$ is a $Y$-graph. Let $Y_0$ be a $Y$-graph in $\mathbb{R}^3$ and $N(Y_0)$ be its neighborhood, namely a genus three handlebody. Then there is an orientation-preserving homeomorphism $h\colon N(Y_0) \to N(G)$ sending $Y_0$ to $G$. Here, we replace $Y_0$ with the six-components framed link drawn in Figure~\ref{fig:YgraphSurgery}. We denote by $M_G$ the manifold obtained from $G$ by surgery along the framed link in $N(G)$ corresponding to the above link, and let $\gamma_G$ denote the image of $\gamma$ in $M_G$. In the general case, we perform surgery along each $Y$-graph of $Y(G)$. \begin{figure}[h] \centering \includegraphics[width=0.8\columnwidth]{YgraphSurgery.pdf} \caption{A $Y$-graph, its neighborhood and the corresponding link} \label{fig:YgraphSurgery} \end{figure}% \begin{definition} Let $k\ge1$. A pair $(M,\gamma)$ is \emph{$Y_k$-equivalent} to $(M',\gamma')$ if there exists a graph clasper $G$ such that the internal degree of each connected component is equal to $k$ and $G$ satisfies $(M_G,\gamma_G)=(M',\gamma')$. \end{definition} It is known that the $Y_k$-equivalence is an equivalence relation (see, for example, \cite[Theorem~3.2]{GGP01}). Let us fix a $Y_1$-equivalence class $\mathcal{M}^0$ of 3-manifolds with tangles (see \cite[Theorem~7.6]{CHM08}). \begin{definition} We define $\mathcal{F}_k(\mathcal{M}^0)$ to be the $\mathbb{Q}$-vector space spanned by $\{[M,G] \mid M \in \mathcal{M}^0,\ \operatorname{i-deg} G=k\}$. Here, $[M,G]$ is defined by \[[M,G] := \sum_{G' \subset G}(-1)^{|G'|}M_{G'} \in \mathbb{Q}\mathcal{M}^0,\] where $G'$ runs over all connected components of $G$, and $|G'|$ is the number of connected components of $G'$. (The above sum consists of $2^{|G|}$ terms.) \end{definition} We have the \emph{$Y$-filtration} \[\mathbb{Q}\mathcal{M}^0 = \mathcal{F}_0(\mathcal{M}^0) \supset \mathcal{F}_1(\mathcal{M}^0) \supset \dotsb,\] and we set $\mathcal{G}_i(\mathcal{M}^0) := \mathcal{F}_i(\mathcal{M}^0)/\mathcal{F}_{i+1}(\mathcal{M}^0)$. Moreover, the associated graded vector space $\mathcal{G}(\mathcal{M}^0)$ is defined by $\mathcal{G}(\mathcal{M}^0) := \prod_{i\geq1}\mathcal{G}_i(\mathcal{M}^0)$. \begin{definition} Let $V$ be a $\mathbb{Q}$-vector space and let $d\ge0$. A map $f \colon \mathcal{M}^0 \to V$ is a ($V$-valued) \emph{finite-type invariant} of \emph{degree} $d$ if the linear extension $\widetilde{f} \colon \mathbb{Q}\mathcal{M}^0 \to V$ is non-trivial on $\mathcal{F}_d(\mathcal{M}^0)$ and trivial on $\mathcal{F}_{d+1}(\mathcal{M}^0)$. \end{definition} \subsection{Proof of the universality}\label{subsec:ProofUniversal} To state Theorem~\ref{thm:Gar02} below, we review the surgery map $\SS$ according to \cite{CHM08}. Fix a $Y_1$-equivalence class $\mathcal{M}^0$ of $\mathcal{LC}\mathit{ob}_q(w,v)$ and let $g:=|w^\bullet|$, $f:=|v^\bullet|$, $n:=|w^\circ|$. Let $D$ be a Jacobi diagram in the vector space $\mathcal{A}^Y_i(\fc{g}^{+}\cup\fc{f}^{-}\cup\fc{n}^0)$. We first construct an oriented surface $S(D)$ by replacing the internal vertices, external vertices and edges to disks, annuli and bands respectively, where the vertex orientation of $D$ induces the orientation of disks, and the bands connect the disks so that the orientations are compatible. For each annulus, $A$, the core of $A$ is defined to be the push-off into $\operatorname{Int} A$ of the outer boundary of $A$ that is the boundary connecting with an edge. Namely, the core is an oriented simple closed curve on $A$. Next, we take a $q$-cobordism $(M,\sigma,m) \in \mathcal{LC}\mathit{ob}_q(w,v)$. The graph clasper $G(D)$ is defined to be the image of an embedding $S(D)$ into $M$ such that each annulus corresponding to the $j^-$- (resp.\ $j^+$-, $j^0$-) colored vertices is the push-off into $\operatorname{Int} M$ of a (unoriented) neighborhood of $m_{-}(\alpha_j) \subset \partial M$ (resp.\ $m_{+}(\beta_j)$, $m_{+}(\delta_j)$) and its core corresponds to the oriented curve $m_{-}(\alpha_j)$ (resp.\ $m_{+}(\beta_j)$, $m_{+}(\delta_j)$). Here, $G(D)$ is called a \emph{topological realization} of a Jacobi diagram $D$. \begin{figure}[h] \centering \includegraphics[width=0.9\columnwidth]{TopologicalRealization.pdf} \caption{A Jacobi diagram, the corresponding graph clasper and its embedding into a 3-manifold} \label{fig:TopologicalRealization} \end{figure} The manifold $M_{G(D)}$ depends on the choice of the cobordism $(M,\sigma,m)$ and the topological realization $G(D)$. However, it is known that $[M,G(D)] \in \mathcal{G}_i(\mathcal{M}^0)$ is independent of the choice, and the following theorem holds. \begin{theorem}[{\cite[Corollary 1.4]{Gar02}}]\label{thm:Gar02} The graded linear map \[\SS \colon \mathcal{A}^Y(\fc{g}^{+}\cup\fc{f}^{-}\cup\fc{n}^0) \to \mathcal{G}(\mathcal{M}^0)\] defined by $D \mapsto [M,G(D)]$ is well-defined and surjective. \end{theorem} Using this theorem, we prove the next theorem saying that $\widetilde{Z}$ is universal among rational-valued finite-type invariants. The following proof is an obvious extension of the proof of \cite[Theorem 7.11]{CHM08}. \begin{theorem}\label{thm:universal} Let $\mathcal{A}^Y:=\mathcal{A}^Y(\fc{g}^{+}\cup\fc{f}^{-}\cup\fc{n}^0)$. The map $\widetilde{Z}^Y_i \colon \mathcal{M}^0 \to \mathcal{A}^Y\times\{\sigma\} = \mathcal{A}^Y$ is a finite-type invariant of degree at most $i$. Moreover, the induced map $\operatorname{Gr}\widetilde{Z}^Y \colon \mathcal{G}(\mathcal{M}^0) \to \mathcal{G}(\mathcal{A}^Y)=\mathcal{A}^Y$ satisfies \[\operatorname{Gr}\widetilde{Z}^Y\circ\SS(D)=(-1)^{i+c+e}D,\] where $i:=\operatorname{i-deg} D$, $c:=|D|$ and $e$ is the number of internal edges of $D$. Consequently, $\SS$ and $\operatorname{Gr}\widetilde{Z}$ are isomorphisms, and $\widetilde{Z}^Y_i$ is a finite-type invariant of degree $i$. \end{theorem} \begin{proof} Let us prove that $\widetilde{Z}^Y_{i-1}$ is a finite-type invariant of degree at most $i-1$. We first take a graph clasper $G$ in a cobordism $M=(M,\sigma,m) \in\mathcal{M}^0$ such that $[M,G] \in\mathcal{F}_i(\mathcal{M}^0)$. Let $i:=\operatorname{i-deg} G$, $c:=|G|$ and $Y(G) = G_1\sqcup\dots\sqcup G_i$. Then we have \begin{align*} [M,Y(G)] &= \sum_{J \subset\fc{i}} (-1)^{|J|} M_{\bigsqcup_{j\in J} G_j} \\ &= \sum_{G'\subset G} (-1)^{|G\setminus G'|+|Y(G\setminus G')|} (-1)^{|Y(G')|} M_{G'} \\ &= \sum_{G'\subset G} (-1)^{c-|G'|+i} M_{G'} = (-1)^{i+c}[M,G], \end{align*} where the second equality follows from the equality \[\prod_{j=1}^{|G\setminus G'|} \left((1+t_j)^{|Y(G_j)|}-t_j^{|Y(G_j)|}\right)\Big|_{t_j=-1} = (-1)^{|G\setminus G'|+|Y(G\setminus G')|}\] and the fact that surgery along a part of a connected component is same as doing nothing. Using the transformation depicted in Figure~\ref{fig:InsertHopfLink}, the bottom-top tangle $\mathrm{D}^{-1}(M_G)$ can be expressed as Figure~\ref{fig:universal}, where $R$ is some cobordism of $\mathcal{LC}\mathit{ob}(w,v\otimes((\bullet\bullet)\bullet)^{\otimes i})$. Moreover, $Y_1,\dots,Y_i$ are the bottom-top tangles illustrated in Figure~\ref{fig:YbyClasper}, which are same as the cobordism $Y$ listed in Table~\ref{tab:Gen}. \begin{figure}[h] \centering \begin{minipage}{0.45\columnwidth} \centering \includegraphics[width=0.6\columnwidth]{InsertHopfLink.pdf} \caption{A transformation} \label{fig:InsertHopfLink} \end{minipage} \begin{minipage}{0.45\columnwidth} \centering \includegraphics[width=0.7\columnwidth]{YbyClasper.pdf} \caption{An expression of $Y$ by using a $Y$-graph} \label{fig:YbyClasper} \end{minipage} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.7\columnwidth]{universal.pdf} \caption{A bottom-top tangle (The thick lines are parts of tangles and claspers.)} \label{fig:universal} \end{figure} Then we have, by the functoriality of $\widetilde{Z}$, \begin{align}\label{eq:Universal} & (-1)^{i+c}\widetilde{Z}([M,G]) = \widetilde{Z}([M,Y(G)]) \\ &= \left(\widetilde{Z}(\mathrm{Id}_v) \otimes \bigotimes_{j=1}^i \left( \widetilde{Z}((\varepsilon\otimes\varepsilon)\otimes\varepsilon) -\widetilde{Z}(Y_j) \right) \right) \circ \widetilde{Z}(R) \notag \\ &= \left( \left[\sum_{k=1}^f \strutgraph{k^+}{k^-}\right] \sqcup \bigsqcup_{j=1}^i \left(-\widetilde{Z}_1(Y_j) +(\operatorname{i-deg}>1)\right) \right) \circ \widetilde{Z}(R). \notag \end{align} Since the composition in ${}^{\mathit{ts}}\!\mathcal{A}$ preserves the filtration $\{{}^{\mathit{ts}}\!\mathcal{A}_{\geq i}\}_{i\geq0}$, we conclude $\widetilde{Z}^Y_{i-1}([M,G])=0$. Next, we prove the equality in the second statement. We take a Jacobi diagram $D \in\mathcal{A}^Y_i$ and define $G$ to be a topological realization $G(D)$ in some cobordism $M$. The goal is to show $\operatorname{Gr}\widetilde{Z}([M,G])=(-1)^{i+c+e}D$. We first define three sets by \begin{align*} \mathcal{L} &:= \bigcup_{j=1}^i \{f+3j-2, f+3j-1, f+3j\} \subset \fc{f+3i}, \\ \mathcal{O} &:= \{l \in \mathcal{L} \mid \text{the $l$th leaf is originally a leaf}\}, \\ \mathcal{T} &:= \{(t_1,t_2) \in \mathcal{L}\times\mathcal{L} \mid \text{the $t_1$th and $t_2$th leaves arise by the fission rule}\}. \end{align*} Since the leaf corresponding to $l \in \mathcal{O}$ arises from some external vertex of $D$, we define $c(l)$ to be its color. On the other hand, $\mathcal{T}$ is divided into $\mathcal{T}_<$ and $\mathcal{T}_>$, where $\mathcal{T}_\lessgtr:=\{(t_1,t_2) \in \mathcal{T} \mid t_1 \lessgtr t_2\}$ and $|\mathcal{T}_\lessgtr|=e$. We now have \begin{align*} \operatorname{Lk}(R) = \operatorname{Lk}(M_G) +\sum_{l\in\mathcal{O}}(E_{l^-,c(l)}+E_{c(l),l^-}) -\sum_{(t_1,t_2)\in\mathcal{T}}E_{t_1^-,t_2^-}\ (=:B), \end{align*} where $E_{i,j}$ is the matrix unit. It follows from this equality, \eqref{eq:Universal} and Lemma~\ref{lem:Composition} that \begin{align*} & (-1)^{i+c}\widetilde{Z}^Y_i([M,G]) \\ &= \ang{ \left( \bigsqcup_{j=1}^i -\widetilde{Z}_1(Y_j) \middle/ \parbox{7.8em}{$k^+ \mapsto k^\ast +B^{--}k^{-} \\\hspace*{2.5em} +B^{+-}k^{-}$} \right), \left( [B^{--}/2+B^{-0}] \middle/ k^{-}\mapsto k^\ast \right) }_{\fc{f+3i}^\ast} \\ &= \ang{ \left( \bigsqcup_{j=1}^i -\widetilde{Z}_1(Y_j) \middle/ \parbox{11em}{$l^+ \mapsto \begin{cases} c(l) & \text{if $l\in\mathcal{O}$}, \\ l^\ast & \text{otherwise.} \end{cases}$} \right), \bigsqcup_{(t_1,t_2)\in\mathcal{T}_<} -\strutgraph{t_2^\ast}{t_1^\ast} }_{\{l^\ast \mid l\in\mathcal{L}\setminus\mathcal{O}\}} \\ &= (-1)^e D, \end{align*} where the second equality follows from the following argument: Since the linear combination on the left in the angular brackets contains no $k^0$-colored vertex, $[B^{-0}]$ can be moved to the left as Lemma~\ref{lem:PreComposition}. \end{proof} \section{Relations with knot theory}\label{sec:Knot} We focus on cobordisms derived from knots with additional information. \subsection{Cobordisms between two annuli}\label{subsec:Annuli} We first review a natural construction of a cobordism from a knot according to \cite[Section~2.2]{CFK11}. Let $Y$ be a homology sphere and let $K$ be an oriented framed knot in $Y$. The boundary of $Y\setminus \operatorname{Int} N(K)$ is a torus with the oriented longitude determined by the orientation and framing of $K$. Here, let $m\colon R^{\circ}_{\circ,\mathrm{Id}} \to \partial(Y\setminus \operatorname{Int} N(K))$ be a homeomorphism that maps the oriented simple closed curves $\delta_1$ and $\widehat{\delta_1}$ depicted in Figure~\ref{fig:CFK11} to the meridian and longitude of $K$, respectively. Then we obtain $\mathrm{C}(Y,K):=(Y\setminus \operatorname{Int} N(K), \mathrm{Id}_{\SS_1}, m) \in \mathcal{LC}\mathit{ob}_q(\circ,\circ)$. \begin{figure}[h] \centering \includegraphics[width=0.2\columnwidth]{CFK11.pdf} \caption{The reference surface $R^{\circ}_{\circ,\mathrm{Id}}$ with two curves $\delta_1$, $\widehat{\delta_1}$} \label{fig:CFK11} \end{figure} \begin{proposition} Let $K \subset Y$ be an oriented framed knot in a homology sphere. Then, under the canonical isomorphism $\mathcal{A}(\downarrow)\cong\mathcal{A}(\circlearrowleft)$, \[\chi_{\fc{1}^0}\widetilde{Z}(\mathrm{C}(Y,K)) = Z^{\mathrm{K\text{-}LMO}}(Y\setminus\operatorname{Int}[-1,1]^3,K) \sharp \nu^{-1}.\] \end{proposition} \begin{proof} Let $(B,\gamma):=\mathrm{D}^{-1}(\mathrm{C}(Y,K))$. Then the closure $(\widehat{B},\widehat{\gamma})$ defined in Section~\ref{subsec:BTtangle} coincides with the pair $(Y,K)$. Therefore, \begin{align*} \chi_{\fc{1}^0}\widetilde{Z}(\mathrm{C}(Y,K)) &= Z^{\mathrm{K\text{-}LMO}}(Y\setminus\operatorname{Int}[-1,1]^3,\gamma) \\ &= Z^{\mathrm{K\text{-}LMO}}(Y\setminus\operatorname{Int}[-1,1]^3,K) \sharp \nu^{-1}. \end{align*} The proof is completed. \end{proof} Since $\widetilde{Z}$ is a functor, we have \[\widetilde{Z}(\mathrm{C}(Y_1,K_1)) \circ \widetilde{Z}(\mathrm{C}(Y_2,K_2)) = \widetilde{Z}(\mathrm{C}(Y_1,K_1) \circ \mathrm{C}(Y_2,K_2)).\] Here, $\mathrm{C}(Y_1,K_1) \circ \mathrm{C}(Y_2,K_2)$ can be interpreted as a certain connected sum defined as follows: Consider an embedding between pairs $\iota_i \colon ([-1,1]^3,0\times0\times[-1,1]) \to (Y_i,K_i)$ that preserves the orientations and framing. Then the above composition is equal to the pair \[\left( (Y_1,K_1)\setminus\operatorname{Int}[-1,1]^3 \sqcup (Y_2,K_2)\setminus\operatorname{Int}[-1,1]^3 \right) / \iota_1(x,y,z)\sim\iota_2(x,y,-z).\] \subsection{Cobordisms derived from a link and its Seifert surfaces}\label{subsec:SeifertSurf} Let $Y$ be an integral homology sphere and let $L$ be an oriented link in $Y$. One can take a Seifert surface $F$ of $L$, and let $g$ be its genus. We fix an orientation-preserving homeomorphism $h\colon F_w \to F$, where $w \in \mathrm{Mon}(\bullet,\circ)$ satisfies $|w^\bullet|=g$, $|w^\circ|=|L|-1$. Using the above information, we construct the cobordism $(M,\mathrm{Id},m) \in \mathcal{C}\mathit{ob}(w,w)$ (that is not necessarily Lagrangian) as follows: Define $M$ to be the manifold obtained from $Y \setminus \operatorname{Int} N(L)$ by cutting along $F$. Now, $\partial M$ is the union of the surface $F_+$ to which the normal vectors point, $F_-$ and the $[-1,1]\times L$. Therefore, we decide $m_+\colon F_w \to F_-$ and $m_-\colon F_w \to F_+$ by $h$, and the rest is uniquely determined by the meridians of $L$. \begin{example} Let $L$ be the negative Hopf link in $S^3$. Since we need a Seifert surface that is easy to see, we represent the pair $(S^3,L)$ by surgery along the $(+1)$-framed unknot depicted in Figure~\ref{fig:HopfLinkSeifert}. One can now easily construct a Seifert surface by avoiding the surgery knot, and we give the parametrization indicated by the line drawn in the middle of Figure~\ref{fig:HopfLinkSeifert}. Then we have the Lagrangian cobordism $M$ illustrated in Figure~\ref{fig:HopfLinkSeifert}, where the oriented closed curve represents $h(\widehat{\delta_1})$. Hence, $\widetilde{Z}(M)$ is exactly equal to $\chi_{\fc{1}^0}^{-1} \exp_\natural \left(\dfrac{1}{2}\ \raisegraph{-1em}{2.5em}{ArrowArc.pdf}\right)$. It follows from \cite[Exercise~5.4]{BLT03} that \begin{align*} \widetilde{Z}(M) &= \Omega \sqcup \left(\exp_\sqcup \left(\frac{1}{2}\dstrutgraph{1^0}{1^0} \right) \right) = \left[\frac{1}{2}\dstrutgraph{1^0}{1^0} +\frac{1}{48}\dPhigraph{1^0}{1^0} +(\operatorname{i-deg}>2)\right]. \end{align*} \end{example} \begin{figure}[h] \centering \includegraphics[width=0.9\columnwidth]{HopfLinkSeifert.pdf} \caption{A Hopf link $L$, its Seifert surface and the cobordism obtained by cutting} \label{fig:HopfLinkSeifert} \end{figure}% \subsection{Milnor invariants of string links}\label{subsec:MilnorInv} Habegger and Masbaum~\cite{HaMa00} proved that the first non-vanishing Milnor invariant of a string link appears in the tree reduction of its Kontsevich invariant. The same is true for the Kontsevich-LMO invariant of string links in a homology cube, which was shown in \cite{Mof06}. Moreover, the same holds for the LMO functor of the cobordisms derived from string links in a homology cube, which was proven in \cite{CHM08}. We prove that the same is true for the extension of the LMO functor. We first review the Milnor invariants of string links along \cite{CHM08}. \begin{definition} A \emph{string link} on $l$ strands is an orientation-reversed bottom-top tangle of type $(w,w,\mathrm{Id}_{\SS_l})$, where $w$ is a word in $\mathrm{Mon}(\circ)$ of length $l$. Namely, each strand runs from $r_i\times(-1)$ to $r_i\times1$. \end{definition} Actually, we are interested in the monoid \[\mathcal{S}_l := \{\text{string link $(B,\sigma)$ on $l$ strands} \mid \text{$B$ is an \emph{integral} homology cube}\}.\] Let $(B,\sigma) \in \mathcal{S}_l$. Then $(S,\mathrm{Id}_{\SS_l},s):=\mathrm{D}(B,\sigma)$ is a cobordism from $F_w$ to $F_w$. However, in this subsection, we use the surface $D_l$ with the oriented simple closed curves $x_1,\dots,x_l$ as in Figure~\ref{fig:D} instead of $F_w$ with the oriented simple closed curves $\delta_1,\dots,\delta_l$ Note that the orientation of $x_i$ differs from the one of $\delta_i$. \begin{figure}[h] \centering \includegraphics[width=0.4\columnwidth]{D.pdf} \vspace{-0.8em} \caption{The surface $D_l = \Sigma_{0,l+1}$ with the closed curves $x_1,\dots,x_l$} \label{fig:D} \end{figure} Let $\varpi$ denote the fundamental group $\pi_1(D_l,\ast)$ and $\varpi_n$ is its lower central series defined by $\varpi_k:=[\varpi_{k-1},\varpi]$, $\varpi_1=\varpi$. Since the continuous maps \[s_{+},\ s_{-}\colon D_l \hookrightarrow \partial(D_l\times[-1,1]) \xrightarrow{s} S\] induce isomorphisms on their first homologies and surjections on their second homologies (\cite[Theorem~5.1]{Sta65}), they induce isomorphisms \[s_{+,\ast},\ s_{-,\ast}\colon \varpi/\varpi_k \to \pi_1(S)/\pi_1(S)_k\] for all $k\ge1$. The $k$th \emph{Artin representation} is the monoid anti-homomorphism $A_k\colon \mathcal{S}_l \to \operatorname{Aut}(\varpi/\varpi_{k+1})$ defined by $A_k(B,\sigma) := s_{+,\ast}^{-1} \circ s_{-,\ast}$. We set $\mathcal{S}_l[k]:=\operatorname{Ker} A_k$, and the filtration \[\mathcal{S}_l=\mathcal{S}_l[1] \supset \mathcal{S}_l[2] \supset \dotsb\] is obtained. From now on, we suppose that each strand of $\sigma$ is 0-framed, and let $\lambda_i$ be the preferred longitude of the $i$th strand of $\sigma$. It is well known that $(B,\sigma) \in \mathcal{S}_l$ belongs to $\mathcal{S}_l[k]$ if and only if $\lambda_i \in \pi_1(S)_k$ for all $i$ (see, for example, \cite[Remark~5.1]{HaMa00}). \begin{definition} The $k$th \emph{Milnor invariant} is the monoid homomorphism $\mu_k\colon \mathcal{S}_l[n] \to \varpi/\varpi_2 \otimes \varpi_k/\varpi_{k+1}$ defined by \[\mu_k(B,\sigma) := \sum_{i=1}^l x_i \otimes s_{+,\ast}^{-1}(\lambda_i).\] \end{definition} The following lemma is well known and directly follows from the above fact. \begin{lemma}\label{lem:mu} Let $(B,\sigma) \in \mathcal{S}_l[k]$. Then $\mu_k(B,\sigma)=0$ if and only if $(B,\sigma)$ belongs to $\mathcal{S}_l[k+1]$. \end{lemma} Next, we regard $\mu_k(B,\sigma)$ as a linear combination of connected tree Jacobi diagrams. The subalgebra of $\mathcal{A}(C)$ generated by tree Jacobi diagrams is denoted by $\mathcal{A}^t(C)$, which is identified with the quotient by the ideal generated by looped Jacobi diagrams, and the image of $x \in \mathcal{A}(C)$ by the quotient map is denoted by $x^t$. Moreover, the subspace of $\mathcal{A}^t(C)$ spanned by connected Jacobi diagrams is denoted by $\mathcal{A}^{t,c}(C)$. The free Lie algebra over $\mathbb{Q}$ generated by a set $C$ is denoted by $\operatorname{Lie}(C)$ and its degree $n$ part is denoted by $\operatorname{Lie}_n(C)$. We define the linear map $\eta_n\colon \mathcal{A}^{t,c}_n(C) \to \mathbb{Q} C \otimes_{\mathbb{Q}} \operatorname{Lie}_n(C)$ by \[\eta_k(D) := \sum_{\text{$v$}} (\text{color of $v$}) \otimes \operatorname{comm}(D_v),\] where $v$ runs over all external vertices in $D$, and $D_v$ is the tree rooted at $v$. Here, $\operatorname{comm}(D_v)$ is defined as follows: \[\operatorname{comm} \left( \raisegraph{-2.3em}{5em}{tree-crop.pdf} \right) = [c_1,[[c_2,c_3],c_4]].\] Moreover, it is well known that the sequence \begin{align}\label{eq:EtaBracket} 0 \to \mathcal{A}^{t,c}_{k-1}(C) \xrightarrow{\eta_{k-1}} \mathbb{Q} C \otimes_\mathbb{Q} \operatorname{Lie}_k(C) \xrightarrow{[-,-]} \operatorname{Lie}_{k+1}(C) \to 0 \end{align} is exact (see, for example, \cite[Theorem 1]{Lev02}). \begin{remark} The map $\eta_k$ is same as in \cite[Section~4.3]{HaMa12}, which differs from \cite{HabN00}, \cite{Lev02}, \cite{Mof06} and \cite{CHM08} by $(-1)^k$. \end{remark} Let $G$ be the free group generated by $C=\{c_1,\dots,c_r\}$ and let $G_k$ be its lower central series. Recall the well-known fact that there exists a natural isomorphism $\alpha_n\colon (G/G_2 \otimes G_k/G_{k+1}) \otimes \mathbb{Q} \to \mathbb{Q} C \otimes_\mathbb{Q} \operatorname{Lie}_k(C)$. In particular, if $C=\fc{l}^\ast$, then one can show that $[-,-] \circ \alpha_k \circ \mu_k = 0$. The exact sequence \eqref{eq:EtaBracket} enable us to define $\mu_k^{\mathcal{A}}(B,\sigma) \in \mathcal{A}^{t,c}_{k-1}(C)$ to be the unique element sent to $\alpha_k(\mu_k(B,\sigma))$ by $\eta_{k-1}$. Finally, we discuss a relation between cobordisms and string links. Let $w \in \mathrm{Mon}(\bullet,\circ)$ and let $g:=|w^\bullet|$, $n:=|w^\circ|$, $l:=2g+n$. Figure~\ref{fig:MilnorJohnson} represents a map from $\{(M,\tau,m) \in \mathcal{LC}\mathit{ob}(w,w) \mid \tau=\mathrm{Id}_{\SS_n}\}$ to $\mathcal{S}_l$, which sends $\mathrm{Id}_w$ to the trivial string link on $l$ strands, which is an extension of the original one defined in \cite{HabN00}, \cite{CHM08}. \begin{figure}[h] \centering \includegraphics[width=0.9\columnwidth]{MilnorJohnson.pdf} \caption{The Milnor-Johnson correspondence} \label{fig:MilnorJohnson} \end{figure} \begin{definition} The \emph{Milnor-Johnson correspondence} $\operatorname{MJ}_w$ is the bijection from the $Y_1$-equivalence class of $\mathrm{Id}_w$ in $\mathcal{LC}\mathit{ob}(w,w)$ to the $Y_1$-equivalence class of the trivial string link in $\mathcal{S}_l$, depicted in Figure~\ref{fig:MilnorJohnson}. \end{definition} Let us introduce a map $R_w$ for $w \in \mathrm{Mon}(\bullet,\circ)$ to state the main theorem in this section. We first define the bijection $\rho_w\colon \fc{g}^{+}\cup\fc{g}^{-}\cup\fc{n}^0 \to \fc{2g+n}^\ast$ by sending the $j$th color in the sequence $\left(w / \text{$i$th}\ \bullet \mapsto i^{-}i^{+},\ \text{$i$th}\ \circ \mapsto i^0 \right)$ to $j^\ast$. Next, $R_w\colon \mathcal{A}(X,\fc{g}^{+}\cup\fc{g}^{-}\cup\fc{n}^0) \to \mathcal{A}(X,\fc{2g+n}^\ast)$ is defined by \[R_w(D) := \left( D \middle/ i^{-}\mapsto -\rho_w(i^-),\ i^{+}\mapsto \rho_w(i^+),\ i^0\mapsto -\rho_w(i^0) \right).\] For example, if $w=\bullet\circ$, then we have \[R_w(D) = \left( D \middle/ 1^{-}\mapsto -1^\ast,\ 1^{+}\mapsto 2^\ast,\ 1^0\mapsto -3^\ast \right).\] \begin{figure}[h] \centering \includegraphics[width=0.8\columnwidth]{Psi.pdf} \caption{An example of a transformation with $w=\bullet\circ\bullet$} \label{fig:Psi} \end{figure} \begin{lemma}\label{lem:Psi} There exists a linear map \[\Psi\colon \mathcal{A}(\rotatebox[origin=c]{180}{$\curvearrowleft$}^{\fc{g}^+} \curvearrowleft^{\fc{g}^-} \downarrow^{\fc{n}^0}) \to \mathcal{A}(\downarrow^{\fc{2g+n}^\ast})\] such that $\Psi(Z(\mathrm{D}^{-1}(M))) = Z^{\mathrm{K\text{-}LMO}}(\operatorname{MJ}_w(M))$. Moreover, $\Psi$ satisfies \[\chi_{\fc{2g+n}^\ast}^{-1}\Psi(E) = \left[\sum_{i=1}^g \strutgraph{\rho_w(i^+)}{-\rho_w(i^-)} \qquad \right] \sqcup R_w(\widetilde{E}) + (\operatorname{i-deg}>\operatorname{i-deg} E),\] for any Jacobi diagram $E \in \mathcal{A}(\rotatebox[origin=c]{180}{$\curvearrowleft$}^{\fc{g}^+} \curvearrowleft^{\fc{g}^-} \downarrow^{\fc{n}^0})$. \end{lemma} \begin{proof} The string link $\operatorname{MJ}_w(M)$ can be transformed as illustrated in Figure~\ref{fig:Psi}. Therefore, we set \[\Psi(E) := Z^{\mathrm{K\text{-}LMO}}(\text{bottom part}) \circ (\mathrm{Id} \otimes S_{\fc{g}^{-}\cup\fc{n}^0}(E) \otimes \mathrm{Id}) \circ Z^{\mathrm{K\text{-}LMO}}(\text{top part}),\] where $S_{\fc{g}^{-}\cup\fc{n}^0}(E)$ is the element obtained from $E$ by applying the orientation reversal map $S$ to the $(\fc{g}^{-}\cup\fc{n}^0)$-indexed components of $E$. Then, $\Psi$ has the desired property. \end{proof} \begin{theorem}\label{thm:MilnorInv} Let $l\ge1$, $k\ge2$, $(B,\sigma) \in \mathcal{S}_l[k]$ and let $w \in \mathrm{Mon}(\bullet,\circ)$ such that $l=2|w^\bullet|+|w^\circ|$. Then $\widetilde{Z}^{Y,t}_{<k}(\operatorname{MJ}_w^{-1}(B,\sigma))$ is equal to \[\varnothing +R_w^{-1}(\mu_k^{\mathcal{A}}(B,\sigma)) \in \mathcal{A}^Y_{<k}(\fc{g}^{+}\cup\fc{g}^{-}\cup\fc{n}^0).\] Conversely, if $(B,\sigma) \in \mathcal{S}_l[2]$ and $\widetilde{Z}^{Y,t}_{<k}(\operatorname{MJ}_w^{-1}(M,\sigma))$ is of the form \[\varnothing + x \quad(\operatorname{i-deg} x=k-1),\] then $(B,\sigma)$ belongs to $\mathcal{S}_l[k]$, and $R_w(x)$ is equal to $\mu_k^{\mathcal{A}}(B,\sigma)$. \end{theorem} \begin{proof} Let $(B,\sigma) \in \mathcal{S}_l[k]$ and let $g=|w^\bullet|$, $n=|w^\circ|$. We set $(M,\mathrm{Id}_{\SS_n},m):=\operatorname{MJ}_w^{-1}(B,\sigma)$. Let us prove the former by induction on $k$. We first assume that it is true up to $k-1$, where $k\ge3$. (The case $k=2$ is mentioned later.) It follows from Lemma~\ref{lem:mu} that $\widetilde{Z}(M)$ is written as \[[I_g^{+-}] \sqcup \left(\varnothing +x +(\text{$\operatorname{i-deg} \ge k$ or looped}) \right),\] where $\deg x=k-1$. Hence, Lemma~\ref{lem:Composition} implies \begin{align*} & \widetilde{Z}(M) \circ_{{}^{\mathit{ts}}\!\mathcal{A}} \mathsf{T}_g^{-1} \\ &= \mathsf{T}_g^{-1} + [I_g^{+-}] \sqcup \ang{\left( x +(\text{$\operatorname{i-deg} \ge k$ or looped}) \middle/ \parbox{3em}{$i^{+}\mapsto \\ i^\ast +i^{+}$}\right), \left( (\mathsf{T}_g^{-1})^Y \middle/ \parbox{3em}{$i^{-}\mapsto \\ i^\ast +i^{-}$} \right)}_{\fc{g}^\ast} \\ &= \mathsf{T}_g^{-1} + [I_g^{+-}] \sqcup x +(\text{$\operatorname{i-deg} \ge k$ or looped}). \end{align*} By applying the composite map $\chi_{\fc{2g+n}^\ast}^{-1} \Psi \chi_{\fc{g}^{+}\cup\fc{g}^{-}\cup\fc{n}^0}$, we have \begin{align*} & \chi_{\fc{2g+n}^\ast}^{-1} \Psi \chi_{\fc{g}^{+}\cup\fc{g}^{-}\cup\fc{n}^0} (\widetilde{Z}(M) \circ \mathsf{T}_g^{-1}) \\ &= \chi^{-1}\Psi\chi(\mathsf{T}_g^{-1}) + \left[\sum_{i=1}^g \strutgraph{\rho_w(i^+)}{\rho_w(i^-)} \qquad \right] \sqcup R_w\left( \widetilde{\chi([I_g^{+-}] \sqcup x)} \right) + \left(\parbox{4.3em}{$\operatorname{i-deg} \ge k$ \\ \text{or looped}}\right) \\ &= \chi^{-1}\Psi(Z(\mathrm{Id}_w)) + \left[\sum_{i=1}^g \strutgraph{\rho_w(i^+)}{\rho_w(i^-)} \qquad \right] \sqcup R_w\left( [I_g^{+-}] \sqcup x \right) +\left(\parbox{4.3em}{$\operatorname{i-deg} \ge k$ \\ \text{or looped}}\right) \\ &= \varnothing +R_w(x) +(\text{$\operatorname{i-deg} \ge k$ or looped}), \end{align*} where we use Lemma~\ref{lem:Psi} and the fact that $\chi^{-1}\Psi\chi(\text{looped}) \subset (\text{looped})$. On the other hand, by \cite[Theorem~3]{Mof06}, \begin{align*} \chi^{-1}\Psi\chi (\widetilde{Z}(M) \circ \mathsf{T}_g^{-1}) &= \chi^{-1}\Psi Z(M) \\ &= \chi^{-1}Z^{\mathrm{K\text{-}LMO}}(\operatorname{MJ}_w(M)) \\ &= \varnothing + \mu_n(B,\sigma) +(\text{$\operatorname{i-deg} \ge k$ or looped}). \end{align*} Comparing the internal degree $k$ parts of each tree reduction, we conclude $\mu_k(B,\sigma)=x$. Here, we have to consider the case $n=2$. In general, $\widetilde{Z}^{Y,t}(M)$ is of the form $\varnothing +x +(\operatorname{i-deg}>1)$ where $\operatorname{i-deg} x=1$. Hence, applying the above argument, we conclude $x=\mu_2(B,\sigma)$. The latter follows from the former and Lemma~\ref{lem:mu}. \end{proof} \begin{example}\label{ex:Borromean} Let us compute the second Milnor invariant of $\sigma=([-1,1]^3,\sigma) \in \mathcal{S}_3[2]$ illustrated in Figure~\ref{fig:Borromean}. \begin{figure}[h] \centering \includegraphics[width=0.9\columnwidth]{Borromean.pdf} \caption{An examples of a bottom-top tangle, the corresponding string link and its closure} \label{fig:Borromean} \end{figure}% First, one can check that $\operatorname{MJ}_{\bullet\circ}^{-1}(\sigma) = \psi_{\circ,\bullet}\circ\psi_{\bullet,\circ}$, where $\psi_{\circ,\bullet}$ and $\psi_{\bullet,\circ}$ are the cobordisms define in Table~\ref{tab:Gen}. It follows from the functoriality of $\widetilde{Z}$, Lemma~\ref{lem:Composition} and Table~\ref{tab:Value} that \begin{align*} \widetilde{Z}^Y(\operatorname{MJ}_{\bullet\circ}^{-1}(\sigma)) &= \chi_{1^0}^{-1}\chi_{1^0,1^{0'}} \left( \varnothing +\frac{1}{2}\rTgraph{1^+}{1^-}{1^0} + \frac{1}{2}\rTgraph{1^+}{1^-}{1^{0'}} +(\operatorname{i-deg}>1) \right) \\ &= \varnothing + \rTgraph{1^+}{1^-}{1^0} + (\operatorname{i-deg}>1). \end{align*} Thus the previous theorem implies $\mu_2^{\mathcal{A}}(\sigma) = (-1)^2\rTgraph{2^\ast}{1^\ast}{3^\ast}$. Therefore, \[\mu_2(\sigma) = x_1\otimes[x_2,x_3] +x_2\otimes[x_3,x_1] +x_3\otimes[x_1,x_2] \in \varpi/\varpi_2 \otimes \varpi_2/\varpi_3.\] Moreover, by \cite[Remark in page~1160]{CHM08}, the Milnor $\overline{\mu}$-invariants of length 3 of the closure of $\sigma$, which is the Borromean rings, are as follows: \[\overline{\mu}_{\widehat{\sigma}}(j_1,j_2;i)= \begin{cases} \operatorname{sgn}(j_1\;j_2\;i) & \text{if $\{j_1,j_2,i\}=\{1,2,3\}$,} \\ 0 & \text{otherwise.} \end{cases}\] \end{example} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'${$'$}

https://arxiv.org/abs/1909.03852

Exploration of Balanced Metrics on Symmetric Positive Definite Matrices

Symmetric Positive Definite (SPD) matrices have been used in many fields of medical data analysis. Many Riemannian metrics have been defined on this manifold but the choice of the Riemannian structure lacks a set of principles that could lead one to choose properly the metric. This drives us to introduce the principle of balanced metrics that relate the affine-invariant metric with the Euclidean and inverse-Euclidean metric, or the Bogoliubov-Kubo-Mori metric with the Euclidean and log-Euclidean metrics. We introduce two new families of balanced metrics, the mixed-power-Euclidean and the mixed-power-affine metrics and we discuss the relation between this new principle of balanced metrics and the concept of dual connections in information geometry.

\section{Introduction} Symmetric Positive Definite (SPD) matrices are used in many applications: for example, they represent covariance matrices in signal or image processing \cite{Barachant13,Deligianni11,Cheng13} and they are diffusion tensors in diffusion tensor imaging \cite{lenglet_statistics_2006,pennec_riemannian_2006,Fletcher07}. Many Riemannian structures have been introduced on the manifold of SPD matrices depending on the problem and showing significantly different results from one another on statistical procedures such as the computation of barycenters or the principal component analysis. Non exhaustively, we can cite Euclidean metrics, power-Euclidean metrics \cite{Dryden10}, log-Euclidean metrics \cite{Arsigny06}, which are flat; affine-invariant metrics \cite{pennec_riemannian_2006,Fletcher07,Lenglet06} which are negatively curved; the Bogoliubov-Kubo-Mori metric \cite{Michor00} whose curvature has a quite complex expression. Are there some relations between them? This question has practical interests. First, understanding the links between these metrics could lead to interesting formulas and allow to perform more efficient algorithms. Second, finding families of metrics that comprise these isolated metrics could allow to perform optimization on the parameters of these families to find a better adapted metric. Some relations already exist. For example, the power-Euclidean metrics \cite{Dryden10} (resp. power-affine metrics \cite{Thanwerdas19}) comprise the Euclidean metric (resp. affine-invariant metric) and tend to the log-Euclidean metric when the power tends to zero. We propose the principle of \textit{balanced metrics} after observing two facts. The affine-invariant metric $g^A_\Sigma(X,Y) = \mathrm{tr}((\Sigma^{-1}X\Sigma^{-1})Y)$ on SPD matrices may be seen as a balanced hybridization of the Euclidean metric $g^E_\Sigma(X,Y)=\mathrm{tr}(XY)$ on one vector and of the inverse-Euclidean metric (the Euclidean metric on precision matrices) $g^I_\Sigma(X,Y)=\mathrm{tr}((\Sigma^{-1}X\Sigma^{-1})(\Sigma^{-1}Y\Sigma^{-1}))$ on the other vector. Moreover, the definition of the Bogoliubov-Kubo-Mori metric can be rewritten as $g^{BKM}_\Sigma(X,Y)=\mathrm{tr}(\partial_X\log(\Sigma)\,Y)$ where it appears as a balance of the Euclidean metric and the log-Euclidean metric $g^{LE}_\Sigma(X,Y)=\mathrm{tr}(\partial_X\log(\Sigma)\,\partial_Y\log(\Sigma))$. These observations raise a few questions. Given two metrics, is it possible to define a balanced bilinear form in general? If yes, is it clear that this bilinear form is symmetric and positive definite? If it is a metric, are the Levi-Civita connections of the two initial metrics dual in the sense of information geometry? In this work, we explore this principle through the affine-invariant metric, the Bogoliubov-Kubo-Mori metric and we define two new families of balanced metrics, the mixed-power-Euclidean and the mixed-power-affine metrics. In section 2, we show that if a balanced metric comes from two flat metrics, the three of them define a dually flat structure. In particular, we show that the balanced structure defined by the Euclidean and the inverse-Euclidean metrics corresponds to the dually flat structure given by the $\pm1$-connections of Fisher information geometry. In section 3, we enlighten the balanced structure of the BKM metric and we generalize it by defining the family of mixed-power-Euclidean metrics. In section 4, we define the family of mixed-power-affine metrics and we discuss the relation between the concepts of balanced metric and dual connections when the two initial metrics are not flat. \section{Affine-invariant metric as a balance of Euclidean and inverse-Euclidean metrics} Because the vocabulary may vary from one community to another, we shall first introduce properly the main geometric tools that we use in the article (Section 2.1). Then we examine in Section 2.2 the principle of balanced metric in the particular case of the pair Euclidean / inverse-Euclidean metrics and we formalize it in the general case of two flat metrics. In Section 2.3, we show that the $\pm 1$-connections of the centered multivariate normal model are exactly the Levi-Civita connections of the Euclidean and inverse-Euclidean metrics. \subsection{Reminder on metrics, connections and parallel transport} On a manifold $\mathcal{M}$, we denote $\mathcal{C}^\infty(\mathcal{M})$ the ring of smooth real functions and $\mathfrak{X}(\mathcal{M})$ the $\mathcal{C}^\infty(\mathcal{M})$-module of vector fields. \subsubsection{Connection} A \textit{connection} is an $\mathbb{R}$-bilinear map $\nabla:\mathfrak{X}(\mathcal{M})\times\mathfrak{X}(\mathcal{M})\longrightarrow\mathfrak{X}(\mathcal{M})$ that is $\mathcal{C}^\infty(\mathcal{M})$-linear in the first variable and satisfies the Leibniz rule in the second variable. It gives notions of parallelism, parallel transport and geodesics. A vector field $V$ is \textit{parallel} to the curve $\gamma$ if $\nabla_{\dot{\gamma}}V=0$. The \textit{parallel transport} of a vector $v$ along a curve $\gamma$ is the unique vector field $V_{\gamma(t)}=\Pi_\gamma^{0\rightarrow t}v$ that extends $v$ and that is parallel to $\gamma$. Thus, the connection is an infinitesimal parallel transport, that is $\Pi_\gamma^{t\rightarrow 0}V_{\gamma(t)}=V_{\gamma(0)}+t\nabla_{\dot{\gamma}}V+o(t)$. The \textit{geodesics} are autoparallel curves, that is curves $\gamma$ satisfying $\nabla_{\dot{\gamma}}\dot{\gamma}=0$. \subsubsection{Levi-Civita connection} Given a metric $g$ on a manifold $\mathcal{M}$, the \textit{Levi-Civita connection} is the unique torsion-free connection $\nabla^g$ compatible with the metric $g$, that is $\nabla^gg=0$ or more explicitly $X(g(Y,Z))=g(\nabla^g_XY,Z)+g(Y,\nabla^g_XZ)$ for all vector fields $X,Y,Z\in\mathfrak{X}(\mathcal{M})$. Thus a metric inherits notions of parallel transport and geodesics. Note that geodesics coincide with distance-minimizing curves with constant speed. \subsubsection{Dual connections} Given a metric $g$ and a connection $\nabla$, the \textit{dual connection} of $\nabla$ with respect to $g$ is the unique connection $\nabla^*$ satisfying the following equality $X(g(Y,Z))=g(\nabla_XY,Z)+g(Y,\nabla^*_XZ)$ for all vector fields $X,Y,Z\in\mathfrak{X}(\mathcal{M})$. It is characterized by Lemma \ref{dualconnections} below. In this sense, the Levi-Civita connection $\nabla^{g}$ is the unique torsion-free self-dual connection with respect to $g$. We say that $(\mathcal{M},g,\nabla,\nabla^*)$ is a \textit{dually-flat manifold} when $\nabla,\nabla^*$ are dual with respect to $g$ and $\nabla$ is flat (then $\nabla^*$ is automatically flat \cite{Amari00}). \begin{lemma}[Characterization of dual connections]\label{dualconnections} Two connections $\nabla,\nabla'$ with parallel transports $\Pi,\Pi'$ are dual with respect to a metric $g$ if and only if the dual parallel transport preserves the metric, i.e. for all vector fields $X,Y\in\mathfrak{X}(\mathcal{M})$ and all curve $\gamma$, $g_{\gamma(t)}(X_{\gamma(t)},Y_{\gamma(t)})=g_{\gamma(0)}(\Pi_\gamma^{t\rightarrow 0}X_{\gamma(t)},(\Pi')_\gamma^{t\rightarrow 0}Y_{\gamma(t)})$. \end{lemma} \begin{proof} The direct sense is proved in \cite{Amari00}. Let us assume that the dual parallel transport preserves the metric and let $X,Y,Z\in\mathfrak{X}(\mathcal{M})$ be vector fields. Let $x\in\mathcal{M}$ and let $\gamma$ be a curve such that $\gamma(0)=x$ and $\dot{\gamma}(0)=X_x$. Using the first order approximation of the parallel transport, our assumption leads to: \begin{align*} g_{\gamma(t)}(Y_{\gamma(t)},Z_{\gamma(t)}) &=g_x(\Pi_\gamma^{t\rightarrow 0}Y_{\gamma(t)},(\Pi')_\gamma^{t\rightarrow 0}Z_{\gamma(t)})\\ &=g_x(Y_x+t\nabla_{\dot{\gamma}}Y+o(t),Z_x+t\nabla'_{\dot{\gamma}}Z+o(t))\\ &=g_x(Y_x,Z_x)+t[g_x(\nabla_{\dot{\gamma}}Y,Z)+g_x(Y,\nabla'_{\dot{\gamma}}Z)]+o(t). \end{align*} So $X_x(g(Y,Z))=g_x(\nabla_{\dot{\gamma}}Y,Z)+g_x(Y,\nabla'_{\dot{\gamma}}Z)$ and $\nabla,\nabla'$ are dual w.r.t. $g$. $\Box$ \end{proof} \subsection{Principle of balanced metrics} \subsubsection{Observation} We denote $\mathcal{M}={SPD}_n$ the manifold of SPD matrices and $N=\dim\mathcal{M}=\frac{n(n+1)}{2}$. The (A)ffine-invariant metric $g^A$ on SPD matrices \cite{pennec_riemannian_2006,Fletcher07,Lenglet06}, i.e. satisfying $g^A_{M\Sigma M^\top}(MXM^\top,MYM^\top)=g^A_\Sigma(X,Y)$ for $M\in{GL}_n$, is defined by: \begin{equation}\label{A} g^A_\Sigma(X,Y) = \mathrm{tr}((\Sigma^{-1}X\Sigma^{-1})Y)=\mathrm{tr}(X(\Sigma^{-1}Y\Sigma^{-1})). \end{equation} The (E)uclidean metric $g^E$ on SPD matrices is the pullback metric by the embedding $id:\mathcal{M}\hookrightarrow(\mathrm{Sym}_n,\dotprod{\cdot}{\cdot}_\mathrm{Frob})$: \begin{equation}\label{E} g^E_\Sigma(X,Y) = \mathrm{tr}(XY). \end{equation} The (I)nverse-Euclidean metric $g^I$ on SPD matrices belongs to the family of power-Euclidean metrics \cite{Dryden10} with power $-1$. If SPD matrices are seen as covariance matrices $\Sigma$, the inverse-Euclidean metric is the Euclidean metric on precision matrices $\Sigma^{-1}$: \begin{equation}\label{I} g^I_\Sigma(X,Y) = \mathrm{tr}(\Sigma^{-2}X\Sigma^{-2}Y) = \mathrm{tr}((\Sigma^{-1}X\Sigma^{-1})(\Sigma^{-1}Y\Sigma^{-1})). \end{equation} Observing these definitions, the affine-invariant metric (\ref{A}) appears as a \textit{balance} of the Euclidean metric (\ref{E}) and the inverse-Euclidean metric (\ref{I}). We formalize this idea thanks to parallel transport. \subsubsection{Formalization} The diffeomorphism $\mathrm{inv}:(\mathcal{M},g^I)\longrightarrow(\mathcal{M},g^E)$ is an isometry. Since these two metrics are flat, the parallel transports do not depend on the curve. On the one hand, the Euclidean parallel transport from $\Sigma$ to $I_n$ is the identity map $\Pi^E:X\in T_\Sigma\mathcal{M}\longmapsto X\in T_{I_n}\mathcal{M}$ since all tangent spaces are identified to the vector space of symmetric matrices $\mathrm{Sym}_n$ by the differential of the embedding $id:\mathcal{M}\hookrightarrow\mathrm{Sym}_n$. On the other hand, the isometry $\mathrm{inv}$ gives the inverse-Euclidean parallel transport from $\Sigma$ to $I_n$, $\Pi^{I}:X\in T_\Sigma\mathcal{M}\longmapsto\Sigma^{-1}X\Sigma^{-1}\in T_{I_n}\mathcal{M}$. We generalize this situation in Definition \ref{balanced}. Given Lemma \ref{dualconnections}, it automatically leads to Theorem \ref{dual}. \begin{definition}[Balanced bilinear form]\label{balanced} Let $g,g^*$ be two flat metrics on ${SPD}_n$ and $\Pi,\Pi^*$ the associated parallel transports that do not depend on the curve. We define the balanced bilinear form $g^0_\Sigma(X,Y)=\mathrm{tr}((\Pi_{\Sigma\rightarrow I_n}X)(\Pi^*_{\Sigma\rightarrow I_n}Y))$. \end{definition} \begin{theorem}[A balanced metric defines a dually flat manifold]\label{dual} Let $g,g^*$ be two flat metrics and let $\nabla,\nabla^*$ be their Levi-Civita connections. If the balanced bilinear form $g^0$ of $g,g^*$ is a metric, then $(\mathcal{M},g^0,\nabla,\nabla^*)$ is a dually flat manifold. \end{theorem} If two connections $\nabla$ and $\nabla^*$ are dual connections with respect to a metric $g^0$, there is no reason for them to be Levi-Civita connections of some metrics. Therefore, the main advantage of the principle of balanced metrics on the concept of dual connections seems to be the metric nature of the dual connections. \begin{corollary}[Euclidean and inverse-Euclidean are dual with respect to affine-invariant] We denote $\nabla^E$ and $\nabla^I$ the Levi-Civita connections of the Euclidean metric $g^E$ and the inverse-Euclidean metric $g^I$. Then $g^A$ is the balanced metric of $g^I,g^E$ and $({SPD}_n,g^A,\nabla^I,\nabla^E)$ is a dually flat manifold. \end{corollary} \subsection{Relation with Fisher information geometry} We know from \cite{Skovgaard84} that the affine-invariant metric is the Fisher metric of the centered multivariate normal model. Information geometry provides a natural one-parameter family of dual connections, called $\alpha$-connections \cite{Amari00}. In the following table, we recall the main quantities characterizing the \textit{centered multivariate normal model} $\mathcal{P}=\{p_\Sigma:\mathbb{R}^n\longrightarrow\mathbb{R}_+^*,\Sigma\in\mathcal{M}\}$, where $\mathcal{M}={SPD}_n$. $$\begin{array}{cc} \mathrm{Densities} & ~~p_\Sigma(x)=\frac{1}{{\sqrt{2\pi}}^n}\frac{1}{\sqrt{\det\Sigma}}\exp\left(\frac{1}{2}x^\top\Sigma^{-1}x\right) \\[1.5mm] \mathrm{Log~likelihood} & ~~l_\Sigma(x)=\log p_\Sigma(x)=\frac{1}{2}\left(-n\log(2\pi)-\log\det\Sigma+x^\top\Sigma^{-1}x\right)\\[1.5mm] \mathrm{Differential} & ~~d_\Sigma l(V)(x)=-\frac{1}{2}\left[\mathrm{tr}(\Sigma^{-1}V)+x^\top\Sigma^{-1}V\Sigma^{-1}x\right]\\[1.5mm] \mathrm{Fisher~metric} & ~~g_\Sigma(V,W)=\frac{1}{2}\mathrm{tr}(\Sigma^{-1}V\Sigma^{-1}W) \end{array}$$ We recall that the \textit{$\alpha$-connections} $\nabla^{(\alpha)}$ \cite{Amari00} of a family of densities $\mathcal{P}$ are defined by their Christoffel symbols $\Gamma_{ijk}=g_{lk}\Gamma_{ij}^l$ in the local basis $(\partial_i)_{1\leqslant i\leqslant N}$ at $\Sigma\in\mathcal{M}$: \begin{equation} \left(\Gamma_{ijk}^{(\alpha)}\right)_\Sigma=\mathbb{E}_\Sigma\left[\left(\partial_i\partial_j l+\frac{1-\alpha}{2}\partial_i l\partial_j l\right)\partial_k l\right]. \end{equation} We give in Theorem \ref{alpha} the expression of the $\alpha$-connections of the centered multivariate normal model and we notice that the Euclidean and inverse-Euclidean Levi-Civita connections belong to this family. \begin{theorem}[$\alpha$-connections of the centered multivariate normal model]\label{alpha} In the global basis of $\mathcal{M}={SPD}_n$ given by the inclusion $\mathcal{M}\hookrightarrow\mathrm{Sym}(n)\simeq\mathbb{R}^N$, writing $\partial_XY=X^i(\partial_i Y^j)\partial_j$, the $\alpha$-connections of the multivariate centered normal model are given by the following formula: \begin{equation}\label{alphaconnection} \nabla^{(\alpha)}_XY=\partial_XY-\frac{1+\alpha}{2}(X\Sigma^{-1}Y+Y\Sigma^{-1}X). \end{equation} The mixture $m$-connection ($\alpha=-1$) is the Levi-Civita connection of the Euclidean metric $g^E_\Sigma(X,Y)=\mathrm{tr}(XY)$. The exponential $e$-connection ($\alpha=1$) is the Levi-Civita connection of the inverse-Euclidean metric, i.e. the pullback of the Euclidean metric by matrix inversion, $g^I_\Sigma(X,Y)=\mathrm{tr}(\Sigma^{-2}X\Sigma^{-2}Y)$. \end{theorem} The formula (\ref{alphaconnection}) can be proved thanks to Lemma \ref{expectations} which gives the results of expressions of type $\int_{\mathbb{R}^n}{x^\top\Sigma^{-1}X\Sigma^{-1}Y\Sigma^{-1}Z\Sigma^{-1}x\exp\left(-\frac{1}{2}x^\top\Sigma^{-1}x\right)dx}$, with $y=\Sigma^{-1/2}x$, $A=\Sigma^{-1/2}X\Sigma^{-1/2}$, $B=\Sigma^{-1/2}Y\Sigma^{-1/2}$ and $C=\Sigma^{-1/2}Z\Sigma^{-1/2}$. If one wants to avoid using the third formula of Lemma \ref{expectations}, one can rely on the formula (\ref{alphaconnection}) in the case $\alpha=0$ which is already known from \cite{Skovgaard84}. \begin{lemma}\label{expectations} For $A,B,C\in\mathrm{Sym}_n$: \begin{align*} \mathbb{E}_{I_n}(y\longmapsto y^\top Ay) &=\mathrm{tr}(A),\\ \mathbb{E}_{I_n}(y\longmapsto y^\top Ayy^\top By) &=\mathrm{tr}(A)\mathrm{tr}(B)+2\mathrm{tr}(AB),\\ \mathbb{E}_{I_n}(y\longmapsto y^\top Ayy^\top Byy^\top Cy) &=\mathrm{tr}(A)\mathrm{tr}(B)\mathrm{tr}(C)+8\mathrm{tr}(ABC)\\ & \quad +2(\mathrm{tr}(AB)\mathrm{tr}(C)+\mathrm{tr}(BC)\mathrm{tr}(A)+\mathrm{tr}(CA)\mathrm{tr}(B)). \end{align*} \end{lemma} \begin{proof}[Theorem \ref{alpha}] Given Lemma \ref{expectations}, the computation of the Christoffel symbols $\left(\Gamma^{(\alpha)}_{ijk}\right)_\Sigma$ leads to $\left(\Gamma^{(\alpha)}_{ijk}\right)_\Sigma X^iY^jZ^k=-\frac{1+\alpha}{4}\mathrm{tr}(\Sigma^{-1}[X\Sigma^{-1}Y+Y\Sigma^{-1}X]\Sigma^{-1}Z)$. On the other hand, the relation $\Gamma_{ijk}=g_{lk}\Gamma_{ij}^l$ between Christoffel symbols gives $\left(\Gamma^{(\alpha)}_{ijk}\right)_\Sigma X^iY^jZ^k=\frac{1}{2}\mathrm{tr}\left(\Sigma^{-1}\left[\left(\Gamma_{ij}^l\right)_\Sigma^{(\alpha)} X^iY^j\partial_l\right]\Sigma^{-1}Z\right)$. So we get: \begin{equation} \nabla^{(\alpha)}_XY=\partial_XY+\left[\left(\Gamma_{ij}^l\right)_\Sigma^{(\alpha)} X^iY^j\partial_l\right]=\partial_XY-\frac{1+\alpha}{2}(X\Sigma^{-1}Y+Y\Sigma^{-1}X). \end{equation} It is clear that the mixture connection ($\alpha=-1$) is the Euclidean connection. The inverse-Euclidean connection can be computed thanks to the Koszul formula. This calculus drives exactly to the exponential connection ($\alpha=1$). $\Box$ \end{proof} In the next section, we apply the principle of balanced metrics to the pairs Euclidean / log-Euclidean (Bogoliubov-Kubo-Mori metric) and power-Euclidean / power-Euclidean (mixed-power-Euclidean metrics). \section{The family of mixed-power-Euclidean metrics} \subsection{Bogoliubov-Kubo-Mori metric} The Bogoliubov-Kubo-Mori metric $g^{BKM}$ is a metric on symmetric positive definite matrices used in quantum physics. It was introduced as $g_\Sigma^{BKM}(X,Y)=\int_0^\infty{\mathrm{tr}((\Sigma+tI_n)^{-1}X(\Sigma+tI_n)^{-1}Y)dt}$ and can be rewritten \cite{Michor00} thanks to the differential of the symmetric matrix logarithm $\log:\mathcal{M}={SPD}_n\longrightarrow\mathrm{Sym}_n$ as: \begin{equation}\label{BKM} g_\Sigma^{BKM}(X,Y)=\mathrm{tr}(\partial_X\log(\Sigma)\,Y)=\mathrm{tr}(X\,\partial_Y\log(\Sigma)). \end{equation} The log-Euclidean metric $g^{LE}$ \cite{Arsigny06} is the pullback metric of the Euclidean metric by the symmetric matrix logarithm $\log:(\mathcal{M},g^{LE})\longrightarrow(\mathrm{Sym}_n,g^E)$: \begin{equation}\label{LE} g^{LE}_\Sigma(X,Y)=\mathrm{tr}(\partial_X\log(\Sigma)\,\partial_Y\log(\Sigma)). \end{equation} Therefore, the BKM metric (\ref{BKM}) appears as the balanced metric of the Euclidean metric (\ref{E}) and the log-Euclidean metric (\ref{LE}). As the Euclidean and log-Euclidean metrics are flat, the parallel transport does not depend on the curve and Theorem \ref{dual} ensures that they form a dually flat manifold. \begin{corollary}[Euclidean and log-Euclidean are dual with respect to BKM]\label{dualBKM} We denote $\nabla^{E}$ and $\nabla^{LE}$ the Levi-Civita connections of the Euclidean metric $g^E$ and the log-Euclidean metric $g^{LE}$. Then $g^{BKM}$ is the balanced metric of $g^{LE},g^E$ and $({SPD}_n,g^{BKM},\nabla^{LE},\nabla^E)$ is a dually flat manifold. \end{corollary} \subsection{Mixed-power-Euclidean} Up to now, we observed that existing metrics (affine-invariant and BKM) were the balanced metrics of pairs of flat metrics (Euclidean / inverse-Euclidean and Euclidean / log-Euclidean). Thus, the symmetry and the positivity of the balanced bilinear forms were obvious. From now on, we build new bilinear forms thanks to the principle of balanced metrics. Therefore, it is not as obvious as before that these bilinear forms are metrics. The family of power-Euclidean metrics $g^{E,\theta}$ \cite{Dryden10} indexed by the power $\theta\ne 0$ is defined by pullback of the Euclidean metric by the power function $\mathrm{pow}_\theta=\exp\circ\,\theta\log:(\mathcal{M},\theta^2g^{E,\theta})\longrightarrow(\mathcal{M},g^E)$: \begin{equation} g^{E,\theta}_\Sigma(X,Y)=\frac{1}{\theta^2}\mathrm{tr}(\partial_X\mathrm{pow}_\theta(\Sigma)\,\partial_Y\mathrm{pow}_\theta(\Sigma)). \end{equation} This family comprise the Euclidean metric for $\theta=1$ and tends to the log-Euclidean metric when the power $\theta$ goes to 0. Therefore, we abusively consider that the log-Euclidean metric belongs to the family and we denote it $g^{E,0}:=g^{LE}$. We define the mixed-power-Euclidean metrics $g^{E,\theta_1,\theta_2}$ as the balanced bilinear form of the power-Euclidean metrics $g^{E,\theta_1}$ and $g^{E,\theta_2}$, where $\theta_1,\theta_2\in\mathbb{R}$: \begin{equation} g^{E,\theta_1,\theta_2}_\Sigma(X,Y)=\frac{1}{\theta_1\theta_2}\mathrm{tr}(\partial_X\mathrm{pow}_{\theta_1}(\Sigma)\,\partial_Y\mathrm{pow}_{\theta_2}(\Sigma)). \end{equation} Note that the family of mixed-power-Euclidean metrics contains the BKM metric for $(\theta_1,\theta_2)=(1,0)$ and the $\theta$-power-Euclidean metric for $(\theta_1,\theta_2)=(\theta,\theta)$, including the Euclidean metric for $\theta=1$ and the log-Euclidean metric for $\theta=0$. At this stage, we do not know that the bilinear form $g^{E,\theta_1,\theta_2}$ is a metric. This is stated by Theorem \ref{MPE}. As the power-Euclidean metrics are flat, Theorem \ref{dual} combined with Theorem \ref{MPE} ensure that the Levi-Civita connections $\nabla^{E,\theta_1}$ and $\nabla^{E,\theta_2}$ of the metrics $g^{E,\theta_1}$ and $g^{E,\theta_2}$ are dual with respect to the $(\theta_1,\theta_2)$-mixed-power-Euclidean metric. This is stated by Corollary \ref{dualMPE}. \begin{theorem}\label{MPE} The bilinear form $g^{E,\theta_1,\theta_2}$ is symmetric and positive definite so it is a metric on ${SPD}_n$. Moreover, the symmetry ensures that $g^{E,\theta_1,\theta_2}=g^{E,\theta_2,\theta_1}$. \end{theorem} \begin{corollary}[$\theta_1$ and $\theta_2$-power-Euclidean are dual with respect to $(\theta_1,\theta_2)$-mixed-power-Euclidean]\label{dualMPE} For $\theta_1,\theta_2\in\mathbb{R}$, we denote $\nabla^{E,\theta_1}$ and $\nabla^{E,\theta_2}$ the Levi-Civita connections of the power-Euclidean metrics $g^{E,\theta_1}$ and $g^{E,\theta_2}$. Then $g^{E,\theta_1,\theta_2}$ is the balanced metric of $g^{E,\theta_1},g^{E,\theta_2}$ and $({SPD}_n,g^{E,\theta_1,\theta_2},\nabla^{E,\theta_1},\nabla^{E,\theta_2})$ is a dually flat manifold. \end{corollary} To prove Theorem \ref{MPE}, we show that for all spectral decomposition $\Sigma=PDP^\top$ of an SPD matrix, there exists a matrix $A$ with positive coefficients $A(i,j)>0$ such that $g_\Sigma^{E,\theta_1,\theta_2}(X,Y)=\mathrm{tr}((A\bullet P^\top XP)(A\bullet P^\top YP))$, where $\bullet$ is the Hadamard product, i.e. $(A\bullet B)(i,j)=A(i,j)B(i,j)$, which is associative, commutative, distributive w.r.t. matrix addition and satisfies $\mathrm{tr}((A\bullet B)C)=\mathrm{tr}(B(A\bullet C))$ for symmetric matrices $A,B,C\in\mathrm{Sym}_n$. The existence of $A$ relies on Lemma \ref{diff}. \begin{lemma}\label{diff} Let $\Sigma=PDP^\top$ be a spectral decomposition of $\Sigma\in\mathcal{M}$, with $P\in O(n)$ and $D$ diagonal. For $f\in\{\exp,\log,\mathrm{pow}_\theta\}$, $\partial_Vf(\Sigma)=P(\delta(f,D)\bullet P^\top VP)P^\top$ where $\delta(f,D)(i,j)=\frac{f(d_i)-f(d_j)}{d_i-d_j}$. Note that $\frac{1}{\theta}\delta(\mathrm{pow}_\theta,D)(i,j)>0$ for all $\theta\in\mathbb{R}^*$ and $\delta(\log,D)(i,j)>0$. \end{lemma} \begin{proof}[Lemma \ref{diff}] Once shown for $f=\exp$, it is easy to get for $f=\log$ by inversion and for $f=\mathrm{pow}_\theta=\exp\circ\,\theta\log$ by composition. But the case $f=\exp$ itself reduces to the case $f=\mathrm{pow}_k$ with $k\in\mathbb{N}$ by linearity, so we focus on this last case. As $\partial_V\mathrm{pow}_k(\Sigma)=\sum_{l=0}^{k-1}{\Sigma^lV\Sigma^{k-1-l}}=P\partial_{P^\top VP}\mathrm{pow}_k(D)P^\top$ and $\partial_{P^\top VP}\mathrm{pow}_k(D)(i,j)=\sum_{l=0}^{k-1}{D^lP^\top VPD^{k-1-l}(i,j)}=\frac{d_i^k-d_j^k}{d_i-d_j}P^\top VP(i,j)$, we get $\partial_V\mathrm{pow}_k(\Sigma)=P(\delta(\mathrm{pow}_k,D)\bullet P^\top VP)P^\top$. \end{proof} \begin{proof}[Theorem \ref{MPE}] Let $\theta_1,\theta_2\in\mathbb{R}^*$. For a spectral decomposition $\Sigma=PDP^\top$, the matrix $A$ defined by $A(i,j)=\sqrt{\frac{1}{\theta_1}\delta(\mathrm{pow}_{\theta_1},D)(i,j)\frac{1}{\theta_2}\delta(\mathrm{pow}_{\theta_2},D)(i,j)}>0$ satisfies $g_\Sigma^{E,\theta_1,\theta_2}(X,Y)=\mathrm{tr}((A\bullet P^\top XP)(A\bullet P^\top YP))$. Symmetry and non-negativity are clear since they come from the Frobenius scalar product. Finally, if $g_\Sigma^{E,\theta_1,\theta_2}(X,X)=0$, then $A\bullet P^\top XP=0$ so $X=0$. So $g^{E,\theta_1,\theta_2}$ is a metric. If $\theta_1=0$, the matrix $A$ defined by $A(i,j)=\sqrt{\delta(\log,D)(i,j)\frac{1}{\theta_2}\delta(\mathrm{pow}_{\theta_2},D)(i,j)}>0$ satisfies the same property and $g^{E,0,\theta_2}$ is a metric. $\Box$ \end{proof} \section{The family of mixed-power-affine metrics} In previous sections, we defined our balanced metric from a pair of two flat metrics and we showed that it corresponded to the duality of (Levi-Civita) connections in information geometry. In this section, we investigate the balanced metric of two non-flat metrics and we observe that the corresponding Levi-Civita connections cannot be dual with respect to this balanced metric. The family of power-affine metrics $g^{A,\theta}$ \cite{Thanwerdas19} indexed by the power $\theta\ne 0$ are defined by pullback of the affine-invariant metric by the power function $\mathrm{pow}_\theta:(\mathcal{M},\theta^2g^{A,\theta})\longrightarrow(\mathcal{M},g^A)$: \begin{align} g^{A,\theta}_\Sigma(X,Y) &=\frac{1}{\theta^2}\mathrm{tr}(\Sigma^{-\theta}\,\partial_X\mathrm{pow}_\theta(\Sigma)\,\Sigma^{-\theta}\,\partial_Y\mathrm{pow}_\theta(\Sigma)) \end{align} This family comprise the affine-invariant metric for $\theta=1$ and tends to the log-Euclidean metric when the power $\theta$ goes to 0. We consider that the log-Euclidean metric belongs to the family and we denote $g^{A,0}:=g^{LE}$. As these metrics have no cut locus because they endow the manifold with a negatively curved Riemannian symmetric structure, there exists a unique geodesic between two given points. Therefore, a canonical parallel transport can be defined along geodesics. This allows to define the balanced bilinear form of two metrics without cut locus. \begin{definition}[Balanced bilinear form]\label{balanced2} Let $g,g^*$ be two metrics without cut locus on ${SPD}_n$ and $\Pi,\Pi^*$ the associated geodesic parallel transports. We define the balanced bilinear form $g^0_\Sigma(X,Y)=\mathrm{tr}((\Pi_{\Sigma\rightarrow I_n}X)(\Pi^*_{\Sigma\rightarrow I_n}Y))$. \end{definition} Given that the geodesic parallel transport on the manifold $(\mathcal{M},g^{A,\theta})$ is $\Pi_{\Sigma\rightarrow I_n}:X\in T_\Sigma\mathcal{M}\longmapsto\frac{1}{\theta}\Sigma^{-\theta/2}\partial_X\mathrm{pow}_\theta(\Sigma)\Sigma^{-\theta/2}\in T_{I_n}\mathcal{M}$, we define the mixed-power-affine metrics $g^{A,\theta_1,\theta_2}$ as the balanced metric of the power-affine metrics $g^{A,\theta_1}$ and $g^{A,\theta_2}$, where $\theta_1,\theta_2\in\mathbb{R}$ and $\theta=(\theta_1+\theta_2)/2$: \begin{equation} g^{A,\theta_1,\theta_2}_\Sigma(X,Y)=\frac{1}{\theta_1\theta_2}\mathrm{tr}(\Sigma^{-\theta}\,\partial_X\mathrm{pow}_{\theta_1}(\Sigma)\,\Sigma^{-\theta}\,\partial_Y\mathrm{pow}_{\theta_2}(\Sigma)). \end{equation} Note that the family of mixed-power-affine metrics contains the $\theta$-power-affine metric for $(\theta_1,\theta_2)=(\theta,\theta)$, including the affine-invariant metric for $\theta=1$ and the log-Euclidean metric for $\theta=0$. This family has two symmetries since $g^{A,\theta_1,\theta_2}=g^{A,\pm\theta_1,\pm\theta_2}$, they come from the inverse-consistency of the affine-invariant metric. This family has a non-empty intersection with the family of mixed-power-Euclidean metrics since $g^{A,\theta_1,-\theta_1}=g^{E,\theta_1,-\theta_1}=g^{A,\theta_1}$ for all $\theta_1\in\mathbb{R}$. The fact that $g^{A,\theta_1,\theta_2}$ is a metric can be shown exactly the same way as in the mixed-power-Euclidean case thanks to the equality $\Sigma^{-\theta/2}\partial_V\mathrm{pow}_\theta(\Sigma)\Sigma^{-\theta/2}=P(\varepsilon(\mathrm{pow}_\theta,D)\bullet P^\top VP)P^\top$ where $\varepsilon(\mathrm{pow}_\theta,D)=(d_id_j)^{-\theta/2}\delta(\mathrm{pow}_\theta,D)$ and where $\delta(\mathrm{pow}_\theta,D)$ has been defined in Lemma \ref{diff}. This is stated in Theorem \ref{MPA}. \begin{theorem}\label{MPA} The bilinear form $g^{A,\theta_1,\theta_2}$ is symmetric and positive definite. Hence it is a metric on ${SPD}_n$ and $g^{A,\theta_1,\theta_2}=g^{A,\theta_2,\theta_1}$. \end{theorem} Power-affine metrics being non-flat, $(\mathcal{M},g^{A,\theta_1,\theta_2},\nabla^{A,\theta_1},\nabla^{A,\theta_2})$, where $\nabla^{A,\theta_1}$ and $\nabla^{A,\theta_2}$ are Levi-Civita connections of $g^{A,\theta_1}$ and $g^{A,\theta_2}$, cannot be a dually-flat manifold. Actually, the two connections are even not dual. It can be understood by comparison with previous sections since the duality was a consequence of the independence of the parallel transport with respect to the chosen curve, which was a consequence of the flatness of the two connections. Moreover, in the Definition \ref{balanced2}, the vectors are parallel transported along two different curves (the geodesics relative to each connection) so it may exists a better definition for the balanced bilinear form of two metrics without cut locus or even of two general metrics. \section{Conclusion} The principle of balanced bilinear form is a procedure on SPD matrices that takes a pair of flat metrics or metrics without cut locus and builds a new metric based on the parallel transport of the initial metrics. When the two initial metrics are flat, we showed that the two Levi-Civita connections are dual with respect to the balanced metric. When the two initial metrics are not flat, the two Levi-Civita connections seem not to be dual, so the principle of balanced metrics does not reduce to the concept of dual Levi-Civita connections. A challenging objective for future works is to define properly this principle for other general pairs of metrics and to find conditions under which the balanced bilinear form is a metric. \paragraph{Acknowledgements.} This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 786854). This work has been supported by the French government, through the UCAJEDI Investments in the Future project managed by the National Research Agency (ANR) with the reference number ANR-15-IDEX-01. \bibliographystyle{unsrt}

https://arxiv.org/abs/2004.12325

Nonlinear variants of a theorem of Kwapień

A famous result of S. Kwapień asserts that a linear operator from a Banach space to a Hilbert space is absolutely $1$-summing whenever its adjoint is absolutely $q$-summing for some $1\leq q<\infty$; this result was recently extended to Lipschitz operators by Chen and Zheng. In the present paper we show that Kwapień's and Chen--Zheng theorems hold in a very relaxed nonlinear environment, under weaker hypotheses. Even when restricted to the original linear case, our result generalizes Kwapień's theorem because it holds when the adjoint is just almost summing. In addition, a variant for $\mathcal{L}_{p}$-spaces, with $p\geq2$, instead of Hilbert spaces is provided.

\section{Introduction} The theory of absolutely summing linear operators was originated from Grothendieck's Resum\'{e} \cite{gro} and since the publication of the papers of Lindenstrauss, Pe{\l }czy{\'{n}}ski and Pietsch \cite{LP, Pie}, it has performed a fundamental role in Banach Space Theory and its applications. In the last decades the path from the linear to the nonlinear theory of absolutely summing operators was encouraged by seminal works of Pietsch \cite{PPPP}, Farmer and Johnson \cite{FJ}, among others. Nowadays, nonlinear variants of absolutely summing operators have been explored in different settings with applications in different fields of pure and applied mathematics (see, for instance, \cite{bayart, perez, vieira} and the references therein). Let $X,Y$ be Banach spaces over $\mathbb{K}=\mathbb{R}$ or $\mathbb{C}$ and, from now on, $B_{X}$ denotes the closed unit ball of $X$, and $X^{\ast}$ represents its topological dual. If $1\leq p<\infty,$ a linear operator $T:X\longrightarrow Y$ is \textit{absolutely} $p$-\textit{summing} if there exists a constant $C$ such tha \begin{equation} \left( {\displaystyle\sum\limits_{i=1}^{n}} \left\Vert T(x_{i})\right\Vert ^{p}\right) ^{1/p}\leq C\Vert(x_{i})_{i=1 ^{n}\Vert_{p,w},\label{s \end{equation} for every $x_{1},...,x_{n}\in X$ and all positive integers $n$, where \[ \Vert(x_{i})_{i=1}^{n}\Vert_{p,w}:=\sup_{\varphi\in B_{X^{\ast}}}\left( {\displaystyle\sum\limits_{i=1}^{n}} |\varphi(x_{i})|^{p}\right) ^{1/p}. \] The class of absolutely $p$-summing linear operators from $X$ to $Y$ will be represented, as it is usual, by $\Pi_{p}\left( X,Y\right) $ and the infimum of all $C$ satisfying (\ref{s}) defines a norm on $\Pi_{p}\left( X,Y\right) $, denoted by $\pi_{p}(T)$. The related notion of almost summing operators will be important for our purposes. According to \cite{BBJ}, a linear operator $T:X\longrightarrow Y$ is \textit{almost} $p$-\textit{summing} if there is a constant $C$ such that \begin{equation} \left( {\int\nolimits_{0}^{1}}\left\Vert \sum_{i=1}^{n}r_{i}(t)T(x_{i )\right\Vert ^{2}dt\right) ^{\frac{1}{2}}\leq C\left\Vert (x_{i})_{i=1 ^{n}\right\Vert _{p,w},\label{ppk \end{equation} for every positive integer $n$ and $x_{1},...,x_{n}\in X$. Above, $r_{i}$ are the Rademacher functions, defined as \ \begin{array} [c]{cccc r_{i}: & \left[ 0,1\right] & \longrightarrow & \mathbb{R}\\ & t & \longmapsto & \text{sign}\left( \sin2^{i}\pi t\right) . \end{array} \] The infimum of the constants $C$ satisfying (\ref{ppk}) is denoted by $\pi_{al.s.p}(T)$. The class of all almost $p$-summing operators is denoted by $\Pi_{al.s.p}$. When $p=2$, these operators are simply called \textit{almost summing} and we write $\Pi_{al.s}$ instead of $\Pi_{al.s.2}$ (see \cite[Chapter 12]{djt}). By \cite[Proposition 12.5]{djt}, \begin{equation} \bigcup_{1\leq q<\infty}\Pi_{q}(X,Y)\subseteq\Pi_{al.s}(X,Y).\label{inccc \end{equation} For details, we refer to the classical book by Diestel, Jarchow and Tonge \cite{djt}. From now on $\mathcal{I}$ denotes an operator ideal, $\mathcal{L}$ denotes the ideal of all bounded linear operators and $u^{\ast}$ represents the adjoint of a linear operator $u$. We denote, as usual, by $\mathcal{I}^{dual}$ the \textit{dual ideal} of $\mathcal{I}$ (see \cite[page 186]{djt}), that is, \[ \mathcal{I}^{dual}(X,Y):=\{u\in\mathcal{L}(X,Y);~u^{\ast}\in\mathcal{I (Y^{\ast},X^{\ast})\}. \] In general, the adjoint of an absolutely $p$-summing operator may fail to be absolutely $p$-summing and \textit{vice-versa}. For instance, considering the formal identity $i:\ell_{2}\longrightarrow c_{0}$, it follows that $i^{\ast}$ is absolutely $p$-summing for all $1\leq p<\infty,$ whereas $i$ and $i^{\ast\ast}$ aren't absolutely $p$-summing. Then, the following problem arises naturally: \begin{problem} For which Banach spaces $X$ and $Y$ and $1\leq p,q<\infty$ does the following inclusion hold \begin{align*} \Pi_{q}^{dual}(X,Y) \subseteq\Pi_{p}(X,Y)? \end{align*} \end{problem} A remarkable result due to Kwapie\'{n} \cite{K} shows that a linear operator $T$ from a Banach space $X$ to a Hilbert space $H$ is absolutely $1$-summing whenever $T^{\ast}$ is absolutely $q$-summing for some $1\leq q<\infty$. In other words, \begin{theorem} $($see \cite[Theorem 2.21]{djt}$)$\label{kuap} If $X$ is a Banach space and $H$ is a Hilbert space, then \begin{align*} \bigcup_{1\leq q<\infty}\Pi_{q}^{dual}(X,H)\subseteq\Pi_{1}(X,H). \end{align*} \end{theorem} Kwapie\'{n}'s Theorem was recently extended to Lipschitz operators by Chen and Zheng \cite[Theorem 3.1]{CZ} as we shall see next. Let $X$ be a pointed metric space with a base point denoted by $0$, and let $Y$ be a Banach space. The Lipschitz space $Lip_{0}(X;Y)$ is the Banach space of all Lipschitz operators $T:X\longrightarrow Y$ such that $T(0)=0$, under the Lipschitz norm defined by \[ Lip(T)=\sup\left\{ \frac{\Vert T(x)-T(y)\Vert}{d(x,y)}~:~x,y\in X,~x\neq y\right\} . \] When $Y=\mathbb{K}$, we write $X^{\#}=Lip_{0}(X;\mathbb{K})$. A cornerstone of the nonlinear theory of absolutely summing operators is the paper of Farmer and Johnson \cite{FJ}, which introduces the notion of absolutely summing operators to the Lipschitz framework as follows: a Lipschitz operator $T\colon X\rightarrow Y$ is \textit{Lipschitz} $p$-\textit{summing i}f there is a constant $C$ such that \begin{equation} \left( \sum_{i=1}^{n}\Vert T(x_{i})-T(q_{i})\Vert^{p}\right) ^{\frac{1}{p }\leq C\sup_{\varphi\in B_{X^{\#}}}\left( \sum_{i=1}^{n}|\varphi (x_{i})-\varphi(q_{i})|^{p}\right) ^{\frac{1}{p}} , \label{765765 \end{equation} for all $x_{1},\ldots,x_{n},q_{1},\ldots,q_{n}\in X$ and all positive integers $n$ (for related papers on nonlinear summing operators we refer to \cite{JA, JAproc, Fer, MM} and the references therein). According to \cite{Sawashima}, the Lipschitz adjoint $T^{\#}$ of $T\in Lip_{0}(X;Y)$ is defined as the continuous linear operator $T^{\#}\colon Y^{\#}\longrightarrow X^{\#}$ given by $T^{\#}(g):=g\circ T$. The variant of Kwapie\'{n}'s Theorem to Lipschitz operators due to Chen and Zheng reads as follows: \begin{theorem} \label{py} $($see \cite[Theorem 3.1]{CZ}$)$ Let $X$ be a pointed metric space and $H$ be a Hilbert space. If $T\in Lip_{0}(X;H)$ is such that $T^{\# |_{H^{\ast}}$ is absolutely $q$-summing for some $1\leq q<\infty$, then $T$ is Lipschitz $1$-summing. \end{theorem} Following this vein, other questions seem natural: \begin{problem} \label{probl3} Are there other nonlinear extensions/generalizations of the Kwapie\'{n} Theorem? \end{problem} \begin{problem} \label{probl4} Are there variants of the Kwapie\'{n} Theorem in which the range is not necessarily a Hilbert space? \end{problem} In this paper we shall answer these problems. To state our main result we shall recall the notion of $\mathcal{L}_{p}$-spaces. If $\lambda>1$ and $1\leq p\leq\infty$, a Banach space $Y$ is an $\mathcal{L}_{p,\lambda} \textit{-space} if every finite dimensional subspace $E$ of $Y$ is contained in a finite dimensional subspace $Y_{0}$ of $Y$ for which there is an isomorphism $v:Y_{0}\longrightarrow\ell_{p}^{\dim(Y_{0})}$ such that $\Vert v\Vert\Vert v^{-1}\Vert\leq\lambda$. When $Y$ is an $\mathcal{L}_{p,\lambda $\textit{-space for a certain }$\lambda$, we just say that $Y$ is an $\mathcal{L}_{p}$-space$.$ It is simple to observe that Hilbert spaces are $\mathcal{L}_{2}$-spaces. As a consequence of the main results of the present paper we conclude that the original result of Kwapie\'{n} can be improved as follows: \begin{theorem} \label{nn11}If $X$ is a Banach space and $Y$ is an $\mathcal{L}_{p}$-space and $2\leq p<\infty$, the \[ \Pi_{al.s.p}^{dual}(X,Y)\subseteq\Pi_{1}(X,Y). \] In particular, when $Y$ is a Hilbert spac \begin{equation} \Pi_{al.s}^{dual}(X,Y)\subseteq\Pi_{1}(X,Y).\label{2222 \end{equation} \end{theorem} Note that from (\ref{inccc}) it is obvious that (\ref{2222}) recovers the statement of Kwapie\'{n}'s Theorem. The paper is organized as follows. In Section 2 we state our main result (Theorem \ref{q1}); in Section 3 we prove Theorem \ref{q1} and in Section 4 we provide applications of the main result; for instance, we shall generalize the Chen--Zheng Theorem (Theorem \ref{py}) following the lines of what is done in Theorem \ref{nn11}. \section{Main result} We start off by recalling that the sequence of Rademacher functions $\left( r_{i}\right) _{i=1}^{\infty}$ is orthonormal; thus \begin{equation} \int_{0}^{1}\left\vert \sum_{i=1}^{n}r_{i}(t)a_{i}\right\vert ^{2 dt=\sum_{i=1}^{n}|a_{i}|^{2} \label{7654 \end{equation} for all $(a_{i})_{i=1}^{n} \in\ell^{n}_{2}$ and all positive integers $n.$ The well-known Kahane Inequality shows that the spaces generated by the Rademacher functions have equivalent $L_{p}$ norms: \begin{theorem} [Kahane inequality]Let $0<p,q<\infty$. Then there is a constant $\mathrm{K _{p,q}>0$ for which \[ \left( \int_{0}^{1}\left\Vert \sum_{i=1}^{n}r_{i}(t)x_{i}\right\Vert ^{q}dt\right) ^{\frac{1}{q}}\leq\mathrm{K}_{p,q}\left( \int_{0 ^{1}\left\Vert \sum_{i=1}^{n}r_{i}(t)x_{i}\right\Vert ^{p}dt\right) ^{\frac{1}{p}} \] holds, regardless of the choice of a Banach space $X$ and of finitely many vectors $x_{1},\dots,x_{n}\in X$. \end{theorem} The previous theorem, combined with (\ref{7654}) recovers the Khinchin inequality (see \cite[page 10]{djt}), when $q=2$ and $X=\mathbb{K}$, and in this case $\mathrm{K}_{p,2}^{-1}$ is usually denoted by $A_{p}$; it is simple to observe that $A_{2}=1$ and it is well-known that $A_{1}=(\sqrt{2})^{-1}$. In the recent years a series of works (\cite{jfa, joed, adv, LONDON, PSDianaE}) have shown that several important results of the theory of summing operators in fact need essentially no linear structure to be valid. The proof of our main result shall rely on the abstract environment presented in \cite{adv}. From now on, unless stated otherwise, $p\in\lbrack1,\infty).$ Let $X$ and $Y$ be non-void sets, $\mathcal{H}\left( X;Y\right) $ be a non-void family of mappings from $X$ to $Y$ and $K$ be a compact Hausdorff space. Let \[ R\colon K\times X\times X\longrightarrow\lbrack0,\infty)~\text{and \mathrm{~}S\colon\mathcal{H}\left( X;Y\right) \times X\times X\longrightarrow\lbrack0,\infty) \] be arbitrary mappings. A mapping $f\in\mathcal{H}\left( X;Y\right) $ is $RS$-\textit{abstract} $p$-\textit{summing} if there is a constant $C$ such that \[ \left( \sum_{i=1}^{n}S(f,x_{i},q_{i})^{p}\right) ^{\frac{1}{p}}\leq C\sup_{\varphi\in K}\left( \sum_{i=1}^{n}R\left( \varphi,x_{i},q_{i}\right) ^{p}\right) ^{\frac{1}{p}}, \] for all $x_{1},\ldots,x_{n},q_{1},\ldots,q_{n}\in X$ and all positive integers $n$. We define \[ {\mathcal{H}}_{RS,p}(X;Y)=\left\{ f\in\mathcal{H}\left( X;Y\right) :\text{ }f\text{ is }RS\text{-abstract }p\text{-summing}\right\} . \] Let $X$ be a pointed metric space and $Y$ be a Banach space. We shall choose a suitable Banach space $X^{d}\subseteq\mathbb{K}^{X}$ containing $X^{\#}$ such that \[ f^{a}\colon Y^{\#}\longrightarrow X^{d}~,f^{a}(h):=h\circ f \] is well-defined for all $f\in\mathcal{H}\left( X;Y\right) $. In general, for our applications, it will be enough to consider $X^{d}=X^{\#}$. Note that, when $f$ is a linear operator between Banach spaces we have $f^{a}(y^{\ast })=f^{\ast}(y^{\ast})$ for all $y^{\ast}\in Y^{\ast}$, i.e., $f^{a}$ is a kind of \textit{abstract adjoint} of $f$. We shall need the following properties of $R$ and $S$: \medskip \textbf{(I)} $S(f,x,q)\leq\left\Vert f(x)-f(q)\right\Vert $ for all $f\in\mathcal{H}\left( X;Y\right) $ and $x,q\in X$. \textbf{(II)} $K\subseteq X^{d}$, $B_{T^{a}(Y^{\ast})}\subseteq K$ and \[ \left\vert g(x)-g(q)\right\vert \leq R(g,x,q)\ \ \text{for all}\ \ g\in B_{T^{a}(Y^{\ast})}\text{ and}\ x,q\in X. \] In the next section we shall present a nonlinear Kwapie\'{n}-type theorem valid for $\mathcal{L}_{p}$-spaces instead of just Hilbert spaces. As an illustration, when the main result of this section is restricted to the linear case, we conclude that if $X$ is a Banach space and $Y$ is an $\mathcal{L _{p}$-space, with $2\leq p<\infty$, then \[ \Pi_{al.s.p}^{dual}(X,Y)\subseteq\Pi_{1}(X,Y). \] Our techniques are, in general, different from the ones used in the proofs of Theorems \ref{kuap} and \ref{py}; for instance, unlike what is done in the aforementioned results, we do not use the Pietsch Domination Theorem. Our main result is the following: \begin{theorem} \label{q1} Let $X$ be a pointed metric space, $Y$ be an $\mathcal{L}_{p $-space, with $2\leq p<\infty$, and $f:X\longrightarrow Y$ an arbitrary map. If $R$ and $S$ satisfy \textbf{(I)} and \textbf{(II)} and $f^{a}|_{Y^{\ast}}$ is almost $p$-summing, then $f$ is $RS$-abstract $1$-summing. \end{theorem} In Section 4, applications of our main result are provided. For instance, choosing suitable $R,S$ we obtain extensions of the theorems of Kwapie\'{n} and Chen--Zheng (see Subsections \ref{ss1} and \ref{ss2}). \section{Proof of the main result} Let $f\in\Pi_{al.s.p}^{dual}(X,Y)$. Fix $x_{1},...,x_{n},q_{1},...,q_{n}\in X$ and a subspace $Y_{0}$ of $Y$ containing $\{f(x_{1}),...,f(x_{n ),f(q_{1}),...,f(q_{n})\}$ and for which there is an isomorphism $v:Y_{0}\longrightarrow\ell_{p}^{m}$ such that $\Vert v\Vert\Vert v^{-1 \Vert\leq\lambda$. Then, using the monotonicity of the $\ell_{p}$ norms, the Khinchin inequality and denoting by $\left( e_{k}\right) _{k=1}^{m}$ the canonical basis of $\ell_{p^{\ast}}^{m}=\left( \ell_{p}^{m}\right) ^{\ast}$, where $p^{\ast}$ is the conjugate of $p$, it follows that \begin{align*} \sum_{i=1}^{n}S(f,x_{i},q_{i}) & \overset{\mathbf{(I)}}{\leq}\sum_{i=1 ^{n}\Vert f(x_{i})-f(q_{i})\Vert\\ & =\sum_{i=1}^{n}\Vert v^{-1}vf(x_{i})-v^{-1}vf(q_{i})\Vert\\ & \leq\Vert v^{-1}\Vert\cdot\sum_{i=1}^{n}\Vert vf(x_{i})-vf(q_{i})\Vert\\ & =\Vert v^{-1}\Vert\cdot\sum_{i=1}^{n}\left( \sum_{k=1}^{m}|e_{k}\left( vf(x_{i})-vf(q_{i}\right) )|^{p}\right) ^{\frac{1}{p}}\\ & \leq\Vert v^{-1}\Vert\cdot\sum_{i=1}^{n}\left( \sum_{k=1}^{m}|e_{k}\left( vf(x_{i})-vf(q_{i})\right) |^{2}\right) ^{\frac{1}{2}}\\ & \leq\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot\sum_{i=1}^{n}{\int\nolimits_{0 ^{1}}\left\vert \sum_{k=1}^{m}e_{k}\left( vf(x_{i})-vf(q_{i}\right) )\cdot r_{k}(t)\right\vert dt\\ & =\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot\sum_{i=1}^{n}{\int\nolimits_{0}^{1 }\left\vert \sum_{k=1}^{m}e_{k}(vf(x_{i}))\cdot r_{k}(t)-\sum_{k=1}^{m e_{k}(vf(q_{i}))\cdot r_{k}(t)\right\vert dt\\ & =\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot\sum_{i=1}^{n}{\int\nolimits_{0}^{1 }\left\vert \sum_{k=1}^{m}v^{\ast}\left( e_{k}\right) (f(x_{i}))\cdot r_{k}(t)-\sum_{k=1}^{m}v^{\ast}\left( e_{k}\right) (f(q_{i}))\cdot r_{k}(t)\right\vert dt. \end{align*} By the Hahn-Banach Theorem we can extend $v^{\ast}(e_{k})$ to $Y^{\ast}$, and thus \begin{align*} \sum_{i=1}^{n}S(f,x_{i},q_{i}) & \leq\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot \sum_{i=1}^{n}{\int\nolimits_{0}^{1}}\left\vert \sum_{k=1}^{m}f^{a}v^{\ast }\left( e_{k}\right) (x_{i})\cdot r_{k}(t)-\sum_{k=1}^{m}f^{a}v^{\ast }\left( e_{k}\right) (q_{i})\cdot r_{k}(t)\right\vert dt\\ & =\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot\sum_{i=1}^{n}{\int\nolimits_{0}^{1 }\left\vert \left( \sum_{k=1}^{m}r_{k}(t)f^{a}v^{\ast}\left( e_{k}\right) \right) (x_{i})-\left( \sum_{k=1}^{m}r_{k}(t)f^{a}v^{\ast}\left( e_{k}\right) \right) (q_{i})\right\vert dt. \end{align*} Combining the previous inequality with \textbf{(II)} we obtain \[ \sum_{i=1}^{n}S(f,x_{i},q_{i})\leq\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot {\int\nolimits_{0}^{1}}\left\Vert \sum_{k=1}^{m}r_{k}(t)f^{a}v^{\ast (e_{k})\right\Vert dt\cdot\sup_{g\in K}\sum_{i=1}^{n}R(g,x_{i},q_{i}) \] and, using the monotonicity of the $L_{p}$ norms, we have \[ \sum_{i=1}^{n}S(f,x_{i},q_{i})\leq\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot\left( {\int\nolimits_{0}^{1}}\left\Vert \sum_{k=1}^{m}r_{k}(t)f^{a}v^{\ast (e_{k})\right\Vert ^{2}dt\right) ^{\frac{1}{2}}\cdot\sup_{g\in K}\sum _{i=1}^{n}R(g,x_{i},q_{i}). \] Since $f^{a}|_{Y^{\ast}}$ is almost $p$-summing, it follows that $f^{a}|_{Y^{\ast}}v^{\ast}$ also is almost $p$-summing and, since $\Vert (e_{k})_{k=1}^{n}\Vert_{p,w}=1$ in $\ell_{p^{\ast}}^{m},$ we have \begin{align*} \sum_{i=1}^{n}S(f,x_{i},q_{i}) & \leq\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot \pi_{al.s.p}(f^{a}|_{Y^{\ast}}v^{\ast})\cdot\sup_{g\in K}\sum_{i=1 ^{n}R(g,x_{i},q_{i})\\ & \leq\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot\Vert v^{\ast}\Vert\cdot\pi _{al.s.p}(f^{a}|_{Y^{\ast}})\cdot\sup_{g\in K}\sum_{i=1}^{n}R(g,x_{i},q_{i})\\ & =\sqrt{2}\cdot\Vert v^{-1}\Vert\cdot\Vert v\Vert\cdot\pi_{al.s.p (f^{a}|_{Y^{\ast}})\cdot\sup_{g\in K}\sum_{i=1}^{n}R(g,x_{i},q_{i})\\ & \leq\sqrt{2}\cdot\lambda\cdot\pi_{al.s.p}(f^{a}|_{Y^{\ast}})\cdot\sup_{g\in K}\sum_{i=1}^{n}R(g,x_{i},q_{i}). \end{align*} As a corollary, since every Hilbert space is an $\mathcal{L}_{2}$-space, we have an abstract generalization of the Kwapie\'{n} Theorem: \begin{corollary} [Abstract Kwapie\'{n}-type Theorem]\label{7777}Let $X$ be a pointed metric space, $H$ be a Hilbert space and $f\in\mathcal{H}(X,H)$. If $R$ and $S$ satisfy \textbf{(I)} and \textbf{(II) }and $f^{a}|_{H^{\ast}}$ is absolutely $q$-summing for some $1\leq q<\infty$, then $T$ is $RS$-abstract $1$-summing. \end{corollary} \section{Applications} In this section we show that Theorem \ref{q1} provides new Kwapie\'{n}-type theorems for several classes of linear and nonlinear summing operators (and recovers the known results). \subsection{Absolutely summing linear operators\label{ss1}} Let $X$ and $Y$ be Banach spaces and $T\colon X\longrightarrow Y$ be a bounded linear operator. Considering the weak-star topology in $B_{X^{\ast}}$, and letting \[ \left( K,\mathcal{H}\left( X;Y\right) \right) =\left( B_{X^{\ast },\mathcal{L}(X;Y)\right) , \ \[ \left\{ \begin{array} [c]{c R\colon K\times X\times X\longrightarrow\lbrack0,\infty)\\ R(\varphi,x,q)=|\varphi(x-q)|, \end{array} \right. \] an \[ \left\{ \begin{array} [c]{c S\colon\mathcal{L}\left( X;Y\right) \times X\times X\longrightarrow \lbrack0,\infty)\\ S(T,x,q)=\Vert T(x-q)\Vert, \end{array} \right. \] it is plain that $T$ is $RS$-abstract $p$-summing if, and only if, there is a constant $C$ such that \[ \left( \sum_{i=1}^{n}\Vert T(x_{i}-q_{i})\Vert^{p}\right) ^{\frac{1}{p}}\leq C\sup_{\varphi\in B_{X^{\ast}}}\left( \sum_{i=1}^{n}|\varphi(x_{i -q_{i})|^{p}\right) ^{\frac{1}{p}},\label{36M \] for all $x_{1},\ldots,x_{n},q_{1},\ldots,q_{n}\in X$ and every positive integer $n$. Thus, $T$ is absolutely $p$-summing if, and only if, it is $RS$-abstract $p$-summing. Taking $X^{d}=X^{\#}$, it follows that $T^{a}|_{Y^{\ast}}=T^{\ast}$ and the hypotheses \textbf{(I)} and \textbf{(II)} are satisfied. If $Y$ is an $\mathcal{L}_{p}$-space, with $2\leq p<\infty$, Theorem \ref{q1} tells us that $T$ is absolutely $1$-summing wherever $T^{\ast}$ almost $p$-summing. So, Corollary \ref{7777} recovers Theorem \ref{nn11}. \subsection{Lipschitz $p$-summing and $p$-dominated operators\label{ss2}} Let $X$ be a pointed metric space with a base point denoted by $0$ and let $Y$ be an $\mathcal{L}_{p}$-space, with $2\leq p<\infty$. We can easily note that $T$ is Lipschitz $p$-summing if, and only if, it is $RS$-abstract $p$-summing. Considering $B_{X^{\#}}$ with the pointwise convergence topology, \[ \left( K,\mathcal{H}\left( X;Y\right) \right) =\left( B_{X^{\#} ,Lip_{0}(X,Y)\right) , \ \[ \left\{ \begin{array} [c]{c R\colon K\times X\times X\longrightarrow\lbrack0,\infty)\\ R(\varphi,x,q)=|\varphi(x)-\varphi(q)|, \end{array} \right. \] and \[ \left\{ \begin{array} [c]{c S\colon Lip_{0}\left( X,Y\right) \times X\times X\longrightarrow \lbrack0,\infty)\\ S(T,x,q)=\Vert T(x)-T(q)\Vert, \end{array} \right. \] and considering $X^{d}=X^{\#}$ we can invoke Theorem \ref{q1} to prove the following extension of the Chen--Zheng Theorem: \begin{theorem} Let $X$ be a pointed metric space and $Y$ be an $\mathcal{L}_{p}$-space, with $2\leq p<\infty$. If $T\in Lip_{0}(X;Y)$ is such that $T^{\#}|_{Y^{\ast}}$ is almost $p$-summing, then $T$ is Lipschitz $1$-summing. \end{theorem} According to Chen and Zheng \cite{ChZh}, a Lipschitz operator $T\colon X\longrightarrow Y$ between Banach spaces is \textit{Lipschitz $p$\textit{-dominated} if there exist a Banach space $Z$ and an absolutely $p$-summing linear operator $L\colon X\longrightarrow Z$ such that \begin{equation} \left\Vert T(x)-T(q)\right\Vert \leq\left\Vert L(x)-L(q)\right\Vert \mathrm{~for~all~}x,q\in X. \label{dominated \end{equation} As a consequence of our main theorem we also conclude that if a Lipschitz operator $T\colon X\longrightarrow Y$ satisfies (\ref{dominated}) for a certain $L:X\longrightarrow Z$, where $Z$ is an $\mathcal{L}_{p}$-space, with $2\leq p<\infty$, then $T$ is Lipschitz $1$-dominated whenever $L^{\ast}$ is almost $p$-summing. \subsection{Absolutely $p$-summing $\Sigma$-operators} According to Angulo-L\'{o}pez and Fern\'{a}ndez-Unzueta \cite{mexicana, segre}, given Banach spaces $X_{1},\ldots,X_{n}$, the set \[ \Sigma_{X_{1},\ldots,X_{n}}:=\{x_{1}\otimes\cdots\otimes x_{n}\in X_{1 \otimes\cdots\otimes X_{n}:x_{i}\in X_{i},i=1,\ldots,n\} \] is the metric space of decomposable tensors endowed with the metric induced by the projective tensor norm. It is called the \textit{metric Segre cone} of $X_{1},\ldots,X_{n}$. If $T: X_{1} \times\cdots\times X_{n} \longrightarrow Y$ is a multilinear mapping and $Y$ is a vector space, we denote by $\hat{T} \in L(X_{1} \otimes\cdots\otimes X_{n};Y)$ the unique linear mapping satisfying that for every $x_{i} \in X_{i}$, with $i \in\{1,\ldots,n\}$ and $T(x_{1},\ldots,x_{n}) = \hat{T}(x_{1}\otimes\cdots\otimes x_{n})$. \begin{definition} [see \cite{mexicana}]If $X_{1},\ldots,X_{n}$ are Banach spaces and $Y$ is a vector space, an operator $f: \Sigma_{X_{1},\ldots,X_{n}} \longrightarrow Y$ is a $\Sigma$-operator if there exists a multilinear operator $T \in\mathcal{L}(X_{1},..., X_{n},Y)$ such that $f=\hat{T}|_{\Sigma _{X_{1},\ldots,X_{n}}}$. \end{definition} We denote \[ L(\Sigma_{X_{1},\ldots,X_{n}};Y) = \{f: \Sigma_{X_{1},\ldots,X_{n}} \longrightarrow Y ~:~ f ~\text{is a}~ \Sigma\text{-operator} \}, \] and by $\mathcal{L}(\Sigma_{X_{1},\ldots,X_{n}})$ we denote the space of scalar-valued continuous $\Sigma$-operators endowed with the Lipschitz norm, which happens to be a dual Banach space in which we shall consider the weak-star topology. \begin{definition} [see \cite{mexicana}]Let $X_{1},\ldots,X_{m},Y$ be Banach spaces. A bounded $\Sigma$-operator $f\colon\Sigma_{X_{1},\dots,X_{m}}\longrightarrow Y$ is \textit{absolutely }$p$\textit{-summing} if there is a constant $C$ so that \[ \left( \sum\limits_{i=1}^{n}\left\Vert f(x_{i})-f(q_{i})\right\Vert ^{p}\right) ^{\frac{1}{p}}\leq C\sup_{\varphi\in B_{\mathcal{L}(\Sigma _{X_{1},\ldots,X_{m}})}}\left( \sum\limits_{i=1}^{n}\left\vert \varphi (x_{i})-\varphi(q_{i})\right\vert ^{p}\right) ^{\frac{1}{p}}, \] for every natural number $n$ and all $x_{i},q_{i}\in\Sigma_{X_{1},\ldots ,X_{m}}$, with $i=1,...,n$. \end{definition} Let us choose, in our abstract framework, $X$ as the pointed metric space $\Sigma_{X_{1},\ldots,X_{m}}$ with the base point $0=0\otimes\cdots\otimes0$, and \[ \left( K,\mathcal{H}\left( X;Y\right) \right) =\left( B_{\mathcal{L (\Sigma_{X_{1},\ldots,X_{m}})},Lip_{0}(X,Y)\right) , \ \[ \left\{ \begin{array} [c]{c R\colon K\times X\times X\longrightarrow\lbrack0,\infty)\\ R(\varphi,x,q)=|\varphi(x)-\varphi(q)|, \end{array} \right. \] and \[ \left\{ \begin{array} [c]{c S\colon Lip_{0}\left( X,Y\right) \times X\times X\longrightarrow \lbrack0,\infty)\\ S(f,x,q)=\Vert f(x)-f(q)\Vert. \end{array} \right. \] Letting $Y$ be an $\mathcal{L}_{p}$-space, with $2\leq p<\infty$, $X^{d}=X^{\#}=\mathcal{L}(\Sigma_{X_{1},\ldots,X_{m}})$ we have that a bounded $\Sigma$-operator $f\in Lip_{0}\left( X;Y\right) $ is absolutely $p$-summing if, and only if, it is $RS$-abstract $p$-summing. As $f^{a}=f^{\#}$, Theorem \ref{q1} and Corollary \ref{7777} provide Kwapie\'{n}-type theorems for this class of operators. \subsection{Lipschitz $p$-summing operators at one point} The next definition is motivated by the notion of absolutely summing arbitrary mappings between Banach spaces (see \cite{joed} and the references therein). \begin{definition} Let $X$ be a normed vector space and $Y$ be a Banach space. A Lipschitz operator $T:X\longrightarrow Y$ is \textit{Lipschitz} $p$-\textit{summing at the point} $w\in X$ if there is a constant $C$ such that \begin{align*} \left( \sum_{i=1}^{n}\left\Vert T(w+x_{i})-T(w+q_{i})\right\Vert ^{p}\right) ^{\frac{1}{p}}\leq C\sup_{\varphi\in B_{X^{\#}}}\left( \sum_{i=1 ^{n}\left\vert \varphi(x_{i})-\varphi(q_{i})\right\vert ^{p}\right) ^{\frac{1}{p}} , \label{mmu333 \end{align*} for every positive integer $n$ and every $x_{1},...,x_{n},q_{1},...,q_{n}\in X.$ \end{definition} Note that Lipschitz $p$-summability at $0$ is precisely the notion of Lipschitz $p$-summability (see (\ref{765765})). Given $T\in Lip(X,Y)$, we define $T_{w}:X\longrightarrow Y$ by \[ T_{w}(x)=T(w+x)-T(w) \] and it is simple to check that $T\in Lip(X,Y)$ if, and only if, $T_{w}\in Lip(X,Y)$ and $Lip(T)=Lip(T_{w})$. Note also that \[ T\ \text{is Lipschitz}\ p\text{-summing at}\ w\ \text{if, and only if, T_{w}\ \text{is Lipschitz }p\text{-summing \] and \[ T\ \text{is Lipschitz}\ p\text{-summing}\ \text{if, and only if, T_{0}\ \text{is Lipschitz }p\text{-summing}. \] Considering \[ w-Lip(X,Y):=\{T_{w}:T\in Lip(X,Y)\}, \] and choosing $K=B_{X^{\#}}$, with the pointwise convergence topology, ${\mathcal{H}}(X;Y)=w-Lip(X,Y)$ and $R$, $S$ defined b \[ \left\{ \begin{array} [c]{c R\colon B_{X^{\#}}\times X\times X\longrightarrow\lbrack0,\infty)\\ R(\varphi,x,q)=|\varphi(x)-\varphi(q)|, \end{array} \right. \] an \[ \left\{ \begin{array} [c]{c S\colon w-Lip(X,Y)\times X\times X\longrightarrow\lbrack0,\infty)\\ S(T_{w},x,q)=\left\Vert T_{w}(x)-T_{w}(q)\right\Vert , \end{array} \right. \] we conclude that a $T_{w}\in w-Lip(X,Y)$ is $RS$-abstract $p$-summing if, and only if, $T_{w}$ is Lipschitz $p$-summing (or equivalently, $T$ is Lipschitz $p$-summing at $w$). If $Y$ is an $\mathcal{L}_{p}$-space, with $2\leq p<\infty$, $X^{d}=X^{\#}$, then the abstract adjoint $(T_{w})^{a}$ coincides with the Lipschitz adjoint $(T_{w})^{\#}$. Consequently, we have the following theorem \begin{theorem} Let $X$ be a normed vector space, $Y$ be an $\mathcal{L}_{p}$-space, with $2\leq p<\infty$, and $T\in Lip(X;Y)$. If $(T_{w})^{\#}|_{Y^{\ast}}$ is almost $p$-summing, then $T$ is Lipschitz $1$-summing at $w$. \end{theorem} Note that the Chen--Zheng theorem generalizes the result of Kwapie\'{n}, since it shows that when $X$ is a normed space we can replace $Lip_{0}(X;Y)$ by $Lip(X;Y).$

https://arxiv.org/abs/1907.01541

A Column Generation Approach to the Discrete Barycenter Problem

The discrete Wasserstein barycenter problem is a minimum-cost mass transport problem for a set of discrete probability measures. Although an exact barycenter is computable through linear programming, the underlying linear program can be extremely large. For worst-case input, a best known linear programming formulation is exponential in the number of variables, but has a low number of constraints, making it an interesting candidate for column generation.In this paper, we devise and study two column generation strategies: a natural one based on a simplified computation of reduced costs, and one through a Dantzig-Wolfe decomposition. For the latter, we produce efficiently solvable subproblems, namely, a pricing problem in the form of a classical transportation problem. The two strategies begin with an efficient computation of an initial feasible solution. While the structure of the constraints leads to the computation of the reduced costs of all remaining variables for setup, both approaches may outperform a computation using the full program in speed, and dramatically so in memory requirement. In our computational experiments, we exhibit that, depending on the input, either strategy can become a best choice.

\section{Introduction}\label{sec:intro} Optimal transport problems involving joint transport to a set of probability measures appear in a variety of fields, including recent work in image processing \cite{jbtcg-19,sa-19}, machine learning \cite{hbccp-19,shbmccps-18,ydpbbgc-19}, and graph theory \cite{uddgn-18}, to name but a few. A {\em barycenter} is another probability measure which minimizes the total distance to all input measures (i.e, images) and acts as an {\em average} distribution in the probability space. Of particular consideration is the squared Wasserstein distance due to the preservation of the geometric structure of the input. The wide scope of optimal transport problems makes challenging even a reasonably comprehensive summary; for recent monographs on the Wasserstein distance and computational optimal transport, we refer the reader to \cite{pc-18} and \cite{pz-19}, in addition to the seminal work of Villani \cite{v-09}. In image processing, the probability measures not only have discrete support (a finite number of points to which positive mass is associated), but are supported on the same structured set (a pixel grid). This highly structured, highly repetitive support is a best-case input. In this setting, a barycenter can be computed exactly in polynomial time \cite{bp-18}, but in practice this cost is still prohibitive. Therefore, considerable activity continues on exact, approximate and heuristic methods of computation, such as alternating minimization algorithms \cite{qp-18,tdgu-20,ylst-19}. State of the art approximation methods solve the entropy regularized optimal mass transport problem introduced in \cite{c-13}. Entropic regularization leads to a strongly convex program, and its smoothing effects give qualitatively different solutions \cite{bccnp-14}. Entropy regularized transport problems can be solved efficiently, with a linear-in-$n$ complexity bound, using iterative Bregman projection algorithms \cite{bccnp-14,cd-14,twk-18}, although work continues on the stability and complexity of these algorithms, e.g., \cite{kddgtu-19}. By contrast, a worst-case setting occurs for input measures with no repetition or known structure, such as in wildfire ignition points or crime locations \cite{bp-18}. This unstructured setting is not as well-understood as the highly active structured setting and few of the efficient approximation methods consider such challenging input. In this paper, we seek to address this gap through the development of an exact algorithm tailored to worst-case input. We begin with a formal definition of a barycenter and its properties. \subsection{The Discrete Barycenter Problem} The discrete barycenter problem is defined as follows. Given a set of probability measures $P_1, \ldots, P_n$, each with a finite set $\text{supp}{(P_i)}$ of support points in $\mathbb{R}^d$, and a set of $n$ nonnegative weights $\lambda_i \in \mathbb{R}$ with $\sum_{i=1}^n \lambda_i =1$, find a probability measure $\bar P$ on $\mathbb{R}^d$, that is, a (Wasserstein) barycenter, satisfying \begin{equation}\label{eqn:barycost} \phi(\bar P):=\sum\limits_{i=1}^n \lambda_i W_2(\bar P, P_i)^2 = \inf\limits_{P\in \mathcal{P}^2(\mathbb{R}^d)} \sum\limits_{i=1}^n \lambda_i W_2(P,P_i)^2, \end{equation} where $W_2$ is the quadratic Wasserstein distance and $\mathcal{P}^2(\mathbb{R}^d)$ is the set of all probability measures on $\mathbb{R}^d$ with finite second moments \cite{ac-11}. Since $P_1, \ldots, P_n$ have finite sets of support points, we call the measures {\em discrete}. Because $P_1, \ldots P_n$ are discrete, the solution measure $\bar P$ also has a finite set of support points, and the Wasserstein distance is the squared Euclidean distance \cite{abm-16,coo-15,v-09}. The discrete barycenter problem is a multi-marginal optimal transport problem and as such is significantly more challenging than the classical two-marginal optimal transport problem referenced later in Section \ref{sec:price}. In fact, multi-marginal optimal transport is no longer a network flow problem \cite{lhcj-19}, and a variant of the problem -- finding optimal solutions with a bound on the size of the support set -- has recently been shown to be NP-hard \cite{bp-19a}. The problem can be solved exactly by linear programming \cite{ac-11,abm-16,coo-15}, but all known LP formulations may require an exponential number of variables, scaling by the product of the sizes of the support sets of the input measures \cite{bp-18}. Some formulations also have an exponential number of constraints, but beneficially for the worst-case setting, the smallest linear programming formulation contains an extremely low number of constraints \cite{bp-18}. These dimensions indicate the linear program is a prime candidate for column generation. Barycenters produced by linear programming have favorable properties, namely, provable sparsity of the support set, and the following {\em non-mass-splitting property} of any optimal transport plan (the support points in each $P_1, \ldots, P_n$ to which each support point in $P$ transports mass, and the amount of mass transported) \cite{ac-11,abm-16}. This non-mass-splitting property is fundamental to the modeling of many physical applications where such a split would be infeasible, and is also the basis for our description of the linear programming model. \begin{definition}[The Non-mass-splitting Property] The mass of each barycenter support point is transported fully to a single support point in each measure; that is, for each $x_k \in \text{supp}(\bar P)$ with corresponding mass $z_k$, $k = 1, \ldots, |\text{supp}(\bar P)|$, there exists exactly one $x_{i} \in \text{supp}(P_i)$, $i = 1, \ldots, n$ to which the entire mass $z_k$ is transported in any optimal transport plan. \end{definition} Since the non-mass-splitting property holds for all barycenters, each support point in a barycenter is associated with a single combination of input support points, consisting of the points to which its mass is transported. The set of combinations of input support points is denoted $S^* = \{ (\textbf{x}_{1}, \ldots, \textbf{x}_{n}) : \textbf{x}_{i} \in \text{supp}(P_i) \text{ for } i = 1, \ldots, n \}$, with elements $s_h = (\textbf{x}_1^h, \textbf{x}_2^h, \ldots, \textbf{x}_n^h)$, $h = 1,\ldots, |S^*|$. Each combination $s_h$ has an associated \emph{weighted mean} $\textbf{x}^h = \sum_{i=1}^n \lambda_{i} \textbf{x}_{i}^h$. The weighted mean $\textbf{x}^h$ is the optimal location for joint mass transport to the points in the combination $s_h$. Therefore the set $S$ of distinct weighted means contains {\em all possible support points} for the barycenter. We can now formally describe the worst-case setting to which our algorithm is tailored: when each combination $s_h$ produces a different weighted mean $\textbf{x}^h$. Then we say the measures $P_1, \ldots, P_n$ are in \textit{general position}, and using $|P_i|$ to denote the size of the support set of $P_i$, the number of distinct $\textbf{x}^h$ is $|S^*| = \prod_{i=1}^n |P_i|$. Thus the number of weighted means is exponential in the number of input measures $n$ -- the worst-case scenario for the number of possible support points for $\bar P$. We provide an example of a discrete measure in $\mathbb{R}^2$ in Figure \ref{fig:measures} (left), and three measures in general position in Figure \ref{fig:measures} (right). For this tiny example, verifying the measures are in general position is elementary but somewhat tedious, as the weighted means of all 27 combinations of support points must be computed and verified as unique; in general, verifying whether a particular set contains the correct possible support points is NP-hard \cite{bp-18}. A barycenter for these measures is displayed in Figure \ref{fig:measwithbary} (left), shown with associated transport in Figure \ref{fig:measwithbary} (right). \begin{figure}[!t] \begin{center} \fbox{\includegraphics[scale=.5]{SimplefigBlue.pdf}} \fbox{\includegraphics[scale=.5]{SimplefigThree.pdf}} \end{center} \caption{(left) A discrete probability measure in $\mathbb{R}^2$ with three support points. The size of the points indicates their associated mass. (right) Three measures in $\mathbb{R}^2$ each with three support points. All combinations $(\textbf{x}_1, \textbf{x}_2, \textbf{x}_3)$ produce a different weighted mean; therefore these measures are in general position.}\label{fig:measures} \end{figure} \begin{figure}[!b] \begin{center} \fbox{\includegraphics[scale=.5]{SimplefigBary.pdf}} \fbox{\includegraphics[scale=.5]{SimplefigTrans.pdf}} \end{center} \caption{(left) Assuming $\lambda_i=\frac{1}{3}$ for $i = 1,2,3$, a barycenter $\bar P$ for the three measures from Figure \ref{fig:measures}. Each support point has mass $\frac{1}{4}$. (right) The mass transport from each barycenter support point to the original measures. Each barycenter support point is the weighted mean of the points to which it transports.}\label{fig:measwithbary} \end{figure} \subsection{A Linear Program for Measures in General Position} In this paper, we use an LP formulation from \cite{bp-18} based on $S^*$ and chosen because, among known LP formulations, it requires the fewest variables and constraints for general position measures. In this formulation, which we call LP (\ref{LPw}), a variable is introduced for each combination of support points; that is, each $s_h$ has a corresponding variable $w_h$ representing the mass assigned to $\textbf{x}^h$ and transported fully to each $\textbf{x}_i^h$, $i = 1, \ldots, n$. The total transport cost of a unit of mass from $\textbf{x}^h$ is given by $c_h = \sum_{i=1}^n ||\textbf{x}^h-\textbf{x}_{i}^h||^2$. Constraints arise from the requirement that the total transport to each support point in each measure is exactly equal to its mass $d_i$. This produces one equality constraint for each $\textbf{x}_i$; that is, \[ \sum_{h:\textbf{x}_i^h = \textbf{x}_i} w_h = d_i, \text{ }\spc\text{ } \forall i = 1, \ldots, n, \text{ }\spc \forall \textbf{x}_i \in \text{supp}(P_i). \] The constraints can be represented as a real $\sum_{i=1}^n |P_i| \times \prod_{i=1}^n |P_i|$ matrix $A$ times the vector $w$. In $A$, column $h$ contains ones in the $n$ rows where $\textbf{x}_i^h = \textbf{x}_i$, and zeroes otherwise. The matrix $A$ is highly structured, a crucial property for an efficient column generation strategy. We postpone the analysis of $A$ to Section \ref{sec:price}. With $c$ as the vector of costs $c_h$ and $w$ as the vector of variable masses $w_h$, a linear program for the discrete barycenter problem is: \begin{equation*}\label{LPw} \begin{array}{crl} \tag{general} \mathrm{min}& c^T \textbf{w} & \nonumber \\ \mathrm{s.t. } &A \textbf{w} = &d \\ &\textbf{w} \geq & 0. \nonumber \end{array} \end{equation*} The number of constraints $\sum_{i=1}^n |P_i|$ scales linearly, equal to the total number of support points in the input measures. Meanwhile, the number of variables $\prod_{i=1}^n |P_i|$ scales exponentially in the number of input measures; this is our motivation for a column generation algorithm \subsection{Column Generation} As a service to the reader, we briefly recall the basics of a column generation strategy for solving a linear program. The process begins with partitioning the constraint matrix as $A = \begin{bmatrix} A_p \\ A_m \\ \end{bmatrix}$. A preselected number of rows are assigned to the matrix $A_p$ for use in a separate \textit{pricing problem}. The pricing problem must contain enough information for its solution to be meaningful, but also remain sufficiently simple to be efficiently solvable. In Section \ref{sec:price}, we discuss how we achieve this balance for the barycenter problem. The remaining rows of $A$ are assigned to the matrix $A_m$ for use in a version of LP (\ref{LPw}) called the \textit{master problem}. The master problem already has fewer constraints than LP (\ref{LPw}) due to the relocation of $A_p$ to the pricing problem, but the primary reduction in size is achieved by restricting the number of columns in the linear program. The resulting linear program is called the {\em restricted master problem}. We implement the restricted master problem using a Dantzig-Wolfe formulation \cite{ccz-14} for LP (\ref{LPw}); for background on this classic linear programming formulation see, for example, \cite{dw-60}. In LP (\ref{LPw}), the bound on the elements of $d$ guarantee the values of $\textbf{w}$ are in $[0,1]$, so that the feasible region is a bounded polytope and can be expressed in terms of its vertices. Let $V$ represent the set of vertices $\textbf{p}$ of the polytope $\{\textbf{w} \in \mathbb{R}^d: A\textbf{w} = d, \textbf{w} \geq 0\}$. Using a subset of $V$ of size $J$, that is, $\{ \textbf{p}_1, \ldots, \textbf{p}_J\}$ with $1 \leq J \leq |V|$, and variables $\mu$, a restricted master problem LP (\ref{LPRM}) is as follows. \begin{equation*}\label{LPRM} \begin{array}{crl} \tag{RM} \mathrm{min} & \sum\limits_{j=1}^{J} (c^T\textbf{p}_j)\mu_j \nonumber \\ \mathrm{s.t.}&\sum\limits_{j=1}^{J} (A_m \textbf{p}_j) \mu_j &= d_m \\ &\sum\limits_{j=1}^{J} \mu_j &= 1 \\ &\mu_j &\geq 0, \forall j=1,\ldots,J\nonumber \end{array} \end{equation*} The structure that makes LP (\ref{LPw}) a prime candidate for column generation is preserved in LP (\ref{LPRM}): the number of constraints remains low, $\sum_{i=1}^n |P_i| +1$, as LP (\ref{LPRM}) has just one additional constraint for convexity. Ideally, only a fraction of the total number of vertices are used in LP (\ref{LPRM}), so that the number of columns and variables remains low. This is the purpose of the pricing problem: to generate meaningful columns for introduction to LP (\ref{LPRM}). The current optimum of LP (\ref{LPRM}) -- specifically, the dual solution -- is used in the objective of the pricing problem to find new columns that may lead to a better optimum when included in LP (\ref{LPRM}). Column generation terminates when the optimal value of the pricing problem is no longer negative. \subsection{Outline} We establish an efficiently solvable pricing problem in Section \ref{sec:price}. Starting a column generation process requires a solution to an initial feasible LP (\ref{LPRM}). To satisfy this requirement, two algorithms which produce a vertex of $\{\textbf{w} \in \mathbb{R}^d: A\textbf{w} = d, \textbf{w} \geq 0\}$ are presented in Section \ref{sec:master}. Section \ref{sec:comp} contains computational experiments demonstrating the practical advantages of our column generation algorithm over direct computations using LP (\ref{LPw}). We finish with some concluding remarks in Section \ref{sec:conc}. \section{Pricing Problem}\label{sec:price} An efficiently solvable pricing problem is key to successful column generation. In this section, we exploit two beneficial aspects of the discrete barycenter problem which lead to an efficient strategy: the structure of the coefficient matrix $A$, and the structure of the pricing problem that results from the choice of exactly two measures for $A_p$. Recall that the pricing problem uses the current optimum of the restricted master problem to produce a new column to introduce to LP (\ref{LPRM}). Specifically, the objective function of the pricing problem requires the dual solution to LP (\ref{LPRM}), so let the dual solution be given by $(\textbf{y}, \sigma)$, where $\textbf{y}$ is the real vector of size $\sum_{i=1}^n |P_i|$ containing the shadow prices associated with the mass transport constraints, and $\sigma \in \mathbb{R}$ is the shadow price associated with the convexity constraint in LP (\ref{LPRM}). Then the base form of the pricing problem is: \begin{equation*}\label{LPprice} \begin{array}{crl} \tag{price} \mathrm{min} &(c^T-&\textbf{y}^TA_m) \textbf{p} + \sigma \nonumber \\ \mathrm{s.t.} & A_p \textbf{p} &= d_p \\ &\textbf{p} &\geq 0. \end{array} \end{equation*} LP (\ref{LPprice}) is still an exponential-sized linear program: The constraint matrix $A_p$ has an exponential number of columns, as does the matrix $A_m$, and the cost vector $c$ has an exponential number of elements. In fact, LP (\ref{LPprice}) contains the same number of variables as LP (\ref{LPw}). We now develop an improved pricing problem using information specific to the barycenter problem \subsection{The Structure of the Coefficient Matrix $A$} Recall that $A$ contains only elements $1$ and $0$: in column $h$, there is a $1$ when $\textbf{x}_{i}$ is in the tuple $s_h$, that is, $\textbf{x}^h_i = \textbf{x}_i$, and $0$ otherwise. In fact, each column contains exactly $n$ nonzero coefficients. The pattern created within the matrix $A$ is displayed in Example \ref{ex:A}: each row has consecutive ones alternating with consecutive zeros. For each measure, the consecutive ones start in the first column for the first constraint in each measure, then start in the second row immediately after the end of the previous consecutive ones, continuing to the last constraint of the measure, forming a block. The width of the block depends on the measure $P_i$ with which the constraints are associated. The number of consecutive ones equals the product of the sizes of the measures with a higher index: the rows of $A$ associated with $P_i$, $1 \leq i < n$, contain $\prod_{l=i+1}^n |P_l|$ consecutive ones. The block for the final measure is the identity matrix. \begin{example}\label{ex:A} The matrix $A$ for four measures with sizes $|P_1| = |P_3| = 2$ and $|P_2| = |P_4|= 3$ contains blocks of ones and zeros. The width of block structure for particular constraints depends on the index $i$ of the corresponding measure $P_i$. Here there are 36 total columns, and the number of consecutive ones for each measure is 16, 6, 3, and 1, respectively. \setcounter{MaxMatrixCols}{50} \begin{center} \begin{tabular}{cc} $ A= \begin{bmatrix} 1 & 1 & 1 & 1 & 1 & 1& 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 &0 &0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 &0 &0 &1 & 1 & 1 & 1 & 1 & 1& 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\ \hline 1 & 1 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 &1 & 1 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 1 & 1& 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 1 & 1& 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 &0 & 0 & 0 & 0 & 0 & 0 &1 & 1 & 1 & 1 & 1 & 1& 0 & 0 & 0 & 0 & 0 & 0 &0 & 0 & 0 & 0 & 0 & 0 &1 & 1 & 1 & 1 & 1 & 1 \\ \hline 1 & 1 & 1 & 0 & 0 & 0 & 1 & 1 & 1 &0 & 0 & 0 &1 & 1 & 1 &0 & 0 & 0 &1 & 1 & 1 & 0 & 0 & 0 & 1 & 1 & 1 &0 & 0 & 0 &1 & 1 & 1 &0 & 0 & 0 \\ 0 & 0 & 0 & 1 & 1 & 1 &0 & 0 & 0 & 1 & 1 & 1 &0 & 0 & 0 & 1 & 1 & 1 & 0 & 0 & 0 & 1 & 1 & 1 &0 & 0 & 0 & 1 & 1 & 1 &0 & 0 & 0 & 1 & 1 & 1 \\ \hline 1 & 0 & 0 & 1 & 0 & 0 & 1& 0 & 0 & 1 & 0 & 0 & 1& 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1& 0 & 0 & 1 & 0 & 0 & 1& 0 & 0 & 1 & 0 & 0 \\ 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0& 0 & 1 & 0& 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0& 0 & 1 & 0& 0 & 1 & 0 \\ 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1& 0 & 0 & 1& 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1& 0 & 0 & 1& 0 & 0 & 1 \\ \end{bmatrix}$ & $\begin{matrix}\multirow{2}{*}{$P_1$} \\ \\ \multirow{3}{*}{$P_2$} \\ \\ \\ \multirow{2}{*}{$P_3$} \\ \\ \multirow{3}{*}{$P_4$} \\ \\ \\ \end{matrix}$ \end{tabular} \\ \end{center} \qed \end{example} Since the number of consecutive ones for each row of $A$ can be generated from the sizes of the support sets $|P_i|$, the columns of $A$ are easily generated solely from the problem input. Formulas for generating a given column $h$ are provided in Algorithm \ref{alg:ACol}. While Algorithm \ref{alg:ACol} is written with the sums and products in their respective formulas, recalculating these values on repeated runs can be avoided by instead storing the number of consecutive ones and the width of a full block for each measure. Because $A$ is easily generated, the matrix $A_m$ is {\em not required in memory}. Instead, in the objective function, $\textbf{y}^TA_m$ is calculated using the $|P_i|$ to determine which dual values should be added. For further computational efficiency, updates to elements of $(c^T-\textbf{y}^T A_m)$ are only required for those values of $\textbf{y}$ which have changed from the previous iteration; due to the sparse nature of $A_m$, many elements may remain unchanged. \begin{algorithm}[h] \caption{Generation of Column $h$}\label{alg:ACol} \begin{algorithmic}[1] \Statex Input:\begin{itemize} \item Column Index $h$, assuming the index of the first column is $0$ \item $|P_i|$ for $i = 1, \ldots, n$ \end{itemize} \Statex Output: Column $h$ of matrix $A$, denoted $A_h$ \State Let $A_h$ be a column of zeros with length $(\sum_{i=1}^n |P_i|)$ \State $j = \lfloor \frac{h}{\prod_{l=2}^n |P_l|} \rfloor$ \State $A_h(j) = 1$ \For {$i=2, \ldots, n-1}$ \State$j = \sum_{l =1}^{i-1} |P_l| + \lfloor \frac{h \pmod{ \prod_{l=i}^n |P_l|}}{\prod_{l=i+1}^n |P_l|} \rfloor$ \State $A_h(j) = 1$ \EndFor \State $j = \sum_{l =1}^{n-1} |P_l| + h \pmod{|P_n|}$ \State $A_h(j) = 1$ \end{algorithmic} \end{algorithm} The pattern of consecutive ones also guarantees that $A_p$ always has {\em many duplicate columns.} Continuing with the matrix $A$ from Example \ref{ex:A}, in Example \ref{ex:A2}, we assign the constraints for the first two measures to $A_p$, resulting in a matrix with six unique columns, each repeated six times. For any number of measures $n$, partitioning the constraints for $k$ measures, $1 \leq k < n$, to $A_p$ results in $n_u = \prod_{i=1}^k |P_i|$ unique columns, while the number of times each column is duplicated is $n_d = \prod_{i=k+1}^n |P_i|$. For a fixed $k$, the number of unique columns is no longer exponential; we justify the choice $k=2$ momentarily. \begin{example}\label{ex:A2} Using the matrix $A$ from Example \ref{ex:A}, a decomposition of all constraints associated with the first two measures into the pricing problem gives this matrix $A_p$. \begin{center} \begin{tabular}{cc} $A_p = \begin{bmatrix} 1 & 1 & 1 & 1 & 1 & 1& 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 &0 &0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 &0 &0 &1 & 1 & 1 & 1 & 1 & 1& 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\ \hline 1 & 1 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 &1 & 1 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 1 & 1& 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 1 & 1& 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 &0 & 0 & 0 & 0 & 0 & 0 &1 & 1 & 1 & 1 & 1 & 1& 0 & 0 & 0 & 0 & 0 & 0 &0 & 0 & 0 & 0 & 0 & 0 &1 & 1 & 1 & 1 & 1 & 1 \\ \end{bmatrix}$ & $\begin{matrix} \multirow{2}{*}{$P_1$} \\ \\ \multirow{3}{*}{$P_2$} \\ \\ \\ \end{matrix}$ \end{tabular} \end{center} Every column is repeated six times: $|P_3|\cdot|P_4|$. The matrix of unique columns is $U_p$. \[U_p = \begin{bmatrix} 1 &1 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 &1 & 1 & 1 \\ 1 & 0& 0 &1 & 0 & 0 \\ 0 & 1 & 0 & 0 & 1 & 0 \\ 0& 0 & 1 & 0 & 0 & 1 \\ \end{bmatrix} \] \end{example}\qed Replacing the constraint matrix $A_p$ in LP (\ref{LPprice}) with the matrix of unique columns $U_p$ requires a corresponding change to the objective function. Noting that LP (\ref{LPprice}) is the minimization of a linear objective, only the most negative coefficient for each unique column is required: in an optimal solution, all mass is assigned to such a column. Thus, it suffices to keep a best-cost vector $b$ for the unique columns. These two substitutions produce LP (\ref{Uprice}). \begin{equation*}\label{Uprice} \begin{array}{crl} \tag{Uprice} \mathrm{min} & b^T \textbf{q}& + \sigma \nonumber \\ \mathrm{s.t.} & U_p \textbf{q} &= d_p \\ &\textbf{q} &\geq 0. \end{array} \end{equation*} Using LP (\ref{Uprice}) improves solvability and memory requirements in two major ways: when $k=2$, LP (\ref{Uprice}) requires just $|P_1| \cdot |P_2|$ variables, a tremendous reduction from $\prod_{i=1}^n |P_i|$. The number of variables {\em does not depend on $n$}. When the input measures have support sets of equal size $|P|$, this eliminates $|P|^{n-2}$ variables. Additionally, the constraint matrix is stored in memory, so the benefit of replacing $A_p$ with $U_p$ is significant. Using LP (\ref{Uprice}) instead of LP (\ref{LPprice}) requires additional preprocessing each iteration to construct the best-cost vector $b$, which does use the exponential-sized vector $(c^T-\textbf{y}^T A_m)$. The preprocessing for LP (\ref{Uprice}), repeated each iteration, is given in Algorithm \ref{alg:PriceSetup}. In particular, Algorithm \ref{alg:PriceSetup} highlights the important selection of which indices $h$ correspond to the unique columns used in LP (\ref{Uprice}). \begin{algorithm}[h] \caption{Setup of LP (\ref{Uprice})}\label{alg:PriceSetup} \begin{algorithmic} \Statex Intialize the vector of indices I of length $n_u$ \State Update $a = c^T-y^T A_m$ \For {$j=1,\ldots, n_u$} \For {$h = 1+n_d \cdot (j-1), \ldots, n_d \cdot j$} \If {$h = 1+n_d \cdot (j-1)$} \State $b_j =a_h$ \State $I(j) = h$ \Else \State $b_j = \min\{b_j,a_h\}$ \State $I(j) = h$ \EndIf \EndFor \EndFor \State Update objective of LP (\ref{Uprice}) \end{algorithmic} \end{algorithm} \subsection{Decomposition of Constraints for Exactly Two Measures} As in Example \ref{ex:A2}, we partition $A$ with $k=2$; that is, we always partition $A$ where $A_p$, and subsequently the matrix of unique columns $U_p$, contains all rows of constraints associated with exactly two measures. LP (\ref{Uprice}) has a linear objective function $b^T \textbf{q} + \sigma$; because of the linear objective and the structure of the constraints for two measures, LP (\ref{Uprice}) is a classical transportation problem \cite{ff-56b,m-16}, a special case of a minimum-cost flow problem. Therefore, LP (\ref{Uprice}) can be solved in strongly polynomial time \cite{abm-16,bp-18}. \begin{theorem} Let $U_p \textbf{q} = d_p$ be the constraints associated with exactly two measures. Then LP (\ref{Uprice}) is a classical transportation problem and can be solved in strongly polynomial time. \end{theorem} Next, we turn to the restricted master problem, and describe two strategies for a feasible start. \section{Master Problem: Constructing a Vertex}\label{sec:master} The process of column generation begins with a restricted master problem, restated here for reference. \begin{equation*}\label{LPRM} \begin{array}{crl} \tag{RM} \mathrm{min} & \sum\limits_{j=1}^{J} (c^T\textbf{p}_j)\mu_j \nonumber \\ \mathrm{s.t.}&\sum\limits_{j=1}^{J} (A_m \textbf{p}_j) \mu_j &= d_m \\ &\sum\limits_{j=1}^{J} \mu_j &= 1 \\ &\mu_j &\geq 0, \forall j=1,\ldots,J\nonumber \end{array} \end{equation*} By design, column generation starts with a low number of variables, then introduces additional columns iteratively. Therefore, the size of the initial LP (\ref{LPRM}) is small, and increases linearly with the number of columns generated. Once the column generation process has begun, the previous pricing problem produces a solution $\textbf{q}$ containing the nonzero elements of a new $\textbf{p}_J$ to be introduced to LP (\ref{LPRM}). This $\textbf{q}$ has at most $|P_1| \cdot |P_2|$ nonzero elements. Recall from Section \ref{sec:price} that $A_m$ is easily generated, so the pricing problem does not require $A_m$ to be stored in memory. The restricted master problem also does not require $A_m$ to be stored; instead, the $|P_i|$ are used to calculate the new column $A_m \textbf{p}_J$. Combined with the small number of nonzero elements of $\textbf{p}_J$, a computation of $A_m \textbf{p}_J$, as well as of $c^T \textbf{p}_J$, can be done efficiently. For additional efficiency, the solver for LP (\ref{LPRM}) uses the previous solution as a warm start. Using the primal simplex method then typically finds a new optimal solution in just a few simplex steps for each update of LP (\ref{LPRM}). We begin with a single column in LP (\ref{LPRM}) by generating an initial $\textbf{p}_1$ so that the first LP (\ref{LPRM}) has feasible solution $v_1=1$, as required by the convexity constraint. Substituting into the other constraints, $\textbf{p}_1$ must be a solution to $A_m \textbf{p} = d_m$. Any feasible solution to the full system $A \textbf{w} = d$ also contains a solution to the equalities $A_m \textbf{p} = d_m$, so we can generate a vertex of either $A_m \textbf{p} = d_m$ or the full system $A \textbf{w} = d$. Generating a vertex of the full system has the benefit of starting with a potential optimum, and requires little additional computational effort. We consider two methods for constructing a vertex of the polytope generated by the constraints $A\textbf{w}=d$: a greedy construction and the 2-approximation algorithm from \cite{b-17}. The following methods also work for $A_m p = d_m$ by simply limiting the input to the relevant constraints. \subsection{Feasibility through Greedy Construction} We first present an algorithm which greedily constructs a solution to $A \textbf{w} = d$. The process begins by generating a combination $s_h = (\textbf{x}_1^h, \textbf{x}_2^h, \ldots, \textbf{x}_n^h) \in S^*$, then assigning to $\textbf{x}^h$ the maximum mass $w_h$ that does not violate the non-mass-splitting property: the minimum mass not yet supplied to the support points $\textbf{x}_i^h$. After marking the mass $w_h$ as supplied, the process is repeated, generating the next combination. This process is given in Algorithm \ref{alg:GreedyFeas}. \begin{algorithm}[h] \caption{Greedy Construction of $\textbf{w}$: $A\textbf{w} = d$ }\label{alg:GreedyFeas} \begin{algorithmic}[1] \Statex Input: vector $n_o$ containing number of consecutive ones for each $i$ \Statex Output: vector $\textbf{w}$ \Statex Let $L = 1$, $m_1 = 0$, and $j_i = 1 \ \ \forall i = 1, \ldots, n$ \Statex For each $P_i$, and for $j_i = 1, \ldots, |P_i|$, let $b_{ij_i}$ be the mass associated with support point $\textbf{x}_{ij_i}$ \While {$\sum_{l=1}^L m_{l} < 1$} \State $m_{L} = \min\{b_{ij_i}\}$ \State $h = \sum_{i=1}^n ((j_i-1)n_o(i))$ \State $w_h = m_L$ \For {$i=1, \ldots, n$} \State $b_{ij_i} = b_{ij_i}-m_{L}$ \If {$b_{ij_i} = 0$} \State $j_i = j_i +1$ \EndIf \EndFor \State $L = L+1$ \EndWhile \end{algorithmic} \end{algorithm} \begin{theorem}\label{thm:GreedyEff} For $n$ probability measures with fixed sizes $|P_1|, \ldots, |P_n|$, Algorithm \ref{alg:GreedyFeas} runs in $\mathcal{O}(n^2)$ in the arithmetic model of computation. \end{theorem} \begin{proof} Assume probability measures $P_1, \ldots, P_n$ have support sets of fixed size. First, we show that the number of nonzero elements produced by Algorithm \ref{alg:GreedyFeas}, which is also the number of repetitions of the loop of Algorithm \ref{alg:GreedyFeas}, is between $\max_{1 \leq i \leq n} \{|P_i|\}$ and $\sum_{i=1}^n |P_i| -n+1.$ The lower bound, $\max_{1 \leq i \leq n} \{|P_i|\}$, is an immediate consequence of the non-mass-splitting property maintained by Algorithm \ref{alg:GreedyFeas}. For the upper bound, $\sum_{i=1}^n |P_i| -n+1$, note that the last iteration must fully supply the mass to $n$ points, one $\textbf{x}_{i}$ from all $P_i$, because the total mass for each $P_i$ is the same (one). In each previous iteration, the minimum number of support points whose index $j_i$ changes is one, for a total of $\sum_{i=1}^n (|P_i|-1) +1 = \sum_{i=1}^n |P_i| -n+1$ iterations. By assumption, the sizes of the support sets $|P_i|$ do not depend on $n$, so the number of iterations of Algorithm \ref{alg:GreedyFeas} is linear in $n$. Since each step inside this loop of Algorithm \ref{alg:GreedyFeas} requires at most linear-in-$n$ elementary operations, we obtain Theorem \ref{thm:GreedyEff}. \qed \end{proof} As an additional consequence of the iteration bound $\sum_{i=1}^n |P_i| -n+1$, the number of nonzero mass elements of $\textbf{w}$, and subsequently the relevant elements of $\textbf{w}$ that solve $A_m \textbf{p} = d_m$, are also bounded. Therefore the setup of the first LP (\ref{LPRM}) is efficient. We now show that Algorithm \ref{alg:GreedyFeas} produces a vertex of the polytope generated by the constraints $A\textbf{w}=d$. \begin{theorem} Algorithm \ref{alg:GreedyFeas} generates a vertex of the polytope $\{\textbf{w} \in \mathbb{R}^d: A\textbf{w} = d, \textbf{w} \geq 0\}$. \end{theorem} \begin{proof} Let $A$, $d$ be given and let $\textbf{w}$ be generated using Algorithm \ref{alg:GreedyFeas}. We show there exists a $c$ such that $\textbf{w}$ is the unique optimal solution to: \begin{equation*} \begin{array}{crl} \mathrm{min}& c^T \textbf{w} & \nonumber \\ \mathrm{s.t. } &A \textbf{w} = &d \\ &\textbf{w} \geq & 0. \nonumber \end{array} \end{equation*} Let $M$ be the set of nonzero elements of $\textbf{w}$, with size $|M| = L$. Order the elements of $M$ in order of construction by Algorithm \ref{alg:GreedyFeas}, $m_1, \ldots, m_L$. Also order the associated indices $h_1, \ldots, h_L$ as calculated by Algorithm \ref{alg:GreedyFeas}. Let $c_{h_1} = 1$, $c_{h_2} = 2$, \ldots, and $c_{h_L} = L$. Let all other $c_h$, those whose $h$-index is not in $h_1, \ldots, h_l$, be $\sum_{i=1}^n |P_i| -n+2$ (Recall: $|M| \leq \sum_{i=1}^n |P_i| -n+1$). By construction, removing mass from a combination with a lower index and assigning it to a combination with higher index in $M$, that is, from $m_j$ to $m_k$ with $j \leq k$, will strictly increase the value of $c^T \textbf{w}$. This includes moving mass to a combination with no mass in $\textbf{w}$, that is, with an index not in $M$. So it suffices to show that mass cannot be reassigned from $m_k$ to $m_j$, $j \leq k$. The mass $m_j$ is chosen such that for at least one $x_{j_i}$, the mass $d_{j_i}$ has been fully supplied. Therefore $m_j$ cannot be increased without violating the constraints $A \textbf{w} = d$. Therefore $\textbf{w}$ minimizes $c^T \textbf{w}$ subject to $A \textbf{w} = d$, since the maximum mass allowable is assigned to the cheapest costs. Furthermore, $\textbf{w}$ does so uniquely, since any change in its elements will strictly increase the value of $c^T \textbf{w}$ due to the construction of $c$. Therefore $\textbf{w}$ is a vertex. \qed \end{proof} \begin{figure}[!b] \begin{center} \fbox{\includegraphics[scale=.45]{figure3meas.pdf}} \fbox{\includegraphics[scale=.45]{figure0_3trans.pdf}} \end{center} \caption{(left) Three measures in general position with 10 or 11 support points and equally distributed mass. (right) A greedily constructed feasible solution. Transport from three sample points -- those constructed first, fifth, and seventeenth -- is shown (arrows). Each support point is the weighted mean of its three destination points. }\label{fig:GenPos} \end{figure} In Figure \ref{fig:GenPos} (left), we display an example with three measures, two with 10 support points and one with 11 support points. Each measure has equally distributed mass. Applying Algorithm \ref{alg:GreedyFeas} results in a feasible solution supported on 20 weighted means of varying mass, displayed in Figure \ref{fig:GenPos} (right), along with the transport for three sample points. \subsection{Feasibility through a 2-Approximation}\label{sec:masterapprox} An alternative to the greedy-start algorithm, the provably efficient 2-approximation algorithm presented in \cite{b-17}, also generates a measure with feasible transport. The motivation for considering this algorithm is the guarantee that the generated transport cost is at most twice the optimal transport cost of a barycenter. This allows the examination of the effects of a feasible start of the column generation with a provably good initial objective value. The non-optimal measure, called an {\em approximate barycenter}, is found by allowing mass to be placed only at the locations of the original support points, $\bigcup_{i=1}^n \text{supp}{(P_i)}$. The points in $\bigcup_{i=1}^n \text{supp}{(P_i)}$ are introduced to an alternative LP from \cite{abm-16,bp-18}. This strategy avoids the possibly exponential number of weighted means by replacing them with the original support points. The approximate barycenter can be calculated in strongly polynomial time \cite{b-17}. In Figure \ref{fig:2app} (left), we show the result of this algorithm applied to the data set from Figure \ref{fig:GenPos} (left). Mass is assigned to 15 of the 31 possible support points from the support sets of the original measures. The approximate barycenter does not immediately translate to a solution to LP (\ref{LPw}) for two reasons. One, this method of approximation does not maintain the non-mass-splitting property. Therefore a support point in the approximate barycenter may not correspond to a single combination $s_h \in S^*$. Two, even if a solution is non-mass-splitting, the points in $\bigcup_{i=1}^n \text{supp}{(P_i)}$ are typically not weighted means of any combination $s_h$. Fortunately, the same process as in Algorithm \ref{alg:GreedyFeas} can address both of these issues. When applied to an approximate barycenter, Algorithm \ref{alg:GreedyFeas} corresponds to the iterative improvement process presented in \cite{b-17}. Each support point in the approximate barycenter is processed as follows: first, a combination in $S^*$ is created by collecting one point from each measure to which the support point transports mass. Then, the minimum of the mass available from the support point of the approximate barycenter and the smallest mass still required by the destination points is assigned to the combination. Applying Algorithm \ref{alg:GreedyFeas} to Figure \ref{fig:2app} (left) results in a measure whose support set contains 24 weighted means, displayed in Figure \ref{fig:2app} (right). \begin{figure}[!t] \begin{center} \fbox{\includegraphics[scale=.45]{figuretwoappalg.pdf}} \fbox{\includegraphics[scale=.45]{twoappspread.pdf}} \end{center} \caption{(left) For the measures of Figure \ref{fig:GenPos}, the support of the measure produced by the two-approximation algorithm consists of 15 of the original support points. (right) Each support point in {\em (left)} is processed by Algorithm \ref{alg:GreedyFeas} to regain the mass-splitting property. The new support points are weighted means.}\label{fig:2app} \end{figure} \section{Computations}\label{sec:comp} The primary goal of these experiments is to demonstrate the computational benefits of our column generation strategy for general position measures over a direct solve of the linear program. To this end, we use event locations from a real-world data set given in longitude and latitude coordinates to construct general position measures; because the event locations occur without known structure, probability measures with these support points are in general position. The measures have varying numbers of support points, displayed in the corresponding tables in this section, and uniformly distributed mass. All computations have been run on a laptop (MacBook Pro, 2.9 GHz Intel Core i7, 16 GB of RAM, SSD). Data processing and the setup of the LPs were implemented in C\texttt{++} and the LPs were solved using Gurobi 8.0. The source code is available at \url{https://github.com/StephanPatterson/Barycenter-Formulations}. Barycenters of general position measures can be produced directly using the previously known LP (\ref{LPw}) only when the measures contain a small number of support points, as LP (\ref{LPw}) already requires millions of variables. As our goal is to compare our column generation algorithm to a direct solution, the following analysis focuses primarily on measures with small support sets; throughout this section, we use the number of variables as a reference label for a particular instance. Our tables include the number of iterations for each instance, which also represents the number of variables introduced in the restricted master problem. To complement the analysis on measures with small support sets, we conclude with some experiments on column generation alone, increasing the size of the support sets. The second goal of these experiments is to examine the practical behavior of variations in implementation strategy. To this end, we compare six variants of column generation for general position data, three using Algorithm \ref{alg:GreedyFeas}, {\em greedy-start}, and three using the 2-approximation algorithm \cite{b-17}, {\em 2-app-start}, to generate the initial vertex. All implementations that use the same initialization algorithm begin with the same vertex. As described in Section \ref{sec:price}, the constraints from two measures are assigned to the pricing problem. Our implementations differ in which two measures are chosen: we consider three options based on the size of the support sets. The variants labeled {\em large} choose the two measures with the largest support sets, resulting in a large pricing problem and small restricted master problem. The {\em small} variants choose the two measures with the smallest support sets, resulting in a small pricing problem and large restricted master problem. Finally, the {\em any} variants just use the first two input measures for the pricing problem, regardless of size; essentially, an arbitrary pair of measures. The choice of the pairs of measures is independent of the initialization strategy. \\ \begin{table}[!t] \begin{center} \begin{tabular}{|c|c|c|c|c|c|c|c|c|}\hlin &\multicolumn{2}{|c|}{ LP (\ref{LPw})} & \multicolumn{6}{|c|}{Greedy-start } \\ \hline &\multicolumn{2}{|c|}{ }& \multicolumn{2}{|c|}{Any }& \multicolumn{2}{|c|}{ Large } & \multicolumn{2}{|c|}{ Small} \\ \hline n & Variables &Time (sec) & Iter & Time (sec) & Iter & Time (sec)& Iter & Time (sec)\\ \hline 12& 55,296 & 0.48 & 132 & \textbf{0.19} & 172 &\textbf{0.19} & 198&0.31 \\ \hline 12&138,240 & 1.25& 239 & \textbf{0.79} & 307 & 0.93 & 300&3.51 \\ \hline 12& 1,990,656 & 23.17& 356 & \textbf{9.75} & 541 & 13.37 &331&10.75 \\ \hline 12 &2,177,280 & 20.32& 478 & 18.21 & 541 & \textbf{13.61} &229&19.48\\ \hline 14&2,654,208 & 36.49 & 328 & \textbf{15.71} & 694 & 22.00 & 328&21.69 \\ \hline 13&3,981,312 & 42.29& 459 & 23.44 & 598 & 28.85 & 340&\textbf{23.19} \\ \hline 15&18,579,456 & 229.17 & 399 & \textbf{133.60} & 704 & 150.86 & 489&226.80 \\ \hline 12& 25,288,704& 592.70& 640 & 213.74 & 541 & \textbf{144.28} &416&191.21 \\ \hline 14 & 28,440,792 & 1029.42 & 545& \textbf{252.37} & 854 & 276.13 & 425&312.28 \\ \hline 14& 32,514,048 & 1248.36 & 601 & 240.06 & 575 & \textbf{217.16} & 450&375.66 \\ \hline 14 & 37,933,056 & * & 759 & 362.06 & 618 & \textbf{268.33} & 479&427.34 \\ \hline 14 & 75,866,112 & * & 595 & 938.63 & 514 & \textbf{454.45} & 466&902.17 \\ \hline 16 & 130,056,192 & * & 513 & \textbf{1265.87} & 872 & 1306.86 & 559&1950.79 \\ \hline 14 &151,732,224 & * & 823 & 1713.78 & 710&\textbf{1307.42} & 559&1625.01\\ \hline \end{tabular} \end{center} \caption{Comparison of greedy-start algorithms for $n$ measures per experiment, with fastest times in bold. Each measure has a small number (between 2 and 12) of points in general position. For larger instances, a direct solution was not possible due to memory constraints (*).}\label{tab:gryrcomps} \end{table} \noindent{\bf Small Measures: Running Times.} We begin with $n$ measures, most of which contain two to four support points, with the largest containing twelve support points. In Table \ref{tab:gryrcomps}, we see that all implementations of greedy-start column generation outperform LP (\ref{LPw}). For most instances, the total running time is about half of a direct solution of LP (\ref{LPw}). The large-choice variant is typically fastest, followed by the any-choice variant. When any-choice is fastest, the running times of large-choice are usually comparable, and both are always better than a direct computation. In Table \ref{tab:tayrcomps}, we display the same instances, now solved using the 2-app-start strategy. Again, all variants significantly outperform the direct solution of LP (\ref{LPw}), and using the large-choice pricing problem typically performs well. The pricing problem that works best for each experiment is the same as for greedy-start in Table \ref{tab:gryrcomps}. Comparing the running times to the greedy-start variants, 2-app-start appears to perform better for small instances, but greedy-start performs better for large instances. \\ \begin{table}[!t] \begin{center} \begin{tabular}{|c|c|c|c|c|c|c|c|c|}\hline &\multicolumn{2}{|c|}{ LP (\ref{LPw})} & \multicolumn{6}{|c|}{ 2-app-start } \\ \hline &\multicolumn{2}{|c|}{ } & \multicolumn{2}{|c|}{ Any}& \multicolumn{2}{|c|}{ Large}& \multicolumn{2}{|c|}{ Small} \\ \hline n&Variables&Time (sec) & Iter & Time (sec) & Iter & Time (sec)& Iter & Time (sec) \\ \hline 12& 55,296&0.48&116 & 0.18 &54&\textbf{0.09} &131 & 0.20 \\ \hline 12& 138,240 &1.25& 179 & \textbf{0.46}&253& 0.59 &137 & 0.54 \\ \hline 12& 1,990,656&23.17&381 & \textbf{9.71} &806 & 19.99& 327 & 10.96 \\ \hline 12& 2,177,280 &20.32&586 & 19.60&601&\textbf{15.61} &212 & 17.55 \\ \hline 14 &2,654,208 & 36.49& 312 & \textbf{13.71} & 428 & 13.75 & 365 & 23.40 \\ \hline 13 &3,981,312 & 42.29 & 623 & 30.55 & 816 & 38.77 & 384 & \textbf{26.97} \\ \hline 15 & 18,579,456 & 229.17& 429 & \textbf{149.30} & 881 & 185.47 & 428 & 194.71 \\ \hline 12&25,288,704 &592.70 & 745 & 264.90& 536&\textbf{157.27} & 474 & 214.33 \\ \hline 14 &28,440,792 & 1029.42 & 636 & 280.12 & 814 & \textbf{258.00} & 406 & 312.40\\ \hline 14 &32,514,048 & 1248.36& 660 & 277.50 & 549 & \textbf{204.41} & 491& 407.67 \\ \hline 14 &37,933,056 & *& 974 & 448.53 & 581 & \textbf{254.73} & 507 & 446.64 \\ \hline 14 &75,866,112 & *& 690 & 1113.12 & 687 & \textbf{595.37} & 487 & 1031.18 \\ \hline 16 & 130,056,192 & *& 554 & \textbf{1382.33} & 1169 & 1722.82 & 545 & 1886.52 \\ \hline 14 &151,732,224 & *&1020 & 2050.50 & 1016 & \textbf{1771.51} & 671 & 2084.94 \\ \hline \end{tabular} \end{center} \caption{Comparison of 2-app-start algorithms for $n$ measures per experiment, with fastest times in bold. Each measure has a small number (between 2 and 12) points in general position. For larger instances, a direct solution was not possible due to memory constraints (*).}\label{tab:tayrcomps} \end{table} \noindent{\bf Small Measures: Memory Usage.} The instances in Tables \ref{tab:gryrcomps} and \ref{tab:tayrcomps} show that, as expected, significantly larger barycenter problems are solvable using column generation. For the four largest instances, a direct computation through LP (\ref{LPw}) is not possible due to memory constraints, but column generation variants reach a solution in only a few additional iterations over the smaller problem instances. In fact, our algorithm requires orders of magnitude less memory. The six variants of column generation do not vary significantly in memory usage; a comparison of memory requirements is available in Table \ref{tab:mem}. \\ \begin{table}[!t] \begin{center} \begin{tabular}{|c|c|c|} \hline \multicolumn{2}{|c|}{ LP (\ref{LPw}) }& Column Generation \\ \hline 1,990,656 & 3.70 GB & \ 95 MB \\ \hline 2,177280 & 4.00 GB & \ 95 MB \\ \hline 2,654,208 & 4.97 GB & 114 MB\\ \hline 3,981,312 & 7.14 GB &164 MB\\ \hline 18,579,456 & 45.2 GB & 724 MB\\ \hline 25,288,704 & 49.6 GB & 977 MB \\ \hline 28,440,792 & 54.8 GB & 1.08 GB\\ \hline 32,514,048 & 62.7 GB &1.23 GB\\ \hline \end{tabular} \end{center} \caption{For the instances of Tables \ref{tab:gryrcomps} and \ref{tab:tayrcomps}, the reduction in memory required is dramatic. \hspace*{2.5cm}}\label{tab:mem} \end{table} \noindent{\bf Small Measures: Running Time Breakdown.} Table \ref{tab:Percentage} displays the percentage of time per iteration spent on each step of the column generation process. The majority of computation time is in the update of $(c^T-\textbf{y}^TA_m)$. This is as expected in light of the discussion in Section \ref{sec:price}: it is the only part of the iteration that still requires exponential effort, due to the exponential size of $(c^T-\textbf{y}^TA_m)$. The computational effort required for solving LP (\ref{LPRM}) and LP (\ref{Uprice}) is very small, relative to the effort required for the coefficient calculation. Since LP (\ref{Uprice}) can be solved in less total time than LP (\ref{LPRM}), assigning more constraints to LP (\ref{Uprice}) instead of LP (\ref{LPRM}) is beneficial to running times; however, LP (\ref{LPRM}) also solves relatively quickly, so overall, using the large-choice variant has a minor beneficial effect on running times. This explains why the large-choice variants perform well, but not tremendously better than the other variants in settings where the measures are fairly uniform in size; we will see more differentiation between the large-choice and any-choice variants in later experiments. \\ \begin{table}[!t \begin{center} \begin{tabular}{|c|c|} \hline Step & Percentage of Computation Time \\ \hline Setup LP (\ref{LPRM}) & $< 0.1 \%$ \\ \hline Solve LP (\ref{LPRM}) & 1.1\% \\ \hline Update $(c^T-\textbf{y}^TA_m)$ & 72.5\% \\ \hline Calculate $b$ & 26.2\% \\ \hline Solve LP (\ref{Uprice})& $< 0.1\%$\\ \hline \end{tabular} \end{center} \caption{Percentage computation time per step of an average iteration of column generation. Most of the effort is spent on the setup of LP (\ref{Uprice}); the computation times for solving LP (\ref{Uprice}) and the setup of the next LP (\ref{LPRM}) contribute negligibly to the total.}\label{tab:Percentage} \end{table} \begin{figure}[!t] \begin{center} \includegraphics[scale=.4]{ErrorIter2015.pdf} \end{center} \caption{A representative graph of the absolute error of the optimum of LP (\ref{LPRM}), given by iteration. The 2-app-start begins with less error, but requires more iterations before improvement.}\label{fig:ErrIter} \end{figure} \noindent{\bf Small Measures: Convergence.} Finally, we measure the absolute error associated with the objective value of LP (\ref{LPRM}) for each iteration. We see similar behavior in each experiment; in Figure \ref{fig:ErrIter} we give a representative graph. The 2-app-start begins with less error as expected. However, starting with less error does not appear to reduce the number of required iterations to the final solution as the value does not begin to improve until after the iteration where the greedy-start strategies have reached similar error. All strategies begin with a phase where the objective value is unchanged. To explain this, we note that the $\textbf{p}$ produced by the pricing problem do not satisfy $A_m \textbf{p} = d_m$, by construction. Several iterations are required to build up enough columns (degrees of freedom) for the restricted master problem to reallocate mass and start improving the objective value. After this initial phase, the objective value does improve steadily, although there may still be consecutive iterations with the same value. For this reason, improvement in the objective value is not a feasible stopping criterion. This is similar to the behavior of the simplex method, where multiple iterations may be required to leave a degenerate vertex. In our experiments, we terminate when the pricing objective is above $-10^{-6}$. The corresponding true error is generally about $10^{-7}$, as shown in Figure \ref{fig:ErrIter}. Of course, higher tolerance levels for stopping the column generation algorithm could be used for earlier termination. For the experiment of Figure \ref{fig:ErrIter}, each variant reaches an absolute error of $10^{-5}$ in approximately 150 fewer iterations than reaching an absolute error of $10^{-7}$. \\ \begin{table}[!b] \begin{center} \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|} \hline &LP (\ref{LPw}) & \multicolumn{4}{|c|}{Greedy-start} & \multicolumn{4}{|c|}{2-app-start} \\ \hline &&\multicolumn{2}{|c|}{ Any }& \multicolumn{2}{|c|}{Large }& \multicolumn{2}{|c|}{ Any }& \multicolumn{2}{|c|}{ Large} \\ \hline n &Variables & Iter & Time (sec)& Iter & Time (sec)& Iter & Time (sec)& Iter & Time (sec)\\ \hline 12 & 11,870,208 & 602 & 154.62 & 553 & 75.36 & 740 & 175.27 & 551 & \textbf{73.83} \\ \hline 14 & 157,593,600 & 960 & 2268.26 & 745 & 1364.09 & 1456 & 3372.14 & 665 &\textbf{1223.45} \\ \hline 14 &163,897,344 & 883 & 3221.68 & 431 & 810.62 & 1191&4482.65 & 340& \textbf{697.83}\\ \hline 15 & 331,776,000 & 1081 & 5512.45 & 1024 & \textbf{3819.14} & 1683& 8251.77& 1017&3901.70 \\ \hline \end{tabular} \end{center} \caption{When the two measures chosen for the pricing problem are significantly larger than the other input measures, the large-choice implementations are strongly advantageous. The largest two measures contain twenty to fifty support points, while all other measures still contain two to four support points. Fastest times are highlighted in bold.}\label{tab:ExtremePrice} \end{table} \noindent{\bf Two Measures of Larger Size.} Next, we consider a few experiments for measures with larger support sets. In the previous experiments, Tables \ref{tab:gryrcomps} and \ref{tab:tayrcomps}, all measures are roughly uniform in size, so the measures moved to the pricing problem in the large-choice variants are not significantly larger than the other measures. When there is a more dramatic difference in size, the benefit of using larger measures in the pricing problem is significant. Experiments of this nature are shown in Table \ref{tab:ExtremePrice}; here, the majority of the measures are the same size as in Tables \ref{tab:gryrcomps} and \ref{tab:tayrcomps}, containing two to four support points. However, the largest two measures contain at least twenty support points, up to fifty points. In this setting, the 2-app-start slightly outperforms the greedy-start, and both benefit greatly from the large-choice partitioning. \\ \begin{table}[!t] \begin{center} \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|} \hline &Direct Solution & \multicolumn{4}{|c|}{Greedy-start} & \multicolumn{4}{|c|}{2-app-start} \\ \hline &&\multicolumn{2}{|c|}{ Any }& \multicolumn{2}{|c|}{ Large }& \multicolumn{2}{|c|}{ Any }& \multicolumn{2}{|c|}{Large} \\ \hline n&Variables & Iter & Time (sec)& Iter & Time (sec)& Iter & Time (sec)& Iter & Time (sec) \\ \hline 7 & 48,380,640 & 1475 &796.60 &1429 &\textbf{761.73} & 3406& 1735.96 & 3501 & 1804.31 \\ \hline 8 &50,738,688 & 1287&724.94 &1049 &\textbf{598.70} & 2406 & 1308.10 & 1893 & 1043.28 \\ \hline 6 &50,828,800 & 3004&1635.49 &2553 & \textbf{1386.27} & 4050 & 2195.26 & 3808 & 2057.95 \\ \hline 6 &65,664,000 & 2845 & 2007.78 & 2883 &\textbf{2000.53} & 5142 & 3722.69 & 4848 & 3487.34 \\ \hline 6 & 81,259,200 & 2627 & \textbf{2421.24} & 3993& 3505.62 &4388 & 3832.54 & 5906 & 5263.16 \\ \hline 6&91,733,292&2360&\textbf{2412.78}& 2694&2608.86 & 4436&4358.10 & 4716&4727.34 \\ \hline \end{tabular} \end{center} \caption{These experiments use fewer measures (6-8), but each measure has larger support (10-30 points). The greedy-start strategy consistently outperforms the 2-app-start strategy, with fastest times highlighted in bold.}\label{tab:mome} \end{table} \noindent{\bf All Measures of Larger Size.} We conclude our experiments by considering measures which all have larger support sets, 10-30 support points per measure. The resulting LP (\ref{LPw}) is too large to solve on the laptop for all these instances. However, all implementations of column generation find a solution, with greedy-start outperforming 2-app-start column generation, shown in Table \ref{tab:mome}. This is not a result of generating the initial vertex; the 2-app-start and greedy-start initialization times remain small, at only a few seconds. Instead, we observe a significantly higher number of iterations in both 2-app-start variants. This may be due to the effect shown in Figure \ref{fig:ErrIter}; 2-app-start implementations require many more iterations to begin improving on the initial objective value. \section{Concluding Remarks}\label{sec:conc} Column generation performs well for the discrete barycenter problem with input data in general position, due to a structured constraint matrix and the ability to solve the pricing problem with a small linear program. Using this strategy, we have been able to significantly reduce the memory requirements for a linear programming approach to a barycenter problem, allowing for computations on much larger instances with significant improvements in total solution times. When an exact solution is desired, we recommend initializing column generation using Algorithm \ref{alg:GreedyFeas}, and using the large-choice variant for the measures for the pricing problem. Further analysis would be required to better understand why starting column generation using the 2-approximation algorithm is outperformed by the simpler greedy algorithm. At this time, we believe it is more difficult for the approach to leave a ``local optimum'', as generated by the 2-approximation algorithm Most algorithms for the discrete barycenter problem, LP-based or not, require an explicit specification of the whole set of support points that may be allocated mass in the computation of an approximate or exact barycenter. The size of this support set typically is a bottleneck for the practical performance. However, the existence of sparse solutions to the problem \cite{abm-16} indicates that strategies which dynamically introduce combinations $s_h$ of support points that may be assigned mass are promising. Essentially, the goal would be to generate columns of LP (\ref{LPw}) directly; in this paper, we generated columns for a corresponding Dantzig-Wolfe formulation. An efficient generation of columns for LP (\ref{LPw}) will require methods that are quite different from classical column generation strategies: While it is possible to find the shadow price for any given combination $s_h$, the challenge lies in finding a combination of best shadow price without an explicit evaluation of each combination in $S^*$. \section*{Acknowledgments} We thank Ethan Anderes for the implementation of a visualization basis for barycenters used in \cite{abm-16}, which we modified to produce the figures of Sections \ref{sec:intro} and \ref{sec:master}. Borgwardt gratefully acknowledges support through the Collaboration Grant for Mathematicians {\em Polyhedral Theory in Data Analytics} of the Simons Foundation.

https://arxiv.org/abs/2110.12471

Convergence Criteria for Dynamic Integer Systems

Criteria are presented for testing whether every trajectory of a dynamic integer system converges to the same fixed point

\section{Introduction} \label{Eins} Natural systems (imagine a fountain or a waterfall) develop dynamically and irreversibly from a present state to following states. A realistic or fictitious mathematical model of such a system is called a {\it dynamic(al) system\/}. Similarly to nature, the model can show convergent, divergent, periodic, stable, random or chaotic behavior. A dynamic system on real values $n$ can be generated by iterative application of a unique function $f(n)$ to form a sequence of {\it successors\/} or {\it values\/} $n_{i+1}=f(n_i)$ of a {\it starting value\/} $n=n_1$. Values $n_j$ with $j<i$ are {\it predecessors\/} of $n_i$. An application of the {\it generating function\/} $f(n)$ is called a {\it step\/} and the sequence of successors is called a {\it trajectory\/} $T(n)=(n_1=n,n_2,\ldots,n_i,\ldots)$. A trajectory is also called an {\it integer sequence\/} if all values $n_i$ are integers. The aim of the present study is to construct sufficient criteria for stating that every trajectory of a dynamic system on a set $\mathbb V$ of {\it admitted\/} integers $n>0$ {\it converges\/} to the same reoccurring (or stopping) value $\widetilde n=f(\widetilde n)$ called the {\it fixed point\/}. An outstanding example is the (original) {\it Collatz dynamic system} which is generated on all integers $n>0$ by the function \begin{equation} f_\text{C}(n)=\begin{cases} ~3n+1\enskip, & \text{if $n$ is odd~;}\\ ~n/2\enskip, & \text{if $n$ is even~.} \end{cases}\label{OneA} \end{equation} The famous {\it Collatz conjecture\/} (which was posed in 1937 and is also called the 3$n$+1 conjecture) refers to the observed, but seemingly yet unproven property of the Collatz system that every trajectory leads to a periodically reoccurring value 1. The function $f_\text{C}(n)$ according to Equation (\ref{OneA}) can be simplified by excluding the needless and hindering values $n$ divisible by 2 or 3 (see Section~\ref{Sieben}). The result of this reduction is \begin{equation} n'=f_\text{C}(n)=(3n+1)/2^\nu \enskip, \quad\text{if $n$ and $n'$ are indivisible by 2 and 3} \label{OneB} \end{equation} where the integer $\nu>0$ is chosen for each $n$ so that $n'=f_\text{C}(n)$ becomes odd. Then, all successors of the value~1 are~1 since $f_\text{C}(1)=1$. This fact means $T(1)=(1,1,\ldots)$ and is also taken as a trajectory stop. The reoccurring (or stopping) value $\widetilde n=1$ is the fixed point where all trajectories are conjectured to stop. In Section~\ref{Zwei}, a general {\it dynamic (integer) system\/} is introduced and represented by a {\it graph\/} of connected {\it knots\/}. These knots are either arbitrarily, but uniquely identified or not at all. If not, then the graph is called the system {\it structure\/}. The structure alone determines the system convergence. There is a large variety of possibly convergent systems (including the Collatz system) with the same, {\it isomorphic\/} structure. Section~\ref{DreiA} explains why knot identifications should be removed to form the system structure that can finally lead to a conjectured convergence. Section~\ref{Drei} deals with system isomorphism and transformation needed for using the {\it convergence criteria\/} defined in Section~\ref{Vier}. These criteria are then applied in Sections~\ref{Fuenf} to~\ref{Sieben} to some systems which are already known or not known to converge. In particular, the Collatz system turns out to converge (see Section~\ref{Sieben}). There is already an ``old" convergence criterion. Accordingly, a system converges to the fixed point $\widetilde n=1$ if every value $n>1$ has a smaller successor. Then, this successor can repeatedly represent and replace $n$ until $n=\widetilde n=1$ is obtained. Another approach could show that {\it cycles\/} or {\it divergent\/} trajectories (see Section~\ref{Zwei}) do not exist, but it was tried in vain to the Collatz conjecture. Therefore, essentially other, much more general criteria are presented. Then, a system converges if it has, e.g., a self-similar structure (see Sections~\ref{Zwei} and~\ref{Vier}). In this study and if not otherwise stated, the term ``number" mainly means a result of counting or marking objects, whereas a number itself is called by a more informative term like ``integer", ``prime" or ``real value". This use avoids phrases like ``... a number of numbers~...". A lower case italic Latin or Greek letter always denotes an element of the set~$\mathbb N^+$ of all positive integers. Every used set or infinity means a countable one. Phrases similar to ``finite or (countably) infinite" or ``set or subset" are abbreviated by ``(in)finite" or ``(sub)set", respectively, a reference to the Collatz case by ``CC". For more details about approaches to the Collatz problem (CC), see, e.g., References \cite{Wir} to \cite{Opf} and the literature cited there. Wirsching's book \cite{Wir} is a comprehensive overview. The articles in Wikipedia \cite{Wik} and by P\"oppe \cite{Ppp} (in German) are first introductions to the matter. Feinstein \cite{Fei} and Opfer \cite{Opf} already presented interesting, but rather complex and seemingly unconfirmed proofs stating that the Collatz conjecture ``is not provable" and ``holds true", respectively (see Section~\ref{Acht}). \section{Dynamic system, its graph and structure} \label{Zwei} A dynamic (integer) system is usually generated, represented and investigated by integers $n'=f(n)$. But this approach can be complemented impressively by knots of a graph introduced as follows. As well known, a (countable) {\it set\/} contains a counted number 0, 1, 2, ..., or $\infty$ of {\it elements\/}. An element is any object or may itself be a (sub)set. Similarly, a (countable) {\it graph\/} is a set of a counted number 0, 1, 2, ..., or $\infty$ of {\it knots\/}. A~knot is any object or may itself be a (sub)graph. The knots are (dis)connected and uniquely identified or not. A graph of unidentified knots is called a {\it structure\/}. The structure is a very important essential property of a dynamic system (say, the system skeleton). It is in general defined and represented by the system graph. Such a graph consists of an (in)finite number of identical knots which are arbitrarily (in)directly connected with one another or not, and (only for the structure) without any knot identification by numbers, names or other marks. It is stressed that the graph and its structure have to be understood purely topologically without any dimensions like space, time, measure, dynamics. In the following, only dynamic systems are considered where each of them is generated by iterative application of a unique function $n'=f(n)$ on positive integers $n,n'\in\mathbb V$. This generating function may not only be given by arithmetic formulas, but also by an algorithm or a list of connections $n\to n'$. The system structure is represented by a {\it directed graph\/} of connected, identical knots $\bigcirc\!\!\!\to$, each with a single connection pointer to only one following knot, the {\it direct successor\/}. As an exception, a knot $\bigcirc$ without a pointer may be used for representing a fixed point. Although knot numbering could diminish the transparency of the system structure, the values $n$ of the dynamic system are later on used as knot numbers (see Section~\ref{Drei}). Trajectories of a dynamic system are {\it connected\/} if they have common values. If all trajectories are connected with one another, then the entire system is connected. A value $\widetilde n=f(\widetilde n)$ is a fixed point. If all trajectories converge to the same fixed point, then the system is {\it convergent\/}. A periodically reoccurring sub-sequence of values of a trajectory is called a {\it cycle\/}. A fixed point could be taken as a cycle with period~1, but this is avoided by the present study. Every trajectory ends at its {\it root\/}, that is, it converges either to a fixed point or to a cycle or it {\it diverges\/} to infinity. Every root characterizes a {\it subsystem\/} disconnected from others. Every value can act as a {\it delegate\/} (even as the fixed point) of all its predecessors. Only these predecessors together with their delegate form a {\it branch\/}. The delegate is the only root of a branch. This fact will become important in Section~\ref{Drei}. As already described in Section~\ref{Eins}, the aim of the present study is to introduce criteria for stating that an entire dynamic system, generated by a function $f(n)$, is connected and converges to a single fixed point $\widetilde n=f(\widetilde n)$. Then, more roots do not exist. Before a convergence criterion can be applied, the system graph has to be constructed by using the function $f(n)$ with $n$ acting as the knot identifying numbers. A following removal of all these numbers cleans the graph, but does not change the system structure at all. The structure alone determines the roots of the system and should therefore be investigated, whereas $f(n)$ itself with all values and trajectories will play a minor role only. It is stressed again that a structure only deals with knots and their connections, but yet not with the knot identification, e.g., by values $n$ of the dynamic system in question. Systems with the same structure are {\it isomorphic\/}. A set of them is called a {\it family\/}. Two structures are isomorphic if there is a complete one-to-one (bijective) correspondence between them, concerning all knots (unidentified), connections, roots, trajectories. There are related structures (typed by $\eta$) which can be used as valuable tools in convergence investigation. Every chosen one of these structures defines an isomorphic and convergent infinite system family. And every knot X of the chosen structure points to its own direct successor and has a given, for all knots the same fixed counted number $\eta>0$ or $\eta=\infty$ of other knots which are the {\it direct predecessors\/} of knot X. Only a single knot does not have a successor, the fixed point of the chosen structure. All knots are (in)directly connected with one another at least at the fixed point. Therefore, other roots (fixed points, cycles and divergent trajectories) do not exist. {\it Self-similarity\/} is another important essential property of every graph (or system) of a family with the just chosen particular structure. One can easily see by inspecting the graph that every knot acts as the fixed point of a branch and every branch has exactly the same structure as the chosen structure of the entire graph. This property is referred to in the following by the adjective ``self-similar (of type~$\eta$)". More generally, only infinite, connected dynamic systems with a single fixed point and no other roots can be self-similar provided that the fixed point and every knot have the same number $\eta>0$ or $\eta=\infty$ of direct predecessors. Otherwise, the structure of the entire system and that one of every branch cannot be isomorphic (see Sections~\ref{Fuenf} to~\ref{Sieben} for examples). \section{Interchanging or removing knot identifications} \label{DreiA} The following story shall explain that knots of a system graph do not necessarily need to be identified and available to determine system roots, e.g., convergence to a fixed point. Any marks used to identify knots can arbitrarily be interchanged or removed without modifying the system structure with all its knots, connections, roots. Consider a town with many sights and other important objects such as places, large and small streets, crooked lanes, churches, monuments and more. Every day, many tourists arrive at the railway station and want to return in the evening with the last train. A problem is that many tourists get lost in the town and cannot be back in time at the hidden station. And there are not enough hotels. What should be done? Two alternative solutions are discussed by the town council:\par\noindent (a) Every tourist gets a detailed town map when he starts at the station for sight-seeing.\par\noindent (b) At every important object of the town there is a signpost ``To the station $\to$". Walking to the station can be taken as a dynamic system with the station as a fixed point and the important objects of the town as the values or knots. The Solutions (a) and (b) are equivalent alternatives. With Solution (a), every object must be uniquely identified (e.g., by a name) itself and on the map as well. Solution (b) represents the structure of the system, a directed graph of the object connections. Object identification is not needed, but it must be required that a pointed way (trajectory) does not end at a cycle or anywhere off the station. This can best be achieved by setting up the signposts (knots) from the station back to the objects. There is no need to identify the signposts, but if they are identified (by names or numbers for maintenance or other purposes), then they can arbitrarily be interchanged or not without any modifying the structure. This sentence is most essential. It expresses a new important finding stated here in general for short: Structure only determines convergence and roots at all ! \section{System isomorphism and transformation} \label{Drei} As just stated above, only the structure can determine convergence of a dynamic system~A to a single fixed point. System~A is proven to converge if it has the same structure as another system~B which is already known to converge (see Section~\ref{Vier} Criterion~1). This {\it isomorphism\/} of A and~B is often missing, but can possibly achieved by system transformation which is {\it equivalent\/} with respect to the main intention, e.g., to prove a conjecture or to reduce or to expand a system. Pay first attention to several important general building blocks for equivalently transforming (reducing or, reversely, expanding as well) a dynamic system which shall be tested for convergence and is represented by a directed graph of knots $\bigcirc\!\!\!\to$ ~and $\bigcirc$. Let these knots be uniquely identified by admitted numbers $n\in\mathbb V$ and each knot be connected and pointing at most to one single direct successor knot $n'\in\mathbb V$ obtained by $n'=f(n)$. The main question is whether the system converges to a single fixed point $\bigcirc$ $\widetilde n=f(\widetilde n)$. This equation must have just one single solution $\widetilde n\in\mathbb V$. Otherwise, the system does not converge. The system structure is formed by removing all knot identifying numbers. But all knots with all their directed connections remain. Every knot can not only be arbitrarily and uniquely identified by a number $n$ or otherwise or not at all. It can itself, if suitable, also be a connected (sub)graph. Reversely, any connected (sub)graph can also act as a single knot. This slightly generalized knot definition easily allows equivalent system transformations, preferably reductions with respect to convergence or, as well, expansions to make systems isomorphic. Notice that such a transformation changes not only the graph and the structure, but also the generating function $f(n)$ and the set $\mathbb V$ of admitted knot numbers. Examples of transformation building blocks for direct or reversal application:\par\noindent {\bf Block (1):~} A chain $\bigcirc\!\!\!\to$\Knot...$\bigcirc\!\!\!\to$ ~with all its {\it inputs\/} (connections from the direct predecessors) can be replaced by a single knot $\bigcirc\!\!\!\to$~with all inputs from outside the chain.\par\noindent {\bf Block (2):~} A cycle can be replaced by a fixed point $\bigcirc$ with all inputs from outside the cycle.\par\noindent {\bf Block (3):~} A knot $\bigcirc\!\!\!\to$~with no input at all can be removed.\par\noindent {\bf Block (4):~} A branch $B(n)$ (knot $n$ $\bigcirc\!\!\!\to$~together with all its predecessors, see also Section~\ref{Zwei}) can be taken as a subsystem converging to knot $n$ and can completely be removed since the branch has only the single root knot $n$ (delegate). But attention! Knot $n$ itself must not be a fixed point, a cycle member, or infinite. Otherwise, the due (possibly questionable) subsystem would disappear and could thus cause a wrong proof of convergence. If $n'>n$, it can sometimes be sufficient to interchange the knot numbers $n$ and $n'$ only instead of removing the whole branch $B(n)$.\par\noindent {\bf Block (5):~} Knot numbers, names, or other knot identifiers can arbitrarily be interchanged (by permutation), replaced or removed without any modifying the structure\par\noindent {\bf Block (6):~} If two trajectories are connected at some knot, then an arbitrarily other knot of both trajectories can instead be chosen for the connection. A short example of using the building blocks is the equivalent transformation of CC from Equation~(\ref{OneA}) to Equation (\ref{OneB}). Blocks (1) and (3) are used to exclude values $n$ divisible by 2 and~3, respectively. For a more elaborate transformation, see CC in Section~\ref{Sieben}. Let now again an (in)finite set $\mathbb V$ be given, the elements of which are all the admitted values $n$ of a dynamic system generated by the unique function $f(n)$. The values $n$ shall serve for uniquely identifying the knots of the system graph. If a person~A needs all $n\in\mathbb V$ for the identification, then a person~B also needs all $n\in\mathbb V$ provided that these elements $n$ differ from one another and each $n$ is used once and only once. Naturally, the distributions of the values $n$ to the knots by person~A and by person~B differ in general by a permutation. Consider again a system as just described with an (in)finite number of knots. Let every knot be uniquely identified by a value $n\in\mathbb V$ and have an own given (in)finite number $\eta(n)\ge0$ or $\eta(n)=\infty$ of direct predecessors forming a set $\mathbb U_n$. All these $\mathbb U_n$ are disjoint. This fact allows altogether a unique serial numbering of all knots since it is well-known that the union of (in)finitely many (in)finite, disjoint sets $\mathbb U_n$ becomes a single (in)finite set $\mathbb U=\bigcup_n^{\eta(n)}\mathbb U_n$. It is stated again that all the admitted $n\in\mathbb V$ have to be different and used once and only once for uniquely identifying the knots. Although knots are usually numbered by iterative application of the unique generating function $n'=f(n)$, there is also another, possibly more practical way to generate numbers for the knots of a conjectured convergent structure. This way is an iterative application of the reverse generating function $n=f^\bullet(n')$ with start at the already known fixed point $n'=\widetilde n$. The reverse function is not unique. Every $n'$ can have an own (in)finite number of results $n$, the direct predecessors of $n'$. These different results $n$ should also be identified themselves by a suitable new parameter $\mu$. The described way identifies all knots of the convergent subsystem with $\widetilde n$ as its root. If all admitted $n\in\mathbb V$ are needed, then no $n$ at all remains for other, disconnected subsystems and, thus, the entire system is proven to converge. \section{Convergence criteria} \label{Vier} After laying the foundations in Sections~\ref{Eins} to \ref{Drei}, six convergence criteria can now easily be constructed and understood. In every system, all the admitted $n\in\mathbb V$ have to be used once and only once for a unique knot identification. The criteria are mainly based on the idea of strictly separating structure from knot identification. For applications of the criteria, see Sections~\ref{Fuenf} to~\ref{Sieben}. The criteria are to be used to test the convergence of a dynamic system to the fixed point $\widetilde n$. The criteria are sufficient. This means that in case of a negative test result, one cannot decide whether or not the system converges. Before a criterion is applied, the system should be suitably reduced, e.g., by using the transformation building blocks of Section~\ref{Drei}. The following criteria together with system transformations can turn out to be related to one another, e.g., Criteria 3 and 4. It is also stressed again that the system to be tested for convergence must have one single fixed point only. \medskip\noindent{\bf Criterion 1:~} If two dynamic systems A and B are isomorphic (have the same structure) and system B is known to converge, then system A also converges.\par\noindent --- Comment: Structure alone determines roots (see Sections~\ref{Zwei} and \ref{DreiA}). This convergence Criterion 1 is quite simple and clear, most basic, very general, and does not need any knot identification. But a convergent system~B must be available for the structure comparison. The criterion could even be more generalized for other structural system properties, e.g., cycles or divergences. The systems~A and B are isomorphic if there is a complete one-to-one correspondence between them, concerning all knots (unidentified), connections, roots, trajectories. If A and B are not isomorphic, then they can possibly equivalently transformed to become isomorphic, e.g., if the structure of A (or~B) contains that of B (or~A) with corresponding fixed points. In particular, this applies to the case of a self-similar A where the entire A and every branch are isomorphic. The branch converges and can act as~B. The best way to test A and B for being isomorphic seems trying to generate the graph of A step by step with the reverse function $n=f^\bullet(n')$ on the known graph of B from the known fixed points $\widetilde n$ of A and B up to all predecessors of every knot $n'$ of A. Then, system A converges if all $n\in\mathbb V$ are needed. (See also Section~\ref{Drei} and Criterion 2.) \medskip\noindent{\bf Criterion 2:~} A system converges if all its knots are (in)directly connected with one another and need all $n\in\mathbb V$ for a unique identification.\par\noindent --- Comment: No disconnected knots remain for trajectories to roots besides the single fixed point (For application, see the last paragraph of Section~\ref{Drei} and also Criterion 1). \medskip\noindent{\bf Criterion 3:~} An infinite system converges if its structure is self-similar. \par\noindent --- Comment: Systems which are finite or have roots besides the single fixed point are never self-similar (see end of Section~\ref{Zwei}). \medskip\noindent{\bf Criterion 4:~} An infinite system converges if all knots have the same number $\eta>0$ or $\eta=\infty$ of direct predecessors.\par\noindent --- Comment: The system is self-similar (see Criterion~3). Moreover, systems with the same $\eta$ of all knots are always isomorphic. \medskip\noindent{\bf Criterion 5:~} A system converges if it completely consists of branches with all their delegates (in)directly connected with one another.\par\noindent --- Comment: The branches can successively be removed without influencing the convergence to the fixed point (see Section~\ref{Drei} Blocks (1),(3),(4)). \medskip\noindent{\bf Criterion 6:~} A system converges to the single fixed point $\widetilde n=\min n\in\mathbb V$ if every knot $n>\widetilde n$ has a successor knot $m<n$.\par\noindent --- Comment: This is the old criterion (see Section~\ref{Eins}). The old criterion can only be applied if every case $n'=f(n)>n$ can be removed by a suitable equivalent system transformation, e.g., by removing the branch $B(n)$ or by interchanging the knot numbers $n$ and $n'$ (only these numbers, never the knots themselves with their directed connection!). Cases $n'>n$ of cycles or divergent trajectories can never be removed at all. (See also Section~\ref{Drei} Blocks (4) and (5).) But such cases need no attention. This is an advantage of the old criterion. A shortcoming seems to be that structure is not taken into account. \section{A simple dynamic system} \label{Fuenf} The convergence criteria are first applied to the simple (but not trivial) dynamic system with the following generating function $f_\text{S}(n)$ and its reverse $f_\text{S}^\bullet(n')$: \begin{equation} n'=f_\text{S}(n)=\begin{cases} ~n\enskip, & \text{if $n=1$~;}\\ ~(n-1)/2\enskip, & \text{if $n>1$ is odd~;}\\ ~n/2\enskip, & \text{if $n$ is even~;} \end{cases} \label{SimA} \end{equation} \begin{equation} n=f_\text{S}^\bullet(n')=2n'\enskip \text{and~} =2n'+1 \label{SimB} \end{equation} with all values $n,n'>0$ admitted and forming the set $\mathbb V=\mathbb N^+$. Since always $n'<n$, except $n'=n=\widetilde n=1$ for the fixed point $\widetilde n$, the system is connected and converges to $\widetilde n$ according to the old Criterion 6. Every $n'$ has $\eta=2$ direct predecessors. This fact makes the system structure self-similar. Accordingly, the system also turns out to converge by using Criteria 3 or 4. For a test of convergence to $\widetilde n$, the knots of a system graph can best also be generated and numbered by $n=f^\bullet(n')$ reversely from $\widetilde n$. If all $n\in\mathbb V$ are needed once and only once, then the system converges since no values $n$ remain for disconnected subsystems of other roots. Indeed, $f_\text{S}^\bullet(n')$ generates $\eta=2$ knots $n$ for every $n'$ already given. Every $n$ is needed once and only once and obtains an own trajectory $T(n)$ different from others. See reversely, e.g., $1\Ra2,3\Ra4,5,6,7\Rightarrow ...\Ra2^\nu,...,2^{\nu+1}\!-\!1 \Rightarrow ...$ ($\Rightarrow$ denotes a bundle of reverse connections from ... to ... by $f^\bullet$). No $n$ is lost and knots with the same number do not exist. The system thus converges. \section{Families of convergent dynamic systems} \label{Sechs} The application of the most general convergence Criterion 1 to a system A requires an available, possibly isomorphic, convergent system B. Such a system B can easily be constructed because every integer $n>1$ can uniquely be represented by multiplied primes or, similarly, by a sum of powers $2^\nu$ with exponents $\nu\in\mathbb N$, or by proceeding one of many other ways. In particular, the generating function $f_\text{P}(n)$ of a large family of isomorphic systems on multiplied primes (MP) and their reverse function $f_\text{P}^\bullet(n')$ read \begin{equation} n'=f_\text{P}(n)=n/p(n) \En;\quad n=f_\text{P}^\bullet(n')=n'p \En;\quad (\enskip p, p(n)>3\enskip) \label{Prim} \end{equation} where $p$ and $p(n)$ are arbitrarily chosen prime factors (of $n$ for $p(n)$). The particular choice $p, p(n)>3$ of Equation~(\ref{Prim}) is suitable only for use in Section~\ref{Sieben}. The systems of the family differ only by the sequences of chosen primes used for generating trajectories, that is, by knot numbering. Accordingly, the systems are isomorphic (all have the same structure). Since always $n'<n$ and by application of Criterion 6, all the systems converge to the only fixed point $\widetilde n=1$. A more detailed example follows. The knot numbers $n$ of the family graphs are obtained by $n=f_\text{P}^\bullet(n')$ with an infinity $\eta=\infty$ of direct predecessors $n$ of every given $n'$, beginning with $n'=\widetilde n=1$. For example with all $p>3$, the direct predecessors of the fixed point $n'=\widetilde n=1$ are all single primes $n=p_1=5,7,11,...$ in arbitrary order; those ones of $n'=5$ are all the products $n=p_1p_2=25,35,...$ of two primes, also in arbitrary order, and so on. Every direct predecessor $n$ of $n'$ has one prime factor more than $n'$ and is indivisible by 2 and 3 so that $n=6k\pm1>1$ for every integer $k>0$ (see also CC in Section~\ref{Sieben}). All these $n$ and $\widetilde n=1$ form the infinite set $\mathbb V$ of admitted values. Every $n$ has to be used as a knot number, but only once for a unique knot numbering. Indeed, all direct predecessors $n$ of some given $n'$ differ from one another due to different prime factors $p$. They also differ from all other values $n$ because of a different number of prime factors. And every knot $n$ can be arrived by a reverse trajectory from $\widetilde n=1$ successively multiplied with all prime factors of $n$. Quite similarly, the generating function $f_2(n)$ of another large system family on sums of powers $2^\nu$ and the reverse $f_2^\bullet(n')$ read \begin{equation} n'=f_2(n)=n-2^\nu \En;\quad n=f_2^\bullet(n')=n'+2^\mu \label{Sum} \end{equation} where $\nu$ is a suitably chosen exponent and $n>\widetilde n=0$. All systems converge according to Criterion 6 since always $n'<n$. Again, every knot $n$ can be arrived by a reverse trajectory from $\widetilde n=0$ successively added with all powers $2^\mu$ of $n$. All systems of both and related other families are self-similar since $\eta=\infty$ of every knot. They also converge according to Criteria 3 or~4. Moreover, the systems are isomorphic, i.e., all have the same structure independent from any unique knot numbering. \section{Collatz dynamic system and conjecture} \label{Sieben} The convergence criteria of Section~\ref{Vier} together with the graphic properties, tools and transformations of Sections~\ref{Zwei} and~\ref{Drei} are now applied to easily prove the convergence of a suitably reduced Collatz system for confirming the Collatz conjecture (CC, see Section~\ref{Eins}). The original CC system graph is first generated by $n'=f_\text{C}(n)$ of Equation~(\ref{OneA}) on all admitted knot numbers $n,n'>0$. The graph has a cycle $T(1)=(1,4,2,1,\ldots)$ with $n=1$ conjectured to be also a value of every trajectory $T(n)$. The graph is then reduced suitably and equivalently with respect to the CC conjecture by using the building blocks of Section~\ref{Drei} for transformations until a criterion can work. {\bf 1st reduction:~} According to Block~(1), every chain of knots, generated by some even $m$ divided $\nu$ times repeatedly by 2, is replaced by the single knot with odd $m/2^\nu$. This transformation removes all knots with even $n$ and leads to $f_\text{C}(n)$ in Equation~(\ref{OneB}) with $\nu>0$. The cycle disappears, but its member knot $n=1$ remains as a conjectured fixed point $\widetilde n=1$ of every trajectory. This fixed point $\widetilde n=1$ together with $\nu=2$ is the only solution of $\widetilde n=f_\text{C}(\widetilde n)=(3\widetilde n+1)/2^\nu$ transformed to $(2^\nu-3)\widetilde n=1$. {\bf 2nd reduction:~} Every knot with $n$ divisible by 3 has no predecessor $m$ and can thus also be removed according to Block~(3). No $n=(3m+1)/2^\nu$ is divisible by 3. {\bf 3rd reduction:~} Block~(4) allows to also remove branches $B(n)$, e.g., if $n'>n$ obtained from $n'=(3n+1)/2^\nu>n$ for $\nu=1$ only (see Equation~(\ref{OneB})). Knot $n'$ and all its direct predecessors $n$ for $\nu>1$ remain with $n'<n$ since they do not belong to $B(n)$. Notice that the reductions remove some at first admitted $n$ and change the original cycle to a fixed point, but do not remove any possibly existing roots. Similarly, this applies also to the branches. The generating function $f_\text{C}(n)$ of the three times reduced CC system graph and its reverse $f_\text{C}^\bullet(n')$ now read \begin{equation} n'=f_\text{C}(n)=(3n+1)/2^\nu \En;\quad n=f_\text{C}^\bullet(n')= (2^\mu n'-1)/3 \label{Coll} \end{equation} with admitted $n$ and $n'$ indivisible by 2 and 3 (or $n,n'=6k\pm1>0$, see below) and with the single fixed point $\widetilde n=1$. The parameter $\nu>1$ (in contrast to $\nu>0$ in Equation~(\ref{OneB})!) serves for always $n'<n$ and makes the old Criterion 6 thus applicable. Reversely, the parameter $\mu>1$ uniquely identifies all the infinite number $\eta=\infty$ of admitted results (direct predecessors) $n=f_\text{C}^\bullet(n')$ for every admitted $n'$ given. Although $\eta=\infty$, by far not all values of $\nu$ and $\mu$ have be used in Equation~(\ref{Coll}) to connect admitted $n$ and $n'$. But this fact does not matter. The CC system is now already proven to converge to the fixed point $\widetilde n=1$ according to the old Criterion 6 since always $n'<n$. It also converges according to Criteria 3 and 4 since it is self-similar because of $\eta=\infty$. See also the self-similar particular structure chosen in Section~\ref{Zwei} and the MP and other structures discussed in Section~\ref{Sechs}. All these structures and the Collatz one are isomorphic to one another. They are taken as B and A, respectively, for Criterion 1. Each of the results here obtained already confirms the Collatz conjecture. Nevertheless, someone may remain unsatisfied possibly by lack of numeric calculations similar to those vain ones in the past. More elaborate arguments shall convince him. Read paragraph ``The knot numbers ..." of Section~\ref{Sechs} and choose any MP system according to Equation~(\ref{Prim}) with all the prime factors $p,p(n)>3$. These factors are needed to generate the entire convergent MP graph reversely by $n=f_\text{P}^\bullet(n')$ from the fixed point $\widetilde n=1$ to all knots uniquely numbered by using the set $\mathbb V$ with all the admitted $n,n',\widetilde n$ indivisible by 2~and~3. Every $n'$ turns out to have $\eta=\infty$ direct predecessor knots. The entire CC graph is now reversely generated by $n=f_\text{C}^\bullet(n')$ strictly in parallel to MP with the same set $\mathbb V$. The same result $\eta=\infty$ of MP and CC then leads with Criteria 3 and 4 to the same self-similar and convergent structures. Criterion 1 confirms the convergence of CC as system A and MP as system B. These systems are isomorphic, the knot number distributions differ only by a permutation. This fact becomes clear since every knot number of CC can as well be uniquely expressed by a knot number of MP and vice versa. This one-to-one correspondence between $q=6k\pm1>0$ of CC and ``${\it\Pi}>0$ indivisible by 2 and 3" of MP is (as required by a criticism) shown as follows. Every $q$ is obviously indivisible by 2 and 3. But can every ${\it\Pi}$ also uniquely be represented by $q$? Here, ${\it\Pi}$ is a product of prime factors $p>3$. Then, ${\it\Pi}=q=6k\pm1>0$ must apply for every ${\it\Pi}$, that is, the unknown $k$ must have a unique solution. With other words, either only ${\it\Pi}+1$ or only ${\it\Pi}-1$ must be divisible by 6. Both are already even, divisible by 2. Every triple of succeeding integers $a-1, a, a+1$ has always one integer divisible by 3. But $a\ne{\it\Pi}$ since ${\it\Pi}$ is already indivisible by 3. Thus, either only ${\it\Pi}-1$ or only ${\it\Pi}+1$ is divisible by 3 and 2. The division by 6 then leads uniquely to $k$. Accordingly, CC can be transformed to MP without modifying the structure and vice versa. CC and MP are isomorphic. This fact can also be stated more easily since every knot $n'$ has an infinity $\eta=\infty$ of direct predecessors $n$. This results for CC from $n=f_\text{C}^\bullet(n')=(2^\nu n'-1)/3$ for an infinity of values $\nu$ and all $n$ and $n'$ indivisible by 2 and 3; for MP from $n=f_\text{P}^\bullet(n')=n'p$ for an infinity of primes $p>3$ and all $n$ and $n'$. But these values do not matter at all. They only show a one-to-one correspondence between the knots of CC and MP. Much more easily, if each of all knots of two systems A and B has the same number $\eta>0$ or $\eta=\infty$ of direct predecessors, then A and B are isomorphic. \section{Concluding remarks} \label{Acht} As the main result of the present study, sufficient criteria are established for testing whether a dynamic system on positive integers is connected and converges to a fixed point. The criteria are based on a quite simple idea. Let, e.g., the structure of a dynamic system in question be represented by an infinite, directed graph of identical knots $\bigcirc\!\!\!\to$, each with a single connection pointer to only one following knot. If this structure is self-similar or the same as that of another system already known to be convergent, then cycles and divergent trajectories do not exist and the system converges to just a single fixed point. Every knot identification, e.g., by numbers does not influence the system structure. The criteria thus are general enough for application to a large variety of related systems and problems. In particular, the Collatz conjecture is easily confirmed. The criteria also allow to avoid knot numbers at all to easily overcome the high barrier between very many logical steps and an infinity of them. One may ask why the properties of the huge number of previously investigated individual trajectories of the Collatz dynamic system (see Wikipedia \cite{Wik} or P\"oppe \cite{Ppp}) were not taken into account. The reason is that the present approach only considers the structure and one characteristic problem of the system in its entirety, namely, its convergence to the fixed point~1. Consider, e.g., the related system generated by the function $f(n)=(n+1)/2^\nu$ ($n$ and $f$ are odd). Here, every trajectory is proven to converge to the fixed point~1, easily and merely by the general and inherent property $n'=f(n)<n$ of $f(n)$ itself $(n>1)$, but not by taking into account infinities of trajectories. Many experienced mathematicians on number theory, inspired laymen, and, previously also this author, tried in vain to prove the Collatz conjecture. Why did they not succeed? Possibly, they followed a mainstream approach in which too much attention was paid to all of the trajectories (e.g., by relying on computer experiments) rather than paying attention to the problem in its entirety, say, its infinite structure. Similarly, Feinstein \cite{Fei} may be right in that an infinity of program lines or computing time is needed to test whether or not all the individual trajectories converge to~1. However, to solve a problem it may be sufficient to investigate an essential property, such as the system structure or a permanent increase of entropy (or loss of information), which allows for a completely different approach that avoids difficulties. One example is the irrational number $\pi$ that can never be determined exactly by numerical calculations, but many essential facts about $\pi$ are already known and always used. The human brain can think better than a computer about abstractions, continuities, infinities and irrationals. Therefore, there are not necessarily contradictions between Feinstein's proof \cite{Fei}, Opfer's proof \cite{Opf} and the present approach. Problems can have several solutions, impractical ones needing an infinity of logical steps or others requiring a large or short finite number of them. See, for instance, $\Sigma_{i=0}^\infty x^i=1/(1-x)$ with $|x|<1$. Then, one can say that the present effort to confirm the Collatz conjecture is short compared with other approaches and could finally finish the research of more than 80 years. \section{Acknowledgments} \label{Neun} The present paper describes results of the author's private theoretical research. Some applied concepts, facts, and methods are well-known in mathematics, especially in number theory of dynamic systems. They are combined or used as tools to establish new, simple convergence criteria for dynamic systems. The author would like to express his thanks to the Physikalisch-Technische Bundes\-anstalt (PTB, Federal Institute of Physics and Technology, Braunschweig, Germany) for years of support, to Prof. Dr. L. Baringhaus (retired from Leibniz University of Hannover, Germany), Dr. M. Matzke (like the author retired from PTB), and Dr. M. Reginatto (PTB) for suggestions and criticisms concerning the convergence problem and also for their patience with the author's several vain trials to solve the problem. He also thanks Dr. M. Matzke, Prof. Dr. R. Michel (also retired from Leibniz University) and Dr. W. W\"oger (also retired from PTB) for decades of successful collaboration in theoretical physics of ionizing radiation and in general measurement, Prof. Dr. L. Baringhaus for valuable advice in stochastics, and P. Harris and the author's granddaughter S. Weise for linguistically inspecting the English text. \section{A motto for dynamic systems} \label{Zehn} The following poem, ``The Roman Fountain" by Conrad Ferdinand Meyer (Swiss novelist and epic poet, 1825--1898), could be a motto for the dynamic systems, each appearing as a fountain of integers. \centerline{\bf Der r\"omische Brunnen} {\obeylines\everypar{\hfil}\parindent=0pt Auf steigt der Strahl und fallend gie\ss t Er voll der Marmorschale Rund, Die, sich verschleiernd, \"uberflie\ss t In einer zweiten Schale Grund; Die zweite gibt, sie wird zu reich, Der dritten wallend ihre Flut, Und jede nimmt und gibt zugleich Und str\"omt und ruht.}

https://arxiv.org/abs/1210.6853

On solving large scale polynomial convex problems by randomized first-order algorithms

One of the most attractive recent approaches to processing well-structured large-scale convex optimization problems is based on smooth convex-concave saddle point reformu-lation of the problem of interest and solving the resulting problem by a fast First Order saddle point method utilizing smoothness of the saddle point cost function. In this paper, we demonstrate that when the saddle point cost function is polynomial, the precise gra-dients of the cost function required by deterministic First Order saddle point algorithms and becoming prohibitively computationally expensive in the extremely large-scale case, can be replaced with incomparably cheaper computationally unbiased random estimates of the gradients. We show that for large-scale problems with favourable geometry, this randomization accelerates, progressively as the sizes of the problem grow, the solution process. This extends significantly previous results on acceleration by randomization, which, to the best of our knowledge, dealt solely with bilinear saddle point problems. We illustrate our theoretical findings by instructive and encouraging numerical experiments.

\section{Introduction} The goal of this paper is to develop {\sl randomized} First Order algorithms for solving large-scale ``well structured'' convex-concave saddle point problems. The background and motivation for our work can be briefly outlined as follows. Theoretically, all of convex programming is within the grasp of polynomial time Interior Point Methods capable of generating high-accuracy solutions at a low iteration count. However, the complexity of an IPM iteration, in general, grows rapidly (as $n^3$) with the design dimension $n$ of the problem, which in numerous applications (like LP's with dense constraint matrices arising in Signal Processing) make IPM's prohibitively time-consuming in the large-scale case. There seemingly is consensus that ``beyond the practical grasp of IPM's,'' one should use the First Order Methods (FOM's) which, under favorable circumstances, allow to get medium-accuracy solutions in (nearly) dimension-independent number of relatively cheap iterations. Over the last decade, there was a significant progress in FOM's; to the best of our understanding, the key to this progress is in discovering a way (Nesterov 2003, see \cite{smoothing}) to utilize problem's structure in order to accelerate FOM algorithms, specifically, to reduce a convex minimization problem $\min_{x\in X} f(x)$ with potentially nonsmooth objective $f$ to a saddle point problem $$ \min_{x\in X}\max_{\sl y\in Y}\phi(x,y),\eqno{(SP)} $$ where $\phi$ is a C$^{1,1}$ convex-concave function such that \begin{equation}\label{ieq1} f(x)=\max_{y\in Y}\phi(x,y). \end{equation} The rationale is as follows: when $f$ is nonsmooth (which indeed is the case in typical applications), the (unimprovable in the large-scale case) rate of convergence of FOM's directly applied to the problem of interest $\min_{x\in X} f(x)$ is as low as $O(1/\sqrt{t})$, so that finding a feasible $\epsilon$-optimal solution takes as much as $O(1/\epsilon^2)$ iterations. Utilizing representation (\ref{ieq1}), this rate can be improved to $O(1/t)$; when $X$, $Y$ are simple, this dramatic acceleration keeps the iteration's complexity basically intact. \par Now, in the original {\sl Nesterov's Smoothing} \cite{smoothing}, (\ref{ieq1}) is used to approximate $f$ by a C$^{1,1}$ function which is further minimized by Nesterov's optimal algorithm for smooth convex minimization originating from \cite{Yura83}. An alternative is to work on $(SP)$ ``as it is,'' by applying to $(SP)$ an $O(1/t)$-converging saddle point FOM, like the Mirror Prox algorithm \cite{DMP}; in what follows, we further develop this alternative. \par When solving $(SP)$ by a FOM, the computational effort per iteration has two components: (a) computing the values of $\nabla\phi$ at $O(1)$ points from $Z=X\times Y$, and (b) ``computational overhead,'' like projecting onto $Z$. Depending on problem's structure and sizes, any one of these two components can become dominating; the approach we are developing in this paper is aimed at the situation where the computational ``expenses'' related to (a) by far dominate those related to (b), so that the ``practical grasp'' of the usual -- deterministic -- saddle point FOMs as applied to $(SP)$ is restricted with the problems where the required number of computations of $\nabla\phi$ (which usually is in the range of hundreds) can be carried out in a reasonable time. An attractive way to lift, to some extent, these restrictions is to pass from the precise values of $\nabla\phi$, which can be prohibitively costly computationally in the large-scale case, to computationally cheap unbiased {\sl random estimates} of these values. This idea (in retrospect, originating from the ad hoc sublinear type matrix game algorithm of Grigoriadis and Khachiyan \cite{GriKha}) has been developed in several papers, see \cite{Arora,NJLS,SMP,Baes,Rand}, \cite[section 6.5.2]{FOM} and references therein. To the best of our knowledge, for the time being ``acceleration via randomization'' was developed solely for the case of saddle point problems with {\sl bilinear} cost function $\phi$. The contribution of this paper is in extending the scope of randomization to the case of when $\phi$ is a {\sl polynomial}. \par The main body of this paper is organized as follows. In section \ref{sect:SitGoals}, we formulate the problem of interest and present the necessary background on our ``working horse'' --- Mirror Prox algorithm. In section \ref{sect:rand}, we develop a general randomization scheme aimed at producing unbiased random estimates of $\nabla \phi$ for a polynomial $\phi$. Theoretical efficiency estimates for the resulting randomized saddle point algorithm are derived in section \ref{sect:complanal}. In section \ref{sectIll}, we illustrate our approach by working out in full details two generic examples: optimizing the maximal eigenvalue of a quadratic matrix pencil, and low dimensional approximation of a finite collection of points. We show theoretically (and illustrate by numerical examples) that in both these cases, in a meaningful range of problem's sizes and $\epsilon$, solving problem within accuracy $\epsilon$ by randomized algorithm is by far less demanding computationally than achieving the same goal with the best known to us deterministic competitors, and the resulting ``acceleration by randomization'' goes to $\infty$ as the problem sizes grow. \section{Situation and Goals}\label{sect:SitGoals} \subsection{Problem Statement}\label{sec:POI} Consider the situation as follows: let $X\subset E_x,y\subset E_y$ be convex compact subsets of Euclidean spaces, and let $\phi(x,y): E:=E_x\times E_y\to\bR$ be a polynomial of degree $d$: \begin{equation}\label{phinew} \phi(\underbrace{x,y}_{z})=\sum_{k=0}^dQ_k(\underbrace{z,...,z}_{k}), \end{equation} where $Q_0$ is a constant, and for $k>0$, $Q_k(z^1,...,z^k)$ is a $k$-linear symmetric form on $E$ . From now on we assume that {\sl $\phi(x,y)$ is convex-concave on $X\times Y$}, that is, convex in $x\in X$ for fixed $y\in Y$, and concave in $y\in Y$ for fixed $x\in X$. Our problem of interest is the saddle point problem \begin{equation}\label{SPP} \SadVal=\min_{x\in X}\max_{y\in Y}\phi(x,y). \end{equation} Let \begin{equation}\label{PD} \begin{array}{rlcr} \Opt(P)&=&\min\limits_{x\in X}\left[\overline{\phi}(x):= \max_{y\in Y}\phi(x,y)\right]&(P)\\ \Opt(D)&=&\max\limits_{y\in Y}\left[\underline{\phi}(y):= \min_{x\in X}\phi(x,y)\right]&(D)\\ \end{array} \end{equation} be the primal-dual pair of convex programs associated with (\ref{SPP}), so that $\Opt(P)=\Opt(D)$, let \begin{equation}\label{DualGap} \DualityGap(x,y)=[\overline{\phi}(x)-\Opt(P)]+[\Opt(D)-\underline{\phi}(y)]=\overline{\phi}(x)-\underline{\phi}(y) \end{equation} be the associated duality gap, and, finally, let \begin{equation}\label{F} F(z:=[x;y])=\left[F_x(x,y)=\phi^\prime_x(x,y);F_y(x,y):=-\phi^\prime_{y}(x,y)\right]:Z:=X\times Y\to E:=E_x\times E_y \end{equation} be the monotone mapping associated with (\ref{SPP}). Our ideal goal is, given tolerance $\epsilon>0$, to find an {\sl $\epsilon$-solution} to (\ref{SPP}), i.s., a point $z_\epsilon=(x_\epsilon,y_\epsilon)\in Z$ such that \begin{equation}\label{goal} \DualityGap(z_\epsilon)\leq\epsilon, \end{equation} whence $x_\epsilon$ is a feasible $\epsilon$-optimal solution to $(P)$, while $y_\epsilon$ is a feasible $\epsilon$-optimal solution to $(D)$. We intend to achieve this goal by utilizing {\sl randomized} First Order saddle point algorithm, specifically, Stochastic Mirror Prox method (SMP) \cite{SMP}. \subsection{Background on Stochastic Mirror Prox algorithm}\label{sec:SMP} {\sl A setup} for SMP as applied to (\ref{SPP}) is given by \begin{itemize} \item a norm $\|\cdot\|$ on the subspace $$L[Z]:=\Lin(Z-Z)$$ in the embedding space $E:=E_x\times E_y$ of the domain $Z:=X\times Y$ of the saddle point problem. The (semi)norm on $E$ conjugate to $\|\cdot\|$ is denoted by $\|\cdot\|_*$: $$ \|\zeta\|_*=\max\limits_{z\in L[Z]}\{\langle \zeta,z\rangle: \|z\|\leq1\}; $$ \item a {\sl distance-generating function} (d.g.-f.) $\omega(z):Z\to\bR$ which should be convex and continuously differentiable on $Z$ and should be {\sl compatible} with $\|\cdot\|$, meaning strong convexity of $\omega(\cdot)$, modulus 1, w.r.t. $\|\cdot\|$: $$\langle \omega'(z)-\omega'(z'),z-z'\rangle \geq \|z-z'\|^2\,\,\forall (z,z'\in Z).$$ \end{itemize} An SMP setup induces several important entities, specifically \begin{itemize} \item $\omega$-center $z_\omega:=\argmin_{z\in Z}\omega(z)$ of $Z$; \item {\sl Bregman distance} $V_z(w):=\omega(w)-\omega(z)-\langle \omega'(z),w-z\rangle$, where $z,w\in Z$. By strong convexity of $\omega$, we have $V_z(w)\geq{1\over 2}\|w-z\|^2$; \item {\sl $\omega$-radius} $\Omega:=\sqrt{2[\max_Z\omega(\cdot)-\min_Z\omega(\cdot)]}$; noting that ${1\over 2}\|w-z_\omega\|^2\leq V_{z_\omega}(w)\leq \omega(w)-\omega(z_\omega)$, we conclude that \begin{equation}\label{raduis} \forall (w\in Z): \|w-z\|\leq\Omega; \end{equation} \item {\sl Prox-mapping} $\Prox_z(\xi)$, $z\in Z$, $\xi\in E$, defined as $$ \Prox_z(\xi)=\argmin_{w\in Z}\left[\langle \xi,w\rangle +V_z(w)\right]=\argmin_{w\in Z}\left[\langle \xi-\omega'(z),w\rangle +\omega(w)\right] $$ \end{itemize} As applied to (\ref{SPP}), SMP operates with {\sl Stochastic Oracle} representation of the vector field $F$ associated with the problem. A {\sl Stochastic Oracle} is a procedure (``black box'') which, at $t$-th call, a point $z_t$ being the input, returns the random vector $$ g(z_t,\xi_t)=F(z_t)+\Delta(z_t,\xi_t)\in E $$ where $\Delta(\cdot,\cdot)$ is a deterministic function, and $\xi_1,\xi_2,...$ is a sequence of i.i.d. ``oracle noises.'' The SMP algorithm is the recurrence \begin{equation}\label{SMP} \begin{array}{lrcl}\\ \hbox{initialization:}&z_1&=&z_\omega;\\ \hbox{search points:}&z_t&\mapsto&w_t=\Prox_{z_t}(\gamma_t g(z_t,\xi_{2t-1}))\mapsto z_{t+1}=\Prox_{z_t}(\gamma_t g(w_t,\xi_{2t}))\mapsto...\\ \hbox{approximate solutions:}&z^t&=&(x^t,y^t)=[\sum_{\tau=1}^t\gamma_\tau]^{-1}\sum_{\tau=1}^t\gamma_\tau w_\tau\\ \end{array} \end{equation} where $\gamma_t>0$ are deterministic stepsizes.\par The main results on SMP we need are as follows (see the case $M=\mu=0$ of \cite[Corollary 1]{SMP}): \begin{theorem}\label{theSMP} Assume that $\cL<\infty$ and $\sigma<\infty$ are such that \begin{equation}\label{aresuch} \begin{array}{ll} (a)&\|F(z)-F(z')\|_*\leq \cL\|z-z'\|\,\,\forall z,z'\in Z\\ (b)&\bE_\xi\{\Delta(z,\xi)\}=0\,\,\forall z\in Z\\ (c)&\bE_\xi\{\|\Delta(z,\xi)\|_*^2\}\leq\sigma^2\,\,\forall z\in Z\\ \end{array} \end{equation} Then for every $t=1,2,...$ the $t$-step SMP with constant stepsizes \begin{equation}\label{stepsizes} \gamma_\tau=\min\left[{1\over\sqrt{3}\cL}, {\Omega\over\sqrt{7}\sigma\sqrt{t}}\right],\,1\leq\tau\leq t \end{equation} ensures that \begin{equation}\label{mean} \bE\{\DualityGap(x^t,y^t)\}\leq K_t:=\max\left[{2\Omega^2\cL\over t},{6\Omega\sigma\over\sqrt{t}}\right]. \end{equation} In addition, strengthening {\rm (\ref{aresuch}.$b$,$c$)} to \begin{equation}\label{strengthening} \bE_\xi\{\Delta(z,\xi)\}=0,\,\bE\{\exp\{\|\Delta(z,\xi)\|_*^2/\sigma^2\}\}\leq\exp\{1\} \end{equation} we have an exponential bound on large deviations: for every $\Lambda>0$, we have \begin{equation}\label{largedev} \Prob\left\{\DualityGap(x^t,y^t)> K_t +\Lambda {7\Omega\sigma\over 2\sqrt{t}}\right\}\leq \exp\{-\Lambda^2/3\}+\exp\{-\Lambda t\}. \end{equation} \end{theorem} \section{Randomization}\label{sect:rand} Problem (\ref{SPP}) by itself is a fully deterministic problem; with ``normal'' representation of the polynomial $\phi(x,y)$ (e.g., by the list of its nonzero coefficients), a precise ($\sigma=0$) deterministic oracle for $F$ is available; utilizing this oracle, a solution of accuracy $\epsilon$ is obtained in $O(1)\Omega^2\cL/\epsilon$ iterations, with computational effort per iteration dominated by the necessity to compute the values of $F$ at two points and the values of two prox-mappings. When $Z$ is ``simple enough,'' the complexity of the second of these two tasks -- computing prox-mappings -- is a tiny fraction of the complexity of precise computation of the values of $F$. Whenever this is the case, it {\sl might} make sense to replace the precise values $F$ (which can be very costly in the large-scale case) with computationally cheap unbiased random estimates of these values. This is the option we intend to investigate in this paper. We start with a general description of the randomization we intend to use. \par Observe, first, that $$ F(z)=D \nabla\phi (z) $$ where $D=\Diag\{\Id_x,-\Id_y\}$, $\Id_x$ and $\Id_y$ being the identity mappings on $E_x$ and $E_y$, respectively. Now, representing the polynomial $\phi(z)$ as \begin{equation}\label{phi} \phi(z)=\sum_{k=0}^dQ_k(\underbrace{z,...,z}_{k}), \end{equation} where $Q_k(z^1,...,z^k)$ is a symmetric $k$-linear form on $E$, differentiating (\ref{phi}) in $z$ and taking into account symmetry of $Q_k$, we have \begin{equation}\label{formulaA} \langle F(z),h\rangle =\langle D\nabla\phi(z),h\rangle=\langle \nabla\phi(z),Dh\rangle=\sum_{k=1}^d kQ_k(Dh,\underbrace{z,...,z}_{k-1}) \end{equation} Now assume that we can associate with every $z\in Z$ a probability distribution $P_z$ on $E$ such that \begin{equation}\label{NoBias} \int \xi dP_z(\xi)=z\,\,\forall z\in E. \end{equation} In order to get an unbiased estimate of $F(z)$, one can act as follows: \begin{itemize} \item given $z$, draw $d-1$ independent samples $z^i\sim P_z$, $i=1,...,d-1$ \item compute the linear form $G=G[z^1,...,z^{d-1}]$ on $E$ given by \begin{equation}\label{base} \forall h\in E: \langle G,h\rangle=\sum_{k=1}^d kQ_k(Dh,z^1,z^2,...,z^{k-1}). \end{equation} thus ensuring that \begin{equation}\label{unbiased} \bE_{(z^1,...,z^{d-1})\sim P_z\times...\times P_z}\{G[z^1,...,z^{d-1}]\}=F(z)\,\,\forall z\in Z. \end{equation} \end{itemize} Note that we can represent a random variable distributed according to $P_z$ as a deterministic function of $z$ and random variable $\xi$ uniformly distributed on $[0,1]$, which makes $G$ a deterministic function of $z$ and $\xi\sim\hbox{Uniform}[0,1]$, as required by our model of a Stochastic Oracle. \par Observe that for a general-type convex-concave polynomial $\phi(x,y)$ of degree $d$, precise deterministic computation of $F(z)$ is as suggested by (\ref{base}) {\sl with $P_z$ being the unit mass sitting at the singleton $z$}, that is, with $z^1=...=z^{d-1}=z$. It follows that {\sl if the distributions $P_z$, for every $z\in Z$ are such that computing the vectors $g_k$ of coefficients of the linear forms $Q_k(Dh,z^1,...,z^{k-1})$ of $h\in E$ is much cheaper than the similar task for the linear forms $Q_k(Dh,z,...,z)$ for a ``general position'' $z\in Z$, then computing the unbiased estimate $G=G[z^1,...,z^{d-1}]$ of $F(z)$ is much cheaper computationally than the precise computation of $F(z)$}, so that there are chances for the outlined randomization to reduce the overall complexity of computing $\epsilon$-solution to (\ref{SPP}). Let us look at a simple preliminary example: \paragraph{Example 1} [``Scalar case'']: $E$ is just the space $\bR^n$ of $n$-dimensional vectors, and we have access to the coefficients of the $k$-linear forms $Q_k(\cdot)$ (e.g., $Q_k$ are given by lists of their nonzero coefficients). In this case, we can specify $P_z$ as follows: \par Given $z\in E=\bR^n\backslash \{0\}$, let $P_z$ be the discrete probability distribution supported on the set $\{f^i=\sign(z_i)\|z\|_1e_i\}_{i=1}^n$, where $e^i$ are the standard basic orths in $E$, with the probability mass of $f^i$ equal to $|z_i|/\|z\|_1$; when $z=0$, let $P_z$ be the unit mass sitting at the origin. We clearly have $\bE_{f\sim P_z} \{f\}=z$, and all realizations of $f\sim P_z$ are extremely sparse --- with at most one nonzero entry. Now, in order to generate $f\sim P_z$, we need preprocessing of $O(1)n$ a.o. aimed to compute $\|z\|_1$ and the ``cumulative distribution'' $s_i=\|z\|_1^{-1}\sum_{j=1}^i|z_j|$, $i=1,...,n$. With this cumulative distribution at hand, to draw a sample $f\sim P_z$ takes just $O(1)\ln(n)$ a.o.: we draw at random a real $\alpha$ uniformly distributed in $[0,1]$ (which for all practical purposed is just $O(1)$ a.o.), find by bisection the smallest $i\in\{1,...,n\}$ such that $\alpha\leq s_i$ ($O(1)\ln(n)$ a.o.) and return the index $i$ and the value $\sign(z_i)\|z\|_1$ of the only nonzero entry in the resulting vector $f$ ($O(1)$ a.o.). Thus, generating $z^1,...,z^{d-1}$ costs $O(1)[n+d\ln(n)]$ a.o. Now, with our ``ultimately sparse'' $z^1,...,z^{d-1}$, computing the $n$ coefficients of the linear form $Q_k(Dh,z^1,...,z^{k-1})$ of $h$ takes at most $O(1)[d+n\cC]$ a.o., where $\cC$ is an upper bound on the cost of extracting a coefficient of the $k$-linear symmetric form $Q_k$, $k\leq d$, given its ``address.'' The bottom line is that the complexity of computing $G[z^1,..,z^{d-1}]$ is $$ \cC_r[P] =O(1)\left[n+d\ln(n)+d[d+n\cC]\right]=O(1)[d^2+dn\cC] \hbox{\ a.o.} $$ On the other hand, computing $F(z)$ exactly costs something like $$ \cC_d[P]=O(1)[n+\sum_{k=1}^d kN_k\cC]\hbox{\ a.o.} $$ where $N_k$ is the total number of nonzero coefficients in $Q_k(\cdot,...,\cdot)$. Assuming that $d=O(1)$, we see that {\sl unless all $Q_k$ are pretty sparse -- just with $N_k=O(n)$ nonzero coefficients, mimicking unbiased Stochastic Oracle takes by orders of magnitude less computations than precise deterministic computation of $F(z)$.} \section{Complexity Analysis}\label{sect:complanal} The discussion in the previous section demonstrates that in some interesting cases unbiased random estimates of the vector field $F$ associated with (\ref{SPP}) are significantly cheaper computationally than the precise values of $F$. This does not mean, however, that in all these cases randomization is profitable --- it well may happen that as far as the overall complexity of $\epsilon$-solution is concerned, expensive high-quality local information is better than cheap low quality one. We intend to analyze the situation in the regime when the degree $d$ of the polynomial $\phi$ is a small integer formally treated as $O(1)$; this allows us to ignore in the sequel the details on how the hidden factors in $O(\cdot)$'s to follow depend on $d$. \subsection{Preliminaries} \paragraph{Standing Assumptions.} Observe that \begin{equation}\label{direct1} L[Z]:=\Lin(Z-Z)=\Lin(X-X)\times\Lin(Y-Y)=L[X]\times L[Y]. \end{equation} Now, the sets $$ X^s={1\over 2}[X-X],\, Y^s={1\over 2}[Y-Y],\, Z^s={1\over 2}[Z-Z]=X^s\times Y^s$$ are unit balls of certain norms $\|\cdot\|_X$ on $L[X]$, $\|\cdot\|_Y$ on $L[Y]$ and $\|\cdot\|$ on $L[Z]$, with \begin{equation}\label{direct2} \|(x,y)\|=\max[\|x\|_X,\|y\|_Y],\,x\in L[X], y\in L[Y]. \end{equation} From now on, we make the following \begin{quote} {\bf Assumption A.} {\sl The just defined norm $\|\cdot\|$ with the unit ball ${1\over 2}[Z-Z]$ is the norm used in the SMP setup, while the d.-g.f $\omega(x,y)$ is of the form $\omega_X(x)+\omega_Y(y)$, where $(\|\cdot\|_X,\omega_X(\cdot))$ and $(\|\cdot\|_Y,\omega_Y(\cdot))$ form SMP setups for $(X,E_x)$ and $(Y,E_y)$ respectively\footnote{Note that such a sum indeed is a d.-g.f. fitting the norm $\|\cdot\|$.}.} \end{quote} Note that \begin{itemize} \item We have \begin{equation}\label{direct3} \|[\xi;\eta]\|_*=\|\xi\|_{X,*}+\|\eta\|_{Y,*}, \end{equation} where $\|\cdot\|_{X,*}$ and $\|\cdot\|_{Y,*}$ are the (semi)norms conjugate to $\|\cdot\|_X$, $\|\cdot\|_Y$, respectively. In particular, we have \begin{equation}\label{unitary} \|F(z)\|_*=\|\nabla\phi(z)\|_*,\,\,\|F(z)-F(z')\|_*=\|\nabla\phi(z)-\nabla\phi(z')\|_*\,\,\forall z,z'\in E. \end{equation} \item The $\omega$-radius $\Omega$ of $Z$ is \begin{equation}\label{Omega} \Omega=\sqrt{\Omega_X^2+\Omega_Y^2},\Omega_X=\sqrt{2[\max_{x\in X}\omega_X(x)-\min_{x\in X}\omega_X(x)]},\Omega_Y=\sqrt{2[\max_{y\in Y}\omega_Y(y)-\min_{y\in Y}\omega_Y(y)]} \end{equation} \end{itemize} \paragraph{Scale factor.} When speaking about complexity of finding $\epsilon$-solution, we shall express it in terms of the {\sl relative accuracy} $\nu=\epsilon/\bV$, where the {\sl scale factor} $\bV$ is defined as follows. Let $\widehat{Z}$ be the convex hull of $\{0\}\cup Z$, and let $$ \widehat{\phi}(z)=\phi(z)-\phi(0)-\langle \phi'(0),z\rangle=\sum_{k=2}^dQ_k(z,...,z). $$ We set \begin{equation}\label{V} \bV=\bV_Z[\phi]:={\max}_{z\in \widehat{Z}} \widehat{\phi}(z)-{\min}_{z\in \widehat{Z}} \widehat{\phi}(z). \end{equation} The importance of this scale factor in our contents stems from the following simple observation (see also Lemma \ref{lemraash} below): \begin{lemma}\label{lem1} For properly chosen positive real $C^{(1)}$ depending solely on $d$, for all $k$, $2\leq k\leq d$ and all collections $z^1,...,z^k$ of vectors from $L[\widehat{Z}]$ one has \begin{equation}\label{eq1} |Q_k(z^1,...,z^k)|\leq C^{(1)}\bV\prod\limits_{i=1}^k\|z^i\|_{\widehat{Z}} \end{equation} In particular, the vector field $F(z)$ associated with {\rm (\ref{SPP})} satisfies {\rm (\ref{aresuch}.$a$)} with \begin{equation}\label{cLis} \cL=C^{(1)}\bV\sum_{k=2}^dk(k-1)2^{k-2}:=C^{(2)}\bV, \end{equation} where $C^{(2)}$ depends solely on $d$. \end{lemma} For proof, see Appendix. \par An immediate question related to the definition of the scaling factor is: a ``shift of the problem by $a\in E$'' -- a simple substitution of variables $z=w-a$ -- changes the factor and thus the complexity estimates, although such a substitution leaves the problem ``the same.'' The answer is as follows: while the ``shift option'' should be kept in mind, such a shift changes the Stochastic Oracle as given by (\ref{base}). Indeed, this oracle is defined in terms of the {\sl homogeneous components in the Taylor decomposition of $\phi(\cdot)$ taken at the origin}, and this is why the origin is participating in the description of $\widehat{Z}$ and thus of $\bV$. Shifting the origin, we, in general, change the \SO \footnote{For example, with $\phi(x,y)\equiv x^3$, the oracle (\ref{base}) is $G=[3x^1x^2;0]$, $z^i=[x^i;0]\sim P_z$. Substituting $x=1+h$, carrying out the construction of the \SO ``in $h$-variable'' and translating the result back to $x$-variable, the resulting \SO turns out to be $G=[3x^1x^2+3x^1-3x^2;0]$, which is not the oracle we started with.}, and thus there is nothing strange that our scaling of the accuracy (and thus -- the efficiency estimates) corresponding to a given $Z$ and a given (implicitly participating in (\ref{base})) \SO is not translation-invariant. \subsection{Complexity Analysis} \paragraph{Preliminaries.} From now on we assume that as applied to (\ref{SPP}), SMP utilizes Stochastic Oracle \SO given according to (\ref{base}) by a family of probability distributions $\cP=\{P_z:z\in Z\}$ on $E$ satisfying (\ref{NoBias}). From now on, we make the following \begin{quote} {\bf Assumption B.} {\sl For some $\rho\geq 0$, all distributions $P_z$, $z\in Z$, are supported on the set $Z+2\rho Z^s\subset\Aff(Z)$, where $Z^s={1\over 2}[Z-Z]$ and $\Aff(Z)$ is the affine hull of $Z$.} \end{quote} In particular, when $P_z$ is supported on $Z$ for all $z\in Z$ (''proper case''), Assumption B is satisfied with $\rho=0$. \par It is time now to note that the \SO we have developed so far gives rise to a {\sl parametric family} of Stochastic Oracles, specifically, as follows. First of all, our basic \SO in fact can be ``split'' into two Stochastic Oracles, $\SOmath^x$ and $\SOmath^y$, providing estimates of the $x$- and the $y$-components $F_x,F_y$ of $F(z)=[F_x(z);F_y(z)]$: the estimates $$ \begin{array}{rcl} E_x\ni G_x&=&G_x[z^1,...,z^{d-1}]: \forall \xi\in E_x: \langle G_x,\xi\rangle=\sum_{k=1}^dkQ_k([\xi;0],z^1,...,z^{k-1}),\\ E_y\ni G_y&=&G_y[z^1,...,z^{d-1}]: \forall \eta\in E_y: \langle G_y,\eta\rangle=-\sum_{k=1}^dkQ_k([0;\eta],z^1,...,z^{k-1}).\\ \end{array} $$ Here, as above, $z^1,...,z^{d-1}$ are, independently of each other, sampled from $P_z$. Now, given two positive integers $k_x,k_y$, we can ``recombine'' our ``partial stochastic oracles'' $\SOmath^x$, $\SOmath^y$ into a new Stochastic Oracle $\SOmath_{k_x,k_y}$ as follows: in order to generate a random estimate of $F(z)$ given $z\in Z$, we generate $(d-1)\max[k_x,k_y]$ independent samples $z^k_\tau\sim P_z$, $1\leq k\leq d-1$, $1\leq\tau\leq k_{xy}:=\max[k_x,k_y]$ and then set \begin{equation}\label{g} g=G^{k_x,k_y}_z\left[\{z^k_\tau\}_{{1\leq k\leq d-1,\atop1\leq\tau\leq k_{xy}}}\right]= \left[{1\over k_x}\sum_{\tau=1}^{k_x}G_x[z^1_\tau,...,z^{d-1}_\tau];{1\over k_y}\sum_{\tau=1}^{k_y}G_y[z^1_\tau,...,z^{d-1}_\tau]\right]. \end{equation} In the sequel, we refer to $k_x$ and $k_y$ as the {\sl $x$- and $y$- multiplicities} of the Stochastic Oracle $\SOmath_{k_x,k_y}$.\par We will make use of the following \begin{lemma}\label{lemraash} Under Assumptions A, B, for all positive integer multiplicities $k_x$, $k_y$, $\SOmath_{k_x,k_y}$ ensures validity of {\rm (\ref{aresuch}.$b$)}, same as the validity of {\rm (\ref{strengthening})} with \begin{equation}\label{sigmais} \sigma=C^{(3)}\bV(1+\rho)^{d-1}\left[\min[1,\Omega_X/\sqrt{k_x}]+\min[1,\Omega_Y/\sqrt{k_y}]\right], \end{equation} where C$^{(3)}$ depends solely on $d$. \end{lemma} For proof, see Appendix. \par We have arrived at the following \begin{theorem}\label{themain} Let $t\geq 1$ be given, let Assumptions A, B be satisfied, and let problem {\rm (\ref{SPP})} be solved by $t$-step SMP utilizing $\SOmath_{k_x,k_y}$, with the parameters $\cL$, $\sigma$ underlying the stepsize policy {\rm (\ref{stepsizes})} given by {\rm (\ref{cLis}), (\ref{sigmais})}. Then, for some $C$ depending solely on $d$, \begin{equation}\label{totalb} \begin{array}{ll} (a)&\bE\{\DualityGap(x^t,y^t)\}\leq K(t):=C\left[{\Omega_X^2+\Omega_Y^2\over t}+{\sqrt{\Omega_X^2+\Omega_Y^2}(1+\rho)^{d-1}\vartheta\over\sqrt{t}}\right]\bV\\ &\multicolumn{1}{r}{\vartheta=\min[1,\Omega_X/\sqrt{k_x}]+\min[1,\Omega_Y/\sqrt{k_y}];}\\ (b)&\Prob\left\{\DualityGap(x^t,y^t)>K(t)+C\Lambda{\sqrt{\Omega_X^2+\Omega_Y^2}(1+\rho)^{d-1}\vartheta\bV\over\sqrt{t}}\right\}\leq\exp\{-\Lambda^2/3\}+\exp\{-\Lambda t\}.\\ &\multicolumn{1}{r}{\forall \Lambda>0.}\\ \end{array} \end{equation} \end{theorem} \section{Illustrations}\label{sectIll} We illustrate the proposed approach by two examples. The first of them is of a purely academic nature, the second can pretend to be of some applied interest. When selecting the examples, our major goal was to illustrate randomization schemes different from the one in Example 1. \subsection{Illustration I: minimizing the maximal eigenvalue of a quadratic matrix pencil}\label{Ill1} \paragraph{The problem} we are interested in is as follows: We are given a symmetric matrix quadratically depending on the ``design variables'' $x_1,...,x_J$ which themselves are matrices: \begin{equation}\label{cA} \cA(x)=\sum_{i=1}^I \left[a_i^Tx_{j(i)}^Tq_ix_{j(i)}a_i+b_i^Tx_{j(i)}c_i+c_i^Tx_{j(i)}^Tb_i\right]+d\in E_y:=\bS^m, \end{equation} where \begin{itemize} \item $\bS^m$ is the space of $m\times m$ symmetric matrices equipped with the Frobenius inner product, \item $x=\{x_j\in\bR^{m_j\times n_j}\}_{j=1}^J$ is a collection of variable matrices which we treat as a block-diagonal rectangular matrix with diagonal blocks $x_j$, $1\leq j\leq J$. We denote the linear space of all these matrices by $E_x$ and equip it with the Frobenius inner product; \item $j(i)\in\{1,...,J\}$, $1\leq i\leq I$, are given integers, \item $\{a_i,b_i,c_i,q_i\}_{i=1}^I$, $d$ are data matrices of appropriate sizes and structures: $$ a_i,c_i\in\bR^{n_{j(i)}\times m},\,b_i\in\bR^{m_{j(i)}\times m},\,q_i\in\bS^{m_{j(i)}},\,d\in\bS^m; $$ in addition, we assume that {\sl all $q_i$ are positive semidefinite}, and that the values $j(i)$, $1\leq i\leq I$, cover the entire range $1\leq j\leq J$, meaning that every one of the blocks $x_j$ indeed participates in $\cA(\cdot)$. \end{itemize} For a matrix $a\in\bR^{p\times q}$, let $\sigma(a)=[\sigma_1(a);...;\sigma_{\min[p,q]}(a)]$ be the vector of singular values of $a$ arranged in the non-ascending order, and let $\|a\|_\nuc=\|\sigma(a)\|_1$ be the nuclear norm of $a$. For a symmetric matrix $a$, let $\lambda_{\max}(a)$ be the maximal eigenvalue of $a$. Finally, let $$X=\{x\in E_x:\|x\|_\nuc\leq1\}.$$ Our goal is to solve the optimization problem \begin{equation}\label{prb} \Opt=\min_{x\in X} \left\{\lambda_{\max}(\cA(x))\right\}, \end{equation} Denoting by $Y$ the standard {\sl spectahedron} in $\bS^m$: $$ Y=\{y\in\bS^m:y\succeq0, \Tr(y)=1\} $$ and observing that $\lambda_{\max}(a)=\max_y\{\Tr(ay):y\in Y\}$, we can convert the problem of interest into the saddle point problem as follows: \begin{equation}\label{SPI} \Opt=\min_{x\in X}\max_{y\in Y}\left[\phi(x,y):=\Tr(y\cA(x))\right]. \end{equation} From $q_i\succeq0$, $i\leq I$, and the fact that $y\succeq0$ for all $y\in Y$ it immediately follows that the restriction of $\phi$ on $y\in Y$ is convex in $x\in E_x$; as a function of $y$, $\phi$ is just linear. Thus, $\phi$ is a convex-concave on $X\times Y$ polynomial of degree $d=3$. The monotone mapping (\ref{F}) associated with (\ref{SPI}) is \begin{equation}\label{FI} \begin{array}{rcl} F_x(x,y)&=&2\Diag\{\sum_{i:j(i)=j}[q_ix_ja_iya_i^T+b_iyc_i^T],1\leq j\leq J\}\in E_x,\\ F_y(x,y)&=&-\cA(x),\\ \end{array} \end{equation} Now let us apply to (\ref{SPI}) the approach we have developed so far. \par {\bf A.} First, let us fix the setup for SMP. We are in the situation when $X^s:={1\over 2}[X-X]$ is $X$ -- the unit ball of the nuclear norm on $E_x$; thus, $\|\cdot\|_X$ is the nuclear norm on $E_x$. The set $Y^s={1\over 2}[Y-Y]$ clearly is contained in the unit nuclear norm ball of $\bS^m$ and contains the concentric nuclear norm ball of radius $1/2$, meaning that $\|\cdot\|_Y$ is within factor 2 of the nuclear norm: $$ 2\|y\|_\nuc\geq \|y\|_Y\geq \|y\|_\nuc\,\,\forall y\in \bS^m=E_y. $$ The best, within $O(1)$ factors, known so far under circumstances choice of the d.-g.f.'s is (see \cite[section 5.7.1]{FOM} or Propositions \ref{propNN2}, \ref{propNN1} in Appendix) \begin{equation}\label{dgfs} \begin{array}{rcl} \omega_X(x=\Diag\{x_1,...,x_J\})&=&O(1)\ln(n)\sum_{j=1}^J\sum_{\ell=1}^{\min[m_j,n_j]}\sigma_\ell^{q({n})}(x_j),\\ &&{n}=\sum_{j=1}^J\min[m_j,n_j],\,q(n)={1\over 2\ln({n})},\\ \omega_Y(y)&=&O(1)\ln(m)\sum_{\ell=1}^m \sigma_\ell^{q(m)}(y),\\ \end{array}\ \ \footnotemark^) \end{equation} \footnotetext{To avoid trivial situations, we assume from now on that $m>1$, $n>1$.}\noindent with explicitly given absolute constants $O(1)$. This choice is reasonably good in terms of the values of the corresponding radii of $X$, $Y$ which turn to be ``quite moderate:'' \begin{equation}\label{radii} \Omega_X\leq O(1)\sqrt{\ln({n})},\,\,\Omega_Y\leq O(1)\sqrt{\ln(m)}. \end{equation} Note that the efficiency estimate (\ref{totalb}) says that we are interested in as small values of $\Omega_X$, $\Omega_Y$ as possible. At the same time, it is immediately seen that if $\omega(\cdot)$ is a d.-g-.f. for $Z$ compatible with the norm generated by $Z$ (i.e., with the unit ball $Z^s={1\over 2}[Z-Z]$, then the $\omega$-radius of $Z$ is {\sl at least} $O(1)$, so that $\Omega_X$, $\Omega_Y$ are ``nearly as good'' as the could be. \par The outlined d.-g.f.'s are also the best known under circumstances in terms of the computational complexity of the associated prox-mapping; it is easily seen that this complexity is dominated by the necessity to carry out singular value decomposition of a matrix from $E_x$ (which takes $O(\sum_jm_j n_j\min[m_j,n_j])$ a.o.) and eigenvalue decomposition of a matrix from $\bS^m$ ($O(m^3)$ a.o.), see below. \par {\bf B.} With our approach, the ``basic'' option when solving (\ref{SPI}) is to use the deterministic version of SMP, i.e., to use as $P_z$ the unit mass sitting at $z$. The corresponding efficiency estimate can be obtained from (\ref{totalb}) by setting $k_x=k_y=\infty$; taking into account (\ref{radii}), the resulting estimate says that a solution to (\ref{SPI}) of a given accuracy $\epsilon\leq\bV$ will be found in course of \begin{equation}\label{Nofeps} N_\det(\epsilon/\bV)=O(1)\ln(m{n}){\bV/\epsilon} \end{equation} iterations. Now let us evaluate the arithmetic complexity of an iteration. From the description of the algorithm it is clear than the computational effort at an iteration is dominated by the necessity to compute exactly $O(1)$ values of the monotone mapping (\ref{FI}) and of $O(1)$ prox mappings. To simplify evaluating the computational cost of an iteration, assume from now on that we are in the {\sl simple case}: $$ m_j=n_j=\nu,\,1\leq j\leq J. $$ In this case, computing $O(1)$ values of the prox mapping costs $$ \cC_\prox=O(1)[m^3+J\nu^3]\hbox{\ a.o.} $$ \begin{quote} {\small Indeed, with our $\omega_X(\cdot)$, computing the $x$-component of prox mapping reduces to solving the optimization problem $\min_{v\in E_x,\|v\|_\nuc\leq1}[\sum_{j=}^J\sum_{\ell=1}^\nu\sigma_\ell^q(v_j) -\Tr(g^Tv)]$ with a given $q\in(1,2]$ and a given $g\in E_x$. To solve the problem, we compute the singular value decompositions of all diagonal blocks $g_j$ in $g$, this getting a representation $g=U\Diag\{\gamma\}V^T$ with block-diagonal orthogonal matrices $U$, $V$, which takes $O(1)J\nu^3$ a.o. It is immediately seen that the problem admits an optimal solution $v$ of the same structure as $g$: $v=U\Diag\{\upsilon\}V^T$. Specifying $\upsilon$ reduces to solving the convex optimization problem $$\min_{\upsilon\in\bR^n:\|\upsilon\|_1\leq1}\left[\sum_j[|\upsilon_j|^q+\gamma_j\upsilon_j]\right];$$ this convex problem with separable objective and a single separable constraint clearly can be solved within machine precision in $O(n)$ a.o. Finally, given $\upsilon$, it takes $O(1)J\nu^3$ operations to compute the $x$-component $U\Diag\{\upsilon\}V^T$ of the prox mapping. Thus, the total cost of the $x$-component of the prox mapping is $O(1)J\nu^3$ a.o. The situation with computing the $y$-component of the mapping is completely similar, and the cost of this component is $O(1)m^3$ a.o.} \end{quote} Looking at (\ref{FI}), we see that computing $O(1)$ values of $F$ at ``general position'' points $z$, assuming all the data matrices dense, is $$ \cC_F=O(1)\nu m (\nu+m) I \hbox{\ a.o.} $$ As a result, the arithmetic cost of finding $\epsilon$-solution to (\ref{SPI}) (and thus -- to (\ref{prb})) by the deterministic version of SMP is \begin{equation}\label{complIdet} \cC_\det (\epsilon)=O(1)\ln(mn)\underbrace{\big[\overbrace{m^3+ J\nu^3}^{\Theta_\prox}+\overbrace{m\nu (m+\nu) I}^{\Theta_F}\big]}_{\Theta}{\bV\over\epsilon}\hbox{\ a.o.} \end{equation} Note that we are not aware of better complexity bounds for large-scale problems (\ref{prb}), at least in the case when in the expression for $\Theta$, the term $m^3$ is dominated by the sum of other terms. \par {\bf C.} Now let us look whether we can reduce the overall arithmetic cost of $\epsilon$-solution to (\ref{prb}) by randomization. An immediate observation is that the only case when it can happen is the one of $\Theta_F\gg\Theta_\prox$. Indeed, comparing the efficiency estimates (\ref{totalb}) and (\ref{Nofeps}), we conclude that randomization can only increase the iteration cost of $\epsilon$-solution; in order to overweigh the growth in the number of iterations, we need to reduce significantly the arithmetic cost of an iteration, and to this end, this cost, in the deterministic case, should be by far dominated by the cost of computing the values of $F$ (the only component of our computational effort which can be reduced by randomization). Assuming $\Theta_F\gg \Theta_\prox$, let us look which kind of randomization could be useful in our context. Note that in order for randomization to be useful, the underlying distributions $P_z$ should be supported on the set of those pairs $(x,y)$ for which computing an estimate $g$ of $F(z)$ according to (\ref{g}) is much cheaper than computing $F$ at a general-type point $(x,y)\in E_x\times E_y$. A natural way to meet this requirement us to use the ``matrix analogy'' of Example 1, where $P_z$ are supported on the set of low rank matrices. Specifically, in order to get an unbiased estimate of $F(z)$, $z=(x,y)\in X\times Y$, let us act as follows: \begin{enumerate} \item We compute the singular value decomposition $x=U\Diag\{\sigma[x]\}V^T$ of $x= \Diag\{x_1,...,x_J\}$ (here $\sigma[x]=[\sigma(x_1);...;\sigma(x_J)]$) and eigenvalue decomposition $y=W\Diag\{\sigma(y)\}W^T$ of $y$ \footnote{Note that the singular values of $y$ are the same as eigenvalues, since $y\succeq0$ due to $y\in Y$.}, where $U,V$ are block-diagonal $n\times n$ orthogonal with $\nu\times\nu$ diagonal blocks, and $W$ is an orthogonal $m\times m$ matrix. \item We specify $P_x$ as the distribution of a random matrix $\xi\in E_x$ with takes the values $$\|\sigma[x]\|_1\Col_\ell[U]\Col_\ell^T[V],\, 1\leq \ell\leq n,$$ with the probabilities $(\sigma[x])_\ell/\|\sigma[x]\|_1$ (when $\sigma[x]=0$, $\xi$ takes value $0$ with probability 1); here $\Col_\ell(A)$ denotes $\ell$-th column of a matrix $A$. \item We specify $P_y$ as the distribution of the random symmetric matrix $\eta$ which takes values $\Col_i[W]\Col_i^T[W]$, $1\leq i\leq m$, with probabilities $\sigma_i(y)$, and specify $P_z$ as the direct product of $P_x$ and $P_y$. \end{enumerate} Observe that the expectation of $\zeta\sim P_z$ is exactly $z$, and that $P_z$, $z\in Z=X\times Y$, is supported on $Z$ due to $\|\sigma(x)\|_1\leq1$, $x\in X$, $\|\sigma(y)\|_1=1$, $y\in Y$. In other words, {\sl assumption B is satisfied with $\rho=0$.} \par Note that with the just defined $P_z$, a realization $\zeta=(\xi,\eta)\sim P_z$ is of very special structure: \begin{equation}\label{structure} \xi=u\times v^T,\,u,v\in\bR^n,\,\,\eta=ww^T,\,w\in\bR^m; \end{equation} moreover, among the $J$ consecutive $\nu$-dimensional blocks $u_j,v_j$, $j=1,...,J$, of every one of the vectors $u,v\in\bR^{n=J\nu}$, all but one blocks are zero, and the nonzero blocks $u_j$, $v_j$ share a common index $j$. \par It is immediately seen that with the just defined distributions $P_z$, the unbiased estimate (\ref{g}) of $F(z)$ is as follows: \begin{equation}\label{gI} \begin{array}{rcl} G_x^{k_x}&=&{1\over k_x}\sum\limits_{\ell=1}^{k_x}\Diag\bigg\{\sum\limits_{i:j(i)=j}\big[q_i u^{2\ell-1}_j[v^{2\ell-1}_j]^Ta_iw^{2\ell}[w^{2\ell}]^Ta_i^T +q_i u^{2\ell}_j[v^{2\ell}_j]^Ta_iw^{2\ell-1}[w^{2\ell-1}]^Ta_i^T\\ &&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad+2b_iw^{2\ell-1}[w^{2\ell-1}]^Tc_i^T\big],j=1,...,J\bigg\}\\ G_y^{k_y}&=&-d-{1\over k_y}\sum\limits_{\ell=1}^{k_y}\sum\limits_{i=1}^I\bigg[{1\over 2}a_i^Tv^{2\ell-1}_{j(i)}[u^{2\ell-1}_{j(i)}]^Tq_iu^{2\ell}_{j(i)}[v^{2\ell}_{j(i)}]^Ta_i+{1\over 2}a_i^Tv^{2\ell}_{j(i)}[u^{2\ell}_{j(i)}]^Tq_iu^{2\ell-1}_{j(i)}[v^{2\ell-1}_{j(i)}]^T\\ &&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad +b_i^Tu^{2\ell-1}_{j(i)}[v^{2\ell-1}_{j(i)}]^Tc_i+c_i^Tv^{2\ell-1}_{j(i)}[u^{2\ell-1}_{j(i)}]^Tb_i\bigg]\\ \end{array} \end{equation} where the collections $$\zeta^\ell=\left([u^\ell_1;...;u^\ell_J][v^\ell_1;...;v^\ell_J]^T,w^\ell[w^\ell]^T\right),\,\ell=1,...,2\max[k_x,k_y]$$ are independently of each other drawn from $P_z$. \par It is immediately seen that the arithmetic cost of computing $(G_x,G_y)$ given $z=(x,y)$ is comprised of the components as follows: \begin{enumerate} \item ``Setup cost'' -- one of computing singular value decomposition of $x$ and eigenvalue decomposition of $y$ \footnote{In fact, this cost is nonexisting: by construction of the method, the points $z$ where one needs to evaluate $F$ are the values of already computed prox-mappings; according to how we compute these values (see above), they go together with their singular value/eigenvalue decompositions.} ($O(1)(m^3+J\nu^3)$ a.o.) plus the cost of computing the ``cumulative distributions'' $S_j(x)=\|\sigma[x]\|_1^{-1}\sum_{\tau=1}^j(\sigma[x])_\tau$, $1\leq j\leq J\nu$, $S_i(y)=\sum_{\tau=1}^i\sigma_\tau(y)$ ($O(1)(m+J\nu)$ a.o.). \item After the setup cost is paid, for every $\ell$\\ --- generating $\zeta^\ell$ costs $O(1)(\ln(m)+\ln(J\nu)+m+\nu)$ a.o.,\\ --- computing the contribution of $(\zeta^{2\ell-1},\zeta^{2\ell})$ to $G_x$ costs no more than $O(1)I\nu(m+\nu)$ a.o. (look at (\ref{gI})), and this cost should be paid $k_x$ times;\\ --- computing the contribution of $(\zeta^{2\ell-1},\zeta^{2\ell})$ to $G_y$ costs at most $O(1)(m+\nu)^2K$ a.o., where $K=\max_{1\leq j\leq J} \Card\{i: j(i)=j\}$ (look at (\ref{gI}) and take into account that the vectors $u^{\ell}$, $v^{\ell}$ have a single nonzero $\nu$-dimensional block each), and this cost should be paid $k_y$ times. \end{enumerate} Thus, the cost of computing $(G_x^{k_x},G_y^{k_y})$ is \begin{equation}\label{costofG} \begin{array}{c} O(1)\left(m^3+J\nu^3+k_x \nu(m+\nu)I+ k_y (m+\nu)^2K+k_y\ln(J)\right)\hbox{\ a.o.},\\ K=\max_{1\leq j\leq J} \Card\{i: j(i)=j\}\\ \end{array} \end{equation} (note that $J\leq I$). To simplify the analysis to follow, assume from now on that $I=J\leq\exp\{(m+\nu)^2K\}$ and $j(\cdot)$ is one-to-one. In this case $K=1$ and the cost of an iteration is \begin{equation}\label{costrand} O(1)\left(m^3+J\nu^3 + k_x\nu(m+\nu)J+k_y(m+\nu)^2\right)\hbox{\ a.o.} \end{equation} Now let us evaluate the overall complexity of finding, with confidence $1-\delta$, $\delta\ll1$, an $\epsilon$-solution by the randomized SMP. We assume from now on that $\epsilon\leq\bV$ (otherwise the problem is trivial, since $\DualityGap(z)\leq\bV$ for every $z\in X\times Y$). For the sake of simplicity, we restrict ourselves with the case of $k_x=k_y=1$. Invoking the efficiency estimate (\ref{totalb}.$b$) with $\Lambda=O(1)\sqrt{\ln(1/\delta)}$ and taking into account (\ref{radii}) and the fact that we are in the situation of $\rho=0$, the number $t$ of iterations which results,with confidence $1-\delta$, in $\DualityGap(x^t,y^t)\leq\epsilon$ is bounded from above by $$ N_{\rand,\delta}(\epsilon)=O(1)\ln(mn)\ln(1/\delta)(\bV/\epsilon)^2, $$ provided that $\ln(1/\delta)\leq O(1)\ln(mn)(\bV/\epsilon)^2$. Thus, the iteration count now is nearly square of the one for the deterministic algorithm, see (\ref{Nofeps}). Taking into account (\ref{costrand}), the overall complexity of achieving our goal with the randomized algorithm does not exceed $$ \cC_{\rand,\delta}(\epsilon)=O(1)\ln(mn)\ln(1/\delta)\left[m^3+J\nu^3+(m+\nu)(m+\nu J)\right](\bV/\epsilon)^2 \hbox{\ a.o.} $$ The ratio of this quantity and the ``deterministic complexity'' (see (\ref{complIdet}) and take into account that we are in the case of $I=J$) is $$ \cR={\cC_{\rand,\delta}(\epsilon)\over\cC_\det(\epsilon)}=O(1)\ln(1/\delta)\underbrace{\left[{m^3+\nu^3J+(m+\nu)(m+\nu J)\over m^3+\nu^3J+m\nu(m+\nu)J}\right]}_{r}\cdot{\bV\over\epsilon}. $$ It is immediately seen that {\sl when $\bV/\epsilon$ and $\delta$ are fixed, and $m,\nu,J$ vary in such a way that $m,n=\nu J$ go to $\infty$ and $\nu/m$, $m/n$ go to 0}, $\cR$ goes to 0 as $O(1/m)$, meaning that eventually the randomized algorithm outperforms its deterministic competitor, and the ``performance ratio'' goes to $\infty$ as the sizes $m,n$ of the problem grow. \paragraph{Numerical illustration.} In the experiment we are about to describe, the sizes of problem (\ref{prb}) were selected as $$ m=300,\,m_j\equiv n_j\equiv=\nu=2,\,I=J=5000,\, j(i)\equiv i, $$ which results in $\dim x=20000$, $\dim y=45150$. The data matrices $q_i\succeq0,a_i,b_i,c_i$ were generated at random and normalized to have spectral norms 1, which ensures $\bV\leq1$. A generated instance was processed as follows: \\ \indent $\bullet$ first, it was solved by the deterministic Mirror Prox algorithm (DMP) with on-line adjustable ``aggressive'' stepsize policy \cite{DMP}; up to this policy, this is nothing but SMP with $P_z$ specified as the unit mass sitting at $z$, $z\in Z$; \\ \indent $\bullet$ next, it was solved by SMP (10 runs) with $k_x=1$, $k_y=100$ \footnote{with our $m,\nu,J$, the coefficient at $k_x$ in the right hand side of (\ref{costofG}) is nearly 30 times larger than the one at $k_y$, this is why we use $k_y\gg k_x$.} and the stepsize policy $$ \gamma_\tau=\alpha\min\left[{1\over\sqrt{3}\cL}, {\sqrt{\Omega_X^2+\Omega_Y^2}\over\sqrt{7}\sigma\sqrt{\tau}}\right],\,\tau=1,2,... $$ with $\cL$ and $\sigma$ given by (\ref{cLis}) (where we replace $\bV$ by its valid upper bound 1) and (\ref{sigmais}) (where we use $\Omega_X,\Omega_Y$ as given by (\ref{radii})). When $\alpha=1$, our stepsize policy becomes the ``rolling horizon'' version of (\ref{stepsizes}); it can be shown that this policy (which does not require the number $t$ of steps to be chosen in advance) is, theoretically, basically as good as its constant stepsizes prototype). The role of the ``acceleration factor'' $\alpha\geq1$ is to allow for larger stepsizes than those given by the worst-case-oriented considerations underlying (\ref{stepsizes}), the option which for DMP is given by the aforementioned on-line adjustable stepsize policy (in our experiments, the latter resulted in stepsizes which, at average, were $\approx 250$ times the ``theoretically safe'' ones). The value of $\alpha$ we used (1000) was selected empirically in a small series of pilot experiments and was never revised in the main series of experiments.\\ \indent $\bullet$ In every experiment, a solution with the duality gap $\leq\epsilon=0.01$ was sought. Since the duality gap is not directly observable, this goal was achieved as follows. From time to time (specifically, after every 30 iterations for DMP and every 50 iterations for SMP) we computed $F(z^t)$ for the current approximate solution $z^t=(x^t,y^t)$ (see (\ref{SMP})), thus getting $g:=\nabla_x\phi(x^t,y^t)$ and $\cA(x^t)=\nabla_y\phi(x^t,y^t)$. We then computed the maximal eigenvalue $\phi^+=\lambda_{\max}(\cA(x^t))$, which is nothing but $\overline{\phi}(x^t)=\max_{y\in Y}\phi(x,y)$, and the quantity $\phi^-=\min_{x\in X}[\phi(x^t,y^t)+\Tr([x-x^t]^Tg)]$, which is a lower bound on $\underline{\phi}(y^t)=\min_{x\in X}\phi(x,y^t)$. The quantity $\Delta=\phi^+-\phi^-$ is an upper bound on $\DualityGap(x^t,y^t)$, and the relation $\Delta\leq\epsilon=0.01$ was used as the termination criterion. \par The results of a typical experiment are presented in table \ref{table1}. We see that while randomization increases essentially the iteration count, it results in overall reduction of the CPU time by a quite significant factor. It makes sense to note that of 2167 sec CPU time for DMP, 91\% (1982 sec) were spent on computing the values of $F$, and just 9\% -- on computing prox-mappings; for SMP, both these components take nearly equal times. \begin{table} \centerline{ \begin{tabular}{|c||c|c|c||c|c|c||} \cline{2-7} \multicolumn{1}{c||}{}&\multicolumn{3}{|c||}{Iteration count}&\multicolumn{3}{|c||}{CPU, sec}\\ \hline Algorithm&$\min$&mean&$\max$&$\min$&mean&$\max$\\ \hline\hline DMP&\multicolumn{3}{c||}{61}&\multicolumn{3}{c||}{2167}\\ \hline SMP&251&281&351&496&571&708\\ \hline\hline \end{tabular}} \caption{\label{table1}. Effect of randomization, problem (\ref{SPI}) ($I=J=5000,m=300,j(i)\equiv i,m_j\equiv n_j\equiv 2$). In the table: DMP/SMP -- Deterministic/Randomized Mirror Prox. Data for SMP are obtained in 10 runs of the algorithm. Running times include those needed to check the termination criterion.} \end{table} \subsection{Illustration II: low dimensional approximation} Consider the problem as follows: we are given $n$ {\sl unit} vectors $a_j\in\bR^m$, $1\leq j\leq n$, and know that for some given $k$, $1< k\leq m/2$, and $\delta\in(0,1)$ all $a_j$'s are at the $\|\cdot\|_2$-distance at most $\delta<1$ form certain $k$-dimensional subspace $L$, common for all points. The problem is to recover this subspace\footnote{Note the difference with the PCA -- Principal Component Analysis: we want to minimize the maximal, over $j$, deviation of $a_j$, from $L$ rather than the sum of squares of these deviations.}, which reduces to solving the problem \begin{equation}\label{intract} \Opt_*=\max_{x\in \cP_k}\min_{y\in Y} \sum_{j=1}^n y_ja_j^Txa_j, \end{equation} where $\cP_k\subset E_x=\bS^m$ is the family of all orthoprojectors of rank $k$ on $\bR^p$, and $Y=\{y\in\bR^n_+:\sum_jy_j=1\}$ is the standard simplex in $E_y=\bR^m$. The set $\cP_k$ is nonconvex; we relax it to the set $$ X=\{x\in\bS^m: I_m\succeq x\succeq0,\Tr(x)=k\}, $$ thus arriving at the relaxed saddle point problem \begin{equation}\label{(SP)} \begin{array}{c} -\Opt=\min_{x\in X}\max_{y\in Y}[\phi(x,y):=-\sum_{j=1}^n y_ja_j^Txa_j]\\ F_x(x,y)=-\sum_{j=1}^n y_ja_ja_j^T,\quad F_y(x,y)=[a_1^Txa_1;...;a_n^Txa_n]\\ \end{array} \end{equation} (we have equivalently transformed the relaxed problem to fit our standard notation). Note that $\phi$ is a polynomial of degree $d=2$ (just bilinear). Let us apply to (\ref{(SP)}) our approach. \paragraph{Scale factor.} We clearly have $\bV\leq1$ (recall that $\|a_j\|_2=1$, $0\preceq x\preceq I_m$ for $x\in X$, and $\|y\|_1\leq1$ for $y\in Y$). \paragraph{Setup.} We set \begin{equation}\label{omegasLDA} \begin{array}{rcl} \omega_X(x)&=&{8\over q(1+q)}\sum_{i=1}^m\lambda_i^{1+q}(x),\,\,q=\min[1,\ln(k)/\ln(m/k)],\\ \omega_Y(y)&=&{8\sqrt{\e}\over p(1+p)}\sum_{j=1}^n y_j^{1+p},\,p=1/(2\ln(n)),\\ \end{array} \end{equation} thus getting d.-g.f.'s for $X$, $Y$ compatible with $\|\cdot\|_X$, $\|\cdot\|_Y$, respectively (Proposition \ref{propNN1} and Remark \ref{remlast}), the corresponding radii of $X$, $Y$ are \begin{equation}\label{radiiNew} \Omega_x\leq O(1)\sqrt{{k\ln(k)/\ln(m/k)}},\,\,\Omega_Y\leq O(1)\sqrt{\ln(n)}, \end{equation} see (\ref{XomegaX}). \paragraph{Deterministic algorithm.} When solving (\ref{(SP)}) within accuracy $\epsilon<1$ by the deterministic algorithm DMP,\\ --- the iteration count is $N_\det(\epsilon)=O(1) {k\ln(k)/\ln(m/k)+\ln(n)\over\epsilon}$,\\ --- the complexity of an iteration is $O(1)(m^3+n)$ a.o. for computing prox-mappings and $O(1)m^2n$ a.o. for computing the values of $F$. \\ Note that {\sl as far as deterministic solution algorithms are concerned}, the outlined bounds result in the best known to us overall arithmetic complexity of finding an $\epsilon$-solution in the large scale case. \par When $n\gg m$, the cost of prox-mapping is much smaller than the one of computing the values of $F$, implying that there might be room for accelerating by randomization. \paragraph{Randomization.} In order to compute, given $z=(x,y)\in X\times Y$, unbiased random estimates of $F_x(x,y)$ and $F_y(x,y)$, we act as follows. \begin{enumerate} \item We associate with $y$ the distribution $P_y$ on $Y$ as follows: $\eta\sim P_y$ takes the values $e^j$ (basic orths in $\bR^n$) with probabilities $y_j$, $1\leq j\leq n$ (cf. Example 1); the corresponding random estimate $G^x$ of $F_x(x,y)$ takes the values $-a_ja_j^T$ with probabilities $y_j$, $1\leq j\leq n$. Generating the estimate requires the ``setup cost'' of $O(n)$ a.o.; after this cost is paid, generating the estimate takes $O(1)[\ln(n)+m^2]$ a.o. \item We associate with $x\in X$ the distribution $P_x$ on $X$ as follows. Given $x$, we compute its eigenvalue decomposition $x=U\Diag\{\xi\}U^T$. The vector $\xi$ belongs to the polytope $Q=\{\xi\in\bR^m:0 \leq \xi_i\leq 1\,\forall i, \sum_i\xi_i=k\}$. Now, there is a simple algorithm \cite[section A.1]{Rand} which allows, given $\xi\in Q$, to represent $\xi$ as a convex combination $\sum_{i=1}^m\lambda_i\xi^i$ of extreme points of $Q$ (which are Boolean vectors with exactly $k$ entries equal to 1); the cost of building this representation is $O(1)km^2$ a.o. We build this representation and define $P_x$ as the distribution of a random symmetric matrix which takes values $U\Diag\{\xi^i\}U^T$ with probabilities $\lambda_i$, $1\leq i\leq m$, so that the random estimate of $F_y(x,y)$ is the vector with the entries $G^y_j=\sum_{\ell\in I_i}(a_j^T\Col_\ell[U])^2$, $1\leq j\leq n$, where $I_i$ is the set of indexes of the $k$ nonzero entries of the Boolean vector $\xi^i$, and $i$ takes values $1,...,m$ with probabilities $\lambda_1,...,\lambda_m$. Finally, we set $P_z=P_x\times P_y$. Note that this distribution is supported on $X\times Y$ (i.e., Assumption B is satisfied with $\rho=0$). The ``setup'' cost of sampling from $P_x$ is $O(1)m^3$ a.o.; after this cost is paid, generating a sample value of $G^y$ costs $O(1)kmn$ a.o. \end{enumerate} With the outlined randomization, the cost of generating a sample value of $G_{k_x,k_y}$ in the range $\ln(n)\leq O(1)m^2$ costs $$ O(1)(m^3+k_xkmn+k_ym^2) \hbox{\ \rm a.o.} $$ When $n\gg m\gg k$ and $k_x$, $k_y$ are moderate, this cost is by far less than the cost $O(1)m^2n$ of deterministic computation of $F(x,y)$, so that our randomization indeed possesses some potential. Analysis completely similar to the one in section \ref{Ill1} shows that our current situation is completely similar to the one in the latter section: while with $k_x=O(1)$, $k_y=O(1)$, the iteration count for the randomized algorithm is proportional to $\epsilon^{-2}$ instead of being proportional to $\epsilon^{-1}$, as for the deterministic algorithm, the growth in this count, in certain meaningful range of values of $k,m,n,\epsilon$ is by far overweight by reduction in the cost of an iteration. As a result, for $\epsilon$ fixed and in the case of appropriate proportion between $k,m,n$, the randomized algorithm progressively outperforms its deterministic competitor as the sizes of the problem grow. \paragraph{Numerical illustration.} In the experiment we are about to describe, the sizes of problem (\ref{(SP)}) were selected as $$ m=100,\,k=10,\,n=300,000. $$ The data points $a_j$ were selected at random in certain ``smart'' way aimed at creating difficult instances; we are not sure that this goal was indeed achieved, but at least the PCA solution (which, with straightforward random generation of $a_j$, turns out to recover perfectly well the approximating subspace) was ``cut off:'' -- the largest, over all $j$, distance of $a_j$'s to the $k=10$-dimensional PCA subspace in our experiments was as large as 0.99. \par Implementation of the approach was completely similar to the one outlined in section \ref{Ill1}; the only specific issue which should be addressed here is the one of termination. Problem (\ref{(SP)}) by its origin is no more than a relaxation of the ``true'' problem (\ref{intract}), so solving it within a given accuracy is of no much interest. Instead, we from time to time (namely, every 10 iterations) took the $x$-component $x^t$ of the current approximate solution, subject it to eigenvalue decomposition and checked straightforwardly what is the largest, over $j\leq n$, $\|\cdot\|_2$-deviation $D$ of $a_j$ from the $k$-dimensional subspace of $\bR^m$ spanned by $k$ principal eigenvectors of $x^t$. We terminated the solution process when this distance was $\leq \delta+\epsilon$, where $\epsilon$ is a prescribed tolerance. \par Typical experimental results are presented in table \ref{table2}. The results look surprisingly good -- the iteration count is quite low and is the same for both deterministic and randomized algorithms. We do not know whether this unexpected phenomenon reflects the intrinsic simplicity of the problem, or our inability to generate really difficult instances, or the fact that we worked with although reasonable, but not ``really small'' values of $\epsilon$; this being said, we again see that randomization reduces the CPU time by a quite significant factor. \begin{table} \centerline{\small\begin{tabular}{|c|c|c|c|c|} \cline{2-5} \multicolumn{1}{c|}{}&Method&\# of steps&CPU, sec& Final deviation $D$\\ \hline\hline $\delta=0.4,\delta+\epsilon=0.45$&DMP&20&{478}&0.401\\ \hline \multicolumn{1}{c|}{}&SMP&20&{104}&0.427\\ \hline\hline $\delta=0.6,\delta+\epsilon=0.65$&DMP&20&{ 504}&0.603\\ \hline \multicolumn{1}{c|}{}&SMP&20&{ 105}&0.620\\ \hline\hline $\delta=0.8,\delta+\epsilon=0.85$&DMP&20&{478}&0.809\\ \hline \multicolumn{1}{c|}{}&SMP&20&{92}&0.819\\ \cline{2-5} \end{tabular}} \caption{\label{table2} Deterministic (DMP) and randomized (SMP, $k_x=1$, $k_y=10$) algorithms on the low dimensional approximation problem. } \end{table}

https://arxiv.org/abs/1510.07966

Analysis of a splitting-differentiation population model leading to cross-diffusion

Starting from the dynamical system model capturing the splitting-differentiation process of populations, we extend this notion to show how the speciation mechanism from a single species leads to the consideration of several well known evolution cross-diffusion partial differential equations.Among the different alternatives for the diffusion terms, we study the model introduced by Busenberg and Travis, for which we prove the existence of solutions in the one-dimensional spatial case. Using a direct parabolic regularization technique, we show that the problem is well posed in the space of bounded variation functions, and demonstrate with a simple example that this is the best regularity expected for solutions.We numerically compare our approach to other alternative regularizations previously introduced in the literature, for the particular case of the contact inhibition problem. Simulation experiments indicate that the numerical scheme arising from the approximation introduced in this article outperforms those of the existent models from the stability point of view.

\section{Mathematical model and main result} \subsection{The splitting-differentiation model in terms of ODEs} In \cite{sanchez-palencia}, S\'anchez-Palencia analyzes the situation in which a species, with population density $U$, splits, due to a number of factors, into two different species with population densities $U_1$ and $U_2$, but still keeping the original ecological behavior. Thus, we assume that $U$ followed a logistics law until time $t^*$, i.e. \begin{equation} \label{sp.1} U'(t)= U(t) \big(\alpha -\beta U(t)\big),\qtext{for } t\in(0,t^*), \quad U(0)=U_0>0. \end{equation} Then, after splitting, $(U_1,U_2)$ is assumed to satisfy the Lotka-Volterra system \begin{align} \label{sp.2} U_i'(t)= U_i(t) \big(\alpha -\beta \big(U_1(t)+U_2(t)\big)\big),\qtext{for } t\in(t^*,T), \quad U_i(t^*)=U_{i0}>0, \end{align} for $i=1,2$, with $U_{10}+U_{20}=U(t^*)$. Note that, under this splitting \emph{without differentiation}, $U_1+U_2$ still satisfies \fer{sp.1} for $t\geq t^*$. Although problem \fer{sp.1} has a unique non-trivial equilibrium, $U_{\infty}=\alpha /\beta$, problem \fer{sp.2} has a continuum set of non-trivial equilibria given by all the combinations of $U_{1\infty}\geq0$ and $U_{2\infty}\geq0$ such that $U_{1\infty}+U_{2\infty}=\alpha /\beta$. In \cite{sanchez-palencia}, the author analyzes how the differentiation of populations $U_1$, $U_2$ after splitting, understood as a perturbation in the Lotka-Volterra coefficients, affects the equilibrium of the system. By splitting \emph{with differentiation} we mean that $(U_1,U_2)$ is a solution of the following problem: \begin{align} \label{sp.3} U_i'(t)= U_i(t) \big(\alpha_i - \big(\beta_{i1}U_1(t)+\beta_{i2}U_2(t)\big)\big),\qtext{for } t\in(t^*,T), \quad U_i(t^*)=U_{i0}>0, \end{align} for $i=1,2$, with $U_{10}+U_{20}=U(t^*)$. Observe that, under this splitting, $U_1+U_2$ does not satisfy, in general, problem \fer{sp.1} for $t\geq t^*$. The main conclusion of \cite{sanchez-palencia} is that the differentiation mechanism selects, in general, a unique solution for the equilibrium system, having therefore a stabilizing effect. \subsection{The splitting model in terms of PDEs: cross-diffusion} In this article, we extend the previous dynamical system models to the case of space dependent population densities. We start considering the dynamics of one single species population satisfying \begin{equation} \left\{ \begin{array}{ll} \partial_t u -\Div J(u) =f(u) & \qtext{in } Q_{(0,t^*)},\\ J(u)\cdot \boldsymbol{\nu}=0 & \qtext{on } \Gamma_{(0,t^*)}, \\ u(0,\cdot)=u_0 \geq 0 & \qtext{on }\O, \end{array} \right. \label{p.u} \end{equation} where $\O\subset \mathbb{R}^N$ is a bounded domain with Lipschitz continuous boundary, $\partial\O$, $Q_{(0,t^*)}=(0,t^*)\times\O$, and $\Gamma_{(0,t^*)}=(0,t^*)\times\partial\O$ is the parabolic boundary of $Q_{(0,t^*)}$. The vector $\boldsymbol{\nu}$ is the outwards canonical normal to $\partial\O$. The growth-competition term is assumed to have the logistic form $f(u)=u(\alpha-\beta u)$, and the flow to be given by $ J(u)=u\nabla u+u\mathbf{q}$. As it is well known, the term $u\nabla u$ captures the individuals aversion to overcrowding, while $\mathbf{q}$ is usually determined by an environmental potential, $\mathbf{q}=-\nabla \varphi$, whose minima represent attracting points for the populations. For biological background and origins of the model see, for instance, \cite{okubo}. After splitting, the new two populations, $u_1$ and $u_2$, satisfy, for $i=1,2$, \begin{equation} \left\{ \begin{array}{ll} \partial_t u_i -\Div J_i(u_1,u_2) =f_i(u_1,u_2) & \qtext{in } Q_{(t^*,T)},\\ J_i(u_1,u_2)\cdot \boldsymbol{\nu}=0 & \qtext{on } \Gamma_{ (t^*,T)}, \\ u_i(t^*,\cdot)=u_{i0} & \qtext{on }\O, \end{array} \right. \label{p.u2} \end{equation} with $u_{i0}$ such that $u_{10}+u_{20}=u(t^*,\cdot)$, and with $J_i$ and $f_i$ to be defined. Assuming, like in the dynamical system model, that the possible differentiation process only takes place through the growth and the inter- and intra-competitive behavior of the new species implies that the split flows must satisfy $J_1(u_1,u_2)+J_2(u_1,u_2)=J(u_1+u_2)$, that is \[ J_1(u_1,u_2)+J_2(u_1,u_2)=(u_1+u_2)\nabla(u_1+u_2) + (u_1+u_2)\mathbf{q}. \] Being clear the way of defining the linear transport term of $J_i$, the nonlinear diffusive term admits several reasonable decompositions. For instance, in \cite{Galiano2012}, the following splitting was considered \begin{equation} \label{def.flows2} J_i(u_1,u_2)=u_i\nabla u_i + b_i \nabla(u_1u_2)+u_i\mathbf{q} , \end{equation} with $b_i\geq 0$, and $b_1+b_2 = 1$. Under this splitting, problem \fer{p.u2} takes the form of the cross-diffusion model introduced by Shigesada et al. \cite{SKT79}, for which a thorough mathematical analysis does exist, see for instance \cite{ggj2,gambino,andreianov11,berres11,Ruiz-Baier2012,Juengel2014} for numerical approaches, \cite{chen,gv,Gambino2012,Gambino2013,Desvillettes2014} for analytical and qualitative results, or \cite{gjv,Galiano2011,Juengel2012} for applications. In this paper, we consider the alternative splitting \begin{equation} \label{def.flows} J_i(u_1,u_2)=u_i\nabla (u_1+u_2)+ u_i\mathbf{q}, \end{equation} which brings problem \fer{p.u2} to the form of the Busenberg and Travis model \cite{busenberg83}. Although apparently simpler than \fer{def.flows2}, no general proof of existence of solutions does exist for problem \fer{p.u2} with flows given by \fer{def.flows}. However, some partial results related to the cell-growth contact-inhibition problem may be found in \cite{bertsch85,bertsch2010,bertsch2012,gsv}, as well as in \cite{Galiano2014,gs} for other specific situations. \subsection{Differentiation after splitting} According to the species behavior after splitting, we consider two problems arising from two different sets of Lotka-Volterra terms: \begin{align} & f_i(u_1,u_2)=u_i(\alpha-\beta (u_1+ u_2)),& &\text{(non-differentiation)}& \label{f.nm}\\ &f_i(u_1,u_2)=u_i \big(\alpha_i - \big(\beta_{i1}u_1+\beta_{i2}u_2\big)\big),& &\text{(differentiation)}\label{f.m}& \end{align} for $i=1,2$. We shall refer to problem \fer{p.u2} with flows given by \fer{def.flows} and with $f_i$ given by \fer{f.nm} and \fer{f.m} as to \textbf{problems (ND) and (D)}, respectively. Observe that these are the PDE versions that generalize the non-differentiation and differentiation ODE problems \fer{sp.2} and \fer{sp.3}, introduced in \cite{sanchez-palencia}. The existence of solutions of problem (ND) was proven in \cite{Galiano2014} in the multi-dimensional case. The proof is based on the construction of solutions, $(u^{\delta}_1 ,u^{\delta}_2)$, to a nonlinear parabolic regularization of problem (ND) such that, although $u^{\delta}_i$ only converges weakly to the solution of the limit problem, $u_i$, the global density, $u^{\delta}_1+u^{\delta}_2$, converges strongly to $u_1+u_2$. Thus, weak and strong convergences are compensated so that the limit problem (ND) is shown to have a solution. However, since no a.e. convergence of $u^{\delta}_i$ to $u_i$ was proven in \cite{Galiano2014}, the case of differentiated Lotka-Volterra terms, i.e. problem (D), may not be handled directly with this technique. Previously to \cite{Galiano2014}, in the one-dimensional setting and for $q=0$, Bertsch et al. \cite{bertsch2010} proved the existence of a solution of problem (D) with the form \begin{equation*} u_1(t,x)=r(t,x)u(t,x),\quad u_2(t,x)=\big(1-r(t,x)\big) u(t,x), \end{equation*} where $u=u_1+u_2$ and $0\leq r \leq 1$ solve certain parabolic-hyperbolic auxiliary problem, see problem (P)$_B$ in Section~\ref{sec:app}. Their proof is based on the parabolic regularization of the auxiliary problem together with the use of the Lagrangian flows (characteristics) associated to $\partial_x u$. In particular, and important for the results proven in this article, they obtained \emph{strong convergence} of the sequence of approximated solutions to a limit, which is identified as a solution of problem~(D). The first aim of this paper is to show the existence of solutions of problem (D) with an alternative proof to that given in \cite{bertsch2010}. Our proof is based on the direct parabolic regularization used for problem (ND) in \cite{Galiano2014}, and takes advantage of the techniques employed in \cite{bertsch2010} to obtain the strong convergence of each component of the regularized problem. In fact, like in \cite{bertsch2010}, the space regularity of solutions of problem (D) is shown to be of bounded variation, $BV(\O)$, which seems to be the optimal expected regularity, see Theorem~\ref{th.easy} in Section~\ref{sec:numerics}, for an example. More explicitly, we prove the following result in Section~\ref{sec:proof}. Here, we retake the usual notation $[0,T]$ for the time domain, replacing $[t^*,T]$. \begin{theorem} \label{th.1} Let $\O\subset\mathbb{R}$ be a bounded interval and $T>0$ be arbitrarily fixed. Let $q\in C^{0,1}(\bar Q_T)$, with $q(t,\cdot)=0$ on $\partial\O$ for all $t\in[0,T]$, and $f_i:\mathbb{R}^2\to\mathbb{R}$ be given by \fer{f.m}. Assume that $u_{10},u_{20}\in BV(\O)$ are non-negative, with $u_0:=u_{10}+u_{20}>0$ in $\bar\O$, $u_0\in C^{0,1}(\bar \O)$ and $\partial_x u_0=0$ on $\partial\O$. Then, there exists a weak solution of problem (D), $(u_1,u_2)$, with, for $i=1,2$, \begin{enumerate} \item $u_i \geq0$ a.e. in $Q_T$. \item $u_i \in L^\infty(0,T;BV(\O))\cap BV(0,T;L^1(\O))$. \item $u:=u_1+u_2 \in C^{0,1}(\bar Q_T) \cap L^2(0,T;H^2(\O))$. \item For all $\varphi \in C^1(0,T;H^1(\O))$ with $\varphi(T,\cdot)=0$ in $\O$, \begin{align*} \int_{Q_T} u_i\partial_t \varphi = \int_{Q_T} \big(u_i\partial_x u + u_iq \big) \partial_x\varphi- \int_{Q_T} f_i(u_1,u_2)\varphi- \int_\O u_{i0}\varphi(0,\cdot). \end{align*} \end{enumerate} \end{theorem} The second aim of this article is to numerically investigate and compare the resulting Finite Element schemes of each regularizing approach, see problems {\text{(P)}$_\delta$}~ and (P)$_B$ in Section~\ref{sec:app}. We focus our attention into two model problems: (i) the \emph{invasion} problem, arising in Tumor theory, as suggested by Bertsch et al. \cite{bertsch2010}, in which an initial small perturbation (tumor) inside a healthy tissue evolves in time keeping a sharp interface; and (ii) a model problem with a Barenblatt-based explicit solution, for which a detailed comparison to the approximated solutions is given, see Section~\ref{sec:numerics}. After the proofs of our results in Section~\ref{sec:proof}, we conclude the article with some conclusions, in Section~\ref{sec:conclusions}. \section{Approximated problems and discretization}\label{sec:app} In this section we compare the approximated solutions to problem (D) constructed via the regularization scheme employed in the proof of Theorem~\ref{th.1}, and those corresponding to the approximated problem introduced in \cite{bertsch2010}. The formulation of these problems is the following, respectively. \begin{equation*} \text{(P)$_\delta$}\left\{ \begin{array}{ll} \partial_t u_i - \partial_x \big(u_i\partial_x (u_1+u_2) + u_iq \big) - \frac{\delta}{2}\partial_{xx}(u_i (u_1+u_2)) = f_i(u_1,u_2) & \text{in } Q_{T},\\ \partial_x u_i=0 & \text{on } \Gamma_{T}, \\ u_i(0,\cdot)=u^{\delta}_{i0} & \text{in }\O, \end{array} \right. \end{equation*} for $\delta>0$, and some non-negative $u^{\delta}_{i0}\in C^1(\bar\O)$ such that $u^{\delta}_{i0}\to u_{i0}$ strongly in $BV(\O)$, for $i=1,2$. \begin{equation*} \text{(P)$_B$}\left\{ \begin{array}{ll} \partial_t u -\partial_x(u (\partial_x u +q)) =F_1(u,r) & \text{in } Q_{T},\\ \partial_t r -(\partial_x u+q) \partial_x r -\delta_B \partial_{xx} r=F_2(u,r) & \text{in } Q_{T},\\ \partial_x u =\partial_x r=0 & \text{on } \Gamma_{T}, \\ u(0,\cdot)=u_{0},\quad r(0,\cdot)=r_0^{\delta_B} & \text{in }\O, \end{array} \right. \end{equation*} for $\delta_B>0$, and for some $r_0^{\delta_B}\in C^1(\bar\O)$ such that $r_0^{\delta_B}\to r_{0}$ strongly in $BV(\O)$, and with \begin{align*} & F_1(u,r)=f_1(ru,(1-r)u)+f_2(ru,(1-r)u), \\ & F_2(u,r)=r(1-r)\Big(\frac{f_1(ru,(1-r)u)}{ru}-\frac{f_2(ru,(1-r)u)}{(1-r)u}\Big). \end{align*} Recall that, according to \cite{bertsch2010}, for any $\delta_B>0$ there exists a regular solution to problem (P)$_B$, $(u^{\delta_B},r^{\delta_B})$, and that this sequence converges strongly in $L^1(Q_T)$, as $\delta_B\to0$, to some $(u,r)$ such that $u_1=ru$, and $u_2=(1-r)u$ are a solution of problem (D). For the numerical discretization of problems (P)$_\delta$ and (P)$_B$, we consider a fully discrete approximation using finite elements in space and backward finite differences in time. The proof of the convergence of the numerical scheme for problem (P)$_\delta$ may be found in \cite{Galiano2014}. We consider a quasi-uniform mesh on the interval $\Omega$, $\{\mathcal{T}_h\} _h$, with $h$ representing step size. We introduce the finite element space of continuous $\mathbb{P}_1$-piecewise elements: $$ S^h = \{ \chi\in \mathcal{C}(\overline{\Omega} ) ; \, \chi |_\kappa \in\mathbb{P}_1\,\text{ for all } \kappa \in\mathcal{T}_h \} . $$ The Lagrange interpolation operator is denoted by $\pi ^h : \mathcal{C}(\overline{\Omega} ) \to S^h$. We also introduce the discrete semi-inner product on $\mathcal{C}(\overline{\Omega} ) $ and its induced discrete seminorm: $$ (\eta_1,\eta_2)^h= \int_{\Omega} \pi^h(\eta_1\eta_2) ,\quad |\eta|_h=\sqrt{(\eta,\eta)^h}. $$ For each $\varepsilon\in (0,1)$ we consider the function \begin{equation} \label{landa} \lambda_\varepsilon(s)=\left\{ \begin{array}{ll} \varepsilon & \text{if }s\leq \varepsilon,\\ s & \text{if } \varepsilon\leq s\leq \varepsilon^{-1},\\ \varepsilon^{-1} & \text{if } s\geq \varepsilon^{-1}, \end{array} \right. \end{equation} and the linear operator $\Lambda _{\varepsilon} : S^h\to L^{\infty}(\Omega)$ given by, for $x_{mp}=(x_j+x_{j+1})/2$ \begin{equation} \label{landamay} \Lambda_\varepsilon(z^h) = \lambda_{\varepsilon}(z^h(x_{mp})), \qtext{in } (x_j,x_{j+1}) . \end{equation} For the time discretization, we take in the experiments a uniform partition of $[0,T]$ of time step $\tau$. For $t=t_0=0$, set $u_{i\varepsilon }^0=u_i^0$, for $i=1,2$. Then, for $n\geq 1$ the full discretization of problem (P)$_\delta$ reads: Find $u_{i\varepsilon }^{n}\in S^h$ such that \begin{align*} \tfrac{1}{\tau}\big( u^n_{i\varepsilon }-u^{n-1}_{i\varepsilon } , \chi )^h + (1+\tfrac{\delta}{2})\big(U_{i\varepsilon}^n \partial_x ( u^n_{1\varepsilon }+u^n_{2\varepsilon }) ,\partial_x \chi \big) + \tfrac{\delta}{2}\big((U_{1\varepsilon}^n+U_{2\varepsilon}^n) \partial_x u^n_{i\varepsilon } ,\partial_x \chi \big) \\ + \big( \pi^h(q) U_{i\varepsilon}^n ,\partial_x \chi \big) = \big( \alpha_i u^n_{i\varepsilon } - \lambda_\varepsilon(u^n_{i\varepsilon })(\beta_{i1}\lambda _{\varepsilon } (u^{n-1}_{1\varepsilon })+\beta_{i2}\lambda _{\varepsilon } (u^{n-1}_{2\varepsilon })), \chi \big)^h , \end{align*} for every $ \chi\in S^h $, where we introduced the notation $U_{i\varepsilon}^n=\Lambda_\varepsilon(u_{i\varepsilon}^n)$. Similarly, the full discretization of problem (P)$_B$ reads: Set $(u_{\varepsilon }^0, r_\varepsilon^0)=(u_0, r_0)$. Then, for $n\geq 1$, find $(u_{\varepsilon }^{n}, r_\varepsilon^n) \in S^h\times S^h$ such that \begin{align*} \tfrac{1}{\tau}\big( u^{n}_{\varepsilon }-u^{n-1}_{\varepsilon } , \chi )^h & + \big(U^{n}_{\varepsilon } \partial_x u^{n}_{\varepsilon } ,\partial_x \chi \big) + \big( \pi^h(q) U^{n}_{\varepsilon } ,\partial_x \chi \big) = \big( F_{1\varepsilon}(u^{n}_{\varepsilon },r^{n}_{\varepsilon }), \chi \big)^h ,\\[2ex] \tfrac{1}{\tau}\big( r^{n}_{\varepsilon }-r^{n-1}_{\varepsilon } , \chi )^h & + \delta_B \big( \partial_x r^{n}_{\varepsilon } ,\partial_x \chi \big) - \big( (\partial_x u^{n}_\varepsilon + \pi^h(q)) \partial_x r^{n}_\varepsilon ,\partial_x \chi \big) \\ & = \big( F_{2\varepsilon}(u^{n}_{\varepsilon },r^{n}_{\varepsilon }), \chi \big)^h . \end{align*} for every $ \chi\in S^h $, where $U_{\varepsilon}^n=\Lambda_\varepsilon(u_{\varepsilon}^n)$, and $F_{i\varepsilon}(s,\sigma)= F_i(\lambda _{\varepsilon } (s),\lambda _{\varepsilon } (\sigma))$. Since the above systems are nonlinear algebraic problems, we use a fixed point argument to approximate their solution at each time slice $t=t_n$, from the previous approximation at $t=t_{n-1}$. Thus, for problem (P)$_\delta$, let $u_{\varepsilon i}^{n,0}=u_{\varepsilon i}^{n-1}$. Then, for $k\geq 1$ the problem is to find $u_{\varepsilon i}^{n,k} \in S^h$ such that for $i=1,2$, and for all $\chi \in S^h$ \begin{align*} \tfrac{1}{\tau}\big( u^{n,k}_{i\varepsilon }-u^{n-1}_{i\varepsilon } , \chi )^h &+ (1+\tfrac{\delta}{2})\big(U_{i\varepsilon}^{n,k-1} \partial_x ( u^{n,k}_{1\varepsilon }+u^{n,k}_{2\varepsilon }),\partial_x \chi \big)\\ & + \tfrac{\delta}{2}\big((U_{1\varepsilon}^{n,k-1}+U_{2\varepsilon}^{n,k-1}) \partial_x u^{n,k}_{i\varepsilon } ,\partial_x \chi \big) + \big( \pi^h(q) U_{i\varepsilon}^{n,k-1} ,\partial_x \chi \big) \\ & = \big( \alpha_i u^{n,k}_{i\varepsilon } - \lambda_\varepsilon(u^{n,k-1}_{i\varepsilon })(\beta_{i1}\lambda _{\varepsilon } (u^{n-1}_{1\varepsilon })+\beta_{i2}\lambda _{\varepsilon } (u^{n-1}_{2\varepsilon })), \chi \big)^h . \end{align*} We then use the stopping criterion $\max _{i=1,2} \nor{u_{\varepsilon,i}^{n,k}-u_{\varepsilon,i}^{n,k-1}}_\infty <\text{tol}$, for values of $\text{tol}$ chosen empirically, and set $u_i^n=u_i^{n,k}$. Similarly, for problem (P)$_B$ we use the scheme \begin{align*} \tfrac{1}{\tau}\big( u^{n,k}_{\varepsilon }-u^{n-1}_{\varepsilon } , \chi )^h & + \big(U^{n,k-1}_{\varepsilon } \partial_x u^{n,k}_{\varepsilon } ,\partial_x \chi \big) + \big( \pi^h(q) U^{n,k-1}_{\varepsilon } ,\partial_x \chi \big) \\ & = \big( F_{1\varepsilon}(u^{n,k-1}_{\varepsilon },r^{n,k-1}_{\varepsilon }), \chi \big)^h ,\\[2ex] \tfrac{1}{\tau}\big( r^{n,k}_{\varepsilon }-r^{n-1}_{\varepsilon } , \chi )^h & + \delta_B \big( \partial_x r^{n,k}_{\varepsilon } ,\partial_x \chi \big) - \big( (\partial_x u^{n,k-1}_\varepsilon + \pi^h(q)) \partial_x r^{n,k-1}_\varepsilon ,\partial_x \chi \big) \\ & = \big( F_{2\varepsilon}(u^{n,k-1}_{\varepsilon },r^{n,k-1}_{\varepsilon }), \chi \big)^h , \end{align*} with an analogous stopping criterion than above. \section{Numerical simulations \label{sec:numerics} In this section, we present numerical simulations for two sets of initial and Lotka-Volterra data aiming to clarify the advantages of approximating the solutions of problem (D) either by the scheme derived from the regularized parabolic-hyperbolic formulation (P)$_B$ of \cite{bertsch2010}, or by the direct viscosity approximation introduced in this article, (P)$_\delta$. The general conclusion is that approximating the solution, $(u_1,u_2)$, of problem (D) by the sequence $(u_{1\delta},u_{2\delta})$ of solutions of problem (P)$_\delta$ is more robust against instabilities produced in discontinuity points than that given by the sequence $(u_{\delta_B},r_{\delta_B})$ corresponding to (P)$_B$. Although we only show the results for two standard examples, we confirmed this conclusion for a variety of data problem. Parameter data was fixed as follows. In all the examples, we consider the spatial domain $\O=(-2,2)$ and investigate the mesh sizes $h=0.04$, $0.013$ and $0.008$, whereas the time step is empirically chosen as $\tau=1.e-3$ (in the first experiment) and $1.e-4$ (in the second). The regularization parameters are taken as $\delta=h^2$, $\delta_B=2h^2$. To avoid negative values of the discrete solutions, we use the functions $\lambda_\varepsilon$ and $\Lambda_\varepsilon$, see \fer{landa} and \fer{landamay}, respectively, setting $\varepsilon=1.e-10$. Finally, the tolerance for the fixed point stopping criterion is set to $\text{tol}=1.e-8$; in all the examples, we observed good convergence properties of the fixed point algorithm, reaching the prescribed tolerance in less than ten iterations. In all the cases, the solutions corresponding to problem (P)$_B$ produce more oscillations than those of problem (P)$_\delta$. To measure these oscillations, we compute an approximation to the zero-crossings of $\partial_x u(t,\cdot)$ in $\O$ by \[ \textrm{osc}(u)(t)= h\sum \abs{\Delta(\textrm{sign}(\Delta u(t,\cdot)))}, \] where $\Delta$ stands for the difference between two consecutive node values, and the sum runs over the spatial nodes. In fact, we may observe that while solutions of (P)$_\delta$ attenuate the oscillations when $t$ increases, those corresponding to solutions of (P)$_B$ are always above some threshold value. In addition, in all the experiments we observe that the size of oscillations diminish according to the mesh size, for both approximations. \subsection{Experiment 1: Invasion} In this example we investigate the question (Q2) stated in \cite{bertsch2010}, concerning the invasion of a population (which simulates mutated abnormal cells) over an initially dominant population (which represents the normal cell). More precisely, we take the initial data $$ u_{10}(x) = 0.22\, \mathrm{exp} (-(x-0.25)^2/0.001)\, , \quad u_{20}(x) = 0.45-u_{10}(x) \qquad\text{for }x\in \O \, . $$ Besides, we deal with the Lotka-Volterra coefficients $$ \alpha_i =1 \, , \quad \beta _{ij}= i \qquad\text{for }i,j=1,2 \, . $$ We conducted the experiments for several choices of the mesh size, $h$. Only the experiments corresponding to $h=0.04$ (101 nodes), $h=0.013$ (301 nodes), and $h=0.008$ (501 nodes) are shown in Figure~\ref{fig_invasion_sol}. We may check that for the finer mesh (last two rows), both approximations give similar results. However, instabilities are already present in the approximation obtained from problem (P)$_B$ for the medium size mesh (row four), which appear amplified for the coarsest mesh (row two). For the printed resolution, instabilities are not visible in the approximation obtained from problem (P)$_\delta$ (rows one and three). This graphical evidences are confirmed through the oscillation measure $\textrm{osc}(u)(t)$, plotted in Figure \ref{fig_invasion_osc}. \begin{figure}[h] \centering {\includegraphics[width=3cm,height=3cm]{figINV_100_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_100_sol_5001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_100_sol_10001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_100_sol_15001.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{figINV_UR_100_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_100_sol_5001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_100_sol_10001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_100_sol_15001.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{figINV_300_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_300_sol_5001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_300_sol_10001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_300_sol_15001.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{figINV_UR_300_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_300_sol_5001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_300_sol_10001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_300_sol_15001.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{figINV_500_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_500_sol_5001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_500_sol_10001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_500_sol_15001.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{figINV_UR_500_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_500_sol_5001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_500_sol_10001.jpg}} {\includegraphics[width=3cm,height=3cm]{figINV_UR_500_sol_15001.jpg}}\\ \caption{ Experiment 1. Odd rows correspond to the approximated solution of problem (P)$_\delta$ for several time slices, $t$, while even rows correspond to the solution of problem (P)$_B$. Rows correspond to different mesh sizes, captured by parameter $h$. Mind the different vertical scales among time slices.} \label{fig_invasion_sol} \end{figure} \begin{figure}[h] \centering {\includegraphics[width=4cm,height=4cm]{figINV_100_osc.jpg}} {\includegraphics[width=4cm,height=4cm]{figINV_300_osc.jpg}} {\includegraphics[width=4cm,height=4cm]{figINV_500_osc.jpg}} \\[2ex] {\includegraphics[width=4cm,height=4cm]{figINV_UR_100_osc.jpg}} {\includegraphics[width=4cm,height=4cm]{figINV_UR_300_osc.jpg}} {\includegraphics[width=4cm,height=4cm]{figINV_UR_500_osc.jpg}} \caption{ Experiment 1. Each row corresponds to the oscillation measure $\textrm{osc}(u)(t)$ of the solutions approximated by (P)$_\delta$ and (P)$_B$, respectively. Columns correspond to different mesh sizes, captured by parameter $h$.} \label{fig_invasion_osc} \end{figure} \subsection{Experiment 2: A Barenblatt-based explicit solution} The contact inhibition problem is a particular case of problem (D) arising in tumor dynamics theory, see \cite{Chaplain2006} for the modeling background. In this problem, the initial density distributions are segregated, satisfying \begin{equation} \label{ci.id} \operatorname{supp} u_{10} \cup \operatorname{supp}u_{20}=\Omega,\quad \operatorname{supp}u_{10}\cap\operatorname{supp}u_{20}=\omega_0, \end{equation} where $\omega_0\subset\Omega$ is the contact hypersurface, one or several points in our one-dimensional simulations. The theory predicts that the initial segregation is kept for later times, that is, no mixing between species is possible. Under our modeling point of view, the initial data segregation can be interpreted as if the process of species splitting (tumor cells and normal cells) took place in coincidence with that of niches segregation: tumor tissue on one side of $\omega_0$, and normal tissue on the other side. For the one-dimensional model, the results of Bertsch et al. \cite{bertsch2010}, or the results proven in the present article include this special case of initial data. For the multi-dimensional model, existence of solutions of problem (D) with initial data satisfying \fer{ci.id} is proven in \cite{bertsch2012,gsv}. These multi-dimensional results are based on the use of suitable Lagrangian formulations of the problem but, as suggested by its parabolic-hyperbolic nature evidenced in the auxiliary formulation introduced in \cite{bertsch2010}, see Problem (P)$_B$, this construction lacks of uniqueness, as proven in \cite{gsv}. Thus, the viscosity approximation used to prove the existence of solutions in the one-dimensional case \cite{Galiano2014} could be a method to select one of the infinitely many solutions obtained by the Lagrangian approach. This, however, is just a conjecture. Finally, notice that the case of general initial data in the multi-dimensional framework is still an open problem. In this example, we consider a particular situation of the contact-inhibition problem where a space discontinuous explicit solution of problem~(ND) may be computed in terms of the Barenblatt explicit solution of the porous medium equation, the Heaviside function and the trajectory of the contact-inhibition point. To be precise, we construct a solution to the problem, \begin{align} & \partial_t u_{i} - \partial_x(u_i \partial_x(u_1+u_2)) =0&& \text{in }Q_T,&\label{eq:s1}\\ &u_i \partial_x(u_1+u_2) = 0 &&\text{on } \Gamma_T,& \label{eq:s2} \end{align} with, for $x,x_0\in\O=(-L,L)$, \begin{equation} \label{def:us} u_{10}(x)=H(x-x_0)B(0,x), \quad u_{20}(x)=H(x_0-x)B(0,x). \end{equation} Here, $H$ is the Heaviside function and $B$ is the Barenblatt solution of the porous medium equation corresponding to the initial datum $B(-t^*,\cdot)=\delta_0$ (Dirac delta function), i.e. \begin{equation*} B(t,x)= 2 (t+t^*)^ {-1/3} \big[1-\frac{1}{12}x^ 2(t+t^*)^ {-2/3}\big]_+, \end{equation*} with $t^* >0$. For simplicity, we consider problem \eqref{eq:s1}-\eqref{def:us} for $T>0$ such that $\rho(T)<L^ 2$, with $\rho(t)=\sqrt{12}(t+t^*)^ {1/3}$, so that $B(t,\pm L)=0$ for all $t\in[0,T]$. The point $x_0$ is the initial contact inhibition point, for which we assume $|x_0|<\rho(0)$, i.e. it belongs to the interior of the support of $B(0,\cdot)$, implying that the initial mass of both populations is positive. \begin{theorem}\label{th.easy} The $L^\infty(0,T;BV(\O))$ functions \begin{equation} \label{def.uib} u_1(t,x)=H(x-\eta(t))B(t,x),\quad u_2(t,x)=H(\eta(t)-x)B(t,x), \end{equation} with $\eta(t)=x_0(1+ \frac{t}{t^*})^{1/3}$, are a weak solution of problem \eqref{eq:s1}-\eqref{def:us} in the following sense: For any $\varphi \in C^1(0,T;H^1(\O))$ with $\varphi(T,\cdot)=0$ in $\O$, \begin{align*} \int_{Q_T} u_i\partial_t \varphi = \int_{Q_T} (u_i\partial_x u ) \partial_x\varphi- \int_{-L}^{L} u_{i0}\varphi(0,\cdot). \end{align*} \end{theorem} With this experiment, we check whether the condition $u_1+u_2>0$ in $Q_T$ needed for proving the convergence of the solutions of problems (P)$_\delta$ and (P)$_B$ to a solution of problem (D) is necessary or not. In particular, we consider $L=2$, $x_0=-0,25$ and $t^*=0.01$, so that $u_1(t,\cdot)+u_2(t,\cdot)=0$ in some regions of $\O=(-2,2)$ for $t<T=0.15$, see Figure~\ref{fig_barenblatt_sol}. In fact, we already have initial conditions such that $u_{10}+u_{20}=0$ in some regions of $\O$. According to this, and on the contrary of problem (P)$_\delta$, the initialization of the data for problem (P)$_B$ requires special care since by construction we have $r_0=u_{10}/(u_{10}+u_{20})$ in $\O$. In this sense, among several choices to define $r_0$ the following gave us the best results: Considering the perturbation, for $\gamma>0$, \begin{equation*} u_{10}^\gamma =\left\{\begin{array}{ll} \gamma+u_{10} & \text{if }x\leq x_0, \\ 0 & \text{if }x > x_0 , \end{array} \right.\quad u_{20}^\gamma =\left\{\begin{array}{ll} 0 & \text{if }x < x_0 ,\\ \gamma+u_{20} & \text{if }x\geq x_0 , \end{array} \right.\ \end{equation*} we get $u_{10}^\gamma+u_{20}^\gamma =\gamma +u_{10}+u_{20} >0$. Then, $r_{0}^\gamma(x)=1$ if $x<x_0$ and $r_{0}^\gamma(x)=0$ if $x>x_0$, and thus, taking $\gamma\to 0$ we obtain the same definition for $r_0$. In Figures~\ref{fig_barenblatt_sol} and \ref{fig_barenblatt_error} we plot the discrete solutions obtained from both approximation schemes together with the explicit solution, and the corresponding oscillation measure. Just as happened in the previous example, in all the cases the solutions corresponding to problem (P)$_B$ produce more oscillations than those of problem (P)$_\delta$, and solutions of (P)$_\delta$ attenuate the oscillations when $t$ increases whereas those corresponding to solutions of (P)$_B$ are always above some threshold value. Moreover, in this example the oscillations come up around the contact inhibition point. It is also interesting to notice that, since this example has an explicitly known solution, we can check the convergence of the numerical solutions to the exact one. Indeed, the relative $L^2$ errors of both solutions are similar, and decreasing with respect to the mesh size, see Figure~\ref{fig_barenblatt_error}, last two rows. \begin{figure}[h] \centering {\includegraphics[width=3cm,height=3cm]{fig_100_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_100_sol_501.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_100_sol_1001.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_100_sol_1501.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{fig_UR_100_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_100_sol_501.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_100_sol_1001.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_100_sol_1501.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{fig_300_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_300_sol_501.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_300_sol_1001.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_300_sol_1501.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{fig_UR_300_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_300_sol_501.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_300_sol_1001.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_300_sol_1501.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{fig_500_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_500_sol_501.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_500_sol_1001.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_500_sol_1501.jpg}}\\ {\includegraphics[width=3cm,height=3cm]{fig_UR_500_sol_1.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_500_sol_501.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_500_sol_1001.jpg}} {\includegraphics[width=3cm,height=3cm]{fig_UR_500_sol_1501.jpg}}\\ \caption{ Experiment 2. Exact solution (dots) against approximated solutions (solid lines). Odd rows correspond to the approximated solution of problem (P)$_\delta$ for several time slices, $t$, while even rows correspond to the solution of problem (P)$_B$. Rows correspond to different mesh sizes, captured by parameter $h$. Mind the different vertical scales among time slices.} \label{fig_barenblatt_sol} \end{figure} \begin{figure}[h] \centering {\includegraphics[width=4cm,height=4cm]{fig_100_osc.jpg}} {\includegraphics[width=4cm,height=4cm]{fig_300_osc.jpg}} {\includegraphics[width=4cm,height=4cm]{fig_500_osc.jpg}}\\ {\includegraphics[width=4cm,height=4cm]{fig_UR_100_osc.jpg}} {\includegraphics[width=4cm,height=4cm]{fig_UR_300_osc.jpg}} {\includegraphics[width=4cm,height=4cm]{fig_UR_500_osc.jpg}}\\ {\includegraphics[width=4cm,height=4cm]{fig_100_error.jpg}} {\includegraphics[width=4cm,height=4cm]{fig_300_error.jpg}} {\includegraphics[width=4cm,height=4cm]{fig_500_error.jpg}}\\ {\includegraphics[width=4cm,height=4cm]{fig_UR_100_error.jpg}} {\includegraphics[width=4cm,height=4cm]{fig_UR_300_error.jpg}} {\includegraphics[width=4cm,height=4cm]{fig_UR_500_error.jpg}} \caption{ Experiment 2. Two first rows correspond to the oscillation measure $\textrm{osc}(u)(t)$ of the solutions approximated by (P)$_\delta$ and (P)$_B$, respectively. Two last rows correspond to the relative error between the exact solution and the solution approximated by (P)$_\delta$ and (P)$_B$, respectively. Columns correspond to different mesh sizes, captured by parameter $h$.} \label{fig_barenblatt_error} \end{figure} \section{Proofs of the theorems}\label{sec:proof} \subsection{Proof of Theorem~\ref{th.1}} We divide the proof in several steps. \noindent\emph{Step 1. An approximation problem. }We consider problem {\text{(P)}$_\delta$}~ for some non-negative $u^{\delta}_{i0}\in C^1(\bar\O)$ with $u^{\delta}_{i0}\to u_{i0}$ strongly in $BV(\O)$, $ \min(u_0)\leq u^{\delta}_{10}+u^{\delta}_{20}\leq \max(u_0)$ in $\O$, and \begin{align} \label{ui0.b} \sqrt{\delta}\nor{\partial_x u^{\delta}_{i0}}_{L^2(\O)} \text{ uniformly bounded.} \end{align} This nonlinear viscosity regularization allows us to use the results in \cite{ggj2,chen} to deduce the existence of a solution $(u^{\delta}_1,u^{\delta}_2)$, with non-negative $u^{\delta}_i \in L^2(0,T;H^1(\O))\cap W^{1,4/3}(0,T;(W^{1,4}(\O))')$, satisfying {\text{(P)}$_\delta$}~ in the weak sense: \begin{align} \int_0^T < \partial_t u^{\delta}_i,\varphi> & +\int_{Q_T} \big(u^{\delta}_i\partial_x u^{\delta} + u^{\delta}_iq +\frac{\delta}{2} \partial_x(u^{\delta}_i u^{\delta}) \big) \partial_x\varphi \nonumber \\ & = \int_{Q_T} f_i(u^{\delta}_1,u^{\delta}_2)\varphi, \label{def.wd} \end{align} for all $\varphi\in L^{4}(0,T;W^{1,4}(\O))\cap L^\infty(Q_T)$. Here, $<\cdot,\cdot>$ denotes the duality product $(W^{1,4}(\O))'\times W^{1,4}(\O)$. Adding the corresponding equations of problem {\text{(P)}$_\delta$}, for $i=1,2$, we find that $u^{\delta}=u^{\delta}_1+u^{\delta}_2$ satisfies, in a weak sense, the following problem {\text{(PS)}$_\delta$} \begin{align*} &\partial_t u^{\delta} -(1+\delta) \partial_x (u^{\delta} \partial_x u^{\delta}) + \partial_x(qu^{\delta} ) = F(u^{\delta}_1,u^{\delta}_2)& & \text{in } Q_T, & \\ &\partial_x u^{\delta} =0 & & \text{on } \Gamma_T, & \\ &u^{\delta}(0,\cdot)=u_{10}^\delta+u_{20}^\delta&& \text{in } \O, & \end{align*} where $F(s_1,s_2)= f_1(s_1,s_2)+f_2(s_1,s_2)$. Using Theorem~2.2 (step 5) of \cite{dgj2001} and the pointwise bounds for $u^{\delta}_{i0}$, we get \begin{align} \label{aux1} \min(u_0)e^{-\lambda t}\leq u^{\delta} (t,\cdot) \leq \max(u_0)e^{\lambda t}\qtext{in }\O,\qtext{for a.e. }t\in (0,T), \end{align} for some $\lambda>0$ independent of $\delta$. Thus, since $u^{\delta}_i\geq 0$ in $\O$, we also get \begin{align} \label{est.uiinf} 0\leq u^{\delta}_i (t,\cdot) \leq \max(u_0)e^{\lambda t}\qtext{in }\O,\qtext{for a.e. }t\in (0,T). \end{align} The uniform estimate \fer{aux1} together with the regularity of $u_0$ and $q$, and the compatibility conditions satisfied by these functions, allow us to apply the general theory for evolution quasilinear equations to deduce that the unique solution of {\text{(PS)}$_\delta$}~is Lipschitz continuous in $Q_T$, and that, for any $1\leq p \leq\infty$, \begin{align} \label{reg.u} \nor{\partial_t u^{\delta}}_{L^p(Q_T)},~\nor{\partial_{xx}u^{\delta}}_{L^p(Q_T)} \text{ are uniformly bounded,} \end{align} see \cite[Theorem 7.20]{lieberman}. Returning to problem {\text{(P)}$_\delta$}, we may see that the additional regularity $u^{\delta}_i\in L^\infty(Q_T)$ given by \fer{est.uiinf} allows us to expand the sense of weak solution to the regularity \begin{align*} u^{\delta}_i \in L^2(0,T;H^1(\O))\cap H^1(0,T;(H^1(\O))') \cap L^\infty(Q_T), \end{align*} and for test functions $\varphi\in L^{2}(0,T;H^1(\O))$. Using the arguments of the proof of \cite[Lemma 3]{Galiano2014}, we easily get \begin{align} \label{est.ui0} \sqrt{\delta}\nor{\partial_x u^{\delta}_i}_{L^2(Q_T)}, ~\nor{\partial_t u^{\delta}_i}_{L^2(0,T;(H^1(\O))')} \qtext{are uniformly bounded.} \end{align} Moreover, the above mentioned regularity $\partial_{xx}u^{\delta}\in L^\infty(Q_T)$ allows us to express \fer{def.wd} as, for all $\varphi\in L^{2}(0,T;H^1(\O))\cap L^\infty(Q_T)$, \begin{align*} \int_0^T <\partial_t u^{\delta}_i , \varphi > & + \frac{\delta}{2} \int_{Q_T} u^{\delta} \partial_xu^{\delta}_i \partial_x \varphi = \int_{Q_T} \psi\varphi, \end{align*} with \begin{align*} \psi = (1+\frac{\delta}{2}) ( \partial_x u^{\delta}_i \partial_x u^{\delta} + u^{\delta}_i \partial_{xx}u^{\delta} ) + u^{\delta}_i \partial_x q +\partial_x u^{\delta}_i q + f_i(u^{\delta}_1,u^{\delta}_2) . \end{align*} Since $\psi\in L^2(Q_T)$, we have that $u^{\delta}_i$ satisfies \begin{align*} \partial_t u^{\delta}_i -\frac{\delta}{2} \partial_{x}(u^{\delta} \partial_xu^{\delta}_i) =\psi \in L^2(Q_T), \end{align*} with $u^{\delta}\in C^{0,1}(\bar Q_T)$ bounded away from zero, see \fer{aux1}. Thus, we have $\partial_t u^{\delta}_i,~\partial_{xx} u^{\delta}_i \in L^2(Q_T)$, implying $\partial_x u^{\delta}_i\in L^4(Q_T)$. We then deduce that, in fact, $\psi \in L^4(Q_T)$, improving in this way the regularity of $u^{\delta}_i$. A boot-strap argument allows us to deduce \begin{align} \label{reg.ui1} \partial_t u^{\delta}_i,~\partial_{xx} u^{\delta}_i \in L^\infty(Q_T), \end{align} implying the Lipschitz continuity of $u^{\delta}_i$ (with norm depending on $\delta$). \bigskip \noindent\emph{Step 2. Uniform estimates for $r^{\delta}=u^{\delta}_1/u^{\delta}$.} For clarity, we omit the super-index $\delta$ in this Step. Due to the regularity \fer{reg.ui1}, the derivatives $\partial_t r$, $\partial_x r$, and $\partial_{xx} r$ are well defined as $L^\infty(Q_T)$ functions. After some computations, we obtain that $r$ satisfies \begin{align} &\partial_t r -\frac{\delta}{2}u \partial_{xx} r - ( q+(1+2\delta) \partial_x u) \partial_x r = G(u,r) & & \text{in } Q_T, & \label{eq.r} \\ &\partial_x r =0 & & \text{on } \Gamma_T, & \nonumber \\ &r(0,\cdot)=u_{10}/u_0&& \text{in } \O,\nonumber & \end{align} with, \begin{align*} G(u,r)= \frac{1-r}{u} f_1(ur,u(1-r))- \frac{r}{u}f_2(ur,u(1-r)). \end{align*} Since $\partial_u G$ and $\partial_r G$ are bounded, and $u$ and $r$ are Lipschitz continuous, we deduce $\partial_x G(u,r)\in L^\infty(Q_T)$. Therefore, the solution $r$ of \fer{eq.r} is regular enough to allow us to differentiate equation \fer{eq.r} with respect to $x$, for a.e. $(t,x)\in Q_T$. We then multiply the differentiated equation by $\xi= \partial_x r /\sqrt{\delta+ (\partial_x r)^2}$ and integrate by parts to get, term by term, \begin{align*} & \int_{Q_T} \partial_x \partial_t r \xi = \int_\O \sqrt{\delta+ (\partial_x r)^2} \Big|_{0}^{T} ,\\ -\frac{\delta}{2} & \int_{Q_T} \partial_x\big(u\partial_{xx}r \big)\xi = \frac{\delta^2}{2} \int_{Q_T} \frac{u(\partial_{xx} r)^2}{(\delta+ (\partial_x r)^2)^{3/2}} ,\\ -& \int_{Q_T} \partial_x \big( ( q+(1+2\delta) \partial_x u) \partial_x r \big) \xi = \delta \int_{Q_T}\frac{\partial_x ( q+(1+2\delta) \partial_x u)}{\sqrt{\delta+ (\partial_x r)^2}} ,\\ & \int_{Q_T} \partial_x G(u,r) \xi = \int_{Q_T} (\partial_u G(u,r) \partial_x u +\partial_r G(u,r) \partial_x r ) \xi. \end{align*} Therefore, thanks to the regularity $\partial_x q \in L^1(Q_T)$ and to the uniform bounds for $\partial_x u$ and $\partial_{xx} u$ in $L^1(Q_T)$ we obtain \begin{align*} \int_\O \sqrt{\delta+ (\partial_x r)^2} \Big|_{0}^{T} \leq c_1 + c_2 \int_{Q_T} \abs{\partial_x r}, \end{align*} with $c_1,c_2$ independent of $\delta$. We then deduce from Gronwall's lemma that \begin{align} \label{est.rx1} \sup_{[0,T]}\int_{\O} \abs{\partial_x r} \qtext{is uniformly bounded with respect to }\delta. \end{align} We finish this step by showing a uniform bound for $\partial_t r$. From equation \fer{eq.r} we get \begin{align} \label{est.rt} \int_{Q_T} \abs{\partial_t r} \leq \frac{\delta}{2}\nor{u}_{L^\infty(Q_T)} \int_{Q_T}\abs{\partial_{xx} r} + c_3 \int_{Q_T} \abs{\partial_x r }+ \int_{Q_T} \abs{G(u,r)}, \end{align} with $c_3=\nor{q}_{L^\infty(Q_T)} +(1+2\delta) \nor{\partial_x u}_{L^\infty(Q_T)}$. We have already shown that the two last terms of the right hand side of \fer{est.rt} are uniformly bounded. For estimating the first term, we multiply equation \fer{eq.r} by $\delta \partial_{xx}r$ and integrate by parts. We obtain \begin{align*} \min(u) \frac{\delta^2}{2}\int_{Q_T} \abs{\partial_{xx} r}^2 \leq & \frac{\delta}{2}\int_\O \abs{\partial_x r(0,\cdot)}^2 + \frac{\delta(1+2\delta)}{2} \int_{Q_T} (\abs{\partial_{xx} u}+\abs{\partial_{x} q}) \abs{\partial_{x} r}^2 \\ &+ \delta \int_{Q_T} \abs{G(u,r)}\abs{\partial_{xx} r}. \end{align*} Using property \fer{ui0.b}, we find that the first term of the right hand side is uniformly bounded. The third term may be handled by H\"older's and Young's inequality to get \begin{align*} \min(u) \frac{\delta^2}{4}\int_{Q_T} \abs{\partial_{xx} r}^2 \leq & c_4 + \frac{\delta(1+2\delta)}{2} \int_{Q_T} (\abs{\partial_{xx} u}+\abs{\partial_{x} q}) \abs{\partial_{x} r}^2 , \end{align*} with $c_4$ independent of $\delta$. Finally, using the uniform lower bound for $u$ in $Q_T$, the uniform bound for $\partial_{xx} u$ in $L^2(Q_T)$, the regularity $\partial_x q \in L^2(Q_T)$ and applying the last argument of point (v) of the proof of Lemma 3.1 of \cite{bertsch2010}, we deduce that $\delta \nor{\partial_{xx} r}_{L^1(Q_T)}$ is uniformly bounded, and therefore, from \fer{est.rt} we also deduce \begin{align} \label{est.rx2} \nor{\partial_{t} r}_{L^1(Q_T)} \qtext{is uniformly bounded with respect to }\delta. \end{align} \noindent\emph{Step 3. Passing to the limit $\delta\to0$. } From the uniform estimates \fer{est.uiinf} and \fer{est.ui0}, we deduce the existence of $u_i\in L^\infty(Q_T) \cap L^2(0,T;(H^1(\O))') $ such that (for a subsequence, not relabeled) \begin{align*} & u^{\delta}_i\rightharpoonup u_i \qtext{weakly * in } L^\infty(Q_T),\\ & \partial_t u^{\delta}_i\rightharpoonup \partial_t u_i \qtext{weakly in } L^2(0,T;(H^1(\O))'),\\ & \delta \partial_x u^{\delta}_i\to 0 \qtext{strongly in } L^2(Q_T), \end{align*} and from \fer{reg.u}, we also deduce the existence of $u\in C^{0,1}(\bar Q_T)\cap L^2(0,T;H^2(\O))$ such that \begin{align*} & u^{\delta}\to u \qtext{uniformly in } C^{0,1}(\bar Q_T),\\ & \partial_x u^{\delta}\to \partial_x u \qtext{strongly in } L^2(Q_T). \end{align*} In particular, the weak and strong convergences of $u^{\delta}_i$ and $u^{\delta}$ imply $u=u_1+u_2$. Using the uniform estimates \fer{est.rx1} and \fer{est.rx2} we deduce the existence of $r\in L^\infty(0,T;BV(\O))\cap BV(0,T;L^1(\O))$ such that, for $1\leq p<\infty$, \begin{align*} & r^{\delta}\to r \text{ strongly in } L^p(Q_T). \end{align*} Since $r^{\delta}=u^{\delta}_1/u^{\delta}$ the weak and strong convergences of $u^{\delta}_i$ and $u^{\delta}$ in $L^2(Q_T)$, respectively, imply $r=u_1/u$. Then, the strong convergences of $r^{\delta}$ and $u^{\delta}$ in $L^p(Q_T)$ and $L^\infty(Q_T)$ imply \begin{align*} u^{\delta}_1=\rdu^{\delta} \to ru=u_1 \text{ strongly in } L^p(Q_T). \end{align*} Similarly, we obtain \begin{align*} u^{\delta}_2=(1-r^{\delta})u^{\delta} \to (1-r)u=u_2 \text{ strongly in } L^p(Q_T). \end{align*} Finally, using again the uniform estimates \fer{est.rx1} and \fer{est.rx2}, we also obtain \begin{align*} \nor{\partial_x u^{\delta}_i}_{L^\infty(0,T;L^1(\O))},\nor{\partial_t u^{\delta}_i}_{L^1(Q_T)}\text{ uniformly bounded,} \end{align*} from where we, additionally, deduce \begin{align*} u_i\in L^\infty(0,T;BV(\O))\cap BV(0,T;L^1(\O)). \end{align*} Thus, the passing to the limit $\delta\to 0$ in \fer{def.wd} is justified, and so we identify the limit $(u_1,u_2)$ as a weak solution of problem (D). $\Box$ \subsection{Proof of Theorem~\ref{th.easy}} \noindent\emph{Proof. } Let $H_\epsilon$ be the regularization of the Heaviside function taking the values $\left\{0,\frac{1}{2}(1+x/\epsilon),1\right\}$ in the intervals $(-L,-\epsilon)$, $(-\epsilon,\epsilon)$ and $(\epsilon,L)$, respectively, for $\epsilon>0$ small. Define the functions $u_i^\epsilon$ as in \fer{def.uib}, with $H$ replaced by $H_\epsilon$, and let $\varphi\in H^ 1(Q_T)$, with $\varphi(T,\cdot)=0$ in $\O$. Using $\varphi_\epsilon (t,x)=\varphi(t,x) H_\epsilon(x-\eta(t)) $ as a test function in the weak formulation of \fer{eq:s1}-\fer{eq:s2} corresponding to the initial datum $B(0,\cdot)$, that is, the problem satisfied by the Barenblatt solution in $Q_T$, we obtain \begin{align} \label{weak.B} \int_{Q_T} H_\epsilon(x-\eta(t)) B(t,x) (\partial_t\varphi(t,x) & - \partial_x B(t,x) \partial_x\varphi(t,x) )dxdt \\ & +\int_{-L}^ {L} H_\epsilon(x-x_0) B(0,x) \varphi(0,x)dx =I^ 1_\epsilon, \nonumber \end{align} with \begin{eqnarray*} I^ 1_\epsilon = \int_{Q_T} \varphi(t,x) B(t,x) H'_\epsilon(x-\eta(t)) \big(\eta'(t) + \partial_x B(t,x)\big) dx dt. \end{eqnarray*} Since $|x_0|<\rho(0)$, we have $\abs{\eta(t)}<L^2-\epsilon$, for $\epsilon$ small enough and $t\in(0,T)$. Using the explicit expression of $\partial_x B$ and $\eta'$ we deduce \begin{equation*} I_\epsilon^1=-\frac{1}{6\epsilon}\int_{0}^T\int_{-\epsilon}^\epsilon y \varphi(t,y+\eta(t))B(t,y+\eta(t))dydt. \end{equation*} Since $\varphi$ and $B$ are uniformly bounded in $L^\infty(Q_T)$, we obtain \begin{equation} \label{est.I} |I^ 1_\epsilon|\leq C \epsilon , \end{equation} with $C>0$ independent of $\epsilon$. The computation using $\varphi(t,x)H_\epsilon(\eta(t)-x)$ as test function gives similar results for some $I_\epsilon^ 2$ satisfying the same estimate \eqref{est.I} than $I_\epsilon^ 1$. Since, by definition, $u_1^\epsilon+u_2^\epsilon=B$, we obtain \begin{equation*} \left|\int_{Q_T} u_{i}^ \epsilon (\partial_t \varphi - \partial_x(u_1^ \epsilon+u_2^ \epsilon ) \partial_x\varphi) + \int_{-L}^ {L} (u_i^{\epsilon} \varphi)(0,\cdot) \right| \leq C\epsilon \end{equation*} for all $\varphi\in H^ 1(Q_T)$. Thus, using that $u_i^ \epsilon \to u_i$ uniformly in $L^ \infty (Q_T)$, and performing the limit $\epsilon\to 0$ in \fer{weak.B}, we deduce that $(u_1,u_2)$ is a solution of problem \fer{eq:s1}-\fer{def:us}. $\Box$ \begin{remark} It is not difficult to extend the above construction to problem~(ND) for non-vanishing $q$ and $f_i$. Indeed, reconsider the solution $u$ of problem \fer{p.u} and the corresponding approximations $u_i^\epsilon$ given in the proof of the previous theorem. Then, to handle the integrals $I_\epsilon^i$, we first observe that for $\epsilon\to0$ we get \begin{eqnarray*} I^ 1_\epsilon \to -\int_0^T \varphi(t,\eta(t)) u(t,\eta(t)) \big(\eta'(t) - b(t,\eta(t))\big) dt, \end{eqnarray*} with $b = -(\partial_x u + q).$ Therefore, if the ODE problem \begin{equation} \label{eq.eta2} \left\{ \begin{array}{ll} \eta'(t)=b(t,\eta(t)) & \text{for }t\in (0,T), \\ \eta(0)=x_0, \end{array} \right. \end{equation} is solvable, a solution for problem (ND) may be constructed as before. Typical conditions on $b$ for \eqref{eq.eta2} to be solvable (in the one-dimensional case) are given in terms of spatial Lipschitz continuity. That is, the solution of problem \fer{p.u} must satisfy $\partial_{xx}u\in L^\infty(Q_T)$, which is true for smooth data. \end{remark} \section{Conclusions}\label{sec:conclusions} The generalization of the dynamical system model for the splitting - differentiation process of populations \cite{sanchez-palencia} to the space dependent situation leads to a family of cross-diffusion PDE problems including several well known segregation models. In a previous work \cite{Galiano2012}, we analyzed the case in which the cross-diffusion is of the type introduced by Shigesada et al. \cite{SKT79}, a model well studied in the literature, see \cite{Galiano2014} and its references. In this article, we turned to another important case, the model of Busenberg and Travis \cite{busenberg83}, in which the segregation effects are stronger, as motivated in \cite{Galiano2014}. The model of Busenberg and Travis was studied by Bertsch et al. in a series of papers \cite{bertsch85,bertsch2010,bertsch2012} focusing in a special case of initial data arising from Tumor modeling which gives rise to the so-called contact-inhibition problem. From the analytical side, we have produced a proof of existence of solutions conceptually simpler than previous proofs \cite{bertsch2010,gsv}, since ours derives from a direct parabolic regularization of the problem meanwhile previous involved a change of unknowns rendering the problem to a parabolic-hyperbolic formulation. Although difficult to ensure, we believe that our approach also gives a way to select a unique \emph{natural} solution of the problem as a limit of vanishing viscosity solutions of the regularized parabolic problems, tackling thus the non-uniqueness issue of the parabolic-hyperbolic formulation \cite{gsv}. From the numerical side, we have introduced a Finite Element discretization of both formulations and compared the results in a series of experiments (only two of them showed in the article). The general observation is that our approach is always more stable in the tricky regions where the solutions exhibit discontinuities. Finally, there are evidences in our numerical experiments showing that the main data restriction of the existence theorems, the positivity of the initial total mass, is not a necessary condition. Thus, future research will focus in the replacement of this condition, as well as in the generalization to the multi-dimensional case.

https://arxiv.org/abs/2201.11004

Actions of nilpotent groups on complex algebraic varieties

We study nilpotent groups acting faithfully on complex algebraic varieties. We use a method of base change. For finite p-groups, we go from $k$, a number field, to a finite field in order to use counting lemmas. We show that a finite $p$-group of polynomial automorphisms of $k^d$ is isomorphic to a subgroup of GL$_d(k)$. For infinite groups, we go from $\mathbb{C}$ to $\mathbb{Z}_p$ and use p-adic analytic tools and the theory of p-adic Lie groups. We show that a finitely generated nilpotent group $H$ acting faithfully on a complex quasiprojective variety $X$ of dimension $d$ can be embedded into a $p$-adic Lie group acting faithfully and analytically on $\mathbb{Z}_p^d$; we deduce that $d$ is larger than the virtual derived length of $H$.

\section{Introduction} \subsection{Minkowski's bound for polynomial automorphisms.}\label{par:Minkowski_Schur} \paragraph{Rational numbers.--} Let $p$ be a prime. A finite \emph{$p$-group} is a group of size $p^\alpha$ for some integer $\alpha \geq 0$. For $d\in \mathbf{Z}_+$, define $M_\mathbf{Q}(d,p)$ to be the integer \[ M_\mathbf{Q}(d,p) = \ent{\frac{d}{p-1}} + \ent{\frac{d}{p(p-1)}} + \ent{\frac{d}{p^2 (p-1)}} + \cdots \] (Here $M$ stands for Minkowski). Let $v_p$ be the $p$-adic valuation; then $M_\mathbf{Q}(d,p) = \ent{\frac{d}{p-1}} + v_p \left( \ent{\frac{d}{p-1}} ! \right)$. \begin{thm}[Minkowski 1887, see \cite{SerreBoundsOrder}]\label{MinkowskiBoundLinear} Let $d$ be a natural number and let $p$ be a prime. If $G$ is a finite $p$-subgroup of ${\mathrm{\mathrm{GL}}}_d (\mathbf{Q})$, then $v_p(\vert G\vert)\leq M_\mathbf{Q}(d,p)$, and this upper bound is optimal: there are groups of order $p^{M_\mathbf{Q}(d,p)}$ in ${\mathrm{\mathrm{GL}}}_d(\mathbf{Q})$. \end{thm} \paragraph{Number fields.--} Schur extended Minkowski's result to the case of number fields. To state Schur's result, let us introduce some notation for cyclotomic extensions. Consider a number field $\k$ and fix an algebraic closure ${\overline{\k}}$ of $\k$. Denote by $z_a\in {\overline{\k}}$ any primitive $a$-th root of unity, for $a$ any positive integer; for instance $z_4={\mathsf{i}}$, a square root of $-1$. \begin{itemize} \item If $p \geq 3$, set $t(\k;p) = [\k(z_p) : \k]$ and let $m(\k;p)$ be the maximal integer $a$ such that $\k(z_p)$ contains $z_{p^a}$; note that $m(\k; p)$ is finite because $\k$ is a finite extension of $\mathbf{Q}$. Then, define \[ M_\k(d,p) := m(\k; p) \cdot \ent{\frac{d}{t(\k;p)}} + \ent{\frac{d}{p \cdot t(\k;p)}} + \ent{\frac{d}{p^2t(\k;p)}} + \cdots . \] \item If $p=2$, set $t(\k; 2) =[\k(z_4): \k]$ and let $m(\k; 2)$ be the largest integer $a$ such that $z_{2^a} \in \k(z_4)$. Define \[ M_\k (d,2) = d + (m(\k; 2)-1) \ent{\frac{d}{t(\k;2)}} + \ent{\frac{d}{2t(\k;2)}} + \ent{\frac{d}{4t(\k;2)}} + \cdots . \] \end{itemize} This definition is consistent with the definition of $M_\mathbf{Q} (d,p)$ given above. \begin{thm}[\cite{schur1973klasse}, \cite{SerreBoundsOrder}]\label{SchurBound} Let $d$ be a natural number, and let $p$ be a prime. If $G$ is a finite $p$-subgroup of ${\mathrm{\mathrm{GL}}}_d (\k)$ then $v_p(\vert G\vert)\leq M_\k(d,p)$ and this bound is optimal. \end{thm} It is not difficult to find a subgroup $G \subset \mathrm{\mathrm{GL}}_d (\k)$ such that $\abs G = p^{M(d,p)}$. We recall how to do so in Proposition \ref{PropBorneOptimale}. \paragraph{Polynomial automorphisms.--} Our first goal is to extend the theorem of Minkowski and Schur to an algebraic, but nonlinear context. Let $\Aut(\mathbf{A}_\k^d)$ be the group of polynomial automorphisms of the affine space $\mathbf{A}^d$, over some number field $\k$. This group contains $\mathrm{GL}_d (\k)$ but it is much more complicated. Surprisingly, we are able to show that the Minkowski-Schur bound still holds for subgroups of $\Aut (\mathbf{A}_\k^d)$ and in fact the same finite subgroups appear. \begin{bigtheorem}\label{SchurBoundPolynomial} Let $\k$ be a number field, let $d$ a natural number, and let $p \geq 3$ be a prime. If $G$ is a finite $p$-subgroup of $\Aut(\mathbf{A}_\k^d)$, then there exists a group embedding $ G \hookrightarrow \mathrm{GL}_d(\k).$ In particular, Schur's bound still holds: \[ v_p (\abs G) \leq M_\k(d,p), \] and this bound is optimal. \end{bigtheorem} The proof first shows the bound on the cardinal of the group $G$ and we then find the group embedding $G \hookrightarrow \mathrm{GL}_d(\k)$ using a Sylow argument. \begin{rmq} \label{remark:CasPImpairPasOptimal} The case $p=2$ is also dealt with in Section 2. But we don't get an optimal bound. For example for $p=2$ and $\k = \mathbf{Q}$, we show that any $2$-subgroup $G$ of $\Aut(\mathbf{A}_\mathbf{Q}^d)$ can be embedded into $\mathrm{GL}_d(\mathbf{Q}(z_4))$ and therefore satisfies $v_2(\abs G) \leq M_\mathbf{Q}(d,2) + \ent{\frac{d}{2}}$. More precisely, Proposition \ref{ImageCaractereCyclotomique} defines three cases $(a), (b)$ and $(c)$ when $p=2$. We get an embedding into $\mathrm{GL}_d(\k)$ in case $(a)$ and $(b)$ (this is the case for example if $\k$ contains $z_4$), but in case $(c)$ we can only get an embedding of $G$ into $\mathrm{GL}_d(\k(z_4))$ and therefore we get the bound $ v_2(\abs G) \leq M_\k(d,2) + \ent{\frac{d}{2}} = M_{\k(z_4)}(d,2) $. See Theorem \ref{BigThmSchurBoundPolynomial} page \pageref{BigThmSchurBoundPolynomial} for the general statement. In fact, Theorem \ref{SchurBoundPolynomial} still holds when $\k$ is a finitely generated field over $\mathbf{Q}$ but the proof is less intuitive so we will show the proof for $\k$ a number field and explain how to extend it to finitely generated field over $\mathbf{Q}$ in Remark \ref{RmqFinitelyGeneratedField}. We then state the complete theorem for finitely generated fields over $\mathbf{Q}$ in Theorem \ref{BigThMSchurBoundPolynomialGeneralCase} page \pageref{BigThMSchurBoundPolynomialGeneralCase}. \end{rmq} Our method of proof follows \cite{SerreBoundsOrder}, in which Serre bounds the order of the finite subgroups of ${\mathrm{H}}(\k)$, for ${\mathrm{H}}$ a semi-simple algebraic group; the phenomenon mentioned in Remark \ref{remark:CasPImpairPasOptimal} also appears for such groups ${\mathrm{H}}$. The general idea is to embed $G$ into a group of linear automorphisms over a finite field, study the finite field case, and use cyclotomic characters to find the optimal bound yield by this method. \paragraph{Birational transformations.--} The problem of the existence of uniform bounds on the size of finite $p$-groups or finite simple groups in infinite dimensional groups such as $\Aut(\mathbf{A}^d)$ or $\Bir(\mathbf{A}^d)$ has been studied extensively during the last decade (see~\cite{Serre09}). For an arbitrary complex projective variety $X$, one cannot expect uniform bounds that would only depend on the dimension of $X$, since every finite group is the group of automorphisms of a complex projective curve (see~\cite{Greenberg}). But precise results have been obtained when $X$ is rationally connected. Recently, Jinsong Xu showed the following optimal result: {\sl{Let $d$ be a natural number and let $p$ be a prime $>d+1$. If $X$ is a rationally connected variety of dimension $d$ over an algebraically closed field of characteristic $0$, and $G$ is a finite $p$-subgroup of $\Bir(X)$, then G is abelian and its rank is at most $d$}} (see~\cite{Jinsong_Xu:CRAS_2020}). Results of this type were first shown by Prokhorov, Shramov and Birkar in \cite{prokhorov_shramov_2014} for birational transformations of any varieties and improvements were made for rationally connected varieties in \cite{prokhorov2016jordan}. These results are deeper than our Theorem \ref{SchurBoundPolynomial}, but our contribution has a few advantages: it may serve as an introduction to the work of Prokhorov and Shramov, the techniques are more elementary, the precise bound we obtain illustrates the interplay between the arithmetic of the field $\k$ and the size of the group, and the proof shows why the upper bound of Minkowski and Schur is still valid in $\Aut(\mathbf{A}_\k^d)$. \begin{rmq} The results of Prokhorov and Shramov rely on the BAB conjecture, which was proved by Birkar in \cite{birkar2016singularities}. The result of J. Xu relies on the work of Haution on equivariant cohomology and fixed points of finite groups (see~\cite{haution_2019}). \end{rmq} \subsection{A bound for the action of finitely generated nilpotent groups} \subsubsection{Nilpotent and solvable groups}\label{par:nilpotent_and_solvable} Let~$H$~be a group. If~$a,b \in H$, we denote by~$[a,b] := ab \inv a \inv b$~their commutator. If~$H_1,H_2$~are two subgroups of~$H$, then we denote by~$H_1 H_2$~the subgroup generated by the set~$\{h_1 h_2: h_1 \in H_1, h_2 \in H_2\}$~and by $[H_1,H_2]$ the subgroup generated by the set~$\{[h_1, h_2]: h_1 \in H_1, h_2 \in H_2\}$. The lower central (resp. derived) series is defined by~$D^{0}(H) = H$~(resp.~$D_0(H) = H$)~and~$D^{i+1}(H) = [H, D^{i}(H)]$~(resp. $D_{i+1}(H) = [D_i(H), D_i (H)]$). A group $H$ is \emph{nilpotent} (resp. \emph{solvable}) when there exists an integer $k$ such that $D^k(H) = {1}$ (resp. $D_k(H) = {1}$). If $H$ is nilpotent, its {\emph{nilpotency class}} $\nilp(H)$~is the lowest integer such that $D^k(H) = {1}$. For a solvable group $H$, denote by $\dl (H)$ its derived length, that is the least integer $k$ such that $D_k(H) = {1}$. The \emph{virtual derived length} is the minimum of $\dl(H_0)$ over finite index subgroups $H_0$ of $H$. Similar definitions and notation will be used for Lie algebras. \subsubsection{Upper bounds on the virtual derived length} Finite $p$-groups are nilpotent. We now look at infinite, finitely generated nilpotent groups, and their actions by automorphisms and birational transformations. In \cite{cantat2014algebraic}, Cantat and Xie used $p$-adic analysis to give information on group actions on complex algebraic varieties by birational transformations, and sketched the proof of the following result. \begin{bigtheorem}\label{BoundNilpotentGroups} Let $H$ be a finitely generated nilpotent group acting faithfully on a quasi-projective variety $X$ by algebraic automorphisms over a field of characteristic zero. Then, \[ \vdl(H)\leq \dim X \] where $\vdl(H)$ is the \emph{virtual derived length} of $H$. Furthermore, this bound is optimal. \end{bigtheorem} Another goal of this paper is to give a complete proof of this result. Again, the main idea is to replace the initial field of definition by another one, here $\mathbf{Q}_p$, and in fact by $\mathbf{Z}_p$, for a suitable prime $p$. Then, the initial action of the discrete group $H$ will be extended to an analytic action of a $p$-adic Lie group over $\mathbf{Z}_p^{\dim X}$, so that tools from $p$-adic analysis will be available, in particular $p$-adic analytic vector fields and $p$-adic Lie algebras. Thus, Theorem \ref{BoundNilpotentGroups} will follow from a similar theorem we prove over $\mathbf{Z}_p$. Section \ref{SecPAdicAnalysis} is dedicated to the construction of $p$-adic analytic tools needed for the proof of Theorem \ref{BoundNilpotentGroups} such as infinite dimensional $p$-adic Lie groups or Tate-analytic diffeomorphisms and Section \ref{SecFinitelyGeneratedNilpotentGroups} is dedicated to the proof of Theorem \ref{BoundNilpotentGroups}. \section{Finite $p$-groups} \subsection{Preliminaries} \paragraph{Primes and $p$-adic numbers} In the rest of the article, $p$ is a prime unless mentioned otherwise, $\mathbf{Z}_p$ denote the ring of $p$-adic integers and $\mathbf{Q}_p$ is the fraction field of $\mathbf{Z}_p$. Recall that Dirichlet's theorem states for any integers $a,n$ such that $gcd(a,n) =1$, there is an infinite amount of prime numbers $\ell$ such that $\ell = a \mod n$. \paragraph{Maximal ideals and reduction} If $q$ is a power of a prime, we denote by $\mathbf{F}_q$ the field with $q$ elements. Let $A$ be a finitely generated $\mathbf{Z}$-algebra. Then for every maximal ideal $\m \subset A$, $A / \m$ is a finite field. This comes from the Nullstellensatz for Jacobson rings which is proven in \cite{bourbaki2007algebre}, chapter 5, $\S 3$, theorem 3 of section 4. \subsection{Groups of linear transformations over $\mathbf{Q}$} To warm up, let us prove the theorem of Minkowski. For a ring $A$, we denote by $A^{\times}$ its subgroup of invertible elements; for any prime $p$ the group $(\mathbf{Z} / p^2 \mathbf{Z})^{\times}$ is cyclic. \begin{prop}\label{PropEmbeddingFinite} Let $G$ be a finite subgroup of $\mathrm{GL}_d (\mathbf{Q})$. For any prime $\ell$ large enough there exists an injective homomorphism $G \hookrightarrow \mathrm{GL}_d (\mathbf{F}_\ell)$. \end{prop} \begin{proof} Since $G$ is finite, there exists an integer $N$ such that $G \subset \mathrm{GL}_d (\mathbf{Z}[1 / N])$. Now, for each $g \in G \setminus \left\{ \id \right\}$ denote by $l(g)$ the largest prime factor that appears in the prime decomposition of the rational numbers given by the coefficients of the matrix $g - \id$; denote by $L$ the maximum of the primes $l(g)$. If $\ell > \max(N,L)$, the homomorphism of reduction modulo $l$ is defined on $G$ and is injective. \end{proof} Thus, if $G \subset \mathrm{GL}_d ( \mathbf{Q})$ is a finite subgroup, $v_p(\abs G) \leq v_p (\abs{\mathrm{GL}_d(\mathbf{F}_\ell)})$ for any $\ell$ given by Proposition \ref{PropEmbeddingFinite}. We know that \begin{equation} \abs{ \mathrm{GL}_d (\mathbf{F}_\ell)} = \ell^{d(d-1)/2} \prod_{i=1}^{d-1} \left( \ell^i -1 \right). \label{EqCardinalGLd} \end{equation} for any prime $\ell$. Let us compute the $p$-adic valuation of such a product. \begin{lemme} \label{CalculValuationsP-Adiques} Suppose $p\neq 2$ and let $\ell$ be a generator of $(\mathbf{Z} / p^2 \mathbf{Z})^\times$. \begin{enumerate} \item If $p$ divides $\ell^i -1$ then $p-1$ divides $i$; \item If $p-1$ divides $i$ then $v_p(\ell^i-1) = 1 + v_p(i)$. \end{enumerate} \end{lemme} \begin{proof} Suppose $p$ divides $\ell^i-1$. Note that $\ell^{ip}-1 = (\ell^i -1) \sum_{j=0}^{p-1} \ell^{ij}$; since $\ell^i \equiv 1 \mod p$, we have $\sum_{j=0}^{p-1} \ell^{ij} \equiv 0 \mod p$, and then $\ell^{ip} \equiv 1 \mod p^2$. Since $\ell$ is of order $p(p-1)$ in $(\mathbf{Z} / p^2 \mathbf{Z})^{\times}$, we have that $p(p-1)$ divides $ip$ therefore, $p-1$ divides $i$, which proves the first assertion. We prove assertion 2 by induction on $v_p(i)$. To initialize the induction assume $v_p(i) =0$. Then $p$ and therefore $p(p-1)$ do not divide $i$; thus $\ell^i \not \equiv 1 \mod p^2$ because $\ell$ is of order $p(p-1)$. Thus, $v_p(\ell^i -1) =1$. Now suppose the assertion true for $v_p(i) = k$ with $k \geq 0$ and suppose $v_p(i) = k+1$. Write $i = (p-1) p^{k+1} m$ with $m$ not divisible by $p$ and suppose the result true for $v_p(i) = k$. Let $s:= \ell^{(p-1)m}$, then \begin{align*} \ell^{i} -1 = s^{p^{k+1}} -1 = (s^{p^k}-1) \sum_{j=0}^{p-1} s^{j p^k}. \end{align*} By induction, $s^{p^k}$ is of the form $s^{p^k} = 1 + u p^{k+1}$ where $u$ is an integer not divisible by $p$. Therefore, for all $1 \leq j \leq p-1, s^{jp^k} = 1 + j p^{k+1} u + v_jp^2$ where $v_j$ is some integer. , therefore we can write \[ \sum_{j=0}^{p-1} s^{jp^k} = p + p^{k+1} \frac{p(p-1)}{2} u + p^2V = p \left(1 + p^{k+1} \frac{p-1}{2} u + pV \right) \] where $V = \sum v_j$. Since $p$ is odd, $\frac{p-1}{2}$ is an integer and this sum has $p$-adic valuation $1$ since $k+1 \geq 1$. \[ s^{p^{k+1}} -1 = (s^{p^k} -1) \cdot p \left(1 + p^k\sum_{j=0}^{p-1} u_j \right) \] since $k \geq 1$, we get $v_p (s^{p^{k+1}} -1) = 1 + v_p (s^k -1 ) = 1 + (k+1)$. \end{proof} Equation \eqref{EqCardinalGLd} and Lemma \ref{CalculValuationsP-Adiques} provide the following corollary. \begin{cor} Let $d$ be an integer, let $p$ be an odd prime, and let $\ell$ be a prime whose image in $(\mathbf{Z} / p^2 \mathbf{Z})^{\times}$ is a generator. Then \[ v_p (\mathrm{GL}_d(\mathbf{F}_\ell)) = M_\mathbf{Q}(d,p). \] This proves also the fact that Theorem \ref{MinkowskiBoundLinear} "is optimal for $\mathrm{GL}_d(\mathbf{F}_\ell)$" by Sylow. \end{cor} To prove Theorem \ref{MinkowskiBoundLinear}, consider a finite group $G \subset \mathrm{GL}_d (\mathbf{Q})$, then apply Dirichlet's theorem and Proposition \ref{PropEmbeddingFinite} to embed $G$ in $\mathrm{GL}_d(\mathbf{F}_\ell)$ for some prime generator $\ell$ of $(\mathbf{Z} / p^2 \mathbf{Z})^*$. The corollary gives the desired upper bound. \begin{rmq}\label{remarkCasPairLinear} The case $p=2$ is also treated by Minkowski and in fact the same bound applies. However the proof is slightly different as it is required to embed $G$ into an orthogonal group over a finite field. Indeed, If $\ell$ is an odd prime then the best bound one can get is $v_2(\mathrm{GL}_d(\mathbf{F}_l)) \leq M(d,2) + \lfloor d/2 \rfloor$ with equality with the right choice of $\ell$ (see Proposition \ref{ConstanteMajoréeParBorneSchur}). To embed a finite group $H$ of matrices over $\mathbf{Q}$ into an orthogonal group over a finite field, one just need to look at the positive definite bilinear form $\psi := \sum_{h \in H} {}^t \! H H$. For any prime $\ell$ large enough such that $\ell$ does not divide $\det \psi$, the group homomorphism of reduction mod $\ell$ induces an embedding of $H$ into an orthogonal group over $\mathbf{F}_l$, however this process does not generalize well when looking at polynomial automorphisms (See Remark \ref{remarkCasPairNeMarchPas}). \end{rmq} \subsection{The Minkowski's bound for finite groups of polynomial automorphisms with rational coefficients} To prove Theorem \ref{SchurBoundPolynomial}, we adapt the proof of the Minkowski bound for linear automorphisms. Actually, to conclude it suffices to show that Proposition \ref{PropEmbeddingFinite} also holds for finite $p$-subgroups of polynomial automorphisms. \begin{prop}\label{EmbeddingPolynomialCase} Let $d$ be an integer. Let $G$ be a finite $p$-subgroup of $\Aut(\mathbf{A}_\mathbf{Q}^d)$. Then, there exists a prime $\ell$ such that \begin{enumerate} \item $\ell$ is a generator of $(\mathbf{Z} / p^2 \mathbf{Z})^{\times}$ \item There is an injective homomorphism $G \hookrightarrow \mathrm{GL}_d (\mathbf{F}_\ell)$ \end{enumerate} \end{prop} \begin{lemme}\label{PointFixeEtDifferentielleInj} Let $d$ be an integer and $p$ a prime. Let $F$ be a finite field with $\Char (F) \neq p$. Let $G$ be a finite subgroup of $\Aut (\mathbf{A}_F^d)$ of order $p^\alpha$. Then $G$ has a fixed point $x_0 \in \mathbf{A}^d (F) = F^d$ and the homomorphism \[ \begin{array}{lrcl} \Phi:& G & \longrightarrow & \mathrm{GL}_d (F)\\ & g & \longmapsto & D_{x_0} g \end{array} \] is injective. \end{lemme} \begin{proof} The group $G$ acts on $F^d$ which is of size $\abs F^d$. Since $\abs G = p^\alpha$ and $p$ does not divide $\abs F$, the class equations gives the existence of at least one trivial $G$-orbit in $F^d$; hence, the existence of a fixed point $x_0 \in F^d$. Up to a translation we can suppose that $x_0 = 0$. Now to show the injectivity of $\Phi$. Take $g$ in $G$ such that $D_{0} g = \id$, then \[ g(x_1,\cdots, x_d) = g(\mathbf x) = \id + \sum_{j \geq 2} A_j (\mathbf x) \] where $A_j$ is the homogeneous part of $g$ of degree $j$. Suppose that $g \neq \id$, let $j_0$ be the lowest index $j \geq 2$ such that $A_j \neq 0$. We rewrite $g$ as $g = \id + A_{j_0} + B$ where $B = \sum_{j > j_0} A_j$ and compute the second iterate \begin{align*} g^2 (\mathbf x) &= g(\mathbf x) + A_{j_0} (g(\mathbf x)) + B (g(\mathbf x)) \\ &= \id + A_{j_0} (\mathbf x) + B(\mathbf x) + A_{j_0} (\mathbf x + A_{j_0} (\mathbf x) + B(\mathbf x) ) + B(g (\mathbf x)) \\ &= \id + 2 A_{j_0} (\mathbf x) + (\text{terms of higher degree}). \end{align*} And for every $k \geq 1$ we obtain \[ g^k (\mathbf x) = \id + k A_{j_0} (\mathbf x) + (\text{terms of higher degree}). \] Since, $g$ is of order $p^t$ for a certain $t>0$, replacing $k$ by $p^t$ in this formula we get $ p^t A_{j_0} (\mathbf x) =0 $, a contradiction since $\Char F \neq p$. \end{proof} \begin{rmq}\label{remarkChar0ResteInj} If $F$ is of characteristic $0$ and $x_0$ is fixed by $G$, then the proof shows also that $\Phi: g \mapsto D_{x_0} g$ is injective. \end{rmq} \begin{proof}[Proof of Theorem \ref{SchurBoundPolynomial} when $\k = \mathbf{Q}$] As in the linear case, we can find an integer $N$ such that $G \subset \Aut (\mathbf{A}_{\mathbf{Z}[1 / N]}^d)$. So, for $\ell > N$ prime , reduction modulo $\ell$ is well defined on $G$. Now, for $\ell$ large enough such that $\ell$ does not divide any coefficient of $g - \id$ for all $g \in G \subset \Aut (\mathbf{A}_{\mathbf{Z} [ 1/N]}^d)$, this homomorphism is injective and we can use Dirichlet's theorem to ensure that $\ell$ is a generator of $(\mathbf{Z} / p^2 \mathbf{Z})^{\times}$. $G$ is now embedded in $\Aut (\mathbf{A}_{\mathbf{F}_\ell}^d)$ and we replace it by its image in $\Aut(\mathbf{A}^d_{\mathbf{F}_\ell})$. By Lemma \ref{PointFixeEtDifferentielleInj}, there is a point $x_0 \in \mathbf{F}_\ell^d$ fixed by $G$ and we have an injective homomorphism $\Phi: G \hookrightarrow \mathrm{GL}_d (\mathbf{F}_\ell)$. This concludes the proof when $p \neq 2$. \end{proof} \subsection{Extension of Minkoswski's bound to number fields} \label{SubsecSchurBoundNumberFields} \paragraph{Strategy.--} This part is dedicated to the proof of Schur's bound for finite $p$-groups of polynomial automorphisms over arbitrary number fields. We will then prove Theorem \ref{SchurBoundPolynomial} using a Sylow argument. As in the previous section, we want to show the \begin{thm}\label{theoremPlongementCorpsFiniBonCardinal} Let $\k$ be a number field, $d$ an integer and $p$ be an odd prime. Let $G$ be a finite $p$-subgroup of $\Aut_\k (\mathbf{A}^d)$, then there exists a finite field $\mathbf{F}$ with $\Char \mathbf{F} \neq p$ and an injective group homomorphism $G \hookrightarrow \mathrm{GL}_d (\mathbf{F})$ such that $v_p (\abs{\mathrm{GL}_d(\mathbf{F})}) \leq M_\k (d,p)$. \end{thm} Indeed, this would prove that $v_p ( \abs G) \leq v_p (\abs{\mathrm{GL}_d (\mathbf{F})}) \leq M_\k (d,p)$. The natural idea is to do an analog of the proof for $\k =\mathbf{Q}$. Replace $\mathbf{Z}$ by the ring of integers $L := \mathcal O_\k$ of $\k$, then for any maximal ideal $\m$ of $L$ lying over a sufficiently large prime, there is an injective homomorphism $G \hookrightarrow \Aut (\mathbf{A}_{L / \m}^d)$. By taking differentials at a fixed point over $L / \m$ we would see $G$ as a subgroup of $\mathrm{GL}_d( L / \m)$ and the order of $\mathrm{GL}_d( L /\m)$ would give a bound $v_p (\abs{G}) \leq \sum_{i=1}^d v_p (\abs{L / \m}^i -1)$. The remaining part is to choose $\m$ wisely so that we get the lowest bound possible. To do this, we use cyclotomic characters. \paragraph{Cyclotomic characters.--}In this part, $\k$ is a finitely generated field over $\mathbf{Q}$. We denote by $\mu_{n}$ the group of $n$-th roots of unity in $\overline \k$. Recall that $\Aut (\mu_{n}) = (\mathbf{Z} / n \mathbf{Z})^{\times}$ because every automorphism $\phi$ is of the form $\phi(\omega) = \omega^a$ where $a \in (\mathbf{Z} / n \mathbf{Z})^\times$. \begin{dfn}[Cyclotomic character] Denote by $\Gamma_\k = \Gal(\overline \k / \k)$ the absolute Galois group of $\k$. For every $n \geq 1, \Gamma_\k$ preserves the group $\mu_{n} \subset \overline \k^{\times}$ of $n$-th roots of unity, this induces a group homomorphism \[ \chi_n : \Gamma_\k \rightarrow \Aut (\mu_n) = (\mathbf{Z} / n \mathbf{Z})^\times \] called the \emph{$n$-th cyclotomic character of $\k$}. In particular, if $p$ is a prime number, since the inclusion $\mu_{p^n} \subset \mu_{p^{n+1}}$ induces a group homomorphism $\Aut (\mu_{p^{n+1}}) = (\mathbf{Z} / p^{n+1} \mathbf{Z})^\times \rightarrow \Aut (\mu_{p^n}) = (\mathbf{Z} / p^n \mathbf{Z})^\times$, we have a compatible family of homomorphisms \[ \chi_{p^n} : \Gamma_\k \rightarrow \Aut (\mu_{p^n} ). \] This family of homomorphisms induces the $p^\infty$-\emph{cyclotomic character} \[ \chi_{p^\infty} : \Gamma_\k \rightarrow \mathbf{Z}_p^{\times} = \lim_{\longleftarrow} (\mathbf{Z} / p^n \mathbf{Z})^{\times} \] where $\mathbf{Z}_p$ is the ring of $p$-adic integers. This homomorphism is continuous with respect to the profinite topologies on $\Gamma_\k$ and $\mathbf{Z}_p^\times$. \end{dfn} We are interested in the image of $\chi_{p^\infty}$ which is a closed subgroup of $\mathbf{Z}_p^{\times}$. Define $t(\k; p)$ and $m(\k; p)$ as in Section \ref{par:Minkowski_Schur}. The number $m(\k; p)$ is always finite if $\k$ is finitely generated over $\mathbf{Q}$ (see \cite{SerreBoundsOrder}, $\S4.3$). If $s$ is an integer, we denote by $C_s$ the cyclic group of order $s$. \begin{prop}[\cite{SerreBoundsOrder}, $\S 4$]\label{ImageCaractereCyclotomique} $\phantom{-}$ \begin{enumerate} \item If $p$ is an odd prime, one has \[ \mathbf{Z}_p^{\times} \simeq C_{p-1} \times (1 + p \cdot \mathbf{Z}_p). \] The group $1 + p \cdot \mathbf{Z}_p$ is a procyclic subgroup generated by $1+p$ as a topological group and isomorphic to the additive group $\mathbf{Z}_p$. Its closed subgroups are the groups $1 + p^j \mathbf{Z}_p$ with $j\geq 1$. Furthermore, one has \[ \im \chi_{p^\infty} = C_{t(\k;p)} \times \left\{ 1 + p^{m(\k;p)} \cdot \mathbf{Z}_p \right\}. \] \item If $p=2$, then $\mathbf{Z}_2^{\times} = C_2 \times \left\{ 1 + 4 \cdot \mathbf{Z}_2 \right\}$. There are 3 possibilities for $\im \chi_{2^\infty}$: \begin{enumerate} \item $\im \chi_{2^\infty} = 1 + 2^{m(\k;p)} \cdot \mathbf{Z}_2$ and then $t(\k;p)=1$. \item $\im \chi_{2^\infty} = \langle -1 + 2^{m(\k;p)-1} \rangle$ (the closure of the group generated by $-1 + 2^{m(\k;p)-1} $) and then $t(\k;p)=2$. \item $\im \chi_{2^\infty} = C_2 \times \left\{ 1 + 2^{m(\k;p)} \mathbf{Z}_2 \right\}$ and then $t(\k;p)=2$. \end{enumerate} \end{enumerate} \end{prop} \begin{rmq} Those 3 cases are distinct when $m(\k,p) \neq \infty$. We will refer as $\k$ being in case (a), (b), or (c) when $\im \chi_{2^\infty}$ is of the form (a),(b) or (c) of Proposition \ref{ImageCaractereCyclotomique}. \end{rmq} Recall that an integral domain $L$ is \emph{normal} if every localisation at a prime ideal of $L$ is integrally closed. Let $L$ be a normal domain that is finitely generated over $\mathbf{Z}$ such that the fraction field of $L$ is $\k$. For any maximal ideal $\m \subset L$, the quotient $L / \m$ is finite by the Nullstellensatz for Jacobson rings and $N(\m) := \abs{L / \m}$ is the \emph{norm} of $\m$. Recall that for a ring $R$, $\Spec R$ denotes the set of prime ideals of $R$ and $\Specmax R$ the set of its maximal ideals both with the Zariski topology. The following theorem is proven in \cite[$\S 6$ Theorem 7]{SerreBoundsOrder}. \begin{thm}\label{OuvertDenseAvecAnneauNormal} Let $L$ be a normal domain finitely generated over $\mathbf{Z}$ such that the fraction field of $L$ is $\k$. Let $n$ be an integer and $c$ an element of $(\mathbf{Z} / n\mathbf{Z})^{\times}$. Denote by $X_c$ the set of elements $x \in \Specmax(L)$ such that $N(x) \equiv c \mod n$. Then: \begin{enumerate} \item If $c \not \in \im \chi_n$, $X_c = \emptyset$. \item If $c \in \im \chi_n$, then $X_c$ is Zariski-dense in $\Specmax (L)$. In particular, $X_c$ is infinite. \end{enumerate} \end{thm} In particular, the ring of integers of a number field is normal because it is integrally closed and this property is stable under localisation. So Theorem \ref{OuvertDenseAvecAnneauNormal} holds for $L$ the ring of integers of a number field. \paragraph{Valuations.--} We define the constant \[ M'_\k (d,p) = \inf_{u \in \im \chi_{p^\infty}} \sum_{i=1}^d v_p (u^i -1). \] The next proposition is adapted from Proposition 4, $\S 6$ of \cite{SerreBoundsOrder} to our context. \begin{prop}\label{ConstanteMajoréeParBorneSchur} One has \begin{enumerate}[label=(\alph*)] \item If $p \neq 2$ or if $p=2$ and $t(\k;p)=1$ ($\k$ is in case (a)), then \[ M'_\k (d,p) = \sum_{\substack{ i=1 \\ t(\k;p) | i}}^d (m(\k;p)+ v_p(i)) = M_\k (d,p). \] \item If $p=2$, $t(\k;p)=2$ and $\k$ is in case (b), one has \[ M'_\k (d,2) = r_1 + (m(\k;p)-1)r_0 + \sum_{i=1}^d v_2 (i) = M_\k (d,2) \] where $r_1$ is the number of odd integers between $1$ and $d$ and $r_0$ the number of even integers in this range. \item If $p=2$, $t(\k;p)=2$ and $\k$ is in case(c), one has \[ M'_\k(d,2) = r_1 + m(\k;p)r_0 + \sum_{i=1}^d v_2 (i) = \ent{\frac{d}{2}} + M_\k (d,2) \] with the same definition for $r_1$ and $r_0$. \end{enumerate} \end{prop} \begin{proof} Set $t= t(\k;p), m = m(\k;p)$. We start with the case $p \neq 2$. First if $t$ divides $i$, then $v_p (u^i -1) \geq m + v_p(i)$. This is because $u$ can be written as $zv$ with $z^t =1$ and $v_p (v-1) \geq m$, so $v_p (u^i -1) = v_p (v^i-1)$. So we have an inequality $M'_\k(d,p) \geq \sum_{\substack{ i=1 \\ t | i}}^d (m+ v_p(i))$. To have the opposite one, choose $u \in \im \chi_{p^\infty}$ such that $u = zx$ with $z$ of order $t$ and $v_p(x -1) = m$. This also works for $p=2$ and $t=1$. Suppose now that $p=2$ and $t=2$, Define $m'= m-1$ in case (b) and $m' = m$ in case (c). Then for every $x \in \im \chi_{2^\infty}$, \begin{align*} v_2(x^i -1) &\geq m' + v_2 (i) \text{ if } i \text{ is even.} \\ v_2(x^i -1)& \geq 1 \text{ if } i \text{ is odd.} \end{align*} This gives \[ M'_\k(d,2) \geq \sum_{i \text{ odd}} 1 + \sum_{i \text{ even}} (m' + v_2 (i) ) = r_1 + m' r_0 + \sum_{i \text{ even}} v_2 (i).\] To show the opposite inequality, we use the fact that $x =-1 + 2^{m'} \in \im \chi_{2^\infty}$ and we check that $\sum_{i=1}^d v_2(x^i-1) = r_1 + m'r_0 + \sum_{i=1}^d v_2 (i)$. Now, to show the different equalities, notice that for (a): \[ M'_\k (d,p) = m \cdot \ent{\frac{d}{t}} + \sum_{i=1}^{\ent{\frac{d}{t}}} v_p(ti). \] Now, since $t$ divides $p-1$, one has $v_p(ti) = v_p(i)$ and the rest of the computation is similar as in the case $\k = \mathbf{Q}$. For (b) and (c), we have $r_0 = \ent{\frac{d}{2}}$ and $r_1 = d - r_0$. \begin{align*} M'_\k(d,2) &\leq d - \ent{\frac{d}{2}} + m' \ent{\frac{d}{2}} + \sum_{i=1}^d v_2(i)\\ &= d + (m' -1) \ent{\frac{d}{2}} + \sum_{i=1}^d v_2 (i) \\ &= d + (m'-1) \ent{\frac{d}{t}} + \sum_{k \geq 1} \ent{\frac{d}{2^k}} \\ &= d + m' \ent{\frac{d}{t}} + \sum_{k\geq 1} \ent{\frac{d}{2^k t}}. \end{align*} \end{proof} We can now state Theorem \ref{theoremPlongementCorpsFiniBonCardinal} without assuming $p$ odd. \begin{thm}\label{theoremPlongementCorpsFiniBonCardinalPGeneral} Let $\k$ be a number field, $d$ an integer and $p$ be prime. Let $G$ be a finite $p$-subgroup of $\Aut_\k (\mathbf{A}^d)$, then there exists a finite field $\mathbf{F}$ with $\Char \mathbf{F} \neq p$ and an injective group homomorphism $G \hookrightarrow \mathrm{GL}_d (\mathbf{F})$ such that $v_p (\abs{\mathrm{GL}_d(\mathbf{F})}) \leq M_\k ' (d,p)$. \end{thm} \paragraph{Proof of Theorem \ref{theoremPlongementCorpsFiniBonCardinalPGeneral}.--} Take $G$ a finite $p$-subgroup of $\Aut (\mathbf{A}_\k^d)$ with $p$ prime. \smallskip {\sl{Step 1. Reduction modulo $\mathfrak l $.-- }} Set $L = \mathcal O_\k$. For every element $a \in \k^\times$ the fractional ideal generated by $a$ is of the form (see \cite{Neukirch}, $\S 3$) \[ a \cdot \mathcal O_\k = (a) = \prod_{\mathfrak l \in \Spec L} \mathfrak l^{v_\mathfrak l (a)} \] and the prime ideals $\mathfrak l$ such that $v_\mathfrak l (a) \neq 0$ are in finite number. For such an $\mathfrak l$ there exists a unique prime $\ell \in \mathbf{Z}_+$ such that $(\ell) \subset \mathfrak l$. We define for $g \in \Aut (\mathbf{A}_\k^d)$ \[ \ell_g := \max_{a \in \text{coeff}(g - \id)} \left\{ \text{prime } \ell \in \mathbf{Z}_+ : \exists \mathfrak l \in \Spec L, (\ell) \subset \mathfrak l, v_\mathfrak l (a) \neq 0 \right\} \] where $\text{coeff} (g- \id)$ is the set of coefficients of the polynomial transformation $g - \id$. Set $M_1 = \max_{g \in G} \ell_g$ ($M_1 < + \infty$ since $G$ is finite) and $M = \max(M_1,p)$, then for every prime $\ell >M$ and for every $\m \in \Specmax(L)$ such that $(\ell) \subset \m$, we have a well-defined injective homomorphism \[ \Psi: G \hookrightarrow \Aut( \mathbf{A}_{\mathbf{F}}^d), \] where $\mathbf{F} = L / \m$. Indeed, the homomorphism of rings $\phi: L \twoheadrightarrow L / \m$ induces the homomorphism $\phi: L_\m := \inv{(L \setminus \m)} L \rightarrow L/ \m$. By construction, $G$ is a subgroup of $\Aut (\mathbf{A}_{L_\m}^d)$, so $\phi: G \rightarrow \Aut (\mathbf{A}_{L / \m}^d)$ is well-defined and it is injective by our definition of $M$. \smallskip {\sl{Step 2. The group $\Psi(G)$.--}} Now, $\Psi(G)$ is a $p$-subgroup of $\Aut (\mathbf{A}_{\mathbf{F}}^d)$. Since $p \not \in \m$, we get $\Char (\mathbf{F}) \neq p$. By Proposition \ref{PointFixeEtDifferentielleInj}, there is a point $x_0$ in $\mathbf{A}^d(\mathbf{F})$ fixed by $\Psi(G)$ and by taking the differentials at $x_0$, we obtain an injective homomorphism $G \hookrightarrow \Psi(G) \hookrightarrow \mathrm{GL}_d (\mathbf{F})$. So, we get \begin{equation}\label{eq1} v_p (\abs G) \leq v_p \left( N(\m)^{\frac{d(d+1)}{2}} \prod_{i=1}^d (N(\m)^i-1) \right) = \sum_{i=1}^d v_p(N(\m)^i -1). \end{equation} Set $X := \left\{ \m \in \Specmax (L) : \m | (s), \text{for some }s > M \text{ prime} \right\}$, then (\ref{eq1}) holds for all $\m \in X$ and we obtain $v_p (\abs G) \leq \inf_{\m \in X} \sum_{i=1}^d v_p(N(\m)^i -1).$ So, to conclude, all we have to prove is \begin{equation} \inf_{\m \in X} \sum_{i=1}^d v_p(N(\m)^i -1) \leq M'_\k(d,p). \label{eqInegalite} \end{equation} \smallskip {\sl{Step 3. Proof of \eqref{eqInegalite}.--}} The set $X$ is open in $\Specmax L$. For, $X = \left( \bigcup_{l \leq M, l \text{ prime}} V(l) \right)^c$ with $V(l) = \left\{ \m \in \Specmax(L) : (l) \subset \m \right\}$ and $V(l)$ is closed. Take $u \in \im \chi_{p^\infty}$. For $j \geq 1$, let $u_j$ be the projection of $u$ in $(\mathbf{Z} / p^j \mathbf{Z})^{\times}$. By Theorem \ref{OuvertDenseAvecAnneauNormal} the set of maximal ideals $\m$ such that $N(\m) \equiv u_j \mod p^j$ is dense, therefore it intersects the open subset $X$, so for every $j\geq 1$, we can find $\m_j \in X$ such that $ N(\m_j) \equiv u_j \mod p^j$. Then, one has $\lim_{j \rightarrow \infty} N( \m_j) = u$ in $\mathbf{Z}_p^{\times}$, therefore $v_p (u^i -1) = \lim_{j \rightarrow \infty} v_p( N(\m_j)^i -1)$ so \[ \inf_{\m \in X} \sum_{i=1}^d v_p(N(\m)^i -1) \leq \sum_{i=1}^d v_p(u^i-1); \] and this holds for every $u \in \im \chi_{p^\infty}$. Using Proposition \ref{ConstanteMajoréeParBorneSchur}, we get \[ \inf_{\m \in X} \sum_{i=1}^d v_p(N(\m)^i -1) \leq \inf_{u \in \im_{\chi_{p^\infty}}} \sum_{i=1}^d v_p(u^i-1) = M'_\k( d,p). \] \paragraph{Proof of Theorem \ref{SchurBoundPolynomial} and comments.--} \begin{bigtheorem}\label{BigThmSchurBoundPolynomial} Let $\k$ be a number field, let $d$ be a natural number, and let $p$ be a prime. Let $G$ be a finite $p$-subgroup of $\Aut( \mathbf{A}^d_\k)$, then \begin{enumerate} \item If $p \geq 3$ or $p=2$ and $\k$ is in case (a) or (b), there exists a group embedding \[ G \hookrightarrow \mathrm{GL}_d (\k). \] \item If $p=2$ and $\k$ is in case (c), there exists a group embedding \[ G \hookrightarrow \mathrm{GL}_d (\k(z_4)). \] \end{enumerate} \end{bigtheorem} \begin{rmq} We do not state a Sylow-like property, saying that $G$ is conjugated to a subgroup of $\mathrm{GL}_d (\k)$, we only state that we can find an isomorphism of abstract groups from $G$ to a subgroup of $\mathrm{GL}_d (\k)$. \end{rmq} \begin{proof} For 1, we know that $v_p(\abs G) \leq M_\k (d,p)$ and that there exists a subgroup $H \subset \mathrm{GL}_d (\k)$ such that $\abs H = p^{M_\k (d,p)}$ by Theorem \ref{SchurBound}. Let $L = \mathcal O_\k$ be the ring of integers of $\k$. The proof of Theorem \ref{theoremPlongementCorpsFiniBonCardinalPGeneral} shows that there exists an infinite number of maximal ideals $\m$ of $L$ such that $v_p (\mathrm{GL}_d (\mathbf{F})) \leq M_\k (d,p)$ where $\mathbf{F} = L / \m$. So for any such maximal ideal $\m \subset L$ lying over a sufficiently large prime, there are embeddings $\Psi_H: H \hookrightarrow \mathrm{GL}_d (\mathbf{F})$ and $\Psi_G : G \hookrightarrow \mathrm{GL}_d (\mathbf{F})$. Looking at the size of $H$, we deduce that $v_p (\mathrm{GL}_d (\mathbf{F})) = M_\k (d,p)$ and $\Psi_H (H)$ is a $p$-Sylow of $\mathrm{GL}_d(\mathbf{F})$. By Sylow's theorems, $\Psi_G(G)$ is conjugated to a subgroup of $\Psi_H(H)$ in $\mathrm{GL}_d(\mathbf{F})$. This implies that $G$ is isomorphic to a subgroup of $H$. For 2, if $\k$ is in case (c) then one can check that $\k(z_4)$ is in case (a) and that $m(\k(z_4); 2) = m(\k; 2)$, therefore $M_{\k(z_4)}(d,2) = M_\k (d,2) + \ent{\frac{d}{2}}$ and the same proof as 1 shows the result. \end{proof} \begin{rmq}\label{RmqFinitelyGeneratedField} Theorem \ref{SchurBoundPolynomial} and \ref{BigThmSchurBoundPolynomial} still hold for $\k$ finitely generated over $\mathbf{Q}$. We just need to explain how the proof of Theorem \ref{theoremPlongementCorpsFiniBonCardinalPGeneral} works in that case. We need to find a normal domain $L$ finitely generated over $\mathbf{Z}$ such that $G$ is defined over $L$ and to define the open subset $X \subset \Specmax L$ used for equation \eqref{eqInegalite}. Here is how to proceed: since $G$ is finite, there exists a finitely generated $\mathbf{Z}$-algebra $R$ such that the elements of $G$ are defined over $R$, we can suppose that $R$ contains $1/p$. By Noether Normalization's Lemma and more precisely by generic freeness (see \cite{eisenbud1995commutative}, Theorem 14.4), there exists $t_1, \dots, t_s \in R$ and an integer $N$ such that $R$ is a finite free module over $\mathbf{Z}[1/N][t_1,\dots, t_s]$. We can then take for $L$ the integral closure of $\mathbf{Z}[1/N][t_1, \dots, t_s]$ in $\k$, $L$ is a normal domain over which $G$ is defined since $R \subset L$. We also have that $L$ is finitely generated over $\mathbf{Z}$ because by \cite[Theorem 4.14]{eisenbud1995commutative} it is a finite module over $\mathbf{Z}[1/N] [t_1, \dots, t_s]$. Now, let $A$ be the set of coefficients of $g - \id$ for $g \in G$. Set $X = \left\{ \m \in \Specmax L : A \cap \m = \emptyset \right\}$. This is an open subset of $\Specmax L$ as $A$ is finite and $X = \bigcap_{a \in A} V(a)^c$. For any $\m \in X$ we have an injective group homomorphism $G \hookrightarrow \Aut (\mathbf{A}^d_{L / \m})$ and Equation \eqref{eq1} holds. The proof of Equation \eqref{eqInegalite} is the same as in the case of number fields. This proves Theorem \ref{theoremPlongementCorpsFiniBonCardinalPGeneral} for finitely generated fields over $\mathbf{Q}$. \end{rmq} To prove Theorem \ref{BigThmSchurBoundPolynomial}, the key ingredient is that there exists subgroups of $\mathrm{GL}_d(\k)$ of size $p^{M_\k(d,p)}$, as Theorem \ref{SchurBoundPolynomial} is stated only for number fields we show for completeness how to construct finite $p$-groups of $\mathrm{GL}_d (\k)$ of size $p^{M_\k(d,p)}$ when $\k$ is finitely generated over $\mathbf{Q}$. The proof of Theorem \ref{BigThmSchurBoundPolynomial} for finitely generated fields over $\mathbf{Q}$ is then similar as in the case of number fields using Noether Normalization Lemma, we leave the details to the reader. \begin{prop} Let $\k$ be a finitely generated field over $\mathbf{Q}$ and let $p$ be a prime, there exists a finite $p$-subgroup of $\mathrm{GL}_d(\k)$ of size $p^{M_\k(d,p)}$. \label{PropBorneOptimale} \end{prop} \begin{proof} Set $t= t(\k; p), m = m(\k;p)$ and $r = \lfloor d/t \rfloor$. \paragraph{The case $p \geq 3$.--} Let $\rho = z_{p^m} \in \k(z_p)$. Then, the group $(\mathbf{Z} / p^m \mathbf{Z})$ acts on $\k(z_p)$ via multiplication by $\rho^k$ for all $k \in \mathbf{Z} / p^m \mathbf{Z}$. Now take $r$ copies of $\k(z_p)$; this is a $\k$-vector space $V$ of dimension $t \cdot r \leq d$ and let $S_r$ be the $r$-th symmetric group, $S_r$ acts on $V$ by permuting the $r$ copies of $\k(z_p)$ and therefore the group \[ G := S_r \ltimes (\mathbf{Z} / p^m \mathbf{Z})^r \] acts faithfully by linear automorphisms on $V$ and has the desired size. Indeed, $v_p (\vert G \vert) = m \cdot \left\lfloor \frac{d}{t} \right\rfloor + v_p (\lfloor \frac{d}{t} \rfloor !)$. \paragraph{The case $p=2$ and $t=1$.--} In that case, $\k = \k (z_4)$, then $M_\k(d,2) = m \cdot \left\lfloor \frac{d}{t} \right\rfloor + v_2 ( \left\lfloor \frac{d}{t} \right\rfloor !)$. Therefore, the proof above works as well, with $\rho = z_{2^m}$ acting on $\k(z_4) = \k$. \paragraph{The case $p=2$ and $t=2$.--} The construction above yields that $(\mathbf{Z} / 2^m \mathbf{Z})$ acts linearly on $\k(z_4)$. We twist this action by the Galois automorphism $\sigma$ that sends $z_4$ to $-z_4$; $\sigma$ is an involution that sends $\rho = z_{2^m}$ to another primitive $2^m$-th root of unity. So we get that the group $H:= \mathbf{Z} / 2 \mathbf{Z} \ltimes \mathbf{Z} / 2^m \mathbf{Z}$ acts faithfully on $\k(z_4)$. Now set $ r = \lfloor d/2 \rfloor$, we have that $G := S_r \ltimes H$ acts faithfully and linearly on a $\k$ vector space $V$ consisting of $r$ copies of $\k(z_4)$. The vector space $V$ has dimension $2 \cdot \lfloor d/2 \rfloor \leq d$. Now, we have \[ v_2 (\vert G \vert) = (m+1) \cdot \lfloor d/2 \rfloor + v_2 (\lfloor d/2 \rfloor !). \] If $d$ is even this is equal to $M_\k (d,2)$ and we are done. If $d$ is odd then $v_2 (\abs G) = M_\k (d,2) -1$ but then $V$ is of dimension $d-1$ so the group $G \times \{ \pm 1 \}$ acts faithfully on $V \oplus \k$ that is of dimension $d$ and this group has the desired size. \end{proof} We can therefore state: \begin{bigtheorem}\label{BigThMSchurBoundPolynomialGeneralCase} Let $\k$ be a finitely generated field over $\mathbf{Q}$, let $d$ be a natural number, and let $p$ be a prime. Let $G$ be a finite $p$-subgroup of $\Aut( \mathbf{A}^d_\k)$, then \begin{enumerate} \item If $p \geq 3$ or $p=2$ and $\k$ is in case (a) or (b), there exists a group embedding $ G \hookrightarrow \mathrm{GL}_d (\k)$ and $ v_p (\vert G \vert) \leq M_\k (d;p)$. \item If $p=2$ and $\k$ is in case (c), there exists a group embedding $ G \hookrightarrow \mathrm{GL}_d (\k(z_4))$ and $v_2(\vert G \vert) \leq M_\k(d, 2) + \lfloor \frac{d}{2} \rfloor $. \end{enumerate} \end{bigtheorem} \begin{rmq}\label{remarkCasPairNeMarchPas} We get the optimal bounds except when $p=2$ and $\k$ is in case (c) (this includes $\k = \mathbf{Q}$). For that case, following Remark \ref{remarkCasPairLinear}, to get the optimal bound one would need a result of the following type: \emph{ Let $\k$ be a number field in case (c) and $G$ a finite subgroup of $\Aut(\mathbf{A}_\k^d)$ of order $2^\alpha$, then for $\m$ in the complement of a finite set of $\Specmax \mathcal O_\k$ the group $G$ embeds into an orthogonal group over $\mathcal O_\k / \m$.} We know that for any maximal ideal $\m$ lying over a large enough prime, there exists an embedding $G \hookrightarrow \mathrm{GL}_d(\mathbf{F})$ and a fixed point $\bar x \in (\mathbf{F})^d$ of $G$ where $\mathbf{F} = \mathcal O_\k / \m$. The problem is to find a symmetric matrix $A$ such that \[ A_G := \sum_{g \in G} {}^t \! D_{\bar x} g \cdot A \cdot D_{\bar x} g \] is non-degenerate. Such an $A$ does not exist for every subgroup of $\mathrm{GL}_d (\mathbf{F})$ precisely because $v_2 (\abs{\mathrm{GL}_d(\mathbf{F}})$ is larger than the 2-adic valuation of the order of any orthogonal group over $\mathbf{F}$. So we have to use that $G$ comes from a group over $\k$ and adapt $\m$ wisely. Here is one way to attack this problem. Pick a fixed point $\overline x$ of $G$ with coordinates in $\overline \mathbf{Q}$; such a point exist because otherwise let $(P_n)$ be the system of polynomial equations stating that $G$ has a fixed point. If this system has no solution over $\overline \mathbf{Q}$ then by Hilbert's Nullstellensatz, there is a relation of the form $1 = \sum Q_i P_i$ for some polynomials $Q_i$. Now take a number field $\k'$ where this relation is defined. By the previous paragraph we can reduce modulo a large enough maximal ideal $\m$ of $\mathcal O_{\k'}$ (i.e lying over a large enough prime) and this would yield an injective group homomorphism $G \hookrightarrow \Aut (\mathbf{A}^d)$ where $F$ is a finite field with $\Char F \neq p$. The relation $1 = \sum Q_i P_i $ still holds in $\mathbf{F}$ but this is absurd since we know that $G$ admits a fixed point over $\mathbf{F}$. Let $\k '$ be the number field generated by the coordinates of $\overline x$ and $\k$. We would like to find $A$ such that $A_G$ is non-degenerate. If $\k ' \subset \mathbf{R}$ we can use argument of positive definiteness to do so, but otherwise a first difficulty occurs. Now, even if such an $A$ could be found, the arithmetic of $\k '$ leads to another difficulty: For any maximal ideal $\m ' \subset \mathcal O_{\k '}$ lying over a large enough maximal ideal $\m \subset \mathcal O_\k$, the image $x'$ of $\overline x$ in $\mathbf{F} ' = \mathcal O_{\k'} / \m'$ is a fixed point of $G$, and the reduction modulo $\m'$ of $A_G$ is an invertible symmetric matrix over $\mathbf{F}'$. But if the degree $[ \mathbf{F}', \mathbf{F}]$ is even, then the 2-adic valuation of any orthogonal group over $\mathbf{F} '$ will be too large to get the optimal bound. \end{rmq} \section{$p$-adic analysis}\label{SecPAdicAnalysis} To prove Theorem \ref{BoundNilpotentGroups}, we will show that any finitely generated nilpotent group acting on a complex quasiprojective variety of dimension~$d$~can be embedded in a finite dimensional~$p$-adic Lie group acting analytically on a~$p$-adic manifold of dimension~$d$. The theorem will follow from a version of Theorem 1.1 of \cite{epstein1979transformation} in a~$p$-adic context. In this section, we introduce all the tools from~$p$-adic analysis and~$p$-adic Lie groups needed for the proof. \subsection{Tate-Analytic Diffeomorphisms}\label{SecAnalyticDiffeo} \subsubsection{Definitions and topology} Let~$p$~be a prime. We denote by~$\mathbf{Z}_p$~the completed ring of~$\mathbf{Z}$~with respect to the~$p$-adic norm defined such that~$\abs p = 1/p$. Denote by~$\mathbf{Q}_p$~the completion of~$\mathbf{Q}$~with respect to this norm. Then~$\mathbf{Q}_p = \Frac(\mathbf{Z}_p)$~and~$\mathbf{Z}_p$~is the set of elements of~$\mathbf{Q}_p$~of absolute value~$\leq 1$. We extend this norm to~$\mathbf{Q}_p^d$~by taking the maximum of the absolute values of the coordinates. We will use explicitly the ring~$\mathbf{Z}_p$~and the field~$\mathbf{Q}_p$~but what follows can be done with any complete valued ring or field of characteristic~$0$. The right setup would be to consider~$\mathbf{C}_p$~the completion of the algebraic closure of~$\mathbf{Q}_p$~and~$\D_p$~the unit ball of $\mathbf{C}_p$. For reference, check \cite{cantat2014algebraic}. We denote by~$B(x,r) = \left\{ y \in \mathbf{Q}_p^d : \norm{x-y} \leq r \right\}$~the closed ball of radius~$r$~and center $x$. It is both open and closed. Such sets will be called \emph{clopen}. \paragraph{Tate analytic maps.--}Classically, a function~$\mathbf{Z}_p^d \rightarrow \mathbf{Q}_p$~is analytic if it can be written locally as a converging power series, we work with \emph{Tate-analytic} functions which are converging power series of radius~$\geq 1$~over~$\mathbf{Z}_p^d$. Take~$\mathbf{Z}_p^d$~with its standard coordinates~$\mathbf x = x_1, \cdots, x_d$. On~$\mathbf{Q}_p[x_1,\cdots,x_d] =: \mathbf{Q}_p [\mathbf x]$~the Gauss norm is defined by \[ \forall g \in \mathbf{Q}_p[\mathbf x], \quad g = \sum_{I \subset \mathbf{Z}_+^d} a_I \mathbf x^I, \quad \norm g := \max_{I} \abs{a_I} \] where $I = (I_1, \cdots, I_d)$ and $\mathbf x^I := x_1^{I_1}\cdots x_d^{I_d}$; we denote by~$\mathbf{Q}_p \langle x_1,\cdots, x_d \rangle =: \mathbf{Q}_p \langle \mathbf x \rangle$~the completion of $\mathbf{Q}_p [x_1,\cdots,x_d]$ with respect to the Gauss norm~$\mathbf{Q}_p \langle \mathbf x \rangle$~is the set of formal power series with coefficients in~$\mathbf{Q}_p$~such that~$a_I \rightarrow 0$~when~$I \rightarrow \infty$~(i.e when $\max(I) \rightarrow \infty$). It is also the set of formal power series with coefficients in~$\mathbf{Q}_p$~converging over~$\mathbf{Z}_p^d$. This shows that~$\mathbf{Q}_p \langle \mathbf x \rangle$ equipped with the Gauss norm is an infinite-dimensional Banach space over~$\mathbf{Q}_p$. For all polynomials~$f,g \in \mathbf{Q}_p [\mathbf x]$, then~$\norm{f \cdot g} \leq \norm f \cdot \norm g$~and this is also true in~$\mathbf{Q}_p \langle \mathbf x \rangle$, therefore~$\mathbf{Q}_p \langle \mathbf x \rangle$~is a Banach algebra over~$\mathbf{Q}_p$, it is the \emph{Tate algebra} over~$\mathbf{Q}_p$~in~$d$~variables (see \cite{robert2013course}). We also define~$\mathbf{Z}_p \langle \mathbf x \rangle$~which is the completion of~$\mathbf{Z}_p [\mathbf x]$~for the gauss norm; it is in fact the set of elements of~$\mathbf{Q}_p \langle \mathbf x \rangle$~of norm~$\leq 1$. \begin{rmq}\label{remarkCoeffEntierAMultiplicationPres} For each~$f \in \mathbf{Q}_p \langle \mathbf x \rangle$~there exists an element~$s \in \mathbf{Z}_p$~such that~$s \cdot f \in \mathbf{Z}_p \langle \mathbf x \rangle$~and if~$g \in \mathbf{Q}_p \langle \mathbf x \rangle$~is such that~$g(0) \in \mathbf{Z}_p$, then there exist an integer~$N>0$~such that~$g (p^N \mathbf x) \in \mathbf{Z}_p \langle \mathbf x \rangle$. Moreover, if~$g \in \mathbf{Q}_p [ [ \mathbf x ] ]$~is a formal power series with coefficients in~$\mathbf{Q}_p$~with a strictly positive radius of convergence, then there exists an integer~$N$~such that~$g (p^N \mathbf x)$~belongs to~$\mathbf{Q}_p \langle \mathbf x \rangle$. \end{rmq} \begin{rmq}\label{remarkPourquoiCoeffEntiers} There exist Tate-analytic maps with non-integer coefficients such that~$f(\mathbf{Z}_p^d) \subset \mathbf{Z}_p$. For example, take \[ f(x) = \frac{x^p - x}{p}. \] Since for all~$x \in \mathbf{Z}_p, x^p \equiv x \mod p$,~$f$~induces a map~$f: \mathbf{Z}_p \rightarrow \mathbf{Z}_p$. However every element~$f \in \mathbf{Q}_p \langle \mathbf x \rangle^d$~induces a map~$f: \D_p^d \rightarrow \mathbf{C}_p$~and we have~$f(\D_p^d) \subset \D_p \Leftrightarrow f \in \mathbf{Z}_p \langle \mathbf x \rangle^d$. This has to do with the residue field of~$\mathbf{Z}_p$~being finite but not the residue field of~$\D_p$~(see \cite{robert2013course}, Proposition of page 240). \end{rmq} For any~$m \geq 0$, elements of~$\mathbf{Q}_p \langle \mathbf x \rangle^m$~are called \emph{Tate-analytic functions}. If~$g \in \mathbf{Q}_p \langle \mathbf x \rangle^d$, then \begin{equation} \forall x, y \in \mathbf{Z}_p^d, \norm{g(x) - g(y)} \leq \norm g \norm{x-y}. \label{EqLipschitz} \end{equation} In particular,~$g$~is~$\norm g$-Lipschitz. \begin{prop}[Strassman's Theorem, see \cite{robert2013course}, chapter 6, section 2.1] \label{PropIsolatedZeroPrinciple} Let~$f \in \mathbf{Q}_p \langle t \rangle$~be a Tate-analytic function in one variable, if~$f$~is not the zero function, then~$f$~has a finite number of zeros over~$\mathbf{Z}_p$. \end{prop} \begin{cor}\label{AnalyticContinuation} Let~$f \in \mathbf{Q}_p \langle \mathbf x \rangle$, if there exists a non-empty open subset ~$\mathcal U \subset \mathbf{Z}_p^d$~such that~$f_{| \mathcal U} \equiv 0$~then~$f$~is the zero function. \end{cor} \begin{rmq} This is not true for analytic functions over~$\mathbf{Z}_p^d$. For example define~$g$~by~$g(y) = 1$~if~$\norm y \leq \abs p$~and~$g(y) = 0$~otherwise. Then,~$g$~is analytic at every point of~$\mathbf{Z}_p^d$~because it is locally constant, it vanishes on the open subset~$\left\{ x \in \mathbf{Z}_p^d : \norm x = 1 \right\}$~but~$g$~is not the zero function. \end{rmq} \begin{proof}[Proof of Corollary \ref{AnalyticContinuation}] Take~$y \in \mathcal U$~and~$x \in \mathbf{Z}_p^d$. Let~$\varphi$~be the function~$\varphi: t \in \mathbf{Z}_p \mapsto f(tx + (1-t)y)$. Then~$\varphi$~belongs to~$\mathbf{Q}_p \langle t \rangle$~and it vanishes for any sufficiently small~$t$. By Proposition \ref{PropIsolatedZeroPrinciple}, we have that~$\varphi$~is the zero function, therefore~$f(x) = 0$. \end{proof} Let~$f,g \in \mathbf{Q}_p \langle \mathbf x \rangle$~and~$c>0$, we write~$f \equiv g \mod p^c$~if $\norm{f -g} \leq \abs p^c$ and we extend such notation componentwise for~$\mathbf{Q}_p \langle \mathbf x \rangle^m$~for every~$m \geq 1$. \begin{ex} \label{ExampleCongruence} If~$c=1$~and~$f,g \in \mathbf{Z}_p \langle \mathbf x \rangle$, then~$f = \sum_I a_I \mathbf x^I \equiv \id(\mathbf x) \mod p$~means that~$\overline f := \sum_{I} \overline{a_I} \mathbf x^I = \id (\mathbf x)$~where~$\overline{a_I} = a_I \mod p$~is the reduction of~$a_i$~mod~$p \mathbf{Z}_p$. \end{ex} \paragraph{Tate analytic diffeomorphisms.--} The composition determines a natural map \[ \begin{array}{cclll} \mathbf{Z}_p \langle X_1,\cdots,X_n \rangle^m & \times & \mathbf{Z}_p \langle Y_1,\cdots, Y_s \rangle^n & \longrightarrow & \mathbf{Z}_p \langle Y_1,\cdots, Y_s \rangle^m \\ (g_1,\cdots.,g_m) & &(h_1,\cdots,h_n) & \longmapsto & (g_1(h_1,\cdots,h_n),\cdots, g_m(h_1,\cdots,h_n)) \end{array} \] If the three integers~$n,m,s$~are equal to the same integer~$d$,~$(\mathbf{Z}_p \langle \mathbf x \rangle^d, \circ)$~becomes a semigroup. The invertible elements of this semigroup are called \emph{Tate-analytic diffeomorphisms} and form a group denoted by $\Diff^{an} (\mathbf{Z}_p^d)$. Using Equation \eqref{EqLipschitz}, we have that~$\Diff^{an}(\mathbf{Z}_p^d)$~acts by isometries on $\mathbf{Z}_p^d$. \begin{rmq} Following Remark \ref{remarkPourquoiCoeffEntiers}, we see that~$\Diff^{an} (\mathbf{Z}_p^d)$~consists exactly of the elements of ~$f \in \mathbf{Q}_p \langle \mathbf x \rangle$~that induces a Tate-analytic diffeomorphisms~$f: \D_p^d \rightarrow \D_p^d$. \end{rmq} The next proposition shows an easy way to construct Tate-analytic diffeomorphisms of small polydisks. \begin{prop}[Local inversion theorem, see \cite{SerreLieGroupsLieAlgebras}]\label{ExistenceInverse} Let~$\Phi \in \mathbf{Z}_p [[X_1,\cdots.,X_d]]^d$~be a power series with a strictly positive radius of convergence. Suppose that~$\Phi(0) = 0$~and~$\det (D_0 \Phi) \neq 0$, then there exists a unique~$\Psi \in \mathbf{Q}_p [[X_1,\cdots.,X_d]]^d$, with a strictly positive radius of convergence, such that~$\Psi(0) = 0$~and \[ \Phi \circ \Psi (\mathbf x) = \Psi \circ \Phi(\mathbf x) = \mathbf x. \] Furthermore,~$\norm{\Psi_n} \leq \max (1, \norm{ \inv {D_0 \Phi}}^n)$, where~$\Psi_n \in \mathbf{Q}_p [X_1,\cdots,X_n]^d$~is the homogeneous part of degree~$n$~of~$\Psi$~and~$\abs{\abs{\cdot}}$~is the Gauss norm over polynomials. Therefore, if~$\Phi$~belongs to~$\mathbf{Z}_p \langle \mathbf x \rangle^d$, then for any~$k$~such that~$~\abs p^k < \norm{D_0 \inv \Phi}$, we have that~$\frac{1}{p^k} \Phi (p^k \mathbf x)$~and~$\frac{1}{p^k} \Psi (p^k \mathbf x)$~are Tate-analytic diffeomorphisms and are inverse of each other. \end{prop} \paragraph{Group topology.--} The following proposition shows that~$\Diff^{an} (\mathbf{Z}_p^d)$~is a topological group with respect to the topology induced by the Gauss norm. \begin{prop}\label{truc1} Let~$f,g,h \in \mathbf{Z}_p \langle \mathbf x \rangle^d$, then \begin{enumerate} \item~$\norm{g \circ f} \leq \norm g$. \item If~$f$~is an element of~$\Diff^{an} (\mathbf{Z}_p^d)$~then~$\norm{ g \circ f} = \norm g$. \item~$~\norm{ g \circ (\id +h) - g } \leq \norm h$. \item~$\norm{\inv f - \id} = \norm{f - \id}$~if~$f$~is a Tate-analytic diffeomorphism. \end{enumerate} \end{prop} \begin{lemme}\label{lemma:PuissanceCongruence} Let~$f$~be an element of~$\Diff^{an}(\mathbf{Z}_p^d)$, if~$f \equiv \id \mod p$~then~$f^{p^c} \equiv \id \mod p^c$. \end{lemme} \begin{cor}\label{corollary:SubgroupsBasisOfNeighbourhoods} Let~$c>0$~be a real number, then the subgroup~$\Diff^{an}_c(\mathbf{Z}_p^d)$~of~$\Diff^{an} (\mathbf{Z}_p^d)$~consisting of all elements~$f \in \Diff^{an} (\mathbf{Z}_p^d)$~such that~$f \equiv \id \mod p^c$~is a normal subgroup of ~$\Diff^{an}(\mathbf{Z}_p^d)$. \end{cor} Proposition \ref{truc1}, Lemma \ref{lemma:PuissanceCongruence} and Corollary \ref{corollary:SubgroupsBasisOfNeighbourhoods} are proven in \cite{cantat2014algebraic}, section 2.1. \subsubsection{Analytic flow and Bell-Poonen theorem} \paragraph{Flows and vector fields.--} As in real or complex geometry, we define vector fields and flows. Let~$d$~be an integer: A \emph{Tate-analytic vector field}~$\mathbf X$~over~$\mathbf{Z}_p^d$~is a vector field of the form \[ \mathbf X(\mathbf x) = \sum_{i=1}^d u_i (\mathbf x) \partial_i \] where each~$u_i$~belongs to~$\mathbf{Q}_p \langle \mathbf x \rangle$. The Lie bracket of two vector fields~$\mathbf X$~and~$\mathbf Y= \sum_{i=1}^d v_i \partial_i$~is the vector field defined by \[ [ \mathbf X, \mathbf Y] = \sum_{j=1}^d w_j (\mathbf x) \partial_j \text{ with } w_j = \sum_{i=1}^d \left(u_i \frac{\partial v_j}{\partial x_i} - v_i \frac{\partial u_j}{\partial x_i}\right). \] The~$\mathbf{Q}_p$-Lie algebra of Tate-analytic vector fields over~$\mathbf{Z}_p^d$~is denoted by~$\Theta(\mathbf{Z}_p^d)$~it is a strict subalgebra of the Lie Algebra of analytic vector fields over~$\mathbf{Z}_p^d$. The Gauss norm of a Tate-analytic vector field ~$\mathbf X = \sum u_i (\mathbf x) \partial_i$~is defined as~$\norm \mathbf X = \max_i \norm {u_i}$~and makes~$\Theta(\mathbf{Z}_p^d)$~a complete Lie Algebra over~$\mathbf{Q}_p$~isomorphic as a Banach space to~$\mathbf{Q}_p \langle \mathbf x \rangle^d$. A \emph{Tate-analytic flow}~$\Phi$~over~$\mathbf{Z}_p^d$~is an element of~$\mathbf{Z}_p \langle X_1, \cdots, X_d, t \rangle^d = \mathbf{Z}_p \langle \mathbf x,t \rangle^d$~which satisfies the following properties \begin{enumerate}[label=(\roman*)] \item~$~\forall \mathbf x \in \mathbf{Z}_p^d, \ \forall s,t \in \mathbf{Z}_p, \quad \Phi(\mathbf x, s+t) = \Phi( \Phi(\mathbf x,s), t).$ \item~$\forall \mathbf x \in \mathbf{Z}_p^d,\quad \Phi(\mathbf x,0) = \id(\mathbf x)$. \end{enumerate} Set~$\Phi_t := \Phi( \cdot, t) \in \mathbf{Z}_p \langle \mathbf x \rangle$. Then,~$\Phi_0 = \id$~and~$\Phi_t \in \Diff^{an}(\mathbf{Z}_p^d)$ since~$\inv{\Phi_t} = \Phi_{-t}$. Then,~$t \in \mathbf{Z}_p \mapsto \Phi_t \in \Diff^{an}(\mathbf{Z}_p^d)$~is a continuous homomorphism of topological groups with respect to the Gauss norm. The main point here is that flows are parametrized by the compact group~$(\mathbf{Z}_p, +)$. \begin{ex} If~$\Phi$~is a Tate-analytic flow, then we can define its associated Tate-analytic vector field~$\mathbf X_\Phi := \frac{\partial \Phi_t}{\partial t}_{|t=0}$. In particular,~$\mathbf X_\Phi$~is~$\Phi_t$-invariant, for all~$t \in \mathbf{Z}_p$. \end{ex} \paragraph{From vector fields to Tate-analytic flows.--} Since a Tate-analytic vector field~$\mathbf X$~is analytic, it is a general fact that it admits local analytic flows over $\mathbf{Z}_p^d$~(see \cite{bourbaki2007varietes} for example), the next proposition shows that if the norm of~$\mathbf X$~is sufficiently small, then it admits a global Tate-analytic flow. \begin{prop}\label{PropExistenceGlobalTateAnalyticFlow} If~$\mathbf X$~is a Tate-analytic flow over~$\mathbf{Z}_p^d$, then for any sufficiently small~$\lambda \in \mathbf{Z}_p$, there exists a unique Tate-analytic flow~$\Phi^\lambda \in \mathbf{Z}_p \langle \mathbf x, t \rangle^d$~such that \[ \frac{\partial \Phi_t^\lambda (\mathbf x)}{\partial t} = \lambda \mathbf X(\Phi_t^\lambda(\mathbf x)). \] In particular, let $c>0$ be such that $c > \frac{1}{p-1}$, then every Tate-analytic vector fields~$\mathbf X$~such that~$\norm \mathbf X \leq \abs p^c$~admits a global Tate-analytic flow. \end{prop} \begin{proof} The strategy is to solve this differential equation in the space of power series~$\mathbf{Q}_p \left[ \left[ \mathbf x, t \right] \right]^d$~and then to show some properties on the radius of convergence of the solution. We first replace~$\mathbf X$~by ~$\mu \mathbf X$~for some~$\mu \in \mathbf{Z}_p$~such that~$\norm \mathbf X \leq 1$. Write~$\mathbf X (\mathbf x) = \sum_i u_i (\mathbf x) \partial_i$~with~$u_i \in \mathbf{Z}_p \langle \mathbf x \rangle$. We look at the differential equations \begin{equation} \frac{\partial}{\partial t } f_i (\mathbf x, t) = u_i(f(\mathbf x,t)) \label{EqDiff} \end{equation} with~$f_i \in \mathbf{Q}_p \left[ \left[ \mathbf x, t \right] \right]$~and~$f = (f_1, \cdots, f_d)$~such that~$f(\mathbf x, 0) = \mathbf x$. Write \[ f_i (\mathbf x ,t) = \sum_{k \geq 0 } a_k^{(i)} (\mathbf x) t^k , \quad a_k^{(i)} \in \mathbf{Q}_p \left[ \left[ \mathbf x \right] \right] \] then, the unique solution of this equation is formally given by the formulas~$a_k^{(i)} (\mathbf x) = \frac{1}{k !} \frac{\partial^k f_i}{\partial t^k} (\mathbf x, 0)$. We show that for all integer~$k \geq 0, \frac{\partial^k f_i}{\partial t^k} (\mathbf x, 0)$~belongs to~$\mathbf{Z}_p \langle \mathbf x \rangle$~by induction on~$k$. We get~$a_0^{(i)} = x_i$~since~$f(\mathbf x, 0) = \id(\mathbf x)$~and~$a_1^{(i)} (\mathbf x) = u_i(\mathbf x)$~by Equation \eqref{EqDiff}. Take~$k \geq 2$~and suppose the result to be true for all~$l < k$. By differentiating both sides of Equation \eqref{EqDiff}~$k-1$~times with respect to~$t$~and taking~$t=0$, we see that~$\frac{\partial^k f_i}{\partial t^k} (\mathbf x, 0)$~is obtained by sum and compositions of differentials of orders~$\leq k - 1$~of the Tate-analytic function~$u_i \in \mathbf{Z}_p \langle \mathbf x \rangle$~and the Tate-analytic functions~$\frac{\partial^l}{\partial t^l} f_i (\mathbf x, 0) \in \mathbf{Z}_p \langle \mathbf x \rangle$~with~$l < k$. So~$\frac{\partial^k f_i}{\partial t^k} (\mathbf x, 0)$~belongs to~$\mathbf{Z}_p \langle \mathbf x \rangle$~by induction. The solution~$f$~is then of the form \[ f(\mathbf x ,t) = \id(\mathbf x) + \sum_{k \geq 1} \frac{\partial^k f}{\partial t^k} (\mathbf x, 0) \frac{t^k}{k !}. \] Now take $\lambda \in \mathbf{Z}_p$, such that $\abs \lambda \leq \abs p ^c$. We have that for all $k \geq 0, \frac{\lambda^k}{k!} \in \mathbf{Z}_p$ and $\lambda^k / k! \rightarrow 0$ in $\mathbf{Z}_p$ when $k \rightarrow \infty$. Then, $\Phi^\lambda_t := f( \cdot, \lambda t)$~is a Tate-analytic flow such that~$\frac{\partial \Phi_\lambda^t}{\partial t} (\mathbf x) = \lambda \mathbf X(\Phi_t^\lambda (\mathbf x))$. For the final statement, take~$\mathbf X$~a Tate-analytic vector field such that~$\norm \mathbf X \leq \abs p^c$ and let $s \in \mathbf{Z}_p$ be such that $\abs s = \norm \mathbf X$, then $\mathbf Y := \frac{1}{s} \mathbf X$ has norm $\leq 1$. The proof shows that there exists a unique Tate-analytic flow~$\Phi$~such that~$\frac{\partial \Phi_t}{\partial t}_{|t=0} = s \mathbf Y = \mathbf X$. \end{proof} \begin{thm}[local linearisation of vector fields]\label{pAdicFrob} Let~$\mathbf X_1,\cdots,\mathbf X_k$~be Tate-analytic vector fields over ~$\mathbf{Z}_p^d$~such that~$[\mathbf X_i, \mathbf X_j] = 0$~for all~$1 \leq i,j \leq k$. Suppose that there exists a point~$m \in \mathbf{Z}_p^d$~such that the vectors~$\mathbf X_i (m)$~are linearly independent. Then, there exists a clopen subset~$\mathcal V \subset \mathbf{Z}_p^d$~containing~$m$~and an analytic diffeomorphism~$\varphi$~from~$\mathbf{Z}_p^d$~onto~$\mathcal V$~such that~$\varphi^* (X_{i|\mathcal V}) = \partial_i$~and such that ~$\varphi^*$~yields an injective Lie Algebra homomorphism~$\Theta(\mathbf{Z}_p^d)_{|\mathcal V} \hookrightarrow \Theta(\mathcal V)$. \end{thm} \begin{rmq} This theorem is well known in~$p$-adic differential geometry with analytic regularity (see \cite{bourbaki2007varietes}), what is important here is that when changing coordinates we keep the Tate-analytic regularity for vector fields. \end{rmq} \begin{proof} By translation, we can suppose that~$m=0$. We pick $Y_0 \subset T_0 \mathbf{Z}_p^d$ such that we have the decomposition $T_0 \mathbf{Z}_p^d = \Vect ( X_1(0),\cdots, X_k (0)) \oplus Y_0$. Let~$e_{1}, \cdots, e_{d-k}$~be a basis of~$Y_0$. Pick local (analytic) coordinates~$(x_1,\cdots,x_k, y_1,\cdots, y_{d-k})$~such that for all~$1 \leq j \leq d-k, \frac{\partial }{\partial y_j} (0) = e_j$. Define :~$f: \mathbf{Z}_p^{d-k} \rightarrow \mathbf{Z}_p^d$~by \[ f(y_1,\cdots,y_{d-k}) =(0,\cdots,0, y_1,\cdots,y_{d-k}). \] Take the local analytic flows~$\varphi^1,\cdots, \varphi^k$~associated to~$\mathbf X_1,\cdots, \mathbf X_k$~at~$0$~(here we do not suppose these flows to be Tate-analytic)~and consider \[ \begin{array}{crcl} g&: \mathbf{Z}_p^k \times \mathbf{Z}_p^{d-k} & \longrightarrow & {\mathbf{Z}_p^d}\\ & {(t_1,\cdots,t_k; y)} & \longmapsto& { \varphi^1_{t_1} \circ \cdots \circ \varphi^k_{t_k} (f(y)).} \end{array} \] The function~$g$~belongs to~$\mathbf{Z}_p \left[ \left[ t_1, \cdots, t_k, \mathbf y \right] \right]^d$~with a radius of convergence~$r_g >0$, satisfies~$g(0) = 0$~and its differential at the point~$(0,0)$~is \[ (x_1,\cdots,x_k; z) \mapsto x_1 \mathbf X_1(0) + \cdots+ x_k \mathbf X_k(0) + \sum_j z_j \frac{\partial}{\partial y_j} (0). \] Therefore it is invertible. By Proposition \ref{ExistenceInverse}~$g$~admits a formal inverse~$h \in \mathbf{Q}_p \left[ \left[ t_1, \cdots, t_k, \mathbf y \right] \right]^d$ with a radius of convergence~$r_h >0$. Denote by~$\mathbf z$~the set of coordinates~$(t_1, \cdots, t_k, y_1, \cdots, y_{d-k})$. Pick integers~$K, L$~such that~$\abs p ^K < r_g$~and~$\abs p^L < r_h$~such that~$g(B(0, \abs p^K)) \subset B(0, \abs p^L)$. Let $\mathcal V$ denote $g(B(0, \abs p^K))$; it is a clopen subset of~$\mathbf{Z}_p^d$~because~$B(0, \abs p^K)$~is clopen. Set~$\varphi := \frac{1}{p^L} g(p^K \mathbf z)$~and~$\psi := \frac{1}{p^K} h(p^L \mathbf z)$, they both belong to ~$\mathbf{Q}_p \langle \mathbf z \rangle^d$~and are inverse of each other and we have~$\varphi^* \mathbf X_i = \partial_i$. Finally, since ~$\varphi \in \mathbf{Q}_p \langle \mathbf z \rangle^d$, the map~$\varphi^*$~preserves Tate-analytic vector fields. \end{proof} \begin{thm}[$p$-adic version of \cite{epstein1979transformation} Theorem 1.1]\label{theoremPAdicEpsteinThurston} Let~$\mathfrak h$~be a nilpotent Lie algebra of Tate-analytic vector fields of~$\mathbf{Z}_p^d$, then~$d \geq \dl (\mathfrak h)$. \end{thm} \begin{proof} We follow the proof of \cite{cantat2014mapping} Proposition 3.10 and proceed by induction on the dimension~$d$. If ~$d=0$, there is nothing to prove. Suppose~$d \geq 1$~and that the result is true in dimension~$d-1$. Since~$\mathfrak h$~is nilpotent, its center is not trivial. Let~$\mathbf X$~be a nonzero central element of~$\mathfrak h$. Let~$m$~be a point where~$\mathbf X(m) \neq 0$, then by Theorem \ref{pAdicFrob}, there exists a small clopen subset~$\mathcal V \subset \mathbf{Z}_p^d$~and an analytic diffeomorphism~$\varphi: \mathcal V \rightarrow \mathbf{Z}_p^d$~that yields coordinates~$x_1, \cdots, x_d$~over~$\mathcal V$~such that~$\varphi_* \mathbf X = \partial_d$~and such that~$\varphi_*$~maps Tate-analytic vector fields to Tate-analytic vector fields. By Proposition \ref{AnalyticContinuation} the morphism of restriction~$\mathfrak h \rightarrow \mathfrak h_{|\mathcal V}$~is an isomorphism of Lie algebras. We replace~$\mathfrak h$~by~$\mathfrak h_{|\mathcal V}$~and work with the coordinates~$x_1, \cdots, x_d$~over ~$\mathcal V$. Every vector field~$\mathbf Y$~of~$\mathfrak h$~must commute with~$\mathbf X = \partial_d$~so it is of the form \[ \mathbf Y = \sum_{i=1}^d u_i(x_1,\cdots, x_{d-1}) \partial_i. \] Let~$\pi: \mathcal V \simeq \mathbf{Z}_p^d \rightarrow \mathbf{Z}_p^{d-1}$~be the projection over the first~$d-1$~coordinates. This yields a Lie algebra homomorphism~$\pi_* : \mathfrak h \rightarrow \Theta(\mathbf{Z}_p^{d-1})$. Denote by~$\mathfrak h_1$~the image of~$\mathfrak h$~under~$\pi_*$~and~$\mathfrak h_0$ its kernel. We have the exact sequence \[ 0 \rightarrow \mathfrak h_0 \rightarrow \mathfrak h \rightarrow \mathfrak h_1 \rightarrow 0. \] Now,~$\mathfrak h_0$~consists of Tate-analytic vector fields of~$\mathfrak h$~of the form~$u(x_1, \ldots, x_{d-1}) \partial_d$~so it is abelian and~$\mathfrak h_1$~is nilpotent because~$\mathfrak h$~is. So we get~$\dl(\mathfrak h) \leq \dl(\mathfrak h_1) + 1$~by the exact sequence and~$\dl(\mathfrak h_1) \leq d-1$~by induction. \end{proof} We discuss the optimality of Theorem \ref{theoremPAdicEpsteinThurston} in Section \ref{SubSecOptimality}. \paragraph{The theorem of Bell and Poonen.--} The following theorem first proven by Bell in \cite{Bell05} then by Poonen in \cite{poonen2014p} gives us an easy way to construct flows from analytic transformations. This is a very strong theorem as it shows that, contrary to~$\mathbf{R}$, over~$\mathbf{Q}_p$~a lot of analytic diffeomorphisms are in a flow. See \cite{Cantat_smf_18} for a more precise discussion on Bell-Poonen theorem. \begin{thm}[Bell-Poonen]\label{theoremBellPoonen} Let~$d \geq 1$~be an integer, and~$f \in \mathbf{Z}_p \langle \mathbf x \rangle^d$. Take~$c > \frac{1}{p-1}$ and suppose that~$f \equiv \id \mod p^c$, then \begin{enumerate} \item~$f$~is a Tate-analytic diffeomorphism. \item There exists a unique Tate-analytic flow~$\Phi \in \mathbf{Z}_p \langle \mathbf x,t \rangle^d$~such that \[ \forall n \in \mathbf{Z}, \quad \Phi(\mathbf x,n) = f^n (\mathbf x). \] In particular,~$\Phi_1 = f$. \end{enumerate} \end{thm} In fact, Poonen showed this theorem for the valuation ring of any ultrametric field~$\mathbf{K}$. So, Bell-Poonen Theorem also holds over~$\D_p$~or over any finite extension of~$\mathbf{Q}_p$~for example. \begin{cor}\label{NoTorsion} Let~$H$~be a subgroup of~$\Diff^{an}_1 (\mathbf{Z}_p^d)$~with~$p \geq 3$, then~$H$~is torsion-free. \end{cor} \begin{proof} Let~$h \in H$, suppose that~$h$~has order~$N < \infty$. By Theorem \ref{theoremBellPoonen}, there exists an Tate-analytic flow~$\Phi$~such that~$\Phi_1 = h$. Then for all ~$\mathbf x \in \mathbf{Z}_p^d$~the function~$t \in \mathbf{Z}_p \mapsto \Phi_t (\mathbf x) - \mathbf x \in \mathbf{Z}_p^d$~is analytic and has an infinite number of zeros, so it is zero everywhere by Proposition \ref{AnalyticContinuation}. Therefore~$\Phi_1 (\mathbf x) = h(\mathbf x) = \mathbf x$~and~$h = \id$. \end{proof} The next proposition won't be used in the proof of Theorem \ref{BoundNilpotentGroups} but it gives useful information on the dynamics of Tate-analytic flows. \begin{prop}\label{StableAIndiceFiniPres} Let~$\Phi \in \mathbf{Z}_p \langle \mathbf x,t \rangle$~be a Tate-analytic flow over~$\mathbf{Z}_p^d$. If~$\mathcal U \subset \mathbf{Z}_p^d$~is a clopen set, then there exists an~$\epsilon >0$~such that \[ \forall t \in \mathbf{Z}_p, \quad \abs t \leq \epsilon \Rightarrow \Phi_t (\mathcal U) = \mathcal U. \] \end{prop} \begin{proof} Fix~$x \in \mathbf{Z}_p^d$~and~$0 < r \leq 1$. Since~$\Phi_t \rightarrow \id$~as~$t \rightarrow 0$~in~$\Diff^{an} (\mathbf{Z}_p)$, there exists~$\epsilon >0$~such that for all~$t \in \mathbf{Z}_p$,~$\abs t \leq \epsilon \Rightarrow \norm{\Phi_t - \id} \leq r$. Now for all~$z \in \mathbf{Z}_p^d, \norm{ \Phi_t (z) - z} \leq \norm{\Phi_t - \id} \leq r$. Then, for all~$y$~such that~$\norm{y -x} \leq r$, \begin{align*} \norm{\Phi_t (y) - x} &= \norm{\Phi_t (y) - y + y -x } \\ &\leq \max( \norm{\Phi_t(y) - y}, \norm{y-x} ) \leq r. \end{align*} So if~$\abs t \leq \epsilon$, we have~$\Phi_t (B(x,r)) \subset B(x,r)$~and~$\Phi_{-t}(B(x,r)) \subset B(x,r)$, so we get the equality. Since~$\mathcal U$~is clopen, by compactness,~$\mathcal U = \bigcup_{i=1}^T B(x_i, r_i)$~for some finite set~$\left\{ x_1, \cdots, x_T \right\} \subset \mathcal U$~and radii~$r_i \in (0, 1]$. Thus, the results follows from the case of one ball. \end{proof} \subsection{Infinite-dimensional analytic manifold over~$\mathbf{Q}_p$}\label{SecInfiniteDimensionalAnalyticmanifolds} The main goal of the next two sections is to show that the topological group~$\Diff^{an}(\mathbf{Z}_p^d)$~is in fact an infinite dimensional Lie group over~$\mathbf{Q}_p$. We refer to \cite{bourbaki2007varietes} for reference on analytic functions and analytic manifolds over a Banach space. In this section~$\k$~is an ultrametric complete field and~$E, F$~are Banach spaces over~$\k$~(potentially of infinite dimension). As we shall see, taking~$\k = \mathbf{Q}_p$~and~$E, F = \mathbf{Q}_p^d$~allows one to recover the definition of converging power series and analytic functions over~$\mathbf{Q}_p^d$. Basically, if~$A$~is a Banach algebra over~$\mathbf{Q}_p$, then any map of the form~$f: A^d \rightarrow A$~such that locally at any point~$x \in A^d$, there is a expression of~$f$~as a converging power series \[ f(x + h) = \sum_{I \subset \mathbf{Z}_+^d} a_I h^I \] with~$a_I \in A, a_I \rightarrow 0$~is an analytic map from~$A^d$~to~$A$. The problem is that if~$A$~is not finite dimensional, this definition is not enough, as for example a continuous linear map is not necessarily described by an expression of this form but still should be analytic. \paragraph{Multi-indices, multi-linear maps.--} If~$\alpha = (\alpha_1, \cdots, \alpha_d) \in \mathbf{Z}_+^d$~is a multi-index, then~$\abs \alpha := \sum_i \alpha_i$. For~$1 \leq j \leq \abs \alpha$, we define \[ \alpha(j) = \max \left\{ k + 1 \in \mathbf{Z}_+ : \alpha_1 + \cdots + \alpha_{k} < j \right\}. \] The sequence~$(\alpha(j))_{1 \leq j \leq \abs \alpha}$~is the increasing sequence consisting of~$\alpha_1$~times the number 1,~$\alpha_2$~times the number 2, \ldots,~$\alpha_d$~times the number~$d$. For example, if~$\alpha = (1, 5 ,7)$, then~$d=3, \abs \alpha = 13$~and \[ (\alpha(j))_{1 \leq j \leq 13} = (1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3). \] For~$1 \leq i \leq d$, we denote by $p_i: E^d \rightarrow E$~the projection to the~$i$-th coordinate. For a multi-index~$\alpha \in \mathbf{Z}_+^d$, we define \[ p_\alpha := (p_{\alpha(j)})_{1 \leq j \leq \abs \alpha}: E^d \rightarrow E^{\abs \alpha}. \] If~$\beta \in \mathbf{Z}_+^d$~is another multi-index, then we write~$\alpha + \beta$~for the multi-index~$(\alpha_i + \beta_i)_{1 \leq i \leq d}$. We write~$\alpha \geq \beta$~if~$\alpha_i \geq \beta_i$~for all~$1 \leq i \leq d$; in that case there is a unique multi-index~$\gamma$~such that~$\alpha = \beta + \gamma$, and we set~$\alpha - \beta := \gamma$. We also define the binomial coefficient~$\binom{\alpha}{\beta} := \binom{\alpha_1}{\beta_1} \cdots \binom{\alpha_d}{\beta_d}$. Finally, if $\mathbf x = (x_1, \cdots, x_d)$, then~$\mathbf x^\alpha := x_1^{\alpha_1} \cdots x_d^{\alpha_d}$~ and if~$\mathbf y = (y_1, \cdots, y_d)$, one has the identity \begin{align*} (\mathbf x + \mathbf y)^\alpha &= (x_1 + y_1)^{\alpha_1} \cdots (x_d + y_d)^{\alpha_d} \\ &= \left(\sum_{\beta_1 = 0}^{\alpha_1} \binom{\alpha_1}{\beta_1}x_1^{\beta_1} y_1^{\alpha_1 - \beta_1} \right) \cdots \left( \sum_{\beta_d = 0}^{\alpha_d} \binom{\alpha_d}{\beta_d}x_d^{\beta_d} y_d^{\alpha_d - \beta_d} \right) \\ &= \sum_{0 \leq \beta_1 \leq \alpha_1} \cdots \sum_{0 \leq \beta_d \leq \alpha_d} \binom{\alpha_1}{\beta_1} \cdots \binom{\alpha_d}{\beta_d} x_1^{\beta_1} \cdots x_d^{\beta_d} y_1^{\alpha_1 - \beta_1} \cdots y_d^{\alpha_d - \beta_d} \\ &= \sum_{\beta \leq \alpha }\binom{\alpha}{\beta} \mathbf x^{\beta} \mathbf y^{\alpha - \beta}. \end{align*} For an integer~$k$, let~$\mathcal L_k (E, F)$~be the set of continuous multilinear maps from~$E^k$~to~$F$~equipped with the topology of uniform convergence over bounded subsets. The norm of an element~$\phi \in \mathcal L_k(E,F)$~is defined by \[ \norm \phi = \inf \left\{ a > 0 : \forall x_1, \cdots, x_k \in E^k, \norm{\phi(x_1, \cdots, x_k)}_F \leq a \norm{x_1}_E \cdots \norm{x_k}_E \right\}. \] \paragraph*{Continuous polynomial maps and power series.--} (\cite{bourbaki2007varietes} Appendix of \S 1-7) A \emph{continuous homogeneous polynomial map of multi degree}~$\alpha$, is a map~$f: E^d \rightarrow F$ such that there exists~$u \in \mathcal L_{\abs{\alpha}}(E,F)$~for which~$f = u \circ p_\alpha$. We denote by~$P_\alpha(E,F)$~the vector space of continuous homogeneous polynomial maps of multi-degree~$\alpha$~equipped with the quotient topology from~$\mathcal L_{\abs {\alpha}}(E,F)$. The norm of a continuous homogeneous polynomial map~$P \in P_\alpha (E,F)$~is defined by \[ \norm P := \inf_{u \in \mathcal L_{\abs{\alpha}} (E,F), P = u \circ p_\alpha } \norm u_{\mathcal L_{\abs{\alpha}}(E,F)}. \] \begin{ex} Set~$E, F = \mathbf{Q}_p \langle \mathbf x \rangle$. Let~$P$~be the monomial~$\mathbf x^\alpha$, then the map~$P: g \in \mathbf{Q}_p \langle \mathbf x \rangle^d \mapsto P(g) \in \mathbf{Q}_p \langle \mathbf x \rangle$~is a continuous homogeneous polynomial map of multi-degree~$\alpha$. Indeed, let~$k = \abs \alpha$~and consider the multilinear map \[ \begin{array}{lccc} {T_{k}}:& {E^{k}}& \longrightarrow & {F}\\ &{(f_1, \cdots, f_{k})} &\longmapsto &{f_1 \cdots f_{k}}; \end{array} \] it is continuous as~$\norm{ T_{k}(f_1, \cdots, f_{k})} \leq \norm{f_1} \cdots \norm{f_{k}}$~and~$~P = T_{k} \circ p_\alpha$. Furthermore, for a multi-index~$\beta$, define~$\phi_\beta: \mathbf{Q}_p \langle \mathbf x \rangle \rightarrow \mathbf{Q}_p \langle \mathbf x \rangle$~such that~$\phi_\beta (g)$~is the homogeneous part of multi-degree~$\beta$~of~$g$. Then,~$\phi_\beta$~is linear and continuous, therefore if~$P(\mathbf x) = \mathbf x^\alpha$, the map~$g \in \mathbf{Q}_p \langle \mathbf x \rangle^d \mapsto P(\phi_{\beta_1}(g_1), \cdots, \phi_{\beta_d}(g_d))$~is a continuous homogeneous polynomial map of multi-degree~$\alpha$~for any multi-index $(\beta_i)_{1 \leq i \leq d}$. \end{ex} For an integer~$k$,~$P_k (E^d,F)$~is the direct sum of the~$P_\alpha(E,F)$~for~$\alpha$~such that~$\abs \alpha = k$, the elements of~$P_k (E^d,F)$~are the \emph{continuous homogeneous polynomial maps of total degree~$k$.} \begin{ex}\label{ExampleOfhomogeneousContinuousPolynomialMap} If~$P \in \mathbf{Q}_p [\mathbf x]$~is a homogeneous polynomial of degree~$k$~in~$d$~variables, then the map~$P: g \in \mathbf{Q}_p \langle \mathbf x \rangle^d \mapsto P(g)$~is a continuous homogeneous polynomial map of total degree~$k$~and for any sequence of multi-index ~$(\beta_i)_{1 \leq i \leq d}$, the map~$g \in \mathbf{Q}_p \langle \mathbf x \rangle^d \mapsto P(\phi_{\beta_1}(g), \cdots, \phi_{\beta_d}(g))$~also is. \end{ex} We denote by~$P(E^d, F)$~the direct sum of the spaces~$P(E^d, F)$, its elements are the \emph{continuous polynomial maps in $d$~variables}. \begin{prop} Set~$E, F = \mathbf{Q}_p \langle \mathbf x \rangle$. Take a polynomial~$P \in \mathbf{Q}_p [\mathbf x ]$. Then,~$P$~induces a continuous polynomial map~$E^d \rightarrow F$~and the linear embedding~$\mathbf{Q}_p[\mathbf x] \hookrightarrow P(E^d, F)$~is an isometry. \end{prop} Finally, the set~$\hat P (E^d, F)$~of \emph{power series} in~$d$~variables over~$E$~is the (infinite) product of the~$P_\alpha(E, F)$~(or of the~$P_k(E^d, F)$) for~$\alpha \in \mathbf{Z}_+^d$~(for~$k \in \mathbf{Z}_+$) equipped with the product topology of the discrete topology over each factor; equivalently if~$f = \sum_\alpha f_\alpha \in \hat P(E^d, F)$, then the order of vanishing at~$0$~of $f$~is~$\ord (f) = \min \left\{ \abs \alpha : f_\alpha \neq 0 \right\}$~and this is the topology induced by the norm~$\norm f := 2^{- \ord(f)}$. The space~$\hat P(E^d, F)$~is complete Hausdorff for this topology. A \emph{converging power series} is an element~$f = \sum_\alpha f_\alpha$~of~$\hat P(E^d, F)$~such that there exists~$R \in (\mathbf{R}_{>0})^d$ satisfying $\sup_\alpha R^{\alpha} \norm{f_\alpha}_{P_\alpha(E,F)} < + \infty$. If~$f = \sum_\alpha f_\alpha$, then the \emph{polyradius of convergence of~$f$} is \[ r(f) := \sup \left\{ R \in (\mathbf{R}_{>0})^d : R^{\alpha} \norm{f_\alpha} \rightarrow 0 \text{ when } \abs \alpha \rightarrow \infty \right\}. \] \begin{dfn} Let~$\mathcal U$~be an open subset of~$E^d$, a map~$f: \mathcal U \rightarrow F$~is \emph{analytic} at a point~$a \in \mathcal U$~if there exists a converging power series~$f_a$~such that for all~$x$~in a small neighbourhood of~$a$~in ~$\mathcal U, f(a +x) = f_a (x)$. The function~$f$~is analytic if it is analytic at every point of~$\mathcal U$. For any integer~$m \geq 1$, a map~$f: \mathcal U \rightarrow F^m$~is analytic if each of its coordinates is analytic. \end{dfn} \begin{ex} Every continuous linear map~$\mathbf{Q}_p \langle \mathbf x \rangle^d \rightarrow \mathbf{Q}_p \langle \mathbf x \rangle^d$~is analytic. \end{ex} \begin{prop}\label{PropCompositionIsAnalytic} The map~$~\text{\emph{Comp}}: (h, f) \in \mathbf{Z}_p \langle \mathbf x \rangle^d \times \mathbf{Z}_p \langle \mathbf x \rangle^d \mapsto h \circ f \in \mathbf{Z}_p \langle \mathbf x \rangle^d$~is analytic. In particular, it is linear in~$h$. \end{prop} \begin{proof} It is enough to show that the map~$\Phi: (h,f) \in \mathbf{Z}_p \langle \mathbf x \rangle \times \mathbf{Z}_p \langle \mathbf x \rangle^d \mapsto h \circ f \in \mathbf{Z}_p \langle \mathbf x \rangle$~is analytic. Let~$(h, f) \in \mathbf{Z}_p \langle \mathbf x \rangle \times \mathbf{Z}_p \langle \mathbf x \rangle^d$, we show that~$\Phi$~is analytic at~$(h, f)$. Let~$g \in \mathbf{Z}_p \langle \mathbf x \rangle^d$~and write~$h(\mathbf x) = \sum_\alpha a_\alpha \mathbf x^\alpha$, then \begin{align*} h \circ (f + g(\mathbf x)) &= \sum_\alpha a_\alpha (f(\mathbf x) + g(\mathbf x))^\alpha \\ &= \sum_\alpha \sum_{\gamma \leq \alpha } a_\alpha \binom{\alpha}{\gamma} f(\mathbf x)^{\alpha - \gamma} g(\mathbf x)^{\gamma} \\ &= \sum_{\beta} \left( \sum_{\alpha \geq \beta} a_\alpha \binom{\alpha}{\beta}f(\mathbf x)^{\alpha - \beta} \right) g(\mathbf x)^\beta \\ &= \sum_\beta Q_{\beta,f}(h)(\mathbf x) \cdot g(\mathbf x)^\beta \end{align*} where~$Q_{\beta,f}: \mathbf{Q}_p \langle \mathbf x \rangle \rightarrow \mathbf{Q}_p \langle \mathbf x \rangle$~is a continuous linear map and $\norm {Q_\beta} \rightarrow 0$ when $\beta \rightarrow \infty$, this is a converging power series in the variables~$(h, g)$~of polyradius of convergence~$(+ \infty, 1)$. Therefore~$\Phi$~is analytic at any point~$(0, f)$~and by linearity in~$h, \Phi$~is analytic at any point~$(h,f)$. \end{proof} \paragraph{Analytic manifolds.--} Let~$\mathbf{K}$~be an ultrametric field and let~$X$~be a topological space. A~$\mathbf{K}$-\emph{chart} of~$X$~is a homeomorphism~$\phi: U \rightarrow \phi(U) \subset E$~where~$U$~in an open subset of~$X$~and~$E$~a Banach space over~$\mathbf{K}$. We say that two $\mathbf{K}$-charts~$\phi: U \rightarrow E, \psi: V \rightarrow F$~are \emph{compatible} if \begin{enumerate} \item~$\phi (U \cap V)$~is open in~$E$~and~$\psi (U \cap V)$~is open in~$F$. \item~$\psi \circ \inv \phi : \phi (U \cap V) \rightarrow F$~is analytic. \item~$\phi \circ \inv \psi: \psi (U \cap V) \rightarrow E$~is analytic. \end{enumerate} An analytic manifold~$X$~over~$\mathbf{K}$~is defined classically as a topological space equipped with an atlas of compatible~$\mathbf{K}$-charts. For a point~$x \in X$, the tangent space at~$x$~is denoted by~$T_x X$. A function~$f: X \rightarrow Y$~between two analytic manifolds is analytic if for every chart~$\phi: U \subset X \rightarrow E, \psi: V \subset Y \rightarrow F$, the map~$\psi \circ f \circ \inv \phi: \inv \phi(U) \rightarrow F$~is analytic. The differential of~$f$~at a point~$x$~will be denoted~$D_x f$. \begin{prop}\label{PropDiffAnIsAnAnalyticVariety} The topological space~$\Diff^{an}(\mathbf{Z}_p^d)$~is an analytic manifold over~$\mathbf{Q}_p$, it is in fact an open subset of the Banach space~$\mathbf{Q}_p \langle \mathbf x \rangle^d$. The subgroups~$\Diff^{an}_c(\mathbf{Z}_p^d)$ for~$c > \frac{1}{p-1}$~are diffeomorphic to~$\mathbf{Z}_p \langle \mathbf x \rangle^d$~and they form a basis of neighbourhood of ~$\id$~in~$\Diff^{an}(\mathbf{Z}_p^d)$. \end{prop} \begin{proof} Theorem \ref{theoremBellPoonen} shows that~$\Diff^{an}_c (\mathbf{Z}_p^d)$~is the ball of center~$\id$~and radius~$\abs p^c$~in~$\mathbf{Z}_p \langle \mathbf x \rangle^d$, using Proposition \ref{truc1} we see that for every~$f \in \Diff^{an}(\mathbf{Z}_p^d)$, the ball of center~$f$~and radius~$\abs p^c$~is included in~$f \circ \Diff^{an}_c(\mathbf{Z}_p^d)$~therefore it is an open set of~$\mathbf{Q}_p \langle \mathbf x \rangle^d$, so~$\Diff^{an}(\mathbf{Z}_p^d)$~is an infinite dimensional analytic manifold over~$\mathbf{Q}_p$. \end{proof} \paragraph{The implicit function theorem.--} Let~$X, Y, Z$~be manifolds over~$\mathbf{K}$~and let~$f: X \times Y \rightarrow Z$~be an analytic map. Let~$(a,b) \in X \times Y$, we write~$D_{(a,b)} f$~the differential map of~$f$~at~$(a,b)$~and let~$D_{(a,b)}^{(1)} f$~be the differential of the partial map~$x \in X \mapsto f(x, b)$~at~$a$~and~$D_{(a,b)}^{(2)} f$~the differential of the partial map~$y \in Y \mapsto f(a, y)$~at~$b$. Then, one has~$T_{(a,b)} X \times Y = T_a X \times T_b Y$ and~$D_{(a,b)}f (u,v) = D_{(a,b)}^{(1)}f \cdot u + D_{(a,b)}^{(2)} f \cdot v$. \begin{thm}[Implicit function theorem, 5.6.1 of \cite{bourbaki2007varietes}] Suppose that~$D_{(a,b)}^{(2)}f$~is bijective, then there exists an open neighbourhood~$U$~of~$a$~in~$X$~and an open neighbourhood~$V$~of~$b$~in~$Y$~and a unique analytic map~$g: U \rightarrow V$~such that \[ \forall x \in U, \quad f(x, g(x)) = f(a,b) \] and the differential of~$g$~at any~$x \in \mathcal U$~is given by \[ D_x g = - \inv{\left( D_{(x,g(x))}^{(2)} f \right)} \circ D_{(x,g(x))}^{(1)} f \] \end{thm} \begin{prop}\label{PropInvIsAnalytic} The inversion map~$\Inv : f \in \Diff^{an}(\mathbf{Z}_p^d) \mapsto \inv f$~is analytic. \end{prop} \begin{proof} We write~$\mathcal U = \Diff^{an}(\mathbf{Z}_p^d)$, we know that~$\mathcal U$~is an analytic manifold over~$\mathbf{Q}_p$~by Proposition \ref{PropDiffAnIsAnAnalyticVariety}. By Proposition \ref{PropCompositionIsAnalytic}, the composition operation is analytic over~$\mathbf{Z}_p \langle \mathbf x \rangle^d \times \mathbf{Z}_p \langle \mathbf x \rangle^d$, therefore it is over~$\mathcal U \times \mathcal U$. To show that~$\Inv$~is analytic we only need to show that it is analytic at~$\id$. Indeed, take~$f \in \mathcal U$, then~$\Inv = L_{\inv f} \circ \Inv \circ R_{\inv f}$~where~$R_{\inv f}$~is composition on the right by~${\inv f}$~and~$L_{\inv f}$~composition on the left. Since~$L_{\inv f}$ and $R_{\inv f}$ are analytic,~$\Inv$~is analytic at~$f$~if and only if it is analytic at~$\id$. To show that~$\Inv$~is analytic at~$\id$, we use the implicit function theorem, since the map~$M: (f,g) \in \mathcal U \times \mathcal U \rightarrow f \circ g \in \mathcal U$ is analytic and the partial differential~$D_{\id, \id}^{(2)} M = \id$, one has the existence of a unique function~$G: \mathcal V \rightarrow \mathcal \mathcal U$~with~$\mathcal V$~an open neighbourhood of~$\id$~such that~$G$~is analytic at ~$\id$~and~$M(f, G(f)) = \id$~for all~$f \in \mathcal V$. Therefore~$\Inv_{|\mathcal V} = G$~and inversion is analytic at~$\id$. \end{proof} \subsection{$p$-adic Lie groups} We refer to \cite{Bourbaki06} for more details on the results provided in this section. A~$p$-adic Lie group~$G$~is a topological group with a structure of a~$p$-adic analytic manifold such that the multiplication map and the inverse map are analytic. The dimension of~$G$~is its dimension as an analytic manifold. It can be infinite. Its \emph{Lie algebra}~$\mathfrak g$~is the tangent space of~$G$~at the neutral element, it is equipped with a Lie bracket~$[\cdot , \cdot]$~defined as follows. Let~$g \in G$~and~$\iota_g: h \in G \mapsto g h \inv g$, then~$\Ad (g) := D_e \iota_g \in \mathrm{GL}(\mathfrak g)$~is the adjoint representation of~$G$. Define~$\ad := D_e \Ad$, then \[ \forall \mathbf X, \mathbf Y \in \mathfrak g, [\mathbf X, \mathbf Y] := \ad(\mathbf X)(\mathbf Y). \] \begin{thm}\label{theoremDiffAnIsALieGroup} The topological group~$\Diff^{an} (\mathbf{Z}_p^d)$~is an infinite-dimensional Lie group over~$\mathbf{Q}_p$. Its Lie Algebra is~$\Theta(\mathbf{Z}_p^d)$. Moreover, the subgroups~$\Diff^{an}_c (\mathbf{Z}_p^d)$~are also Lie groups for~$c > \frac{1}{p-1}$~and they form a basis of neighbourhood of~$\id$~in~$\Diff^{an}_c(\mathbf{Z}_p^d)$. \end{thm} \begin{proof} The fact that~$\Diff^{an} (\mathbf{Z}_p^d)$~is a Lie group over~$\mathbf{Q}_p$~follows from Propositions \ref{PropCompositionIsAnalytic}, \ref{PropDiffAnIsAnAnalyticVariety} and \ref{PropInvIsAnalytic} where it was shown that it was an analytic manifold and that composition and inversion are analytic maps. The statement for ~$\Diff^{an}_c(\mathbf{Z}_p^d)$~follows from the same propositions. The tangent space at~$\id$~is~$\mathbf{Q}_p \langle \mathbf x \rangle^d~$~that we identify with~$\Theta(\mathbf{Z}_p^d)$~and under this identification the Lie bracket between two Tate-analytic vector fields corresponds to the Lie bracket of the Lie algebra of the Lie group~$\Diff^{an}(\mathbf{Z}_p^d)$~because if~$\mathbf X, \mathbf Y$~are of norm~$\leq \abs p^c$ with $c > \frac{1}{p-1}$, then they admit global Tate-analytic flows~$\Phi^\mathbf X$~and~$\Phi^\mathbf Y$~by Proposition \ref{PropExistenceGlobalTateAnalyticFlow} and \begin{eqnarray*} [\mathbf X, \mathbf Y] &= \frac{\partial }{\partial_s}_{|s=0} \frac{\partial }{\partial t}_{|t=0} \Phi^\mathbf X_{-s} \circ \Phi^\mathbf Y_t \circ \Phi^\mathbf X_s \\ &= \frac{\partial }{\partial_s}_{|s=0} \frac{\partial }{\partial t}_{|t=0} \iota_{\Phi^\mathbf X_s} (\Phi^\mathbf Y_t) \\ &= D_{\id} \Ad(\mathbf X) (\mathbf Y) = \ad(\mathbf X) (\mathbf Y). \end{eqnarray*} On the other hand, if~$f,g \in \Diff^{an}_c(\mathbf{Z}_p^d)$~with~$c > \frac{1}{p-1}$ , then~$\frac{\partial}{\partial s}_{|s=0} \frac{\partial}{\partial t}_{|t=0} \Phi^f_{-s} \circ \Phi^g_t \circ \Phi^f_s = [\mathbf X_f, \mathbf X_g] = \ad\mathbf X_f (\mathbf X_g)$. \end{proof} \begin{rmq} Since Bell-Poonen theorem holds for any ultrametric field, the same proof shows that $\Diff^{an}(\D_p^d)$ is a Lie group over~$\mathbf{C}_p$. In fact, for any complete extension~$\mathbf{K}$~of~$\mathbf{Q}_p$~with unit ball~$\mathbf A$, the group~$\Diff^{an}(\mathbf A^d)$~is a Lie group over~$\mathbf{K}$. \end{rmq} \begin{thm}[\cite{Bourbaki06}, \S 8, Theorem 1]\label{PropContinousMorphismIsAnalytic} Let~$G, H$~be Lie groups over~$\mathbf{Q}_p$~and~$\phi: G \rightarrow H$~be a continuous homomorphism of topological groups. Then,~$\phi$~is analytic and therefore a homomorphism of Lie groups. \end{thm} \begin{rmq} The proof relies heavily on~$\mathbf{Q}$~being dense in~$\mathbf{Q}_p$~and the theorem is false if we replace ~$\mathbf{Q}_p$~by any finite extension of~$\mathbf{Q}_p$. Indeed, suppose for example that~$K = \mathbf{Q}_p (\sqrt \alpha)$~is a quadratic extension. Any element~$z$~of~$\mathbf{K}$~is of the form~$z = x + \sqrt \alpha y$. Then, the function \[ f: z = x + \sqrt \alpha y \mapsto x - \sqrt \alpha y \] is a continuous group homomorphism, it is~$\mathbf{Q}_p$-analytic but not~$\mathbf{K}$-analytic as~$f_{|1 \cdot \mathbf{Q}_p} = \id$~and~$f_{|\sqrt \alpha \cdot \mathbf{Q}_p} = - \id$. \end{rmq} Let~$\Gamma$~be a finitely generated group, the pro-$p$ completion~$\Gamma_p$~of~$\Gamma$~is the projective limit of the quotient of~$\Gamma$~that are finite~$p$-groups, it is a topological group with respect to the profinite topology. In particular, for any~$\gamma \in \Gamma$, the group homomorphism~$n \in \mathbf{Z} \mapsto \gamma^n \in \Gamma$~extends uniquely to a continuous group homomorphism~$t \in \mathbf{Z}_p \mapsto \gamma^t \in \Gamma_p$. In the context of Tate-analytic diffeomorphisms, if~$p \geq 3$~and~$f \equiv \id \mod p$, then the extension $n \in \mathbf{Z} \mapsto f^n \in \Diff^{an}_1 (\mathbf{Z}_p^d)$~is the Tate-analytic flow~$t \in \mathbf{Z}_p \mapsto \Phi_t^f \in \Diff^{an}(\mathbf{Z}_p^d)$~associated to~$f$~given by Bell-Poonen theorem. \begin{prop}\label{PropEmbeddingOfProPcompletion} Let~$p$~be a prime, let $c>0$ be such that $c > \frac{1}{p-1}$ and let~$G$~be a compact Lie group over~$\mathbf{Q}_p$. Let $\Gamma$~be a finitely generated subgroup of~$G$~such that~$G$~is the pro-$p$-completion of~$\Gamma$ and let~$\iota: \Gamma \rightarrow \Diff^{an}_c (\mathbf{Z}_p^d)$~be a group homomorphism, then~$\iota$~extends uniquely to a Lie group homomorphism~$\iota: G \rightarrow \Diff^{an}_c (\mathbf{Z}_p^d)$~such that for all~$t \in \mathbf{Z}_p$, all~$g \in \Gamma$,~$\iota(g^t) = \iota(g)^t$~and the map~$(t, \mathbf x) \in \mathbf{Z}_p \times \mathbf{Z}_p^d \mapsto \iota(g)^t (\mathbf x)$~is analytic. \end{prop} \begin{proof} Theorem 2.11 of \cite{cantat2014algebraic} shows that~$\iota$~extends uniquely to a continuous map. In \cite{cantat2014algebraic} this is only shown when $p \geq 3$ and $c = 1$ but the proof is identical with $p \geq 2$ and $c > \frac{1}{p-1}$ at it is only required that the image of the elements of $\Gamma$ admits a Tate-analytic flow. Since~$G$~and~$\Diff^{an}_c (\mathbf{Z}_p^d)$~are both Lie groups over~$\mathbf{Q}_p$,~$\iota$~is automatically a Lie group homomorphism by Theorem \ref{PropContinousMorphismIsAnalytic}. \end{proof} \begin{thm}[\cite{Bourbaki06}, \S 8, Theorem 2]\label{theoremClosedSubgroupAreLieGroups} Let~$G$~be a finite-dimensional Lie group over~$\mathbf{Q}_p$, then every closed subgroup of~$G$~is a Lie subgroup of~$G$. \end{thm} \begin{prop}[\cite{Bourbaki06}, \S 9, Corollary of Proposition 6] \label{PropOpenSubgroupDerivedSeries} Let~$G$~be a finite-dimensional Lie group over~$\mathbf{Q}_p$~and $\mathfrak g$ its Lie algebra, there exists an open subgroup ~$G_0$~of~$G$~such that for all~$i \geq 0$, the subgroups~$D^i(G_0)$~and~$D_i(G_0)$ are Lie subgroups with Lie algebra ~$\mathcal D^i (\mathfrak h)$~and~$\mathcal D_i (\mathfrak h)$~respectively. \end{prop} \subsection{Nilpotent groups and embedding into~$p$-adic Lie groups.} \subsubsection{Nilpotent groups}\label{SubSubSecNilpotentGroups} The main goal of this section is to show that if~$H$~is a finitely generated nilpotent group with generators~$h_1, \ldots, h_s$, then for any~$m \geq 1$~the subgroup~$H_m$~of~$H$~generated by~$h_1^m, \ldots, h_s^m$~is a finite index subgroup of~$H$. This will be useful in the proof of Theorem \ref{BoundNilpotentGroups} because if~$H \subset \Diff^{an}_1 (\mathbf{Z}_p^d)$~we will need to consider such a subgroup~$H_m$~to get the desired result. Recall the notation introduced in \S~\ref{par:nilpotent_and_solvable} for nilpotent and solvable groups and Lie algebras. We shall say that an expression that involves~$k$~commutator brackets is a commutator of length~$k$; for instance~$[ [a, [b,c]], d]$~is a commutator of length 3 and a single element can be viewed as a commutator of length 0. For~$k \geq 1$, we denote by~$[a_1; \cdots; a_k]$~the commutator~$[a_1, [a_2, \cdots, [a_{k-1}, a_k] \cdots]$; its length is~$k$. Let~$G, G', G''$~be groups, a map~$\phi: G \times G' \rightarrow G''$~is \emph{bilinear} if for every~$g \in G, g' \in G'$, the maps~$\phi(g, \cdot)$~and $\phi( \cdot, g')$ are group homomorphisms. More generally, a map~$G_1 \times \cdots \times G_m \rightarrow G$~is~$m$-linear if fixing~$m-1$~coordinates yields a group homomorphism. For any triple of elements~$x,y,z$~in~$G$, we have \begin{itemize} \item~$\inv{[x,y]} = [y,x]$. \item~$[x,yz] = [x,y] [ y, [x,z]] [x,z]$. \item~$[xy,z] = [x,[y,z]] [y,z] [x,z]$. \end{itemize} The image of the map~$(a,b) \mapsto [a,b]$~from~$G \times D^{k-1}(G)$~to~$D^k(G)$~generates $D^k (G)$. It follows from the last three formulas that, for every~$k \geq 1$, this map induces a bilinear map \[ \mathrm{co}_k : G \times D^{k-1}(G) \mapsto D^k(G) / D^{k+1} (G) \] and the image $\im \mathrm{co}_k$ generates $D^k(G) / D^{k+1}(G)$. \begin{prop}\label{PropGeneratorsOfLowerCentralSeries} Let~$G$~be a group and~$S$~a set of generators of~$G$. \begin{enumerate} \item for every integer~$k \geq 0$, the subgroup~$D^{k}(G) / D^{k+1}(G)$~is generated by the commutators of length~$k$~consisting of elements of~$S$. \item if~$G$~is finitely generated, then~$D^{k}(G) / D^{k+1}(G)$~is finitely generated for every~$k \geq 0$. \item If~$G$~is nilpotent, then~$D^{\nilp(G)-1} (G)$~is generated by the commutators of length~$\nilp(G) -1$~in elements of~$S$. \end{enumerate} \end{prop} \begin{proof} Let us prove the first assertion by induction on~$k$. Let~$X_k$~be the set of commutators of length~$k$ in elements of~$S$. The initialization~$k=0$~follows from~$X_0 = S$~and the fact that~$S$~generates ~$G$. Now, suppose~$k \geq 1$~and that~$X_{k-1}$~generates~$D^{k-1} (G) / D^{k}(G)$. The image of the map~$\mathrm{co}_k$~generates $D^k(G) / D^{k+1} (G)$; by induction and since~$\mathrm{co}_k (a,b)$~is a homomorphism with respect to~$a$~and with respect to~$b$, the elements~$[s, x_{k-1}]$~for~$s$~in~$S$~and $x_{k-1} \in X_{k-1}$~generate~$D^k(G) / D^{k+1}(G)$, and these elements are exactly the commutators of length~$k$~in the elements of~$S$. The second and third assertions follow from the first one. \end{proof} \begin{prop}\label{PropSubgroupIsFinitelyGenerated} Let~$H$~be a finitely generated nilpotent group, then every subgroup~$H_0$~of~$H$~is finitely generated. \end{prop} For a proof see \cite{segal2005polycyclic} where this is actually shown for polycyclic groups, the result follows since finitely generated nilpotent groups are polycyclic. \begin{prop}\label{CorMapNLin} Let~$H$~be a nilpotent group of nilpotency class~$t$. \begin{enumerate} \item the map~$\mathrm{Br}_t : H^t \rightarrow D^{t-1}, (h_1, \cdots, h_t) \mapsto [h_1; h_2; \cdots; h_t]$~is multilinear. \item If~$\left\{ h_1, \cdots, h_s \right\}$~generates~$H$, then for every~$m \geq 1$, the subgroup generated by~$\left\{ h_1^m, \cdots, h_s^m \right\}$~is of finite index in~$H$. \end{enumerate} \end{prop} \begin{proof}[Proof of the first assertion] Let us do an induction on~$t$. The case~$t=1$~being trivial, suppose the result true for a nilpotent group of class~$t-1$~and consider~$H$~a nilpotent group of class~$t$. Since~$D^t(H) = 0$, one has that the map~$\mathrm{co}_{t-1}: (h_1,h) \in H \times D^{t-2}(H) / D^{t-1}(H) \mapsto [h;x] \in D^{t-1}(H)$~is bilinear; thus,~$\mathrm{Br}_t$~is a homomorphism with respect to the first factor~$h_1 \in H$. Let us show that~$\mathrm{Br}_t$~is a homomorphism in the second coordinates~$h_2$, the other coordinates are dealt with in the same way. By induction, the map \[ \mathrm{Br}_{t-1}^{H / D^{t-1}(H)}: (H / D^{t-1}(H))^{t-1}\rightarrow D^{t-2}(H) / D^{t-1}(H) \] is~multilinear. Take~$h_1, h_2, h_2', h_3, \cdots, h_{t-1} \in H$, the multilinearity of~$\mathrm{Br}_{t-1}^{H / D^{t-1}(H)}$~provides an element~$g \in D^{t-1}(H)$~such that \[ [h_1; h_2 h_2'; \cdots ; h_{t-1}] = [h_1, [h_2; \cdots; h_{t-1}]\cdot [h_2 '; \cdots; h_{t-1}] \cdot g] \] and the bilinearity of~$\mathrm{co}_{t-1}$~gives the result since~$[h_1, g] =0$. \end{proof} \begin{proof}[Proof of the second assertion] We set~$S=\left\{ h_1, \cdots, h_s \right\}$~ and we denote by~$H_{S,m}$~the subgroup of~$H$~generated by the set ~$\left\{ s^m : s \in S \right\}$. We show by induction on~$t=\nilp(H)$~that~$H_{S,m}$~is of finite index in~$H$. If~$t=1$~then~$H$~is abelian and there is a unique surjective group homomorphism~$\mathbf{Z}^s \rightarrow H$~sending the canonical basis to~$S = (h_1, \cdots, h_s)$. The subgroup~$H_{S,m}$~is the image of~$m \mathbf{Z}^s$. Therefore, there is a surjective group homomorphism~$\mathbf{Z}^s / m \mathbf{Z}^s \twoheadrightarrow H / H_{S,m}$~and we get that~$H / H_{S,m}$~has at most~$m^s$~elements. Now suppose the result true for a group of nilpotency class~$t-1$~and assume~$\nilp (H) = t$, with~$t \geq 2$. Set~$T := D^{t-1}(H)$,~$T$~is central in~$H$. One has the exact sequence \[ 1 \rightarrow T \rightarrow H \rightarrow H/ T \rightarrow 1. \] By induction, the image of~$H_{S,m}$~in~$H / T$~is of finite index; thus, one can fix a finite set~$A \subset H$~such that~$H = \bigsqcup_{h \in A} h H_{S,m} T$. To conclude, we only need to show that the index of~$T \cap H_{S,m}$~in~$T$~is finite. Since,~$T \cap H_{S,m}$~contains the subgroup of~$t-1$~commutators~$D^{t-1}(H_{S,m})$~it suffices to show that the index of~$D^{t-1}(H_{S,m})$ in~$T$~is finite. By Proposition \ref{PropGeneratorsOfLowerCentralSeries},~$T$~is generated by the set~$S' = \left\{ [x_1; \cdots; x_{t-1}]: x_i \in S \right\}$~and~$D^{t-1}(H_{S,m})$~is generated by the set~$S'' = \left\{[x_1^m; \cdots; x_{t-1}^m] : x_i \in S \right\}$~furthermore, the first assertion shows that~$S''$~consists exactly of the elements of~$S'$~raised to the power~$m^{t-1}$. So by the abelian case,~$D^{t-1}(H_{S,m})$~is of finite index in~$T$. \end{proof} \subsubsection{Malcev's completion of nilpotent torsion-free finitely generated group} Denote by~$\hat \mathbf{Z} = \prod_{p \text{ prime}} \mathbf{Z}_p$~equipped with the product topology (the adelic topology). It is the profinite completion of~$\mathbf{Z}$. Let~$H$~be a nilpotent torsion-free finitely generated group. It is known that~$H$~embeds into ~$\Tri_1(n, \mathbf{Z})$~the group of upper triangular matrices with integer coefficients and 1's on the diagonal for some integer~$n$~(see for example \cite{segal2005polycyclic} Theorem 2 of Chapter 5). For the rest of this section, we fix an embedding~$\iota: H \hookrightarrow \Tri_1(n, \mathbf{Z})$. There are two topologies that one can consider on~$\iota(H)$. First the adelic topology induced by the inclusion~$\Tri_1(n, \mathbf{Z}) \subset \Tri_1(n, \hat \mathbf{Z})$, and second, the profinite topology where a basis of neighbourhood for the neutral element are the subgroups of finite index in~$\iota(H)$. \begin{prop}\label{PropAdelicTopologyAndProfiniteTopologyAreTheSame} Let~$G \subset \Tri_1(n, \mathbf{Z})$~be a subgroup of matrices with integer coefficients and 1's on the diagonal, then the profinite topology and the adelic topology on~$G$~are the same. In particular, the profinite completion of $G$ coincides with the closure of $G$ in $\Tri_1 (n , \hat \mathbf{Z})$. \end{prop} \begin{proof} First, let~$K$~be a subgroup of~$\mathrm{GL}_n(\mathbf{Z})$~of the form~$K =\left\{ A \in \mathrm{GL}_n(\mathbf{Z}) : A \equiv \id \mod m \right\}$~for some integer~$m$, such groups~$K$~form a basis of open neighbourhood of~$\id$~for the adelic topology. It is a normal subgroup of~$\mathrm{GL}_n(\mathbf{Z})$~with finite quotient, therefore~$G \cap K$~is a finite index subgroup of~$G$. Therefore the adelic topology is finer than the profinite topology. Conversely,~$G$~is a unipotent group of matrices over~$\mathbf{Q}$, therefore it is arithmetic (see \cite{segal2005polycyclic} Exercise 13 of Chapter 6). By the affirmative solution to the congruence subgroup problem for arithmetic soluble groups (see \cite{chahal1980}), we get that~$G$~is a congruence subgroup. This means that every finite index subgroup of~$G$ contains a subgroup of the form~$G \cap \left\{ A \in \mathrm{GL}_n (\mathbf{Z}) : A \equiv \id \mod m \right\}$~for some integer~$m$. Therefore, the profinite topology is finer than the adelic topology; thus, they are the same. \end{proof} A consequence of this proposition is that the profinite completion of~$\iota(H)$~is exactly the closure of ~$\iota(H)$~in $\Tri_1(n, \hat \mathbf{Z})$. \begin{prop}\label{PropCompletionIsProP} Let~$G$~be a nilpotent subgroup of~$\Tri_1(n, \mathbf{Z})$. The closure of~$G$~in~$\Tri_1(n, \mathbf{Z}_p)$~is the pro-$p$-completion of~$G$, in particular it is a~$p$-adic Lie group. \end{prop} \begin{proof} Denote by~$\hat G$~the profinite completion of~$G$~and for a prime~$\ell$,~$G_\ell$~the pro-$\ell$-completion of~$G$. Since~$G$~is nilpotent and a finite nilpotent group is a product of~$\ell$-groups for some primes~$\ell$~(see \cite{bourbaki1970algebre} chapter 1, \S 7, Theorem~4) we have that~$\hat G = \prod_\ell G_\ell$. By Proposition \ref{PropAdelicTopologyAndProfiniteTopologyAreTheSame}, we have a continuous injective homomorphism of topological groups \[ \hat G = \prod_\ell G_\ell \hookrightarrow \Tri_1(n, \hat \mathbf{Z}) = \prod_\ell \Tri_1(n, \mathbf{Z}_\ell). \] For a prime~$p$, this induces a continuous group homomorphism~$G_p \hookrightarrow \prod_\ell \Tri_1(n, \mathbf{Z}_\ell)$. But,~$G_p$~is a pro-$p$-group and for every prime~$\ell$,~$\Tri_1(n, \mathbf{Z}_\ell) = \varprojlim \Tri_1(n, \mathbf{Z} /\ell^k \mathbf{Z})$~is a pro-$\ell$-group. Therefore,~$G_p$~can be identified with the image of~$\hat G$~in~$\Tri_1(n, \mathbf{Z}_p)$; this is exactly the completion of~$G$~in~$\Tri_1(n, \mathbf{Z}_p)$, meaning that~$G_p$~is a closed subgroup of the~$p$-adic Lie group~$\Tri_1(n, \mathbf{Z}_p)$, so it is a Lie group by Theorem \ref{theoremClosedSubgroupAreLieGroups}. \end{proof} \begin{thm}\label{BigtheoremPropClosureIsALieGroup}\label{MinorationpAdic2} Let $c>0$ be such that $c > \frac{1}{p-1}$ and let~$H$~be a finitely generated nilpotent subgroup of~$\Diff^{an}_c(\mathbf{Z}_p^d)$, then the closure~$\bar H$~of ~$H$~in~$\Diff^{an}(\mathbf{Z}_p^d)$~is a finite-dimensional nilpotent Lie group. Furthermore, denote by~$\mathfrak h$~the Lie algebra of~$\bar H$, then~$\mathfrak h$~is a finite-dimensional nilpotent Lie algebra and ~$\dl(\mathfrak h) \geq \vdl(H)$. \end{thm} \begin{proof} Set~$G = \iota(H)$~and~$\psi := \inv \iota: G \rightarrow \Diff^{an}_c(\mathbf{Z}_p^d)$. By Proposition \ref{PropCompletionIsProP} and Proposition \ref{PropEmbeddingOfProPcompletion},~$\psi$ extends to a Lie group homomorphism~$\psi: G_p \rightarrow \Diff^{an}_c(\mathbf{Z}_p^d)$~where~$G_p$~is the closure of~$G$~in~$\Tri_1 (n, \mathbf{Z}_p)$; we show that the image of~$\psi$~is the closure of~$H$~in~$\Diff^{an}(\mathbf{Z}_p^d)$. Let~$K$~be the image of~$\psi$. Since~$\Tri_1(n,\mathbf{Z}_p)$~is compact and~$G_p$~is closed,~$G_p$~is also compact and so is~$K$. This implies that the closure~$\overline H$~of~$H$~is included in~$K$. And~$K$~is included in~$\overline H$~because of the continuity of~$\psi$. This shows that~$\overline H$~is a finite dimensional Lie group isomorphic to~$G_p / \ker \psi$. Now, we show the statement for~$\mathfrak h$. By Proposition \ref{PropOpenSubgroupDerivedSeries}, there exists an open subgroup~$H_1$~of~$\overline H$, such that~$D^i (H_1)$~is a Lie subgroup of~$\overline H$~with Lie algebra~$\mathcal D^i (\mathfrak h)$. Since~$H_1$~is open, by Theorem \ref{theoremDiffAnIsALieGroup} there exists an integer~$c >0$~such that~$\Diff^{an}_c (\mathbf{Z}_p^d) \cap H \subset H_1$. Take~$f_1, \cdots, f_s$~generators of~$H$. Then by Proposition \ref{CorMapNLin} the subgroup~$H'$~generated by the~$f_i^{p^c}$'s is a finite index subgroup of~$H$~and it is included in~$H_1$~by Lemma \ref{lemma:PuissanceCongruence}, therefore~$\dl (\mathfrak h) = \dl(H_1) \geq \dl(H') \geq \vdl(H)$. \end{proof} \section{Finitely generated nilpotent groups}\label{SecFinitelyGeneratedNilpotentGroups} \subsection{Base change from~$\mathbf{C}$~to~$\mathbf{Z}_p$: Good models} To prove Theorem \ref{BoundNilpotentGroups}, we shall ultimately apply Theorem \ref{BigtheoremPropClosureIsALieGroup}. Thus, we need a method to transfer problems regarding groups of automorphisms defined over~$\mathbf{C}$~to similar problems on groups of Tate analytic diffeomorphisms over~$\mathbf{Z}_p$, for certain primes~$p$. \begin{thm}[Lech, see \cite{Lech53}] \label{theoremLechPlongementPadique} Let~$\mathbf{K}$~be a finitely generated field over~$\mathbf{Q}$ and let~$S$~be a finite subset of~$\mathbf{K}$. Then there exists an infinite number of prime numbers~$p$~with an embedding~$\mathbf{K} \hookrightarrow \mathbf{Q}_p$~such that all elements of~$S$~are mapped to~$\mathbf{Z}_p$. \end{thm} Let~$X$~be an irreducible quasiprojective variety over~$\mathbf{C}$~and~$\Gamma$~a finitely generated subgroup of~$\Aut (X_\mathbf{C})$. \begin{itemize} \item Let~$R$~be an integral domain. We say that~$(X,\Gamma)$~is \emph{defined over }~$R$, if there exists an irreducible separated reduced scheme~$X_R$~over~$R$~and an injective homomorphism~$\Gamma \hookrightarrow \Aut_\mathbf{R} (X_\mathbf{R})$ such that~$X$~and~$\Gamma$~are obtained by the base change~$X = X_R \times_{\Spec R} \Spec \mathbf{C}$. \item Let~$p$~be a prime number. A \emph{model} of~$(X,\Gamma)$~over~$\mathbf{Z}_p$~is the data of \begin{enumerate}[label=(\roman*)] \item A ring~$R \subset \mathbf{C}$~over which~$(X, \Gamma)$~is defined and an embedding~$R \hookrightarrow \mathbf{Z}_p$. \item An irreducible variety~$\mathcal X$~over~$\mathbf{Z}_p$~and an injective homomorphism~$\rho: \Gamma \hookrightarrow \Aut_{\mathbf{Z}_p} (\mathcal X)$~such that \[ \mathcal X \simeq X_R \times_{\Spec R} \Spec \mathbf{Z}_p. \] is the base change of~$X_R$~and for all~$f \in \Gamma$, ~$\rho (f)$~is the base change of~$f$. \end{enumerate} \item A \emph{good model} over~$\mathbf{Z}_p$~of~$(X,\Gamma)$~is the data of a model of~$(X,\Gamma)$~with the additional condition that the special fiber~$\mathcal X_{\mathbf{F}_p} = \mathcal X \times_{\Spec \mathbf{Z}_p} \Spec \mathbf{F}_p$~is geometrically reduced and irreducible and of dimension \[ \dim_{\mathbf{F}_p} (\mathcal X_{\mathbf{F}_p}) = \dim_{\mathbf{Q}_p} (\mathcal X \times_{\Spec R} \Spec \mathbf{Q}_p). \] \end{itemize} \begin{prop}[Proposition 4.4 of \cite{bell2010dynamical}, Proposition 3.2 of \cite{cantat2014algebraic}]\label{FromCtoZp} Let~$X$~be an irreducible complex quasi-projective variety,~$\alpha \in X(\mathbf{C})$~and~$\Gamma$~be a finitely generated subgroup of~$\Aut_\mathbf{C} (X)$. Then, there exists an infinite number of primes~$p \geq 3$~such that~$(X,\Gamma)$~has a good model~$\mathcal X$~over~$\mathbf{Z}_p$~and such that~$\alpha$~extends to a section~$\alpha: \Spec \mathbf{Z}_p \rightarrow \mathcal X$. \end{prop} \begin{ex} For simplicity, suppose~$X$~is the affine space~$~\mathbf{A}^d_\mathbf{C}$~with its standard coordinates~$x_1,\cdots, x_d$~and~$\Gamma \subset \Aut(\mathbf{A}^d_\mathbf{C})$~is a finitely generated group of polynomial automorphisms. This is already an interesting example. Let~$S$~be a finite symmetrical~$(\inv S = S)$~set of generators of~$\Gamma$. Let~$R$~be the ring generated by all the coefficients of the elements of~$S$~and the coordinates of~$\alpha$. Then,~$(X, \Gamma)$~is defined over~$R$. Plus, by Theorem \ref{theoremLechPlongementPadique} there exists a prime~$p$~and an embedding~$\iota: R \hookrightarrow \mathbf{Z}_p$. Using this embedding, the base change~$\mathcal X = \mathbf{A}^d_{\mathbf{Z}_p}$~and~$\rho: \Gamma \hookrightarrow \Aut (\mathbf{A}^d_{\mathbf{Z}_p})$~show that~$(\mathbf{A}^d, \Gamma)$~is a good model over~$\mathbf{Z}_p$~and~$\alpha$~extends to a ~$\mathbf{Z}_p$-point of~$\mathcal X$. \end{ex} \subsection{From algebraic automorphisms to analytic diffeomorphisms over~$\mathbf{Z}_p$} In this section, we consider a scheme ~$\mathcal X$~of dimension~$d$~over~$\mathbf{Z}_p$, where~$p\geq 3$~is a prime number, such that \begin{itemize} \item~$\mathcal X$~is a quasi-projective variety over~$\mathbf{Z}_p$, and its generic fiber is geometrically irreducible over ~$\mathbf{Q}_p$. \item~$\overline \mathcal X = \mathcal X \times_{\Spec \mathbf{Z}_p} \Spec \mathbf{F}_p$~is the special fiber of~$\mathcal X$~and is geometrically irreducible over~$\mathbf{F}_p$. \item~$f: \mathcal X \rightarrow \mathcal X$~is an automorphism of~$\mathbf{Z}_p$-schemes. \item~$\overline f : \overline \mathcal X \rightarrow \overline \mathcal X$~is the restriction of~$\mathcal X$~to the special fiber. \item~$r: \mathcal X (\mathbf{Z}_p) \rightarrow \overline{\mathcal X} (\mathbf{F}_p)$~is the reduction map. \item~$x$~is a smooth~$\mathbf{F}_p$-point and there exists~$\alpha \in \mathcal X(\mathbf{Z}_p)$~such that~$r(\alpha) = x$. \end{itemize} For the two next propositions, we refer to \cite{bell2010dynamical}. They will enable us to go from algebraic automorphisms to analytic diffeomorphisms. \begin{prop}\label{PropExistenceOfIota} Let~$\mathcal X$~be a quasi-projective scheme over~$\mathbf{Z}_p$. There exists a function~$\iota: \mathbf{Z}_p^d \rightarrow \mathcal X(\mathbf{Z}_p)$~which induces an analytic bijection between~$\mathbf{Z}_p^d$~and the open subset of~$\mathcal X (\mathbf{Z}_p)$~consisting of the points~$\beta$~such that~$r(\beta) = x$. \end{prop} \begin{prop}\label{PropConjugaisonDiffeoAnalytique} Suppose that~$\bar f (x) = x$. Let~$\iota: \mathbf{Z}_p^d \rightarrow \mathcal X (\mathbf{Z}_p)$~be the function defined in Proposition \ref{PropExistenceOfIota}. Then there exist analytic functions~$F_1,\cdots,F_d \in \mathbf{Z}_p \langle T_1,\cdots, T_d \rangle$~such that \begin{enumerate}[label = (\roman*)] \item One has \[ \inv \iota \circ f \circ \iota = (F_1,\cdots,F_d) =: \mathcal F \in \mathbf{Z}_p \langle T_1,\cdots,T_d \rangle^d. \] \item if~$\bar{\mathcal F}$~is the reduction mod~$p$~of~$\mathcal F$, then~$\bar{\mathcal F} = \mathcal F_0 + \mathcal F_1$~with~$\mathcal F_0 \in (\mathbf{Z}/p\mathbf{Z})^d$~and~$\mathcal F_1 \in \mathrm{GL}_d(\mathbf{Z} / p \mathbf{Z})$. \end{enumerate} Furthermore~$\mathcal F$~is a Tate-analytic diffeomorphism because~$f$~is an automorphism. \end{prop} \begin{ex} Propositions \ref{PropExistenceOfIota} and \ref{PropConjugaisonDiffeoAnalytique} are proven in \cite{bell2010dynamical}. We only do the proof in the case~$\mathcal X = \mathbf{A}^d_{\mathbf{Z}_p}$. Take standard coordinates~$\mathbf x = x_1, \cdots, x_d$~over~$\mathcal X$. Then,~$\mathcal X = \Spec \mathbf{Z}_p[\mathbf x]$~and~$\overline \mathcal X = \Spec \mathbf{F}_p[\mathbf x]$. The reduction map~$r: \mathcal X(\mathbf{Z}_p) = \mathbf{Z}_p^d \rightarrow \overline \mathcal X(\mathbf{F}_p) = \mathbf{F}_p^d$~is the reduction mod~$p$~coordinates by coordinates. Take~$x \in \mathbf{F}_p^d$~and~$z \in \mathbf{Z}_p^d$~such that~$r(z) = x$, then the open subset of~$\mathcal X(\mathbf{Z}_p)$~of elements~$\beta$ such that~$r(\beta) =x$~is the ball of center~$z$~and radius~$1 / p$. The analytic bijection~$\iota$~is given by ~$\iota: m \in \mathbf{Z}_p^d \mapsto z + p \cdot m \in \mathcal X(\mathbf{Z}_p) = \mathbf{Z}_p^d$. This proves Proposition \ref{PropExistenceOfIota}. Now, take a polynomial automorphism~$f$, the map~$\overline f$~is the polynomial automorphism over~$\mathbf{F}_p^d$~obtained when taking the coefficients of~$f \mod p$. Take a point~$x \in \mathbf{F}_p^d$~such that~$\bar f (x) = x$, up to a conjugation by a translation (which does not change the result), we can suppose that~$x = 0 \in \mathbf{F}_p^d$. This means that~$f$~preserves the ball of center 0 and radius~$1/p$~in~$\mathbf{Z}_p^d$. Writing~$f$~in coordinates, we have \[ f(\mathbf x) = p a_0 + A_1 (\mathbf x) + A_2(\mathbf x) + \cdots \] where~$a_0 \in \mathbf{Z}_p^d$~and~$A_i$~is the homogeneous part of degree~$i$~of~$f$. Then, \[ \inv \iota \circ f \circ \iota (\mathbf x) = \frac{1}{p} f(p \mathbf x) = a_0 + A_1(\mathbf x) + \sum_{k \geq 2} p^{k-1} A_k (\mathbf x). \] This is indeed an element of~$\mathbf{Z}_p \langle \mathbf x \rangle^d$~and~$\overline{\frac{1}{p}f(p \mathbf x)}$~is an invertible affine transformation of~$\mathbf{F}_p^d$, this proves Proposition \ref{PropConjugaisonDiffeoAnalytique}. \end{ex} \begin{prop}\label{FromAlgAutoToDIffAnal}[Proposition 3.3 of \cite{cantat2014algebraic}] Let~$\Gamma$~be a finitely generated subgroup of~$\Aut_{\mathbf{Z}_p} (\mathcal X)$. There exists a finite index subgroup~$\Gamma_0 \subset \Gamma$~and an open subset~$\mathcal U \subset \mathcal X (\mathbf{Z}_p)$ analytically diffeomorphic to~$\mathbf{Z}_p^d$~such that~$\mathcal U$~is stable by the action of~$\Gamma_0$~on~$\mathcal X$~and this action over~$\mathcal U$~is conjugated to the action of a subgroup of~$\Diff^{an}_1 (\mathcal U)$. \end{prop} \begin{proof} Since~$r(\alpha) =x \in \overline \mathcal X (\mathbf{F}_p)$, the set~$\overline \mathcal X(\mathbf{F}_p)$~is not empty and since~$\overline \mathcal X$~has finitely many~$\mathbf{F}_p$-points, there exists a finite index subgroup~$\Gamma_1 \subset \Gamma$~that acts trivially on~$\overline \mathcal X (\mathbf{F}_p)$. The point~$x$~is fixed by~$\Gamma_1$, let~$\iota$~be as in Proposition \ref{PropExistenceOfIota} and~$\mathcal U$~the open subset of~$\mathcal X(\mathbf{Z}_p)$ consisting of the points~$\beta$~such that~$r(\beta) = x$. Therefore,~$\Gamma_1$~preserves~$\mathcal U$~and by applying Proposition \ref{PropConjugaisonDiffeoAnalytique} to the elements of~$\Gamma_1$, we get that conjugation by~$\iota$~induces a group homomorphism~$\Gamma_1 \hookrightarrow \Diff^{an}(\mathbf{Z}_p^d)$. Composing this embedding with the homomorphism of reduction$\mod p$~induces a group homomorphism from~$\Gamma_1$~to the finite group of affine transformations of~$(\mathbf{Z} / p \mathbf{Z})^d$. Denote by~$\Gamma_0$~the kernel of this homomorphism and the theorem is proven. \end{proof} \subsection{Proof of Theorem \ref{BoundNilpotentGroups}} Take~$H$~a finitely generated nilpotent group acting by algebraic automorphisms on a quasi-projective variety~$X$~over a field of characteristic zero. We are first going to show that we can suppose~$X$~to be irreducible in order to work on a~$\mathbf{Z}_p$-scheme:~$X$~has a finite number of irreducible components and~$H$~permutes them. So there exists a finite index subgroup~$H' \subset H$ that stabilizes every irreducible component~$X_i$~of~$X$. Call~$H_i$~the restriction of~$H'$~to~$X_i$, then~$H' = \prod H_i$~and~$\vdl (H') = \min \vdl (H_i)$. We replace~$X$~by one of its irreducible component of maximal dimension and~$H$~by~$H'$~restricted to this component,~$H'$~is also finitely generated by Proposition \ref{PropSubgroupIsFinitelyGenerated}. Let~$\alpha \in X(\mathbf{C})$,~$X$~is then an irreducible complex quasi-projective variety of dimension~$d$, by proposition \ref{FromCtoZp}, there exists a prime number~$p \geq 3$~such that~$(X,H)$~admits a good model~$\mathcal X$~over~$\mathbf{Z}_p$~and such that~$\alpha$~extends to a~$\mathbf{Z}_p$-point of~$\mathcal X$. Now, by Proposition \ref{FromAlgAutoToDIffAnal}, there exists a finite index subgroup~$H_0 \subset H$~which is isomorphic to a subgroup of~$\Diff_1^{an} (\mathcal U)$, for~$\mathcal U$~an open subset of~$\mathcal X(\mathbf{Z}_p)$~analytically diffeomorphic to~$\mathbf{Z}_p^d$. By Proposition \ref{PropSubgroupIsFinitelyGenerated},~$H_0$~is a finitely generated nilpotent subgroup of~$\Diff^{an}_1(\mathbf{Z}_p^d)$. Using Theorem \ref{MinorationpAdic2}, we get that the Lie algebra~$\mathfrak h$~associated to~${H_0}$~is nilpotent and~$\dl(\mathfrak h) \geq \vdl(H_0) \geq \vdl(H)$. Applying Theorem \ref{theoremPAdicEpsteinThurston}, we get~$d \geq \vdl(H)$. \subsection{Optimality of Theorem \ref{BoundNilpotentGroups}}\label{SubSecOptimality} \paragraph{An example from \cite{epstein1979transformation}.--} We will use the construction from \cite{epstein1979transformation} to find groups where Theorem \ref{BoundNilpotentGroups} is optimal. Let~$n$~be an integer and let~$A$~be the matrix such that~$A(e_i) = e_{i+1}, 1 <i \leq n$~where~$e_i$~is the canonical basis. Consider the subgroup of affine transformations~$G = \left\{ x \in \mathbf{R}^n \mapsto \exp(tA) x + b : t \in \mathbf{R}, b \in \mathbf{R}^n \right\}$, we will write~$(t;b)$~for the element~$(x \mapsto \exp (tA) x + b)$. This is a real Lie group of dimension~$n+1$~of nilpotency class~$n$~and derived length 2, diffeomorphic to ~$\mathbf{R}^{n+1}$. The group law is given by \[ (t;b) (s;c) = (t+s; b + e^{tA}c). \] Notice that the group law is given by polynomials with rational coefficients in~$s,t$~and the coordinates of~$b$~and ~$c$; thus~$G$~is in fact an algebraic group. \begin{lemme}\label{LemmeCrochetPolynomial} Recall the notation of \ref{SubSubSecNilpotentGroups}. Let~$k < n$~be an integer. The map \[ \left( (t_0;b_0), \cdots, (t_k; b_k) \right) \in G^{k+1} = \mathbf{R}^{(n+1)(k+1)} \mapsto \mathrm{Br}_{k+1} \left( (t_0; b_0), \cdots, (t_k; b_k) \right) \in G = \mathbf{R}^{n+1} \] is a nonconstant polynomial map with rational coefficients from~$\mathbf{R}^{(n+1)(k+1)}$~to~$~\mathbf{R}^{n+1}$. \end{lemme} \begin{proof} The map is polynomial with rational coefficients because the group law is, and this map is not constant because~$\nilp (H) = n > k$. \end{proof} Consider the vector space generated by the translations~$T_{e_i}, 2 \leq i \leq n$. The Lie group~$S$~acts on the variety~$G$~on the left and~$G / S$~is a variety diffeomorphic to~$\mathbf{R}^2$. The diffeomorphisms are given by \[ [(t;b)] \in G / S \mapsto (t, b_1) \in \mathbf{R}^2 \] and \[ (x,y) \in \mathbf{R}^2 \mapsto \left[ (x; y e_1) \right] \in G / S \] where the brackets mean that we take the orbit under the action of~$S$. The group~$G$~acts by right composition on~$G /S$~and this action is faithful. The formulas are given by \[ \forall (t;b) \in G, \forall (x,y) \in \mathbf{R}^2 = G / S , \quad (x,y) \cdot (t; b) = \left(x+ t, y + \sum_{k=1}^n \frac{t^{k-1}}{(k-1)!} b_k \right). \] We see that the action is therefore by polynomial automorphisms. We will write $(t;b)$ on the left even though the action is on the right because we view it as a polynomial automorphism of $\mathbf{A}^2_\mathbf{C}$. \paragraph{A group where theorem \ref{BoundNilpotentGroups} is optimal.--} Now, take~$H$~a finitely generated subgroup of~$G$~such that~$\nilp(H) = n$~and~$H$~contains two elements~$(t;b), (s;c)$~such that~$t,s$~and all the coordinates of~$b,c$~are algebraically independent over~$\mathbf{Q}$. The group~$H$~satisfies the condition of Theorem \ref{BoundNilpotentGroups}, it acts faithfully on the quasiprojective variety~$\mathbf{A}^2_\mathbf{C}$~and we have~$\vdl(H) =2$. Indeed, if~$H$~admits an abelian finite index subgroup, then there exists an integer ~$N$~such that~$(t;b)^N$~and~$(s;c)^N$~commute. But this would give a non-trivial polynomial relation over ~$\mathbf{Q}$~between~$s,t$~and the coordinates of~$b,c$~by Lemma \ref{LemmeCrochetPolynomial}, this is absurd. Thus, the bound in Theorem \ref{BoundNilpotentGroups} is optimal for~$H$. \paragraph{Derived length versus nilpotency class.--} In Theorem \ref{BoundNilpotentGroups} we suppose that~$H$~is nilpotent. One might wonder if the bound can be improved using the virtual nilpotency class, i.e the minimum of~$\nilp (H')$~for~$H'$~of finite index in~$H$. We show that this is not possible with a similar counterexample as above. Take~$H$~a finitely generated subgroup of~$G$ such that~$H$~contains~$(t_0; b_0), \cdots, (t_{n-1}; b_{n-1}) \in G^{n}$~such that all the~$t_i$'s and the coordinates of the $b_i$'s are algebraically independent over~$\mathbf{Q}$. We show that every finite index subgroup~$H'$~of~$H$~has a nilpotency class equal to~$n$. Indeed, there exists an integer~$N$~such that for all~$0 \leq i \leq n-1, h_i := (t_i; b_i)^N \in H'$. The coordinates of the~$h_i$'s are still algebraically independent over~$\mathbf{Q}$~because the group law is given by polynomials with rational coefficients and by Lemma \ref{LemmeCrochetPolynomial}, the bracket~$[h_0; \cdots; h_{n-1}]$~of length~$n$~is not the identity, because that would give a nontrivial polynomial relation between the coordinates of the~$h_i$'s. \paragraph{Optimality of Theorem \ref{theoremPAdicEpsteinThurston}.--} We show that in Theorem \ref{theoremPAdicEpsteinThurston} we can't replace the derived length with the nilpotency class and that the theorem is optimal. In fact, the counterexample of \cite{epstein1979transformation} can be adapted over~$\mathbf{Z}_p$~as follows. Consider the group~$G$~given by \[ G := \left\{ \mathbf x \in \mathbf{Z}_p^n \mapsto \exp (p \cdot t A) \mathbf x + b : t \in \mathbf{Z}_p, b \in \mathbf{Z}_p^n \right\}. \] The group law is now given by polynomials with coefficients in~$\mathbf{Z}_p$~and Lemma \ref{LemmeCrochetPolynomial} still holds but the polynomials are with coefficients in~$\mathbf{Z}_p$. Then,~$G /S$~is analytically diffeomorphic to~$\mathbf{Z}_p^2$~and we have an embedding of Lie groups~$G \hookrightarrow \Diff^{an} (\mathbf{Z}_p^2)$~given by \[ \forall (t;b) \in G, \quad (t;b) (x,y) = \left(x + t, y + \sum_{k=1}^n \frac{p^{k-1}t^{k-1}}{(k-1)!} b_k \right). \] Let~$\mathfrak g \subset \Theta (\mathbf{Z}_p^2)$~be the Lie algebra of~$G$,~$\mathfrak g$~is nilpotent and we show that~$\nilp (\mathfrak g) = n$. Let~$k = \nilp (\mathfrak g)$, then by Proposition \ref{PropOpenSubgroupDerivedSeries}, there exists a small subgroup~$G'$~of~$G$~which is a neighbourhood of $\id$ such that~$\nilp (G') = k$. Therefore~$k \leq n$, suppose~$k < n$. By Lemma \ref{LemmeCrochetPolynomial} the map \[ (t_0; b_0), \cdots, (t_k; b_k), (x,y) \in \mathbf{Z}_p^{(n+1)(k+1)} \times \mathbf{Z}_p^2 \mapsto \mathrm{Br}_{k+1} ( (t_0; b_0), \cdots, (t_k; b_k)) (x,y) \in \mathbf{Z}_p^2 \] is polynomial. Let~$P_1 (\mathbf w), P_2(\mathbf w)$~be the first and second coordinate of this map where~$\mathbf w$~is a multivariate variable representing all the variables~$t_i, b_i, x,y$. Since,~$\nilp (G) > k$, the polynomials~$Q_1(\mathbf w) = P_1 (\mathbf w)- x$, $Q_2 (\mathbf w)= P_2 (\mathbf w)- y$~are not zero. Notice that if $(t;b) \in G$, then the Gauss norm of~$(t;b) - \id \in \mathbf{Z}_p \langle x, y \rangle^2$~is bounded by the norm of~$(t;b) \in \mathbf{Z}_p^{n+1}$, therefore there exists an integer~$N >0$~such that for all~$(t;b) \in G, (p^N t; p^N b) \in G'$; thus \[ Q_1 (p^N \mathbf w) \equiv 0, \quad Q_2 (p^N \mathbf w) \equiv 0 \] and this implies that~$Q_1 = 0, Q_2 = 0$, this is a contradiction. By a similar argument, we can show there are no small abelian subgroups~$G' \subset G$ neighbourhood of the identity therefore~$\dl(\mathfrak g) = 2$~by Proposition \ref{PropOpenSubgroupDerivedSeries} and Theorem \ref{theoremPAdicEpsteinThurston} is also optimal. \paragraph*{Acknowledgements.--} I would like to thank my advisor Serge Cantat for his help. He gave me helpful advice whenever I needed them. I would also like to thank Junyi Xie for his suggestions. Finally, I would like to thank the reviewer for his/her very useful observations and detailed advice. \bibliographystyle{alpha}

https://arxiv.org/abs/math/0501410

The First Eigenvalue of the Dirac Operator on Compact Spin Symmetric Spaces

We give a formula for the first eigenvalue of the Dirac operator acting on spinor fields of a spin compact irreducible symmetric space $G/K$.

\section{Introduction} It is well-known that symmetric spaces provide examples where detailed information on the spectrum of Laplace or Dirac operators can be obtained. Indeed, for those manifolds, the computation of the spectrum can be (theoretically) done using group theoretical methods. However the explicit computation is far from being simple in general and only few examples are known. On the other hand, many results require some information about the first (nonzero) eigenvalue, so it seems interesting to get this eigenvalue without computing all the spectrum. In that direction, the aim of this paper is to prove the following formula for the first eigenvalue of the Dirac operator: \begin{theo} Let $G/K$ be a compact, simply-connected, $n$-dimensional irreducible symmetric space with $G$ compact and simply-connected, endowed with the metric induced by the Killing form of $G$ sign-changed. Assume that $G$ and $K$ have same rank and that $G/K$ has a spin structure. Let $\beta_{k}$, $k=1,\ldots, p$, be the $K$-dominant weights occurring in the decomposition into irreducible components of the spin representation under the action of $K$. Then the square of the first eigenvalue of the Dirac operator is \begin{equation}\label{sqvp} 2\,\min_{1\leq k\leq p}\|\beta_{k}\|^{2}+n/8\,, \end{equation} where $\|\cdot\|$ is the norm associated to the scalar product $<\,,\,>$ induced by the Killing form of $G$ sign-changed. \end{theo} \begin{remark} The proof uses a lemma of R. Parthasarathy in \cite{Par}, which allows to express (\ref{sqvp}) in the following way. Let $T$ be a fixed common maximal torus of $G$ and $K$. Let $\Phi$ be the set of non-zero roots of $G$ with respect to $T$. Let $\delta_{G}$, (resp. $\delta_{K}$) be the half-sum of the positive roots of $G$, (resp. $K$), with respect to a fixed lexicographic ordering in $\Phi$. Then the square of the first eigenvalue of the Dirac operator is given by \begin{equation} 2\,\|\delta_{G}\|^{2}+2\, \|\delta_{K}\|^{2}-4\, \max_{w\in W}<w\cdot\delta_{G},\delta_{K}>+n/8\,, \end{equation} where $W$ is a certain (well-defined) subset of the Weyl group of $G$. \end{remark} \section{The Dirac Operator on a Spin Compact Symmetric Space} We first review some results about the Dirac operator on a spin symmetric space, cf.~for instance \cite{CFG89} or \cite{Bar91}. A detailed survey on the subject may be found, among other topics, in the reference \cite{BHMM}. Let $G/K$ be a spin compact symmetric space. We assume that $G/K$ is simply connected, so $G$ may be chosen to be compact and simply connected and $K$ is the connected subgroup formed by the fixed elements of an involution $\sigma$ of $G$, cf. \cite{Hel}. This involution induces the Cartan decomposition of the Lie algebra $\mathfrak{G}$ of $G$ into $$\mathfrak{G}=\mathfrak{K}\oplus\mathfrak{P}\,,$$ where $\mathfrak{K}$ is the Lie algebra of $K$ and $\mathfrak{P}$ is the vector space $\{X\in\mathfrak{G}\,;\, \sigma_{*}\cdot X=-X\}$. This space $\mathfrak{P}$ is canonically identified with the tangent space to $G/K$ at the point $o$, $o$ being the class of the neutral element of $G$. We also assume that the symmetric space $G/K$ is irreducible, so all the $G$-invariant scalar products on $\mathfrak{P}$, hence all the $G$-invariant Riemannian metrics on $G/K$ are proportional. We consider the metric induced by the Killing form of $G$ sign-changed. With this metric, $G/K$ is an Einstein space with scalar curvature $\mathrm{Scal}=n/2$. The spin condition implies that the homomorphism $\alpha: K\rightarrow \mathrm{SO}(\mathfrak{P})\simeq \mathrm{SO}_{n}$, $k\mapsto \mathrm{Ad}_{G}(k)_{|\mathfrak{P}}$ lifts to a homomorphism $\widetilde{\alpha}:K\rightarrow \mathrm{Spin}_{n}$, cf.~\cite{CG}. Let $\rho:\mathrm{Spin}_{n}\rightarrow \mathrm{Hom}_{\mathbb{C}}(\Sigma,\Sigma)$ be the spin representation. The composition $\rho\circ\widetilde{\alpha}$ defines a ``{spin}'' representation of $K$ which is denoted $\rho_{K}$. The spinor bundle is then isomorphic to the vector bundle $$\boldsymbol{\Sigma}:=G\times_{\rho_{K}}\Sigma\,.$$ Spinor fields on $G/K$ are then viewed as $K$-equivariant functions $G\rightarrow \Sigma$, i.e. functions: $$\Psi:G\rightarrow\Sigma\quad \mathrm{s.t.}\quad \forall g\in G\;,\;\forall k\in K\;,\; \Psi(gk)=\rho_{K}(k^{-1})\cdot\Psi(g)\,.$$ Let $L_{K}^{2}(G,\Sigma)$ be the Hilbert space of $L^{2}$ $K$-equivariant functions $G\rightarrow \Sigma$. The Dirac operator $\mathcal{D}$ extends to a self-adjoint operator on $L_{K}^{2}(G,\Sigma)$. Since it is an elliptic operator, it has a (real) discrete spectrum. Now if the spinor field $\Psi$ is an eigenvector of $\mathcal{D}$ for the eigenvalue $\lambda$, then the spinor field $\sigma^{*}\cdot\Psi$ is an eigenvector for the eigenvalue $-\lambda$, hence the spectrum of the Dirac operator is symmetric with respect to the origin. Thus the spectrum of $\mathcal{D}$ may be deduced from the spectrum of its square $\mathcal{D}^{2}$. By the Peter-Weyl theorem, the natural unitary representation of $G$ on the Hilbert space $L_{K}^{2}(G,\Sigma)$ decomposes into the Hilbert sum $$ \mathop{\oplus}_{\gamma \in \widehat{G}}V_{\gamma}\otimes \mathrm{Hom}_{K}(V_{\gamma},\Sigma)\,,$$ where $\widehat{G}$ is the set of equivalence classes of irreducible unitary complex representations of $G$, $(\rho_{\gamma},V_{\gamma})$ represents an element $\gamma\in\widehat{G}$ and $\mathrm{Hom}_{K}(V_{\gamma},\Sigma)$ is the vector space of $K$-equivariant homomorphisms $V_{\gamma}\rightarrow \Sigma$, i.e. $$\mathrm{Hom}_{K}(V_{\gamma},\Sigma)=\{ A\in \mathrm{Hom}(V_{\gamma},\Sigma)\; \mathrm{s.t.}\; \forall k\in K\,, A\circ \rho_{\gamma}(k)=\rho_{K}(k)\circ A\}\,.$$ The injection $ V_{\gamma}\otimes \mathrm{Hom}_{K}(V_{\gamma},\Sigma)\hookrightarrow L_{K}^{2}(G,\Sigma)$ is given by $$v\otimes A \mapsto \Big(g\mapsto (A\circ\rho_{\gamma}(g^{-1})\,)\cdot v\Big)\,.$$ Note that $V_{\gamma}\otimes \mathrm{Hom}_{K}(V_{\gamma},\Sigma)$ consists of $\mathcal{C}^{\infty}$ spinor fields to which the Dirac operator can be applied. The restriction of $\mathcal{D}^{2}$ to the space $V_{\gamma}\otimes \mathrm{Hom}_{K}(V_{\gamma},\Sigma)$ is given by the Parthasaraty formula, \cite{Par}: \begin{equation}\label{Par} \mathcal{D}^{2}(v\otimes A)= v\otimes (A\circ \mathcal{C}_{\gamma})+\frac{\mathrm{Scal}}{8}\,v\otimes A\,, \end{equation} where $\mathcal{C}_{\gamma}$ is the Casimir operator of the representation $(\rho_{\gamma},V_{\gamma})$. Now since the representation is irreducible, the Casimir operator is a scalar multiple of identity, $\mathcal{C}_{\gamma}=c_{\gamma}\, \mathrm{id}$, where the eigenvalue $c_{\gamma}$ only depends of $\gamma\in\widehat{G}$. Hence if $\mathrm{Hom}_{K}(V_{\gamma},\Sigma)\neq\{0\}$, $c_{\gamma}+n/16$ belongs to the spectrum of $\mathcal{D}^{2}$. Let $\rho_{K}=\oplus \rho_{K,k}$ be the decomposition of the spin representation $K\rightarrow\Sigma$ into irreducible components. Denote by $\mathrm{m}({\rho_{\gamma}}_{|K},\rho_{K,k})$ the multiplicity of the irreducible $K$-representation $\rho_{K,k}$ in the representation $\rho_{\gamma}$ restricted to $K$. Then $$\dim\, \mathrm{Hom}_{K}(V_{\gamma},\Sigma)=\sum_{k} \mathrm{m}({\rho_{\gamma}}_{|K},\rho_{K,k})\,.$$ So the spectrum of the square of the Dirac operator is \begin{equation}\label{spec} \mathrm{Spec}(\mathcal{D}^{2})=\{c_{\gamma}+n/16\;;\; \gamma\in\widehat{G}\;\mathrm{s.t.}\;\exists k\;\mathrm{s.t.}\; \mathrm{m}({\rho_{\gamma}}_{|K},\rho_{K,k})\neq 0\}\,.\end{equation} \section{Proof of the result} We assume that $G$ and $K$ have same rank. Let $T$ be a fixed common maximal torus. Let $\Phi$ be the set of non-zero roots of the group $G$ with respect to $T$. According to a classical terminology, a root $\theta$ is called compact if the corresponding root space is contained in $\mathfrak{K}_{\mathbb{C}}$ (that is, $\theta$ is a root of $K$ with respect to $T$) and noncompact if the root space is contained in $\mathfrak{P}_{\mathbb{C}}$. Let $\Phi_{G}^{+}$ be the set of positive roots of $G$, $\Phi_{K}^{+}$ be the set of positive roots of $K$, and $\Phi_{n}^{+}$ be the set of positive noncompact roots with respect to a fixed lexicographic ordering in $\Phi$. The half-sums of the positive roots of $G$ and $K$ are respectively denoted $\delta_{G}$ and $\delta_{K}$ and the half-sum of noncompact positive roots is denoted by $\delta_{n}$. The Weyl group of $G$ is denoted $W_{G}$. The space of weights is endowed with the $W_{G}$-invariant scalar product $<\,,\,>$ induced by the Killing form of $G$ sign-changed.\\ Let \begin{equation}W:=\{w\in W_{G}\;;\; w\cdot \Phi_{G}^{+}\supset \Phi_{K}^{+}\}\,. \end{equation} By a result of R. Parthasaraty, cf.~lemma~2.2 in \cite{Par}, the spin representation $\rho_{K}$ of $K$ decomposes into the irreducible sum \begin{equation} \rho_{K}=\bigoplus_{w\in W}\rho_{K,w}\;, \end{equation} where $\rho_{K,w}$ has for dominant weight \begin{equation} \beta_{w}:= w\cdot \delta_{G}-\delta_{K}\,. \end{equation} Now define $w_{0}\in W$ such that \begin{equation}\label{w0} \|\beta_{w_{0}}\|^{2}=\min_{w\in W}\|\beta_{w}\|^{2}\,, \end{equation} and \begin{equation}\label{w1} \mbox{if there exists a } w_{1}\neq w_{0}\in W \mbox{ such that } \|\beta_{w_{1}}\|^{2}=\min_{w\in W}\|\beta_{w}\|^{2}\, ,\mbox { then } \beta_{w_{1}}\prec \beta_{w_{0}}\,, \end{equation} where $\prec$ is the usual ordering on weights. \begin{lem} The weight $$\beta_{w_{0}}^{G}:=w_{0}^{-1}\cdot\beta_{w_{0}}=\delta_{G}-w_{0}^{-1}\cdot\delta_{K}\,,$$ is $G$-dominant. \end{lem} \begin{proof} Let $\Pi_{G}=\{\theta_{1},\ldots,\theta_{r}\}\subset \Phi_{G}^{+}$ be the set of simple roots. It is sufficient to prove that $2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}$ is a non-negative integer for any simple root $\theta_{i}$. Since $T$ is a maximal common torus of $G$ and $K$, $\beta_{w_{0}}$, which is an integral weight for $K$ is also an integral weight for $G$. Now since the Weyl group $W_{G}$ permutes the weights, $\beta_{w_{0}}^{G}=w_{0}^{-1}\cdot\beta_{w_{0}}$ is also a integral weight for $G$, hence $2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}$ is an integer for any simple root $\theta_{i}$. So we only have to prove that this integer is non-negative.\\ Let $\theta_{i}$ be a simple root. Since $2\,\frac{<\delta_{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}=1$, (see for instance \S~10.2 in \cite{Hum}) and since the scalar product $<\cdot ,\cdot>$ is $W_{G}$-invariant, one gets \begin{equation} 2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}=1- 2\,\frac{<\delta_{K},w_{0}\cdot\theta_{i}>}{<\theta_{i},\theta_{i}>}\,. \end{equation} Suppose first that $w_{0}\cdot\theta_{i}\in \Phi_{K}$. If $w_{0}\cdot\theta_{i}$ is positive then $w_{0}\cdot\theta_{i}$ is necessarily a $K$-simple root. Indeed let $\Pi_{K}=\{\theta_{1}',\ldots,\theta_{l}'\}\subset \Phi_{K}^{+}$ be the set of $K$-simple roots. One has $w_{0}\cdot\theta_{i}=\sum_{j=1}^{l} b_{ij}\, \theta_{j}'$, where the $b_{ij}$ are non-negative integers. But since $w_{0}\in W$, there are $l$ positive roots $\alpha_{1},\ldots,\alpha_{l}$ in $\Phi_{G}^{+}$ such that $w_{0}\cdot \alpha_{j}=\theta_{j}'$, $j=1,\ldots, l$. So $ \theta_{i}=\sum_{j=1}^{l} b_{ij}\,\alpha_{j}$. Now each $\alpha_{j}$ is a sum of simple roots $\sum_{k=1}^{r}a_{jk}\,\theta_{k}$, where the $a_{jk}$ are non-negative integers. So $\theta_{i}=\sum_{j,k} b_{ij}\, a_{jk}\,\theta_{k}$. By the linear independence of simple roots, one gets $\sum_{j} b_{ij}\, a_{jk}=0$ if $k\neq i$, and $\sum_{j} b_{ij}\, a_{ji}=1$. Hence there exists a $j_{0}$ such that $b_{ij_{0}}=a_{j_{0}i}=1$, the other coefficients being zero. So $w_{0}\cdot\theta_{i}=\theta_{j_{0}}'$ is a $K$-simple root. Now since $2\,\frac{<\delta_{K},w_{0}\cdot\theta_{i}>}{<\theta_{i},\theta_{i}>}= 2\,\frac{<\delta_{K},w_{0}\cdot\theta_{i}>} {<w_{0}\cdot\theta_{i},w_{0}\cdot\theta_{i}>}=1$, one gets $2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}=0$, hence $2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}\geq 0$. Now, the same conclusion holds if $w_{0}\cdot\theta_{i}$ is a negative root of $K$, since $2\,\frac{<\delta_{K},w_{0}\cdot\theta_{i}>}{<\theta_{i},\theta_{i}>}=-2\, \frac{<\delta_{K},-w_{0}\cdot\theta_{i}>}{<\theta_{i},\theta_{i}>}=-1$, hence $2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}=2$.\\ Suppose now that $w_{0}\cdot\theta_{i}\notin \Phi_{K}$, that is $w_{0}\cdot\theta_{i}$ is a noncompact root. This implies that $w_{0}\sigma_{i}$, where $\sigma_{i}$ is the reflection across the hyperplane $\theta_{i}^{\bot}$, is an element of $W$. Let $ \alpha_{1},\ldots, \alpha_{m}$ be the positive roots in $\Phi_{G}^{+}$ such that $w_{0}\cdot \alpha_{j}=\alpha'_{j}$, where the $\alpha'_{j}$, $j=1,\ldots, m$ are the positive roots of $K$. Since $\sigma_{i}$ permutes the positive roots other than $\theta_{i}$, (cf. for instance Lemma~B, \S~10.2 in \cite{Hum}), and since $\theta_{i}$ can not be one of the roots $\alpha_{1},\ldots,\alpha_{m}$ (otherwise $w_{0}\cdot\theta_{i}\in \Phi_{K}^{+}$), each root $\sigma_{i}\cdot \alpha_{j}$ is positive. So $w_{0}\sigma_{i}\in W$ since $w_{0}\sigma_{i}\cdot(\sigma_{i}\cdot \alpha_{j})=\alpha_{j}'$, $j=1,\ldots,m$.\\ We now claim that $2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}<0$, which is equivalent to $2\,\frac{<\delta_{K},w_{0}\cdot\theta_{i}>}{<\theta_{i},\theta_{i}>}>1$, is impossible.\\ Suppose that \begin{equation} 2\,\frac{<\delta_{K},w_{0}\cdot\theta_{i}>}{<\theta_{i},\theta_{i}>}>1\,. \end{equation} Since $\delta_{K}$ can be expressed as $\delta_{K}=\sum_{i=1}^{l} c_{i}\,\theta_{i}'$, where the $c_{i}$ are nonnegative, there exists a $K$-simple root $\theta_{j}'$ such that $<\theta_{j}',w_{0}\cdot\theta_{i}>0$, and since $2\,\frac{<\theta_{j}',w_{0}\cdot\theta_{i}>}{<\theta_{j}',\theta_{j}'>}$ is an integer, this implies that \begin{equation}\label{cond1} 2\,\frac{<\theta_{j}',w_{0}\cdot\theta_{i}>}{<\theta_{j}',\theta_{j}'>}\geq 1\,. \end{equation} So $\theta_{j}'-w_{0}\cdot\theta_{i}$ is a root (cf. for instance \S~9.4 in \cite{Hum}). Moreover, from the bracket relation $[\mathfrak{K},\mathfrak{P}]\subset\mathfrak{P}$, it is a noncompact root. Now $\pm (\theta_{j}'-w_{0}\cdot\theta_{i})$ is a positive noncompact root, so by the description of the weights of the spin representation $\rho_{K}$, (they are of the form: $\delta_{n}-$(a sum of distinct positive noncompact roots), cf. \S2~in \cite{Par}), $$(w_{0}\cdot\delta_{G}-\delta_{K})\pm (\theta_{j}'-w_{0}\cdot\theta_{i}) \mbox{ is a weight of } \rho_{K}\,.$$ Now, $(w_{0}\cdot\delta_{G}-\delta_{K})+ (\theta_{j}'-w_{0}\cdot\theta_{i})$ can not be a weight of $\rho_{K}$. Otherwise since $\sigma_{i}\cdot\delta_{G}=\delta_{G}-\theta_{i}$, $(w_{0}\sigma_{i}\cdot\delta_{G}-\delta_{K})+ \theta_{j}'$ is a weight of $\rho_{K}$. But since $w_{0}\sigma_{i}\in W$, $\mu:=w_{0}\sigma_{i}\cdot\delta_{G}-\delta_{K}$ is a dominant weight of $\rho_{K}$. So $\mu$ is a dominant weight but not the highest weight of an irreducible component of $\rho_{K}$. Hence there exists an irreducible representation of $\rho_{K}$ with dominant weight $\lambda=w\cdot\delta_{G}-\delta_{K}$, $w\in W$, whose set of weights $\Pi$ contains $\mu$. Furthermore $\mu\prec\lambda$. Now since $\mu \in \Pi$$, \|\mu+\delta_{K}\|^{2}\leq \|\lambda+\delta_{K}\|^{2}$, with equality only if $\mu=\lambda$, (cf. for instance Lemma~C, \S 13.4 in \cite{Hum}). But $\|\mu+\delta_{K}\|^{2}=\|\delta_{G}\|^{2}= \|\lambda+\delta_{K}\|^{2}$, so $\mu=\lambda$, contradicting the fact that $\mu\prec\lambda$. \\ Thus only \begin{equation}\mu_{0}:=(w_{0}\cdot\delta_{G}-\delta_{K})- (\theta_{j}'-w_{0}\cdot\theta_{i})\,, \end{equation} can be a weight of $\rho_{K}$. Now one has $$ \begin{array}{rl} \|\mu_{0}\|^{2}= &\|w_{0}\cdot\delta_{G}-\delta_{K}+ w_{0}\cdot\theta_{i}\|^{2}\\ &-2\,<w_{0}\cdot\delta_{G}-\delta_{K}+w_{0}\cdot\theta_{i},\theta_{j}'> +\|\theta_{j}'\|^{2}\,. \end{array}$$ Since $w_{0}\cdot\delta_{G}-\delta_{K}$ is a dominant weight, $<w_{0}\cdot\delta_{G}-\delta_{K},\theta_{j}'>\geq 0$, and from (\ref{cond1}), $2\,<w_{0}\cdot\theta_{i},\theta_{j}'> -\|\theta_{j}'\|^{2}\geq 0$, so $$ \|\mu_{0}\|^{2}\leq \|(w_{0}\cdot\delta_{G}-\delta_{K})+w_{0}\cdot\theta_{i}\|^{2}\,.$$ Now $$ \begin{array}{rl} \|(w_{0}\cdot\delta_{G}-\delta_{K})+w_{0}\cdot\theta_{i}\|^{2}= &\|w_{0}\cdot\delta_{G}-\delta_{K}\|^{2}\\ &+2\,<\delta_{G}-w_{0}^{-1}\cdot\delta_{K},\theta_{i}> +\|\theta_{i}\|^{2}\,. \end{array}$$ But, as we supposed $2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}<0$, one has $\frac{2\,<\delta_{G}-w_{0}^{-1}\cdot\delta_{K},\theta_{i}>}{\|\theta_{i}\|^{2}}\leq -1$, so\\ $2\,<\delta_{G}-w_{0}^{-1}\cdot\delta_{K},\theta_{i}> +\|\theta_{i}\|^{2}\leq 0$, hence $$\|(w_{0}\cdot\delta_{G}-\delta_{K})+w_{0}\cdot\theta_{i}\|^{2}\leq \|w_{0}\cdot\delta_{G}-\delta_{K}\|^{2}\,,$$ so $$\|\mu_{0}\|^{2}\leq \|w_{0}\cdot\delta_{G}-\delta_{K}\|^{2}\,.$$ Now, being a weight of $\rho_{K}$, $\mu_{0}$ is conjugate under the Weyl group of $K$ to a dominant weight of $\rho_{K}$, say $w_{1}\cdot\delta_{G}-\delta_{K}$, with $w_{1}\in W$. Note that $w_{1}\neq w_{0}$, otherwise since $\mu_{0}\prec w_{1}\cdot\delta_{G}-\delta_{K}$, (cf. Lemma~A, \S~13.2 in \cite{Hum}), the noncompact root $\theta_{j}'-w_{0}\cdot\theta_{i}$ should be a linear combination with integral coefficients of compact simple roots. But, by the bracket relation $[\mathfrak{K},\mathfrak{K}]\subset\mathfrak{K}$, that is impossible. Thus, by the definition of $w_{0}$, cf. (\ref{w0}), $\|w_{0}\cdot\delta_{G}-\delta_{K}\|^{2}\leq \|w_{1}\cdot\delta_{G}-\delta_{K}\|^{2}=\|\mu_{0}\|^{2}$, so $$ \|\mu_{0}\|^{2}=\|w_{1}\cdot\delta_{G}-\delta_{K}\|^{2} =\|w_{0}\cdot\delta_{G}-\delta_{K}\|^{2}\,.$$ But by the condition~(\ref{w1}), the last equality is impossible, otherwise since $\mu_{0}\prec w_{1}\cdot\delta_{G}-\delta_{K}$ and $w_{1}\cdot\delta_{G}-\delta_{K}\prec w_{0}\cdot\delta_{G}-\delta_{K}$, the noncompact root $\theta_{j}'-w_{0}\cdot\theta_{i}$ should be a linear combination with integral coefficients of compact simple roots. Hence $2\,\frac{<\beta_{w_{0}}^{G},\theta_{i}>}{<\theta_{i},\theta_{i}>}\geq 0$ also if $w_{0}\cdot\theta_{i}\notin \Phi_{K}$. \end{proof} Now let $(\rho_{0},V_{0})$ be an irreducible representation of $G$ with dominant weight $\beta_{w_{0}}^{G}$. The fact that $\beta_{w_{0}}=w_{0}\cdot \beta_{w_{0}}^{G}$ is a weight of $\rho_{0}$ is an indication that ${\rho_{0}}_{|K}$ may contain the irreducible representation $\rho_{K,w_{0}}$. This is actually true: \begin{lem} With the notations above, $$\mathrm{m}({\rho_{0}}_{|K},\rho_{K,w_{0}})\geq 1\,.$$ \end{lem} \begin{proof} Let $v_{0}$ be the maximal vector in $V_{0}$, (it is unique up to a nonzero scalar multiple). Let $g_{0}\in T$ be a representative of $w_{0}$. Then $g_{0}\cdot v_{0}$ is a weight vector for the weight $\beta_{w_{0}}$, since for any $X$ in the Lie algebra $\mathfrak{T}$ of $T$: $$ \begin{array}{rll} X\cdot (g_{0}\cdot v_{0}) &=\frac{d}{dt}\Big((\exp(tX)\, g_{0})\cdot v_{0}\Big)_{|t=0} & =\frac{d}{dt}\Big(\Big(g_{0}\,g_{0}^{-1}\exp(tX)\, g_{0}\Big)\cdot v_{0}\Big)_{|t=0}\\ &= g_{0}\cdot\Big( \Big(\mathrm{Ad}(g_{0}^{-1})\cdot X\Big)\cdot v_{0}\Big)&=\beta_{w_{0}}^{G}(w_{0}^{-1}\cdot X)\,(g_{0}\cdot v_{0})\\ &=(w_{0}\cdot \beta_{w_{0}}^{G})(X)\,(g_{0}\cdot v_{0})&=\beta_{w_{0}}(X)\, (g_{0}\cdot v_{0})\,. \end{array}$$ In order to prove the result, we only have to prove that $g_{0}\cdot v_{0}$ is a maximal vector (for the action $K$), hence is killed by root-vectors corresponding to simple roots of $K$. So let $\theta_{i}'$ be a simple root of $K$ and $E_{i}'$ be a root-vector corresponding to that simple root. Since $w_{0}\in W$, there exists a positive root $\alpha_{i}\in \Phi_{G}^{+}$ such that $w_{0}\cdot\alpha_{i}=\theta_{i}'$. Then $ E_{i}:= \mathrm{Ad}(g_{0}^{-1})(E_{i}')$ is a root-vector corresponding to the root $\alpha_{i}$ since for any $X$ in $\mathfrak{T}$ $$ \begin{array}{rll} [X, E_{i}]&= [X,\mathrm{Ad}(g_{0}^{-1})(E_{i}')] & =\mathrm{Ad}(g_{0}^{-1})\cdot [\mathrm{Ad}(g_{0})(X),E_{i}'] \\ &=\mathrm{Ad}(g_{0}^{-1})\cdot[w_{0}\cdot X,E_{i}'] &=\Big((w_{0}^{-1}\cdot\theta_{i}')(X)\Big)\;\mathrm{Ad}(g_{0}^{-1})\cdot E_{i}'\\ &=\alpha_{i}(X)\; E_{i}\,.& \end{array}$$ But since $v_{0}$ is killed by the action of the root-vectors corresponding to positive roots in $\Phi_{G}^{+}$, one gets $$ \begin{array}{rll} E_{i}'\cdot(g_{0}\cdot v_{0}) & =\frac{d}{dt}\Big(\Big(g_{0}g_{0}^{-1}\exp(t\,E_{i}') g_{0}\Big)\cdot v_{0}\Big)_{|t=0} \\&=\frac{d}{dt}\Big(\Big(g_{0}\exp\Big(t\,\mathrm{Ad}(g_{0}^{-1})\cdot E_{i}'\Big) \Big)\cdot v_{0}\Big)_{|t=0}\\&= g_{0}\cdot\Big( E_{i}\cdot v_{0}\Big)\\ &=0\,. \end{array}$$ Hence the result. \end{proof} From the result~(\ref{spec}), we may then conclude: \begin{lem} $$2\,\|\beta_{w_{0}}\|^{2}+n/8\,,$$ is an eigenvalue of the square of the Dirac operator. \end{lem} \begin{proof} By the Freudenthal's formula, the Casimir eigenvalue $c_{\gamma_{0}}$ of the representation $(\rho_{0},V_{0})$ is given by $$\|\beta_{w_{0}}^{G}+\delta_{G}\|^{2}-\|\delta_{G}\|^{2}=3\,\|\delta_{G}\|^{2} +\|\delta_{K}\|^{2}-4\,<w_{0}\cdot\delta_{G},\delta_{K}>\,.$$ On the other hand $$\|\beta_{w_{0}}\|^{2}=\|\delta_{G}\|^{2} +\|\delta_{K}\|^{2}-2\,<w_{0}\cdot\delta_{G},\delta_{K}>\,.$$ Hence $$c_{\gamma_{0}}=2\,\|\beta_{w_{0}}\|^{2}+\|\delta_{G}\|^{2}-\|\delta_{K}\|^{2}\,.$$ Now, the Casimir operator of $\mathfrak{K}$ acts on the spin representation $\rho_{K}$ as scalar multiplication by $ \|\delta_{G}\|^{2}-\|\delta_{K}\|^{2}$, (cf.~lemma~2.2 in \cite{Par}). Indeed, each dominant weight of $\rho_{K}$ being of the form $w\cdot\delta_{G}-\delta_{K}$, $w\in W$, the eigenvalue of the Casimir operator on each irreducible component is given by: $$\|(w\cdot\delta_{G}-\delta_{K})+\delta_{K}\|^{2}-\|\delta_{K}\|^{2}= \|w\cdot\delta_{G}\|^{2}-\|\delta_{K}\|^{2}=\|\delta_{G}\|^{2}-\|\delta_{K}\|^{2}\,.$$ On the other hand, the proof of the formula (\ref{Par}) shows that the Casimir operator of $\mathfrak{K}$ acts on the spin representation $\rho_{K}$ as scalar multiplication by $\frac{\mathrm{Scal}}{8}=n/16$ (cf. \cite{Sul}), hence \begin{equation}\label{sf} \|\delta_{G}\|^{2}-\|\delta_{K}\|^{2}=n/16\,. \end{equation} So $$c_{\gamma_{0}}+n/16=2\,\|\beta_{w_{0}}\|^{2}+n/8\,.$$ \end{proof} In order to conclude, we have to prove that \begin{lem} $$2\,\|\beta_{w_{0}}\|^{2}+n/8\,,$$ is the lowest eigenvalue of the square of the Dirac operator. \end{lem} \begin{proof} Let $\gamma\in\widehat{G}$ such that there exists $w\in W$ such that $\mathrm{m}({\rho_{\gamma}}_{|K},\rho_{K,w})\geq 1$. Let $\beta_{\gamma}$ be the dominant weight of ${\rho_{\gamma}}$. First, since the Weyl group permutes the weights of ${\rho_{\gamma}}$, $w^{-1}\cdot\beta_{w}=\delta_{G}-w^{-1}\cdot\delta_{K}$ is a weight of ${\rho_{\gamma}}$. Hence $$\|\beta_{\gamma}+\delta_{G}\|^{2}\geq \|w^{-1}\cdot\beta_{w}+\delta_{G}\|^{2}\,,$$ (cf. for instance Lemma~C, \S 13.4 in \cite{Hum}). So, from the Freudenthal formula, $$c_{\gamma}=\|\beta_{\gamma}+\delta_{G}\|^{2}-\|\delta_{G}\|^{2}\geq \|w^{-1}\cdot\beta_{w}+\delta_{G}\|^{2}-\|\delta_{G}\|^{2}\,.$$ But, using (\ref{sf}) $$\|w^{-1}\cdot\beta_{w}+\delta_{G}\|^{2}-\|\delta_{G}\|^{2}= 2\, \|\beta_{w}\|^{2}+\|\delta_{G}\|^{2}-\|\delta_{K}\|^{2}=2\, \|\beta_{w}\|^2+n/16\,.$$ Hence by the definition of $\beta_{w_{0}}$, $$c_{\gamma}\geq 2\, \|\beta_{w}\|^2+n/16\geq 2\,\|\beta_{w_{0}}\|^2+n/16\,.$$ Hence the result. \end{proof}

https://arxiv.org/abs/2004.08323

A polynomial-degree-robust a posteriori error estimator for Nédélec discretizations of magnetostatic problems

We present an equilibration-based a posteriori error estimator for Nédélec element discretizations of the magnetostatic problem. The estimator is obtained by adding a gradient correction to the estimator for Nédélec elements of arbitrary degree presented in [J. Gedicke, S. Geevers, and I. Perugia. An equilibrated a posteriori error estimator for arbitrary-order Nédélec elements for magnetostatic problems. Journal of Scientific Computing, 83:1-23, 2020]. This new estimator is proven to be reliable, with reliability constant 1, and efficient, with an efficiency constant that is independent of the polynomial degree of the approximation. These properties are demonstrated in a series of numerical experiments on three-dimensional test problems.

\section{Introduction} Magnetostatic equations of the form $\nabla\times(\mu^{-1}\nabla\times\vct{u}) = \vct{j}$ are often approximated using N\'ed\'elec elements. To control the error of the N\'ed\'elec finite element approximation, a wide variety of \textit{a posteriori} error estimators are available, including residual-type error estimators \cite{monk98,beck00}, hierarchical error estimators \cite{beck99}, Zienkiewicz--Zhu-type error estimators \cite{nicaise05}, equilibration-based error estimators \cite{braess08,tang13,creuse17,creuse19b}, and functional estimates \cite{neittaanmaki10}. Of particular interest are the localised equilibration-based error estimators, since \emph{(i)} they provide an explicit upper bound on the error without any unknown constant involved \cite{braess08}, \emph{(ii)} they are efficient with an efficiency constant that is typically independent of the polynomial degree \cite{braess09}, and \emph{(iii)} they only require solving small local problems. For an overview of these type of error estimators, see, for example, \cite{ern15} and the references therein. The first localised equilibration-based error estimator for the magnetostatic problem was introduced in \cite{braess08}. That estimator was designed for N\'ed\'elec element approximations of lowest order only, and requires the solution of local problems on vertex patches. In \cite{GGP_arXiv}, an alternative localised equilibration-based error estimator was presented that is applicable to N\'ed\'elec element approximations of arbitrary degree. That method requires the solution of local problems on single elements, on single faces, and on small sets of nodes. While it was proven in \cite{GGP_arXiv} that the estimator satisfies bounds of the form \begin{equation*} C_{\text{eff}}\cdot\text{estimator} \leq \text{error} \leq C_{\text{rel}}\cdot\text{estimator} \end{equation*} up to some higher-order data oscillation terms, with reliability constant $C_{\text{rel}}=1$ and efficiency constant $C_{\text{eff}}>0$ independent of the mesh size, numerical experiments showed that the efficiency constant $C_{\text{eff}}$ still mildly depends on the polynomial degree. In this paper, a new error estimator is constructed by adding a gradient correction to the estimator of~\cite{GGP_arXiv}, resulting in an efficiency index that is now also independent of the polynomial degree. The proof of reliability (Theorem~\ref{thm:reliability}) is a slight modification of the corresponding one developed in~\cite{GGP_arXiv}, whereas the proof of efficiency with a constant independent of the polynomial degree (Theorem~\ref{thm:efficiency}) is significantly more involved. Unlike in~\cite{GGP_arXiv}, we can no longer rely on the efficiency of the residual error estimator, since this error estimator is not polynomial-degree robust. Instead, the efficiency proof is based on a new decomposition of the error and relies on the stability property of the regularized Poincar\'e integral operator proven in~\cite{costabel10}, and on the stable broken $H^1$ polynomial extensions presented in \cite{ern19}. The new estimator and its analysis are presented for the case of piecewise constant magnetic permeability. The extension to the case of piecewise smooth magnetic permeability is discussed in Remark~\ref{rem:variablemu}. The outline of this paper is as follows. In Section \ref{sec:intro}, the considered model problem and its N\'ed\'elec finite element discretization is presented. In Section \ref{sec:errEst}, the new error estimator is introduced, and the main theorems on reliability and efficiency are stated. In Section \ref{sec:analysis}, the efficiency of the estimator is proven. Numerical examples are presented in Section \ref{sec:numerics}, and the main results are summarised in Section~\ref{sec:conclusion}. \section{Model problem and notation}\label{sec:intro} In this section, we define the same model problem and notation as in~\cite{GGP_arXiv}. We consider the linear magnetostatic problem in the unknown magnetic field $\vct{H}$: \begin{align*} \nabla\times\vct{H} &= \vct{j} &&\text{in }\Omega, \\ \nabla\cdot\mu\vct{H} &= 0 &&\text{in }\Omega, \\ \hat{\vct{n}}\cdot\mu\vct{H} &= 0 &&\text{on }\partial\Omega, \end{align*} where $\Omega\subset\mathbb{R}^3$ is an open, bounded, simply connected, polyhedral domain with a connected Lipschitz boundary $\partial\Omega$ with outward pointing unit normal vector $\hat{\vct{n}}$, $\mu$ is a scalar magnetic permeability, $\vct{j}$ a given divergence-free current density, and $\nabla$, $\nabla\times$ and $\nabla\cdot$ denote the gradient, the curl, and the divergence operator, respectively. We assume that $\mu:\Omega\rightarrow\mathbb{R}^+$, and $\mu_0\leq\mu\leq\mu_1$, for some positive constants $\mu_0$ and $\mu_1$. In terms of a vector potential $\vct{u}$ such that $\vct{H}=\mu^{-1}\nabla\times\vct{u}$, the problem can be rewritten as the following system: \begin{subequations} \label{eq:curlcurl} \begin{align} \nabla\times(\mu^{-1}\nabla\times \vct{u}) &= \vct{j} &&\text{in }\Omega, \\ \nabla\cdot\vct{u} &= 0 &&\text{in }\Omega, \\ \hat{\vct{n}}\times\vct{u} &= \vct{0} &&\text{on }\partial\Omega, \end{align} \end{subequations} where the uniqueness of $\vct{u}$ is imposed by the second equation (Coulomb's gauge). Let $D\in\mathbb{R}^3$ be any given domain. We denote by $L^2(D)^m$ the standard space of square-integrable functions $\vct{u}:D\rightarrow\mathbb{R}^m$ endowed with norm $\|\vct{u}\|_D^2:=\int_D \vct{u}\cdot\vct{u} \;\mathrm{d}{\vct{x}}$ and inner product $(\vct{u},\vct{w})_D=\int_D \vct{u}\cdot\vct{w}\;\mathrm{d}{\vct{x}}$. We also define the following functional spaces: \begin{align*} H^1(D) &:= \{\phi\in L^2(D) \;|\; \nabla \phi\in L^2(D)^3 \}, \\ {H_{0}^1(D)} &:= \{\phi\in H^1(D) \;|\; \phi=0 \text{ on }\partial\Omega \}, \\ H(\mathrm{curl}; D) &:= \{\vct{u}\in L^2(D)^3 \;|\; \nabla\times\vct{u}\in L^2(\Omega)^3 \}, \\ {H_{0}(\mathrm{curl}; D)} &:= \{\vct{u}\in H(\mathrm{curl};D) \;|\; \hat{\vct{n}}\times\vct{u}=\vct{0} \text{ on }\partial\Omega \}, \\ H(\mathrm{div}; D) &:= \{\vct{u}\in L^2(D)^3 \;|\; \nabla\cdot\vct{u}\in L^2(\Omega) \}, \\ H(\mathrm{div}^0; D) &:= \{\vct{u}\in L^2(D)^3 \;|\; \nabla\cdot\vct{u}=0 \},\\ H_{0,\Gamma}^1(D) &:= \{\phi\in H^1(D) \;|\; \phi=0 \text{ on }\Gamma \}, \end{align*} where, in the last definition, $\Gamma$ is any two-dimensional manifold $\Gamma\subset\partial D$. If $\mathcal{T}_D$ is any tessellation of $D$, we set \begin{align*} H^1(\mathcal{T}_D) &:= \{\phi\in L^2(D) \;|\; \phi|_T\in H^1(T) \text{ for all }T\in\mathcal{T}_D\}, \\ H(\mathrm{curl};\mathcal{T}_D) &:= \{\vct{u}\in L^2(D) \;|\; \vct{u}|_T\in H(\mathrm{curl};T) \text{ for all }T\in\mathcal{T}_D\}, \\ H(\mathrm{div};\mathcal{T}_D) &:= \{\vct{u}\in L^2(D) \;|\; \vct{u}|_T\in H(\mathrm{div};T) \text{ for all }T\in\mathcal{T}_D\}. \end{align*} The variational formulation of problem \eqref{eq:curlcurl} reads as follows: Find $\vct{u}\in H_0(\mathrm{curl};\Omega)\cap H(\mathrm{div}^0;\Omega)$ such that \begin{align} \label{eq:WF} (\mu^{-1}\nabla\times \vct{u},\nabla\times\vct{w})_{\Omega} &= (\vct{j},\vct{w})_{\Omega} &&\forall \vct{w}\in H_0(\mathrm{curl};\Omega). \end{align} Before introducing a finite element approximation of~\eqref{eq:WF}, we define the following polynomial spaces. For any $D\subset\mathbb{R}^3$, let $P_k(D)$ denote the space of polynomials of degree $k$ or less. Moreover, for any tetrahedron $T$, let $R_k(T)$ and $D_k(T)$ denote the first-kind N\'ed\'elec space and the Raviart-Thomas space, respectively: \begin{align*} R_k(T) &:= \{\vct{u}\in P_k(T)^3 \;|\; \vct{u}(\vct{x}) = \vct{v}(\vct{x}) + \vct{x}\times\vct{w}(\vct{x}) \text{ for some }\vct{v},\vct{w}\in P_{k-1}(T)^3\}, \\ D_k(T) &:= \{\vct{u}\in P_k(T)^3 \;|\; \vct{u}(\vct{x}) = \vct{v}(\vct{x}) + \vct{x} w(\vct{x}) \text{ for some }\vct{v}\in P_{k-1}(T)^3, w\in P_{k-1}(T)\}. \end{align*} Furthermore, for any domain $D\subset\mathbb{R}^3$ with a tessellation $\mathcal{T}_D$, we define the \emph{discontinuous} spaces \begin{align*} P_k^{-1}(\mathcal{T}_D) &:= \{\phi\in L^2(D) \;|\; \phi|_T \in P_k(T) \text{ for all }T\in\mathcal{T}_D \}, \\ R_k^{-1}(\mathcal{T}_D) &:= \{\vct{u}\in L^2(D)^3 \;|\; \vct{u}|_T \in R_k(T) \text{ for all }T\in\mathcal{T}_D \}, \\ D_k^{-1}(\mathcal{T}_D) &:= \{\vct{u}\in L^2(D)^3 \;|\; \vct{u}|_T \in D_k(T) \text{ for all }T\in\mathcal{T}_D \}, \end{align*} and the \emph{conforming} spaces \begin{align*} P_k(\mathcal{T}_D)&:=P^{-1}_k(\mathcal{T}_h)\cap H^1(D), & {P_{k,0}(\mathcal{T}_D)} &:=P^{-1}_k(\mathcal{T}_D)\cap H^1_{0}(D),\\ R_k(\mathcal{T}_D)&:=R^{-1}_k(\mathcal{T}_h)\cap H(\mathrm{curl};D), & {R_{k,0}(\mathcal{T}_D)} &:=R^{-1}_k(\mathcal{T}_D)\cap H_{0}(\mathrm{curl};D),\\ D_k(\mathcal{T}_D)&:=D^{-1}_k(\mathcal{T}_h)\cap H(\mathrm{div};D). & \end{align*}% For any two-dimensional manifold $\Gamma_D\subset \partial D$, also define \begin{align*} P_{k,0,\Gamma_D}(\mathcal{T}_D) &:=P^{-1}_k(\mathcal{T}_D)\cap H^1_{0,\Gamma_D}(C). \end{align*} We consider the following finite element approximation of~\eqref{eq:WF} on a tetrahedral mesh $\mathcal{T}_h$ of $\Omega$ of granularity $h$: Find $\vct{u}_h\in R_{k,0}(\mathcal{T}_h)$ such that \begin{subequations} \label{eq:FEM} \begin{align} (\mu^{-1}\nabla\times \vct{u}_h,\nabla\times\vct{w})_{\Omega} &= (\vct{j},\vct{w})_{\Omega} &&\forall \vct{w}\in R_{k,0}(\mathcal{T}_h), \label{eq:FEMa}\\ (\vct{u}_h,\nabla\psi)_{\Omega} &=0 &&\forall \psi\in P_{k,0}(\mathcal{T}_h). \label{eq:FEMb} \end{align} \end{subequations} The approximated of the magnetic field is then defined by \begin{align*} \vct{H}_h:=\mu^{-1}\nabla\times\vct{u}_h. \end{align*} For the well-posedness of the continuous problem~\eqref{eq:WF} and the $h$-convergence of the finite element method~\eqref{eq:FEM}, see, e.g., \cite[Theorems~5.9 and~5.10]{hiptmair02}. \section{A polynomial-degree-robust {a posteriori} error estimator}\label{sec:errEst} As in~\cite{braess08,GGP_arXiv}, the equilibrated \emph{a posteriori} error estimator we are going to introduce is based on the following result (\cite[Theorem 10]{braess08}, \cite[Corollary 3.3]{GGP_arXiv}): \begin{thm} \label{thm:estimator} Let $\vct{u}$ be the solution to (\ref{eq:WF}), let $\vct{u}_h$ be the solution of (\ref{eq:FEM}), set $\vct{H}:=\mu^{-1}\nabla\times\vct{u}$ and $\vct{H}_h:=\mu^{-1}\nabla\times\vct{u}_h$, and let $\vct{j}_h:=\nabla\times\vct{H}_h$ be the discrete current distribution. If $\vtH^{\Delta}\in L^2(\Omega)^3$ satisfies the (residual) equilibrium condition \begin{align} \label{eq:vtHdel} \nabla\times \vtH^{\Delta} &= \vct{j}-\vct{j}_h \end{align} in a distributional sense, then \begin{align} \label{eq:errEst2} \|\mu^{1/2}(\vct{H}-\vct{H}_h)\|_{\Omega} &\leq \|\mu^{1/2}\vtH^{\Delta}\|_{\Omega}. \end{align} \end{thm} To construct a field $\vtH^{\Delta}$ that satisfies \eqref{eq:vtHdel}, we use polynomial function spaces of a fixed degree $k'\geq k$ and make the following two assumptions: \begin{itemize} \item[A1.] The magnetic permeability $\mu$ is piecewise constant and the mesh is assumed to be chosen in such a way that $\mu$ is constant within each element. \item[A2.] The current density $\vct{j}$ is in $D_{k'}(\mathcal{T}_h)\cap H(\mathrm{div}^0;\Omega)$. \end{itemize}\medskip The case of a piecewise smooth instead of a piecewise constant magnetic permeability is discussed in Remark \ref{rem:variablemu} below. \begin{rem} In case assumption A2 is not satisfied, $\vct{j}$ can be replaced by a suitable projection $\pi_h\vct{j}$ such as, for instance, the standard Raviart-Thomas interpolate in $D_{k'}(\mathcal{T}_h)$. As observed in~\cite[Section~3.1]{GGP_arXiv}, the error then satisfies $\|\mu^{1/2}(\vct{H}-\vct{H}_h)\|_{\Omega}\le \|\mu^{1/2}\vtH^{\Delta}\|_{\Omega} + \|\mu^{1/2}(\vct{H}-\vct{H}^{\prime})\|_{\Omega}$, with $\vct{H}^{\prime}$ the solution to~\eqref{eq:WF} with $\pi_h\vct{j}$ instead of $\vct{j}$. As proven in~\cite[Appendix~A]{GGP_arXiv}, whenever $\vct{j}$ admits a compactly supported extension $\vct{j}^*\in H(\mathrm{div};\mathbb{R}^3)\cap H^{k'}(\mathbb{R}^3)^3$, the term $\|\mu^{1/2}(\vct{H}-\vct{H}') \|_{\Omega}$ is of order $h^{k'+1}$ and therefore of higher order than $\| \mu^{1/2}(\vct{H}-\vct{H}_h) \|_{\Omega}$. \end{rem} For the construction of a field $\vtH^{\Delta}$ satisfying \eqref{eq:vtHdel}, we proceed as in \cite{GGP_arXiv}, but perform one additional step (Step 4). \emph{Step 1.} We compute $\hat{\tilde{\vH}}^{\Delta}\in R^{-1}_{k'}(\mathcal{T}_h)$ from the datum $\vct{j}$ and the numerical solution~$\vct{H}_h$ by solving \begin{subequations} \label{eq:vthHdel} \begin{align} \nabla\times\hat{\tilde{\vH}}^{\Delta} |_T &= \vct{j}^{\Delta}_T := \vct{j}|_T-\nabla\times\vct{H}_h|_T, && \\ (\hat{\tilde{\vH}}^{\Delta},\nabla\psi)_T &= 0 &&\forall \psi\in P_{k'}(T) \label{eq:vthHdel1b} \end{align} \end{subequations} for each $T\in\mathcal{T}_h$. \emph{Step 2.} For each internal face $f\in\mathcal{F}_{h}^{I}$, let $T^+$ and $T^-$ denote the two adjacent elements, let $\hat{\vct{n}}^{\pm}$ denote the normal unit vector pointing outward of $T^{\pm}$, let $\vct{H}^{\pm}:=\vct{H}|_{T^{\pm}}$ denote the vector field restricted to $T^{\pm}$, let $\jut{\vct{H}}|_f := (\hat{\vct{n}}^+\times\vct{H}^+ + \hat{\vct{n}}^-\times\vct{H}^-)|_{f}$ denote the tangential jump operator, and let $\nabla_f$ denote the gradient operator restricted to the face $f$. We set $\hat{\vct{n}}_f:=\hat{\vct{n}}^+|_f$ and compute $\tilde{\lambda}_f\in P_{k'}(f)$ by solving \begin{subequations} \label{eq:tlambda} \begin{align} -\hat{\vct{n}}_f\times\nabla_f\tilde{\lambda}_f &= \hat{\tilde{\vjj}}^{\Delta}_f := \jut{\vct{H}_h + \hat{\tilde{\vH}}^{\Delta}}|_f , \\ (\tilde{\lambda}_f,1)_f &=0 \label{eq:tlambda1b} \end{align} \end{subequations} for each internal face $f\in\mathcal{F}_{h}^{I}$. \emph{Step 3.} Let $\mathcal{Q}_h$ denote the set of standard Lagrangian nodes corresponding to the finite element space $P_{k'}(\mathcal{T}_h)$. We compute $\tilde{\phi}\in P^{-1}_{k'}(\mathcal{T}_h)$ by solving, for each $\vct{x}\in\mathcal{Q}_h$, the small set of degrees of freedom $\{\tilde{\phi}_{T,\vct{x}}\}_{T:\overline{T}\ni\vct{x}}$ such that \begin{subequations} \label{eq:tphi} \begin{align} \tilde{\phi}_{T^+,\vct{x}} - \tilde{\phi}_{T^-,\vct{x}} &= \tilde{\lambda}_{f}(\vct{x}) &&\forall f\in\mathcal{F}_{h}^{I}:\partial f\ni\vct{x}, \label{eq:tphi1a} \\ \sum_{T:\overline{T}\ni\vct{x}} \tilde{\phi}_{T,\vct{x}} &= 0, && \end{align} \end{subequations} where $\tilde{\phi}_{T,\vct{x}}$ denotes the value of $\tilde{\phi}|_T$ at node $\vct{x}$. \emph{Step 4.} Let $\mathcal{V}_h$ denote the set of all mesh vertices and, for each $\nu\in\mathcal{V}_h$, let $\mathcal{T}_\nu$ denote the element patch consisting of all elements adjacent to $\nu$, set $\overline{\omega_\nu}:=\bigcup_{T\in\mathcal{T}_\nu} \overline{T}$, and set $\Gamma_\nu :=\partial\omega_\nu$ whenever $\nu$ is an interior vertex and $\Gamma_\nu:=\partial\omega_\nu\setminus\partial\Omega$ whenever $\nu$ is a vertex on the boundary $\partial\Omega$. For each vertex $\nu\in\mathcal{V}_h$, we compute a continuous scalar field $\tilde{\alpha}_\nu\in P_{k'+1,0,\Gamma_\nu}(\mathcal{T}_\nu)$ such that \begin{align} \label{eq:talpha_nu} (\mu\nabla\tilde{\alpha}_\nu,\nabla\psi)_{\omega_\nu} &= (\mu\nabla_h(\theta_\nu\tilde{\phi}),\nabla\psi)_{\omega_\nu} &&\forall \psi\in P_{k'+1,0,\Gamma_\nu}(\mathcal{T}_\nu), \end{align} where $\nabla_h$ denotes the element-wise gradient operator and $\theta_\nu$ denotes the hat function corresponding to vertex $\nu$. We then extend $\tilde{\alpha}_\nu$ by zero to the rest of the domain $\Omega$ and set $\tilde{\alpha} := \sum_{\nu\in\mathcal{V}_h} \tilde{\alpha}_\nu$. \emph{Step 5.} We compute the field \begin{align*} \vtH^{\Delta} = \hat{\tilde{\vH}}^{\Delta} + \nabla_h\tilde{\phi} - \nabla\tilde{\alpha}, \end{align*} and compute the error estimator \begin{align} \label{eq:eta} \eta_h :=\|\mu^{1/2}\vtH^{\Delta}\|_{\Omega} = \left(\sum_{T\in \mathcal{T}_h}\eta_T^2\right)^{1/2}, &\quad\text{where}\quad \eta_T :=\|\mu^{1/2}\vtH^{\Delta}\|_T. \end{align} The resulting error estimator is reliable and provides an explicit upper bound on the error, i.e. the upper bound does not involve any unknown constants. \begin{thm}[reliability]\label{thm:reliability} Let $\vct{u}$ be the solution to (\ref{eq:WF}), let $\vct{u}_h$ be the solution to (\ref{eq:FEM}), and set $\vct{H}:=\mu^{-1}\nabla\times\vct{u}$ and $\vct{H}_h:=\mu^{-1}\nabla\times\vct{u}_h$. Also, fix $k'\geq k$ and assume that assumptions A1 and A2 hold true. Then the problems in Steps~1--4 are all well-defined and have a unique solution. Furthermore, if $\vtH^{\Delta}$ is computed by following Steps~1--5, then \begin{align*} \|\mu^{1/2}(\vct{H}-\vct{H}_h)\|_{\Omega} &\leq \|\mu^{1/2}\vtH^{\Delta} \|_{\Omega}=\eta_h. \end{align*} \end{thm} \begin{proof} In \cite{GGP_arXiv}, it was shown that Steps~1--3 are well-defined and result in a field $\hat{\tilde{\vH}}^{\Delta}+\nabla_h\tilde{\phi}$ that satisfies \begin{align*} \nabla\times(\hat{\tilde{\vH}}^{\Delta}+\nabla_h\tilde{\phi}) &= \vct{j}-\vct{j}_h \end{align*} in a distributional sense. Step 4 is also well-defined, since $\tilde{\alpha}_\nu$ is the unique discrete finite element approximation for the elliptic problem $-\nabla\cdot\mu\nabla\alpha_\nu=\nabla\cdot\mu\nabla_h(\theta_\nu\tilde{\phi})$ in $\omega_\nu$ and $\alpha_\nu|_{\Gamma_\nu}=0$. Since $\nabla\times\nabla\tilde{\alpha}=\vct{0}$, we have that \begin{align*} \nabla\times\vtH^{\Delta} = \nabla\times(\hat{\tilde{\vH}}^{\Delta}+\nabla_h\tilde{\phi}-\nabla\tilde{\alpha})=\nabla\times(\hat{\tilde{\vH}}^{\Delta}+\nabla_h\tilde{\phi}) &= \vct{j}-\vct{j}_h. \end{align*} The theorem then follows from Theorem \ref{thm:estimator}. \end{proof} The following theorem is the main result of this paper. It states that the error estimator is efficient and that the efficiency index is bounded by a constant that is independent of the polynomial degree. \begin{thm}[local efficiency] \label{thm:efficiency} Let $\vct{u}$ be the solution to (\ref{eq:WF}), let $\vct{u}_h$ be the solution to (\ref{eq:FEM}), and set $\vct{H}:=\mu^{-1}\nabla\times\vct{u}$ and $\vct{H}_h:=\mu^{-1}\nabla\times\vct{u}_h$. Also, fix $k'\geq k$ and assume that assumptions A1 and A2 hold true. If $\vtH^{\Delta}$ is computed by following Steps~1--5, then \begin{align} \label{eq:efficiency} \eta_T &= \|\mu^{1/2}\vtH^{\Delta}\|_T \leq C\sum_{T':\overline{T'}\cap\overline{T}\neq\emptyset} \|\mu^{1/2}(\vct{H}-\vct{H}_h)\|_{T'} \end{align} for all $T\in\mathcal{T}_h$, where $C$ is some positive constant that depends on the magnetic permeability $\mu$ and the shape-regularity of the mesh, but not on the mesh width $h$ or the polynomial degree $k'$. \end{thm} The proof of Theorem \ref{thm:efficiency} is given in the next section. \begin{rem} As observed in \cite[Remark~3.4]{GGP_arXiv}, for the case $k'=1$, this algorithm requires solving local problems with 6 unknowns per element in Step~1, 3 unknowns per face in Step~2, and $(\# T\in\mathcal{T}_\nu)\approx 24$ unknowns per vertex $\nu$ in Step~3. The problem in the additional Step~4 involves $1+(\# e:\overline{e}\ni\nu) \approx 15$ unknowns per vertex when $k'=1$. \end{rem} \begin{rem}\label{rem:variablemu} \newcommand{\Pi_h^{k'-1}}{\Pi_h^{k'-1}} Assume that $\mu$ is \emph{piecewise smooth}, and that the mesh is chosen in such a way that $\mu$ is smooth within each element. The definition of the error estimator $\eta_h$ can be extended to this case as follows. Define $\vct{H}_h^\ast:=\Pi_h^{k'-1}\vct{H}_h = \Pi_h^{k'-1} (\mu^{-1}\nabla\times\vct{u}_h)$, where $\Pi_h^{k'-1}$ is the weighted $ L^2(\Omega)^3$ projection onto $P_{k'-1}^{-1}(\mathcal{T}_h)^3$ such that $(\mu\Pi_h^{k'-1} \vct{H}_h,\vct{w})_\Omega = (\mu\vct{H}_h,\vct{w})_{\Omega}$ for all $\vct{w}\in P_{k'-1}^{-1}(\mathcal{T}_h)^3$. Set $\vct{j}_h:=\nabla\times \vct{H}_h$ and $\vct{j}_h^*:=\nabla\times\vct{H}_h^*$. Then, compute $\vtH^{\Delta}$ such that $\nabla\times\vtH^{\Delta} = \vct{j}-\vct{j}_h^*$ by following Steps 1-5 with $\vct{H}_h$ replaced by $\vct{H}_h^*$. One can prove, in a way analogous to \cite[Section 3.2]{GGP_arXiv} and the proof of Theorem \ref{thm:reliability}, that the problems in Steps 1-4 with $\vct{H}_h$ replaced by $\vct{H}_h^*$ are well-posed, and thus $\vtH^{\Delta}$ is well-defined. Then, the new local and global error estimators are defined as \begin{align*} \eta_T &:= \left(\|\mu^{1/2}\vtH^{\Delta} \|_{T}^2 +\| \mu^{1/2}(\vct{H}_h-\vct{H}_h^\ast)\|_{T}^2\right)^{1/2},\\ \eta_h &:= \left(\|\mu^{1/2}\vtH^{\Delta} \|_{\Omega}^2 +\| \mu^{1/2}(\vct{H}_h-\vct{H}_h^\ast)\|_{\Omega}^2\right)^{1/2} = \left(\sum_{T\in\mathcal{T}_h} \eta_T^2\right)^{1/2}. \end{align*} Clearly, for piecewise constant $\mu$, the new estimators coincide with the old ones. The reliability bound $\|\vct{H}-\vct{H}_h\|_{\Omega}\le \eta_h$ follows from Theorem~\ref{thm:estimator} and the fact that $\vtH^{\Delta}+\vct{H}_h^*-\vct{H}_h$ satisfies the residual equilibrium condition \begin{align*} \nabla\times(\vtH^{\Delta}+\vct{H}_h^*-\vct{H}_h) = \vct{j}-\vct{j}_h. \end{align*} For the local efficiency bound, one can check that Theorem~\ref{thm:efficiency} still holds true when replacing $\vct{H}_h$ by $\vct{H}_h^*$. We then only need to prove efficiency of the additional term~$\|\mu^{1/2}(\vct{H}_h-\vct{H}_h^*)\|_T$ for each $T\in\mathcal{T}_h$. We have \[ \begin{split} \|\mu^{1/2}(\vct{H}_h-\vct{H}_h^*)\|_{T} &=\|\mu^{1/2}({\mathrm I}-\Pi_h^{k'-1})\vct{H}_h \|_{T} \\ &\le \|\mu^{1/2}({\mathrm I}-\Pi_h^{k'-1})(\vct{H}-\vct{H}_h)\|_{T} + \|\mu^{1/2}({\mathrm I}-\Pi_h^{k'-1})\vct{H}\|_{T}\\ &\le \| \mu^{1/2}(\vct{H}-\vct{H}_h)\|_{T} + \| \mu^{1/2}({\mathrm I}-\Pi_h^{k'-1}) \vct{H} \|_{T}, \end{split} \] for all $T\in\mathcal{T}_h$, where ${\mathrm I}$ denotes the identity operator and where the last inequality follows from the $L^2$ stability of the weighted $L^2$ projection. The behaviour of the second term on the right-hand side depends on the smoothness of $\vct{H}$ and on the mesh grading towards possible solution singularities. Therefore, it behaves similarly to the actual error $\|\mu^{1/2}(\vct{H}-\vct{H}_h)\|_T$. \end{rem} \section{Proof of Theorem \ref{thm:efficiency}}\label{sec:analysis} In this section, we let $\vct{u}$, $\vct{u}_h$, $\vct{H}$, $\vct{H}_h$, and $\vtH^{\Delta}$ be the fields as defined in Theorem \ref{thm:efficiency} and let $\hat{\tilde{\vH}}^{\Delta}$, $\tilde{\phi}$, $\tilde{\alpha}_\nu$, and $\tilde{\alpha}$ as described in Steps~1--5 of Section \ref{sec:errEst}. We will also always let $C$ denote some positive constant that may depend on the magnetic permeability $\mu$ and the shape regularity of the mesh, but not on the mesh width $h$ or the polynomial degree $k'$. In Section \ref{sec:errDec}, we introduce a vector field $\hat{\vH}^{\Delta}\in H(\mathrm{curl};\mathcal{T}_h)$ and scalar fields $\phi\in H^1(\mathcal{T}_h)$ and $\alpha\in H^1(\Omega)$, and show that the error $\vct{H}^{\Delta}:=\vct{H}-\vct{H}_h$ can be written as $\vct{H}^{\Delta}=\hat{\vH}^{\Delta} + \nabla_h\phi - \nabla\alpha$. We also show there that \begin{subequations} \label{eq:vHdelBound1}% \begin{align} \| \mu^{1/2} \hat{\vH}^{\Delta} \|_T \leq \| \mu^{1/2}\vct{H}^{\Delta}\|_T &&\forall T\in\mathcal{T}_h \text{ (Section \ref{sec:errDec})}, \\ \| \mu^{1/2} \nabla(\phi-\alpha) \|_T \leq \| \mu^{1/2}\vct{H}^{\Delta}\|_T &&\forall T\in\mathcal{T}_h \text{ (Section \ref{sec:errDec})}. \end{align} \end{subequations} In Sections \ref{sec:vthHdelBound} and \ref{sec:tphiBound} we then prove that \begin{subequations} \label{eq:vtHdelBound2} \begin{align} \|\mu^{1/2} \hat{\tilde{\vH}}^{\Delta}\|_T &\leq C\|\mu^{1/2} \vct{H}^{\Delta}\|_T &&\forall T\in\mathcal{T}_h \text{ (Section \ref{sec:vthHdelBound})}, \label{eq:vthHdelBound} \\ \| \mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi}-\tilde{\alpha}_\nu)\|_{\omega_\nu} &\leq C\| \mu^{1/2} \vct{H}^{\Delta} \|_{\omega_\nu} &&\forall \nu\in\mathcal{V}_h\text{ (Section \ref{sec:tphiBound})} \label{eq:tphiBound}. \end{align} \end{subequations} Since \begin{align*} \vtH^{\Delta}|_T &= \hat{\tilde{\vH}}^{\Delta}|_T + \sum_{\nu:\nu\subset\partial_T} \nabla(\theta_\nu\tilde{\phi}-\tilde{\alpha}_\nu)|_T, \end{align*} we can use the triangle inequality and \eqref{eq:vtHdelBound2} to obtain \begin{align*} \|\mu^{1/2}\vtH^{\Delta}\|_T &\leq \|\mu^{1/2} \hat{\tilde{\vH}}^{\Delta} \|_T + \sum_{\nu:\nu\subset\partial_T} \|\mu^{1/2} \nabla(\theta_\nu\tilde{\phi}-\tilde{\alpha}_\nu)\|_T \\ &\leq \|\mu^{1/2} \hat{\tilde{\vH}}^{\Delta} \|_T + \sum_{\nu:\nu\subset\partial_T} \|\mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi}-\tilde{\alpha}_\nu)\|_{\omega_\nu} \\ &\leq C\sum_{T':\overline{T'}\cap\overline{T}\neq\emptyset} \|\mu^{1/2}\vct{H}^{\Delta}\|_{T'}, \end{align*} which completes the proof of Theorem \ref{thm:efficiency}. It thus remains to prove \eqref{eq:vHdelBound1} and \eqref{eq:vtHdelBound2}. \subsection{Decomposition of the error and proof of~\eqref{eq:vHdelBound1}} \label{sec:errDec} Define $\hat{\vH}^{\Delta}\in H(\mathrm{curl};\mathcal{T}_h)$ as the unique solution of \begin{subequations} \label{eq:vhHdel} \begin{align} \nabla\times\hat{\vH}^{\Delta} |_T &= \vct{j}^{\Delta}_T = \vct{j}|_T-\nabla\times\vct{H}_h|_T, && \label{eq:vhHdel1a}\\ (\mu\hat{\vH}^{\Delta},\nabla\psi)_T &= 0 &&\forall \psi\in H^1(T), \label{eq:vhHdel1b} \end{align} \end{subequations} Since $\nabla\times(\hat{\vH}^{\Delta}-\hat{\tilde{\vH}}^{\Delta})|_T=\vct{j}^{\Delta}_T-\vct{j}^{\Delta}_T = 0$ for each $T\in\mathcal{T}_h$, we can define $\phi^\Delta\in H^1(\mathcal{T}_h)$ such that \begin{subequations} \label{eq:phidel} \begin{align} -\nabla_h\phi^\Delta &= \hat{\vH}^{\Delta}-\hat{\tilde{\vH}}^{\Delta}, &&\\ (\phi^\Delta,1)_T &= 0 &&\forall T\in\mathcal{T}_h. \end{align} \end{subequations} Now, set $\phi:=\tilde{\phi}+\phi^\Delta\in H^1(\mathcal{T}_h)^3$. We can then write $\vtH^{\Delta}=\hat{\vH}^{\Delta}+\nabla_h\phi-\nabla\tilde{\alpha}$. Finally, since $\nabla\times(\vct{H}^{\Delta}-\vtH^{\Delta})=\vct{j}^{\Delta}-\vct{j}^{\Delta}=0$, we can define $\alpha^\Delta\in H^1(\Omega)$ such that \begin{subequations} \label{eq:alpha} \begin{align} -\nabla\alpha^\Delta &= \vct{H}^{\Delta}-\vtH^{\Delta}, \\ (\alpha,1)_{\Omega} &= 0. \end{align} \end{subequations} If we now set $\alpha=\tilde{\alpha}+\alpha^\Delta\in H^1(\Omega)$, we obtain the following decomposition of the error: \begin{align*} \vct{H}^{\Delta} = \hat{\vH}^{\Delta} + \nabla_h\phi - \nabla\alpha. \end{align*} Note that, because of (\ref{eq:vhHdel1b}), we have that $(\mu\hat{\vH}^{\Delta},\nabla(\phi-\alpha))_T=0$ for all $T\in\mathcal{T}_h$. From Pythagoras' theorem, it then follows that $\|\mu^{1/2}\vct{H}^{\Delta}\|^2_T=\|\mu^{1/2}\hat{\vH}^{\Delta}\|_T^2+\|\mu^{1/2}\nabla(\phi-\alpha)\|_T^2$, which proves the bounds in \eqref{eq:vHdelBound1}. \subsection{Upper bound on $\|\mu^{1/2} \hat{\tilde{\vH}}^{\Delta}\|_T$ in terms of $\|\mu^{1/2} \vct{H}^{\Delta}\|_T$ (proof of~\eqref{eq:vthHdelBound})} \label{sec:vthHdelBound} Firstly, observe that \begin{align} \label{eq:vthHdelMin} \|\hat{\tilde{\vH}}^{\Delta}\|_T &= \inf_{{\vct{H}' \in R_{k'}(T), \nabla\times\vct{H}'=\vct{j}^{\Delta}_T}} \| \vct{H}'\|_T, \end{align} for each $T\in\mathcal{T}_h$. Indeed, let $\vct{H}' \in R_{k'}(T)$ with $\nabla\times\vct{H}'=\vct{j}^{\Delta}_T$. Then $\nabla\times(\hat{\tilde{\vH}}^{\Delta}|_T-\vct{H}')=\vct{j}_T-\vct{j}_T=0$ and so we can write $\vct{H}'-\hat{\tilde{\vH}}^{\Delta}|_T =\nabla\phi$ for some $\phi\in P_{k'}(T)$. From \eqref{eq:vthHdel1b}, it then follows that $(\hat{\tilde{\vH}}^{\Delta}, \vct{H}'-\hat{\tilde{\vH}}^{\Delta})_T=(\hat{\tilde{\vH}}^{\Delta},\nabla\phi)_T=0$ and from Pythagoras' theorem, it then follows that $\|\vct{H}'\|^2_T = \|\hat{\tilde{\vH}}^{\Delta} \|^2_T + \|\vct{H}'-\hat{\tilde{\vH}}^{\Delta}\|_T^2\geq \|\hat{\tilde{\vH}}^{\Delta}\|_T^2$. In an analogous way, we can show that \begin{align} \label{eq:vhHdelMin} \|\mu^{1/2}\hat{\vH}^{\Delta}\|_T &= \inf_{{\vct{H}' \in H(\mathrm{curl};T), \nabla\times\vct{H}'=\vct{j}^{\Delta}_T}} \|\mu^{1/2}\vct{H}'\|_T. \end{align} We also need the following result, which follows from the stability of the regularised Poincar\'e integral operator that was proven in \cite{costabel10}. \begin{lem} \label{lem:invCurlBound} Let $T$ be a tetrahedron. For any $\vct{r}\inD_{k'}(T)\cap H(\mathrm{div}^0;T)$, there exists a $\vct{G}\inR_{k'}(T)$ such that $\nabla\times\vct{G}=\vct{r}$ and \begin{align*} \|\vct{G}\|_T &\leq C\inf_{\vct{G}'\in H(\mathrm{curl};T), \nabla\times\vct{G}'=\vct{r}} \|\vct{G}'\|_T. \end{align*} \end{lem} \begin{proof} Let $\hat{T}$ denote the reference tetrahedron. We will construct an operator $\hat{\vct{R}}:H(\mathrm{div};\hat{T})\rightarrow H(\mathrm{curl};\hat T)$, independent of $k'$, such that \begin{enumerate} \item[C1.] $\nabla\times\hat{\vct{R}}\hat{\vct{r}}=\hat{\vct{r}}$ whenever $\hat{\vct{r}}\in H(\mathrm{div}^0,\hat T)$. \item[C2.] $\hat{\vct{R}}\hat{\vct{r}}\inR_{k'}(\hat T)$ whenever $\hat{\vct{r}}\inD_{k'}\cap H(\mathrm{div}^0;\hat{T})$. \end{enumerate} To construct such an operator, define $\hat{\vct{R}}_{\hat{\vct{z}}}:\mathcal{C}^{\infty}(\overline{\hat{T}})\rightarrow H(\mathrm{curl};\hat T)$, for any $\hat{\vct{z}}\in\hat{T}$, as the following Poincar\'e integral operator: \begin{align*} \hat{\vct{R}}_{\hat{\vct{z}}}\hat{\vct{r}}(\hat{\vct{x}}) &:= -(\hat{\vct{x}}-\hat{\vct{z}})\times \int_0^1 \tau\hat{\vct{r}}(\tau(\hat{\vct{x}}-\hat{\vct{z}})+\hat{\vct{z}})\;\mathrm{d}\tau. \end{align*} The operator $\hat{\vct{R}}_{\hat{\vct{z}}}$ can be extended to $H(\mathrm{div};\hat T)$ and satisfies conditions C1 and C2 \cite[Theorem 2.1 and Proposition 3.1]{gopalakrishnan04}. Now, let $B$ be an open ball in $\hat{T}$ and let $\vartheta\in\mathcal{C}_0^\infty(\hat T)$ be an analytic function with support on $B$ such that $\int_{\hat T} \vartheta(\hat{\vct{x}})\;\mathrm{d}{\hat{\vct{x}}} = 1$. We then define $\hat{\vct{R}}:\mathcal{C}^{\infty}(\overline{\hat{T}})\rightarrow H(\mathrm{curl};\hat T)$ as the following regularised Poincar\'e integral operator: \begin{align*} \hat{\vct{R}}\hat{\vct{r}}(\hat{\vct{x}}):= \int_{\hat T} \vartheta(\hat{\vct{z}})\hat{\vct{R}}_{\hat{\vct{z}}}\hat{\vct{r}}(\hat{\vct{x}}) \;\mathrm{d}{\hat{\vct{z}}}. \end{align*} Since $\hat{\vct{R}}_{\hat{\vct{z}}}$, for every $\hat{\vct{z}}\in\hat T$, can be extended to $H(\mathrm{div};\hat T)$ and satisfies conditions C1 and C2, so does $\hat{\vct{R}}$. By applying the coordinate transformations $\hat{\vct{y}}=\tau(\hat{\vct{x}}-\hat{\vct{z}})+\hat{\vct{z}}$ and $t=(1-\tau)^{-1}$, the above can be rewritten as \begin{align*} \hat{\vct{R}}\hat{\vct{r}}(\hat{\vct{x}}) = \int_{\hat T} -(\hat{\vct{x}}-\hat{\vct{y}})\times\hat{\vct{r}}(\hat{\vct{y}}) \left(\int_{1}^\infty t(t-1)\vartheta(\hat{\vct{x}}+t(\hat{\vct{y}}-\hat{\vct{x}})) \;\mathrm{d}t\right) \;\mathrm{d}{\hat{\vct{y}}}, \end{align*} where $\vartheta$ and $\hat{\vct{r}}$ are extended by zero to $\mathbb{R}^3$. This is exactly the operator $R_2$ of \cite[Definition 3.1]{costabel10}. By taking $s=-1$ in \cite[Corollary 3.4]{costabel10}, it follows that \begin{align} \label{eq:invCurlBound1} \| \hat{\vct{R}}\hat{\vct{r}} \|_{\hat T} &\leq C \|\hat{\vct{r}}\|_{H^{-1}(\hat T)} &&\forall \hat{\vct{r}}\in H(\mathrm{div}^0,\hat T), \end{align} where we stress once more that the operator $\hat{\vct{R}}$ and the constant $C$ are independent of $k'$. Now, let $\hat{\vG}'\in H(\mathrm{curl};\hat T)$ and $\hat{\vct{r}}\in H(\mathrm{div}^0,\hat T)$ be two functions such that $\nabla\times\hat{\vG}'=\hat{\vct{r}}$. Then \begin{align*} \|\hat{\vct{r}}\|_{H^{-1}(\hat T)^3} &= \sup_{\vct{w}\in H_0^1(\hat T)^3\setminus\{0\}} \frac{(\hat{\vct{r}},\vct{w})_{\hat T}}{\|\vct{w}\|_{H^1(\hat T)^3}} \\ &= \sup_{\vct{w}\in H_0^1(\hat T)^3 \setminus\{0\}} \frac{(\nabla\times\hat{\vG}',\vct{w})_{\hat T}}{\|\vct{w}\|_{H^1(\hat T)^3}} \\ &= \sup_{\vct{w}\in H_0^1(\hat T)^3 \setminus\{0\}} \frac{(\hat{\vG}',\nabla\times\vct{w})_{\hat T}}{\|\vct{w}\|_{H^1(\hat T)^3}} \\ &\leq \sup_{\vct{w}\in H_0^1(\hat T)^3 \setminus\{0\}} \frac{\|\hat{\vG}'\|_{\hat T} \|\nabla\times\vct{w}\|_{\hat T}}{\|\vct{w}\|_{H^1(\hat T)^3}} \\ &\leq \sqrt{2} \|\hat{\vG}'\|_{\hat T} \end{align*} where $\|\vct{w}\|_{H^1(\hat T)^3}^2:= \|\vct{w}\|_{\hat T}^2+\|\nabla\vct{w}\|_{\hat T}^2$ and where the fourth line follows from the Cauchy--Schwarz inequality and the last line from the fact that $\|\nabla\times\vct{w}\|_{\hat T}\leq \sqrt{2} \|\vct{w}\|_{H^1(\hat T)^3}$. From \eqref{eq:invCurlBound1}, it then follows that \begin{align} \label{eq:invCurlBound2} \|\hat{\vct{R}}\hat{\vct{r}}\|_{\hat T} &\leq C \inf_{\hat{\vG}'\in H(\mathrm{curl};\hat T), \nabla\times\hat{\vG}'=\hat{\vct{r}}} \|\hat{\vG}'\|_{\hat T}. \end{align} Now, let $\boldsymbol{\varphi}_T:\hat{T}\rightarrow T$ denote the affine element mapping and let $J_T := [\frac{\partial \boldsymbol{\varphi}_T}{\partial\hat{x}_1}\, \frac{\partial \boldsymbol{\varphi}_T}{\partial\hat{x}_2}\, \frac{\partial \boldsymbol{\varphi}_T}{\partial\hat{x}_3}]$ be the Jacobian of $\boldsymbol{\varphi}_T$, with $\frac{\partial \boldsymbol{\varphi}_T}{\partial\hat{x}_i}$ column vectors. We define the covariant transformation $\vct{T}_{T,\mathrm{curl}}:H(\mathrm{curl};\hat T)\rightarrow H(\mathrm{curl};T)$ and the Piola contravariant transformation $\vct{T}_{T,\mathrm{div}}:H(\mathrm{div};\hat T)\rightarrow H(\mathrm{div};T)$ such that \begin{align*} \vct{T}_{T,\mathrm{curl}}\hat{\vG}\circ \boldsymbol{\varphi}_T &:=J_T^{-t} \hat{\vG}, \qquad \vct{T}_{T,\mathrm{div}}\hat{\vct{r}}\circ \boldsymbol{\varphi}_T := \frac{1}{\mathrm{det}(J_T)}J_T \hat{\vct{r}}, \end{align*} where $J_T^{-t}$ denotes the transposed of the inverse of $J_T$ and $\mathrm{det}(J_T)$ denotes the determinant of $J_T$. We set $\vct{G}=\vct{T}_{T,\mathrm{curl}}\hat{\vct{R}}\vct{T}_{T,\mathrm{div}}^{-1}\vct{r}$. Then $\nabla\times\vct{G}=\vct{r}$. For any $\vct{G}'\in H(\mathrm{curl};T)$ that satisfies $\nabla\times\hat{\vG}'=\vct{r}$, we can then derive \begin{align*} \|\vct{G}\|_T &\leq Ch_T^{3/2} \|\vct{T}_{T,\mathrm{curl}}^{-1}\vct{G}\|_{\hat T} \\ &= Ch_T^{3/2} \|\hat{\vct{R}}\vct{T}_{T,\mathrm{div}}^{-1}\vct{r} \|_{\hat T} \\ &\leq Ch_T^{3/2} \| \vct{T}_{T,\mathrm{curl}}^{-1}\vct{G}' \|_{\hat T} \\ &\leq C \| \vct{G}' \|_{T}, \end{align*} where $h_T$ denotes the diameter of $T$, where the first and last lines follow from standard scaling arguments, and where the third line follows from \eqref{eq:invCurlBound2} and the fact that $\nabla\times \vct{T}_{T,\mathrm{curl}}^{-1}\vct{G}' = \vct{T}_{T,\mathrm{div}}^{-1}\nabla\times\vct{G}' = \vct{T}_{T,\mathrm{div}}^{-1}\vct{r}$. This then proves the lemma. \end{proof} From \eqref{eq:vthHdelMin}, Lemma \ref{lem:invCurlBound}, and \eqref{eq:vHdelBound1}, it follows that \begin{align*} \|\mu^{1/2} \hat{\tilde{\vH}}^{\Delta}\|_T \leq C \| \mu^{1/2} \hat{\vH}^{\Delta}\|_T \leq C \| \mu^{1/2} \vct{H}^{\Delta} \|_T \end{align*} for all $T\in\mathcal{T}_h$, which proves \eqref{eq:vthHdelBound}. \subsection{Upper bound on $\| \mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi}-\tilde{\alpha}_\nu)\|_{\omega_\nu}$ in terms of $\| \mu^{1/2} \vct{H}^{\Delta} \|_{\omega_\nu}$ (proof of~\eqref{eq:tphiBound})} \label{sec:tphiBound} For all $\nu\in\mathcal{V}_h$, define $\mathcal{F}_{\nu}^I:=\{f\in\mathcal{F}_{h}^{I} \;|\; \partial f\ni \nu\}$ as the set of all internal faces that are connected to $\nu$. Observe that \begin{align} \label{eq:talpha_nu2} \|\mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi}-\tilde{\alpha}_\nu)\|_{\omega_\nu} = \inf_{\substack{u'\in P_{k'+1}^{-1}(\mathcal{T}_\nu), \\ \ju{u'}_f=\ju{\theta_\nu\tilde{\phi}}_f \;\forall f\in\mathcal{F}_{\nu}^I, \\ u'|_f=0 \;\forall f\subset\Gamma_\nu}} \| \mu^{1/2} \nabla_h u'\|_{\omega_\nu} \end{align} for all $\nu\in\mathcal{V}_h$. Indeed, let $u'\in P^{-1}_{k'+1}(\mathcal{T}_\nu)$ such that $\ju{u'}_f=\ju{\theta_\nu\tilde{\phi}}_f$ for all $f\in\mathcal{F}_{\nu}^I$ and $u'|_f=0$ for all $f\subset{\Gamma_\nu}$. Then $u'-\theta_\nu\tilde{\phi} \in P_{k'+1,0,\Gamma_\nu}(\omega_\nu)$ and so $w:=u'- (\theta_\nu\tilde{\phi} -\tilde{\alpha}_\nu) \in P_{k'+1,0,\Gamma_\nu}(\omega_\nu)$. Using \eqref{eq:talpha_nu}, we can then derive \begin{align*} \big(\mu\nabla_h(\theta_\nu\tilde{\phi} -\tilde{\alpha}_\nu),\nabla_hw \big)_{\omega_\nu} = \big(\mu\nabla_h(\theta_\nu\tilde{\phi})-\mu\nabla\tilde{\alpha}_\nu,\nabla w\big)_{\omega_\nu} = 0. \end{align*} From Pythagoras' theorem it then follows that \begin{align*} \|\mu^{1/2} \nabla_hu'\|^2_{\omega_\nu} &= \|\mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi} -\tilde{\alpha}_\nu) \|^2_{\omega_\nu} + \|\mu^{1/2} \nabla w\|^2_{\omega_\nu} \geq \| \mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi} -\tilde{\alpha}_\nu) \|^2_{\omega_\nu}, \end{align*} which proves \eqref{eq:talpha_nu2}. Now, define $\alpha_\nu\in H^1_{0,\Gamma_\nu}(\omega_\nu)$ such that \begin{align*} (\mu\nabla\alpha_\nu,\nabla w)_{\omega_\nu} &= (\mu\nabla_h(\theta_\nu\tilde{\phi}),\nabla w)_{\omega_\nu} &&\forall w\in H^1_{0,\Gamma_\nu}(\omega_\nu). \end{align*} In a similar way as for the discrete case \eqref{eq:talpha_nu2}, one can prove that \begin{align} \label{eq:alpha_nu} \|\mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi}-\alpha_\nu)\|_{\omega_\nu} = \inf_{\substack{u'\in H^1(\mathcal{T}_\nu), \\ \ju{u'}_f=\ju{\theta_\nu\tilde{\phi}}_f \;\forall f\in\mathcal{F}_{\nu}^I, \\ u'|_f=0 \;\forall f\subset\Gamma_\nu}} \|\mu^{1/2} \nabla_h u'\|_{\omega_\nu} \end{align} for all $\nu\in\mathcal{V}_h$. From \cite[Theorem 2.4]{ern19}, it follows that \begin{align*} \inf_{\substack{u'\in P_{k'+1}^{-1}(\mathcal{T}_\nu), \\ \ju{u'}_f=\ju{\theta_\nu\tilde{\phi}}_f \;\forall f\in\mathcal{F}_{\nu}^I, \\ u'|_f=0 \;\forall f\subset\Gamma_\nu}} \|\nabla_h u'\|_{\omega_\nu} &\leq C\inf_{\substack{u'\in H^1(\mathcal{T}_\nu), \\ \ju{u'}_f=\ju{\theta_\nu\tilde{\phi}}_f \;\forall f\in\mathcal{F}_{\nu}^I, \\ u'|_f=0 \;\forall f\subset\Gamma_\nu}} \|\nabla_h u'\|_{\omega_\nu} \end{align*} for all $\nu\in\mathcal{V}_h$. From \eqref{eq:talpha_nu2}, \eqref{eq:alpha_nu}, and the above, it then follows that \begin{align} \label{eq:talphaBound} \|\mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi}-\tilde{\alpha}_\nu)\|_{\omega_\nu} &\leq C \| \mu^{1/2} \nabla_h(\theta_\nu\tilde{\phi}-\alpha_\nu)\|_{\omega_\nu} \end{align} for all $\nu\in\mathcal{V}_h$. Properties \eqref{eq:talpha_nu2}, \eqref{eq:alpha_nu}, and \eqref{eq:talphaBound} are also a consequence of \cite[Corollary 3.1, Remark 3.2]{ern19}. It now remains to derive an upper bound on $\|\nabla_h(\theta_\nu\tilde{\phi}-\alpha_\nu)\|_{\omega_\nu}$ in terms of $\|\vct{H}^{\Delta}\|_{\omega_\nu}$. To do this, we need the following result, which follows immediately from \cite[Theorem 5.1, Remark 5.3]{brenner03}; for completeness, we report a proof of it in Appendix~\ref{appA}. \begin{prp} \label{prp:phiProjErr} For every $u\in H^1(\mathcal{T}_\nu)$, with $(\ju{u},1)_f=0$ for each $f\in\mathcal{F}_{\nu}^I$, we have that \begin{align*} \|u-\overline{u}^{\omega_\nu} \|_{\omega_\nu} \leq Ch_\nu \| \nabla_hu\|_{\omega_\nu}, \end{align*} where $\overline{u}^{\omega_\nu}$ denotes the average of $u$ in $\omega_\nu$. \end{prp} Now, note that $\ju{\theta_\nu\tilde{\phi}-\alpha_\nu}_f = \ju{\theta_\nu\tilde{\phi}}_f = \ju{\theta_\nu(\tilde{\phi} - \alpha - \overline{(\tilde{\phi} - \alpha )}^{\omega_\nu})}_f$ for all $f\in\mathcal{F}_{\nu}^I$. Using \eqref{eq:alpha_nu}, we can then derive \begin{align*} &\|\mu^{1/2}\nabla_h(\theta_\nu\tilde{\phi}-\alpha_\nu)\|_{\omega_\nu} \leq \| \mu^{1/2}\nabla_h (\theta_\nu(\tilde{\phi} - \alpha - \overline{(\tilde{\phi} - \alpha )}^{\omega_\nu})) \|_{\omega_\nu} \\ &\qquad= \|\mu^{1/2}(\nabla\theta_\nu) (\tilde{\phi} - \alpha - \overline{(\tilde{\phi} - \alpha )}^{\omega_\nu}) + \mu^{1/2}\theta_\nu \nabla_h(\tilde{\phi} - \alpha - \overline{(\tilde{\phi} - \alpha )}^{\omega_\nu}) \|_{\omega_\nu} \\ &\qquad\leq \| \mu^{1/2}(\nabla\theta_\nu) (\tilde{\phi} - \alpha - \overline{(\tilde{\phi} - \alpha )}^{\omega_\nu}) \|_{\omega_\nu} + \| \mu^{1/2}\theta_\nu \nabla_h(\tilde{\phi} - \alpha - \overline{(\tilde{\phi} - \alpha )}^{\omega_\nu}) \|_{\omega_\nu} \\ &\qquad\leq Ch_{\nu}^{-1} \| \tilde{\phi} - \alpha - \overline{(\tilde{\phi} - \alpha )}^{\omega_\nu} \|_{\omega_\nu} + \| \nabla_h(\tilde{\phi} - \alpha - \overline{(\tilde{\phi} - \alpha )}^{\omega_\nu}) \|_{\omega_\nu} \\ &\qquad\leq C\| \nabla_h(\tilde{\phi} - \alpha) \|_{\omega_\nu} \\ &\qquad\leq C\| \mu^{1/2} \nabla_h(\tilde{\phi} - \alpha) \|_{\omega_\nu} \end{align*} for all $\nu\in\mathcal{V}_h$, where the fifth line follows from Proposition~\ref{prp:phiProjErr}, \eqref{eq:tlambda1b}, and \eqref{eq:tphi1a}. Now, recall that $\tilde{\phi}-\alpha=\phi-\alpha - \phi^\Delta$ and $\nabla_h\phi^\Delta=\hat{\tilde{\vH}}^{\Delta}-\hat{\vH}^{\Delta}$ (see Section \ref{sec:errDec}). We can use the triangle inequality, \eqref{eq:vthHdelBound}, and \eqref{eq:vHdelBound1} to derive \begin{align*} \| \mu^{1/2} \nabla_h(\tilde{\phi}-\alpha) \|_{\omega_\nu} &\leq \|\mu^{1/2} \nabla_h(\phi-\alpha)\|_{\omega_\nu} + \| \mu^{1/2} \nabla_h\phi^\Delta\|_{\omega_\nu} \\ &\leq \|\mu^{1/2} \nabla_h(\phi-\alpha)\|_{\omega_\nu} + \|\mu^{1/2} \hat{\vH}^{\Delta}\|_{\omega_\nu} + \|\mu^{1/2}\hat{\tilde{\vH}}^{\Delta} \|_{\omega_\nu} \\ &\leq C\| \mu^{1/2} \vct{H}^{\Delta} \|_{\omega_\nu} \end{align*} for all $\nu\in\mathcal{V}_h$. Inequality \eqref{eq:tphiBound} then follows from the last two inequalities and~\eqref{eq:talphaBound}. \section{Numerical experiments}\label{sec:numerics} In the following, we investigate the reliability, efficiency, and polynomial-degree robustness of the equilibrated {\em a posteriori} error estimator $\eta_h$ constructed following Steps 1-5 in Section \ref{sec:errEst}. We present numerical experiments for the unit cube and the L-brick domain on the same test problems as in our previous work~\cite{GGP_arXiv}. In all experiments we choose, for efficiency of the computations, $k'=k$, and set $\mu=1$, unless stated otherwise. As in~\cite{GGP_arXiv}, we do not project the right hand side $\mathbf{j}$ onto $D_k(\mathcal{T}_h)\cap H(\mathrm{div}^0;\Omega)$. This introduces small compatibility errors in Steps 1-3 that can be neglected. We investigate the reliability and efficiency of $\eta_h$ for uniformly refined and adaptively refined meshes. For the adaptive mesh refinement, we employ the standard adaptive finite element loop, \emph{solve}, \emph{estimate}, \emph{mark}, and \emph{refine}. We use a multigrid preconditioned conjugate gradient solver \cite{H1999}, choose $\theta=0.5$ in the bulk marking strategy \cite{Doerfler1996}, and refine the mesh using a bisection strategy \cite{AMP2000}. In order to ensure that the discretisation of $\vct{j}$ is compatible, we add a small gradient correction term following \cite[Section 4.1]{creuse19a}. \subsection{Unit cube examples} \begin{figure}[t] \includegraphics[width=0.49\textwidth]{Cube3dconvergence} \includegraphics[width=0.49\textwidth]{Cube3defficiency} \caption{Error and efficiency indices for the unit cube example with polynomial solution and uniformly refined meshes.} \label{fig:unit:cube} \end{figure} \begin{figure}[t] \includegraphics[width=0.49\textwidth]{Cube3dpolynomialrobustness} \includegraphics[width=0.49\textwidth]{Cube3dpolynomialrobustness2} \caption{Polynomial robustness of the equilibrated {\em a posteriori} error estimator $\eta_h$ in comparison to the residual {\em a posteriori} error estimator $\mu_h$ (left) and the equilibrated estimator $\tilde\eta_h$ (right) for the second unit cube example and a quasi-uniform mesh with 24 elements.} \label{fig:unit:cube:2} \end{figure} In this example, we solve the Maxwell problem on the unit cube $\Omega=(0,1)^3$ with $\hat{\vct{n}}\cdot\vct{H} = 0$ on $\partial \Omega$, for two different right hand sides. \par Firstly, we choose the right-hand side $\mathbf{j}$ according to the polynomial solution \[ \vct{H}=\nabla\times\vct{u}, \qquad \mathbf{u}(x,y,z) = \left( \begin{array}{c} y(1-y)z(1-z)\\ x(1-x)z(1-z) \\x(1-x)y(1-y) \end{array}\right). \] The errors $\|\vct{H}-\vct{H}_h \|_{\Omega}$ and efficiency indices $\eta_h/\| \vct{H}-\vct{H}_h \|_{\Omega}$ are presented in Figure~\ref{fig:unit:cube} for $k=1,2,3$ and uniformly refined meshes. We observe optimal rates $\mathcal{O}(h^k) = \mathcal{O}(N_h^{-k/3})$, $N_h = \operatorname{dim}(R_k(\mathcal{T}_h))$, for the convergence of the errors, and efficiency indices between $1$ and $2$. Note that, for $k=3$, $\mathbf{j}\in D_k(\mathcal{T}_h)\cap H(\mathrm{div}^0;\Omega)$, hence in that case there is no compatibility error. \par For the investigation of the robustness with respect to the polynomial degree $k$, we consider the right-hand side $\mathbf{j}$ according to the non-polynomial solution \[ \vct{H}=\nabla\times\vct{u}, \qquad \mathbf{u}(x,y,z) = \left( \begin{array}{c} \sin(\pi y)\sin(\pi z)\\ \sin(\pi x)\sin(\pi z)\\ \sin(\pi x)\sin(\pi y) \end{array}\right). \] We compare the efficiency indices for $k$-refinement of $\eta_h$ to those of the residual {\em a posteriori} error estimator \cite{BHHW2000} \[ \mu_h^2 := \sum_{T\in\mathcal{T}_h} \frac{h_T^2}{k^2} \| \vct{j} - \nabla\times\vct{H}_h \|_{T}^2 + \sum_{f\in\mathcal{F}_{h}^{I}}\frac{h_f}{k} \| \jut{ \vct{H}_h } \|_{f}^2, \] and to those of the equilibrated {\em a posteriori} error estimator \[ \tilde\eta_h := \| \hat{\tilde{\vH}}^{\Delta} +\nabla_h\tilde\phi\|_{\Omega} \] of our previous work \cite{GGP_arXiv}, which does not include the computation of $\tilde\alpha$. We observe in Figure~\ref{fig:unit:cube:2} that both the efficiency indices for $\mu_h$ and $\tilde\eta_h$ grow in $k$ (although $\tilde\eta_h$ remains confined to small values, for all tested polynomial degrees), while those of $\eta_h$ are stable in $k$. \subsection{L-brick example} \begin{figure}[t] \includegraphics[width=0.49\textwidth]{Lshape3dconvergence} \includegraphics[width=0.49\textwidth]{Lshape3defficiency} \caption{Error and efficiency indices for adaptive mesh refinement for the L-brick example.} \label{fig:L-brick} \end{figure} \begin{figure}[t] \includegraphics[width=0.49\textwidth]{Lshape3dpolynomialrobustness} \includegraphics[width=0.49\textwidth]{Lshape3dpolynomialrobustness2} \caption{Polynomial robustness of the equilibrated {\em a posteriori} error estimator $\eta_h$ in comparison to the residual {\em a posteriori} error estimator $\mu_h$ (left) and the equilibrated estimator $\tilde\eta_h$ (right) for the L-brick example and a quasi-uniform mesh with 36 elements.} \label{fig:L-brick:robustness} \end{figure} In this example, we consider the homogeneous Maxwell problem on the (nonconvex) domain \[ \Omega = (-1,1)\times(-1,1)\times(0,1)\backslash \left( [0,1]\times[-1,0]\times[0,1]\right). \] We choose the right-hand side $\mathbf{j}$ according to the singular solution \[ \vct{H}=\nabla\times\vct{u}, \quad \mathbf{u}(x,y,z) = \nabla \times \left( \begin{array}{c} 0\\ 0 \\(1-x^2)^2(1-y^2)^2((1-z)z)^2r^{2/3}\cos(\frac{2}{3}\varphi) \end{array}\right), \] where $(r,\varphi)$ are the two dimensional polar coordinates in the $x$-$y$-plane. In Figure~\ref{fig:L-brick}, we observe suboptimal convergence rates of asymptotically $\mathcal{O}(N_h^{-2/9})$ for uniform mesh refinement and $k=2$, due to the edge singularity. For adaptive mesh refinement, we observe improved convergence rates of $\mathcal{O}({N_h}^{-1/3})$ for $k=1$, close to $\mathcal{O}(({N_h}/\ln({N_h}))^{-2/3})$ for $k=2$, and of $\mathcal{O}({N_h}^{-2/3})$ for $k\geq3$, which are in fact the best possible rates one can get with isotropic mesh refinement, cf. \cite[section 4.2.3]{A1999}. Again, we observe efficiency indices between 1 and 2. To investigate the $k$-robustness of the estimator $\eta_h$, we compare the efficiency indices for $k$-refinement of $\eta_h$ to those of $\mu_h$, and $\tilde\eta_h$. In Figure~\ref{fig:L-brick:robustness}, we observe the same as for the unit cube example, namely, that the new estimator $\eta_h$ is robust with respect to the polynomial degree $k$, while the residual estimator $\mu_h$, as well as the equilibrated estimator $\tilde\eta_h$, is not robust in $k$. \subsection{Example with discontinuous permeability} \begin{figure}[t] \includegraphics[width=0.49\textwidth]{CubeVarCoeff3dconvergence} \includegraphics[width=0.49\textwidth]{CubeVarCoeff3defficiency} \caption{Error and efficiency indices for adaptive mesh refinement for the example with discontinuous permeability.} \label{fig:VarCoeff} \end{figure} For the last example, we choose a discontinuous permeability \[ \mu(x,y,z) = \left\{ \begin{array}{ll} \mu_1 & \text{if } y<1/2 \text{ and } z < 1/2,\\ \mu_2 & \text{otherwise}, \end{array} \right. \] on the unit cube $\Omega = (0,1)^3$, and the right hand side $\mathbf{j}=(1,0,0)^t$. We choose $k=2$, $\mu_1=1$, and vary $\mu_2= 10^\ell$ for $\ell=1,2,3$. Since the exact solution is unknown, we approximate the error by comparing the numerical approximations to a reference solution, which is obtained from the last numerical approximation by 8 more adaptive mesh refinements. In this example, the adaptive algorithm refines strongly along the edge with endpoints $(0,1/2,1/2)^t$ and $(1,1/2,1/2)^t$, similarly to what is shown in~\cite[Figure~6]{GGP_arXiv}. The errors in Figure~\ref{fig:VarCoeff} converge with about $\mathcal{O}((N_h/\ln(N_h))^{-2/3})$, which is optimal for isotropic adaptive mesh refinement, and the efficiency indices are robust with respect to the contrast of the permeability. \section{Conclusions}\label{sec:conclusion} We have introduced and analyzed an \emph{a posteriori} error estimator for arbitrary-degree N\'ed\'elec discretizations of the magnetostatic problem based on an equilibration principle. This estimator is constructed by adding a localized gradient correction to the estimator introduced in~\cite{GGP_arXiv}, and is proven to be reliable with reliability constant 1, and uniformly efficient, not only in the mesh size, but also in the degree of the polynomial approximation. The computation of the new gradient term requires solving local problems on vertex patches. The polynomial-degree robustness of the new estimator has been numerically demonstrated on test problems with smooth as well as singular solutions, and for problems with a discontinuous magnetic permeability. \bibliographystyle{abbrv}

https://arxiv.org/abs/1211.6435

The topology of toric origami manifolds

A folded symplectic form on a manifold is a closed 2-form with the mildest possible degeneracy along a hypersurface. A special class of folded symplectic manifolds are the origami symplectic manifolds, studied by Cannas da Silva, Guillemin and Pires, who classified toric origami manifolds by combinatorial origami templates. In this paper, we examine the topology of toric origami manifolds that have acyclic origami template and co-orientable folding hypersurface. We prove that the cohomology is concentrated in even degrees, and that the equivariant cohomology satisfies the GKM description. Finally we show that toric origami manifolds with co-orientable folding hypersurface provide a class of examples of Masuda and Panov's torus manifolds.

\section{Introduction}\label{sec:intro} Toric symplectic manifolds are a useful class of examples for testing general theories and making explicit computations. Statements and proofs of important theorems often simplify in the case of toric manifolds. Delzant's classification of toric symplectic manifolds in terms of convex polytopes allows the translation of geometric and topological questions into combinatorial ones. In this paper, we study toric actions in the category of folded symplectic manifolds. Relaxing the requirement that the manifold be symplectic broadens the class of manifolds with toric actions. The mildest degeneracy is to allow the $2$-form to be zero along a hypersurface. In this instance, there remains enough geometric structure to be able to classify such toric origami manifolds combinatorially. In this paper, we study the topology of a particular class of toric origami manifolds, those with acyclic template and co\"orientable fold. For such manifolds, we prove that the ordinary cohomology is concentrated in even degrees (Theorem~\ref{thm:even cohomology}). This allows us to deduce a variety of facts about the equivariant cohomology of these manifolds, and in particular to describe the equivariant cohomology ring combinatorially (Theorem~\ref{thm:origami GKM}). Our class of toric origami manifolds does fit into the framework of torus manifolds (Theorem~\ref{thm:locstd}). The origami structure allows us to give explicit inductive proofs. We plan to use similar geometric techniques to study the non-co\"orientable and non-acyclic cases. We hope that this approach will also generalize to a class of torus manifolds that arise from combinatorial origami templates, in the same way that some torus manifolds arise from combinatorial polytopes. The remainder of this paper is organized as follows. In Section~\ref{sec:background}, we review the symplectic and folded symplectic geometry underlying our work. We then provide a framework for computing the ordinary and equivariant cohomology of origami manifolds with co\"orientable folding hypersurface and acyclic template in Sections~\ref{se:even} and \ref{se:eq-coh}. In Section~\ref{sec:std} we describe the relationship of our work with the toric topology literature. \smallskip \noindent {\bf Acknowledgements.} We are grateful to Jean-Claude Hausmann for his help and patience when we were sorting out the commutativity of diagram~\eqref{eq:comm diag PD}; and to Nick Sheridan for his suggestions regarding Definition~\ref{def:template}. We would also like to thank Ana Cannas da Silva, Victor Guillemin, Allen Hatcher, Yael Karshon, Allen Knutson, Tomoo Matsumura, and Milena Pabiniak for many helpful conversations. We are very grateful for the comments from the anonymous referees, which led to several improvements of this article. \section{Origami manifolds}\label{sec:background} \subsection{Symplectic manifolds.}\label{se:symplectic} We begin with a very quick review of symplectic geometry, following \cite{ca:book}. Let $M$ be a manifold equipped with a {\bf symplectic form} $\omega\in \Omega^2(M)$: that is, $\omega$ is closed ($d\omega = 0$) and non-degenerate. In particular, the non-degeneracy condition implies that $M$ must be an even-dimensional manifold. The simplest examples include \begin{enumerate} \item $M=\SS^2 = {\mathbb{C}} \P^1$ with $\omega_p ({\mathcal{X}},{\mathcal{Y}})=$ signed area of the parallelogram spanned by ${\mathcal{X}}$ and ${\mathcal{Y}}$; \item $M$ any compact orientable surface with $\omega$ the area form; and \item $M= {\mathbb{R}}^{2d}$ with $\omega = \sum dx_i\wedge dy_i$. The Darboux Theorem says that every symplectic manifold has local co\"ordinates so that $\omega$ is of this standard form. \end{enumerate} Suppose that a compact connected abelian Lie group ${\mathbb{T}}= (\SS^1)^n$ acts on $M$ preserving $\omega$. The action is {\bf weakly Hamiltonian} if for every vector $\xi\in\mathfrak{t}$ in the Lie algebra $\mathfrak{t}$ of ${\mathbb{T}}$, the vector field $$ {\mathcal{X}}_\xi(p) = \frac{d}{dt}\Big[ \exp (t\xi)\cdot p \Big] \bigg|_{t=0} $$ is a {\bf Hamiltonian vector field}. That is, we require $\omega({\mathcal{X}}_\xi, \cdot )$ to be an exact one-form\footnote{\, The one-form $\omega({\mathcal{X}}_\xi, \cdot )$ is automatically closed because the action preserves $\omega$.}: \begin{equation}\label{eq:mmap} \omega({\mathcal{X}}_\xi, \cdot ) = d\phi^\xi. \end{equation} Thus each $\phi^\xi$ is a smooth function on $M$ defined by the differential equation \eqref{eq:mmap}, so determined up to a constant. Taking them together, we may define a {\bf moment map} $$ \begin{array}{rcc} \Phi: M & \to & \mathfrak{t}^* \\ p & \mapsto & \left(\begin{array}{rcl} \mathfrak{t} & \longrightarrow & {\mathbb{R}} \\ \xi & \mapsto & \phi^\xi(p) \end{array}\right). \end{array} $$ The action is {\bf Hamiltonian} if the moment map $\Phi$ can be chosen to be a ${\mathbb{T}}$-invariant map. Atiyah and Guillemin-Sternberg have shown that when $M$ is a compact Hamiltonian ${\mathbb{T}}$-manifold, the image $\Phi(M)$ is a convex polytope, and is the convex hull of the images of the fixed points $\Phi(M^{{\mathbb{T}}})$ \cite{at:convexity, gu-st:convexity}. For an {\bf effective}\footnote{ \, An action is effective if no non-trivial subgroup acts trivially.} Hamiltonian ${\mathbb{T}}$ action on $M$, $ \dim({\mathbb{T}})\leq \frac{1}{2}\dim(M). $ We say that the action is {\bf toric} if this inequality is in fact an equality. A symplectic manifold $M$ with a toric Hamiltonian ${\mathbb{T}}$ action is called a {\bf symplectic toric manifold}. Delzant used the moment polytope to classify symplectic toric manifolds. A polytope $\Delta$ in ${\mathbb{R}}^n$ is {\bf simple} if there are $n$ edges incident to each vertex, and it is {\bf rational} if each edge vector has rational slope: it lies in ${\mathbb{Q}}^n\subset {\mathbb{R}}^n$. A simple polytope is {\bf smooth at a vertex} if the $n$ primitive vectors parallel to the edges at the vertex span the lattice ${\mathbb{Z}}^n\subseteq{\mathbb{R}}^n$ over ${\mathbb{Z}}$. It is {\bf smooth} if it is smooth at each vertex. A simple rational smooth convex polytope is called a {\bf Delzant polytope}. We may now state Delzant's result. \begin{theorem}[Delzant \cite{de:hamiltoniens}] There is a one-to-one correspondence $$ \left\{\begin{array}{c} \mbox{compact toric}\\ \mbox{symplectic manifolds}\\ \end{array}\right\} \leftrightsquigarrow \left\{\begin{array}{c} \mbox{Delzant polytopes} \end{array}\right\} , $$ up to equivariant symplectomorphism on the left-hand side and affine equivalence on the right-hand side. \end{theorem} \subsection{Origami manifolds.} We now relax the non-degeneracy condition on $\omega$, following \cite{CGP:origami}. A {\bf folded symplectic form} on a $2n$-dimensional manifold $M$ is a $2$-form $\omega\in \Omega^2(M)$ that is closed ($d\omega = 0$), whose top power $\omega^n$ intersects the zero section transversely on a subset $Z$ and whose restriction to points in $Z$ has maximal rank. The transversality forces $Z$ to be a codimension $1$ embedded submanifold of $M$. We call $Z$ the {\bf folding hypersurface} or {\bf fold}. The simplest examples of folded symplectic manifolds include the following. \begin{enumerate} \item Euclidean space $M= {\mathbb{R}}^{2d}$ has folded symplectic form $\omega = x_1 dx_1\wedge dy_1 + \sum_{i=2}^d dx_i\wedge dy_i$. The Folded Darboux Theorem says that at points in $Z=\{ x_1=0\}$, every folded symplectic manifold has local co\"ordinates so that $\omega$ is of this standard form \cite[IIIA.4.2.2]{ma:formes}. \item Any even-dimensional sphere $M=\SS^{2n}\subset {\mathbb{C}}^{n}\oplus {\mathbb{R}}$ may be equipped with the form $\omega_{{\mathbb{C}}^n}\oplus 0$. The folding hypersurface is the equator $Z = \SS^{2n-1}\subset {\mathbb{C}}^n\oplus \{ 0\}$. \item Any compact surface $M$ can be equipped with a folded symplectic form with $Z$ a union of circles. See, for instance, Example 3.19 of \cite{CGP:origami}, and use Remark 2.33 of the same paper together with the classification of closed surfaces. This includes non-orientable surfaces. For example, ${\mathbb{R}} P^2$ can be equipped with a folded symplectic form so that $Z$ is a single circle. \end{enumerate} Let $i:Z\hookrightarrow M$ be the inclusion of $Z$ as a submanifold of $M$. Our assumptions imply that $i^*\omega$ has a $1$-dimensional kernel on $Z$. This line field is called the {\bf null foliation} on $Z$. An {\bf origami manifold} is a folded symplectic manifold $(M, \omega)$ whose null foliation is fibrating: $Z\stackrel{\pi}{\to} B$ is a fiber bundle with orientable circle fibers over a compact base $B$. The form $\omega$ is called an {\bf origami form} and the bundle $\pi$ is called the {\bf null fibration}. A diffeomorphism between two origami manifolds which intertwines the origami forms is called an {\bf origami-symplectomorphism}. In the examples above, the first is not origami because the fibers are ${\mathbb{R}}$ rather than $\SS^1$, but the second and third are origami. In the second example, the null fibration is the Hopf bundle $\SS^{2n-1}\to {\mathbb{C}} P^{n-1}$, and in the third example, the base $B$ consists of isolated points. The definition of a Hamiltonian action only depends on $\omega$ being closed. Thus, in the folded framework, we may define moment maps and toric actions exactly as in Section~\ref{se:symplectic}. For example, the action ${\mathbb{T}}^2\mathbin{\raisebox{-.5pt}{\reflectbox{\begin{sideways \SS^4\subset {\mathbb{C}}^2\oplus {\mathbb{R}}$ given by rotation on the ${\mathbb{C}}^2$ co\"ordinates is Hamiltonian with moment map $$ \Phi ( z_1, z_1, t) = \left( |z_1|^2,|z_2|^2\right) . $$ The image of this map is shown in Figure~\ref{fig:S4} below. \begin{figure}[ht] \centering { \includegraphics[scale=0.4]{4-sphere.pdf} } \caption{The moment map image for the ${\mathbb{T}}^2$ action on $\SS^4$. The image consists of two overlapping copies of a triangle, which we have slightly unfolded. The red hypotenuse is the image of the equator $\SS^3$. Every other point in the image has two connected components mapping to it, one from the northern hemisphere and the other from the southern. } \label{fig:S4} \end{figure} An oriented origami manifold $M$ with fold $Z$ may be \textbf{unfolded} into a symplectic manifold as follows. Consider the closures of the connected components of $M\setminus Z$, a manifold with boundary which consists of two copies of $Z$. We collapse the fibers of the null fibration by identifying the boundary points that are in the same fiber of the null fibration of each individual copy of $Z$. The result, $M_0:=(M\setminus Z) \cup B_1 \cup B_2$, is a (disconnected) smooth manifold that can be naturally endowed with a symplectic form which on $M_0\setminus (B_1 \cup B_2)$ coincides with the origami form on $M\setminus Z$. Because this can be achieved using symplectic cutting techniques, the resulting manifold $M_0$ is called the \textbf{symplectic cut space} (and its connected components the \textbf{symplectic cut pieces}), and the process is also called \textbf{cutting}. An example of cutting a $2$-torus is shown in Figure~\ref{fig:cut-torus}. The symplectic cut space of a nonorientable origami manifold is the ${\mathbb{Z}}_2$-quotient of the symplectic cut space of its orientable double cover. \begin{center} \begin{figure}[h] \includegraphics[height=0.65in]{torus-example-unfolding-1.pdf} \hskip 0.1in \includegraphics[height=0.65in]{torus-example-unfolding-2.pdf} \includegraphics[height=0.65in]{torus-example-unfolding-3.pdf} \hskip 0.1in \includegraphics[height=0.65in]{torus-example-unfolding-4.pdf} \includegraphics[height=0.65in]{torus-example-unfolding-5.pdf} \caption{ The torus, with fold $Z=\SS^1\cup \SS^1$ in purple; the middle step before collapsing, the two copies of $Z$ are in blue and purple; and the final cut space $M_0=\SS^2\cup \SS^2$ with $B_1$ in red and $B_2$ in blue.}\label{fig:cut-torus} \end{figure} \end{center} In the example shown in Figure~\ref{fig:S4}, unfolding the origami $\SS^4$ yields ${\mathbb{C}}\P^2\sqcup\overline{{\mathbb{C}}\P^2}$. This is suggested by the image of the moment map: the moment image of each toric ${\mathbb{C}}\P^2$ (regardless of orientation) is a triangle. The cut space $M_0$ of an oriented origami manifold $(M,\omega)$ inherits a natural orientation. It is the orientation on $M_0$ induced from the orientation on $M$ that matches the symplectic orientation on the symplectic cut pieces corresponding to the subset of $M\setminus Z$ where $\omega^n>0$ and the opposite orientation on those pieces where $\omega^n<0$. In this way, we can associate a $+$ or $-$ sign to each of the symplectic cut pieces of an orientend origami manifold, as well as to the corresponding connected components of $M\setminus Z$. \begin{remark}\label{rmk:orientable} In this paper we restrict to origami manifolds whose fold is {\bf co\"orientable}: that is, the fold has an orientable neighborhood. Note that this not imply that the manifold is orientable. Indeed, for an orientable $M$, the condition that $\omega^n$ intersects the zero section transversally implies that the connected components of $M\setminus Z$ which are adjacent in $M$ have opposite signs. Since $M$ is connected, picking a sign for one connected component of $M\setminus Z$ determines the signs for all other components. As a consequence, an origami manifold $M$ with co\"orientable fold is orientable if and only if it is possible to make such a global choice of signs for the connected components of $M\setminus Z$. The moment image of a non-orientable origami manifold that nevertheless has co\"orientable fold is given in Figure~\ref{fig:non-acyclic}. \end{remark} \begin{proposition}[\!\! {\cite[Props.\ 2.5 \& 2.7]{CGP:origami}}] \label{prop:model} Let $M$ be a (possibly disconnected) symplectic manifold with a codimension two symplectic submanifold $B$ and a symplectic involution $\gamma$ of a tubular neighborhood $\mathcal{U}$ of $B$ which preserves $B$\footnote{\, In the nonco\"orientable case, the involution must satisfy additional conditions, see \cite[Def.\ 2.23]{CGP:origami}. In the co\"orientable case, we have $B=B_1\cup B_2$ and the involution $\gamma$ maps a tubular neighborhood of $B_1$ to one of $B_2$ and vice versa.}. Then there is an origami manifold $\widetilde{M}$ such that $M$ is the symplectic cut space of $\widetilde{M}$. Moreover, this manifold is unique up to origami-symplectomorphism. \end{proposition} This newly-created fold $Z\subset\widetilde{M}$ involves the radial projectivized normal bundle of $B\subset M$, so we call the origami manifold $\widetilde{M}$ the \textbf{radial blow-up} of $M$ through $(\gamma,B)$. The cutting operation and the radial blow-up operation are in the following sense inverse to each other. \begin{proposition}[\!\! {\cite[Prop.\ 2.37]{CGP:origami}}] Let $M$ be an origami manifold with cut space $M_0$. The radial blow-up $\widetilde{M_0}$ is origami-symplectomorphic to $M$. \end{proposition} There exist Hamiltonian versions of these two operations which may be used to see that the moment map $\Phi$ for an origami manifold $M$ coincides, on each connected component of $M \setminus Z$ with the induced moment map $\Phi_i$ on the corresponding symplectic cut piece $M_i$. As a result, the moment image $\Phi(M)$ is the union of convex polytopes $\Delta_i$. Furthermore, if the circle fibers of the null fibration for a connected component $\mathcal{Z}$ of the fold $Z$ are orbits for a circle subgroup $\SS^1\subset {\mathbb{T}}$, then $\Phi(\mathcal{Z})$ is a facet of each of the two polytopes corresponding to neighboring components of $M\setminus Z$. Let us denote these two polytopes $\Delta_1$ and $\Delta_2$. We note that they must \textbf{agree} near $\Phi(\mathcal{Z})$: there is a neighborhood $\mathcal{V}$ of $\Phi(\mathcal{Z})$ in ${\mathbb{R}}^n$ such that $\Delta_1\cap\mathcal{V}=\Delta_2\cap\mathcal{V}$. The condition that the circle fibers are orbits is automatically satisfied when the action is toric, and in that case there is a classification theorem in terms of the moment data. The moment data of a toric origami manifold can be encoded in the form of an origami template{, originally defined in~\cite[Def.\ 3.12]{CGP:origami}. Definition~\ref{def:template} below is a refinement of that original definition. The reasons for this refinement are explained in Remark~\ref{rmk:reasons}. Following~\cite[p.\ 5]{book:graph}, a \textbf{graph} $G$ consists of a nonempty set $V$ of \textbf{vertices} and a set $E$ of \textbf{edges} together with an incidence relation that associates an edge with its two \textbf{end vertices}, which need not be distinct. Note that this allows for the existence of (distinguishable) multiple edges with the same two end vertices, and of \textbf{loops} whose two end vertices are equal. We introduce some additional notation: let $\mathcal{D}_n$ be the set of all Delzant polytopes in ${\mathbb{R}}^n$ and $\mathcal{E}_n$ the set of all subsets of ${\mathbb{R}}^n$ which are facets of elements of $\mathcal{D}_n$. \begin{definition}\label{def:template} An $n$-dimensional \textbf{origami template} consists of a graph $G$, called the \textbf{template graph}, and a pair of maps $\Psi_V: V\to\mathcal{D}_n$ and $\Psi_E:E\to\mathcal{E}_n$ such that: \begin{enumerate} \item if $e$ is an edge of $G$ with end vertices $u$ and $v$, then $\Psi_E(e)$ is a facet of each of the polytopes $\Psi_V(u)$ and $\Psi_V(v)$, and these polytopes agree near $\Psi_E(e)$; and \item if $v$ is an end vertex of each of the two distinct edges $e$ and $f$, then $\Psi_E(e)\cap\Psi_E(f)=\emptyset$. \end{enumerate} \end{definition} The polytopes in the image of the map $\Psi_V$ are the Delzant polytopes of the symplectic cut pieces. For each edge $e$, the set $\Psi_E(e)$ is a facet of the polytope(s) corresponding to the end vertices of $e$. We refer to such a set as a {\bf fold facet}, as it is the image of the connected components of the folding hypersurface\footnote{ \, A nonco\"orientable connected component of the folding hypersurface corresponds to a loop edge $e$.}. In the example of Figure~\ref{fig:S4}, the template graph $G$ has two vertices and one edge joining them. Both vertices are mapped to the same isosceles right angle triangle under $\Psi_V$, and the edge is mapped to the hypotenuse of that triangle under $\Psi_E$. \begin{remark}\label{rmk:reasons} In the original definition of origami template, Definition 3.12 in~\cite{CGP:origami}, a template consisted of a pair $(\mathcal{P},\mathcal{F})$. The set $\mathcal{P}$ was a collection of Delzant polytopes and $\mathcal{F}$ was a collection of pairs or singletons of facets of polytopes in $\mathcal{P}$, satisfying certain conditions. Roughly speaking, $\mathcal{P}$ is the image of $\Psi_V$ and the sets in $\mathcal{F}$ assigned identifications of facets of polytopes in $\mathcal{P}$ in a way similar to that of the map $\Psi_E$. To understand the problem with this old definition we turn again to the example of Figure~\ref{fig:S4}: the collection $\mathcal{P}$ would contain two identical triangles, and $\mathcal{F}$ would contain one pair, consisting of the hypotenuses of each of the triangles. However, $\mathcal{P}$ is a set, and therefore if it consists of two identical elements it actually consists of only one such element. The same issue exists with the pairs in $\mathcal{F}$ and in other examples, with $\mathcal{F}$ itself. Simply replacing the word set by the word multiset to allow for multiple instances of the same element gives rise to a different type of problem. \vskip 0.1in \noindent We thank an anonymous referee for bringing this problem to our attention. \end{remark} } \noindent With these combinatorial data in place, we may now state the classification theorem. \begin{theorem}[\!\! {\cite[Theorem 3.13]{CGP:origami}}]\label{thm:origamiDelzant} There is a one-to-one correspondence $$ \left\{\begin{array}{c} \mbox{compact toric}\\ \mbox{origami manifolds} \end{array}\right\} \leftrightsquigarrow \left\{\begin{array}{c} \mbox{origami templates} \end{array}\right\} , $$ up to equivariant origami-symplectomorphism on the left-hand side, and affine equivalence of the image of the template in $\mathbb{R}^n$ on the right-hand side. \end{theorem} The orbit space $M/{\mathbb{T}} $ of a toric origami manifold is closely related to the origami template. When $M$ is a toric symplectic manifold, then the orbit space may be identified with the corresponding Delzant polytope; this identification is achieved by the moment map. For a toric origami manifold, the orbit space is realized as the topological space obtained by gluing the polytopes in $\Psi_V(V)$ along the fold facets as specified by the map $\Psi_E$. More precisely, the orbit space is the quotient \begin{equation}\label{eq:orbitspace} M/{\mathbb{T}} = \bigsqcup_{v\in V} (v,\Psi_V(v)) \Big/ \thicksim \ , \end{equation} where we identify $(u,x)\thicksim(v,y)$ if there exists an edge $e$ with endpoints $u$ and $v$ and the points $x=y\in\Psi_E(e)\subset {\mathbb{R}}^n$. Again, this identification is achieved by the moment map. In simple low-dimensional examples, we can visualize the orbit space by superimposing the polytopes $\Psi_V(v)$ in $\mathbb{R}^n$ and indicating which of their facets to identify; see for instance Figures~\ref{fig:S4}, \ref{fig:non-acyclic}, \ref{fig:torictorus} and ~\ref{fig:folded-hirz}. We will see in Section~\ref{sec:std} that there is a deformation retraction from orbit space $M/{\mathbb{T}}$ to the template graph. There is a natural description of the faces of $M/{\mathbb{T}}$. The facets of a polytope are well-understood. The set of facets of $M/{\mathbb{T}}$ is $$ \bigsqcup_{\mathclap{\substack{v\in V \\ F \text{ facet of } \Psi_V(v)\\ F \text{ not a fold facet}}}} \,(v,F) \Big/ \thicksim \ , $$ \vskip 0.05in \noindent where the equivalence relation is induced by the one in (\ref{eq:orbitspace}). The faces of $M/{\mathbb{T}}$ are non-empty intersections of facets in $M/{\mathbb{T}}$, together with $M/{\mathbb{T}}$ itself. This notion of face of the orbit space agrees with Masuda and Panov's definition mentioned in Section~\ref{sec:std}. \section{Cohomology concentrated in even degrees} \label{se:even} We say that the origami template is {\bf acyclic} if the template graph is acyclic, and therefore a tree. In this case, the {\bf leaves} of the origami template are the polytopes which are images under $\Psi_V$ of the leaves of the template graph. In light of Remark~\ref{rmk:orientable}, a toric origami manifold with co\"orientable folding hypersurface is orientable exactly when the template graph has no odd cycles. In particular, if $M$ has an acyclic origami template, then $M$ is automatically orientable. Two non-acyclic origami templates are shown in Figure~\ref{fig:non-acyclic}, one corresponding to an orientable origami manifold and the other to a non-orientable one. \begin{figure}[h] \begin{center} \includegraphics[height=1.6in]{cropped-two-mobius.pdf} \caption{The orbit spaces corresponding to two non-acyclic templates of origami manifolds with co\"orientable fold. Such manifolds are orientable exactly when there exists a consistent choice of signs for the polytopes such that the sign changes whenever we traverse a fold facet. The one on the left corresponds to a non-orientable manifold and the one on the right to an orientable manifold.} \label{fig:non-acyclic} \end{center} \end{figure} The proof of the main theorem in this section will involve induction on the number of vertices of the template graph. To prove the inductive hypothesis, we need some auxiliary spaces. We focus on a connected component $\mathcal{Z}$ of the fold $Z$ such that $M\smallsetminus\mathcal{Z}$ is the union of one open symplectic manifold $W_-$ and one open origami manifold $W_+$. The corresponding closed manifolds with boundary are $M_-=W_-\cup \mathcal{Z}$ and $M_+=W_+\cup \mathcal{Z}$. Combinatorially, $M_-$ corresponds to a leaf of the origami template for $M$. Collapsing the fibers of the null-foliation on $\mathcal{Z}$ results in a toric symplectic manifold $\mathcal{B}=\mathcal{Z}/\SS^1$ of dimension $\dim(\mathcal{B})=\dim(M)-2$. Cutting $M$ along $\mathcal{Z}$ yields one toric symplectic manifold $C_-$ and one toric origami manifold $C_+$ with one fewer connected component of the fold $Z\smallsetminus\mathcal{Z}$. Finally, we use the space $C=C_+\cup_{\mathcal{B}} C_-$, which is not a manifold. This notation is illustrated in the Figure~\ref{fig:notation} and summarized in Table~\ref{table:notation}. \begin{center} \begin{figure}[h] \includegraphics[width=0.75in]{M.pdf} \hskip 0.4in \includegraphics[width=0.75in]{MplusMminus.pdf} \hskip 0.4in \includegraphics[width=0.75in]{CplusCminus.pdf} \hskip 0.4in \includegraphics[width=0.75in]{C.pdf} \caption{From left to right, the spaces $M$, $M_+\sqcup M_-$, $C_+\sqcup C_-$ and $C$.}\label{fig:notation} \end{figure} \end{center} \renewcommand{\arraystretch}{1.3} \begin{center} \begin{table}[h]\caption{Summary of notation} \begin{tabular}{c|l} \hline Notation & Description \\ \hline $M$ & Toric origami manifold, ${\mathbb{T}}\mathbin{\raisebox{-.5pt}{\reflectbox{\begin{sideways M$ \\ \hline $\mathcal{Z}\subset Z$ & Connected component $\mathcal{Z}$ of the fold $Z$ \\ \hline $\mathcal{B}\subset B$ & Toric symplectic manifold $\mathcal{B}=\mathcal{Z}/\SS^1$ and union of such $B=Z/\SS^1$ \\ \hline $W_+$ & Connected component of $M\setminus \mathcal{Z}$ that is an open origami manifold \\ \hline $M_+$ & $W_+\cup \mathcal{Z}$, an origami manifold with boundary \\ \hline $C_+$ & $W_+\cup \mathcal{B}$, an origami manifold with one fewer vertex in its template graph \\ \hline $W_-$ & Connected component of $M\setminus \mathcal{Z}$ that is an open symplectic manifold \\ \hline $M_-$ & $W_-\cup \mathcal{Z}$, a symplectic manifold with boundary \\ \hline $C_-$ & $W_-\cup \mathcal{B}$, a toric symplectic manifold \\ \hline $C$ & $W_+ \cup \mathcal{B}\cup W_-=C_+\cup_{\mathcal{B}} C_-$ (a ${\mathbb{T}}$-space, but not a manifold) \\ \hline \end{tabular}\label{table:notation} \end{table} \end{center} \renewcommand{\arraystretch}{1.1} \begin{lemma}\label{le:magic} Suppose that $M$ is a compact symplectic toric manifold with moment polytope $\Delta_M$. Let $\mathcal{B}$ be a codimension $k$ ${\mathbb{T}}$-invariant symplectic submanifold whose moment map image $\Delta_\mathcal{B}$ is a $k$-dimensional face of $\Delta_M$. Then the inclusion $i:\mathcal{B}\hookrightarrow M$ induces a surjection $$ i^*:H^*(M;{\mathbb{Z}})\twoheadrightarrow H^*(\mathcal{B};{\mathbb{Z}}). $$ \end{lemma} \begin{remark}\label{rem:to-masuda-panov} Though it holds in more generality, we will only use this Lemma when the submanifold $\mathcal{B}$ is of codimension $2$. Just as \cite[Lemma 2.3]{masuda-panov} allows Masuda and Panov to make inductive arguments, our Lemma~\ref{le:magic} will be the crucial ingredient when we build the cohomology of $M$ from its related toric pieces. \end{remark} \begin{proof} The manifold $\mathcal{B}$ is itself a symplectic toric manifold. Its cohomology is generated in degree $2$, with one class for each facet $F$ of $\Delta_\mathcal{B}$. Such a facet $F$ is the intersection of a facet $\widetilde{F}$ of $\Delta_M$ with $\Delta_\mathcal{B}$. Under the restriction map $i^*$, the generator corresponding to $\widetilde{F}$ maps to the generator corresponding to $F$. Thus, $i^*$ is surjective. \end{proof} \begin{theorem}\label{thm:even cohomology} Let ${\mathbb{T}}\mathbin{\raisebox{-.5pt}{\reflectbox{\begin{sideways M$ be a compact toric origami with acyclic origami template and co\"orientable folding hypersurface. Then the cohomology $H^*(M;{\mathbb{Z}})$ is concentrated in even degrees. \end{theorem} \begin{proof} We proceed by induction on the number $n$ of vertices of the template graph, or equivalently, of connected components of $M\setminus Z$. The base case is when $n=1$ and $M$ is a compact toric symplectic manifold. In this case, the fact that $H^*(M)$ is generated in degree 2, and hence concentrated in even degrees is well-known. For example, see \cite{danilov, jur:toric}. The case of a connected folding hypersurface is when $n=2$, and concentration in even degrees is proven in \cite[Corollary~5.1]{CGP:origami}. For the inductive step, we assume that every compact toric origami manifold with co\"orientable folding hypersurface and acyclic origami template with at most $(n-1)$ vertices has cohomology concentrated in even degrees. Let $M$ be a compact toric origami manifold with co\"orientable folding hypersurface and acyclic origami template with $n$ vertices. Choose a leaf of the origami template, and let $\mathcal{Z}$ be the connected component of the folding hypersurface that corresponds to the facet separating the leaf from the rest of the origami template. We use the notation $M_-$, $M_+$, $C_-$, $C_+$, $C$ and $\mathcal{B}$ as listed in Table~\ref{table:notation}. In particular, we note that $C_-$ is actually a compact toric symplectic manifold and $C_+$ is a compact toric origami manifold with co\"orientable folding hypersurface and acyclic origami template with $(n-1)$ vertices. Let $\mathcal{Z}\stackrel{\pi}{\longrightarrow} \mathcal{B}$ be the quotient by the null-fibration. Then $\pi$ induces maps $$ M\stackrel{p}{\longrightarrow} C \mbox{ and } M_-\stackrel{p_-}{\longrightarrow} C_-. $$ \noindent We begin by studying the cohomology of $C$. \begin{Claim}\label{cl:H*C} The cohomology ring $H^*(C;{\mathbb{Z}})$ is concentrated in even degrees. \end{Claim} \noindent {\bf Proof of Claim~\ref{cl:H*C}.} We may choose ${\mathbb{T}}$-invariant collar neighborhoods of $C_-$ and $C_+$ in $C$ that deformation retract to $C_-$ and $C_+$ respectively. This is analogous to choosing a collar neighborhood of $Z$ in $M$, as described in the remarks just before Proposition~\ref{prop:model} above. The intersection of these neighborhoods is a collar neighborhood of $\mathcal{B}$ and deformation retracts onto $\mathcal{B}$. The Mayer-Vietoris sequence for these collar neighborhoods induces a long exact sequence, in cohomology with integer coefficients \begin{eqnarray}\label{MayerVietoris} \xymatrix{ \cdots \ar[r] & H^*(C) \ar[r] & H^*(C_+)\oplus H^*(C_-) \ar[r] & H^*(\mathcal{B}) \ar[r] & \cdots }. \end{eqnarray} As $C_-$ is a compact toric symplectic manifold, Lemma~\ref{le:magic} implies that $H^*(C_-)\rightarrow H^*(B)$ is a surjection. Thus the long exact \eqref{MayerVietoris} splits into short exact sequences (again with integer coefficients) \begin{eqnarray}\label{ShortMayerVietoris} \xymatrix{ 0 \ar[r] & H^*(C) \ar[r] & H^*(C_+)\oplus H^*(C_-) \ar[r] & H^*(\mathcal{B}) \ar[r] & 0 }. \end{eqnarray} Note that the cohomology of $C_-$ and $\mathcal{B}$ is concentrated in even degrees because $C_-$ and $\mathcal{B}$ are compact toric symplectic manifolds. By the induction hypothesis, the cohomology of $C_+$ is concentrated in even degrees. We conclude from \eqref{ShortMayerVietoris} in odd degrees that $H^*(C;{\mathbb{Z}})$ must be zero in odd degrees. \hfill \ding{52} \medskip \noindent We now look at the relationship between the cohomology of $C_-$ and that of $M_-$. \begin{Claim}\label{C- surjects} The quotient map $p_-:M_-\longrightarrow C_-$ induces a surjection in cohomology $$ p_-^*: H^*(C_-;{\mathbb{Z}})\twoheadrightarrow H^*(M_-;{\mathbb{Z}}). $$ In particular, $H^*(C_-;{\mathbb{Z}})$ is concentrated in even degrees, and so $H^*(M_-;{\mathbb{Z}})$ is as well. \end{Claim} \noindent {\bf Proof of Claim~\ref{C- surjects}.} This is an argument based on \cite[Proof of Proposition~1.3]{hk:coh-cuts}, with corrections following \cite{hausmann:personal} and adjustments for integer coefficients. Consider long exact sequence in homology with integer coefficients of the pair $(C_-,\mathcal{B})$ \begin{eqnarray}\label{eq:LES homology} \xymatrix{ \cdots \ar[r] & H_*(\mathcal{B}) \ar[r]^{i_*} & H_*(C_-) \ar[r]^{j_*} & H_*(C_-,\mathcal{B}) \ar[r] & \cdots, } \end{eqnarray} where $i:\mathcal{B}\hookrightarrow C_-$ is inclusion and $j:({C_-},\emptyset)\to (C_-,\mathcal{B})$ is inclusion of the pair. We may apply Poincar\'e duality to Lemma~\ref{le:magic} to establish that $i_*$ is an injection in homology with integer coefficients. Thus the long exact sequence \eqref{eq:LES homology} splits into short exact sequences. We then have a commutative diagram, with integer coefficients, \begin{eqnarray}\label{eq:comm diag PD} \begin{array}{c} \xymatrix{ & H^{*-2}(\mathcal{B}) \ar[dd]_{\cong}^{\mbox{\ding{172}}} \ar[r]^{i_!} & H^*(C_-) \ar[dd]_{\cong}^{\mbox{\ding{173}}} \ar[r]^{p_-^*} & H^*(M_-) \ar[d]_{\cong}^{\mbox{\ding{174}}} \\ & & & H_{d-*}(M_-,\mathcal{Z}) \ar[d]_{\cong}^{\mbox{\ding{175}}}\\ 0 \ar[r] & H_{d-*}(\mathcal{B}) \ar[r]^{i_*} & H_{d-*}(C_-) \ar[r]^{j_*} & H_{d-*}(C_-,\mathcal{B}) \ar[r] & 0. }\end{array} \end{eqnarray} In this diagram, the manifold $C_-$ has dimension $d$, and $\mathcal{B}$ has dimension $d-2$. The maps \ding{172} and \ding{173} are Poincar\'e duality for the manifolds $\mathcal{B}$ and $C_-$ respectively, and \ding{174} is Poincar\'e duality for the manifold $M_-$ with boundary $\mathcal{Z}$. Finally, the map \ding{175} is $(p_-)_*$ and is an isomorphism by excision. The left square commutes because it is the definition of the push-forward map $i_!$. We now check that the right square commutes. We use the fact that the Poincar\'e duality isomorphism is the cap product with the fundamental class. So we need to show that for any $a\in H^*(C_-)$, $$ (p_-)_*\big( p^*_-(a)\frown [M_-]\big) = j_*\big( a\frown [C_-]\big). $$ But now, using the properties of the cap product as developed in \cite[\S 3.3]{hatcher:AT}, we have $$ \begin{array}{rcll} (p_-)_*\big( p^*_-(a)\frown [M_-]\big) & = & a\frown (p_-)_*\big( [M_-]\big) & \comeq{by naturality of the cap product}\\ & = & a\frown j_*\big( [C_-]\big) & \comeq{because $(p_-)_*\big( [M_-]\big) =j_*\big( [C_-]\big)$}\\ & = & j_*\big(a\frown [C_-]\big) & \comeq{by relative naturality of the cap product, and $j^*(a)=a$.} \end{array} $$ \noindent Thus, the diagram commutes and we may now conclude that $p_-^*$ is a surjection. \hfill \ding{52} \medskip \noindent Finally, we turn to the relationship between the cohomology of $C$ and that of $M$. \begin{Claim}\label{M to C surjects} The quotient map $p:M\longrightarrow C$ induces a surjection in cohomology $$ p^*: H^*(C;{\mathbb{Z}})\twoheadrightarrow H^*(M;{\mathbb{Z}}). $$ \end{Claim} \noindent {\bf Proof of Claim~\ref{M to C surjects}.} We have long exact sequences in cohomology with integer coefficients for the pairs $(M,M_-)$ and $(C,C_-)$ that fit into a commutative diagram \begin{eqnarray*} \begin{array}{c} \xymatrix{ \cdots \ar[r] & H^*(C,C_-) \ar[d]_{\cong}^{\mbox{\ding{172}}} \ar[r] & H^*(C) \ar[d]^{\mbox{\ding{173}}}_{p^*} \ar[r] & H^*(C_-) \ar@{->>}[d]^{\mbox{\ding{174}}}_{p_-^*} \ar[r] & H^{*+1}(C,C_-) \ar[d]_{\cong}^{\mbox{\ding{175}}} \ar[r] & \cdots \\ \cdots \ar[r] & H^*(M,M_-) \ar[r] & H^*(M) \ar[r] & H^*(M_-) \ar[r] & H^{*+1}(M,M_-) \ar[r] & \cdots .\\ }\end{array} \end{eqnarray*} Note that the maps \ding{172} and \ding{175} are isomorphisms by excision, and the map \ding{174} is onto by Claim~\ref{C- surjects}. The Four Lemma (the ``onto" half of the Five Lemma) states that if \ding{172} and \ding{174} are onto and \ding{175} is one-to-one, then \ding{173} must be onto. We have this for each degree, completing the proof. \hfill \ding{52} \medskip Claim~\ref{cl:H*C} guarantees that the cohomology of $C$ is concentrated in even degrees. Claim~\ref{M to C surjects} tells us that $H^*(C;{\mathbb{Z}})\stackrel{p^*}{\to} H^*(M;{\mathbb{Z}})$ is surjective, and so $H^*(M;{\mathbb{Z}})$ is necessarily concentrated in even degrees. \end{proof} Next we see how the conclusion of Theorem~\ref{thm:even cohomology} can fail in the non-acyclic case. \begin{nonexample}\label{eg:toric-torus} The torus ${\mathbb{T}}^2$ is a toric origami manifold. The (toric) circle action is rotation along one of the coordinate circles. The folding hypersurface consists of two disjoint circles, as shown in Figure~\ref{fig:torictorus}. The orbit space consists of two superimposed identical intervals, glued to one another at each end. The template graph has two vertices (one for each of the intervals) connected to one another by two edges (one for the top fold facet and one for the bottom fold facet), and therefore the template is not acyclic. \begin{figure}[h] \begin{center} \includegraphics[height=1.5in]{torictorus.pdf} \caption{The moment map for $\SS^1$ acting on ${\mathbb{T}}^2$.}\label{fig:torictorus} \end{center} \end{figure} \noindent It is not hard to compute that $$ H^k({\mathbb{T}}^2;{\mathbb{Z}}) = \left\{\begin{array}{cl} {\mathbb{Z}} & k=0,2\\ {\mathbb{Z}}\oplus {\mathbb{Z}} & k=1\\ 0 & \mbox{else} \end{array}\right. , $$ and so the conclusion of Theorem~\ref{thm:even cohomology} fails. \end{nonexample} \section{Equivariant cohomology} \label{se:eq-coh} Equivariant cohomology is a generalized cohomology theory in the equivariant category. We use the Borel model to compute equivariant cohomology. For the torus ${\mathbb{T}}$, we let $E{\mathbb{T}}$ be a contractible space on which ${\mathbb{T}}$ acts freely. Explicitly, for a circle, we may choose $E\SS^1$ to be the unit sphere $\SS^\infty$ in a Banach space. This is well-known to be contractible. Since ${\mathbb{T}}=\SS^1\times\cdots\times\SS^1$ is a product, we may let $E{\mathbb{T}}$ be a product of infinite-dimensional spheres. For any ${\mathbb{T}}$-space $X$, the diagonal action of ${\mathbb{T}}$ on $X\times E{\mathbb{T}}$ is free, and $$ X_{{\mathbb{T}}} = (X\times E{\mathbb{T}})/{\mathbb{T}} $$ is the {\bf Borel mixing space} or {\bf homotopy quotient} of $X$. We define the (Borel) equivariant cohomology ring to be $$ H_{{\mathbb{T}}}^*(X;R) := H^*(X_{{\mathbb{T}}};R), $$ where $H^*(-;R)$ denotes singular cohomology with coefficients in the commutative ring $R$. Thus, when $X$ is a free ${{\mathbb{T}}}$-space, we may identify $$ H_{{\mathbb{T}}}^*(X;R) \cong H^*(X/{{\mathbb{T}}};R). $$ At the other extreme, if ${{\mathbb{T}}}$ acts trivially on $X$, then $$ H_{{\mathbb{T}}}^*(X;R) \cong H^*(X\times B{{\mathbb{T}}};R), $$ where $B{{\mathbb{T}}} = E{{\mathbb{T}}}/{{\mathbb{T}}}$ is the {\bf classifying space} of ${{\mathbb{T}}}$. Note that the cohomology of the classifying space, $H^*(B{{\mathbb{T}}};R) \cong H_{{\mathbb{T}}}^*(pt;R)$, is the equivariant cohomology ring of a point. For any ${\mathbb{T}}$-space $X$, we have the fibration \begin{equation}\label{eq:fib} X \hookrightarrow X_{{\mathbb{T}}} \longrightarrow B{\mathbb{T}}. \end{equation} The projection $X_{{\mathbb{T}}} \longrightarrow B{\mathbb{T}}$ induces the map $ H_{{\mathbb{T}}}^*(pt;R) \to H_{{\mathbb{T}}}^*(X;R)$, which turns $H_{{\mathbb{T}}}^*(X;R)$ into an $H_{{\mathbb{T}}}^*(pt;R)$-module. Natural maps in equivariant cohomology preserve this module structure. A common tool in the computation of equivariant cohomology is the Serre spectral sequence applied to the fibration \eqref{eq:fib}. This has $E_2$-page $$ E_2^{p,q} = H^p(B{\mathbb{T}};H^q(X;R)). $$ This spectral sequence converges to $H_{{\mathbb{T}}}^*(X;R)$. When $X$ has cohomology concentrated in even degrees, then this spectral sequence is $0$ in every other row and column, and automatically collapses. In particular, the equivariant cohomology is also concentrated in even degrees. Goresky, Kottwitz and MacPherson call a ${\mathbb{T}}$-space $X$ {\bf equivariantly formal} if the Serre spectral sequence collapses at the $E^2$-page \cite{GKM}. This spectral sequence does collapse for a compact toric origami manifold with acyclic origami template and co\"orientable folding hypersurface, because the cohomology is concentrated in even degrees (Theorem~\ref{thm:even cohomology}). Historically, the term ``formal'' has been used in rational homotopy theory, and so equivariantly formal has multiple interpretations. Scull describes the relationships between these interpretations \cite{scull}. To avoid further confusion, we will not use this term in the remainder of this paper. Suppose that a torus ${\mathbb{T}}$ acts on a compact manifold $M$. Then the inclusion of the fixed points $I:M^{{\mathbb{T}}}\to M$ induces a map in equivariant cohomology, \begin{equation}\label{eq:fixed-pts} I^*:H_{{\mathbb{T}}}^*(M;R)\to H_{{\mathbb{T}}}^*(M^{{\mathbb{T}}};R). \end{equation} A classical result of Borel establishes that the kernel and cokernel of $I^*$ are torsion submodules \cite{Bo:transf}. Our first step is to prove that in our set-up, $I^*$ is injective. We can deduce this in a variety of ways. We supply a constructive proof here that we hope adds geometric intuition in the origami setting. \begin{theorem}\label{thm:origami inj} Let ${\mathbb{T}}\mathbin{\raisebox{-.5pt}{\reflectbox{\begin{sideways M$ be a compact toric origami with acyclic origami template and co\"orientable folding hypersurface. Then the inclusion $I:M^{\mathbb{T}}\hookrightarrow M$ induces an injection in equivariant cohomology $$ I^*:H_{\mathbb{T}}^*(M;{\mathbb{Z}})\to H_{\mathbb{T}}^*(M^{\mathbb{T}};{\mathbb{Z}}). $$ \end{theorem} \begin{proof} We proceed by induction on the number of vertices in the template graph. \vskip 0.1in \noindent {\bf Base Case}: Suppose the template graph has a single vertex. Then $M$ is a toric symplectic manifold. In particular, $M$ is K\"ahler and has isolated fixed points. Frankel showed that $H^*(M;{\mathbb{Z}})$ is torsion free in this situation \cite[Corollary~2]{frankel}. The Serre spectral sequence then has no torsion at the $E_2$ page, where it collapses, so we may conclude that $H^*_{\mathbb{T}}(M;{\mathbb{Z}})$ is torsion free. As the fixed points are isolated, $H_{\mathbb{T}}^*(M^{\mathbb{T}};{\mathbb{Z}})$ is also torsion free, and so Borel's classical result now implies injectivity. \vskip 0.1in \noindent {\bf Inductive Step}: We now assume that the statement holds for any acyclic toric origami manifold with co\"orientable fold with at most $(n-1)$ vertices in its template graph. As in the previous section, we choose a leaf of the origami template, and let $\mathcal{Z}$ be the connected component of the folding hypersurface that corresponds to the facet separating the leaf from the rest of the origami template. We continue to use the auxiliary spaces $M_-$, $M_+$, $C_-$, $C_+$, $C$ and $\mathcal{B}$ as listed in Table~\ref{table:notation}. \begin{Claim}\label{GKM for C} The inclusion $C^{\mathbb{T}} \to C$ induces an injection $$ H_{\mathbb{T}}^*(C;{\mathbb{Z}})\to H_{\mathbb{T}}^*(C^{\mathbb{T}};{\mathbb{Z}}). $$ \end{Claim} \vskip 0.1in \noindent {\bf Proof of Claim~\ref{GKM for C}.} We note that $C_-$ is a toric symplectic manifold, and $C_+$ is a toric origami manifold with fewer vertices in its template graph. Thus, in equivariant cohomology with integer coefficients, $$ H_{\mathbb{T}}^*(C_-)\stackrel{I_-^*}{\to} H_{\mathbb{T}}^*(C^{\mathbb{T}}_-) \ \mbox{ and } \ H_{\mathbb{T}}^*(C_+)\stackrel{I_+^*}{\longrightarrow} H_{\mathbb{T}}^*(C^{\mathbb{T}}_+) $$ are both injective. We now consider the equivariant Mayer-Vietoris long exact sequence for ${\mathbb{T}}$-invariant neighborhoods of $C=C_+\cup C_-$. The spaces $C$, $C_+$, $C_-$ and $\mathcal{B}$ each have ordinary cohomology only in even degrees, and hence equivariant cohomology only in even degrees. Thus, the equivariant Mayer-Vietoris long exact sequence splits into short exact sequences. We then have a commutative diagram, with integer coefficients, \begin{eqnarray*} \begin{array}{c} \xymatrix{ 0 \ar[r] & H^*_{\mathbb{T}}(C) \ar[r]^(0.35){\mbox{\ding{173}}} \ar[d]_{\mbox{\ding{172}}} & H_{\mathbb{T}}^*(C_+)\oplus H_{\mathbb{T}}^*(C_-) \ar[r]\ar[d]_{\mbox{\ding{174}}} & H_{\mathbb{T}}^*(\mathcal{B}) \ar[r]\ar[d] & 0\\ 0 \ar[r] & H_{\mathbb{T}}^*(C^{\mathbb{T}}) \ar[r] \ar[r]_(0.35){\mbox{\ding{175}}} & H_{\mathbb{T}}^*(C_+^{\mathbb{T}})\oplus H_{\mathbb{T}}^*(C_-^{\mathbb{T}}) \ar[r] & H_{\mathbb{T}}^*(\mathcal{B}^{\mathbb{T}}) \ar[r] & 0 } \end{array}. \end{eqnarray*} The map \ding{173} is injective because the top row is short exact. The map \ding{174} is $I_-^*\oplus I_+^*$, and is thus injective. Therefore, $\mbox{\ding{174}}\circ\mbox{\ding{173}}$ is injective. But $\mbox{\ding{174}}\circ\mbox{\ding{173}} = \mbox{\ding{175}}\circ\mbox{\ding{172}}$. Hence, \ding{172} must be injective. \hfill \ding{52} \begin{Claim}\label{even injectivity} In even degrees, the map $$ H^{2*}_{\mathbb{T}}(C,C_-) \to H_{\mathbb{T}}^{2*}(C^{\mathbb{T}},C_-^{\mathbb{T}}) $$ is injective. \end{Claim} \vskip 0.1in \noindent {\bf Proof of Claim~\ref{even injectivity}.} The pair $(C,C_-)$ is ${\mathbb{T}}$-invariant, so we consider the long exact sequence of the pair in equivariant cohomology. By Claim~\ref{cl:H*C}, the cohomology of $C$ is concentrated in even degrees. The space $C_-$ is a toric symplectic manifold, so its cohomology is also concentrated in even degrees. Thus the long exact sequence splits into a $4$-term short exact sequence. This induces a commutative diagram \begin{eqnarray*} \begin{array}{c} \xymatrix{ 0 \ar[r] & H^{2*}_{\mathbb{T}}(C,C_-) \ar[r]^(0.55){\mbox{\ding{173}}} \ar[d]_{\mbox{\ding{172}}} & H_{\mathbb{T}}^{2*}(C) \ar[r]\ar[d]_{\mbox{\ding{174}}} & H_{\mathbb{T}}^{2*}(C_-) \ar[r]\ar[d]& H^{2*+1}_{\mathbb{T}}(C,C_-) \ar[r] \ar[d] & 0 \\ 0 \ar[r] & H^{2*}_{\mathbb{T}}(C^{\mathbb{T}},C^{\mathbb{T}}_-) \ar[r]^(0.55){\mbox{\ding{175}}}& H_{\mathbb{T}}^{2*}(C^{\mathbb{T}}) \ar[r] & H_{\mathbb{T}}^{2*}(C^{\mathbb{T}}_-) \ar[r] & H^{2*+1}_{\mathbb{T}}(C^{\mathbb{T}},C^{\mathbb{T}}_-) \ar[r] & 0 \\ } \end{array}. \end{eqnarray*} The map \ding{173} is injective because the top row is exact. The map \ding{174} is injective by Claim~\ref{GKM for C}. Therefore, $\mbox{\ding{174}}\circ\mbox{\ding{173}}$ is injective. But $\mbox{\ding{174}}\circ\mbox{\ding{173}} = \mbox{\ding{175}}\circ\mbox{\ding{172}}$. Hence, \ding{172} must be injective. \hfill \ding{52} \begin{Claim}\label{cl:M_- injectivity} The inclusion $M_-^{\mathbb{T}}\hookrightarrow M_-$ induces an injection $ H_{\mathbb{T}}^*(M_-)\hookrightarrow H_{\mathbb{T}}^*(M_-^{\mathbb{T}}). $ \end{Claim} \vskip 0.1in \noindent {\bf Proof of Claim~\ref{cl:M_- injectivity}.} Recall that $C_-$ is a toric symplectic manifold. Let $f : C_- \to {\mathbb{R}}$ be the component of its moment map that attains its maximum value on $\mathcal{B}$. Let $f(\mathcal{B})=b\in {\mathbb{R}}$. Let $g: M_-\to {\mathbb{R}}$ be the composition $M_- \stackrel{p_-}{\to} C_-\stackrel{f}{\to} {\mathbb{R}}$. Choose $\varepsilon>0$ such that there is no critical value in between $b-\varepsilon$ and $b$, and so that $g^{-1}((b-\varepsilon, b])$ is contained in the intersection of $M_-$ with a Moser neighborhood of $Z$ in $M$. The fact that $f$ is a Morse-Bott function on $C_-$ with no critical values between $b-\varepsilon$ and $b$ guarantees that $f^{-1}((-\infty, b))$ and $f^{-1}((-\infty, b-\frac{\varepsilon}{2}])$ are homotopy equivalent. In addition, the fact that $g^{-1}((b-\varepsilon, b])$ is contained in the intersection of $M_-$ with a Moser neighborhood of $\mathcal{Z}$ in $M$ guarantees that $f^{-1}((-\infty, b-\frac{\varepsilon}{2}])$ is homotopy equivalent to $M_-$. We now appeal to a standard argument from equivariant symplectic geometry to conclude that $$ M_-^{\mathbb{T}} = f^{-1}\left(\left(-\infty, b-\frac{\varepsilon}{2}\right]\right)^{\mathbb{T}} \hookrightarrow f^{-1}\left(\left(-\infty, b-\frac{\varepsilon}{2}\right]\right) \simeq M_- $$ induces an injection in equivariant cohomology. This is an inductive argument on the critical set of $f$, and can be copied verbatim from the proof of \cite[Theorem 2]{TW:hamTsp}. \hfill \ding{52} \vskip 0.1in We now consider the long exact sequence in equivariant cohomology for the pair $(M,M_-)$. We have shown that $M_-$ and $M$ have cohomology and thus equivariant cohomology concentrated in even degrees. Thus the long exact sequence splits into a $4$-term short exact sequence. This induces a commutative diagram \begin{eqnarray*} \begin{array}{c} \xymatrix{ 0 \ar[r]\ar[d]^{\mbox{\ding{172}}} & H_{\mathbb{T}}^{2*}(M,M_-) \ar[d]^{\mbox{\ding{173}}} \ar[r] & H_{\mathbb{T}}^{2*}(M) \ar[d]^{\mbox{\ding{174}}} \ar[r] & H_{\mathbb{T}}^{2*}(M_-) \ar[d]^{\mbox{\ding{175}}} \ar[r] & H_{\mathbb{T}}^{2*+1}(M,M_-) \ar[r]\ar[d] & 0\\ 0 \ar[r] & H_{\mathbb{T}}^{2*}(M^{\mathbb{T}},M^{\mathbb{T}}_-) \ar[r] & H_{\mathbb{T}}^{2*}(M^{\mathbb{T}}) \ar[r] & H_{\mathbb{T}}^{2*}(M^{\mathbb{T}}_-) \ar[r] & H_{\mathbb{T}}^{2*+1}(M,M_-) \ar[r] & 0\\ }\end{array}. \end{eqnarray*} We want to show that \ding{174} is injective. The Four Lemma (the ``injectivity" half of the Five Lemma) states that if \ding{173} and \ding{175} are injective and \ding{172} is surjective, then \ding{174} must be injective. We first note that $H_{\mathbb{T}}^{*}(M,M_-)\cong H_{\mathbb{T}}^{*}(C,C_-)$, and $H_{\mathbb{T}}^{*}(M^{\mathbb{T}},M^{\mathbb{T}}_-)=H_{\mathbb{T}}^{*}(C^{\mathbb{T}},C^{\mathbb{T}}_-)$. Thus, the map \ding{173} is injective (in even degrees) by Claim~\ref{even injectivity}. The map \ding{175} is injective by Claim~\ref{cl:M_- injectivity}. The map \ding{172} is obviously surjective. Thus, by the Four Lemma, the map \ding{174} must be injective, as desired. \end{proof} \begin{remark} We may also derive Theorem~\ref{thm:origami inj} from work of Franz and Puppe \cite{franz-puppe}. We describe this approach, and its further applications, in the proof of Theorem~\ref{thm:origami GKM} below. \end{remark} We now identify the image of $I^*$. Goresky, Kottwitz, and MacPherson proved that the equivariant cohomology of certain spaces may be described combinatorially as $n$-tuples of polynomials with divisibility conditions on pairs of the polynomials \cite[Theorem 1.22]{GKM}. The description applies, for example, to toric varieties \cite[\S 2.2]{brion}, hypertoric varieties \cite[Proposition~3.2]{haho}, and coadjoint orbits \cite[\S 7.8]{GKM}. In this section, we prove that the description also applies to any compact toric origami manifold with acyclic origami template and co\"orientable folding hypersurface. We begin by recalling the assumptions and results from \cite{GKM}. The two key assumptions are \begin{enumerate} \item[(A)] The fixed point set $M^{{\mathbb{T}}}$ consists of isolated points; and \item[(B)] The {\bf one-skeleton} $M_{1} = \{ p\in M \ | \ \dim({\mathbb{T}}\cdot p)\leq 1\}$ is $2$-dimensional. \end{enumerate} The first assumption simplifies what $H_{{\mathbb{T}}}^*(M^{{\mathbb{T}}};{\mathbb{Z}})$ can be. When the fixed point set consists of isolated points, this ring is a direct product of copies of $$ H_{\mathbb{T}}^*(pt;{\mathbb{Z}})\cong {\mathbb{Z}}[x_1,\dots,x_n], $$ one for each fixed point. Thus, every class can be represented as a tuple of polynomials, and the ring structure is the component-wise product of polynomials. When $M$ is a compact Hamiltonian ${\mathbb{T}}$-space, the second assumption ensures that the one-skeleton must consist of $2$-spheres intersecting one another at the isolated fixed points. Moreover, the ${\mathbb{T}}$-action preserves $M_1$, and the action rotates each $\SS^2$ about an axis. The image of $M_1$ under the moment map is an immersed graph $\Phi(M_1) = \Gamma$ called the {\bf moment graph}\footnote{ \, The moment graph $\Gamma$ is sometimes called the {\bf GKM graph}. It is not the template graph.} whose vertices correspond to the fixed points $M^{{\mathbb{T}}}$ and whose edges correspond to the embedded $\SS^2$'s. Each edge $e$ in $\Gamma$ is labeled by the weight \footnote{ \, This is well-defined up to a sign, which is sufficient for our purposes.} $\alpha_e\in\mathfrak{t}^*$ by which ${\mathbb{T}}$ acts on $e$. Indeed, the moment map sends the corresponding $\SS^2$ to a line segment parallel to the weight $\alpha_e$. The embedding of the graph $\Gamma$ encodes in this way the isotropy data, denoted $\alpha$. In this framework, we have the following description of $H_{{\mathbb{T}}}^*(M;{\mathbb{Q}})$. \begin{theorem}[Goresky-Kottwitz-MacPherson \cite{GKM}]\label{thm:gkm} Suppose $M$ is a compact Hamiltonian ${\mathbb{T}}$-space satisfying conditions {\rm (A)} and {\rm (B)} above. Then $I^*$ is injective $$ I^*:H_{{\mathbb{T}}}^*(M;{\mathbb{Q}})\hookrightarrow H_{{\mathbb{T}}}^*(M^{{\mathbb{T}}};{\mathbb{Q}})\cong \bigoplus_{p\in M^{{\mathbb{T}}}} H_{{\mathbb{T}}}^*(pt;{\mathbb{Q}}), $$ and its image consists of \begin{equation}\label{eq:gkm} \Big\{ \, (f_p)\in \bigoplus_{p\in M^{{\mathbb{T}}}} H_{{\mathbb{T}}}^*(pt;{\mathbb{Q}})\ \Big|\ \alpha_e \big| (f_p-f_q) \mbox{ for each edge } e=(p,q) \mbox{ in } \Gamma\, \Big\}. \end{equation} We will refer to these divisibility conditions as the {\bf GKM description}. \end{theorem} \begin{remark} For a Hamiltonian ${\mathbb{T}}$-space, assumption (A) guarantees that $I^*$ is injective in equivariant cohomology with integer coefficients. We may strengthen assumption (B) to guarantee that the GKM description holds over ${\mathbb{Z}}$. A stronger set of assumptions are described in \cite[\S3]{HHH}; they include the existence of a cell decomposition of the manifold. In particular, for Hamiltonian ${\mathbb{T}}$-spaces with isolated points, Morse theory can be applied to a generic component of the moment map to establish that these stronger assumptions boil down to local topological properties that must be checked at the fixed points. These can then be verified for symplectic toric manifolds and for coadjoint orbits. As we have seen, the moment map for a toric origami manifold $M$ does not necessarily produce Morse functions on $M$. We do not know if there is a cell decomposition of a toric origami manifold that would allow us to apply techniques from \cite{HHH}. \end{remark} A key technical tool in the proof of Theorem~\ref{thm:gkm} is the Chang-Skjelbred Lemma \cite[Lemma~2.3]{CS:injectivity}. Let $J: M^{{\mathbb{T}}}\to M_1$ denote the inclusion of the fixed points into the one-skeleton. The Chang-Skjelbred Lemma states that $I^*(H_{{\mathbb{T}}}^*(M)) = J^*(H_{{\mathbb{T}}}^*(M_1))$. Since the one-skeleton consists of $\SS^2$'s, we must understand $H_{\mathbb{T}}^*(\SS^2)$. It is a simple calculation to check that each $\SS^2$ contributes one of the divisibility conditions in \eqref{eq:gkm}. Now let ${\mathbb{T}}\mathbin{\raisebox{-.5pt}{\reflectbox{\begin{sideways M$ be a compact toric origami manifold with acyclic origami template and with co\"orientable folding hypersurface. The fixed points $M^{\mathbb{T}}$ correspond to the $0$-dimensional faces of the orbit space $M/{\mathbb{T}}$. Just as for toric symplectic manifolds, these are isolated fixed points. The one-skeleton corresponds to the (possibly folded) edges ($1$-dimensional faces) of the orbit space. These are the $1$-dimensional faces of the polytopes of the symplectic cut pieces that are not entirely contained in a fold. The corresponding subsets of $M$ are symplectic or origami $2$-spheres. Therefore the one-skeleton is $2$-dimensional. An example is shown in Figure~\ref{fig:folded-hirz}. \begin{figure}[h] \begin{center} \includegraphics[width=5.75in]{cropped-folded-hirzebruch.pdf} \caption{ The orbit space and the GKM graph for a toric origami structure on the Hirzebruch surface. The GKM graph has four vertices and four edges, two of which are folded. }\label{fig:folded-hirz} \end{center} \end{figure} Thus, assumptions (A) and (B) are satisfied in the case of toric origami manifolds, and indeed the GKM theorem generalizes to our set-up. \begin{theorem}\label{thm:origami GKM} Let ${\mathbb{T}}\mathbin{\raisebox{-.5pt}{\reflectbox{\begin{sideways M$ be a compact toric origami with acyclic origami template and co\"orientable folding hypersurface. Then $I^*$ is injective $$ I^*:H_{{\mathbb{T}}}^*(M;{\mathbb{Z}})\hookrightarrow H_{{\mathbb{T}}}^*(M^{{\mathbb{T}}};{\mathbb{Z}})\cong \bigoplus_{p\in M^{{\mathbb{T}}}} H_{{\mathbb{T}}}^*(pt;{\mathbb{Z}}), $$ and the image consists of \begin{equation}\label{eq:origami gkm} \Big\{ (f_p)\in \bigoplus_{p\in M^{{\mathbb{T}}}} H_{{\mathbb{T}}}^*(pt;{\mathbb{Z}})\ \Big|\ \alpha_e \big| (f_p-f_q) \mbox{ for each edge } e=(p,q) \mbox{ in } \Gamma\Big\}, \end{equation} where $\alpha_e$ is the weight of the action ${\mathbb{T}}\mathbin{\raisebox{-.5pt}{\reflectbox{\begin{sideways \SS^2_e$ on the $2$-sphere corresponding to $e$. \end{theorem} \begin{proof} In Theorem~\ref{thm:origami inj}, we have established that $I^*$ is injective (over ${\mathbb{Z}}$). This can also be derived from an algebraic result of Franz and Puppe. In \cite[Theorem~1.1]{franz-puppe}, for a ${\mathbb{T}}$-space $X$ with connected stabilizers, they show that five conditions are equivalent. Their condition (ii) is that the Serre spectral sequence collapses at the $E_2$-page. Their condition (v) gives a long exact sequence. A consequence of the origami template classification of toric origami manifolds is that the stabilizer of a point is a connected subtorus of ${\mathbb{T}}$. Thus, we may appeal to Franz and Puppe's theorem. Our Theorem~\ref{thm:even cohomology} implies that the Serre spectral sequence collapses at the $E_2$-page, assertion (ii) in \cite[Theorem~1.1]{franz-puppe}. This is then equivalent to assertion (v) which gives a long exact sequence, the first few terms of which are $$ 0 \to H^*_{\mathbb{T}}(M;{\mathbb{Z}}) \stackrel{\mbox{\ding{172}}}{\to} H^*_{\mathbb{T}}(M_0;{\mathbb{Z}}) \stackrel{\mbox{\ding{173}}}{\to} H^{*+1}_{\mathbb{T}}(M_1, M_0;{\mathbb{Z}}). $$ The content of our Theorem~\ref{thm:origami inj} is that \ding{172} (which is $I^*$) is injective. That the sequence is exact then means that the image of \ding{172} is equal to the kernel of \ding{173}. The map \ding{173} is the boundary map in the long exact sequence of the pair $(M_1,M_0)$. Thus we have $$ \cdots \to H_{\mathbb{T}}^*(M_1^*;{\mathbb{Z}}) \stackrel{\mbox{\ding{174}}}{\to} H_{\mathbb{T}}^*(M_0;{\mathbb{Z}}) \stackrel{\mbox{\ding{173}}}{\to} H_{\mathbb{T}}^{*+1}(M_1,M_0;{\mathbb{Z}})\to\cdots. $$ The kernel of \ding{173} is then equal to the image of \ding{174}, which is the image of the equivariant cohomology of the one-skeleton in $H_{\mathbb{T}}^*(M_0;{\mathbb{Z}})$. The fact that the one-skeleton consists of symplectic and origami $2$-spheres means that each $\SS^2$ contributes one of the divisibility conditions in \eqref{eq:origami gkm}. \end{proof} In Section~\ref{se:even}, we proved that $H^*(M;{\mathbb{Z}})$ is concentrated in even degrees. We do not have a Morse function on $M$ that would allow us to compute the ranks of these cohomology groups. With our explicit description of $H_{\mathbb{T}}^*(M;{\mathbb{Z}})$, it is possible in examples to determine the ranks and ring structure of $H^*(M;{\mathbb{Z}})$. This is a consequence of the collapse of the Serre spectral sequence, which implies that $$ H^*(M;{\mathbb{Z}}) \cong H_{\mathbb{T}}^*(M;{\mathbb{Z}})\otimes_{H^*_{\mathbb{T}}(pt;{\mathbb{Z}})} {\mathbb{Z}}. $$ \begin{example}\label{eg:sphere} The $2n$-sphere $\SS^{2n}$ may be endowed with toric origami structure whose template graph has two vertices and a single edge between them. Each of the two vertices maps to a the $n$-simplex in ${\mathbb{R}}^n$ with an orthogonal corner at the origin; that is, a simplex with vertices the origin and the standard basis vectors $e_i=(0,\dots,1,\dots,0)$ with a single $1$ in the $i^{th}$ co\"ordinate and $0$s elsewhere. The edge maps to the fold facet by which these two polytopes are glued together: the $(n-1)$-simplex with vertices the $e_i$, opposite the origin. The orbit space for $\SS^4$ is shown in Figure~\ref{fig:S4}. Thus the toric action has $2$ fixed points, which we denote $N$ and $S$ (for the north and south poles). There are $n$ edges in the GKM graph, each joining $\Phi(N)$ and $\Phi(S)$. We can identify $H_{\mathbb{T}}^*(pt;{\mathbb{Z}}) ={\mathbb{Z}}[x_1,\dots,x_n]$. From the representation of the orbit space in ${\mathbb{R}}^n$ we can see that the ${\mathbb{T}}$-action on the sphere mapping to the $i^{th}$ co\"ordinate line in ${\mathbb{R}}^{n}$ has weight $x_i$. Theorem~\ref{thm:origami GKM} states that $$ I^*(H_{\mathbb{T}}^*(\SS^{2n};{\mathbb{Z}})) = \Big\{ \, (f_N,f_S)\in {\mathbb{Z}}[x_1,\dots,x_n]\oplus{\mathbb{Z}}[x_1,\dots,x_n]\, \Big| \, x_i\big| (f_N-f_S) \, \mbox{ for } i=1,\dots,n\,\,\Big\}. $$ From this, we can find a module basis (for $H_{\mathbb{T}}^*(\SS^{2n};{\mathbb{Z}})$ as an $H_{\mathbb{T}}^*(pt;{\mathbb{Z}})$-module) with two elements $$ I^*(\mathds{1}) = (1,1) \mbox{ and } I^*(\pi) = (x_1\cdots x_n,0), $$ where $\mathds{1}\in H_{\mathbb{T}}^0(\SS^{2n};{\mathbb{Z}})$ and $\pi\in H_{\mathbb{T}}^{2n}(\SS^{2n};{\mathbb{Z}})$. \end{example} \begin{nonexample} We revisit Nonexample \ref{eg:toric-torus}, of a toric circle action on a torus. The circle action is free, and so has no fixed points. Nevertheless, we may compute $$ H^k_{\SS^1}({\mathbb{T}}^2;{\mathbb{Z}})= H^k({\mathbb{T}}^2/\SS^1;{\mathbb{Z}}) = H^k(\SS^1) = \left\{\begin{array}{cl} {\mathbb{Z}} & k=0,1\\ 0 & \mbox{else} \end{array}\right. . $$ In particular, the conclusion of Theorem~\ref{thm:origami inj} cannot hold. \end{nonexample} \section{Toric origami manifolds are locally standard}\label{sec:std} Toric topology is the study of topological analogues of toric symplectic manifolds and toric varieties. The symplectic or algebraic structure is dropped, and the focus is the existence of an effective smooth action of a torus half the dimension of the manifold. Examples of such topological analogues, from most restrictive to most general, are toric manifolds \cite{davisjanus} (referred to by some authors as quasitoric manifolds), topological toric manifolds \cite{ishidafukukawamasuda} and torus manifolds \cite{masu99}. We now show that acyclic toric origami manifolds fit into the framework of torus manifolds, and that Theorem~\ref{thm:even cohomology} also follows from the work of Masuda and Panov on the cohomology of torus manifolds \cite{masuda-panov}. Their theory is more general and their proofs algebraic. A {\bf torus manifold} is a $2n$-dimensional closed connected orientable smooth manifold $M$ with an effective smooth action of an $n$-dimensional torus ${\mathbb{T}}^n$ with non-empty fixed set. A torus manifold $M$ is said to be {\bf locally standard} if every point in $M$ has an invariant neighbourhood $U$ weakly equivariantly diffeomorphic to an open subset $W\subset{\mathbb{C}}^n$ invariant under the standard ${\mathbb{T}}^n$-action on ${\mathbb{C}}^n$. The adverb `weakly' means that there is an automorphism $\rho\colon {\mathbb{T}}\to {\mathbb{T}}$ and a diffeomorphism $f\colon U\to W$ such that $$f(ty)=\rho(t)f(y)$$ for all $t\in {\mathbb{T}}$, \ $y\in U$. Compact symplectic toric manifolds are locally standard \cite[Proof of Lemme 2.4]{de:hamiltoniens}. Next we will prove that toric origami manifolds with co\"orientable folding hypersurface are also locally standard. Toric origami manifolds with non-co\"orientable components of the fold are not locally standard. Indeed, an invariant neighborhood of a point on a non-co\"orientable component of the fold is a bundle of M\"obius bands over the corresponding connected component of $B$ \cite[Rmk.~2.26]{CGP:origami}, which is not equivariantly diffeomorphic to an invariant open subset of ${\mathbb{T}}^n\mathbin{\raisebox{-.5pt}{\reflectbox{\begin{sideways\mathbb{C}^n$. \begin{lemma}\label{le:loc-std} Suppose that $(M,Z,\omega,\Phi,{\mathbb{T}})$ is a toric origami manifold with co\"orientable folding hypersurface. Then $M$ is locally standard. \end{lemma} \begin{proof} The argument used in \cite[Proof of Lemme 2.4]{de:hamiltoniens} to prove that compact symplectic toric manifolds are locally standard does not use compacteness of the manifold, and therefore applies directly to the manifold $M\setminus Z$. Next, we check the `locally standard' condition on a point $p\in Z$ on the fold. We use a Moser model, as defined in \cite[Def.\ 2.12]{CGP:origami}, for a neighborhood of $p$. As remarked in \cite{CGP:origami}, such Moser models exist for orientable origami manifolds. What is necessary for the local existence of the Moser model near a single component of the fold is simply the co\"orientability of that piece of the fold. Thus, we may assume that $p\in Z$ has a neighborhood with a Moser model. Let $\mathcal{Z}_p$ denote the connected component of $Z$ containing $p$. The local Moser model is an equivariant diffeomorphism $$ \varphi: \mathcal{Z}_p \times (-\varepsilon,\varepsilon) \to \mathcal{U}, $$ where $\varepsilon>0$ and $\mathcal{U}$ is a tubular neighborhood of $\mathcal{Z}_p$, such that $\varphi(x,0)=x$ for all $x\in \mathcal{Z}_p$. The symplectic form can be written in these co\"ordinates, but we do not need that here. We now consider the null-fibration $\SS^1\hookrightarrow \mathcal{Z}_p\stackrel{\pi}{\longrightarrow} B_p$. This is a principal $\SS^1$-bundle, and the base space is a compact symplectic toric manifold of dimension $(2n-2)$. Let $b=\pi(p)$. Compact toric symplectic manifolds are locally standard. Choose a neighborhood $V$ of $b\in B_p$ that is weakly equivariantly diffeomorphic to an open subset $W\subset {\mathbb{C}}^{n-1}$ that is invariant with respect to the standard ${\mathbb{T}}^{n-1}$-action on ${\mathbb{C}}^{n-1}$. By possibly passing to a smaller neighborhood of $b$, we may assume that the bundle over $V$ is trivial, $V\times \SS^1\stackrel{\pi}{\longrightarrow} V$. Thus, we have an equivariant neighborhood $$ V\times \SS^1 \times (-\varepsilon,\varepsilon) $$ of $p\in \mathcal{Z}_p$. Under this identification, the action of ${\mathbb{T}}^n$ splits into the ${\mathbb{T}}^{n-1}$ action on $V$, and $\SS^1$ acting on itself by multiplication on the $\SS^1$. We may embed $\SS^1\times (-\varepsilon,\varepsilon)$ as an open annulus $A\subset {\mathbb{C}}$ by equivariant diffeomorphism. Therefore $V\times \SS^1 \times (-\varepsilon,\varepsilon)$ is weakly equivariantly diffeomorphic to an open subset $W\times A\subset {\mathbb{C}}^{n-1}\times {\mathbb{C}}$ that is invariant with respect to the co\"ordinate ${\mathbb{T}}^{n}$-action on the vector space ${\mathbb{C}}^{n}$. \footnote{\, An alternative proof for this Lemma was pointed out to us by one of the referees: it uses the fact that any submanifold of $M$ consisting of points with the same isotropy subgroup is tranverse to the folding hypersurface $Z$. This fact relies strongly on the co\"orientability hypothesis.} \end{proof} A key player in Masuda and Panov's work on torus manifolds is the orbit space $Q =M/{\mathbb{T}}$, which in the origami framework is closely related to the origami template, as explained at the end of Section~\ref{sec:background}. Masuda and Panov define the faces of the orbit space using their notion of characteristic submanifold. The orbit space is then called {\bf face-acyclic} if every face $F$ (including $Q$ itself) is {\bf acyclic}: that is, it has $\widetilde{H}^*(F)=0$. Note that the orbit space $M/{\mathbb{T}}$ deformation retracts onto the template graph: each polytope $\Psi_V(v)$ deformation retracts onto a point in its center and rays from that point to each of the fold facets of that polytope. This can be done so that when two polytopes are glued along a fold facet, the rays from the center points of the two polytopes join at the fold facet: the two rays now form a line between the center points of the two polytopes. Viewing the center points of the polytopes as vertices of a graph and the lines joining them as edges, we recover the template graph. An example is provided in Figure~\ref{fig:retracts}. \begin{figure}[ht] \centering{ \includegraphics[height=1.5in]{retracts1.pdf} \hskip 0.7in \includegraphics[height=1.6in]{retracts2.pdf}\hskip 0.1in \includegraphics[height=1.6in]{retracts3.pdf} } \caption{Each polytope deformation retracts onto a central point and rays towards the fold facets. The orbit space $M/{\mathbb{T}}$ deformation retracts onto the template graph.} \label{fig:retracts} \end{figure} \begin{theorem}\label{thm:locstd} Suppose that $(M,Z,\omega,\Phi,{\mathbb{T}})$ is a toric origami manifold such that each connected component of the folding hypersurface is co\"orientable. Then $M/{\mathbb{T}}$ is face-acyclic if and only if the origami template is acyclic. \end{theorem} \begin{proof} The orbit space $M/{\mathbb{T}}$ deformation retracts onto the template graph, and any face $F$ of $M/{\mathbb{T}}$ deformation retracts onto a subgraph: the vertices of this subgraph correpond to the polytopes $\Psi_V(v)$ which have non-empty intersection with $F$, its edges are the fold facets $\Psi_E(e)$ which have non-empty intersection with $F$. Being homotopy equivalent to a (sub)graph, a face of $M/{\mathbb{T}}$ will be acyclic if and only if that (sub)graph has no cycles, and therefore $M/{\mathbb{T}}$ is face-acyclic exactly when the template graph has no cycles. \end{proof} We now can derive our Theorem~\ref{thm:even cohomology} from Masuda and Panov's work: they prove that face-acyclic locally standard torus manifolds have no odd-degree cohomology \cite[Theorem 9.3]{masuda-panov}. While our proofs have very different flavors, it is interesting to note that a crucial ingredient in their proof is their \cite[Lemma 2.3]{masuda-panov}, which is closely related to our Lemma~\ref{le:magic}, as described in Remark~\ref{rem:to-masuda-panov}.

https://arxiv.org/abs/1407.2148

Open book decompositions versus prime factorizations of closed, oriented 3-manifolds

Let $M$ be a closed, oriented, connected 3--manifold and $(B,\pi)$ an open book decomposition on $M$ with page $\Sigma$ and monodromy $\varphi$. It is easy to see that the first Betti number of $\Sigma$ is bounded below by the number of $S^2\times S^1$--factors in the prime factorization of $M$. Our main result is that equality is realized if and only if $\varphi$ is trivial and $M$ is a connected sum of $S^2\times S^1$'s. We also give some applications of our main result, such as a new proof of the result by Birman and Menasco that if the closure of a braid with $n$ strands is the unlink with $n$ components then the braid is trivial.

\section{Introduction}\label{s:intro} An {\em abstract open book} is a pair $(\Sigma, \varphi)$, where $\Sigma$ is a connected, oriented surface with $\partial\Sigma\neq\emptyset$ and the {\em monodromy} $\varphi$ is an element of the group $\rm Diff^+(\Sigma,\partial\Sigma)$ of orientation--preserving diffeomorphisms of $\Sigma$ which restrict to the identity on a neighborhood of the boundary. We say that the monodromy $\varphi$ is {\em trivial} if it is isotopic to the identity of $\Sigma$ via diffeomorphisms which fix $\partial\Sigma$ pointwise. Let $N_{\varphi}$ denote the mapping torus \[ N_\varphi = \Sigma\times [0,1]/(p,1) \sim (\varphi(p),0). \] To the open book $(\Sigma,\varphi)$ one can associate a closed, oriented, connected 3--manifold $M_{(\Sigma,\varphi)}$ by using the natural identification of $\partial N_\varphi = \partial\Sigma\times S^1$ with the boundary of $\partial\Sigma\times D^2$: \[ M_{(\Sigma,\varphi)} := N_\varphi\cup_{\partial} \partial\Sigma\times D^2. \] The link $B:=\partial \Sigma\times \{0\}\subset M_{(\Sigma,\varphi)}$ is fibered, with fibration $\pi\thinspace\colon M_{(\Sigma,\varphi)}\setminus B\to S^1$ given by the obvious extension of the natural projection \[ N_\varphi = \Sigma\times [0,1]/(p,1) \sim (\varphi(p),0)\to S^1=[0,1]/1\sim 0 \] and monodromy equal to $\varphi$. In other words, the pair $(B,\pi)$ is an {\em open book decomposition} of $M=M_{(\Sigma,h)}$ with {\em binding} $B$, pages $\Sigma_\th:= \overline{\pi^{-1}(\th)}$, $\th\in S^1$ and monodromy $\varphi$. We will always identify $N_{\varphi}$ with the complement of a tubular neighborhood of $B$ in $M$. If $(B,\pi)$ is an open book decomposition of $M$ with page $\Sigma$, it is easy to see that $M$ has a Heegaard splitting of genus $b_1(\Sigma)$. Since $M$ is obtained from each handlebody of the splitting by attaching 2--disks and 3--balls, this immediately implies the inequality \begin{equation}\label{e:b1ineq} b_1(M)\leq b_1(\Sigma). \end{equation} We will provide a refinement of Inequality~\eqref{e:b1ineq} with Proposition~\ref{p:intersection}. The following theorem is our main result. Its proof is based on well--known results due to Reidemeister~\cite{Re33}, Singer~\cite{Si33} and Haken~\cite{Ha68} (see Section~\ref{s:proof}). Recall that each closed, oriented, connected 3--manifold $M$ has a prime factorization, unique up to order of the factors, of the form \begin{equation}\label{e:factor} M = M_1\#\cdots\# M_h\# S^2\times S^1\#\stackrel{(k)}{\cdots}\# S^2\times S^1, \end{equation} where each $M_i$ is irreducible (see~e.g.~\cite{He76}). \begin{thm}\label{t:main} Let $(B,\pi)$ be an open book decomposition of a closed, oriented, connected 3--manifold $M$ with page $\Sigma$ and monodromy $\varphi$. Then, $b_1(\Sigma)$ is equal to the number of $S^2\times S^1$--factors in the prime factorization of $M$ if and only if $\varphi$ is trivial and $M$ is a connected sum of $S^2\times S^1$'s. \end{thm} Theorem~\ref{t:main} immediately implies the following corollary, which is also proved in~\cite[Proof of Theorem~1.3]{Ni10} and~\cite[Theorem~2]{GW13} using the fact that finitely generated free groups are not isomorphic to any of their nontrivial quotients. \begin{cor}\label{c:main} Any open book decomposition of $\#^k S^2\times S^1$ whose page $\Sigma$ satisfies $b_1(\Sigma)=k$ must have trivial monodromy. \end{cor} Corollary~\ref{c:main} implies Corollary~\ref{c:bir-men}, which was obtained previously by Birman--Menasco as an application of their braid foliation techniques~\cite[Theorem~1]{BM92}. Grigsby and Wehrli gave two further proofs of Corollary~\ref{c:bir-men}, one using the fact that finitely generated free groups are not isomorphic to any of their nontrivial quotients, and the other using Khovanov homology~\cite{GW13}. \begin{cor}\label{c:bir-men} Let $b \in B_n$ be a braid on $n$ strands such that its closure $\hat b$ is the trivial link $U_n$ with $n$ components. Then, $b$ is the identity. \end{cor} \begin{proof} Put $\hat b$ in braid form with respect to the binding of the trivial open book decomposition of $S^3$ and consider the two--fold branched cover $\Sigma(\hat b)$ along $\hat b$. Then, \[ \Sigma(\hat{b}) = \Sigma(U_n) = \#^{n-1} S^2 \times S^1. \] Pulling back the trivial open book of $S^3$ to $\Sigma(\hat{b})$ we obtain an open book decomposition of $ \#^{n-1} S^2 \times S^1$, whose page is a surface $\Sigma$ with $b_1(\Sigma)=n-1$, which we view as a 2--fold branched cover of the disk with $n$ branch points. Under the identificaton of $B_n$ with the subgroup of the mapping class group of $\Sigma$ given by the elements commuting with the covering involution~\cite{BH73}, the monodromy of the open book is equal to $b$. By Corollary~\ref{c:main}, the braid $b$ must be the identity in $B_n$. \end{proof} Let $\Sigma$ and $\Sigma'$ be two orientable surfaces. By performing a boundary connected sum between them we obtain a surface $\Sigma\natural\Sigma'$. If $\varphi$ is a diffeomorphism of $\Sigma$, $\psi$ is a diffeomorphism of $\Sigma'$ and both $\varphi$ and $\psi$ are the identity on a neighborhood of the boundary, we can form a diffeomorphism $\varphi \natural \psi$ of $\Sigma\natural\Sigma'$. This geometric operation yields a homomorphism \[ \Gamma_\Sigma \times \Gamma_{\Sigma'} \to \Gamma_{\Sigma\natural\Sigma'}, \] which we will call {\em boundary connected sum homomorphism.} A combination of Inequality~\eqref{e:b1ineq} with Corollary~\ref{c:main} yields the following Corollary~\ref{c:injectivity}, which can also be proved e.g.~applying~\cite[Corollary~4.2 (iii)]{PR00}. \begin{cor}\label{c:injectivity} Let $\Gamma_\Sigma$ be the mapping class group of the orientable surface $\Sigma$. Then, the boundary connected sum homomorphism \[ \Gamma_\Sigma \times \Gamma_{\Sigma'} \to \Gamma_{\Sigma\natural\Sigma'} \] is injective. \end{cor} \begin{proof} Under the map $(\Sigma,\varphi)\to M_{(\Sigma,\varphi)}$ described above, boundary connected sum of abstract open books corresponds to connected sum of $3$-manifolds: \[ M_{(\Sigma\natural\Sigma', \varphi \natural \psi)} = M_{(\Sigma,\varphi)} \# M_{(\Sigma',\psi)}. \] Observe that $b_1(\Sigma\natural\Sigma') = b_1(\Sigma) + b_1(\Sigma')$. Therefore, if $\varphi \natural \psi$ is isotopic to the identity relative to the boundary then $M_{(\Sigma\natural\Sigma', \varphi \natural \psi)}$ is diffeomorphic to $\#^{b_1(\Sigma) + b_1(\Sigma')} S^2 \times S^1$. The uniqueness of the prime factorization for $3$--manifolds~\cite{He76} implies that $M_{(\Sigma,\varphi)} = \#^k S^2 \times S^1$ and $M_{(\Sigma',\psi)} = \#^l S^2 \times S^1$ for some non--negative integers $k, l$ such that $k+l = b_1(\Sigma)+b_1(\Sigma')$. By Inequality~\eqref{e:b1ineq} we have $k \le b_1(\Sigma)$ and $l \le b_1(\Sigma')$, which forces $k=b_1(\Sigma)$ and $l=b_1(\Sigma')$ as the only possibility. Corollary~\ref{c:main} implies that $\varphi$ and $\psi$ are isotopic to the identity. \end{proof} The rest of the paper is organized as follows. In Section~\ref{s:further} we recall two well known results independent of Theorem~\ref{t:main}, i.e.~Propositions~\ref{p:spheres-B} and~\ref{p:intersection}. Proposition~\ref{p:spheres-B} shows that any embedded 2--sphere disjoint from the binding of an open book decomposition is homologically trivial. Proposition~\ref{p:intersection} is a refinement of Inequality~\eqref{e:b1ineq} and can be viewed as saying that the homology of a closed, oriented, connected 3--manifold $M$ puts homological constrains on the monodromy of any open book decomposition of $M$. In Section~\ref{s:proof} we prove Theorem~\ref{t:main}. {\bf Acknowledgements:} the authors wish to thank the anonymous referees for valuable comments. The present work is part of the authors' activities within CAST, a Research Network Program of the European Science Foundation. The first author was partially supported by the ERC grant ``Geodycon''. The second author was partially supported by the PRIN--MIUR research project 2010--2011 ``Variet\`a reali e complesse: geometria, topologia e analisi armonica''. \section{Non--separating 2--spheres and a refinement of Inequality~\eqref{e:b1ineq}}\label{s:further} Given a closed, oriented, connected 3--manifold $M$ endowed with an open book decomposition $(B,\pi)$ and having a prime factorization as in~\eqref{e:factor}, one of the first questions one could ask is how a non--separating 2--sphere $S$ in $M$ can be positioned with respect to the binding $B$. Since $B$ is homologically trivial in $M$, the following proposition implies that, possibly after a small isotopy, each such $S$ must intersect $B$ transversally at least twice. \begin{prop} \label{p:spheres-B} Let $(B,\pi)$ be an open book decomposition with page $\Sigma$ and monodromy $\varphi$ of a closed, oriented, connected 3--manifold $M$. Then, each embedded 2--sphere $S \subset M\setminus B$ bounds an embedded ball in $M \setminus B$ and, in particular, is homologically trivial in $M$. \end{prop} \begin{proof} Recall that $M = N_{\varphi} \cup V$, where $V$ is a tubular neighborhood of the binding. Up to an isotopy of $S$, we can assume $S \subset N_\varphi$. The universal cover of $N_\varphi$ is homeomorphic to $\mathbb R^3$ and from this the triviality of $[S]$ in $H_2(M \setminus B)$, and therefore in $H_2(M)$, follows immediately. In order to prove that $S$ bounds a ball in $M \setminus B$ we need to use some basic results in three--dimensional topology. In fact $\mathbb R^3$ is irreducible~\cite[Theorem~1.1]{Ha00} and this implies~\cite[Proposition~1.6]{Ha00} that $N_\varphi$ is also irreducible, therefore $S$ bounds an embedded ball in $N_\varphi$. \end{proof} We now establish a result which refines Inequality~\eqref{e:b1ineq}. Proposition~\ref{p:intersection} below can be viewed as saying that the homology of a closed, oriented, connected 3--manifold $M$ puts homological constraints on the monodromy of any open book decomposition of $M$. For the rest of this section all homology groups will be taken with coefficients in the field $\mathbb Q$ of rational numbers unless specified otherwise. Let $H_1(\Sigma, \partial \Sigma)^{\varphi}$ denote the subspace of $H_1(\Sigma,\partial\Sigma)$ consisting of the elements fixed by the map \[ \varphi_* \colon H_1(\Sigma, \partial \Sigma) \to H_1(\Sigma, \partial \Sigma) \] induced by the monodromy $\varphi\thinspace\colon\Sigma\to\Sigma$. \begin{prop}\label{p:intersection} Let $(B,\pi)$ be an open book decomposition with page $\Sigma$ and monodromy $\varphi$ of a closed, oriented, connected 3--manifold $M$. Then, \[ b_1(M) = \dim_\mathbb Q H_1(\Sigma,\partial\Sigma)^\varphi. \] More precisely, there is an isomorphism $H_2(M)\cong H_1(\Sigma, \partial \Sigma)^\varphi$ induced by a well--defined map $H_2(M;\Z) \to H_1(\Sigma, \partial \Sigma;\Z)^\varphi$ given by $\alpha\mapsto [F\cap\Sigma]$, where $F\subset M$ is any closed, oriented and properly embedded surface which represents $\alpha$ and intersects the page $\Sigma \times \{ 0 \}$ transversally. \end{prop} \begin{proof} We can view $N_{\varphi}$ as the union of $\Sigma \times [0, 1/2]$ and $\Sigma \times [1/2, 1]$ with $(x, 1)$ identified to $(\varphi(x), 0)$. Using the fact that $\Sigma$ times an interval is homotopically equivalent to $\Sigma$, the (relative) Mayer--Vietoris sequence for this splitting gives the following exact sequence: \[ H_2(\Sigma, \partial \Sigma)^2 \stackrel{f_1} \longrightarrow H_2(N_{\varphi}, \partial N_{\varphi}) \stackrel{f_2} \longrightarrow H_1(\Sigma, \partial \Sigma)^2 \stackrel{f_3} \longrightarrow H_1(\Sigma, \partial \Sigma)^2. \] The map $f_3$ is given by the matrix \[ \left ( \begin{matrix} Id & Id \\ \varphi_* & Id \end{matrix} \right )\in M_2(\End(H_1(\Sigma,\partial\Sigma))). \] This immediately implies that the image of $f_2$ is isomorphic to $H_1(\Sigma, \partial \Sigma)^{\varphi}$. Recall the decomposition $M = N_{\varphi} \cup V$, where $V$ is a tubular neighborhood of the binding. Since $H_2(V)=\{0\}$, the homology exact sequence for the pair $(M, V)$ implies that the map $g\thinspace\colon H_2(M) \to H_2(M, V)$ induced by the inclusion map is injective. On the other hand, by excision the inclusion $N_\varphi \subset M$ induces an isomorphism $\psi\thinspace\colon H_2(N_{\varphi}, \partial N_{\varphi})\stackrel{\cong} \longrightarrow H_2(M, V)$. Moreover, it is easy to see that the image of the map $\psi\circ f_1$ maps injectively to $H_1(V)$ under the next map $\delta\thinspace\colon H_2(M,V)\to H_1(V)$ in the exact sequence of the pair, while the image of $g$ maps trivially. This shows that the images of $f_1$ and of $\psi^{-1}\circ g$ have trivial intersection. Therefore the composition $f_2\circ\psi^{-1}\circ g$ sends $H_2(M)$ injectively into the image of the map $f_2$, which, as we have just shown, is isomorphic to $H_2(\Sigma, \partial \Sigma)^{\varphi}$. We claim that $f_2\circ\psi^{-1}\circ g$ sends $H_2(M)$ also surjectively onto the image of $f_2$. In order to verify this, we argue by induction. Assume first that $\partial\Sigma$ is connected. In this situation the map $\delta\circ\psi\circ f_1$ is clearly surjective. Therefore, if $x\in H_2(N_\varphi, \partial N_\varphi)$ with $f_2(x)\neq 0$, there exists $y\in H_2(\Sigma, \partial \Sigma)^2$ with $\delta\circ\psi\circ f_1(y) = \delta\circ\psi(x)$. It follows that setting $x' = x - f_1(y)$ we have $f_2(x') = f_2(x)$ and $\delta\circ\psi(x') = 0$; therefore $x'$ is in the image of $\psi^{-1}\circ g$, and the claim is proved when $\partial\Sigma$ is connected. Now assume $\partial\Sigma$ is disconnected and denote by $|\partial\Sigma|$ the number of its connected components. By the inductive hypothesis we assume that the claim holds for open books with $|\partial \Sigma|-1$ binding components. Let $(\widehat\Sigma, \widehat\varphi)$ be another abstract open book, constructed as follows. The connected, oriented surface $\widehat\Sigma$ is obtained by attaching a 2--dimensional 1--handle $h$ to $\partial\Sigma$ so that $|\partial\widehat\Sigma| = |\partial\Sigma|-1$, while $\widehat\varphi$ is defined by first extending $\varphi$ as the identity over $h$, and then composing with a (positive or negative) Dehn twist along a simple closed curve in $\widehat\Sigma$ which intersects the cocore $c$ of $h$ transversely once. It is a well--known fact that the open book decomposition $(\widehat B,\widehat\pi)$ associated to $(\widehat\Sigma, \widehat\varphi)$ is obtained from the open book decomposition $(B,\pi)$ associated to $(\Sigma, \varphi)$ by plumbing with a Hopf band, and that $M_{(\widehat\Sigma,\widehat\varphi)}$ is diffeomorphic to $M$ (see e.g.~\cite{GG06}). We can choose a basis $[c_1],\ldots, [c_{b_1(\Sigma)}]$ of $H_1(\Sigma,\partial\Sigma)$ such that each $c_i\subset\Sigma$ is a properly embedded arc disjoint from $\gamma\cap\Sigma$, and so that, viewing the classes $[c_i]$ in $H_1(\widehat\Sigma,\partial\widehat\Sigma)$, when we add $[c]$ we obtain a basis of $H_1(\widehat\Sigma,\partial\widehat\Sigma)$. Using this basis one can easily check that the natural inclusion map $H_1(\Sigma,\partial\Sigma)\to H_1(\widehat\Sigma,\partial\widehat\Sigma)$ restricts to an isomorphism \[ H_1(\Sigma,\partial\Sigma)^\varphi\cong H_1(\widehat\Sigma,\partial\widehat\Sigma)^{\widehat\varphi}. \] Since $|\partial\widehat\Sigma| = |\partial\Sigma|-1$, by the inductive assumption we have $b_1(M) = \dim_{\mathbb Q} H_1(\widehat\Sigma,\partial\widehat\Sigma)^{\widehat\varphi}$. This proves the claim in full generality. Finally, observe that the maps $f_2$ and $f_2\circ\psi^{-1}\circ g$ are well--defined over the integers. If we represent homology classes in $H_2(N_{\varphi}, \partial N_{\varphi};\Z)$ and $H_2(M;\Z)$ by oriented, properly embedded surfaces intersecting the page $\Sigma \times \{ 0 \}$ transversally and we follow the construction of the connecting homomorphism, we see that the maps $f_2$ and $f_2\circ\psi^{-1}\circ g$ are both realized geometrically by intersecting with $\Sigma \times \{ 0 \}$. This concludes the proof. \end{proof} \section{The proof of Theorem~\ref{t:main}}\label{s:proof} We start by recalling a basic result of Reidemeister and Singer about collections of compressing disks in a handlebody. We refer to~\cite{Jo06} for a modern presentation of this material. Let $H_g$ be a 3--dimensional handlebody of genus $g$. A properly embedded disk $D\subset H_g$ is {\em essential} if $\partial D$ does not bound a disk in $\partial H_g$. \begin{defn}\label{d:minsysdis} A collection $\{D_1,\ldots, D_g\}\subset H_g$ of $g$ properly embedded, pairwise disjoint essential disks is a {\em minimal system of disks} for $H_g$ if the complement of a regular neighborhood of $\bigcup_i D_i$ in $H_g$ is homeomorphic to a 3--dimensional ball. \end{defn} Let $D_1, D_2\subset H$ be properly embedded, essential disks in the handlebody $H_g$. Let $a\subset\partial H$ be an embedded arc with one endpoint on $\partial D_1$ and the other endpoint on $\partial D_2$. Let $N$ be the closure of a regular neighborhood of $D_1\cup D_2\cup a$ in $H$. Then, $N$ is homeomorphic to a closed $3$--ball, and it intersects $\partial H_g$ in a subset of $\partial N$ homeomorphic to a three--punctured 2--sphere. The complement $\partial N\setminus \partial H_g$ of this subset consists of the disjoint union of three disks, two of which are isotopic to $D_1$ and $D_2$ respectively, and the third one is denoted by ${D_1}*_a D_2$. See Figure~\ref{f:disk-slide}. \begin{figure}[ht] \labellist \hair 2pt \pinlabel $D_1$ at 155 270 \pinlabel $D_2$ at 560 270 \pinlabel ${D_1}{*_a}D_2$ at 360 215 \pinlabel $a$ at 360 287 \endlabellist \centering \includegraphics[scale=0.4]{slide} \caption{A disk slide} \label{f:disk-slide} \end{figure} Let ${\bf D} = \{D_1,\ldots, D_g\}$ be a minimal system of disks for a handlebody $H_g$, $a\subset H_g$ an embedded arc with one endpoint on $\partial D_i$, the other endpoint on $\partial D_j$, with $i\neq j$, and the interior of $a$ disjoint from $\bigcup_i \partial D_i$. Then, removing either $D_i$ or $D_j$ from ${\bf D}$ and adding $D_i *_a D_j$ yields a new minimal system of disks ${\bf D'}$ for $H_g$, well--defined up to isotopy~\cite[Corollary~2.11]{Jo06}. In this situation we say that ${\bf D'}$ is obtained from ${\bf D}$ by a {\em disk slide}. \begin{defn}\label{d:slideequiv} Two minimal systems of disks for $H_g$ are {\em slide equivalent} if they are connected by a finite sequence ${\bf D}_1,\ldots,{\bf D}_m$ such that ${\bf D}_{i+1}$ is obtained from ${\bf D}_i$ by a disk slide for each $i$. \end{defn} To prove Theorem~\ref{t:main} we need the following result (see~\cite[Theorem~2.13]{Jo06} for a modern exposition). \begin{thm}[\cite{Re33, Si33}]\label{t:RS} Any two minimal systems of disks for a handlebody are slide equivalent. \qed\end{thm} We can now start the formal proof of Theorem~\ref{t:main}. The first step is to normalize the position of certain non--separating $2$--spheres with respect to a Heegaard splitting. This will be done in the following lemma. \begin{lemma}\label{standard spheres} Let $M= H \cup H'$ be a Heegaard splitting of a $3$-manifold $M$ which admits a prime factorization \begin{equation}\label{e:factor2} M = M_1\#\cdots\# M_h\# S^2\times S^1\#\stackrel{(k)}{\cdots}\# S^2\times S^1 \end{equation} with $b_1(M)=k$. Then, there are pairwise disjoint, embedded 2--spheres $S_1, \ldots, S_k$ in $M$ such that each $S_i$ intersects the Heegaard surface $\partial H$ in a single circle $C_i$. Moreover, after choosing an orientation of each $S_i$, the corresponding 2--homology classes $[S_i]$ generate $H_2(M; \mathbb Q)$. \end{lemma} \begin{proof} Suppose that $M'=M_1 \# \cdots \# M_h$ where each $M_i$ is irreducible. By definition any embedded 2--sphere $S\subset M_i$ bounds a 3--ball. Therefore, if we denote by $S_1',\ldots, S_{h-1}'\subset M'$ the separating spheres along which the connected sums are performed and $S_h' \subset M'$ is any smoothly embedded 2--sphere disjoint from $S_1',\ldots, S_{h-1}'$, then the closure of some component of $M'\setminus\bigcup_{i=1}^h S_i'$ is a punctured 3--ball. In the terminology of Haken~\cite{Ha68}, a collection of pairwise disjoint, embedded 2--spheres with such a property is called a {\em complete system of spheres}. Thus, the collection $S_1',\ldots, S_{h-1}'$ is a complete system of spheres for $M'$. If we view each sphere $S_i'$ as contained in $M$ and denote by $S_{h-1+i}'\subset M$, for $i=1,\ldots, k$, the embedded 2--sphere corresponding to $S^2\times \{ 1 \}$ in the $i$--th $S^2\times S^1$--factor of the factorization~\eqref{e:factor2}, the whole collection $S_1',\ldots, S_{h-1}', S'_h,\ldots, S_{h-1+k}'$ is a complete system of spheres for $M$. Observe that, since $b_1(M)=k$, $b_1(M')=0$. Then, after choosing orientations, the homology classes $[S_{h-1+i}']\in H_2(M;\mathbb Q)$ generate $H_2(M;\mathbb Q)$ as a $\mathbb Q$--vector space, and {\it a fortiori} the same is true for the classes $[S_1'],\ldots, [S_{h-1+k}']$. Now, according to the lemma on page 84 of~\cite{Ha68}, the system of spheres $S_1',\ldots, S_{h-1+k}'$ may be transformed by a finite sequence of isotopies and ``$\rho$--operations'' (see~\cite{Ha68} for the definition) into a collection of pairwise disjoint, incompressible 2--spheres $S_1,\ldots, S_t$, $t\geq h-1+k$, such that each $S_i$ intersects the Heegaard surface $\partial H$ in a single circle $C_i = S_i\cap\partial H$, and moreover the classes $[S_i]$ still generate $H_2(M;\mathbb Q)$. Since $\dim_\mathbb Q H_2(M;\mathbb Q)=k$, up to renaming the spheres we may assume that $[S_1],\ldots, [S_k]$ are generators of $H_2(M;\mathbb Q)$. This finishes the proof of the lemma. \end{proof} \begin{proof}[Proof of Theorem~\ref{t:main}] Let $(B,\pi)$ be an open book decomposition of a closed, oriented, connected 3--manifold $M$ with page $\Sigma$ and monodromy $\varphi$. If $\varphi$ is trivial then it is easy to check that $M$ is homeomorphic to the connected sum of $b_1(\Sigma)$ copies of $S^2\times S^1$. This proves one direction of the statement. For the other direction, suppose that $M$ factorizes as in~\eqref{e:factor}. In view of Proposition~\ref{p:intersection} or Inequality~\eqref{e:b1ineq} we have \[ b_1(\Sigma)\geq b_1(M) \geq k. \] If $b_1(\Sigma)=k$, the above inequality implies $b_1(M) = k$ and therefore if we set \[ M':= M_1\#\cdots\# M_h \] we have $b_1(M')=0$. Denote by $H_{b_1(\Sigma)}\subset M$ the handlebody of genus $b_1(\Sigma)$ consisting of a regular neighborhood of $\Sigma$ in $M$. Since $\Sigma$ is the fiber of a fibration, the closure of the complement $M\setminus H_{b_1(\Sigma)}$ is a handlebody as well, which we denote by $H'_{b_1(\Sigma)}$. It follows that $M$ admits the Heegaard splitting \begin{equation}\label{e:hdec} M=H_{b_1(\Sigma)} \cup H'_{b_1(\Sigma)}. \end{equation} By Lemma \ref{standard spheres} there are pairwise disjoint embedded spheres $S_1, \ldots, S_k \subset M$ which generate $H_2(M;\mathbb Q)$ and such that each $S_i$ intersects the Heegaard surface $\partial H_{b_1(\Sigma)}$ in a single circle $C_i$. Observe that each circle $C_i$ bounds the disk $D_i = S_i\cap H_{b_1(\Sigma)}$ inside $H_{b_1(\Sigma)}$ and the disk $S_i\cap H'_{b_1(\Sigma)}$ inside $H'_{b_1(\Sigma)}$. Since the map \[ H_2(M;\mathbb Q)\to H_1(\partial H_{b_1(\Sigma)};\mathbb Q) \] appearing in the Mayer--Vietoris sequence associated with the decomposition~\eqref{e:hdec} is injective, after choosing orientations we see that the induced homology classes $[C_i]$ generate a half-dimensional subspace of $H_1(\partial H_{b_1(\Sigma)};\mathbb Q)$ which is Lagrangian for the intersection form on $H_1(\partial H_{b_1(\Sigma)};\mathbb Q)$ because the $C_i$'s are pairwise disjoint. We now claim that the $D_i$'s are a minimal system of compressing disks for $H_{b_1(\Sigma)}$. To see this we can argue by induction on $b_1(\Sigma)$. If $b_1(\Sigma)=0$ there is nothing to prove, so we may assume $b_1(\Sigma)>0$. Let $N$ be an open regular neighborhood of $D_1$. Since $[C_1]\neq 0$, $H_{b_1(\Sigma)}\setminus N$ is connected and therefore by e.g.~\cite[Proposition~5.18]{Jo06} it is a handlebody. Moreover, the remaining homology classes $[C_i]$, $i\geq 2$, generate a Lagrangian subspace in the first homology group of the boundary of $H_{b_1(\Sigma)}\setminus N$. By the inductive assumption the disks $D_i$, for $i\geq 2$, are a minimal system of compressing disks for $H_{b_1(\Sigma)}\setminus N$, which proves the claim. Recall that, by construction, the curves $C_i = \partial D_i$ bound compressing disks in $H'_{b_1(\Sigma)}$. Arguing as for $H_{b_1(\Sigma)}$ shows that such disks constitute a minimal system for $H'_{b_1(\Sigma)}$. Thus, surgering $M$ along the spheres $S_1,\ldots, S_k$ yields a 3--manifold having a genus--0 Heegaard splitting, i.e.~$S^3$. This implies that $M$ is a connected sum of $k$ copies of $S^2\times S^1$, and we are left to show that the monodromy $\varphi$ is trivial. Now we choose a system of arcs for $\Sigma$, i.e.~a collection of properly embedded, pairwise disjoint oriented arcs $a_1,\ldots, a_{b_1(\Sigma)}\subset\Sigma$ whose associated homology classes $[a_i]\in H_1(\Sigma,\partial\Sigma;\mathbb Q)$ generate the $\mathbb Q$--vector space $H_1(\Sigma,\partial\Sigma;\mathbb Q)$. Then, after fixing an identification $H_{b_1(\Sigma)} = \Sigma\times I$, the disks $a_i\times I\subset \Sigma\times I$ yield another minimal system of disks $\{D'_i\}_{i=1}^g$ for $H_{b_1(\Sigma)}$. Thus, according to Theorem~\ref{t:RS}, the system $\{D_i\}_{i=1}^g$ is slide equivalent to the system $\{D'_i\}_{i=1}^g$. But recall that, by construction, each curve $C_i = \partial D_i$ bounds a compressing disk in $H'_{b_1(\Sigma)}$, and a moment's reflection shows that any disk slide among the $D_i$'s gives rise to a disk $D_i *_a D_j$ whose boundary also bounds a compressing disk in $H'_{b_1(\Sigma)}$. By induction we conclude that any minimal system of disks $\{\tilde D_i\}_{i=1}^g$ obtained from $\{D_i\}_{i=1}^g$ by a finite sequence of isotopies and disk slides still has the property that each curve $\partial\tilde D_i$ bounds a compressing disk in $H'_{b_1(\Sigma)}$. In particular, this conclusion applies to the system $\{D'_i\}_{i=1}^g$, showing that each of the circles $\partial D'_i$ bounds a compressing disk in $H'_{b_1(\Sigma)}$. Since the splitting~\eqref{e:hdec} is induced by the open book decomposition $(B,\pi)$, we can choose an identification $H'_{b_1(\Sigma)} = \Sigma\times [0,1]$ such that each $\partial D'_i$ is of the form \[ a_i\times\{0\}\bigcup \varphi(a_i)\times\{1\}, \] where $\varphi$ is the monodromy of $(B,\pi)$. The fact that $\partial D'_i$ bounds a disk in $H'_{b_1(\Sigma)}$ says that there is a family of arcs in $\Sigma\times I$ interpolating between $a_i\times\{0\}$ and $\varphi(a_i)\times\{1\}$. Mapping such family to $\Sigma$ via the projection $\Sigma\times I\to\Sigma$ shows that each $a_i$ is homotopic to $\varphi(a_i)$ (with fixed endpoints), and therefore by~\cite{Ep66} each $a_i$ is isotopic to $\varphi(a_i)$ via an isotopy which keeps the endpoints fixed. Since $\{a_i\}$ is a system of arcs for $\Sigma$, a standard argument based on the Alexander lemma~\cite[Lemma~2.1]{FM11} implies that $\varphi$ is isotopic to the identity of $\Sigma$ via diffeomorphisms which fix $\partial\Sigma$ pointwise. This concludes the proof of Theorem~\ref{t:main}. \end{proof} \bibliographystyle{amsplain}

https://arxiv.org/abs/1106.3309

Firing map of an almost periodic input function

In mathematical biology and the theory of electric networks the firing map of an integrate-and-fire system is a notion of importance. In order to prove useful properties of this map authors of previous papers assumed that the stimulus function f of the system \dot{x}= f(t,x) is continuous and usually periodic in the time variable. In this work we show that the required properties of the firing map for the simplified model \dot{x}=f(t) still hold if f \in L_{loc}^1(R) and f is an almost periodic function. Moreover, in this way we prepare a formal framework for next study of a discrete dynamics of the firing map arising from almost periodic stimulus that gives information on consecutive resets (spikes).

\section{{section}{1}{\centering\Large\scshape}} \newcommand{\hbx}{\hfill$\Box$} \newcommand{\ep}{\varepsilon} \newcommand{\eps}[1]{{#1}_{\varepsilon}} \begin{document} \title{Firing map of an almost periodic input function} \author{Wac{\l}aw Marzantowicz and Justyna Signerska} \date{} \maketitle \medskip \bigskip \begin{abstract} In mathematical biology and the theory of electric networks the firing map of an integrate-and-fire system is a notion of importance. In order to prove useful properties of this map authors of previous papers assumed that the stimulus function $f$ of the system $\dot{x}= f(t,x)$ is continuous and usually periodic in the time variable. In this work we show that the required properties of the firing map for the simplified model $\dot{x}=f(t)$ still hold if $f \in L_{loc}^1(R)$ and $f$ is an almost periodic function. Moreover, in this way we prepare a formal framework for next study of a discrete dynamics of the firing map arising from almost periodic stimulus that gives information on consecutive resets (spikes). \end{abstract} \section*{Introduction} The integrate-and-fire models are used mainly in neuroscience to describe nerve-membrane voltage response to a given input. In these, usually one-dimensional models, \begin{equation}\label{row ogolne} \dot{x}=f(t,x), \;\;\; f:\mathbb{R}^2\to\mathbb{R}\, \end{equation} the continuous dynamics govern by a differential equation is interrupted by threshold - reset behavior which is supposed to mimic spiking in real neurons. Precisely, the system evolves from an initial condition according to the differential equation as long as the threshold-value of a dynamical variable is achieved. Then there is an immediate resetting to the resting-value and the dynamics continues from the new initial condition. The resting and threshold values are typically set to $0$ and $1$, respectively. The sequence of consecutive spikes can be described by the iterates of the \emph{firing map}. So far an analytical approach was carried mainly towards models with periodic input, where the firing map is a lift of a degree one circle map and the tools of rotation theory can be used (\cite{brette1}, \cite{Car-Ong}, \cite{gedeon}). Properties of the firing map were also investigated with the help of numerical simulations (\cite{mode},\cite{keener1}). In analytical survey it was always assumed that the input function is smooth enough. In this work we study the case when the function on the right hand side of a differential equation is in general not continuous and not periodic but almost periodic. \section*{Properties of the firing map}\label{firing map} \begin{definition} \rm For the equation (\ref{row ogolne}) we define a map $\Phi: \mathbb{R} \to \mathbb{R}$, $$\Phi(t)=\inf\{s>t: \ \, x(s;t,0)=1\}\,,$$ where $x(\cdot;t,0)$ is a solution of (\ref{row ogolne}) satisfying the initial condition $(t,0)$. \end{definition} \noindent A natural domain of $\Phi$ is the set $ D_{\Phi}=\{t\in \mathbb{R}: \quad \exists_{s>t} \ x(s;t,0)=1\}$. The mapping $\Phi$ assigns to each $t\in D_{\Phi}$ the value $\Phi(t)$, equal to the time after which a trajectory starting from $x=0$ at the moment $t$ reaches the line $x=1$ for the first time. This is the time at which we have the \emph{firing} (alternatively called also a \emph{spike}). Immediately after that the solution $x(t)$ is ''reset'' to zero and next the system given by (\ref{row ogolne}) evolves from new starting point $(\Phi(t),0)$ until a moment $\Phi^2(t)$ satisfying $x(\Phi^2(t);\Phi(t),0)=1$. Thus the iterates $\Phi^n(t)$ of the firing map can be defined recursively as \begin{displaymath} \Phi^{n}(t)=\inf\{s>\Phi^{n-1}(t): \ x(s;\Phi^{n-1}(t),0)=1\}. \end{displaymath} The constant functions $x_r=0$ and $x_\tau=1$ are called {\it the reset} and {\it threshold} functions respectively. One can investigate varying threshold and reset functions (e.g.\cite{gedeon}) but the general cases reduces to the above (\cite{brette1}). \subsection*{Perfect Integrator Model} Consider the system called \emph{Perfect Integrator Model} (after \cite{brette1}), i.e. the scalar ODE: \begin{equation}\label{row1} \dot{x}=f(t), \end{equation} with ''resetting mechanism''. Throughout the rest of this paper we assume that $f\in L^1_{\textrm{loc}}(\mathbb{R})$, unless some additional assumptions are stated, and write simply $\int_{a}^{b}f(u)\,du$ meaning the Lebesque integral $\int_{A}f(u)\,du$ for $A=[a,b]$. We will denote $x(t;t_0,x_0):=\int_{t_0}^tf(u)\,du+x_0$ and call $x(\cdot;t_0,x_0)$ the solution of (\ref{row1}) (satisfying the initial condition $(t_0,x_0)$) or a trajectory of (\ref{row1}). A locally integrable function has desired properties: the function $F(t)=\int_{t_0}^{t}f(u)\,du$ is a continuous function $F: (t_0,\infty)\to \mathbb{R}$ and it assures that the system (\ref{row1}) produces only a finite number of spikes in any finite interval, as we will see after Corollary \ref{uwagadopisana}. Moreover, this is a significantly larger class of functions than was studied before (\cite{brette1}, \cite{Car-Ong}, \cite{mode}, \cite{gedeon}, \cite{keener1}): it includes, for instance, non-continuous piece-wise constant stimulus functions (e.g. the Haar wavelets), often used in neuroscience (see e.g \cite{izykiewicz} and references therein). Since every two trajectories of (\ref{row1}) differ by a constant and the solution $x(t;a,x_0)=\int_{a}^{t}f(u)\,du+x_0$ is continuous for $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$, we get the following \begin{corollary}\label{uwagadopisana} If $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$, then for the equation (\ref{row1}) the consecutive iterates of the firing map are equal to \begin{equation}\label{szybkieiteracje} \Phi^n(t)=\min\{s>t: \ x(s;t,0)=n\} \end{equation} and there is only a finite number of firings in every bounded interval. \end{corollary} Now we state necessary and sufficient conditions under which the firing map $\Phi:\mathbb{R}\to\mathbb{R}$ is properly defined, it is when $\textrm{D}_{\Phi}=\mathbb{R}$. In this case $\Phi$ induces a semi-dynamical system on $\mathbb{R}$. \begin{lemma}\label{lematdopisany} For the model $\dot{x}=f(t)$ the firing map $\Phi:\mathbb{R}\to\mathbb{R}$ is well defined if and only if \begin{equation}\label{warunekkonidost} \limsup_{t\to\infty}\int_{0}^{t}f(u) d u=\infty. \end{equation} In particular, for any $t_0\in\mathbb{R}$ (\ref{warunekkonidost}) is the necessary and sufficient condition for all the iterations $\{\Phi^n(t_0)\}_{n=1}^{\infty}$ to be finite real valued. \end{lemma} \textbf{Proof.} Choose $t_0\in\mathbb{R}$. We want all the iterations $\{\Phi(t_0),\Phi^2(t_0),\Phi^3(t_0), ...\}$ to be well-defined according to (\ref{szybkieiteracje}), that is we demand that for all $n\geq 1$ there exists $s>\Phi^{n-1}(t_0)$ such that $x(s)=n$. However, this is equivalent to demanding that the solution $x:[t_0,\infty)\to\mathbb{R}$ with $x(t_0)=0$ is unbounded from above. But \begin{displaymath} \limsup_{t\to\infty}x(t)=\infty\iff\limsup_{t\to\infty}\int_{t_0}^{t}f(s) d s=\infty\iff\limsup_{t\to\infty}\int_{0}^{t}f(s) d s=\infty. \end{displaymath} Obviously, $\limsup_{t\to\infty}\int_0^t f(s) d s=\infty$ asserts that all the solutions of (\ref{row1}) are unbounded from above. \hbx \vspace{0.5cm} Throughout the rest of the paper we assume that (\ref{warunekkonidost}) is satisfied. By using a property of the Lebesque integral which says that if the integral $\int_{A} f(t) d t$ of a non-negative function on a measurable set $A \subset \mathbb{R}$, $\mu(A)>0$, is positive then there exist $t_1 < t_2$ such that $\int_{t_1}^{t_2} f(t) dt >0$, one obtains \begin{lemma}\label{monotonicity of solutions} Any solution $x(t)$ of the problem (\ref{row1}) is non-decreasing iff $f(t)\geq 0$ and is increasing iff $f(t)>0$ a.e. in $ \mathbb{R}$. \end{lemma} \begin{lemma}\label{monotonicity of firing} Let $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$ and $\Phi: \mathbb{R} \to \mathbb{R} $ be the firing map associated with (\ref{row1}). Then $\Phi$ is increasing, correspondingly, non-decreasing, iff $f(t)>0$, or $f(t)\geq 0$ respectively, a.e. in $\mathbb{R}$. \end{lemma} \textbf{Proof.} The proof that $f(t)\geq 0$ ($f(t)>0$) a.e. induces non-decreasing (increasing) firing map $\Phi$ is almost immediate from the previous lemma. Conversely, suppose that there exists a non-zero measure set $A$ such that $f(t)<0$ for $t\in A$. From Lemma \ref{monotonicity of solutions} there exist $t_1,t_2$, $t_1<t_2$, such that $x(t_2)-x(t_1)=\int_{t_1}^{t_2}f(u)\, du<0$ where $x(\cdot)$ is an arbitrary solution of (\ref{row1}). Suppose that $\Phi(t_1)\geq t_2$. Then $1=\int_{t_1}^{t_2}f(u)\, du+\int_{t_2}^{\Phi(t_1)}f(u)\, du$ implies $\int_{t_2}^{\Phi(t_1)}f(u)\, du>1.$ Consequently $\Phi(t_2)<\Phi(t_1)$. If $\Phi(t_1)<t_2$, let then $n\in\mathbb{N}$ be such that $\Phi^{n}(t_1)<t_2$ but $\Phi^{n+1}\geq t_2$ (existence of such $n$ is ensured by the fact that there can be only a finite number of firings in the interval $[t_1,t_2]$). Then, as before, we obtain that $\Phi(t_2)<\Phi(t_*)$, where $t_*=\Phi^{n}(t)<t_2$ because in this case $\int_{t_*}^{t_2}f(u)\,du<0$. Analogously we show that when $\Phi$ is strictly increasing, then $f$ is positive, perhaps with the exception of some zero-measure set. \hbx \vspace{0.5cm} Some general results concerning continuity and injectivity of the firing map for the models of the type $\dot{x}=f(t,x)$ when $f$ is smooth enough are given in \cite{Car-Ong}. However, for the Perfect Integrator we obtain a detailed description of continuity if only $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$. Let $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$, $f(t)\geq 0$ almost everywhere. For $a\in \mathbb{R}$ consider $\Phi^{-1}(a) \subset \mathbb{R}$. Let $\bar{a}= \sup\{ t\in \Phi^{-1}(a)\}$. Note that $\bar{a} < \infty$, because $\bar{a}<a$ since for any $t\in\mathbb{R}$ $t<\Phi(t)$. Moreover, there exists an interval $[\bar{a}, \bar{a} +\bar{\delta}]$ such that for every $t$ satisfying $\bar{a} < t \leq \bar{\delta}$ we have $\int_{\bar{a}}^t f(u) \,du >0$, or equivalently $\mu(\{u: \, f(u)>0\} \cap [\bar{a}, \bar{a} + t]) >0$. Indeed, if there exists $t$ such that $a> t>\bar{a}$ and $\int_{\bar{a}}^tf(u)\, du = 0$ then $ \int_{t}^{a} f(u)\, du = \int_{\bar{a}}^{a} f(u)\,du =1$ and hence $t\in\Phi^{-1}(a)$ contrary to the definition of $\bar{a}$. \begin{proposition}\label{discontunity} If $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$ and $f(t)\geq 0$ almost everywhere. Then \begin{itemize} \item[i)] {$\Phi$ is left continuous,} \item[ii)] {$\Phi$ is not right continuous at every point $\bar{a} \in \Phi^{-1}(a)$ for which there exists $\delta_0>0$ such that $f(t)=0$ almost everywhere in $[a, a+\delta_0]$. Moreover such points are the only points of discontinuity of $\Phi$. } \end{itemize} \end{proposition} \textbf{Proof.} For the first part of the statement take a sequence $t_n \to t_0$, $t_n <t_0$. Then $1 =\int_{t_n}^{\Phi(t_n)} f(u)\, du $ and $1 = \int_{t_0}^{\Phi(t_0)} f(u)\,du $ implies $\int_{t_n}^{t_0} f(u) \,du = \int_{\Phi(t_n)}^{\Phi(t_0)} f(u)\, du $. It follows that $\int_{\Phi(t_n)}^{\Phi(t_0)} f(u)\, du \to 0$ if $n\to \infty$. We have to prove that $\Phi(t_n) \to \Phi(t_0)$. Note that since $\Phi$ is non-decreasing, $\Phi(t_n)\leq\Phi(t_0)$. If for some $\delta_0>0$ $f(t) = 0 $ almost everywhere in the interval $[\Phi(t_0) - \delta_0,\Phi(t_0)]$ then $ \int_{t_0}^{\Phi(t_0) - \delta_0} f(u)\, du=1$ which gives a contradiction to the definition of $\Phi$. Thus $f(t)>0 $ on a set of positive measure in any small interval $[\Phi(t_0) - \delta,\Phi(t_0)]$ and consequently $\Phi(t_n) \to \Phi(t_0)$. To prove the second part of the statement let us take $t_n\to \bar{a}, \; t_n>\bar{a}$. We have already showed that $\int_{\bar{a}}^{t_n} f(u) \,du \, >0 $. Then $\Phi(t_n) \geq \Phi(\bar{a}) +\delta_0 $ for large $n$, because for every $b$, $ \Phi(\bar{a}) \leq b \leq \Phi(\bar{a}) +\delta_0$ we have $$ \int_{t_n}^b f(u)\,du = \int_{t_n}^{\Phi(\bar{a})} f(u) \,du =\, \int_{\bar{a}}^{\Phi(\bar{a})} f(u)\,du - \int_{\bar{a}}^{t_n} f(u) \,du \, < 1.$$ It follows that $\Phi(\bar{a}_+) \geq \Phi(\bar{a}) +\delta_0$ and consequently $\Phi$ is not right continuous at $\bar{a}$. Conversely, let $t_n \to t_0$, $t_n> t_0$ and $\Phi(t_n) \geq \Phi(t_0) +\bar{\delta}$. Then $\int_{t_0}^{t_n} f(u)\, du >0$, since otherwise $1= \int_{t_0}^{\Phi(t_0)} f(u)\, du = \int_{t_n}^{\Phi(t_0)} f(u)\, du $, which gives $\Phi(t_n)=\Phi(t_0)$ contrary to the supposition. Next, if for every $\delta_0 >0$ we have $\mu(\{f(u)>0\} \cap [\Phi(t_0), \Phi(t_0) +\delta_0]) >0$, then $\int_{\Phi(t_0)}^{\Phi(t_0)+\delta_0} f(u)\, du >0$. We can take $\delta_0 < \bar{\delta}$. For large $n$ $\int_{t_0}^{t_n}f(u)\, du $ is small, say smaller than $\int_{\Phi(t_0)}^{\Phi(t_0)+\delta_0} f(u)\, du$. This gives $ \int_{t_n}^{\Phi(t_0) +\delta_0} f(u)\,du$ $= \int_{t_0}^{\Phi(t_0)} f(u) \,du -\int_{t_o}^{t_n} f(u)\,du + \int_{\Phi(t_0)}^{\Phi(t_0) +\delta_0} f(u)\,du >1$, thus $\Phi(t_n) < \Phi(t_0) + \delta_0$ contrary to our supposition that $\Phi(t_n) \geq \Phi(t_0) +\bar{\delta}$. \hbx \vspace{0.5cm} As a direct consequence of Proposition \ref{discontunity} we get the following: \begin{corollary}\label{continuity} If $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$ and $f(t)>0$ almost everywhere, then $\Phi$ is continuous. \end{corollary} \section*{Firing rate} \begin{definition}\label{def_firing} For every $t\in \mathbb{R}$ one defines the \emph{firing rate}, denoted by $r(t)$, as \begin{equation}\label{row2} r(t):=\lim_{n\to\infty}\frac{n}{\Phi^{n}(t)}. \end{equation} \end{definition} In general for an equation $\dot{x}=f(t,x)$ the limit of (\ref{row2}) might not exist for some $t$. Moreover, even if the limit exists for all $t$, it might depend on $t$. However, for the simplified model (\ref{row1}) we have the following theorem, which was proved in \cite{brette1} and the proof is valid for $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$: \begin{theorem}\label{tw1} Suppose that for the model (\ref{row1}) there exists a finite limit \begin{equation}\label{row3} r=\lim_{t\to\infty}\frac{1}{t}\int_{0}^{t}f(u)d u. \end{equation} Then for every point $t_0\in\mathbb{R}$ the firing rate $r(t_0)$ exists and is given by the formula (\ref{row3}). In particular the firing rate $r(t)$ does not depend on $t$. \end{theorem} \subsection*{Almost periodic functions} In next we consider the model (\ref{row1}) with the function $f:\mathbb{R}\to\mathbb{R}$ almost periodic. There are various classes of almost periodic functions, e.g. in the sense of Bohr, Stepanov, Weyl and Besicovitch. They are defined as the closure of the space of generalized trigonometric polynomials under corresponding norms. The notion of almost periodicity generally holds for functions $f:\mathbb{R}\to\mathbb{C}$. However, we will only consider almost periodic functions taking real values. \begin{definition}\label{Bohr alm per} A continuous function $f:\mathbb{R}\to\mathbb{C}$ is \emph{almost periodic function in the sense of Bohr} (alternatively called also uniformly almost periodic) if for any $\ep>0$ the set $E\{\ep,f(t)\}$ of all the numbers $\tau$ such that for all $t\in\mathbb{R}$ $|f(t+\tau)-f(t)|<\ep$ is relatively dense, i.e. there exists $l_{\ep}>0$ such that in any interval of length $l_{\ep}$ there is at least one $\tau$ satisfying this inequality. The set $E\{\ep,f(t)\}$ is called the set of $\ep$-almost periods of $f$. \end{definition} Equivalently we can say that $f$ is Bohr almost periodic if it is the limit of a sequence of generalized trigonometric polynomials $P_n(t)=\sum_{i=1}^{k(n)}c_i\exp(\imath \lambda_i t)$, $c_i\in\mathbb{C}$ and $\lambda_i\in\mathbb{R}$, under the norm $\|f\|=\sup_{t\in\mathbb{R}}|f(t)|$. Any Bohr almost periodic function is uniformly continuous and bounded (\cite{bezykowicz},\cite{corduneanu}). \begin{definition}\label{stepanov alm per} A function $f:\mathbb{R}\to\mathbb{C}$, $f\in L^{p}_{\textrm{loc}}(\mathbb{R})$, is \emph{Stepanov almost periodic} if for any $\ep>0$ the set $SE\{\ep,f(t)\}$ of all the numbers $\tau$ such that $\|f(t+\tau)-f(t)\|_{\textrm{St.},r,p}<\ep$ is relatively dense, where $$\|f\|_{\textrm{St.},r,p}:=\sup_{t\in\mathbb{R}}[\frac{1}{r}\int_{t}^{t+r}|f(u)|^p\,du]^{1/p}, \quad r>0, \ 1\leq p<\infty.$$ The set $SE\{\ep,f(t)\}$ is called the set of $\ep$-Stepanov almost periods of $f$. \end{definition} The space of Stepanov almost periodic functions can be obtained also as the closure of generalized trigonometric polynomials under the norm $\|\cdot\|_{\textrm{St.},r,p}$. Any Stepanov almost periodic function which is uniformly continuous is Bohr almost periodic. Different values of $r$ in $\|\cdot\|_{\textrm{St.},r,p}$ give obviously different norms but the same topology. Thus we will consider Stepanov almost periodic functions with $r=1$. If $f$ is a $p_1$-Stepanov almost periodic function, then it also $p_2$-Stepanov almost periodic for $p_2<p_1$ with fixed $r$ (\cite{andres}, \cite{bezykowicz}). One can also consider almost periodic functions in the sense of Weyl or Besicovitch, where the class of Besicovitch almost periodic functions with $p=1$ is larger and includes all the other mentioned types of almost periodicity. For an almost periodic function $f:\mathbb{R}\to\mathbb{C}$ of any of these types there exists a finite \emph{mean} (\cite{andres},\cite{bezykowicz}, \cite{corduneanu}): \[ \mathcal{M}\{f(t)\}=\lim_{t\to\infty}\frac{1}{t}\int_{0}^{t}f(u)\,du. \] From this we conclude that: \begin{corollary}\label{maly wniosek1} For the model (\ref{row1}), where $f:\mathbb{R} \to \mathbb{R}$ is almost periodic in any of the above sense the firing rate exists and equals the mean of a function, provided that the mean is nonnegative (otherwise the firing rate is zero). \end{corollary} Note that if $\mathcal{M}\{f(t)\}<0$, then (\ref{warunekkonidost}) does not hold and there is no firing. \begin{definition}\label{def limit periodic} A function $f:\mathbb{R}\to\mathbb{C}$ is limit-periodic (in the sense of Bohr) if it is a limit of uniformly converging sequence $\{f_n\}$ of continuous periodic functions. \end{definition} Note that generalized trigonometric polynomials are in general not periodic (consider, for example, $P(t)=\sin(\sqrt{2}t)+\sin(2t)$) and that a limit-periodic function is an uniformly almost periodic function. One can read in \cite{andres} in details about the theory of almost periodic functions. We will write shortly that $f$ is u.a.p. or S.a.p., meaning uniform (Bohr) or Stepanov (with $r=1$ and $p=1$) almost periodicity, respectively. The mean $\mathcal{M}$ is continuous \cite{bezykowicz}, meaning that $\mathcal{M}\{f_n\}\to\mathcal{M}\{f\}$ if $f_n\to f$ in any of the ``almost-periodic norms" (i.e. the norms with respect to which a given space of almost periodic functions can be obtained as the closure of the space of generalized trigonometric polynomials). Thus the following holds: \begin{theorem}\label{firing rate for almost periodic} Let $(f_n)_{n=1}^{\infty}$ be a sequence of almost-periodic functions $f_n:\mathbb{R}\to\mathbb{R}$ of the Bohr, Stepanov, Weyl or Besicovitch type converging to $f$ in uniform, Stepanov, Weyl or Besicovitch norm, respectively. Let $r_n$ be the firing rate of the equation $\dot{x}=f_n(t)$, for $n=1,2,3,...$, and $r$ the firing rate for (\ref{row1}). Then $r=\lim_{n\to\infty}r_n$. \end{theorem} We end this section with propositions showing when the existence of the positive mean value of $f$ is a necessary and sufficient condition for existence of the firing map $\Phi:\mathbb{R}\to\mathbb{R}$. By elementary calculations one proves \begin{proposition}\label{mean implies firing} Let $f\in L^{1}_\textrm{loc}(\mathbb{R})$. If $\mathcal{M}\{f\}>0$ then $ \limsup \int_0^{t} f(u)\,du = \infty$. Consequently the firing map $\Phi$ of (\ref{row1}) is well-defined for all $t\in \mathbb{R}$. Conversely, if $f(t)= \sum _{j=1}^k c_j\exp(\imath \lambda_j \, t)$, then $\limsup_{t\to\infty} \int_0^t f(u)\, du =\infty $ implies $ \mathcal{M}\{f\}>0$. \end{proposition} In the second case it must hold that $\mathcal{M}\{f\}=c_{j_*}>0$, where $c_{j_*}$ corresponds to $\lambda_{j_*}=0$. The mean has the following property, which in general does not hold for functions that have finite $\sup$- or Stepanov norms but are not almost periodic: \begin{lemma}\label{o sredniej dla nieujemnej} Let $f:\mathbb{R}\to\mathbb{R}$ be a real non-negative (a.e.) u.a.p. (/S.a.p.) function. If $\mathcal{M}\{f(t)\}=0$, then $f(t)=0$ for all (/almost all) $t\in\mathbb{R}$. \end{lemma} \textbf{Proof.} The proof of the statement for u.a.p. functions can be found in \cite{bezykowicz}. We will prove the statement for $f$ being a S.a.p. function. Suppose on the contrary that there exists a non-zero measure set $A$ such that for all $t\in A$ we have $f(t)>0$. Then there exists $a\in\mathbb{R}$ such that $ \int_{a}^{a+1}f(u)\,du>\lambda$ for some $\lambda>0$ as follows from Lemma \ref{monotonicity of firing}. Since $f$ is S.a.p. the set of all the $\tau$ such that for all $t\in\mathbb{R}$ $\int_{t}^{t+1}|f(u+\tau)-f(u)|<\frac{\lambda}{2}$ is relatively dense. It follows that there exists $l_{\frac{\lambda}{2}}>0$ such that in any interval $I$ of length $l_{\frac{\lambda}{2}}$ we can find $\tau\in I$ such that $\int_{a+\tau}^{a+\tau+1}f(u)\,du=\int_{a}^{a+1}f(u+\tau)\,du>\frac{\lambda}{2}$: Indeed, if $\int_{a}^{a+1}|f(u+\tau)-f(u)|\,du<\frac{\lambda}{2}$ and $\int_{a}^{a+1}f(u)\,du>\lambda$, then $\frac{\lambda}{2}<\int_{a}^{a+1}f(u)\,du-\int_{a}^{a+1}|f(u+\tau)-f(u)|\,du\leq\int_{a}^{a+1}f(u+\tau)\,du$. We can assume that $l_{\frac{\lambda}{2}}>1$. Then we can find a sequence $\{\tau_n\}$ where $\tau_{n}\in[2n l_{\frac{\lambda}{2}},2(n+1)l_{\frac{\lambda}{2}}-1]$ such that \[ \int_{a+2kl_{\frac{\lambda}{2}}}^{a+2(k+1)l_{\frac{\lambda}{2}}}f(u)\,du>\int_{a+\tau_{k}}^{a+\tau_{k}+1}f(u)\,du>\frac{\lambda}{2} \] for any $k\in\mathbb{N}\cup\{0\}$ because $a+\tau_{k},a+\tau_{k}+1\in[a+2kl_{\frac{\lambda}{2}},a+2(k+1)l_{\frac{\lambda}{2}}]$. Consequently, \begin{equation} \begin{split} 0=\mathcal{M}\{f\} =&\lim_{t\to\infty}\frac{1}{t}\int_{0}^{t}f(u)\, du=\lim_{t\to\infty}\frac{1}{t}\int_{a}^{a+t}f(u)\, du\nonumber\\ =&\lim_{n\to\infty}\frac{1}{2nl_{\frac{\lambda}{2}}}\int_{a}^{a+2nl_{\frac{\lambda}{2}}}f(u)\,du>\frac{1}{2nl_{\frac{\lambda}{2}}}\,n\,\frac{\lambda}{2}\,=\,\frac{\lambda}{4l_{\frac{\lambda}{2}}} \nonumber \end{split} \end{equation} and we obtain a contradiction. \hbx \vspace{0.5cm} \begin{proposition}\label{mean and firing dla nonnegative} Suppose that $f(t)\geq 0$ (a.e.) is an u.a.p. (/S.a.p.) function. Then $\mathcal{M}\{f\}>0$ if and only if the firing map is well-defined. \end{proposition} \textbf{Proof.} Sufficiency was covered in the previous proposition, necessity follows from Lemma \ref{o sredniej dla nieujemnej}. Indeed, if $\mathcal{M}\{f\}=0$ for non-negative (a.e.) u.a.p. (/S.a.p.) function $f$, then $f(t)=0$ for all $t\in\mathbb{R}$ (/a.e. in $\mathbb{R}$). In both cases the condition $\limsup_{t\to\infty}\int_{0}^t f(u)\,du=\lim_{t\to\infty}\int_{0}^t f(u)\,du=\infty$ is not satisfied. \hbx \vspace{0.5cm} \begin{remark}\label{welvets} The assumption that a stimulus function $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$ allows us also to take as $f$ a linear combination of the Haar wavelets which is more natural for many problems with discontinuous stimulus function. \end{remark} \section*{Displacement map} \begin{definition}\label{displacement map}{\rm Let $\Phi: \mathbb{R} \to \mathbb{R}$ be a map of the real line. The {\em displacement map} $\Psi: \mathbb{R} \to \mathbb{R} $ of $\Phi$ is defined as $$ \Psi(t)\,:= \Phi(t)-t \,.$$} \end{definition} If $\Phi$ is a firing map then the displacement map $\Psi(t)$ says how long we have to wait for a next firing if we know that there was a firing at the time $t$. Firstly, we will consider the firing map and its displacement for the general case, i.e. for the equation (\ref{row ogolne}). The following observation was first made in \cite{keener1}: \begin{proposition}\label{displacement okresowe} If the function $f$ in (\ref{row ogolne}) is periodic in $t$ (that is, there exists $T$ such that for all $x$ and $t$ we have $f(t,x)=f(t+T,x)$), then the firing map $\Phi$ has periodic displacement. In particular for $T=1$ we have $\Phi(t+1)=\Phi(t)+1$ and thus $\Phi$ is a lift of a degree one circle map under the standard projection $\mathfrak{p}: t\mapsto \exp(2\pi\imath t)$. \end{proposition} Therefore for the models (\ref{row ogolne}) with $f$ periodic in $t$, the firing map $\Phi$ has periodic displacement and induces a degree one circle map (this map is sometimes in the literature referred to as the \emph{firing phase map}). In this case the tools of the rotation theory (\cite{misiurewicz}) can be used to study the properties of $\Phi$. Especially, the rotation number (which is simply the reciprocal of the firing rate, if the unique non-zero firing rate exists) or the rotation intervals are closely related to the so-called phase-locking phenomena (\cite{brette1}, \cite{keener1}). There arises a natural question of whether for the map $f$ almost periodic in $t$ we obtain the firing map with almost periodic displacement. We will tackle this problem for the Perfect Integrator Model. Firstly, let us consider the case of limit-periodic input: \begin{theorem} \label{twierdzenie o limitperiodic} Let $\Phi:\mathbb{R}\to \mathbb{R}$ be the firing map induced by (\ref{row1}). If $f:\mathbb{R}\to\mathbb{R}$ is limit periodic and $f(t)>\delta$ for some $\delta>0$, then the firing map $\Phi$ has limit periodic displacement. \end{theorem} \textbf{Proof.} Denote the displacement of $\Phi$ by $\Psi$. Suppose, without the loss of generality, that $0<\delta<1$ and let $\ep>0$ be any small number. We can assume that $\delta>\ep/2>0$. Since $f$ is limit periodic, there exists continuous periodic function $\widetilde{f}:\mathbb{R}\to\mathbb{R}$ such that $ |\widetilde{f}(t)-f(t)|<\delta^2\ep/4$ for all $t$. Choose the initial condition $(t_0,0)$. By $x(t)$ denote the solution of (\ref{row1}) satisfying $x(t_0)=0$ and by $\widetilde{x}(t)$ the solution of \begin{equation}\label{eq5} \dot{x}=\widetilde{f}(t) \end{equation} with the same initial condition. Equation (\ref{eq5}) gives rise to the firing map $\widetilde{\Phi}(t)$ with the displacement $\widetilde{\Psi}(t)$. We will show that $|\Psi(t_0)-\widetilde{\Psi}(t_0)|<\ep$, which is equivalent to $|\Phi(t_0)-\widetilde{\Phi}(t_0)|<\ep$, and then the theorem will follow from the fact that the firing map $\widetilde{\Phi}$ of periodically forced Perfect Integrator Model has periodic displacement, which is covered by Proposition \ref{displacement okresowe}. Since $\widetilde{f}$ approximates $f$ in the uniform norm, we have that for all $t$ $\widetilde{f}(t) \geq \delta/2$. We have to consider two possibilities: $\Phi(t_0)>\widetilde{\Phi}(t_0)$ and $\widetilde{\Phi}(t_0)>\Phi(t_0)$. Suppose that $\Phi(t_0)>\widetilde{\Phi}(t_0)$, i.e. the solution $\widetilde{x}(t)$ fires before $x(t)$. Then $x(\Phi(t_0))=1$ and $x({\widetilde{\Phi}}(t_0))<1$ by the definition of $\widetilde{\Phi}$ and $\Phi$. From the Mean Value Theorem \begin{equation}\label{eq7} 1-x(\widetilde{\Phi}(t_0))=f(\alpha)(\Phi(t_0)-\widetilde{\Phi}(t_0)) \end{equation} for some $\alpha\in(\widetilde{\Phi}(t_0),\Phi(t_0))$. Now define $y(t):=\widetilde{x}(t)-x(t)$. Then $y(t)-y(t_0)=y^{\prime}(\xi_t)(t-t_0)$ for any $t>t_0$ and some $\xi_t\in(t_0,t)$. Since $y(t_0)=0$ and $y^{\prime}(\xi_t)=\widetilde{x}^{\prime}(\xi_t)-x^{\prime}(\xi_t)=\widetilde{f}(\xi_t)-f(\xi_t)\in(-\delta^2\ep/4, \delta^2\ep/4)$, we get \begin{equation}\label{eq8} \widetilde{x}(\widetilde{\Phi}(t_0))-x(\widetilde{\Phi}(t_0))<(\widetilde{\Phi}(t_0)-t_0)\frac{\delta^2\ep}{4}. \end{equation} Equations (\ref{eq7}) and (\ref{eq8}) yield $\Phi(t_0)-\widetilde{\Phi}(t_0)<\frac{ \delta^2\ep}{4 f(\alpha)}(\widetilde{\Phi}(t_0)-t_0)$, where the difference $\widetilde{\Phi}(t_0)-t_0\leq\frac{2}{\delta}$ because $1=\widetilde{x}(\widetilde{\Phi}(t_0))-\widetilde{x}(t_0)=\widetilde{f}(\beta)(\widetilde{\Phi}(t_0)-t_0)$ for some $\beta\in(t_0,\widetilde{\Phi}(t_0))$. Finally, $\Phi(t_0)-\widetilde{\Phi}(t_0)<\frac{ \delta^2\ep}{4 f(\alpha)\widetilde{f}(\beta)}< \ep$ since $f(\alpha),\widetilde{f}(\beta)>\frac{\delta}{2}$. Analogously we prove that $\widetilde{\Phi}(t_0)-\Phi(t_0)<\ep$ for $\widetilde{\Phi}(t_0)>\Phi(t_0)$. \hbx \vspace{0.5cm} From the proof of Theorem \ref{twierdzenie o limitperiodic} follows more general \begin{proposition}\label{twierdzenie o przyblizaniu} If $f:\mathbb{R}\to\mathbb{R}$, $f(t)>\delta>0$ is a limit of uniformly convergent sequence of continuous functions $f_n:\mathbb{R}\to\mathbb{R}$ inducing firing maps $\Phi_n$ with displacements $\Psi_n$, the firing map $\Phi$ and the displacement $\Psi$ of $f$ can be obtained as the uniform limits of $\Phi_n$ and $\Psi_n$, respectively. \end{proposition} The above means that if the system $\dot{x}=f_{\gamma}(t)$ depends on the parameter $\gamma\in\mathbb{R}$ continuously with respect to the uniform norm, i.e. $\sup_{t\in\mathbb{R}}|f_{\gamma_1}(t)-f_{\gamma_2}(t)|$ is arbitrary small provided that the difference $|\gamma_1-\gamma_2|$ is small enough, then the corresponding firing maps $\Phi_{\gamma}(t)$ also change continuously with respect to the uniform topology. Furthermore we show that \begin{theorem}\label{new almost periodic displacement} Let $f:\mathbb{R}\to\mathbb{R}$ be a S.a.p. function such that $f(t)>\delta$ almost everywhere for some $\delta>0$. Then the firing map $\Phi$ induced by the equation (\ref{row1}) has u.a.p. displacement. \end{theorem} \textbf{Proof.} We are to show that for any $\ep>0$ the set $E\{\ep,\Psi(t)\}$ of all the numbers $\tau\in\mathbb{R}$ such that $\sup_{t\in\mathbb{R}}|\Psi(t+\tau)-\Psi(t)|=\sup_{t\in\mathbb{R}}|\Phi(t+\tau)-\tau -\Phi(t)|<\ep$ is relatively dense. Without the loss of generality we assume that $\delta<1$. Choose then $\ep>0$ and let $\tau\in SE\{\frac{\delta^2\ep}{2},f(t)\}$. Take $t\in\mathbb{R}$. Then by the assumption on $f$ $\int_{t}^{t+1}|f(u+\tau)-f(u)|\,du<\frac{\delta^2 \ep}{2}$. Suppose that $\min\{\Phi(t),\Phi(t+\tau)-\tau\}=\Phi(t)$. By the definition of the firing map $\int_{t+\tau}^{\Phi(t+\tau)}f(u)\,du=\int_{t}^{\Phi(t)}f(u)\,du$ and it follows that \begin{equation} \begin{split} 0 =& \int_{t+\tau}^{\Phi(t+\tau)}f(u)\,du-\int_{t}^{\Phi(t)}f(u)\,du= \int_{t}^{\Phi(t+\tau)-\tau}f(u+\tau)\,du-\int_{t}^{\Phi(t)}f(u)\,du=\nonumber \\ =& \int_{t}^{\Phi(t)}f(u+\tau)-f(u)\,du+\int_{\Phi(t)}^{\Phi(t+\tau)-\tau}f(u+\tau)\,du.\nonumber \end{split} \end{equation} Thus $|\int_t^{\Phi(t)}f(u+\tau)-f(u)\,du|=|\int_{\Phi(t)}^{\Phi(t+\tau)-\tau}f(u+\tau)\,du|$. Since \nolinebreak[10]$\tau\nolinebreak[10]\in\nolinebreak[10]SE\{\frac{\delta^2\varepsilon}{2}, f(t)\}$, $ |\int_{t}^{\Phi(t)}f(u+\tau)-f(u)\,du| \leq \int_{t}^{\Phi(t)}|f(u+\tau)-f(u)|\,du\leq \int_{t}^{t+k}|f(u+\tau)-f(u)|\,du < k \delta^2 \ep/2, $ where $k\in\mathbb{N}$ is the smallest integer such that $\Phi(t)\leq t+k$. However, as $f(u)>\delta$ almost everywhere, $\int_{\Phi(t)}^{\Phi(t+\tau)-\tau}f(u+\tau)\,du>\delta(\Phi(t+\tau)-\tau-\Phi(t))$. Finally \[ \Phi(t+\tau)-\tau-\Phi(t)<\frac{k \delta\ep}{2}<\frac{(\frac{1}{\delta}+1)\delta\ep}{2}=\frac{(1+\delta)\ep}{2}<\ep \] because $\Phi(u)-u<1/\delta$ for any $u\in\mathbb{R}$, which can be estimated as in the proof of Theorem \ref{twierdzenie o limitperiodic}, and thus $k<(1/\delta+1)$. If $\min\{\Phi(t),\Phi(t+\tau)-\tau\}=\Phi(t+\tau)-\tau$ in a similar way we obtain that $\Phi(t)-\Phi(t+\tau)+\tau<k \delta\ep/2$ where $k$ is the smallest integer such that $\Phi(t+\tau)-(t+\tau)<k$. Thus in any case $|\Phi(t+\tau)-\Phi(t)-\tau|<\ep$. It follows that $\tau\in E\{\ep,\Psi(t)\}$. Consequently the set $E\{\ep,\Psi(t)\}$ is relatively dense because it contains relatively dense set $SE\{\frac{\delta^2\varepsilon}{2},f(t)\}$. \hbx \vspace{0.5cm} Note that if the displacement $\Psi(t)=\Phi(t)-t$ is uniformly almost periodic then it is uniformly continuous and the following conclusion is immediate: \begin{corollary} Under the assumptions of Theorem \ref{new almost periodic displacement} the firing map $\Phi:\mathbb{R}\to\mathbb{R}$ is uniformly continuous. \end{corollary} \begin{remark}\label{Kwapiszwork} Rotation numbers and sets for maps of the real line with almost periodic displacement were investigated by J. Kwapisz in \cite{kwapisz}. Observe that Theorems \ref{twierdzenie o limitperiodic} and \ref{new almost periodic displacement} ensure that the displacement of a firing map is almost periodic if $f$ is so. On the other hand if a given almost periodic function is a firing map then the thesis of main theorems of \cite{kwapisz} can be deduced easier by a direct argument. Anyway, it would give an opportunity to compare geometrical definitions of rotation intervals of \cite{misiurewicz}, used in \cite{kwapisz}, with analytic formulas derived here. \end{remark} \section*{Prospects} As we have already said, our aim was also to establish a framework for study dynamical properties of the firing map $\Phi$ and its displacement $\Psi$. Now we state a conjecture that the differences of consecutive spikes form an asymptotically semi-periodic, or respectively almost periodic sequence if the input function is so (cf. \cite{berg} for definitions of semi-periodic, and almost-periodic sequence). Let $\Phi: \mathbb{R} \to \mathbb{R}$ be the firing map of the equation $\dot{x}=f(t)$ with $f\in L^{1}_{\textrm{loc}}(\mathbb{R})$ and $\Psi:\mathbb{R}\to \mathbb{R}$ its displacement. For any $t\in \mathbb{R}$, we consider a sequence $$ \eta_n(t)= \Phi^n(t) - \Phi^{n-1}(t) = \Psi(\Phi^{n-1}(t)) $$ \begin{conjecture} If $f$ is periodic, correspondingly almost-periodic, then for every $t$ the sequence $\eta_n(t)$ is asymptotically semi-periodic, respectively asymptotically almost-periodic. \end{conjecture}

https://arxiv.org/abs/1309.4667

Volatility occupation times

We propose nonparametric estimators of the occupation measure and the occupation density of the diffusion coefficient (stochastic volatility) of a discretely observed Itô semimartingale on a fixed interval when the mesh of the observation grid shrinks to zero asymptotically. In a first step we estimate the volatility locally over blocks of shrinking length, and then in a second step we use these estimates to construct a sample analogue of the volatility occupation time and a kernel-based estimator of its density. We prove the consistency of our estimators and further derive bounds for their rates of convergence. We use these results to estimate nonparametrically the quantiles associated with the volatility occupation measure.

\section{Introduction} Continuous-time It\^o semimartingales are used widely to model stochastic processes in various areas such as finance. The general It\^o semimartingale process is given by \begin{equation} \label{eqXintro} X_{t} = X_{0}+\int_{0}^{t}b_{s}\,ds+ \int_{0}^{t}\sigma_{s}\,dW_{s}+J_{t}, \end{equation} where $b_t$ and $\sigma_t$ are processes with c\`{a}dl\`{a}g paths, $W_t$ is a Brownian motion and $J_t$ is a jump process; formal conditions are given in the next section. Inference for model~(\ref {eqXintro}) in the general case (either in a parametric or a nonparametric context) is quite complicated because of the many ``layers of latency,'' for example, as typical in financial applications, $\sigma_t$ and $J_t$ can have randomness not captured by~$X_t$. When $X$ is sampled at discrete times but with the mesh of the observation grid shrinking to zero, that is, high-frequency data of $X$ are available, the distinct pathwise behavior of the components in (\ref{eqXintro}) can be used to nonparametrically separate them. Indeed, various techniques have been already proposed to estimate nonparametrically the integrated variance $\int_0^T\sigma_s^2\,ds$ over a specific interval $[0,T]$ (see, e.g.,~\cite{BNS06} and \cite{Ma1}), and more generally integrated variance measures of the form $\int_{0}^{T}g(\sigma_{s}^2)\,ds$, where $g(\cdot)$ is a continuous function with polynomial growth [and there are more smoothness requirements on $g(\cdot)$ for determining the rate of convergence]; see Theorems 3.4.1 and 9.4.1 in~\cite{jacodprotter2012} and the recent work of~\cite{JR}. This paper extends the existing literature on high-frequency nonparametric volatility estimation by developing a nonparametric jump-robust estimator of the occupation time of the latent volatility process $(V_{t})_{t\geq 0}\equiv(\sigma_{t}^2)_{t\geq0}$ where the volatility occupation time is defined by \begin{equation} \label{eqot} F_{t}(x)=\int_{0}^{t}1_{\{V_{s}\leq x\} }\,ds\qquad \forall x>0,\ t\in [ 0,T ]. \end{equation} Evidently, the right-hand side of (\ref{eqot}) is of the form $\int_0^tg(V_s)\,ds$ with $g(v) = 1_{\{v\leq x\}}$, which unlike earlier work is a discontinuous function. If $F_{t} ( \cdot ) $ is absolutely continuous with respect to the Lebesgue measure, its derivative $f_{t} ( \cdot ) $, that is, the volatility occupation density, is well-defined. By the Lebesgue differentiation theorem, the occupation density can be equivalently defined as \begin{equation} \label{eqod} f_{t}(x)=\lim_{\varepsilon\downarrow0}\frac{1}{2\varepsilon} \bigl(F_{t}(x+\varepsilon)-F_{t}(x-\varepsilon) \bigr). \end{equation} In addition to estimating $F_t(x)$, in this paper we also develop a consistent estimator for the volatility occupation density $f_t(x)$ using the high-frequency record of $X$. The occupation measure of the volatility process ``summarizes'' in a convenient way the information regarding the volatility behavior over the given time interval. Indeed, for any bounded (or nonnegative) Borel function $g(\cdot)$ (see, e.g., Theorem~6.4 of \cite {GH80}), we have \begin{equation} \label{eqotheorem} \int_{0}^{t}g(V_{s})\,ds= \int_{\mathbb{R}_{+}}g(x)f_{t}(x)\,dx=\int_{\mathbb {R}_{+}}g(x)\,dF_{t}(x). \end{equation} Thus, the occupation time and its density can be considered as the pathwise analogues of the cumulative distribution function and its density. Our interest in occupation times stems from the fact that they are natural measures of risk, particularly in nonstationary settings where invariant distributions do not exist; see, for example, the discussion in~\cite{BP03}. Indeed, there has been a significant interest (both theoretically and in practice) in pricing options based on the occupation times of an underlying asset; see for example, \cite {Dassios} and~\cite{Yor} and references therein. Here, we show how to measure nonparametrically\vadjust{\goodbreak} occupation times associated with the volatility risk of the price process. As a by-product, we also estimate the corresponding quantiles of the actual path of the volatility process over the fixed time interval. Since the pathwise volatility quantiles are preserved under monotone transformations, they provide a convenient way of studying the variability of the volatility and the relationship of the latter with the volatility process itself.\looseness=-1 We summarize our estimation procedure as follows. We first split the fixed time interval into blocks of decreasing length and form local estimates of the unobserved stochastic variance over each of the blocks. The volatility estimates over the blocks are truncated variations (see, e.g.,~\cite{Ma1} and~\cite{jacodprotter2012}), and we further allow for adaptive choice of the truncation level that makes use of some preliminary estimates of the stochastic variance. Then, our estimator of the volatility occupation time is simply the empirical cumulative distribution function of the local volatility estimates over the blocks. Analogously, we estimate the volatility occupation density from the local volatility estimates using kernel smoothing. Our estimation problem can be compared with the recent work of \cite {JR}. Jacod and Rosenbaum~\cite{JR}~show that an estimator of $\int_0^Tg(\sigma_s^2)\,ds$, for $g(\cdot)$ a $C^3$ function, formed by plugging in local variance estimates formed over blocks of decreasing length, can achieve the efficient $\Delta_n^{-1/2}$ rate of convergence (for $\Delta_n$ being the length of the high-frequency intervals). Similar to~\cite{JR}, our estimator here is formed by plugging local variance estimates in our function of interest. The main difference between the current work and~\cite{JR} is that in our case the function $g(\cdot)$ in (\ref{eqotheorem}) is discontinuous. As a result, the precision of estimating the volatility occupation time depends on the uniform rate of recovering the volatility process outside of the times of the ``big'' volatility jumps (with the size of the ``big'' jumps shrinking asymptotically to zero). Therefore, in the basic case when $X$ and $V$ are continuous, the rate of convergence of the volatility occupation time estimator is (almost) $\Delta_n^{-1/4}$ which, as we show in the paper, is the optimal uniform rate for recovering the volatility trajectory from high-frequency observations. By contrast,~\cite{JR} derive a central limit theorem for the convergence of their estimator to $\int_0^Tg(\sigma_s^2)\,ds$ by making use of the assumed smoothness of $g$ and applying second-order Taylor expansion of the function $g$ evaluated at the local volatility estimator in their bias-correction and asymptotic negligibility arguments. Finally, our inference for the volatility occupation time and its density can be compared with the estimation of occupation time and density of a recurrent Markov diffusion process from discrete observations of the process; see, for example,~\cite{FZ93} and \cite {BP03}. The main difference is that here the state vector, and therefore the stochastic volatility, is not fully observed. Hence, we first need to recover nonparametrically the unobserved volatility trajectory, and the error associated with recovering the volatility trajectory determines the asymptotic behavior of our estimators.\vadjust{\goodbreak} The paper is organized as follows. In Section~\ref{secsetup} we introduce the formal setup and state our assumptions. In Section~\ref{sec-c} we develop our estimator of the volatility occupation measure and prove its consistency. In Section~\ref{sec-r} we derive bounds for the rate of convergence of the volatility occupation time estimator. Section~\ref{sec-ker} derives a consistent estimator for the volatility occupation density. Section~\ref{secmc} reports results from a Monte Carlo study of our estimation technique. Section~\ref{secconcl} concludes. Section~\ref{sec-pf} contains all proofs. \section{Setup and assumptions}\label{secsetup} We start with introducing the formal setup and stating our assumptions about $X$. The process $X$ in (\ref{eqXintro}) is defined on a filtered space $(\Omega,\mathcal{F},(\mathcal{F _{t})_{t\geq0},\mathbb{P})$ with $b_t$ and $\sigma_t$ being adapted to the filtration. Further, the jump component $J_t$ is defined as \begin{eqnarray}\label{eqj} J_{t}&=&\int _{0}^{t}\int_{\mathbb{R}}\delta ( s,z ) 1_{ \{ \llvert \delta ( s,z ) \rrvert \leq1 \} } ( \mu -\nu ) ( ds,dz ) \nonumber \\[-8pt] \\[-8pt] \nonumber &&{}+\int_{0}^{t}\int _{\mathbb{R}}\delta ( s,z ) 1_{ \{ \llvert \delta ( s,z ) \rrvert >1 \} }\mu ( ds,dz ), \end{eqnarray} where $\mu$ is a Poisson measure on $\mathbb{R}_{+}\times\mathbb{R}$ with compensator $\nu$ of the form $\nu ( dt,dz ) =dt\otimes \lambda (dz ) $ for some $\sigma$-finite measure $\lambda$ on $\mathbb{R}$ and $\delta\dvtx \Omega\times\mathbb{R}_{+}\times \mathbb {R}\mapsto\mathbb{R}$ is a predictable function. Regularity conditions on $X_{t}$ are collected below. \renewcommand{\theass}{\Alph{ass}} \setcounter{ass}{0} \begin{ass}\label{assA} Let $r\in [ 0,2 ] $ be a constant. The process $X$ is an It\^{o} semimartingale given by (\ref {eqXintro}) and (\ref{eqj}), with $b_{t}$ locally bounded and $\sigma_{t}$ c\` {a}d \`{a}g. Moreover $\llvert \delta ( \omega,t,z ) \rrvert \wedge1\leq\Gamma_{m} ( z ) $ for all $ ( \omega,t,z ) $ with $t\leq\tau_{m} ( \omega ) $, where $ ( \tau _{m} ) $ is a localizing sequence of stopping times, and each function \Gamma_{m}$ on $\mathbb{R}$ satisfies $\int_{\mathbb{R}}\Gamma _{m} ( z ) ^{r}\lambda ( dz ) <\infty$. \end{ass} Assumption~\ref{assA} can be viewed as a regularity-type condition. The coefficient $r$ in Assumption~\ref{assA} controls the degree of activity of the jump component $J$ and will play an important role in the rate of convergence of our estimator. We note that $r$ provides an upper bound for the (generalized) Blumenthal--Getoor index of $J_t$; see, for example, Lemma~3.2.1 in~\cite{jacodprotter2012}. We next state our assumption for the volatility occupation time. \begin{ass}\label{assB} Fix $x\geq0$. We have $F_{T} ( \cdot ) $ a.s. differentiable with derivative $f_{T} ( \cdot ) $ in a neighborhood $\mathcal{N}_{x}$ containing $x$. Moreover, $\sup_{z\in \mathcal{N}_{x}}\mathbb{E}[f_{T} ( z ) ]<\infty$. \end{ass} Assumption~\ref{assB} is mainly concerned with the pathwise smoothness of the occupation time. The differentiability condition amounts to the existence of occupation density $f_{T} ( \cdot ) $, which is not a strong requirement; see, for example,~\cite{GH80} and~\cite{Protter}. The condition $\sup_{z\in\mathcal{N}_{x}}\mathbb{E [f_{T} ( z ) ]<\infty$ only requires the temporal average (over [ 0,T ] $) of the probability density of $V_{t}$ uniformly bounded in the neighborhood~$\mathcal{N}_{x}$, which is satisfied by most stochastic volatility models. This condition is of course much weaker than requiring $\mathbb{E[}\sup_{z\in\mathcal{N}_{x}}f_{T} ( z ) ]<\infty$, as the latter would demand more on the pathwise regularity of the occupation density. \begin{example*} Let $V_t$ solve the following stochastic differential equation: \[ dV_t = a_t^{V}\,dt+s(V_t)\,dW_t^{V}+dJ_t^{V}, \] where $a_t^{V}$ is a locally bounded process, $W_t^{V}$ is a Brownian motion, $J_t^{V}$ is a finite-variational jump process and $s (\cdot )$ has twice continuously differentiable reciprocal. Assume further that $V$ has an invariant distribution which is $C^1$ in a neighborhood of $x$. Then Assumption~\ref{assB} holds. This follows from an application of It\^o's formula and Theorem IV.75 in~\cite{Protter}. This example includes many parametric models of interest like the square-root diffusion model and the more general constant elasticity of variance model. \end{example*} For some of the results we will need a stronger condition on the volatility occupation density, mainly its continuity which we state formally in the next assumption. \renewcommand{\theass}{\Alph{ass}$^{\prime}$} \setcounter{ass}{1} \begin{ass}\label{assBprime} $F_{T} ( \cdot ) $ is a.s. continuously differentiable on $\mathbb{R}$ with derivative $f_{T} (\cdot ) $. \end{ass} Assumption~\ref{assBprime} is harder to verify than Assumption~\ref{assB}. Necessary and sufficient conditions for the continuity of the occupation density (local time) of a Borel right Markov process are discussed in~\cite{EK07}. We finally state a slightly stronger condition on the volatility process that we will need for deriving the rate of convergence of our estimator. \renewcommand{\theass}{\Alph{ass}} \setcounter{ass}{2} \begin{ass}\label{assC} The process $\sigma_{t}$ is an It\^{o} semimartingale with the form \[ \sigma_{t}=\sigma_{0}+\int_{0}^{t} \tilde{b}_{s}\,ds+\int_{0}^{t}\tilde { \sigm }_{s}\,dW_{s}+\int_{0}^{t} \tilde{\sigma}_{s}^{\prime}\,dW_{s}^{\prime }+\int _{0}^{t}\int_{\mathbb{R}}\tilde{ \delta} ( s,z ) ( \mu -\nu ) ( ds,dz ), \] where the processes $\tilde{b}$, $\tilde{\sigma}$, $\tilde{\sigma }^{\prime}$ are locally bounded and adapted, $W^{\prime}$ is a Brownian motion orthogonal to $W$, and $\tilde \delta} ( \cdot ) $ is a predictable function. Moreover $|\tilde \delta} ( \omega,t,z ) |\wedge1\leq\tilde{\Gamma }_{m} ( z ) $ for all $ ( \omega,t,z ) $ with $t\leq\tau _{m} ( \omega ) $, where $ ( \tau_{m} ) $ is a localizing sequence of stopping times, and for some $\tilde{r}\in(0,2]$, each function $\tilde \Gamma}_{m}$ on $\mathbb{R}$ satisfies $\int_{\mathbb{R}}\tilde {\Gamma _{m} ( z ) ^{\tilde{r}}\lambda ( dz ) <\infty$. \end{ass} Assumption~\ref{assC} assumes that $\sigma_t$ is an It\^o semimartingale, an assumption that is satisfied by most stochastic volatility models. We impose no restriction on the activity of the volatility jumps as well as the dependence between $\sigma_t$ and $X_t$, a~generality that is important in practical applications (particularly in finance). \section{The estimator and its consistency} \label{sec-c} We next introduce our estimator of the volatility occupation time and derive its consistency. We suppose that the process $X_t$ is observed at discrete times $i\Delta_n$, $i=0,1,\ldots,$ on $[0,T]$ for a fixed $T>0$ with the time lag $\Delta_n\rightarrow0$ when $n\rightarrow \infty $. The assumption for equidistant observations is merely for simplicity, and the theoretical results that follow (except Theorem~\ref {thmevt}) will continue to hold in the case of irregular (but nonrandom) sampling with $\Delta_n$ replaced by the mesh of the irregular observation grid. In what follows the high-frequency increment of any process $Y$ is denoted as $\Delta_{i}^{n}Y=Y_{i\Delta _{n}}-Y_{ ( i-1 ) \Delta_{n}}$. Our strategy of estimating $F_{T} ( \cdot ) $ is to first form an approximation of the volatility trajectory and then use the latter to form a sample analogue of $F_{T} ( \cdot ) $. To recover the volatility trajectory we construct local approximations for the spot variance process $V$ over blocks of shrinking length. To this end, let $k_{n}$ be a sequence of integers with $k_{n}\rightarrow\infty$ and $k_{n}\Delta_{n}\rightarrow0$. Henceforth we use the shorthand notation $u_{n}=k_{n}\Delta_{n}$. We also set a truncation process $v_{n,t}$ verifying the following assumption, which is maintained throughout the paper without further mention. \begin{ass}\label{assD} We have $v_{n,t}=\alpha_{n,t}\Delta _{n}^{\varpi}$, where $\varpi\in ( 0,1/2 ) $ is constant, and \alpha_{n,t}$ is a strictly positive real-valued process such that for some localizing sequence of stopping times $ ( \tau_{m} )$, (\sup_{t\in [ 0,T ] }(\alpha_{n,t\wedge\tau_{m}}\vee \alpha _{n,t\wedge\tau_{m}}^{-1}))_{n\geq1}$ is tight for each $m\geq1$. With this notation, for each $i=0,\ldots, \lfloor T/\Delta _{n} \rfloor-k_{n}$, we set \begin{equation}\label{eqvh} \cases{ \displaystyle\widehat{V}_{i\Delta_{n}}^{\ast}= \frac{1}{u_{n}}\sum_{j=1}^{k_{n}} \bigl( \Delta_{i+j}^{n}X \bigr) ^{2},\vspace*{2pt}\cr \displaystyle\widehat{V}_{i\Delta_{n}}=\frac{1} u_{n}}\sum_{j=1}^{k_{n}} \bigl( \Delta_{i+j}^{n}X \bigr) ^{2}1_{ \{ \llvert \Delta_{i+j}^{n}X\rrvert \leq v_{n,i\Delta_n} \} }.} \end{equation} Here, $\widehat{V}_{i\Delta_{n}}^{\ast}$ is a local approximation of $ V_{i\Delta_{n}}$ when $X$ is continuous, while $\widehat{V}_{i\Delta_{n}}$ serves the same purpose but is robust to the presence of jumps in~$X$. As a generalization to the standard truncation-based methods (see e.g., Chapter~9 of~\cite{jacodprotter2012}), we allow explicitly the truncation parameter $\alpha_{n,t}$ to be time-varying and depend on $\{X_{i\Delta_n}\}_{i=1,\ldots,\lfloor T/\Delta _n\rfloor}$. For example, one convenient and commonly used choice is to set $\alpha _{n,t} = c \overline{\sigma}_n$, where $c$ is a constant [typically in the range $(3,5)$], and $\overline {\sigma}_n$ is a preliminary estimate of the average volatility over $[0,T]$. Assumption~\ref{assD} is verified as soon as $\overline{\sigma}_n$ and $\overline{\sigma}_n^{-1}$ are tight. Another possibility is to make $\alpha_{n,t}$ adaptive by setting $\alpha_{n,t} = c \hat{\sigma}_{i}$ for $t\in[(i-1)u_n, iu_n)$, where $\hat{\sigma}_{i}$ is a preliminary estimate for the volatility in the local window $[(i-1)u_n, iu_n)$. For example, one may take $\hat{\sigma}_{i}$ to be a localized version of the Bipower variation estimator of~\cite{BNS04a}: $\hat{\sigma}_{i} = ((\pi/2)u_n^{-1}\sum_{j=1}^{k_n} |\Delta^n_{i+j} X||\Delta^n_{i+j+1} X|)^{1/2}$.\vadjust{\goodbreak} The tightness requirement in Assumption~\ref{assD} can be easily fulfilled by replacing $\hat{\sigma}_{i}$ with $(\hat{\sigma}_{i}\vee(1/C))\wedge C$ for some pre-specified regularization constant $C\geq1$. Finally, in the above two examples for $\alpha_{n,t}$, we can further replace the constant $c$ with a deterministic sequence $c_n$ increasing at a logarithmic rate as $\Delta _n\rightarrow0$. We will use $\widehat{V}_{i\Delta_{n}}^{\ast}$ and $\widehat {V}_{i\Delta_{n}}$ to approximate for the volatility trajectory within the block. That is for $0\leq i\leq \lfloor T/u_{n} \rfloor-1$, \begin{equation}\label{eqvp} \cases{ \widehat{V}_{t}^{\ast}= \widehat{V}_{iu_{n}}^{\ast }\quad\mbox{and}\quad\widehat{V _{t}=\widehat{V}_{iu_{n}},&\quad $t\in\bigl\lbrack iu_{n}, ( i+1 ) u_{n}\bigr)$, \vspace*{2pt}\cr \widehat{V}_{t}^{\ast}=\widehat{V}_{ ( \lfloor T/u_{n} \rfloor-1 ) u_{n}}^{\ast} \quad\mbox{and}\quad\widehat {V}_{t}=\widehat{V} _{ ( \lfloor T/u_{n} \rfloor-1 ) u_{n}}, &\quad $\lfloor T/u_{n} \rfloor u_{n}\leq t\leq T$. }\hspace*{-35pt} \end{equation} \end{ass} \begin{remark} We can alternatively define local estimators of volatility for each $i=1,\ldots,\lfloor T/\Delta_n\rfloor$ by averaging the $k_n$ past squared increments below the threshold. All the results in the paper, except for Theorem~\ref{thmevt} below, will hold for this alternative way of recovering the spot volatility. \end{remark} Using $\widehat{V}_{t}^{\ast}$ and $\widehat{V}_{t}$, our proposed estimators of $F_{T} ( \cdot ) $ are defined as \[ \widehat{F}_{n,T}^{\ast} ( x ) =\int_{0}^{T}1_{ \{ \widehat{V _{s}^{\ast}\leq x \} }\,ds,\qquad \widehat{F}_{n,T} ( x ) =\int_{0}^{T}1_{ \{ \widehat{V}_{s}\leq x \} }\,ds,\qquad x\in\mathbb{R}. \] We first consider the pointwise consistency of $\widehat {F}_{n,T}^{\ast } ( x ) $ and $\widehat{F}_{n,T} ( x ) $. As a matter of fact, it is not much harder to prove a more general result as follows. \begin{lemma} \label{lem-1}Let $g\dvtx \mathbb{R}_{+}\mapsto [ 0,1 ] $ be a measurable function and $D_{g}$ be the collection of discontinuity points of $g$. Suppose: \begin{enumerate}[(ii)] \item[(i)] Assumption~\ref{assA} holds for $r=2$; \item[(ii)] for Lebesgue a.e. $t\i [ 0,T ] $, $\mathbb{P} ( V_{t}\in D_{g} ) =0$. Then we have \begin{enumerate} \item[(a)] \begin{equation} \int_{0}^{T}g(\widehat{V}_{s})\,ds \stackrel{\mathbb {P}} {\longrightarrow \int_{0}^{T}g ( V_{s} ) \,ds; \label{thm-1a} \end{equation} \item[(b)] if, in addition, $X$ is continuous, then (\ref{thm-1a}) also holds when replacing $\widehat{V}$ with $\widehat{V}^{\ast}$. \end{enumerate} \end{enumerate} \end{lemma} Lemma~\ref{lem-1} extends Theorem~9.4.1 in~\cite{jacodprotter2012} by allowing for discontinuities in the test function $g ( \cdot ) $. For fixed $x\geq0$, the pointwise consistency of $\widehat {F}_{n,T} ( x ) $, and $\widehat{F}_{n,T}^{\ast} ( x ) $ if $X$ is continuous, follows immediately [with $g ( \cdot ) =1_{ \{ \cdot \leq x \} }$] provided that $\mathbb{P} ( V_{t}=x ) =0$ for Lebesgue a.e. $t\in [ 0,T ] $. The uniform consistency of \widehat{F}_{n,T}^{\ast} ( \cdot ) $ and $\widehat {F}_{n,T} ( \cdot ) $ is available if the occupation time $F_{T} ( \cdot ) $ is a.s. continuous, as shown below. \begin{theorem} \label{thm-1}Suppose Assumption~\ref{assA} holds for $r=2$ and $F_{T} ( \cdot ) $ is a.s. continuous. We have:\vadjust{\goodbreak} \begin{longlist}[(a)] \item[(a)] $\sup_{x\in\mathbb{R}} \widehat{F}_{n,T} ( x ) -F_{T} ( x ) |\stackrel {\mathbb {P}} \longrightarrow}0$; \item[(b)] if, in addition, $X$ is continuous, then \textup{(a)} still holds when replacing $\widehat{F}_{n,T}$ with $\widehat {F}_{n,T}^{\ast}$. \end{longlist} \end{theorem} Analogous to the classical notion of quantile for cumulative distribution functions, the quantile of the occupation time $F_{T} ( \cdot ) $ is naturally defined as the left-continuous functional inverse of F_{T} ( \cdot ) $: for $\alpha\in ( 0,T ) $, we set Q_{T} ( \alpha ) =\inf\{x\in\mathbb{R}\dvtx F_{T} ( x ) \geq \alpha\}$. A natural estimator for $Q_{T} ( \alpha ) $ is $ \widehat{Q}_{n,T} ( \alpha ) =\inf\{x\in\mathbb {R}\dvtx \widehat {F _{n,T} ( x ) \geq\alpha\}$, and $\widehat{Q}_{n,T}^{\ast } ( \alpha ) $ can be defined analogously for $\widehat{F}_{n,T}^{\ast } ( \cdot ) $. The consistency of the quantile estimators is given by the next corollary. \begin{corollary} \label{cor-Q}Suppose Assumption~\ref{assA} holds for $r=2$ and $F_{T} ( \cdot ) $ is a.s. continuous. Let $\mathcal{Q}\equiv\{\alpha \in ( 0,T ) \dvtx Q_{T} ( \cdot ) $ is continuous at $\alpha$ a.s.$\}$. We have for each $\alpha\in\mathcal{Q}$: \begin{longlist}[(a)] \item[(a)] $\widehat{Q _{n,T} ( \alpha ) \stackrel{\mathbb{P}}{\longrightarrow Q_{T} ( \alpha ) $; \item[(b)]if, in addition, $X$ is continuous, then \widehat{Q}_{n,T}^{\ast} ( \alpha ) \stackrel{\mathbb {P}} \longrightarrow}Q_{T} ( \alpha ) $. \end{longlist} \end{corollary} \section{\texorpdfstring{Rate of convergence of $\widehat{F}_{n,T}$} {Rate of convergence of F n,T}}\label{sec-r} We next study the rate of convergence of $\widehat{F}_{n,T} ( x )$. We first consider in Section~\ref{sec-rc} the case when $V$ is continuous and then the general case with discontinuous $V$ is studied in Section~\ref{sec-rd}, where the uniform rate of convergence of $\widehat{F}_{n,T}(\cdot)$ is also considered. \subsection{The continuous volatility case and the uniform approximation of $V$} \label{sec-rc} In the continuous volatility case we can link the rate of convergence of our volatility occupation time estimators with the rate of convergence of $\widehat{V}_{t}^{\ast}$ and $\widehat{V}_{t}$ toward $V_t$ on the space of c\`{a}dl\`{a}g functions equipped with the uniform norm. We denote the latter as \[ \eta_{n}^{\ast}=\sup_{t\in [ 0,T ] }\bigl\llvert \widehat {V _{t}^{\ast}-V_{t}\bigr \rrvert, \qquad \eta_{n}=\sup_{t\in [ 0, ] }\llvert \widehat{V}_{t}-V_{t}\rrvert. \] The rate of convergence\vspace*{1pt} of $\widehat{F}_{n,T}$ and $\widehat{Q}_{n,T}$ is then related with $\eta_{n}$ through Lemma~\ref{lem-fu} below; an analogous result holds for $\widehat{F}_{n,T}^{\ast}$, $\widehat {Q}_{n,T}^{\ast}$ and $\eta_{n}^{\ast}$, but is omitted here for brevity. \begin{lemma}\label{lem-fu} \textup{(a)} For any $x\geq0$ and $\alpha\in ( 0,T ) $, we have $|\widehat{F}_{n,T} ( x ) -F_{T} ( x ) |\leq F_{T}(x+\eta_{n})-F_{T}(x-\eta_{n})$ and $|\widehat{Q}_{n,T}(\alpha )-Q_{T}(\alpha)|\leq\eta_{n}$. \textup{(b)} Suppose Assumption~\ref{assB} and $\eta_{n}=O_{p} ( a_{n} ) $ for some nonrandom sequence $a_{n}\rightarrow0$. Then $\widehat {F}_{n,T} (x ) -F_{T} ( x ) =O_{p} ( a_{n} ) $. \end{lemma} In view of Lemma~\ref{lem-fu}, bounding the rate of convergence of the occupation time estimators boils down to\vadjust{\goodbreak} establishing the asymptotic order of magnitude of $\eta_{n}$ and~$\eta_{n}^{\ast}$. Our main result concerning the uniform approximation of the $V$ process is given by the following theorem. \begin{theorem} \label{thm-u}Suppose Assumptions~\ref{assA} and~\ref{assC} with $\Delta V_s=0$ for $s\in [0,T]$. Let $k_{n}\asymp\Delta _{n}^{-\gamma}$ for some $\gamma\in(r\varpi+(1\vee r)(1-2\varpi ),1)$, \varpi\in((1\vee r-1)/(2 ( 1\vee r ) -r),1/2)$ and $\iota >0$ be arbitrarily small but fixed. We have: \begin{longlist}[(a)] \item[(a)] $\eta_{n}=O_{p} ( a_{n} ) $, where \begin{equation}\label{an} \qquad a_{n}=\cases{ \Delta_{n}^{\gamma-1+ ( 2-r ) \varpi} \vee\Delta _{n}^{\gamma /2-\iota}\vee\Delta_{n}^{ ( 1-\gamma ) /2-\iota}, &\quad $\mbox {if } r\leq1$, \vspace*{2pt}\cr \Delta_{n}^{\gamma/r- ( 1-\varpi ) -\iota}\vee\Delta _{n}^{\gamma/2-\iota} \vee\Delta_{n}^{ ( 1-\gamma ) /2-\iota}, & \quad$\mbox{if } r>1$, \vspace*{2pt}\cr \Delta_{n}^{\gamma/2-\iota}\vee\Delta_{n}^{ ( 1-\gamma ) /2-\iota}, &\quad $\mbox{if } X\mbox{ is continuous;}$} \end{equation} \item[(b)] if $X$ is continuous, we also have $\eta_{n}^{\ast}=O_{p} (a_{n} ) $. \end{longlist} \end{theorem} When $X$ is continuous, the terms $\Delta_{n}^{\gamma/2-\iota}$ and $\Delta_{n}^{ ( 1-\gamma ) /2-\iota}$ capture, respectively, the sampling variability and the discretization bias in the approximation of the spot variance. When $X$ is discontinuous, $a_{n}$ contains an additional term arising from the elimination of jumps, which of course depends on the concentration of ``small'' jumps through $r$ (recall Assumption~\ref{assA}). The conditions on $\varpi$ and $\gamma$ imply $a_n\to0$. In particular, when $r$ is close to $2$, $\varpi$ and $\gamma$ need to be chosen close to $1/2$ and $1$, respectively, to ensure that $a_{n}\rightarrow0$, rendering the rate of convergence arbitrarily slow. Nonetheless, if $X$ is discontinuous, $\eta_{n}$ still has the same rate of convergence as in the continuous case, that is, $\Delta_{n}^{1/4-\iota}$, provided $r\in ( 0,1/2 ) $. This rate can be achieved by setting $\varpi\in (3/ ( 8-4r ),1/2 ) $ and $\gamma=1/2$. Of course Theorem~\ref{thm-u} provides only a bound for the rate of convergence of $\eta_n$ and $\eta_n^*$. The following theorem, however, establishes the exact asymptotic distributions of $\eta_n$ and $\eta^*_n$ in a simple model with constant volatility. \begin{theorem}\label{thmevt} Suppose: \begin{longlist} \item[(i)] Assumption~\ref{assA} holds with $V_t$ constant and $b_t=0$ on $[0,T]$; \item[(ii)] $r<1/2$ and $\varpi\in(3/(8-4r),1/2)$. Then \begin{equation} \label{thmevt1} \sqrt{\log\bigl(\lfloor T/u_n\rfloor\bigr)} ( \sqrt{k_n}\eta_n - \sqrt {2}Vm_{n} ) \stackrel{\mathcal{L}} {\longrightarrow} V\times \Lambda, \end{equation} provided $k_n \asymp\Delta_n^{-1/2}$ and where $\Lambda$ is a random variable with c.d.f.\break $\exp(-2\exp(-x))$, and \begin{equation} \label{thmevt2} m_{n} = \sqrt{2\log\bigl(\lfloor T/u_n \rfloor\bigr)} - \frac{\log(\log (\lfloor T/u_n\rfloor))+\log(4\pi)}{2\sqrt{2\log(\lfloor T/u_n\rfloor)}}. \end{equation} If we further assume \textup{(iii)} $X_t$ is continuous, then (\ref{thmevt1}) still holds with $\eta_n$ replaced by $\eta_n^*$. \end{longlist} \end{theorem} \begin{remark} Theorem~\ref{thmevt} shows that the rates given in (\ref{an}) are almost optimal when the jumps of $X_t$ are not very active ($r<1/2$). To be precise, we observe that (\ref{an}) suggests $\eta_n = O_p(\Delta_n^{1/4-\iota})$ for $\iota>0$ fixed but arbitrarily small, while the optimal rate in Theorem~\ref{thmevt} provides a slightly sharper bound $\eta_n = O_p(\Delta_n^{1/4}\log (\lfloor T\Delta_n^{-1/2}\rfloor)^{1/2})$. \end{remark} The rate of convergence of our volatility occupation time estimators and their quantiles is a direct corollary of Lemma~\ref{lem-fu} and Theorem~\ref{thm-u}; the proof is omitted for brevity. \begin{corollary} \label{cor-ratec}Let $x\geq0$ and $\alpha\in(0,T)$. Suppose Assumption~\ref{assB} and the same setting as in Theorem~\ref{thm-u}. Then $\widehat {F}_{n,T} ( x ) -F_{T} ( x ) $ and $\widehat {Q}_{n,T} ( \alpha ) -Q_{T} ( \alpha ) $ are $O_{p} ( a_{n} ) . If $X$ is continuous and $k_{n}\asymp\Delta_{n}^{-1/2}$, then $\widehat{ }_{n,T} ( x ) -F_{T} ( x ) $, $\widehat {F}_{n,T}^{\ast } ( x ) -F_{T} ( x ) $, $\widehat {Q}_{n,T}(\alpha )-Q_{T} ( \alpha ) $ and $\widehat{Q}_{n,T}^{\ast}(\alpha )-Q_{T} ( \alpha ) $ are $O_{p}(\Delta_{n}^{1/4-\iota})$ for \iota>0$ arbitrarily small but fixed. \end{corollary} We should point out that in the trivial cases when $x<\inf_{t\in [0,T]}V_t$ or $x>\sup_{t\in[0,T]}V_t$ on a given path, the error in recovering the occupation time will become identically zero for $n$ sufficiently high (up to taking a subsequence). \begin{remark} More generally, we can use Theorem~\ref{thm-u} to show in the setting of the theorem that if $\mathcal{L}\dvtx \mathcal{D}([0,T])\rightarrow \mathbb{R}$, where $\mathcal{D}([0,T])$ is the space of c\`{a}dl\`{a}g functions on the interval $[0,T]$ equipped with the uniform topology, is a continuous function, we have $\mathcal{L}(\widehat{V})\stackrel {\mathbb{P}}{\longrightarrow}\mathcal{L}(V)$. If further $|\mathcal {L}(f)-\mathcal{L}(g)|\leq K\sup_{s\in[0,T]}|f_s-g_s|$ for any elements $f, g\in\mathcal{D}([0,T])$ and some positive constant $K$, then the rate of convergence of $\mathcal{L}(\widehat{V})$ to $\mathcal{L}(V)$ is bounded by the order of magnitude of $\eta_n$ (under the conditions of Theorem~\ref{thm-u}). An example of such a function is $\mathcal {L}(f) = \sup_{s\in[0,T]}f_s$. \end{remark} \subsection{The discontinuous volatility case} \label{sec-rd} We now turn to the general case when the volatility process contains jumps. When $V$ is discontinuous, bounding the rate of convergence of the volatility occupation time estimators is much less straightforward. Lemma~\ref{lem-fu} is still valid, however, and the uniform approximation error $\eta_{n}$ no longer vanishes asymptotically, due to the discontinuity in $V$. This is true even if we consider the ideal case where $\widehat{V}_{iu_{n}}=V_{iu_{n}}$, that is, perfect pointwise approximation is available. Indeed, the lack of uniform approximation for a discontinuous process with its discretized version is well known in the study of convergence of processes. Our strategy of bounding the rate of convergence of $\widehat {F}_{n,T} ( x )$ and $\widehat{F}_{n,T}^* ( x )$ is to pick out the ``big'' jumps in the $V$ process and then consider the uniform rate of approximation to the ``remainder'' process. The idea is best illustrated in the basic case where the jumps in $V$ is finitely active. In this case, the volatility jumps only occur within finitely many time blocks with the form $[iu_{n}, ( i+1 ) u_{n})$ and hence their total effect on the estimation is $O_{p} ( u_{n} )$. On time blocks not containing the jumps of $V$, $V$ is continuous, so Theorem~\ref{thm-u} can be used to provide uniform bound. The situation becomes considerably more complicated when $V$ has infinitely active, or even infinite variational, jumps. In this case, one needs to compute the trade-off between picking out a smaller number of big jumps with less accurate uniform approximation to the remainder process, and picking out a larger number of big jumps with more accurate uniform approximation to the remainder process. The end result of this calculation is Theorem~\ref{thm-rd} below. \begin{theorem} \label{thm-rd}Suppose Assumptions~\ref{assA},~\ref{assB} and~\ref{assC}. Let $k_{n}\asymp\Delta _{n}^{-\gamma}$ for some $\gamma\in(r\varpi+(1\vee r)(1-2\varpi ),1)$, \varpi\in((1\vee r-1)/(2 ( 1\vee r ) -r),1/2)$ and $\iota >0$ be arbitrarily small but fixed. We have: \begin{longlist}[(a)] \item[(a)] $\widehat{F}_{n,T} ( x ) -F_{T} ( x ) =O_{p} ( d_{n} ) $, where $d_{n}=a_{n}\vee \Delta_{n}^{ ( 1-\gamma ) / ( 1+\tilde{r} ) -\iota}$ and $a_{n}$ is given by (\ref{an}); \item[(b)] if $X$ is continuous, we also have \widehat{F}_{n,T}^{\ast} ( x ) -F_{T} ( x ) =O_{p} ( d_{n} ) $. \end{longlist} \end{theorem} Theorem~\ref{thm-rd} establishes an upper bound for the pointwise rate of convergence of the occupation time estimators. We remind the reader that the rate $d_n$ depends on $r$ through $a_n$; recall (\ref{an}) and the discussion following Theorem~\ref{thm-u}. Whether the rate is optimal or not is an open question. The rate optimality for jump-robust estimation of the integrated variance, that is, $\int_0^T V_s\, ds = \int_0^{\infty} x F_T(dx)$, is studied by~\cite{jacodreiss2012}. In order to establish a uniform bound, we invoke the stronger Assumption~\ref{assBprime} which assumes continuity of the volatility occupation density. \begin{theorem} \label{thm-rdu}Consider the same setting as in Theorem~\ref{thm-rd} except with Assumption~\ref{assBprime} replacing Assumption~\ref{assBprime}. Then \textup{(a)} and \textup{(b)} in Theorem~\re {thm-rd} hold uniformly in $x\in\mathbb{R}$. Moreover, for $\alpha \in(0,T) $ with $f_{T}(Q_{T}(\alpha))>0$ a.s., we have $|\widehat {Q}_{n,T}(\alpha )-Q_{T}(\alpha)|=O_{p} ( d_{n} ) $ and, if $X$ is continuous, $ \widehat{Q}_{n,T}^{\ast}(\alpha)-Q_{T}(\alpha)|=O_{p} ( d_{n} ) $. \end{theorem} Theorem~\ref{thm-rdu} establishes a bound for the uniform rate of convergence of the occupation time estimators; the rate of convergence of the quantiles follows as a consequence. \section{Estimation of the volatility occupation density} \label{sec-ker} We now turn to estimating the volatility occupation density $f_{T}(x)$. Clearly, the uniform convergence of the occupation time (Theorem~\ref {thm-1 ) does not directly lead to valid estimation for the occupation density. While the focus of the current paper is on the occupation time, we consider the estimation of $f_{T}(\cdot)$ theoretically complementary and empirically relevant. Our occupation density estimator is based on the local volatility estimates $\widehat{V}_{t}$ and $\widehat{V}_{t}^{\ast}$ and kernel smoothing. In particular, we propose the following kernel estimator of Nadaraya--Watson type: \[ \widehat{f}_{n,T} ( x ) \equiv\int_{0}^{T} \frac {1}{h_{n}}\kappa \biggl(\frac{\widehat{V}_{s}-x}{h_{n} \biggr)\,ds, \] where $h_{n}\rightarrow0$ is a bandwidth sequence and the kernel function \kappa\dvtx \mathbb{R}\mapsto\mathbb{R}_{+}$ is bounded $C^{1}$ with bounded derivative and $\int_{\mathbb{R}}\kappa ( x ) \,dx=1$. We can define $\widehat{f}_{n,T}^{\ast}(x)$ similarly but with $\widehat {V}_{s}^{\ast}$ replacing $\widehat{V}_{s}$. Below, we consider a weight function $w\dvtx \mathbb{R}\mapsto\mathbb{R}_{+}$ with $\int_{\mathbb{R}}w ( x ) \,dx<\infty$. For generic real-valued functions $g_{1}$ and $g_{2}$ on $\mathbb{R}$, we denote \[ \llVert g_{1}-g_{2}\rrVert _{w}\equiv\int _{\mathbb{R}}\bigl\llvert g_{1} ( x ) -g_{2} ( x ) \bigr\rrvert w ( x ) \,dx. \] \begin{theorem} \label{thm-kd}Suppose: \begin{longlist}[(iii)] \item[(i)] Assumptions~\ref{assA},~\ref{assBprime} and~\ref{assC}; \item[(ii)] $r\in ( 0,2 ) $ and $\varpi\in((1\vee r-1)/(2(1\vee r)-r),1/2)$; \item[(iii)] $k_{n}\asymp\Delta_{n}^{-\gamma}$ for some $\gamma\in ( 0,1 ) $; \item[(iv)] $V_{t}^{-1}$ is locally bounded; \item[(v)] for some $\beta\in(0,1]$ and any compact $\mathcal{K}\subset ( 0,\infty ) $, there exists a constant $C_{\mathcal{K}}>0$, such that for all $x,y\in\mathcal{K}$, \mathbb{E}|f_{T}(x)-f_{T}(y)|\leq C_{\mathcal{K}}|x-y|^{\beta}$; \item[(vi)] \int_{\mathbb{R}}\kappa ( z ) \llvert z\rrvert ^{\beta }\,dz<\infty$. \end{longlist} We set \[ \bar{a}_{n}^{\ast}\equiv\Delta_{n}^{\gamma/2} \vee\Delta _{n}^{ ( 1-\gamma ) /2},\qquad \bar{a}_{n}\equiv \bar{a}_{n}^{\ast}\vee \Delta _{n}^{{(1-r\varpi-\theta)}/{(1\vee r)}- ( 1-2\varpi ) }, \] where $\theta=0$ when $r\leq1$ and $\theta>0$ is arbitrarily fixed when r>1$. Then for each $x\geq0$, we have: \begin{longlist}[(a)] \item[(a)] $\widehat{f}_{n,T} ( x ) -f_{T} ( x ) $ and $\Vert\widehat{f}_{n,T}-f_{T}\Vert_{w}$ are O_{p}(h_{n}^{-2}\bar{a}_{n}\vee h_{n}^{\beta})$; \item[(b)] if $X$ is continuous, \widehat{f}_{n,T} ( x ) -f_{n,T} ( x ) $ and $\Vert \widehat{f}_{n,T}-f_{T}\Vert_{w}$ are $O_{p}(h_{n}^{-2}\bar {a}_{n}^{\ast }\vee h_{n}^{\beta})$ and moreover, the results still hold with $\widehat{f _{n,T}(\cdot)$ replaced by $\widehat{f}_{n,T}^{\ast}(\cdot)$. \end{longlist} \end{theorem} \begin{remark} Condition (v) in Theorem~\ref{thm-kd} requires the occupation density of V_{t}$ to be H\"{o}lder continuous on compacta with exponent $\beta$ under the $L_1$-norm. We preclude the analysis for cases in which $f_{T}(\cdot)$ is differentiable, or in a H\"{o}lder class of higher order, because occupation densities of semimartingales in general do not enjoy such higher-order smoothness; recall from Assumption~\ref{assC} that $V_t$ is a semimartingale. For example, the occupation density of a one-dimensional Brownian motion is H\"{o}lder continuous in $L_1$ with exponent $\beta= 1/2$; see Exercise VI.1.32 in~\cite{revuzyor}. That being said, occupation densities of other processes, such as certain Gaussian processes (see, e.g., Table~2 in~\cite{GH80}), may enjoy higher-order smoothness. Such models have rarely been studied in the analysis of high-frequency financial data and are not directly compatible with Assumption~\ref{assC}, so we do not pursue further results here. Notice that the rate $\bar {a}_{n}^{\ast}$ is optimized by setting $\gamma=1/2$, resulting in $\bar{a}_{n}^{\ast}=\Delta_{n}^{1/4}$. Furthermore, when $X$ is continuous, the estimation error of the occupation density is $O_{p}(\Delta_{n}^{\beta/4(2+\beta)})$, which is achieved by setting $h_{n}\asymp\Delta_{n}^{1/4(2+\beta)}$. Not surprisingly, the smoother the occupation density (larger $\beta$), the faster the rate of convergence. \end{remark} \begin{remark} Theorem~\ref{thm-kd}(a) implies $\widehat{f}_{n,T} ( x ) -f_{T} ( x ) \stackrel{\mathbb{P}}{\longrightarrow}0$ and $\Vert \widehat{f}_{n,T}-f_{T}\Vert_{w}\stackrel{\mathbb {P}}{\longrightarrow}0$, provided that $h_{n}\rightarrow0$ and $h_{n}^{-2}\bar {a}_{n}\rightarrow0$. These results can be shown directly without conditions (iv)--(vi) using a very similar proof; the details are omitted for brevity. A similar comment applies to Theorem~\ref{thm-kd}(b). \end{remark} \setcounter{equation}{0} \section{Monte Carlo}\label{secmc} We test the performance of our nonparametric procedures on two popular stochastic volatility models. The first is the square-root diffusion volatility model, given by \begin{equation} \label{mc1} dX_t = \sqrt{V_t}\,dW_t,\qquad dV_t = 0.03(1.0-V_t)\,dt+0.2\sqrt{V_t}\,dB_t, \end{equation} $W_t$ and $B_t$ are two independent Brownian motions. Our second model is a jump-diffusion volatility model in which the log-volatility is a L\'{e}vy-driven Ornstein--Uhlenbeck (OU) process, that is, \begin{equation} \label{mc2} dX_t = e^{V_t-1}\,dW_t,\qquad dV_t = -0.03V_t\,dt+\,dL_t, \end{equation} where $L_t$ is a L\'{e}vy martingale uniquely defined by the marginal law of $V_t$ which in turn has a selfdecomposable distribution (see Theorem~17.4 of~\cite{SATO}) with characteristic triplet (Definition~8.2 of \cite{SATO}) of $(0,1,\nu)$ for $\nu(dx) = \frac{2.33e^{-2.0|x|}}{|x|^{1+0.5}}1_{\{x>0\}}\,dx$ with respect to the identity truncation function. The mean and persistence of both volatility specifications are calibrated realistically to observed financial data, and the two models differ in the presence of volatility jumps as well as in the modeling of the volatility of volatility: for model (\ref{mc1}), the transformation $\sqrt{V_t}$ is with constant diffusion coefficient while for (\ref{mc2}) this is the case for the transformation $\log{V_t}$. In the Monte Carlo we fix the time span to $T=22$ days (our unit of time is a day), equivalent to one calendar month, and we consider $n=80$ and $n=400$, which correspond to $5$-minute and $1$-minute, respectively, of intraday observations of $X$ in a $6.5$-hour trading day. We set $k_n=20$ for $n=80$ and we increase it to $k_n=40$ when $n=400$, which, respectively, correspond to $4$ and $10$ blocks per unit of time. We finally set the truncation process at $v_{n,t} = 3\sqrt {BV_j}\Delta_n^{0.49}$ for $t\in[j-1,j)$ and where $BV_j = \frac{\pi }{2}\sum_{i=\lfloor(j-1)/\Delta_n\rfloor+2}^{\lfloor j/\Delta _n\rfloor }|\Delta_{i-1}^nX||\Delta_{i}^nX|$\vspace*{2pt} is the Bipower Variation on the unit interval $[j-1,j)$. For each realization we compute the $25$th, $50$th and $75$th volatility qunatiles over the interval $[0,T]$. The results from the Monte Carlo are summarized in Table~\ref{tbmc}. Overall, the performance of our volatility quantile estimator is satisfactory. The highest bias arises for the square-root diffusion volatility model when volatility was started from a high value (the $75$th quantile of its invariant distribution). Intuitively, in this case volatility drifts toward its unconditional mean, and this results in its larger variation over $[0,T]$, which in turn is more difficult to accurately disentangle from the Gaussian noise in the price process, that is, the Brownian motion $W_t$ in~$X_t$. Consistently with our asymptotic results, the biases and the mean absolute deviations of all volatility quantiles shrink as we increase the sampling frequency from $n=80$ to $n=400$ in all considered scenarios. \begin{sidewaystable} \tablewidth=\textwidth \caption{Monte Carlo results} \label{tbmc} \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}lccccccccc@{}} \hline & \multicolumn{3}{c}{$\bolds{\widehat{Q}_{T,n}(0.25)}$} & \multicolumn{3}{c}{$\bolds{\widehat{Q}_{T,n}(0.50)}$} & \multicolumn {3}{c}{$\bolds{\widehat{Q}_{T,n}(0.75)}$}\\[-6pt] & \multicolumn{3}{c}{\hrulefill} & \multicolumn{3}{c}{\hrulefill} & \multicolumn {3}{c@{}}{\hrulefill}\\ \textbf{Start value}& \textbf{True} & \textbf{Bias} & \textbf{MAD} & \textbf{True} & \textbf{Bias} & \textbf{MAD} & \textbf{True} & \textbf{Bias} & \textbf{MAD}\\ \hline \multicolumn{10}{c}{Panel A: Square-root volatility model, $n=80$} \\ $V_0 = Q^V(0.25)$ & $0.3798$ & $-0.0536$ & $0.0547$ & $0.5394$ & $-0.0478$ & $0.0514$ & $0.7324$ & $-0.0190$ & $0.0473$ \\ $V_0 = Q^V(0.50)$ & $0.6223$ & $-0.0916$ & $0.0929$ & $0.8170$ & $-0.0651$ & $0.0703$ & $1.0513$ & $-0.0081$ & $0.0626$ \\ $V_0 = Q^V(0.75)$ & $0.9865$ & $-0.1516$ & $0.1525$ & $1.2359$ & $-0.0949$ & $0.1027$ & $1.5310$ & $ 0.0110$ & $0.0911$ \\[3pt] \multicolumn{10}{c}{Panel B: Square-root volatility model, $n=400$} \\ $V_0 = Q^V(0.25)$ & $0.3798$ & $-0.0305$ & $0.0315$ & $0.5394$ & $-0.0304$ & $0.0327$ & $0.7324$ & $-0.0178$ & $0.0293$ \\ $V_0 = Q^V(0.50)$ & $0.6223$ & $-0.0519$ & $0.0529$ & $0.8170$ & $-0.0412$ & $0.0453$ & $1.0513$ & $-0.0146$ & $0.0375$ \\ $V_0 = Q^V(0.75)$ & $0.9865$ & $-0.0868$ & $0.0882$ & $1.2359$ & $-0.0596$ & $0.0654$ & $1.5310$ & $-0.0043$ & $0.0554$ \\[3pt] \multicolumn{10}{c}{Panel C: Log-volatility model, $n=80$} \\ $V_0 = Q^V(0.25)$ & $0.1737$ & $-0.0231$ & $0.0249$ & $0.2860$ & $-0.0269$ & $0.0302$ & $0.4519$ & $-0.0171$ & $0.0358$ \\ $V_0 = Q^V(0.50)$ & $0.3293$ & $-0.0428$ & $0.0455$ & $0.5243$ & $-0.0460$ & $0.0524$ & $0.8069$ & $-0.0245$ & $0.0610$ \\ $V_0 = Q^V(0.75)$ & $0.6337$ & $-0.0809$ & $0.0866$ & $0.9945$ & $-0.0807$ & $0.0968$ & $1.5162$ & $-0.0434$ & $0.1117$ \\[6pt] \multicolumn{10}{c}{Panel D: Log-volatility model, $n=400$} \\ $V_0 = Q^V(0.25)$ & $0.1737$ & $-0.0131$ & $0.0142$ & $0.2860$ & $-0.0158$ & $0.0180$ & $0.4519$ & $-0.0116$ & $0.0224$ \\ $V_0 = Q^V(0.50)$ & $0.3293$ & $-0.0248$ & $0.0268$ & $0.5243$ & $-0.0276$ & $0.0318$ & $0.8069$ & $-0.0169$ & $0.0358$ \\ $V_0 = Q^V(0.75)$ & $0.6337$ & $-0.0452$ & $0.0490$ & $0.9945$ & $-0.0480$ & $0.0575$ & $1.5162$ & $-0.0305$ & $0.0682$ \\ \hline \end{tabular*} \tabnotetext[]{}{\textit{Notes}: In all simulated scenarios $T=22$, and we set $k_n = 20$ for $n=80$ and $k_n = 40$ for $n=400$. In each of the cases, the volatility is started from a fixed point being the $25$th, $50$th and $75$th quantile of the invariant distribution of the volatility process, denoted correspondingly as $Q^V(0.25)$, $Q^V(0.50)$ and $Q^V(0.75)$. The columns ``True'' report the average value (across the Monte Carlo simulations) of the true variance quantile that is estimated; MAD stands for mean absolute deviation around true value. The Monte Carlo replica is $1000$.} \end{sidewaystable} \section{Conclusion} \label{secconcl} In this paper we propose nonparametric estimators of the volatility occupation time and its density from discrete observations of the process over a fixed time interval with asymptotically shrinking mesh of the observation grid. We derive the asymptotic properties of our volatility occupation time estimator and further invert it to estimate the corresponding quantiles of the volatility path over the fixed time interval. Monte Carlo shows satisfactory performance of the proposed estimation techniques. \section{Proofs}\label{sec-pf} This section contains all proofs. Throughout the proof, we denote by $K$ a generic constant that may change from line to line. We sometimes emphasize its dependence on some parameter $p$ by writing $K_{p}$. As is typical in this kind of problem, by a standard localization procedure, Assumptions~\ref{assA}, \ref{assC} and~\ref{assD} can be strengthened into the following stronger versions without loss of generality. \renewcommand{\theass}{S\Alph{ass}} \setcounter{ass}{0} \begin{ass}\label{assSA} We have Assumption~\ref{assA}. The processes $ b_{t}$ and $\sigma_{t}$ are bounded, and for some bounded nonnegative function $\Gamma$ on $\mathbb{R}$, $|\delta (\omega,t,z)|\leq\Gamma(z)$ and $\int_{\mathbb{R}}\Gamma ( z ) ^{r}\lambda ( dz ) <\infty$. \end{ass} \renewcommand{\theass}{S\Alph{ass}} \setcounter{ass}{2} \begin{ass}\label{assSC}We have Assumption~\ref{assC}. The processes $ \tilde{b}_{t}$, $\tilde{\sigma}_{t}$ and $\tilde \sigma}_{t}^{\prime}$ are bounded, and for some bounded nonnegative function $\Gamma_{\sigma}$ on $\mathbb{R} , $|\tilde{\delta}(\omega,t,z)|\leq\Gamma_{\sigma}(z)$ and $\int_{\mathbb{R}}\Gamma_{\sigma} ( z ) ^{\tilde{r}}\lambda ( dz ) <\infty$. \end{ass} \renewcommand{\theass}{S\Alph{ass}} \begin{ass}\label{assSD} We have Assumption~\ref{assD}. Moreover, $\alpha _{n,t}$ and $\alpha_{n,t}^{-1}$ are uniformly bounded for all $n,t$. \end{ass} \subsection{\texorpdfstring{Proofs in Section \protect\ref{sec-c}}{Proofs in Section 3}} \mbox{} \begin{pf*}{Proof of Lemma~\ref{lem-1}} (a) We set $\widehat {V _{t}^{+}=\widehat{V}_{iu_{n}}$ for $t\in\lbrack ( i-1 ) u_{n},iu_{n})$. Denote the left-hand side of (\ref{thm-1a}) by $S_{n}$ and T_{n}= \lfloor T/u_{n} \rfloor u_{n}$. We have \[ S_{n}=\int_{0}^{ ( \lfloor T/u_{n} \rfloor-1 ) u_{n}}g\bigl \widehat{V}_{s}^{+}\bigr)\,ds+\int_{0}^{u_{n}}g( \widehat{V}_{s})\,ds \int_{T_{n}}^{T}g( \widehat{V}_{s})\,ds. \] Since $g$ is bounded \begin{equation}\quad \mathbb{E}\biggl\llvert S_{n}-\int_{0}^{T}g ( V_{s} ) \,ds\biggr\rrvert \leq Ku_{n}+\int _{0}^{ ( \lfloor T/u_{n} \rfloor -1 ) u_{n}}\mathbb{E}\bigl\llvert g\bigl( \widehat{V}_{s}^{+}\bigr)-g(V_{s})\bigr\rrvert \,ds. \label{thm-1-101} \end{equation} Observe that for each $s\in\lbrack0, ( \lfloor T/u_{n} \rfloor-1 ) u_{n})$, $\widehat {V}_{s}^{+}\stackrel {\mathbb P}}{\longrightarrow}V_{s}$. To see this, we recall from Assumption~\ref{assSD} that $\alpha_{n,t}\in[\underline{\alpha},\overline{\alpha}]$ for some constant $\overline{\alpha}\geq\underline{\alpha}>0$. Let $\widehat{V}_{s}^{+}(\overline{\alpha})$ and $\widehat{V}_{s}^{+}(\underline{\alpha})$ be defined as $\widehat {V}_{s}^{+}$ except with $\alpha_{n,t}$ replaced, respectively, by $\overline{\alpha}$ and $\underline{\alpha}$. By Theorem~9.3.2 in~\cite{jacodprotter2012}, the right continuity\vspace*{1pt} of $V$ and $u_{n}\rightarrow0$, $\widehat {V}_{s}^{+}(\overline{\alpha})$ and $\widehat{V}_{s}^{+}(\underline{\alpha})$ converge in probability to $V_s$. The claim then follows $\widehat{V}_{s}^{+}(\underline{\alpha})\leq\widehat{V}_{s}^{+} \leq \widehat{V}_{s}^{+}(\overline{\alpha})$. Hence, by condition (ii) and bounded convergence, for Lebesgue a.e. $s\in [ 0,T ] $, \mathbb{E}|g(\widehat{V}_{s}^{+})-g(V_{s})|=\mathbb{E}|(g(\widehat {V _{s}^{+})-g(V_{s}))1_{ \{ V_{s}\notin D_{g} \} }|\rightarrow0$. Applying bounded convergence on (\ref{thm-1-101}), we readily obtain (\re {thm-1a}). Part (b) can be shown similarly. \end{pf*} \begin{pf*}{Proof of Theorem~\ref{thm-1}} (a) For each $x\geq 0$, F_{T} ( x ) =F_{T} ( x- ) $ a.s. by the continuity of F_{T} ( \cdot ) $. Hence, \[ \int_{0}^{T}\mathbb{P} ( V_{s}=x ) \,ds = \mathbb{E} \bigl[ F_{T} ( x ) -F_{T} ( x- ) \bigr] = 0. \] Therefore, $\mathbb{P} ( V_{s}=x ) =0$ for Lebesgue a.e. $s\in [ 0,T ] $. By Lemma~\ref{lem-1} with $g ( \cdot ) =1_{ \{ \cdot \leq x \} }$, $\widehat{F}_{n,T} ( x ) \stackrel{\mathbb{P}}{\longrightarrow}F_{T} ( x ) $. Since \widehat{F}_{n,T} ( \cdot ) $ and $F_{T} ( \cdot ) $ are increasing, and $F_{T} ( \cdot ) $ is continuous, this convergence also holds locally uniformly. Since $V$ is c\`{a}dl\`{a}g, $\bar {V}\equiv \sup_{t\in [ 0,T ] }V_{t}=O_{p} ( 1 ) $. For any $\eta>0 , there exists some $M>0$, such that $\mathbb{P} ( \bar {V}>M ) <\eta$, yielding $\mathbb{P} ( T\not=F_{T} ( M ) ) <\eta $. Hence, for any $\varepsilon>0$, \begin{eqnarray*} &&\limsup_{n\rightarrow\infty}\mathbb{P} \Bigl( \sup_{x\in\mathbb {R } \bigl\llvert \widehat{F}_{n,T} ( x ) -F_{T} ( x ) \bigr \rrvert >\varepsilon \Bigr) \\ &&\qquad\leq\limsup_{n\rightarrow\infty}\mathbb{P} \Bigl( \sup _{0\leq x\leq M}\bigl\llvert \widehat{F}_{n,T} ( x ) -F_{T} ( x ) \bigr\rrvert >\varepsilon \Bigr) \\ &&\qquad\quad{}+\limsup_{n\rightarrow\infty}\mathbb{P} \Bigl( \sup_{x\geq M} \bigl\llvert \widehat{F}_{n,T} ( x ) -F_{T} ( x ) \bigr \rrvert >\varepsilon \Bigr) \\ &&\qquad\leq\limsup_{n\rightarrow\infty}\mathbb{P} \bigl( \bigl\llvert \widehat{F _{n,T} ( M ) -F_{T} ( M ) \bigr\rrvert >\varepsilon /2 \bigr) +\mathbb{P} \bigl( T-F_{T} ( M ) >\varepsilon /2 \bigr) \\ &&\qquad<\eta. \end{eqnarray*} Sending $\eta\rightarrow0$, we readily derive the assertion in part (a). Part (b) can be proved similarly. \end{pf*} \begin{pf*}{Proof of Corollary~\ref{cor-Q}} By Theorem~\ref {thm-1}, \widehat{F}_{n,T} ( \cdot ) \stackrel{\mathbb {P}}{\longrightarrow F_{T} ( \cdot ) $ uniformly. By a subsequence argument, we can then assume $\widehat{F}_{n,T} ( \cdot ) \stackrel{\mathrm{a.s.}} \longrightarrow}F_{T} ( \cdot ) $ uniformly without loss. The assertion in part (a) then follows Lemma~21.2 of \cite {vandervaart1998}. The proof of part~(b) is similar. \end{pf*} \subsection{\texorpdfstring{Proofs in Section \protect\ref{sec-rc}}{Proofs in Section 4.1}} \mbox{} \begin{pf*}{Proof of Lemma~\ref{lem-fu}} (a) Observ \begin{eqnarray}\label{lem-fu-101} F_{T} ( x-\eta_{n} ) &=&\int_{0}^{T}1_{ \{ V_{s}\leq x-\eta_{n} \} }\,ds \nonumber \\[-8pt] \\[-8pt] \nonumber &\leq&\widehat{F}_{n,T} ( x ) \leq \int_{0}^{T}1_{ \{ V_{s}\leq x+\eta_{n} \} }\,ds=F_{T} ( x+\eta_{n} ) \end{eqnarray} Since $F_{T} ( x-\eta_{n} ) \leq F_{T} ( x ) \leq F_{T} ( x+\eta_{n} ) $, the first assertion in part (a) readily follows. Now consider the quantiles. By definition, $\widehat{F}_{n,T}( \widehat{Q}_{n,T}(\alpha))\geq\alpha$. By~(\ref{lem-fu-101}), $F_{T}( \widehat{Q}_{n,T}(\alpha)+\eta_{n})\geq\alpha$. Therefore, $Q_{T}(\alpha )\leq\widehat{Q}_{n,T}(\alpha)+\eta_{n}$. For any $\varepsilon>0$, by \ref{lem-fu-101}), we have $F_{T}(\widehat{Q}_{n,T}(\alpha)-\eta _{n}-\varepsilon)\leq\widehat{F}_{n,T}(\widehat{Q}_{n,T}(\alpha )-\varepsilon)<\alpha$. Hence, $\widehat{Q}_{n,T}(\alpha)-\eta _{n}-\varepsilon<Q_{T}(\alpha)$. Since $\varepsilon>0$ is arbitrary, \widehat{Q}_{n,T}(\alpha)-\eta_{n}\leq Q_{T}(\alpha)$. The second assertion of part (a) is then obvious. (b) Fix some $\varepsilon>0$. There exists $M>0$ such that $\mathbb{P} ( \eta_{n}\geq Ma_{n} ) <\varepsilon/2$ for $n$ sufficiently large. Since $a_{n}\rightarrow0$, $ [ x-\eta_{n},x+\eta _{n} ] $ is contained in $\mathcal{N}_{x}$ with probability approaching one (w.p.a.1) and by part (a), \[ \bigl\llvert \widehat{F}_{n,T} ( x ) -F_{T} ( x ) \bigr \rrvert \leq\int_{x-\eta_{n}}^{x+\eta_{n}}f_{T} ( z ) \,dz. \] Let $M^{\prime}=4M\sup_{z\in\mathcal{N}_{x}}\mathbb{E} [ f_{T} ( z ) ] /\varepsilon$. We have for $n$ sufficiently large, \begin{eqnarray*} &&\mathbb{P} \biggl( \int_{x-\eta_{n}}^{x+\eta_{n}}f_{T} ( z ) \,dz>M^{\prime}a_{n} \biggr) \\ &&\qquad\leq\mathbb{P} \biggl( \int_{x-Ma_{n}}^{x+Ma_{n}}f_{T} ( z ) \,dz>M^{\prime}a_{n} \biggr) +\mathbb{P} ( \eta_{n}\geq Ma_{n} ) \\ &&\qquad<\frac{2M\sup_{z\in\mathcal{N}_{x}}\mathbb{E} [ f_{T} ( z ) ] }{M^{\prime}}+ \varepsilon/2 \\ &&\qquad\leq\varepsilon \end{eqnarray*} Hence, $\int_{x-\eta_{n}}^{x+\eta_{n}}f_{T} ( z ) \,dz=O_{p} ( a_{n} ) $. The assertion in part (b) then readily follows. \end{pf*} We now prove Theorem~\ref{thm-u}, starting with two lemmas. Below, $\Vert \cdot \Vert_p$ denotes the $L_p$ norm. \begin{lemma} \label{lem-vstar}Let $p\geq1$ be a constant and $k_{n}\asymp\Delta _{n}^{-\gamma}$ for some $\gamma\in ( 0,1 ) $. Suppose Assumption~\ref{assSA} holds with $X$ continuous and Assumption~\ref{assSD}. Then for each $0\leq i\leq \lfloor T/u_{n} \rfloor-1$ \begin{equation} \bigl\llvert \widehat{V}_{iu_{n}}^{\ast}-V_{iu_{n}}\bigr \rrvert \vee\llvert \widehat{V}_{iu_{n}}-V_{iu_{n}} \rrvert \leq \xi_{n,i}+ \sup_{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\llvert V_{s}-V_{iu_{n}}\rrvert, \label{lem-vstar-a} \end{equation} where the variable $\xi_{n,i}$ satisfies $\Vert\xi_{n,i} \Vert_p \leq K_p k_n^{-1/2}$. If we further have Assumption~\ref{assSC} with $\Delta V_s =0$ for $s\in[0,T]$, then the majorant side of the above can be bounded by $K_{p}(k_{n}^{-1/2}+u_{n}^{1/2})$ in $L_p$. \end{lemma} \begin{pf} By It\^o's formula, $\widehat {V}_{iu_{n}}^{\ast}-V_{iu_{n}} =\zeta_{n,i}^{\prime}+\zeta _{n,i}^{\prime\prime}$, where \begin{eqnarray*} \zeta_{n,i}^{\prime} &=&\frac{2}{u_{n}}\int _{iu_{n}}^{ ( i+1 ) u_{n}} ( X_{s}-X_{n,s} ) \,dX_{s}, \\ \zeta_{n,i}^{\prime\prime} &=&\frac{1}{u_{n}}\int _{iu_{n}}^{ ( i+1 ) u_{n}} ( V_{s}-V_{iu_{n}} ) \,ds, \end{eqnarray*} and $X_{n,s}$ is the discretized process given by $X_{n,s}=X_{iu_{n}+ ( j-1 ) \Delta_{n}}$ when $s\in \lbrack iu_{n}+ ( j-1 ) \Delta_{n},iu_{n}+j\Delta_{n})$. By classical estimates (note that $X$ is continuous), \begin{eqnarray*} \mathbb{E}\biggl\llvert \frac{2}{u_{n}}\int_{iu_{n}}^{ ( i+1 ) u_{n}} ( X_{s}-X_{n,s} ) b_{s}\,ds\biggr\rrvert ^{p} &\leq &K_{p}\Delta _{n}^{p/2}, \\ \mathbb{E}\biggl\llvert \frac{2}{u_{n}}\int_{iu_{n}}^{ ( i+1 ) u_{n}} ( X_{s}-X_{n,s} ) \sigma_{s}\,dW_{s} \biggr\rrvert ^{p} &\leq &K_{p}k_{n}^{-p/2}. \end{eqnarray*} Since $k_{n}=o ( \Delta_{n}^{-1} ) $, we have $\Vert\zeta _{n,i}^{\prime}\Vert_{p}\leq K_{p}k_{n}^{-1/2}$ by Minkowski's inequality. We also observe $\vert\zeta_{n,i}^{\prime\prime}\vert \leq\sup_{s\in\lbrack iu_{n},( i+1) u_{n})}\vert V_{s}-V_{iu_{n}}\vert$. Now note that $\widehat{V}_{iu_{n}}^{\ast}-\widehat {V}_{iu_{n}}=k_{n}^{-1 \sum_{j=1}^{k_{n}}(\Delta_{ik_{n}+j}^{n}X/\Delta _{n}^{1/2})^{2}1_{\{|\Delta_{ik_{n}+j}^{n}X|>v_{n,iu_n}\}}$. Under Assumption~\ref{assSD}, $\alpha_{n,t}\geq\underline{\alpha}$ for some constant $\underline {\alpha}>0$. Hence, $v_{n,t}\geq\underline{v}_n\equiv\underline{\alpha} \Delta _{n}^{\varpi}$. Since $X$ is continuous, for any $q\geq0$, \begin{eqnarray*} &&\bigl\llVert \bigl(\Delta_{ik_{n}+j}^{n}X/\Delta_{n}^{1/2} \bigr)^{2}1_{ \{ \llvert \Delta_{ik_{n}+j}^{n}X\rrvert >v_{n,iu_n} \} }\bigr\rrVert _{p}\\ &&\qquad\leq \biggl( \frac{\mathbb{E}\llvert \Delta _{ik_{n}+j}^{n}X/\Delta _{n}^{1/2}\rrvert ^{2p+q}}{ ( \underline{v}_{n}/\Delta _{n}^{1/2} ) ^{q} \biggr) ^{1/p} \\ &&\qquad\leq K_{p,q}\Delta_{n}^{q ( 1/2-\varpi ) /p}. \end{eqnarray*} Since $\varpi\in ( 0,1/2 ) $, when $q$ is taken sufficiently large, terms in the above display can be further bounded by K_{p}k_{n}^{-1/2}$. Hence, $\Vert\widehat{V}_{iu_{n}}^{\ast }-\widehat {V _{iu_{n}}\Vert_{p}\leq K_{p}k_{n}^{-1/2}$. The first assertion then readily follows by setting $\xi_{n,i} = |\zeta^{\prime}_{n,i}| + \vert \widehat {V}_{iu_{n}}^{\ast}-\widehat{V _{iu_{n}} \vert$. Now, suppose Assumption~\ref{assSC} together with $V$ being continuous. By standard estimates, for each $p\geq1$, the second term on the right-hand side of (\ref {lem-vstar-a ) can be bounded by $K_{p}u_{n}^{1/2}$ in $L_p$. The second assertion of the lemma is then obvious. \end{pf} Under Assumption~\ref{assSA}, we set \begin{eqnarray*} X_{t}^{\prime} &=&X_{t}-X_{t}^{\prime\prime}, \\ X_{t}^{\prime\prime} &=&\cases{ \displaystyle \int _{0}^{t}\int_{\mathbb{R}}\delta ( s,z ) \mu ( ds,dz ), & \quad$\mbox{if } r\leq1$, \vspace*{2pt}\cr \displaystyle\int_{0}^{t}\int_{\mathbb{R}} \delta ( s,z ) ( \mu -\nu ) ( ds,dz ), &\quad $\mbox{if } r>1.$} \end{eqnarray*} We define $\widehat{V}^{\prime}$ as $\widehat{V}^{\ast}$ in (\ref{eqvh}) but with $X^{\prime}$ in place of $X$; in particular, $\widehat{V _{iu_{n}}^{\prime}\equiv u_{n}^{-1}\sum_{j=1}^{k_{n}}(\Delta _{ik_{n}+j}^{n}X^{\prime})^{2}$. \begin{lemma} \label{lem-eoj}Suppose that Assumption~\ref{assSA} holds with some $r\in(0,2)$ and Assumption~\ref{assSD}. Let p\geq r\vee1$ and $\varpi\in ( \frac{p-1}{2p-r},\frac {1}{2} ) $. Let $\theta\in ( 0,\infty ) $ be arbitrarily fixed if $r>1$ and \theta=0$ if $r\leq1$. We have for each $i$, \[ \bigl\llVert \widehat{V}_{iu_{n}}-\widehat{V}_{iu_{n}}^{\prime} \bigr\rrVert _{p}\leq K_{p}\Delta_{n}^{{(1-r\varpi-\theta)}/{p}- ( 1-2\varpi ) }. \] \end{lemma} \begin{pf} Under Assumption~\ref{assSD}, $\alpha_{n,t}\in \lbrack \underline{\alpha},\overline{\alpha}]$ for constants $\overline {\alpha \geq\underline{\alpha}>0$. We set $\bar{v}_{n}=\overline{\alpha }\Delta _{n}^{\varpi}$ and $\underline{v}_{n}=\underline{\alpha}\Delta _{n}^{\varpi}$. By applying Lemma~13.2.6 in~\cite{jacodprotter2012} [with s=1$, $s^{\prime}=2$, $m=p$, $p^{\prime}=1$, $k=1$, $F ( x ) =x^{2}$], we have \begin{eqnarray*} &&\mathbb{E} \bigl\vert \bigl( \Delta _{ik_{n}+j}^{n}X/\Delta _{n}^{1/2} \bigr) ^{2}1_{ \{ \llvert \Delta _{ik_{n}+j}^{n}X\rrvert \leq\bar{v}_{n} \} } - \bigl( \Delta _{ik_{n}+j}^{n}X^{\prime}/ \Delta_{n}^{1/2} \bigr) ^{2}1_{ \{ \llvert \Delta_{ik_{n}+j}^{n}X^{\prime}\rrvert \leq\bar{v _{n} \} } \bigr\vert^{p} \\ &&\qquad\leq K_{p}\Delta_{n}^{{(2-r)}/{2}-\theta }+K_{p} \Delta_{n}^{1-r\varpi -p ( 1-2\varpi ) -\theta} \\ &&\qquad\leq K_{p}\Delta_{n}^{1-r\varpi-p ( 1-2\varpi ) -\theta}. \end{eqnarray*} By a similar argument as in the proof of Lemma~\ref{lem-vstar}, \[ \bigl\llVert \bigl(\Delta_{ik_{n}+j}^{n}X^{\prime}/ \Delta_{n}^{1/2}\bigr)^{2}1_{\{|\Delta _{ik_{n}+j}^{n}X^{\prime}|>\bar{v}_{n}\}}\bigr \rrVert _{p}\leq K_{p,q}\Delta _{n}^{q ( 1/2-\varpi ) /p} \] for any $q\geq0$. Taking $q$ sufficiently large, we then deriv \begin{eqnarray \label{lem-eoj-101} &&\bigl\llVert \bigl( \Delta_{ik_{n}+j}^{n}X/\Delta_{n}^{1/2} \bigr) ^{2}1_{ \{ \llvert \Delta_{ik_{n}+j}^{n}X\rrvert \leq \bar {v _{n} \} }- \bigl( \Delta_{ik_{n}+j}^{n}X^{\prime}/ \Delta _{n}^{1/2} \bigr) ^{2}\bigr\rrVert _{p} \nonumber \\[-8pt] \\[-8pt] \nonumber &&\qquad\leq K_{p} \Delta_{n}^{{(1-r\varpi)}/{p}- ( 1-2\varpi ) {\theta}/{p}}. \end{eqnarray} By a similar argument, we can derive (\ref{lem-eoj-101}) when $\bar{v}_{n}$ is replaced with $\underline{v}_{n}$. Since $\underline{v}_{n}\leq v_{n,iu_n}\leq \bar{v}_{n}$, (\ref{lem-eoj-101}) also holds when $\bar{v}_{n}$ is replaced with $v_{n,iu_n}$. The assertion of the lemma then follows from Minkowski's inequality.\vadjust{\goodbreak} \end{pf} \begin{pf*}{Proof of Theorem~\ref{thm-u}} \textit{Step} 1. We first suppose $X$ is continuous. Observe that \[ \eta_{n}\leq\sup_{0\leq i\leq \lfloor T/u_{n} \rfloor -1}\llvert \widehat{V}_{iu_{n}}-V_{iu_{n}}\rrvert +2\sup_{0\leq i\leq \lfloor T/u_{n} \rfloor} \sup_{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\llvert V_{s}-V_{iu_{n}}\rrvert. \] Since $V$ is continuous, $\Vert\sup_{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\llvert V_{s}-V_{iu_{n}}\rrvert \Vert_{p}\leq K_{p}u_{n}^{1/2} $ for any $p\geq1$ by standard estimates. By Lemma~\re {lem-vstar} and the maximal inequality (e.g., Lemma~2.2.2 in \cit {vandervaartwellner}), for any $p\geq1$, $\llVert \eta_{n}\rrVert _{p}\leq K_{p}u_{n}^{-1/p}(k_{n}^{-1/2}+u_{n}^{1/2})$. Since $k_{n}\asymp \Delta_{n}^{-\gamma}$ by assumption, we derive $\Vert\eta_{n}\Vert _{p}\leq K\Delta_{n}^{ ( \gamma\wedge ( 1-\gamma ) ) /2-\iota} $ by taking $p$ sufficiently large. The same argument yields \Vert\eta_{n}^{\ast}\Vert_{p}\leq K\Delta_{n}^{ ( \gamma \wedge ( 1-\gamma ) ) /2-\iota}$. This finishes the proof of part (a) with $X$ continuous, as well as part (b). \textit{Step} 2. We now consider part (a) allowing $X$ to be discontinuous. Let $\widehat{V}^{\prime}$ be defined as in Lemma~\ref{lem-eoj}. Observe that \begin{eqnarray*} \eta_{n} &\leq&\sup_{0\leq i\leq \lfloor T/u_{n} \rfloor -1}\bigl\llvert \widehat{V}_{iu_{n}}-\widehat{V}_{iu_{n}}^{\prime }\bigr\rrvert +\eta_{n}^{\prime},\qquad \mbox{where} \\ \eta_{n}^{\prime} &=&\sup_{0\leq i\leq \lfloor T/u_{n} \rfloor -1}\bigl\llvert \widehat{V}_{iu_{n}}^{\prime}-V_{iu_{n}}\bigr\rrvert +2\sup _{0\leq i\leq \lfloor T/u_{n} \rfloor}\sup_{s\in \lbrack iu_{n}, ( i+1 ) u_{n})}\llvert V_{s}-V_{iu_{n}} \rrvert. \end{eqnarray*} A similar argument as in part (a) yields $\eta_{n}^{\prime}= O_p(\Delta_{n}^{ ( \gamma\wedge ( 1-\gamma ) ) /2-\iota})$. By the maximal inequality and Lemma~\ref{lem-eoj} for $p=1\vee r $, \begin{eqnarray*} \Bigl\llVert \sup_{0\leq i\leq \lfloor T/u_{n} \rfloor -1}\bigl\llvert \widehat{V}_{iu_{n}}- \widehat{V}_{iu_{n}}^{\prime}\bigr\rrvert \Bigr\rrVert _{p} &\leq&Ku_{n}^{-1/p}\Delta_{n}^{{(1-r\varpi-\theta) }/{p}- ( 1-2\varpi ) } \\ &\leq&K\Delta_{n}^{{(\gamma-r\varpi-\theta)}/{p}- ( 1-2\varpi ) } \\ &\leq&\cases{ K\Delta_{n}^{\gamma-r\varpi- ( 1-2\varpi ) }, & \quad$\mbox {if } r\leq1$, \vspace*{2pt}\cr K\Delta_{n}^{ ( \gamma-\theta ) /r- ( 1-\varpi ) }, & \quad$\mbox{if } r>1.$} \end{eqnarray*} Taking $\theta$ sufficiently small in the $r>1$ case, we readily derive the assertion in part~(a). \end{pf*} \begin{pf*}{Proof of Theorem~\ref{thmevt}} \textit{Step} 1. We first prove the assertion on $\eta_n^*$, so condition (iii) is in force. In the constant volatility setting of the theorem, we have $\sqrt{k_n}\eta_n^* = \sqrt{2}V\times M_n$, where we denote \begin{eqnarray} M_n &=& \sup_{i=0,\ldots,\lfloor{T}/{u_n}\rfloor-1}\bigl\llvert Z_i^n \bigr\rrvert, \nonumber \\[-8pt] \\[-8pt] \nonumber Z_i^n &=& \frac{\sqrt{k_n}}{\sqrt{2}V} \bigl( \widehat{V}^*_{iu_n} -V \bigr),\qquad i=0,\ldots,\lfloor T/u_n \rfloor-1. \end{eqnarray} Under our constant volatility assumption $\{Z_i^n\}_i$ are independent and identically distributed with distribution which is approximately standard normal. Therefore, we can use Edgeworth expansion of the c.d.f. together with extreme value theory to pin down the limit distribution of $M_n$. To this end, we set \begin{equation} \cases{ \displaystyle c_{n} = \bigl(2 \log(b_n)\bigr)^{-1/2},\qquad m_{n} = \sqrt{2 \log(b_n)} - \frac {\log (\log(b_n))+\log(4\pi)}{2\sqrt{2\log(b_n)}}, \vspace*{2pt}\cr \tau_{n}(x) = c_{n}x+m_{n}, \qquad b_n = \lfloor T/u_n\rfloor,\qquad x\in\mathbb{R}_+. }\hspace*{-35pt} \end{equation} Note that $\tau_{n}(x)\asymp\sqrt{2\log(b_n)}$ and hence increases to infinity as the number of blocks increases to infinity for every fixed $x$. Using second-order Edgeworth expansion and denoting with $\Phi(\cdot)$, the c.d.f. of standard normal random variable (see Theorem~2.2 and Lemma~5.4 of~\cite{Hall}), we have \begin{equation} \mathbb{P} \bigl(\bigl|Z_i^n\bigr|\leq\tau_n(x) \bigr) = \Phi\bigl(\tau _n(x)\bigr)-\Phi \bigl(-\tau_n(x) \bigr) + \frac{(\log(b_n))^4}{b_n\sqrt{k_n}}K(x)+o \biggl(\frac {1}{k_n} \biggr)\hspace*{-35pt} \end{equation} for any $x$ where $K(x)$ is a polynomial of $x$. Then we have \begin{equation} \label{eqproofevtt} \lim_{n\rightarrow\infty} \biggl[\frac{\mathbb{P} (|Z_i^n|\leq\tau _n(x) )}{\Phi(\tau_n(x))-\Phi(-\tau_n(x))} \biggr]^{b_n} = 1, \end{equation} provided $k_n \asymp\Delta_n^{-1/2}$. This assumption on the rate of growth of $k_n$ guarantees that the distribution of $Z_i^n$ is ``sufficiently close'' to standard normal. Now we can use~(\ref{eqproofevtt}) to get \begin{eqnarray} \mathbb{P}\bigl(c_{n}^{-1}(M_n-m_{n}) \leq x\bigr) &=& \bigl[\mathbb{P} \bigl(\bigl|Z_i^n\bigr|\leq \tau_n(x) \bigr)\bigr]^{b_n} \nonumber \\[-8pt] \\[-8pt] \nonumber &\sim&\bigl[\Phi\bigl( \tau_n(x)\bigr)-\Phi\bigl(-\tau_n(x)\bigr) \bigr]^{b_n} \end{eqnarray} as $n\rightarrow\infty$. From here, using the results for the maximum domain of attraction of the Gumbel distribution (see e.g., Example~1.1.7 of~\cite{Haan}), we have \begin{eqnarray} \bigl[\Phi\bigl(\tau_n(x)\bigr)-\Phi\bigl(-\tau_n(x) \bigr)\bigr]^{b_n} &=& \bigl[2\Phi\bigl(\tau _n(x)\bigr)-1 \bigr]^{b_n} \nonumber \\[-8pt] \\[-8pt] \nonumber & \longrightarrow &\exp\bigl(-2\exp(-x)\bigr) \qquad \forall x, \end{eqnarray} and hence \begin{equation} c_{n}^{-1}(M_n-m_{n}) \stackrel{ \mathcal{L}} {\longrightarrow } \Lambda \end{equation} for $\Lambda$ being a random variable with c.d.f. $\exp(-2\exp(-x))$. From here the result in (\ref{thmevt1}) (with $\eta_n$ replaced by $\eta _n^*$) follows. \textit{Step} 2. We now prove (\ref{thmevt1}) with condition (iii) relaxed. Let $\eta_n^{\prime\ast}$ be defined as $\eta_n^*$ but with $\widehat {V}_t^{\ast}$ replaced by $\widehat{V}_t^{\prime}$. By step 1, (\ref {thmevt1}) holds with $\eta_n$ replaced by $\eta_n^{\prime\ast}$. It remains to show that \begin{equation} \log\bigl(\lfloor T/u_n\rfloor\bigr)^{1/2} \Delta_n^{-1/4} \bigl(\eta_n - \eta _n^{\prime\ast}\bigr) = o_p(1).\label{eta301} \end{equation} Note that $|\eta_n-\eta_n^{\prime\ast}| \leq\sup_{0\leq i \leq \lfloor T/u_n \rfloor-1} |\widehat{V}_{iu_n} - \widehat{V}^{\prime }_{iu_n}| = O_p(\Delta_n^{(2-r)\varpi-1/2})$,\vspace*{2pt} where the stochastic order is shown in step 2 of the proof of Theorem~\ref{thm-u}. (\ref {eta301}) then follows condition (ii). This completes the proof. \end{pf*} \subsection{\texorpdfstring{Proofs in Section \protect\ref{sec-rd}}{Proofs in Section 4.2}} Under Assumption~\ref{assC}, by It\^{o}'s formula, we can represent $V$ as \begin{eqnarray*} V_{t} &=&V_{0}+\int_{0}^{t} \tilde{b}_{V,s}\,ds+\int_{0}^{t}\tilde { \sigma _{V,s}\,dW_{s}+\int_{0}^{t} \tilde{\sigma}_{V,s}^{\prime }\,dW_{s}^{\prime} \\ &&{}+\int_{0}^{t}\int_{\mathbb{R}} \tilde{\delta}_{V} ( s,z ) ( \mu-\nu ) ( ds,dz ), \end{eqnarray*} where, by localization, we can assume without loss that the coefficients \tilde{b}_{V}$, $\tilde{\sigma}_{V}$, $\tilde{\sigma}_{V}^{\prime }$ are bounded, and $|\tilde{\delta}_{V} ( \omega,s,z ) |\leq \tilde{ \Gamma} ( z ) $ for any $ ( \omega,s,z ) $, where $\tilde \Gamma} ( \cdot ) $ is bounded and deterministic, and satisfies \int_{\mathbb{R}}\tilde{\Gamma} ( z ) ^{\tilde {r}}\lambda ( dz ) <\infty$. We consider the following decomposition: for $q>0$, \begin{eqnarray*} V_{t}&=&V_{t}^{\prime} ( q ) +V_{t}^{\prime\prime} ( q ),\qquad\mbox{where } \\ V_{t}^{c}&=&V_{0}+ \int_{0}^{t}\tilde{b}_{V,s}\,ds+\int _{0}^{t}\tilde {\sigma} _{V,s}\,dW_{s}+\int_{0}^{t} \tilde{\sigma}_{V,s}^{\prime }\,dW_{s}^{\prime }, \\ V_{t}^{\prime} ( q ) &=&V_{t}^{c}+\int_{0}^{t} \int_{ \{ z\dvtx \tilde \Gamma} ( z ) \leq q \} }\tilde{\delta}_{V} ( s,z ) ( \mu-\nu ) ( ds,dz ) \\ &&{}-\int _{0}^{t}\int_{ \{ z\dvtx \tilde \Gamma} ( z ) >q \} }\tilde{ \delta}_{V} ( s,z ) \nu ( ds,dz ), \\ V_{t}^{\prime\prime} ( q ) &=&\int_{0}^{t}\int_{ \{ z\dvtx \tilde \Gamma} ( z ) >q \} } \tilde{\delta}_{V} ( s,z ) \mu ( ds,dz ) \end{eqnarray*} Denote $I ( n,i ) =[iu_{n}, ( i+1 ) u_{n})$. We also set \mathcal{I}_{n} ( q ) =\{0 \leq i\leq \lfloor T/u_{n} \rfloor-1\dvtx \mu(I ( n,i ) \times\{z\dvtx \tilde {\Gamma ( z ) >q\})=0\}$ and $\mathcal{T}_{n} ( q ) =\bigcup_{i\in\mathcal{I}_{n} ( q ) }I ( n,i ) $. Here, \mathcal{I}_{n} ( q ) $ collects indices of intervals not containing ``big'' jumps. We can decompose \widehat{F}_{n,T} ( x ) =\widehat{F}_{n,T} ( x;q ) \widehat{R}_{n,T} ( x;q )$ where \[ \widehat{F}_{n,T} ( x;q ) =u_{n}\sum _{i\in\mathcal {I}_{n} ( q ) }1_{ \{ \widehat{V}_{iu_{n}}\leq x \} },\qquad \widehat{R}_{n,T} ( x;q ) = \int_{ [ 0,T ] \setminus \mathcal{T}_{n} ( q ) }1_{ \{ \widehat{V}_{s}\leq x \} }\,ds \] Analogously, we have $F_{T}(x)=F_{n,T}(x;q)+R_{n,T}(x;q)$, where \[ F_{n,T} ( x;q ) =\int_{\mathcal{T}_{n} ( q ) }1_{ \{ V_{s}\leq x \} }\,ds,\qquad R_{n,T} ( x;q ) =\int_{ [ 0,T ] \setminus\mathcal{T}_{n} ( q ) }1_{ \{ V_{s}\leq x \} }\,ds. \] Finally, we set \begin{eqnarray*} \hat{\eta}_{n} ( q ) &=&\sup_{i\in\mathcal{I}_{n} ( q ) }\llvert \widehat{V}_{iu_{n}}-V_{iu_{n}}\rrvert, \\ \eta_{n}^{\prime} ( q ) &=&\sup_{0\leq i\leq \lfloor T/u_{n} \rfloor}\sup _{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\bigl\llvert V_{s}^{\prime} ( q ) -V_{iu_{n}}^{\prime } ( q ) \bigr\rrvert, \\ \eta_{n} ( q ) &=&\hat{\eta}_{n} ( q ) +\eta _{n}^{\prime} ( q ). \end{eqnarray*} We now generalize Lemma~\ref{lem-fu} as follows.\vadjust{\goodbreak} \begin{lemma} \label{lem-rd}Suppose $\eta_{n} ( q_{n} ) =O_{p} ( w_{n} ) $ for some nonrandom sequences $q_{n}\rightarrow0$ and w_{n}\rightarrow0$ and Assumption~\ref{assSA} with $r=2$. Then \textup{(a)} under Assumption~\ref{assB}, $\widehat{F}_{n,T} ( x ) -F_{T} ( x ) =O_{p} ( w_{n} ) +O_{p} ( u_{n}q_{n}^{-\tilde{r}} ) $; \textup{(b)} under Assumption~\ref{assBprime}, the assertion in \textup{(a)} holds uniformly in $x\in\mathbb{R}$ and moreover \textup{(c)} $|\widehat {Q}_{n,T}(\alpha) -Q_{T}(\alpha)|\leq\xi_{n}\sup_{x\in\mathbb{R }|\widehat{F}_{n,T}(x)-F_{T}(x)|$ for some tight sequence of variables $\xi _{n}$, provided $f_{T}(Q_{T}(\alpha))>0$ a.s. \end{lemma} \begin{pf} (a) Observe that $\mathbb{E}[\int_{ [ 0, ] \setminus\mathcal{T}_{n} ( q_{n} ) }\,ds]\leq Ku_{n}q_{n}^{ \tilde{r}}$, yieldin \begin{equation} \qquad\sup_{x\in\mathbb{R}}R_{n,T} ( x;q_{n} ) =O_{p} \bigl( u_{n}q_{n}^{ \tilde{r}} \bigr), \qquad\sup_{x\in\mathbb{R}}\widehat{R _{n,T} ( x;q_{n} ) =O_{p} \bigl( u_{n}q_{n}^{-\tilde {r}} \bigr). \label{lem-rd1-101} \end{equation} By definition, $\widehat{F}_{n,T} ( x;q_{n} ) =\int_{\mathcal {T _{n} ( q_{n} ) }1_{ \{ \widehat{V}_{s}\leq x \} }\,ds$. Note that over $\mathcal{T}_{n} ( q_{n} ) $, the process $V_{t}^{\prime \prime} ( q_{n} ) $ is identically zero. Hence, $\sup_{t\in \mathcal{T}_{n} ( q_{n} ) }|\widehat{V}_{t}-V_{t}|\leq \eta _{n} ( q_{n} ) $. By a similar argument as in (\ref{lem-fu-101}), we deduc \begin{eqnarray}\label{lem-rd1-201} &&\bigl\llvert \widehat{F}_{n,T} ( x;q_{n} ) -F_{n,T} ( x;q_{n} ) \bigr\rrvert \nonumber\\ &&\qquad\leq F_{n,T} \bigl( x+\eta_{n} ( q_{n} );q_{n} \bigr) -F_{n,T} \bigl( x-\eta_{n} ( q_{n} );q_{n} \bigr) \\ &&\qquad\leq F_{T} \bigl( x+\eta_{n} ( q_{n} ) \bigr) -F_{T} \bigl( x-\eta_{n} ( q_{n} ) \bigr).\nonumber \end{eqnarray} By an argument similar to part (b) of Lemma~\ref{lem-fu}, we derive F_{T} ( x+\eta_{n} ( q_{n} ) ) -F_{T} ( x-\eta _{n} ( q_{n} ) ) =O_{p} ( w_{n} ) $. The assertion of part (a) then follows (\ref{lem-rd1-101}) and~(\ref{lem-rd1-201}). (b) By localization, we can suppose that $V$ is bounded and thus f_{T} ( \cdot ) $ is compactly supported. Since $f_{T} ( \cdot ) $ is continuous, $\sup_{x\in\mathbb{R}}f_{T} ( x ) =O_{p} ( 1 ) $. By~(\ref{lem-rd1-201}) \begin{equation} \sup_{x\in\mathbb{R}}\bigl|\widehat{F}_{n,T} ( x;q_{n} ) -F_{n,T} ( x;q_{n} ) \bigr|\leq2\eta_{n} ( q_{n} ) \sup_{z\in \mathbb{R} }f_{T} ( z ). \label{lem-rd-203} \end{equation} The assertion then readily follows (\ref{lem-rd1-101}) and (\ref {lem-rd-203 ). (c) Observe that $\widehat{F}_{n,T}(\widehat{Q}_{n,T}(\alpha))\geq \alpha =F_{T}(Q_{T}(\alpha))$, where the inequality follows the definition of quantiles and the equality is due to the continuity of $F_{T}(\cdot)$. Hence \begin{eqnarray}\label{lem-rd-304} F_{T}\bigl(Q_{T}(\alpha)\bigr)-F_{T}\bigl( \widehat{Q}_{n,T}(\alpha)\bigr) & \leq& \widehat{F}_{n,T}\bigl( \widehat{Q}_{n,T}(\alpha)\bigr)-F_{T}\bigl(\widehat {Q}_{n,T}(\alpha )\bigr) \nonumber \\[-8pt] \\[-8pt] \nonumber & \leq& \sup_{x\in\mathbb{R}} \bigl\llvert \widehat {F}_{n,T}(x)-F_{T}(x)\bigr\rrvert \end{eqnarray} For any $\varepsilon>0$, $\widehat{F}_{n,T}(\widehat{Q}_{n,T}(\alpha )-\varepsilon)<\alpha=F_{T}(Q_{T}(\alpha))$, yielding \begin{eqnarray*} F_{T}\bigl(\widehat{Q}_{n,T}(\alpha)-\varepsilon \bigr)-F_{T}\bigl(Q_{T}(\alpha)\bigr) &<&F_{T} \bigl(\widehat{Q}_{n,T}(\alpha)-\varepsilon\bigr)-\widehat {F}_{n,T}\bigl(\widehat Q}_{n,T}(\alpha)- \varepsilon\bigr) \\ &\leq&\sup_{x\in\mathbb{R}}\bigl\vert \widehat {F}_{n,T}(x)-F_{T}(x) \bigr\vert. \end{eqnarray*} Since $F_{T} ( \cdot ) $ is continuous, by sending $\varepsilon \downarrow0$ we deduc \begin{equation} F_{T}\bigl(\widehat{Q}_{n,T}(\alpha)\bigr)-F_{T} \bigl(Q_{T}(\alpha)\bigr)\leq\sup_{x\in \mathbb{R}}\bigl\llvert \widehat{F}_{n,T}(x)-F_{T}(x)\bigr\rrvert. \label{lem-rd-305} \end{equation} Let $\mathcal{K}_{n,T} ( \alpha ) $ be the closed interval with endpoints $Q_{T}(\alpha)$ and $\widehat{Q}_{n,T}(\alpha)$. Set $\xi _{n}\equiv\sup_{x\in\mathcal{K}_{n,T} ( \alpha ) }f_{T}^{-1} ( x ) $. By a mean-value expansion \begin{equation} \qquad\bigl\llvert F_{T}\bigl(\widehat{Q}_{n,T}(\alpha) \bigr)-F_{T}\bigl(Q_{T}(\alpha )\bigr)\bigr\rrvert \geq\inf _{x\in\mathcal{K}_{n,T} ( \alpha ) }f_{T} ( x ) \bigl\llvert \widehat{Q}_{n,T}( \alpha )-Q_{T}(\alpha )\bigr\rrvert. \label{lem-rd-306} \end{equation} Since $f_{T}(Q_{T}(\alpha))>0$ a.s., $\widehat{Q}_{n,T}(\alpha )\stackrel \mathbb{P}}{\longrightarrow}Q_{T}(\alpha)$ by Corollary~\ref {cor-Q}. Since $f_{T} ( \cdot ) $ is continuous, $\inf_{x\in\mathcal{K _{n,T} ( \alpha ) }f_{T} ( x ) \stackrel {\mathbb{P}} \longrightarrow}f_{T}(Q_{T}(\alpha))>0$; hence $\xi_{n}$ is tight. The assertion then follows (\ref{lem-rd-304}), (\ref{lem-rd-305}) and (\re {lem-rd-306}). \end{pf} \begin{pf*}{Proof of Theorem~\ref{thm-rd}} \textit{Step} 1. We first consider \eta_{n}^{\prime} ( q_{n} ) $. For each $i$, \begin{eqnarray*} &&\sup_{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\bigl\llvert V_{s}^{\prime } ( q_{n} ) -V_{iu_{n}}^{\prime} ( q_{n} ) \bigr \rrvert \leq\zeta_{n,i}+\zeta_{n,i}^{\prime}+ \zeta_{n,i}^{\prime\prime },\qquad\mbox{where} \\ &&\qquad\zeta_{n,i}=\sup _{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\biggl\llvert \int_{iu_{n}}^{s} \int_{ \{ z\dvtx \tilde{\Gamma} ( z ) >q_{n} \} }\tilde{\delta}_{V} ( s,z ) \nu ( ds,dz ) \biggr\rrvert, \\ &&\qquad\zeta_{n,i}^{\prime}= \sup_{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\biggl\llvert \int_{iu_{n}}^{s} \int_{ \{ z\dvtx \tilde{\Gamma } ( z ) \leq q_{n} \} }\tilde{\delta}_{V} ( s,z ) ( \mu -\nu ) ( ds,dz ) \biggr\rrvert, \\ &&\qquad\zeta_{n,i}^{\prime\prime}= \sup_{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\bigl\llvert V_{s}^{c}-V_{iu_{n}}^{c} \bigr\rrvert. \end{eqnarray*} For $\zeta_{n,i}$, observe tha \begin{eqnarray}\label{thm-rd-101}\quad \Bigl\llvert \sup_{0\leq i\leq \lfloor T/u_{n} \rfloor }\zeta _{n,i}\Bigr\rrvert & \leq& \sup_{0\leq i\leq \lfloor T/u_{n} \rfloor} \int_{iu_{n}}^{ ( i+1 ) u_{n}}\int_{ \{ z\dvtx \tilde{\Gamma} ( z ) >q_{n} \} }\tilde{ \Gamma} ( z ) \nu ( ds,dz ) \nonumber \\[-8pt] \\[-8pt] \nonumber \quad & \leq& Ku_{n}q_{n}^{ ( 1-\tilde{r} ) \wedge0}. \end{eqnarray} Now turn to $\zeta_{n,i}^{\prime}$. Let $p\geq2$. For each $i\geq 0$, by Lemma~2.1.5 in~\cite{jacodprotter2012}, \begin{eqnarray*} \mathbb{E}\bigl\llvert \zeta_{n,i}^{\prime}\bigr\rrvert ^{p} &\leq &K_{p}u_{n}\int_{ \{ z\dvtx \tilde{\Gamma} ( z ) \leq q_{n} \} \tilde{\Gamma} ( z ) ^{p}\lambda ( dz ) \\ &&{}+K_{p}u_{n}^{p/2} \biggl( \frac{1}{u_{n}}\int _{I ( n,i ) }\,ds\int_{ \{ z\dvtx \tilde{\Gamma} ( z ) \leq q_{n} \} }\tilde \Gamma} ( z ) ^{2}\lambda ( dz ) \biggr) ^{p/2} \\ &\leq&K_{p}u_{n}q_{n}^{p-\tilde{r}}+K_{p}u_{n}^{p/2}q_{n}^{ ( 2-\tilde{ } ) p/2}. \end{eqnarray*} Hence, $\Vert\zeta_{n,i}^{\prime}\Vert_{p}\leq K_{p}u_{n}^{1/p}q_{n}^{1-\tilde{r}/p}+K_{p}u_{n}^{1/2}q_{n}^{1-\tilde{r}/2}$. By the maximal inequality (Lemma~2.2.2 in~\cite{vandervaartwellner}) \begin{equation} \Bigl\llVert \sup_{0\leq i\leq \lfloor T/u_{n} \rfloor }\zeta _{n,i}^{\prime} \Bigr\rrVert _{p}\leq K_{p}q_{n}^{1-\tilde{r /p}+K_{p}u_{n}^{1/2-1/p}q_{n}^{1-\tilde{r}/2}. \label{thm-rd-102} \end{equation} Since $V^{c}$ is continuous, by a standard estimate for $\zeta _{n,i}^{\prime\prime}$ and the maximal inequality \begin{equation} \Bigl\llVert \sup_{0\leq i\leq \lfloor T/u_{n} \rfloor }\zeta _{n,i}^{\prime\prime} \Bigr\rrVert _{p}\leq K_{p}u_{n}^{1/2-1/p}. \label{thm-rd-103} \end{equation} Combining (\ref{thm-rd-101}), (\ref{thm-rd-102}) and (\ref {thm-rd-103}), we derive for $p\geq2$, $\Vert\eta_{n}^{\prime}( q_{n}) \Vert_{p} \leq K_{p}a_{n,p}^{\prime}$, where \[ a_{n,p}^{\prime} \equiv u_{n}q_{n}^{ ( 1-\tilde{r} ) \wedge0} \vee q_{n}^{1-\tilde{r}/p}\vee u_{n}^{1/2-1/p}. \] \textit{Step} 2. Observe tha \begin{equation} \hat{\eta}_{n} ( q_{n} ) \leq\sup_{i\in\mathcal {I}_{n} ( q_{n} ) } \bigl\llvert \widehat{V}_{iu_{n}}-\widehat {V}_{iu_{n}}^{\prime } \bigr\rrvert +\sup_{i\in\mathcal{I}_{n} ( q_{n} ) }\bigl\llvert \widehat{V}_{iu_{n}}^{\prime}-V_{iu_{n}} \bigr\rrvert. \label{thm-rd-201} \end{equation} By Lemma~\ref{lem-vstar}, for $i\in\mathcal{I}_{n} ( q_{n} ) $, for some $\xi_{n,i}$ with $\Vert\xi_{n,i} \Vert_p \leq K_p k_n^{-1/2}$, \[ \bigl\llvert \widehat{V}_{iu_{n}}^{\prime}-V_{iu_{n}}\bigr \rrvert \leq \xi_{n,i}+\sup_{s\in\lbrack iu_{n}, ( i+1 ) u_{n})}\bigl\llvert V_{s}^{\prime} ( q_{n} ) -V_{iu_{n}}^{\prime } ( q_{n} ) \bigr\rrvert. \] Therefore, by a similar argument as in step 1 \begin{equation} \Bigl\llVert \sup_{i\in\mathcal{I}_{n} ( q_{n} ) }\bigl\llvert \widehat{V}_{iu_{n}}^{\prime}-V_{iu_{n}} \bigr\rrvert \Bigr\rrVert _{p}\leq K_{p}u_{n}^{-1/p}k_{n}^{-1/2}+K_{p}a_{n,p}^{\prime}. \label{thm-rd-202} \end{equation} Similarly as in step 2 of the proof of Theorem~\ref{thm-u} [recall that $\theta =0$ if $r\leq1$ and $\theta\in ( 0,\infty ) $ can be arbitrarily fixed when $r>1$ \begin{equation} \sup_{i\in\mathcal{I}_{n} ( q_{n} ) }\bigl\llvert \widehat {V _{iu_{n}}-\widehat{V}_{iu_{n}}^{\prime}\bigr\rrvert =O_{p} \bigl( \Delta _{n}^{{(\gamma-r\varpi-\theta)}/{(1\vee r)}- ( 1-2\varpi ) } \bigr). \label{thm-rd-203} \end{equation} Combining (\ref{thm-rd-201})--(\ref{thm-rd-203}), we derive $\hat {\eta _{n} ( q_{n} ) =O_{p} ( w_{n,p} ) $ for p\geq2$, where \[ w_{n,p} \equiv\Delta_{n}^{{(\gamma-r\varpi-\theta)}/{(1\vee r) - ( 1-2\varpi ) }\vee u_{n}^{-1/p}k_{n}^{-1/2}\vee a_{n,p}^{\prime }. \] By step 1, we further derive $\eta_{n} ( q_{n} ) =O_{p} ( w_{n,p} ) $. By Lemma~\ref{lem-rd}(a), we have \begin{eqnarray*} &&\widehat{F}_{n,T} ( x ) -F_{T} ( x ) \\ &&\qquad=O_{p} ( w_{n,p} ) +O_{p} \bigl( u_{n}q_{n}^{ \tilde{r}} \bigr) \\ &&\qquad=O_{p} \bigl( \Delta_{n}^{{(\gamma-r\varpi-\theta)}/{(1\vee r)} - ( 1-2\varpi ) }\vee u_{n}^{-1/p}k_{n}^{-1/2} \vee q_{n}^{1-\tilde r}/p}\vee u_{n}^{1/2-1/p}\\ &&\hspace*{280pt}{}\vee u_{n}q_{n}^{-\tilde{r}} \bigr). \end{eqnarray*} Taking $q_{n}=u_{n}^{1/ ( 1+\tilde{r}-\tilde{r}/p ) }$ and recalling $k_{n}\asymp\Delta_{n}^{-\gamma}$, we have \begin{eqnarray*} &&\widehat{F}_{n,T} ( x ) -F_{T} ( x ) \\ &&\qquad=O_{p} \bigl( \Delta_{n}^{{(\gamma-r\varpi-\theta)}/{(1\vee r)} - ( 1-2\varpi ) }\vee\Delta_{n}^{{\gamma}/{2}- {(1-\gamma)}/{p}}\vee \Delta_{n}^{ ( 1-\gamma ) ( {1}/{2}-{1}/{p})}\\ &&\hspace*{193pt}{}\vee\Delta _{n}^{{((1-\gamma)(1-\tilde{r}/p))}/{(1+\tilde{r}-\tilde {r}/p)}} \bigr) \end{eqnarray*} and by taking $p$ sufficiently large, \begin{eqnarray*} &&\widehat{F}_{n,T} ( x ) -F_{T} ( x ) \\ &&\qquad=O_{p} \bigl( \Delta_{n}^{{(\gamma-r\varpi-\theta)}/{(1\vee r)} - ( 1-2\varpi ) }\vee\Delta_{n}^{\gamma/2-\iota}\vee \Delta _{n}^{ ( 1-\gamma ) /2-\iota}\vee\Delta_{n}^{ ( 1-\gamma ) / ( 1+\tilde{r} ) -\iota} \bigr). \end{eqnarray*} The discontinuous case in part (a) then readily follows the definitions of~$\theta$, $a_{n}$ [see (\ref{an})] and $d_{n}$. The continuous case in part (a), as well as part (b), can be proved in a similar (but simpler) way. \end{pf*} \begin{pf*}{Proof of Theorem~\ref{thm-rdu}} We first show that \sup_{x\in\mathbb{R}}|\widehat{F}_{n,T}(x)-F_{T}(x)|=O_{p}(d_{n})$. The proof is similar as that of Theorem~\ref{thm-rd}, except in step 2 of the proof, we use Lemma~\ref{lem-rd}(b) instead of Lemma~\ref{lem-rd}(a). The result for $\widehat{F}_{n,T}^{\ast}$ can be proved similarly. The assertion concerning $\widehat{Q}_{n,T}(\alpha)$ then follows from Lemma~\ref{lem-rd (c). The assertion on $\widehat{Q}^{*}_{n,T}(\alpha)$ can be proved in a similar (but simpler) way; the details are omitted for brevity. \end{pf*} \subsection{\texorpdfstring{Proofs in Section \protect\ref{sec-ker}}{Proofs in Section 5}} \mbox{} \begin{pf*}{Proof of Theorem~\ref{thm-kd}} (a) By localization and condition (iv), we can assume that $V_{t}$ takes value in some compact $ \mathcal{K}\subset ( 0,\infty ) $, and thus $f_{T} ( \cdot ) $ is supported on~$\mathcal{K}$. We set $f_{n,T}(x) \int_{0}^{T}h_{n}^{-1}\kappa( h_{n}^{-1}(V_{s}-x)) \,ds$. For each $x\in\mathbb{R}$, \begin{eqnarray*} \mathbb{E}\bigl\vert \widehat{f}_{n,T} ( x ) -f_{n,T} ( x ) \bigr\vert &\leq&\mathbb{E} \biggl[ h_{n}^{-1}\int _{0}^{T \biggl\vert\kappa \biggl(\frac{\widehat{V}_{s}-x}{h_{n} \biggr)-\kappa \biggl(\frac{V_{s}-x}{h_{n} \biggr)\biggr\vert \,ds \biggr] \\ &\leq&Kh_{n}^{-2}\mathbb{E} \biggl[ \int _{0}^{T}\vert \widehat {V _{s}-V_{s}\rrvert ds \biggr]. \end{eqnarray*} By Lemmas~\ref{lem-vstar} and~\ref{lem-eoj}, $\mathbb{E}|\widehat{V} _{s}-V_{s}|\leq K\bar{a}_{n}$. Hence, \begin{equation} \mathbb{E}\bigl\llvert \widehat{f}_{n,T} ( x ) -f_{n,T} ( x ) \bigr\rrvert \leq Kh_{n}^{-2}\bar{a}_{n},\qquad \mathbb {E} \bigl[ \Vert\widehat{f}_{n,T}-f_{n,T} \Vert_{w} \bigr] \leq Kh_{n}^{-2}\bar{a}_{n}, \label{ker-1}\hspace*{-35pt} \end{equation} where $K$ does not depend on $x$. Now observe that $f_{n,T}(x)=\int_{\mathbb R}}h_{n}^{-1}\kappa( h_{n}^{-1}(y-x)) f_{T}(y)\,dy$. By a change of variable, $f_{n,T}(x)=\int_{\mathbb{R}}\kappa ( z ) f_{T}(x+h_{n}z)\,dz$. Hence, \begin{eqnarray*} \mathbb{E}\bigl\llvert f_{n,T}(x)-f_{T} ( x ) \bigr\rrvert &\leq &\int_{\mathbb{R}}\kappa ( z ) \mathbb{E}\bigl\llvert f_{T} ( x+h_{n}z ) -f_{T} ( x ) \bigr\rrvert \,dz \\ &\leq&Kh_{n}^{\beta}\int_{\mathbb{R}}\kappa ( z ) \llvert z\rrvert ^{\beta}\,dz\leq Kh_{n}^{\beta}, \end{eqnarray*} which further implies $\mathbb{E}[\Vert f_{n,T}-f_{T}\Vert_{w}]\leq Kh_{n}^{\beta}$. Combining these estimates with~(\ref{ker-1}) completes the proof of part (a). Part (b) can be proved in a similar way. \end{pf*} \section*{Acknowledgments} We would like to thank Tim Bollerslev, Nathalie Eisenbaum, Jean Jacod, Andrew Patton and Philip Protter for helpful discussions as well as an Associate Editor and two referees for very helpful suggestions. We are particularly grateful to Markus Reiss for suggesting the direct estimation approach adopted in the paper and the link with the uniform error in estimating the volatility path given in Lemma~\ref{lem-fu} of the paper.